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 *
10 * The SPL is free software; you can redistribute it and/or modify it
11 * under the terms of the GNU General Public License as published by the
12 * Free Software Foundation; either version 2 of the License, or (at your
13 * option) any later version.
14 *
15 * The SPL is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 * for more details.
19 *
20 * You should have received a copy of the GNU General Public License along
21 * with the SPL. If not, see <http://www.gnu.org/licenses/>.
22 *
23 * Solaris Porting Layer (SPL) Task Queue Implementation.
24 */
25
26 #include <sys/timer.h>
27 #include <sys/taskq.h>
28 #include <sys/kmem.h>
29 #include <sys/tsd.h>
30 #include <sys/trace_spl.h>
31 #ifdef HAVE_CPU_HOTPLUG
32 #include <linux/cpuhotplug.h>
33 #endif
34
35 static int spl_taskq_thread_bind = 0;
36 module_param(spl_taskq_thread_bind, int, 0644);
37 MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default");
38
39
40 static int spl_taskq_thread_dynamic = 1;
41 module_param(spl_taskq_thread_dynamic, int, 0444);
42 MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads");
43
44 static int spl_taskq_thread_priority = 1;
45 module_param(spl_taskq_thread_priority, int, 0644);
46 MODULE_PARM_DESC(spl_taskq_thread_priority,
47 "Allow non-default priority for taskq threads");
48
49 static uint_t spl_taskq_thread_sequential = 4;
50 /* BEGIN CSTYLED */
51 module_param(spl_taskq_thread_sequential, uint, 0644);
52 /* END CSTYLED */
53 MODULE_PARM_DESC(spl_taskq_thread_sequential,
54 "Create new taskq threads after N sequential tasks");
55
56 /*
57 * Global system-wide dynamic task queue available for all consumers. This
58 * taskq is not intended for long-running tasks; instead, a dedicated taskq
59 * should be created.
60 */
61 taskq_t *system_taskq;
62 EXPORT_SYMBOL(system_taskq);
63 /* Global dynamic task queue for long delay */
64 taskq_t *system_delay_taskq;
65 EXPORT_SYMBOL(system_delay_taskq);
66
67 /* Private dedicated taskq for creating new taskq threads on demand. */
68 static taskq_t *dynamic_taskq;
69 static taskq_thread_t *taskq_thread_create(taskq_t *);
70
71 #ifdef HAVE_CPU_HOTPLUG
72 /* Multi-callback id for cpu hotplugging. */
73 static int spl_taskq_cpuhp_state;
74 #endif
75
76 /* List of all taskqs */
77 LIST_HEAD(tq_list);
78 struct rw_semaphore tq_list_sem;
79 static uint_t taskq_tsd;
80
81 static int
82 task_km_flags(uint_t flags)
83 {
84 if (flags & TQ_NOSLEEP)
85 return (KM_NOSLEEP);
86
87 if (flags & TQ_PUSHPAGE)
88 return (KM_PUSHPAGE);
89
90 return (KM_SLEEP);
91 }
92
93 /*
94 * taskq_find_by_name - Find the largest instance number of a named taskq.
95 */
96 static int
97 taskq_find_by_name(const char *name)
98 {
99 struct list_head *tql = NULL;
100 taskq_t *tq;
101
102 list_for_each_prev(tql, &tq_list) {
103 tq = list_entry(tql, taskq_t, tq_taskqs);
104 if (strcmp(name, tq->tq_name) == 0)
105 return (tq->tq_instance);
106 }
107 return (-1);
108 }
109
110 /*
111 * NOTE: Must be called with tq->tq_lock held, returns a list_t which
112 * is not attached to the free, work, or pending taskq lists.
113 */
114 static taskq_ent_t *
115 task_alloc(taskq_t *tq, uint_t flags, unsigned long *irqflags)
116 {
117 taskq_ent_t *t;
118 int count = 0;
119
120 ASSERT(tq);
121 retry:
122 /* Acquire taskq_ent_t's from free list if available */
123 if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
124 t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
125
126 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
127 ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL));
128 ASSERT(!timer_pending(&t->tqent_timer));
129
130 list_del_init(&t->tqent_list);
131 return (t);
132 }
133
134 /* Free list is empty and memory allocations are prohibited */
135 if (flags & TQ_NOALLOC)
136 return (NULL);
137
138 /* Hit maximum taskq_ent_t pool size */
139 if (tq->tq_nalloc >= tq->tq_maxalloc) {
140 if (flags & TQ_NOSLEEP)
141 return (NULL);
142
143 /*
144 * Sleep periodically polling the free list for an available
145 * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
146 * but we cannot block forever waiting for an taskq_ent_t to
147 * show up in the free list, otherwise a deadlock can happen.
148 *
149 * Therefore, we need to allocate a new task even if the number
150 * of allocated tasks is above tq->tq_maxalloc, but we still
151 * end up delaying the task allocation by one second, thereby
152 * throttling the task dispatch rate.
153 */
154 spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
155 schedule_timeout(HZ / 100);
156 spin_lock_irqsave_nested(&tq->tq_lock, *irqflags,
157 tq->tq_lock_class);
158 if (count < 100) {
159 count++;
160 goto retry;
161 }
162 }
163
164 spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
165 t = kmem_alloc(sizeof (taskq_ent_t), task_km_flags(flags));
166 spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class);
167
168 if (t) {
169 taskq_init_ent(t);
170 tq->tq_nalloc++;
171 }
172
173 return (t);
174 }
175
176 /*
177 * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t
178 * to already be removed from the free, work, or pending taskq lists.
179 */
180 static void
181 task_free(taskq_t *tq, taskq_ent_t *t)
182 {
183 ASSERT(tq);
184 ASSERT(t);
185 ASSERT(list_empty(&t->tqent_list));
186 ASSERT(!timer_pending(&t->tqent_timer));
187
188 kmem_free(t, sizeof (taskq_ent_t));
189 tq->tq_nalloc--;
190 }
191
192 /*
193 * NOTE: Must be called with tq->tq_lock held, either destroys the
194 * taskq_ent_t if too many exist or moves it to the free list for later use.
