granicus.if.org Git - zfs/blob - 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 typedef struct spl_task {
  42         spinlock_t              t_lock;
  43         struct list_head        t_list;
  44         taskqid_t               t_id;
  45         task_func_t             *t_func;
  46         void                    *t_arg;
  47 } spl_task_t;
  48
  49 /*
  50  * NOTE: Must be called with tq->tq_lock held, returns a list_t which
  51  * is not attached to the free, work, or pending taskq lists.
  52  */
  53 static spl_task_t *
  54 task_alloc(taskq_t *tq, uint_t flags)
  55 {
  56         spl_task_t *t;
  57         int count = 0;
  58         SENTRY;
  59
  60         ASSERT(tq);
  61         ASSERT(flags & (TQ_SLEEP | TQ_NOSLEEP));               /* One set */
  62         ASSERT(!((flags & TQ_SLEEP) && (flags & TQ_NOSLEEP))); /* Not both */
  63         ASSERT(spin_is_locked(&tq->tq_lock));
  64 retry:
  65         /* Acquire spl_task_t's from free list if available */
  66         if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
  67                 t = list_entry(tq->tq_free_list.next, spl_task_t, t_list);
  68                 list_del_init(&t->t_list);
  69                 SRETURN(t);
  70         }
  71
  72         /* Free list is empty and memory allocations are prohibited */
  73         if (flags & TQ_NOALLOC)
  74                 SRETURN(NULL);
  75
  76         /* Hit maximum spl_task_t pool size */
  77         if (tq->tq_nalloc >= tq->tq_maxalloc) {
  78                 if (flags & TQ_NOSLEEP)
  79                         SRETURN(NULL);
  80
  81                 /* Sleep periodically polling the free list for an available
  82                  * spl_task_t.  If a full second passes and we have not found
  83                  * one gives up and return a NULL to the caller. */
  84                 if (flags & TQ_SLEEP) {
  85                         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
  86                         schedule_timeout(HZ / 100);
  87                         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
  88                         if (count < 100)
  89                                 SGOTO(retry, count++);
  90
  91                         SRETURN(NULL);
  92                 }
  93
  94                 /* Unreachable, Neither TQ_SLEEP or TQ_NOSLEEP set */
  95                 PANIC("Neither TQ_SLEEP or TQ_NOSLEEP set");
  96         }
  97
  98         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
  99         t = kmem_alloc(sizeof(spl_task_t), flags & (TQ_SLEEP | TQ_NOSLEEP));
 100         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 101
 102         if (t) {
 103                 spin_lock_init(&t->t_lock);
 104                 INIT_LIST_HEAD(&t->t_list);
 105                 t->t_id = 0;
 106                 t->t_func = NULL;
 107                 t->t_arg = NULL;
 108                 tq->tq_nalloc++;
 109         }
 110
 111         SRETURN(t);
 112 }
 113
 114 /*
 115  * NOTE: Must be called with tq->tq_lock held, expects the spl_task_t
 116  * to already be removed from the free, work, or pending taskq lists.
 117  */
 118 static void
 119 task_free(taskq_t *tq, spl_task_t *t)
 120 {
 121         SENTRY;
 122
 123         ASSERT(tq);
 124         ASSERT(t);
 125         ASSERT(spin_is_locked(&tq->tq_lock));
 126         ASSERT(list_empty(&t->t_list));
 127
 128         kmem_free(t, sizeof(spl_task_t));
 129         tq->tq_nalloc--;
 130
 131         SEXIT;
 132 }
 133
 134 /*
 135  * NOTE: Must be called with tq->tq_lock held, either destroys the
 136  * spl_task_t if too many exist or moves it to the free list for later use.
