]>
Commit | Line | Data |
---|---|---|
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://zfsonlinux.org/>. | |
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 | ||
30 | int spl_taskq_thread_bind = 0; | |
31 | module_param(spl_taskq_thread_bind, int, 0644); | |
32 | MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default"); | |
33 | ||
34 | ||
35 | int spl_taskq_thread_dynamic = 1; | |
36 | module_param(spl_taskq_thread_dynamic, int, 0644); | |
37 | MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads"); | |
38 | ||
39 | int spl_taskq_thread_priority = 1; | |
40 | module_param(spl_taskq_thread_priority, int, 0644); | |
41 | MODULE_PARM_DESC(spl_taskq_thread_priority, | |
42 | "Allow non-default priority for taskq threads"); | |
43 | ||
44 | int spl_taskq_thread_sequential = 4; | |
45 | module_param(spl_taskq_thread_sequential, int, 0644); | |
46 | MODULE_PARM_DESC(spl_taskq_thread_sequential, | |
47 | "Create new taskq threads after N sequential tasks"); | |
48 | ||
49 | /* Global system-wide dynamic task queue available for all consumers */ | |
50 | taskq_t *system_taskq; | |
51 | EXPORT_SYMBOL(system_taskq); | |
52 | ||
53 | /* Private dedicated taskq for creating new taskq threads on demand. */ | |
54 | static taskq_t *dynamic_taskq; | |
55 | static taskq_thread_t *taskq_thread_create(taskq_t *); | |
56 | ||
57 | static int | |
58 | task_km_flags(uint_t flags) | |
59 | { | |
60 | if (flags & TQ_NOSLEEP) | |
61 | return KM_NOSLEEP; | |
62 | ||
63 | if (flags & TQ_PUSHPAGE) | |
64 | return KM_PUSHPAGE; | |
65 | ||
66 | return KM_SLEEP; | |
67 | } | |
68 | ||
69 | /* | |
70 | * NOTE: Must be called with tq->tq_lock held, returns a list_t which | |
71 | * is not attached to the free, work, or pending taskq lists. | |
72 | */ | |
73 | static taskq_ent_t * | |
74 | task_alloc(taskq_t *tq, uint_t flags) | |
75 | { | |
76 | taskq_ent_t *t; | |
77 | int count = 0; | |
78 | ||
79 | ASSERT(tq); | |
80 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
81 | retry: | |
82 | /* Acquire taskq_ent_t's from free list if available */ | |
83 | if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) { | |
84 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); | |
85 | ||
86 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
87 | ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL)); | |
88 | ASSERT(!timer_pending(&t->tqent_timer)); | |
89 | ||
90 | list_del_init(&t->tqent_list); | |
91 | return (t); | |
92 | } | |
93 | ||
94 | /* Free list is empty and memory allocations are prohibited */ | |
95 | if (flags & TQ_NOALLOC) | |
96 | return (NULL); | |
97 | ||
98 | /* Hit maximum taskq_ent_t pool size */ | |
99 | if (tq->tq_nalloc >= tq->tq_maxalloc) { | |
100 | if (flags & TQ_NOSLEEP) | |
101 | return (NULL); | |
102 | ||
103 | /* | |
104 | * Sleep periodically polling the free list for an available | |
105 | * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed | |
106 | * but we cannot block forever waiting for an taskq_ent_t to | |
107 | * show up in the free list, otherwise a deadlock can happen. | |
108 | * | |
109 | * Therefore, we need to allocate a new task even if the number | |
110 | * of allocated tasks is above tq->tq_maxalloc, but we still | |
111 | * end up delaying the task allocation by one second, thereby | |
112 | * throttling the task dispatch rate. | |
113 | */ | |
114 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
115 | schedule_timeout(HZ / 100); | |
116 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
117 | if (count < 100) { | |
118 | count++; | |
119 | goto retry; | |
120 | } | |
121 | } | |
122 | ||
123 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
124 | t = kmem_alloc(sizeof(taskq_ent_t), task_km_flags(flags)); | |
125 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
126 | ||
127 | if (t) { | |
128 | taskq_init_ent(t); | |
129 | tq->tq_nalloc++; | |
130 | } | |
131 | ||
132 | return (t); | |
133 | } | |
134 | ||
135 | /* | |
136 | * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t | |
137 | * to already be removed from the free, work, or pending taskq lists. | |
138 | */ | |
139 | static void | |
140 | task_free(taskq_t *tq, taskq_ent_t *t) | |
141 | { | |
142 | ASSERT(tq); | |
143 | ASSERT(t); | |
144 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
145 | ASSERT(list_empty(&t->tqent_list)); | |
146 | ASSERT(!timer_pending(&t->tqent_timer)); | |
147 | ||
148 | kmem_free(t, sizeof(taskq_ent_t)); | |
149 | tq->tq_nalloc--; | |
150 | } | |
151 | ||
152 | /* | |
153 | * NOTE: Must be called with tq->tq_lock held, either destroys the | |
