]>
Commit | Line | Data |
---|---|---|
2c4332cf | 1 | /* |
716154c5 BB |
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 BB |
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 BB |
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 | 22 | * with the SPL. If not, see <http://www.gnu.org/licenses/>. |
2c4332cf | 23 | * |
716154c5 | 24 | * Solaris Porting Layer (SPL) Task Queue Implementation. |
2c4332cf | 25 | */ |
715f6251 | 26 | |
f4b37741 | 27 | #include <sys/taskq.h> |
3d061e9d | 28 | #include <sys/kmem.h> |
16522ac2 | 29 | #include <sys/tsd.h> |
937879f1 | 30 | |
703371d8 AV |
31 | int spl_taskq_thread_bind = 0; |
32 | module_param(spl_taskq_thread_bind, int, 0644); | |
33 | MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default"); | |
34 | ||
f7a973d9 BB |
35 | |
36 | int spl_taskq_thread_dynamic = 1; | |
37 | module_param(spl_taskq_thread_dynamic, int, 0644); | |
38 | MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads"); | |
39 | ||
62aa81a5 BB |
40 | int spl_taskq_thread_priority = 1; |
41 | module_param(spl_taskq_thread_priority, int, 0644); | |
42 | MODULE_PARM_DESC(spl_taskq_thread_priority, | |
2c4332cf | 43 | "Allow non-default priority for taskq threads"); |
62aa81a5 | 44 | |
f7a973d9 BB |
45 | int spl_taskq_thread_sequential = 4; |
46 | module_param(spl_taskq_thread_sequential, int, 0644); | |
47 | MODULE_PARM_DESC(spl_taskq_thread_sequential, | |
2c4332cf | 48 | "Create new taskq threads after N sequential tasks"); |
f7a973d9 | 49 | |
e9cb2b4f BB |
50 | /* Global system-wide dynamic task queue available for all consumers */ |
51 | taskq_t *system_taskq; | |
52 | EXPORT_SYMBOL(system_taskq); | |
f200b836 CC |
53 | /* Global dynamic task queue for long delay */ |
54 | taskq_t *system_delay_taskq; | |
55 | EXPORT_SYMBOL(system_delay_taskq); | |
e9cb2b4f | 56 | |
f7a973d9 BB |
57 | /* Private dedicated taskq for creating new taskq threads on demand. */ |
58 | static taskq_t *dynamic_taskq; | |
59 | static taskq_thread_t *taskq_thread_create(taskq_t *); | |
60 | ||
200366f2 TC |
61 | /* List of all taskqs */ |
62 | LIST_HEAD(tq_list); | |
63 | DECLARE_RWSEM(tq_list_sem); | |
16522ac2 | 64 | static uint_t taskq_tsd; |
200366f2 | 65 | |
9b51f218 BB |
66 | static int |
67 | task_km_flags(uint_t flags) | |
68 | { | |
69 | if (flags & TQ_NOSLEEP) | |
2c4332cf | 70 | return (KM_NOSLEEP); |
9b51f218 BB |
71 | |
72 | if (flags & TQ_PUSHPAGE) | |
2c4332cf | 73 | return (KM_PUSHPAGE); |
9b51f218 | 74 | |
2c4332cf | 75 | return (KM_SLEEP); |
9b51f218 BB |
76 | } |
77 | ||
200366f2 TC |
78 | /* |
79 | * taskq_find_by_name - Find the largest instance number of a named taskq. | |
80 | */ | |
81 | static int | |
82 | taskq_find_by_name(const char *name) | |
83 | { | |
84 | struct list_head *tql; | |
85 | taskq_t *tq; | |
86 | ||
87 | list_for_each_prev(tql, &tq_list) { | |
88 | tq = list_entry(tql, taskq_t, tq_taskqs); | |
89 | if (strcmp(name, tq->tq_name) == 0) | |
90 | return tq->tq_instance; | |
91 | } | |
92 | return (-1); | |
93 | } | |
94 | ||
82387586 BB |
95 | /* |
96 | * NOTE: Must be called with tq->tq_lock held, returns a list_t which | |
bcd68186 | 97 | * is not attached to the free, work, or pending taskq lists. |
f1ca4da6 | 98 | */ |
046a70c9 | 99 | static taskq_ent_t * |
066b89e6 | 100 | task_alloc(taskq_t *tq, uint_t flags, unsigned long *irqflags) |
bcd68186 | 101 | { |
472a34ca BB |
102 | taskq_ent_t *t; |
103 | int count = 0; | |
bcd68186 | 104 | |
472a34ca BB |
105 | ASSERT(tq); |
106 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
bcd68186 | 107 | retry: |
472a34ca BB |
108 | /* Acquire taskq_ent_t's from free list if available */ |
109 | if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) { | |
110 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); | |
111 | ||
112 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
d9acd930 BB |
113 | ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL)); |
114 | ASSERT(!timer_pending(&t->tqent_timer)); | |
472a34ca BB |
115 | |
116 | list_del_init(&t->tqent_list); | |
8d9a23e8 | 117 | return (t); |
472a34ca BB |
118 | } |
119 | ||
120 | /* Free list is empty and memory allocations are prohibited */ | |
121 | if (flags & TQ_NOALLOC) | |
8d9a23e8 | 122 | return (NULL); |
472a34ca BB |
123 | |
124 | /* Hit maximum taskq_ent_t pool size */ | |
125 | if (tq->tq_nalloc >= tq->tq_maxalloc) { | |
126 | if (flags & TQ_NOSLEEP) | |
8d9a23e8 | 127 | return (NULL); |
472a34ca BB |
128 | |
129 | /* | |
130 | * Sleep periodically polling the free list for an available | |
131 | * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed | |
132 | * but we cannot block forever waiting for an taskq_ent_t to | |
133 | * show up in the free list, otherwise a deadlock can happen. | |
134 | * | |
135 | * Therefore, we need to allocate a new task even if the number | |
136 | * of allocated tasks is above tq->tq_maxalloc, but we still | |
137 | * end up delaying the task allocation by one second, thereby | |
138 | * throttling the task dispatch rate. | |
139 | */ | |
066b89e6 | 140 | spin_unlock_irqrestore(&tq->tq_lock, *irqflags); |
472a34ca | 141 | schedule_timeout(HZ / 100); |
066b89e6 | 142 | spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, |
326172d8 | 143 | tq->tq_lock_class); |
8d9a23e8 BB |
144 | if (count < 100) { |
145 | count++; | |
146 | goto retry; | |
147 | } | |
472a34ca BB |
148 | } |
149 | ||
066b89e6 | 150 | spin_unlock_irqrestore(&tq->tq_lock, *irqflags); |
2c4332cf | 151 | t = kmem_alloc(sizeof (taskq_ent_t), task_km_flags(flags)); |
066b89e6 | 152 | spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class); |
472a34ca BB |
153 | |
154 | if (t) { | |
155 | taskq_init_ent(t); | |
156 | tq->tq_nalloc++; | |
157 | } | |
158 | ||
8d9a23e8 | 159 | return (t); |
bcd68186 BB |
160 | } |
161 | ||
82387586 | 162 | /* |
046a70c9 | 163 | * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t |
bcd68186 BB |
164 | * to already be removed from the free, work, or pending taskq lists. |
165 | */ | |
166 | static void | |
046a70c9 | 167 | task_free(taskq_t *tq, taskq_ent_t *t) |
bcd68186 | 168 | { |
472a34ca BB |
169 | ASSERT(tq); |
170 | ASSERT(t); | |
bcd68186 | 171 | ASSERT(spin_is_locked(&tq->tq_lock)); |
