xref: /freebsd/sys/contrib/openzfs/module/zfs/zthr.c (revision a0409676120c1e558d0ade943019934e0f15118d)
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  * or
88  *     zthr_t *zthr_pointer = zthr_create_timer(checkfunc, func,
89  *         args, max_sleep);
90  *
91  * After that you should be able to wakeup, cancel, and resume the
92  * zthr from another thread using the zthr_pointer.
93  *
94  * NOTE: ZTHR threads could potentially wake up spuriously and the
95  * user should take this into account when writing a checkfunc.
96  * [see ZTHR state transitions]
97  *
98  * == ZTHR wakeup
99  *
100  * ZTHR wakeup should be used when new work is added for the zthr. The
101  * sleeping zthr will wakeup, see that it has more work to complete
102  * and proceed. This can be invoked from open or syncing context.
103  *
104  * To wakeup a zthr:
105  *     zthr_wakeup(zthr_t *t)
106  *
107  * == ZTHR cancellation and resumption
108  *
109  * ZTHR threads must be cancelled when their SPA is being exported
110  * or when they need to be paused so they don't interfere with other
111  * operations.
112  *
113  * To cancel a zthr:
114  *     zthr_cancel(zthr_pointer);
115  *
116  * To resume it:
117  *     zthr_resume(zthr_pointer);
118  *
119  * ZTHR cancel and resume should be invoked in open context during the
120  * lifecycle of the pool as it is imported, exported or destroyed.
121  *
122  * A zthr will implicitly check if it has received a cancellation
123  * signal every time func returns and every time it wakes up [see
124  * ZTHR state transitions below].
125  *
126  * At times, waiting for the zthr's func to finish its job may take
127  * time. This may be very time-consuming for some operations that
128  * need to cancel the SPA's zthrs (e.g spa_export). For this scenario
129  * the user can explicitly make their ZTHR function aware of incoming
130  * cancellation signals using zthr_iscancelled(). A common pattern for
131  * that looks like this:
132  *
133  * int
134  * func_name(void *args, zthr_t *t)
135  * {
136  *     ... <unpack args> ...
137  *     while (!work_done && !zthr_iscancelled(t)) {
138  *         ... <do more work> ...
139  *     }
140  * }
141  *
142  * == ZTHR cleanup
143  *
144  * Cancelling a zthr doesn't clean up its metadata (internal locks,
145  * function pointers to func and checkfunc, etc..). This is because
146  * we want to keep them around in case we want to resume the execution
147  * of the zthr later. Similarly for zthrs that exit themselves.
148  *
149  * To completely cleanup a zthr, cancel it first to ensure that it
150  * is not running and then use zthr_destroy().
151  *
152  * == ZTHR state transitions
153  *
154  *    zthr creation
155  *      +
156  *      |
157  *      |      woke up
158  *      |   +--------------+ sleep
159  *      |   |                  ^
160  *      |   |                  |
161  *      |   |                  | FALSE
162  *      |   |                  |
163  *      v   v     FALSE        +
164  *   cancelled? +---------> checkfunc?
165  *      +   ^                  +
166  *      |   |                  |
167  *      |   |                  | TRUE
168  *      |   |                  |
169  *      |   |  func returned   v
170  *      |   +---------------+ func
171  *      |
172  *      | TRUE
173  *      |
174  *      v
175  *   zthr stopped running
176  *
177  * == Implementation of ZTHR requests
178  *
179  * ZTHR cancel and resume are requests on a zthr to change its
180  * internal state. These requests are serialized using the
181  * zthr_request_lock, while changes in its internal state are
182  * protected by the zthr_state_lock. A request will first acquire
183  * the zthr_request_lock and then immediately acquire the
184  * zthr_state_lock. We do this so that incoming requests are
185  * serialized using the request lock, while still allowing us
186  * to use the state lock for thread communication via zthr_cv.
187  *
188  * ZTHR wakeup broadcasts to zthr_cv, causing sleeping threads
189  * to wakeup. It acquires the zthr_state_lock but not the
190  * zthr_request_lock, so that a wakeup on a zthr in the middle
191  * of being cancelled will not block.
