2
0
mirror of https://gitlab.isc.org/isc-projects/bind9 synced 2025-08-23 02:28:55 +00:00
bind/lib/isc/netmgr/tcpdns.c

1491 lines
36 KiB
C
Raw Normal View History

/*
* Copyright (C) Internet Systems Consortium, Inc. ("ISC")
*
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, you can obtain one at https://mozilla.org/MPL/2.0/.
*
* See the COPYRIGHT file distributed with this work for additional
* information regarding copyright ownership.
*/
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
#include <libgen.h>
#include <unistd.h>
#include <uv.h>
#include <isc/atomic.h>
#include <isc/barrier.h>
#include <isc/buffer.h>
#include <isc/condition.h>
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
#include <isc/errno.h>
#include <isc/log.h>
#include <isc/magic.h>
#include <isc/mem.h>
#include <isc/netmgr.h>
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
#include <isc/quota.h>
#include <isc/random.h>
#include <isc/refcount.h>
#include <isc/region.h>
#include <isc/result.h>
#include <isc/sockaddr.h>
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
#include <isc/stdtime.h>
#include <isc/thread.h>
#include <isc/util.h>
#include "netmgr-int.h"
#include "uv-compat.h"
/*%<
*
* Maximum number of simultaneous handles in flight supported for a single
* connected TCPDNS socket. This value was chosen arbitrarily, and may be
* changed in the future.
*/
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
static atomic_uint_fast32_t last_tcpdnsquota_log = ATOMIC_VAR_INIT(0);
static bool
can_log_tcpdns_quota(void) {
isc_stdtime_t now, last;
isc_stdtime_get(&now);
last = atomic_exchange_relaxed(&last_tcpdnsquota_log, now);
if (now != last) {
return (true);
}
return (false);
}
static isc_result_t
tcpdns_connect_direct(isc_nmsocket_t *sock, isc__nm_uvreq_t *req);
2020-03-02 12:10:26 +01:00
static void
tcpdns_close_direct(isc_nmsocket_t *sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
static void
tcpdns_connect_cb(uv_connect_t *uvreq, int status);
static void
tcpdns_connection_cb(uv_stream_t *server, int status);
static void
tcpdns_close_cb(uv_handle_t *uvhandle);
static isc_result_t
accept_connection(isc_nmsocket_t *ssock, isc_quota_t *quota);
static void
quota_accept_cb(isc_quota_t *quota, void *sock0);
static void
stop_tcpdns_parent(isc_nmsocket_t *sock);
static void
stop_tcpdns_child(isc_nmsocket_t *sock);
static isc_result_t
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
tcpdns_connect_direct(isc_nmsocket_t *sock, isc__nm_uvreq_t *req) {
isc__networker_t *worker = NULL;
isc_result_t result = ISC_R_UNSET;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
int r;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(VALID_UVREQ(req));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(isc__nm_in_netthread());
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
worker = &sock->mgr->workers[sock->tid];
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->connecting, true);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_tcp_init(&worker->loop, &sock->uv_handle.tcp);
RUNTIME_CHECK(r == 0);
uv_handle_set_data(&sock->uv_handle.handle, sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_timer_init(&worker->loop, &sock->timer);
RUNTIME_CHECK(r == 0);
if (isc__nm_closing(sock)) {
result = ISC_R_CANCELED;
goto error;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_tcp_open(&sock->uv_handle.tcp, sock->fd);
if (r != 0) {
isc__nm_closesocket(sock->fd);
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_OPENFAIL]);
goto done;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_OPEN]);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (req->local.length != 0) {
r = uv_tcp_bind(&sock->uv_handle.tcp, &req->local.type.sa, 0);
/*
* In case of shared socket UV_EINVAL will be returned and needs
* to be ignored
*/
if (r != 0 && r != UV_EINVAL) {
isc__nm_incstats(sock->mgr,
sock->statsindex[STATID_BINDFAIL]);
goto done;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
uv_handle_set_data(&req->uv_req.handle, req);
r = uv_tcp_connect(&req->uv_req.connect, &sock->uv_handle.tcp,
&req->peer.type.sa, tcpdns_connect_cb);
if (r != 0) {
isc__nm_incstats(sock->mgr,
sock->statsindex[STATID_CONNECTFAIL]);
goto done;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_CONNECT]);
uv_handle_set_data((uv_handle_t *)&sock->timer, &req->uv_req.connect);
isc__nmsocket_timer_start(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->connected, true);
done:
result = isc__nm_uverr2result(r);
error:
LOCK(&sock->lock);
sock->result = result;
SIGNAL(&sock->cond);
if (!atomic_load(&sock->active)) {
WAIT(&sock->scond, &sock->lock);
}
INSIST(atomic_load(&sock->active));
UNLOCK(&sock->lock);
return (result);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
void
isc__nm_async_tcpdnsconnect(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnsconnect_t *ievent =
(isc__netievent_tcpdnsconnect_t *)ev0;
isc_nmsocket_t *sock = ievent->sock;
isc__nm_uvreq_t *req = ievent->req;
isc_result_t result = ISC_R_SUCCESS;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
UNUSED(worker);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->type == isc_nm_tcpdnssocket);
REQUIRE(sock->iface != NULL);
REQUIRE(sock->parent == NULL);
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = tcpdns_connect_direct(sock, req);
if (result != ISC_R_SUCCESS) {
isc__nmsocket_clearcb(sock);
isc__nm_connectcb(sock, req, result, true);
atomic_store(&sock->active, false);
isc__nm_tcpdns_close(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* The sock is now attached to the handle.