046a70c9 | 172 | ASSERT(list_empty(&t->tqent_list)); |
d9acd930 | 173 | ASSERT(!timer_pending(&t->tqent_timer)); |
bcd68186 | 174 | |
2c4332cf | 175 | kmem_free(t, sizeof (taskq_ent_t)); |
472a34ca | 176 | tq->tq_nalloc--; |
bcd68186 BB |
177 | } |
178 | ||
82387586 BB |
179 | /* |
180 | * NOTE: Must be called with tq->tq_lock held, either destroys the | |
046a70c9 | 181 | * taskq_ent_t if too many exist or moves it to the free list for later use. |
bcd68186 | 182 | */ |
f1ca4da6 | 183 | static void |
046a70c9 | 184 | task_done(taskq_t *tq, taskq_ent_t *t) |
f1ca4da6 | 185 | { |
bcd68186 BB |
186 | ASSERT(tq); |
187 | ASSERT(t); | |
188 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
189 | ||
d9acd930 BB |
190 | /* Wake tasks blocked in taskq_wait_id() */ |
191 | wake_up_all(&t->tqent_waitq); | |
192 | ||
046a70c9 | 193 | list_del_init(&t->tqent_list); |
f1ca4da6 | 194 | |
472a34ca | 195 | if (tq->tq_nalloc <= tq->tq_minalloc) { |
cbba7146 | 196 | t->tqent_id = TASKQID_INVALID; |
046a70c9 PS |
197 | t->tqent_func = NULL; |
198 | t->tqent_arg = NULL; | |
44217f7a | 199 | t->tqent_flags = 0; |
8f2503e0 | 200 | |
472a34ca | 201 | list_add_tail(&t->tqent_list, &tq->tq_free_list); |
bcd68186 BB |
202 | } else { |
203 | task_free(tq, t); | |
204 | } | |
f1ca4da6 BB |
205 | } |
206 | ||
82387586 | 207 | /* |
d9acd930 BB |
208 | * When a delayed task timer expires remove it from the delay list and |
209 | * add it to the priority list in order for immediate processing. | |
bcd68186 | 210 | */ |
d9acd930 BB |
211 | static void |
212 | task_expire(unsigned long data) | |
bcd68186 | 213 | { |
d9acd930 BB |
214 | taskq_ent_t *w, *t = (taskq_ent_t *)data; |
215 | taskq_t *tq = t->tqent_taskq; | |
216 | struct list_head *l; | |
066b89e6 | 217 | unsigned long flags; |
7257ec41 | 218 | |
066b89e6 | 219 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
d9acd930 BB |
220 | |
221 | if (t->tqent_flags & TQENT_FLAG_CANCEL) { | |
222 | ASSERT(list_empty(&t->tqent_list)); | |
066b89e6 | 223 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 BB |
224 | return; |
225 | } | |
226 | ||
8f3b403a | 227 | t->tqent_birth = jiffies; |
d9acd930 BB |
228 | /* |
229 | * The priority list must be maintained in strict task id order | |
230 | * from lowest to highest for lowest_id to be easily calculable. | |
231 | */ | |
232 | list_del(&t->tqent_list); | |
233 | list_for_each_prev(l, &tq->tq_prio_list) { | |
234 | w = list_entry(l, taskq_ent_t, tqent_list); | |
235 | if (w->tqent_id < t->tqent_id) { | |
236 | list_add(&t->tqent_list, l); | |
237 | break; | |
238 | } | |
239 | } | |
240 | if (l == &tq->tq_prio_list) | |
241 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
242 | ||
066b89e6 | 243 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
7257ec41 | 244 | |
d9acd930 BB |
245 | wake_up(&tq->tq_work_waitq); |
246 | } | |
247 | ||
248 | /* | |
249 | * Returns the lowest incomplete taskqid_t. The taskqid_t may | |
250 | * be queued on the pending list, on the priority list, on the | |
251 | * delay list, or on the work list currently being handled, but | |
252 | * it is not 100% complete yet. | |
253 | */ | |
254 | static taskqid_t | |
255 | taskq_lowest_id(taskq_t *tq) | |
256 | { | |
257 | taskqid_t lowest_id = tq->tq_next_id; | |
258 | taskq_ent_t *t; | |
259 | taskq_thread_t *tqt; | |
d9acd930 BB |
260 | |
261 | ASSERT(tq); | |
262 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
263 | ||
264 | if (!list_empty(&tq->tq_pend_list)) { | |
265 | t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list); | |
266 | lowest_id = MIN(lowest_id, t->tqent_id); | |
267 | } | |
268 | ||
269 | if (!list_empty(&tq->tq_prio_list)) { | |
270 | t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list); | |
271 | lowest_id = MIN(lowest_id, t->tqent_id); | |
272 | } | |
273 | ||
274 | if (!list_empty(&tq->tq_delay_list)) { | |
275 | t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list); | |
276 | lowest_id = MIN(lowest_id, t->tqent_id); | |
277 | } | |
278 | ||
279 | if (!list_empty(&tq->tq_active_list)) { | |
280 | tqt = list_entry(tq->tq_active_list.next, taskq_thread_t, | |
281 | tqt_active_list); | |
cbba7146 | 282 | ASSERT(tqt->tqt_id != TASKQID_INVALID); |
d9acd930 BB |
283 | lowest_id = MIN(lowest_id, tqt->tqt_id); |
284 | } | |
285 | ||
8d9a23e8 | 286 | return (lowest_id); |
d9acd930 BB |
287 | } |
288 | ||
289 | /* | |
290 | * Insert a task into a list keeping the list sorted by increasing taskqid. | |
291 | */ | |
292 | static void | |
293 | taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt) | |
294 | { | |
295 | taskq_thread_t *w; | |
296 | struct list_head *l; | |
297 | ||
d9acd930 BB |
298 | ASSERT(tq); |
299 | ASSERT(tqt); | |
300 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
301 | ||
302 | list_for_each_prev(l, &tq->tq_active_list) { | |
303 | w = list_entry(l, taskq_thread_t, tqt_active_list); | |
304 | if (w->tqt_id < tqt->tqt_id) { | |
305 | list_add(&tqt->tqt_active_list, l); | |
306 | break; | |
307 | } | |
308 | } | |
309 | if (l == &tq->tq_active_list) | |
310 | list_add(&tqt->tqt_active_list, &tq->tq_active_list); | |
d9acd930 BB |
311 | } |
312 | ||
313 | /* | |
314 | * Find and return a task from the given list if it exists. The list | |
315 | * must be in lowest to highest task id order. | |
316 | */ | |
317 | static taskq_ent_t * | |
318 | taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id) | |
319 | { | |
320 | struct list_head *l; | |
321 | taskq_ent_t *t; | |
d9acd930 BB |
322 | |
323 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
324 | ||
325 | list_for_each(l, lh) { | |
326 | t = list_entry(l, taskq_ent_t, tqent_list); | |
327 | ||
328 | if (t->tqent_id == id) | |
8d9a23e8 | 329 | return (t); |
d9acd930 BB |
330 | |
331 | if (t->tqent_id > id) | |
332 | break; | |
333 | } | |
334 | ||
8d9a23e8 | 335 | return (NULL); |
bcd68186 BB |
336 | } |
337 | ||
d9acd930 BB |
338 | /* |
339 | * Find an already dispatched task given the task id regardless of what | |
cce83ba0 CC |
340 | * state it is in. If a task is still pending it will be returned. |
341 | * If a task is executing, then -EBUSY will be returned instead. | |
342 | * If the task has already been run then NULL is returned. | |
