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 "list.h"
  30 #include "netdev-dpdk.h"
  31 #include "ovs-rcu.h"
  32 #include "poll-loop.h"
  33 #include "seq.h"
  34 #include "socket-util.h"
  35 #include "util.h"
  36
  37 #ifdef __CHECKER__
  38 /* Omit the definitions in this file because they are somewhat difficult to
  39  * write without prompting "sparse" complaints, without ugliness or
  40  * cut-and-paste.  Since "sparse" is just a checker, not a compiler, it
  41  * doesn't matter that we don't define them. */
  42 #else
  43 #include "openvswitch/vlog.h"
  44
  45 VLOG_DEFINE_THIS_MODULE(ovs_thread);
  46
  47 /* If there is a reason that we cannot fork anymore (unless the fork will be
  48  * immediately followed by an exec), then this points to a string that
  49  * explains why. */
  50 static const char *must_not_fork;
  51
  52 /* True if we created any threads beyond the main initial thread. */
  53 static bool multithreaded;
  54
  55 #define LOCK_FUNCTION(TYPE, FUN) \
  56     void \
  57     ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
  58                             const char *where) \
  59         OVS_NO_THREAD_SAFETY_ANALYSIS \
  60     { \
  61         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
  62         int error; \
  63  \
  64         /* Verify that 'l' was initialized. */ \
  65         if (OVS_UNLIKELY(!l->where)) { \
  66             ovs_abort(0, "%s: %s() passed uninitialized ovs_"#TYPE, \
  67                       where, __func__); \
  68         } \
  69  \
  70         error = pthread_##TYPE##_##FUN(&l->lock); \
  71         if (OVS_UNLIKELY(error)) { \
  72             ovs_abort(error, "%s: pthread_%s_%s failed", where, #TYPE, #FUN); \
  73         } \
  74         l->where = where; \
  75  }
  76 LOCK_FUNCTION(mutex, lock);
  77 LOCK_FUNCTION(rwlock, rdlock);
  78 LOCK_FUNCTION(rwlock, wrlock);
  79
  80 #define TRY_LOCK_FUNCTION(TYPE, FUN) \
  81     int \
  82     ovs_##TYPE##_##FUN##_at(const struct ovs_##TYPE *l_, \
  83                             const char *where) \
  84         OVS_NO_THREAD_SAFETY_ANALYSIS \
  85     { \
  86         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
  87         int error; \
  88  \
  89         /* Verify that 'l' was initialized. */ \
  90         if (OVS_UNLIKELY(!l->where)) { \
  91             ovs_abort(0, "%s: %s() passed uninitialized ovs_"#TYPE, \
  92                       where, __func__); \
  93         } \
  94  \
  95         error = pthread_##TYPE##_##FUN(&l->lock); \
  96         if (OVS_UNLIKELY(error) && error != EBUSY) { \
  97             ovs_abort(error, "%s: pthread_%s_%s failed", where, #TYPE, #FUN); \
  98         } \
  99         if (!error) { \
 100             l->where = where; \
 101         } \
 102         return error; \
 103     }
 104 TRY_LOCK_FUNCTION(mutex, trylock);
 105 TRY_LOCK_FUNCTION(rwlock, tryrdlock);
 106 TRY_LOCK_FUNCTION(rwlock, trywrlock);
 107
 108 #define UNLOCK_FUNCTION(TYPE, FUN, WHERE) \
 109     void \
 110     ovs_##TYPE##_##FUN(const struct ovs_##TYPE *l_) \
 111         OVS_NO_THREAD_SAFETY_ANALYSIS \
 112     { \
 113         struct ovs_##TYPE *l = CONST_CAST(struct ovs_##TYPE *, l_); \
 114         int error; \
 115  \
