xref: /freebsd/sys/contrib/openzfs/module/os/linux/spl/spl-taskq.c (revision 17aab35a77a1b1bf02fc85bb8ffadccb0ca5006d)
1 /*
2  *  Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3  *  Copyright (C) 2007 The Regents of the University of California.
4  *  Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5  *  Written by Brian Behlendorf <behlendorf1@llnl.gov>.
6  *  UCRL-CODE-235197
7  *
8  *  This file is part of the SPL, Solaris Porting Layer.
9  *
10  *  The SPL is free software; you can redistribute it and/or modify it
11  *  under the terms of the GNU General Public License as published by the
12  *  Free Software Foundation; either version 2 of the License, or (at your
13  *  option) any later version.
14  *
15  *  The SPL is distributed in the hope that it will be useful, but WITHOUT
16  *  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17  *  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
18  *  for more details.
19  *
20  *  You should have received a copy of the GNU General Public License along
21  *  with the SPL.  If not, see <http://www.gnu.org/licenses/>.
22  *
23  *  Solaris Porting Layer (SPL) Task Queue Implementation.
24  */
25 /*
26  * Copyright (c) 2024, Klara Inc.
27  * Copyright (c) 2024, Syneto
28  */
29 
30 #include <sys/timer.h>
31 #include <sys/taskq.h>
32 #include <sys/kmem.h>
33 #include <sys/tsd.h>
34 #include <sys/trace_spl.h>
35 #include <sys/time.h>
36 #include <sys/atomic.h>
37 #include <sys/kstat.h>
38 #include <linux/cpuhotplug.h>
39 
40 typedef struct taskq_kstats {
41 	/* static values, for completeness */
42 	kstat_named_t tqks_threads_max;
43 	kstat_named_t tqks_entry_pool_min;
44 	kstat_named_t tqks_entry_pool_max;
45 
46 	/* gauges (inc/dec counters, current value) */
47 	kstat_named_t tqks_threads_active;
48 	kstat_named_t tqks_threads_idle;
49 	kstat_named_t tqks_threads_total;
50 	kstat_named_t tqks_tasks_pending;
51 	kstat_named_t tqks_tasks_priority;
52 	kstat_named_t tqks_tasks_total;
53 	kstat_named_t tqks_tasks_delayed;
54 	kstat_named_t tqks_entries_free;
55 
56 	/* counters (inc only, since taskq creation) */
57 	kstat_named_t tqks_threads_created;
58 	kstat_named_t tqks_threads_destroyed;
59 	kstat_named_t tqks_tasks_dispatched;
60 	kstat_named_t tqks_tasks_dispatched_delayed;
61 	kstat_named_t tqks_tasks_executed_normal;
62 	kstat_named_t tqks_tasks_executed_priority;
63 	kstat_named_t tqks_tasks_executed;
64 	kstat_named_t tqks_tasks_delayed_requeued;
65 	kstat_named_t tqks_tasks_cancelled;
66 	kstat_named_t tqks_thread_wakeups;
67 	kstat_named_t tqks_thread_wakeups_nowork;
68 	kstat_named_t tqks_thread_sleeps;
69 } taskq_kstats_t;
70 
71 static taskq_kstats_t taskq_kstats_template = {
72 	{ "threads_max",		KSTAT_DATA_UINT64 },
73 	{ "entry_pool_min",		KSTAT_DATA_UINT64 },
74 	{ "entry_pool_max",		KSTAT_DATA_UINT64 },
75 	{ "threads_active",		KSTAT_DATA_UINT64 },
76 	{ "threads_idle",		KSTAT_DATA_UINT64 },
77 	{ "threads_total",		KSTAT_DATA_UINT64 },
78 	{ "tasks_pending",		KSTAT_DATA_UINT64 },
79 	{ "tasks_priority",		KSTAT_DATA_UINT64 },
80 	{ "tasks_total",		KSTAT_DATA_UINT64 },
81 	{ "tasks_delayed",		KSTAT_DATA_UINT64 },
82 	{ "entries_free",		KSTAT_DATA_UINT64 },
83 
84 	{ "threads_created",		KSTAT_DATA_UINT64 },
85 	{ "threads_destroyed",		KSTAT_DATA_UINT64 },
86 	{ "tasks_dispatched",		KSTAT_DATA_UINT64 },
87 	{ "tasks_dispatched_delayed",	KSTAT_DATA_UINT64 },
88 	{ "tasks_executed_normal",	KSTAT_DATA_UINT64 },
89 	{ "tasks_executed_priority",	KSTAT_DATA_UINT64 },
90 	{ "tasks_executed",		KSTAT_DATA_UINT64 },
91 	{ "tasks_delayed_requeued",	KSTAT_DATA_UINT64 },
92 	{ "tasks_cancelled",		KSTAT_DATA_UINT64 },
93 	{ "thread_wakeups",		KSTAT_DATA_UINT64 },
94 	{ "thread_wakeups_nowork",	KSTAT_DATA_UINT64 },
95 	{ "thread_sleeps",		KSTAT_DATA_UINT64 },
96 };
97 
98 #define	TQSTAT_INC(tq, stat)	wmsum_add(&tq->tq_sums.tqs_##stat, 1)
99 #define	TQSTAT_DEC(tq, stat)	wmsum_add(&tq->tq_sums.tqs_##stat, -1)
100 
101 #define	_TQSTAT_MOD_LIST(mod, tq, t) do { \
102 	switch (t->tqent_flags & TQENT_LIST_MASK) {			\
103 	case TQENT_LIST_NONE: ASSERT(list_empty(&t->tqent_list)); break;\
104 	case TQENT_LIST_PENDING: mod(tq, tasks_pending); break;		\
105 	case TQENT_LIST_PRIORITY: mod(tq, tasks_priority); break;	\
106 	case TQENT_LIST_DELAY: mod(tq, tasks_delayed); break;		\
107 	}								\
108 } while (0)
109 #define	TQSTAT_INC_LIST(tq, t)	_TQSTAT_MOD_LIST(TQSTAT_INC, tq, t)
110 #define	TQSTAT_DEC_LIST(tq, t)	_TQSTAT_MOD_LIST(TQSTAT_DEC, tq, t)
111 
112 #define	TQENT_SET_LIST(t, l)	\
113 	t->tqent_flags = (t->tqent_flags & ~TQENT_LIST_MASK) | l;
114 
115 static int spl_taskq_thread_bind = 0;
116 module_param(spl_taskq_thread_bind, int, 0644);
117 MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default");
118 
119 static uint_t spl_taskq_thread_timeout_ms = 5000;
120 module_param(spl_taskq_thread_timeout_ms, uint, 0644);
121 MODULE_PARM_DESC(spl_taskq_thread_timeout_ms,
122 	"Minimum idle threads exit interval for dynamic taskqs");
123 
124 static int spl_taskq_thread_dynamic = 1;
125 module_param(spl_taskq_thread_dynamic, int, 0444);
126 MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads");
127 
128 static int spl_taskq_thread_priority = 1;
129 module_param(spl_taskq_thread_priority, int, 0644);
130 MODULE_PARM_DESC(spl_taskq_thread_priority,
131 	"Allow non-default priority for taskq threads");
132 
133 static uint_t spl_taskq_thread_sequential = 4;
134 module_param(spl_taskq_thread_sequential, uint, 0644);
135 MODULE_PARM_DESC(spl_taskq_thread_sequential,
136 	"Create new taskq threads after N sequential tasks");
137 
138 /*
139  * Global system-wide dynamic task queue available for all consumers. This
140  * taskq is not intended for long-running tasks; instead, a dedicated taskq
141  * should be created.
142  */
143 taskq_t *system_taskq;
144 EXPORT_SYMBOL(system_taskq);
145 /* Global dynamic task queue for long delay */
146 taskq_t *system_delay_taskq;
147 EXPORT_SYMBOL(system_delay_taskq);
148 
149 /* Private dedicated taskq for creating new taskq threads on demand. */
150 static taskq_t *dynamic_taskq;
151 static taskq_thread_t *taskq_thread_create(taskq_t *);
152 
153 /* Multi-callback id for cpu hotplugging. */
154 static int spl_taskq_cpuhp_state;
155 
156 /* List of all taskqs */
157 LIST_HEAD(tq_list);
158 struct rw_semaphore tq_list_sem;
159 static uint_t taskq_tsd;
160 
161 static int
task_km_flags(uint_t flags)162 task_km_flags(uint_t flags)
163 {
164 	if (flags & TQ_NOSLEEP)
165 		return (KM_NOSLEEP);
166 
167 	if (flags & TQ_PUSHPAGE)
168 		return (KM_PUSHPAGE);
169 
170 	return (KM_SLEEP);
171 }
172 
173 /*
174  * taskq_find_by_name - Find the largest instance number of a named taskq.
175  */
176 static int
taskq_find_by_name(const char * name)177 taskq_find_by_name(const char *name)
178 {
179 	struct list_head *tql = NULL;
180 	taskq_t *tq;
181 
182 	list_for_each_prev(tql, &tq_list) {
183 		tq = list_entry(tql, taskq_t, tq_taskqs);
184 		if (strcmp(name, tq->tq_name) == 0)
185 			return (tq->tq_instance);
186 	}
187 	return (-1);
188 }
189 
190 /*
191  * NOTE: Must be called with tq->tq_lock held, returns a list_t which
192  * is not attached to the free, work, or pending taskq lists.
