]>
Commit | Line | Data |
---|---|---|
ec06701b | 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 | 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 | 22 | * with the SPL. If not, see <http://www.gnu.org/licenses/>. |
ec06701b | 23 | * |
716154c5 | 24 | * Solaris Porting Layer (SPL) Task Queue Implementation. |
ec06701b | 25 | */ |
715f6251 | 26 | |
f4b37741 | 27 | #include <sys/taskq.h> |
3d061e9d | 28 | #include <sys/kmem.h> |
ec06701b | 29 | #include <sys/tsd.h> |
937879f1 | 30 | |
9e4fb5c2 LG |
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 | ||
f6188ddd | 35 | |
ec06701b | 36 | int spl_taskq_thread_dynamic = 1; |
f6188ddd AX |
37 | module_param(spl_taskq_thread_dynamic, int, 0644); |
38 | MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads"); | |
39 | ||
40 | int spl_taskq_thread_priority = 1; | |
41 | module_param(spl_taskq_thread_priority, int, 0644); | |
42 | MODULE_PARM_DESC(spl_taskq_thread_priority, | |
ec06701b | 43 | "Allow non-default priority for taskq threads"); |
f6188ddd AX |
44 | |
45 | int spl_taskq_thread_sequential = 4; | |
46 | module_param(spl_taskq_thread_sequential, int, 0644); | |
47 | MODULE_PARM_DESC(spl_taskq_thread_sequential, | |
ec06701b | 48 | "Create new taskq threads after N sequential tasks"); |
f6188ddd | 49 | |
e9cb2b4f BB |
50 | /* Global system-wide dynamic task queue available for all consumers */ |
51 | taskq_t *system_taskq; | |
52 | EXPORT_SYMBOL(system_taskq); | |
ec06701b AX |
53 | /* Global dynamic task queue for long delay */ |
54 | taskq_t *system_delay_taskq; | |
55 | EXPORT_SYMBOL(system_delay_taskq); | |
e9cb2b4f | 56 | |
f6188ddd AX |
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 | ||
ec06701b AX |
61 | /* List of all taskqs */ |
62 | LIST_HEAD(tq_list); | |
63 | DECLARE_RWSEM(tq_list_sem); | |
64 | static uint_t taskq_tsd; | |
65 | ||
9b51f218 BB |
66 | static int |
67 | task_km_flags(uint_t flags) | |
68 | { | |
69 | if (flags & TQ_NOSLEEP) | |
ec06701b | 70 | return (KM_NOSLEEP); |
9b51f218 BB |
71 | |
72 | if (flags & TQ_PUSHPAGE) | |
ec06701b | 73 | return (KM_PUSHPAGE); |
9b51f218 | 74 | |
ec06701b AX |
75 | return (KM_SLEEP); |
76 | } | |
77 | ||
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) | |
34ee731f | 90 | return (tq->tq_instance); |
ec06701b AX |
91 | } |
92 | return (-1); | |
9b51f218 BB |
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 * |
ec06701b | 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 | 105 | ASSERT(tq); |
bcd68186 | 106 | retry: |
472a34ca BB |
107 | /* Acquire taskq_ent_t's from free list if available */ |
108 | if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) { | |
109 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); | |
110 | ||
111 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
d9acd930 BB |
112 | ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL)); |
113 | ASSERT(!timer_pending(&t->tqent_timer)); | |
472a34ca BB |
114 | |
115 | list_del_init(&t->tqent_list); | |
10946b02 | 116 | return (t); |
472a34ca BB |
117 | } |
118 | ||
119 | /* Free list is empty and memory allocations are prohibited */ | |
120 | if (flags & TQ_NOALLOC) | |
10946b02 | 121 | return (NULL); |
472a34ca BB |
122 | |
123 | /* Hit maximum taskq_ent_t pool size */ | |
124 | if (tq->tq_nalloc >= tq->tq_maxalloc) { | |
125 | if (flags & TQ_NOSLEEP) | |
10946b02 | 126 | return (NULL); |
472a34ca BB |
127 | |
128 | /* | |
129 | * Sleep periodically polling the free list for an available | |
130 | * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed | |
131 | * but we cannot block forever waiting for an taskq_ent_t to | |
132 | * show up in the free list, otherwise a deadlock can happen. | |
133 | * | |
134 | * Therefore, we need to allocate a new task even if the number | |
135 | * of allocated tasks is above tq->tq_maxalloc, but we still | |
136 | * end up delaying the task allocation by one second, thereby | |
137 | * throttling the task dispatch rate. | |
138 | */ | |
ec06701b | 139 | spin_unlock_irqrestore(&tq->tq_lock, *irqflags); |
472a34ca | 140 | schedule_timeout(HZ / 100); |
ec06701b AX |
141 | spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, |
142 | tq->tq_lock_class); | |
10946b02 AX |
143 | if (count < 100) { |
144 | count++; | |
145 | goto retry; | |
146 | } | |
472a34ca BB |
147 | } |
148 | ||
ec06701b AX |
149 | spin_unlock_irqrestore(&tq->tq_lock, *irqflags); |
150 | t = kmem_alloc(sizeof (taskq_ent_t), task_km_flags(flags)); | |
151 | spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class); | |
472a34ca BB |
152 | |
153 | if (t) { | |
154 | taskq_init_ent(t); | |
155 | tq->tq_nalloc++; | |
156 | } | |
157 | ||
10946b02 | 158 | return (t); |
bcd68186 | 159 | } |
160 | ||
82387586 | 161 | /* |
046a70c9 | 162 | * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t |
bcd68186 | 163 | * to already be removed from the free, work, or pending taskq lists. |
164 | */ | |
165 | static void | |
046a70c9 | 166 | task_free(taskq_t *tq, taskq_ent_t *t) |
bcd68186 | 167 | { |
472a34ca BB |
168 | ASSERT(tq); |
169 | ASSERT(t); | |
046a70c9 | 170 | ASSERT(list_empty(&t->tqent_list)); |
d9acd930 | 171 | ASSERT(!timer_pending(&t->tqent_timer)); |
bcd68186 | 172 | |
ec06701b | 173 | kmem_free(t, sizeof (taskq_ent_t)); |
472a34ca | 174 | tq->tq_nalloc--; |
bcd68186 | 175 | } |
176 | ||
82387586 BB |
177 | /* |
178 | * NOTE: Must be called with tq->tq_lock held, either destroys the | |
046a70c9 | 179 | * taskq_ent_t if too many exist or moves it to the free list for later use. |
bcd68186 | 180 | */ |
f1ca4da6 | 181 | static void |
046a70c9 | 182 | task_done(taskq_t *tq, taskq_ent_t *t) |
f1ca4da6 | 183 | { |
bcd68186 | 184 | ASSERT(tq); |
185 | ASSERT(t); | |
bcd68186 | 186 | |
d9acd930 BB |
187 | /* Wake tasks blocked in taskq_wait_id() */ |
188 | wake_up_all(&t->tqent_waitq); | |
189 | ||
046a70c9 | 190 | list_del_init(&t->tqent_list); |
f1ca4da6 | 191 | |
472a34ca | 192 | if (tq->tq_nalloc <= tq->tq_minalloc) { |
ec06701b | 193 | t->tqent_id = TASKQID_INVALID; |
046a70c9 PS |
194 | t->tqent_func = NULL; |
195 | t->tqent_arg = NULL; | |
44217f7a | 196 | t->tqent_flags = 0; |
8f2503e0 | 197 | |