 137  */
 138 static void
 139 task_done(taskq_t *tq, spl_task_t *t)
 140 {
 141         SENTRY;
 142         ASSERT(tq);
 143         ASSERT(t);
 144         ASSERT(spin_is_locked(&tq->tq_lock));
 145
 146         list_del_init(&t->t_list);
 147
 148         if (tq->tq_nalloc <= tq->tq_minalloc) {
 149                 t->t_id = 0;
 150                 t->t_func = NULL;
 151                 t->t_arg = NULL;
 152                 list_add_tail(&t->t_list, &tq->tq_free_list);
 153         } else {
 154                 task_free(tq, t);
 155         }
 156
 157         SEXIT;
 158 }
 159
 160 /*
 161  * As tasks are submitted to the task queue they are assigned a
 162  * monotonically increasing taskqid and added to the tail of the pending
 163  * list.  As worker threads become available the tasks are removed from
 164  * the head of the pending or priority list, giving preference to the
 165  * priority list.  The tasks are then added to the work list, preserving
 166  * the ordering by taskqid.  Finally, as tasks complete they are removed
 167  * from the work list.  This means that the pending and work lists are
 168  * always kept sorted by taskqid.  Thus the lowest outstanding
 169  * incomplete taskqid can be determined simply by checking the min
 170  * taskqid for each head item on the pending, priority, and work list.
 171  * This value is stored in tq->tq_lowest_id and only updated to the new
 172  * lowest id when the previous lowest id completes.  All taskqids lower
 173  * than tq->tq_lowest_id must have completed.  It is also possible
 174  * larger taskqid's have completed because they may be processed in
 175  * parallel by several worker threads.  However, this is not a problem
 176  * because the behavior of taskq_wait_id() is to block until all
 177  * previously submitted taskqid's have completed.
 178  *
 179  * XXX: Taskqid_t wrapping is not handled.  However, taskqid_t's are
 180  * 64-bit values so even if a taskq is processing 2^24 (16,777,216)
 181  * taskqid_ts per second it will still take 2^40 seconds, 34,865 years,
 182  * before the wrap occurs.  I can live with that for now.
 183  */
 184 static int
 185 taskq_wait_check(taskq_t *tq, taskqid_t id)
 186 {
 187         int rc;
 188
 189         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 190         rc = (id < tq->tq_lowest_id);
 191         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 192
 193         SRETURN(rc);
 194 }
 195
 196 void
 197 __taskq_wait_id(taskq_t *tq, taskqid_t id)
 198 {
 199         SENTRY;
 200         ASSERT(tq);
 201
 202         wait_event(tq->tq_wait_waitq, taskq_wait_check(tq, id));
 203
 204         SEXIT;
 205 }
 206 EXPORT_SYMBOL(__taskq_wait_id);
 207
 208 void
 209 __taskq_wait(taskq_t *tq)
 210 {
 211         taskqid_t id;
 212         SENTRY;
 213         ASSERT(tq);
 214
 215         /* Wait for the largest outstanding taskqid */
 216         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 217         id = tq->tq_next_id - 1;
 218         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 219
 220         __taskq_wait_id(tq, id);
 221
 222         SEXIT;
 223
 224 }
 225 EXPORT_SYMBOL(__taskq_wait);
 226
 227 int
 228 __taskq_member(taskq_t *tq, void *t)
 229 {
 230         int i;
 231         SENTRY;
 232
 233         ASSERT(tq);
 234         ASSERT(t);
 235
 236         for (i = 0; i < tq->tq_nthreads; i++)
 237                 if (tq->tq_threads[i] == (struct task_struct *)t)
 238                         SRETURN(1);
 239
 240         SRETURN(0);
 241 }
 242 EXPORT_SYMBOL(__taskq_member);
 243
 244 taskqid_t
 245 __taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
 246 {
 247         spl_task_t *t;
 248         taskqid_t rc = 0;
 249         SENTRY;
 250
 251         ASSERT(tq);
 252         ASSERT(func);
 253
 254         /* Solaris assumes TQ_SLEEP if not passed explicitly */
 255         if (!(flags & (TQ_SLEEP | TQ_NOSLEEP)))
 256                 flags |= TQ_SLEEP;