 116         /* Verify that 'l' was initialized. */ \
 117         ovs_assert(l->where); \
 118  \
 119         l->where = WHERE; \
 120         error = pthread_##TYPE##_##FUN(&l->lock); \
 121         if (OVS_UNLIKELY(error)) { \
 122             ovs_abort(error, "pthread_%s_%s failed", #TYPE, #FUN); \
 123         } \
 124     }
 125 UNLOCK_FUNCTION(mutex, unlock, "<unlocked>");
 126 UNLOCK_FUNCTION(mutex, destroy, NULL);
 127 UNLOCK_FUNCTION(rwlock, unlock, "<unlocked>");
 128 UNLOCK_FUNCTION(rwlock, destroy, NULL);
 129
 130 #define XPTHREAD_FUNC1(FUNCTION, PARAM1)                \
 131     void                                                \
 132     x##FUNCTION(PARAM1 arg1)                            \
 133     {                                                   \
 134         int error = FUNCTION(arg1);                     \
 135         if (OVS_UNLIKELY(error)) {                      \
 136             ovs_abort(error, "%s failed", #FUNCTION);   \
 137         }                                               \
 138     }
 139 #define XPTHREAD_FUNC2(FUNCTION, PARAM1, PARAM2)        \
 140     void                                                \
 141     x##FUNCTION(PARAM1 arg1, PARAM2 arg2)               \
 142     {                                                   \
 143         int error = FUNCTION(arg1, arg2);               \
 144         if (OVS_UNLIKELY(error)) {                      \
 145             ovs_abort(error, "%s failed", #FUNCTION);   \
 146         }                                               \
 147     }
 148 #define XPTHREAD_FUNC3(FUNCTION, PARAM1, PARAM2, PARAM3)\
 149     void                                                \
 150     x##FUNCTION(PARAM1 arg1, PARAM2 arg2, PARAM3 arg3)  \
 151     {                                                   \
 152         int error = FUNCTION(arg1, arg2, arg3);         \
 153         if (OVS_UNLIKELY(error)) {                      \
 154             ovs_abort(error, "%s failed", #FUNCTION);   \
 155         }                                               \
 156     }
 157
 158 XPTHREAD_FUNC1(pthread_mutex_lock, pthread_mutex_t *);
 159 XPTHREAD_FUNC1(pthread_mutex_unlock, pthread_mutex_t *);
 160 XPTHREAD_FUNC1(pthread_mutexattr_init, pthread_mutexattr_t *);
 161 XPTHREAD_FUNC1(pthread_mutexattr_destroy, pthread_mutexattr_t *);
 162 XPTHREAD_FUNC2(pthread_mutexattr_settype, pthread_mutexattr_t *, int);
 163 XPTHREAD_FUNC2(pthread_mutexattr_gettype, pthread_mutexattr_t *, int *);
 164
 165 XPTHREAD_FUNC1(pthread_rwlockattr_init, pthread_rwlockattr_t *);
 166 XPTHREAD_FUNC1(pthread_rwlockattr_destroy, pthread_rwlockattr_t *);
 167 #ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
 168 XPTHREAD_FUNC2(pthread_rwlockattr_setkind_np, pthread_rwlockattr_t *, int);
 169 #endif
 170
 171 XPTHREAD_FUNC2(pthread_cond_init, pthread_cond_t *, pthread_condattr_t *);
 172 XPTHREAD_FUNC1(pthread_cond_destroy, pthread_cond_t *);
 173 XPTHREAD_FUNC1(pthread_cond_signal, pthread_cond_t *);
 174 XPTHREAD_FUNC1(pthread_cond_broadcast, pthread_cond_t *);
 175
 176 XPTHREAD_FUNC2(pthread_join, pthread_t, void **);
 177
 178 typedef void destructor_func(void *);
 179 XPTHREAD_FUNC2(pthread_key_create, pthread_key_t *, destructor_func *);
