2 * Copyright (c) 2013, 2014 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "ovs-thread.h"
25 #include "poll-loop.h"
26 #include "socket-util.h"
30 /* Omit the definitions in this file because they are somewhat difficult to
31 * write without prompting "sparse" complaints, without ugliness or
32 * cut-and-paste. Since "sparse" is just a checker, not a compiler, it
33 * doesn't matter that we don't define them. */
37 VLOG_DEFINE_THIS_MODULE(ovs_thread);
39 /* If there is a reason that we cannot fork anymore (unless the fork will be
40 * immediately followed by an exec), then this points to a string that
42 static const char *must_not_fork;
44 /* True if we created any threads beyond the main initial thread. */
45 static bool multithreaded;
47 #define LOCK_FUNCTION(TYPE, FUN) \
49 ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
51 OVS_NO_THREAD_SAFETY_ANALYSIS \
53 struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
54 int error = pthread_##TYPE##_##FUN(&l->lock); \
55 if (OVS_UNLIKELY(error)) { \
56 ovs_abort(error, "pthread_%s_%s failed", #TYPE, #FUN); \
60 LOCK_FUNCTION(mutex, lock);
61 LOCK_FUNCTION(rwlock, rdlock);
62 LOCK_FUNCTION(rwlock, wrlock);
64 #define TRY_LOCK_FUNCTION(TYPE, FUN) \
66 ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
68 OVS_NO_THREAD_SAFETY_ANALYSIS \
70 struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
71 int error = pthread_##TYPE##_##FUN(&l->lock); \
72 if (OVS_UNLIKELY(error) && error != EBUSY) { \
73 ovs_abort(error, "pthread_%s_%s failed", #TYPE, #FUN); \
80 TRY_LOCK_FUNCTION(mutex, trylock);
81 TRY_LOCK_FUNCTION(rwlock, tryrdlock);
82 TRY_LOCK_FUNCTION(rwlock, trywrlock);
84 #define UNLOCK_FUNCTION(TYPE, FUN) \
86 ovs_##TYPE##_##FUN(const struct ovs_##TYPE *l_) \
87 OVS_NO_THREAD_SAFETY_ANALYSIS \
89 struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
92 error = pthread_##TYPE##_##FUN(&l->lock); \
93 if (OVS_UNLIKELY(error)) { \
94 ovs_abort(error, "pthread_%s_%sfailed", #TYPE, #FUN); \
97 UNLOCK_FUNCTION(mutex, unlock);
98 UNLOCK_FUNCTION(mutex, destroy);
99 UNLOCK_FUNCTION(rwlock, unlock);
100 UNLOCK_FUNCTION(rwlock, destroy);
102 #define XPTHREAD_FUNC1(FUNCTION, PARAM1) \
104 x##FUNCTION(PARAM1 arg1) \
106 int error = FUNCTION(arg1); \
107 if (OVS_UNLIKELY(error)) { \
108 ovs_abort(error, "%s failed", #FUNCTION); \
111 #define XPTHREAD_FUNC2(FUNCTION, PARAM1, PARAM2) \
113 x##FUNCTION(PARAM1 arg1, PARAM2 arg2) \
115 int error = FUNCTION(arg1, arg2); \
116 if (OVS_UNLIKELY(error)) { \
117 ovs_abort(error, "%s failed", #FUNCTION); \
121 XPTHREAD_FUNC1(pthread_mutex_lock, pthread_mutex_t *);
122 XPTHREAD_FUNC1(pthread_mutex_unlock, pthread_mutex_t *);
123 XPTHREAD_FUNC1(pthread_mutexattr_init, pthread_mutexattr_t *);
124 XPTHREAD_FUNC1(pthread_mutexattr_destroy, pthread_mutexattr_t *);
125 XPTHREAD_FUNC2(pthread_mutexattr_settype, pthread_mutexattr_t *, int);
126 XPTHREAD_FUNC2(pthread_mutexattr_gettype, pthread_mutexattr_t *, int *);
128 XPTHREAD_FUNC2(pthread_cond_init, pthread_cond_t *, pthread_condattr_t *);
129 XPTHREAD_FUNC1(pthread_cond_destroy, pthread_cond_t *);
130 XPTHREAD_FUNC1(pthread_cond_signal, pthread_cond_t *);
131 XPTHREAD_FUNC1(pthread_cond_broadcast, pthread_cond_t *);
133 XPTHREAD_FUNC2(pthread_join, pthread_t, void **);
135 typedef void destructor_func(void *);
136 XPTHREAD_FUNC2(pthread_key_create, pthread_key_t *, destructor_func *);
137 XPTHREAD_FUNC1(pthread_key_delete, pthread_key_t);
138 XPTHREAD_FUNC2(pthread_setspecific, pthread_key_t, const void *);
141 ovs_mutex_init__(const struct ovs_mutex *l_, int type)