 257
 258         if (unlikely(in_atomic() && (flags & TQ_SLEEP)))
 259                 PANIC("May schedule while atomic: %s/0x%08x/%d\n",
 260                     current->comm, preempt_count(), current->pid);
 261
 262         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 263
 264         /* Taskq being destroyed and all tasks drained */
 265         if (!(tq->tq_flags & TQ_ACTIVE))
 266                 SGOTO(out, rc = 0);
 267
 268         /* Do not queue the task unless there is idle thread for it */
 269         ASSERT(tq->tq_nactive <= tq->tq_nthreads);
 270         if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads))
 271                 SGOTO(out, rc = 0);
 272
 273         if ((t = task_alloc(tq, flags)) == NULL)
 274                 SGOTO(out, rc = 0);
 275
 276         spin_lock(&t->t_lock);
 277
 278         /* Queue to the priority list instead of the pending list */
 279         if (flags & TQ_FRONT)
 280                 list_add_tail(&t->t_list, &tq->tq_prio_list);
 281         else
 282                 list_add_tail(&t->t_list, &tq->tq_pend_list);
 283
 284         t->t_id = rc = tq->tq_next_id;
 285         tq->tq_next_id++;
 286         t->t_func = func;
 287         t->t_arg = arg;
 288         spin_unlock(&t->t_lock);
 289
 290         wake_up(&tq->tq_work_waitq);
 291 out:
 292         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 293         SRETURN(rc);
 294 }
 295 EXPORT_SYMBOL(__taskq_dispatch);
 296
 297 /*
 298  * Returns the lowest incomplete taskqid_t.  The taskqid_t may
 299  * be queued on the pending list, on the priority list,  or on
 300  * the work list currently being handled, but it is not 100%
 301  * complete yet.
 302  */
 303 static taskqid_t
 304 taskq_lowest_id(taskq_t *tq)
 305 {
 306         taskqid_t lowest_id = tq->tq_next_id;
 307         spl_task_t *t;
 308         SENTRY;
 309
 310         ASSERT(tq);
 311         ASSERT(spin_is_locked(&tq->tq_lock));
 312
 313         if (!list_empty(&tq->tq_pend_list)) {
 314                 t = list_entry(tq->tq_pend_list.next, spl_task_t, t_list);
 315                 lowest_id = MIN(lowest_id, t->t_id);
 316         }
 317
 318         if (!list_empty(&tq->tq_prio_list)) {
 319                 t = list_entry(tq->tq_prio_list.next, spl_task_t, t_list);
 320                 lowest_id = MIN(lowest_id, t->t_id);
 321         }
 322
 323         if (!list_empty(&tq->tq_work_list)) {
 324                 t = list_entry(tq->tq_work_list.next, spl_task_t, t_list);
 325                 lowest_id = MIN(lowest_id, t->t_id);
 326         }
 327
 328         SRETURN(lowest_id);
 329 }
 330
 331 /*
 332  * Insert a task into a list keeping the list sorted by increasing
 333  * taskqid.
 334  */
 335 static void
 336 taskq_insert_in_order(taskq_t *tq, spl_task_t *t)
 337 {
 338         spl_task_t *w;
 339         struct list_head *l;
 340
 341         SENTRY;
 342         ASSERT(tq);
 343         ASSERT(t);
 344         ASSERT(spin_is_locked(&tq->tq_lock));
 345
 346         list_for_each_prev(l, &tq->tq_work_list) {
 347                 w = list_entry(l, spl_task_t, t_list);
 348                 if (w->t_id < t->t_id) {
 349                         list_add(&t->t_list, l);
 350                         break;
 351                 }
 352         }
 353         if (l == &tq->tq_work_list)
 354                 list_add(&t->t_list, &tq->tq_work_list);
 355
 356         SEXIT;
 357 }
 358
 359 static int
 360 taskq_thread(void *args)
 361 {
 362         DECLARE_WAITQUEUE(wait, current);
 363         sigset_t blocked;
 364         taskqid_t id;
 365         taskq_t *tq = args;
 366         spl_task_t *t;
 367         struct list_head *pend_list;
 368         SENTRY;
 369
 370         ASSERT(tq);
 371         current->flags |= PF_NOFREEZE;
 372
 373         sigfillset(&blocked);
 374         sigprocmask(SIG_BLOCK, &blocked, NULL);
 375         flush_signals(current);
 376
 377         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 378         tq->tq_nthreads++;
