xref: /freebsd/sys/contrib/openzfs/lib/libtpool/thread_pool.c (revision e8d8bef961a50d4dc22501cde4fb9fb0be1b2532)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <stdlib.h>
28 #include <signal.h>
29 #include <errno.h>
30 #include <assert.h>
31 #include "thread_pool_impl.h"
32 
33 static pthread_mutex_t thread_pool_lock = PTHREAD_MUTEX_INITIALIZER;
34 static tpool_t *thread_pools = NULL;
35 
36 static void
37 delete_pool(tpool_t *tpool)
38 {
39 	tpool_job_t *job;
40 
41 	ASSERT(tpool->tp_current == 0 && tpool->tp_active == NULL);
42 
43 	/*
44 	 * Unlink the pool from the global list of all pools.
45 	 */
46 	(void) pthread_mutex_lock(&thread_pool_lock);
47 	if (thread_pools == tpool)
48 		thread_pools = tpool->tp_forw;
49 	if (thread_pools == tpool)
50 		thread_pools = NULL;
51 	else {
52 		tpool->tp_back->tp_forw = tpool->tp_forw;
53 		tpool->tp_forw->tp_back = tpool->tp_back;
54 	}
55 	pthread_mutex_unlock(&thread_pool_lock);
56 
57 	/*
58 	 * There should be no pending jobs, but just in case...
59 	 */
60 	for (job = tpool->tp_head; job != NULL; job = tpool->tp_head) {
61 		tpool->tp_head = job->tpj_next;
62 		free(job);
63 	}
64 	(void) pthread_attr_destroy(&tpool->tp_attr);
65 	free(tpool);
66 }
67 
68 /*
69  * Worker thread is terminating.
70  */
71 static void
72 worker_cleanup(void *arg)
73 {
74 	tpool_t *tpool = (tpool_t *)arg;
75 
76 	if (--tpool->tp_current == 0 &&
77 	    (tpool->tp_flags & (TP_DESTROY | TP_ABANDON))) {
78 		if (tpool->tp_flags & TP_ABANDON) {
79 			pthread_mutex_unlock(&tpool->tp_mutex);
80 			delete_pool(tpool);
81 			return;
82 		}
83 		if (tpool->tp_flags & TP_DESTROY)
84 			(void) pthread_cond_broadcast(&tpool->tp_busycv);
85 	}
86 	pthread_mutex_unlock(&tpool->tp_mutex);
87 }
88 
89 static void
90 notify_waiters(tpool_t *tpool)
91 {
92 	if (tpool->tp_head == NULL && tpool->tp_active == NULL) {
93 		tpool->tp_flags &= ~TP_WAIT;
94 		(void) pthread_cond_broadcast(&tpool->tp_waitcv);
95 	}
96 }
97 
98 /*
99  * Called by a worker thread on return from a tpool_dispatch()d job.
100  */
101 static void
102 job_cleanup(void *arg)
103 {
104 	tpool_t *tpool = (tpool_t *)arg;
105 
106 	pthread_t my_tid = pthread_self();
107 	tpool_active_t *activep;
108 	tpool_active_t **activepp;
109 
110 	pthread_mutex_lock(&tpool->tp_mutex);
111 	/* CSTYLED */
112 	for (activepp = &tpool->tp_active;; activepp = &activep->tpa_next) {
113 		activep = *activepp;
114 		if (activep->tpa_tid == my_tid) {
115 			*activepp = activep->tpa_next;
116 			break;
117 		}
118 	}
119 	if (tpool->tp_flags & TP_WAIT)
120 		notify_waiters(tpool);
121 }
122 
123 static void *
124 tpool_worker(void *arg)
125 {
126 	tpool_t *tpool = (tpool_t *)arg;
127 	int elapsed;
128 	tpool_job_t *job;
129 	void (*func)(void *);
130 	tpool_active_t active;
131 
132 	pthread_mutex_lock(&tpool->tp_mutex);
133 	pthread_cleanup_push(worker_cleanup, tpool);
134 
135 	/*
136 	 * This is the worker's main loop.
137 	 * It will only be left if a timeout or an error has occurred.
