When shutting down TCP sockets, the read callback calling logic was
flawed, it would call either one less callback or one extra. Fix the
logic in the way:
1. When isc_nm_read() has been called but isc_nm_read_stop() hasn't on
the handle, the read callback will be called with ISC_R_CANCELED to
cancel active reading from the socket/handle.
2. When isc_nm_read() has been called and isc_nm_read_stop() has been
called on the on the handle, the read callback will be called with
ISC_R_SHUTTINGDOWN to signal that the dormant (not-reading) socket
is being shut down.
3. The .reading and .recv_read flags are little bit tricky. The
.reading flag indicates if the outer layer is reading the data (that
would be uv_tcp_t for TCP and isc_nmsocket_t (TCP) for TLSStream),
the .recv_read flag indicates whether somebody is interested in the
data read from the socket.
Usually, you would expect that the .reading should be false when
.recv_read is false, but it gets even more tricky with TLSStream as
the TLS protocol might need to read from the socket even when sending
data.
Fix the usage of the .recv_read and .reading flags in the TLSStream
to their true meaning - which mostly consist of using .recv_read
everywhere and then wrapping isc_nm_read() and isc_nm_read_stop()
with the .reading flag.
4. The TLS failed read helper has been modified to resemble the TCP code
as much as possible, clearing and re-setting the .recv_read flag in
the TCP timeout code has been fixed and .recv_read is now cleared
when isc_nm_read_stop() has been called on the streaming socket.
5. The use of Network Manager in the named_controlconf, isccc_ccmsg, and
isc_httpd units have been greatly simplified due to the improved design.
6. More unit tests for TCP and TLS testing the shutdown conditions have
been added.
Co-authored-by: Ondřej Surý <ondrej@isc.org>
Co-authored-by: Artem Boldariev <artem@isc.org>
In e18541287231b721c9cdb7e492697a2a80fd83fc, the TCP accept quota code
became broken in a subtle way - the quota would get initialized on the
first accept for the server socket and then deleted from the server
socket, so it would never get applied again.
Properly fixing this required a bigger refactoring of the isc_quota API
code to make it much simpler. The new code decouples the ownership of
the quota and acquiring/releasing the quota limit.
After (during) the refactoring it became more clear that we need to use
the callback from the child side of the accepted connection, and not the
server side.
If the quota callback is called on a thread matching the socket, call
the TCP accept function directly instead of using isc_async_run() which
allocates-deallocates memory.
This is a simple replacement using the semantic patch from the previous
commit and as added bonus, one removal of previously undetected unused
variable in named/server.c.
When accepting a TCP connection in the higher layers (tlsstream,
streamdns, and http) attach to the socket the connection was accepted
on, and use this socket instead of the parent listening socket.
This has an advantage - accessing the sock->listener now doesn't break
the thread boundaries, so we can properly check whether the socket is
being closed without requiring .closing member to be atomic_bool.
The last atomic_bool variable sock->active was converted to non-atomic
bool by properly handling the listening socket case where we were
checking parent socket instead of children sockets.
This is no longer necessary as we properly set the .active to false on
the children sockets.
Additionally, cleanup the .rchildren - the atomic variable was used for
mutex+condition to block until all children were listening, but that's
now being handled by a barrier.
Finally, just remove dead .self and .active_child_connections members of
the netmgr socket.
Now that everything runs on their own loop and we don't cross the thread
boundaries (with few exceptions), most of the atomic_bool variables used
to track the socket state have been unatomicized because they are always
accessed from the matching thread.
The remaining few have been relaxed: a) the sock->active is now using
acquire/release memory ordering; b) the various global limits are now
using relaxed memory ordering - we don't really care about the
synchronization for those.
Change the isc__nm_uvreq_t to have the idle callback as a separate
member as we always need to use it to properly close the uvreq.