193  */
194 static taskq_ent_t *
task_alloc(taskq_t * tq,uint_t flags,unsigned long * irqflags)195 task_alloc(taskq_t *tq, uint_t flags, unsigned long *irqflags)
196 {
197 	taskq_ent_t *t;
198 	int count = 0;
199 
200 	ASSERT(tq);
201 retry:
202 	/* Acquire taskq_ent_t's from free list if available */
203 	if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
204 		t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
205 
206 		ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
207 		ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL));
208 		ASSERT(!timer_pending(&t->tqent_timer));
209 
210 		list_del_init(&t->tqent_list);
211 		TQSTAT_DEC(tq, entries_free);
212 		return (t);
213 	}
214 
215 	/* Free list is empty and memory allocations are prohibited */
216 	if (flags & TQ_NOALLOC)
217 		return (NULL);
218 
219 	/* Hit maximum taskq_ent_t pool size */
220 	if (tq->tq_nalloc >= tq->tq_maxalloc) {
221 		if (flags & TQ_NOSLEEP)
222 			return (NULL);
223 
224 		/*
225 		 * Sleep periodically polling the free list for an available
226 		 * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
227 		 * but we cannot block forever waiting for an taskq_ent_t to
228 		 * show up in the free list, otherwise a deadlock can happen.
229 		 *
230 		 * Therefore, we need to allocate a new task even if the number
231 		 * of allocated tasks is above tq->tq_maxalloc, but we still
232 		 * end up delaying the task allocation by one second, thereby
233 		 * throttling the task dispatch rate.
234 		 */
235 		spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
236 		schedule_timeout_interruptible(HZ / 100);
237 		spin_lock_irqsave_nested(&tq->tq_lock, *irqflags,
238 		    tq->tq_lock_class);
239 		if (count < 100) {
240 			count++;
241 			goto retry;
242 		}
243 	}
244 
245 	spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
246 	t = kmem_alloc(sizeof (taskq_ent_t), task_km_flags(flags));
247 	spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class);
248 
249 	if (t) {
250 		taskq_init_ent(t);
251 		tq->tq_nalloc++;
252 	}
253 
254 	return (t);
255 }
256 
257 /*
258  * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t
259  * to already be removed from the free, work, or pending taskq lists.
260  */
261 static void
task_free(taskq_t * tq,taskq_ent_t * t)262 task_free(taskq_t *tq, taskq_ent_t *t)
263 {
264 	ASSERT(tq);
265 	ASSERT(t);
266 	ASSERT(list_empty(&t->tqent_list));
267 	ASSERT(!timer_pending(&t->tqent_timer));
268 
269 	kmem_free(t, sizeof (taskq_ent_t));
270 	tq->tq_nalloc--;
271 }
272 
273 /*
274  * NOTE: Must be called with tq->tq_lock held, either destroys the
275  * taskq_ent_t if too many exist or moves it to the free list for later use.
276  */
277 static void
task_done(taskq_t * tq,taskq_ent_t * t)278 task_done(taskq_t *tq, taskq_ent_t *t)
279 {
280 	ASSERT(tq);
281 	ASSERT(t);
282 	ASSERT(list_empty(&t->tqent_list));
283 
284 	/* Wake tasks blocked in taskq_wait_id() */
285 	wake_up_all(&t->tqent_waitq);
286 
287 	if (tq->tq_nalloc <= tq->tq_minalloc) {
288 		t->tqent_id = TASKQID_INVALID;
289 		t->tqent_func = NULL;
290 		t->tqent_arg = NULL;
291 		t->tqent_flags = 0;
292 
293 		list_add_tail(&t->tqent_list, &tq->tq_free_list);
294 		TQSTAT_INC(tq, entries_free);
295 	} else {
296 		task_free(tq, t);
297 	}
298 }
299 
300 /*
301  * When a delayed task timer expires remove it from the delay list and
302  * add it to the priority list in order for immediate processing.
303  */
304 static void
task_expire_impl(taskq_ent_t * t)305 task_expire_impl(taskq_ent_t *t)
306 {
307 	taskq_ent_t *w;
308 	taskq_t *tq = t->tqent_taskq;
309 	struct list_head *l = NULL;
310 	unsigned long flags;
311 
312 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
313 
314 	if (t->tqent_flags & TQENT_FLAG_CANCEL) {
315 		ASSERT(list_empty(&t->tqent_list));
316 		spin_unlock_irqrestore(&tq->tq_lock, flags);
317 		return;
318 	}
319 
320 	t->tqent_birth = jiffies;
321 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
322 
323 	/*
324 	 * The priority list must be maintained in strict task id order
325 	 * from lowest to highest for lowest_id to be easily calculable.
326 	 */
327 	list_del(&t->tqent_list);
328 	list_for_each_prev(l, &tq->tq_prio_list) {
329 		w = list_entry(l, taskq_ent_t, tqent_list);
330 		if (w->tqent_id < t->tqent_id) {
331 			list_add(&t->tqent_list, l);
332 			break;
333 		}
334 	}
335 	if (l == &tq->tq_prio_list)
336 		list_add(&t->tqent_list, &tq->tq_prio_list);
337 
338 	spin_unlock_irqrestore(&tq->tq_lock, flags);
339 
340 	wake_up(&tq->tq_work_waitq);
341 
342 	TQSTAT_INC(tq, tasks_delayed_requeued);
343 }
344 
345 static void
task_expire(struct timer_list * tl)346 task_expire(struct timer_list *tl)
347 {
348 	struct timer_list *tmr = (struct timer_list *)tl;
349 	taskq_ent_t *t = from_timer(t, tmr, tqent_timer);
350 	task_expire_impl(t);
351 }
352 
353 /*
354  * Returns the lowest incomplete taskqid_t.  The taskqid_t may
355  * be queued on the pending list, on the priority list, on the
356  * delay list, or on the work list currently being handled, but
357  * it is not 100% complete yet.
358  */
359 static taskqid_t
taskq_lowest_id(taskq_t * tq)360 taskq_lowest_id(taskq_t *tq)
361 {
362 	taskqid_t lowest_id = tq->tq_next_id;
363 	taskq_ent_t *t;
364 	taskq_thread_t *tqt;
365 
366 	if (!list_empty(&tq->tq_pend_list)) {
367 		t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
368 		lowest_id = MIN(lowest_id, t->tqent_id);
369 	}
370 
371 	if (!list_empty(&tq->tq_prio_list)) {
372 		t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
373 		lowest_id = MIN(lowest_id, t->tqent_id);
374 	}
375 
376 	if (!list_empty(&tq->tq_delay_list)) {
377 		t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list);
378 		lowest_id = MIN(lowest_id, t->tqent_id);
379 	}
380 
381 	if (!list_empty(&tq->tq_active_list)) {
382 		tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
383 		    tqt_active_list);
384 		ASSERT(tqt->tqt_id != TASKQID_INVALID);
385 		lowest_id = MIN(lowest_id, tqt->tqt_id);
386 	}
387 
388 	return (lowest_id);
389 }
390 
391 /*
392  * Insert a task into a list keeping the list sorted by increasing taskqid.
393  */
394 static void
taskq_insert_in_order(taskq_t * tq,taskq_thread_t * tqt)395 taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
396 {
397 	taskq_thread_t *w;
398 	struct list_head *l = NULL;
399 
400 	ASSERT(tq);
401 	ASSERT(tqt);
402 
403 	list_for_each_prev(l, &tq->tq_active_list) {
404 		w = list_entry(l, taskq_thread_t, tqt_active_list);
405 		if (w->tqt_id < tqt->tqt_id) {
406 			list_add(&tqt->tqt_active_list, l);
407 			break;
408 		}
409 	}
410 	if (l == &tq->tq_active_list)
411 		list_add(&tqt->tqt_active_list, &tq->tq_active_list);
412 }
413 
414 /*
415  * Find and return a task from the given list if it exists.  The list
416  * must be in lowest to highest task id order.
417  */
418 static taskq_ent_t *
taskq_find_list(taskq_t * tq,struct list_head * lh,taskqid_t id)419 taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id)
420 {
421 	struct list_head *l = NULL;
422 	taskq_ent_t *t;
423 
424 	list_for_each(l, lh) {
425 		t = list_entry(l, taskq_ent_t, tqent_list);
426 
427 		if (t->tqent_id == id)
428 			return (t);
429 
430 		if (t->tqent_id > id)
431 			break;
432 	}
433 
434 	return (NULL);
435 }
436 
437 /*
438  * Find an already dispatched task given the task id regardless of what
439  * state it is in.  If a task is still pending it will be returned.
440  * If a task is executing, then -EBUSY will be returned instead.
441  * If the task has already been run then NULL is returned.
442  */
443 static taskq_ent_t *
taskq_find(taskq_t * tq,taskqid_t id)444 taskq_find(taskq_t *tq, taskqid_t id)
445 {
446 	taskq_thread_t *tqt;
447 	struct list_head *l = NULL;
448 	taskq_ent_t *t;
449 
450 	t = taskq_find_list(tq, &tq->tq_delay_list, id);
451 	if (t)
452 		return (t);
453 
454 	t = taskq_find_list(tq, &tq->tq_prio_list, id);
455 	if (t)
456 		return (t);
457 
458 	t = taskq_find_list(tq, &tq->tq_pend_list, id);
459 	if (t)
460 		return (t);
461 
462 	list_for_each(l, &tq->tq_active_list) {
463 		tqt = list_entry(l, taskq_thread_t, tqt_active_list);
464 		if (tqt->tqt_id == id) {
465 			/*
466 			 * Instead of returning tqt_task, we just return a non
467 			 * NULL value to prevent misuse, since tqt_task only
468 			 * has two valid fields.
