libgomp/team.c

   1 /* Copyright (C) 2005, 2006, 2007, 2008, 2009, 2011, 2012
   2    Free Software Foundation, Inc.
   3    Contributed by Richard Henderson <rth@redhat.com>.
   4
   5    This file is part of the GNU OpenMP Library (libgomp).
   6
   7    Libgomp is free software; you can redistribute it and/or modify it
   8    under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3, or (at your option)
  10    any later version.
  11
  12    Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
  13    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
  14    FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  15    more details.
  16
  17    Under Section 7 of GPL version 3, you are granted additional
  18    permissions described in the GCC Runtime Library Exception, version
  19    3.1, as published by the Free Software Foundation.
  20
  21    You should have received a copy of the GNU General Public License and
  22    a copy of the GCC Runtime Library Exception along with this program;
  23    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  24    <http://www.gnu.org/licenses/>.  */
  25
  26 /* This file handles the maintainence of threads in response to team
  27    creation and termination.  */
  28
  29 #include "libgomp.h"
  30 #include <stdlib.h>
  31 #include <string.h>
  32
  33 /* This attribute contains PTHREAD_CREATE_DETACHED.  */
  34 pthread_attr_t gomp_thread_attr;
  35
  36 /* This key is for the thread destructor.  */
  37 pthread_key_t gomp_thread_destructor;
  38
  39
  40 /* This is the libgomp per-thread data structure.  */
  41 #ifdef HAVE_TLS
  42 __thread struct gomp_thread gomp_tls_data;
  43 #else
  44 pthread_key_t gomp_tls_key;
  45 #endif
  46
  47
  48 /* This structure is used to communicate across pthread_create.  */
  49
  50 struct gomp_thread_start_data
  51 {
  52   void (*fn) (void *);
  53   void *fn_data;
  54   struct gomp_team_state ts;
  55   struct gomp_task *task;
  56   struct gomp_thread_pool *thread_pool;
  57   bool nested;
  58 };
  59
  60
  61 /* This function is a pthread_create entry point.  This contains the idle
  62    loop in which a thread waits to be called up to become part of a team.  */
  63
  64 static void *
  65 gomp_thread_start (void *xdata)
  66 {
  67   struct gomp_thread_start_data *data = xdata;
  68   struct gomp_thread *thr;
  69   struct gomp_thread_pool *pool;
  70   void (*local_fn) (void *);
  71   void *local_data;
  72
  73 #ifdef HAVE_TLS
  74   thr = &gomp_tls_data;
  75 #else
  76   struct gomp_thread local_thr;
  77   thr = &local_thr;
  78   pthread_setspecific (gomp_tls_key, thr);
  79 #endif
  80   gomp_sem_init (&thr->release, 0);
  81
  82   /* Extract what we need from data.  */
  83   local_fn = data->fn;
  84   local_data = data->fn_data;
  85   thr->thread_pool = data->thread_pool;
  86   thr->ts = data->ts;
  87   thr->task = data->task;
  88
  89   thr->ts.team->ordered_release[thr->ts.team_id] = &thr->release;
  90
  91   /* Make thread pool local. */
  92   pool = thr->thread_pool;
  93
  94   if (data->nested)
  95     {
  96       struct gomp_team *team = thr->ts.team;
  97       struct gomp_task *task = thr->task;
  98
  99       gomp_barrier_wait (&team->barrier);
 100
 101       local_fn (local_data);
 102       gomp_team_barrier_wait (&team->barrier);
 103       gomp_finish_task (task);
 104       gomp_barrier_wait_last (&team->barrier);
 105     }
 106   else
 107     {
 108       pool->threads[thr->ts.team_id] = thr;
 109
 110       gomp_barrier_wait (&pool->threads_dock);
 111       do
 112         {
 113           struct gomp_team *team = thr->ts.team;
 114           struct gomp_task *task = thr->task;
 115
 116           local_fn (local_data);
 117           gomp_team_barrier_wait (&team->barrier);
 118           gomp_finish_task (task);
 119
 120           gomp_barrier_wait (&pool->threads_dock);