195 */
196 static void
197 task_done(taskq_t *tq, taskq_ent_t *t)
198 {
199 ASSERT(tq);
200 ASSERT(t);
201
202 /* Wake tasks blocked in taskq_wait_id() */
203 wake_up_all(&t->tqent_waitq);
204
205 list_del_init(&t->tqent_list);
206
207 if (tq->tq_nalloc <= tq->tq_minalloc) {
208 t->tqent_id = TASKQID_INVALID;
209 t->tqent_func = NULL;
210 t->tqent_arg = NULL;
211 t->tqent_flags = 0;
212
213 list_add_tail(&t->tqent_list, &tq->tq_free_list);
214 } else {
215 task_free(tq, t);
216 }
217 }
218
219 /*
220 * When a delayed task timer expires remove it from the delay list and
221 * add it to the priority list in order for immediate processing.
222 */
223 static void
224 task_expire_impl(taskq_ent_t *t)
225 {
226 taskq_ent_t *w;
227 taskq_t *tq = t->tqent_taskq;
228 struct list_head *l = NULL;
229 unsigned long flags;
230
231 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
232
233 if (t->tqent_flags & TQENT_FLAG_CANCEL) {
234 ASSERT(list_empty(&t->tqent_list));
235 spin_unlock_irqrestore(&tq->tq_lock, flags);
236 return;
237 }
238
239 t->tqent_birth = jiffies;
240 DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
241
242 /*
243 * The priority list must be maintained in strict task id order
244 * from lowest to highest for lowest_id to be easily calculable.
245 */
246 list_del(&t->tqent_list);
247 list_for_each_prev(l, &tq->tq_prio_list) {
248 w = list_entry(l, taskq_ent_t, tqent_list);
249 if (w->tqent_id < t->tqent_id) {
250 list_add(&t->tqent_list, l);
251 break;
252 }
253 }
254 if (l == &tq->tq_prio_list)
255 list_add(&t->tqent_list, &tq->tq_prio_list);
256
257 spin_unlock_irqrestore(&tq->tq_lock, flags);
258
259 wake_up(&tq->tq_work_waitq);
260 }
261
262 static void
263 task_expire(spl_timer_list_t tl)
264 {
265 struct timer_list *tmr = (struct timer_list *)tl;
266 taskq_ent_t *t = from_timer(t, tmr, tqent_timer);
267 task_expire_impl(t);
268 }
269
270 /*
271 * Returns the lowest incomplete taskqid_t. The taskqid_t may
272 * be queued on the pending list, on the priority list, on the
273 * delay list, or on the work list currently being handled, but
274 * it is not 100% complete yet.
275 */
276 static taskqid_t
277 taskq_lowest_id(taskq_t *tq)
278 {
279 taskqid_t lowest_id = tq->tq_next_id;
280 taskq_ent_t *t;
281 taskq_thread_t *tqt;
282
283 if (!list_empty(&tq->tq_pend_list)) {
284 t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
285 lowest_id = MIN(lowest_id, t->tqent_id);
286 }
287
288 if (!list_empty(&tq->tq_prio_list)) {
289 t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
290 lowest_id = MIN(lowest_id, t->tqent_id);
291 }
292
293 if (!list_empty(&tq->tq_delay_list)) {
294 t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list);
295 lowest_id = MIN(lowest_id, t->tqent_id);
296 }
297
298 if (!list_empty(&tq->tq_active_list)) {
299 tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
300 tqt_active_list);
301 ASSERT(tqt->tqt_id != TASKQID_INVALID);
302 lowest_id = MIN(lowest_id, tqt->tqt_id);
303 }
304
305 return (lowest_id);
306 }
307
308 /*
309 * Insert a task into a list keeping the list sorted by increasing taskqid.
310 */
311 static void
312 taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
313 {
314 taskq_thread_t *w;
315 struct list_head *l = NULL;
316
317 ASSERT(tq);
318 ASSERT(tqt);
319
320 list_for_each_prev(l, &tq->tq_active_list) {
321 w = list_entry(l, taskq_thread_t, tqt_active_list);
322 if (w->tqt_id < tqt->tqt_id) {
323 list_add(&tqt->tqt_active_list, l);
324 break;
325 }
326 }
327 if (l == &tq->tq_active_list)
328 list_add(&tqt->tqt_active_list, &tq->tq_active_list);
329 }
330
331 /*
332 * Find and return a task from the given list if it exists. The list
333 * must be in lowest to highest task id order.
334 */
335 static taskq_ent_t *
336 taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id)
337 {
338 struct list_head *l = NULL;
339 taskq_ent_t *t;
340
341 list_for_each(l, lh) {
342 t = list_entry(l, taskq_ent_t, tqent_list);
343
344 if (t->tqent_id == id)
345 return (t);
346
347 if (t->tqent_id > id)
348 break;
349 }
350
351 return (NULL);
352 }
353
354 /*
355 * Find an already dispatched task given the task id regardless of what
356 * state it is in. If a task is still pending it will be returned.
357 * If a task is executing, then -EBUSY will be returned instead.
358 * If the task has already been run then NULL is returned.
359 */
360 static taskq_ent_t *
361 taskq_find(taskq_t *tq, taskqid_t id)
362 {
363 taskq_thread_t *tqt;
364 struct list_head *l = NULL;
365 taskq_ent_t *t;
366
367 t = taskq_find_list(tq, &tq->tq_delay_list, id);
368 if (t)
369 return (t);
370
371 t = taskq_find_list(tq, &tq->tq_prio_list, id);
372 if (t)
373 return (t);
374
375 t = taskq_find_list(tq, &tq->tq_pend_list, id);
376 if (t)
377 return (t);
378
379 list_for_each(l, &tq->tq_active_list) {
380 tqt = list_entry(l, taskq_thread_t, tqt_active_list);
381 if (tqt->tqt_id == id) {
382 /*
383 * Instead of returning tqt_task, we just return a non
384 * NULL value to prevent misuse, since tqt_task only
385 * has two valid fields.
386 */
387 return (ERR_PTR(-EBUSY));
388 }
389 }
390
391 return (NULL);
392 }
393
394 /*
395 * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and
396 * taskq_wait() functions below.
