lib/ovs-thread.c

   1 /*
   2  * Copyright (c) 2013, 2014, 2015, 2016 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "ovs-thread.h"
  19 #include <errno.h>
  20 #include <poll.h>
  21 #ifndef _WIN32
  22 #include <signal.h>
  23 #endif
  24 #include <stdlib.h>
  25 #include <unistd.h>
  26 #include "compiler.h"
  27 #include "fatal-signal.h"
  28 #include "hash.h"
  29 #include "openvswitch/list.h"
  30 #include "ovs-rcu.h"
  31 #include "openvswitch/poll-loop.h"
  32 #include "seq.h"
  33 #include "socket-util.h"
  34 #include "util.h"
  35
  36 #ifdef __CHECKER__
  37 /* Omit the definitions in this file because they are somewhat difficult to
  38  * write without prompting "sparse" complaints, without ugliness or
  39  * cut-and-paste.  Since "sparse" is just a checker, not a compiler, it
  40  * doesn't matter that we don't define them. */
  41 #else
  42 #include "openvswitch/vlog.h"
  43
  44 VLOG_DEFINE_THIS_MODULE(ovs_thread);
  45
  46 /* If there is a reason that we cannot fork anymore (unless the fork will be
  47  * immediately followed by an exec), then this points to a string that
  48  * explains why. */
  49 static const char *must_not_fork;
  50
  51 /* True if we created any threads beyond the main initial thread. */
  52 static bool multithreaded;
  53
  54 #define LOCK_FUNCTION(TYPE, FUN) \
  55     void \
  56     ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
  57                             const char *where) \
  58         OVS_NO_THREAD_SAFETY_ANALYSIS \
  59     { \
  60         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
  61         int error; \
  62  \
  63         /* Verify that 'l' was initialized. */ \
  64         if (OVS_UNLIKELY(!l->where)) { \
  65             ovs_abort(0, "%s: %s() passed uninitialized ovs_"#TYPE, \
  66                       where, __func__); \
  67         } \
  68  \
  69         error = pthread_##TYPE##_##FUN(&l->lock); \
  70         if (OVS_UNLIKELY(error)) { \
  71             ovs_abort(error, "%s: pthread_%s_%s failed", where, #TYPE, #FUN); \
  72         } \
  73         l->where = where; \
  74  }
  75 LOCK_FUNCTION(mutex, lock);
  76 LOCK_FUNCTION(rwlock, rdlock);
  77 LOCK_FUNCTION(rwlock, wrlock);
  78 #ifdef HAVE_PTHREAD_SPIN_LOCK
  79 LOCK_FUNCTION(spin, lock);
  80 #endif
  81
  82 #define TRY_LOCK_FUNCTION(TYPE, FUN) \
  83     int \
  84     ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
  85                             const char *where) \
  86         OVS_NO_THREAD_SAFETY_ANALYSIS \
  87     { \
  88         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
  89         int error; \
  90  \
  91         /* Verify that 'l' was initialized. */ \
  92         if (OVS_UNLIKELY(!l->where)) { \
  93             ovs_abort(0, "%s: %s() passed uninitialized ovs_"#TYPE, \
  94                       where, __func__); \
  95         } \
  96  \
  97         error = pthread_##TYPE##_##FUN(&l->lock); \
  98         if (OVS_UNLIKELY(error) && error != EBUSY) { \
  99             ovs_abort(error, "%s: pthread_%s_%s failed", where, #TYPE, #FUN); \
 100         } \
 101         if (!error) { \
 102             l->where = where; \
 103         } \
