]>
Commit | Line | Data |
---|---|---|
716154c5 BB |
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>. | |
715f6251 | 6 | * UCRL-CODE-235197 |
7 | * | |
716154c5 | 8 | * This file is part of the SPL, Solaris Porting Layer. |
3d6af2dd | 9 | * For details, see <http://zfsonlinux.org/>. |
715f6251 | 10 | * |
716154c5 BB |
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 | |
715f6251 | 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 | |
716154c5 BB |
22 | * with the SPL. If not, see <http://www.gnu.org/licenses/>. |
23 | ***************************************************************************** | |
24 | * Solaris Porting Layer (SPL) Task Queue Implementation. | |
25 | \*****************************************************************************/ | |
715f6251 | 26 | |
f4b37741 | 27 | #include <sys/taskq.h> |
3d061e9d | 28 | #include <sys/kmem.h> |
937879f1 | 29 | |
9e4fb5c2 LG |
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 | ||
e9cb2b4f BB |
34 | /* Global system-wide dynamic task queue available for all consumers */ |
35 | taskq_t *system_taskq; | |
36 | EXPORT_SYMBOL(system_taskq); | |
37 | ||
9b51f218 BB |
38 | static int |
39 | task_km_flags(uint_t flags) | |
40 | { | |
41 | if (flags & TQ_NOSLEEP) | |
42 | return KM_NOSLEEP; | |
43 | ||
44 | if (flags & TQ_PUSHPAGE) | |
45 | return KM_PUSHPAGE; | |
46 | ||
47 | return KM_SLEEP; | |
48 | } | |
49 | ||
82387586 BB |
50 | /* |
51 | * NOTE: Must be called with tq->tq_lock held, returns a list_t which | |
bcd68186 | 52 | * is not attached to the free, work, or pending taskq lists. |
f1ca4da6 | 53 | */ |
046a70c9 | 54 | static taskq_ent_t * |
bcd68186 | 55 | task_alloc(taskq_t *tq, uint_t flags) |
56 | { | |
472a34ca BB |
57 | taskq_ent_t *t; |
58 | int count = 0; | |
bcd68186 | 59 | |
472a34ca BB |
60 | ASSERT(tq); |
61 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
bcd68186 | 62 | retry: |
472a34ca BB |
63 | /* Acquire taskq_ent_t's from free list if available */ |
64 | if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) { | |
65 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); | |
66 | ||
67 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
d9acd930 BB |
68 | ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL)); |
69 | ASSERT(!timer_pending(&t->tqent_timer)); | |
472a34ca BB |
70 | |
71 | list_del_init(&t->tqent_list); | |
10946b02 | 72 | return (t); |
472a34ca BB |
73 | } |
74 | ||
75 | /* Free list is empty and memory allocations are prohibited */ | |
76 | if (flags & TQ_NOALLOC) | |
10946b02 | 77 | return (NULL); |
472a34ca BB |
78 | |
79 | /* Hit maximum taskq_ent_t pool size */ | |
80 | if (tq->tq_nalloc >= tq->tq_maxalloc) { | |
81 | if (flags & TQ_NOSLEEP) | |
10946b02 | 82 | return (NULL); |
472a34ca BB |
83 | |
84 | /* | |
85 | * Sleep periodically polling the free list for an available | |
86 | * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed | |
87 | * but we cannot block forever waiting for an taskq_ent_t to | |
88 | * show up in the free list, otherwise a deadlock can happen. | |
89 | * | |
90 | * Therefore, we need to allocate a new task even if the number | |
91 | * of allocated tasks is above tq->tq_maxalloc, but we still | |
92 | * end up delaying the task allocation by one second, thereby | |
93 | * throttling the task dispatch rate. | |
94 | */ | |
95 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
96 | schedule_timeout(HZ / 100); | |
97 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
10946b02 AX |
98 | if (count < 100) { |
99 | count++; | |
100 | goto retry; | |
101 | } | |
472a34ca BB |
102 | } |
103 | ||
104 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
105 | t = kmem_alloc(sizeof(taskq_ent_t), task_km_flags(flags)); | |
106 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
107 | ||
108 | if (t) { | |
109 | taskq_init_ent(t); | |
110 | tq->tq_nalloc++; | |
111 | } | |
112 | ||
10946b02 | 113 | return (t); |
bcd68186 | 114 | } |
115 | ||
82387586 | 116 | /* |
046a70c9 | 117 | * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t |
bcd68186 | 118 | * to already be removed from the free, work, or pending taskq lists. |
119 | */ | |
120 | static void | |
046a70c9 | 121 | task_free(taskq_t *tq, taskq_ent_t *t) |
bcd68186 | 122 | { |
472a34ca BB |