154 | * taskq_ent_t if too many exist or moves it to the free list for later use. | |
155 | */ | |
156 | static void | |
157 | task_done(taskq_t *tq, taskq_ent_t *t) | |
158 | { | |
159 | ASSERT(tq); | |
160 | ASSERT(t); | |
161 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
162 | ||
163 | /* Wake tasks blocked in taskq_wait_id() */ | |
164 | wake_up_all(&t->tqent_waitq); | |
165 | ||
166 | list_del_init(&t->tqent_list); | |
167 | ||
168 | if (tq->tq_nalloc <= tq->tq_minalloc) { | |
169 | t->tqent_id = 0; | |
170 | t->tqent_func = NULL; | |
171 | t->tqent_arg = NULL; | |
172 | t->tqent_flags = 0; | |
173 | ||
174 | list_add_tail(&t->tqent_list, &tq->tq_free_list); | |
175 | } else { | |
176 | task_free(tq, t); | |
177 | } | |
178 | } | |
179 | ||
180 | /* | |
181 | * When a delayed task timer expires remove it from the delay list and | |
182 | * add it to the priority list in order for immediate processing. | |
183 | */ | |
184 | static void | |
185 | task_expire(unsigned long data) | |
186 | { | |
187 | taskq_ent_t *w, *t = (taskq_ent_t *)data; | |
188 | taskq_t *tq = t->tqent_taskq; | |
189 | struct list_head *l; | |
190 | ||
191 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
192 | ||
193 | if (t->tqent_flags & TQENT_FLAG_CANCEL) { | |
194 | ASSERT(list_empty(&t->tqent_list)); | |
195 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
196 | return; | |
197 | } | |
198 | ||
199 | /* | |
200 | * The priority list must be maintained in strict task id order | |
201 | * from lowest to highest for lowest_id to be easily calculable. | |
202 | */ | |
203 | list_del(&t->tqent_list); | |
204 | list_for_each_prev(l, &tq->tq_prio_list) { | |
205 | w = list_entry(l, taskq_ent_t, tqent_list); | |
206 | if (w->tqent_id < t->tqent_id) { | |
207 | list_add(&t->tqent_list, l); | |
208 | break; | |
209 | } | |
210 | } | |
211 | if (l == &tq->tq_prio_list) | |
212 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
213 | ||
214 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
215 | ||
216 | wake_up(&tq->tq_work_waitq); | |
217 | } | |
218 | ||
219 | /* | |
220 | * Returns the lowest incomplete taskqid_t. The taskqid_t may | |
221 | * be queued on the pending list, on the priority list, on the | |
222 | * delay list, or on the work list currently being handled, but | |
223 | * it is not 100% complete yet. | |
224 | */ | |
225 | static taskqid_t | |
226 | taskq_lowest_id(taskq_t *tq) | |
227 | { | |
228 | taskqid_t lowest_id = tq->tq_next_id; | |
229 | taskq_ent_t *t; | |
230 | taskq_thread_t *tqt; | |
231 | ||
232 | ASSERT(tq); | |
233 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
234 | ||
235 | if (!list_empty(&tq->tq_pend_list)) { | |
236 | t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list); | |
237 | lowest_id = MIN(lowest_id, t->tqent_id); | |
238 | } | |
239 | ||
240 | if (!list_empty(&tq->tq_prio_list)) { | |
241 | t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list); | |
242 | lowest_id = MIN(lowest_id, t->tqent_id); | |
243 | } | |
244 | ||
245 | if (!list_empty(&tq->tq_delay_list)) { | |
246 | t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list); | |
247 | lowest_id = MIN(lowest_id, t->tqent_id); | |
248 | } | |
249 | ||
250 | if (!list_empty(&tq->tq_active_list)) { | |
251 | tqt = list_entry(tq->tq_active_list.next, taskq_thread_t, | |
252 | tqt_active_list); | |
253 | ASSERT(tqt->tqt_id != 0); | |
254 | lowest_id = MIN(lowest_id, tqt->tqt_id); | |
255 | } | |
256 | ||
257 | return (lowest_id); | |
258 | } | |
259 | ||
260 | /* | |
261 | * Insert a task into a list keeping the list sorted by increasing taskqid. | |
262 | */ | |
263 | static void | |
264 | taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt) | |
265 | { | |
266 | taskq_thread_t *w; | |
267 | struct list_head *l; | |
268 | ||
269 | ASSERT(tq); | |
270 | ASSERT(tqt); | |
271 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
272 | ||
273 | list_for_each_prev(l, &tq->tq_active_list) { | |
274 | w = list_entry(l, taskq_thread_t, tqt_active_list); | |
275 | if (w->tqt_id < tqt->tqt_id) { | |
276 | list_add(&tqt->tqt_active_list, l); | |
277 | break; | |
278 | } | |
279 | } | |
280 | if (l == &tq->tq_active_list) | |
281 | list_add(&tqt->tqt_active_list, &tq->tq_active_list); | |
282 | } | |
283 | ||
284 | /* | |
285 | * Find and return a task from the given list if it exists. The list | |
286 | * must be in lowest to highest task id order. | |
287 | */ | |
288 | static taskq_ent_t * | |
289 | taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id) | |
