xref: /freebsd/sys/contrib/openzfs/module/zfs/zthr.c (revision da5137abdf463bb5fee85061958a14dd12bc043e)
1 /*
2  * CDDL HEADER START
3  *
4  * This file and its contents are supplied under the terms of the
5  * Common Development and Distribution License ("CDDL"), version 1.0.
6  * You may only use this file in accordance with the terms of version
7  * 1.0 of the CDDL.
8  *
9  * A full copy of the text of the CDDL should have accompanied this
10  * source. A copy of the CDDL is also available via the Internet at
11  * http://www.illumos.org/license/CDDL.
12  *
13  * CDDL HEADER END
14  */
15 
16 /*
17  * Copyright (c) 2017, 2020 by Delphix. All rights reserved.
18  */
19 
20 /*
21  * ZTHR Infrastructure
22  * ===================
23  *
24  * ZTHR threads are used for isolated operations that span multiple txgs
25  * within a SPA. They generally exist from SPA creation/loading and until
26  * the SPA is exported/destroyed. The ideal requirements for an operation
27  * to be modeled with a zthr are the following:
28  *
29  * 1] The operation needs to run over multiple txgs.
30  * 2] There is be a single point of reference in memory or on disk that
31  *    indicates whether the operation should run/is running or has
32  *    stopped.
33  *
34  * If the operation satisfies the above then the following rules guarantee
35  * a certain level of correctness:
36  *
37  * 1] Any thread EXCEPT the zthr changes the work indicator from stopped
38  *    to running but not the opposite.
39  * 2] Only the zthr can change the work indicator from running to stopped
40  *    (e.g. when it is done) but not the opposite.
41  *
42  * This way a normal zthr cycle should go like this:
43  *
44  * 1] An external thread changes the work indicator from stopped to
45  *    running and wakes up the zthr.
46  * 2] The zthr wakes up, checks the indicator and starts working.
47  * 3] When the zthr is done, it changes the indicator to stopped, allowing
48  *    a new cycle to start.
49  *
50  * Besides being awakened by other threads, a zthr can be configured
51  * during creation to wakeup on its own after a specified interval
52  * [see zthr_create_timer()].
53  *
54  * Note: ZTHR threads are NOT a replacement for generic threads! Please
55  * ensure that they fit your use-case well before using them.
56  *
57  * == ZTHR creation
58  *
59  * Every zthr needs four inputs to start running:
60  *
61  * 1] A user-defined checker function (checkfunc) that decides whether
62  *    the zthr should start working or go to sleep. The function should
63  *    return TRUE when the zthr needs to work or FALSE to let it sleep,
64  *    and should adhere to the following signature:
65  *    boolean_t checkfunc_name(void *args, zthr_t *t);
66  *
67  * 2] A user-defined ZTHR function (func) which the zthr executes when
68  *    it is not sleeping. The function should adhere to the following
69  *    signature type:
70  *    void func_name(void *args, zthr_t *t);
71  *
72  * 3] A void args pointer that will be passed to checkfunc and func
73  *    implicitly by the infrastructure.
74  *
75  * 4] A name for the thread. This string must be valid for the lifetime
76  *    of the zthr.
77  *
78  * The reason why the above API needs two different functions,
79  * instead of one that both checks and does the work, has to do with
80  * the zthr's internal state lock (zthr_state_lock) and the allowed
81  * cancellation windows. We want to hold the zthr_state_lock while
82  * running checkfunc but not while running func. This way the zthr
83  * can be cancelled while doing work and not while checking for work.
84  *
85  * To start a zthr:
86  *     zthr_t *zthr_pointer = zthr_create(checkfunc, func, args,
87  *         pri);
88  * or
89  *     zthr_t *zthr_pointer = zthr_create_timer(checkfunc, func,
90  *         args, max_sleep, pri);
91  *
92  * After that you should be able to wakeup, cancel, and resume the
93  * zthr from another thread using the zthr_pointer.
94  *
95  * NOTE: ZTHR threads could potentially wake up spuriously and the
96  * user should take this into account when writing a checkfunc.
97  * [see ZTHR state transitions]
98  *
99  * == ZTHR wakeup
100  *
101  * ZTHR wakeup should be used when new work is added for the zthr. The
102  * sleeping zthr will wakeup, see that it has more work to complete
103  * and proceed. This can be invoked from open or syncing context.
