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
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.
17 * Copyright (c) 2017, 2020 by Delphix. All rights reserved.
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:
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
34 * If the operation satisfies the above then the following rules guarantee
35 * a certain level of correctness:
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.
42 * This way a normal zthr cycle should go like this:
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.
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()].
54 * Note: ZTHR threads are NOT a replacement for generic threads! Please
55 * ensure that they fit your use-case well before using them.
59 * Every zthr needs three inputs to start running:
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);
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
70 * void func_name(void *args, zthr_t *t);
72 * 3] A void args pointer that will be passed to checkfunc and func
73 * implicitly by the infrastructure.
75 * The reason why the above API needs two different functions,
76 * instead of one that both checks and does the work, has to do with
77 * the zthr's internal state lock (zthr_state_lock) and the allowed
78 * cancellation windows. We want to hold the zthr_state_lock while
79 * running checkfunc but not while running func. This way the zthr
80 * can be cancelled while doing work and not while checking for work.
83 * zthr_t *zthr_pointer = zthr_create(checkfunc, func, args);
85 * zthr_t *zthr_pointer = zthr_create_timer(checkfunc, func,
88 * After that you should be able to wakeup, cancel, and resume the
89 * zthr from another thread using the zthr_pointer.
91 * NOTE: ZTHR threads could potentially wake up spuriously and the
92 * user should take this into account when writing a checkfunc.
93 * [see ZTHR state transitions]
97 * ZTHR wakeup should be used when new work is added for the zthr. The
98 * sleeping zthr will wakeup, see that it has more work to complete
99 * and proceed. This can be invoked from open or syncing context.
102 * zthr_wakeup(zthr_t *t)
104 * == ZTHR cancellation and resumption
106 * ZTHR threads must be cancelled when their SPA is being exported
107 * or when they need to be paused so they don't interfere with other
111 * zthr_cancel(zthr_pointer);
114 * zthr_resume(zthr_pointer);
116 * ZTHR cancel and resume should be invoked in open context during the
117 * lifecycle of the pool as it is imported, exported or destroyed.
119 * A zthr will implicitly check if it has received a cancellation
120 * signal every time func returns and every time it wakes up [see
121 * ZTHR state transitions below].
123 * At times, waiting for the zthr's func to finish its job may take
124 * time. This may be very time-consuming for some operations that
125 * need to cancel the SPA's zthrs (e.g spa_export). For this scenario
126 * the user can explicitly make their ZTHR function aware of incoming
127 * cancellation signals using zthr_iscancelled(). A common pattern for
128 * that looks like this:
131 * func_name(void *args, zthr_t *t)
133 * ... <unpack args> ...
134 * while (!work_done && !zthr_iscancelled(t)) {
135 * ... <do more work> ...
141 * Cancelling a zthr doesn't clean up its metadata (internal locks,
142 * function pointers to func and checkfunc, etc..). This is because
143 * we want to keep them around in case we want to resume the execution
144 * of the zthr later. Similarly for zthrs that exit themselves.
146 * To completely cleanup a zthr, cancel it first to ensure that it
147 * is not running and then use zthr_destroy().
149 * == ZTHR state transitions
155 * | +--------------+ sleep
161 * cancelled? +---------> checkfunc?
166 * | | func returned v
167 * | +---------------+ func
172 * zthr stopped running
174 * == Implementation of ZTHR requests
176 * ZTHR cancel and resume are requests on a zthr to change its
177 * internal state. These requests are serialized using the
178 * zthr_request_lock, while changes in its internal state are
179 * protected by the zthr_state_lock. A request will first acquire
180 * the zthr_request_lock and then immediately acquire the
181 * zthr_state_lock. We do this so that incoming requests are
182 * serialized using the request lock, while still allowing us
183 * to use the state lock for thread communication via zthr_cv.
185 * ZTHR wakeup broadcasts to zthr_cv, causing sleeping threads
186 * to wakeup. It acquires the zthr_state_lock but not the
187 * zthr_request_lock, so that a wakeup on a zthr in the middle
188 * of being cancelled will not block.
191 #include <sys/zfs_context.h>
192 #include <sys/zthr.h>
195 /* running thread doing the work */
196 kthread_t *zthr_thread;
198 /* lock protecting internal data & invariants */
199 kmutex_t zthr_state_lock;
201 /* mutex that serializes external requests */
202 kmutex_t zthr_request_lock;
204 /* notification mechanism for requests */
207 /* flag set to true if we are canceling the zthr */
208 boolean_t zthr_cancel;
210 /* flag set to true if we are waiting for the zthr to finish */
211 boolean_t zthr_haswaiters;
212 kcondvar_t zthr_wait_cv;
214 * maximum amount of time that the zthr is spent sleeping;
215 * if this is 0, the thread doesn't wake up until it gets
218 hrtime_t zthr_sleep_timeout;
220 /* consumer-provided callbacks & data */
221 zthr_checkfunc_t *zthr_checkfunc;
222 zthr_func_t *zthr_func;
227 zthr_procedure(void *arg)
231 mutex_enter(&t->zthr_state_lock);
232 ASSERT3P(t->zthr_thread, ==, curthread);
234 while (!t->zthr_cancel) {
235 if (t->zthr_checkfunc(t->zthr_arg, t)) {
236 mutex_exit(&t->zthr_state_lock);
237 t->zthr_func(t->zthr_arg, t);
238 mutex_enter(&t->zthr_state_lock);
241 * cv_wait_sig() is used instead of cv_wait() in
242 * order to prevent this process from incorrectly
243 * contributing to the system load average when idle.
