]>
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 | ||
f6188ddd AX |
34 | |
35 | int spl_taskq_thread_dynamic = 0; | |
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 | ||
e9cb2b4f BB |
49 | /* Global system-wide dynamic task queue available for all consumers */ |
50 | taskq_t *system_taskq; | |
51 | EXPORT_SYMBOL(system_taskq); | |
52 | ||
f6188ddd AX |
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 | ||
9b51f218 BB |
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 | ||
82387586 BB |
69 | /* |
70 | * NOTE: Must be called with tq->tq_lock held, returns a list_t which | |
bcd68186 | 71 | * is not attached to the free, work, or pending taskq lists. |
f1ca4da6 | 72 | */ |
046a70c9 | 73 | static taskq_ent_t * |
bcd68186 | 74 | task_alloc(taskq_t *tq, uint_t flags) |
75 | { | |
472a34ca BB |
76 | taskq_ent_t *t; |
77 | int count = 0; | |
bcd68186 | 78 | |
472a34ca BB |
79 | ASSERT(tq); |
80 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
bcd68186 | 81 | retry: |
472a34ca BB |
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)); | |
d9acd930 BB |
87 | ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL)); |
88 | ASSERT(!timer_pending(&t->tqent_timer)); | |
472a34ca BB |
89 | |
90 | list_del_init(&t->tqent_list); | |
10946b02 | 91 | return (t); |
472a34ca BB |
92 | } |
93 | ||
94 | /* Free list is empty and memory allocations are prohibited */ | |
95 | if (flags & TQ_NOALLOC) | |
10946b02 | 96 | return (NULL); |
472a34ca BB |
97 | |
98 | /* Hit maximum taskq_ent_t pool size */ | |
99 | if (tq->tq_nalloc >= tq->tq_maxalloc) { | |
100 | if (flags & TQ_NOSLEEP) | |
10946b02 | 101 | return (NULL); |
472a34ca BB |
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); | |
10946b02 AX |
117 | if (count < 100) { |
118 | count++; | |
119 | goto retry; | |
120 | } | |
472a34ca BB |
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 | ||
10946b02 | 132 | return (t); |
bcd68186 | 133 | } |
134 | ||
82387586 | 135 | /* |
046a70c9 | 136 | * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t |
bcd68186 | 137 | * to already be removed from the free, work, or pending taskq lists. |
138 | */ | |
139 | static void | |
046a70c9 | 140 | task_free(taskq_t *tq, taskq_ent_t *t) |
bcd68186 | 141 | { |
472a34ca BB |
142 | ASSERT(tq); |
143 | ASSERT(t); | |
bcd68186 | 144 | ASSERT(spin_is_locked(&tq->tq_lock)); |
046a70c9 | 145 | ASSERT(list_empty(&t->tqent_list)); |
d9acd930 | 146 | ASSERT(!timer_pending(&t->tqent_timer)); |
bcd68186 | 147 | |
472a34ca BB |
148 | kmem_free(t, sizeof(taskq_ent_t)); |
149 | tq->tq_nalloc--; | |
bcd68186 | 150 | } |
151 | ||
82387586 BB |
152 | /* |
153 | * NOTE: Must be called with tq->tq_lock held, either destroys the | |
046a70c9 | 154 | * taskq_ent_t if too many exist or moves it to the free list for later use. |
bcd68186 | 155 | */ |
f1ca4da6 | 156 | static void |
046a70c9 | 157 | task_done(taskq_t *tq, taskq_ent_t *t) |
f1ca4da6 | 158 | { |
bcd68186 | 159 | ASSERT(tq); |
160 | ASSERT(t); | |
161 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
162 | ||
d9acd930 BB |
163 | /* Wake tasks blocked in taskq_wait_id() */ |
164 | wake_up_all(&t->tqent_waitq); | |
165 | ||
046a70c9 | 166 | list_del_init(&t->tqent_list); |
f1ca4da6 | 167 | |
472a34ca | 168 | if (tq->tq_nalloc <= tq->tq_minalloc) { |
046a70c9 PS |
169 | t->tqent_id = 0; |
170 | t->tqent_func = NULL; | |
171 | t->tqent_arg = NULL; | |
44217f7a | 172 | t->tqent_flags = 0; |
8f2503e0 | 173 | |
472a34ca | 174 | list_add_tail(&t->tqent_list, &tq->tq_free_list); |
bcd68186 | 175 | } else { |
176 | task_free(tq, t); | |
177 | } | |
f1ca4da6 | 178 | } |
179 | ||
82387586 | 180 | /* |
d9acd930 BB |
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. | |
bcd68186 | 183 | */ |
d9acd930 BB |
184 | static void |
185 | task_expire(unsigned long data) | |
bcd68186 | 186 | { |
d9acd930 BB |
187 | taskq_ent_t *w, *t = (taskq_ent_t *)data; |
188 | taskq_t *tq = t->tqent_taskq; | |
189 | struct list_head *l; | |
7257ec41 BB |
190 | |
191 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
d9acd930 BB |
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 | ||
7257ec41 BB |
214 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