123 | ASSERT(tq); |
124 | ASSERT(t); | |
bcd68186 | 125 | ASSERT(spin_is_locked(&tq->tq_lock)); |
046a70c9 | 126 | ASSERT(list_empty(&t->tqent_list)); |
d9acd930 | 127 | ASSERT(!timer_pending(&t->tqent_timer)); |
bcd68186 | 128 | |
472a34ca BB |
129 | kmem_free(t, sizeof(taskq_ent_t)); |
130 | tq->tq_nalloc--; | |
bcd68186 | 131 | } |
132 | ||
82387586 BB |
133 | /* |
134 | * NOTE: Must be called with tq->tq_lock held, either destroys the | |
046a70c9 | 135 | * taskq_ent_t if too many exist or moves it to the free list for later use. |
bcd68186 | 136 | */ |
f1ca4da6 | 137 | static void |
046a70c9 | 138 | task_done(taskq_t *tq, taskq_ent_t *t) |
f1ca4da6 | 139 | { |
bcd68186 | 140 | ASSERT(tq); |
141 | ASSERT(t); | |
142 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
143 | ||
d9acd930 BB |
144 | /* Wake tasks blocked in taskq_wait_id() */ |
145 | wake_up_all(&t->tqent_waitq); | |
146 | ||
046a70c9 | 147 | list_del_init(&t->tqent_list); |
f1ca4da6 | 148 | |
472a34ca | 149 | if (tq->tq_nalloc <= tq->tq_minalloc) { |
046a70c9 PS |
150 | t->tqent_id = 0; |
151 | t->tqent_func = NULL; | |
152 | t->tqent_arg = NULL; | |
44217f7a | 153 | t->tqent_flags = 0; |
8f2503e0 | 154 | |
472a34ca | 155 | list_add_tail(&t->tqent_list, &tq->tq_free_list); |
bcd68186 | 156 | } else { |
157 | task_free(tq, t); | |
158 | } | |
f1ca4da6 | 159 | } |
160 | ||
82387586 | 161 | /* |
d9acd930 BB |
162 | * When a delayed task timer expires remove it from the delay list and |
163 | * add it to the priority list in order for immediate processing. | |
bcd68186 | 164 | */ |
d9acd930 BB |
165 | static void |
166 | task_expire(unsigned long data) | |
bcd68186 | 167 | { |
d9acd930 BB |
168 | taskq_ent_t *w, *t = (taskq_ent_t *)data; |
169 | taskq_t *tq = t->tqent_taskq; | |
170 | struct list_head *l; | |
7257ec41 BB |
171 | |
172 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
d9acd930 BB |
173 | |
174 | if (t->tqent_flags & TQENT_FLAG_CANCEL) { | |
175 | ASSERT(list_empty(&t->tqent_list)); | |
176 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
177 | return; | |
178 | } | |
179 | ||
180 | /* | |
181 | * The priority list must be maintained in strict task id order | |
182 | * from lowest to highest for lowest_id to be easily calculable. | |
183 | */ | |
184 | list_del(&t->tqent_list); | |
185 | list_for_each_prev(l, &tq->tq_prio_list) { | |
186 | w = list_entry(l, taskq_ent_t, tqent_list); | |
187 | if (w->tqent_id < t->tqent_id) { | |
188 | list_add(&t->tqent_list, l); | |
189 | break; | |
190 | } | |
191 | } | |
192 | if (l == &tq->tq_prio_list) | |
193 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
194 | ||
7257ec41 BB |
195 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
196 | ||
d9acd930 BB |
197 | wake_up(&tq->tq_work_waitq); |
198 | } | |
199 | ||
200 | /* | |
201 | * Returns the lowest incomplete taskqid_t. The taskqid_t may | |
202 | * be queued on the pending list, on the priority list, on the | |
203 | * delay list, or on the work list currently being handled, but | |
204 | * it is not 100% complete yet. | |
205 | */ | |
206 | static taskqid_t | |
207 | taskq_lowest_id(taskq_t *tq) | |
208 | { | |
209 | taskqid_t lowest_id = tq->tq_next_id; | |
210 | taskq_ent_t *t; | |
211 | taskq_thread_t *tqt; | |
d9acd930 BB |
212 | |
213 | ASSERT(tq); | |
214 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
215 | ||
216 | if (!list_empty(&tq->tq_pend_list)) { | |
217 | t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list); | |
218 | lowest_id = MIN(lowest_id, t->tqent_id); | |
219 | } | |
220 | ||
221 | if (!list_empty(&tq->tq_prio_list)) { | |
222 | t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list); | |
223 | lowest_id = MIN(lowest_id, t->tqent_id); | |
224 | } | |
225 | ||
226 | if (!list_empty(&tq->tq_delay_list)) { | |
227 | t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list); | |
228 | lowest_id = MIN(lowest_id, t->tqent_id); | |
229 | } | |
230 | ||
231 | if (!list_empty(&tq->tq_active_list)) { | |
232 | tqt = list_entry(tq->tq_active_list.next, taskq_thread_t, | |
233 | tqt_active_list); | |
234 | ASSERT(tqt->tqt_id != 0); | |
235 | lowest_id = MIN(lowest_id, tqt->tqt_id); | |
236 | } | |
237 | ||
10946b02 | 238 | return (lowest_id); |
d9acd930 BB |
239 | } |
