]>
git.proxmox.com Git - mirror_spl.git/blob - module/spl/spl-taskq.c
5c22544b8b14c53b9234da28ba0af5e59e463ae6
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>.
8 * This file is part of the SPL, Solaris Porting Layer.
9 * For details, see <http://github.com/behlendorf/spl/>.
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.
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
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 \*****************************************************************************/
27 #include <sys/taskq.h>
29 #include <spl-debug.h>
31 #ifdef SS_DEBUG_SUBSYS
32 #undef SS_DEBUG_SUBSYS
35 #define SS_DEBUG_SUBSYS SS_TASKQ
37 /* Global system-wide dynamic task queue available for all consumers */
38 taskq_t
*system_taskq
;
39 EXPORT_SYMBOL(system_taskq
);
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.
46 task_alloc(taskq_t
*tq
, uint_t flags
)
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
));
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 list_del_init(&t
->tqent_list
);
64 /* Free list is empty and memory allocations are prohibited */
65 if (flags
& TQ_NOALLOC
)
68 /* Hit maximum taskq_ent_t pool size */
69 if (tq
->tq_nalloc
>= tq
->tq_maxalloc
) {
70 if (flags
& TQ_NOSLEEP
)
74 * Sleep periodically polling the free list for an available
75 * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
76 * but we cannot block forever waiting for an taskq_entq_t to
77 * show up in the free list, otherwise a deadlock can happen.
79 * Therefore, we need to allocate a new task even if the number
80 * of allocated tasks is above tq->tq_maxalloc, but we still
81 * end up delaying the task allocation by one second, thereby
82 * throttling the task dispatch rate.
84 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
85 schedule_timeout(HZ
/ 100);
86 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
88 SGOTO(retry
, count
++);
91 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
92 t
= kmem_alloc(sizeof(taskq_ent_t
), flags
& (TQ_SLEEP
| TQ_NOSLEEP
));
93 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
96 spin_lock_init(&t
->tqent_lock
);
97 INIT_LIST_HEAD(&t
->tqent_list
);
108 * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t
109 * to already be removed from the free, work, or pending taskq lists.
112 task_free(taskq_t
*tq
, taskq_ent_t
*t
)
118 ASSERT(spin_is_locked(&tq
->tq_lock
));
119 ASSERT(list_empty(&t
->tqent_list
));
121 kmem_free(t
, sizeof(taskq_ent_t
));
128 * NOTE: Must be called with tq->tq_lock held, either destroys the
129 * taskq_ent_t if too many exist or moves it to the free list for later use.
132 task_done(taskq_t
*tq
, taskq_ent_t
*t
)
137 ASSERT(spin_is_locked(&tq
->tq_lock
));
139 list_del_init(&t
->tqent_list
);
141 if (tq
->tq_nalloc
<= tq
->tq_minalloc
) {
143 t
->tqent_func
= NULL
;
145 list_add_tail(&t
->tqent_list
, &tq
->tq_free_list
);
154 * As tasks are submitted to the task queue they are assigned a
155 * monotonically increasing taskqid and added to the tail of the pending
156 * list. As worker threads become available the tasks are removed from
157 * the head of the pending or priority list, giving preference to the
158 * priority list. The tasks are then added to the work list, preserving
159 * the ordering by taskqid. Finally, as tasks complete they are removed
160 * from the work list. This means that the pending and work lists are
161 * always kept sorted by taskqid. Thus the lowest outstanding
162 * incomplete taskqid can be determined simply by checking the min
163 * taskqid for each head item on the pending, priority, and work list.
164 * This value is stored in tq->tq_lowest_id and only updated to the new
165 * lowest id when the previous lowest id completes. All taskqids lower
166 * than tq->tq_lowest_id must have completed. It is also possible
167 * larger taskqid's have completed because they may be processed in
168 * parallel by several worker threads. However, this is not a problem
169 * because the behavior of taskq_wait_id() is to block until all
170 * previously submitted taskqid's have completed.
172 * XXX: Taskqid_t wrapping is not handled. However, taskqid_t's are
173 * 64-bit values so even if a taskq is processing 2^24 (16,777,216)
174 * taskqid_ts per second it will still take 2^40 seconds, 34,865 years,
175 * before the wrap occurs. I can live with that for now.