138 	 */
139 	active.tpa_tid = pthread_self();
140 	for (;;) {
141 		elapsed = 0;
142 		tpool->tp_idle++;
143 		if (tpool->tp_flags & TP_WAIT)
144 			notify_waiters(tpool);
145 		while ((tpool->tp_head == NULL ||
146 		    (tpool->tp_flags & TP_SUSPEND)) &&
147 		    !(tpool->tp_flags & (TP_DESTROY | TP_ABANDON))) {
148 			if (tpool->tp_current <= tpool->tp_minimum ||
149 			    tpool->tp_linger == 0) {
150 				(void) pthread_cond_wait(&tpool->tp_workcv,
151 				    &tpool->tp_mutex);
152 			} else {
153 				struct timespec ts;
154 
155 				clock_gettime(CLOCK_REALTIME, &ts);
156 				ts.tv_sec += tpool->tp_linger;
157 
158 				if (pthread_cond_timedwait(&tpool->tp_workcv,
159 				    &tpool->tp_mutex, &ts) != 0) {
160 					elapsed = 1;
161 					break;
162 				}
163 			}
164 		}
165 		tpool->tp_idle--;
166 		if (tpool->tp_flags & TP_DESTROY)
167 			break;
168 		if (tpool->tp_flags & TP_ABANDON) {
169 			/* can't abandon a suspended pool */
170 			if (tpool->tp_flags & TP_SUSPEND) {
171 				tpool->tp_flags &= ~TP_SUSPEND;
172 				(void) pthread_cond_broadcast(
173 				    &tpool->tp_workcv);
174 			}
175 			if (tpool->tp_head == NULL)
176 				break;
177 		}
178 		if ((job = tpool->tp_head) != NULL &&
179 		    !(tpool->tp_flags & TP_SUSPEND)) {
180 			elapsed = 0;
181 			func = job->tpj_func;
182 			arg = job->tpj_arg;
183 			tpool->tp_head = job->tpj_next;
184 			if (job == tpool->tp_tail)
185 				tpool->tp_tail = NULL;
186 			tpool->tp_njobs--;
187 			active.tpa_next = tpool->tp_active;
188 			tpool->tp_active = &active;
189 			pthread_mutex_unlock(&tpool->tp_mutex);
190 			pthread_cleanup_push(job_cleanup, tpool);
191 			free(job);
192 
193 			sigset_t maskset;
194 			(void) pthread_sigmask(SIG_SETMASK, NULL, &maskset);
195 
196 			/*
197 			 * Call the specified function.
198 			 */
199 			func(arg);
200 			/*
201 			 * We don't know what this thread has been doing,
202 			 * so we reset its signal mask and cancellation
203 			 * state back to the values prior to calling func().
204 			 */
205 			(void) pthread_sigmask(SIG_SETMASK, &maskset, NULL);
206 			(void) pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED,
207 			    NULL);
208 			(void) pthread_setcancelstate(PTHREAD_CANCEL_ENABLE,
209 			    NULL);
210 			pthread_cleanup_pop(1);
211 		}
212 		if (elapsed && tpool->tp_current > tpool->tp_minimum) {
213 			/*
214 			 * We timed out and there is no work to be done
215 			 * and the number of workers exceeds the minimum.
216 			 * Exit now to reduce the size of the pool.
217 			 */
218 			break;
219 		}
220 	}
221 	pthread_cleanup_pop(1);
222 	return (arg);
223 }
224 
225 /*
226  * Create a worker thread, with default signals blocked.
227  */
228 static int
229 create_worker(tpool_t *tpool)
230 {
231 	pthread_t thread;
232 	sigset_t oset;
233 	int error;
234 
235 	(void) pthread_sigmask(SIG_SETMASK, NULL, &oset);
236 	error = pthread_create(&thread, &tpool->tp_attr, tpool_worker, tpool);
237 	(void) pthread_sigmask(SIG_SETMASK, &oset, NULL);
238 	return (error);
239 }
240 
241 
242 /*
243  * pthread_attr_clone: make a copy of a pthread_attr_t.  When old_attr
244  * is NULL initialize the cloned attr using default values.