240 | ||
241 | /* | |
242 | * Insert a task into a list keeping the list sorted by increasing taskqid. | |
243 | */ | |
244 | static void | |
245 | taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt) | |
246 | { | |
247 | taskq_thread_t *w; | |
248 | struct list_head *l; | |
249 | ||
d9acd930 BB |
250 | ASSERT(tq); |
251 | ASSERT(tqt); | |
252 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
253 | ||
254 | list_for_each_prev(l, &tq->tq_active_list) { | |
255 | w = list_entry(l, taskq_thread_t, tqt_active_list); | |
256 | if (w->tqt_id < tqt->tqt_id) { | |
257 | list_add(&tqt->tqt_active_list, l); | |
258 | break; | |
259 | } | |
260 | } | |
261 | if (l == &tq->tq_active_list) | |
262 | list_add(&tqt->tqt_active_list, &tq->tq_active_list); | |
d9acd930 BB |
263 | } |
264 | ||
265 | /* | |
266 | * Find and return a task from the given list if it exists. The list | |
267 | * must be in lowest to highest task id order. | |
268 | */ | |
269 | static taskq_ent_t * | |
270 | taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id) | |
271 | { | |
272 | struct list_head *l; | |
273 | taskq_ent_t *t; | |
d9acd930 BB |
274 | |
275 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
276 | ||
277 | list_for_each(l, lh) { | |
278 | t = list_entry(l, taskq_ent_t, tqent_list); | |
279 | ||
280 | if (t->tqent_id == id) | |
10946b02 | 281 | return (t); |
d9acd930 BB |
282 | |
283 | if (t->tqent_id > id) | |
284 | break; | |
285 | } | |
286 | ||
10946b02 | 287 | return (NULL); |
bcd68186 | 288 | } |
289 | ||
d9acd930 BB |
290 | /* |
291 | * Find an already dispatched task given the task id regardless of what | |
292 | * state it is in. If a task is still pending or executing it will be | |
293 | * returned and 'active' set appropriately. If the task has already | |
294 | * been run then NULL is returned. | |
295 | */ | |
296 | static taskq_ent_t * | |
297 | taskq_find(taskq_t *tq, taskqid_t id, int *active) | |
298 | { | |
299 | taskq_thread_t *tqt; | |
300 | struct list_head *l; | |
301 | taskq_ent_t *t; | |
d9acd930 BB |
302 | |
303 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
304 | *active = 0; | |
305 | ||
306 | t = taskq_find_list(tq, &tq->tq_delay_list, id); | |
307 | if (t) | |
10946b02 | 308 | return (t); |
d9acd930 BB |
309 | |
310 | t = taskq_find_list(tq, &tq->tq_prio_list, id); | |
311 | if (t) | |
10946b02 | 312 | return (t); |
d9acd930 BB |
313 | |
314 | t = taskq_find_list(tq, &tq->tq_pend_list, id); | |
315 | if (t) | |
10946b02 | 316 | return (t); |
d9acd930 BB |
317 | |
318 | list_for_each(l, &tq->tq_active_list) { | |
319 | tqt = list_entry(l, taskq_thread_t, tqt_active_list); | |
320 | if (tqt->tqt_id == id) { | |
321 | t = tqt->tqt_task; | |
322 | *active = 1; | |
10946b02 | 323 | return (t); |
d9acd930 BB |
324 | } |
325 | } | |
326 | ||
10946b02 | 327 | return (NULL); |
d9acd930 BB |
328 | } |
329 | ||
80093b6f AX |
330 | static int |
331 | taskq_wait_id_check(taskq_t *tq, taskqid_t id) | |
f1ca4da6 | 332 | { |
d9acd930 | 333 | int active = 0; |
80093b6f | 334 | int rc; |
bcd68186 | 335 | |
d9acd930 | 336 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
80093b6f | 337 | rc = (taskq_find(tq, id, &active) == NULL); |
d9acd930 BB |
338 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
339 | ||
80093b6f AX |
340 | return (rc); |
341 | } | |
bcd68186 | 342 | |
80093b6f AX |
343 | /* |
344 | * The taskq_wait_id() function blocks until the passed task id completes. | |
345 | * This does not guarantee that all lower task ids have completed. | |
346 | */ | |
347 | void | |
348 | taskq_wait_id(taskq_t *tq, taskqid_t id) | |
349 | { | |
350 | wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id)); | |
bcd68186 | 351 | } |
aed8671c | 352 | EXPORT_SYMBOL(taskq_wait_id); |
bcd68186 | 353 | |
d9acd930 BB |
354 | /* |
355 | * The taskq_wait() function will block until all previously submitted | |
356 | * tasks have been completed. A previously submitted task is defined as | |
357 | * a task with a lower task id than the current task queue id. Note that | |
358 | * all task id's are assigned monotonically at dispatch time. | |
359 | * | |
360 | * Waiting for all previous tasks to complete is accomplished by tracking | |
361 | * the lowest outstanding task id. As tasks are dispatched they are added | |