 121
 122           local_fn = thr->fn;
 123           local_data = thr->data;
 124           thr->fn = NULL;
 125         }
 126       while (local_fn);
 127     }
 128
 129   gomp_sem_destroy (&thr->release);
 130   return NULL;
 131 }
 132
 133
 134 /* Create a new team data structure.  */
 135
 136 struct gomp_team *
 137 gomp_new_team (unsigned nthreads)
 138 {
 139   struct gomp_team *team;
 140   size_t size;
 141   int i;
 142
 143   size = sizeof (*team) + nthreads * (sizeof (team->ordered_release[0])
 144                                       + sizeof (team->implicit_task[0]));
 145   team = gomp_malloc (size);
 146
 147   team->work_share_chunk = 8;
 148 #ifdef HAVE_SYNC_BUILTINS
 149   team->single_count = 0;
 150 #else
 151   gomp_mutex_init (&team->work_share_list_free_lock);
 152 #endif
 153   gomp_init_work_share (&team->work_shares[0], false, nthreads);
 154   team->work_shares[0].next_alloc = NULL;
 155   team->work_share_list_free = NULL;
 156   team->work_share_list_alloc = &team->work_shares[1];
 157   for (i = 1; i < 7; i++)
 158     team->work_shares[i].next_free = &team->work_shares[i + 1];
 159   team->work_shares[i].next_free = NULL;
 160
 161   team->nthreads = nthreads;
 162   gomp_barrier_init (&team->barrier, nthreads);
 163
 164   gomp_sem_init (&team->master_release, 0);
 165   team->ordered_release = (void *) &team->implicit_task[nthreads];
 166   team->ordered_release[0] = &team->master_release;
 167
 168   gomp_mutex_init (&team->task_lock);
 169   team->task_queue = NULL;
 170   team->task_count = 0;
 171   team->task_running_count = 0;
 172
 173   return team;
 174 }
 175
 176
 177 /* Free a team data structure.  */
 178
 179 static void
 180 free_team (struct gomp_team *team)
 181 {
 182   gomp_barrier_destroy (&team->barrier);
 183   gomp_mutex_destroy (&team->task_lock);
 184   free (team);
 185 }
 186
 187 /* Allocate and initialize a thread pool. */
 188
 189 static struct gomp_thread_pool *gomp_new_thread_pool (void)
 190 {
 191   struct gomp_thread_pool *pool
 192     = gomp_malloc (sizeof(struct gomp_thread_pool));
 193   pool->threads = NULL;
 194   pool->threads_size = 0;
 195   pool->threads_used = 0;
 196   pool->last_team = NULL;
 197   return pool;
 198 }
 199
 200 static void
 201 gomp_free_pool_helper (void *thread_pool)
 202 {
 203   struct gomp_thread_pool *pool
 204     = (struct gomp_thread_pool *) thread_pool;
 205   gomp_barrier_wait_last (&pool->threads_dock);
 206   gomp_sem_destroy (&gomp_thread ()->release);
 207   pthread_exit (NULL);
 208 }
 209
 210 /* Free a thread pool and release its threads. */
 211
 212 static void
 213 gomp_free_thread (void *arg __attribute__((unused)))
 214 {
 215   struct gomp_thread *thr = gomp_thread ();
 216   struct gomp_thread_pool *pool = thr->thread_pool;
 217   if (pool)
 218     {
 219       if (pool->threads_used > 0)
 220         {
 221           int i;
 222           for (i = 1; i < pool->threads_used; i++)
 223             {
 224               struct gomp_thread *nthr = pool->threads[i];
 225               nthr->fn = gomp_free_pool_helper;
 226               nthr->data = pool;
 227             }
 228           /* This barrier undocks threads docked on pool->threads_dock.  */
 229           gomp_barrier_wait (&pool->threads_dock);
 230           /* And this waits till all threads have called gomp_barrier_wait_last
 231              in gomp_free_pool_helper.  */
 232           gomp_barrier_wait (&pool->threads_dock);
 233           /* Now it is safe to destroy the barrier and free the pool.  */
 234           gomp_barrier_destroy (&pool->threads_dock);
 235
 236 #ifdef HAVE_SYNC_BUILTINS
 237           __sync_fetch_and_add (&gomp_managed_threads,
 238                                 1L - pool->threads_used);
 239 #else
 240           gomp_mutex_lock (&gomp_remaining_threads_lock);