397 *
398 * Taskq waiting is accomplished by tracking the lowest outstanding task
399 * id and the next available task id. As tasks are dispatched they are
400 * added to the tail of the pending, priority, or delay lists. As worker
401 * threads become available the tasks are removed from the heads of these
402 * lists and linked to the worker threads. This ensures the lists are
403 * kept sorted by lowest to highest task id.
404 *
405 * Therefore the lowest outstanding task id can be quickly determined by
406 * checking the head item from all of these lists. This value is stored
407 * with the taskq as the lowest id. It only needs to be recalculated when
408 * either the task with the current lowest id completes or is canceled.
409 *
410 * By blocking until the lowest task id exceeds the passed task id the
411 * taskq_wait_outstanding() function can be easily implemented. Similarly,
412 * by blocking until the lowest task id matches the next task id taskq_wait()
413 * can be implemented.
414 *
415 * Callers should be aware that when there are multiple worked threads it
416 * is possible for larger task ids to complete before smaller ones. Also
417 * when the taskq contains delay tasks with small task ids callers may
418 * block for a considerable length of time waiting for them to expire and
419 * execute.
420 */
421 static int
422 taskq_wait_id_check(taskq_t *tq, taskqid_t id)
423 {
424 int rc;
425 unsigned long flags;
426
427 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
428 rc = (taskq_find(tq, id) == NULL);
429 spin_unlock_irqrestore(&tq->tq_lock, flags);
430
431 return (rc);
432 }
433
434 /*
435 * The taskq_wait_id() function blocks until the passed task id completes.
436 * This does not guarantee that all lower task ids have completed.
437 */
438 void
439 taskq_wait_id(taskq_t *tq, taskqid_t id)
440 {
441 wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id));
442 }
443 EXPORT_SYMBOL(taskq_wait_id);
444
445 static int
446 taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id)
447 {
448 int rc;
449 unsigned long flags;
450
451 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
452 rc = (id < tq->tq_lowest_id);
453 spin_unlock_irqrestore(&tq->tq_lock, flags);
454
455 return (rc);
456 }
457
458 /*
459 * The taskq_wait_outstanding() function will block until all tasks with a
460 * lower taskqid than the passed 'id' have been completed. Note that all
461 * task id's are assigned monotonically at dispatch time. Zero may be
462 * passed for the id to indicate all tasks dispatch up to this point,
463 * but not after, should be waited for.
464 */
465 void
466 taskq_wait_outstanding(taskq_t *tq, taskqid_t id)
467 {
468 id = id ? id : tq->tq_next_id - 1;
469 wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id));
470 }
471 EXPORT_SYMBOL(taskq_wait_outstanding);
472
473 static int
474 taskq_wait_check(taskq_t *tq)
475 {
476 int rc;
477 unsigned long flags;
478
479 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
480 rc = (tq->tq_lowest_id == tq->tq_next_id);
481 spin_unlock_irqrestore(&tq->tq_lock, flags);
482
483 return (rc);
484 }
485
486 /*
487 * The taskq_wait() function will block until the taskq is empty.
488 * This means that if a taskq re-dispatches work to itself taskq_wait()
489 * callers will block indefinitely.
490 */
491 void
492 taskq_wait(taskq_t *tq)
493 {
494 wait_event(tq->tq_wait_waitq, taskq_wait_check(tq));
495 }
496 EXPORT_SYMBOL(taskq_wait);
497
498 int
499 taskq_member(taskq_t *tq, kthread_t *t)
500 {
501 return (tq == (taskq_t *)tsd_get_by_thread(taskq_tsd, t));
502 }
503 EXPORT_SYMBOL(taskq_member);
504
505 taskq_t *
506 taskq_of_curthread(void)
507 {
508 return (tsd_get(taskq_tsd));
509 }
510 EXPORT_SYMBOL(taskq_of_curthread);
511
512 /*
513 * Cancel an already dispatched task given the task id. Still pending tasks
514 * will be immediately canceled, and if the task is active the function will
515 * block until it completes. Preallocated tasks which are canceled must be
516 * freed by the caller.
517 */
518 int
519 taskq_cancel_id(taskq_t *tq, taskqid_t id)
520 {
521 taskq_ent_t *t;
522 int rc = ENOENT;
523 unsigned long flags;
524
525 ASSERT(tq);
526
527 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
528 t = taskq_find(tq, id);
529 if (t && t != ERR_PTR(-EBUSY)) {
530 list_del_init(&t->tqent_list);
531 t->tqent_flags |= TQENT_FLAG_CANCEL;
532
533 /*
534 * When canceling the lowest outstanding task id we
535 * must recalculate the new lowest outstanding id.
536 */
537 if (tq->tq_lowest_id == t->tqent_id) {
538 tq->tq_lowest_id = taskq_lowest_id(tq);
539 ASSERT3S(tq->tq_lowest_id, >, t->tqent_id);
540 }
541
542 /*
543 * The task_expire() function takes the tq->tq_lock so drop
544 * drop the lock before synchronously cancelling the timer.