472a34ca | 198 | list_add_tail(&t->tqent_list, &tq->tq_free_list); |
bcd68186 | 199 | } else { |
200 | task_free(tq, t); | |
201 | } | |
f1ca4da6 | 202 | } |
203 | ||
82387586 | 204 | /* |
d9acd930 BB |
205 | * When a delayed task timer expires remove it from the delay list and |
206 | * add it to the priority list in order for immediate processing. | |
bcd68186 | 207 | */ |
d9acd930 | 208 | static void |
f3bc16ff | 209 | task_expire_impl(taskq_ent_t *t) |
bcd68186 | 210 | { |
f3bc16ff | 211 | taskq_ent_t *w; |
d9acd930 BB |
212 | taskq_t *tq = t->tqent_taskq; |
213 | struct list_head *l; | |
ec06701b | 214 | unsigned long flags; |
7257ec41 | 215 | |
ec06701b | 216 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
d9acd930 BB |
217 | |
218 | if (t->tqent_flags & TQENT_FLAG_CANCEL) { | |
219 | ASSERT(list_empty(&t->tqent_list)); | |
ec06701b | 220 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 BB |
221 | return; |
222 | } | |
223 | ||
ec06701b | 224 | t->tqent_birth = jiffies; |
d9acd930 BB |
225 | /* |
226 | * The priority list must be maintained in strict task id order | |
227 | * from lowest to highest for lowest_id to be easily calculable. | |
228 | */ | |
229 | list_del(&t->tqent_list); | |
230 | list_for_each_prev(l, &tq->tq_prio_list) { | |
231 | w = list_entry(l, taskq_ent_t, tqent_list); | |
232 | if (w->tqent_id < t->tqent_id) { | |
233 | list_add(&t->tqent_list, l); | |
234 | break; | |
235 | } | |
236 | } | |
237 | if (l == &tq->tq_prio_list) | |
238 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
239 | ||
ec06701b | 240 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
7257ec41 | 241 | |
d9acd930 BB |
242 | wake_up(&tq->tq_work_waitq); |
243 | } | |
244 | ||
f3bc16ff AX |
245 | #ifdef HAVE_KERNEL_TIMER_FUNCTION_TIMER_LIST |
246 | static void | |
247 | task_expire(struct timer_list *tl) | |
248 | { | |
249 | taskq_ent_t *t = from_timer(t, tl, tqent_timer); | |
250 | task_expire_impl(t); | |
251 | } | |
252 | #else | |
253 | static void | |
254 | task_expire(unsigned long data) | |
255 | { | |
256 | task_expire_impl((taskq_ent_t *)data); | |
257 | } | |
258 | #endif | |
259 | ||
d9acd930 BB |
260 | /* |
261 | * Returns the lowest incomplete taskqid_t. The taskqid_t may | |
262 | * be queued on the pending list, on the priority list, on the | |
263 | * delay list, or on the work list currently being handled, but | |
264 | * it is not 100% complete yet. | |
265 | */ | |
266 | static taskqid_t | |
267 | taskq_lowest_id(taskq_t *tq) | |
268 | { | |
269 | taskqid_t lowest_id = tq->tq_next_id; | |
270 | taskq_ent_t *t; | |
271 | taskq_thread_t *tqt; | |
d9acd930 BB |
272 | |
273 | ASSERT(tq); | |
d9acd930 BB |
274 | |
275 | if (!list_empty(&tq->tq_pend_list)) { | |
276 | t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list); | |
277 | lowest_id = MIN(lowest_id, t->tqent_id); | |
278 | } | |
279 | ||
280 | if (!list_empty(&tq->tq_prio_list)) { | |
281 | t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list); | |
282 | lowest_id = MIN(lowest_id, t->tqent_id); | |
283 | } | |
284 | ||
285 | if (!list_empty(&tq->tq_delay_list)) { | |
286 | t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list); | |
287 | lowest_id = MIN(lowest_id, t->tqent_id); | |
288 | } | |
289 | ||
290 | if (!list_empty(&tq->tq_active_list)) { | |
291 | tqt = list_entry(tq->tq_active_list.next, taskq_thread_t, | |
292 | tqt_active_list); | |
ec06701b | 293 | ASSERT(tqt->tqt_id != TASKQID_INVALID); |
d9acd930 BB |
294 | lowest_id = MIN(lowest_id, tqt->tqt_id); |
295 | } | |
296 | ||
10946b02 | 297 | return (lowest_id); |
d9acd930 BB |
298 | } |
299 | ||
300 | /* | |
301 | * Insert a task into a list keeping the list sorted by increasing taskqid. | |
302 | */ | |
303 | static void | |
304 | taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt) | |
305 | { | |
306 | taskq_thread_t *w; | |
307 | struct list_head *l; | |
308 | ||
d9acd930 BB |
309 | ASSERT(tq); |
310 | ASSERT(tqt); | |
d9acd930 BB |
311 | |
312 | list_for_each_prev(l, &tq->tq_active_list) { | |
313 | w = list_entry(l, taskq_thread_t, tqt_active_list); | |
314 | if (w->tqt_id < tqt->tqt_id) { | |
315 | list_add(&tqt->tqt_active_list, l); | |
316 | break; | |
317 | } | |
318 | } | |
319 | if (l == &tq->tq_active_list) | |
320 | list_add(&tqt->tqt_active_list, &tq->tq_active_list); | |
d9acd930 BB |
321 | } |
322 | ||
323 | /* | |
324 | * Find and return a task from the given list if it exists. The list | |
325 | * must be in lowest to highest task id order. | |
326 | */ | |
327 | static taskq_ent_t * | |
328 | taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id) | |
329 | { | |
330 | struct list_head *l; | |
331 | taskq_ent_t *t; | |
d9acd930 | 332 | |
d9acd930 BB |
333 | list_for_each(l, lh) { |
334 | t = list_entry(l, taskq_ent_t, tqent_list); | |
335 | ||
336 | if (t->tqent_id == id) | |
10946b02 | 337 | return (t); |
d9acd930 BB |
338 | |
339 | if (t->tqent_id > id) | |
340 | break; | |
341 | } | |
342 | ||
10946b02 | 343 | return (NULL); |
bcd68186 | 344 | } |
345 | ||
d9acd930 BB |
346 | /* |
347 | * Find an already dispatched task given the task id regardless of what | |
99d6d8dc AX |
348 | * state it is in. If a task is still pending it will be returned. |
349 | * If a task is executing, then -EBUSY will be returned instead. | |
350 | * If the task has already been run then NULL is returned. | |
d9acd930 BB |
351 | */ |
352 | static taskq_ent_t * | |
99d6d8dc | 353 | taskq_find(taskq_t *tq, taskqid_t id) |
d9acd930 BB |
354 | { |
355 | taskq_thread_t *tqt; | |
356 | struct list_head *l; | |
357 | taskq_ent_t *t; | |
d9acd930 | 358 | |
d9acd930 BB |
359 | t = taskq_find_list(tq, &tq->tq_delay_list, id); |
360 | if (t) | |
10946b02 | 361 | return (t); |
d9acd930 BB |
362 | |
363 | t = taskq_find_list(tq, &tq->tq_prio_list, id); | |
364 | if (t) | |
10946b02 | 365 | return (t); |
d9acd930 BB |
366 | |
367 | t = taskq_find_list(tq, &tq->tq_pend_list, id); | |
368 | if (t) | |
10946b02 | 369 | return (t); |
d9acd930 BB |
370 | |
371 | list_for_each(l, &tq->tq_active_list) { | |
372 | tqt = list_entry(l, taskq_thread_t, tqt_active_list); | |
373 | if (tqt->tqt_id == id) { | |