143 struct ovs_mutex *l = CONST_CAST(struct ovs_mutex *, l_);
144 pthread_mutexattr_t attr;
148 xpthread_mutexattr_init(&attr);
149 xpthread_mutexattr_settype(&attr, type);
150 error = pthread_mutex_init(&l->lock, &attr);
151 if (OVS_UNLIKELY(error)) {
152 ovs_abort(error, "pthread_mutex_init failed");
154 xpthread_mutexattr_destroy(&attr);
157 /* Initializes 'mutex' as a normal (non-recursive) mutex. */
159 ovs_mutex_init(const struct ovs_mutex *mutex)
161 ovs_mutex_init__(mutex, PTHREAD_MUTEX_ERRORCHECK);
164 /* Initializes 'mutex' as a recursive mutex. */
166 ovs_mutex_init_recursive(const struct ovs_mutex *mutex)
168 ovs_mutex_init__(mutex, PTHREAD_MUTEX_RECURSIVE);
172 ovs_rwlock_init(const struct ovs_rwlock *l_)
174 struct ovs_rwlock *l = CONST_CAST(struct ovs_rwlock *, l_);
178 error = pthread_rwlock_init(&l->lock, NULL);
179 if (OVS_UNLIKELY(error)) {
180 ovs_abort(error, "pthread_rwlock_init failed");
185 ovs_mutex_cond_wait(pthread_cond_t *cond, const struct ovs_mutex *mutex_)
187 struct ovs_mutex *mutex = CONST_CAST(struct ovs_mutex *, mutex_);
188 int error = pthread_cond_wait(cond, &mutex->lock);
189 if (OVS_UNLIKELY(error)) {
190 ovs_abort(error, "pthread_cond_wait failed");
194 DEFINE_EXTERN_PER_THREAD_DATA(ovsthread_id, 0);
196 struct ovsthread_aux {
197 void *(*start)(void *);
202 ovsthread_wrapper(void *aux_)
204 static atomic_uint next_id = ATOMIC_VAR_INIT(1);
206 struct ovsthread_aux *auxp = aux_;
207 struct ovsthread_aux aux;
210 atomic_add(&next_id, 1, &id);
211 *ovsthread_id_get() = id;
216 return aux.start(aux.arg);
220 xpthread_create(pthread_t *threadp, pthread_attr_t *attr,
221 void *(*start)(void *), void *arg)
223 struct ovsthread_aux *aux;
227 forbid_forking("multiple threads exist");
228 multithreaded = true;
230 aux = xmalloc(sizeof *aux);
234 error = pthread_create(threadp ? threadp : &thread, attr,
235 ovsthread_wrapper, aux);
237 ovs_abort(error, "pthread_create failed");
242 ovsthread_once_start__(struct ovsthread_once *once)
244 ovs_mutex_lock(&once->mutex);
245 if (!ovsthread_once_is_done__(once)) {
248 ovs_mutex_unlock(&once->mutex);
253 ovsthread_once_done(struct ovsthread_once *once)
255 atomic_store(&once->done, true);
256 ovs_mutex_unlock(&once->mutex);
259 /* Asserts that the process has not yet created any threads (beyond the initial
262 * ('where' is used in logging. Commonly one would use
263 * assert_single_threaded() to automatically provide the caller's source file
264 * and line number for 'where'.) */
266 assert_single_threaded_at(const char *where)
269 VLOG_FATAL("%s: attempted operation not allowed when multithreaded",
274 /* Forks the current process (checking that this is allowed). Aborts with
275 * VLOG_FATAL if fork() returns an error, and otherwise returns the value
276 * returned by fork().
278 * ('where' is used in logging. Commonly one would use xfork() to
279 * automatically provide the caller's source file and line number for
282 xfork_at(const char *where)
287 VLOG_FATAL("%s: attempted to fork but forking not allowed (%s)",
288 where, must_not_fork);
293 VLOG_FATAL("%s: fork failed (%s)", where, ovs_strerror(errno));
298 /* Notes that the process must not call fork() from now on, for the specified
299 * 'reason'. (The process may still fork() if it execs itself immediately
302 forbid_forking(const char *reason)
304 ovs_assert(reason != NULL);
305 must_not_fork = reason;
308 /* Returns true if the process is allowed to fork, false otherwise. */
312 return !must_not_fork;
315 /* ovsthread_counter.