192  */
193 
194 #include <sys/zfs_context.h>
195 #include <sys/zthr.h>
196 
197 struct zthr {
198 	/* running thread doing the work */
199 	kthread_t	*zthr_thread;
200 
201 	/* lock protecting internal data & invariants */
202 	kmutex_t	zthr_state_lock;
203 
204 	/* mutex that serializes external requests */
205 	kmutex_t	zthr_request_lock;
206 
207 	/* notification mechanism for requests */
208 	kcondvar_t	zthr_cv;
209 
210 	/* flag set to true if we are canceling the zthr */
211 	boolean_t	zthr_cancel;
212 
213 	/* flag set to true if we are waiting for the zthr to finish */
214 	boolean_t	zthr_haswaiters;
215 	kcondvar_t	zthr_wait_cv;
216 	/*
217 	 * maximum amount of time that the zthr is spent sleeping;
218 	 * if this is 0, the thread doesn't wake up until it gets
219 	 * signaled.
220 	 */
221 	hrtime_t	zthr_sleep_timeout;
222 
223 	/* consumer-provided callbacks & data */
224 	zthr_checkfunc_t	*zthr_checkfunc;
225 	zthr_func_t	*zthr_func;
226 	void		*zthr_arg;
227 	const char	*zthr_name;
228 };
229 
230 static void
231 zthr_procedure(void *arg)
232 {
233 	zthr_t *t = arg;
234 
235 	mutex_enter(&t->zthr_state_lock);
236 	ASSERT3P(t->zthr_thread, ==, curthread);
237 
238 	while (!t->zthr_cancel) {
239 		if (t->zthr_checkfunc(t->zthr_arg, t)) {
240 			mutex_exit(&t->zthr_state_lock);
241 			t->zthr_func(t->zthr_arg, t);
242 			mutex_enter(&t->zthr_state_lock);
243 		} else {
244 			if (t->zthr_sleep_timeout == 0) {
245 				cv_wait_idle(&t->zthr_cv, &t->zthr_state_lock);
246 			} else {
247 				(void) cv_timedwait_idle_hires(&t->zthr_cv,
248 				    &t->zthr_state_lock, t->zthr_sleep_timeout,
249 				    MSEC2NSEC(1), 0);
250 			}
251 		}
252 		if (t->zthr_haswaiters) {
253 			t->zthr_haswaiters = B_FALSE;
254 			cv_broadcast(&t->zthr_wait_cv);
255 		}
256 	}
257 
258 	/*
259 	 * Clear out the kernel thread metadata and notify the
260 	 * zthr_cancel() thread that we've stopped running.
261 	 */
262 	t->zthr_thread = NULL;
263 	t->zthr_cancel = B_FALSE;
264 	cv_broadcast(&t->zthr_cv);
265 
266 	mutex_exit(&t->zthr_state_lock);
267 	thread_exit();
268 }
269 
270 zthr_t *
271 zthr_create(const char *zthr_name, zthr_checkfunc_t *checkfunc,
272     zthr_func_t *func, void *arg)
273 {
274 	return (zthr_create_timer(zthr_name, checkfunc,
275 	    func, arg, (hrtime_t)0));
276 }
277 
278 /*
279  * Create a zthr with specified maximum sleep time.  If the time
280  * in sleeping state exceeds max_sleep, a wakeup(do the check and
281  * start working if required) will be triggered.
282  */
283 zthr_t *
284 zthr_create_timer(const char *zthr_name, zthr_checkfunc_t *checkfunc,
285     zthr_func_t *func, void *arg, hrtime_t max_sleep)
286 {
287 	zthr_t *t = kmem_zalloc(sizeof (*t), KM_SLEEP);
288 	mutex_init(&t->zthr_state_lock, NULL, MUTEX_DEFAULT, NULL);
289 	mutex_init(&t->zthr_request_lock, NULL, MUTEX_DEFAULT, NULL);
290 	cv_init(&t->zthr_cv, NULL, CV_DEFAULT, NULL);
291 	cv_init(&t->zthr_wait_cv, NULL, CV_DEFAULT, NULL);
292 
293 	mutex_enter(&t->zthr_state_lock);
294 	t->zthr_checkfunc = checkfunc;
295 	t->zthr_func = func;
296 	t->zthr_arg = arg;
297 	t->zthr_sleep_timeout = max_sleep;
298 	t->zthr_name = zthr_name;
299 
300 	t->zthr_thread = thread_create_named(zthr_name, NULL, 0,
301 	    zthr_procedure, t, 0, &p0, TS_RUN, minclsyspri);
302 
303 	mutex_exit(&t->zthr_state_lock);
304 
305 	return (t);
306 }
307 
308 void
309 zthr_destroy(zthr_t *t)
310 {
311 	ASSERT(!MUTEX_HELD(&t->zthr_state_lock));
312 	ASSERT(!MUTEX_HELD(&t->zthr_request_lock));
313 	VERIFY3P(t->zthr_thread, ==, NULL);
314 	mutex_destroy(&t->zthr_request_lock);
315 	mutex_destroy(&t->zthr_state_lock);
316 	cv_destroy(&t->zthr_cv);
317 	cv_destroy(&t->zthr_wait_cv);
318 	kmem_free(t, sizeof (*t));
319 }
320 
321 /*
322  * Wake up the zthr if it is sleeping. If the thread has been cancelled
323  * or is in the process of being cancelled, this is a no-op.