215 | ||
d9acd930 BB |
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; | |
d9acd930 BB |
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 | ||
10946b02 | 257 | return (lowest_id); |
d9acd930 BB |
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 | ||
d9acd930 BB |
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); | |
d9acd930 BB |
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; | |
d9acd930 BB |
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) | |
10946b02 | 300 | return (t); |
d9acd930 BB |
301 | |
302 | if (t->tqent_id > id) | |
303 | break; | |
304 | } | |
305 | ||
10946b02 | 306 | return (NULL); |
bcd68186 | 307 | } |
308 | ||
d9acd930 BB |
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; | |
d9acd930 BB |
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) | |
10946b02 | 327 | return (t); |
d9acd930 BB |
328 | |
329 | t = taskq_find_list(tq, &tq->tq_prio_list, id); | |
330 | if (t) | |
10946b02 | 331 | return (t); |
d9acd930 BB |
332 | |
333 | t = taskq_find_list(tq, &tq->tq_pend_list, id); | |
334 | if (t) | |
10946b02 | 335 | return (t); |
d9acd930 BB |
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; | |
10946b02 | 342 | return (t); |
d9acd930 BB |
343 | } |
344 | } | |
345 | ||
10946b02 | 346 | return (NULL); |
d9acd930 BB |
347 | } |
348 | ||
f6188ddd AX |
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 | */ | |
80093b6f AX |
376 | static int |
377 | taskq_wait_id_check(taskq_t *tq, taskqid_t id) | |
f1ca4da6 | 378 | { |
d9acd930 | 379 | int active = 0; |
80093b6f | 380 | int rc; |
bcd68186 | 381 | |
d9acd930 | 382 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
80093b6f | 383 | rc = (taskq_find(tq, id, &active) == NULL); |
d9acd930 BB |
384 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
385 | ||
80093b6f AX |
386 | return (rc); |
387 | } | |
bcd68186 | 388 | |
80093b6f AX |
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)); | |
bcd68186 | 397 | } |
aed8671c | 398 | EXPORT_SYMBOL(taskq_wait_id); |
bcd68186 | 399 | |
d9acd930 | 400 | static int |
f6188ddd | 401 | taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id) |
d9acd930 BB |
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 | ||
10946b02 | 409 | return (rc); |
d9acd930 BB |
410 | } |
411 | ||
f6188ddd AX |
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 | */ | |
d9acd930 | 419 | void |
f6188ddd | 420 | taskq_wait_outstanding(taskq_t *tq, taskqid_t id) |
d9acd930 | 421 | { |
0f836a62 AX |
422 | id = id ? id : tq->tq_next_id - 1; |
423 | wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id)); | |
d9acd930 | 424 | } |
f6188ddd | 425 | EXPORT_SYMBOL(taskq_wait_outstanding); |
d9acd930 | 426 | |
f6188ddd AX |
427 | static int |
428 | taskq_wait_check(taskq_t *tq) | |
bcd68186 | 429 | { |
f6188ddd | 430 | int rc; |
bcd68186 | 431 | |
749045bb | 432 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
f6188ddd | 433 | rc = (tq->tq_lowest_id == tq->tq_next_id); |
749045bb | 434 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 435 | |
f6188ddd AX |
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)); | |
bcd68186 | 448 | } |
aed8671c | 449 | EXPORT_SYMBOL(taskq_wait); |
bcd68186 | 450 | |
f6188ddd AX |
451 | static int |
452 | taskq_member_impl(taskq_t *tq, void *t) | |
bcd68186 | 453 | { |
2c02b71b PS |
454 | struct list_head *l; |
455 | taskq_thread_t *tqt; | |
f6188ddd | 456 | int found = 0; |
bcd68186 | 457 | |
458 | ASSERT(tq); | |
472a34ca | 459 | ASSERT(t); |
f6188ddd | 460 | ASSERT(spin_is_locked(&tq->tq_lock)); |
bcd68186 | 461 | |
2c02b71b PS |
462 | list_for_each(l, &tq->tq_thread_list) { |
463 | tqt = list_entry(l, taskq_thread_t, tqt_thread_list); | |
f6188ddd AX |
464 | if (tqt->tqt_thread == (struct task_struct *)t) { |
465 | found = 1; | |
466 | break; | |
467 | } | |
2c02b71b | 468 | } |
f6188ddd AX |
469 | return (found); |
470 | } | |
bcd68186 | 471 | |
f6188ddd AX |
472 | int |
473 | taskq_member(taskq_t *tq, void *t) | |
474 | { | |
475 | int found; | |
476 | ||
477 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
478 | found = taskq_member_impl(tq, t); | |
479 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
480 | ||
481 | return (found); | |