469 			 */
470 			return (ERR_PTR(-EBUSY));
471 		}
472 	}
473 
474 	return (NULL);
475 }
476 
477 /*
478  * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and
479  * taskq_wait() functions below.
480  *
481  * Taskq waiting is accomplished by tracking the lowest outstanding task
482  * id and the next available task id.  As tasks are dispatched they are
483  * added to the tail of the pending, priority, or delay lists.  As worker
484  * threads become available the tasks are removed from the heads of these
485  * lists and linked to the worker threads.  This ensures the lists are
486  * kept sorted by lowest to highest task id.
487  *
488  * Therefore the lowest outstanding task id can be quickly determined by
489  * checking the head item from all of these lists.  This value is stored
490  * with the taskq as the lowest id.  It only needs to be recalculated when
491  * either the task with the current lowest id completes or is canceled.
492  *
493  * By blocking until the lowest task id exceeds the passed task id the
494  * taskq_wait_outstanding() function can be easily implemented.  Similarly,
495  * by blocking until the lowest task id matches the next task id taskq_wait()
496  * can be implemented.
497  *
498  * Callers should be aware that when there are multiple worked threads it
499  * is possible for larger task ids to complete before smaller ones.  Also
500  * when the taskq contains delay tasks with small task ids callers may
501  * block for a considerable length of time waiting for them to expire and
502  * execute.
503  */
504 static int
taskq_wait_id_check(taskq_t * tq,taskqid_t id)505 taskq_wait_id_check(taskq_t *tq, taskqid_t id)
506 {
507 	int rc;
508 	unsigned long flags;
509 
510 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
511 	rc = (taskq_find(tq, id) == NULL);
512 	spin_unlock_irqrestore(&tq->tq_lock, flags);
513 
514 	return (rc);
515 }
516 
517 /*
518  * The taskq_wait_id() function blocks until the passed task id completes.
519  * This does not guarantee that all lower task ids have completed.
520  */
521 void
taskq_wait_id(taskq_t * tq,taskqid_t id)522 taskq_wait_id(taskq_t *tq, taskqid_t id)
523 {
524 	wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id));
525 }
526 EXPORT_SYMBOL(taskq_wait_id);
527 
528 static int
taskq_wait_outstanding_check(taskq_t * tq,taskqid_t id)529 taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id)
530 {
531 	int rc;
532 	unsigned long flags;
533 
534 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
535 	rc = (id < tq->tq_lowest_id);
536 	spin_unlock_irqrestore(&tq->tq_lock, flags);
537 
538 	return (rc);
539 }
540 
541 /*
542  * The taskq_wait_outstanding() function will block until all tasks with a
543  * lower taskqid than the passed 'id' have been completed.  Note that all
544  * task id's are assigned monotonically at dispatch time.  Zero may be
545  * passed for the id to indicate all tasks dispatch up to this point,
546  * but not after, should be waited for.
547  */
548 void
taskq_wait_outstanding(taskq_t * tq,taskqid_t id)549 taskq_wait_outstanding(taskq_t *tq, taskqid_t id)
550 {
551 	id = id ? id : tq->tq_next_id - 1;
552 	wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id));
553 }
554 EXPORT_SYMBOL(taskq_wait_outstanding);
555 
556 static int
taskq_wait_check(taskq_t * tq)557 taskq_wait_check(taskq_t *tq)
558 {
559 	int rc;
560 	unsigned long flags;
561 
562 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
563 	rc = (tq->tq_lowest_id == tq->tq_next_id);
564 	spin_unlock_irqrestore(&tq->tq_lock, flags);
565 
566 	return (rc);
567 }
568 
569 /*
570  * The taskq_wait() function will block until the taskq is empty.
571  * This means that if a taskq re-dispatches work to itself taskq_wait()
572  * callers will block indefinitely.
573  */
574 void
taskq_wait(taskq_t * tq)575 taskq_wait(taskq_t *tq)
576 {
577 	wait_event(tq->tq_wait_waitq, taskq_wait_check(tq));
578 }
579 EXPORT_SYMBOL(taskq_wait);
580 
581 int
taskq_member(taskq_t * tq,kthread_t * t)582 taskq_member(taskq_t *tq, kthread_t *t)
583 {
584 	return (tq == (taskq_t *)tsd_get_by_thread(taskq_tsd, t));
585 }
586 EXPORT_SYMBOL(taskq_member);
587 
588 taskq_t *
taskq_of_curthread(void)589 taskq_of_curthread(void)
590 {
591 	return (tsd_get(taskq_tsd));
592 }
593 EXPORT_SYMBOL(taskq_of_curthread);
594 
595 /*
596  * Cancel an already dispatched task given the task id.  Still pending tasks
597  * will be immediately canceled, and if the task is active the function will
598  * block until it completes.  Preallocated tasks which are canceled must be
599  * freed by the caller.
600  */
601 int
taskq_cancel_id(taskq_t * tq,taskqid_t id)602 taskq_cancel_id(taskq_t *tq, taskqid_t id)
603 {
604 	taskq_ent_t *t;
605 	int rc = ENOENT;
606 	unsigned long flags;
607 
608 	ASSERT(tq);
609 
610 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
611 	t = taskq_find(tq, id);
612 	if (t && t != ERR_PTR(-EBUSY)) {
613 		list_del_init(&t->tqent_list);
614 		TQSTAT_DEC_LIST(tq, t);
615 		TQSTAT_DEC(tq, tasks_total);
616 
617 		t->tqent_flags |= TQENT_FLAG_CANCEL;
618 		TQSTAT_INC(tq, tasks_cancelled);
619 
620 		/*
621 		 * When canceling the lowest outstanding task id we
622 		 * must recalculate the new lowest outstanding id.
623 		 */
624 		if (tq->tq_lowest_id == t->tqent_id) {
625 			tq->tq_lowest_id = taskq_lowest_id(tq);
626 			ASSERT3S(tq->tq_lowest_id, >, t->tqent_id);
627 		}
628 
629 		/*
630 		 * The task_expire() function takes the tq->tq_lock so drop
631 		 * drop the lock before synchronously cancelling the timer.
632 		 */
633 		if (timer_pending(&t->tqent_timer)) {
634 			spin_unlock_irqrestore(&tq->tq_lock, flags);
635 			del_timer_sync(&t->tqent_timer);
636 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
637 			    tq->tq_lock_class);
638 		}
639 
640 		if (!(t->tqent_flags & TQENT_FLAG_PREALLOC))
641 			task_done(tq, t);
642 
643 		rc = 0;
644 	}
645 	spin_unlock_irqrestore(&tq->tq_lock, flags);
646 
647 	if (t == ERR_PTR(-EBUSY)) {
648 		taskq_wait_id(tq, id);
649 		rc = EBUSY;
650 	}
651 
652 	return (rc);
653 }
654 EXPORT_SYMBOL(taskq_cancel_id);
655 
656 static int taskq_thread_spawn(taskq_t *tq);
657 
658 taskqid_t
taskq_dispatch(taskq_t * tq,task_func_t func,void * arg,uint_t flags)659 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
660 {
661 	taskq_ent_t *t;
662 	taskqid_t rc = TASKQID_INVALID;
663 	unsigned long irqflags;
664 
665 	ASSERT(tq);
666 	ASSERT(func);
667 
668 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
669 
670 	/* Taskq being destroyed and all tasks drained */
671 	if (!(tq->tq_flags & TASKQ_ACTIVE))
672 		goto out;
673 
674 	/* Do not queue the task unless there is idle thread for it */
675 	ASSERT(tq->tq_nactive <= tq->tq_nthreads);
676 	if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
677 		/* Dynamic taskq may be able to spawn another thread */
678 		if (taskq_thread_spawn(tq) == 0)
679 			goto out;
680 	}
681 
682 	if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
683 		goto out;
684 
685 	spin_lock(&t->tqent_lock);
686 
687 	/* Queue to the front of the list to enforce TQ_NOQUEUE semantics */
688 	if (flags & TQ_NOQUEUE) {
689 		TQENT_SET_LIST(t, TQENT_LIST_PRIORITY);
690 		list_add(&t->tqent_list, &tq->tq_prio_list);
691 	/* Queue to the priority list instead of the pending list */
692 	} else if (flags & TQ_FRONT) {
693 		TQENT_SET_LIST(t, TQENT_LIST_PRIORITY);
694 		list_add_tail(&t->tqent_list, &tq->tq_prio_list);
695 	} else {
696 		TQENT_SET_LIST(t, TQENT_LIST_PENDING);