 241           gomp_managed_threads -= pool->threads_used - 1L;
 242           gomp_mutex_unlock (&gomp_remaining_threads_lock);
 243 #endif
 244         }
 245       free (pool->threads);
 246       if (pool->last_team)
 247         free_team (pool->last_team);
 248       free (pool);
 249       thr->thread_pool = NULL;
 250     }
 251   if (thr->task != NULL)
 252     {
 253       struct gomp_task *task = thr->task;
 254       gomp_end_task ();
 255       free (task);
 256     }
 257 }
 258
 259 /* Launch a team.  */
 260
 261 void
 262 gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
 263                  struct gomp_team *team)
 264 {
 265   struct gomp_thread_start_data *start_data;
 266   struct gomp_thread *thr, *nthr;
 267   struct gomp_task *task;
 268   struct gomp_task_icv *icv;
 269   bool nested;
 270   struct gomp_thread_pool *pool;
 271   unsigned i, n, old_threads_used = 0;
 272   pthread_attr_t thread_attr, *attr;
 273   unsigned long nthreads_var;
 274
 275   thr = gomp_thread ();
 276   nested = thr->ts.team != NULL;
 277   if (__builtin_expect (thr->thread_pool == NULL, 0))
 278     {
 279       thr->thread_pool = gomp_new_thread_pool ();
 280       pthread_setspecific (gomp_thread_destructor, thr);
 281     }
 282   pool = thr->thread_pool;
 283   task = thr->task;
 284   icv = task ? &task->icv : &gomp_global_icv;
 285
 286   /* Always save the previous state, even if this isn't a nested team.
 287      In particular, we should save any work share state from an outer
 288      orphaned work share construct.  */
 289   team->prev_ts = thr->ts;
 290
 291   thr->ts.team = team;
 292   thr->ts.team_id = 0;
 293   ++thr->ts.level;
 294   if (nthreads > 1)
 295     ++thr->ts.active_level;
 296   thr->ts.work_share = &team->work_shares[0];
 297   thr->ts.last_work_share = NULL;
 298 #ifdef HAVE_SYNC_BUILTINS
 299   thr->ts.single_count = 0;
 300 #endif
 301   thr->ts.static_trip = 0;
 302   thr->task = &team->implicit_task[0];
 303   nthreads_var = icv->nthreads_var;
 304   if (__builtin_expect (gomp_nthreads_var_list != NULL, 0)
 305       && thr->ts.level < gomp_nthreads_var_list_len)
 306     nthreads_var = gomp_nthreads_var_list[thr->ts.level];
 307   gomp_init_task (thr->task, task, icv);
 308   team->implicit_task[0].icv.nthreads_var = nthreads_var;
 309
 310   if (nthreads == 1)
 311     return;
 312
 313   i = 1;
 314
 315   /* We only allow the reuse of idle threads for non-nested PARALLEL
 316      regions.  This appears to be implied by the semantics of
 317      threadprivate variables, but perhaps that's reading too much into
 318      things.  Certainly it does prevent any locking problems, since
 319      only the initial program thread will modify gomp_threads.  */
 320   if (!nested)
 321     {
 322       old_threads_used = pool->threads_used;
 323
 324       if (nthreads <= old_threads_used)
 325         n = nthreads;
 326       else if (old_threads_used == 0)
 327         {
 328           n = 0;
 329           gomp_barrier_init (&pool->threads_dock, nthreads);
 330         }
 331       else
 332         {
 333           n = old_threads_used;
 334
 335           /* Increase the barrier threshold to make sure all new
 336              threads arrive before the team is released.  */
 337           gomp_barrier_reinit (&pool->threads_dock, nthreads);
 338         }
 339
 340       /* Not true yet, but soon will be.  We're going to release all
 341          threads from the dock, and those that aren't part of the
 342          team will exit.  */
 343       pool->threads_used = nthreads;
 344
 345       /* Release existing idle threads.  */
 346       for (; i < n; ++i)
 347         {
 348           nthr = pool->threads[i];
 349           nthr->ts.team = team;
 350           nthr->ts.work_share = &team->work_shares[0];
 351           nthr->ts.last_work_share = NULL;