362 | * added to the tail of the pending, priority, or delay lists. And as | |
363 | * worker threads become available the tasks are removed from the heads | |
364 | * of these lists and linked to the worker threads. This ensures the | |
365 | * lists are kept in lowest to highest task id order. | |
366 | * | |
367 | * Therefore the lowest outstanding task id can be quickly determined by | |
368 | * checking the head item from all of these lists. This value is stored | |
369 | * with the task queue as the lowest id. It only needs to be recalculated | |
370 | * when either the task with the current lowest id completes or is canceled. | |
371 | * | |
372 | * By blocking until the lowest task id exceeds the current task id when | |
373 | * the function was called we ensure all previous tasks have completed. | |
374 | * | |
375 | * NOTE: When there are multiple worked threads it is possible for larger | |
376 | * task ids to complete before smaller ones. Conversely when the task | |
377 | * queue contains delay tasks with small task ids, you may block for a | |
378 | * considerable length of time waiting for them to expire and execute. | |
379 | */ | |
380 | static int | |
381 | taskq_wait_check(taskq_t *tq, taskqid_t id) | |
382 | { | |
383 | int rc; | |
384 | ||
385 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
386 | rc = (id < tq->tq_lowest_id); | |
387 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
388 | ||
10946b02 | 389 | return (rc); |
d9acd930 BB |
390 | } |
391 | ||
392 | void | |
393 | taskq_wait_all(taskq_t *tq, taskqid_t id) | |
394 | { | |
395 | wait_event(tq->tq_wait_waitq, taskq_wait_check(tq, id)); | |
396 | } | |
397 | EXPORT_SYMBOL(taskq_wait_all); | |
398 | ||
bcd68186 | 399 | void |
aed8671c | 400 | taskq_wait(taskq_t *tq) |
bcd68186 | 401 | { |
402 | taskqid_t id; | |
10946b02 | 403 | |
bcd68186 | 404 | ASSERT(tq); |
405 | ||
7257ec41 | 406 | /* Wait for the largest outstanding taskqid */ |
749045bb | 407 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
7257ec41 | 408 | id = tq->tq_next_id - 1; |
749045bb | 409 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 410 | |
d9acd930 | 411 | taskq_wait_all(tq, id); |
bcd68186 | 412 | } |
aed8671c | 413 | EXPORT_SYMBOL(taskq_wait); |
bcd68186 | 414 | |
415 | int | |
aed8671c | 416 | taskq_member(taskq_t *tq, void *t) |
bcd68186 | 417 | { |
2c02b71b PS |
418 | struct list_head *l; |
419 | taskq_thread_t *tqt; | |
bcd68186 | 420 | |
421 | ASSERT(tq); | |
472a34ca | 422 | ASSERT(t); |
bcd68186 | 423 | |
2c02b71b PS |
424 | list_for_each(l, &tq->tq_thread_list) { |
425 | tqt = list_entry(l, taskq_thread_t, tqt_thread_list); | |
426 | if (tqt->tqt_thread == (struct task_struct *)t) | |
10946b02 | 427 | return (1); |
2c02b71b | 428 | } |
bcd68186 | 429 | |
10946b02 | 430 | return (0); |
bcd68186 | 431 | } |
aed8671c | 432 | EXPORT_SYMBOL(taskq_member); |
bcd68186 | 433 | |
d9acd930 BB |
434 | /* |
435 | * Cancel an already dispatched task given the task id. Still pending tasks | |
436 | * will be immediately canceled, and if the task is active the function will | |
437 | * block until it completes. Preallocated tasks which are canceled must be | |
438 | * freed by the caller. | |
439 | */ | |
440 | int | |
441 | taskq_cancel_id(taskq_t *tq, taskqid_t id) | |
442 | { | |
443 | taskq_ent_t *t; | |
444 | int active = 0; | |
445 | int rc = ENOENT; | |
d9acd930 BB |
446 | |
447 | ASSERT(tq); | |
448 | ||
449 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
450 | t = taskq_find(tq, id, &active); | |
451 | if (t && !active) { | |
452 | list_del_init(&t->tqent_list); | |
453 | t->tqent_flags |= TQENT_FLAG_CANCEL; | |
454 | ||
455 | /* | |
456 | * When canceling the lowest outstanding task id we | |
457 | * must recalculate the new lowest outstanding id. | |
458 | */ | |
459 | if (tq->tq_lowest_id == t->tqent_id) { | |
460 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
461 | ASSERT3S(tq->tq_lowest_id, >, t->tqent_id); | |
462 | } | |
463 | ||
464 | /* | |
465 | * The task_expire() function takes the tq->tq_lock so drop | |
466 | * drop the lock before synchronously cancelling the timer. | |
467 | */ | |
468 | if (timer_pending(&t->tqent_timer)) { | |
469 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