d9acd930 BB |
343 | */ |
344 | static taskq_ent_t * | |
cce83ba0 | 345 | taskq_find(taskq_t *tq, taskqid_t id) |
d9acd930 BB |
346 | { |
347 | taskq_thread_t *tqt; | |
348 | struct list_head *l; | |
349 | taskq_ent_t *t; | |
d9acd930 BB |
350 | |
351 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
d9acd930 BB |
352 | |
353 | t = taskq_find_list(tq, &tq->tq_delay_list, id); | |
354 | if (t) | |
8d9a23e8 | 355 | return (t); |
d9acd930 BB |
356 | |
357 | t = taskq_find_list(tq, &tq->tq_prio_list, id); | |
358 | if (t) | |
8d9a23e8 | 359 | return (t); |
d9acd930 BB |
360 | |
361 | t = taskq_find_list(tq, &tq->tq_pend_list, id); | |
362 | if (t) | |
8d9a23e8 | 363 | return (t); |
d9acd930 BB |
364 | |
365 | list_for_each(l, &tq->tq_active_list) { | |
366 | tqt = list_entry(l, taskq_thread_t, tqt_active_list); | |
367 | if (tqt->tqt_id == id) { | |
cce83ba0 CC |
368 | /* |
369 | * Instead of returning tqt_task, we just return a non | |
370 | * NULL value to prevent misuse, since tqt_task only | |
371 | * has two valid fields. | |
372 | */ | |
373 | return (ERR_PTR(-EBUSY)); | |
d9acd930 BB |
374 | } |
375 | } | |
376 | ||
8d9a23e8 | 377 | return (NULL); |
d9acd930 BB |
378 | } |
379 | ||
a876b030 CD |
380 | /* |
381 | * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and | |
382 | * taskq_wait() functions below. | |
383 | * | |
384 | * Taskq waiting is accomplished by tracking the lowest outstanding task | |
385 | * id and the next available task id. As tasks are dispatched they are | |
386 | * added to the tail of the pending, priority, or delay lists. As worker | |
387 | * threads become available the tasks are removed from the heads of these | |
388 | * lists and linked to the worker threads. This ensures the lists are | |
389 | * kept sorted by lowest to highest task id. | |
390 | * | |
391 | * Therefore the lowest outstanding task id can be quickly determined by | |
392 | * checking the head item from all of these lists. This value is stored | |
393 | * with the taskq as the lowest id. It only needs to be recalculated when | |
394 | * either the task with the current lowest id completes or is canceled. | |
395 | * | |
396 | * By blocking until the lowest task id exceeds the passed task id the | |
397 | * taskq_wait_outstanding() function can be easily implemented. Similarly, | |
398 | * by blocking until the lowest task id matches the next task id taskq_wait() | |
399 | * can be implemented. | |
400 | * | |
401 | * Callers should be aware that when there are multiple worked threads it | |
402 | * is possible for larger task ids to complete before smaller ones. Also | |
403 | * when the taskq contains delay tasks with small task ids callers may | |
404 | * block for a considerable length of time waiting for them to expire and | |
405 | * execute. | |
406 | */ | |
99c452bb BB |
407 | static int |
408 | taskq_wait_id_check(taskq_t *tq, taskqid_t id) | |
f1ca4da6 | 409 | { |
99c452bb | 410 | int rc; |
066b89e6 | 411 | unsigned long flags; |
bcd68186 | 412 | |
066b89e6 | 413 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
cce83ba0 | 414 | rc = (taskq_find(tq, id) == NULL); |
066b89e6 | 415 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 416 | |
99c452bb BB |
417 | return (rc); |
418 | } | |
bcd68186 | 419 | |
99c452bb BB |
420 | /* |
421 | * The taskq_wait_id() function blocks until the passed task id completes. | |
422 | * This does not guarantee that all lower task ids have completed. | |
423 | */ | |
424 | void | |
425 | taskq_wait_id(taskq_t *tq, taskqid_t id) | |
426 | { | |
427 | wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id)); | |
bcd68186 | 428 | } |
aed8671c | 429 | EXPORT_SYMBOL(taskq_wait_id); |
bcd68186 | 430 | |
d9acd930 | 431 | static int |
a876b030 | 432 | taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id) |
d9acd930 BB |
433 | { |
434 | int rc; | |
066b89e6 | 435 | unsigned long flags; |
d9acd930 | 436 | |
066b89e6 | 437 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
d9acd930 | 438 | rc = (id < tq->tq_lowest_id); |
066b89e6 | 439 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 440 | |
8d9a23e8 | 441 | return (rc); |
d9acd930 BB |
442 | } |
443 | ||
a876b030 CD |
444 | /* |
445 | * The taskq_wait_outstanding() function will block until all tasks with a | |
446 | * lower taskqid than the passed 'id' have been completed. Note that all | |
447 | * task id's are assigned monotonically at dispatch time. Zero may be | |
448 | * passed for the id to indicate all tasks dispatch up to this point, | |
449 | * but not after, should be waited for. | |
450 | */ | |
d9acd930 | 451 | void |
a876b030 | 452 | taskq_wait_outstanding(taskq_t *tq, taskqid_t id) |
d9acd930 | 453 | { |
b3a22a0a CC |
454 | id = id ? id : tq->tq_next_id - 1; |
455 | wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id)); | |
d9acd930 | 456 | } |
a876b030 | 457 | EXPORT_SYMBOL(taskq_wait_outstanding); |
d9acd930 | 458 | |
a876b030 CD |
459 | static int |
460 | taskq_wait_check(taskq_t *tq) | |
bcd68186 | 461 | { |
a876b030 | 462 | int rc; |
066b89e6 | 463 | unsigned long flags; |
bcd68186 | 464 | |
066b89e6 | 465 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
a876b030 | 466 | rc = (tq->tq_lowest_id == tq->tq_next_id); |
066b89e6 | 467 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 468 | |
a876b030 CD |
469 | return (rc); |
470 | } | |
471 | ||
472 | /* | |
473 | * The taskq_wait() function will block until the taskq is empty. | |
474 | * This means that if a taskq re-dispatches work to itself taskq_wait() | |
475 | * callers will block indefinitely. | |
476 | */ | |
477 | void | |
478 | taskq_wait(taskq_t *tq) | |
479 | { | |
480 | wait_event(tq->tq_wait_waitq, taskq_wait_check(tq)); | |
bcd68186 | 481 | } |
aed8671c | 482 | EXPORT_SYMBOL(taskq_wait); |
bcd68186 | 483 | |
c5a8b1e1 | 484 | int |
16522ac2 | 485 | taskq_member(taskq_t *tq, kthread_t *t) |
c5a8b1e1 | 486 | { |
16522ac2 | 487 | return (tq == (taskq_t *)tsd_get_by_thread(taskq_tsd, t)); |
c5a8b1e1 BB |
488 | } |
489 | EXPORT_SYMBOL(taskq_member); | |
490 | ||
d9acd930 BB |
491 | /* |
492 | * Cancel an already dispatched task given the task id. Still pending tasks | |