178 taskq_wait_check(taskq_t
*tq
, taskqid_t id
)
182 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
183 rc
= (id
< tq
->tq_lowest_id
);
184 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
190 __taskq_wait_id(taskq_t
*tq
, taskqid_t id
)
195 wait_event(tq
->tq_wait_waitq
, taskq_wait_check(tq
, id
));
199 EXPORT_SYMBOL(__taskq_wait_id
);
202 __taskq_wait(taskq_t
*tq
)
208 /* Wait for the largest outstanding taskqid */
209 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
210 id
= tq
->tq_next_id
- 1;
211 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
213 __taskq_wait_id(tq
, id
);
218 EXPORT_SYMBOL(__taskq_wait
);
221 __taskq_member(taskq_t
*tq
, void *t
)
230 list_for_each(l
, &tq
->tq_thread_list
) {
231 tqt
= list_entry(l
, taskq_thread_t
, tqt_thread_list
);
232 if (tqt
->tqt_thread
== (struct task_struct
*)t
)
238 EXPORT_SYMBOL(__taskq_member
);
241 __taskq_dispatch(taskq_t
*tq
, task_func_t func
, void *arg
, uint_t flags
)
250 /* Solaris assumes TQ_SLEEP if not passed explicitly */
251 if (!(flags
& (TQ_SLEEP
| TQ_NOSLEEP
)))
254 if (unlikely(in_atomic() && (flags
& TQ_SLEEP
)))
255 PANIC("May schedule while atomic: %s/0x%08x/%d\n",
256 current
->comm
, preempt_count(), current
->pid
);
258 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
260 /* Taskq being destroyed and all tasks drained */
261 if (!(tq
->tq_flags
& TQ_ACTIVE
))
264 /* Do not queue the task unless there is idle thread for it */
265 ASSERT(tq
->tq_nactive
<= tq
->tq_nthreads
);
266 if ((flags
& TQ_NOQUEUE
) && (tq
->tq_nactive
== tq
->tq_nthreads
))
269 if ((t
= task_alloc(tq
, flags
)) == NULL
)
272 spin_lock(&t
->tqent_lock
);
274 /* Queue to the priority list instead of the pending list */
275 if (flags
& TQ_FRONT
)
276 list_add_tail(&t
->tqent_list
, &tq
->tq_prio_list
);
278 list_add_tail(&t
->tqent_list
, &tq
->tq_pend_list
);
280 t
->tqent_id
= rc
= tq
->tq_next_id
;
282 t
->tqent_func
= func
;
284 spin_unlock(&t
->tqent_lock
);
286 wake_up(&tq
->tq_work_waitq
);
288 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
291 EXPORT_SYMBOL(__taskq_dispatch
);
293 * Returns the lowest incomplete taskqid_t. The taskqid_t may
294 * be queued on the pending list, on the priority list, or on
295 * the work list currently being handled, but it is not 100%
299 taskq_lowest_id(taskq_t
*tq
)
301 taskqid_t lowest_id
= tq
->tq_next_id
;
307 ASSERT(spin_is_locked(&tq
->tq_lock
));
309 if (!list_empty(&tq
->tq_pend_list
)) {
310 t
= list_entry(tq
->tq_pend_list
.next
, taskq_ent_t
, tqent_list
);
311 lowest_id
= MIN(lowest_id
, t
->tqent_id
);
314 if (!list_empty(&tq
->tq_prio_list
)) {
315 t
= list_entry(tq
->tq_prio_list
.next
, taskq_ent_t
, tqent_list
);
316 lowest_id
= MIN(lowest_id
, t
->tqent_id
);
319 if (!list_empty(&tq
->tq_active_list
)) {
320 tqt
= list_entry(tq
->tq_active_list
.next
, taskq_thread_t
,
322 ASSERT(tqt
->tqt_ent
!= NULL
);
323 lowest_id
= MIN(lowest_id
, tqt
->tqt_ent
->tqent_id
);
330 * Insert a task into a list keeping the list sorted by increasing
334 taskq_insert_in_order(taskq_t
*tq
, taskq_thread_t
*tqt
)
342 ASSERT(spin_is_locked(&tq