Slightly refactor uvreq_put and uvreq_get to remove the unneeded
arguments - in uvreq_get(), we always use sock->worker, and in
uvreq_put, we always use req->sock, so there's not reason to pass those
extra arguments.
when isc_nm_listenstreamdns() is called with a local port of 0,
a random port is chosen. call uv_getsockname() to determine what
the port is as soon as the socket is bound, and add a function
isc_nmsocket_getaddr() to retrieve it, so that the caller can
connect to the listening socket. this will be used in cases
where the same process is acting as both client and server.
add a public function ns_interface_create() allowing the caller
to set up a listening interface directly without having to set
up listen-on and scan network interfaces.
Simplify the acception the new TCP connection by using the isc_async API
from the loopmgr instead of using the asychronous netievent mechanism in
the netmgr.
Simplify the stopping of the TCP children by using the isc_async API
from the loopmgr instead of using the asychronous netievent mechanism in
the netmgr.
Simplify the starting of the TCP children by using the isc_async API
from the loopmgr instead of using the asychronous netievent mechanism in
the netmgr.
The old code didn't handle race conditions and errors on systems
with non load balancing sockets gracefully. Look for an error on
any child socket and if found close all the child sockets and return
an error.
We already have a synchronization mechanism when starting the UDP and
TCP listener children - barriers. Change how we start the first-born
child (tid == 0), so we don't have to race for sock->parent->result and
sock->parent->fd.
Always track the per-worker sockets in the .active_sockets field in the
isc__networker_t struct and always track the per-socket handles in the
.active_handles field ian the isc_nmsocket_t struct.
On some platforms, when a synchronizing barrier is cleared, one
thread can progress while other threads are still in the process
of releasing the barrier. If a barrier is reused by the progressing
thread during this window, it can cause a deadlock. This can occur if,
for example, we stop listening immediately after we start, because the
stop and listen functions both use socket->barrier. This has been
addressed by using separate barrier objects for stop and listen.
The new internal function works in the same way as isc_nm_send()
except that it sends a DNS message size ahead of the DNS message
data (the format used in DNS over TCP).
The intention is to provide a fast path for sending DNS messages over
streams protocols - that is, without allocating any intermediate
memory buffers.
This commit adds a manual read timer control mode to the TCP
code (adding isc__nmhandle_set_manual_timer() as the interface to it).
Manual read timer control mode suppresses read timer restarting the
read timer when receiving any amount of data. This way the read timer
can be controlled manually using:
* isc__nmsocket_timer_start();
* isc__nmsocket_timer_stop();
* isc__nmsocket_timer_restart().
The change is required to make it possible to implement more
sophisticated read timer control policies in DNS transports, built on
top of TCP.
Add internal logging functions isc__netmgr_log, isc__nmsocket_log(), and
isc__nmhandle_log() that can be used to add logging messages to the
netmgr, and change all direct use of isc_log_write() to use those
logging functions to properly prefix them with netmgr, nmsocket and
nmsocket+nmhandle.
This commit make TCP code use uv_try_write() on best effort basis,
just like TCP DNS and TLS DNS code does.
This optimisation was added in
'caa5b6548a11da6ca772d6f7e10db3a164a18f8d' but, similar change was
mistakenly omitted for generic TCP code. This commit fixes that.
Previously, the send callback would be synchronous only on success. Add
an option (similar to what other callbacks have) to decide whether we
need the asynchronous send callback on a higher level.
On a general level, we need the asynchronous callbacks to happen only
when we are invoking the callback from the public API. If the path to
the callback went through the libuv callback or netmgr callback, we are
already on asynchronous path, and there's no need to make the call to
the callback asynchronous again.
For the send callback, this means we need the asynchronous path for
failure paths inside the isc_nm_send() (which calls isc__nm_udp_send(),
isc__nm_tcp_send(), etc...) - all other invocations of the send callback
could be synchronous, because those are called from the respective libuv
send callbacks.
Previously, the read callback would be synchronous only on success or
timeout. Add an option (similar to what other callbacks have) to decide
whether we need the asynchronous read callback on a higher level.
On a general level, we need the asynchronous callbacks to happen only
when we are invoking the callback from the public API. If the path to
the callback went through the libuv callback or netmgr callback, we are
already on asynchronous path, and there's no need to make the call to
the callback asynchronous again.