99d6d8dc AX |
374 | /* |
375 | * Instead of returning tqt_task, we just return a non | |
376 | * NULL value to prevent misuse, since tqt_task only | |
377 | * has two valid fields. | |
378 | */ | |
379 | return (ERR_PTR(-EBUSY)); | |
d9acd930 BB |
380 | } |
381 | } | |
382 | ||
10946b02 | 383 | return (NULL); |
d9acd930 BB |
384 | } |
385 | ||
f6188ddd AX |
386 | /* |
387 | * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and | |
388 | * taskq_wait() functions below. | |
389 | * | |
390 | * Taskq waiting is accomplished by tracking the lowest outstanding task | |
391 | * id and the next available task id. As tasks are dispatched they are | |
392 | * added to the tail of the pending, priority, or delay lists. As worker | |
393 | * threads become available the tasks are removed from the heads of these | |
394 | * lists and linked to the worker threads. This ensures the lists are | |
395 | * kept sorted by lowest to highest task id. | |
396 | * | |
397 | * Therefore the lowest outstanding task id can be quickly determined by | |
398 | * checking the head item from all of these lists. This value is stored | |
399 | * with the taskq as the lowest id. It only needs to be recalculated when | |
400 | * either the task with the current lowest id completes or is canceled. | |
401 | * | |
402 | * By blocking until the lowest task id exceeds the passed task id the | |
403 | * taskq_wait_outstanding() function can be easily implemented. Similarly, | |
404 | * by blocking until the lowest task id matches the next task id taskq_wait() | |
405 | * can be implemented. | |
406 | * | |
407 | * Callers should be aware that when there are multiple worked threads it | |
408 | * is possible for larger task ids to complete before smaller ones. Also | |
409 | * when the taskq contains delay tasks with small task ids callers may | |
410 | * block for a considerable length of time waiting for them to expire and | |
411 | * execute. | |
412 | */ | |
80093b6f AX |
413 | static int |
414 | taskq_wait_id_check(taskq_t *tq, taskqid_t id) | |
f1ca4da6 | 415 | { |
80093b6f | 416 | int rc; |
ec06701b | 417 | unsigned long flags; |
bcd68186 | 418 | |
ec06701b | 419 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
99d6d8dc | 420 | rc = (taskq_find(tq, id) == NULL); |
ec06701b | 421 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 422 | |
80093b6f AX |
423 | return (rc); |
424 | } | |
bcd68186 | 425 | |
80093b6f AX |
426 | /* |
427 | * The taskq_wait_id() function blocks until the passed task id completes. | |
428 | * This does not guarantee that all lower task ids have completed. | |
429 | */ | |
430 | void | |
431 | taskq_wait_id(taskq_t *tq, taskqid_t id) | |
432 | { | |
433 | wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id)); | |
bcd68186 | 434 | } |
aed8671c | 435 | EXPORT_SYMBOL(taskq_wait_id); |
bcd68186 | 436 | |
d9acd930 | 437 | static int |
f6188ddd | 438 | taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id) |
d9acd930 BB |
439 | { |
440 | int rc; | |
ec06701b | 441 | unsigned long flags; |
d9acd930 | 442 | |
ec06701b | 443 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
d9acd930 | 444 | rc = (id < tq->tq_lowest_id); |
ec06701b | 445 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 446 | |
10946b02 | 447 | return (rc); |
d9acd930 BB |
448 | } |
449 | ||
f6188ddd AX |
450 | /* |
451 | * The taskq_wait_outstanding() function will block until all tasks with a | |
452 | * lower taskqid than the passed 'id' have been completed. Note that all | |
453 | * task id's are assigned monotonically at dispatch time. Zero may be | |
454 | * passed for the id to indicate all tasks dispatch up to this point, | |
455 | * but not after, should be waited for. | |
456 | */ | |
d9acd930 | 457 | void |
f6188ddd | 458 | taskq_wait_outstanding(taskq_t *tq, taskqid_t id) |
d9acd930 | 459 | { |
0f836a62 AX |
460 | id = id ? id : tq->tq_next_id - 1; |
461 | wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id)); | |
d9acd930 | 462 | } |
f6188ddd | 463 | EXPORT_SYMBOL(taskq_wait_outstanding); |
d9acd930 | 464 | |
f6188ddd AX |
465 | static int |
466 | taskq_wait_check(taskq_t *tq) | |
bcd68186 | 467 | { |
f6188ddd | 468 | int rc; |
ec06701b | 469 | unsigned long flags; |
bcd68186 | 470 | |
ec06701b | 471 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
f6188ddd | 472 | rc = (tq->tq_lowest_id == tq->tq_next_id); |
ec06701b | 473 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 474 | |
f6188ddd AX |
475 | return (rc); |
476 | } | |
477 | ||
478 | /* | |
479 | * The taskq_wait() function will block until the taskq is empty. | |
480 | * This means that if a taskq re-dispatches work to itself taskq_wait() | |
481 | * callers will block indefinitely. | |
482 | */ | |
483 | void | |
484 | taskq_wait(taskq_t *tq) | |
485 | { | |
486 | wait_event(tq->tq_wait_waitq, taskq_wait_check(tq)); | |
bcd68186 | 487 | } |
aed8671c | 488 | EXPORT_SYMBOL(taskq_wait); |
bcd68186 | 489 | |
f6188ddd | 490 | int |
ec06701b | 491 | taskq_member(taskq_t *tq, kthread_t *t) |
f6188ddd | 492 | { |
ec06701b | 493 | return (tq == (taskq_t *)tsd_get_by_thread(taskq_tsd, t)); |
bcd68186 | 494 | } |
aed8671c | 495 | EXPORT_SYMBOL(taskq_member); |
bcd68186 | 496 | |
d9acd930 BB |
497 | /* |
498 | * Cancel an already dispatched task given the task id. Still pending tasks | |
499 | * will be immediately canceled, and if the task is active the function will | |
500 | * block until it completes. Preallocated tasks which are canceled must be | |
501 | * freed by the caller. | |
502 | */ | |
503 | int | |
504 | taskq_cancel_id(taskq_t *tq, taskqid_t id) | |
505 | { | |
506 | taskq_ent_t *t; | |
d9acd930 | 507 | int rc = ENOENT; |
ec06701b | 508 | unsigned long flags; |
d9acd930 BB |
509 | |
510 | ASSERT(tq); | |
511 | ||
ec06701b | 512 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
99d6d8dc AX |
513 | t = taskq_find(tq, id); |
514 | if (t && t != ERR_PTR(-EBUSY)) { | |
d9acd930 BB |
515 | list_del_init(&t->tqent_list); |
516 | t->tqent_flags |= TQENT_FLAG_CANCEL; | |
517 | ||
518 | /* | |
519 | * When canceling the lowest outstanding task id we | |