*/
isc__nmsocket_detach(&sock);
}
static void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
tcpdns_connect_cb(uv_connect_t *uvreq, int status) {
isc_result_t result;
isc__nm_uvreq_t *req = NULL;
isc_nmsocket_t *sock = uv_handle_get_data((uv_handle_t *)uvreq->handle);
struct sockaddr_storage ss;
int r;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
isc__nmsocket_timer_stop(sock);
uv_handle_set_data((uv_handle_t *)&sock->timer, sock);
if (!atomic_load(&sock->connecting)) {
return;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
req = uv_handle_get_data((uv_handle_t *)uvreq);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_UVREQ(req));
REQUIRE(VALID_NMHANDLE(req->handle));
if (isc__nmsocket_closing(sock)) {
/* Socket was closed midflight by isc__nm_tcpdns_shutdown() */
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = ISC_R_CANCELED;
goto error;
} else if (status == UV_ETIMEDOUT) {
/* Timeout status code here indicates hard error */
result = ISC_R_CANCELED;
goto error;
} else if (status != 0) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = isc__nm_uverr2result(status);
goto error;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_CONNECT]);
r = uv_tcp_getpeername(&sock->uv_handle.tcp, (struct sockaddr *)&ss,
&(int){ sizeof(ss) });
if (r != 0) {
result = isc__nm_uverr2result(r);
goto error;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->connecting, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = isc_sockaddr_fromsockaddr(&sock->peer, (struct sockaddr *)&ss);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
isc__nm_connectcb(sock, req, ISC_R_SUCCESS, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
error:
isc__nm_failed_connect_cb(sock, req, result, false);
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nm_tcpdnsconnect(isc_nm_t *mgr, isc_nmiface_t *local, isc_nmiface_t *peer,
isc_nm_cb_t cb, void *cbarg, unsigned int timeout,
size_t extrahandlesize) {
isc_result_t result = ISC_R_SUCCESS;
isc_nmsocket_t *sock = NULL;
isc__netievent_tcpdnsconnect_t *ievent = NULL;
isc__nm_uvreq_t *req = NULL;
sa_family_t sa_family;
REQUIRE(VALID_NM(mgr));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(local != NULL);
REQUIRE(peer != NULL);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sa_family = peer->addr.type.sa.sa_family;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock = isc_mem_get(mgr->mctx, sizeof(*sock));
isc__nmsocket_init(sock, mgr, isc_nm_tcpdnssocket, local);
sock->extrahandlesize = extrahandlesize;
sock->connect_timeout = timeout;
sock->result = ISC_R_UNSET;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_init(&sock->client, true);
req = isc__nm_uvreq_get(mgr, sock);
req->cb.connect = cb;
req->cbarg = cbarg;
req->peer = peer->addr;
req->local = local->addr;
req->handle = isc__nmhandle_get(sock, &req->peer, &sock->iface->addr);
result = isc__nm_socket(sa_family, SOCK_STREAM, 0, &sock->fd);
if (result != ISC_R_SUCCESS) {
if (isc__nm_in_netthread()) {
sock->tid = isc_nm_tid();
}
isc__nmsocket_clearcb(sock);
isc__nm_connectcb(sock, req, result, true);
atomic_store(&sock->closed, true);
isc__nmsocket_detach(&sock);
return;
}
/* 2 minute timeout */
result = isc__nm_socket_connectiontimeout(sock->fd, 120 * 1000);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
ievent = isc__nm_get_netievent_tcpdnsconnect(mgr, sock, req);
if (isc__nm_in_netthread()) {
atomic_store(&sock->active, true);
sock->tid = isc_nm_tid();
isc__nm_async_tcpdnsconnect(&mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
isc__nm_put_netievent_tcpdnsconnect(mgr, ievent);
} else {
atomic_init(&sock->active, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock->tid = isc_random_uniform(mgr->nworkers);
isc__nm_enqueue_ievent(&mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
LOCK(&sock->lock);
while (sock->result == ISC_R_UNSET) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
WAIT(&sock->cond, &sock->lock);
}
atomic_store(&sock->active, true);
BROADCAST(&sock->scond);
UNLOCK(&sock->lock);
}
static uv_os_sock_t
isc__nm_tcpdns_lb_socket(sa_family_t sa_family) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_result_t result;
uv_os_sock_t sock;
result = isc__nm_socket(sa_family, SOCK_STREAM, 0, &sock);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
(void)isc__nm_socket_incoming_cpu(sock);
/* FIXME: set mss */
result = isc__nm_socket_reuse(sock);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
#if HAVE_SO_REUSEPORT_LB
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = isc__nm_socket_reuse_lb(sock);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
#endif
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return (sock);
}
static void
start_tcpdns_child(isc_nm_t *mgr, isc_nmiface_t *iface, isc_nmsocket_t *sock,
uv_os_sock_t fd, int tid) {
isc__netievent_tcpdnslisten_t *ievent = NULL;
isc_nmsocket_t *csock = &sock->children[tid];
isc__nmsocket_init(csock, mgr, isc_nm_tcpdnssocket, iface);
csock->parent = sock;
csock->accept_cb = sock->accept_cb;
csock->accept_cbarg = sock->accept_cbarg;
csock->recv_cb = sock->recv_cb;
csock->recv_cbarg = sock->recv_cbarg;
csock->extrahandlesize = sock->extrahandlesize;
csock->backlog = sock->backlog;
csock->tid = tid;
/*
* We don't attach to quota, just assign - to avoid
* increasing quota unnecessarily.
*/
csock->pquota = sock->pquota;
isc_quota_cb_init(&csock->quotacb, quota_accept_cb, csock);
#if HAVE_SO_REUSEPORT_LB || defined(WIN32)
UNUSED(fd);
csock->fd = isc__nm_tcpdns_lb_socket(iface->addr.type.sa.sa_family);
#else
csock->fd = dup(fd);
#endif
REQUIRE(csock->fd >= 0);
ievent = isc__nm_get_netievent_tcpdnslisten(mgr, csock);
isc__nm_maybe_enqueue_ievent(&mgr->workers[tid],
(isc__netievent_t *)ievent);
}
isc_result_t
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nm_listentcpdns(isc_nm_t *mgr, isc_nmiface_t *iface,
isc_nm_recv_cb_t recv_cb, void *recv_cbarg,
isc_nm_accept_cb_t accept_cb, void *accept_cbarg,
size_t extrahandlesize, int backlog, isc_quota_t *quota,
isc_nmsocket_t **sockp) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_result_t result = ISC_R_SUCCESS;
isc_nmsocket_t *sock = NULL;
size_t children_size = 0;
uv_os_sock_t fd = -1;
REQUIRE(VALID_NM(mgr));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock = isc_mem_get(mgr->mctx, sizeof(*sock));
isc__nmsocket_init(sock, mgr, isc_nm_tcpdnslistener, iface);
atomic_init(&sock->rchildren, 0);
#if defined(WIN32)
sock->nchildren = 1;
#else
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock->nchildren = mgr->nworkers;
#endif
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
children_size = sock->nchildren * sizeof(sock->children[0]);
sock->children = isc_mem_get(mgr->mctx, children_size);
memset(sock->children, 0, children_size);
sock->result = ISC_R_UNSET;
sock->accept_cb = accept_cb;
sock->accept_cbarg = accept_cbarg;
sock->recv_cb = recv_cb;
sock->recv_cbarg = recv_cbarg;
sock->extrahandlesize = extrahandlesize;
sock->backlog = backlog;
sock->pquota = quota;
if (isc__nm_in_netthread()) {
sock->tid = isc_nm_tid();
} else {
sock->tid = isc_random_uniform(sock->nchildren);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock->fd = -1;
#if !HAVE_SO_REUSEPORT_LB && !defined(WIN32)
fd = isc__nm_tcpdns_lb_socket(iface->addr.type.sa.sa_family);
#endif
isc_barrier_init(&sock->startlistening, sock->nchildren);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
for (size_t i = 0; i < sock->nchildren; i++) {
if ((int)i == isc_nm_tid()) {
continue;
}
start_tcpdns_child(mgr, iface, sock, fd, i);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (isc__nm_in_netthread()) {
start_tcpdns_child(mgr, iface, sock, fd, isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
#if !HAVE_SO_REUSEPORT_LB && !defined(WIN32)
isc__nm_closesocket(fd);
#endif
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
LOCK(&sock->lock);
while (atomic_load(&sock->rchildren) != sock->nchildren) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
WAIT(&sock->cond, &sock->lock);
}
result = sock->result;
atomic_store(&sock->active, true);
UNLOCK(&sock->lock);
INSIST(result != ISC_R_UNSET);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (result == ISC_R_SUCCESS) {
REQUIRE(atomic_load(&sock->rchildren) == sock->nchildren);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
*sockp = sock;
} else {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->active, false);
isc_nm_stoplistening(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmsocket_close(&sock);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return (result);
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_async_tcpdnslisten(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnslisten_t *ievent =
(isc__netievent_tcpdnslisten_t *)ev0;
isc_nmiface_t *iface = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sa_family_t sa_family;
int r;
int flags = 0;
isc_nmsocket_t *sock = NULL;
isc_result_t result = ISC_R_UNSET;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(ievent->sock));
REQUIRE(ievent->sock->tid == isc_nm_tid());
REQUIRE(VALID_NMSOCK(ievent->sock->parent));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock = ievent->sock;
iface = sock->iface;
sa_family = iface->addr.type.sa.sa_family;
REQUIRE(sock->type == isc_nm_tcpdnssocket);
REQUIRE(sock->iface != NULL);
REQUIRE(sock->parent != NULL);
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/* TODO: set min mss */
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_tcp_init(&worker->loop, &sock->uv_handle.tcp);
RUNTIME_CHECK(r == 0);
uv_handle_set_data(&sock->uv_handle.handle, sock);
/* This keeps the socket alive after everything else is gone */
isc__nmsocket_attach(sock, &(isc_nmsocket_t *){ NULL });
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_timer_init(&worker->loop, &sock->timer);
RUNTIME_CHECK(r == 0);
uv_handle_set_data((uv_handle_t *)&sock->timer, sock);
LOCK(&sock->parent->lock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = uv_tcp_open(&sock->uv_handle.tcp, sock->fd);
if (r < 0) {
isc__nm_closesocket(sock->fd);
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_OPENFAIL]);
goto done;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_OPEN]);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sa_family == AF_INET6) {
flags = UV_TCP_IPV6ONLY;
}
#if HAVE_SO_REUSEPORT_LB || defined(WIN32)
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
r = isc_uv_tcp_freebind(&sock->uv_handle.tcp,
&sock->iface->addr.type.sa, flags);
if (r < 0) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_BINDFAIL]);
goto done;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
#else
if (sock->parent->fd == -1) {
r = isc_uv_tcp_freebind(&sock->uv_handle.tcp,
&sock->iface->addr.type.sa, flags);
if (r < 0) {
isc__nm_incstats(sock->mgr,
sock->statsindex[STATID_BINDFAIL]);
goto done;
}
sock->parent->uv_handle.tcp.flags = sock->uv_handle.tcp.flags;
sock->parent->fd = sock->fd;
} else {
/* The socket is already bound, just copy the flags */
sock->uv_handle.tcp.flags = sock->parent->uv_handle.tcp.flags;
}
#endif
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* The callback will run in the same thread uv_listen() was called
* from, so a race with tcpdns_connection_cb() isn't possible.