324  */
325 void
326 zthr_wakeup(zthr_t *t)
327 {
328 	mutex_enter(&t->zthr_state_lock);
329 
330 	/*
331 	 * There are 5 states that we can find the zthr when issuing
332 	 * this broadcast:
333 	 *
334 	 * [1] The common case of the thread being asleep, at which
335 	 *     point the broadcast will wake it up.
336 	 * [2] The thread has been cancelled. Waking up a cancelled
337 	 *     thread is a no-op. Any work that is still left to be
338 	 *     done should be handled the next time the thread is
339 	 *     resumed.
340 	 * [3] The thread is doing work and is already up, so this
341 	 *     is basically a no-op.
342 	 * [4] The thread was just created/resumed, in which case the
343 	 *     behavior is similar to [3].
344 	 * [5] The thread is in the middle of being cancelled, which
345 	 *     will be a no-op.
346 	 */
347 	cv_broadcast(&t->zthr_cv);
348 
349 	mutex_exit(&t->zthr_state_lock);
350 }
351 
352 /*
353  * Sends a cancel request to the zthr and blocks until the zthr is
354  * cancelled. If the zthr is not running (e.g. has been cancelled
355  * already), this is a no-op. Note that this function should not be
356  * called from syncing context as it could deadlock with the zthr_func.
357  */
358 void
359 zthr_cancel(zthr_t *t)
360 {
361 	mutex_enter(&t->zthr_request_lock);
362 	mutex_enter(&t->zthr_state_lock);
363 
364 	/*
365 	 * Since we are holding the zthr_state_lock at this point
366 	 * we can find the state in one of the following 4 states:
367 	 *
368 	 * [1] The thread has already been cancelled, therefore
369 	 *     there is nothing for us to do.
370 	 * [2] The thread is sleeping so we set the flag, broadcast
371 	 *     the CV and wait for it to exit.
372 	 * [3] The thread is doing work, in which case we just set
373 	 *     the flag and wait for it to finish.
374 	 * [4] The thread was just created/resumed, in which case
375 	 *     the behavior is similar to [3].
376 	 *
377 	 * Since requests are serialized, by the time that we get
378 	 * control back we expect that the zthr is cancelled and
379 	 * not running anymore.
380 	 */
381 	if (t->zthr_thread != NULL) {
382 		t->zthr_cancel = B_TRUE;
383 
384 		/* broadcast in case the zthr is sleeping */
385 		cv_broadcast(&t->zthr_cv);
386 
387 		while (t->zthr_thread != NULL)
388 			cv_wait(&t->zthr_cv, &t->zthr_state_lock);
389 
390 		ASSERT(!t->zthr_cancel);
391 	}
392 
393 	mutex_exit(&t->zthr_state_lock);
394 	mutex_exit(&t->zthr_request_lock);
395 }
396 
397 /*
398  * Sends a resume request to the supplied zthr. If the zthr is already
399  * running this is a no-op. Note that this function should not be
400  * called from syncing context as it could deadlock with the zthr_func.
401  */
402 void
403 zthr_resume(zthr_t *t)
404 {
405 	mutex_enter(&t->zthr_request_lock);
406 	mutex_enter(&t->zthr_state_lock);
407 
408 	ASSERT3P(&t->zthr_checkfunc, !=, NULL);
409 	ASSERT3P(&t->zthr_func, !=, NULL);
410 	ASSERT(!t->zthr_cancel);
411 	ASSERT(!t->zthr_haswaiters);
412 
413 	/*
414 	 * There are 4 states that we find the zthr in at this point
415 	 * given the locks that we hold:
416 	 *
417 	 * [1] The zthr was cancelled, so we spawn a new thread for
418 	 *     the zthr (common case).
419 	 * [2] The zthr is running at which point this is a no-op.
420 	 * [3] The zthr is sleeping at which point this is a no-op.
421 	 * [4] The zthr was just spawned at which point this is a
422 	 *     no-op.