For the read callback, this means we need the asynchronous path for
failure paths inside the isc_nm_read() (which calls isc__nm_udp_read(),
isc__nm_tcp_read(), etc...) - all other invocations of the read callback
could be synchronous, because those are called from the respective libuv
or netmgr read callbacks.
This commit introduces a primitive isc__nmsocket_stop() which performs
shutting down on a multilayered socket ensuring the proper order of
the operations.
The shared data within the socket object can be destroyed after the
call completed, as it is guaranteed to not be used from within the
context of other worker threads.
The check is left from when tcp_connect_direct() called isc__nm_socket()
and it was uncertain whether it had succeeded, but now isc__nm_socket()
is called before tcp_connect_direct(), so sock->fd cannot be -1.
*** CID 357292: (REVERSE_NEGATIVE)
/lib/isc/netmgr/tcp.c: 309 in isc_nm_tcpconnect()
303
304 atomic_store(&sock->active, true);
305
306 result = tcp_connect_direct(sock, req);
307 if (result != ISC_R_SUCCESS) {
308 atomic_store(&sock->active, false);
>>> CID 357292: (REVERSE_NEGATIVE)
>>> You might be using variable "sock->fd" before verifying that it is >= 0.
309 if (sock->fd != (uv_os_sock_t)(-1)) {
310 isc__nm_tcp_close(sock);
311 }
312 isc__nm_connectcb(sock, req, result, true);
313 }
314
Add new semantic patch to replace the straightfoward uses of:
ptr = isc_mem_{get,allocate}(..., size);
memset(ptr, 0, size);
with the new API call:
ptr = isc_mem_{get,allocate}x(..., size, ISC_MEM_ZERO);
After the loopmgr work has been merged, we can now cleanup the TCP and
TLS protocols a little bit, because there are stronger guarantees that
the sockets will be kept on the respective loops/threads. We only need
asynchronous call for listening sockets (start, stop) and reading from
the TCP (because the isc_nm_read() might be called from read callback
again.
This commit does the following changes (they are intertwined together):
1. Cleanup most of the asynchronous events in the TCP code, and add
comments for the events that needs to be kept asynchronous.
2. Remove isc_nm_resumeread() from the netmgr API, and replace
isc_nm_resumeread() calls with existing isc_nm_read() calls.
3. Remove isc_nm_pauseread() from the netmgr API, and replace
isc_nm_pauseread() calls with a new isc_nm_read_stop() call.
4. Disable the isc_nm_cancelread() for the streaming protocols, only the
datagram-like protocols can use isc_nm_cancelread().
5. Add isc_nmhandle_close() that can be used to shutdown the socket
earlier than after the last detach. Formerly, the socket would be
closed only after all reading and sending would be finished and the
last reference would be detached. The new isc_nmhandle_close() can
be used to close the underlying socket earlier, so all the other
asynchronous calls would call their respective callbacks immediately.
Co-authored-by: Ondřej Surý <ondrej@isc.org>
Co-authored-by: Artem Boldariev <artem@isc.org>
Simplify the closing code - during the loopmgr implementation, it was
discovered that the various lists used by the uv_loop_t aren't FIFO, but
LIFO. See doc/dev/libuv.md for more details.
With this knowledge, we can close the protocol handles (uv_udp_t and
uv_tcp_t) and uv_timer_t at the same time by reordering the uv_close()
calls, and thus making sure that after calling the
isc__nm_stoplistening(), the code will not issue any additional callback
calls (accept, read) on the socket that stopped listening.
This might help with the TLS and DoH shutting down sequence as described
in the [GL #3509] as we now stop the reading, stop the timer and call
the uv_close() as earliest as possible.
When we are closing the listening sockets, there's a time window in
which the TCP connection could be accepted although the respective
stoplistening function has already returned to control to the caller.
Clear the accept callback function early, so it doesn't get called when
we are not interested in the incoming connections anymore.
Previously:
* applications were using isc_app as the base unit for running the
application and signal handling.