*/
r = uv_listen((uv_stream_t *)&sock->uv_handle.tcp, sock->backlog,
tcpdns_connection_cb);
if (r != 0) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_log_write(isc_lctx, ISC_LOGCATEGORY_GENERAL,
ISC_LOGMODULE_NETMGR, ISC_LOG_ERROR,
"uv_listen failed: %s",
isc_result_totext(isc__nm_uverr2result(r)));
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_BINDFAIL]);
goto done;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->listening, true);
done:
result = isc__nm_uverr2result(r);
if (result != ISC_R_SUCCESS) {
sock->pquota = NULL;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_fetch_add(&sock->parent->rchildren, 1);
if (sock->parent->result == ISC_R_UNSET) {
sock->parent->result = result;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
SIGNAL(&sock->parent->cond);
UNLOCK(&sock->parent->lock);
isc_barrier_wait(&sock->parent->startlistening);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
static void
tcpdns_connection_cb(uv_stream_t *server, int status) {
isc_nmsocket_t *ssock = uv_handle_get_data((uv_handle_t *)server);
isc_result_t result;
isc_quota_t *quota = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (status != 0) {
result = isc__nm_uverr2result(status);
goto done;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(ssock));
REQUIRE(ssock->tid == isc_nm_tid());
if (isc__nmsocket_closing(ssock)) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = ISC_R_CANCELED;
goto done;
}
if (ssock->pquota != NULL) {
result = isc_quota_attach_cb(ssock->pquota, &quota,
&ssock->quotacb);
if (result == ISC_R_QUOTA) {
isc__nm_incstats(ssock->mgr,
ssock->statsindex[STATID_ACCEPTFAIL]);
return;
}
}
result = accept_connection(ssock, quota);
done:
if (result != ISC_R_SUCCESS && result != ISC_R_NOCONN) {
if ((result != ISC_R_QUOTA && result != ISC_R_SOFTQUOTA) ||
can_log_tcpdns_quota())
{
isc_log_write(isc_lctx, ISC_LOGCATEGORY_GENERAL,
ISC_LOGMODULE_NETMGR, ISC_LOG_ERROR,
"TCP connection failed: %s",
isc_result_totext(result));
}
}
}
static void
enqueue_stoplistening(isc_nmsocket_t *sock) {
isc__netievent_tcpdnsstop_t *ievent =
isc__nm_get_netievent_tcpdnsstop(sock->mgr, sock);
isc__nm_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
}
void
isc__nm_tcpdns_stoplistening(isc_nmsocket_t *sock) {
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->type == isc_nm_tcpdnslistener);
if (!atomic_compare_exchange_strong(&sock->closing, &(bool){ false },
true)) {
INSIST(0);
ISC_UNREACHABLE();
}
if (!isc__nm_in_netthread()) {
enqueue_stoplistening(sock);
} else if (!isc__nm_acquire_interlocked(sock->mgr)) {
enqueue_stoplistening(sock);
} else {
stop_tcpdns_parent(sock);
isc__nm_drop_interlocked(sock->mgr);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
void
isc__nm_async_tcpdnsstop(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnsstop_t *ievent =
(isc__netievent_tcpdnsstop_t *)ev0;
isc_nmsocket_t *sock = ievent->sock;
UNUSED(worker);
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
if (sock->parent != NULL) {
stop_tcpdns_child(sock);
return;
}
/*
* If network manager is paused, re-enqueue the event for later.
*/
if (!isc__nm_acquire_interlocked(sock->mgr)) {
enqueue_stoplistening(sock);
} else {
stop_tcpdns_parent(sock);
isc__nm_drop_interlocked(sock->mgr);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
void
isc__nm_tcpdns_failed_read_cb(isc_nmsocket_t *sock, isc_result_t result) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(result != ISC_R_SUCCESS);
isc__nmsocket_timer_stop(sock);
isc__nm_stop_reading(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (!sock->recv_read) {
goto destroy;
}
sock->recv_read = false;
if (sock->recv_cb != NULL) {
isc__nm_uvreq_t *req = isc__nm_get_read_req(sock, NULL);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nmsocket_clearcb(sock);
isc__nm_readcb(sock, req, result);
}
destroy:
isc__nmsocket_prep_destroy(sock);
/*
* We need to detach from quota after the read callback function had a
* chance to be executed.