 180 XPTHREAD_FUNC1(pthread_key_delete, pthread_key_t);
 181 XPTHREAD_FUNC2(pthread_setspecific, pthread_key_t, const void *);
 182
 183 #ifndef _WIN32
 184 XPTHREAD_FUNC3(pthread_sigmask, int, const sigset_t *, sigset_t *);
 185 #endif
 186
 187 static void
 188 ovs_mutex_init__(const struct ovs_mutex *l_, int type)
 189 {
 190     struct ovs_mutex *l = CONST_CAST(struct ovs_mutex *, l_);
 191     pthread_mutexattr_t attr;
 192     int error;
 193
 194     l->where = "<unlocked>";
 195     xpthread_mutexattr_init(&attr);
 196     xpthread_mutexattr_settype(&attr, type);
 197     error = pthread_mutex_init(&l->lock, &attr);
 198     if (OVS_UNLIKELY(error)) {
 199         ovs_abort(error, "pthread_mutex_init failed");
 200     }
 201     xpthread_mutexattr_destroy(&attr);
 202 }
 203
 204 /* Initializes 'mutex' as a normal (non-recursive) mutex. */
 205 void
 206 ovs_mutex_init(const struct ovs_mutex *mutex)
 207 {
 208     ovs_mutex_init__(mutex, PTHREAD_MUTEX_ERRORCHECK);
 209 }
 210
 211 /* Initializes 'mutex' as a recursive mutex. */
 212 void
 213 ovs_mutex_init_recursive(const struct ovs_mutex *mutex)
 214 {
 215     ovs_mutex_init__(mutex, PTHREAD_MUTEX_RECURSIVE);
 216 }
 217
 218 /* Initializes 'mutex' as a recursive mutex. */
 219 void
 220 ovs_mutex_init_adaptive(const struct ovs_mutex *mutex)
 221 {
 222 #ifdef PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP
 223     ovs_mutex_init__(mutex, PTHREAD_MUTEX_ADAPTIVE_NP);
 224 #else
 225     ovs_mutex_init(mutex);
 226 #endif
 227 }
 228
 229 void
 230 ovs_rwlock_init(const struct ovs_rwlock *l_)
 231 {
 232     struct ovs_rwlock *l = CONST_CAST(struct ovs_rwlock *, l_);
 233     pthread_rwlockattr_t attr;
 234     int error;
 235
 236     l->where = "<unlocked>";
 237
 238     xpthread_rwlockattr_init(&attr);
 239 #ifdef PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP
 240     xpthread_rwlockattr_setkind_np(
 241         &attr, PTHREAD_RWLOCK_PREFER_WRITER_NONRECURSIVE_NP);
 242 #endif
 243     error = pthread_rwlock_init(&l->lock, NULL);
 244     if (OVS_UNLIKELY(error)) {
 245         ovs_abort(error, "pthread_rwlock_init failed");
 246     }
 247     xpthread_rwlockattr_destroy(&attr);
 248 }
 249
 250 void
 251 ovs_mutex_cond_wait(pthread_cond_t *cond, const struct ovs_mutex *mutex_)
 252 {
 253     struct ovs_mutex *mutex = CONST_CAST(struct ovs_mutex *, mutex_);
 254     int error;
 255
 256     ovsrcu_quiesce_start();
 257     error = pthread_cond_wait(cond, &mutex->lock);
 258     ovsrcu_quiesce_end();
 259
 260     if (OVS_UNLIKELY(error)) {
 261         ovs_abort(error, "pthread_cond_wait failed");
 262     }
 263 }
 264
 265 /* Initializes the 'barrier'.  'size' is the number of threads
 266  * expected to hit the barrier. */
 267 void
 268 ovs_barrier_init(struct ovs_barrier *barrier, uint32_t size)
 269 {
 270     barrier->size = size;
 271     atomic_count_init(&barrier->count, 0);
 272     barrier->seq = seq_create();
 273 }
 274
 275 /* Destroys the 'barrier'. */
 276 void
 277 ovs_barrier_destroy(struct ovs_barrier *barrier)
 278 {
 279     seq_destroy(barrier->seq);
 280 }
 281
 282 /* Makes the calling thread block on the 'barrier' until all
 283  * 'barrier->size' threads hit the barrier.