* networking was handled in the netmgr layer, which would start a
number of threads, each with a uv_loop event loop.
* task/event handling was done in the isc_task unit, which used
netmgr event loops to run the isc_event calls.
In this refactoring:
* the network manager now uses isc_loop instead of maintaining its
own worker threads and event loops.
* the taskmgr that manages isc_task instances now also uses isc_loopmgr,
and every isc_task runs on a specific isc_loop bound to the specific
thread.
* applications have been updated as necessary to use the new API.
* new ISC_LOOP_TEST macros have been added to enable unit tests to
run isc_loop event loops. unit tests have been updated to use this
where needed.
The STATID_CONNECT and STATID_CONNECTFAIL statistics were used
incorrectly. The STATID_CONNECT was incremented twice (once in
the *_connect_direct() and once in the callback) and STATID_CONNECTFAIL
would not be incremented at all if the failure happened in the callback.
Closes: #3452
On FreeBSD (and perhaps other *BSD) systems, the TCP connect() call (via
uv_tcp_connect()) can fail with transient UV_EADDRINUSE error. The UDP
code already handles this by trying three times (is a charm) before
giving up. Add a code for the TCP, TCPDNS and TLSDNS layers to also try
three times before giving up by calling uv_tcp_connect() from the
callback two more time on UV_EADDRINUSE error.
Additionally, stop the timer only if we succeed or on hard error via
isc__nm_failed_connect_cb().
Under specific rare timing circumstances the uv_read_start() could
fail with UV_EINVAL when the connection is reset between the connect (or
accept) and the uv_read_start() call on the nmworker loop. Handle such
situation gracefully by propagating the errors from uv_read_start() into
upper layers, so the socket can be internally closed().
Setting the sock->write_timeout from the TCP, TCPDNS, and TLSDNS send
functions could lead to (harmless) data race when setting the value for
the first time when the isc_nm_send() function would be called from
thread not-matching the socket we are sending to. Move the setting the
sock->write_timeout to the matching async function which is always
called from the matching thread.
As we are going to use libuv outside of the netmgr, we need the shims to
be readily available for the rest of the codebase.
Move the "netmgr/uv-compat.h" to <isc/uv.h> and netmgr/uv-compat.c to
uv.c, and as a rule of thumb, the users of libuv should include
<isc/uv.h> instead of <uv.h> directly.
Additionally, merge netmgr/uverr2result.c into uv.c and rename the
single function from isc__nm_uverr2result() to isc_uverr2result().
Move the netmgr socket related functions from netmgr/netmgr.c and
netmgr/uv-compat.c to netmgr/socket.c, so they are all present all in
the same place. Adjust the names of couple interal functions
accordingly.
For some applications, it's useful to not listen on full battery of
threads. Add workers argument to all isc_nm_listen*() functions and
convenience ISC_NM_LISTEN_ONE and ISC_NM_LISTEN_ALL macros.
Previously, the option to enable kernel load balancing of the sockets
was always enabled when supported by the operating system (SO_REUSEPORT
on Linux and SO_REUSEPORT_LB on FreeBSD).
It was reported that in scenarios where the networking threads are also
responsible for processing long-running tasks (like RPZ processing, CATZ
processing or large zone transfers), this could lead to intermitten
brownouts for some clients, because the thread assigned by the operating
system might be busy. In such scenarious, the overall performance would
be better served by threads competing over the sockets because the idle
threads can pick up the incoming traffic.
Add new configuration option (`load-balance-sockets`) to allow enabling
or disabling the load balancing of the sockets.
Previously, it was possible to assign a bit of memory space in the
nmhandle to store the client data. This was complicated and prevents
further refactoring of isc_nmhandle_t caching (future work).
Instead of caching the data in the nmhandle, allocate the hot-path
ns_client_t objects from per-thread clientmgr memory context and just
assign it to the isc_nmhandle_t via isc_nmhandle_set().