290 | { | |
291 | struct list_head *l; | |
292 | taskq_ent_t *t; | |
293 | ||
294 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
295 | ||
296 | list_for_each(l, lh) { | |
297 | t = list_entry(l, taskq_ent_t, tqent_list); | |
298 | ||
299 | if (t->tqent_id == id) | |
300 | return (t); | |
301 | ||
302 | if (t->tqent_id > id) | |
303 | break; | |
304 | } | |
305 | ||
306 | return (NULL); | |
307 | } | |
308 | ||
309 | /* | |
310 | * Find an already dispatched task given the task id regardless of what | |
311 | * state it is in. If a task is still pending or executing it will be | |
312 | * returned and 'active' set appropriately. If the task has already | |
313 | * been run then NULL is returned. | |
314 | */ | |
315 | static taskq_ent_t * | |
316 | taskq_find(taskq_t *tq, taskqid_t id, int *active) | |
317 | { | |
318 | taskq_thread_t *tqt; | |
319 | struct list_head *l; | |
320 | taskq_ent_t *t; | |
321 | ||
322 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
323 | *active = 0; | |
324 | ||
325 | t = taskq_find_list(tq, &tq->tq_delay_list, id); | |
326 | if (t) | |
327 | return (t); | |
328 | ||
329 | t = taskq_find_list(tq, &tq->tq_prio_list, id); | |
330 | if (t) | |
331 | return (t); | |
332 | ||
333 | t = taskq_find_list(tq, &tq->tq_pend_list, id); | |
334 | if (t) | |
335 | return (t); | |
336 | ||
337 | list_for_each(l, &tq->tq_active_list) { | |
338 | tqt = list_entry(l, taskq_thread_t, tqt_active_list); | |
339 | if (tqt->tqt_id == id) { | |
340 | t = tqt->tqt_task; | |
341 | *active = 1; | |
342 | return (t); | |
343 | } | |
344 | } | |
345 | ||
346 | return (NULL); | |
347 | } | |
348 | ||
349 | /* | |
350 | * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and | |
351 | * taskq_wait() functions below. | |
352 | * | |
353 | * Taskq waiting is accomplished by tracking the lowest outstanding task | |
354 | * id and the next available task id. As tasks are dispatched they are | |
355 | * added to the tail of the pending, priority, or delay lists. As worker | |
356 | * threads become available the tasks are removed from the heads of these | |
357 | * lists and linked to the worker threads. This ensures the lists are | |
358 | * kept sorted by lowest to highest task id. | |
359 | * | |
360 | * Therefore the lowest outstanding task id can be quickly determined by | |
361 | * checking the head item from all of these lists. This value is stored | |
362 | * with the taskq as the lowest id. It only needs to be recalculated when | |
363 | * either the task with the current lowest id completes or is canceled. | |
364 | * | |
365 | * By blocking until the lowest task id exceeds the passed task id the | |
366 | * taskq_wait_outstanding() function can be easily implemented. Similarly, | |
367 | * by blocking until the lowest task id matches the next task id taskq_wait() | |
368 | * can be implemented. | |
369 | * | |
370 | * Callers should be aware that when there are multiple worked threads it | |
371 | * is possible for larger task ids to complete before smaller ones. Also | |
372 | * when the taskq contains delay tasks with small task ids callers may | |
373 | * block for a considerable length of time waiting for them to expire and | |
374 | * execute. | |
375 | */ | |
376 | static int | |
377 | taskq_wait_id_check(taskq_t *tq, taskqid_t id) | |
378 | { | |
379 | int active = 0; | |
380 | int rc; | |
381 | ||
382 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
383 | rc = (taskq_find(tq, id, &active) == NULL); | |
384 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
385 | ||
386 | return (rc); | |
387 | } | |
388 | ||
389 | /* | |
390 | * The taskq_wait_id() function blocks until the passed task id completes. | |
391 | * This does not guarantee that all lower task ids have completed. | |
392 | */ | |
393 | void | |
394 | taskq_wait_id(taskq_t *tq, taskqid_t id) | |
395 | { | |
396 | wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id)); | |
397 | } | |
398 | EXPORT_SYMBOL(taskq_wait_id); | |
399 | ||
400 | static int | |
401 | taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id) | |
402 | { | |
403 | int rc; | |
404 | ||
405 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
406 | rc = (id < tq->tq_lowest_id); | |
407 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
408 | ||
409 | return (rc); | |
410 | } | |
411 | ||
412 | /* | |
413 | * The taskq_wait_outstanding() function will block until all tasks with a | |
414 | * lower taskqid than the passed 'id' have been completed. Note that all | |