470 | del_timer_sync(&t->tqent_timer); | |
471 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
472 | } | |
473 | ||
474 | if (!(t->tqent_flags & TQENT_FLAG_PREALLOC)) | |
475 | task_done(tq, t); | |
476 | ||
477 | rc = 0; | |
478 | } | |
479 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
480 | ||
481 | if (active) { | |
482 | taskq_wait_id(tq, id); | |
483 | rc = EBUSY; | |
484 | } | |
485 | ||
10946b02 | 486 | return (rc); |
d9acd930 BB |
487 | } |
488 | EXPORT_SYMBOL(taskq_cancel_id); | |
489 | ||
bcd68186 | 490 | taskqid_t |
aed8671c | 491 | taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags) |
bcd68186 | 492 | { |
472a34ca | 493 | taskq_ent_t *t; |
bcd68186 | 494 | taskqid_t rc = 0; |
f1ca4da6 | 495 | |
472a34ca BB |
496 | ASSERT(tq); |
497 | ASSERT(func); | |
d05ec4b4 | 498 | |
472a34ca | 499 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
f1ca4da6 | 500 | |
bcd68186 | 501 | /* Taskq being destroyed and all tasks drained */ |
502 | if (!(tq->tq_flags & TQ_ACTIVE)) | |
10946b02 | 503 | goto out; |
f1ca4da6 | 504 | |
bcd68186 | 505 | /* Do not queue the task unless there is idle thread for it */ |
506 | ASSERT(tq->tq_nactive <= tq->tq_nthreads); | |
507 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) | |
10946b02 | 508 | goto out; |
bcd68186 | 509 | |
472a34ca | 510 | if ((t = task_alloc(tq, flags)) == NULL) |
10946b02 | 511 | goto out; |
f1ca4da6 | 512 | |
046a70c9 | 513 | spin_lock(&t->tqent_lock); |
f0d8bb26 NB |
514 | |
515 | /* Queue to the priority list instead of the pending list */ | |
516 | if (flags & TQ_FRONT) | |
046a70c9 | 517 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); |
f0d8bb26 | 518 | else |
046a70c9 | 519 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); |
f0d8bb26 | 520 | |
046a70c9 | 521 | t->tqent_id = rc = tq->tq_next_id; |
bcd68186 | 522 | tq->tq_next_id++; |
472a34ca BB |
523 | t->tqent_func = func; |
524 | t->tqent_arg = arg; | |
d9acd930 BB |
525 | t->tqent_taskq = tq; |
526 | t->tqent_timer.data = 0; | |
527 | t->tqent_timer.function = NULL; | |
528 | t->tqent_timer.expires = 0; | |
44217f7a PS |
529 | |
530 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
531 | ||
046a70c9 | 532 | spin_unlock(&t->tqent_lock); |
0bb43ca2 NB |
533 | |
534 | wake_up(&tq->tq_work_waitq); | |
bcd68186 | 535 | out: |
749045bb | 536 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
10946b02 | 537 | return (rc); |
f1ca4da6 | 538 | } |
aed8671c | 539 | EXPORT_SYMBOL(taskq_dispatch); |
44217f7a | 540 | |
d9acd930 BB |
541 | taskqid_t |
542 | taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, | |
543 | uint_t flags, clock_t expire_time) | |
544 | { | |
d9acd930 | 545 | taskqid_t rc = 0; |
10946b02 | 546 | taskq_ent_t *t; |
d9acd930 BB |
547 | |
548 | ASSERT(tq); | |
549 | ASSERT(func); | |
550 | ||
551 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
552 | ||
553 | /* Taskq being destroyed and all tasks drained */ | |
554 | if (!(tq->tq_flags & TQ_ACTIVE)) | |
10946b02 | 555 | goto out; |
d9acd930 BB |
556 | |
557 | if ((t = task_alloc(tq, flags)) == NULL) | |
10946b02 | 558 | goto out; |
d9acd930 BB |
559 | |
560 | spin_lock(&t->tqent_lock); | |
561 | ||
562 | /* Queue to the delay list for subsequent execution */ | |
563 | list_add_tail(&t->tqent_list, &tq->tq_delay_list); | |
564 | ||
565 | t->tqent_id = rc = tq->tq_next_id; | |
566 | tq->tq_next_id++; | |
567 | t->tqent_func = func; | |
568 | t->tqent_arg = arg; | |
569 | t->tqent_taskq = tq; | |
570 | t->tqent_timer.data = (unsigned long)t; | |
571 | t->tqent_timer.function = task_expire; | |
572 | t->tqent_timer.expires = (unsigned long)expire_time; | |
573 | add_timer(&t->tqent_timer); | |
574 | ||
575 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
576 | ||
577 | spin_unlock(&t->tqent_lock); | |
578 | out: | |
579 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
10946b02 | 580 | return (rc); |
d9acd930 BB |
581 | } |
582 | EXPORT_SYMBOL(taskq_dispatch_delay); | |
583 | ||
44217f7a | 584 | void |
aed8671c | 585 | taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags, |
44217f7a PS |
586 | taskq_ent_t *t) |
587 | { | |
44217f7a PS |
588 | ASSERT(tq); |
589 | ASSERT(func); | |
590 | ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC)); | |