697 		list_add_tail(&t->tqent_list, &tq->tq_pend_list);
698 	}
699 	TQSTAT_INC_LIST(tq, t);
700 	TQSTAT_INC(tq, tasks_total);
701 
702 	t->tqent_id = rc = tq->tq_next_id;
703 	tq->tq_next_id++;
704 	t->tqent_func = func;
705 	t->tqent_arg = arg;
706 	t->tqent_taskq = tq;
707 	t->tqent_timer.function = NULL;
708 	t->tqent_timer.expires = 0;
709 
710 	t->tqent_birth = jiffies;
711 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
712 
713 	ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
714 
715 	spin_unlock(&t->tqent_lock);
716 
717 	wake_up(&tq->tq_work_waitq);
718 
719 	TQSTAT_INC(tq, tasks_dispatched);
720 
721 	/* Spawn additional taskq threads if required. */
722 	if (!(flags & TQ_NOQUEUE) && tq->tq_nactive == tq->tq_nthreads)
723 		(void) taskq_thread_spawn(tq);
724 out:
725 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
726 	return (rc);
727 }
728 EXPORT_SYMBOL(taskq_dispatch);
729 
730 taskqid_t
taskq_dispatch_delay(taskq_t * tq,task_func_t func,void * arg,uint_t flags,clock_t expire_time)731 taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
732     uint_t flags, clock_t expire_time)
733 {
734 	taskqid_t rc = TASKQID_INVALID;
735 	taskq_ent_t *t;
736 	unsigned long irqflags;
737 
738 	ASSERT(tq);
739 	ASSERT(func);
740 
741 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
742 
743 	/* Taskq being destroyed and all tasks drained */
744 	if (!(tq->tq_flags & TASKQ_ACTIVE))
745 		goto out;
746 
747 	if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
748 		goto out;
749 
750 	spin_lock(&t->tqent_lock);
751 
752 	/* Queue to the delay list for subsequent execution */
753 	list_add_tail(&t->tqent_list, &tq->tq_delay_list);
754 	TQENT_SET_LIST(t, TQENT_LIST_DELAY);
755 	TQSTAT_INC_LIST(tq, t);
756 	TQSTAT_INC(tq, tasks_total);
757 
758 	t->tqent_id = rc = tq->tq_next_id;
759 	tq->tq_next_id++;
760 	t->tqent_func = func;
761 	t->tqent_arg = arg;
762 	t->tqent_taskq = tq;
763 	t->tqent_timer.function = task_expire;
764 	t->tqent_timer.expires = (unsigned long)expire_time;
765 	add_timer(&t->tqent_timer);
766 
767 	ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
768 
769 	spin_unlock(&t->tqent_lock);
770 
771 	TQSTAT_INC(tq, tasks_dispatched_delayed);
772 
773 	/* Spawn additional taskq threads if required. */
774 	if (tq->tq_nactive == tq->tq_nthreads)
775 		(void) taskq_thread_spawn(tq);
776 out:
777 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
778 	return (rc);
779 }
780 EXPORT_SYMBOL(taskq_dispatch_delay);
781 
782 void
taskq_dispatch_ent(taskq_t * tq,task_func_t func,void * arg,uint_t flags,taskq_ent_t * t)783 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
784     taskq_ent_t *t)
785 {
786 	unsigned long irqflags;
787 	ASSERT(tq);
788 	ASSERT(func);
789 
790 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
791 	    tq->tq_lock_class);
792 
793 	/* Taskq being destroyed and all tasks drained */
794 	if (!(tq->tq_flags & TASKQ_ACTIVE)) {
795 		t->tqent_id = TASKQID_INVALID;
796 		goto out;
797 	}
798 
799 	if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
800 		/* Dynamic taskq may be able to spawn another thread */
801 		if (taskq_thread_spawn(tq) == 0)
802 			goto out;
803 		flags |= TQ_FRONT;
804 	}
805 
806 	spin_lock(&t->tqent_lock);
807 
808 	/*
809 	 * Make sure the entry is not on some other taskq; it is important to
810 	 * ASSERT() under lock
811 	 */
812 	ASSERT(taskq_empty_ent(t));
813 
814 	/*
815 	 * Mark it as a prealloc'd task.  This is important
816 	 * to ensure that we don't free it later.
817 	 */
818 	t->tqent_flags |= TQENT_FLAG_PREALLOC;
819 
820 	/* Queue to the priority list instead of the pending list */
821 	if (flags & TQ_FRONT) {
822 		TQENT_SET_LIST(t, TQENT_LIST_PRIORITY);
823 		list_add_tail(&t->tqent_list, &tq->tq_prio_list);
824 	} else {
825 		TQENT_SET_LIST(t, TQENT_LIST_PENDING);
826 		list_add_tail(&t->tqent_list, &tq->tq_pend_list);
827 	}
828 	TQSTAT_INC_LIST(tq, t);
829 	TQSTAT_INC(tq, tasks_total);
830 
831 	t->tqent_id = tq->tq_next_id;
832 	tq->tq_next_id++;
833 	t->tqent_func = func;
834 	t->tqent_arg = arg;
835 	t->tqent_taskq = tq;
836 
837 	t->tqent_birth = jiffies;
838 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
839 
840 	spin_unlock(&t->tqent_lock);
841 
842 	wake_up(&tq->tq_work_waitq);
843 
844 	TQSTAT_INC(tq, tasks_dispatched);
845 
846 	/* Spawn additional taskq threads if required. */
847 	if (tq->tq_nactive == tq->tq_nthreads)
848 		(void) taskq_thread_spawn(tq);
849 out:
850 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
851 }
852 EXPORT_SYMBOL(taskq_dispatch_ent);
853 
854 int
taskq_empty_ent(taskq_ent_t * t)855 taskq_empty_ent(taskq_ent_t *t)
856 {
857 	return (list_empty(&t->tqent_list));
858 }
859 EXPORT_SYMBOL(taskq_empty_ent);
860 
861 void
taskq_init_ent(taskq_ent_t * t)862 taskq_init_ent(taskq_ent_t *t)
863 {
864 	spin_lock_init(&t->tqent_lock);
865 	init_waitqueue_head(&t->tqent_waitq);
866 	timer_setup(&t->tqent_timer, NULL, 0);
867 	INIT_LIST_HEAD(&t->tqent_list);
868 	t->tqent_id = 0;
869 	t->tqent_func = NULL;
870 	t->tqent_arg = NULL;
871 	t->tqent_flags = 0;
872 	t->tqent_taskq = NULL;
873 }
874 EXPORT_SYMBOL(taskq_init_ent);
875 
876 /*
877  * Return the next pending task, preference is given to tasks on the
878  * priority list which were dispatched with TQ_FRONT.
879  */
880 static taskq_ent_t *
taskq_next_ent(taskq_t * tq)881 taskq_next_ent(taskq_t *tq)
882 {
883 	struct list_head *list;
884 
885 	if (!list_empty(&tq->tq_prio_list))
886 		list = &tq->tq_prio_list;
887 	else if (!list_empty(&tq->tq_pend_list))
888 		list = &tq->tq_pend_list;
889 	else
890 		return (NULL);
891 
892 	return (list_entry(list->next, taskq_ent_t, tqent_list));
893 }
894 
895 /*
896  * Spawns a new thread for the specified taskq.
897  */
898 static void
taskq_thread_spawn_task(void * arg)899 taskq_thread_spawn_task(void *arg)
900 {
901 	taskq_t *tq = (taskq_t *)arg;
902 	unsigned long flags;
903 
904 	if (taskq_thread_create(tq) == NULL) {
905 		/* restore spawning count if failed */
906 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
907 		    tq->tq_lock_class);
908 		tq->tq_nspawn--;
909 		spin_unlock_irqrestore(&tq->tq_lock, flags);
910 	}
911 }
912 
913 /*
914  * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current
915  * number of threads is insufficient to handle the pending tasks.  These
916  * new threads must be created by the dedicated dynamic_taskq to avoid
917  * deadlocks between thread creation and memory reclaim.  The system_taskq
918  * which is also a dynamic taskq cannot be safely used for this.
919  */
920 static int
taskq_thread_spawn(taskq_t * tq)921 taskq_thread_spawn(taskq_t *tq)
922 {
923 	int spawning = 0;
924 
925 	if (!(tq->tq_flags & TASKQ_DYNAMIC))
926 		return (0);
927 
928 	tq->lastspawnstop = jiffies;
929 	if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) &&
930 	    (tq->tq_flags & TASKQ_ACTIVE)) {
931 		spawning = (++tq->tq_nspawn);
932 		taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task,
933 		    tq, TQ_NOSLEEP);
934 	}
935 
936 	return (spawning);
937 }
938 
939 /*
940  * Threads in a dynamic taskq may exit once there is no more work to do.
941  * To prevent threads from being created and destroyed too often limit
942  * the exit rate to one per spl_taskq_thread_timeout_ms.
943  *
944  * The first thread is the thread list is treated as the primary thread.
945  * There is nothing special about the primary thread but in order to avoid
946  * all the taskq pids from changing we opt to make it long running.
