2
0
mirror of https://github.com/openvswitch/ovs synced 2025-08-22 09:58:01 +00:00
ovs/lib/ovs-thread.h

529 lines
24 KiB
C
Raw Normal View History

/*
* Copyright (c) 2013, 2014 Nicira, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at:
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef OVS_THREAD_H
#define OVS_THREAD_H 1
#include <pthread.h>
#include <stddef.h>
#include <sys/types.h>
#include "ovs-atomic.h"
#include "ovs-rcu.h"
#include "openvswitch/thread.h"
#include "util.h"
struct seq;
/* Poll-block()-able barrier similar to pthread_barrier_t. */
struct ovs_barrier_impl;
struct ovs_barrier {
OVSRCU_TYPE(struct ovs_barrier_impl *) impl;
};
/* Wrappers for pthread_mutexattr_*() that abort the process on any error. */
void xpthread_mutexattr_init(pthread_mutexattr_t *);
void xpthread_mutexattr_destroy(pthread_mutexattr_t *);
void xpthread_mutexattr_settype(pthread_mutexattr_t *, int type);
void xpthread_mutexattr_gettype(pthread_mutexattr_t *, int *typep);
/* Read-write lock.
*
* An ovs_rwlock does not support recursive readers, because POSIX allows
* taking the reader lock recursively to deadlock when a thread is waiting on
* the write-lock. (NetBSD does deadlock.) glibc rwlocks in their default
* configuration do not deadlock, but ovs_rwlock_init() initializes rwlocks as
* non-recursive (which will deadlock) for two reasons:
*
* - glibc only provides fairness to writers in this mode.
*
* - It's better to find bugs in the primary Open vSwitch target rather
* than exposing them only to porters. */
struct OVS_LOCKABLE ovs_rwlock {
pthread_rwlock_t lock;
const char *where; /* NULL if and only if uninitialized. */
};
/* Initializer. */
#ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
#define OVS_RWLOCK_INITIALIZER \
{ PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP, "<unlocked>" }
#else
#define OVS_RWLOCK_INITIALIZER { PTHREAD_RWLOCK_INITIALIZER, "<unlocked>" }
#endif
/* ovs_rwlock functions analogous to pthread_rwlock_*() functions.
*
* Most of these functions abort the process with an error message on any
* error. The "trylock" functions are exception: they pass through a 0 or
* EBUSY return value to the caller and abort on any other error. */
void ovs_rwlock_init(const struct ovs_rwlock *);
void ovs_rwlock_destroy(const struct ovs_rwlock *);
void ovs_rwlock_unlock(const struct ovs_rwlock *rwlock) OVS_RELEASES(rwlock);
/* Wrappers for pthread_rwlockattr_*() that abort the process on any error. */
void xpthread_rwlockattr_init(pthread_rwlockattr_t *);
void xpthread_rwlockattr_destroy(pthread_rwlockattr_t *);
#ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
void xpthread_rwlockattr_setkind_np(pthread_rwlockattr_t *, int kind);
#endif
void ovs_rwlock_wrlock_at(const struct ovs_rwlock *rwlock, const char *where)
OVS_ACQ_WRLOCK(rwlock);
#define ovs_rwlock_wrlock(rwlock) \
ovs_rwlock_wrlock_at(rwlock, OVS_SOURCE_LOCATOR)
int ovs_rwlock_trywrlock_at(const struct ovs_rwlock *rwlock, const char *where)
OVS_TRY_WRLOCK(0, rwlock);
#define ovs_rwlock_trywrlock(rwlock) \
ovs_rwlock_trywrlock_at(rwlock, OVS_SOURCE_LOCATOR)
void ovs_rwlock_rdlock_at(const struct ovs_rwlock *rwlock, const char *where)
OVS_ACQ_RDLOCK(rwlock);
#define ovs_rwlock_rdlock(rwlock) \
ovs_rwlock_rdlock_at(rwlock, OVS_SOURCE_LOCATOR)
int ovs_rwlock_tryrdlock_at(const struct ovs_rwlock *rwlock, const char *where)
OVS_TRY_RDLOCK(0, rwlock);
#define ovs_rwlock_tryrdlock(rwlock) \
ovs_rwlock_tryrdlock_at(rwlock, OVS_SOURCE_LOCATOR)
/* ovs_barrier functions analogous to pthread_barrier_*() functions. */
void ovs_barrier_init(struct ovs_barrier *, uint32_t count);
void ovs_barrier_destroy(struct ovs_barrier *);
void ovs_barrier_block(struct ovs_barrier *);
/* Wrappers for xpthread_cond_*() that abort the process on any error.
