module/spl/spl-taskq.c

   1 /*****************************************************************************\
   2  *  Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
   3  *  Copyright (C) 2007 The Regents of the University of California.
   4  *  Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
   5  *  Written by Brian Behlendorf <behlendorf1@llnl.gov>.
   6  *  UCRL-CODE-235197
   7  *
   8  *  This file is part of the SPL, Solaris Porting Layer.
   9  *  For details, see <http://github.com/behlendorf/spl/>.
  10  *
  11  *  The SPL is free software; you can redistribute it and/or modify it
  12  *  under the terms of the GNU General Public License as published by the
  13  *  Free Software Foundation; either version 2 of the License, or (at your
  14  *  option) any later version.
  15  *
  16  *  The SPL is distributed in the hope that it will be useful, but WITHOUT
  17  *  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  18  *  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  19  *  for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License along
  22  *  with the SPL.  If not, see <http://www.gnu.org/licenses/>.
  23  *****************************************************************************
  24  *  Solaris Porting Layer (SPL) Task Queue Implementation.
  25 \*****************************************************************************/
  26
  27 #include <sys/taskq.h>
  28 #include <sys/kmem.h>
  29 #include <spl-debug.h>
  30
  31 #ifdef SS_DEBUG_SUBSYS
  32 #undef SS_DEBUG_SUBSYS
  33 #endif
  34
  35 #define SS_DEBUG_SUBSYS SS_TASKQ
  36
  37 /* Global system-wide dynamic task queue available for all consumers */
  38 taskq_t *system_taskq;
  39 EXPORT_SYMBOL(system_taskq);
  40
  41 /*
  42  * NOTE: Must be called with tq->tq_lock held, returns a list_t which
  43  * is not attached to the free, work, or pending taskq lists.
  44  */
  45 static taskq_ent_t *
  46 task_alloc(taskq_t *tq, uint_t flags)
  47 {
  48         taskq_ent_t *t;
  49         int count = 0;
  50         SENTRY;
  51
  52         ASSERT(tq);
  53         ASSERT(flags & (TQ_SLEEP | TQ_NOSLEEP));               /* One set */
  54         ASSERT(!((flags & TQ_SLEEP) && (flags & TQ_NOSLEEP))); /* Not both */
  55         ASSERT(spin_is_locked(&tq->tq_lock));
  56 retry:
  57         /* Acquire taskq_ent_t's from free list if available */
  58         if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
  59                 t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
  60
  61                 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
  62
  63                 list_del_init(&t->tqent_list);
  64                 SRETURN(t);
  65         }
  66
  67         /* Free list is empty and memory allocations are prohibited */
  68         if (flags & TQ_NOALLOC)
  69                 SRETURN(NULL);
  70
  71         /* Hit maximum taskq_ent_t pool size */
  72         if (tq->tq_nalloc >= tq->tq_maxalloc) {
  73                 if (flags & TQ_NOSLEEP)
  74                         SRETURN(NULL);
  75
  76                 /*
  77                  * Sleep periodically polling the free list for an available
  78                  * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
  79                  * but we cannot block forever waiting for an taskq_entq_t to
  80                  * show up in the free list, otherwise a deadlock can happen.
  81                  *
  82                  * Therefore, we need to allocate a new task even if the number
  83                  * of allocated tasks is above tq->tq_maxalloc, but we still
  84                  * end up delaying the task allocation by one second, thereby
  85                  * throttling the task dispatch rate.
  86                  */
  87                  spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
  88                  schedule_timeout(HZ / 100);
  89                  spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
  90                  if (count < 100)
  91                         SGOTO(retry, count++);
  92         }
  93
  94         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
  95         t = kmem_alloc(sizeof(taskq_ent_t), flags & (TQ_SLEEP | TQ_NOSLEEP));
  96         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
  97
  98         if (t) {
  99                 taskq_init_ent(t);
 100                 tq->tq_nalloc++;
 101         }
 102
 103         SRETURN(t);
 104 }
 105
 106 /*
 107  * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t
 108  * to already be removed from the free, work, or pending taskq lists.