bcd68186 | 482 | } |
aed8671c | 483 | EXPORT_SYMBOL(taskq_member); |
bcd68186 | 484 | |
d9acd930 BB |
485 | /* |
486 | * Cancel an already dispatched task given the task id. Still pending tasks | |
487 | * will be immediately canceled, and if the task is active the function will | |
488 | * block until it completes. Preallocated tasks which are canceled must be | |
489 | * freed by the caller. | |
490 | */ | |
491 | int | |
492 | taskq_cancel_id(taskq_t *tq, taskqid_t id) | |
493 | { | |
494 | taskq_ent_t *t; | |
495 | int active = 0; | |
496 | int rc = ENOENT; | |
d9acd930 BB |
497 | |
498 | ASSERT(tq); | |
499 | ||
500 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
501 | t = taskq_find(tq, id, &active); | |
502 | if (t && !active) { | |
503 | list_del_init(&t->tqent_list); | |
504 | t->tqent_flags |= TQENT_FLAG_CANCEL; | |
505 | ||
506 | /* | |
507 | * When canceling the lowest outstanding task id we | |
508 | * must recalculate the new lowest outstanding id. | |
509 | */ | |
510 | if (tq->tq_lowest_id == t->tqent_id) { | |
511 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
512 | ASSERT3S(tq->tq_lowest_id, >, t->tqent_id); | |
513 | } | |
514 | ||
515 | /* | |
516 | * The task_expire() function takes the tq->tq_lock so drop | |
517 | * drop the lock before synchronously cancelling the timer. | |
518 | */ | |
519 | if (timer_pending(&t->tqent_timer)) { | |
520 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
521 | del_timer_sync(&t->tqent_timer); | |
522 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
523 | } | |
524 | ||
525 | if (!(t->tqent_flags & TQENT_FLAG_PREALLOC)) | |
526 | task_done(tq, t); | |
527 | ||
528 | rc = 0; | |
529 | } | |
530 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
531 | ||
532 | if (active) { | |
533 | taskq_wait_id(tq, id); | |
534 | rc = EBUSY; | |
535 | } | |
536 | ||
10946b02 | 537 | return (rc); |
d9acd930 BB |
538 | } |
539 | EXPORT_SYMBOL(taskq_cancel_id); | |
540 | ||
879bbbc7 | 541 | static int taskq_thread_spawn(taskq_t *tq); |
f6188ddd | 542 | |
bcd68186 | 543 | taskqid_t |
aed8671c | 544 | taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags) |
bcd68186 | 545 | { |
472a34ca | 546 | taskq_ent_t *t; |
bcd68186 | 547 | taskqid_t rc = 0; |
f1ca4da6 | 548 | |
472a34ca BB |
549 | ASSERT(tq); |
550 | ASSERT(func); | |
d05ec4b4 | 551 | |
472a34ca | 552 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
f1ca4da6 | 553 | |
bcd68186 | 554 | /* Taskq being destroyed and all tasks drained */ |
f6188ddd | 555 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
10946b02 | 556 | goto out; |
f1ca4da6 | 557 | |
bcd68186 | 558 | /* Do not queue the task unless there is idle thread for it */ |
559 | ASSERT(tq->tq_nactive <= tq->tq_nthreads); | |
560 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) | |
10946b02 | 561 | goto out; |
bcd68186 | 562 | |
472a34ca | 563 | if ((t = task_alloc(tq, flags)) == NULL) |
10946b02 | 564 | goto out; |
f1ca4da6 | 565 | |
046a70c9 | 566 | spin_lock(&t->tqent_lock); |
f0d8bb26 NB |
567 | |
568 | /* Queue to the priority list instead of the pending list */ | |
569 | if (flags & TQ_FRONT) | |
046a70c9 | 570 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); |
f0d8bb26 | 571 | else |
046a70c9 | 572 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); |
f0d8bb26 | 573 | |
046a70c9 | 574 | t->tqent_id = rc = tq->tq_next_id; |
bcd68186 | 575 | tq->tq_next_id++; |
472a34ca BB |
576 | t->tqent_func = func; |
577 | t->tqent_arg = arg; | |
d9acd930 BB |
578 | t->tqent_taskq = tq; |
579 | t->tqent_timer.data = 0; | |
580 | t->tqent_timer.function = NULL; | |
581 | t->tqent_timer.expires = 0; | |
44217f7a PS |
582 | |
583 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
584 | ||
046a70c9 | 585 | spin_unlock(&t->tqent_lock); |
0bb43ca2 NB |
586 | |
587 | wake_up(&tq->tq_work_waitq); | |
bcd68186 | 588 | out: |
f6188ddd | 589 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
590 | if (tq->tq_nactive == tq->tq_nthreads) |
591 | (void) taskq_thread_spawn(tq); | |
f6188ddd | 592 | |
749045bb | 593 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
10946b02 | 594 | return (rc); |
f1ca4da6 | 595 | } |