545 */
546 if (timer_pending(&t->tqent_timer)) {
547 spin_unlock_irqrestore(&tq->tq_lock, flags);
548 del_timer_sync(&t->tqent_timer);
549 spin_lock_irqsave_nested(&tq->tq_lock, flags,
550 tq->tq_lock_class);
551 }
552
553 if (!(t->tqent_flags & TQENT_FLAG_PREALLOC))
554 task_done(tq, t);
555
556 rc = 0;
557 }
558 spin_unlock_irqrestore(&tq->tq_lock, flags);
559
560 if (t == ERR_PTR(-EBUSY)) {
561 taskq_wait_id(tq, id);
562 rc = EBUSY;
563 }
564
565 return (rc);
566 }
567 EXPORT_SYMBOL(taskq_cancel_id);
568
569 static int taskq_thread_spawn(taskq_t *tq);
570
571 taskqid_t
572 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
573 {
574 taskq_ent_t *t;
575 taskqid_t rc = TASKQID_INVALID;
576 unsigned long irqflags;
577
578 ASSERT(tq);
579 ASSERT(func);
580
581 spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
582
583 /* Taskq being destroyed and all tasks drained */
584 if (!(tq->tq_flags & TASKQ_ACTIVE))
585 goto out;
586
587 /* Do not queue the task unless there is idle thread for it */
588 ASSERT(tq->tq_nactive <= tq->tq_nthreads);
589 if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
590 /* Dynamic taskq may be able to spawn another thread */
591 if (!(tq->tq_flags & TASKQ_DYNAMIC) ||
592 taskq_thread_spawn(tq) == 0)
593 goto out;
594 }
595
596 if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
597 goto out;
598
599 spin_lock(&t->tqent_lock);
600
601 /* Queue to the front of the list to enforce TQ_NOQUEUE semantics */
602 if (flags & TQ_NOQUEUE)
603 list_add(&t->tqent_list, &tq->tq_prio_list);
604 /* Queue to the priority list instead of the pending list */
605 else if (flags & TQ_FRONT)
606 list_add_tail(&t->tqent_list, &tq->tq_prio_list);
607 else
608 list_add_tail(&t->tqent_list, &tq->tq_pend_list);
609
610 t->tqent_id = rc = tq->tq_next_id;
611 tq->tq_next_id++;
612 t->tqent_func = func;
613 t->tqent_arg = arg;
614 t->tqent_taskq = tq;
615 t->tqent_timer.function = NULL;
616 t->tqent_timer.expires = 0;
617
618 t->tqent_birth = jiffies;
619 DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
620
621 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
622
623 spin_unlock(&t->tqent_lock);
624
625 wake_up(&tq->tq_work_waitq);
626 out:
627 /* Spawn additional taskq threads if required. */
628 if (!(flags & TQ_NOQUEUE) && tq->tq_nactive == tq->tq_nthreads)
629 (void) taskq_thread_spawn(tq);
630
631 spin_unlock_irqrestore(&tq->tq_lock, irqflags);
632 return (rc);
633 }
634 EXPORT_SYMBOL(taskq_dispatch);
635
636 taskqid_t
637 taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
638 uint_t flags, clock_t expire_time)
639 {
640 taskqid_t rc = TASKQID_INVALID;
641 taskq_ent_t *t;
642 unsigned long irqflags;
643
644 ASSERT(tq);
645 ASSERT(func);
646
647 spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
648
649 /* Taskq being destroyed and all tasks drained */
650 if (!(tq->tq_flags & TASKQ_ACTIVE))
651 goto out;
652
653 if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
654 goto out;
655
656 spin_lock(&t->tqent_lock);
657
658 /* Queue to the delay list for subsequent execution */
659 list_add_tail(&t->tqent_list, &tq->tq_delay_list);
660
661 t->tqent_id = rc = tq->tq_next_id;
662 tq->tq_next_id++;
663 t->tqent_func = func;
664 t->tqent_arg = arg;
665 t->tqent_taskq = tq;
666 t->tqent_timer.function = task_expire;
667 t->tqent_timer.expires = (unsigned long)expire_time;
668 add_timer(&t->tqent_timer);
669
670 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
671
672 spin_unlock(&t->tqent_lock);
673 out:
674 /* Spawn additional taskq threads if required. */
675 if (tq->tq_nactive == tq->tq_nthreads)
676 (void) taskq_thread_spawn(tq);
677 spin_unlock_irqrestore(&tq->tq_lock, irqflags);
678 return (rc);
679 }
680 EXPORT_SYMBOL(taskq_dispatch_delay);
681
682 void
683 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
684 taskq_ent_t *t)
685 {
686 unsigned long irqflags;
687 ASSERT(tq);
688 ASSERT(func);
689
690 spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
691 tq->tq_lock_class);
692
693 /* Taskq being destroyed and all tasks drained */
694 if (!(tq->tq_flags & TASKQ_ACTIVE)) {
695 t->tqent_id = TASKQID_INVALID;
696 goto out;
697 }
698
699 if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
700 /* Dynamic taskq may be able to spawn another thread */
701 if (!(tq->tq_flags & TASKQ_DYNAMIC) ||
702 taskq_thread_spawn(tq) == 0)
703 goto out2;
704 flags |= TQ_FRONT;
705 }
706
707 spin_lock(&t->tqent_lock);
708
709 /*
710 * Make sure the entry is not on some other taskq; it is important to
711 * ASSERT() under lock
712 */
713 ASSERT(taskq_empty_ent(t));
714
715 /*
716 * Mark it as a prealloc'd task. This is important
717 * to ensure that we don't free it later.
718 */
719 t->tqent_flags |= TQENT_FLAG_PREALLOC;
720
721 /* Queue to the priority list instead of the pending list */
722 if (flags & TQ_FRONT)
723 list_add_tail(&t->tqent_list, &tq->tq_prio_list);
724 else
725 list_add_tail(&t->tqent_list, &tq->tq_pend_list);
726
727 t->tqent_id = tq->tq_next_id;
728 tq->tq_next_id++;
729 t->tqent_func = func;
730 t->tqent_arg = arg;
731 t->tqent_taskq = tq;
732
733 t->tqent_birth = jiffies;
734 DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
735
736 spin_unlock(&t->tqent_lock);
737
738 wake_up(&tq->tq_work_waitq);
739 out:
740 /* Spawn additional taskq threads if required. */
741 if (tq->tq_nactive == tq->tq_nthreads)
742 (void) taskq_thread_spawn(tq);
743 out2:
744 spin_unlock_irqrestore(&tq->tq_lock, irqflags);
745 }
746 EXPORT_SYMBOL(taskq_dispatch_ent);
747
748 int
749 taskq_empty_ent(taskq_ent_t *t)
750 {
751 return (list_empty(&t->tqent_list));
752 }
753 EXPORT_SYMBOL(taskq_empty_ent);
754
755 void
756 taskq_init_ent(taskq_ent_t *t)
757 {
758 spin_lock_init(&t->tqent_lock);
759 init_waitqueue_head(&t->tqent_waitq);
760 timer_setup(&t->tqent_timer, NULL, 0);
761 INIT_LIST_HEAD(&t->tqent_list);
762 t->tqent_id = 0;
763 t->tqent_func = NULL;
764 t->tqent_arg = NULL;
765 t->tqent_flags = 0;
766 t->tqent_taskq = NULL;
767 }
768 EXPORT_SYMBOL(taskq_init_ent);
769
770 /*
771 * Return the next pending task, preference is given to tasks on the
772 * priority list which were dispatched with TQ_FRONT.