520 | * must recalculate the new lowest outstanding id. | |
521 | */ | |
522 | if (tq->tq_lowest_id == t->tqent_id) { | |
523 | tq->tq_lowest_id = taskq_lowest_id(tq); | |
524 | ASSERT3S(tq->tq_lowest_id, >, t->tqent_id); | |
525 | } | |
526 | ||
527 | /* | |
528 | * The task_expire() function takes the tq->tq_lock so drop | |
529 | * drop the lock before synchronously cancelling the timer. | |
530 | */ | |
531 | if (timer_pending(&t->tqent_timer)) { | |
ec06701b | 532 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 533 | del_timer_sync(&t->tqent_timer); |
ec06701b AX |
534 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
535 | tq->tq_lock_class); | |
d9acd930 BB |
536 | } |
537 | ||
538 | if (!(t->tqent_flags & TQENT_FLAG_PREALLOC)) | |
539 | task_done(tq, t); | |
540 | ||
541 | rc = 0; | |
542 | } | |
ec06701b | 543 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
d9acd930 | 544 | |
99d6d8dc | 545 | if (t == ERR_PTR(-EBUSY)) { |
d9acd930 BB |
546 | taskq_wait_id(tq, id); |
547 | rc = EBUSY; | |
548 | } | |
549 | ||
10946b02 | 550 | return (rc); |
d9acd930 BB |
551 | } |
552 | EXPORT_SYMBOL(taskq_cancel_id); | |
553 | ||
879bbbc7 | 554 | static int taskq_thread_spawn(taskq_t *tq); |
f6188ddd | 555 | |
bcd68186 | 556 | taskqid_t |
aed8671c | 557 | taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags) |
bcd68186 | 558 | { |
472a34ca | 559 | taskq_ent_t *t; |
ec06701b AX |
560 | taskqid_t rc = TASKQID_INVALID; |
561 | unsigned long irqflags; | |
f1ca4da6 | 562 | |
472a34ca BB |
563 | ASSERT(tq); |
564 | ASSERT(func); | |
d05ec4b4 | 565 | |
ec06701b | 566 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class); |
f1ca4da6 | 567 | |
bcd68186 | 568 | /* Taskq being destroyed and all tasks drained */ |
f6188ddd | 569 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
10946b02 | 570 | goto out; |
f1ca4da6 | 571 | |
bcd68186 | 572 | /* Do not queue the task unless there is idle thread for it */ |
573 | ASSERT(tq->tq_nactive <= tq->tq_nthreads); | |
ec06701b AX |
574 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) { |
575 | /* Dynamic taskq may be able to spawn another thread */ | |
34ee731f AX |
576 | if (!(tq->tq_flags & TASKQ_DYNAMIC) || |
577 | taskq_thread_spawn(tq) == 0) | |
ec06701b AX |
578 | goto out; |
579 | } | |
bcd68186 | 580 | |
ec06701b | 581 | if ((t = task_alloc(tq, flags, &irqflags)) == NULL) |
10946b02 | 582 | goto out; |
f1ca4da6 | 583 | |
046a70c9 | 584 | spin_lock(&t->tqent_lock); |
f0d8bb26 | 585 | |
ec06701b AX |
586 | /* Queue to the front of the list to enforce TQ_NOQUEUE semantics */ |
587 | if (flags & TQ_NOQUEUE) | |
588 | list_add(&t->tqent_list, &tq->tq_prio_list); | |
f0d8bb26 | 589 | /* Queue to the priority list instead of the pending list */ |
ec06701b | 590 | else if (flags & TQ_FRONT) |
046a70c9 | 591 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); |
f0d8bb26 | 592 | else |
046a70c9 | 593 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); |
f0d8bb26 | 594 | |
046a70c9 | 595 | t->tqent_id = rc = tq->tq_next_id; |
bcd68186 | 596 | tq->tq_next_id++; |
472a34ca BB |
597 | t->tqent_func = func; |
598 | t->tqent_arg = arg; | |
d9acd930 | 599 | t->tqent_taskq = tq; |
f3bc16ff | 600 | #ifndef HAVE_KERNEL_TIMER_FUNCTION_TIMER_LIST |
d9acd930 | 601 | t->tqent_timer.data = 0; |
f3bc16ff | 602 | #endif |
d9acd930 BB |
603 | t->tqent_timer.function = NULL; |
604 | t->tqent_timer.expires = 0; | |
ec06701b | 605 | t->tqent_birth = jiffies; |
44217f7a PS |
606 | |
607 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
608 | ||
046a70c9 | 609 | spin_unlock(&t->tqent_lock); |
0bb43ca2 NB |
610 | |
611 | wake_up(&tq->tq_work_waitq); | |
bcd68186 | 612 | out: |
f6188ddd | 613 | /* Spawn additional taskq threads if required. */ |
ec06701b | 614 | if (!(flags & TQ_NOQUEUE) && tq->tq_nactive == tq->tq_nthreads) |
879bbbc7 | 615 | (void) taskq_thread_spawn(tq); |
f6188ddd | 616 | |
ec06701b | 617 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
10946b02 | 618 | return (rc); |
f1ca4da6 | 619 | } |
aed8671c | 620 | EXPORT_SYMBOL(taskq_dispatch); |
44217f7a | 621 | |
d9acd930 BB |
622 | taskqid_t |
623 | taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg, | |
624 | uint_t flags, clock_t expire_time) | |
625 | { | |
ec06701b | 626 | taskqid_t rc = TASKQID_INVALID; |
10946b02 | 627 | taskq_ent_t *t; |
ec06701b | 628 | unsigned long irqflags; |
d9acd930 BB |
629 | |
630 | ASSERT(tq); | |
631 | ASSERT(func); | |
632 | ||
ec06701b | 633 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class); |
d9acd930 BB |
634 | |
635 | /* Taskq being destroyed and all tasks drained */ | |
f6188ddd | 636 | if (!(tq->tq_flags & TASKQ_ACTIVE)) |
10946b02 | 637 | goto out; |
d9acd930 | 638 | |
ec06701b | 639 | if ((t = task_alloc(tq, flags, &irqflags)) == NULL) |
10946b02 | 640 | goto out; |
d9acd930 BB |
641 | |
642 | spin_lock(&t->tqent_lock); | |
643 | ||
644 | /* Queue to the delay list for subsequent execution */ | |
645 | list_add_tail(&t->tqent_list, &tq->tq_delay_list); | |
646 | ||
647 | t->tqent_id = rc = tq->tq_next_id; | |
648 | tq->tq_next_id++; | |
649 | t->tqent_func = func; | |
650 | t->tqent_arg = arg; | |
651 | t->tqent_taskq = tq; | |
f3bc16ff | 652 | #ifndef HAVE_KERNEL_TIMER_FUNCTION_TIMER_LIST |
d9acd930 | 653 | t->tqent_timer.data = (unsigned long)t; |
f3bc16ff | 654 | #endif |
d9acd930 BB |
655 | t->tqent_timer.function = task_expire; |
656 | t->tqent_timer.expires = (unsigned long)expire_time; | |
657 | add_timer(&t->tqent_timer); | |
658 | ||
659 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
660 | ||
661 | spin_unlock(&t->tqent_lock); | |
662 | out: | |
f6188ddd | 663 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
664 | if (tq->tq_nactive == tq->tq_nthreads) |
665 | (void) taskq_thread_spawn(tq); | |
ec06701b | 666 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
10946b02 | 667 | return (rc); |
d9acd930 BB |
668 | } |
669 | EXPORT_SYMBOL(taskq_dispatch_delay); | |
670 | ||
44217f7a | 671 | void |
aed8671c | 672 | taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags, |
ec06701b | 673 | taskq_ent_t *t) |
44217f7a | 674 | { |
ec06701b | 675 | unsigned long irqflags; |