493 | * will be immediately canceled, and if the task is active the function will | |
494 | * block until it completes. Preallocated tasks which are canceled must be | |
495 | * freed by the caller. | |
496 | */ | |
497 | int | |
498 | taskq_cancel_id(taskq_t *tq, taskqid_t id) | |
499 | { | |
500 | taskq_ent_t *t; | |
d9acd930 | 501 | int rc = ENOENT; |
066b89e6 | 502 | unsigned long flags; |
d9acd930 BB |
503 | |
504 | ASSERT(tq); | |
505 | ||
066b89e6 | 506 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
cce83ba0 CC |
507 | t = taskq_find(tq, id); |
508 | if (t && t != ERR_PTR(-EBUSY)) { | |
d9acd930 BB |
509 | list_del_init(&t->tqent_list); |
510 | t->tqent_flags |= TQENT_FLAG_CANCEL; | |
511 | ||
512 | /* | |
513 | * When canceling the lowest outstanding task id we | |
514 | * must recalculate the new lowest outstanding id. | |
515 | */ | |
516 | if (tq->tq_lowest_id == t->tqent_id) { | |
517 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
518 | ASSERT3S(tq->tq_lowest_id, >, t->tqent_id); | |
519 | } | |
520 | ||
521 | /* | |
522 | * The task_expire() function takes the tq->tq_lock so drop | |
523 | * drop the lock before synchronously cancelling the timer. | |
524 | */ | |
525 | if (timer_pending(&t->tqent_timer)) { | |
066b89e6 | 526 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 527 | del_timer_sync(&t->tqent_timer); |
066b89e6 CC |
528 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
529 | tq->tq_lock_class); | |
d9acd930 BB |
530 | } |
531 | ||
532 | if (!(t->tqent_flags & TQENT_FLAG_PREALLOC)) | |
533 | task_done(tq, t); | |
534 | ||
535 | rc = 0; | |
536 | } | |
066b89e6 | 537 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 538 | |
cce83ba0 | 539 | if (t == ERR_PTR(-EBUSY)) { |
d9acd930 BB |
540 | taskq_wait_id(tq, id); |
541 | rc = EBUSY; | |
542 | } | |
543 | ||
8d9a23e8 | 544 | return (rc); |
d9acd930 BB |
545 | } |
546 | EXPORT_SYMBOL(taskq_cancel_id); | |
547 | ||
f5f2b87d | 548 | static int taskq_thread_spawn(taskq_t *tq); |
a64e5575 | 549 | |
bcd68186 | 550 | taskqid_t |
aed8671c | 551 | taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags) |
bcd68186 | 552 | { |
472a34ca | 553 | taskq_ent_t *t; |
cbba7146 | 554 | taskqid_t rc = TASKQID_INVALID; |
066b89e6 | 555 | unsigned long irqflags; |
f1ca4da6 | 556 | |
472a34ca BB |
557 | ASSERT(tq); |
558 | ASSERT(func); | |
d05ec4b4 | 559 | |
066b89e6 | 560 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class); |
f1ca4da6 | 561 | |
bcd68186 | 562 | /* Taskq being destroyed and all tasks drained */ |
f7a973d9 | 563 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
8d9a23e8 | 564 | goto out; |
f1ca4da6 | 565 | |
bcd68186 BB |
566 | /* Do not queue the task unless there is idle thread for it */ |
567 | ASSERT(tq->tq_nactive <= tq->tq_nthreads); | |
7bb5d92d TC |
568 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) { |
569 | /* Dynamic taskq may be able to spawn another thread */ | |
570 | if (!(tq->tq_flags & TASKQ_DYNAMIC) || taskq_thread_spawn(tq) == 0) | |
571 | goto out; | |
572 | } | |
bcd68186 | 573 | |
066b89e6 | 574 | if ((t = task_alloc(tq, flags, &irqflags)) == NULL) |
8d9a23e8 | 575 | goto out; |
f1ca4da6 | 576 | |
046a70c9 | 577 | spin_lock(&t->tqent_lock); |
f0d8bb26 | 578 | |
7bb5d92d TC |
579 | /* Queue to the front of the list to enforce TQ_NOQUEUE semantics */ |
580 | if (flags & TQ_NOQUEUE) | |
581 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
f0d8bb26 | 582 | /* Queue to the priority list instead of the pending list */ |
7bb5d92d | 583 | else if (flags & TQ_FRONT) |
046a70c9 | 584 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); |
f0d8bb26 | 585 | else |
046a70c9 | 586 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); |
f0d8bb26 | 587 | |
046a70c9 | 588 | t->tqent_id = rc = tq->tq_next_id; |
bcd68186 | 589 | tq->tq_next_id++; |
472a34ca BB |
590 | t->tqent_func = func; |
591 | t->tqent_arg = arg; | |
d9acd930 BB |
592 | t->tqent_taskq = tq; |
593 | t->tqent_timer.data = 0; | |
594 | t->tqent_timer.function = NULL; | |
595 | t->tqent_timer.expires = 0; | |
8f3b403a | 596 | t->tqent_birth = jiffies; |
44217f7a PS |
597 | |
598 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
599 | ||
046a70c9 | 600 | spin_unlock(&t->tqent_lock); |
0bb43ca2 NB |
601 | |
602 | wake_up(&tq->tq_work_waitq); | |
bcd68186 | 603 | out: |
a64e5575 | 604 | /* Spawn additional taskq threads if required. */ |
7bb5d92d | 605 | if (!(flags & TQ_NOQUEUE) && tq->tq_nactive == tq->tq_nthreads) |
f5f2b87d | 606 | (void) taskq_thread_spawn(tq); |
a64e5575 | 607 | |
066b89e6 | 608 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
8d9a23e8 | 609 | return (rc); |
f1ca4da6 | 610 | } |
aed8671c | 611 | EXPORT_SYMBOL(taskq_dispatch); |
44217f7a | 612 | |
d9acd930 BB |
613 | taskqid_t |
614 | taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, | |
615 | uint_t flags, clock_t expire_time) | |
616 | { | |
cbba7146 | 617 | taskqid_t rc = TASKQID_INVALID; |
8d9a23e8 | 618 | taskq_ent_t *t; |
066b89e6 | 619 | unsigned long irqflags; |
d9acd930 BB |
620 | |
621 | ASSERT(tq); | |
622 | ASSERT(func); | |
623 | ||
066b89e6 | 624 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class); |
d9acd930 BB |
625 | |
626 | /* Taskq being destroyed and all tasks drained */ | |
f7a973d9 | 627 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
8d9a23e8 | 628 | goto out; |
d9acd930 | 629 | |
066b89e6 | 630 | if ((t = task_alloc(tq, flags, &irqflags)) == NULL) |
8d9a23e8 | 631 | goto out; |
d9acd930 BB |
632 | |
633 | spin_lock(&t->tqent_lock); | |
634 | ||
635 | /* Queue to the delay list for subsequent execution */ | |
636 | list_add_tail(&t->tqent_list, &tq->tq_delay_list); | |
637 | ||
638 | t->tqent_id = rc = tq->tq_next_id; | |
639 | tq->tq_next_id++; | |
640 | t->tqent_func = func; | |
641 | t->tqent_arg = arg; | |
642 | t->tqent_taskq = tq; | |
643 | t->tqent_timer.data = (unsigned long)t; | |
644 | t->tqent_timer.function = task_expire; | |
645 | t->tqent_timer.expires = (unsigned long)expire_time; | |
646 | add_timer(&t->tqent_timer); | |
647 | ||
648 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
649 | ||
650 | spin_unlock(&t->tqent_lock); | |
651 | out: | |
a64e5575 | 652 | /* Spawn additional taskq threads if required. */ |