*
* Use ovs_mutex_cond_wait() to wait for a condition. */
void xpthread_cond_init(pthread_cond_t *, pthread_condattr_t *);
void xpthread_cond_destroy(pthread_cond_t *);
void xpthread_cond_signal(pthread_cond_t *);
void xpthread_cond_broadcast(pthread_cond_t *);
void xpthread_key_create(pthread_key_t *, void (*destructor)(void *));
void xpthread_key_delete(pthread_key_t);
void xpthread_setspecific(pthread_key_t, const void *);
#ifndef _WIN32
void xpthread_sigmask(int, const sigset_t *, sigset_t *);
#endif
pthread_t ovs_thread_create(const char *name, void *(*)(void *), void *);
void xpthread_join(pthread_t, void **);
/* Per-thread data.
*
*
* Standard Forms
* ==============
*
* Multiple forms of standard per-thread data exist, each with its own pluses
* and minuses. In general, if one of these forms is appropriate, then it's a
* good idea to use it:
*
* - POSIX per-thread data via pthread_key_t is portable to any pthreads
* implementation, and allows a destructor function to be defined. It
* only (directly) supports per-thread pointers, which are always
* initialized to NULL. It requires once-only allocation of a
* pthread_key_t value. It is relatively slow. Typically few
* "pthread_key_t"s are available (POSIX requires only at least 128,
* glibc supplies only 1024).
*
* - The thread_local feature newly defined in C11 <threads.h> works with
* any data type and initializer, and it is fast. thread_local does not
* require once-only initialization like pthread_key_t. C11 does not
* define what happens if one attempts to access a thread_local object
* from a thread other than the one to which that object belongs. There
* is no provision to call a user-specified destructor when a thread
* ends. Typical implementations allow for an arbitrary amount of
* thread_local storage, but statically allocated only.
*
* - The __thread keyword is a GCC extension similar to thread_local but
* with a longer history. __thread is not portable to every GCC version
* or environment. __thread does not restrict the use of a thread-local
* object outside its own thread.
*
* Here's a handy summary:
*
* pthread_key_t thread_local __thread
* ------------- ------------ -------------
* portability high low medium
* speed low high high
* supports destructors? yes no no
* needs key allocation? yes no no
* arbitrary initializer? no yes yes
* cross-thread access? yes no yes
* amount available? few arbitrary arbitrary
* dynamically allocated? yes no no
*
*
* Extensions
* ==========
*
* OVS provides some extensions and wrappers:
*
* - In a situation where the performance of thread_local or __thread is
* desirable, but portability is required, DEFINE_STATIC_PER_THREAD_DATA
* and DECLARE_EXTERN_PER_THREAD_DATA/DEFINE_EXTERN_PER_THREAD_DATA may
* be appropriate (see below).
*
* - DEFINE_PER_THREAD_MALLOCED_DATA can be convenient for simple
* per-thread malloc()'d buffers.
*
* - struct ovs_tsd provides an alternative to pthread_key_t that isn't
* limited to a small number of keys.
*/
/* For static data, use this macro in a source file:
*
* DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, INITIALIZER).
*
* For global data, "declare" the data in the header and "define" it in
* the source file, with:
*
* DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME).
* DEFINE_EXTERN_PER_THREAD_DATA(NAME, INITIALIZER).
*
* One should prefer to use POSIX per-thread data, via pthread_key_t, when its
* performance is acceptable, because of its portability (see the table above).
* This macro is an alternatives that takes advantage of thread_local (and
* __thread), for its performance, when it is available, and falls back to
* POSIX per-thread data otherwise.
*
* Defines per-thread variable NAME with the given TYPE, initialized to
* INITIALIZER (which must be valid as an initializer for a variable with
* static lifetime).
*
* The public interface to the variable is:
*
* TYPE *NAME_get(void)
* TYPE *NAME_get_unsafe(void)
*
* Returns the address of this thread's instance of NAME.