 104         return error; \
 105     }
 106 TRY_LOCK_FUNCTION(mutex, trylock);
 107 TRY_LOCK_FUNCTION(rwlock, tryrdlock);
 108 TRY_LOCK_FUNCTION(rwlock, trywrlock);
 109 #ifdef HAVE_PTHREAD_SPIN_LOCK
 110 TRY_LOCK_FUNCTION(spin, trylock);
 111 #endif
 112
 113 #define UNLOCK_FUNCTION(TYPE, FUN, WHERE) \
 114     void \
 115     ovs_##TYPE##_##FUN(const struct ovs_##TYPE *l_) \
 116         OVS_NO_THREAD_SAFETY_ANALYSIS \
 117     { \
 118         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
 119         int error; \
 120  \
 121         /* Verify that 'l' was initialized. */ \
 122         ovs_assert(l->where); \
 123  \
 124         l->where = WHERE; \
 125         error = pthread_##TYPE##_##FUN(&l->lock); \
 126         if (OVS_UNLIKELY(error)) { \
 127             ovs_abort(error, "pthread_%s_%s failed", #TYPE, #FUN); \
 128         } \
 129     }
 130 UNLOCK_FUNCTION(mutex, unlock, "<unlocked>");
 131 UNLOCK_FUNCTION(mutex, destroy, NULL);
 132 UNLOCK_FUNCTION(rwlock, unlock, "<unlocked>");
 133 UNLOCK_FUNCTION(rwlock, destroy, NULL);
 134 #ifdef HAVE_PTHREAD_SPIN_LOCK
 135 UNLOCK_FUNCTION(spin, unlock, "<unlocked>");
 136 UNLOCK_FUNCTION(spin, destroy, NULL);
 137 #endif
 138
 139 #define XPTHREAD_FUNC1(FUNCTION, PARAM1)                \
 140     void                                                \
 141     x##FUNCTION(PARAM1 arg1)                            \
 142     {                                                   \
 143         int error = FUNCTION(arg1);                     \
 144         if (OVS_UNLIKELY(error)) {                      \
 145             ovs_abort(error, "%s failed", #FUNCTION);   \
 146         }                                               \
 147     }
 148 #define XPTHREAD_FUNC2(FUNCTION, PARAM1, PARAM2)        \
 149     void                                                \
 150     x##FUNCTION(PARAM1 arg1, PARAM2 arg2)               \
 151     {                                                   \
 152         int error = FUNCTION(arg1, arg2);               \
 153         if (OVS_UNLIKELY(error)) {                      \
 154             ovs_abort(error, "%s failed", #FUNCTION);   \
 155         }                                               \
 156     }
 157 #define XPTHREAD_FUNC3(FUNCTION, PARAM1, PARAM2, PARAM3)\
 158     void                                                \
 159     x##FUNCTION(PARAM1 arg1, PARAM2 arg2, PARAM3 arg3)  \
 160     {                                                   \
 161         int error = FUNCTION(arg1, arg2, arg3);         \
 162         if (OVS_UNLIKELY(error)) {                      \
 163             ovs_abort(error, "%s failed", #FUNCTION);   \
 164         }                                               \
 165     }
 166
 167 XPTHREAD_FUNC1(pthread_mutexattr_init, pthread_mutexattr_t *);
 168 XPTHREAD_FUNC1(pthread_mutexattr_destroy, pthread_mutexattr_t *);
 169 XPTHREAD_FUNC2(pthread_mutexattr_settype, pthread_mutexattr_t *, int);
 170 XPTHREAD_FUNC2(pthread_mutexattr_gettype, pthread_mutexattr_t *, int *);
 171
 172 XPTHREAD_FUNC1(pthread_rwlockattr_init, pthread_rwlockattr_t *);
 173 XPTHREAD_FUNC1(pthread_rwlockattr_destroy, pthread_rwlockattr_t *);