 109  */
 110 static void
 111 task_free(taskq_t *tq, taskq_ent_t *t)
 112 {
 113         SENTRY;
 114
 115         ASSERT(tq);
 116         ASSERT(t);
 117         ASSERT(spin_is_locked(&tq->tq_lock));
 118         ASSERT(list_empty(&t->tqent_list));
 119
 120         kmem_free(t, sizeof(taskq_ent_t));
 121         tq->tq_nalloc--;
 122
 123         SEXIT;
 124 }
 125
 126 /*
 127  * NOTE: Must be called with tq->tq_lock held, either destroys the
 128  * taskq_ent_t if too many exist or moves it to the free list for later use.
 129  */
 130 static void
 131 task_done(taskq_t *tq, taskq_ent_t *t)
 132 {
 133         SENTRY;
 134         ASSERT(tq);
 135         ASSERT(t);
 136         ASSERT(spin_is_locked(&tq->tq_lock));
 137
 138         /* For prealloc'd tasks, we don't free anything. */
 139         if ((!(tq->tq_flags & TASKQ_DYNAMIC)) &&
 140             (t->tqent_flags & TQENT_FLAG_PREALLOC))
 141                 return;
 142
 143         list_del_init(&t->tqent_list);
 144
 145         if (tq->tq_nalloc <= tq->tq_minalloc) {
 146                 t->tqent_id = 0;
 147                 t->tqent_func = NULL;
 148                 t->tqent_arg = NULL;
 149                 t->tqent_flags = 0;
 150                 list_add_tail(&t->tqent_list, &tq->tq_free_list);
 151         } else {
 152                 task_free(tq, t);
 153         }
 154
 155         SEXIT;
 156 }
 157
 158 /*
 159  * As tasks are submitted to the task queue they are assigned a
 160  * monotonically increasing taskqid and added to the tail of the pending
 161  * list.  As worker threads become available the tasks are removed from
 162  * the head of the pending or priority list, giving preference to the
 163  * priority list.  The tasks are then added to the work list, preserving
 164  * the ordering by taskqid.  Finally, as tasks complete they are removed
 165  * from the work list.  This means that the pending and work lists are
 166  * always kept sorted by taskqid.  Thus the lowest outstanding
 167  * incomplete taskqid can be determined simply by checking the min
 168  * taskqid for each head item on the pending, priority, and work list.
 169  * This value is stored in tq->tq_lowest_id and only updated to the new
 170  * lowest id when the previous lowest id completes.  All taskqids lower
 171  * than tq->tq_lowest_id must have completed.  It is also possible
 172  * larger taskqid's have completed because they may be processed in
 173  * parallel by several worker threads.  However, this is not a problem
 174  * because the behavior of taskq_wait_id() is to block until all
 175  * previously submitted taskqid's have completed.
 176  *
 177  * XXX: Taskqid_t wrapping is not handled.  However, taskqid_t's are
 178  * 64-bit values so even if a taskq is processing 2^24 (16,777,216)
 179  * taskqid_ts per second it will still take 2^40 seconds, 34,865 years,
 180  * before the wrap occurs.  I can live with that for now.