*
* Use NAME_get() in a context where this might be the first use of the
* per-thread variable in the program. Use NAME_get_unsafe(), which
* avoids a conditional test and is thus slightly faster, in a context
* where one knows that NAME_get() has already been called previously.
*
* There is no "NAME_set()" (or "NAME_set_unsafe()") function. To set the
* value of the per-thread variable, dereference the pointer returned by
* TYPE_get() or TYPE_get_unsafe(), e.g. *TYPE_get() = 0.
*/
#if HAVE_THREAD_LOCAL || HAVE___THREAD
#if HAVE_THREAD_LOCAL
#include <threads.h>
#elif HAVE___THREAD
#define thread_local __thread
#else
#error
#endif
#define DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, ...) \
typedef TYPE NAME##_type; \
\
static NAME##_type * \
NAME##_get_unsafe(void) \
{ \
static thread_local NAME##_type var = __VA_ARGS__; \
return &var; \
} \
\
static NAME##_type * \
NAME##_get(void) \
{ \
return NAME##_get_unsafe(); \
}
#define DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME) \
typedef TYPE NAME##_type; \
extern thread_local NAME##_type NAME##_var; \
\
static inline NAME##_type * \
NAME##_get_unsafe(void) \
{ \
return (NAME##_type *)&NAME##_var; \
} \
\
static inline NAME##_type * \
NAME##_get(void) \
{ \
return NAME##_get_unsafe(); \
}
#define DEFINE_EXTERN_PER_THREAD_DATA(NAME, ...) \
thread_local NAME##_type NAME##_var = __VA_ARGS__;
#else /* no C implementation support for thread-local storage */
#define DEFINE_STATIC_PER_THREAD_DATA(TYPE, NAME, ...) \
typedef TYPE NAME##_type; \
static pthread_key_t NAME##_key; \
\
static NAME##_type * \
NAME##_get_unsafe(void) \
{ \
return pthread_getspecific(NAME##_key); \
} \
\
static void \
NAME##_once_init(void) \
{ \
if (pthread_key_create(&NAME##_key, free)) { \
abort(); \
} \
} \
\
static NAME##_type * \
NAME##_get(void) \
{ \
static pthread_once_t once = PTHREAD_ONCE_INIT; \
NAME##_type *value; \
\
pthread_once(&once, NAME##_once_init); \
value = NAME##_get_unsafe(); \
if (!value) { \
static const NAME##_type initial_value = __VA_ARGS__; \
\
value = xmalloc__(sizeof *value); \
*value = initial_value; \
xpthread_setspecific(NAME##_key, value); \
} \
return value; \
}
#define DECLARE_EXTERN_PER_THREAD_DATA(TYPE, NAME) \
typedef TYPE NAME##_type; \
static pthread_key_t NAME##_key; \
\
static inline NAME##_type * \
NAME##_get_unsafe(void) \
{ \
return (NAME##_type *)pthread_getspecific(NAME##_key); \
} \
\
NAME##_type *NAME##_get(void);
#define DEFINE_EXTERN_PER_THREAD_DATA(NAME, ...) \
static void \
NAME##_once_init(void) \
{ \
if (pthread_key_create(&NAME##_key, free)) { \
abort(); \
} \
} \
\
NAME##_type * \
NAME##_get(void) \
{ \
static pthread_once_t once = PTHREAD_ONCE_INIT; \
NAME##_type *value; \
\
pthread_once(&once, NAME##_once_init); \
value = NAME##_get_unsafe(); \
if (!value) { \
static const NAME##_type initial_value = __VA_ARGS__; \
\
value = xmalloc__(sizeof *value); \
*value = initial_value; \
xpthread_setspecific(NAME##_key, value); \
} \
return value; \
}
#endif
/* DEFINE_PER_THREAD_MALLOCED_DATA(TYPE, NAME).
*
* This is a simple wrapper around POSIX per-thread data primitives. It
* defines per-thread variable NAME with the given TYPE, which must be a
* pointer type. In each thread, the per-thread variable is initialized to
* NULL. When a thread terminates, the variable is freed with free().
*
* The public interface to the variable is:
*
* TYPE NAME_get(void)
* TYPE NAME_get_unsafe(void)
*
* Returns the value of per-thread variable NAME in this thread.
*
* Use NAME_get() in a context where this might be the first use of the
* per-thread variable in the program. Use NAME_get_unsafe(), which
* avoids a conditional test and is thus slightly faster, in a context
* where one knows that NAME_get() has already been called previously.