415 | * task id's are assigned monotonically at dispatch time. Zero may be | |
416 | * passed for the id to indicate all tasks dispatch up to this point, | |
417 | * but not after, should be waited for. | |
418 | */ | |
419 | void | |
420 | taskq_wait_outstanding(taskq_t *tq, taskqid_t id) | |
421 | { | |
422 | wait_event(tq->tq_wait_waitq, | |
423 | taskq_wait_outstanding_check(tq, id ? id : tq->tq_next_id - 1)); | |
424 | } | |
425 | EXPORT_SYMBOL(taskq_wait_outstanding); | |
426 | ||
427 | static int | |
428 | taskq_wait_check(taskq_t *tq) | |
429 | { | |
430 | int rc; | |
431 | ||
432 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
433 | rc = (tq->tq_lowest_id == tq->tq_next_id); | |
434 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
435 | ||
436 | return (rc); | |
437 | } | |
438 | ||
439 | /* | |
440 | * The taskq_wait() function will block until the taskq is empty. | |
441 | * This means that if a taskq re-dispatches work to itself taskq_wait() | |
442 | * callers will block indefinitely. | |
443 | */ | |
444 | void | |
445 | taskq_wait(taskq_t *tq) | |
446 | { | |
447 | wait_event(tq->tq_wait_waitq, taskq_wait_check(tq)); | |
448 | } | |
449 | EXPORT_SYMBOL(taskq_wait); | |
450 | ||
451 | /* | |
452 | * Cancel an already dispatched task given the task id. Still pending tasks | |
453 | * will be immediately canceled, and if the task is active the function will | |
454 | * block until it completes. Preallocated tasks which are canceled must be | |
455 | * freed by the caller. | |
456 | */ | |
457 | int | |
458 | taskq_cancel_id(taskq_t *tq, taskqid_t id) | |
459 | { | |
460 | taskq_ent_t *t; | |
461 | int active = 0; | |
462 | int rc = ENOENT; | |
463 | ||
464 | ASSERT(tq); | |
465 | ||
466 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
467 | t = taskq_find(tq, id, &active); | |
468 | if (t && !active) { | |
469 | list_del_init(&t->tqent_list); | |
470 | t->tqent_flags |= TQENT_FLAG_CANCEL; | |
471 | ||
472 | /* | |
473 | * When canceling the lowest outstanding task id we | |
474 | * must recalculate the new lowest outstanding id. | |
475 | */ | |
476 | if (tq->tq_lowest_id == t->tqent_id) { | |
477 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
478 | ASSERT3S(tq->tq_lowest_id, >, t->tqent_id); | |
479 | } | |
480 | ||
481 | /* | |
482 | * The task_expire() function takes the tq->tq_lock so drop | |
483 | * drop the lock before synchronously cancelling the timer. | |
484 | */ | |
485 | if (timer_pending(&t->tqent_timer)) { | |
486 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
487 | del_timer_sync(&t->tqent_timer); | |
488 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
489 | } | |
490 | ||
491 | if (!(t->tqent_flags & TQENT_FLAG_PREALLOC)) | |
492 | task_done(tq, t); | |
493 | ||
494 | rc = 0; | |
495 | } | |
496 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
497 | ||
498 | if (active) { | |
499 | taskq_wait_id(tq, id); | |
500 | rc = EBUSY; | |
501 | } | |
502 | ||
503 | return (rc); | |
504 | } | |
505 | EXPORT_SYMBOL(taskq_cancel_id); | |
506 | ||
507 | static int taskq_thread_spawn(taskq_t *tq); | |
508 | ||
509 | taskqid_t | |
510 | taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags) | |
511 | { | |
512 | taskq_ent_t *t; | |
513 | taskqid_t rc = 0; | |
514 | ||
515 | ASSERT(tq); | |
516 | ASSERT(func); | |
517 | ||
518 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
519 | ||
520 | /* Taskq being destroyed and all tasks drained */ | |
521 | if (!(tq->tq_flags & TASKQ_ACTIVE)) | |
522 | goto out; | |
523 | ||
524 | /* Do not queue the task unless there is idle thread for it */ | |
525 | ASSERT(tq->tq_nactive <= tq->tq_nthreads); | |
526 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) | |
527 | goto out; | |
528 | ||
529 | if ((t = task_alloc(tq, flags)) == NULL) | |
530 | goto out; | |
531 | ||
532 | spin_lock(&t->tqent_lock); | |
533 | ||
534 | /* Queue to the priority list instead of the pending list */ | |
535 | if (flags & TQ_FRONT) | |
536 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
537 | else | |
538 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
539 | ||
540 | t->tqent_id = rc = tq->tq_next_id; | |
541 | tq->tq_next_id++; | |
542 | t->tqent_func = func; | |
543 | t->tqent_arg = arg; | |
544 | t->tqent_taskq = tq; | |
545 | t->tqent_timer.data = 0; | |
546 | t->tqent_timer.function = NULL; | |
547 | t->tqent_timer.expires = 0; | |
548 | ||
549 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
550 | ||
551 | spin_unlock(&t->tqent_lock); | |
552 | ||
553 | wake_up(&tq->tq_work_waitq); | |
554 | out: | |