 284  * ovs_barrier provides the necessary acquire-release semantics to make
 285  * the effects of prior memory accesses of all the participating threads
 286  * visible on return and to prevent the following memory accesses to be
 287  * reordered before the ovs_barrier_block(). */
 288 void
 289 ovs_barrier_block(struct ovs_barrier *barrier)
 290 {
 291     uint64_t seq = seq_read(barrier->seq);
 292     uint32_t orig;
 293
 294     orig = atomic_count_inc(&barrier->count);
 295     if (orig + 1 == barrier->size) {
 296         atomic_count_set(&barrier->count, 0);
 297         /* seq_change() serves as a release barrier against the other threads,
 298          * so the zeroed count is visible to them as they continue. */
 299         seq_change(barrier->seq);
 300     } else {
 301         /* To prevent thread from waking up by other event,
 302          * keeps waiting for the change of 'barrier->seq'. */
 303         while (seq == seq_read(barrier->seq)) {
 304             seq_wait(barrier->seq, seq);
 305             poll_block();
 306         }
 307     }
 308 }
 309 \f
 310 DEFINE_EXTERN_PER_THREAD_DATA(ovsthread_id, 0);
 311
 312 struct ovsthread_aux {
 313     void *(*start)(void *);
 314     void *arg;
 315     char name[16];
 316 };
 317
 318 static void *
 319 ovsthread_wrapper(void *aux_)
 320 {
 321     static atomic_count next_id = ATOMIC_COUNT_INIT(1);
 322
 323     struct ovsthread_aux *auxp = aux_;
 324     struct ovsthread_aux aux;
 325     unsigned int id;
 326
 327     id = atomic_count_inc(&next_id);
 328     *ovsthread_id_get() = id;
 329
 330     aux = *auxp;
 331     free(auxp);
 332
 333     /* The order of the following calls is important, because
 334      * ovsrcu_quiesce_end() saves a copy of the thread name. */
 335     char *subprogram_name = xasprintf("%s%u", aux.name, id);
 336     set_subprogram_name(subprogram_name);
 337     free(subprogram_name);
 338     ovsrcu_quiesce_end();
 339
 340     return aux.start(aux.arg);
 341 }
 342
 343 static void
 344 set_min_stack_size(pthread_attr_t *attr, size_t min_stacksize)
 345 {
 346     size_t stacksize;
 347     int error;
 348
 349     error = pthread_attr_getstacksize(attr, &stacksize);
 350     if (error) {
 351         ovs_abort(error, "pthread_attr_getstacksize failed");
 352     }
 353
 354     if (stacksize < min_stacksize) {
 355         error = pthread_attr_setstacksize(attr, min_stacksize);
 356         if (error) {
 357             ovs_abort(error, "pthread_attr_setstacksize failed");
 358         }
 359     }
 360 }
 361
 362 /* Starts a thread that calls 'start(arg)'.  Sets the thread's name to 'name'
 363  * (suffixed by its ovsthread_id()).  Returns the new thread's pthread_t. */
 364 pthread_t
 365 ovs_thread_create(const char *name, void *(*start)(void *), void *arg)
 366 {
 367     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 368     struct ovsthread_aux *aux;
 369     pthread_t thread;
 370     int error;
 371
 372     forbid_forking("multiple threads exist");
 373     multithreaded = true;
 374
 375     if (ovsthread_once_start(&once)) {
 376         /* The first call to this function has to happen in the main thread.
 377          * Before the process becomes multithreaded we make sure that the
 378          * main thread is considered non quiescent.
 379          *
 380          * For other threads this is done in ovs_thread_wrapper(), but the
 381          * main thread has no such wrapper.