*/
if (sock->quota != NULL) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_quota_detach(&sock->quota);
}
}
void
isc__nm_tcpdns_read(isc_nmhandle_t *handle, isc_nm_recv_cb_t cb, void *cbarg) {
REQUIRE(VALID_NMHANDLE(handle));
REQUIRE(VALID_NMSOCK(handle->sock));
isc_nmsocket_t *sock = handle->sock;
isc__netievent_tcpdnsread_t *ievent = NULL;
REQUIRE(sock->type == isc_nm_tcpdnssocket);
REQUIRE(sock->statichandle == handle);
REQUIRE(sock->tid == isc_nm_tid());
REQUIRE(!sock->recv_read);
sock->recv_cb = cb;
sock->recv_cbarg = cbarg;
sock->recv_read = true;
if (sock->read_timeout == 0) {
Fix the data race in accessing the isc_nm_t timers The following TSAN report about accessing the mgr timers (mgr->init, mgr->idle, mgr->keepalive and mgr->advertised) has been fixed in this commit: ================== WARNING: ThreadSanitizer: data race (pid=2746) Read of size 4 at 0x7b440008a948 by thread T18: #0 isc__nm_tcpdns_read /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 (libisc.so.1706+0x2ba0f) #1 isc_nm_read /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1679:3 (libisc.so.1706+0x22258) #2 tcpdns_connect_connect_cb /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:363:2 (tcpdns_test+0x4bc5fb) #3 isc__nm_async_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1816:2 (libisc.so.1706+0x228c9) #4 isc__nm_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1791:3 (libisc.so.1706+0x22713) #5 tcpdns_connect_cb /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:343:2 (libisc.so.1706+0x2d89d) #6 uv__stream_connect /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1381:5 (libuv.so.1+0x27c18) #7 uv__stream_io /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1298:5 (libuv.so.1+0x25977) #8 uv__io_poll /home/ondrej/Projects/tsan/libuv/src/unix/linux-core.c:462:11 (libuv.so.1+0x2e795) #9 uv_run /home/ondrej/Projects/tsan/libuv/src/unix/core.c:385:5 (libuv.so.1+0x158ec) #10 nm_thread /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:530:11 (libisc.so.1706+0x1c94a) Previous write of size 4 at 0x7b440008a948 by main thread: #0 isc_nm_settimeouts /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:490:12 (libisc.so.1706+0x1dda5) #1 tcpdns_recv_two /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:601:2 (tcpdns_test+0x4bad0e) #2 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70be) #3 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Location is heap block of size 281 at 0x7b440008a840 allocated by main thread: #0 malloc <null> (tcpdns_test+0x42864b) #1 default_memalloc /home/ondrej/Projects/bind9/lib/isc/mem.c:713:8 (libisc.so.1706+0x6d261) #2 mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:622:8 (libisc.so.1706+0x69b9c) #3 isc___mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:1044:9 (libisc.so.1706+0x6d379) #4 isc__mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:2432:10 (libisc.so.1706+0x6889e) #5 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:203:8 (libisc.so.1706+0x1c219) #6 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #7 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #8 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Thread T18 'isc-net-0000' (tid=3513, running) created by main thread at: #0 pthread_create <null> (tcpdns_test+0x429e7b) #1 isc_thread_create /home/ondrej/Projects/bind9/lib/isc/pthreads/thread.c:73:8 (libisc.so.1706+0x8476a) #2 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:271:3 (libisc.so.1706+0x1c66a) #3 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #4 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #5 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) SUMMARY: ThreadSanitizer: data race /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 in isc__nm_tcpdns_read ================== ThreadSanitizer: reported 1 warnings
2020-12-02 09:52:39 +01:00
sock->read_timeout =
(atomic_load(&sock->keepalive)
? atomic_load(&sock->mgr->keepalive)
: atomic_load(&sock->mgr->idle));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
ievent = isc__nm_get_netievent_tcpdnsread(sock->mgr, sock);
/*
* This MUST be done asynchronously, no matter which thread we're
* in. The callback function for isc_nm_read() often calls
* isc_nm_read() again; if we tried to do that synchronously
* we'd clash in processbuffer() and grow the stack indefinitely.
*/
isc__nm_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
return;
}
void
isc__nm_async_tcpdnsread(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnsread_t *ievent =
(isc__netievent_tcpdnsread_t *)ev0;
isc_nmsocket_t *sock = ievent->sock;
UNUSED(worker);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
if (isc__nmsocket_closing(sock)) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
sock->reading = true;
isc__nm_failed_read_cb(sock, ISC_R_CANCELED, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return;
}
isc__nm_process_sock_buffer(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
/*
* Process a single packet from the incoming buffer.
*
* Return ISC_R_SUCCESS and attach 'handlep' to a handle if something
* was processed; return ISC_R_NOMORE if there isn't a full message
* to be processed.
*
* The caller will need to unreference the handle.
*/
isc_result_t
isc__nm_tcpdns_processbuffer(isc_nmsocket_t *sock) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
size_t len;
isc__nm_uvreq_t *req = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmhandle_t *handle = NULL;
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
if (isc__nmsocket_closing(sock)) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return (ISC_R_CANCELED);
}
/*
* If we don't even have the length yet, we can't do
* anything.
*/
if (sock->buf_len < 2) {
return (ISC_R_NOMORE);
}
/*
* Process the first packet from the buffer, leaving
* the rest (if any) for later.
*/
len = ntohs(*(uint16_t *)sock->buf);
if (len > sock->buf_len - 2) {
return (ISC_R_NOMORE);
}
req = isc__nm_get_read_req(sock, NULL);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_UVREQ(req));
/*
* We need to launch the resume_processing after the buffer has
* been consumed, thus we need to delay the detaching the handle.
*/
isc_nmhandle_attach(req->handle, &handle);
/*
* The callback will be called synchronously because the
* result is ISC_R_SUCCESS, so we don't need to have
* the buffer on the heap
*/
req->uvbuf.base = (char *)sock->buf + 2;
req->uvbuf.len = len;
/*
* If isc__nm_tcpdns_read() was called, it will be satisfied by single
* DNS message in the next call.
*/
sock->recv_read = false;
/*
* The assertion failure here means that there's a errnoneous extra
* nmhandle detach happening in the callback and resume_processing gets
* called while we are still processing the buffer.
*/
REQUIRE(sock->processing == false);
sock->processing = true;
isc__nm_readcb(sock, req, ISC_R_SUCCESS);
sock->processing = false;
len += 2;
sock->buf_len -= len;
if (sock->buf_len > 0) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
memmove(sock->buf, sock->buf + len, sock->buf_len);
}
isc_nmhandle_detach(&handle);
return (ISC_R_SUCCESS);
}
void
isc__nm_tcpdns_read_cb(uv_stream_t *stream, ssize_t nread,
const uv_buf_t *buf) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmsocket_t *sock = uv_handle_get_data((uv_handle_t *)stream);
uint8_t *base = NULL;
size_t len;
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
REQUIRE(sock->reading);
REQUIRE(buf != NULL);
if (isc__nmsocket_closing(sock)) {
isc__nm_failed_read_cb(sock, ISC_R_CANCELED, true);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
goto free;
}
if (nread < 0) {
if (nread != UV_EOF) {
isc__nm_incstats(sock->mgr,
sock->statsindex[STATID_RECVFAIL]);
}
isc__nm_failed_read_cb(sock, isc__nm_uverr2result(nread), true);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
goto free;
}
base = (uint8_t *)buf->base;
len = nread;
/*
* FIXME: We can avoid the memmove here if we know we have received full
* packet; e.g. we should be smarter, a.s. there are just few situations
*
* The tcp_alloc_buf should be smarter and point the uv_read_start to
* the position where previous read has ended in the sock->buf, that way
* the data could be read directly into sock->buf.