555 | /* Spawn additional taskq threads if required. */ | |
556 | if (tq->tq_nactive == tq->tq_nthreads) | |
557 | (void) taskq_thread_spawn(tq); | |
558 | ||
559 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
560 | return (rc); | |
561 | } | |
562 | EXPORT_SYMBOL(taskq_dispatch); | |
563 | ||
564 | taskqid_t | |
565 | taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, | |
566 | uint_t flags, clock_t expire_time) | |
567 | { | |
568 | taskqid_t rc = 0; | |
569 | taskq_ent_t *t; | |
570 | ||
571 | ASSERT(tq); | |
572 | ASSERT(func); | |
573 | ||
574 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
575 | ||
576 | /* Taskq being destroyed and all tasks drained */ | |
577 | if (!(tq->tq_flags & TASKQ_ACTIVE)) | |
578 | goto out; | |
579 | ||
580 | if ((t = task_alloc(tq, flags)) == NULL) | |
581 | goto out; | |
582 | ||
583 | spin_lock(&t->tqent_lock); | |
584 | ||
585 | /* Queue to the delay list for subsequent execution */ | |
586 | list_add_tail(&t->tqent_list, &tq->tq_delay_list); | |
587 | ||
588 | t->tqent_id = rc = tq->tq_next_id; | |
589 | tq->tq_next_id++; | |
590 | t->tqent_func = func; | |
591 | t->tqent_arg = arg; | |
592 | t->tqent_taskq = tq; | |
593 | t->tqent_timer.data = (unsigned long)t; | |
594 | t->tqent_timer.function = task_expire; | |
595 | t->tqent_timer.expires = (unsigned long)expire_time; | |
596 | add_timer(&t->tqent_timer); | |
597 | ||
598 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
599 | ||
600 | spin_unlock(&t->tqent_lock); | |
601 | out: | |
602 | /* Spawn additional taskq threads if required. */ | |
603 | if (tq->tq_nactive == tq->tq_nthreads) | |
604 | (void) taskq_thread_spawn(tq); | |
605 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
606 | return (rc); | |
607 | } | |
608 | EXPORT_SYMBOL(taskq_dispatch_delay); | |
609 | ||
610 | void | |
611 | taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags, | |
612 | taskq_ent_t *t) | |
613 | { | |
614 | ASSERT(tq); | |
615 | ASSERT(func); | |
616 | ||
617 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
618 | ||
619 | /* Taskq being destroyed and all tasks drained */ | |
620 | if (!(tq->tq_flags & TASKQ_ACTIVE)) { | |
621 | t->tqent_id = 0; | |
622 | goto out; | |
623 | } | |
624 | ||
625 | spin_lock(&t->tqent_lock); | |
626 | ||
627 | /* | |
628 | * Mark it as a prealloc'd task. This is important | |
629 | * to ensure that we don't free it later. | |
630 | */ | |
631 | t->tqent_flags |= TQENT_FLAG_PREALLOC; | |
632 | ||
633 | /* Queue to the priority list instead of the pending list */ | |
634 | if (flags & TQ_FRONT) | |
635 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
636 | else | |
637 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
638 | ||
639 | t->tqent_id = tq->tq_next_id; | |
640 | tq->tq_next_id++; | |
641 | t->tqent_func = func; | |
642 | t->tqent_arg = arg; | |
643 | t->tqent_taskq = tq; | |
644 | ||
645 | spin_unlock(&t->tqent_lock); | |
646 | ||
647 | wake_up(&tq->tq_work_waitq); | |
648 | out: | |
649 | /* Spawn additional taskq threads if required. */ | |
650 | if (tq->tq_nactive == tq->tq_nthreads) | |
651 | (void) taskq_thread_spawn(tq); | |
652 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
653 | } | |
654 | EXPORT_SYMBOL(taskq_dispatch_ent); | |
655 | ||
656 | int | |
657 | taskq_empty_ent(taskq_ent_t *t) | |
658 | { | |
659 | return list_empty(&t->tqent_list); | |
660 | } | |
661 | EXPORT_SYMBOL(taskq_empty_ent); | |
662 | ||
663 | void | |
664 | taskq_init_ent(taskq_ent_t *t) | |
665 | { | |
666 | spin_lock_init(&t->tqent_lock); | |
667 | init_waitqueue_head(&t->tqent_waitq); | |
668 | init_timer(&t->tqent_timer); | |
669 | INIT_LIST_HEAD(&t->tqent_list); | |
670 | t->tqent_id = 0; | |
671 | t->tqent_func = NULL; | |
672 | t->tqent_arg = NULL; | |
673 | t->tqent_flags = 0; | |
674 | t->tqent_taskq = NULL; | |
675 | } | |
676 | EXPORT_SYMBOL(taskq_init_ent); | |
677 | ||
678 | /* | |
679 | * Return the next pending task, preference is given to tasks on the | |
680 | * priority list which were dispatched with TQ_FRONT. | |
681 | */ | |
682 | static taskq_ent_t * | |
683 | taskq_next_ent(taskq_t *tq) | |
684 | { | |
685 | struct list_head *list; | |
686 | ||
687 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
688 | ||
689 | if (!list_empty(&tq->tq_prio_list)) | |
690 | list = &tq->tq_prio_list; | |
691 | else if (!list_empty(&tq->tq_pend_list)) | |
692 | list = &tq->tq_pend_list; | |
693 | else | |
694 | return (NULL); | |
695 | ||
696 | return (list_entry(list->next, taskq_ent_t, tqent_list)); | |
697 | } | |
698 | ||
699 | /* | |