f5f2b87d | 653 | if (tq->tq_nactive == tq->tq_nthreads) |
654 | (void) taskq_thread_spawn(tq); | |
066b89e6 | 655 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
8d9a23e8 | 656 | return (rc); |
d9acd930 BB |
657 | } |
658 | EXPORT_SYMBOL(taskq_dispatch_delay); | |
659 | ||
44217f7a | 660 | void |
aed8671c | 661 | taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags, |
2c4332cf | 662 | taskq_ent_t *t) |
44217f7a | 663 | { |
066b89e6 | 664 | unsigned long irqflags; |
44217f7a PS |
665 | ASSERT(tq); |
666 | ASSERT(func); | |
44217f7a | 667 | |
066b89e6 | 668 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, |
326172d8 | 669 | tq->tq_lock_class); |
44217f7a PS |
670 | |
671 | /* Taskq being destroyed and all tasks drained */ | |
f7a973d9 | 672 | if (!(tq->tq_flags & TASKQ_ACTIVE)) { |
cbba7146 | 673 | t->tqent_id = TASKQID_INVALID; |
44217f7a PS |
674 | goto out; |
675 | } | |
676 | ||
7bb5d92d TC |
677 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) { |
678 | /* Dynamic taskq may be able to spawn another thread */ | |
679 | if (!(tq->tq_flags & TASKQ_DYNAMIC) || taskq_thread_spawn(tq) == 0) | |
680 | goto out2; | |
681 | flags |= TQ_FRONT; | |
682 | } | |
683 | ||
44217f7a PS |
684 | spin_lock(&t->tqent_lock); |
685 | ||
9243b0fb BP |
686 | /* |
687 | * Make sure the entry is not on some other taskq; it is important to | |
688 | * ASSERT() under lock | |
689 | */ | |
690 | ASSERT(taskq_empty_ent(t)); | |
691 | ||
44217f7a PS |
692 | /* |
693 | * Mark it as a prealloc'd task. This is important | |
694 | * to ensure that we don't free it later. | |
695 | */ | |
696 | t->tqent_flags |= TQENT_FLAG_PREALLOC; | |
697 | ||
698 | /* Queue to the priority list instead of the pending list */ | |
699 | if (flags & TQ_FRONT) | |
700 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
701 | else | |
702 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
703 | ||
704 | t->tqent_id = tq->tq_next_id; | |
705 | tq->tq_next_id++; | |
706 | t->tqent_func = func; | |
707 | t->tqent_arg = arg; | |
d9acd930 | 708 | t->tqent_taskq = tq; |
8f3b403a | 709 | t->tqent_birth = jiffies; |
44217f7a PS |
710 | |
711 | spin_unlock(&t->tqent_lock); | |
712 | ||
713 | wake_up(&tq->tq_work_waitq); | |
714 | out: | |
a64e5575 | 715 | /* Spawn additional taskq threads if required. */ |
f5f2b87d | 716 | if (tq->tq_nactive == tq->tq_nthreads) |
717 | (void) taskq_thread_spawn(tq); | |
7bb5d92d | 718 | out2: |
066b89e6 | 719 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
44217f7a | 720 | } |
aed8671c | 721 | EXPORT_SYMBOL(taskq_dispatch_ent); |
44217f7a PS |
722 | |
723 | int | |
aed8671c | 724 | taskq_empty_ent(taskq_ent_t *t) |
44217f7a | 725 | { |
2c4332cf | 726 | return (list_empty(&t->tqent_list)); |
44217f7a | 727 | } |
aed8671c | 728 | EXPORT_SYMBOL(taskq_empty_ent); |
44217f7a PS |
729 | |
730 | void | |
aed8671c | 731 | taskq_init_ent(taskq_ent_t *t) |
44217f7a PS |
732 | { |
733 | spin_lock_init(&t->tqent_lock); | |
d9acd930 BB |
734 | init_waitqueue_head(&t->tqent_waitq); |
735 | init_timer(&t->tqent_timer); | |
44217f7a PS |
736 | INIT_LIST_HEAD(&t->tqent_list); |
737 | t->tqent_id = 0; | |
738 | t->tqent_func = NULL; | |
739 | t->tqent_arg = NULL; | |
740 | t->tqent_flags = 0; | |
d9acd930 | 741 | t->tqent_taskq = NULL; |
44217f7a | 742 | } |
aed8671c | 743 | EXPORT_SYMBOL(taskq_init_ent); |
44217f7a | 744 | |
f7a973d9 BB |
745 | /* |
746 | * Return the next pending task, preference is given to tasks on the | |
747 | * priority list which were dispatched with TQ_FRONT. | |
748 | */ | |
749 | static taskq_ent_t * | |
750 | taskq_next_ent(taskq_t *tq) | |
751 | { | |
752 | struct list_head *list; | |
753 | ||
754 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
755 | ||
756 | if (!list_empty(&tq->tq_prio_list)) | |
757 | list = &tq->tq_prio_list; | |
758 | else if (!list_empty(&tq->tq_pend_list)) | |
759 | list = &tq->tq_pend_list; | |
760 | else | |
761 | return (NULL); | |
762 | ||
763 | return (list_entry(list->next, taskq_ent_t, tqent_list)); | |
764 | } | |
765 | ||
766 | /* | |
767 | * Spawns a new thread for the specified taskq. | |
768 | */ | |
769 | static void | |
770 | taskq_thread_spawn_task(void *arg) | |
771 | { | |
772 | taskq_t *tq = (taskq_t *)arg; | |
066b89e6 | 773 | unsigned long flags; |
f7a973d9 | 774 | |
5ce028b0 CC |
775 | if (taskq_thread_create(tq) == NULL) { |
776 | /* restore spawning count if failed */ | |
777 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); | |
778 | tq->tq_nspawn--; | |
779 | spin_unlock_irqrestore(&tq->tq_lock, flags); | |
780 | } | |
f7a973d9 BB |
781 | } |
782 | ||
783 | /* | |
326172d8 | 784 | * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current |
f7a973d9 BB |
785 | * number of threads is insufficient to handle the pending tasks. These |
786 | * new threads must be created by the dedicated dynamic_taskq to avoid | |
787 | * deadlocks between thread creation and memory reclaim. The system_taskq | |
788 | * which is also a dynamic taskq cannot be safely used for this. | |
789 | */ | |
790 | static int | |
f5f2b87d | 791 | taskq_thread_spawn(taskq_t *tq) |
f7a973d9 BB |
792 | { |
793 | int spawning = 0; | |
794 | ||
795 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
796 | return (0); | |
797 | ||
f5f2b87d | 798 | if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) && |
f7a973d9 BB |
799 | (tq->tq_flags & TASKQ_ACTIVE)) { |
800 | spawning = (++tq->tq_nspawn); | |
801 | taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task, | |
802 | tq, TQ_NOSLEEP); | |
803 | } | |
804 | ||
805 | return (spawning); | |
806 | } | |
807 | ||
808 | /* | |
809 | * Threads in a dynamic taskq should only exit once it has been completely | |
810 | * drained and no other threads are actively servicing tasks. This prevents | |
811 | * threads from being created and destroyed more than is required. | |
812 | * | |
813 | * The first thread is the thread list is treated as the primary thread. | |
814 | * There is nothing special about the primary thread but in order to avoid | |
815 | * all the taskq pids from changing we opt to make it long running. | |
816 | */ | |
817 | static int | |
818 | taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt) | |