104  *
105  * To wakeup a zthr:
106  *     zthr_wakeup(zthr_t *t)
107  *
108  * == ZTHR cancellation and resumption
109  *
110  * ZTHR threads must be cancelled when their SPA is being exported
111  * or when they need to be paused so they don't interfere with other
112  * operations.
113  *
114  * To cancel a zthr:
115  *     zthr_cancel(zthr_pointer);
116  *
117  * To resume it:
118  *     zthr_resume(zthr_pointer);
119  *
120  * ZTHR cancel and resume should be invoked in open context during the
121  * lifecycle of the pool as it is imported, exported or destroyed.
122  *
123  * A zthr will implicitly check if it has received a cancellation
124  * signal every time func returns and every time it wakes up [see
125  * ZTHR state transitions below].
126  *
127  * At times, waiting for the zthr's func to finish its job may take
128  * time. This may be very time-consuming for some operations that
129  * need to cancel the SPA's zthrs (e.g spa_export). For this scenario
130  * the user can explicitly make their ZTHR function aware of incoming
131  * cancellation signals using zthr_iscancelled(). A common pattern for
132  * that looks like this:
133  *
134  * int
135  * func_name(void *args, zthr_t *t)
136  * {
137  *     ... <unpack args> ...
138  *     while (!work_done && !zthr_iscancelled(t)) {
139  *         ... <do more work> ...
140  *     }
141  * }
142  *
143  * == ZTHR cleanup
144  *
145  * Cancelling a zthr doesn't clean up its metadata (internal locks,
146  * function pointers to func and checkfunc, etc..). This is because
147  * we want to keep them around in case we want to resume the execution
148  * of the zthr later. Similarly for zthrs that exit themselves.
149  *
150  * To completely cleanup a zthr, cancel it first to ensure that it
151  * is not running and then use zthr_destroy().
152  *
153  * == ZTHR state transitions
154  *
155  *    zthr creation
156  *      +
157  *      |
158  *      |      woke up
159  *      |   +--------------+ sleep
160  *      |   |                  ^
161  *      |   |                  |
162  *      |   |                  | FALSE
163  *      |   |                  |
164  *      v   v     FALSE        +
165  *   cancelled? +---------> checkfunc?
166  *      +   ^                  +
167  *      |   |                  |
168  *      |   |                  | TRUE
169  *      |   |                  |
170  *      |   |  func returned   v
171  *      |   +---------------+ func
172  *      |
173  *      | TRUE
174  *      |
175  *      v
176  *   zthr stopped running
177  *
178  * == Implementation of ZTHR requests
179  *
180  * ZTHR cancel and resume are requests on a zthr to change its
181  * internal state. These requests are serialized using the
182  * zthr_request_lock, while changes in its internal state are
183  * protected by the zthr_state_lock. A request will first acquire
184  * the zthr_request_lock and then immediately acquire the
185  * zthr_state_lock. We do this so that incoming requests are
186  * serialized using the request lock, while still allowing us
187  * to use the state lock for thread communication via zthr_cv.
188  *
189  * ZTHR wakeup broadcasts to zthr_cv, causing sleeping threads
190  * to wakeup. It acquires the zthr_state_lock but not the
191  * zthr_request_lock, so that a wakeup on a zthr in the middle
192  * of being cancelled will not block.
193  */
194 
195 #include <sys/zfs_context.h>
196 #include <sys/zthr.h>
197 
198 struct zthr {
199 	/* running thread doing the work */
200 	kthread_t	*zthr_thread;
201 
202 	/* lock protecting internal data & invariants */
203 	kmutex_t	zthr_state_lock;
204 
205 	/* mutex that serializes external requests */
206 	kmutex_t	zthr_request_lock;
207 
208 	/* notification mechanism for requests */
209 	kcondvar_t	zthr_cv;
210 
211 	/* flag set to true if we are canceling the zthr */
212 	boolean_t	zthr_cancel;
213 
214 	/* flag set to true if we are waiting for the zthr to finish */
215 	boolean_t	zthr_haswaiters;
216 	kcondvar_t	zthr_wait_cv;
217 	/*
218 	 * maximum amount of time that the zthr is spent sleeping;
219 	 * if this is 0, the thread doesn't wake up until it gets
220 	 * signaled.