*
* TYPE NAME_set(TYPE new_value)
* TYPE NAME_set_unsafe(TYPE new_value)
*
* Sets the value of per-thread variable NAME to 'new_value' in this
* thread, and returns its previous value.
*
* Use NAME_set() in a context where this might be the first use of the
* per-thread variable in the program. Use NAME_set_unsafe(), which
* avoids a conditional test and is thus slightly faster, in a context
* where one knows that NAME_set() has already been called previously.
*/
#define DEFINE_PER_THREAD_MALLOCED_DATA(TYPE, NAME) \
static pthread_key_t NAME##_key; \
\
static void \
NAME##_once_init(void) \
{ \
if (pthread_key_create(&NAME##_key, free)) { \
abort(); \
} \
} \
\
static void \
NAME##_init(void) \
{ \
static pthread_once_t once = PTHREAD_ONCE_INIT; \
pthread_once(&once, NAME##_once_init); \
} \
\
static TYPE \
NAME##_get_unsafe(void) \
{ \
return pthread_getspecific(NAME##_key); \
} \
\
static OVS_UNUSED TYPE \
NAME##_get(void) \
{ \
NAME##_init(); \
return NAME##_get_unsafe(); \
} \
\
static TYPE \
NAME##_set_unsafe(TYPE value) \
{ \
TYPE old_value = NAME##_get_unsafe(); \
xpthread_setspecific(NAME##_key, value); \
return old_value; \
} \
\
static OVS_UNUSED TYPE \
NAME##_set(TYPE value) \
{ \
NAME##_init(); \
return NAME##_set_unsafe(value); \
}
/* Dynamically allocated thread-specific data with lots of slots.
*
* pthread_key_t can provide as few as 128 pieces of thread-specific data (even
* glibc is limited to 1,024). Thus, one must be careful to allocate only a
* few keys globally. One cannot, for example, allocate a key for every
* instance of a data structure if there might be an arbitrary number of those
* data structures.
*
* This API is similar to the pthread one (simply search and replace pthread_
* by ovsthread_) but it a much larger limit that can be raised if necessary
* (by recompiling). Thus, one may more freely use this form of
* thread-specific data.
*
* ovsthread_key_t also differs from pthread_key_t in the following ways:
*
* - Destructors must not access thread-specific data (via ovsthread_key).
*
* - The pthread_key_t API allows concurrently exiting threads to start
* executing the destructor after pthread_key_delete() returns. The
* ovsthread_key_t API guarantees that, when ovsthread_key_delete()
* returns, all destructors have returned and no new ones will start
* execution.
*/
typedef struct ovsthread_key *ovsthread_key_t;
void ovsthread_key_create(ovsthread_key_t *, void (*destructor)(void *));
void ovsthread_key_delete(ovsthread_key_t);
void ovsthread_setspecific(ovsthread_key_t, const void *);
void *ovsthread_getspecific(ovsthread_key_t);
/* Thread ID.
*
* pthread_t isn't so nice for some purposes. Its size and representation are
* implementation dependent, which means that there is no way to hash it.
* This thread ID avoids the problem.