 382          *
 383          * There's no reason to call ovsrcu_quiesce_end() in subsequent
 384          * invocations of this function and it might introduce problems
 385          * for other threads. */
 386         ovsrcu_quiesce_end();
 387         ovsthread_once_done(&once);
 388     }
 389
 390     aux = xmalloc(sizeof *aux);
 391     aux->start = start;
 392     aux->arg = arg;
 393     ovs_strlcpy(aux->name, name, sizeof aux->name);
 394
 395     /* Some small systems use a default stack size as small as 80 kB, but OVS
 396      * requires approximately 384 kB according to the following analysis:
 397      * http://openvswitch.org/pipermail/dev/2016-January/065049.html
 398      *
 399      * We use 512 kB to give us some margin of error. */
 400     pthread_attr_t attr;
 401     pthread_attr_init(&attr);
 402     set_min_stack_size(&attr, 512 * 1024);
 403
 404     error = pthread_create(&thread, &attr, ovsthread_wrapper, aux);
 405     if (error) {
 406         ovs_abort(error, "pthread_create failed");
 407     }
 408     pthread_attr_destroy(&attr);
 409     return thread;
 410 }
 411 \f
 412 bool
 413 ovsthread_once_start__(struct ovsthread_once *once)
 414 {
 415     ovs_mutex_lock(&once->mutex);
 416     /* Mutex synchronizes memory, so we get the current value of 'done'. */
 417     if (!once->done) {
 418         return true;
 419     }
 420     ovs_mutex_unlock(&once->mutex);
 421     return false;
 422 }
 423
 424 void
 425 ovsthread_once_done(struct ovsthread_once *once)
 426 {
 427     /* We need release semantics here, so that the following store may not
 428      * be moved ahead of any of the preceding initialization operations.
 429      * A release atomic_thread_fence provides that prior memory accesses
 430      * will not be reordered to take place after the following store. */
 431     atomic_thread_fence(memory_order_release);
 432     once->done = true;
 433     ovs_mutex_unlock(&once->mutex);
 434 }
 435 \f
 436 bool
 437 single_threaded(void)
 438 {
 439     return !multithreaded;
 440 }
 441
 442 /* Asserts that the process has not yet created any threads (beyond the initial
 443  * thread).
 444  *
 445  * ('where' is used in logging.  Commonly one would use
 446  * assert_single_threaded() to automatically provide the caller's source file
 447  * and line number for 'where'.) */
 448 void
 449 assert_single_threaded_at(const char *where)
 450 {
 451     if (multithreaded) {
 452         VLOG_FATAL("%s: attempted operation not allowed when multithreaded",
 453                    where);
 454     }
 455 }
 456
 457 #ifndef _WIN32
 458 /* Forks the current process (checking that this is allowed).  Aborts with
 459  * VLOG_FATAL if fork() returns an error, and otherwise returns the value
 460  * returned by fork().
 461  *
 462  * ('where' is used in logging.  Commonly one would use xfork() to
 463  * automatically provide the caller's source file and line number for
 464  * 'where'.) */
 465 pid_t
 466 xfork_at(const char *where)
 467 {
 468     pid_t pid;
 469
 470     if (must_not_fork) {
 471         VLOG_FATAL("%s: attempted to fork but forking not allowed (%s)",
 472                    where, must_not_fork);
 473     }
 474
 475     pid = fork();
 476     if (pid < 0) {
 477         VLOG_FATAL("%s: fork failed (%s)", where, ovs_strerror(errno));
 478     }
 479     return pid;
 480 }
 481 #endif
 482
 483 /* Notes that the process must not call fork() from now on, for the specified
 484  * 'reason'.  (The process may still fork() if it execs itself immediately
 485  * afterward.) */
 486 void
 487 forbid_forking(const char *reason)
 488 {
 489     ovs_assert(reason != NULL);
 490     must_not_fork = reason;
 491 }
 492
 493 /* Returns true if the process is allowed to fork, false otherwise. */
 494 bool
 495 may_fork(void)
 496 {
 497     return !must_not_fork;
 498 }
 499 \f
 500 /* ovsthread_stats. */
 501
 502 void
 503 ovsthread_stats_init(struct ovsthread_stats *stats)
 504 {
 505     int i;
 506
 507     ovs_mutex_init(&stats->mutex);
 508     for (i = 0; i < ARRAY_SIZE(stats->buckets); i++) {
 509         stats->buckets[i] = NULL;
 510     }
 511 }
 512
 513 void
 514 ovsthread_stats_destroy(struct ovsthread_stats *stats)
 515 {
 516     ovs_mutex_destroy(&stats->mutex);
 517 }
 518
 519 void *
 520 ovsthread_stats_bucket_get(struct ovsthread_stats *stats,
 521                            void *(*new_bucket)(void))
 522 {
 523     unsigned int idx = ovsthread_id_self() & (ARRAY_SIZE(stats->buckets) - 1);
 524     void *bucket = stats->buckets[idx];
 525     if (!bucket) {
 526         ovs_mutex_lock(&stats->mutex);
 527         bucket = stats->buckets[idx];
 528         if (!bucket) {
 529             bucket = stats->buckets[idx] = new_bucket();
 530         }
 531         ovs_mutex_unlock(&stats->mutex);
 532     }
 533     return bucket;
 534 }
 535
 536 size_t
 537 ovs_thread_stats_next_bucket(const struct ovsthread_stats *stats, size_t i)
 538 {
 539     for (; i < ARRAY_SIZE(stats->buckets); i++) {
 540         if (stats->buckets[i]) {
 541             break;
 542         }
 543     }
 544     return i;
 545 }
 546
 547 \f
 548 /* Parses /proc/cpuinfo for the total number of physical cores on this system
 549  * across all CPU packages, not counting hyper-threads.