 174 #ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
 175 XPTHREAD_FUNC2(pthread_rwlockattr_setkind_np, pthread_rwlockattr_t *, int);
 176 #endif
 177
 178 XPTHREAD_FUNC2(pthread_cond_init, pthread_cond_t *, pthread_condattr_t *);
 179 XPTHREAD_FUNC1(pthread_cond_destroy, pthread_cond_t *);
 180 XPTHREAD_FUNC1(pthread_cond_signal, pthread_cond_t *);
 181 XPTHREAD_FUNC1(pthread_cond_broadcast, pthread_cond_t *);
 182
 183 XPTHREAD_FUNC2(pthread_join, pthread_t, void **);
 184
 185 typedef void destructor_func(void *);
 186 XPTHREAD_FUNC2(pthread_key_create, pthread_key_t *, destructor_func *);
 187 XPTHREAD_FUNC1(pthread_key_delete, pthread_key_t);
 188 XPTHREAD_FUNC2(pthread_setspecific, pthread_key_t, const void *);
 189
 190 #ifndef _WIN32
 191 XPTHREAD_FUNC3(pthread_sigmask, int, const sigset_t *, sigset_t *);
 192 #endif
 193
 194 static void
 195 ovs_mutex_init__(const struct ovs_mutex *l_, int type)
 196 {
 197     struct ovs_mutex *l = CONST_CAST(struct ovs_mutex *, l_);
 198     pthread_mutexattr_t attr;
 199     int error;
 200
 201     l->where = "<unlocked>";
 202     xpthread_mutexattr_init(&attr);
 203     xpthread_mutexattr_settype(&attr, type);
 204     error = pthread_mutex_init(&l->lock, &attr);
 205     if (OVS_UNLIKELY(error)) {
 206         ovs_abort(error, "pthread_mutex_init failed");
 207     }
 208     xpthread_mutexattr_destroy(&attr);
 209 }
 210
 211 /* Initializes 'mutex' as a normal (non-recursive) mutex. */
 212 void
 213 ovs_mutex_init(const struct ovs_mutex *mutex)
 214 {
 215     ovs_mutex_init__(mutex, PTHREAD_MUTEX_ERRORCHECK);
 216 }
 217
 218 /* Initializes 'mutex' as a recursive mutex. */
 219 void
 220 ovs_mutex_init_recursive(const struct ovs_mutex *mutex)
 221 {
 222     ovs_mutex_init__(mutex, PTHREAD_MUTEX_RECURSIVE);
 223 }
 224
 225 /* Initializes 'mutex' as a recursive mutex. */
 226 void
 227 ovs_mutex_init_adaptive(const struct ovs_mutex *mutex)
 228 {
 229 #ifdef PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP
 230     ovs_mutex_init__(mutex, PTHREAD_MUTEX_ADAPTIVE_NP);
 231 #else
 232     ovs_mutex_init(mutex);
 233 #endif
 234 }
 235
 236 void
 237 ovs_rwlock_init(const struct ovs_rwlock *l_)
 238 {
 239     struct ovs_rwlock *l = CONST_CAST(struct ovs_rwlock *, l_);
 240     int error;
 241
 242     l->where = "<unlocked>";
 243
 244 #ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
 245     pthread_rwlockattr_t attr;
 246     xpthread_rwlockattr_init(&attr);
 247     xpthread_rwlockattr_setkind_np(
 248         &attr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP);
 249     error = pthread_rwlock_init(&l->lock, &attr);
 250     xpthread_rwlockattr_destroy(&attr);
 251 #else
 252     /* It is important to avoid passing a rwlockattr in this case because
 253      * Windows pthreads 2.9.1 (and earlier) fail and abort if passed one, even
 254      * one without any special attributes. */
 255     error = pthread_rwlock_init(&l->lock, NULL);
 256 #endif
 257
 258     if (OVS_UNLIKELY(error)) {
 259         ovs_abort(error, "pthread_rwlock_init failed");
 260     }
 261 }
 262
 263 /* Provides an error-checking wrapper around pthread_cond_wait().