*/
ovs-thread: Fix thread id for threads not started with ovs_thread_create() When ping-pong'in a live VM migration between two machines running OVS-DPDK every now and then the ping misses would increase dramatically. For example: Acked-by: Ilya Maximets <i.maximets@samsung.com> ===========Stream Rate: 3Mpps=========== No Stream_Rate Downtime Totaltime Ping_Loss Moongen_Loss 0 3Mpps 128 13974 115 7168374 1 3Mpps 145 13620 17 1169770 2 3Mpps 140 14499 116 7141175 3 3Mpps 142 13358 16 1150606 4 3Mpps 136 14004 16 1124020 5 3Mpps 139 15494 214 13170452 6 3Mpps 136 15610 217 13282413 7 3Mpps 146 13194 17 1167512 8 3Mpps 148 12871 16 1162655 9 3Mpps 137 15615 214 13170656 I identified this issue being introduced in OVS commit, f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") and more specific due to DPDK commit, af1475918124 ("vhost: introduce API to start a specific driver"). The combined changes no longer have OVS start the vhost socket polling thread at startup, but DPDK will do it on its own when the first vhost client is started. Figuring out the reason why this happens kept me puzzled for quite some time... What happens is that the callbacks called from the vhost thread are calling ovsrcu_synchronize() as part of destroy_device(). This will end-up calling seq_wait__(). By default, all created threads outside of OVS will get thread id 0, which is equal to the main ovs thread. So for example in the seq_wait__() function above if the main thread is waiting already we won't add ourselves as a waiter. The fix below assigns OVSTHREAD_ID_UNSET to none OVS created threads, which will get updated to a valid ID on the first call to ovsthread_id_self(). Signed-off-by: Eelco Chaudron <echaudro@redhat.com> Fixes: f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") Acked-by: Ilya Maximets <i.maximets@samsung.com> Signed-off-by: Ian Stokes <ian.stokes@intel.com>
2018-06-04 10:07:36 +02:00
#define OVSTHREAD_ID_UNSET UINT_MAX
DECLARE_EXTERN_PER_THREAD_DATA(unsigned int, ovsthread_id);
ovs-thread: Fix thread id for threads not started with ovs_thread_create() When ping-pong'in a live VM migration between two machines running OVS-DPDK every now and then the ping misses would increase dramatically. For example: Acked-by: Ilya Maximets <i.maximets@samsung.com> ===========Stream Rate: 3Mpps=========== No Stream_Rate Downtime Totaltime Ping_Loss Moongen_Loss 0 3Mpps 128 13974 115 7168374 1 3Mpps 145 13620 17 1169770 2 3Mpps 140 14499 116 7141175 3 3Mpps 142 13358 16 1150606 4 3Mpps 136 14004 16 1124020 5 3Mpps 139 15494 214 13170452 6 3Mpps 136 15610 217 13282413 7 3Mpps 146 13194 17 1167512 8 3Mpps 148 12871 16 1162655 9 3Mpps 137 15615 214 13170656 I identified this issue being introduced in OVS commit, f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") and more specific due to DPDK commit, af1475918124 ("vhost: introduce API to start a specific driver"). The combined changes no longer have OVS start the vhost socket polling thread at startup, but DPDK will do it on its own when the first vhost client is started. Figuring out the reason why this happens kept me puzzled for quite some time... What happens is that the callbacks called from the vhost thread are calling ovsrcu_synchronize() as part of destroy_device(). This will end-up calling seq_wait__(). By default, all created threads outside of OVS will get thread id 0, which is equal to the main ovs thread. So for example in the seq_wait__() function above if the main thread is waiting already we won't add ourselves as a waiter. The fix below assigns OVSTHREAD_ID_UNSET to none OVS created threads, which will get updated to a valid ID on the first call to ovsthread_id_self(). Signed-off-by: Eelco Chaudron <echaudro@redhat.com> Fixes: f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") Acked-by: Ilya Maximets <i.maximets@samsung.com> Signed-off-by: Ian Stokes <ian.stokes@intel.com>
2018-06-04 10:07:36 +02:00
/* Initializes the unique per thread identifier */
unsigned int ovsthread_id_init(void);
/* Returns a per-thread identifier unique within the lifetime of the
* process. */
static inline unsigned int
ovsthread_id_self(void)
{
ovs-thread: Fix thread id for threads not started with ovs_thread_create() When ping-pong'in a live VM migration between two machines running OVS-DPDK every now and then the ping misses would increase dramatically. For example: Acked-by: Ilya Maximets <i.maximets@samsung.com> ===========Stream Rate: 3Mpps=========== No Stream_Rate Downtime Totaltime Ping_Loss Moongen_Loss 0 3Mpps 128 13974 115 7168374 1 3Mpps 145 13620 17 1169770 2 3Mpps 140 14499 116 7141175 3 3Mpps 142 13358 16 1150606 4 3Mpps 136 14004 16 1124020 5 3Mpps 139 15494 214 13170452 6 3Mpps 136 15610 217 13282413 7 3Mpps 146 13194 17 1167512 8 3Mpps 148 12871 16 1162655 9 3Mpps 137 15615 214 13170656 I identified this issue being introduced in OVS commit, f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") and more specific due to DPDK commit, af1475918124 ("vhost: introduce API to start a specific driver"). The combined changes no longer have OVS start the vhost socket polling thread at startup, but DPDK will do it on its own when the first vhost client is started. Figuring out the reason why this happens kept me puzzled for quite some time... What happens is that the callbacks called from the vhost thread are calling ovsrcu_synchronize() as part of destroy_device(). This will end-up calling seq_wait__(). By default, all created threads outside of OVS will get thread id 0, which is equal to the main ovs thread. So for example in the seq_wait__() function above if the main thread is waiting already we won't add ourselves as a waiter. The fix below assigns OVSTHREAD_ID_UNSET to none OVS created threads, which will get updated to a valid ID on the first call to ovsthread_id_self(). Signed-off-by: Eelco Chaudron <echaudro@redhat.com> Fixes: f3e7ec254738 ("Update relevant artifacts to add support for DPDK 17.05.1.") Acked-by: Ilya Maximets <i.maximets@samsung.com> Signed-off-by: Ian Stokes <ian.stokes@intel.com>
2018-06-04 10:07:36 +02:00
unsigned int id = *ovsthread_id_get();
if (OVS_UNLIKELY(id == OVSTHREAD_ID_UNSET)) {
id = ovsthread_id_init();
}
return id;
}
/* Simulated global counter.