773 */
774 static taskq_ent_t *
775 taskq_next_ent(taskq_t *tq)
776 {
777 struct list_head *list;
778
779 if (!list_empty(&tq->tq_prio_list))
780 list = &tq->tq_prio_list;
781 else if (!list_empty(&tq->tq_pend_list))
782 list = &tq->tq_pend_list;
783 else
784 return (NULL);
785
786 return (list_entry(list->next, taskq_ent_t, tqent_list));
787 }
788
789 /*
790 * Spawns a new thread for the specified taskq.
791 */
792 static void
793 taskq_thread_spawn_task(void *arg)
794 {
795 taskq_t *tq = (taskq_t *)arg;
796 unsigned long flags;
797
798 if (taskq_thread_create(tq) == NULL) {
799 /* restore spawning count if failed */
800 spin_lock_irqsave_nested(&tq->tq_lock, flags,
801 tq->tq_lock_class);
802 tq->tq_nspawn--;
803 spin_unlock_irqrestore(&tq->tq_lock, flags);
804 }
805 }
806
807 /*
808 * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current
809 * number of threads is insufficient to handle the pending tasks. These
810 * new threads must be created by the dedicated dynamic_taskq to avoid
811 * deadlocks between thread creation and memory reclaim. The system_taskq
812 * which is also a dynamic taskq cannot be safely used for this.
813 */
814 static int
815 taskq_thread_spawn(taskq_t *tq)
816 {
817 int spawning = 0;
818
819 if (!(tq->tq_flags & TASKQ_DYNAMIC))
820 return (0);
821
822 if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) &&
823 (tq->tq_flags & TASKQ_ACTIVE)) {
824 spawning = (++tq->tq_nspawn);
825 taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task,
826 tq, TQ_NOSLEEP);
827 }
828
829 return (spawning);
830 }
831
832 /*
833 * Threads in a dynamic taskq should only exit once it has been completely
834 * drained and no other threads are actively servicing tasks. This prevents
835 * threads from being created and destroyed more than is required.
836 *
837 * The first thread is the thread list is treated as the primary thread.
838 * There is nothing special about the primary thread but in order to avoid
839 * all the taskq pids from changing we opt to make it long running.
840 */
841 static int
842 taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt)
843 {
844 if (!(tq->tq_flags & TASKQ_DYNAMIC))
845 return (0);
846
847 if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t,
848 tqt_thread_list) == tqt)
849 return (0);
850
851 return
852 ((tq->tq_nspawn == 0) && /* No threads are being spawned */
853 (tq->tq_nactive == 0) && /* No threads are handling tasks */
854 (tq->tq_nthreads > 1) && /* More than 1 thread is running */
855 (!taskq_next_ent(tq)) && /* There are no pending tasks */
856 (spl_taskq_thread_dynamic)); /* Dynamic taskqs are allowed */
857 }
858
859 static int
860 taskq_thread(void *args)
861 {
862 DECLARE_WAITQUEUE(wait, current);
863 sigset_t blocked;
864 taskq_thread_t *tqt = args;
865 taskq_t *tq;
866 taskq_ent_t *t;
867 int seq_tasks = 0;
868 unsigned long flags;
869 taskq_ent_t dup_task = {};
870
871 ASSERT(tqt);
872 ASSERT(tqt->tqt_tq);
873 tq = tqt->tqt_tq;
874 current->flags |= PF_NOFREEZE;
875
876 (void) spl_fstrans_mark();
877
878 sigfillset(&blocked);
879 sigprocmask(SIG_BLOCK, &blocked, NULL);
880 flush_signals(current);
881
882 tsd_set(taskq_tsd, tq);
883 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
884 /*
885 * If we are dynamically spawned, decrease spawning count. Note that
886 * we could be created during taskq_create, in which case we shouldn't
887 * do the decrement. But it's fine because taskq_create will reset
888 * tq_nspawn later.
889 */
890 if (tq->tq_flags & TASKQ_DYNAMIC)
891 tq->tq_nspawn--;
892
893 /* Immediately exit if more threads than allowed were created. */
894 if (tq->tq_nthreads >= tq->tq_maxthreads)
895 goto error;
896
897 tq->tq_nthreads++;
898 list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list);
899 wake_up(&tq->tq_wait_waitq);
900 set_current_state(TASK_INTERRUPTIBLE);
901
902 while (!kthread_should_stop()) {
903
904 if (list_empty(&tq->tq_pend_list) &&
905 list_empty(&tq->tq_prio_list)) {
906
907 if (taskq_thread_should_stop(tq, tqt)) {
908 wake_up_all(&tq->tq_wait_waitq);
909 break;
910 }
911
912 add_wait_queue_exclusive(&tq->tq_work_waitq, &wait);
913 spin_unlock_irqrestore(&tq->tq_lock, flags);
914
915 schedule();
916 seq_tasks = 0;
917
918 spin_lock_irqsave_nested(&tq->tq_lock, flags,
919 tq->tq_lock_class);
920 remove_wait_queue(&tq->tq_work_waitq, &wait);
921 } else {
922 __set_current_state(TASK_RUNNING);
923 }
924
925 if ((t = taskq_next_ent(tq)) != NULL) {
926 list_del_init(&t->tqent_list);
927
928 /*
929 * A TQENT_FLAG_PREALLOC task may be reused or freed
930 * during the task function call. Store tqent_id and
931 * tqent_flags here.
932 *
933 * Also use an on stack taskq_ent_t for tqt_task
934 * assignment in this case; we want to make sure
935 * to duplicate all fields, so the values are
936 * correct when it's accessed via DTRACE_PROBE*.