 264  *
 265  * If the wait can take a significant amount of time, consider bracketing this
 266  * call with calls to ovsrcu_quiesce_start() and ovsrcu_quiesce_end().  */
 267 void
 268 ovs_mutex_cond_wait(pthread_cond_t *cond, const struct ovs_mutex *mutex_)
 269     OVS_NO_THREAD_SAFETY_ANALYSIS
 270 {
 271     struct ovs_mutex *mutex = CONST_CAST(struct ovs_mutex *, mutex_);
 272     int error;
 273
 274     error = pthread_cond_wait(cond, &mutex->lock);
 275
 276     if (OVS_UNLIKELY(error)) {
 277         ovs_abort(error, "pthread_cond_wait failed");
 278     }
 279 }
 280
 281 #ifdef HAVE_PTHREAD_SPIN_LOCK
 282 static void
 283 ovs_spin_init__(const struct ovs_spin *l_, int pshared)
 284 {
 285     struct ovs_spin *l = CONST_CAST(struct ovs_spin *, l_);
 286     int error;
 287
 288     l->where = "<unlocked>";
 289     error = pthread_spin_init(&l->lock, pshared);
 290     if (OVS_UNLIKELY(error)) {
 291         ovs_abort(error, "pthread_spin_init failed");
 292     }
 293 }
 294
 295 void
 296 ovs_spin_init(const struct ovs_spin *spin)
 297 {
 298     ovs_spin_init__(spin, PTHREAD_PROCESS_PRIVATE);
 299 }
 300 #endif
 301
 302 /* Initializes the 'barrier'.  'size' is the number of threads
 303  * expected to hit the barrier. */
 304 void
 305 ovs_barrier_init(struct ovs_barrier *barrier, uint32_t size)
 306 {
 307     barrier->size = size;
 308     atomic_count_init(&barrier->count, 0);
 309     barrier->seq = seq_create();
 310 }
 311
 312 /* Destroys the 'barrier'. */
 313 void
 314 ovs_barrier_destroy(struct ovs_barrier *barrier)
 315 {
 316     seq_destroy(barrier->seq);
 317 }
 318
 319 /* Makes the calling thread block on the 'barrier' until all
 320  * 'barrier->size' threads hit the barrier.
 321  * ovs_barrier provides the necessary acquire-release semantics to make
 322  * the effects of prior memory accesses of all the participating threads
 323  * visible on return and to prevent the following memory accesses to be
 324  * reordered before the ovs_barrier_block(). */
 325 void
 326 ovs_barrier_block(struct ovs_barrier *barrier)
 327 {
 328     uint64_t seq = seq_read(barrier->seq);
 329     uint32_t orig;
 330
 331     orig = atomic_count_inc(&barrier->count);
 332     if (orig + 1 == barrier->size) {
 333         atomic_count_set(&barrier->count, 0);
 334         /* seq_change() serves as a release barrier against the other threads,
 335          * so the zeroed count is visible to them as they continue. */
 336         seq_change(barrier->seq);
 337     } else {
 338         /* To prevent thread from waking up by other event,
 339          * keeps waiting for the change of 'barrier->seq'. */
 340         while (seq == seq_read(barrier->seq)) {
 341             seq_wait(barrier->seq, seq);
 342             poll_block();
 343         }
 344     }
 345 }
 346 \f
 347 DEFINE_EXTERN_PER_THREAD_DATA(ovsthread_id, OVSTHREAD_ID_UNSET);
 348
 349 struct ovsthread_aux {
 350     void *(*start)(void *);
 351     void *arg;
 352     char name[16];
 353 };
 354
 355 unsigned int
 356 ovsthread_id_init(void)
 357 {
 358     static atomic_count next_id = ATOMIC_COUNT_INIT(0);
 359
 360     ovs_assert(*ovsthread_id_get() == OVSTHREAD_ID_UNSET);
 361     return *ovsthread_id_get() = atomic_count_inc(&next_id);
 362 }
 363
 364 static void *
 365 ovsthread_wrapper(void *aux_)
 366 {
 367     struct ovsthread_aux *auxp = aux_;
 368     struct ovsthread_aux aux;
 369     unsigned int id;
 370
 371     id = ovsthread_id_init();
 372
 373     aux = *auxp;
 374     free(auxp);
 375
 376     /* The order of the following calls is important, because
 377      * ovsrcu_quiesce_end() saves a copy of the thread name. */
 378     char *subprogram_name = xasprintf("%s%u", aux.name, id);
 379     set_subprogram_name(subprogram_name);
 380     free(subprogram_name);
 381     ovsrcu_quiesce_end();
 382
 383     return aux.start(aux.arg);
 384 }
 385
 386 static void
 387 set_min_stack_size(pthread_attr_t *attr, size_t min_stacksize)
 388 {
 389     size_t stacksize;
 390     int error;
 391
 392     error = pthread_attr_getstacksize(attr, &stacksize);
 393     if (error) {
 394         ovs_abort(error, "pthread_attr_getstacksize failed");
 395     }
 396
 397     if (stacksize < min_stacksize) {
 398         error = pthread_attr_setstacksize(attr, min_stacksize);
 399         if (error) {
 400             ovs_abort(error, "pthread_attr_setstacksize failed");
 401         }
 402     }
 403 }
 404
 405 /* Starts a thread that calls 'start(arg)'.  Sets the thread's name to 'name'
 406  * (suffixed by its ovsthread_id()).  Returns the new thread's pthread_t. */
 407 pthread_t
 408 ovs_thread_create(const char *name, void *(*start)(void *), void *arg)
 409 {
 410     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 411     struct ovsthread_aux *aux;
 412     pthread_t thread;
 413     int error;
 414
 415     forbid_forking("multiple threads exist");
 416
 417     if (ovsthread_once_start(&once)) {
 418         /* The first call to this function has to happen in the main thread.