819 | { | |
820 | ASSERT(spin_is_locked(&tq->tq_lock)); | |
821 | ||
822 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
823 | return (0); | |
824 | ||
825 | if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t, | |
826 | tqt_thread_list) == tqt) | |
827 | return (0); | |
828 | ||
829 | return | |
830 | ((tq->tq_nspawn == 0) && /* No threads are being spawned */ | |
831 | (tq->tq_nactive == 0) && /* No threads are handling tasks */ | |
832 | (tq->tq_nthreads > 1) && /* More than 1 thread is running */ | |
833 | (!taskq_next_ent(tq)) && /* There are no pending tasks */ | |
2c4332cf | 834 | (spl_taskq_thread_dynamic)); /* Dynamic taskqs are allowed */ |
f7a973d9 BB |
835 | } |
836 | ||
bcd68186 BB |
837 | static int |
838 | taskq_thread(void *args) | |
839 | { | |
472a34ca BB |
840 | DECLARE_WAITQUEUE(wait, current); |
841 | sigset_t blocked; | |
2c02b71b | 842 | taskq_thread_t *tqt = args; |
472a34ca BB |
843 | taskq_t *tq; |
844 | taskq_ent_t *t; | |
f7a973d9 | 845 | int seq_tasks = 0; |
066b89e6 | 846 | unsigned long flags; |
cce83ba0 | 847 | taskq_ent_t dup_task = {}; |
bcd68186 | 848 | |
472a34ca | 849 | ASSERT(tqt); |
326172d8 | 850 | ASSERT(tqt->tqt_tq); |
2c02b71b | 851 | tq = tqt->tqt_tq; |
472a34ca | 852 | current->flags |= PF_NOFREEZE; |
bcd68186 | 853 | |
b4ad50ac | 854 | (void) spl_fstrans_mark(); |
d4bf6d84 | 855 | |
472a34ca BB |
856 | sigfillset(&blocked); |
857 | sigprocmask(SIG_BLOCK, &blocked, NULL); | |
858 | flush_signals(current); | |
bcd68186 | 859 | |
16522ac2 | 860 | tsd_set(taskq_tsd, tq); |
066b89e6 | 861 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
5ce028b0 CC |
862 | /* |
863 | * If we are dynamically spawned, decrease spawning count. Note that | |
864 | * we could be created during taskq_create, in which case we shouldn't | |
865 | * do the decrement. But it's fine because taskq_create will reset | |
866 | * tq_nspawn later. | |
867 | */ | |
868 | if (tq->tq_flags & TASKQ_DYNAMIC) | |
869 | tq->tq_nspawn--; | |
f7a973d9 BB |
870 | |
871 | /* Immediately exit if more threads than allowed were created. */ | |
872 | if (tq->tq_nthreads >= tq->tq_maxthreads) | |
873 | goto error; | |
874 | ||
472a34ca | 875 | tq->tq_nthreads++; |
f7a973d9 | 876 | list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list); |
472a34ca BB |
877 | wake_up(&tq->tq_wait_waitq); |
878 | set_current_state(TASK_INTERRUPTIBLE); | |
bcd68186 | 879 | |
472a34ca | 880 | while (!kthread_should_stop()) { |
bcd68186 | 881 | |
f0d8bb26 NB |
882 | if (list_empty(&tq->tq_pend_list) && |
883 | list_empty(&tq->tq_prio_list)) { | |
f7a973d9 BB |
884 | |
885 | if (taskq_thread_should_stop(tq, tqt)) { | |
886 | wake_up_all(&tq->tq_wait_waitq); | |
887 | break; | |
888 | } | |
889 | ||
3c6ed541 | 890 | add_wait_queue_exclusive(&tq->tq_work_waitq, &wait); |
066b89e6 | 891 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
f7a973d9 | 892 | |
bcd68186 | 893 | schedule(); |
f7a973d9 BB |
894 | seq_tasks = 0; |
895 | ||
066b89e6 CC |
896 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
897 | tq->tq_lock_class); | |
3c6ed541 | 898 | remove_wait_queue(&tq->tq_work_waitq, &wait); |
bcd68186 BB |
899 | } else { |
900 | __set_current_state(TASK_RUNNING); | |
901 | } | |
902 | ||
f7a973d9 | 903 | if ((t = taskq_next_ent(tq)) != NULL) { |
472a34ca | 904 | list_del_init(&t->tqent_list); |
8f2503e0 | 905 | |
2c4332cf | 906 | /* |
cce83ba0 CC |
907 | * A TQENT_FLAG_PREALLOC task may be reused or freed |
908 | * during the task function call. Store tqent_id and | |
909 | * tqent_flags here. | |
910 | * | |
911 | * Also use an on stack taskq_ent_t for tqt_task | |
912 | * assignment in this case. We only populate the two | |
913 | * fields used by the only user in taskq proc file. | |
2c4332cf | 914 | */ |
e7e5f78e | 915 | tqt->tqt_id = t->tqent_id; |
8f2503e0 PS |
916 | tqt->tqt_flags = t->tqent_flags; |
917 | ||
cce83ba0 CC |
918 | if (t->tqent_flags & TQENT_FLAG_PREALLOC) { |
919 | dup_task.tqent_func = t->tqent_func; | |
920 | dup_task.tqent_arg = t->tqent_arg; | |
921 | t = &dup_task; | |
922 | } | |
923 | tqt->tqt_task = t; | |
924 | ||
2c02b71b | 925 | taskq_insert_in_order(tq, tqt); |
472a34ca | 926 | tq->tq_nactive++; |
066b89e6 | 927 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 BB |
928 | |
929 | /* Perform the requested task */ | |
472a34ca | 930 | t->tqent_func(t->tqent_arg); |
bcd68186 | 931 | |
066b89e6 CC |
932 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
933 | tq->tq_lock_class); | |
472a34ca | 934 | tq->tq_nactive--; |
2c02b71b | 935 | list_del_init(&tqt->tqt_active_list); |
d9acd930 | 936 | tqt->tqt_task = NULL; |
8f2503e0 PS |
937 | |
938 | /* For prealloc'd tasks, we don't free anything. */ | |
f7a973d9 | 939 | if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC)) |
8f2503e0 | 940 | task_done(tq, t); |
bcd68186 | 941 | |
2c4332cf BB |
942 | /* |
943 | * When the current lowest outstanding taskqid is | |
944 | * done calculate the new lowest outstanding id | |
945 | */ | |
e7e5f78e | 946 | if (tq->tq_lowest_id == tqt->tqt_id) { |
bcd68186 | 947 | tq->tq_lowest_id = taskq_lowest_id(tq); |
e7e5f78e | 948 | ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id); |
bcd68186 BB |
949 | } |
950 | ||
f7a973d9 | 951 | /* Spawn additional taskq threads if required. */ |
f5f2b87d | 952 | if ((++seq_tasks) > spl_taskq_thread_sequential && |
953 | taskq_thread_spawn(tq)) | |
f7a973d9 BB |
954 | seq_tasks = 0; |
955 | ||
cbba7146 | 956 | tqt->tqt_id = TASKQID_INVALID; |
8f2503e0 | 957 | tqt->tqt_flags = 0; |
472a34ca | 958 | wake_up_all(&tq->tq_wait_waitq); |
f7a973d9 BB |
959 | } else { |
960 | if (taskq_thread_should_stop(tq, tqt)) | |
961 | break; | |
bcd68186 BB |
962 | } |
963 | ||
964 | set_current_state(TASK_INTERRUPTIBLE); | |
965 | ||
472a34ca | 966 | } |
bcd68186 BB |
967 | |
968 | __set_current_state(TASK_RUNNING); | |
472a34ca | 969 | tq->tq_nthreads--; |
2c02b71b | 970 | list_del_init(&tqt->tqt_thread_list); |
f7a973d9 BB |
971 | error: |
972 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
066b89e6 | 973 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 974 | |
16522ac2 CC |
975 | tsd_set(taskq_tsd, NULL); |
976 | ||
8d9a23e8 | 977 | return (0); |
bcd68186 BB |