 181  */
 182 static int
 183 taskq_wait_check(taskq_t *tq, taskqid_t id)
 184 {
 185         int rc;
 186
 187         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 188         rc = (id < tq->tq_lowest_id);
 189         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 190
 191         SRETURN(rc);
 192 }
 193
 194 void
 195 __taskq_wait_id(taskq_t *tq, taskqid_t id)
 196 {
 197         SENTRY;
 198         ASSERT(tq);
 199
 200         wait_event(tq->tq_wait_waitq, taskq_wait_check(tq, id));
 201
 202         SEXIT;
 203 }
 204 EXPORT_SYMBOL(__taskq_wait_id);
 205
 206 void
 207 __taskq_wait(taskq_t *tq)
 208 {
 209         taskqid_t id;
 210         SENTRY;
 211         ASSERT(tq);
 212
 213         /* Wait for the largest outstanding taskqid */
 214         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 215         id = tq->tq_next_id - 1;
 216         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 217
 218         __taskq_wait_id(tq, id);
 219
 220         SEXIT;
 221
 222 }
 223 EXPORT_SYMBOL(__taskq_wait);
 224
 225 int
 226 __taskq_member(taskq_t *tq, void *t)
 227 {
 228         struct list_head *l;
 229         taskq_thread_t *tqt;
 230         SENTRY;
 231
 232         ASSERT(tq);
 233         ASSERT(t);
 234
 235         list_for_each(l, &tq->tq_thread_list) {
 236                 tqt = list_entry(l, taskq_thread_t, tqt_thread_list);
 237                 if (tqt->tqt_thread == (struct task_struct *)t)
 238                         SRETURN(1);
 239         }
 240
 241         SRETURN(0);
 242 }
 243 EXPORT_SYMBOL(__taskq_member);
 244
 245 taskqid_t
 246 __taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
 247 {
 248         taskq_ent_t *t;
 249         taskqid_t rc = 0;
 250         SENTRY;
 251
 252         ASSERT(tq);
 253         ASSERT(func);
 254
 255         /* Solaris assumes TQ_SLEEP if not passed explicitly */
 256         if (!(flags & (TQ_SLEEP | TQ_NOSLEEP)))
 257                 flags |= TQ_SLEEP;
 258
 259         if (unlikely(in_atomic() && (flags & TQ_SLEEP)))
 260                 PANIC("May schedule while atomic: %s/0x%08x/%d\n",
 261                     current->comm, preempt_count(), current->pid);
 262
 263         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 264
 265         /* Taskq being destroyed and all tasks drained */
 266         if (!(tq->tq_flags & TQ_ACTIVE))
 267                 SGOTO(out, rc = 0);
 268
 269         /* Do not queue the task unless there is idle thread for it */
 270         ASSERT(tq->tq_nactive <= tq->tq_nthreads);
 271         if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads))
 272                 SGOTO(out, rc = 0);
 273
 274         if ((t = task_alloc(tq, flags)) == NULL)
 275                 SGOTO(out, rc = 0);
 276
 277         spin_lock(&t->tqent_lock);
 278
 279         /* Queue to the priority list instead of the pending list */
 280         if (flags & TQ_FRONT)
 281                 list_add_tail(&t->tqent_list, &tq->tq_prio_list);
 282         else
 283                 list_add_tail(&t->tqent_list, &tq->tq_pend_list);
 284
 285         t->tqent_id = rc = tq->tq_next_id;
 286         tq->tq_next_id++;
 287         t->tqent_func = func;
 288         t->tqent_arg = arg;
 289
 290         ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
 291
 292         spin_unlock(&t->tqent_lock);
 293
 294         wake_up(&tq->tq_work_waitq);
 295 out:
 296         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 297         SRETURN(rc);
 298 }
 299 EXPORT_SYMBOL(__taskq_dispatch);
 300
 301 void
 302 __taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
 303    taskq_ent_t *t)
 304 {
 305         SENTRY;
 306
 307         ASSERT(tq);
 308         ASSERT(func);
 309         ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
 310
 311         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 312
 313         /* Taskq being destroyed and all tasks drained */
 314         if (!(tq->tq_flags & TQ_ACTIVE)) {
 315                 t->tqent_id = 0;
 316                 goto out;
 317         }
 318
 319         spin_lock(&t->tqent_lock);
 320
 321         /*
 322          * Mark it as a prealloc'd task.  This is important
 323          * to ensure that we don't free it later.
 324          */
 325         t->tqent_flags |= TQENT_FLAG_PREALLOC;
 326
 327         /* Queue to the priority list instead of the pending list */
 328         if (flags & TQ_FRONT)
 329                 list_add_tail(&t->tqent_list, &tq->tq_prio_list);
 330         else
 331                 list_add_tail(&t->tqent_list, &tq->tq_pend_list);
 332
 333         t->tqent_id = tq->tq_next_id;
 334         tq->tq_next_id++;
 335         t->tqent_func = func;
 336         t->tqent_arg = arg;
 337
 338         spin_unlock(&t->tqent_lock);
 339
 340         wake_up(&tq->tq_work_waitq);
 341 out:
 342         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 343         SEXIT;
 344 }
 345 EXPORT_SYMBOL(__taskq_dispatch_ent);
 346
 347 int
 348 __taskq_empty_ent(taskq_ent_t *t)
 349 {
 350         return list_empty(&t->tqent_list);
 351 }
 352 EXPORT_SYMBOL(__taskq_empty_ent);
 353
 354 void
 355 __taskq_init_ent(taskq_ent_t *t)
 356 {
 357         spin_lock_init(&t->tqent_lock);
 358         INIT_LIST_HEAD(&t->tqent_list);
 359         t->tqent_id = 0;
 360         t->tqent_func = NULL;
 361         t->tqent_arg = NULL;
 362         t->tqent_flags = 0;
 363 }
 364 EXPORT_SYMBOL(__taskq_init_ent);
 365
 366 /*
 367  * Returns the lowest incomplete taskqid_t.  The taskqid_t may
 368  * be queued on the pending list, on the priority list,  or on
 369  * the work list currently being handled, but it is not 100%
 370  * complete yet.