*/
if (sock->buf_len + len > sock->buf_size) {
isc__nm_alloc_dnsbuf(sock, sock->buf_len + len);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
memmove(sock->buf + sock->buf_len, base, len);
sock->buf_len += len;
if (!atomic_load(&sock->client)) {
Fix the data race in accessing the isc_nm_t timers The following TSAN report about accessing the mgr timers (mgr->init, mgr->idle, mgr->keepalive and mgr->advertised) has been fixed in this commit: ================== WARNING: ThreadSanitizer: data race (pid=2746) Read of size 4 at 0x7b440008a948 by thread T18: #0 isc__nm_tcpdns_read /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 (libisc.so.1706+0x2ba0f) #1 isc_nm_read /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1679:3 (libisc.so.1706+0x22258) #2 tcpdns_connect_connect_cb /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:363:2 (tcpdns_test+0x4bc5fb) #3 isc__nm_async_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1816:2 (libisc.so.1706+0x228c9) #4 isc__nm_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1791:3 (libisc.so.1706+0x22713) #5 tcpdns_connect_cb /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:343:2 (libisc.so.1706+0x2d89d) #6 uv__stream_connect /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1381:5 (libuv.so.1+0x27c18) #7 uv__stream_io /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1298:5 (libuv.so.1+0x25977) #8 uv__io_poll /home/ondrej/Projects/tsan/libuv/src/unix/linux-core.c:462:11 (libuv.so.1+0x2e795) #9 uv_run /home/ondrej/Projects/tsan/libuv/src/unix/core.c:385:5 (libuv.so.1+0x158ec) #10 nm_thread /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:530:11 (libisc.so.1706+0x1c94a) Previous write of size 4 at 0x7b440008a948 by main thread: #0 isc_nm_settimeouts /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:490:12 (libisc.so.1706+0x1dda5) #1 tcpdns_recv_two /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:601:2 (tcpdns_test+0x4bad0e) #2 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70be) #3 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Location is heap block of size 281 at 0x7b440008a840 allocated by main thread: #0 malloc <null> (tcpdns_test+0x42864b) #1 default_memalloc /home/ondrej/Projects/bind9/lib/isc/mem.c:713:8 (libisc.so.1706+0x6d261) #2 mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:622:8 (libisc.so.1706+0x69b9c) #3 isc___mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:1044:9 (libisc.so.1706+0x6d379) #4 isc__mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:2432:10 (libisc.so.1706+0x6889e) #5 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:203:8 (libisc.so.1706+0x1c219) #6 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #7 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #8 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Thread T18 'isc-net-0000' (tid=3513, running) created by main thread at: #0 pthread_create <null> (tcpdns_test+0x429e7b) #1 isc_thread_create /home/ondrej/Projects/bind9/lib/isc/pthreads/thread.c:73:8 (libisc.so.1706+0x8476a) #2 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:271:3 (libisc.so.1706+0x1c66a) #3 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #4 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #5 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) SUMMARY: ThreadSanitizer: data race /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 in isc__nm_tcpdns_read ================== ThreadSanitizer: reported 1 warnings
2020-12-02 09:52:39 +01:00
sock->read_timeout = atomic_load(&sock->mgr->idle);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
isc__nm_process_sock_buffer(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
free:
isc__nm_free_uvbuf(sock, buf);
}
static void
quota_accept_cb(isc_quota_t *quota, void *sock0) {
isc_nmsocket_t *sock = (isc_nmsocket_t *)sock0;
REQUIRE(VALID_NMSOCK(sock));
/*
* Create a tcpdnsaccept event and pass it using the async channel.
*/
isc__netievent_tcpdnsaccept_t *ievent =
isc__nm_get_netievent_tcpdnsaccept(sock->mgr, sock, quota);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_maybe_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
}
/*
* This is called after we get a quota_accept_cb() callback.
*/
void
isc__nm_async_tcpdnsaccept(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnsaccept_t *ievent =
(isc__netievent_tcpdnsaccept_t *)ev0;
isc_result_t result;
UNUSED(worker);
REQUIRE(VALID_NMSOCK(ievent->sock));
REQUIRE(ievent->sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
result = accept_connection(ievent->sock, ievent->quota);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (result != ISC_R_SUCCESS && result != ISC_R_NOCONN) {
if ((result != ISC_R_QUOTA && result != ISC_R_SOFTQUOTA) ||
can_log_tcpdns_quota())
{
isc_log_write(isc_lctx, ISC_LOGCATEGORY_GENERAL,
ISC_LOGMODULE_NETMGR, ISC_LOG_ERROR,
"TCP connection failed: %s",
isc_result_totext(result));
}
}
}
static isc_result_t
accept_connection(isc_nmsocket_t *ssock, isc_quota_t *quota) {
isc_nmsocket_t *csock = NULL;
isc__networker_t *worker = NULL;
int r;
isc_result_t result;
struct sockaddr_storage peer_ss;
struct sockaddr_storage local_ss;
isc_sockaddr_t local;
isc_nmhandle_t *handle = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(ssock));
REQUIRE(ssock->tid == isc_nm_tid());
if (isc__nmsocket_closing(ssock)) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (quota != NULL) {
isc_quota_detach(&quota);
}
return (ISC_R_CANCELED);
}
REQUIRE(ssock->accept_cb != NULL);
csock = isc_mem_get(ssock->mgr->mctx, sizeof(isc_nmsocket_t));
isc__nmsocket_init(csock, ssock->mgr, isc_nm_tcpdnssocket,
ssock->iface);
csock->tid = ssock->tid;
csock->extrahandlesize = ssock->extrahandlesize;
isc__nmsocket_attach(ssock, &csock->server);
csock->recv_cb = ssock->recv_cb;
csock->recv_cbarg = ssock->recv_cbarg;
csock->quota = quota;
csock->accepting = true;
worker = &csock->mgr->workers[csock->tid];
r = uv_tcp_init(&worker->loop, &csock->uv_handle.tcp);
RUNTIME_CHECK(r == 0);
uv_handle_set_data(&csock->uv_handle.handle, csock);
r = uv_timer_init(&worker->loop, &csock->timer);
RUNTIME_CHECK(r == 0);
uv_handle_set_data((uv_handle_t *)&csock->timer, csock);
r = uv_accept(&ssock->uv_handle.stream, &csock->uv_handle.stream);
if (r != 0) {
result = isc__nm_uverr2result(r);
goto failure;
}
r = uv_tcp_getpeername(&csock->uv_handle.tcp,
(struct sockaddr *)&peer_ss,
&(int){ sizeof(peer_ss) });
if (r != 0) {
result = isc__nm_uverr2result(r);
goto failure;
}
result = isc_sockaddr_fromsockaddr(&csock->peer,
(struct sockaddr *)&peer_ss);
if (result != ISC_R_SUCCESS) {
goto failure;
}
r = uv_tcp_getsockname(&csock->uv_handle.tcp,
(struct sockaddr *)&local_ss,
&(int){ sizeof(local_ss) });
if (r != 0) {
result = isc__nm_uverr2result(r);
goto failure;
}
result = isc_sockaddr_fromsockaddr(&local,
(struct sockaddr *)&local_ss);
if (result != ISC_R_SUCCESS) {
goto failure;
}
/*
* The handle will be either detached on acceptcb failure or in the
* readcb.
*/
handle = isc__nmhandle_get(csock, NULL, &local);
result = ssock->accept_cb(handle, ISC_R_SUCCESS, ssock->accept_cbarg);
if (result != ISC_R_SUCCESS) {
isc_nmhandle_detach(&handle);
goto failure;
}
csock->accepting = false;
isc__nm_incstats(csock->mgr, csock->statsindex[STATID_ACCEPT]);
Fix the data race in accessing the isc_nm_t timers The following TSAN report about accessing the mgr timers (mgr->init, mgr->idle, mgr->keepalive and mgr->advertised) has been fixed in this commit: ================== WARNING: ThreadSanitizer: data race (pid=2746) Read of size 4 at 0x7b440008a948 by thread T18: #0 isc__nm_tcpdns_read /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 (libisc.so.1706+0x2ba0f) #1 isc_nm_read /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1679:3 (libisc.so.1706+0x22258) #2 tcpdns_connect_connect_cb /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:363:2 (tcpdns_test+0x4bc5fb) #3 isc__nm_async_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1816:2 (libisc.so.1706+0x228c9) #4 isc__nm_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1791:3 (libisc.so.1706+0x22713) #5 tcpdns_connect_cb /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:343:2 (libisc.so.1706+0x2d89d) #6 uv__stream_connect /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1381:5 (libuv.so.1+0x27c18) #7 uv__stream_io /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1298:5 (libuv.so.1+0x25977) #8 uv__io_poll /home/ondrej/Projects/tsan/libuv/src/unix/linux-core.c:462:11 (libuv.so.1+0x2e795) #9 uv_run /home/ondrej/Projects/tsan/libuv/src/unix/core.c:385:5 (libuv.so.1+0x158ec) #10 nm_thread /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:530:11 (libisc.so.1706+0x1c94a) Previous write of size 4 at 0x7b440008a948 by main thread: #0 isc_nm_settimeouts /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:490:12 (libisc.so.1706+0x1dda5) #1 tcpdns_recv_two /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:601:2 (tcpdns_test+0x4bad0e) #2 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70be) #3 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Location is heap block of size 281 at 0x7b440008a840 allocated by main thread: #0 malloc <null> (tcpdns_test+0x42864b) #1 default_memalloc /home/ondrej/Projects/bind9/lib/isc/mem.c:713:8 (libisc.so.1706+0x6d261) #2 mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:622:8 (libisc.so.1706+0x69b9c) #3 isc___mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:1044:9 (libisc.so.1706+0x6d379) #4 isc__mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:2432:10 (libisc.so.1706+0x6889e) #5 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:203:8 (libisc.so.1706+0x1c219) #6 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #7 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #8 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Thread T18 'isc-net-0000' (tid=3513, running) created by main thread at: #0 pthread_create <null> (tcpdns_test+0x429e7b) #1 isc_thread_create /home/ondrej/Projects/bind9/lib/isc/pthreads/thread.c:73:8 (libisc.so.1706+0x8476a) #2 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:271:3 (libisc.so.1706+0x1c66a) #3 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #4 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #5 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) SUMMARY: ThreadSanitizer: data race /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 in isc__nm_tcpdns_read ================== ThreadSanitizer: reported 1 warnings
2020-12-02 09:52:39 +01:00
csock->read_timeout = atomic_load(&csock->mgr->init);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
csock->closehandle_cb = isc__nm_resume_processing;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* We need to keep the handle alive until we fail to read or connection
* is closed by the other side, it will be detached via
* prep_destroy()->tcpdns_close_direct().