978 | } |
979 | ||
f7a973d9 BB |
980 | static taskq_thread_t * |
981 | taskq_thread_create(taskq_t *tq) | |
982 | { | |
983 | static int last_used_cpu = 0; | |
984 | taskq_thread_t *tqt; | |
985 | ||
986 | tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE); | |
987 | INIT_LIST_HEAD(&tqt->tqt_thread_list); | |
988 | INIT_LIST_HEAD(&tqt->tqt_active_list); | |
989 | tqt->tqt_tq = tq; | |
cbba7146 | 990 | tqt->tqt_id = TASKQID_INVALID; |
f7a973d9 BB |
991 | |
992 | tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt, | |
993 | "%s", tq->tq_name); | |
994 | if (tqt->tqt_thread == NULL) { | |
995 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
996 | return (NULL); | |
997 | } | |
998 | ||
999 | if (spl_taskq_thread_bind) { | |
1000 | last_used_cpu = (last_used_cpu + 1) % num_online_cpus(); | |
1001 | kthread_bind(tqt->tqt_thread, last_used_cpu); | |
1002 | } | |
1003 | ||
62aa81a5 BB |
1004 | if (spl_taskq_thread_priority) |
1005 | set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri)); | |
1006 | ||
f7a973d9 BB |
1007 | wake_up_process(tqt->tqt_thread); |
1008 | ||
1009 | return (tqt); | |
1010 | } | |
1011 | ||
f1ca4da6 | 1012 | taskq_t * |
aed8671c | 1013 | taskq_create(const char *name, int nthreads, pri_t pri, |
472a34ca | 1014 | int minalloc, int maxalloc, uint_t flags) |
f1ca4da6 | 1015 | { |
472a34ca | 1016 | taskq_t *tq; |
2c02b71b | 1017 | taskq_thread_t *tqt; |
f7a973d9 | 1018 | int count = 0, rc = 0, i; |
066b89e6 | 1019 | unsigned long irqflags; |
bcd68186 | 1020 | |
472a34ca | 1021 | ASSERT(name != NULL); |
472a34ca BB |
1022 | ASSERT(minalloc >= 0); |
1023 | ASSERT(maxalloc <= INT_MAX); | |
f7a973d9 | 1024 | ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */ |
bcd68186 | 1025 | |
915404bd BB |
1026 | /* Scale the number of threads using nthreads as a percentage */ |
1027 | if (flags & TASKQ_THREADS_CPU_PCT) { | |
1028 | ASSERT(nthreads <= 100); | |
1029 | ASSERT(nthreads >= 0); | |
1030 | nthreads = MIN(nthreads, 100); | |
1031 | nthreads = MAX(nthreads, 0); | |
1032 | nthreads = MAX((num_online_cpus() * nthreads) / 100, 1); | |
1033 | } | |
1034 | ||
f7a973d9 | 1035 | tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE); |
472a34ca | 1036 | if (tq == NULL) |
8d9a23e8 | 1037 | return (NULL); |
bcd68186 | 1038 | |
472a34ca | 1039 | spin_lock_init(&tq->tq_lock); |
472a34ca BB |
1040 | INIT_LIST_HEAD(&tq->tq_thread_list); |
1041 | INIT_LIST_HEAD(&tq->tq_active_list); | |
2c4332cf BB |
1042 | tq->tq_name = strdup(name); |
1043 | tq->tq_nactive = 0; | |
1044 | tq->tq_nthreads = 0; | |
1045 | tq->tq_nspawn = 0; | |
f7a973d9 | 1046 | tq->tq_maxthreads = nthreads; |
2c4332cf BB |
1047 | tq->tq_pri = pri; |
1048 | tq->tq_minalloc = minalloc; | |
1049 | tq->tq_maxalloc = maxalloc; | |
1050 | tq->tq_nalloc = 0; | |
1051 | tq->tq_flags = (flags | TASKQ_ACTIVE); | |
cbba7146 BB |
1052 | tq->tq_next_id = TASKQID_INITIAL; |
1053 | tq->tq_lowest_id = TASKQID_INITIAL; | |
472a34ca BB |
1054 | INIT_LIST_HEAD(&tq->tq_free_list); |
1055 | INIT_LIST_HEAD(&tq->tq_pend_list); | |
1056 | INIT_LIST_HEAD(&tq->tq_prio_list); | |
d9acd930 | 1057 | INIT_LIST_HEAD(&tq->tq_delay_list); |
472a34ca BB |
1058 | init_waitqueue_head(&tq->tq_work_waitq); |
1059 | init_waitqueue_head(&tq->tq_wait_waitq); | |
326172d8 | 1060 | tq->tq_lock_class = TQ_LOCK_GENERAL; |
200366f2 | 1061 | INIT_LIST_HEAD(&tq->tq_taskqs); |
bcd68186 | 1062 | |
f7a973d9 | 1063 | if (flags & TASKQ_PREPOPULATE) { |
066b89e6 | 1064 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, |
326172d8 | 1065 | tq->tq_lock_class); |
f7a973d9 | 1066 | |
472a34ca | 1067 | for (i = 0; i < minalloc; i++) |
066b89e6 CC |
1068 | task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW, |
1069 | &irqflags)); | |
6e605b6e | 1070 | |
066b89e6 | 1071 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
f7a973d9 BB |
1072 | } |
1073 | ||
1074 | if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) | |
1075 | nthreads = 1; | |
6e605b6e | 1076 | |
2c02b71b | 1077 | for (i = 0; i < nthreads; i++) { |
f7a973d9 BB |
1078 | tqt = taskq_thread_create(tq); |
1079 | if (tqt == NULL) | |
2c02b71b | 1080 | rc = 1; |
f7a973d9 BB |
1081 | else |
1082 | count++; | |
2c02b71b | 1083 | } |
bcd68186 | 1084 | |
472a34ca | 1085 | /* Wait for all threads to be started before potential destroy */ |
f7a973d9 | 1086 | wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count); |
5ce028b0 CC |
1087 | /* |
1088 | * taskq_thread might have touched nspawn, but we don't want them to | |
1089 | * because they're not dynamically spawned. So we reset it to 0 | |
1090 | */ | |
1091 | tq->tq_nspawn = 0; | |
bcd68186 | 1092 | |
472a34ca | 1093 | if (rc) { |
aed8671c | 1094 | taskq_destroy(tq); |
472a34ca | 1095 | tq = NULL; |
200366f2 TC |
1096 | } else { |
1097 | down_write(&tq_list_sem); | |
1098 | tq->tq_instance = taskq_find_by_name(name) + 1; | |
1099 | list_add_tail(&tq->tq_taskqs, &tq_list); | |
1100 | up_write(&tq_list_sem); | |
472a34ca | 1101 | } |
bcd68186 | 1102 | |
8d9a23e8 | 1103 | return (tq); |
f1ca4da6 | 1104 | } |
aed8671c | 1105 | EXPORT_SYMBOL(taskq_create); |
b123971f BB |
1106 | |
1107 | void | |
aed8671c | 1108 | taskq_destroy(taskq_t *tq) |
b123971f | 1109 | { |
2c02b71b PS |
1110 | struct task_struct *thread; |
1111 | taskq_thread_t *tqt; | |
046a70c9 | 1112 | taskq_ent_t *t; |
066b89e6 | 1113 | unsigned long flags; |
b123971f | 1114 | |
bcd68186 | 1115 | ASSERT(tq); |
066b89e6 | 1116 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
f7a973d9 | 1117 | tq->tq_flags &= ~TASKQ_ACTIVE; |
066b89e6 | 1118 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 1119 | |
f7a973d9 BB |
1120 | /* |
1121 | * When TASKQ_ACTIVE is clear new tasks may not be added nor may | |
1122 | * new worker threads be spawned for dynamic taskq. | |
1123 | */ | |
1124 | if (dynamic_taskq != NULL) | |
1125 | taskq_wait_outstanding(dynamic_taskq, 0); | |
1126 | ||
aed8671c | 1127 | taskq_wait(tq); |
bcd68186 | 1128 | |
200366f2 TC |
1129 | /* remove taskq from global list used by the kstats */ |
1130 | down_write(&tq_list_sem); | |
1131 | list_del(&tq->tq_taskqs); | |
1132 | up_write(&tq_list_sem); | |
1133 | ||
066b89e6 | 1134 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