700 | * Spawns a new thread for the specified taskq. | |
701 | */ | |
702 | static void | |
703 | taskq_thread_spawn_task(void *arg) | |
704 | { | |
705 | taskq_t *tq = (taskq_t *)arg; | |
706 | ||
707 | (void) taskq_thread_create(tq); | |
708 | ||
709 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
710 | tq->tq_nspawn--; | |
711 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
712 | } | |
713 | ||
714 | /* | |
715 | * Spawn addition threads for dynamic taskqs (TASKQ_DYNMAIC) the current | |
716 | * number of threads is insufficient to handle the pending tasks. These | |
717 | * new threads must be created by the dedicated dynamic_taskq to avoid | |
718 | * deadlocks between thread creation and memory reclaim. The system_taskq | |
719 | * which is also a dynamic taskq cannot be safely used for this. | |
720 | */ | |
721 | static int | |
722 | taskq_thread_spawn(taskq_t *tq) | |
723 | { | |
724 | int spawning = 0; | |
725 | ||
726 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
727 | return (0); | |
728 | ||
729 | if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) && | |
730 | (tq->tq_flags & TASKQ_ACTIVE)) { | |
731 | spawning = (++tq->tq_nspawn); | |
732 | taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task, | |
733 | tq, TQ_NOSLEEP); | |
734 | } | |
735 | ||
736 | return (spawning); | |
737 | } | |
738 | ||
739 | /* | |
740 | * Threads in a dynamic taskq should only exit once it has been completely | |
741 | * drained and no other threads are actively servicing tasks. This prevents | |
742 | * threads from being created and destroyed more than is required. | |
743 | * | |
744 | * The first thread is the thread list is treated as the primary thread. | |
745 | * There is nothing special about the primary thread but in order to avoid | |
746 | * all the taskq pids from changing we opt to make it long running. | |
747 | */ | |
748 | static int | |
749 | taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt) | |
750 | { | |
751 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
752 | ||
753 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
754 | return (0); | |
755 | ||
756 | if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t, | |
757 | tqt_thread_list) == tqt) | |
758 | return (0); | |
759 | ||
760 | return | |
761 | ((tq->tq_nspawn == 0) && /* No threads are being spawned */ | |
762 | (tq->tq_nactive == 0) && /* No threads are handling tasks */ | |
763 | (tq->tq_nthreads > 1) && /* More than 1 thread is running */ | |
764 | (!taskq_next_ent(tq)) && /* There are no pending tasks */ | |
765 | (spl_taskq_thread_dynamic));/* Dynamic taskqs are allowed */ | |
766 | } | |
767 | ||
768 | static int | |
769 | taskq_thread(void *args) | |
770 | { | |
771 | DECLARE_WAITQUEUE(wait, current); | |
772 | sigset_t blocked; | |
773 | taskq_thread_t *tqt = args; | |
774 | taskq_t *tq; | |
775 | taskq_ent_t *t; | |
776 | int seq_tasks = 0; | |
777 | ||
778 | ASSERT(tqt); | |
779 | tq = tqt->tqt_tq; | |
780 | current->flags |= PF_NOFREEZE; | |
781 | current->journal_info = tq; | |
782 | ||
783 | #if defined(PF_MEMALLOC_NOIO) | |
784 | (void) memalloc_noio_save(); | |
785 | #endif | |
786 | ||
787 | sigfillset(&blocked); | |
788 | sigprocmask(SIG_BLOCK, &blocked, NULL); | |
789 | flush_signals(current); | |
790 | ||
791 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
792 | ||
793 | /* Immediately exit if more threads than allowed were created. */ | |
794 | if (tq->tq_nthreads >= tq->tq_maxthreads) | |
795 | goto error; | |
796 | ||
797 | tq->tq_nthreads++; | |
798 | list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list); | |
799 | wake_up(&tq->tq_wait_waitq); | |
800 | set_current_state(TASK_INTERRUPTIBLE); | |
801 | ||
802 | while (!kthread_should_stop()) { | |
803 | ||
804 | if (list_empty(&tq->tq_pend_list) && | |
805 | list_empty(&tq->tq_prio_list)) { | |
806 | ||
807 | if (taskq_thread_should_stop(tq, tqt)) { | |
808 | wake_up_all(&tq->tq_wait_waitq); | |
809 | break; | |
810 | } | |
811 | ||
812 | add_wait_queue_exclusive(&tq->tq_work_waitq, &wait); | |
813 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
814 | ||
815 | schedule(); | |
816 | seq_tasks = 0; | |
817 | ||
818 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
819 | remove_wait_queue(&tq->tq_work_waitq, &wait); | |
820 | } else { | |
821 | __set_current_state(TASK_RUNNING); | |
822 | } | |
823 | ||
824 | if ((t = taskq_next_ent(tq)) != NULL) { | |
825 | list_del_init(&t->tqent_list); | |
826 | ||
827 | /* In order to support recursively dispatching a | |