aed8671c | 596 | EXPORT_SYMBOL(taskq_dispatch); |
44217f7a | 597 | |
d9acd930 BB |
598 | taskqid_t |
599 | taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, | |
600 | uint_t flags, clock_t expire_time) | |
601 | { | |
d9acd930 | 602 | taskqid_t rc = 0; |
10946b02 | 603 | taskq_ent_t *t; |
d9acd930 BB |
604 | |
605 | ASSERT(tq); | |
606 | ASSERT(func); | |
607 | ||
608 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
609 | ||
610 | /* Taskq being destroyed and all tasks drained */ | |
f6188ddd | 611 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
10946b02 | 612 | goto out; |
d9acd930 BB |
613 | |
614 | if ((t = task_alloc(tq, flags)) == NULL) | |
10946b02 | 615 | goto out; |
d9acd930 BB |
616 | |
617 | spin_lock(&t->tqent_lock); | |
618 | ||
619 | /* Queue to the delay list for subsequent execution */ | |
620 | list_add_tail(&t->tqent_list, &tq->tq_delay_list); | |
621 | ||
622 | t->tqent_id = rc = tq->tq_next_id; | |
623 | tq->tq_next_id++; | |
624 | t->tqent_func = func; | |
625 | t->tqent_arg = arg; | |
626 | t->tqent_taskq = tq; | |
627 | t->tqent_timer.data = (unsigned long)t; | |
628 | t->tqent_timer.function = task_expire; | |
629 | t->tqent_timer.expires = (unsigned long)expire_time; | |
630 | add_timer(&t->tqent_timer); | |
631 | ||
632 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
633 | ||
634 | spin_unlock(&t->tqent_lock); | |
635 | out: | |
f6188ddd | 636 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
637 | if (tq->tq_nactive == tq->tq_nthreads) |
638 | (void) taskq_thread_spawn(tq); | |
d9acd930 | 639 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
10946b02 | 640 | return (rc); |
d9acd930 BB |
641 | } |
642 | EXPORT_SYMBOL(taskq_dispatch_delay); | |
643 | ||
44217f7a | 644 | void |
aed8671c | 645 | taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags, |
44217f7a PS |
646 | taskq_ent_t *t) |
647 | { | |
44217f7a PS |
648 | ASSERT(tq); |
649 | ASSERT(func); | |
44217f7a PS |
650 | |
651 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
652 | ||
653 | /* Taskq being destroyed and all tasks drained */ | |
f6188ddd | 654 | if (!(tq->tq_flags & TASKQ_ACTIVE)) { |
44217f7a PS |
655 | t->tqent_id = 0; |
656 | goto out; | |
657 | } | |
658 | ||
659 | spin_lock(&t->tqent_lock); | |
660 | ||
661 | /* | |
662 | * Mark it as a prealloc'd task. This is important | |
663 | * to ensure that we don't free it later. | |
664 | */ | |
665 | t->tqent_flags |= TQENT_FLAG_PREALLOC; | |
666 | ||
667 | /* Queue to the priority list instead of the pending list */ | |
668 | if (flags & TQ_FRONT) | |
669 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
670 | else | |
671 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
672 | ||
673 | t->tqent_id = tq->tq_next_id; | |
674 | tq->tq_next_id++; | |
675 | t->tqent_func = func; | |
676 | t->tqent_arg = arg; | |
d9acd930 | 677 | t->tqent_taskq = tq; |
44217f7a PS |
678 | |
679 | spin_unlock(&t->tqent_lock); | |
680 | ||
681 | wake_up(&tq->tq_work_waitq); | |
682 | out: | |
f6188ddd | 683 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
684 | if (tq->tq_nactive == tq->tq_nthreads) |
685 | (void) taskq_thread_spawn(tq); | |
0bb43ca2 | 686 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
44217f7a | 687 | } |
aed8671c | 688 | EXPORT_SYMBOL(taskq_dispatch_ent); |
44217f7a PS |
689 | |
690 | int | |
aed8671c | 691 | taskq_empty_ent(taskq_ent_t *t) |
44217f7a PS |
692 | { |
693 | return list_empty(&t->tqent_list); | |
694 | } | |
aed8671c | 695 | EXPORT_SYMBOL(taskq_empty_ent); |
44217f7a PS |
696 | |
697 | void | |
aed8671c | 698 | taskq_init_ent(taskq_ent_t *t) |
44217f7a PS |
699 | { |
700 | spin_lock_init(&t->tqent_lock); | |
d9acd930 BB |
701 | init_waitqueue_head(&t->tqent_waitq); |
702 | init_timer(&t->tqent_timer); | |
44217f7a PS |
703 | INIT_LIST_HEAD(&t->tqent_list); |
704 | t->tqent_id = 0; | |
705 | t->tqent_func = NULL; | |
706 | t->tqent_arg = NULL; | |
707 | t->tqent_flags = 0; | |
d9acd930 | 708 | t->tqent_taskq = NULL; |
44217f7a | 709 | } |
aed8671c | 710 | EXPORT_SYMBOL(taskq_init_ent); |
44217f7a | 711 | |
f6188ddd AX |
712 | /* |
713 | * Return the next pending task, preference is given to tasks on the | |
714 | * priority list which were dispatched with TQ_FRONT. | |