5ce028b0 CC |
1135 | /* wait for spawning threads to insert themselves to the list */ |
1136 | while (tq->tq_nspawn) { | |
1137 | spin_unlock_irqrestore(&tq->tq_lock, flags); | |
1138 | schedule_timeout_interruptible(1); | |
1139 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); | |
1140 | } | |
bcd68186 | 1141 | |
2c02b71b PS |
1142 | /* |
1143 | * Signal each thread to exit and block until it does. Each thread | |
1144 | * is responsible for removing itself from the list and freeing its | |
1145 | * taskq_thread_t. This allows for idle threads to opt to remove | |
1146 | * themselves from the taskq. They can be recreated as needed. | |
1147 | */ | |
1148 | while (!list_empty(&tq->tq_thread_list)) { | |
1149 | tqt = list_entry(tq->tq_thread_list.next, | |
f7a973d9 | 1150 | taskq_thread_t, tqt_thread_list); |
2c02b71b | 1151 | thread = tqt->tqt_thread; |
066b89e6 | 1152 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
2c02b71b PS |
1153 | |
1154 | kthread_stop(thread); | |
1155 | ||
066b89e6 | 1156 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
326172d8 | 1157 | tq->tq_lock_class); |
2c02b71b PS |
1158 | } |
1159 | ||
472a34ca | 1160 | while (!list_empty(&tq->tq_free_list)) { |
046a70c9 | 1161 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); |
44217f7a PS |
1162 | |
1163 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
1164 | ||
472a34ca BB |
1165 | list_del_init(&t->tqent_list); |
1166 | task_free(tq, t); | |
1167 | } | |
bcd68186 | 1168 | |
f7a973d9 BB |
1169 | ASSERT0(tq->tq_nthreads); |
1170 | ASSERT0(tq->tq_nalloc); | |
1171 | ASSERT0(tq->tq_nspawn); | |
472a34ca BB |
1172 | ASSERT(list_empty(&tq->tq_thread_list)); |
1173 | ASSERT(list_empty(&tq->tq_active_list)); | |
1174 | ASSERT(list_empty(&tq->tq_free_list)); | |
1175 | ASSERT(list_empty(&tq->tq_pend_list)); | |
1176 | ASSERT(list_empty(&tq->tq_prio_list)); | |
d9acd930 | 1177 | ASSERT(list_empty(&tq->tq_delay_list)); |
bcd68186 | 1178 | |
066b89e6 | 1179 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
2c02b71b | 1180 | |
f7a973d9 BB |
1181 | strfree(tq->tq_name); |
1182 | kmem_free(tq, sizeof (taskq_t)); | |
b123971f | 1183 | } |
aed8671c | 1184 | EXPORT_SYMBOL(taskq_destroy); |
e9cb2b4f | 1185 | |
8f3b403a CC |
1186 | |
1187 | static unsigned int spl_taskq_kick = 0; | |
1188 | ||
1189 | /* | |
1190 | * 2.6.36 API Change | |
1191 | * module_param_cb is introduced to take kernel_param_ops and | |
1192 | * module_param_call is marked as obsolete. Also set and get operations | |
1193 | * were changed to take a 'const struct kernel_param *'. | |
1194 | */ | |
1195 | static int | |
1196 | #ifdef module_param_cb | |
1197 | param_set_taskq_kick(const char *val, const struct kernel_param *kp) | |
1198 | #else | |
1199 | param_set_taskq_kick(const char *val, struct kernel_param *kp) | |
1200 | #endif | |
1201 | { | |
1202 | int ret; | |
1203 | taskq_t *tq; | |
1204 | taskq_ent_t *t; | |
1205 | unsigned long flags; | |
1206 | ||
1207 | ret = param_set_uint(val, kp); | |
1208 | if (ret < 0 || !spl_taskq_kick) | |
1209 | return (ret); | |
1210 | /* reset value */ | |
1211 | spl_taskq_kick = 0; | |
1212 | ||
1213 | down_read(&tq_list_sem); | |
1214 | list_for_each_entry(tq, &tq_list, tq_taskqs) { | |
1215 | spin_lock_irqsave_nested(&tq->tq_lock, flags, | |
1216 | tq->tq_lock_class); | |
1217 | /* Check if the first pending is older than 5 seconds */ | |
1218 | t = taskq_next_ent(tq); | |
1219 | if (t && time_after(jiffies, t->tqent_birth + 5*HZ)) { | |
1220 | (void) taskq_thread_spawn(tq); | |
1221 | printk(KERN_INFO "spl: Kicked taskq %s/%d\n", | |
1222 | tq->tq_name, tq->tq_instance); | |
1223 | } | |
1224 | spin_unlock_irqrestore(&tq->tq_lock, flags); | |
1225 | } | |
1226 | up_read(&tq_list_sem); | |
1227 | return (ret); | |
1228 | } | |
1229 | ||
1230 | #ifdef module_param_cb | |
1231 | static const struct kernel_param_ops param_ops_taskq_kick = { | |
1232 | .set = param_set_taskq_kick, | |
1233 | .get = param_get_uint, | |
1234 | }; | |
1235 | module_param_cb(spl_taskq_kick, ¶m_ops_taskq_kick, &spl_taskq_kick, 0644); | |
1236 | #else | |
1237 | module_param_call(spl_taskq_kick, param_set_taskq_kick, param_get_uint, | |
1238 | &spl_taskq_kick, 0644); | |
1239 | #endif | |
1240 | MODULE_PARM_DESC(spl_taskq_kick, | |
1241 | "Write nonzero to kick stuck taskqs to spawn more threads"); | |
1242 | ||
e9cb2b4f BB |
1243 | int |
1244 | spl_taskq_init(void) | |
1245 | { | |
16522ac2 CC |
1246 | tsd_create(&taskq_tsd, NULL); |
1247 | ||
3c82160f | 1248 | system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64), |
9dc5ffbe | 1249 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); |
e9cb2b4f | 1250 | if (system_taskq == NULL) |
8d9a23e8 | 1251 | return (1); |
e9cb2b4f | 1252 | |
f200b836 CC |
1253 | system_delay_taskq = taskq_create("spl_delay_taskq", MAX(boot_ncpus, 4), |
1254 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); | |
1255 | if (system_delay_taskq == NULL) { | |
1256 | taskq_destroy(system_taskq); | |
1257 | return (1); | |
1258 | } | |
1259 | ||
f7a973d9 | 1260 | dynamic_taskq = taskq_create("spl_dynamic_taskq", 1, |
9dc5ffbe | 1261 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE); |
f7a973d9 BB |
1262 | if (dynamic_taskq == NULL) { |
1263 | taskq_destroy(system_taskq); | |
f200b836 | 1264 | taskq_destroy(system_delay_taskq); |
f7a973d9 BB |
1265 | return (1); |
1266 | } | |
1267 | ||
2c4332cf BB |
1268 | /* |
1269 | * This is used to annotate tq_lock, so | |
1270 | * taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch | |
326172d8 OF |
1271 | * does not trigger a lockdep warning re: possible recursive locking |
1272 | */ | |
1273 | dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC; | |
1274 | ||
8d9a23e8 | 1275 | return (0); |
e9cb2b4f BB |
1276 | } |
1277 | ||
1278 | void | |
1279 | spl_taskq_fini(void) | |
1280 | { | |
f7a973d9 BB |
1281 | taskq_destroy(dynamic_taskq); |
1282 | dynamic_taskq = NULL; | |
1283 | ||
f200b836 CC |
1284 | taskq_destroy(system_delay_taskq); |
1285 | system_delay_taskq = NULL; | |
1286 | ||
e9cb2b4f | 1287 | taskq_destroy(system_taskq); |
f7a973d9 | 1288 | system_taskq = NULL; |
16522ac2 CC |
1289 | |
1290 | tsd_destroy(&taskq_tsd); | |
e9cb2b4f | 1291 | } |