*/
isc_nmhandle_attach(handle, &csock->recv_handle);
isc__nm_process_sock_buffer(csock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* The initial timer has been set, update the read timeout for the next
* reads.
*/
csock->read_timeout = (atomic_load(&csock->keepalive)
Fix the data race in accessing the isc_nm_t timers The following TSAN report about accessing the mgr timers (mgr->init, mgr->idle, mgr->keepalive and mgr->advertised) has been fixed in this commit: ================== WARNING: ThreadSanitizer: data race (pid=2746) Read of size 4 at 0x7b440008a948 by thread T18: #0 isc__nm_tcpdns_read /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 (libisc.so.1706+0x2ba0f) #1 isc_nm_read /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1679:3 (libisc.so.1706+0x22258) #2 tcpdns_connect_connect_cb /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:363:2 (tcpdns_test+0x4bc5fb) #3 isc__nm_async_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1816:2 (libisc.so.1706+0x228c9) #4 isc__nm_connectcb /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:1791:3 (libisc.so.1706+0x22713) #5 tcpdns_connect_cb /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:343:2 (libisc.so.1706+0x2d89d) #6 uv__stream_connect /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1381:5 (libuv.so.1+0x27c18) #7 uv__stream_io /home/ondrej/Projects/tsan/libuv/src/unix/stream.c:1298:5 (libuv.so.1+0x25977) #8 uv__io_poll /home/ondrej/Projects/tsan/libuv/src/unix/linux-core.c:462:11 (libuv.so.1+0x2e795) #9 uv_run /home/ondrej/Projects/tsan/libuv/src/unix/core.c:385:5 (libuv.so.1+0x158ec) #10 nm_thread /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:530:11 (libisc.so.1706+0x1c94a) Previous write of size 4 at 0x7b440008a948 by main thread: #0 isc_nm_settimeouts /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:490:12 (libisc.so.1706+0x1dda5) #1 tcpdns_recv_two /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:601:2 (tcpdns_test+0x4bad0e) #2 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70be) #3 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Location is heap block of size 281 at 0x7b440008a840 allocated by main thread: #0 malloc <null> (tcpdns_test+0x42864b) #1 default_memalloc /home/ondrej/Projects/bind9/lib/isc/mem.c:713:8 (libisc.so.1706+0x6d261) #2 mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:622:8 (libisc.so.1706+0x69b9c) #3 isc___mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:1044:9 (libisc.so.1706+0x6d379) #4 isc__mem_get /home/ondrej/Projects/bind9/lib/isc/mem.c:2432:10 (libisc.so.1706+0x6889e) #5 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:203:8 (libisc.so.1706+0x1c219) #6 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #7 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #8 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) Thread T18 'isc-net-0000' (tid=3513, running) created by main thread at: #0 pthread_create <null> (tcpdns_test+0x429e7b) #1 isc_thread_create /home/ondrej/Projects/bind9/lib/isc/pthreads/thread.c:73:8 (libisc.so.1706+0x8476a) #2 isc_nm_start /home/ondrej/Projects/bind9/lib/isc/netmgr/netmgr.c:271:3 (libisc.so.1706+0x1c66a) #3 nm_setup /home/ondrej/Projects/bind9/lib/isc/tests/tcpdns_test.c:244:11 (tcpdns_test+0x4baaa4) #4 cmocka_run_one_test_or_fixture <null> (libcmocka.so.0+0x70fd) #5 __libc_start_main /build/glibc-vjB4T1/glibc-2.28/csu/../csu/libc-start.c:308:16 (libc.so.6+0x2409a) SUMMARY: ThreadSanitizer: data race /home/ondrej/Projects/bind9/lib/isc/netmgr/tcpdns.c:849:25 in isc__nm_tcpdns_read ================== ThreadSanitizer: reported 1 warnings
2020-12-02 09:52:39 +01:00
? atomic_load(&csock->mgr->keepalive)
: atomic_load(&csock->mgr->idle));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmhandle_detach(&handle);
/*
* sock is now attached to the handle.
*/
isc__nmsocket_detach(&csock);
return (ISC_R_SUCCESS);
failure:
atomic_store(&csock->active, false);
isc__nm_failed_accept_cb(csock, result);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nmsocket_prep_destroy(csock);
isc__nmsocket_detach(&csock);
return (result);
}
void
isc__nm_tcpdns_send(isc_nmhandle_t *handle, isc_region_t *region,
isc_nm_cb_t cb, void *cbarg) {
REQUIRE(VALID_NMHANDLE(handle));
REQUIRE(VALID_NMSOCK(handle->sock));
isc_nmsocket_t *sock = handle->sock;
isc__netievent_tcpdnssend_t *ievent = NULL;
isc__nm_uvreq_t *uvreq = NULL;
REQUIRE(sock->type == isc_nm_tcpdnssocket);
uvreq = isc__nm_uvreq_get(sock->mgr, sock);
*(uint16_t *)uvreq->tcplen = htons(region->length);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
uvreq->uvbuf.base = (char *)region->base;
uvreq->uvbuf.len = region->length;
isc_nmhandle_attach(handle, &uvreq->handle);
uvreq->cb.send = cb;
uvreq->cbarg = cbarg;
ievent = isc__nm_get_netievent_tcpdnssend(sock->mgr, sock, uvreq);
isc__nm_maybe_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
return;
}
static void
tcpdns_send_cb(uv_write_t *req, int status) {
isc__nm_uvreq_t *uvreq = (isc__nm_uvreq_t *)req->data;
isc_nmsocket_t *sock = uvreq->sock;
REQUIRE(VALID_UVREQ(uvreq));
REQUIRE(VALID_NMHANDLE(uvreq->handle));
if (status < 0) {
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_SENDFAIL]);
isc__nm_failed_send_cb(sock, uvreq,
isc__nm_uverr2result(status));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return;
}
isc__nm_sendcb(sock, uvreq, ISC_R_SUCCESS, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
/*
* Handle 'tcpsend' async event - send a packet on the socket
*/
void
isc__nm_async_tcpdnssend(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnssend_t *ievent =
(isc__netievent_tcpdnssend_t *)ev0;
REQUIRE(ievent->sock->type == isc_nm_tcpdnssocket);
REQUIRE(ievent->sock->tid == isc_nm_tid());
REQUIRE(VALID_NMSOCK(ievent->sock));
REQUIRE(VALID_UVREQ(ievent->req));
REQUIRE(ievent->sock->tid == isc_nm_tid());
isc_result_t result;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmsocket_t *sock = ievent->sock;
isc__nm_uvreq_t *uvreq = ievent->req;
uv_buf_t bufs[2] = { { .base = uvreq->tcplen, .len = 2 },
{ .base = uvreq->uvbuf.base,
.len = uvreq->uvbuf.len } };
int nbufs = 2;
int r;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
UNUSED(worker);
if (isc__nmsocket_closing(sock)) {
result = ISC_R_CANCELED;
goto fail;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
r = uv_try_write(&sock->uv_handle.stream, bufs, nbufs);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (r == (int)(bufs[0].len + bufs[1].len)) {
/* Wrote everything */
isc__nm_sendcb(sock, uvreq, ISC_R_SUCCESS, true);
return;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (r == 1) {
/* Partial write of DNSMSG length */
bufs[0].base = uvreq->tcplen + 1;
bufs[0].len = 1;
} else if (r > 0) {
/* Partial write of DNSMSG */
nbufs = 1;
bufs[0].base = uvreq->uvbuf.base + (r - 2);
bufs[0].len = uvreq->uvbuf.len - (r - 2);
} else if (r == UV_ENOSYS || r == UV_EAGAIN) {
/* uv_try_write not support, send asynchronously */
} else {
/* error sending data */
result = isc__nm_uverr2result(r);
goto fail;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
r = uv_write(&uvreq->uv_req.write, &sock->uv_handle.stream, bufs, nbufs,
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
tcpdns_send_cb);
if (r < 0) {
result = isc__nm_uverr2result(r);
goto fail;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
return;
fail:
if (result != ISC_R_SUCCESS) {
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_SENDFAIL]);
isc__nm_failed_send_cb(sock, uvreq, result);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
static void
tcpdns_stop_cb(uv_handle_t *handle) {
isc_nmsocket_t *sock = uv_handle_get_data(handle);
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
REQUIRE(atomic_load(&sock->closing));