->tq_lock
));
344 list_for_each_prev(l
, &tq
->tq_active_list
) {
345 w
= list_entry(l
, taskq_thread_t
, tqt_active_list
);
346 if (w
->tqt_ent
->tqent_id
< tqt
->tqt_ent
->tqent_id
) {
347 list_add(&tqt
->tqt_active_list
, l
);
351 if (l
== &tq
->tq_active_list
)
352 list_add(&tqt
->tqt_active_list
, &tq
->tq_active_list
);
358 taskq_thread(void *args
)
360 DECLARE_WAITQUEUE(wait
, current
);
363 taskq_thread_t
*tqt
= args
;
366 struct list_head
*pend_list
;
371 current
->flags
|= PF_NOFREEZE
;
373 /* Disable the direct memory reclaim path */
374 if (tq
->tq_flags
& TASKQ_NORECLAIM
)
375 current
->flags
|= PF_MEMALLOC
;
377 sigfillset(&blocked
);
378 sigprocmask(SIG_BLOCK
, &blocked
, NULL
);
379 flush_signals(current
);
381 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
383 wake_up(&tq
->tq_wait_waitq
);
384 set_current_state(TASK_INTERRUPTIBLE
);
386 while (!kthread_should_stop()) {
388 add_wait_queue(&tq
->tq_work_waitq
, &wait
);
389 if (list_empty(&tq
->tq_pend_list
) &&
390 list_empty(&tq
->tq_prio_list
)) {
391 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
393 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
395 __set_current_state(TASK_RUNNING
);
398 remove_wait_queue(&tq
->tq_work_waitq
, &wait
);
400 if (!list_empty(&tq
->tq_prio_list
))
401 pend_list
= &tq
->tq_prio_list
;
402 else if (!list_empty(&tq
->tq_pend_list
))
403 pend_list
= &tq
->tq_pend_list
;
408 t
= list_entry(pend_list
->next
, taskq_ent_t
, tqent_list
);
409 list_del_init(&t
->tqent_list
);
411 taskq_insert_in_order(tq
, tqt
);
413 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
415 /* Perform the requested task */
416 t
->tqent_func(t
->tqent_arg
);
418 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
420 list_del_init(&tqt
->tqt_active_list
);
425 /* When the current lowest outstanding taskqid is
426 * done calculate the new lowest outstanding id */
427 if (tq
->tq_lowest_id
== id
) {
428 tq
->tq_lowest_id
= taskq_lowest_id(tq
);
429 ASSERT(tq
->tq_lowest_id
> id
);
432 wake_up_all(&tq
->tq_wait_waitq
);
435 set_current_state(TASK_INTERRUPTIBLE
);
439 __set_current_state(TASK_RUNNING
);
441 list_del_init(&tqt
->tqt_thread_list
);
442 kmem_free(tqt
, sizeof(taskq_thread_t
));
444 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
450 __taskq_create(const char *name
, int nthreads
, pri_t pri
,
451 int minalloc
, int maxalloc
, uint_t flags
)
455 int rc
= 0, i
, j
= 0;
458 ASSERT(name
!= NULL
);
459 ASSERT(pri
<= maxclsyspri
);
460 ASSERT(minalloc
>= 0);
461 ASSERT(maxalloc
<= INT_MAX
);
462 ASSERT(!(flags
& (TASKQ_CPR_SAFE
| TASKQ_DYNAMIC
))); /* Unsupported */
464 /* Scale the number of threads using nthreads as a percentage */
465 if (flags
& TASKQ_THREADS_CPU_PCT
) {
466 ASSERT(nthreads
<= 100);
467 ASSERT(nthreads
>= 0);
468 nthreads
= MIN(nthreads
, 100);
469 nthreads
= MAX(nthreads
, 0);
470 nthreads
= MAX((num_online_cpus() * nthreads
) / 100, 1);
473 tq
= kmem_alloc(sizeof(*tq
), KM_SLEEP
);
477 spin_lock_init(&tq
->tq_lock
);
478 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
479 INIT_LIST_HEAD(&tq
->tq_thread_list
);
480 INIT_LIST_HEAD(&tq