 352           nthr->ts.team_id = i;
 353           nthr->ts.level = team->prev_ts.level + 1;
 354           nthr->ts.active_level = thr->ts.active_level;
 355 #ifdef HAVE_SYNC_BUILTINS
 356           nthr->ts.single_count = 0;
 357 #endif
 358           nthr->ts.static_trip = 0;
 359           nthr->task = &team->implicit_task[i];
 360           gomp_init_task (nthr->task, task, icv);
 361           team->implicit_task[i].icv.nthreads_var = nthreads_var;
 362           nthr->fn = fn;
 363           nthr->data = data;
 364           team->ordered_release[i] = &nthr->release;
 365         }
 366
 367       if (i == nthreads)
 368         goto do_release;
 369
 370       /* If necessary, expand the size of the gomp_threads array.  It is
 371          expected that changes in the number of threads are rare, thus we
 372          make no effort to expand gomp_threads_size geometrically.  */
 373       if (nthreads >= pool->threads_size)
 374         {
 375           pool->threads_size = nthreads + 1;
 376           pool->threads
 377             = gomp_realloc (pool->threads,
 378                             pool->threads_size
 379                             * sizeof (struct gomp_thread_data *));
 380         }
 381     }
 382
 383   if (__builtin_expect (nthreads > old_threads_used, 0))
 384     {
 385       long diff = (long) nthreads - (long) old_threads_used;
 386
 387       if (old_threads_used == 0)
 388         --diff;
 389
 390 #ifdef HAVE_SYNC_BUILTINS
 391       __sync_fetch_and_add (&gomp_managed_threads, diff);
 392 #else
 393       gomp_mutex_lock (&gomp_remaining_threads_lock);
 394       gomp_managed_threads += diff;
 395       gomp_mutex_unlock (&gomp_remaining_threads_lock);
 396 #endif
 397     }
 398
 399   attr = &gomp_thread_attr;
 400   if (__builtin_expect (gomp_cpu_affinity != NULL, 0))
 401     {
 402       size_t stacksize;
 403       pthread_attr_init (&thread_attr);
 404       pthread_attr_setdetachstate (&thread_attr, PTHREAD_CREATE_DETACHED);
 405       if (! pthread_attr_getstacksize (&gomp_thread_attr, &stacksize))
 406         pthread_attr_setstacksize (&thread_attr, stacksize);
 407       attr = &thread_attr;
 408     }
 409
 410   start_data = gomp_alloca (sizeof (struct gomp_thread_start_data)
 411                             * (nthreads-i));
 412
 413   /* Launch new threads.  */
 414   for (; i < nthreads; ++i, ++start_data)
 415     {
 416       pthread_t pt;
 417       int err;
 418
 419       start_data->fn = fn;
 420       start_data->fn_data = data;
 421       start_data->ts.team = team;
 422       start_data->ts.work_share = &team->work_shares[0];
 423       start_data->ts.last_work_share = NULL;
 424       start_data->ts.team_id = i;
 425       start_data->ts.level = team->prev_ts.level + 1;
 426       start_data->ts.active_level = thr->ts.active_level;
 427 #ifdef HAVE_SYNC_BUILTINS
 428       start_data->ts.single_count = 0;
 429 #endif
 430       start_data->ts.static_trip = 0;
 431       start_data->task = &team->implicit_task[i];
 432       gomp_init_task (start_data->task, task, icv);
 433       team->implicit_task[i].icv.nthreads_var = nthreads_var;
 434       start_data->thread_pool = pool;
 435       start_data->nested = nested;
 436
 437       if (gomp_cpu_affinity != NULL)
 438         gomp_init_thread_affinity (attr);
 439
 440       err = pthread_create (&pt, attr, gomp_thread_start, start_data);
 441       if (err != 0)
 442         gomp_fatal ("Thread creation failed: %s", strerror (err));
 443     }
 444
 445   if (__builtin_expect (gomp_cpu_affinity != NULL, 0))
 446     pthread_attr_destroy (&thread_attr);
 447
 448  do_release:
 449   gomp_barrier_wait (nested ? &team->barrier : &pool->threads_dock);
 450
 451   /* Decrease the barrier threshold to match the number of threads
 452      that should arrive back at the end of this team.  The extra