937 */
938 tqt->tqt_id = t->tqent_id;
939 tqt->tqt_flags = t->tqent_flags;
940
941 if (t->tqent_flags & TQENT_FLAG_PREALLOC) {
942 dup_task = *t;
943 t = &dup_task;
944 }
945 tqt->tqt_task = t;
946
947 taskq_insert_in_order(tq, tqt);
948 tq->tq_nactive++;
949 spin_unlock_irqrestore(&tq->tq_lock, flags);
950
951 DTRACE_PROBE1(taskq_ent__start, taskq_ent_t *, t);
952
953 /* Perform the requested task */
954 t->tqent_func(t->tqent_arg);
955
956 DTRACE_PROBE1(taskq_ent__finish, taskq_ent_t *, t);
957
958 spin_lock_irqsave_nested(&tq->tq_lock, flags,
959 tq->tq_lock_class);
960 tq->tq_nactive--;
961 list_del_init(&tqt->tqt_active_list);
962 tqt->tqt_task = NULL;
963
964 /* For prealloc'd tasks, we don't free anything. */
965 if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
966 task_done(tq, t);
967
968 /*
969 * When the current lowest outstanding taskqid is
970 * done calculate the new lowest outstanding id
971 */
972 if (tq->tq_lowest_id == tqt->tqt_id) {
973 tq->tq_lowest_id = taskq_lowest_id(tq);
974 ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
975 }
976
977 /* Spawn additional taskq threads if required. */
978 if ((++seq_tasks) > spl_taskq_thread_sequential &&
979 taskq_thread_spawn(tq))
980 seq_tasks = 0;
981
982 tqt->tqt_id = TASKQID_INVALID;
983 tqt->tqt_flags = 0;
984 wake_up_all(&tq->tq_wait_waitq);
985 } else {
986 if (taskq_thread_should_stop(tq, tqt))
987 break;
988 }
989
990 set_current_state(TASK_INTERRUPTIBLE);
991
992 }
993
994 __set_current_state(TASK_RUNNING);
995 tq->tq_nthreads--;
996 list_del_init(&tqt->tqt_thread_list);
997 error:
998 kmem_free(tqt, sizeof (taskq_thread_t));
999 spin_unlock_irqrestore(&tq->tq_lock, flags);
1000
1001 tsd_set(taskq_tsd, NULL);
1002 thread_exit();
1003
1004 return (0);
1005 }
1006
1007 static taskq_thread_t *
1008 taskq_thread_create(taskq_t *tq)
1009 {
1010 static int last_used_cpu = 0;
1011 taskq_thread_t *tqt;
1012
1013 tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE);
1014 INIT_LIST_HEAD(&tqt->tqt_thread_list);
1015 INIT_LIST_HEAD(&tqt->tqt_active_list);
1016 tqt->tqt_tq = tq;
1017 tqt->tqt_id = TASKQID_INVALID;
1018
1019 tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt,
1020 "%s", tq->tq_name);
1021 if (tqt->tqt_thread == NULL) {
1022 kmem_free(tqt, sizeof (taskq_thread_t));
1023 return (NULL);
1024 }
1025
1026 if (spl_taskq_thread_bind) {
1027 last_used_cpu = (last_used_cpu + 1) % num_online_cpus();
1028 kthread_bind(tqt->tqt_thread, last_used_cpu);
1029 }
1030
1031 if (spl_taskq_thread_priority)
1032 set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri));
1033
1034 wake_up_process(tqt->tqt_thread);
1035
1036 return (tqt);
1037 }
1038
1039 taskq_t *
1040 taskq_create(const char *name, int threads_arg, pri_t pri,
1041 int minalloc, int maxalloc, uint_t flags)
1042 {
1043 taskq_t *tq;
1044 taskq_thread_t *tqt;
1045 int count = 0, rc = 0, i;
1046 unsigned long irqflags;
1047 int nthreads = threads_arg;
1048
1049 ASSERT(name != NULL);
1050 ASSERT(minalloc >= 0);
1051 ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */
1052
1053 /* Scale the number of threads using nthreads as a percentage */
1054 if (flags & TASKQ_THREADS_CPU_PCT) {
1055 ASSERT(nthreads <= 100);
1056 ASSERT(nthreads >= 0);
1057 nthreads = MIN(threads_arg, 100);
1058 nthreads = MAX(nthreads, 0);
1059 nthreads = MAX((num_online_cpus() * nthreads) /100, 1);
1060 }
1061
1062 tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE);
1063 if (tq == NULL)
1064 return (NULL);
1065
1066 tq->tq_hp_support = B_FALSE;
1067 #ifdef HAVE_CPU_HOTPLUG
1068 if (flags & TASKQ_THREADS_CPU_PCT) {
1069 tq->tq_hp_support = B_TRUE;
1070 if (cpuhp_state_add_instance_nocalls(spl_taskq_cpuhp_state,
1071 &tq->tq_hp_cb_node) != 0) {
1072 kmem_free(tq, sizeof (*tq));
1073 return (NULL);
1074 }
1075 }
1076 #endif
1077
1078 spin_lock_init(&tq->tq_lock);
1079 INIT_LIST_HEAD(&tq->tq_thread_list);
1080 INIT_LIST_HEAD(&tq->tq_active_list);
1081 tq->tq_name = kmem_strdup(name);
1082 tq->tq_nactive = 0;
1083 tq->tq_nthreads = 0;
1084 tq->tq_nspawn = 0;
1085 tq->tq_maxthreads = nthreads;
1086 tq->tq_cpu_pct = threads_arg;
1087 tq->tq_pri = pri;
1088 tq->tq_minalloc = minalloc;
1089 tq->tq_maxalloc = maxalloc;
1090 tq->tq_nalloc = 0;
1091 tq->tq_flags = (flags | TASKQ_ACTIVE);
1092 tq->tq_next_id = TASKQID_INITIAL;
1093 tq->tq_lowest_id = TASKQID_INITIAL;
1094 INIT_LIST_HEAD(&tq->tq_free_list);