245 if (t->zthr_sleep_timeout == 0) {
246 cv_wait_sig(&t->zthr_cv, &t->zthr_state_lock);
248 (void) cv_timedwait_sig_hires(&t->zthr_cv,
249 &t->zthr_state_lock, t->zthr_sleep_timeout,
253 if (t->zthr_haswaiters) {
254 t->zthr_haswaiters = B_FALSE;
255 cv_broadcast(&t->zthr_wait_cv);
260 * Clear out the kernel thread metadata and notify the
261 * zthr_cancel() thread that we've stopped running.
263 t->zthr_thread = NULL;
264 t->zthr_cancel = B_FALSE;
265 cv_broadcast(&t->zthr_cv);
267 mutex_exit(&t->zthr_state_lock);
272 zthr_create(const char *zthr_name, zthr_checkfunc_t *checkfunc,
273 zthr_func_t *func, void *arg)
275 return (zthr_create_timer(zthr_name, checkfunc,
276 func, arg, (hrtime_t)0));
280 * Create a zthr with specified maximum sleep time. If the time
281 * in sleeping state exceeds max_sleep, a wakeup(do the check and
282 * start working if required) will be triggered.
285 zthr_create_timer(const char *zthr_name, zthr_checkfunc_t *checkfunc,
286 zthr_func_t *func, void *arg, hrtime_t max_sleep)
288 zthr_t *t = kmem_zalloc(sizeof (*t), KM_SLEEP);
289 mutex_init(&t->zthr_state_lock, NULL, MUTEX_DEFAULT, NULL);
290 mutex_init(&t->zthr_request_lock, NULL, MUTEX_DEFAULT, NULL);
291 cv_init(&t->zthr_cv, NULL, CV_DEFAULT, NULL);
292 cv_init(&t->zthr_wait_cv, NULL, CV_DEFAULT, NULL);
294 mutex_enter(&t->zthr_state_lock);
295 t->zthr_checkfunc = checkfunc;
298 t->zthr_sleep_timeout = max_sleep;
300 t->zthr_thread = thread_create_named(zthr_name, NULL, 0,
301 zthr_procedure, t, 0, &p0, TS_RUN, minclsyspri);
303 mutex_exit(&t->zthr_state_lock);
309 zthr_destroy(zthr_t *t)
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));
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.
326 zthr_wakeup(zthr_t *t)
328 mutex_enter(&t->zthr_state_lock);
331 * There are 5 states that we can find the zthr when issuing
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
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
347 cv_broadcast(&t->zthr_cv);
349 mutex_exit(&t->zthr_state_lock);
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.
359 zthr_cancel(zthr_t *t)
361 mutex_enter(&t->zthr_request_lock);
362 mutex_enter(&t->zthr_state_lock);
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:
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].
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.
381 if (t->zthr_thread != NULL) {
382 t->zthr_cancel = B_TRUE;
384 /* broadcast in case the zthr is sleeping */
385 cv_broadcast(&t->zthr_cv);
387 while (t->zthr_thread != NULL)
388 cv_wait(&t->zthr_cv, &t->zthr_state_lock);
390 ASSERT(!t->zthr_cancel);
393 mutex_exit(&t->zthr_state_lock);
394 mutex_exit(&t->zthr_request_lock);
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.
403 zthr_resume(zthr_t *t)
405 mutex_enter(&t->zthr_request_lock);
406 mutex_enter(&t->zthr_state_lock);
408 ASSERT3P(&t->zthr_checkfunc, !=, NULL);
409 ASSERT3P(&t->zthr_func, !=, NULL);
410 ASSERT(!t->zthr_cancel);
411 ASSERT(!t->zthr_haswaiters);
414 * There are 4 states that we find the zthr in at this point
415 * given the locks that we hold:
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
424 if (t->zthr_thread == NULL) {
425 t->zthr_thread = thread_create(NULL, 0, zthr_procedure, t,
426 0, &p0, TS_RUN, minclsyspri);
429 mutex_exit(&t->zthr_state_lock);
430 mutex_exit(&t->zthr_request_lock);
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.
439 * returns TRUE if we are in the middle of trying to cancel
442 * returns FALSE otherwise.
445 zthr_iscancelled(zthr_t *t)
447 ASSERT3P(t->zthr_thread, ==, curthread);
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.
458 * By grabbing only the zthr_state_lock this function is allowed
459 * to run concurrently with a zthr_cancel() request.
461 mutex_enter(&t->zthr_state_lock);
462 boolean_t cancelled = t->zthr_cancel;
463 mutex_exit(&t->zthr_state_lock);
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.
474 zthr_wait_cycle_done(zthr_t *t)
476 mutex_enter(&t->zthr_state_lock);
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:
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.
494 * Since requests are serialized, by the time that we get
495 * control back we expect that the zthr has completed it's
498 if (t->zthr_thread != NULL) {
499 t->zthr_haswaiters = B_TRUE;
501 /* broadcast in case the zthr is sleeping */
502 cv_broadcast(&t->zthr_cv);
504 while ((t->zthr_haswaiters) && (t->zthr_thread != NULL))
505 cv_wait(&t->zthr_wait_cv, &t->zthr_state_lock);
507 ASSERT(!t->zthr_haswaiters);
510 mutex_exit(&t->zthr_state_lock);
514 * This function is intended to be used by the zthr itself
515 * to check if another thread is waiting on it to finish
517 * returns TRUE if we have been asked to finish.
519 * returns FALSE otherwise.
522 zthr_has_waiters(zthr_t *t)
524 ASSERT3P(t->zthr_thread, ==, curthread);
526 mutex_enter(&t->zthr_state_lock);
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.
533 boolean_t has_waiters = t->zthr_haswaiters;
534 mutex_exit(&t->zthr_state_lock);
535 return (has_waiters);