44217f7a PS |
676 | ASSERT(tq); |
677 | ASSERT(func); | |
44217f7a | 678 | |
ec06701b AX |
679 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, |
680 | tq->tq_lock_class); | |
44217f7a PS |
681 | |
682 | /* Taskq being destroyed and all tasks drained */ | |
f6188ddd | 683 | if (!(tq->tq_flags & TASKQ_ACTIVE)) { |
ec06701b | 684 | t->tqent_id = TASKQID_INVALID; |
44217f7a PS |
685 | goto out; |
686 | } | |
687 | ||
ec06701b AX |
688 | if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) { |
689 | /* Dynamic taskq may be able to spawn another thread */ | |
34ee731f AX |
690 | if (!(tq->tq_flags & TASKQ_DYNAMIC) || |
691 | taskq_thread_spawn(tq) == 0) | |
ec06701b AX |
692 | goto out2; |
693 | flags |= TQ_FRONT; | |
694 | } | |
695 | ||
44217f7a PS |
696 | spin_lock(&t->tqent_lock); |
697 | ||
ec06701b AX |
698 | /* |
699 | * Make sure the entry is not on some other taskq; it is important to | |
700 | * ASSERT() under lock | |
701 | */ | |
702 | ASSERT(taskq_empty_ent(t)); | |
703 | ||
44217f7a PS |
704 | /* |
705 | * Mark it as a prealloc'd task. This is important | |
706 | * to ensure that we don't free it later. | |
707 | */ | |
708 | t->tqent_flags |= TQENT_FLAG_PREALLOC; | |
709 | ||
710 | /* Queue to the priority list instead of the pending list */ | |
711 | if (flags & TQ_FRONT) | |
712 | list_add_tail(&t->tqent_list, &tq->tq_prio_list); | |
713 | else | |
714 | list_add_tail(&t->tqent_list, &tq->tq_pend_list); | |
715 | ||
716 | t->tqent_id = tq->tq_next_id; | |
717 | tq->tq_next_id++; | |
718 | t->tqent_func = func; | |
719 | t->tqent_arg = arg; | |
d9acd930 | 720 | t->tqent_taskq = tq; |
ec06701b | 721 | t->tqent_birth = jiffies; |
44217f7a PS |
722 | |
723 | spin_unlock(&t->tqent_lock); | |
724 | ||
725 | wake_up(&tq->tq_work_waitq); | |
726 | out: | |
f6188ddd | 727 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
728 | if (tq->tq_nactive == tq->tq_nthreads) |
729 | (void) taskq_thread_spawn(tq); | |
ec06701b AX |
730 | out2: |
731 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); | |
44217f7a | 732 | } |
aed8671c | 733 | EXPORT_SYMBOL(taskq_dispatch_ent); |
44217f7a PS |
734 | |
735 | int | |
aed8671c | 736 | taskq_empty_ent(taskq_ent_t *t) |
44217f7a | 737 | { |
ec06701b | 738 | return (list_empty(&t->tqent_list)); |
44217f7a | 739 | } |
aed8671c | 740 | EXPORT_SYMBOL(taskq_empty_ent); |
44217f7a PS |
741 | |
742 | void | |
aed8671c | 743 | taskq_init_ent(taskq_ent_t *t) |
44217f7a PS |
744 | { |
745 | spin_lock_init(&t->tqent_lock); | |
d9acd930 | 746 | init_waitqueue_head(&t->tqent_waitq); |
f3bc16ff AX |
747 | #ifdef HAVE_KERNEL_TIMER_FUNCTION_TIMER_LIST |
748 | timer_setup(&t->tqent_timer, NULL, 0); | |
749 | #else | |
d9acd930 | 750 | init_timer(&t->tqent_timer); |
f3bc16ff | 751 | #endif |
44217f7a PS |
752 | INIT_LIST_HEAD(&t->tqent_list); |
753 | t->tqent_id = 0; | |
754 | t->tqent_func = NULL; | |
755 | t->tqent_arg = NULL; | |
756 | t->tqent_flags = 0; | |
d9acd930 | 757 | t->tqent_taskq = NULL; |
44217f7a | 758 | } |
aed8671c | 759 | EXPORT_SYMBOL(taskq_init_ent); |
44217f7a | 760 | |
f6188ddd AX |
761 | /* |
762 | * Return the next pending task, preference is given to tasks on the | |
763 | * priority list which were dispatched with TQ_FRONT. | |
764 | */ | |
765 | static taskq_ent_t * | |
766 | taskq_next_ent(taskq_t *tq) | |
767 | { | |
768 | struct list_head *list; | |
769 | ||
f6188ddd AX |
770 | if (!list_empty(&tq->tq_prio_list)) |
771 | list = &tq->tq_prio_list; | |
772 | else if (!list_empty(&tq->tq_pend_list)) | |
773 | list = &tq->tq_pend_list; | |
774 | else | |
775 | return (NULL); | |
776 | ||
777 | return (list_entry(list->next, taskq_ent_t, tqent_list)); | |
778 | } | |
779 | ||
780 | /* | |
781 | * Spawns a new thread for the specified taskq. | |
782 | */ | |
783 | static void | |
784 | taskq_thread_spawn_task(void *arg) | |
785 | { | |
786 | taskq_t *tq = (taskq_t *)arg; | |
ec06701b | 787 | unsigned long flags; |
f6188ddd | 788 | |
0f836a62 AX |
789 | if (taskq_thread_create(tq) == NULL) { |
790 | /* restore spawning count if failed */ | |
34ee731f AX |
791 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
792 | tq->tq_lock_class); | |
0f836a62 | 793 | tq->tq_nspawn--; |
ec06701b | 794 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
0f836a62 | 795 | } |
f6188ddd AX |
796 | } |
797 | ||
798 | /* | |
ec06701b | 799 | * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current |
f6188ddd AX |
800 | * number of threads is insufficient to handle the pending tasks. These |
801 | * new threads must be created by the dedicated dynamic_taskq to avoid | |
802 | * deadlocks between thread creation and memory reclaim. The system_taskq | |
803 | * which is also a dynamic taskq cannot be safely used for this. | |
804 | */ | |
805 | static int | |
879bbbc7 | 806 | taskq_thread_spawn(taskq_t *tq) |
f6188ddd AX |
807 | { |
808 | int spawning = 0; | |
809 | ||
810 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) | |
811 | return (0); | |
812 | ||
879bbbc7 | 813 | if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) && |
f6188ddd AX |
814 | (tq->tq_flags & TASKQ_ACTIVE)) { |
815 | spawning = (++tq->tq_nspawn); | |
816 | taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task, | |
817 | tq, TQ_NOSLEEP); | |
818 | } | |
819 | ||
820 | return (spawning); | |
821 | } | |
822 | ||
823 | /* | |
824 | * Threads in a dynamic taskq should only exit once it has been completely | |
825 | * drained and no other threads are actively servicing tasks. This prevents | |
826 | * threads from being created and destroyed more than is required. | |
827 | * | |
828 | * The first thread is the thread list is treated as the primary thread. | |
829 | * There is nothing special about the primary thread but in order to avoid | |
830 | * all the taskq pids from changing we opt to make it long running. | |
831 | */ | |
832 | static int | |
833 | taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt) | |
834 | { | |
f6188ddd AX |
835 | if (!(tq->tq_flags & TASKQ_DYNAMIC)) |
836 | return (0); | |
837 | ||
838 | if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t, | |
839 | tqt_thread_list) == tqt) | |