221 	 */
222 	hrtime_t	zthr_sleep_timeout;
223 
224 	/* Thread priority */
225 	pri_t		zthr_pri;
226 
227 	/* consumer-provided callbacks & data */
228 	zthr_checkfunc_t	*zthr_checkfunc;
229 	zthr_func_t	*zthr_func;
230 	void		*zthr_arg;
231 	const char	*zthr_name;
232 };
233 
234 static __attribute__((noreturn)) void
zthr_procedure(void * arg)235 zthr_procedure(void *arg)
236 {
237 	zthr_t *t = arg;
238 
239 	mutex_enter(&t->zthr_state_lock);
240 	ASSERT3P(t->zthr_thread, ==, curthread);
241 
242 	while (!t->zthr_cancel) {
243 		if (t->zthr_checkfunc(t->zthr_arg, t)) {
244 			mutex_exit(&t->zthr_state_lock);
245 			t->zthr_func(t->zthr_arg, t);
246 			mutex_enter(&t->zthr_state_lock);
247 		} else {
248 			if (t->zthr_sleep_timeout == 0) {
249 				cv_wait_idle(&t->zthr_cv, &t->zthr_state_lock);
250 			} else {
251 				(void) cv_timedwait_idle_hires(&t->zthr_cv,
252 				    &t->zthr_state_lock, t->zthr_sleep_timeout,
253 				    MSEC2NSEC(1), 0);
254 			}
255 		}
256 		if (t->zthr_haswaiters) {
257 			t->zthr_haswaiters = B_FALSE;
258 			cv_broadcast(&t->zthr_wait_cv);
259 		}
260 	}
261 
262 	/*
263 	 * Clear out the kernel thread metadata and notify the
264 	 * zthr_cancel() thread that we've stopped running.
265 	 */
266 	t->zthr_thread = NULL;
267 	t->zthr_cancel = B_FALSE;
268 	cv_broadcast(&t->zthr_cv);
269 
270 	mutex_exit(&t->zthr_state_lock);
271 	thread_exit();
272 }
273 
274 zthr_t *
zthr_create(const char * zthr_name,zthr_checkfunc_t * checkfunc,zthr_func_t * func,void * arg,pri_t pri)275 zthr_create(const char *zthr_name, zthr_checkfunc_t *checkfunc,
276     zthr_func_t *func, void *arg, pri_t pri)
277 {
278 	return (zthr_create_timer(zthr_name, checkfunc,
279 	    func, arg, (hrtime_t)0, pri));
280 }
281 
282 /*
283  * Create a zthr with specified maximum sleep time.  If the time
284  * in sleeping state exceeds max_sleep, a wakeup(do the check and
285  * start working if required) will be triggered.
286  */
287 zthr_t *
zthr_create_timer(const char * zthr_name,zthr_checkfunc_t * checkfunc,zthr_func_t * func,void * arg,hrtime_t max_sleep,pri_t pri)288 zthr_create_timer(const char *zthr_name, zthr_checkfunc_t *checkfunc,
289     zthr_func_t *func, void *arg, hrtime_t max_sleep, pri_t pri)
290 {
291 	zthr_t *t = kmem_zalloc(sizeof (*t), KM_SLEEP);
292 	mutex_init(&t->zthr_state_lock, NULL, MUTEX_DEFAULT, NULL);
293 	mutex_init(&t->zthr_request_lock, NULL, MUTEX_DEFAULT, NULL);
294 	cv_init(&t->zthr_cv, NULL, CV_DEFAULT, NULL);
295 	cv_init(&t->zthr_wait_cv, NULL, CV_DEFAULT, NULL);
296 
297 	mutex_enter(&t->zthr_state_lock);
298 	t->zthr_checkfunc = checkfunc;
299 	t->zthr_func = func;
300 	t->zthr_arg = arg;
301 	t->zthr_sleep_timeout = max_sleep;
302 	t->zthr_name = zthr_name;
303 	t->zthr_pri = pri;
304 
305 	t->zthr_thread = thread_create_named(zthr_name, NULL, 0,
306 	    zthr_procedure, t, 0, &p0, TS_RUN, pri);
307 
308 	mutex_exit(&t->zthr_state_lock);
309 
310 	return (t);
311 }
312 
313 void
zthr_destroy(zthr_t * t)314 zthr_destroy(zthr_t *t)
315 {
316 	ASSERT(!MUTEX_HELD(&t->zthr_state_lock));
317 	ASSERT(!MUTEX_HELD(&t->zthr_request_lock));
318 	VERIFY3P(t->zthr_thread, ==, NULL);
319 	mutex_destroy(&t->zthr_request_lock);
320 	mutex_destroy(&t->zthr_state_lock);
321 	cv_destroy(&t->zthr_cv);
322 	cv_destroy(&t->zthr_wait_cv);
323 	kmem_free(t, sizeof (*t));
324 }
325 
326 /*
327  * Wake up the zthr if it is sleeping. If the thread has been cancelled
328  * or is in the process of being cancelled, this is a no-op.