947  */
948 static int
taskq_thread_should_stop(taskq_t * tq,taskq_thread_t * tqt)949 taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt)
950 {
951 	ASSERT(!taskq_next_ent(tq));
952 	if (!(tq->tq_flags & TASKQ_DYNAMIC) || !spl_taskq_thread_dynamic)
953 		return (0);
954 	if (!(tq->tq_flags & TASKQ_ACTIVE))
955 		return (1);
956 	if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t,
957 	    tqt_thread_list) == tqt)
958 		return (0);
959 	ASSERT3U(tq->tq_nthreads, >, 1);
960 	if (tq->tq_nspawn != 0)
961 		return (0);
962 	if (time_before(jiffies, tq->lastspawnstop +
963 	    msecs_to_jiffies(spl_taskq_thread_timeout_ms)))
964 		return (0);
965 	tq->lastspawnstop = jiffies;
966 	return (1);
967 }
968 
969 static int
taskq_thread(void * args)970 taskq_thread(void *args)
971 {
972 	DECLARE_WAITQUEUE(wait, current);
973 	sigset_t blocked;
974 	taskq_thread_t *tqt = args;
975 	taskq_t *tq;
976 	taskq_ent_t *t;
977 	int seq_tasks = 0;
978 	unsigned long flags;
979 	taskq_ent_t dup_task = {};
980 
981 	ASSERT(tqt);
982 	ASSERT(tqt->tqt_tq);
983 	tq = tqt->tqt_tq;
984 	current->flags |= PF_NOFREEZE;
985 
986 	(void) spl_fstrans_mark();
987 
988 	sigfillset(&blocked);
989 	sigprocmask(SIG_BLOCK, &blocked, NULL);
990 	flush_signals(current);
991 
992 	tsd_set(taskq_tsd, tq);
993 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
994 	/*
995 	 * If we are dynamically spawned, decrease spawning count. Note that
996 	 * we could be created during taskq_create, in which case we shouldn't
997 	 * do the decrement. But it's fine because taskq_create will reset
998 	 * tq_nspawn later.
999 	 */
1000 	if (tq->tq_flags & TASKQ_DYNAMIC)
1001 		tq->tq_nspawn--;
1002 
1003 	/* Immediately exit if more threads than allowed were created. */
1004 	if (tq->tq_nthreads >= tq->tq_maxthreads)
1005 		goto error;
1006 
1007 	tq->tq_nthreads++;
1008 	list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list);
1009 	wake_up(&tq->tq_wait_waitq);
1010 	set_current_state(TASK_INTERRUPTIBLE);
1011 
1012 	TQSTAT_INC(tq, threads_total);
1013 
1014 	while (!kthread_should_stop()) {
1015 
1016 		if (list_empty(&tq->tq_pend_list) &&
1017 		    list_empty(&tq->tq_prio_list)) {
1018 
1019 			if (taskq_thread_should_stop(tq, tqt))
1020 				break;
1021 
1022 			add_wait_queue_exclusive(&tq->tq_work_waitq, &wait);
1023 			spin_unlock_irqrestore(&tq->tq_lock, flags);
1024 
1025 			TQSTAT_INC(tq, thread_sleeps);
1026 			TQSTAT_INC(tq, threads_idle);
1027 
1028 			schedule();
1029 			seq_tasks = 0;
1030 
1031 			TQSTAT_DEC(tq, threads_idle);
1032 			TQSTAT_INC(tq, thread_wakeups);
1033 
1034 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
1035 			    tq->tq_lock_class);
1036 			remove_wait_queue(&tq->tq_work_waitq, &wait);
1037 		} else {
1038 			__set_current_state(TASK_RUNNING);
1039 		}
1040 
1041 		if ((t = taskq_next_ent(tq)) != NULL) {
1042 			list_del_init(&t->tqent_list);
1043 			TQSTAT_DEC_LIST(tq, t);
1044 			TQSTAT_DEC(tq, tasks_total);
1045 
1046 			/*
1047 			 * A TQENT_FLAG_PREALLOC task may be reused or freed
1048 			 * during the task function call. Store tqent_id and
1049 			 * tqent_flags here.
1050 			 *
1051 			 * Also use an on stack taskq_ent_t for tqt_task
1052 			 * assignment in this case; we want to make sure
1053 			 * to duplicate all fields, so the values are
1054 			 * correct when it's accessed via DTRACE_PROBE*.
1055 			 */
1056 			tqt->tqt_id = t->tqent_id;
1057 			tqt->tqt_flags = t->tqent_flags;
1058 
1059 			if (t->tqent_flags & TQENT_FLAG_PREALLOC) {
1060 				dup_task = *t;
1061 				t = &dup_task;
1062 			}
1063 			tqt->tqt_task = t;
1064 
1065 			taskq_insert_in_order(tq, tqt);
1066 			tq->tq_nactive++;
1067 			spin_unlock_irqrestore(&tq->tq_lock, flags);
1068 
1069 			TQSTAT_INC(tq, threads_active);
1070 			DTRACE_PROBE1(taskq_ent__start, taskq_ent_t *, t);
1071 
1072 			/* Perform the requested task */
1073 			t->tqent_func(t->tqent_arg);
1074 
1075 			DTRACE_PROBE1(taskq_ent__finish, taskq_ent_t *, t);
1076 
1077 			TQSTAT_DEC(tq, threads_active);
1078 			if ((t->tqent_flags & TQENT_LIST_MASK) ==
1079 			    TQENT_LIST_PENDING)
1080 				TQSTAT_INC(tq, tasks_executed_normal);
1081 			else
1082 				TQSTAT_INC(tq, tasks_executed_priority);
1083 			TQSTAT_INC(tq, tasks_executed);
1084 
1085 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
1086 			    tq->tq_lock_class);
1087 
1088 			tq->tq_nactive--;
1089 			list_del_init(&tqt->tqt_active_list);
1090 			tqt->tqt_task = NULL;
1091 
1092 			/* For prealloc'd tasks, we don't free anything. */
1093 			if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
1094 				task_done(tq, t);
1095 
1096 			/*
1097 			 * When the current lowest outstanding taskqid is
1098 			 * done calculate the new lowest outstanding id
1099 			 */
1100 			if (tq->tq_lowest_id == tqt->tqt_id) {
1101 				tq->tq_lowest_id = taskq_lowest_id(tq);
1102 				ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
1103 			}
1104 
1105 			/* Spawn additional taskq threads if required. */
1106 			if ((++seq_tasks) > spl_taskq_thread_sequential &&
1107 			    taskq_thread_spawn(tq))
1108 				seq_tasks = 0;
1109 
1110 			tqt->tqt_id = TASKQID_INVALID;
1111 			tqt->tqt_flags = 0;
1112 			wake_up_all(&tq->tq_wait_waitq);
1113 		} else
1114 			TQSTAT_INC(tq, thread_wakeups_nowork);
1115 
1116 		set_current_state(TASK_INTERRUPTIBLE);
1117 
1118 	}
1119 
1120 	__set_current_state(TASK_RUNNING);
1121 	tq->tq_nthreads--;
1122 	list_del_init(&tqt->tqt_thread_list);
1123 
1124 	TQSTAT_DEC(tq, threads_total);
1125 	TQSTAT_INC(tq, threads_destroyed);
1126 
1127 error:
1128 	kmem_free(tqt, sizeof (taskq_thread_t));
1129 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1130 
1131 	tsd_set(taskq_tsd, NULL);
1132 	thread_exit();
1133 
1134 	return (0);
1135 }
1136 
1137 static taskq_thread_t *
taskq_thread_create(taskq_t * tq)1138 taskq_thread_create(taskq_t *tq)
1139 {
1140 	static int last_used_cpu = 0;
1141 	taskq_thread_t *tqt;
1142 
1143 	tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE);
1144 	INIT_LIST_HEAD(&tqt->tqt_thread_list);
1145 	INIT_LIST_HEAD(&tqt->tqt_active_list);
1146 	tqt->tqt_tq = tq;
1147 	tqt->tqt_id = TASKQID_INVALID;
1148 
1149 	tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt,
1150 	    "%s", tq->tq_name);
1151 	if (tqt->tqt_thread == NULL) {
1152 		kmem_free(tqt, sizeof (taskq_thread_t));
1153 		return (NULL);
1154 	}
1155 
1156 	if (spl_taskq_thread_bind) {
1157 		last_used_cpu = (last_used_cpu + 1) % num_online_cpus();
1158 		kthread_bind(tqt->tqt_thread, last_used_cpu);
1159 	}
1160 
1161 	if (spl_taskq_thread_priority)
1162 		set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri));
1163 
1164 	wake_up_process(tqt->tqt_thread);
1165 
1166 	TQSTAT_INC(tq, threads_created);
1167 
1168 	return (tqt);
1169 }
1170 
1171 static void
taskq_stats_init(taskq_t * tq)1172 taskq_stats_init(taskq_t *tq)
1173 {
1174 	taskq_sums_t *tqs = &tq->tq_sums;
1175 	wmsum_init(&tqs->tqs_threads_active, 0);
1176 	wmsum_init(&tqs->tqs_threads_idle, 0);
1177 	wmsum_init(&tqs->tqs_threads_total, 0);
1178 	wmsum_init(&tqs->tqs_tasks_pending, 0);
1179 	wmsum_init(&tqs->tqs_tasks_priority, 0);
1180 	wmsum_init(&tqs->tqs_tasks_total, 0);
1181 	wmsum_init(&tqs->tqs_tasks_delayed, 0);
1182 	wmsum_init(&tqs->tqs_entries_free, 0);
1183 	wmsum_init(&tqs->tqs_threads_created, 0);
1184 	wmsum_init(&tqs->tqs_threads_destroyed, 0);
1185 	wmsum_init(&tqs->tqs_tasks_dispatched, 0);
1186 	wmsum_init(&tqs->tqs_tasks_dispatched_delayed, 0);
1187 	wmsum_init(&tqs->tqs_tasks_executed_normal, 0);
1188 	wmsum_init(&tqs->tqs_tasks_executed_priority, 0);
1189 	wmsum_init(&tqs->tqs_tasks_executed, 0);
1190 	wmsum_init(&tqs->tqs_tasks_delayed_requeued, 0);
1191 	wmsum_init(&tqs->tqs_tasks_cancelled, 0);
1192 	wmsum_init(&tqs->tqs_thread_wakeups, 0);