 419          * Before the process becomes multithreaded we make sure that the
 420          * main thread is considered non quiescent.
 421          *
 422          * For other threads this is done in ovs_thread_wrapper(), but the
 423          * main thread has no such wrapper.
 424          *
 425          * There's no reason to call ovsrcu_quiesce_end() in subsequent
 426          * invocations of this function and it might introduce problems
 427          * for other threads. */
 428         ovsrcu_quiesce_end();
 429         ovsthread_once_done(&once);
 430     }
 431
 432     multithreaded = true;
 433     aux = xmalloc(sizeof *aux);
 434     aux->start = start;
 435     aux->arg = arg;
 436     ovs_strlcpy(aux->name, name, sizeof aux->name);
 437
 438     /* Some small systems use a default stack size as small as 80 kB, but OVS
 439      * requires approximately 384 kB according to the following analysis:
 440      * https://mail.openvswitch.org/pipermail/ovs-dev/2016-January/308592.html
 441      *
 442      * We use 512 kB to give us some margin of error. */
 443     pthread_attr_t attr;
 444     pthread_attr_init(&attr);
 445     set_min_stack_size(&attr, 512 * 1024);
 446
 447     error = pthread_create(&thread, &attr, ovsthread_wrapper, aux);
 448     if (error) {
 449         ovs_abort(error, "pthread_create failed");
 450     }
 451     pthread_attr_destroy(&attr);
 452     return thread;
 453 }
 454 \f
 455 bool
 456 ovsthread_once_start__(struct ovsthread_once *once)
 457 {
 458     ovs_mutex_lock(&once->mutex);
 459     /* Mutex synchronizes memory, so we get the current value of 'done'. */
 460     if (!once->done) {
 461         return true;
 462     }
 463     ovs_mutex_unlock(&once->mutex);
 464     return false;
 465 }
 466
 467 void
 468 ovsthread_once_done(struct ovsthread_once *once)
 469 {
 470     /* We need release semantics here, so that the following store may not
 471      * be moved ahead of any of the preceding initialization operations.
 472      * A release atomic_thread_fence provides that prior memory accesses
 473      * will not be reordered to take place after the following store. */
 474     atomic_thread_fence(memory_order_release);
 475     once->done = true;
 476     ovs_mutex_unlock(&once->mutex);
 477 }
 478 \f
 479 bool
 480 single_threaded(void)
 481 {
 482     return !multithreaded;
 483 }
 484
 485 /* Asserts that the process has not yet created any threads (beyond the initial
 486  * thread).
 487  *
 488  * ('where' is used in logging.  Commonly one would use
 489  * assert_single_threaded() to automatically provide the caller's source file
 490  * and line number for 'where'.) */
 491 void
 492 assert_single_threaded_at(const char *where)
 493 {
 494     if (multithreaded) {
 495         VLOG_FATAL("%s: attempted operation not allowed when multithreaded",
 496                    where);
 497     }
 498 }
 499
 500 #ifndef _WIN32
 501 /* Forks the current process (checking that this is allowed).  Aborts with
 502  * VLOG_FATAL if fork() returns an error, and otherwise returns the value
 503  * returned by fork().