329  */
330 void
zthr_wakeup(zthr_t * t)331 zthr_wakeup(zthr_t *t)
332 {
333 	mutex_enter(&t->zthr_state_lock);
334 
335 	/*
336 	 * There are 5 states that we can find the zthr when issuing
337 	 * this broadcast:
338 	 *
339 	 * [1] The common case of the thread being asleep, at which
340 	 *     point the broadcast will wake it up.
341 	 * [2] The thread has been cancelled. Waking up a cancelled
342 	 *     thread is a no-op. Any work that is still left to be
343 	 *     done should be handled the next time the thread is
344 	 *     resumed.
345 	 * [3] The thread is doing work and is already up, so this
346 	 *     is basically a no-op.
347 	 * [4] The thread was just created/resumed, in which case the
348 	 *     behavior is similar to [3].
349 	 * [5] The thread is in the middle of being cancelled, which
350 	 *     will be a no-op.
351 	 */
352 	cv_broadcast(&t->zthr_cv);
353 
354 	mutex_exit(&t->zthr_state_lock);
355 }
356 
357 /*
358  * Sends a cancel request to the zthr and blocks until the zthr is
359  * cancelled. If the zthr is not running (e.g. has been cancelled
360  * already), this is a no-op. Note that this function should not be
361  * called from syncing context as it could deadlock with the zthr_func.
362  */
363 void
zthr_cancel(zthr_t * t)364 zthr_cancel(zthr_t *t)
365 {
366 	mutex_enter(&t->zthr_request_lock);
367 	mutex_enter(&t->zthr_state_lock);
368 
369 	/*
370 	 * Since we are holding the zthr_state_lock at this point
371 	 * we can find the state in one of the following 4 states:
372 	 *
373 	 * [1] The thread has already been cancelled, therefore
374 	 *     there is nothing for us to do.
375 	 * [2] The thread is sleeping so we set the flag, broadcast
376 	 *     the CV and wait for it to exit.
377 	 * [3] The thread is doing work, in which case we just set
378 	 *     the flag and wait for it to finish.
379 	 * [4] The thread was just created/resumed, in which case
380 	 *     the behavior is similar to [3].
381 	 *
382 	 * Since requests are serialized, by the time that we get
383 	 * control back we expect that the zthr is cancelled and
384 	 * not running anymore.
385 	 */
386 	if (t->zthr_thread != NULL) {
387 		t->zthr_cancel = B_TRUE;
388 
389 		/* broadcast in case the zthr is sleeping */
390 		cv_broadcast(&t->zthr_cv);
391 
392 		while (t->zthr_thread != NULL)
393 			cv_wait(&t->zthr_cv, &t->zthr_state_lock);
394 
395 		ASSERT(!t->zthr_cancel);
396 	}
397 
398 	mutex_exit(&t->zthr_state_lock);
399 	mutex_exit(&t->zthr_request_lock);
400 }
401 
402 /*
403  * Sends a resume request to the supplied zthr. If the zthr is already
404  * running this is a no-op. Note that this function should not be
405  * called from syncing context as it could deadlock with the zthr_func.
406  */
407 void
zthr_resume(zthr_t * t)408 zthr_resume(zthr_t *t)
409 {
410 	mutex_enter(&t->zthr_request_lock);
411 	mutex_enter(&t->zthr_state_lock);
412 
413 	ASSERT3P(&t->zthr_checkfunc, !=, NULL);
414 	ASSERT3P(&t->zthr_func, !=, NULL);
415 	ASSERT(!t->zthr_cancel);
416 	ASSERT(!t->zthr_haswaiters);
417 
418 	/*
419 	 * There are 4 states that we find the zthr in at this point
420 	 * given the locks that we hold:
421 	 *
422 	 * [1] The zthr was cancelled, so we spawn a new thread for
423 	 *     the zthr (common case).
424 	 * [2] The zthr is running at which point this is a no-op.
425 	 * [3] The zthr is sleeping at which point this is a no-op.
426 	 * [4] The zthr was just spawned at which point this is a
427 	 *     no-op.