 550  *
 551  * Sets *n_cores to the total number of cores on this system, or 0 if the
 552  * number cannot be determined. */
 553 static void
 554 parse_cpuinfo(long int *n_cores)
 555 {
 556     static const char file_name[] = "/proc/cpuinfo";
 557     char line[128];
 558     uint64_t cpu = 0; /* Support up to 64 CPU packages on a single system. */
 559     long int cores = 0;
 560     FILE *stream;
 561
 562     stream = fopen(file_name, "r");
 563     if (!stream) {
 564         VLOG_DBG("%s: open failed (%s)", file_name, ovs_strerror(errno));
 565         return;
 566     }
 567
 568     while (fgets(line, sizeof line, stream)) {
 569         unsigned int id;
 570
 571         /* Find the next CPU package. */
 572         if (ovs_scan(line, "physical id%*[^:]: %u", &id)) {
 573             if (id > 63) {
 574                 VLOG_WARN("Counted over 64 CPU packages on this system. "
 575                           "Parsing %s for core count may be inaccurate.",
 576                           file_name);
 577                 cores = 0;
 578                 break;
 579             }
 580
 581             if (cpu & (1ULL << id)) {
 582                 /* We've already counted this package's cores. */
 583                 continue;
 584             }
 585             cpu |= 1ULL << id;
 586
 587             /* Find the number of cores for this package. */
 588             while (fgets(line, sizeof line, stream)) {
 589                 int count;
 590
 591                 if (ovs_scan(line, "cpu cores%*[^:]: %u", &count)) {
 592                     cores += count;
 593                     break;
 594                 }
 595             }
 596         }
 597     }
 598     fclose(stream);
 599
 600     *n_cores = cores;
 601 }
 602
 603 /* Returns the total number of cores on this system, or 0 if the number cannot
 604  * be determined.
 605  *
 606  * Tries not to count hyper-threads, but may be inaccurate - particularly on
 607  * platforms that do not provide /proc/cpuinfo, but also if /proc/cpuinfo is
 608  * formatted different to the layout that parse_cpuinfo() expects. */
 609 int
 610 count_cpu_cores(void)
 611 {
 612     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 613     static long int n_cores;
 614
 615     if (ovsthread_once_start(&once)) {
 616 #ifndef _WIN32
 617         parse_cpuinfo(&n_cores);
 618         if (!n_cores) {
 619             n_cores = sysconf(_SC_NPROCESSORS_ONLN);
 620         }
 621 #else
 622         SYSTEM_INFO sysinfo;
 623         GetSystemInfo(&sysinfo);
 624         n_cores = sysinfo.dwNumberOfProcessors;
 625 #endif
 626         ovsthread_once_done(&once);
 627     }
 628
 629     return n_cores > 0 ? n_cores : 0;
 630 }
 631
 632 /* Returns 'true' if current thread is PMD thread. */
 633 bool
 634 thread_is_pmd(void)
 635 {
 636     const char *name = get_subprogram_name();
 637     return !strncmp(name, "pmd", 3);
 638 }
 639
 640 \f
 641 /* ovsthread_key. */
 642
 643 #define L1_SIZE 1024
 644 #define L2_SIZE 1024
 645 #define MAX_KEYS (L1_SIZE * L2_SIZE)
 646
 647 /* A piece of thread-specific data. */
 648 struct ovsthread_key {
 649     struct ovs_list list_node;  /* In 'inuse_keys' or 'free_keys'. */
 650     void (*destructor)(void *); /* Called at thread exit. */
 651
 652     /* Indexes into the per-thread array in struct ovsthread_key_slots.