828 | * preallocated taskq_ent_t, tqent_id must be | |
829 | * stored prior to executing tqent_func. */ | |
830 | tqt->tqt_id = t->tqent_id; | |
831 | tqt->tqt_task = t; | |
832 | ||
833 | /* We must store a copy of the flags prior to | |
834 | * servicing the task (servicing a prealloc'd task | |
835 | * returns the ownership of the tqent back to | |
836 | * the caller of taskq_dispatch). Thus, | |
837 | * tqent_flags _may_ change within the call. */ | |
838 | tqt->tqt_flags = t->tqent_flags; | |
839 | ||
840 | taskq_insert_in_order(tq, tqt); | |
841 | tq->tq_nactive++; | |
842 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
843 | ||
844 | /* Perform the requested task */ | |
845 | t->tqent_func(t->tqent_arg); | |
846 | ||
847 | ASSERT3P(tq, ==, current->journal_info); | |
848 | ||
849 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
850 | tq->tq_nactive--; | |
851 | list_del_init(&tqt->tqt_active_list); | |
852 | tqt->tqt_task = NULL; | |
853 | ||
854 | /* For prealloc'd tasks, we don't free anything. */ | |
855 | if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC)) | |
856 | task_done(tq, t); | |
857 | ||
858 | /* When the current lowest outstanding taskqid is | |
859 | * done calculate the new lowest outstanding id */ | |
860 | if (tq->tq_lowest_id == tqt->tqt_id) { | |
861 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
862 | ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id); | |
863 | } | |
864 | ||
865 | /* Spawn additional taskq threads if required. */ | |
866 | if ((++seq_tasks) > spl_taskq_thread_sequential && | |
867 | taskq_thread_spawn(tq)) | |
868 | seq_tasks = 0; | |
869 | ||
870 | tqt->tqt_id = 0; | |
871 | tqt->tqt_flags = 0; | |
872 | wake_up_all(&tq->tq_wait_waitq); | |
873 | } else { | |
874 | if (taskq_thread_should_stop(tq, tqt)) | |
875 | break; | |
876 | } | |
877 | ||
878 | set_current_state(TASK_INTERRUPTIBLE); | |
879 | ||
880 | } | |
881 | ||
882 | __set_current_state(TASK_RUNNING); | |
883 | tq->tq_nthreads--; | |
884 | list_del_init(&tqt->tqt_thread_list); | |
885 | error: | |
886 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
887 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
888 | ||
889 | return (0); | |
890 | } | |
891 | ||
892 | static taskq_thread_t * | |
893 | taskq_thread_create(taskq_t *tq) | |
894 | { | |
895 | static int last_used_cpu = 0; | |
896 | taskq_thread_t *tqt; | |
897 | ||
898 | tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE); | |
899 | INIT_LIST_HEAD(&tqt->tqt_thread_list); | |
900 | INIT_LIST_HEAD(&tqt->tqt_active_list); | |
901 | tqt->tqt_tq = tq; | |
902 | tqt->tqt_id = 0; | |
903 | ||
904 | tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt, | |
905 | "%s", tq->tq_name); | |
906 | if (tqt->tqt_thread == NULL) { | |
907 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
908 | return (NULL); | |
909 | } | |
910 | ||
911 | if (spl_taskq_thread_bind) { | |
912 | last_used_cpu = (last_used_cpu + 1) % num_online_cpus(); | |
913 | kthread_bind(tqt->tqt_thread, last_used_cpu); | |
914 | } | |
915 | ||
916 | if (spl_taskq_thread_priority) | |
917 | set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri)); | |
918 | ||
919 | wake_up_process(tqt->tqt_thread); | |
920 | ||
921 | return (tqt); | |
922 | } | |
923 | ||
924 | taskq_t * | |
925 | taskq_create(const char *name, int nthreads, pri_t pri, | |
926 | int minalloc, int maxalloc, uint_t flags) | |
927 | { | |
928 | taskq_t *tq; | |
929 | taskq_thread_t *tqt; | |
930 | int count = 0, rc = 0, i; | |
931 | ||
932 | ASSERT(name != NULL); | |
933 | ASSERT(minalloc >= 0); | |
934 | ASSERT(maxalloc <= INT_MAX); | |
935 | ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */ | |
936 | ||
937 | /* Scale the number of threads using nthreads as a percentage */ | |
938 | if (flags & TASKQ_THREADS_CPU_PCT) { | |
939 | ASSERT(nthreads <= 100); | |
940 | ASSERT(nthreads >= 0); | |
941 | nthreads = MIN(nthreads, 100); | |
942 | nthreads = MAX(nthreads, 0); | |
943 | nthreads = MAX((num_online_cpus() * nthreads) / 100, 1); | |
944 | } | |
945 | ||
946 | tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE); | |
947 | if (tq == NULL) | |
948 | return (NULL); | |
949 | ||
950 | spin_lock_init(&tq->tq_lock); | |
951 | INIT_LIST_HEAD(&tq->tq_thread_list); | |
952 | INIT_LIST_HEAD(&tq->tq_active_list); | |
953 | tq->tq_name = strdup(name); | |
954 | tq->tq_nactive = 0; | |
955 | tq->tq_nthreads = 0; | |
956 | tq->tq_nspawn = 0; | |
957 | tq->tq_maxthreads = nthreads; | |
958 | tq->tq_pri = pri; | |
959 | tq->tq_minalloc = minalloc; | |