715 | */ | |
716 | static taskq_ent_t * | |
717 | taskq_next_ent(taskq_t *tq) | |
718 | { | |
719 | struct list_head *list; | |
720 | ||
721 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
722 | ||
723 | if (!list_empty(&tq->tq_prio_list)) | |
724 | list = &tq->tq_prio_list; | |
725 | else if (!list_empty(&tq->tq_pend_list)) | |
726 | list = &tq->tq_pend_list; | |
727 | else | |
728 | return (NULL); | |
729 | ||
730 | return (list_entry(list->next, taskq_ent_t, tqent_list)); | |
731 | } | |
732 | ||
733 | /* | |
734 | * Spawns a new thread for the specified taskq. | |
735 | */ | |
736 | static void | |
737 | taskq_thread_spawn_task(void *arg) | |
738 | { | |
739 | taskq_t *tq = (taskq_t *)arg; | |
740 | ||
0f836a62 AX |
741 | if (taskq_thread_create(tq) == NULL) { |
742 | /* restore spawning count if failed */ | |
743 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
744 | tq->tq_nspawn--; | |
745 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
746 | } | |
f6188ddd AX |
747 | } |
748 | ||
749 | /* | |
750 | * Spawn addition threads for dynamic taskqs (TASKQ_DYNMAIC) the current | |
751 | * number of threads is insufficient to handle the pending tasks. These | |
752 | * new threads must be created by the dedicated dynamic_taskq to avoid | |
753 | * deadlocks between thread creation and memory reclaim. The system_taskq | |
754 | * which is also a dynamic taskq cannot be safely used for this. | |
755 | */ | |
756 | static int | |
879bbbc7 | 757 | taskq_thread_spawn(taskq_t *tq) |
f6188ddd AX |
758 | { |
759 | int spawning = 0; | |
760 | ||
761 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
762 | return (0); | |
763 | ||
879bbbc7 | 764 | if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) && |
f6188ddd AX |
765 | (tq->tq_flags & TASKQ_ACTIVE)) { |
766 | spawning = (++tq->tq_nspawn); | |
767 | taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task, | |
768 | tq, TQ_NOSLEEP); | |
769 | } | |
770 | ||
771 | return (spawning); | |
772 | } | |
773 | ||
774 | /* | |
775 | * Threads in a dynamic taskq should only exit once it has been completely | |
776 | * drained and no other threads are actively servicing tasks. This prevents | |
777 | * threads from being created and destroyed more than is required. | |
778 | * | |
779 | * The first thread is the thread list is treated as the primary thread. | |
780 | * There is nothing special about the primary thread but in order to avoid | |
781 | * all the taskq pids from changing we opt to make it long running. | |
782 | */ | |
783 | static int | |
784 | taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt) | |
785 | { | |
786 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
787 | ||
788 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
789 | return (0); | |
790 | ||
791 | if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t, | |
792 | tqt_thread_list) == tqt) | |
793 | return (0); | |
794 | ||
795 | return | |
796 | ((tq->tq_nspawn == 0) && /* No threads are being spawned */ | |
797 | (tq->tq_nactive == 0) && /* No threads are handling tasks */ | |
798 | (tq->tq_nthreads > 1) && /* More than 1 thread is running */ | |
799 | (!taskq_next_ent(tq)) && /* There are no pending tasks */ | |
800 | (spl_taskq_thread_dynamic));/* Dynamic taskqs are allowed */ | |
801 | } | |
802 | ||
bcd68186 | 803 | static int |
804 | taskq_thread(void *args) | |
805 | { | |
472a34ca BB |
806 | DECLARE_WAITQUEUE(wait, current); |
807 | sigset_t blocked; | |
2c02b71b | 808 | taskq_thread_t *tqt = args; |
472a34ca BB |
809 | taskq_t *tq; |
810 | taskq_ent_t *t; | |
f6188ddd | 811 | int seq_tasks = 0; |
bcd68186 | 812 | |
472a34ca | 813 | ASSERT(tqt); |
2c02b71b | 814 | tq = tqt->tqt_tq; |
472a34ca | 815 | current->flags |= PF_NOFREEZE; |
bcd68186 | 816 | |
879bbbc7 | 817 | (void) spl_fstrans_mark(); |
f6188ddd | 818 | |
472a34ca BB |
819 | sigfillset(&blocked); |
820 | sigprocmask(SIG_BLOCK, &blocked, NULL); | |
821 | flush_signals(current); | |
bcd68186 | 822 | |
472a34ca | 823 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
0f836a62 AX |
824 | /* |
825 | * If we are dynamically spawned, decrease spawning count. Note that | |