uv_handle_set_data(handle, NULL);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (!atomic_compare_exchange_strong(&sock->closed, &(bool){ false },
true)) {
INSIST(0);
ISC_UNREACHABLE();
}
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_CLOSE]);
atomic_store(&sock->listening, false);
isc__nmsocket_detach(&sock);
}
static void
tcpdns_close_sock(isc_nmsocket_t *sock) {
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(atomic_load(&sock->closing));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (!atomic_compare_exchange_strong(&sock->closed, &(bool){ false },
true)) {
INSIST(0);
ISC_UNREACHABLE();
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_incstats(sock->mgr, sock->statsindex[STATID_CLOSE]);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sock->server != NULL) {
isc__nmsocket_detach(&sock->server);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->connected, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nmsocket_prep_destroy(sock);
}
static void
tcpdns_close_cb(uv_handle_t *handle) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nmsocket_t *sock = uv_handle_get_data(handle);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
uv_handle_set_data(handle, NULL);
tcpdns_close_sock(sock);
}
static void
timer_close_cb(uv_handle_t *timer) {
isc_nmsocket_t *sock = uv_handle_get_data(timer);
uv_handle_set_data(timer, NULL);
REQUIRE(VALID_NMSOCK(sock));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sock->parent) {
uv_close(&sock->uv_handle.handle, tcpdns_stop_cb);
} else if (uv_is_closing(&sock->uv_handle.handle)) {
tcpdns_close_sock(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
} else {
uv_close(&sock->uv_handle.handle, tcpdns_close_cb);
}
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
static void
stop_tcpdns_child(isc_nmsocket_t *sock) {
REQUIRE(sock->type == isc_nm_tcpdnssocket);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (!atomic_compare_exchange_strong(&sock->closing, &(bool){ false },
true)) {
return;
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
tcpdns_close_direct(sock);
atomic_fetch_sub(&sock->parent->rchildren, 1);
Refactor taskmgr to run on top of netmgr This commit changes the taskmgr to run the individual tasks on the netmgr internal workers. While an effort has been put into keeping the taskmgr interface intact, couple of changes have been made: * The taskmgr has no concept of universal privileged mode - rather the tasks are either privileged or unprivileged (normal). The privileged tasks are run as a first thing when the netmgr is unpaused. There are now four different queues in in the netmgr: 1. priority queue - netievent on the priority queue are run even when the taskmgr enter exclusive mode and netmgr is paused. This is needed to properly start listening on the interfaces, free resources and resume. 2. privileged task queue - only privileged tasks are queued here and this is the first queue that gets processed when network manager is unpaused using isc_nm_resume(). All netmgr workers need to clean the privileged task queue before they all proceed normal operation. Both task queues are processed when the workers are finished. 3. task queue - only (traditional) task are scheduled here and this queue along with privileged task queues are process when the netmgr workers are finishing. This is needed to process the task shutdown events. 4. normal queue - this is the queue with netmgr events, e.g. reading, sending, callbacks and pretty much everything is processed here. * The isc_taskmgr_create() now requires initialized netmgr (isc_nm_t) object. * The isc_nm_destroy() function now waits for indefinite time, but it will print out the active objects when in tracing mode (-DNETMGR_TRACE=1 and -DNETMGR_TRACE_VERBOSE=1), the netmgr has been made a little bit more asynchronous and it might take longer time to shutdown all the active networking connections. * Previously, the isc_nm_stoplistening() was a synchronous operation. This has been changed and the isc_nm_stoplistening() just schedules the child sockets to stop listening and exits. This was needed to prevent a deadlock as the the (traditional) tasks are now executed on the netmgr threads. * The socket selection logic in isc__nm_udp_send() was flawed, but fortunatelly, it was broken, so we never hit the problem where we created uvreq_t on a socket from nmhandle_t, but then a different socket could be picked up and then we were trying to run the send callback on a socket that had different threadid than currently running.
2021-04-09 11:31:19 +02:00
isc_barrier_wait(&sock->parent->stoplistening);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
static void
stop_tcpdns_parent(isc_nmsocket_t *sock) {
isc_nmsocket_t *csock = NULL;
REQUIRE(VALID_NMSOCK(sock));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(sock->type == isc_nm_tcpdnslistener);
isc_barrier_init(&sock->stoplistening, sock->nchildren);
for (size_t i = 0; i < sock->nchildren; i++) {
csock = &sock->children[i];
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(csock));
if ((int)i == isc_nm_tid()) {
/*
* We need to schedule closing the other sockets first
*/
continue;
}
atomic_store(&csock->active, false);
enqueue_stoplistening(csock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
csock = &sock->children[isc_nm_tid()];
atomic_store(&csock->active, false);
stop_tcpdns_child(csock);
atomic_store(&sock->closed, true);
isc__nmsocket_prep_destroy(sock);
}
static void
2020-02-13 14:44:37 -08:00
tcpdns_close_direct(isc_nmsocket_t *sock) {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(atomic_load(&sock->closing));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sock->quota != NULL) {
isc_quota_detach(&sock->quota);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sock->recv_handle != NULL) {
isc_nmhandle_detach(&sock->recv_handle);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nmsocket_timer_stop(sock);
isc__nm_stop_reading(sock);
uv_handle_set_data((uv_handle_t *)&sock->timer, sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
uv_close((uv_handle_t *)&sock->timer, timer_close_cb);
}
void
2020-02-13 14:44:37 -08:00
isc__nm_tcpdns_close(isc_nmsocket_t *sock) {
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->type == isc_nm_tcpdnssocket);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(!isc__nmsocket_active(sock));
if (!atomic_compare_exchange_strong(&sock->closing, &(bool){ false },
true)) {
return;
}
if (sock->tid == isc_nm_tid()) {
tcpdns_close_direct(sock);
} else {
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* We need to create an event and pass it using async channel
*/
isc__netievent_tcpdnsclose_t *ievent =
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_get_netievent_tcpdnsclose(sock->mgr, sock);
isc__nm_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
}
}
void
2020-02-13 14:44:37 -08:00
isc__nm_async_tcpdnsclose(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnsclose_t *ievent =
(isc__netievent_tcpdnsclose_t *)ev0;
isc_nmsocket_t *sock = ievent->sock;
UNUSED(worker);
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
tcpdns_close_direct(sock);
}
static void
tcpdns_close_connect_cb(uv_handle_t *handle) {
isc_nmsocket_t *sock = uv_handle_get_data(handle);
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(isc__nm_in_netthread());
REQUIRE(sock->tid == isc_nm_tid());
isc__nmsocket_prep_destroy(sock);
isc__nmsocket_detach(&sock);
}
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
void
isc__nm_tcpdns_shutdown(isc_nmsocket_t *sock) {
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
REQUIRE(sock->type == isc_nm_tcpdnssocket);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* If the socket is active, mark it inactive and
* continue. If it isn't active, stop now.