1095 INIT_LIST_HEAD(&tq->tq_pend_list);
1096 INIT_LIST_HEAD(&tq->tq_prio_list);
1097 INIT_LIST_HEAD(&tq->tq_delay_list);
1098 init_waitqueue_head(&tq->tq_work_waitq);
1099 init_waitqueue_head(&tq->tq_wait_waitq);
1100 tq->tq_lock_class = TQ_LOCK_GENERAL;
1101 INIT_LIST_HEAD(&tq->tq_taskqs);
1102
1103 if (flags & TASKQ_PREPOPULATE) {
1104 spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
1105 tq->tq_lock_class);
1106
1107 for (i = 0; i < minalloc; i++)
1108 task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW,
1109 &irqflags));
1110
1111 spin_unlock_irqrestore(&tq->tq_lock, irqflags);
1112 }
1113
1114 if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic)
1115 nthreads = 1;
1116
1117 for (i = 0; i < nthreads; i++) {
1118 tqt = taskq_thread_create(tq);
1119 if (tqt == NULL)
1120 rc = 1;
1121 else
1122 count++;
1123 }
1124
1125 /* Wait for all threads to be started before potential destroy */
1126 wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count);
1127 /*
1128 * taskq_thread might have touched nspawn, but we don't want them to
1129 * because they're not dynamically spawned. So we reset it to 0
1130 */
1131 tq->tq_nspawn = 0;
1132
1133 if (rc) {
1134 taskq_destroy(tq);
1135 tq = NULL;
1136 } else {
1137 down_write(&tq_list_sem);
1138 tq->tq_instance = taskq_find_by_name(name) + 1;
1139 list_add_tail(&tq->tq_taskqs, &tq_list);
1140 up_write(&tq_list_sem);
1141 }
1142
1143 return (tq);
1144 }
1145 EXPORT_SYMBOL(taskq_create);
1146
1147 void
1148 taskq_destroy(taskq_t *tq)
1149 {
1150 struct task_struct *thread;
1151 taskq_thread_t *tqt;
1152 taskq_ent_t *t;
1153 unsigned long flags;
1154
1155 ASSERT(tq);
1156 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1157 tq->tq_flags &= ~TASKQ_ACTIVE;
1158 spin_unlock_irqrestore(&tq->tq_lock, flags);
1159
1160 #ifdef HAVE_CPU_HOTPLUG
1161 if (tq->tq_hp_support) {
1162 VERIFY0(cpuhp_state_remove_instance_nocalls(
1163 spl_taskq_cpuhp_state, &tq->tq_hp_cb_node));
1164 }
1165 #endif
1166 /*
1167 * When TASKQ_ACTIVE is clear new tasks may not be added nor may
1168 * new worker threads be spawned for dynamic taskq.
1169 */
1170 if (dynamic_taskq != NULL)
1171 taskq_wait_outstanding(dynamic_taskq, 0);
1172
1173 taskq_wait(tq);
1174
1175 /* remove taskq from global list used by the kstats */
1176 down_write(&tq_list_sem);
1177 list_del(&tq->tq_taskqs);
1178 up_write(&tq_list_sem);
1179
1180 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1181 /* wait for spawning threads to insert themselves to the list */
1182 while (tq->tq_nspawn) {
1183 spin_unlock_irqrestore(&tq->tq_lock, flags);
1184 schedule_timeout_interruptible(1);
1185 spin_lock_irqsave_nested(&tq->tq_lock, flags,
1186 tq->tq_lock_class);
1187 }
1188
1189 /*
1190 * Signal each thread to exit and block until it does. Each thread
1191 * is responsible for removing itself from the list and freeing its
1192 * taskq_thread_t. This allows for idle threads to opt to remove
1193 * themselves from the taskq. They can be recreated as needed.
1194 */
1195 while (!list_empty(&tq->tq_thread_list)) {
1196 tqt = list_entry(tq->tq_thread_list.next,
1197 taskq_thread_t, tqt_thread_list);
1198 thread = tqt->tqt_thread;
1199 spin_unlock_irqrestore(&tq->tq_lock, flags);
1200
1201 kthread_stop(thread);
1202
1203 spin_lock_irqsave_nested(&tq->tq_lock, flags,
1204 tq->tq_lock_class);
1205 }
1206
1207 while (!list_empty(&tq->tq_free_list)) {
1208 t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
1209
1210 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
1211
1212 list_del_init(&t->tqent_list);
1213 task_free(tq, t);
1214 }
1215
1216 ASSERT0(tq->tq_nthreads);
1217 ASSERT0(tq->tq_nalloc);
1218 ASSERT0(tq->tq_nspawn);
1219 ASSERT(list_empty(&tq->tq_thread_list));
1220 ASSERT(list_empty(&tq->tq_active_list));
1221 ASSERT(list_empty(&tq->tq_free_list));
1222 ASSERT(list_empty(&tq->tq_pend_list));
1223 ASSERT(list_empty(&tq->tq_prio_list));
1224 ASSERT(list_empty(&tq->tq_delay_list));
1225
1226 spin_unlock_irqrestore(&tq->tq_lock, flags);
1227
1228 kmem_strfree(tq->tq_name);
1229 kmem_free(tq, sizeof (taskq_t));
1230 }
1231 EXPORT_SYMBOL(taskq_destroy);
1232
1233 static unsigned int spl_taskq_kick = 0;
1234
1235 /*
1236 * 2.6.36 API Change
1237 * module_param_cb is introduced to take kernel_param_ops and
1238 * module_param_call is marked as obsolete. Also set and get operations
1239 * were changed to take a 'const struct kernel_param *'.