591 | ||
592 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
593 | ||
594 | /* Taskq being destroyed and all tasks drained */ | |
595 | if (!(tq->tq_flags & TQ_ACTIVE)) { | |
596 | t->tqent_id = 0; | |
597 | goto out; | |
598 | } | |
599 | ||
600 | spin_lock(&t->tqent_lock); | |
601 | ||
602 | /* | |
603 | * Mark it as a prealloc'd task. This is important | |
604 | * to ensure that we don't free it later. | |
605 | */ | |
606 | t->tqent_flags |= TQENT_FLAG_PREALLOC; | |
607 | ||
608 | /* Queue to the priority list instead of the pending list */ | |
609 | if (flags & TQ_FRONT) | |
610 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
611 | else | |
612 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
613 | ||
614 | t->tqent_id = tq->tq_next_id; | |
615 | tq->tq_next_id++; | |
616 | t->tqent_func = func; | |
617 | t->tqent_arg = arg; | |
d9acd930 | 618 | t->tqent_taskq = tq; |
44217f7a PS |
619 | |
620 | spin_unlock(&t->tqent_lock); | |
621 | ||
622 | wake_up(&tq->tq_work_waitq); | |
623 | out: | |
0bb43ca2 | 624 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
44217f7a | 625 | } |
aed8671c | 626 | EXPORT_SYMBOL(taskq_dispatch_ent); |
44217f7a PS |
627 | |
628 | int | |
aed8671c | 629 | taskq_empty_ent(taskq_ent_t *t) |
44217f7a PS |
630 | { |
631 | return list_empty(&t->tqent_list); | |
632 | } | |
aed8671c | 633 | EXPORT_SYMBOL(taskq_empty_ent); |
44217f7a PS |
634 | |
635 | void | |
aed8671c | 636 | taskq_init_ent(taskq_ent_t *t) |
44217f7a PS |
637 | { |
638 | spin_lock_init(&t->tqent_lock); | |
d9acd930 BB |
639 | init_waitqueue_head(&t->tqent_waitq); |
640 | init_timer(&t->tqent_timer); | |
44217f7a PS |
641 | INIT_LIST_HEAD(&t->tqent_list); |
642 | t->tqent_id = 0; | |
643 | t->tqent_func = NULL; | |
644 | t->tqent_arg = NULL; | |
645 | t->tqent_flags = 0; | |
d9acd930 | 646 | t->tqent_taskq = NULL; |
44217f7a | 647 | } |
aed8671c | 648 | EXPORT_SYMBOL(taskq_init_ent); |
44217f7a | 649 | |
bcd68186 | 650 | static int |
651 | taskq_thread(void *args) | |
652 | { | |
472a34ca BB |
653 | DECLARE_WAITQUEUE(wait, current); |
654 | sigset_t blocked; | |
2c02b71b | 655 | taskq_thread_t *tqt = args; |
472a34ca BB |
656 | taskq_t *tq; |
657 | taskq_ent_t *t; | |
f0d8bb26 | 658 | struct list_head *pend_list; |
bcd68186 | 659 | |
472a34ca | 660 | ASSERT(tqt); |
2c02b71b | 661 | tq = tqt->tqt_tq; |
472a34ca | 662 | current->flags |= PF_NOFREEZE; |
bcd68186 | 663 | |
472a34ca BB |
664 | sigfillset(&blocked); |
665 | sigprocmask(SIG_BLOCK, &blocked, NULL); | |
666 | flush_signals(current); | |
bcd68186 | 667 | |
472a34ca BB |
668 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
669 | tq->tq_nthreads++; | |
670 | wake_up(&tq->tq_wait_waitq); | |
671 | set_current_state(TASK_INTERRUPTIBLE); | |
bcd68186 | 672 | |
472a34ca | 673 | while (!kthread_should_stop()) { |
bcd68186 | 674 | |
f0d8bb26 NB |
675 | if (list_empty(&tq->tq_pend_list) && |
676 | list_empty(&tq->tq_prio_list)) { | |
3c6ed541 | 677 | add_wait_queue_exclusive(&tq->tq_work_waitq, &wait); |
749045bb | 678 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 679 | schedule(); |
749045bb | 680 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
3c6ed541 | 681 | remove_wait_queue(&tq->tq_work_waitq, &wait); |
bcd68186 | 682 | } else { |
683 | __set_current_state(TASK_RUNNING); | |
684 | } | |
685 | ||
f0d8bb26 NB |
686 | |
687 | if (!list_empty(&tq->tq_prio_list)) | |
688 | pend_list = &tq->tq_prio_list; | |
689 | else if (!list_empty(&tq->tq_pend_list)) | |
690 | pend_list = &tq->tq_pend_list; | |
691 | else | |
692 | pend_list = NULL; | |
693 | ||
694 | if (pend_list) { | |
472a34ca BB |
695 | t = list_entry(pend_list->next,taskq_ent_t,tqent_list); |
696 | list_del_init(&t->tqent_list); | |
8f2503e0 | 697 | |
44217f7a PS |
698 | /* In order to support recursively dispatching a |
699 | * preallocated taskq_ent_t, tqent_id must be | |
700 | * stored prior to executing tqent_func. */ | |
e7e5f78e | 701 | tqt->tqt_id = t->tqent_id; |
d9acd930 | 702 | tqt->tqt_task = t; |
8f2503e0 PS |
703 | |
704 | /* We must store a copy of the flags prior to | |
705 | * servicing the task (servicing a prealloc'd task | |
706 | * returns the ownership of the tqent back to | |
707 | * the caller of taskq_dispatch). Thus, | |