*/
if (!isc__nmsocket_deactivate(sock)) {
return;
}
if (sock->accepting) {
return;
}
if (atomic_load(&sock->connecting)) {
isc_nmsocket_t *tsock = NULL;
isc__nmsocket_attach(sock, &tsock);
uv_close(&sock->uv_handle.handle, tcpdns_close_connect_cb);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return;
}
if (sock->statichandle != NULL) {
isc__nm_failed_read_cb(sock, ISC_R_CANCELED, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
return;
}
/*
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
* Otherwise, we just send the socket to abyss...
*/
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
if (sock->parent == NULL) {
isc__nmsocket_prep_destroy(sock);
}
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_tcpdns_cancelread(isc_nmhandle_t *handle) {
isc_nmsocket_t *sock = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__netievent_tcpdnscancel_t *ievent = NULL;
REQUIRE(VALID_NMHANDLE(handle));
sock = handle->sock;
REQUIRE(VALID_NMSOCK(sock));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(sock->type == isc_nm_tcpdnssocket);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
ievent = isc__nm_get_netievent_tcpdnscancel(sock->mgr, sock, handle);
isc__nm_enqueue_ievent(&sock->mgr->workers[sock->tid],
(isc__netievent_t *)ievent);
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc__nm_async_tcpdnscancel(isc__networker_t *worker, isc__netievent_t *ev0) {
isc__netievent_tcpdnscancel_t *ievent =
(isc__netievent_tcpdnscancel_t *)ev0;
isc_nmsocket_t *sock = ievent->sock;
UNUSED(worker);
REQUIRE(VALID_NMSOCK(sock));
REQUIRE(sock->tid == isc_nm_tid());
isc__nm_failed_read_cb(sock, ISC_R_EOF, false);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nm_tcpdns_sequential(isc_nmhandle_t *handle) {
isc_nmsocket_t *sock = NULL;
REQUIRE(VALID_NMHANDLE(handle));
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMSOCK(handle->sock));
REQUIRE(handle->sock->type == isc_nm_tcpdnssocket);
sock = handle->sock;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
/*
* We don't want pipelining on this connection. That means
* that we need to pause after reading each request, and
* resume only after the request has been processed. This
* is done in resume_processing(), which is the socket's
* closehandle_cb callback, called whenever a handle
* is released.
*/
isc__nmsocket_timer_stop(sock);
isc__nm_stop_reading(sock);
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->sequential, true);
}
void
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
isc_nm_tcpdns_keepalive(isc_nmhandle_t *handle, bool value) {
isc_nmsocket_t *sock = NULL;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
REQUIRE(VALID_NMHANDLE(handle));
REQUIRE(VALID_NMSOCK(handle->sock));
REQUIRE(handle->sock->type != isc_nm_tcpdnssocket);
sock = handle->sock;
Refactor netmgr and add more unit tests This is a part of the works that intends to make the netmgr stable, testable, maintainable and tested. It contains a numerous changes to the netmgr code and unfortunately, it was not possible to split this into smaller chunks as the work here needs to be committed as a complete works. NOTE: There's a quite a lot of duplicated code between udp.c, tcp.c and tcpdns.c and it should be a subject to refactoring in the future. The changes that are included in this commit are listed here (extensively, but not exclusively): * The netmgr_test unit test was split into individual tests (udp_test, tcp_test, tcpdns_test and newly added tcp_quota_test) * The udp_test and tcp_test has been extended to allow programatic failures from the libuv API. Unfortunately, we can't use cmocka mock() and will_return(), so we emulate the behaviour with #define and including the netmgr/{udp,tcp}.c source file directly. * The netievents that we put on the nm queue have variable number of members, out of these the isc_nmsocket_t and isc_nmhandle_t always needs to be attached before enqueueing the netievent_<foo> and detached after we have called the isc_nm_async_<foo> to ensure that the socket (handle) doesn't disappear between scheduling the event and actually executing the event. * Cancelling the in-flight TCP connection using libuv requires to call uv_close() on the original uv_tcp_t handle which just breaks too many assumptions we have in the netmgr code. Instead of using uv_timer for TCP connection timeouts, we use platform specific socket option. * Fix the synchronization between {nm,async}_{listentcp,tcpconnect} When isc_nm_listentcp() or isc_nm_tcpconnect() is called it was waiting for socket to either end up with error (that path was fine) or to be listening or connected using condition variable and mutex. Several things could happen: 0. everything is ok 1. the waiting thread would miss the SIGNAL() - because the enqueued event would be processed faster than we could start WAIT()ing. In case the operation would end up with error, it would be ok, as the error variable would be unchanged. 2. the waiting thread miss the sock->{connected,listening} = `true` would be set to `false` in the tcp_{listen,connect}close_cb() as the connection would be so short lived that the socket would be closed before we could even start WAIT()ing * The tcpdns has been converted to using libuv directly. Previously, the tcpdns protocol used tcp protocol from netmgr, this proved to be very complicated to understand, fix and make changes to. The new tcpdns protocol is modeled in a similar way how tcp netmgr protocol. Closes: #2194, #2283, #2318, #2266, #2034, #1920 * The tcp and tcpdns is now not using isc_uv_import/isc_uv_export to pass accepted TCP sockets between netthreads, but instead (similar to UDP) uses per netthread uv_loop listener. This greatly reduces the complexity as the socket is always run in the associated nm and uv loops, and we are also not touching the libuv internals. There's an unfortunate side effect though, the new code requires support for load-balanced sockets from the operating system for both UDP and TCP (see #2137). If the operating system doesn't support the load balanced sockets (either SO_REUSEPORT on Linux or SO_REUSEPORT_LB on FreeBSD 12+), the number of netthreads is limited to 1. * The netmgr has now two debugging #ifdefs: 1. Already existing NETMGR_TRACE prints any dangling nmsockets and nmhandles before triggering assertion failure. This options would reduce performance when enabled, but in theory, it could be enabled on low-performance systems. 2. New NETMGR_TRACE_VERBOSE option has been added that enables extensive netmgr logging that allows the software engineer to precisely track any attach/detach operations on the nmsockets and nmhandles. This is not suitable for any kind of production machine, only for debugging. * The tlsdns netmgr protocol has been split from the tcpdns and it still uses the old method of stacking the netmgr boxes on top of each other. We will have to refactor the tlsdns netmgr protocol to use the same approach - build the stack using only libuv and openssl. * Limit but not assert the tcp buffer size in tcp_alloc_cb Closes: #2061
2020-11-12 10:32:18 +01:00
atomic_store(&sock->keepalive, value);
}