1240 */
1241 static int
1242 #ifdef module_param_cb
1243 param_set_taskq_kick(const char *val, const struct kernel_param *kp)
1244 #else
1245 param_set_taskq_kick(const char *val, struct kernel_param *kp)
1246 #endif
1247 {
1248 int ret;
1249 taskq_t *tq = NULL;
1250 taskq_ent_t *t;
1251 unsigned long flags;
1252
1253 ret = param_set_uint(val, kp);
1254 if (ret < 0 || !spl_taskq_kick)
1255 return (ret);
1256 /* reset value */
1257 spl_taskq_kick = 0;
1258
1259 down_read(&tq_list_sem);
1260 list_for_each_entry(tq, &tq_list, tq_taskqs) {
1261 spin_lock_irqsave_nested(&tq->tq_lock, flags,
1262 tq->tq_lock_class);
1263 /* Check if the first pending is older than 5 seconds */
1264 t = taskq_next_ent(tq);
1265 if (t && time_after(jiffies, t->tqent_birth + 5*HZ)) {
1266 (void) taskq_thread_spawn(tq);
1267 printk(KERN_INFO "spl: Kicked taskq %s/%d\n",
1268 tq->tq_name, tq->tq_instance);
1269 }
1270 spin_unlock_irqrestore(&tq->tq_lock, flags);
1271 }
1272 up_read(&tq_list_sem);
1273 return (ret);
1274 }
1275
1276 #ifdef module_param_cb
1277 static const struct kernel_param_ops param_ops_taskq_kick = {
1278 .set = param_set_taskq_kick,
1279 .get = param_get_uint,
1280 };
1281 module_param_cb(spl_taskq_kick, ¶m_ops_taskq_kick, &spl_taskq_kick, 0644);
1282 #else
1283 module_param_call(spl_taskq_kick, param_set_taskq_kick, param_get_uint,
1284 &spl_taskq_kick, 0644);
1285 #endif
1286 MODULE_PARM_DESC(spl_taskq_kick,
1287 "Write nonzero to kick stuck taskqs to spawn more threads");
1288
1289 #ifdef HAVE_CPU_HOTPLUG
1290 /*
1291 * This callback will be called exactly once for each core that comes online,
1292 * for each dynamic taskq. We attempt to expand taskqs that have
1293 * TASKQ_THREADS_CPU_PCT set. We need to redo the percentage calculation every
1294 * time, to correctly determine whether or not to add a thread.
1295 */
1296 static int
1297 spl_taskq_expand(unsigned int cpu, struct hlist_node *node)
1298 {
1299 taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1300 unsigned long flags;
1301 int err = 0;
1302
1303 ASSERT(tq);
1304 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1305
1306 if (!(tq->tq_flags & TASKQ_ACTIVE)) {
1307 spin_unlock_irqrestore(&tq->tq_lock, flags);
1308 return (err);
1309 }
1310
1311 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1312 int nthreads = MIN(tq->tq_cpu_pct, 100);
1313 nthreads = MAX(((num_online_cpus() + 1) * nthreads) / 100, 1);
1314 tq->tq_maxthreads = nthreads;
1315
1316 if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1317 tq->tq_maxthreads > tq->tq_nthreads) {
1318 spin_unlock_irqrestore(&tq->tq_lock, flags);
1319 taskq_thread_t *tqt = taskq_thread_create(tq);
1320 if (tqt == NULL)
1321 err = -1;
1322 return (err);
1323 }
1324 spin_unlock_irqrestore(&tq->tq_lock, flags);
1325 return (err);
1326 }
1327
1328 /*
1329 * While we don't support offlining CPUs, it is possible that CPUs will fail
1330 * to online successfully. We do need to be able to handle this case
1331 * gracefully.
1332 */
1333 static int
1334 spl_taskq_prepare_down(unsigned int cpu, struct hlist_node *node)
1335 {
1336 taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1337 unsigned long flags;
1338
1339 ASSERT(tq);
1340 spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1341
1342 if (!(tq->tq_flags & TASKQ_ACTIVE))
1343 goto out;
1344
1345 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1346 int nthreads = MIN(tq->tq_cpu_pct, 100);
1347 nthreads = MAX(((num_online_cpus()) * nthreads) / 100, 1);
1348 tq->tq_maxthreads = nthreads;
1349
1350 if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1351 tq->tq_maxthreads < tq->tq_nthreads) {
1352 ASSERT3U(tq->tq_maxthreads, ==, tq->tq_nthreads - 1);
1353 taskq_thread_t *tqt = list_entry(tq->tq_thread_list.next,
1354 taskq_thread_t, tqt_thread_list);
1355 struct task_struct *thread = tqt->tqt_thread;
1356 spin_unlock_irqrestore(&tq->tq_lock, flags);
1357
1358 kthread_stop(thread);
1359
1360 return (0);
1361 }
1362
1363 out:
1364 spin_unlock_irqrestore(&tq->tq_lock, flags);
1365 return (0);
1366 }
1367 #endif
1368
1369 int
1370 spl_taskq_init(void)
1371 {
1372 init_rwsem(&tq_list_sem);
1373 tsd_create(&taskq_tsd, NULL);
1374
1375 #ifdef HAVE_CPU_HOTPLUG
1376 spl_taskq_cpuhp_state = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1377 "fs/spl_taskq:online", spl_taskq_expand, spl_taskq_prepare_down);
1378 #endif
1379
1380 system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64),
1381 maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1382 if (system_taskq == NULL)
1383 return (-ENOMEM);
1384
1385 system_delay_taskq = taskq_create("spl_delay_taskq", MAX(boot_ncpus, 4),
1386 maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1387 if (system_delay_taskq == NULL) {
1388 #ifdef HAVE_CPU_HOTPLUG
1389 cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1390 #endif
1391 taskq_destroy(system_taskq);
1392 return (-ENOMEM);
1393 }
1394
1395 dynamic_taskq = taskq_create("spl_dynamic_taskq", 1,
1396 maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE);
1397 if (dynamic_taskq == NULL) {
1398 #ifdef HAVE_CPU_HOTPLUG
1399 cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1400 #endif
1401 taskq_destroy(system_taskq);
1402 taskq_destroy(system_delay_taskq);
1403 return (-ENOMEM);
1404 }
1405
1406 /*
1407 * This is used to annotate tq_lock, so
1408 * taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch
1409 * does not trigger a lockdep warning re: possible recursive locking
1410 */
1411 dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC;
1412
1413 return (0);
1414 }
1415
1416 void
1417 spl_taskq_fini(void)
1418 {
1419 taskq_destroy(dynamic_taskq);
1420 dynamic_taskq = NULL;
1421
1422 taskq_destroy(system_delay_taskq);
1423 system_delay_taskq = NULL;
1424
1425 taskq_destroy(system_taskq);
1426 system_taskq = NULL;
1427
1428 tsd_destroy(&taskq_tsd);
1429
1430 #ifdef HAVE_CPU_HOTPLUG
1431 cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1432 spl_taskq_cpuhp_state = 0;
1433 #endif
1434 }
Cache object: 9017603e4816759dde23c7396c0155f7
|