826 | * we could be created during taskq_create, in which case we shouldn't | |
827 | * do the decrement. But it's fine because taskq_create will reset | |
828 | * tq_nspawn later. | |
829 | */ | |
830 | if (tq->tq_flags & TASKQ_DYNAMIC) | |
831 | tq->tq_nspawn--; | |
f6188ddd AX |
832 | |
833 | /* Immediately exit if more threads than allowed were created. */ | |
834 | if (tq->tq_nthreads >= tq->tq_maxthreads) | |
835 | goto error; | |
836 | ||
472a34ca | 837 | tq->tq_nthreads++; |
f6188ddd | 838 | list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list); |
472a34ca BB |
839 | wake_up(&tq->tq_wait_waitq); |
840 | set_current_state(TASK_INTERRUPTIBLE); | |
bcd68186 | 841 | |
472a34ca | 842 | while (!kthread_should_stop()) { |
bcd68186 | 843 | |
f0d8bb26 NB |
844 | if (list_empty(&tq->tq_pend_list) && |
845 | list_empty(&tq->tq_prio_list)) { | |
f6188ddd AX |
846 | |
847 | if (taskq_thread_should_stop(tq, tqt)) { | |
848 | wake_up_all(&tq->tq_wait_waitq); | |
849 | break; | |
850 | } | |
851 | ||
3c6ed541 | 852 | add_wait_queue_exclusive(&tq->tq_work_waitq, &wait); |
749045bb | 853 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
f6188ddd | 854 | |
bcd68186 | 855 | schedule(); |
f6188ddd AX |
856 | seq_tasks = 0; |
857 | ||
749045bb | 858 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
3c6ed541 | 859 | remove_wait_queue(&tq->tq_work_waitq, &wait); |
bcd68186 | 860 | } else { |
861 | __set_current_state(TASK_RUNNING); | |
862 | } | |
863 | ||
f6188ddd | 864 | if ((t = taskq_next_ent(tq)) != NULL) { |
472a34ca | 865 | list_del_init(&t->tqent_list); |
8f2503e0 | 866 | |
44217f7a PS |
867 | /* In order to support recursively dispatching a |
868 | * preallocated taskq_ent_t, tqent_id must be | |
869 | * stored prior to executing tqent_func. */ | |
e7e5f78e | 870 | tqt->tqt_id = t->tqent_id; |
d9acd930 | 871 | tqt->tqt_task = t; |
8f2503e0 PS |
872 | |
873 | /* We must store a copy of the flags prior to | |
874 | * servicing the task (servicing a prealloc'd task | |
875 | * returns the ownership of the tqent back to | |
876 | * the caller of taskq_dispatch). Thus, | |
877 | * tqent_flags _may_ change within the call. */ | |
878 | tqt->tqt_flags = t->tqent_flags; | |
879 | ||
2c02b71b | 880 | taskq_insert_in_order(tq, tqt); |
472a34ca | 881 | tq->tq_nactive++; |
749045bb | 882 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 883 | |
884 | /* Perform the requested task */ | |
472a34ca | 885 | t->tqent_func(t->tqent_arg); |
bcd68186 | 886 | |
749045bb | 887 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
472a34ca | 888 | tq->tq_nactive--; |
2c02b71b | 889 | list_del_init(&tqt->tqt_active_list); |
d9acd930 | 890 | tqt->tqt_task = NULL; |
8f2503e0 PS |
891 | |
892 | /* For prealloc'd tasks, we don't free anything. */ | |
f6188ddd | 893 | if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC)) |
8f2503e0 | 894 | task_done(tq, t); |
bcd68186 | 895 | |
7257ec41 BB |
896 | /* When the current lowest outstanding taskqid is |
897 | * done calculate the new lowest outstanding id */ | |
e7e5f78e | 898 | if (tq->tq_lowest_id == tqt->tqt_id) { |
bcd68186 | 899 | tq->tq_lowest_id = taskq_lowest_id(tq); |
e7e5f78e | 900 | ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id); |
bcd68186 | 901 | } |
902 | ||
f6188ddd | 903 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
904 | if ((++seq_tasks) > spl_taskq_thread_sequential && |
905 | taskq_thread_spawn(tq)) | |
f6188ddd AX |
906 | seq_tasks = 0; |
907 | ||
e7e5f78e | 908 | tqt->tqt_id = 0; |
8f2503e0 | 909 | tqt->tqt_flags = 0; |
472a34ca | 910 | wake_up_all(&tq->tq_wait_waitq); |
f6188ddd AX |
911 | } else { |
912 | if (taskq_thread_should_stop(tq, tqt)) | |
913 | break; | |
bcd68186 | 914 | } |
915 | ||
916 | set_current_state(TASK_INTERRUPTIBLE); | |
917 | ||
472a34ca | 918 | } |
bcd68186 | 919 | |
920 | __set_current_state(TASK_RUNNING); | |
472a34ca | 921 | tq->tq_nthreads--; |
2c02b71b | 922 | list_del_init(&tqt->tqt_thread_list); |
f6188ddd AX |
923 | error: |
924 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
472a34ca | 925 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 926 | |
10946b02 | 927 | return (0); |
bcd68186 | 928 | } |
929 | ||
f6188ddd AX |