 504  *
 505  * ('where' is used in logging.  Commonly one would use xfork() to
 506  * automatically provide the caller's source file and line number for
 507  * 'where'.) */
 508 pid_t
 509 xfork_at(const char *where)
 510 {
 511     pid_t pid;
 512
 513     if (must_not_fork) {
 514         VLOG_FATAL("%s: attempted to fork but forking not allowed (%s)",
 515                    where, must_not_fork);
 516     }
 517
 518     pid = fork();
 519     if (pid < 0) {
 520         VLOG_FATAL("%s: fork failed (%s)", where, ovs_strerror(errno));
 521     }
 522     return pid;
 523 }
 524 #endif
 525
 526 /* Notes that the process must not call fork() from now on, for the specified
 527  * 'reason'.  (The process may still fork() if it execs itself immediately
 528  * afterward.) */
 529 void
 530 forbid_forking(const char *reason)
 531 {
 532     ovs_assert(reason != NULL);
 533     must_not_fork = reason;
 534 }
 535
 536 /* Returns true if the process is allowed to fork, false otherwise. */
 537 bool
 538 may_fork(void)
 539 {
 540     return !must_not_fork;
 541 }
 542 \f
 543 /* ovsthread_stats. */
 544
 545 void
 546 ovsthread_stats_init(struct ovsthread_stats *stats)
 547 {
 548     int i;
 549
 550     ovs_mutex_init(&stats->mutex);
 551     for (i = 0; i < ARRAY_SIZE(stats->buckets); i++) {
 552         stats->buckets[i] = NULL;
 553     }
 554 }
 555
 556 void
 557 ovsthread_stats_destroy(struct ovsthread_stats *stats)
 558 {
 559     ovs_mutex_destroy(&stats->mutex);
 560 }
 561
 562 void *
 563 ovsthread_stats_bucket_get(struct ovsthread_stats *stats,
 564                            void *(*new_bucket)(void))
 565 {
 566     unsigned int idx = ovsthread_id_self() & (ARRAY_SIZE(stats->buckets) - 1);
 567     void *bucket = stats->buckets[idx];
 568     if (!bucket) {
 569         ovs_mutex_lock(&stats->mutex);
 570         bucket = stats->buckets[idx];
 571         if (!bucket) {
 572             bucket = stats->buckets[idx] = new_bucket();
 573         }
 574         ovs_mutex_unlock(&stats->mutex);
 575     }
 576     return bucket;
 577 }
 578
 579 size_t
 580 ovs_thread_stats_next_bucket(const struct ovsthread_stats *stats, size_t i)
 581 {
 582     for (; i < ARRAY_SIZE(stats->buckets); i++) {
 583         if (stats->buckets[i]) {
 584             break;
 585         }
 586     }
 587     return i;
 588 }
 589
 590 \f
 591 /* Returns the total number of cores available to this process, or 0 if the
 592  * number cannot be determined. */
 593 int
 594 count_cpu_cores(void)
 595 {
 596     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 597     static long int n_cores;
 598
 599     if (ovsthread_once_start(&once)) {
 600 #ifndef _WIN32
 601         n_cores = sysconf(_SC_NPROCESSORS_ONLN);
 602 #ifdef __linux__
 603         if (n_cores > 0) {
 604             cpu_set_t *set = CPU_ALLOC(n_cores);
 605
 606             if (set) {
 607                 size_t size = CPU_ALLOC_SIZE(n_cores);
 608
 609                 if (!sched_getaffinity(0, size, set)) {
 610                     n_cores = CPU_COUNT_S(size, set);
 611                 }
 612                 CPU_FREE(set);
 613             }
 614         }
 615 #endif
 616 #else
 617         SYSTEM_INFO sysinfo;
 618         GetSystemInfo(&sysinfo);
 619         n_cores = sysinfo.dwNumberOfProcessors;
 620 #endif
 621         ovsthread_once_done(&once);
 622     }
 623
 624     return n_cores > 0 ? n_cores : 0;
 625 }
 626
 627 /* Returns 'true' if current thread is PMD thread. */
 628 bool
 629 thread_is_pmd(void)
 630 {
 631     const char *name = get_subprogram_name();
 632     return !strncmp(name, "pmd", 3);
 633 }
 634
 635 \f
 636 /* ovsthread_key. */
 637
 638 #define L1_SIZE 1024
 639 #define L2_SIZE 1024
 640 #define MAX_KEYS (L1_SIZE * L2_SIZE)
 641
 642 /* A piece of thread-specific data. */
 643 struct ovsthread_key {
 644     struct ovs_list list_node;  /* In 'inuse_keys' or 'free_keys'. */
 645     void (*destructor)(void *); /* Called at thread exit. */
 646
 647     /* Indexes into the per-thread array in struct ovsthread_key_slots.