->tq_active_list
);
485 tq
->tq_minalloc
= minalloc
;
486 tq
->tq_maxalloc
= maxalloc
;
488 tq
->tq_flags
= (flags
| TQ_ACTIVE
);
490 tq
->tq_lowest_id
= 1;
491 INIT_LIST_HEAD(&tq
->tq_free_list
);
492 INIT_LIST_HEAD(&tq
->tq_pend_list
);
493 INIT_LIST_HEAD(&tq
->tq_prio_list
);
494 init_waitqueue_head(&tq
->tq_work_waitq
);
495 init_waitqueue_head(&tq
->tq_wait_waitq
);
497 if (flags
& TASKQ_PREPOPULATE
)
498 for (i
= 0; i
< minalloc
; i
++)
499 task_done(tq
, task_alloc(tq
, TQ_SLEEP
| TQ_NEW
));
501 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
503 for (i
= 0; i
< nthreads
; i
++) {
504 tqt
= kmem_alloc(sizeof(*tqt
), KM_SLEEP
);
505 INIT_LIST_HEAD(&tqt
->tqt_thread_list
);
506 INIT_LIST_HEAD(&tqt
->tqt_active_list
);
510 tqt
->tqt_thread
= kthread_create(taskq_thread
, tqt
,
512 if (tqt
->tqt_thread
) {
513 list_add(&tqt
->tqt_thread_list
, &tq
->tq_thread_list
);
514 kthread_bind(tqt
->tqt_thread
, i
% num_online_cpus());
515 set_user_nice(tqt
->tqt_thread
, PRIO_TO_NICE(pri
));
516 wake_up_process(tqt
->tqt_thread
);
519 kmem_free(tqt
, sizeof(taskq_thread_t
));
524 /* Wait for all threads to be started before potential destroy */
525 wait_event(tq
->tq_wait_waitq
, tq
->tq_nthreads
== j
);
534 EXPORT_SYMBOL(__taskq_create
);
537 __taskq_destroy(taskq_t
*tq
)
539 struct task_struct
*thread
;
545 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
546 tq
->tq_flags
&= ~TQ_ACTIVE
;
547 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
549 /* TQ_ACTIVE cleared prevents new tasks being added to pending */
552 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
555 * Signal each thread to exit and block until it does. Each thread
556 * is responsible for removing itself from the list and freeing its
557 * taskq_thread_t. This allows for idle threads to opt to remove
558 * themselves from the taskq. They can be recreated as needed.
560 while (!list_empty(&tq
->tq_thread_list
)) {
561 tqt
= list_entry(tq
->tq_thread_list
.next
,
562 taskq_thread_t
, tqt_thread_list
);
563 thread
= tqt
->tqt_thread
;
564 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
566 kthread_stop(thread
);
568 spin_lock_irqsave(&tq
->tq_lock
, tq
->tq_lock_flags
);
571 while (!list_empty(&tq
->tq_free_list
)) {
572 t
= list_entry(tq
->tq_free_list
.next
, taskq_ent_t
, tqent_list
);
573 list_del_init(&t
->tqent_list
);
577 ASSERT(tq
->tq_nthreads
== 0);
578 ASSERT(tq
->tq_nalloc
== 0);
579 ASSERT(list_empty(&tq
->tq_thread_list
));
580 ASSERT(list_empty(&tq
->tq_active_list
));
581 ASSERT(list_empty(&tq
->tq_free_list
));
582 ASSERT(list_empty(&tq
->tq_pend_list
));
583 ASSERT(list_empty(&tq
->tq_prio_list
));
585 spin_unlock_irqrestore(&tq
->tq_lock
, tq
->tq_lock_flags
);
587 kmem_free(tq
, sizeof(taskq_t
));
591 EXPORT_SYMBOL(__taskq_destroy
);
598 /* Solaris creates a dynamic taskq of up to 64 threads, however in
599 * a Linux environment 1 thread per-core is usually about right */
600 system_taskq
= taskq_create("spl_system_taskq", num_online_cpus(),
601 minclsyspri
, 4, 512, TASKQ_PREPOPULATE
);
602 if (system_taskq
== NULL
)
612 taskq_destroy(system_taskq
);