 371  */
 372 static taskqid_t
 373 taskq_lowest_id(taskq_t *tq)
 374 {
 375         taskqid_t lowest_id = tq->tq_next_id;
 376         taskq_ent_t *t;
 377         taskq_thread_t *tqt;
 378         SENTRY;
 379
 380         ASSERT(tq);
 381         ASSERT(spin_is_locked(&tq->tq_lock));
 382
 383         if (!list_empty(&tq->tq_pend_list)) {
 384                 t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
 385                 lowest_id = MIN(lowest_id, t->tqent_id);
 386         }
 387
 388         if (!list_empty(&tq->tq_prio_list)) {
 389                 t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
 390                 lowest_id = MIN(lowest_id, t->tqent_id);
 391         }
 392
 393         if (!list_empty(&tq->tq_active_list)) {
 394                 tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
 395                                  tqt_active_list);
 396                 ASSERT(tqt->tqt_id != 0);
 397                 lowest_id = MIN(lowest_id, tqt->tqt_id);
 398         }
 399
 400         SRETURN(lowest_id);
 401 }
 402
 403 /*
 404  * Insert a task into a list keeping the list sorted by increasing
 405  * taskqid.
 406  */
 407 static void
 408 taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
 409 {
 410         taskq_thread_t *w;
 411         struct list_head *l;
 412
 413         SENTRY;
 414         ASSERT(tq);
 415         ASSERT(tqt);
 416         ASSERT(spin_is_locked(&tq->tq_lock));
 417
 418         list_for_each_prev(l, &tq->tq_active_list) {
 419                 w = list_entry(l, taskq_thread_t, tqt_active_list);
 420                 if (w->tqt_id < tqt->tqt_id) {
 421                         list_add(&tqt->tqt_active_list, l);
 422                         break;
 423                 }
 424         }
 425         if (l == &tq->tq_active_list)
 426                 list_add(&tqt->tqt_active_list, &tq->tq_active_list);
 427
 428         SEXIT;
 429 }
 430
 431 static int
 432 taskq_thread(void *args)
 433 {
 434         DECLARE_WAITQUEUE(wait, current);
 435         sigset_t blocked;
 436         taskq_thread_t *tqt = args;
 437         taskq_t *tq;
 438         taskq_ent_t *t;
 439         struct list_head *pend_list;
 440         SENTRY;
 441
 442         ASSERT(tqt);
 443         tq = tqt->tqt_tq;
 444         current->flags |= PF_NOFREEZE;
 445
 446         /* Disable the direct memory reclaim path */
 447         if (tq->tq_flags & TASKQ_NORECLAIM)
 448                 current->flags |= PF_MEMALLOC;
 449
 450         sigfillset(&blocked);
 451         sigprocmask(SIG_BLOCK, &blocked, NULL);
 452         flush_signals(current);
 453
 454         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 455         tq->tq_nthreads++;
 456         wake_up(&tq->tq_wait_waitq);
 457         set_current_state(TASK_INTERRUPTIBLE);
 458
 459         while (!kthread_should_stop()) {
 460
 461                 add_wait_queue(&tq->tq_work_waitq, &wait);
 462                 if (list_empty(&tq->tq_pend_list) &&
 463                     list_empty(&tq->tq_prio_list)) {
 464                         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 465                         schedule();
 466                         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 467                 } else {
 468                         __set_current_state(TASK_RUNNING);
 469                 }
 470
 471                 remove_wait_queue(&tq->tq_work_waitq, &wait);
 472
 473                 if (!list_empty(&tq->tq_prio_list))
 474                         pend_list = &tq->tq_prio_list;
 475                 else if (!list_empty(&tq->tq_pend_list))
 476                         pend_list = &tq->tq_pend_list;
 477                 else
 478                         pend_list = NULL;
 479
 480                 if (pend_list) {
 481                         t = list_entry(pend_list->next, taskq_ent_t, tqent_list);
 482                         list_del_init(&t->tqent_list);
 483                         /* In order to support recursively dispatching a