 648      * This key's data is stored in p1[index / L2_SIZE][index % L2_SIZE]. */
 649     unsigned int index;
 650 };
 651
 652 /* Per-thread data structure. */
 653 struct ovsthread_key_slots {
 654     struct ovs_list list_node;  /* In 'slots_list'. */
 655     void **p1[L1_SIZE];
 656 };
 657
 658 /* Contains "struct ovsthread_key_slots *". */
 659 static pthread_key_t tsd_key;
 660
 661 /* Guards data structures below. */
 662 static struct ovs_mutex key_mutex = OVS_MUTEX_INITIALIZER;
 663
 664 /* 'inuse_keys' holds "struct ovsthread_key"s that have been created and not
 665  * yet destroyed.
 666  *
 667  * 'free_keys' holds "struct ovsthread_key"s that have been deleted and are
 668  * ready for reuse.  (We keep them around only to be able to easily locate
 669  * free indexes.)
 670  *
 671  * Together, 'inuse_keys' and 'free_keys' hold an ovsthread_key for every index
 672  * from 0 to n_keys - 1, inclusive. */
 673 static struct ovs_list inuse_keys OVS_GUARDED_BY(key_mutex)
 674     = OVS_LIST_INITIALIZER(&inuse_keys);
 675 static struct ovs_list free_keys OVS_GUARDED_BY(key_mutex)
 676     = OVS_LIST_INITIALIZER(&free_keys);
 677 static unsigned int n_keys OVS_GUARDED_BY(key_mutex);
 678
 679 /* All existing struct ovsthread_key_slots. */
 680 static struct ovs_list slots_list OVS_GUARDED_BY(key_mutex)
 681     = OVS_LIST_INITIALIZER(&slots_list);
 682
 683 static void *
 684 clear_slot(struct ovsthread_key_slots *slots, unsigned int index)
 685 {
 686     void **p2 = slots->p1[index / L2_SIZE];
 687     if (p2) {
 688         void **valuep = &p2[index % L2_SIZE];
 689         void *value = *valuep;
 690         *valuep = NULL;
 691         return value;
 692     } else {
 693         return NULL;
 694     }
 695 }
 696
 697 static void
 698 ovsthread_key_destruct__(void *slots_)
 699 {
 700     struct ovsthread_key_slots *slots = slots_;
 701     struct ovsthread_key *key;
 702     unsigned int n;
 703     int i;
 704
 705     ovs_mutex_lock(&key_mutex);
 706     ovs_list_remove(&slots->list_node);
 707     LIST_FOR_EACH (key, list_node, &inuse_keys) {
 708         void *value = clear_slot(slots, key->index);
 709         if (value && key->destructor) {
 710             key->destructor(value);
 711         }
 712     }
 713     n = n_keys;
 714     ovs_mutex_unlock(&key_mutex);
 715
 716     for (i = 0; i < DIV_ROUND_UP(n, L2_SIZE); i++) {
 717         free(slots->p1[i]);
 718     }
 719     free(slots);
 720 }
 721
 722 /* Cancels the callback to ovsthread_key_destruct__().
 723  *
 724  * Cancelling the call to the destructor during the main thread exit
 725  * is needed while using pthreads-win32 library in Windows. It has been
 726  * observed that in pthreads-win32, a call to the destructor during
 727  * main thread exit causes undefined behavior. */
 728 static void
 729 ovsthread_cancel_ovsthread_key_destruct__(void *aux OVS_UNUSED)
 730 {
 731     pthread_setspecific(tsd_key, NULL);
 732 }
 733
 734 /* Initializes '*keyp' as a thread-specific data key.  The data items are
 735  * initially null in all threads.