930 | static taskq_thread_t * |
931 | taskq_thread_create(taskq_t *tq) | |
932 | { | |
933 | static int last_used_cpu = 0; | |
934 | taskq_thread_t *tqt; | |
935 | ||
936 | tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE); | |
937 | INIT_LIST_HEAD(&tqt->tqt_thread_list); | |
938 | INIT_LIST_HEAD(&tqt->tqt_active_list); | |
939 | tqt->tqt_tq = tq; | |
940 | tqt->tqt_id = 0; | |
941 | ||
942 | tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt, | |
943 | "%s", tq->tq_name); | |
944 | if (tqt->tqt_thread == NULL) { | |
945 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
946 | return (NULL); | |
947 | } | |
948 | ||
949 | if (spl_taskq_thread_bind) { | |
950 | last_used_cpu = (last_used_cpu + 1) % num_online_cpus(); | |
951 | kthread_bind(tqt->tqt_thread, last_used_cpu); | |
952 | } | |
953 | ||
954 | if (spl_taskq_thread_priority) | |
955 | set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri)); | |
956 | ||
957 | wake_up_process(tqt->tqt_thread); | |
958 | ||
959 | return (tqt); | |
960 | } | |
961 | ||
f1ca4da6 | 962 | taskq_t * |
aed8671c | 963 | taskq_create(const char *name, int nthreads, pri_t pri, |
472a34ca | 964 | int minalloc, int maxalloc, uint_t flags) |
f1ca4da6 | 965 | { |
472a34ca | 966 | taskq_t *tq; |
2c02b71b | 967 | taskq_thread_t *tqt; |
f6188ddd | 968 | int count = 0, rc = 0, i; |
bcd68186 | 969 | |
472a34ca | 970 | ASSERT(name != NULL); |
472a34ca BB |
971 | ASSERT(minalloc >= 0); |
972 | ASSERT(maxalloc <= INT_MAX); | |
f6188ddd | 973 | ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */ |
bcd68186 | 974 | |
915404bd BB |
975 | /* Scale the number of threads using nthreads as a percentage */ |
976 | if (flags & TASKQ_THREADS_CPU_PCT) { | |
977 | ASSERT(nthreads <= 100); | |
978 | ASSERT(nthreads >= 0); | |
979 | nthreads = MIN(nthreads, 100); | |
980 | nthreads = MAX(nthreads, 0); | |
981 | nthreads = MAX((num_online_cpus() * nthreads) / 100, 1); | |
982 | } | |
983 | ||
f6188ddd | 984 | tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE); |
472a34ca | 985 | if (tq == NULL) |
10946b02 | 986 | return (NULL); |
bcd68186 | 987 | |
472a34ca | 988 | spin_lock_init(&tq->tq_lock); |
472a34ca BB |
989 | INIT_LIST_HEAD(&tq->tq_thread_list); |
990 | INIT_LIST_HEAD(&tq->tq_active_list); | |
f6188ddd AX |
991 | tq->tq_name = strdup(name); |
992 | tq->tq_nactive = 0; | |
993 | tq->tq_nthreads = 0; | |
994 | tq->tq_nspawn = 0; | |
995 | tq->tq_maxthreads = nthreads; | |
996 | tq->tq_pri = pri; | |
997 | tq->tq_minalloc = minalloc; | |
998 | tq->tq_maxalloc = maxalloc; | |
999 | tq->tq_nalloc = 0; | |
1000 | tq->tq_flags = (flags | TASKQ_ACTIVE); | |
1001 | tq->tq_next_id = 1; | |
1002 | tq->tq_lowest_id = 1; | |
472a34ca BB |
1003 | INIT_LIST_HEAD(&tq->tq_free_list); |
1004 | INIT_LIST_HEAD(&tq->tq_pend_list); | |
1005 | INIT_LIST_HEAD(&tq->tq_prio_list); | |
d9acd930 | 1006 | INIT_LIST_HEAD(&tq->tq_delay_list); |
472a34ca BB |
1007 | init_waitqueue_head(&tq->tq_work_waitq); |
1008 | init_waitqueue_head(&tq->tq_wait_waitq); | |
bcd68186 | 1009 | |
f6188ddd AX |
1010 | if (flags & TASKQ_PREPOPULATE) { |
1011 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
1012 | ||
472a34ca BB |
1013 | for (i = 0; i < minalloc; i++) |
1014 | task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW)); | |
6e605b6e | 1015 | |
f6188ddd AX |
1016 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
1017 | } | |
1018 | ||
1019 | if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) | |
1020 | nthreads = 1; | |
6e605b6e | 1021 | |
2c02b71b | 1022 | for (i = 0; i < nthreads; i++) { |
f6188ddd AX |
1023 | tqt = taskq_thread_create(tq); |
1024 | if (tqt == NULL) | |
2c02b71b | 1025 | rc = 1; |
f6188ddd AX |
1026 | else |
1027 | count++; | |
2c02b71b | 1028 | } |
bcd68186 | 1029 | |
472a34ca | 1030 | /* Wait for all threads to be started before potential destroy */ |
f6188ddd | 1031 | wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count); |
0f836a62 AX |
1032 | /* |
1033 | * taskq_thread might have touched nspawn, but we don't want them to | |
1034 | * because they're not dynamically spawned. So we reset it to 0 | |
1035 | */ | |
1036 | tq->tq_nspawn = 0; | |
bcd68186 | 1037 | |
472a34ca | 1038 | if (rc) { |