 653      * This key's data is stored in p1[index / L2_SIZE][index % L2_SIZE]. */
 654     unsigned int index;
 655 };
 656
 657 /* Per-thread data structure. */
 658 struct ovsthread_key_slots {
 659     struct ovs_list list_node;  /* In 'slots_list'. */
 660     void **p1[L1_SIZE];
 661 };
 662
 663 /* Contains "struct ovsthread_key_slots *". */
 664 static pthread_key_t tsd_key;
 665
 666 /* Guards data structures below. */
 667 static struct ovs_mutex key_mutex = OVS_MUTEX_INITIALIZER;
 668
 669 /* 'inuse_keys' holds "struct ovsthread_key"s that have been created and not
 670  * yet destroyed.
 671  *
 672  * 'free_keys' holds "struct ovsthread_key"s that have been deleted and are
 673  * ready for reuse.  (We keep them around only to be able to easily locate
 674  * free indexes.)
 675  *
 676  * Together, 'inuse_keys' and 'free_keys' hold an ovsthread_key for every index
 677  * from 0 to n_keys - 1, inclusive. */
 678 static struct ovs_list inuse_keys OVS_GUARDED_BY(key_mutex)
 679     = OVS_LIST_INITIALIZER(&inuse_keys);
 680 static struct ovs_list free_keys OVS_GUARDED_BY(key_mutex)
 681     = OVS_LIST_INITIALIZER(&free_keys);
 682 static unsigned int n_keys OVS_GUARDED_BY(key_mutex);
 683
 684 /* All existing struct ovsthread_key_slots. */
 685 static struct ovs_list slots_list OVS_GUARDED_BY(key_mutex)
 686     = OVS_LIST_INITIALIZER(&slots_list);
 687
 688 static void *
 689 clear_slot(struct ovsthread_key_slots *slots, unsigned int index)
 690 {
 691     void **p2 = slots->p1[index / L2_SIZE];
 692     if (p2) {
 693         void **valuep = &p2[index % L2_SIZE];
 694         void *value = *valuep;
 695         *valuep = NULL;
 696         return value;
 697     } else {
 698         return NULL;
 699     }
 700 }
 701
 702 static void
 703 ovsthread_key_destruct__(void *slots_)
 704 {
 705     struct ovsthread_key_slots *slots = slots_;
 706     struct ovsthread_key *key;
 707     unsigned int n;
 708     int i;
 709
 710     ovs_mutex_lock(&key_mutex);
 711     list_remove(&slots->list_node);
 712     LIST_FOR_EACH (key, list_node, &inuse_keys) {
 713         void *value = clear_slot(slots, key->index);
 714         if (value && key->destructor) {
 715             key->destructor(value);
 716         }
 717     }
 718     n = n_keys;
 719     ovs_mutex_unlock(&key_mutex);
 720
 721     for (i = 0; i < DIV_ROUND_UP(n, L2_SIZE); i++) {
 722         free(slots->p1[i]);
 723     }
 724     free(slots);
 725 }
 726
 727 /* Cancels the callback to ovsthread_key_destruct__().
 728  *
 729  * Cancelling the call to the destructor during the main thread exit
 730  * is needed while using pthreads-win32 library in Windows. It has been
 731  * observed that in pthreads-win32, a call to the destructor during
 732  * main thread exit causes undefined behavior. */
 733 static void
 734 ovsthread_cancel_ovsthread_key_destruct__(void *aux OVS_UNUSED)
 735 {
 736     pthread_setspecific(tsd_key, NULL);
 737 }
 738
 739 /* Initializes '*keyp' as a thread-specific data key.  The data items are
 740  * initially null in all threads.