317 * We implement the counter as an array of N_COUNTERS individual counters, each
318 * with its own lock. Each thread uses one of the counters chosen based on a
319 * hash of the thread's ID, the idea being that, statistically, different
320 * threads will tend to use different counters and therefore avoid
321 * interfering with each other.
323 * Undoubtedly, better implementations are possible. */
325 /* Basic counter structure. */
326 struct ovsthread_counter__ {
327 struct ovs_mutex mutex;
328 unsigned long long int value;
331 /* Pad the basic counter structure to 64 bytes to avoid cache line
333 struct ovsthread_counter {
334 struct ovsthread_counter__ c;
335 char pad[ROUND_UP(sizeof(struct ovsthread_counter__), 64)
336 - sizeof(struct ovsthread_counter__)];
339 #define N_COUNTERS 16
341 struct ovsthread_counter *
342 ovsthread_counter_create(void)
344 struct ovsthread_counter *c;
347 c = xmalloc(N_COUNTERS * sizeof *c);
348 for (i = 0; i < N_COUNTERS; i++) {
349 ovs_mutex_init(&c[i].c.mutex);
356 ovsthread_counter_destroy(struct ovsthread_counter *c)
361 for (i = 0; i < N_COUNTERS; i++) {
362 ovs_mutex_destroy(&c[i].c.mutex);
369 ovsthread_counter_inc(struct ovsthread_counter *c, unsigned long long int n)
371 c = &c[hash_int(ovsthread_id_self(), 0) % N_COUNTERS];
373 ovs_mutex_lock(&c->c.mutex);
375 ovs_mutex_unlock(&c->c.mutex);
378 unsigned long long int
379 ovsthread_counter_read(const struct ovsthread_counter *c)
381 unsigned long long int sum;
385 for (i = 0; i < N_COUNTERS; i++) {
386 ovs_mutex_lock(&c[i].c.mutex);
388 ovs_mutex_unlock(&c[i].c.mutex);
393 /* Parses /proc/cpuinfo for the total number of physical cores on this system
394 * across all CPU packages, not counting hyper-threads.
396 * Sets *n_cores to the total number of cores on this system, or 0 if the
397 * number cannot be determined. */
399 parse_cpuinfo(long int *n_cores)
401 static const char file_name[] = "/proc/cpuinfo";
403 uint64_t cpu = 0; /* Support up to 64 CPU packages on a single system. */
407 stream = fopen(file_name, "r");
409 VLOG_DBG("%s: open failed (%s)", file_name, ovs_strerror(errno));
413 while (fgets(line, sizeof line, stream)) {
416 /* Find the next CPU package. */
417 if (ovs_scan(line, "physical id%*[^:]: %u", &id)) {
419 VLOG_WARN("Counted over 64 CPU packages on this system. "
420 "Parsing %s for core count may be inaccurate.",
426 if (cpu & (1 << id)) {
427 /* We've already counted this package's cores. */
432 /* Find the number of cores for this package. */
433 while (fgets(line, sizeof line, stream)) {
436 if (ovs_scan(line, "cpu cores%*[^:]: %u", &count)) {
448 /* Returns the total number of cores on this system, or 0 if the number cannot
451 * Tries not to count hyper-threads, but may be inaccurate - particularly on
452 * platforms that do not provide /proc/cpuinfo, but also if /proc/cpuinfo is
453 * formatted different to the layout that parse_cpuinfo() expects. */
455 count_cpu_cores(void)
457 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
458 static long int n_cores;
460 if (ovsthread_once_start(&once)) {
461 parse_cpuinfo(&n_cores);
463 n_cores = sysconf(_SC_NPROCESSORS_ONLN);
465 ovsthread_once_done(&once);
468 return n_cores > 0 ? n_cores : 0;
475 #define MAX_KEYS (L1_SIZE * L2_SIZE)
477 /* A piece of thread-specific data. */
478 struct ovsthread_key {
479 struct list list_node; /* In 'inuse_keys' or 'free_keys'. */
480 void (*destructor)(void *); /* Called at thread exit. */
482 /* Indexes into the per-thread array in struct ovsthread_key_slots.