 453      threads should be exiting.  Note that we arrange for this test
 454      to never be true for nested teams.  */
 455   if (__builtin_expect (nthreads < old_threads_used, 0))
 456     {
 457       long diff = (long) nthreads - (long) old_threads_used;
 458
 459       gomp_barrier_reinit (&pool->threads_dock, nthreads);
 460
 461 #ifdef HAVE_SYNC_BUILTINS
 462       __sync_fetch_and_add (&gomp_managed_threads, diff);
 463 #else
 464       gomp_mutex_lock (&gomp_remaining_threads_lock);
 465       gomp_managed_threads += diff;
 466       gomp_mutex_unlock (&gomp_remaining_threads_lock);
 467 #endif
 468     }
 469 }
 470
 471
 472 /* Terminate the current team.  This is only to be called by the master
 473    thread.  We assume that we must wait for the other threads.  */
 474
 475 void
 476 gomp_team_end (void)
 477 {
 478   struct gomp_thread *thr = gomp_thread ();
 479   struct gomp_team *team = thr->ts.team;
 480
 481   /* This barrier handles all pending explicit threads.  */
 482   gomp_team_barrier_wait (&team->barrier);
 483   gomp_fini_work_share (thr->ts.work_share);
 484
 485   gomp_end_task ();
 486   thr->ts = team->prev_ts;
 487
 488   if (__builtin_expect (thr->ts.team != NULL, 0))
 489     {
 490 #ifdef HAVE_SYNC_BUILTINS
 491       __sync_fetch_and_add (&gomp_managed_threads, 1L - team->nthreads);
 492 #else
 493       gomp_mutex_lock (&gomp_remaining_threads_lock);
 494       gomp_managed_threads -= team->nthreads - 1L;
 495       gomp_mutex_unlock (&gomp_remaining_threads_lock);
 496 #endif
 497       /* This barrier has gomp_barrier_wait_last counterparts
 498          and ensures the team can be safely destroyed.  */
 499       gomp_barrier_wait (&team->barrier);
 500     }
 501
 502   if (__builtin_expect (team->work_shares[0].next_alloc != NULL, 0))
 503     {
 504       struct gomp_work_share *ws = team->work_shares[0].next_alloc;
 505       do
 506         {
 507           struct gomp_work_share *next_ws = ws->next_alloc;
 508           free (ws);
 509           ws = next_ws;
 510         }
 511       while (ws != NULL);
 512     }
 513   gomp_sem_destroy (&team->master_release);
 514 #ifndef HAVE_SYNC_BUILTINS
 515   gomp_mutex_destroy (&team->work_share_list_free_lock);
 516 #endif
 517
 518   if (__builtin_expect (thr->ts.team != NULL, 0)
 519       || __builtin_expect (team->nthreads == 1, 0))
 520     free_team (team);
 521   else
 522     {
 523       struct gomp_thread_pool *pool = thr->thread_pool;
 524       if (pool->last_team)
 525         free_team (pool->last_team);
 526       pool->last_team = team;
 527     }
 528 }
 529
 530
 531 /* Constructors for this file.  */
 532
 533 static void __attribute__((constructor))
 534 initialize_team (void)
 535 {
 536   struct gomp_thread *thr;
 537
 538 #ifndef HAVE_TLS
 539   static struct gomp_thread initial_thread_tls_data;
 540
 541   pthread_key_create (&gomp_tls_key, NULL);
 542   pthread_setspecific (gomp_tls_key, &initial_thread_tls_data);
 543 #endif
 544
 545   if (pthread_key_create (&gomp_thread_destructor, gomp_free_thread) != 0)
 546     gomp_fatal ("could not create thread pool destructor.");
 547
 548 #ifdef HAVE_TLS
 549   thr = &gomp_tls_data;
 550 #else
 551   thr = &initial_thread_tls_data;
 552 #endif
 553   gomp_sem_init (&thr->release, 0);
 554 }
 555
 556 static void __attribute__((destructor))
 557 team_destructor (void)
 558 {
 559   /* Without this dlclose on libgomp could lead to subsequent
 560      crashes.  */
 561   pthread_key_delete (gomp_thread_destructor);
 562 }
 563
 564 struct gomp_task_icv *
 565 gomp_new_icv (void)
 566 {
 567   struct gomp_thread *thr = gomp_thread ();
 568   struct gomp_task *task = gomp_malloc (sizeof (struct gomp_task));
 569   gomp_init_task (task, NULL, &gomp_global_icv);
 570   thr->task = task;
 571   pthread_setspecific (gomp_thread_destructor, thr);
 572   return &task->icv;
 573 }