 379         wake_up(&tq->tq_wait_waitq);
 380         set_current_state(TASK_INTERRUPTIBLE);
 381
 382         while (!kthread_should_stop()) {
 383
 384                 add_wait_queue(&tq->tq_work_waitq, &wait);
 385                 if (list_empty(&tq->tq_pend_list) &&
 386                     list_empty(&tq->tq_prio_list)) {
 387                         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 388                         schedule();
 389                         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 390                 } else {
 391                         __set_current_state(TASK_RUNNING);
 392                 }
 393
 394                 remove_wait_queue(&tq->tq_work_waitq, &wait);
 395
 396                 if (!list_empty(&tq->tq_prio_list))
 397                         pend_list = &tq->tq_prio_list;
 398                 else if (!list_empty(&tq->tq_pend_list))
 399                         pend_list = &tq->tq_pend_list;
 400                 else
 401                         pend_list = NULL;
 402
 403                 if (pend_list) {
 404                         t = list_entry(pend_list->next, spl_task_t, t_list);
 405                         list_del_init(&t->t_list);
 406                         taskq_insert_in_order(tq, t);
 407                         tq->tq_nactive++;
 408                         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 409
 410                         /* Perform the requested task */
 411                         t->t_func(t->t_arg);
 412
 413                         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 414                         tq->tq_nactive--;
 415                         id = t->t_id;
 416                         task_done(tq, t);
 417
 418                         /* When the current lowest outstanding taskqid is
 419                          * done calculate the new lowest outstanding id */
 420                         if (tq->tq_lowest_id == id) {
 421                                 tq->tq_lowest_id = taskq_lowest_id(tq);
 422                                 ASSERT(tq->tq_lowest_id > id);
 423                         }
 424
 425                         wake_up_all(&tq->tq_wait_waitq);
 426                 }
 427
 428                 set_current_state(TASK_INTERRUPTIBLE);
 429
 430         }
 431
 432         __set_current_state(TASK_RUNNING);
 433         tq->tq_nthreads--;
 434         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 435
 436         SRETURN(0);
 437 }
 438
 439 taskq_t *
 440 __taskq_create(const char *name, int nthreads, pri_t pri,
 441                int minalloc, int maxalloc, uint_t flags)
 442 {
 443         taskq_t *tq;
 444         struct task_struct *t;
 445         int rc = 0, i, j = 0;
 446         SENTRY;
 447
 448         ASSERT(name != NULL);
 449         ASSERT(pri <= maxclsyspri);
 450         ASSERT(minalloc >= 0);
 451         ASSERT(maxalloc <= INT_MAX);
 452         ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */
 453
 454         /* Scale the number of threads using nthreads as a percentage */
 455         if (flags & TASKQ_THREADS_CPU_PCT) {
 456                 ASSERT(nthreads <= 100);
 457                 ASSERT(nthreads >= 0);
 458                 nthreads = MIN(nthreads, 100);
 459                 nthreads = MAX(nthreads, 0);
 460                 nthreads = MAX((num_online_cpus() * nthreads) / 100, 1);
 461         }
 462
 463         tq = kmem_alloc(sizeof(*tq), KM_SLEEP);
 464         if (tq == NULL)
 465                 SRETURN(NULL);
 466
 467         tq->tq_threads = kmem_alloc(nthreads * sizeof(t), KM_SLEEP);
 468         if (tq->tq_threads == NULL) {
 469                 kmem_free(tq, sizeof(*tq));
 470                 SRETURN(NULL);
 471         }
 472
 473         spin_lock_init(&tq->tq_lock);
 474         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 475         tq->tq_name      = name;
 476         tq->tq_nactive   = 0;