483 * This key's data is stored in p1[index / L2_SIZE][index % L2_SIZE]. */
487 /* Per-thread data structure. */
488 struct ovsthread_key_slots {
489 struct list list_node; /* In 'slots_list'. */
493 /* Contains "struct ovsthread_key_slots *". */
494 static pthread_key_t tsd_key;
496 /* Guards data structures below. */
497 static struct ovs_mutex key_mutex = OVS_MUTEX_INITIALIZER;
499 /* 'inuse_keys' holds "struct ovsthread_key"s that have been created and not
502 * 'free_keys' holds "struct ovsthread_key"s that have been deleted and are
503 * ready for reuse. (We keep them around only to be able to easily locate
506 * Together, 'inuse_keys' and 'free_keys' hold an ovsthread_key for every index
507 * from 0 to n_keys - 1, inclusive. */
508 static struct list inuse_keys OVS_GUARDED_BY(key_mutex)
509 = LIST_INITIALIZER(&inuse_keys);
510 static struct list free_keys OVS_GUARDED_BY(key_mutex)
511 = LIST_INITIALIZER(&free_keys);
512 static unsigned int n_keys OVS_GUARDED_BY(key_mutex);
514 /* All existing struct ovsthread_key_slots. */
515 static struct list slots_list OVS_GUARDED_BY(key_mutex)
516 = LIST_INITIALIZER(&slots_list);
519 clear_slot(struct ovsthread_key_slots *slots, unsigned int index)
521 void **p2 = slots->p1[index / L2_SIZE];
523 void **valuep = &p2[index % L2_SIZE];
524 void *value = *valuep;
533 ovsthread_key_destruct__(void *slots_)
535 struct ovsthread_key_slots *slots = slots_;
536 struct ovsthread_key *key;
540 ovs_mutex_lock(&key_mutex);
541 list_remove(&slots->list_node);
542 LIST_FOR_EACH (key, list_node, &inuse_keys) {
543 void *value = clear_slot(slots, key->index);
544 if (value && key->destructor) {
545 key->destructor(value);
549 ovs_mutex_unlock(&key_mutex);
551 for (i = 0; i < n / L2_SIZE; i++) {
557 /* Initializes '*keyp' as a thread-specific data key. The data items are
558 * initially null in all threads.
560 * If a thread exits with non-null data, then 'destructor', if nonnull, will be
561 * called passing the final data value as its argument. 'destructor' must not
562 * call any thread-specific data functions in this API.
564 * This function is similar to xpthread_key_create(). */
566 ovsthread_key_create(ovsthread_key_t *keyp, void (*destructor)(void *))
568 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
569 struct ovsthread_key *key;
571 if (ovsthread_once_start(&once)) {
572 xpthread_key_create(&tsd_key, ovsthread_key_destruct__);
573 ovsthread_once_done(&once);
576 ovs_mutex_lock(&key_mutex);
577 if (list_is_empty(&free_keys)) {
578 key = xmalloc(sizeof *key);
579 key->index = n_keys++;
580 if (key->index >= MAX_KEYS) {
584 key = CONTAINER_OF(list_pop_back(&free_keys),
585 struct ovsthread_key, list_node);
587 list_push_back(&inuse_keys, &key->list_node);
588 key->destructor = destructor;
589 ovs_mutex_unlock(&key_mutex);
594 /* Frees 'key'. The destructor supplied to ovsthread_key_create(), if any, is
597 * This function is similar to xpthread_key_delete(). */
599 ovsthread_key_delete(ovsthread_key_t key)
601 struct ovsthread_key_slots *slots;
603 ovs_mutex_lock(&key_mutex);
605 /* Move 'key' from 'inuse_keys' to 'free_keys'. */
606 list_remove(&key->list_node);
607 list_push_back(&free_keys, &key->list_node);
609 /* Clear this slot in all threads. */
610 LIST_FOR_EACH (slots, list_node, &slots_list) {
611 clear_slot(slots, key->index);
614 ovs_mutex_unlock(&key_mutex);
618 ovsthread_key_lookup__(const struct ovsthread_key *key)
620 struct ovsthread_key_slots *slots;
623 slots = pthread_getspecific(tsd_key);
625 slots = xzalloc(sizeof *slots);
627 ovs_mutex_lock(&key_mutex);
628 pthread_setspecific(tsd_key, slots);
629 list_push_back(&slots_list, &slots->list_node);
630 ovs_mutex_unlock(&key_mutex);
633 p2 = slots->p1[key->index / L2_SIZE];
635 p2 = xzalloc(L2_SIZE * sizeof *p2);
636 slots->p1[key->index / L2_SIZE] = p2;
639 return &p2[key->index % L2_SIZE];
642 /* Sets the value of thread-specific data item 'key', in the current thread, to
645 * This function is similar to pthread_setspecific(). */
647 ovsthread_setspecific(ovsthread_key_t key, const void *value)
649 *ovsthread_key_lookup__(key) = CONST_CAST(void *, value);
652 /* Returns the value of thread-specific data item 'key' in the current thread.
654 * This function is similar to pthread_getspecific(). */
656 ovsthread_getspecific(ovsthread_key_t key)
658 return *ovsthread_key_lookup__(key);