423 	 */
424 	if (t->zthr_thread == NULL) {
425 		t->zthr_thread = thread_create_named(t->zthr_name, NULL, 0,
426 		    zthr_procedure, t, 0, &p0, TS_RUN, minclsyspri);
427 	}
428 
429 	mutex_exit(&t->zthr_state_lock);
430 	mutex_exit(&t->zthr_request_lock);
431 }
432 
433 /*
434  * This function is intended to be used by the zthr itself
435  * (specifically the zthr_func callback provided) to check
436  * if another thread has signaled it to stop running before
437  * doing some expensive operation.
438  *
439  * returns TRUE if we are in the middle of trying to cancel
440  *     this thread.
441  *
442  * returns FALSE otherwise.
443  */
444 boolean_t
445 zthr_iscancelled(zthr_t *t)
446 {
447 	ASSERT3P(t->zthr_thread, ==, curthread);
448 
449 	/*
450 	 * The majority of the functions here grab zthr_request_lock
451 	 * first and then zthr_state_lock. This function only grabs
452 	 * the zthr_state_lock. That is because this function should
453 	 * only be called from the zthr_func to check if someone has
454 	 * issued a zthr_cancel() on the thread. If there is a zthr_cancel()
455 	 * happening concurrently, attempting to grab the request lock
456 	 * here would result in a deadlock.
457 	 *
458 	 * By grabbing only the zthr_state_lock this function is allowed
459 	 * to run concurrently with a zthr_cancel() request.
460 	 */
461 	mutex_enter(&t->zthr_state_lock);
462 	boolean_t cancelled = t->zthr_cancel;
463 	mutex_exit(&t->zthr_state_lock);
464 	return (cancelled);
465 }
466 
467 /*
468  * Wait for the zthr to finish its current function. Similar to
469  * zthr_iscancelled, you can use zthr_has_waiters to have the zthr_func end
470  * early. Unlike zthr_cancel, the thread is not destroyed. If the zthr was
471  * sleeping or cancelled, return immediately.
472  */
473 void
474 zthr_wait_cycle_done(zthr_t *t)
475 {
476 	mutex_enter(&t->zthr_state_lock);
477 
478 	/*
479 	 * Since we are holding the zthr_state_lock at this point
480 	 * we can find the state in one of the following 5 states:
481 	 *
482 	 * [1] The thread has already cancelled, therefore
483 	 *     there is nothing for us to do.
484 	 * [2] The thread is sleeping so we set the flag, broadcast
485 	 *     the CV and wait for it to exit.
486 	 * [3] The thread is doing work, in which case we just set
487 	 *     the flag and wait for it to finish.
488 	 * [4] The thread was just created/resumed, in which case
489 	 *     the behavior is similar to [3].
490 	 * [5] The thread is the middle of being cancelled, which is
491 	 *     similar to [3]. We'll wait for the cancel, which is
492 	 *     waiting for the zthr func.
493 	 *
494 	 * Since requests are serialized, by the time that we get
495 	 * control back we expect that the zthr has completed it's
496 	 * zthr_func.
497 	 */
498 	if (t->zthr_thread != NULL) {
499 		t->zthr_haswaiters = B_TRUE;
500 
501 		/* broadcast in case the zthr is sleeping */
502 		cv_broadcast(&t->zthr_cv);
503 
504 		while ((t->zthr_haswaiters) && (t->zthr_thread != NULL))
505 			cv_wait(&t->zthr_wait_cv, &t->zthr_state_lock);
506 
507 		ASSERT(!t->zthr_haswaiters);
508 	}
509 
510 	mutex_exit(&t->zthr_state_lock);
511 }
512 
513 /*
514  * This function is intended to be used by the zthr itself
515  * to check if another thread is waiting on it to finish
516  *
517  * returns TRUE if we have been asked to finish.
518  *
519  * returns FALSE otherwise.
520  */
521 boolean_t
522 zthr_has_waiters(zthr_t *t)
523 {
524 	ASSERT3P(t->zthr_thread, ==, curthread);
525 
526 	mutex_enter(&t->zthr_state_lock);
527 
528 	/*
529 	 * Similarly to zthr_iscancelled(), we only grab the
530 	 * zthr_state_lock so that the zthr itself can use this
531 	 * to check for the request.
532 	 */
533 	boolean_t has_waiters = t->zthr_haswaiters;
534 	mutex_exit(&t->zthr_state_lock);
535 	return (has_waiters);
536 }
537