1193 	wmsum_init(&tqs->tqs_thread_wakeups_nowork, 0);
1194 	wmsum_init(&tqs->tqs_thread_sleeps, 0);
1195 }
1196 
1197 static void
taskq_stats_fini(taskq_t * tq)1198 taskq_stats_fini(taskq_t *tq)
1199 {
1200 	taskq_sums_t *tqs = &tq->tq_sums;
1201 	wmsum_fini(&tqs->tqs_threads_active);
1202 	wmsum_fini(&tqs->tqs_threads_idle);
1203 	wmsum_fini(&tqs->tqs_threads_total);
1204 	wmsum_fini(&tqs->tqs_tasks_pending);
1205 	wmsum_fini(&tqs->tqs_tasks_priority);
1206 	wmsum_fini(&tqs->tqs_tasks_total);
1207 	wmsum_fini(&tqs->tqs_tasks_delayed);
1208 	wmsum_fini(&tqs->tqs_entries_free);
1209 	wmsum_fini(&tqs->tqs_threads_created);
1210 	wmsum_fini(&tqs->tqs_threads_destroyed);
1211 	wmsum_fini(&tqs->tqs_tasks_dispatched);
1212 	wmsum_fini(&tqs->tqs_tasks_dispatched_delayed);
1213 	wmsum_fini(&tqs->tqs_tasks_executed_normal);
1214 	wmsum_fini(&tqs->tqs_tasks_executed_priority);
1215 	wmsum_fini(&tqs->tqs_tasks_executed);
1216 	wmsum_fini(&tqs->tqs_tasks_delayed_requeued);
1217 	wmsum_fini(&tqs->tqs_tasks_cancelled);
1218 	wmsum_fini(&tqs->tqs_thread_wakeups);
1219 	wmsum_fini(&tqs->tqs_thread_wakeups_nowork);
1220 	wmsum_fini(&tqs->tqs_thread_sleeps);
1221 }
1222 
1223 static int
taskq_kstats_update(kstat_t * ksp,int rw)1224 taskq_kstats_update(kstat_t *ksp, int rw)
1225 {
1226 	if (rw == KSTAT_WRITE)
1227 		return (EACCES);
1228 
1229 	taskq_t *tq = ksp->ks_private;
1230 	taskq_kstats_t *tqks = ksp->ks_data;
1231 
1232 	tqks->tqks_threads_max.value.ui64 = tq->tq_maxthreads;
1233 	tqks->tqks_entry_pool_min.value.ui64 = tq->tq_minalloc;
1234 	tqks->tqks_entry_pool_max.value.ui64 = tq->tq_maxalloc;
1235 
1236 	taskq_sums_t *tqs = &tq->tq_sums;
1237 
1238 	tqks->tqks_threads_active.value.ui64 =
1239 	    wmsum_value(&tqs->tqs_threads_active);
1240 	tqks->tqks_threads_idle.value.ui64 =
1241 	    wmsum_value(&tqs->tqs_threads_idle);
1242 	tqks->tqks_threads_total.value.ui64 =
1243 	    wmsum_value(&tqs->tqs_threads_total);
1244 	tqks->tqks_tasks_pending.value.ui64 =
1245 	    wmsum_value(&tqs->tqs_tasks_pending);
1246 	tqks->tqks_tasks_priority.value.ui64 =
1247 	    wmsum_value(&tqs->tqs_tasks_priority);
1248 	tqks->tqks_tasks_total.value.ui64 =
1249 	    wmsum_value(&tqs->tqs_tasks_total);
1250 	tqks->tqks_tasks_delayed.value.ui64 =
1251 	    wmsum_value(&tqs->tqs_tasks_delayed);
1252 	tqks->tqks_entries_free.value.ui64 =
1253 	    wmsum_value(&tqs->tqs_entries_free);
1254 	tqks->tqks_threads_created.value.ui64 =
1255 	    wmsum_value(&tqs->tqs_threads_created);
1256 	tqks->tqks_threads_destroyed.value.ui64 =
1257 	    wmsum_value(&tqs->tqs_threads_destroyed);
1258 	tqks->tqks_tasks_dispatched.value.ui64 =
1259 	    wmsum_value(&tqs->tqs_tasks_dispatched);
1260 	tqks->tqks_tasks_dispatched_delayed.value.ui64 =
1261 	    wmsum_value(&tqs->tqs_tasks_dispatched_delayed);
1262 	tqks->tqks_tasks_executed_normal.value.ui64 =
1263 	    wmsum_value(&tqs->tqs_tasks_executed_normal);
1264 	tqks->tqks_tasks_executed_priority.value.ui64 =
1265 	    wmsum_value(&tqs->tqs_tasks_executed_priority);
1266 	tqks->tqks_tasks_executed.value.ui64 =
1267 	    wmsum_value(&tqs->tqs_tasks_executed);
1268 	tqks->tqks_tasks_delayed_requeued.value.ui64 =
1269 	    wmsum_value(&tqs->tqs_tasks_delayed_requeued);
1270 	tqks->tqks_tasks_cancelled.value.ui64 =
1271 	    wmsum_value(&tqs->tqs_tasks_cancelled);
1272 	tqks->tqks_thread_wakeups.value.ui64 =
1273 	    wmsum_value(&tqs->tqs_thread_wakeups);
1274 	tqks->tqks_thread_wakeups_nowork.value.ui64 =
1275 	    wmsum_value(&tqs->tqs_thread_wakeups_nowork);
1276 	tqks->tqks_thread_sleeps.value.ui64 =
1277 	    wmsum_value(&tqs->tqs_thread_sleeps);
1278 
1279 	return (0);
1280 }
1281 
1282 static void
taskq_kstats_init(taskq_t * tq)1283 taskq_kstats_init(taskq_t *tq)
1284 {
1285 	char name[TASKQ_NAMELEN+5]; /* 5 for dot, 3x instance digits, null */
1286 	snprintf(name, sizeof (name), "%s.%d", tq->tq_name, tq->tq_instance);
1287 
1288 	kstat_t *ksp = kstat_create("taskq", 0, name, "misc",
1289 	    KSTAT_TYPE_NAMED, sizeof (taskq_kstats_t) / sizeof (kstat_named_t),
1290 	    KSTAT_FLAG_VIRTUAL);
1291 
1292 	if (ksp == NULL)
1293 		return;
1294 
1295 	ksp->ks_private = tq;
1296 	ksp->ks_update = taskq_kstats_update;
1297 	ksp->ks_data = kmem_alloc(sizeof (taskq_kstats_t), KM_SLEEP);
1298 	memcpy(ksp->ks_data, &taskq_kstats_template, sizeof (taskq_kstats_t));
1299 	kstat_install(ksp);
1300 
1301 	tq->tq_ksp = ksp;
1302 }
1303 
1304 static void
taskq_kstats_fini(taskq_t * tq)1305 taskq_kstats_fini(taskq_t *tq)
1306 {
1307 	if (tq->tq_ksp == NULL)
1308 		return;
1309 
1310 	kmem_free(tq->tq_ksp->ks_data, sizeof (taskq_kstats_t));
1311 	kstat_delete(tq->tq_ksp);
1312 
1313 	tq->tq_ksp = NULL;
1314 }
1315 
1316 taskq_t *
taskq_create(const char * name,int threads_arg,pri_t pri,int minalloc,int maxalloc,uint_t flags)1317 taskq_create(const char *name, int threads_arg, pri_t pri,
1318     int minalloc, int maxalloc, uint_t flags)
1319 {
1320 	taskq_t *tq;
1321 	taskq_thread_t *tqt;
1322 	int count = 0, rc = 0, i;
1323 	unsigned long irqflags;
1324 	int nthreads = threads_arg;
1325 
1326 	ASSERT(name != NULL);
1327 	ASSERT(minalloc >= 0);
1328 	ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */
1329 
1330 	/* Scale the number of threads using nthreads as a percentage */
1331 	if (flags & TASKQ_THREADS_CPU_PCT) {
1332 		ASSERT(nthreads <= 100);
1333 		ASSERT(nthreads >= 0);
1334 		nthreads = MIN(threads_arg, 100);
1335 		nthreads = MAX(nthreads, 0);
1336 		nthreads = MAX((num_online_cpus() * nthreads) /100, 1);
1337 	}
1338 
1339 	tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE);
1340 	if (tq == NULL)
1341 		return (NULL);
1342 
1343 	tq->tq_hp_support = B_FALSE;
1344 
1345 	if (flags & TASKQ_THREADS_CPU_PCT) {
1346 		tq->tq_hp_support = B_TRUE;
1347 		if (cpuhp_state_add_instance_nocalls(spl_taskq_cpuhp_state,
1348 		    &tq->tq_hp_cb_node) != 0) {
1349 			kmem_free(tq, sizeof (*tq));
1350 			return (NULL);
1351 		}
1352 	}
1353 
1354 	spin_lock_init(&tq->tq_lock);
1355 	INIT_LIST_HEAD(&tq->tq_thread_list);
1356 	INIT_LIST_HEAD(&tq->tq_active_list);
1357 	tq->tq_name = kmem_strdup(name);
1358 	tq->tq_nactive = 0;
1359 	tq->tq_nthreads = 0;
1360 	tq->tq_nspawn = 0;
1361 	tq->tq_maxthreads = nthreads;
1362 	tq->tq_cpu_pct = threads_arg;
1363 	tq->tq_pri = pri;
1364 	tq->tq_minalloc = minalloc;
1365 	tq->tq_maxalloc = maxalloc;
1366 	tq->tq_nalloc = 0;
1367 	tq->tq_flags = (flags | TASKQ_ACTIVE);
1368 	tq->tq_next_id = TASKQID_INITIAL;
1369 	tq->tq_lowest_id = TASKQID_INITIAL;
1370 	tq->lastspawnstop = jiffies;
1371 	INIT_LIST_HEAD(&tq->tq_free_list);
1372 	INIT_LIST_HEAD(&tq->tq_pend_list);
1373 	INIT_LIST_HEAD(&tq->tq_prio_list);
1374 	INIT_LIST_HEAD(&tq->tq_delay_list);
1375 	init_waitqueue_head(&tq->tq_work_waitq);
1376 	init_waitqueue_head(&tq->tq_wait_waitq);
1377 	tq->tq_lock_class = TQ_LOCK_GENERAL;
1378 	INIT_LIST_HEAD(&tq->tq_taskqs);
1379 	taskq_stats_init(tq);
1380 
1381 	if (flags & TASKQ_PREPOPULATE) {
1382 		spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
1383 		    tq->tq_lock_class);
1384 
1385 		for (i = 0; i < minalloc; i++)
1386 			task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW,
1387 			    &irqflags));
1388 
1389 		spin_unlock_irqrestore(&tq->tq_lock, irqflags);
1390 	}
1391 
1392 	if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic)
1393 		nthreads = 1;
1394 
1395 	for (i = 0; i < nthreads; i++) {
1396 		tqt = taskq_thread_create(tq);
1397 		if (tqt == NULL)
1398 			rc = 1;
1399 		else
1400 			count++;
1401 	}
1402 
1403 	/* Wait for all threads to be started before potential destroy */
1404 	wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count);
1405 	/*
1406 	 * taskq_thread might have touched nspawn, but we don't want them to
1407 	 * because they're not dynamically spawned. So we reset it to 0
1408 	 */