708 | * tqent_flags _may_ change within the call. */ | |
709 | tqt->tqt_flags = t->tqent_flags; | |
710 | ||
2c02b71b | 711 | taskq_insert_in_order(tq, tqt); |
472a34ca | 712 | tq->tq_nactive++; |
749045bb | 713 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 714 | |
715 | /* Perform the requested task */ | |
472a34ca | 716 | t->tqent_func(t->tqent_arg); |
bcd68186 | 717 | |
749045bb | 718 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
472a34ca | 719 | tq->tq_nactive--; |
2c02b71b | 720 | list_del_init(&tqt->tqt_active_list); |
d9acd930 | 721 | tqt->tqt_task = NULL; |
8f2503e0 PS |
722 | |
723 | /* For prealloc'd tasks, we don't free anything. */ | |
724 | if ((tq->tq_flags & TASKQ_DYNAMIC) || | |
725 | !(tqt->tqt_flags & TQENT_FLAG_PREALLOC)) | |
726 | task_done(tq, t); | |
bcd68186 | 727 | |
7257ec41 BB |
728 | /* When the current lowest outstanding taskqid is |
729 | * done calculate the new lowest outstanding id */ | |
e7e5f78e | 730 | if (tq->tq_lowest_id == tqt->tqt_id) { |
bcd68186 | 731 | tq->tq_lowest_id = taskq_lowest_id(tq); |
e7e5f78e | 732 | ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id); |
bcd68186 | 733 | } |
734 | ||
e7e5f78e | 735 | tqt->tqt_id = 0; |
8f2503e0 | 736 | tqt->tqt_flags = 0; |
472a34ca | 737 | wake_up_all(&tq->tq_wait_waitq); |
bcd68186 | 738 | } |
739 | ||
740 | set_current_state(TASK_INTERRUPTIBLE); | |
741 | ||
472a34ca | 742 | } |
bcd68186 | 743 | |
744 | __set_current_state(TASK_RUNNING); | |
472a34ca | 745 | tq->tq_nthreads--; |
2c02b71b PS |
746 | list_del_init(&tqt->tqt_thread_list); |
747 | kmem_free(tqt, sizeof(taskq_thread_t)); | |
748 | ||
472a34ca | 749 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 750 | |
10946b02 | 751 | return (0); |
bcd68186 | 752 | } |
753 | ||
f1ca4da6 | 754 | taskq_t * |
aed8671c | 755 | taskq_create(const char *name, int nthreads, pri_t pri, |
472a34ca | 756 | int minalloc, int maxalloc, uint_t flags) |
f1ca4da6 | 757 | { |
9e4fb5c2 | 758 | static int last_used_cpu = 0; |
472a34ca | 759 | taskq_t *tq; |
2c02b71b | 760 | taskq_thread_t *tqt; |
472a34ca | 761 | int rc = 0, i, j = 0; |
bcd68186 | 762 | |
472a34ca BB |
763 | ASSERT(name != NULL); |
764 | ASSERT(pri <= maxclsyspri); | |
765 | ASSERT(minalloc >= 0); | |
766 | ASSERT(maxalloc <= INT_MAX); | |
767 | ASSERT(!(flags & (TASKQ_CPR_SAFE | TASKQ_DYNAMIC))); /* Unsupported */ | |
bcd68186 | 768 | |
915404bd BB |
769 | /* Scale the number of threads using nthreads as a percentage */ |
770 | if (flags & TASKQ_THREADS_CPU_PCT) { | |
771 | ASSERT(nthreads <= 100); | |
772 | ASSERT(nthreads >= 0); | |
773 | nthreads = MIN(nthreads, 100); | |
774 | nthreads = MAX(nthreads, 0); | |
775 | nthreads = MAX((num_online_cpus() * nthreads) / 100, 1); | |
776 | } | |
777 | ||
472a34ca BB |
778 | tq = kmem_alloc(sizeof(*tq), KM_PUSHPAGE); |
779 | if (tq == NULL) | |
10946b02 | 780 | return (NULL); |
bcd68186 | 781 | |
472a34ca BB |
782 | spin_lock_init(&tq->tq_lock); |
783 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
784 | INIT_LIST_HEAD(&tq->tq_thread_list); | |
785 | INIT_LIST_HEAD(&tq->tq_active_list); | |
786 | tq->tq_name = name; | |
787 | tq->tq_nactive = 0; | |
bcd68186 | 788 | tq->tq_nthreads = 0; |
472a34ca BB |
789 | tq->tq_pri = pri; |
790 | tq->tq_minalloc = minalloc; | |
791 | tq->tq_maxalloc = maxalloc; | |
bcd68186 | 792 | tq->tq_nalloc = 0; |
472a34ca | 793 | tq->tq_flags = (flags | TQ_ACTIVE); |
bcd68186 | 794 | tq->tq_next_id = 1; |
795 | tq->tq_lowest_id = 1; | |
472a34ca BB |
796 | INIT_LIST_HEAD(&tq->tq_free_list); |
797 | INIT_LIST_HEAD(&tq->tq_pend_list); | |
798 | INIT_LIST_HEAD(&tq->tq_prio_list); | |
d9acd930 | 799 | INIT_LIST_HEAD(&tq->tq_delay_list); |
472a34ca BB |
800 | init_waitqueue_head(&tq->tq_work_waitq); |
801 | init_waitqueue_head(&tq->tq_wait_waitq); | |
bcd68186 | 802 | |
472a34ca BB |
803 | if (flags & TASKQ_PREPOPULATE) |
804 | for (i = 0; i < minalloc; i++) | |
805 | task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW)); | |
6e605b6e | 806 | |
472a34ca | 807 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
6e605b6e | 808 | |
2c02b71b | 809 | for (i = 0; i < nthreads; i++) { |
9b51f218 | 810 | tqt = kmem_alloc(sizeof(*tqt), KM_PUSHPAGE); |
2c02b71b PS |
811 | INIT_LIST_HEAD(&tqt->tqt_thread_list); |