245  */
246 static int
247 pthread_attr_clone(pthread_attr_t *attr, const pthread_attr_t *old_attr)
248 {
249 	int error;
250 
251 	error = pthread_attr_init(attr);
252 	if (error || (old_attr == NULL))
253 		return (error);
254 
255 #ifdef __GLIBC__
256 	cpu_set_t cpuset;
257 	size_t cpusetsize = sizeof (cpuset);
258 	error = pthread_attr_getaffinity_np(old_attr, cpusetsize, &cpuset);
259 	if (error == 0)
260 		error = pthread_attr_setaffinity_np(attr, cpusetsize, &cpuset);
261 	if (error)
262 		goto error;
263 #endif /* __GLIBC__ */
264 
265 	int detachstate;
266 	error = pthread_attr_getdetachstate(old_attr, &detachstate);
267 	if (error == 0)
268 		error = pthread_attr_setdetachstate(attr, detachstate);
269 	if (error)
270 		goto error;
271 
272 	size_t guardsize;
273 	error = pthread_attr_getguardsize(old_attr, &guardsize);
274 	if (error == 0)
275 		error = pthread_attr_setguardsize(attr, guardsize);
276 	if (error)
277 		goto error;
278 
279 	int inheritsched;
280 	error = pthread_attr_getinheritsched(old_attr, &inheritsched);
281 	if (error == 0)
282 		error = pthread_attr_setinheritsched(attr, inheritsched);
283 	if (error)
284 		goto error;
285 
286 	struct sched_param param;
287 	error = pthread_attr_getschedparam(old_attr, &param);
288 	if (error == 0)
289 		error = pthread_attr_setschedparam(attr, &param);
290 	if (error)
291 		goto error;
292 
293 	int policy;
294 	error = pthread_attr_getschedpolicy(old_attr, &policy);
295 	if (error == 0)
296 		error = pthread_attr_setschedpolicy(attr, policy);
297 	if (error)
298 		goto error;
299 
300 	int scope;
301 	error = pthread_attr_getscope(old_attr, &scope);
302 	if (error == 0)
303 		error = pthread_attr_setscope(attr, scope);
304 	if (error)
305 		goto error;
306 
307 	void *stackaddr;
308 	size_t stacksize;
309 	error = pthread_attr_getstack(old_attr, &stackaddr, &stacksize);
310 	if (error == 0)
311 		error = pthread_attr_setstack(attr, stackaddr, stacksize);
312 	if (error)
313 		goto error;
314 
315 	return (0);
316 error:
317 	pthread_attr_destroy(attr);
318 	return (error);
319 }
320 
321 tpool_t	*
322 tpool_create(uint_t min_threads, uint_t max_threads, uint_t linger,
323     pthread_attr_t *attr)
324 {
325 	tpool_t	*tpool;
326 	void *stackaddr;
327 	size_t stacksize;
328 	size_t minstack;
329 	int error;
330 
331 	if (min_threads > max_threads || max_threads < 1) {
332 		errno = EINVAL;
333 		return (NULL);
334 	}
335 	if (attr != NULL) {
336 		if (pthread_attr_getstack(attr, &stackaddr, &stacksize) != 0) {
337 			errno = EINVAL;
338 			return (NULL);
339 		}
340 		/*
341 		 * Allow only one thread in the pool with a specified stack.
342 		 * Require threads to have at least the minimum stack size.
343 		 */
344 		minstack = PTHREAD_STACK_MIN;
345 		if (stackaddr != NULL) {
346 			if (stacksize < minstack || max_threads != 1) {
347 				errno = EINVAL;
348 				return (NULL);
349 			}
350 		} else if (stacksize != 0 && stacksize < minstack) {
351 			errno = EINVAL;
352 			return (NULL);
353 		}
354 	}
355 
356 	tpool = calloc(1, sizeof (*tpool));
357 	if (tpool == NULL) {
358 		errno = ENOMEM;
359 		return (NULL);
360 	}
361 	(void) pthread_mutex_init(&tpool->tp_mutex, NULL);
362 	(void) pthread_cond_init(&tpool->tp_busycv, NULL);
363 	(void) pthread_cond_init(&tpool->tp_workcv, NULL);
364 	(void) pthread_cond_init(&tpool->tp_waitcv, NULL);
365 	tpool->tp_minimum = min_threads;
366 	tpool->tp_maximum = max_threads;
367 	tpool->tp_linger = linger;
368 
369 	/*
370 	 * We cannot just copy the attribute pointer.
371 	 * We need to initialize a new pthread_attr_t structure
372 	 * with the values from the user-supplied pthread_attr_t.