840 | return (0); | |
841 | ||
842 | return | |
843 | ((tq->tq_nspawn == 0) && /* No threads are being spawned */ | |
844 | (tq->tq_nactive == 0) && /* No threads are handling tasks */ | |
845 | (tq->tq_nthreads > 1) && /* More than 1 thread is running */ | |
846 | (!taskq_next_ent(tq)) && /* There are no pending tasks */ | |
ec06701b | 847 | (spl_taskq_thread_dynamic)); /* Dynamic taskqs are allowed */ |
f6188ddd AX |
848 | } |
849 | ||
bcd68186 | 850 | static int |
851 | taskq_thread(void *args) | |
852 | { | |
472a34ca BB |
853 | DECLARE_WAITQUEUE(wait, current); |
854 | sigset_t blocked; | |
2c02b71b | 855 | taskq_thread_t *tqt = args; |
472a34ca BB |
856 | taskq_t *tq; |
857 | taskq_ent_t *t; | |
f6188ddd | 858 | int seq_tasks = 0; |
ec06701b | 859 | unsigned long flags; |
99d6d8dc | 860 | taskq_ent_t dup_task = {}; |
bcd68186 | 861 | |
472a34ca | 862 | ASSERT(tqt); |
ec06701b | 863 | ASSERT(tqt->tqt_tq); |
2c02b71b | 864 | tq = tqt->tqt_tq; |
472a34ca | 865 | current->flags |= PF_NOFREEZE; |
bcd68186 | 866 | |
879bbbc7 | 867 | (void) spl_fstrans_mark(); |
f6188ddd | 868 | |
472a34ca BB |
869 | sigfillset(&blocked); |
870 | sigprocmask(SIG_BLOCK, &blocked, NULL); | |
871 | flush_signals(current); | |
bcd68186 | 872 | |
ec06701b AX |
873 | tsd_set(taskq_tsd, tq); |
874 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); | |
0f836a62 AX |
875 | /* |
876 | * If we are dynamically spawned, decrease spawning count. Note that | |
877 | * we could be created during taskq_create, in which case we shouldn't | |
878 | * do the decrement. But it's fine because taskq_create will reset | |
879 | * tq_nspawn later. | |
880 | */ | |
881 | if (tq->tq_flags & TASKQ_DYNAMIC) | |
882 | tq->tq_nspawn--; | |
f6188ddd AX |
883 | |
884 | /* Immediately exit if more threads than allowed were created. */ | |
885 | if (tq->tq_nthreads >= tq->tq_maxthreads) | |
886 | goto error; | |
887 | ||
472a34ca | 888 | tq->tq_nthreads++; |
f6188ddd | 889 | list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list); |
472a34ca BB |
890 | wake_up(&tq->tq_wait_waitq); |
891 | set_current_state(TASK_INTERRUPTIBLE); | |
bcd68186 | 892 | |
472a34ca | 893 | while (!kthread_should_stop()) { |
bcd68186 | 894 | |
f0d8bb26 NB |
895 | if (list_empty(&tq->tq_pend_list) && |
896 | list_empty(&tq->tq_prio_list)) { | |
f6188ddd AX |
897 | |
898 | if (taskq_thread_should_stop(tq, tqt)) { | |
899 | wake_up_all(&tq->tq_wait_waitq); | |
900 | break; | |
901 | } | |
902 | ||
3c6ed541 | 903 | add_wait_queue_exclusive(&tq->tq_work_waitq, &wait); |
ec06701b | 904 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
f6188ddd | 905 | |
bcd68186 | 906 | schedule(); |
f6188ddd AX |
907 | seq_tasks = 0; |
908 | ||
ec06701b AX |
909 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
910 | tq->tq_lock_class); | |
3c6ed541 | 911 | remove_wait_queue(&tq->tq_work_waitq, &wait); |
bcd68186 | 912 | } else { |
913 | __set_current_state(TASK_RUNNING); | |
914 | } | |
915 | ||
f6188ddd | 916 | if ((t = taskq_next_ent(tq)) != NULL) { |
472a34ca | 917 | list_del_init(&t->tqent_list); |
8f2503e0 | 918 | |
ec06701b | 919 | /* |
99d6d8dc AX |
920 | * A TQENT_FLAG_PREALLOC task may be reused or freed |
921 | * during the task function call. Store tqent_id and | |
922 | * tqent_flags here. | |
923 | * | |
924 | * Also use an on stack taskq_ent_t for tqt_task | |
925 | * assignment in this case. We only populate the two | |
926 | * fields used by the only user in taskq proc file. | |
ec06701b | 927 | */ |
e7e5f78e | 928 | tqt->tqt_id = t->tqent_id; |
8f2503e0 PS |
929 | tqt->tqt_flags = t->tqent_flags; |
930 | ||
99d6d8dc AX |
931 | if (t->tqent_flags & TQENT_FLAG_PREALLOC) { |
932 | dup_task.tqent_func = t->tqent_func; | |
933 | dup_task.tqent_arg = t->tqent_arg; | |
934 | t = &dup_task; | |
935 | } | |
936 | tqt->tqt_task = t; | |
937 | ||
2c02b71b | 938 | taskq_insert_in_order(tq, tqt); |
472a34ca | 939 | tq->tq_nactive++; |
ec06701b | 940 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 941 | |
942 | /* Perform the requested task */ | |
472a34ca | 943 | t->tqent_func(t->tqent_arg); |
bcd68186 | 944 | |
ec06701b AX |
945 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
946 | tq->tq_lock_class); | |
472a34ca | 947 | tq->tq_nactive--; |
2c02b71b | 948 | list_del_init(&tqt->tqt_active_list); |
d9acd930 | 949 | tqt->tqt_task = NULL; |
8f2503e0 PS |
950 | |
951 | /* For prealloc'd tasks, we don't free anything. */ | |
f6188ddd | 952 | if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC)) |
8f2503e0 | 953 | task_done(tq, t); |
bcd68186 | 954 | |
ec06701b AX |
955 | /* |
956 | * When the current lowest outstanding taskqid is | |
957 | * done calculate the new lowest outstanding id | |
958 | */ | |
e7e5f78e | 959 | if (tq->tq_lowest_id == tqt->tqt_id) { |
bcd68186 | 960 | tq->tq_lowest_id = taskq_lowest_id(tq); |
e7e5f78e | 961 | ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id); |
bcd68186 | 962 | } |
963 | ||
f6188ddd | 964 | /* Spawn additional taskq threads if required. */ |
879bbbc7 AX |
965 | if ((++seq_tasks) > spl_taskq_thread_sequential && |
966 | taskq_thread_spawn(tq)) | |
f6188ddd AX |
967 | seq_tasks = 0; |
968 | ||
ec06701b | 969 | tqt->tqt_id = TASKQID_INVALID; |
8f2503e0 | 970 | tqt->tqt_flags = 0; |
472a34ca | 971 | wake_up_all(&tq->tq_wait_waitq); |
f6188ddd AX |
972 | } else { |
973 | if (taskq_thread_should_stop(tq, tqt)) | |
974 | break; | |
bcd68186 | 975 | } |
976 | ||
977 | set_current_state(TASK_INTERRUPTIBLE); | |
978 | ||
472a34ca | 979 | } |
bcd68186 | 980 | |
981 | __set_current_state(TASK_RUNNING); | |
472a34ca | 982 | tq->tq_nthreads--; |
2c02b71b | 983 | list_del_init(&tqt->tqt_thread_list); |
f6188ddd AX |
984 | error: |
985 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
ec06701b AX |
986 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
987 | ||
988 | tsd_set(taskq_tsd, NULL); | |
bcd68186 | 989 | |
10946b02 | 990 | return (0); |
bcd68186 | 991 | } |
992 | ||
f6188ddd AX |
993 | static taskq_thread_t * |
994 | taskq_thread_create(taskq_t *tq) | |
995 | { | |
996 | static int last_used_cpu = 0; | |
997 | taskq_thread_t *tqt; | |
998 | ||