812 | INIT_LIST_HEAD(&tqt->tqt_active_list); | |
813 | tqt->tqt_tq = tq; | |
e7e5f78e | 814 | tqt->tqt_id = 0; |
2c02b71b | 815 | |
9e4fb5c2 | 816 | tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt, |
472a34ca | 817 | "%s/%d", name, i); |
2c02b71b PS |
818 | if (tqt->tqt_thread) { |
819 | list_add(&tqt->tqt_thread_list, &tq->tq_thread_list); | |
9e4fb5c2 LG |
820 | if (spl_taskq_thread_bind) { |
821 | last_used_cpu = (last_used_cpu + 1) % num_online_cpus(); | |
822 | kthread_bind(tqt->tqt_thread, last_used_cpu); | |
823 | } | |
2c02b71b PS |
824 | set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(pri)); |
825 | wake_up_process(tqt->tqt_thread); | |
bcd68186 | 826 | j++; |
2c02b71b PS |
827 | } else { |
828 | kmem_free(tqt, sizeof(taskq_thread_t)); | |
829 | rc = 1; | |
830 | } | |
831 | } | |
bcd68186 | 832 | |
472a34ca | 833 | /* Wait for all threads to be started before potential destroy */ |
bcd68186 | 834 | wait_event(tq->tq_wait_waitq, tq->tq_nthreads == j); |
835 | ||
472a34ca | 836 | if (rc) { |
aed8671c | 837 | taskq_destroy(tq); |
472a34ca BB |
838 | tq = NULL; |
839 | } | |
bcd68186 | 840 | |
10946b02 | 841 | return (tq); |
f1ca4da6 | 842 | } |
aed8671c | 843 | EXPORT_SYMBOL(taskq_create); |
b123971f | 844 | |
845 | void | |
aed8671c | 846 | taskq_destroy(taskq_t *tq) |
b123971f | 847 | { |
2c02b71b PS |
848 | struct task_struct *thread; |
849 | taskq_thread_t *tqt; | |
046a70c9 | 850 | taskq_ent_t *t; |
b123971f | 851 | |
bcd68186 | 852 | ASSERT(tq); |
749045bb | 853 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
472a34ca | 854 | tq->tq_flags &= ~TQ_ACTIVE; |
749045bb | 855 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 856 | |
857 | /* TQ_ACTIVE cleared prevents new tasks being added to pending */ | |
aed8671c | 858 | taskq_wait(tq); |
bcd68186 | 859 | |
472a34ca | 860 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 861 | |
2c02b71b PS |
862 | /* |
863 | * Signal each thread to exit and block until it does. Each thread | |
864 | * is responsible for removing itself from the list and freeing its | |
865 | * taskq_thread_t. This allows for idle threads to opt to remove | |
866 | * themselves from the taskq. They can be recreated as needed. | |
867 | */ | |
868 | while (!list_empty(&tq->tq_thread_list)) { | |
869 | tqt = list_entry(tq->tq_thread_list.next, | |
870 | taskq_thread_t, tqt_thread_list); | |
871 | thread = tqt->tqt_thread; | |
872 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
873 | ||
874 | kthread_stop(thread); | |
875 | ||
472a34ca | 876 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
2c02b71b PS |
877 | } |
878 | ||
472a34ca | 879 | while (!list_empty(&tq->tq_free_list)) { |
046a70c9 | 880 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); |
44217f7a PS |
881 | |
882 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
883 | ||
472a34ca BB |
884 | list_del_init(&t->tqent_list); |
885 | task_free(tq, t); | |
886 | } | |
bcd68186 | 887 | |
472a34ca BB |
888 | ASSERT(tq->tq_nthreads == 0); |
889 | ASSERT(tq->tq_nalloc == 0); | |
890 | ASSERT(list_empty(&tq->tq_thread_list)); | |
891 | ASSERT(list_empty(&tq->tq_active_list)); | |
892 | ASSERT(list_empty(&tq->tq_free_list)); | |
893 | ASSERT(list_empty(&tq->tq_pend_list)); | |
894 | ASSERT(list_empty(&tq->tq_prio_list)); | |
d9acd930 | 895 | ASSERT(list_empty(&tq->tq_delay_list)); |
bcd68186 | 896 | |
472a34ca | 897 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
2c02b71b | 898 | |
472a34ca | 899 | kmem_free(tq, sizeof(taskq_t)); |
b123971f | 900 | } |
aed8671c | 901 | EXPORT_SYMBOL(taskq_destroy); |
e9cb2b4f BB |
902 | |
903 | int | |
904 | spl_taskq_init(void) | |
905 | { | |
f220894e BB |
906 | /* Solaris creates a dynamic taskq of up to 64 threads, however in |
907 | * a Linux environment 1 thread per-core is usually about right */ | |
472a34ca | 908 | system_taskq = taskq_create("spl_system_taskq", num_online_cpus(), |
f220894e | 909 | minclsyspri, 4, 512, TASKQ_PREPOPULATE); |
e9cb2b4f | 910 | if (system_taskq == NULL) |
10946b02 | 911 | return (1); |
e9cb2b4f | 912 | |
10946b02 | 913 | return (0); |
e9cb2b4f BB |
914 | } |
915 | ||
916 | void | |
917 | spl_taskq_fini(void) | |
918 | { | |
e9cb2b4f | 919 | taskq_destroy(system_taskq); |
e9cb2b4f | 920 | } |