 484                          * preallocated taskq_ent_t, tqent_id must be
 485                          * stored prior to executing tqent_func. */
 486                         tqt->tqt_id = t->tqent_id;
 487                         taskq_insert_in_order(tq, tqt);
 488                         tq->tq_nactive++;
 489                         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 490
 491                         /* Perform the requested task */
 492                         t->tqent_func(t->tqent_arg);
 493
 494                         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 495                         tq->tq_nactive--;
 496                         list_del_init(&tqt->tqt_active_list);
 497                         task_done(tq, t);
 498
 499                         /* When the current lowest outstanding taskqid is
 500                          * done calculate the new lowest outstanding id */
 501                         if (tq->tq_lowest_id == tqt->tqt_id) {
 502                                 tq->tq_lowest_id = taskq_lowest_id(tq);
 503                                 ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
 504                         }
 505
 506                         tqt->tqt_id = 0;
 507                         wake_up_all(&tq->tq_wait_waitq);
 508                 }
 509
 510                 set_current_state(TASK_INTERRUPTIBLE);
 511
 512         }
 513
 514         __set_current_state(TASK_RUNNING);
 515         tq->tq_nthreads--;
 516         list_del_init(&tqt->tqt_thread_list);
 517         kmem_free(tqt, sizeof(taskq_thread_t));
 518
 519         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 520
 521         SRETURN(0);
 522 }
 523
 524 taskq_t *
 525 __taskq_create(const char *name, int nthreads, pri_t pri,
 526                int minalloc, int maxalloc, uint_t flags)
 527 {
 528         taskq_t *tq;
 529         taskq_thread_t *tqt;
 530         int rc = 0, i, j = 0;
 531         SENTRY;
 532
 533         ASSERT(name != NULL);
 534         ASSERT(pri <= maxclsyspri);
 535         ASSERT(minalloc >= 0);
 536         ASSERT(maxalloc <= INT_MAX);
 537         ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */
 538
 539         /* Scale the number of threads using nthreads as a percentage */
 540         if (flags & TASKQ_THREADS_CPU_PCT) {
 541                 ASSERT(nthreads <= 100);
 542                 ASSERT(nthreads >= 0);
 543                 nthreads = MIN(nthreads, 100);
 544                 nthreads = MAX(nthreads, 0);
 545                 nthreads = MAX((num_online_cpus() * nthreads) / 100, 1);
 546         }
 547
 548         tq = kmem_alloc(sizeof(*tq), KM_SLEEP);
 549         if (tq == NULL)
 550                 SRETURN(NULL);
 551
 552         spin_lock_init(&tq->tq_lock);
 553         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 554         INIT_LIST_HEAD(&tq->tq_thread_list);
 555         INIT_LIST_HEAD(&tq->tq_active_list);
 556         tq->tq_name      = name;
 557         tq->tq_nactive   = 0;
 558         tq->tq_nthreads  = 0;
 559         tq->tq_pri       = pri;
 560         tq->tq_minalloc  = minalloc;
 561         tq->tq_maxalloc  = maxalloc;
 562         tq->tq_nalloc    = 0;
 563         tq->tq_flags     = (flags | TQ_ACTIVE);
 564         tq->tq_next_id   = 1;
 565         tq->tq_lowest_id = 1;
 566         INIT_LIST_HEAD(&tq->tq_free_list);
 567         INIT_LIST_HEAD(&tq->tq_pend_list);
 568         INIT_LIST_HEAD(&tq->tq_prio_list);
 569         init_waitqueue_head(&tq->tq_work_waitq);
 570         init_waitqueue_head(&tq->tq_wait_waitq);
 571
 572         if (flags & TASKQ_PREPOPULATE)
 573                 for (i = 0; i < minalloc; i++)
 574                         task_done(tq, task_alloc(tq, TQ_SLEEP | TQ_NEW));
 575
 576         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 577
 578         for (i = 0; i < nthreads; i++) {
 579                 tqt = kmem_alloc(sizeof(*tqt), KM_SLEEP);
 580                 INIT_LIST_HEAD(&tqt->tqt_thread_list);
 581                 INIT_LIST_HEAD(&tqt->tqt_active_list);