428 	 */
429 	if (t->zthr_thread == NULL) {
430 		t->zthr_thread = thread_create_named(t->zthr_name, NULL, 0,
431 		    zthr_procedure, t, 0, &p0, TS_RUN, t->zthr_pri);
432 	}
433 
434 	mutex_exit(&t->zthr_state_lock);
435 	mutex_exit(&t->zthr_request_lock);
436 }
437 
438 /*
439  * This function is intended to be used by the zthr itself
440  * (specifically the zthr_func callback provided) to check
441  * if another thread has signaled it to stop running before
442  * doing some expensive operation.
443  *
444  * returns TRUE if we are in the middle of trying to cancel
445  *     this thread.
446  *
447  * returns FALSE otherwise.
448  */
449 boolean_t
zthr_iscancelled(zthr_t * t)450 zthr_iscancelled(zthr_t *t)
451 {
452 	ASSERT3P(t->zthr_thread, ==, curthread);
453 
454 	/*
455 	 * The majority of the functions here grab zthr_request_lock
456 	 * first and then zthr_state_lock. This function only grabs
457 	 * the zthr_state_lock. That is because this function should
458 	 * only be called from the zthr_func to check if someone has
459 	 * issued a zthr_cancel() on the thread. If there is a zthr_cancel()
460 	 * happening concurrently, attempting to grab the request lock
461 	 * here would result in a deadlock.
462 	 *
463 	 * By grabbing only the zthr_state_lock this function is allowed
464 	 * to run concurrently with a zthr_cancel() request.
465 	 */
466 	mutex_enter(&t->zthr_state_lock);
467 	boolean_t cancelled = t->zthr_cancel;
468 	mutex_exit(&t->zthr_state_lock);
469 	return (cancelled);
470 }
471 
472 boolean_t
zthr_iscurthread(zthr_t * t)473 zthr_iscurthread(zthr_t *t)
474 {
475 	return (t->zthr_thread == curthread);
476 }
477 
478 /*
479  * Wait for the zthr to finish its current function. Similar to
480  * zthr_iscancelled, you can use zthr_has_waiters to have the zthr_func end
481  * early. Unlike zthr_cancel, the thread is not destroyed. If the zthr was
482  * sleeping or cancelled, return immediately.
483  */
484 void
zthr_wait_cycle_done(zthr_t * t)485 zthr_wait_cycle_done(zthr_t *t)
486 {
487 	mutex_enter(&t->zthr_state_lock);
488 
489 	/*
490 	 * Since we are holding the zthr_state_lock at this point
491 	 * we can find the state in one of the following 5 states:
492 	 *
493 	 * [1] The thread has already cancelled, therefore
494 	 *     there is nothing for us to do.
495 	 * [2] The thread is sleeping so we set the flag, broadcast
496 	 *     the CV and wait for it to exit.
497 	 * [3] The thread is doing work, in which case we just set
498 	 *     the flag and wait for it to finish.
499 	 * [4] The thread was just created/resumed, in which case
500 	 *     the behavior is similar to [3].
501 	 * [5] The thread is the middle of being cancelled, which is
502 	 *     similar to [3]. We'll wait for the cancel, which is
503 	 *     waiting for the zthr func.
504 	 *
505 	 * Since requests are serialized, by the time that we get
506 	 * control back we expect that the zthr has completed it's
507 	 * zthr_func.
508 	 */
509 	if (t->zthr_thread != NULL) {
510 		t->zthr_haswaiters = B_TRUE;
511 
512 		/* broadcast in case the zthr is sleeping */
513 		cv_broadcast(&t->zthr_cv);
514 
515 		while ((t->zthr_haswaiters) && (t->zthr_thread != NULL))
516 			cv_wait(&t->zthr_wait_cv, &t->zthr_state_lock);
517 
518 		ASSERT(!t->zthr_haswaiters);
519 	}
520 
521 	mutex_exit(&t->zthr_state_lock);
522 }
523 
524 /*
525  * This function is intended to be used by the zthr itself
526  * to check if another thread is waiting on it to finish
527  *
528  * returns TRUE if we have been asked to finish.
529  *
530  * returns FALSE otherwise.
531  */
532 boolean_t
zthr_has_waiters(zthr_t * t)533 zthr_has_waiters(zthr_t *t)
534 {
535 	ASSERT3P(t->zthr_thread, ==, curthread);
536 
537 	mutex_enter(&t->zthr_state_lock);
538 
539 	/*
540 	 * Similarly to zthr_iscancelled(), we only grab the
541 	 * zthr_state_lock so that the zthr itself can use this
542 	 * to check for the request.
543 	 */
544 	boolean_t has_waiters = t->zthr_haswaiters;
545 	mutex_exit(&t->zthr_state_lock);
546 	return (has_waiters);
547 }
548