Previously, the unreachable code paths would have to be tagged with:
INSIST(0);
ISC_UNREACHABLE();
There was also older parts of the code that used comment annotation:
/* NOTREACHED */
Unify the handling of unreachable code paths to just use:
UNREACHABLE();
The UNREACHABLE() macro now asserts when reached and also uses
__builtin_unreachable(); when such builtin is available in the compiler.
Previously, there was a single per-socket write timer that would get
restarted for every new write. This turned out to be insufficient
because the other side could keep reseting the timer, and never reading
back the responses.
Change the single write timer to per-send timer which would in turn
reset the TCP connection on the first send timeout.
The C17 standard deprecated ATOMIC_VAR_INIT() macro (see [1]). Follow
the suite and remove the ATOMIC_VAR_INIT() usage in favor of simple
assignment of the value as this is what all supported stdatomic.h
implementations do anyway:
* MacOSX.plaform: #define ATOMIC_VAR_INIT(__v) {__v}
* Gcc stdatomic.h: #define ATOMIC_VAR_INIT(VALUE) (VALUE)
1. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1138r0.pdf
Previously the socket code would set the TCPv6 maximum segment size to
minimum value to prevent IP fragmentation for TCP. This was not yet
implemented for the network manager.
Implement network manager functions to set and use minimum MTU socket
option and set the TCP_MAXSEG socket option for both IPv4 and IPv6 and
use those to clamp the TCP maximum segment size for TCP, TCPDNS and
TLSDNS layers in the network manager to 1220 bytes, that is 1280 (IPv6
minimum link MTU) minus 40 (IPv6 fixed header) minus 20 (TCP fixed
header)
We already rely on a similar value for UDP to prevent IP fragmentation
and it make sense to use the same value for IPv4 and IPv6 because the
modern networks are required to support IPv6 packet sizes. If there's
need for small TCP segment values, the MTU on the interfaces needs to be
properly configured.
The IPV6_USE_MIN_MTU socket option directs the IP layer to limit the
IPv6 packet size to the minimum required supported MTU from the base
IPv6 specification, i.e. 1280 bytes. Many implementations of TCP
running over IPv6 neglect to check the IPV6_USE_MIN_MTU value when
performing MSS negotiation and when constructing a TCP segment despite
MSS being defined to be the MTU less the IP and TCP header sizes (60
bytes for IPv6). This leads to oversized IPv6 packets being sent
resulting in unintended Path Maximum Transport Unit Discovery (PMTUD)
being performed and to fragmented IPv6 packets being sent.
Add and use a function to set socket option to limit the MTU on IPv6
sockets to the minimum MTU (1280) both for UDP and TCP.
When the TCP, TCPDNS or TLSDNS connection times out, the isc__nm_uvreq_t
would be pushed into sock->inactivereqs before the uv_tcp_connect()
callback finishes. Because the isc__nmsocket_t keeps the list of
inactive isc__nm_uvreq_t, this would cause use-after-free only when the
sock->inactivereqs is full (which could never happen because the failure
happens in connection timeout callback) or when the sock->inactivereqs
mechanism is completely removed (f.e. when running under Address or
Thread Sanitizer).
Delay isc__nm_uvreq_t deallocation to the connection callback and only
signal the connection callback should be called by shutting down the
libuv socket from the connection timeout callback.
When the outgoing TCP write buffers are full because the other party is
not reading the data, the uv_write() could wait indefinitely on the
uv_loop and never calling the callback. Add a new write timer that uses
the `tcp-idle-timeout` value to interrupt the TCP connection when we are
not able to send data for defined period of time.
Before adding the write timer, we have to remove the generic sock->timer
to sock->read_timer. We don't touch the function names to limit the
impact of the refactoring.
When isc_quota_attach_cb() API returns ISC_R_QUOTA (meaning hard quota
was reached) the accept_connection() would return without logging a
message about quota reached.
Change the connection callback to log the quota reached message.
Some operating systems (OpenBSD and DragonFly BSD) don't restrict the
IPv6 sockets to sending and receiving IPv6 packets only. Explicitly
enable the IPV6_V6ONLY socket option on the IPv6 sockets to prevent
failures from using the IPv4-mapped IPv6 address.