373 	 * If the attribute pointer is NULL, we need to initialize
374 	 * the new pthread_attr_t structure with default values.
375 	 */
376 	error = pthread_attr_clone(&tpool->tp_attr, attr);
377 	if (error) {
378 		free(tpool);
379 		errno = error;
380 		return (NULL);
381 	}
382 
383 	/* make all pool threads be detached daemon threads */
384 	(void) pthread_attr_setdetachstate(&tpool->tp_attr,
385 	    PTHREAD_CREATE_DETACHED);
386 
387 	/* insert into the global list of all thread pools */
388 	pthread_mutex_lock(&thread_pool_lock);
389 	if (thread_pools == NULL) {
390 		tpool->tp_forw = tpool;
391 		tpool->tp_back = tpool;
392 		thread_pools = tpool;
393 	} else {
394 		thread_pools->tp_back->tp_forw = tpool;
395 		tpool->tp_forw = thread_pools;
396 		tpool->tp_back = thread_pools->tp_back;
397 		thread_pools->tp_back = tpool;
398 	}
399 	pthread_mutex_unlock(&thread_pool_lock);
400 
401 	return (tpool);
402 }
403 
404 /*
405  * Dispatch a work request to the thread pool.
406  * If there are idle workers, awaken one.
407  * Else, if the maximum number of workers has
408  * not been reached, spawn a new worker thread.
409  * Else just return with the job added to the queue.
410  */
411 int
412 tpool_dispatch(tpool_t *tpool, void (*func)(void *), void *arg)
413 {
414 	tpool_job_t *job;
415 
416 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
417 
418 	if ((job = calloc(1, sizeof (*job))) == NULL)
419 		return (-1);
420 	job->tpj_next = NULL;
421 	job->tpj_func = func;
422 	job->tpj_arg = arg;
423 
424 	pthread_mutex_lock(&tpool->tp_mutex);
425 
426 	if (tpool->tp_head == NULL)
427 		tpool->tp_head = job;
428 	else
429 		tpool->tp_tail->tpj_next = job;
430 	tpool->tp_tail = job;
431 	tpool->tp_njobs++;
432 
433 	if (!(tpool->tp_flags & TP_SUSPEND)) {
434 		if (tpool->tp_idle > 0)
435 			(void) pthread_cond_signal(&tpool->tp_workcv);
436 		else if (tpool->tp_current < tpool->tp_maximum &&
437 		    create_worker(tpool) == 0)
438 			tpool->tp_current++;
439 	}
440 
441 	pthread_mutex_unlock(&tpool->tp_mutex);
442 	return (0);
443 }
444 
445 static void
446 tpool_cleanup(void *arg)
447 {
448 	tpool_t *tpool = (tpool_t *)arg;
449 
450 	pthread_mutex_unlock(&tpool->tp_mutex);
451 }
452 
453 /*
454  * Assumes: by the time tpool_destroy() is called no one will use this
455  * thread pool in any way and no one will try to dispatch entries to it.
456  * Calling tpool_destroy() from a job in the pool will cause deadlock.
457  */
458 void
459 tpool_destroy(tpool_t *tpool)
460 {
461 	tpool_active_t *activep;
462 
463 	ASSERT(!tpool_member(tpool));
464 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
465 
466 	pthread_mutex_lock(&tpool->tp_mutex);
467 	pthread_cleanup_push(tpool_cleanup, tpool);
468 
469 	/* mark the pool as being destroyed; wakeup idle workers */
470 	tpool->tp_flags |= TP_DESTROY;
471 	tpool->tp_flags &= ~TP_SUSPEND;
472 	(void) pthread_cond_broadcast(&tpool->tp_workcv);
473 
474 	/* cancel all active workers */
475 	for (activep = tpool->tp_active; activep; activep = activep->tpa_next)
476 		(void) pthread_cancel(activep->tpa_tid);
477 
478 	/* wait for all active workers to finish */
479 	while (tpool->tp_active != NULL) {
480 		tpool->tp_flags |= TP_WAIT;
481 		(void) pthread_cond_wait(&tpool->tp_waitcv, &tpool->tp_mutex);
482 	}
483 
484 	/* the last worker to terminate will wake us up */
485 	while (tpool->tp_current != 0)
486 		(void) pthread_cond_wait(&tpool->tp_busycv, &tpool->tp_mutex);
487 
488 	pthread_cleanup_pop(1);	/* pthread_mutex_unlock(&tpool->tp_mutex); */
489 	delete_pool(tpool);
490 }
491 
492 /*
493  * Like tpool_destroy(), but don't cancel workers or wait for them to finish.