aed8671c | 1039 | taskq_destroy(tq); |
472a34ca BB |
1040 | tq = NULL; |
1041 | } | |
bcd68186 | 1042 | |
10946b02 | 1043 | return (tq); |
f1ca4da6 | 1044 | } |
aed8671c | 1045 | EXPORT_SYMBOL(taskq_create); |
b123971f | 1046 | |
1047 | void | |
aed8671c | 1048 | taskq_destroy(taskq_t *tq) |
b123971f | 1049 | { |
2c02b71b PS |
1050 | struct task_struct *thread; |
1051 | taskq_thread_t *tqt; | |
046a70c9 | 1052 | taskq_ent_t *t; |
b123971f | 1053 | |
bcd68186 | 1054 | ASSERT(tq); |
749045bb | 1055 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
f6188ddd | 1056 | tq->tq_flags &= ~TASKQ_ACTIVE; |
749045bb | 1057 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
bcd68186 | 1058 | |
f6188ddd AX |
1059 | /* |
1060 | * When TASKQ_ACTIVE is clear new tasks may not be added nor may | |
1061 | * new worker threads be spawned for dynamic taskq. | |
1062 | */ | |
1063 | if (dynamic_taskq != NULL) | |
1064 | taskq_wait_outstanding(dynamic_taskq, 0); | |
1065 | ||
aed8671c | 1066 | taskq_wait(tq); |
bcd68186 | 1067 | |
472a34ca | 1068 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
0f836a62 AX |
1069 | /* wait for spawning threads to insert themselves to the list */ |
1070 | while (tq->tq_nspawn) { | |
1071 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
1072 | schedule_timeout_interruptible(1); | |
1073 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); | |
1074 | } | |
bcd68186 | 1075 | |
2c02b71b PS |
1076 | /* |
1077 | * Signal each thread to exit and block until it does. Each thread | |
1078 | * is responsible for removing itself from the list and freeing its | |
1079 | * taskq_thread_t. This allows for idle threads to opt to remove | |
1080 | * themselves from the taskq. They can be recreated as needed. | |
1081 | */ | |
1082 | while (!list_empty(&tq->tq_thread_list)) { | |
1083 | tqt = list_entry(tq->tq_thread_list.next, | |
f6188ddd | 1084 | taskq_thread_t, tqt_thread_list); |
2c02b71b PS |
1085 | thread = tqt->tqt_thread; |
1086 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); | |
1087 | ||
1088 | kthread_stop(thread); | |
1089 | ||
472a34ca | 1090 | spin_lock_irqsave(&tq->tq_lock, tq->tq_lock_flags); |
2c02b71b PS |
1091 | } |
1092 | ||
472a34ca | 1093 | while (!list_empty(&tq->tq_free_list)) { |
046a70c9 | 1094 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); |
44217f7a PS |
1095 | |
1096 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
1097 | ||
472a34ca BB |
1098 | list_del_init(&t->tqent_list); |
1099 | task_free(tq, t); | |
1100 | } | |
bcd68186 | 1101 | |
f6188ddd AX |
1102 | ASSERT0(tq->tq_nthreads); |
1103 | ASSERT0(tq->tq_nalloc); | |
1104 | ASSERT0(tq->tq_nspawn); | |
472a34ca BB |
1105 | ASSERT(list_empty(&tq->tq_thread_list)); |
1106 | ASSERT(list_empty(&tq->tq_active_list)); | |
1107 | ASSERT(list_empty(&tq->tq_free_list)); | |
1108 | ASSERT(list_empty(&tq->tq_pend_list)); | |
1109 | ASSERT(list_empty(&tq->tq_prio_list)); | |
d9acd930 | 1110 | ASSERT(list_empty(&tq->tq_delay_list)); |
bcd68186 | 1111 | |
472a34ca | 1112 | spin_unlock_irqrestore(&tq->tq_lock, tq->tq_lock_flags); |
2c02b71b | 1113 | |
f6188ddd AX |
1114 | strfree(tq->tq_name); |
1115 | kmem_free(tq, sizeof (taskq_t)); | |
b123971f | 1116 | } |
aed8671c | 1117 | EXPORT_SYMBOL(taskq_destroy); |
e9cb2b4f BB |
1118 | |
1119 | int | |
1120 | spl_taskq_init(void) | |
1121 | { | |
f6188ddd AX |
1122 | system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64), |
1123 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); | |
e9cb2b4f | 1124 | if (system_taskq == NULL) |
10946b02 | 1125 | return (1); |
e9cb2b4f | 1126 | |
f6188ddd AX |
1127 | dynamic_taskq = taskq_create("spl_dynamic_taskq", 1, |
1128 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE); | |
1129 | if (dynamic_taskq == NULL) { | |
1130 | taskq_destroy(system_taskq); | |
1131 | return (1); | |
1132 | } | |
1133 | ||
10946b02 | 1134 | return (0); |
e9cb2b4f BB |
1135 | } |
1136 | ||
1137 | void | |
1138 | spl_taskq_fini(void) | |
1139 | { | |
f6188ddd AX |
1140 | taskq_destroy(dynamic_taskq); |
1141 | dynamic_taskq = NULL; | |
1142 | ||
e9cb2b4f | 1143 | taskq_destroy(system_taskq); |
f6188ddd | 1144 | system_taskq = NULL; |
e9cb2b4f | 1145 | } |