 741  *
 742  * If a thread exits with non-null data, then 'destructor', if nonnull, will be
 743  * called passing the final data value as its argument.  'destructor' must not
 744  * call any thread-specific data functions in this API.
 745  *
 746  * This function is similar to xpthread_key_create(). */
 747 void
 748 ovsthread_key_create(ovsthread_key_t *keyp, void (*destructor)(void *))
 749 {
 750     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 751     struct ovsthread_key *key;
 752
 753     if (ovsthread_once_start(&once)) {
 754         xpthread_key_create(&tsd_key, ovsthread_key_destruct__);
 755         fatal_signal_add_hook(ovsthread_cancel_ovsthread_key_destruct__,
 756                               NULL, NULL, true);
 757         ovsthread_once_done(&once);
 758     }
 759
 760     ovs_mutex_lock(&key_mutex);
 761     if (list_is_empty(&free_keys)) {
 762         key = xmalloc(sizeof *key);
 763         key->index = n_keys++;
 764         if (key->index >= MAX_KEYS) {
 765             abort();
 766         }
 767     } else {
 768         key = CONTAINER_OF(list_pop_back(&free_keys),
 769                             struct ovsthread_key, list_node);
 770     }
 771     list_push_back(&inuse_keys, &key->list_node);
 772     key->destructor = destructor;
 773     ovs_mutex_unlock(&key_mutex);
 774
 775     *keyp = key;
 776 }
 777
 778 /* Frees 'key'.  The destructor supplied to ovsthread_key_create(), if any, is
 779  * not called.
 780  *
 781  * This function is similar to xpthread_key_delete(). */
 782 void
 783 ovsthread_key_delete(ovsthread_key_t key)
 784 {
 785     struct ovsthread_key_slots *slots;
 786
 787     ovs_mutex_lock(&key_mutex);
 788
 789     /* Move 'key' from 'inuse_keys' to 'free_keys'. */
 790     list_remove(&key->list_node);
 791     list_push_back(&free_keys, &key->list_node);
 792
 793     /* Clear this slot in all threads. */
 794     LIST_FOR_EACH (slots, list_node, &slots_list) {
 795         clear_slot(slots, key->index);
 796     }
 797
 798     ovs_mutex_unlock(&key_mutex);
 799 }
 800
 801 static void **
 802 ovsthread_key_lookup__(const struct ovsthread_key *key)
 803 {
 804     struct ovsthread_key_slots *slots;
 805     void **p2;
 806
 807     slots = pthread_getspecific(tsd_key);
 808     if (!slots) {
 809         slots = xzalloc(sizeof *slots);
 810
 811         ovs_mutex_lock(&key_mutex);
 812         pthread_setspecific(tsd_key, slots);
 813         list_push_back(&slots_list, &slots->list_node);
 814         ovs_mutex_unlock(&key_mutex);
 815     }
 816
 817     p2 = slots->p1[key->index / L2_SIZE];
 818     if (!p2) {
 819         p2 = xzalloc(L2_SIZE * sizeof *p2);
 820         slots->p1[key->index / L2_SIZE] = p2;
 821     }
 822
 823     return &p2[key->index % L2_SIZE];
 824 }
 825
 826 /* Sets the value of thread-specific data item 'key', in the current thread, to
 827  * 'value'.
 828  *
 829  * This function is similar to pthread_setspecific(). */
 830 void
 831 ovsthread_setspecific(ovsthread_key_t key, const void *value)
 832 {
 833     *ovsthread_key_lookup__(key) = CONST_CAST(void *, value);
 834 }
 835
 836 /* Returns the value of thread-specific data item 'key' in the current thread.
 837  *
 838  * This function is similar to pthread_getspecific(). */
 839 void *
 840 ovsthread_getspecific(ovsthread_key_t key)
 841 {
 842     return *ovsthread_key_lookup__(key);
 843 }
 844 #endif