494  * The last worker to terminate will delete the pool.
495  */
496 void
497 tpool_abandon(tpool_t *tpool)
498 {
499 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
500 
501 	pthread_mutex_lock(&tpool->tp_mutex);
502 	if (tpool->tp_current == 0) {
503 		/* no workers, just delete the pool */
504 		pthread_mutex_unlock(&tpool->tp_mutex);
505 		delete_pool(tpool);
506 	} else {
507 		/* wake up all workers, last one will delete the pool */
508 		tpool->tp_flags |= TP_ABANDON;
509 		tpool->tp_flags &= ~TP_SUSPEND;
510 		(void) pthread_cond_broadcast(&tpool->tp_workcv);
511 		pthread_mutex_unlock(&tpool->tp_mutex);
512 	}
513 }
514 
515 /*
516  * Wait for all jobs to complete.
517  * Calling tpool_wait() from a job in the pool will cause deadlock.
518  */
519 void
520 tpool_wait(tpool_t *tpool)
521 {
522 	ASSERT(!tpool_member(tpool));
523 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
524 
525 	pthread_mutex_lock(&tpool->tp_mutex);
526 	pthread_cleanup_push(tpool_cleanup, tpool);
527 	while (tpool->tp_head != NULL || tpool->tp_active != NULL) {
528 		tpool->tp_flags |= TP_WAIT;
529 		(void) pthread_cond_wait(&tpool->tp_waitcv, &tpool->tp_mutex);
530 		ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
531 	}
532 	pthread_cleanup_pop(1);	/* pthread_mutex_unlock(&tpool->tp_mutex); */
533 }
534 
535 void
536 tpool_suspend(tpool_t *tpool)
537 {
538 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
539 
540 	pthread_mutex_lock(&tpool->tp_mutex);
541 	tpool->tp_flags |= TP_SUSPEND;
542 	pthread_mutex_unlock(&tpool->tp_mutex);
543 }
544 
545 int
546 tpool_suspended(tpool_t *tpool)
547 {
548 	int suspended;
549 
550 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
551 
552 	pthread_mutex_lock(&tpool->tp_mutex);
553 	suspended = (tpool->tp_flags & TP_SUSPEND) != 0;
554 	pthread_mutex_unlock(&tpool->tp_mutex);
555 
556 	return (suspended);
557 }
558 
559 void
560 tpool_resume(tpool_t *tpool)
561 {
562 	int excess;
563 
564 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
565 
566 	pthread_mutex_lock(&tpool->tp_mutex);
567 	if (!(tpool->tp_flags & TP_SUSPEND)) {
568 		pthread_mutex_unlock(&tpool->tp_mutex);
569 		return;
570 	}
571 	tpool->tp_flags &= ~TP_SUSPEND;
572 	(void) pthread_cond_broadcast(&tpool->tp_workcv);
573 	excess = tpool->tp_njobs - tpool->tp_idle;
574 	while (excess-- > 0 && tpool->tp_current < tpool->tp_maximum) {
575 		if (create_worker(tpool) != 0)
576 			break;		/* pthread_create() failed */
577 		tpool->tp_current++;
578 	}
579 	pthread_mutex_unlock(&tpool->tp_mutex);
580 }
581 
582 int
583 tpool_member(tpool_t *tpool)
584 {
585 	pthread_t my_tid = pthread_self();
586 	tpool_active_t *activep;
587 
588 	ASSERT(!(tpool->tp_flags & (TP_DESTROY | TP_ABANDON)));
589 
590 	pthread_mutex_lock(&tpool->tp_mutex);
591 	for (activep = tpool->tp_active; activep; activep = activep->tpa_next) {
592 		if (activep->tpa_tid == my_tid) {
593 			pthread_mutex_unlock(&tpool->tp_mutex);
594 			return (1);
595 		}
596 	}
597 	pthread_mutex_unlock(&tpool->tp_mutex);
598 	return (0);
599 }
600