 736  *
 737  * If a thread exits with non-null data, then 'destructor', if nonnull, will be
 738  * called passing the final data value as its argument.  'destructor' must not
 739  * call any thread-specific data functions in this API.
 740  *
 741  * This function is similar to xpthread_key_create(). */
 742 void
 743 ovsthread_key_create(ovsthread_key_t *keyp, void (*destructor)(void *))
 744 {
 745     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 746     struct ovsthread_key *key;
 747
 748     if (ovsthread_once_start(&once)) {
 749         xpthread_key_create(&tsd_key, ovsthread_key_destruct__);
 750         fatal_signal_add_hook(ovsthread_cancel_ovsthread_key_destruct__,
 751                               NULL, NULL, true);
 752         ovsthread_once_done(&once);
 753     }
 754
 755     ovs_mutex_lock(&key_mutex);
 756     if (ovs_list_is_empty(&free_keys)) {
 757         key = xmalloc(sizeof *key);
 758         key->index = n_keys++;
 759         if (key->index >= MAX_KEYS) {
 760             abort();
 761         }
 762     } else {
 763         key = CONTAINER_OF(ovs_list_pop_back(&free_keys),
 764                             struct ovsthread_key, list_node);
 765     }
 766     ovs_list_push_back(&inuse_keys, &key->list_node);
 767     key->destructor = destructor;
 768     ovs_mutex_unlock(&key_mutex);
 769
 770     *keyp = key;
 771 }
 772
 773 /* Frees 'key'.  The destructor supplied to ovsthread_key_create(), if any, is
 774  * not called.
 775  *
 776  * This function is similar to xpthread_key_delete(). */
 777 void
 778 ovsthread_key_delete(ovsthread_key_t key)
 779 {
 780     struct ovsthread_key_slots *slots;
 781
 782     ovs_mutex_lock(&key_mutex);
 783
 784     /* Move 'key' from 'inuse_keys' to 'free_keys'. */
 785     ovs_list_remove(&key->list_node);
 786     ovs_list_push_back(&free_keys, &key->list_node);
 787
 788     /* Clear this slot in all threads. */
 789     LIST_FOR_EACH (slots, list_node, &slots_list) {
 790         clear_slot(slots, key->index);
 791     }
 792
 793     ovs_mutex_unlock(&key_mutex);
 794 }
 795
 796 static void **
 797 ovsthread_key_lookup__(const struct ovsthread_key *key)
 798 {
 799     struct ovsthread_key_slots *slots;
 800     void **p2;
 801
 802     slots = pthread_getspecific(tsd_key);
 803     if (!slots) {
 804         slots = xzalloc(sizeof *slots);
 805
 806         ovs_mutex_lock(&key_mutex);
 807         pthread_setspecific(tsd_key, slots);
 808         ovs_list_push_back(&slots_list, &slots->list_node);
 809         ovs_mutex_unlock(&key_mutex);
 810     }
 811
 812     p2 = slots->p1[key->index / L2_SIZE];
 813     if (!p2) {
 814         p2 = xzalloc(L2_SIZE * sizeof *p2);
 815         slots->p1[key->index / L2_SIZE] = p2;
 816     }
 817
 818     return &p2[key->index % L2_SIZE];
 819 }
 820
 821 /* Sets the value of thread-specific data item 'key', in the current thread, to
 822  * 'value'.
 823  *
 824  * This function is similar to pthread_setspecific(). */
 825 void
 826 ovsthread_setspecific(ovsthread_key_t key, const void *value)
 827 {
 828     *ovsthread_key_lookup__(key) = CONST_CAST(void *, value);
 829 }
 830
 831 /* Returns the value of thread-specific data item 'key' in the current thread.
 832  *
 833  * This function is similar to pthread_getspecific(). */
 834 void *
 835 ovsthread_getspecific(ovsthread_key_t key)
 836 {
 837     return *ovsthread_key_lookup__(key);
 838 }
 839 #endif