999 | tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE); | |
1000 | INIT_LIST_HEAD(&tqt->tqt_thread_list); | |
1001 | INIT_LIST_HEAD(&tqt->tqt_active_list); | |
1002 | tqt->tqt_tq = tq; | |
ec06701b | 1003 | tqt->tqt_id = TASKQID_INVALID; |
f6188ddd AX |
1004 | |
1005 | tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt, | |
1006 | "%s", tq->tq_name); | |
1007 | if (tqt->tqt_thread == NULL) { | |
1008 | kmem_free(tqt, sizeof (taskq_thread_t)); | |
1009 | return (NULL); | |
1010 | } | |
1011 | ||
1012 | if (spl_taskq_thread_bind) { | |
1013 | last_used_cpu = (last_used_cpu + 1) % num_online_cpus(); | |
1014 | kthread_bind(tqt->tqt_thread, last_used_cpu); | |
1015 | } | |
1016 | ||
1017 | if (spl_taskq_thread_priority) | |
1018 | set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri)); | |
1019 | ||
1020 | wake_up_process(tqt->tqt_thread); | |
1021 | ||
1022 | return (tqt); | |
1023 | } | |
1024 | ||
f1ca4da6 | 1025 | taskq_t * |
aed8671c | 1026 | taskq_create(const char *name, int nthreads, pri_t pri, |
472a34ca | 1027 | int minalloc, int maxalloc, uint_t flags) |
f1ca4da6 | 1028 | { |
472a34ca | 1029 | taskq_t *tq; |
2c02b71b | 1030 | taskq_thread_t *tqt; |
f6188ddd | 1031 | int count = 0, rc = 0, i; |
ec06701b | 1032 | unsigned long irqflags; |
bcd68186 | 1033 | |
472a34ca | 1034 | ASSERT(name != NULL); |
472a34ca BB |
1035 | ASSERT(minalloc >= 0); |
1036 | ASSERT(maxalloc <= INT_MAX); | |
f6188ddd | 1037 | ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */ |
bcd68186 | 1038 | |
915404bd BB |
1039 | /* Scale the number of threads using nthreads as a percentage */ |
1040 | if (flags & TASKQ_THREADS_CPU_PCT) { | |
1041 | ASSERT(nthreads <= 100); | |
1042 | ASSERT(nthreads >= 0); | |
1043 | nthreads = MIN(nthreads, 100); | |
1044 | nthreads = MAX(nthreads, 0); | |
1045 | nthreads = MAX((num_online_cpus() * nthreads) / 100, 1); | |
1046 | } | |
1047 | ||
f6188ddd | 1048 | tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE); |
472a34ca | 1049 | if (tq == NULL) |
10946b02 | 1050 | return (NULL); |
bcd68186 | 1051 | |
472a34ca | 1052 | spin_lock_init(&tq->tq_lock); |
472a34ca BB |
1053 | INIT_LIST_HEAD(&tq->tq_thread_list); |
1054 | INIT_LIST_HEAD(&tq->tq_active_list); | |
ec06701b AX |
1055 | tq->tq_name = strdup(name); |
1056 | tq->tq_nactive = 0; | |
1057 | tq->tq_nthreads = 0; | |
1058 | tq->tq_nspawn = 0; | |
f6188ddd | 1059 | tq->tq_maxthreads = nthreads; |
ec06701b AX |
1060 | tq->tq_pri = pri; |
1061 | tq->tq_minalloc = minalloc; | |
1062 | tq->tq_maxalloc = maxalloc; | |
1063 | tq->tq_nalloc = 0; | |
1064 | tq->tq_flags = (flags | TASKQ_ACTIVE); | |
1065 | tq->tq_next_id = TASKQID_INITIAL; | |
1066 | tq->tq_lowest_id = TASKQID_INITIAL; | |
472a34ca BB |
1067 | INIT_LIST_HEAD(&tq->tq_free_list); |
1068 | INIT_LIST_HEAD(&tq->tq_pend_list); | |
1069 | INIT_LIST_HEAD(&tq->tq_prio_list); | |
d9acd930 | 1070 | INIT_LIST_HEAD(&tq->tq_delay_list); |
472a34ca BB |
1071 | init_waitqueue_head(&tq->tq_work_waitq); |
1072 | init_waitqueue_head(&tq->tq_wait_waitq); | |
ec06701b AX |
1073 | tq->tq_lock_class = TQ_LOCK_GENERAL; |
1074 | INIT_LIST_HEAD(&tq->tq_taskqs); | |
bcd68186 | 1075 | |
f6188ddd | 1076 | if (flags & TASKQ_PREPOPULATE) { |
ec06701b AX |
1077 | spin_lock_irqsave_nested(&tq->tq_lock, irqflags, |
1078 | tq->tq_lock_class); | |
f6188ddd | 1079 | |
472a34ca | 1080 | for (i = 0; i < minalloc; i++) |
ec06701b AX |
1081 | task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW, |
1082 | &irqflags)); | |
6e605b6e | 1083 | |
ec06701b | 1084 | spin_unlock_irqrestore(&tq->tq_lock, irqflags); |
f6188ddd AX |
1085 | } |
1086 | ||
1087 | if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) | |
1088 | nthreads = 1; | |
6e605b6e | 1089 | |
2c02b71b | 1090 | for (i = 0; i < nthreads; i++) { |
f6188ddd AX |
1091 | tqt = taskq_thread_create(tq); |
1092 | if (tqt == NULL) | |
2c02b71b | 1093 | rc = 1; |
f6188ddd AX |
1094 | else |
1095 | count++; | |
2c02b71b | 1096 | } |
bcd68186 | 1097 | |
472a34ca | 1098 | /* Wait for all threads to be started before potential destroy */ |
f6188ddd | 1099 | wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count); |
0f836a62 AX |
1100 | /* |
1101 | * taskq_thread might have touched nspawn, but we don't want them to | |
1102 | * because they're not dynamically spawned. So we reset it to 0 | |
1103 | */ | |
1104 | tq->tq_nspawn = 0; | |
bcd68186 | 1105 | |
472a34ca | 1106 | if (rc) { |
aed8671c | 1107 | taskq_destroy(tq); |
472a34ca | 1108 | tq = NULL; |
ec06701b AX |
1109 | } else { |
1110 | down_write(&tq_list_sem); | |
1111 | tq->tq_instance = taskq_find_by_name(name) + 1; | |
1112 | list_add_tail(&tq->tq_taskqs, &tq_list); | |
1113 | up_write(&tq_list_sem); | |
472a34ca | 1114 | } |
bcd68186 | 1115 | |
10946b02 | 1116 | return (tq); |
f1ca4da6 | 1117 | } |
aed8671c | 1118 | EXPORT_SYMBOL(taskq_create); |
b123971f | 1119 | |
1120 | void | |
aed8671c | 1121 | taskq_destroy(taskq_t *tq) |
b123971f | 1122 | { |
2c02b71b PS |
1123 | struct task_struct *thread; |
1124 | taskq_thread_t *tqt; | |
046a70c9 | 1125 | taskq_ent_t *t; |
ec06701b | 1126 | unsigned long flags; |
b123971f | 1127 | |
bcd68186 | 1128 | ASSERT(tq); |
ec06701b | 1129 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); |
f6188ddd | 1130 | tq->tq_flags &= ~TASKQ_ACTIVE; |
ec06701b | 1131 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
bcd68186 | 1132 | |
f6188ddd AX |
1133 | /* |
1134 | * When TASKQ_ACTIVE is clear new tasks may not be added nor may | |
1135 | * new worker threads be spawned for dynamic taskq. | |
1136 | */ | |
1137 | if (dynamic_taskq != NULL) | |
1138 | taskq_wait_outstanding(dynamic_taskq, 0); | |
1139 | ||
aed8671c | 1140 | taskq_wait(tq); |
bcd68186 | 1141 | |
ec06701b AX |
1142 | /* remove taskq from global list used by the kstats */ |
1143 | down_write(&tq_list_sem); | |
1144 | list_del(&tq->tq_taskqs); | |
1145 | up_write(&tq_list_sem); | |
1146 | ||
1147 | spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class); | |
0f836a62 AX |
1148 | /* wait for spawning threads to insert themselves to the list */ |
1149 | while (tq->tq_nspawn) { | |
ec06701b | 1150 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