*
* Incrementing such a counter is meant to be cheaper than incrementing a
* global counter protected by a lock. It is probably more expensive than
* incrementing a truly thread-local variable, but such a variable has no
* straightforward way to get the sum.
*
*
* Thread-safety
* =============
*
* Fully thread-safe. */
struct ovsthread_stats {
struct ovs_mutex mutex;
void *volatile buckets[16];
};
void ovsthread_stats_init(struct ovsthread_stats *);
void ovsthread_stats_destroy(struct ovsthread_stats *);
void *ovsthread_stats_bucket_get(struct ovsthread_stats *,
void *(*new_bucket)(void));
#define OVSTHREAD_STATS_FOR_EACH_BUCKET(BUCKET, IDX, STATS) \
for ((IDX) = ovs_thread_stats_next_bucket(STATS, 0); \
((IDX) < ARRAY_SIZE((STATS)->buckets) \
? ((BUCKET) = (STATS)->buckets[IDX], true) \
: false); \
(IDX) = ovs_thread_stats_next_bucket(STATS, (IDX) + 1))
size_t ovs_thread_stats_next_bucket(const struct ovsthread_stats *, size_t);
bool single_threaded(void);
void assert_single_threaded_at(const char *where);
#define assert_single_threaded() assert_single_threaded_at(OVS_SOURCE_LOCATOR)
#ifndef _WIN32
pid_t xfork_at(const char *where);
#define xfork() xfork_at(OVS_SOURCE_LOCATOR)
#endif
void forbid_forking(const char *reason);
bool may_fork(void);
/* Useful functions related to threading. */
int count_cpu_cores(void);
handlers: Create additional handler threads when using CPU isolation. Additional threads are required to service upcalls when we have CPU isolation (in per-cpu dispatch mode). The reason additional threads are required is because it creates a more fair distribution. With more threads we decrease the load of each thread as more threads would decrease the number of cores each threads is assigned. Adding additional threads also increases the chance OVS utilizes all cores available to use. Some RPS schemas might make some handler threads get all the workload while others get no workload. This tends to happen when the handler thread count is low. An example would be an RPS that sends traffic on all even cores on a system with only the lower half of the cores available for OVS to use. In this example we have as many handlers threads as there are available cores. In this case 50% of the handler threads get all the workload while the other 50% get no workload. Not only that, but OVS is only utilizing half of the cores that it can use. This is the worst case scenario. The ideal scenario is to have as many threads as there are cores - in this case we guarantee that all cores OVS can use are utilized But, adding as many threads are there are cores could have a performance hit when the number of active cores (which all threads have to share) is very low. For this reason we avoid creating as many threads as there are cores and instead meet somewhere in the middle. The formula used to calculate the number of handler threads to create is as follows: handlers_n = min(next_prime(active_cores+1), total_cores) Assume default behavior when total_cores <= 2, that is do not create additional threads when we have less than 2 total cores on the system Fixes: b1e517bd2f81 ("dpif-netlink: Introduce per-cpu upcall dispatch.") Signed-off-by: Michael Santana <msantana@redhat.com> Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
2022-08-09 03:18:14 -04:00
int count_total_cores(void);
bool thread_is_pmd(void);
#endif /* ovs-thread.h */