 477         tq->tq_nthreads  = 0;
 478         tq->tq_pri       = pri;
 479         tq->tq_minalloc  = minalloc;
 480         tq->tq_maxalloc  = maxalloc;
 481         tq->tq_nalloc    = 0;
 482         tq->tq_flags     = (flags | TQ_ACTIVE);
 483         tq->tq_next_id   = 1;
 484         tq->tq_lowest_id = 1;
 485         INIT_LIST_HEAD(&tq->tq_free_list);
 486         INIT_LIST_HEAD(&tq->tq_work_list);
 487         INIT_LIST_HEAD(&tq->tq_pend_list);
 488         INIT_LIST_HEAD(&tq->tq_prio_list);
 489         init_waitqueue_head(&tq->tq_work_waitq);
 490         init_waitqueue_head(&tq->tq_wait_waitq);
 491
 492         if (flags & TASKQ_PREPOPULATE)
 493                 for (i = 0; i < minalloc; i++)
 494                         task_done(tq, task_alloc(tq, TQ_SLEEP | TQ_NEW));
 495
 496         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 497
 498         for (i = 0; i < nthreads; i++) {
 499                 t = kthread_create(taskq_thread, tq, "%s/%d", name, i);
 500                 if (t) {
 501                         tq->tq_threads[i] = t;
 502                         kthread_bind(t, i % num_online_cpus());
 503                         set_user_nice(t, PRIO_TO_NICE(pri));
 504                         wake_up_process(t);
 505                         j++;
 506                 } else {
 507                         tq->tq_threads[i] = NULL;
 508                         rc = 1;
 509                 }
 510         }
 511
 512         /* Wait for all threads to be started before potential destroy */
 513         wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j);
 514
 515         if (rc) {
 516                 __taskq_destroy(tq);
 517                 tq = NULL;
 518         }
 519
 520         SRETURN(tq);
 521 }
 522 EXPORT_SYMBOL(__taskq_create);
 523
 524 void
 525 __taskq_destroy(taskq_t *tq)
 526 {
 527         spl_task_t *t;
 528         int i, nthreads;
 529         SENTRY;
 530
 531         ASSERT(tq);
 532         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 533         tq->tq_flags &= ~TQ_ACTIVE;
 534         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 535
 536         /* TQ_ACTIVE cleared prevents new tasks being added to pending */
 537         __taskq_wait(tq);
 538
 539         nthreads = tq->tq_nthreads;
 540         for (i = 0; i < nthreads; i++)
 541                 if (tq->tq_threads[i])
 542                         kthread_stop(tq->tq_threads[i]);
 543
 544         spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags);
 545
 546         while (!list_empty(&tq->tq_free_list)) {
 547                 t = list_entry(tq->tq_free_list.next, spl_task_t, t_list);
 548                 list_del_init(&t->t_list);
 549                 task_free(tq, t);
 550         }
 551
 552         ASSERT(tq->tq_nthreads == 0);
 553         ASSERT(tq->tq_nalloc == 0);
 554         ASSERT(list_empty(&tq->tq_free_list));
 555         ASSERT(list_empty(&tq->tq_work_list));
 556         ASSERT(list_empty(&tq->tq_pend_list));
 557         ASSERT(list_empty(&tq->tq_prio_list));
 558
 559         spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags);
 560         kmem_free(tq->tq_threads, nthreads * sizeof(spl_task_t *));
 561         kmem_free(tq, sizeof(taskq_t));
 562
 563         SEXIT;
 564 }
 565 EXPORT_SYMBOL(__taskq_destroy);
 566
 567 int
 568 spl_taskq_init(void)
 569 {
 570         SENTRY;
 571
 572         /* Solaris creates a dynamic taskq of up to 64 threads, however in
 573          * a Linux environment 1 thread per-core is usually about right */
 574         system_taskq = taskq_create("spl_system_taskq", num_online_cpus(),
 575                                     minclsyspri, 4, 512, TASKQ_PREPOPULATE);
 576         if (system_taskq == NULL)
 577                 SRETURN(1);
 578
 579         SRETURN(0);
 580 }
 581
 582 void
 583 spl_taskq_fini(void)
 584 {
 585         SENTRY;
 586         taskq_destroy(system_taskq);
 587         SEXIT;
 588 }