960 | tq->tq_maxalloc = maxalloc; | |
961 | tq->tq_nalloc = 0; | |
962 | tq->tq_flags = (flags | TASKQ_ACTIVE); | |
963 | tq->tq_next_id = 1; | |
964 | tq->tq_lowest_id = 1; | |
965 | INIT_LIST_HEAD(&tq->tq_free_list); | |
966 | INIT_LIST_HEAD(&tq->tq_pend_list); | |
967 | INIT_LIST_HEAD(&tq->tq_prio_list); | |
968 | INIT_LIST_HEAD(&tq->tq_delay_list); | |
969 | init_waitqueue_head(&tq->tq_work_waitq); | |
970 | init_waitqueue_head(&tq->tq_wait_waitq); | |
971 | ||
972 | if (flags & TASKQ_PREPOPULATE) { | |
973 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
974 | ||
975 | for (i = 0; i < minalloc; i++) | |
976 | task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW)); | |
977 | ||
978 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
979 | } | |
980 | ||
981 | if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) | |
982 | nthreads = 1; | |
983 | ||
984 | for (i = 0; i < nthreads; i++) { | |
985 | tqt = taskq_thread_create(tq); | |
986 | if (tqt == NULL) | |
987 | rc = 1; | |
988 | else | |
989 | count++; | |
990 | } | |
991 | ||
992 | /* Wait for all threads to be started before potential destroy */ | |
993 | wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count); | |
994 | ||
995 | if (rc) { | |
996 | taskq_destroy(tq); | |
997 | tq = NULL; | |
998 | } | |
999 | ||
1000 | return (tq); | |
1001 | } | |
1002 | EXPORT_SYMBOL(taskq_create); | |
1003 | ||
1004 | void | |
1005 | taskq_destroy(taskq_t *tq) | |
1006 | { | |
1007 | struct task_struct *thread; | |
1008 | taskq_thread_t *tqt; | |
1009 | taskq_ent_t *t; | |
1010 | ||
1011 | ASSERT(tq); | |
1012 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
1013 | tq->tq_flags &= ~TASKQ_ACTIVE; | |
1014 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
1015 | ||
1016 | /* | |
1017 | * When TASKQ_ACTIVE is clear new tasks may not be added nor may | |
1018 | * new worker threads be spawned for dynamic taskq. | |
1019 | */ | |
1020 | if (dynamic_taskq != NULL) | |
1021 | taskq_wait_outstanding(dynamic_taskq, 0); | |
1022 | ||
1023 | taskq_wait(tq); | |
1024 | ||
1025 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
1026 | ||
1027 | /* | |
1028 | * Signal each thread to exit and block until it does. Each thread | |
1029 | * is responsible for removing itself from the list and freeing its | |
1030 | * taskq_thread_t. This allows for idle threads to opt to remove | |
1031 | * themselves from the taskq. They can be recreated as needed. | |
1032 | */ | |
1033 | while (!list_empty(&tq->tq_thread_list)) { | |
1034 | tqt = list_entry(tq->tq_thread_list.next, | |
1035 | taskq_thread_t, tqt_thread_list); | |
1036 | thread = tqt->tqt_thread; | |
1037 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
1038 | ||
1039 | kthread_stop(thread); | |
1040 | ||
1041 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
1042 | } | |
1043 | ||
1044 | while (!list_empty(&tq->tq_free_list)) { | |
1045 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); | |
1046 | ||
1047 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
1048 | ||
1049 | list_del_init(&t->tqent_list); | |
1050 | task_free(tq, t); | |
1051 | } | |
1052 | ||
1053 | ASSERT0(tq->tq_nthreads); | |
1054 | ASSERT0(tq->tq_nalloc); | |
1055 | ASSERT0(tq->tq_nspawn); | |
1056 | ASSERT(list_empty(&tq->tq_thread_list)); | |
1057 | ASSERT(list_empty(&tq->tq_active_list)); | |
1058 | ASSERT(list_empty(&tq->tq_free_list)); | |
1059 | ASSERT(list_empty(&tq->tq_pend_list)); | |
1060 | ASSERT(list_empty(&tq->tq_prio_list)); | |
1061 | ASSERT(list_empty(&tq->tq_delay_list)); | |
1062 | ||
1063 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
1064 | ||
1065 | strfree(tq->tq_name); | |
1066 | kmem_free(tq, sizeof (taskq_t)); | |
1067 | } | |
1068 | EXPORT_SYMBOL(taskq_destroy); | |
1069 | ||
1070 | int | |
1071 | spl_taskq_init(void) | |
1072 | { | |
1073 | system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64), | |
1074 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); | |
1075 | if (system_taskq == NULL) | |
1076 | return (1); | |
1077 | ||
1078 | dynamic_taskq = taskq_create("spl_dynamic_taskq", 1, | |
1079 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE); | |
1080 | if (dynamic_taskq == NULL) { | |
1081 | taskq_destroy(system_taskq); | |
1082 | return (1); | |
1083 | } | |
1084 | ||
1085 | return (0); | |
1086 | } | |
1087 | ||
1088 | void | |
1089 | spl_taskq_fini(void) | |
1090 | { | |
1091 | taskq_destroy(dynamic_taskq); | |
1092 | dynamic_taskq = NULL; | |
1093 | ||
1094 | taskq_destroy(system_taskq); | |
1095 | system_taskq = NULL; | |
1096 | } |