1409 	tq->tq_nspawn = 0;
1410 
1411 	if (rc) {
1412 		taskq_destroy(tq);
1413 		return (NULL);
1414 	}
1415 
1416 	down_write(&tq_list_sem);
1417 	tq->tq_instance = taskq_find_by_name(name) + 1;
1418 	list_add_tail(&tq->tq_taskqs, &tq_list);
1419 	up_write(&tq_list_sem);
1420 
1421 	/* Install kstats late, because the name includes tq_instance */
1422 	taskq_kstats_init(tq);
1423 
1424 	return (tq);
1425 }
1426 EXPORT_SYMBOL(taskq_create);
1427 
1428 void
taskq_destroy(taskq_t * tq)1429 taskq_destroy(taskq_t *tq)
1430 {
1431 	struct task_struct *thread;
1432 	taskq_thread_t *tqt;
1433 	taskq_ent_t *t;
1434 	unsigned long flags;
1435 
1436 	ASSERT(tq);
1437 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1438 	tq->tq_flags &= ~TASKQ_ACTIVE;
1439 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1440 
1441 	if (tq->tq_hp_support) {
1442 		VERIFY0(cpuhp_state_remove_instance_nocalls(
1443 		    spl_taskq_cpuhp_state, &tq->tq_hp_cb_node));
1444 	}
1445 
1446 	/*
1447 	 * When TASKQ_ACTIVE is clear new tasks may not be added nor may
1448 	 * new worker threads be spawned for dynamic taskq.
1449 	 */
1450 	if (dynamic_taskq != NULL)
1451 		taskq_wait_outstanding(dynamic_taskq, 0);
1452 
1453 	taskq_wait(tq);
1454 
1455 	taskq_kstats_fini(tq);
1456 
1457 	/* remove taskq from global list used by the kstats */
1458 	down_write(&tq_list_sem);
1459 	list_del(&tq->tq_taskqs);
1460 	up_write(&tq_list_sem);
1461 
1462 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1463 	/* wait for spawning threads to insert themselves to the list */
1464 	while (tq->tq_nspawn) {
1465 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1466 		schedule_timeout_interruptible(1);
1467 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1468 		    tq->tq_lock_class);
1469 	}
1470 
1471 	/*
1472 	 * Signal each thread to exit and block until it does.  Each thread
1473 	 * is responsible for removing itself from the list and freeing its
1474 	 * taskq_thread_t.  This allows for idle threads to opt to remove
1475 	 * themselves from the taskq.  They can be recreated as needed.
1476 	 */
1477 	while (!list_empty(&tq->tq_thread_list)) {
1478 		tqt = list_entry(tq->tq_thread_list.next,
1479 		    taskq_thread_t, tqt_thread_list);
1480 		thread = tqt->tqt_thread;
1481 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1482 
1483 		kthread_stop(thread);
1484 
1485 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1486 		    tq->tq_lock_class);
1487 	}
1488 
1489 	while (!list_empty(&tq->tq_free_list)) {
1490 		t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
1491 
1492 		ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
1493 
1494 		list_del_init(&t->tqent_list);
1495 		task_free(tq, t);
1496 	}
1497 
1498 	ASSERT0(tq->tq_nthreads);
1499 	ASSERT0(tq->tq_nalloc);
1500 	ASSERT0(tq->tq_nspawn);
1501 	ASSERT(list_empty(&tq->tq_thread_list));
1502 	ASSERT(list_empty(&tq->tq_active_list));
1503 	ASSERT(list_empty(&tq->tq_free_list));
1504 	ASSERT(list_empty(&tq->tq_pend_list));
1505 	ASSERT(list_empty(&tq->tq_prio_list));
1506 	ASSERT(list_empty(&tq->tq_delay_list));
1507 
1508 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1509 
1510 	taskq_stats_fini(tq);
1511 	kmem_strfree(tq->tq_name);
1512 	kmem_free(tq, sizeof (taskq_t));
1513 }
1514 EXPORT_SYMBOL(taskq_destroy);
1515 
1516 /*
1517  * Create a taskq with a specified number of pool threads. Allocate
1518  * and return an array of nthreads kthread_t pointers, one for each
1519  * thread in the pool. The array is not ordered and must be freed
1520  * by the caller.
1521  */
1522 taskq_t *
taskq_create_synced(const char * name,int nthreads,pri_t pri,int minalloc,int maxalloc,uint_t flags,kthread_t *** ktpp)1523 taskq_create_synced(const char *name, int nthreads, pri_t pri,
1524     int minalloc, int maxalloc, uint_t flags, kthread_t ***ktpp)
1525 {
1526 	taskq_t *tq;
1527 	taskq_thread_t *tqt;
1528 	int i = 0;
1529 	kthread_t **kthreads = kmem_zalloc(sizeof (*kthreads) * nthreads,
1530 	    KM_SLEEP);
1531 
1532 	flags &= ~(TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT | TASKQ_DC_BATCH);
1533 
1534 	/* taskq_create spawns all the threads before returning */
1535 	tq = taskq_create(name, nthreads, minclsyspri, nthreads, INT_MAX,
1536 	    flags | TASKQ_PREPOPULATE);
1537 	VERIFY(tq != NULL);
1538 	VERIFY(tq->tq_nthreads == nthreads);
1539 
1540 	list_for_each_entry(tqt, &tq->tq_thread_list, tqt_thread_list) {
1541 		kthreads[i] = tqt->tqt_thread;
1542 		i++;
1543 	}
1544 
1545 	ASSERT3S(i, ==, nthreads);
1546 	*ktpp = kthreads;
1547 
1548 	return (tq);
1549 }
1550 EXPORT_SYMBOL(taskq_create_synced);
1551 
1552 static kstat_t *taskq_summary_ksp = NULL;
1553 
1554 static int
spl_taskq_kstat_headers(char * buf,size_t size)1555 spl_taskq_kstat_headers(char *buf, size_t size)
1556 {
1557 	size_t n = snprintf(buf, size,
1558 	    "%-20s | %-17s | %-23s\n"
1559 	    "%-20s | %-17s | %-23s\n"
1560 	    "%-20s | %-17s | %-23s\n",
1561 	    "", "threads", "tasks on queue",
1562 	    "taskq name", "tot [act idl] max", " pend [ norm  high] dly",
1563 	    "--------------------", "-----------------",
1564 	    "-----------------------");
1565 	return (n >= size ? ENOMEM : 0);
1566 }
1567 
1568 static int
spl_taskq_kstat_data(char * buf,size_t size,void * data)1569 spl_taskq_kstat_data(char *buf, size_t size, void *data)
1570 {
1571 	struct list_head *tql = NULL;
1572 	taskq_t *tq;
1573 	char name[TASKQ_NAMELEN+5]; /* 5 for dot, 3x instance digits, null */
1574 	char threads[25];
1575 	char tasks[30];
1576 	size_t n;
1577 	int err = 0;
1578 
1579 	down_read(&tq_list_sem);
1580 	list_for_each_prev(tql, &tq_list) {
1581 		tq = list_entry(tql, taskq_t, tq_taskqs);
1582 
1583 		mutex_enter(tq->tq_ksp->ks_lock);
1584 		taskq_kstats_update(tq->tq_ksp, KSTAT_READ);
1585 		taskq_kstats_t *tqks = tq->tq_ksp->ks_data;
1586 
1587 		snprintf(name, sizeof (name), "%s.%d", tq->tq_name,
1588 		    tq->tq_instance);
1589 		snprintf(threads, sizeof (threads), "%3llu [%3llu %3llu] %3llu",
1590 		    tqks->tqks_threads_total.value.ui64,
1591 		    tqks->tqks_threads_active.value.ui64,
1592 		    tqks->tqks_threads_idle.value.ui64,
1593 		    tqks->tqks_threads_max.value.ui64);
1594 		snprintf(tasks, sizeof (tasks), "%5llu [%5llu %5llu] %3llu",
1595 		    tqks->tqks_tasks_total.value.ui64,
1596 		    tqks->tqks_tasks_pending.value.ui64,
1597 		    tqks->tqks_tasks_priority.value.ui64,
1598 		    tqks->tqks_tasks_delayed.value.ui64);
1599 
1600 		mutex_exit(tq->tq_ksp->ks_lock);
1601 
1602 		n = snprintf(buf, size, "%-20s | %-17s | %-23s\n",
1603 		    name, threads, tasks);
1604 		if (n >= size) {
1605 			err = ENOMEM;
1606 			break;
1607 		}
1608 
1609 		buf = &buf[n];
1610 		size -= n;
1611 	}
1612 
1613 	up_read(&tq_list_sem);
1614 
1615 	return (err);
1616 }
1617 
1618 static void
spl_taskq_kstat_init(void)1619 spl_taskq_kstat_init(void)
1620 {
1621 	kstat_t *ksp = kstat_create("taskq", 0, "summary", "misc",
1622 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
1623 
1624 	if (ksp == NULL)
1625 		return;
1626 
1627 	ksp->ks_data = (void *)(uintptr_t)1;
1628 	ksp->ks_ndata = 1;
1629 	kstat_set_raw_ops(ksp, spl_taskq_kstat_headers,
1630 	    spl_taskq_kstat_data, NULL);
1631 	kstat_install(ksp);
1632 
1633 	taskq_summary_ksp = ksp;
1634 }
1635 
1636 static void
spl_taskq_kstat_fini(void)1637 spl_taskq_kstat_fini(void)
1638 {
1639 	if (taskq_summary_ksp == NULL)
1640 		return;
1641 
1642 	kstat_delete(taskq_summary_ksp);
1643 	taskq_summary_ksp = NULL;
1644 }
1645 
1646 static unsigned int spl_taskq_kick = 0;
1647 
1648 /*
1649  * 2.6.36 API Change
1650  * module_param_cb is introduced to take kernel_param_ops and
1651  * module_param_call is marked as obsolete. Also set and get operations
1652  * were changed to take a 'const struct kernel_param *'.