0f836a62 | 1151 | schedule_timeout_interruptible(1); |
34ee731f AX |
1152 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
1153 | tq->tq_lock_class); | |
0f836a62 | 1154 | } |
bcd68186 | 1155 | |
2c02b71b PS |
1156 | /* |
1157 | * Signal each thread to exit and block until it does. Each thread | |
1158 | * is responsible for removing itself from the list and freeing its | |
1159 | * taskq_thread_t. This allows for idle threads to opt to remove | |
1160 | * themselves from the taskq. They can be recreated as needed. | |
1161 | */ | |
1162 | while (!list_empty(&tq->tq_thread_list)) { | |
1163 | tqt = list_entry(tq->tq_thread_list.next, | |
f6188ddd | 1164 | taskq_thread_t, tqt_thread_list); |
2c02b71b | 1165 | thread = tqt->tqt_thread; |
ec06701b | 1166 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
2c02b71b PS |
1167 | |
1168 | kthread_stop(thread); | |
1169 | ||
ec06701b AX |
1170 | spin_lock_irqsave_nested(&tq->tq_lock, flags, |
1171 | tq->tq_lock_class); | |
2c02b71b PS |
1172 | } |
1173 | ||
472a34ca | 1174 | while (!list_empty(&tq->tq_free_list)) { |
046a70c9 | 1175 | t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list); |
44217f7a PS |
1176 | |
1177 | ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC)); | |
1178 | ||
472a34ca BB |
1179 | list_del_init(&t->tqent_list); |
1180 | task_free(tq, t); | |
1181 | } | |
bcd68186 | 1182 | |
f6188ddd AX |
1183 | ASSERT0(tq->tq_nthreads); |
1184 | ASSERT0(tq->tq_nalloc); | |
1185 | ASSERT0(tq->tq_nspawn); | |
472a34ca BB |
1186 | ASSERT(list_empty(&tq->tq_thread_list)); |
1187 | ASSERT(list_empty(&tq->tq_active_list)); | |
1188 | ASSERT(list_empty(&tq->tq_free_list)); | |
1189 | ASSERT(list_empty(&tq->tq_pend_list)); | |
1190 | ASSERT(list_empty(&tq->tq_prio_list)); | |
d9acd930 | 1191 | ASSERT(list_empty(&tq->tq_delay_list)); |
bcd68186 | 1192 | |
ec06701b | 1193 | spin_unlock_irqrestore(&tq->tq_lock, flags); |
2c02b71b | 1194 | |
f6188ddd AX |
1195 | strfree(tq->tq_name); |
1196 | kmem_free(tq, sizeof (taskq_t)); | |
b123971f | 1197 | } |
aed8671c | 1198 | EXPORT_SYMBOL(taskq_destroy); |
e9cb2b4f | 1199 | |
ec06701b AX |
1200 | |
1201 | static unsigned int spl_taskq_kick = 0; | |
1202 | ||
1203 | /* | |
1204 | * 2.6.36 API Change | |
1205 | * module_param_cb is introduced to take kernel_param_ops and | |
1206 | * module_param_call is marked as obsolete. Also set and get operations | |
1207 | * were changed to take a 'const struct kernel_param *'. | |
1208 | */ | |
1209 | static int | |
1210 | #ifdef module_param_cb | |
1211 | param_set_taskq_kick(const char *val, const struct kernel_param *kp) | |
1212 | #else | |
1213 | param_set_taskq_kick(const char *val, struct kernel_param *kp) | |
1214 | #endif | |
1215 | { | |
1216 | int ret; | |
1217 | taskq_t *tq; | |
1218 | taskq_ent_t *t; | |
1219 | unsigned long flags; | |
1220 | ||
1221 | ret = param_set_uint(val, kp); | |
1222 | if (ret < 0 || !spl_taskq_kick) | |
1223 | return (ret); | |
1224 | /* reset value */ | |
1225 | spl_taskq_kick = 0; | |
1226 | ||
1227 | down_read(&tq_list_sem); | |
1228 | list_for_each_entry(tq, &tq_list, tq_taskqs) { | |
1229 | spin_lock_irqsave_nested(&tq->tq_lock, flags, | |
1230 | tq->tq_lock_class); | |
1231 | /* Check if the first pending is older than 5 seconds */ | |
1232 | t = taskq_next_ent(tq); | |
1233 | if (t && time_after(jiffies, t->tqent_birth + 5*HZ)) { | |
1234 | (void) taskq_thread_spawn(tq); | |
1235 | printk(KERN_INFO "spl: Kicked taskq %s/%d\n", | |
1236 | tq->tq_name, tq->tq_instance); | |
1237 | } | |
1238 | spin_unlock_irqrestore(&tq->tq_lock, flags); | |
1239 | } | |
1240 | up_read(&tq_list_sem); | |
1241 | return (ret); | |
1242 | } | |
1243 | ||
1244 | #ifdef module_param_cb | |
1245 | static const struct kernel_param_ops param_ops_taskq_kick = { | |
34ee731f AX |
1246 | .set = param_set_taskq_kick, |
1247 | .get = param_get_uint, | |
ec06701b AX |
1248 | }; |
1249 | module_param_cb(spl_taskq_kick, ¶m_ops_taskq_kick, &spl_taskq_kick, 0644); | |
1250 | #else | |
1251 | module_param_call(spl_taskq_kick, param_set_taskq_kick, param_get_uint, | |
34ee731f | 1252 | &spl_taskq_kick, 0644); |
ec06701b AX |
1253 | #endif |
1254 | MODULE_PARM_DESC(spl_taskq_kick, | |
34ee731f | 1255 | "Write nonzero to kick stuck taskqs to spawn more threads"); |
ec06701b | 1256 | |
e9cb2b4f BB |
1257 | int |
1258 | spl_taskq_init(void) | |
1259 | { | |
ec06701b AX |
1260 | tsd_create(&taskq_tsd, NULL); |
1261 | ||
f6188ddd AX |
1262 | system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64), |
1263 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); | |
e9cb2b4f | 1264 | if (system_taskq == NULL) |
10946b02 | 1265 | return (1); |
e9cb2b4f | 1266 | |
ec06701b AX |
1267 | system_delay_taskq = taskq_create("spl_delay_taskq", MAX(boot_ncpus, 4), |
1268 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC); | |
1269 | if (system_delay_taskq == NULL) { | |
1270 | taskq_destroy(system_taskq); | |
1271 | return (1); | |
1272 | } | |
1273 | ||
f6188ddd AX |
1274 | dynamic_taskq = taskq_create("spl_dynamic_taskq", 1, |
1275 | maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE); | |
1276 | if (dynamic_taskq == NULL) { | |
1277 | taskq_destroy(system_taskq); | |
ec06701b | 1278 | taskq_destroy(system_delay_taskq); |
f6188ddd AX |
1279 | return (1); |
1280 | } | |
1281 | ||
ec06701b AX |
1282 | /* |
1283 | * This is used to annotate tq_lock, so | |
1284 | * taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch | |
1285 | * does not trigger a lockdep warning re: possible recursive locking | |
1286 | */ | |
1287 | dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC; | |
1288 | ||
10946b02 | 1289 | return (0); |
e9cb2b4f BB |
1290 | } |
1291 | ||
1292 | void | |
1293 | spl_taskq_fini(void) | |
1294 | { | |
f6188ddd AX |
1295 | taskq_destroy(dynamic_taskq); |
1296 | dynamic_taskq = NULL; | |
1297 | ||
ec06701b AX |
1298 | taskq_destroy(system_delay_taskq); |
1299 | system_delay_taskq = NULL; | |
1300 | ||
e9cb2b4f | 1301 | taskq_destroy(system_taskq); |
f6188ddd | 1302 | system_taskq = NULL; |
ec06701b AX |
1303 | |
1304 | tsd_destroy(&taskq_tsd); | |
e9cb2b4f | 1305 | } |