 582                 tqt->tqt_tq = tq;
 583                 tqt->tqt_id = 0;
 584
 585                 tqt->tqt_thread = kthread_create(taskq_thread, tqt,
 586                                                  "%s/%d", name, i);
 587                 if (tqt->tqt_thread) {
 588                         list_add(&tqt->tqt_thread_list, &tq->tq_thread_list);
 589                         kthread_bind(tqt->tqt_thread, i % num_online_cpus());
 590                         set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(pri));
 591                         wake_up_process(tqt->tqt_thread);
 592                         j++;
 593                 } else {
 594                         kmem_free(tqt, sizeof(taskq_thread_t));
 595                         rc = 1;
 596                 }
 597         }
 598
 599         /* Wait for all threads to be started before potential destroy */
 600         wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j);
 601
 602         if (rc) {
 603                 __taskq_destroy(tq);
 604                 tq = NULL;
 605         }
 606
 607         SRETURN(tq);
 608 }
 609 EXPORT_SYMBOL(__taskq_create);
 610
 611 void
 612 __taskq_destroy(taskq_t *tq)
 613 {
 614         struct task_struct *thread;
 615         taskq_thread_t *tqt;
 616         taskq_ent_t *t;
 617         SENTRY;
 618
 619         ASSERT(tq);
 620         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 621         tq->tq_flags &= ~TQ_ACTIVE;
 622         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 623
 624         /* TQ_ACTIVE cleared prevents new tasks being added to pending */
 625         __taskq_wait(tq);
 626
 627         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 628
 629         /*
 630          * Signal each thread to exit and block until it does.  Each thread
 631          * is responsible for removing itself from the list and freeing its
 632          * taskq_thread_t.  This allows for idle threads to opt to remove
 633          * themselves from the taskq.  They can be recreated as needed.
 634          */
 635         while (!list_empty(&tq->tq_thread_list)) {
 636                 tqt = list_entry(tq->tq_thread_list.next,
 637                                  taskq_thread_t, tqt_thread_list);
 638                 thread = tqt->tqt_thread;
 639                 spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 640
 641                 kthread_stop(thread);
 642
 643                 spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 644         }
 645
 646         while (!list_empty(&tq->tq_free_list)) {
 647                 t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
 648
 649                 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
 650
 651                 list_del_init(&t->tqent_list);
 652                 task_free(tq, t);
 653         }
 654
 655         ASSERT(tq->tq_nthreads == 0);
 656         ASSERT(tq->tq_nalloc == 0);
 657         ASSERT(list_empty(&tq->tq_thread_list));
 658         ASSERT(list_empty(&tq->tq_active_list));
 659         ASSERT(list_empty(&tq->tq_free_list));
 660         ASSERT(list_empty(&tq->tq_pend_list));
 661         ASSERT(list_empty(&tq->tq_prio_list));
 662
 663         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 664
 665         kmem_free(tq, sizeof(taskq_t));
 666
 667         SEXIT;
 668 }
 669 EXPORT_SYMBOL(__taskq_destroy);
 670
 671 int
 672 spl_taskq_init(void)
 673 {
 674         SENTRY;
 675
 676         /* Solaris creates a dynamic taskq of up to 64 threads, however in
 677          * a Linux environment 1 thread per-core is usually about right */
 678         system_taskq = taskq_create("spl_system_taskq", num_online_cpus(),
 679                                     minclsyspri, 4, 512, TASKQ_PREPOPULATE);
 680         if (system_taskq == NULL)
 681                 SRETURN(1);
 682
 683         SRETURN(0);
 684 }
 685
 686 void
 687 spl_taskq_fini(void)
 688 {
 689         SENTRY;
 690         taskq_destroy(system_taskq);
 691         SEXIT;
 692 }