1653  */
1654 static int
1655 #ifdef module_param_cb
param_set_taskq_kick(const char * val,const struct kernel_param * kp)1656 param_set_taskq_kick(const char *val, const struct kernel_param *kp)
1657 #else
1658 param_set_taskq_kick(const char *val, struct kernel_param *kp)
1659 #endif
1660 {
1661 	int ret;
1662 	taskq_t *tq = NULL;
1663 	taskq_ent_t *t;
1664 	unsigned long flags;
1665 
1666 	ret = param_set_uint(val, kp);
1667 	if (ret < 0 || !spl_taskq_kick)
1668 		return (ret);
1669 	/* reset value */
1670 	spl_taskq_kick = 0;
1671 
1672 	down_read(&tq_list_sem);
1673 	list_for_each_entry(tq, &tq_list, tq_taskqs) {
1674 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1675 		    tq->tq_lock_class);
1676 		/* Check if the first pending is older than 5 seconds */
1677 		t = taskq_next_ent(tq);
1678 		if (t && time_after(jiffies, t->tqent_birth + 5*HZ)) {
1679 			(void) taskq_thread_spawn(tq);
1680 			printk(KERN_INFO "spl: Kicked taskq %s/%d\n",
1681 			    tq->tq_name, tq->tq_instance);
1682 		}
1683 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1684 	}
1685 	up_read(&tq_list_sem);
1686 	return (ret);
1687 }
1688 
1689 #ifdef module_param_cb
1690 static const struct kernel_param_ops param_ops_taskq_kick = {
1691 	.set = param_set_taskq_kick,
1692 	.get = param_get_uint,
1693 };
1694 module_param_cb(spl_taskq_kick, &param_ops_taskq_kick, &spl_taskq_kick, 0644);
1695 #else
1696 module_param_call(spl_taskq_kick, param_set_taskq_kick, param_get_uint,
1697 	&spl_taskq_kick, 0644);
1698 #endif
1699 MODULE_PARM_DESC(spl_taskq_kick,
1700 	"Write nonzero to kick stuck taskqs to spawn more threads");
1701 
1702 /*
1703  * This callback will be called exactly once for each core that comes online,
1704  * for each dynamic taskq. We attempt to expand taskqs that have
1705  * TASKQ_THREADS_CPU_PCT set. We need to redo the percentage calculation every
1706  * time, to correctly determine whether or not to add a thread.
1707  */
1708 static int
spl_taskq_expand(unsigned int cpu,struct hlist_node * node)1709 spl_taskq_expand(unsigned int cpu, struct hlist_node *node)
1710 {
1711 	taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1712 	unsigned long flags;
1713 	int err = 0;
1714 
1715 	ASSERT(tq);
1716 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1717 
1718 	if (!(tq->tq_flags & TASKQ_ACTIVE)) {
1719 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1720 		return (err);
1721 	}
1722 
1723 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1724 	int nthreads = MIN(tq->tq_cpu_pct, 100);
1725 	nthreads = MAX(((num_online_cpus() + 1) * nthreads) / 100, 1);
1726 	tq->tq_maxthreads = nthreads;
1727 
1728 	if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1729 	    tq->tq_maxthreads > tq->tq_nthreads) {
1730 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1731 		taskq_thread_t *tqt = taskq_thread_create(tq);
1732 		if (tqt == NULL)
1733 			err = -1;
1734 		return (err);
1735 	}
1736 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1737 	return (err);
1738 }
1739 
1740 /*
1741  * While we don't support offlining CPUs, it is possible that CPUs will fail
1742  * to online successfully. We do need to be able to handle this case
1743  * gracefully.
1744  */
1745 static int
spl_taskq_prepare_down(unsigned int cpu,struct hlist_node * node)1746 spl_taskq_prepare_down(unsigned int cpu, struct hlist_node *node)
1747 {
1748 	taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1749 	unsigned long flags;
1750 
1751 	ASSERT(tq);
1752 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1753 
1754 	if (!(tq->tq_flags & TASKQ_ACTIVE))
1755 		goto out;
1756 
1757 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1758 	int nthreads = MIN(tq->tq_cpu_pct, 100);
1759 	nthreads = MAX(((num_online_cpus()) * nthreads) / 100, 1);
1760 	tq->tq_maxthreads = nthreads;
1761 
1762 	if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1763 	    tq->tq_maxthreads < tq->tq_nthreads) {
1764 		ASSERT3U(tq->tq_maxthreads, ==, tq->tq_nthreads - 1);
1765 		taskq_thread_t *tqt = list_entry(tq->tq_thread_list.next,
1766 		    taskq_thread_t, tqt_thread_list);
1767 		struct task_struct *thread = tqt->tqt_thread;
1768 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1769 
1770 		kthread_stop(thread);
1771 
1772 		return (0);
1773 	}
1774 
1775 out:
1776 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1777 	return (0);
1778 }
1779 
1780 int
spl_taskq_init(void)1781 spl_taskq_init(void)
1782 {
1783 	init_rwsem(&tq_list_sem);
1784 	tsd_create(&taskq_tsd, NULL);
1785 
1786 	spl_taskq_cpuhp_state = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1787 	    "fs/spl_taskq:online", spl_taskq_expand, spl_taskq_prepare_down);
1788 
1789 	system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64),
1790 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1791 	if (system_taskq == NULL)
1792 		return (-ENOMEM);
1793 
1794 	system_delay_taskq = taskq_create("spl_delay_taskq", MAX(boot_ncpus, 4),
1795 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1796 	if (system_delay_taskq == NULL) {
1797 		cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1798 		taskq_destroy(system_taskq);
1799 		return (-ENOMEM);
1800 	}
1801 
1802 	dynamic_taskq = taskq_create("spl_dynamic_taskq", 1,
1803 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE);
1804 	if (dynamic_taskq == NULL) {
1805 		cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1806 		taskq_destroy(system_taskq);
1807 		taskq_destroy(system_delay_taskq);
1808 		return (-ENOMEM);
1809 	}
1810 
1811 	/*
1812 	 * This is used to annotate tq_lock, so
1813 	 *   taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch
1814 	 * does not trigger a lockdep warning re: possible recursive locking
1815 	 */
1816 	dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC;
1817 
1818 	spl_taskq_kstat_init();
1819 
1820 	return (0);
1821 }
1822 
1823 void
spl_taskq_fini(void)1824 spl_taskq_fini(void)
1825 {
1826 	spl_taskq_kstat_fini();
1827 
1828 	taskq_destroy(dynamic_taskq);
1829 	dynamic_taskq = NULL;
1830 
1831 	taskq_destroy(system_delay_taskq);
1832 	system_delay_taskq = NULL;
1833 
1834 	taskq_destroy(system_taskq);
1835 	system_taskq = NULL;
1836 
1837 	tsd_destroy(&taskq_tsd);
1838 
1839 	cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1840 	spl_taskq_cpuhp_state = 0;
1841 }
1842