xref: /freebsd/sys/contrib/openzfs/module/zfs/txg.c (revision b985c9cafd2aedac5cf92428c0211485ea4ede24)
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 https://opensource.org/licenses/CDDL-1.0.
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  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Portions Copyright 2011 Martin Matuska
24  * Copyright (c) 2012, 2019 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/txg_impl.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/spa_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dsl_pool.h>
33 #include <sys/dsl_scan.h>
34 #include <sys/zil.h>
35 #include <sys/callb.h>
36 #include <sys/trace_zfs.h>
37 
38 /*
39  * ZFS Transaction Groups
40  * ----------------------
41  *
42  * ZFS transaction groups are, as the name implies, groups of transactions
43  * that act on persistent state. ZFS asserts consistency at the granularity of
44  * these transaction groups. Each successive transaction group (txg) is
45  * assigned a 64-bit consecutive identifier. There are three active
46  * transaction group states: open, quiescing, or syncing. At any given time,
47  * there may be an active txg associated with each state; each active txg may
48  * either be processing, or blocked waiting to enter the next state. There may
49  * be up to three active txgs, and there is always a txg in the open state
50  * (though it may be blocked waiting to enter the quiescing state). In broad
51  * strokes, transactions -- operations that change in-memory structures -- are
52  * accepted into the txg in the open state, and are completed while the txg is
53  * in the open or quiescing states. The accumulated changes are written to
54  * disk in the syncing state.
55  *
56  * Open
57  *
58  * When a new txg becomes active, it first enters the open state. New
59  * transactions -- updates to in-memory structures -- are assigned to the
60  * currently open txg. There is always a txg in the open state so that ZFS can
61  * accept new changes (though the txg may refuse new changes if it has hit
62  * some limit). ZFS advances the open txg to the next state for a variety of
63  * reasons such as it hitting a time or size threshold, or the execution of an
64  * administrative action that must be completed in the syncing state.
65  *
66  * Quiescing
67  *
68  * After a txg exits the open state, it enters the quiescing state. The
69  * quiescing state is intended to provide a buffer between accepting new
70  * transactions in the open state and writing them out to stable storage in
71  * the syncing state. While quiescing, transactions can continue their
72  * operation without delaying either of the other states. Typically, a txg is
73  * in the quiescing state very briefly since the operations are bounded by
74  * software latencies rather than, say, slower I/O latencies. After all
75  * transactions complete, the txg is ready to enter the next state.
76  *
77  * Syncing
78  *
79  * In the syncing state, the in-memory state built up during the open and (to
80  * a lesser degree) the quiescing states is written to stable storage. The
81  * process of writing out modified data can, in turn modify more data. For
82  * example when we write new blocks, we need to allocate space for them; those
83  * allocations modify metadata (space maps)... which themselves must be
84  * written to stable storage. During the sync state, ZFS iterates, writing out
85  * data until it converges and all in-memory changes have been written out.
86  * The first such pass is the largest as it encompasses all the modified user
87  * data (as opposed to filesystem metadata). Subsequent passes typically have
88  * far less data to write as they consist exclusively of filesystem metadata.
89  *
90  * To ensure convergence, after a certain number of passes ZFS begins
91  * overwriting locations on stable storage that had been allocated earlier in
92  * the syncing state (and subsequently freed). ZFS usually allocates new
93  * blocks to optimize for large, continuous, writes. For the syncing state to
94  * converge however it must complete a pass where no new blocks are allocated
95  * since each allocation requires a modification of persistent metadata.
96  * Further, to hasten convergence, after a prescribed number of passes, ZFS
97  * also defers frees, and stops compressing.
98  *
99  * In addition to writing out user data, we must also execute synctasks during
100  * the syncing context. A synctask is the mechanism by which some
101  * administrative activities work such as creating and destroying snapshots or
102  * datasets. Note that when a synctask is initiated it enters the open txg,
103  * and ZFS then pushes that txg as quickly as possible to completion of the
104  * syncing state in order to reduce the latency of the administrative
105  * activity. To complete the syncing state, ZFS writes out a new uberblock,
106  * the root of the tree of blocks that comprise all state stored on the ZFS
107  * pool. Finally, if there is a quiesced txg waiting, we signal that it can
108  * now transition to the syncing state.
109  */
110 
111 static __attribute__((noreturn)) void txg_sync_thread(void *arg);
112 static __attribute__((noreturn)) void txg_quiesce_thread(void *arg);
113 
114 uint_t zfs_txg_timeout = 5;	/* max seconds worth of delta per txg */
115 
116 /*
117  * Prepare the txg subsystem.
118  */
119 void
120 txg_init(dsl_pool_t *dp, uint64_t txg)
121 {
122 	tx_state_t *tx = &dp->dp_tx;
123 	int c;
124 	memset(tx, 0, sizeof (tx_state_t));
125 
126 	tx->tx_cpu = vmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
127 
128 	for (c = 0; c < max_ncpus; c++) {
129 		int i;
130 
131 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
132 		mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_NOLOCKDEP,
133 		    NULL);
134 		for (i = 0; i < TXG_SIZE; i++) {
135 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
136 			    NULL);
137 			list_create(&tx->tx_cpu[c].tc_callbacks[i],
138 			    sizeof (dmu_tx_callback_t),
139 			    offsetof(dmu_tx_callback_t, dcb_node));
140 		}
141 	}
142 
143 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
144 
145 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
146 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
147 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
148 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
149 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
150 
151 	tx->tx_open_txg = txg;
152 }
153 
154 /*
155  * Close down the txg subsystem.
156  */
157 void
158 txg_fini(dsl_pool_t *dp)
159 {
160 	tx_state_t *tx = &dp->dp_tx;
161 	int c;
162 
163 	ASSERT0(tx->tx_threads);
164 
165 	mutex_destroy(&tx->tx_sync_lock);
166 
167 	cv_destroy(&tx->tx_sync_more_cv);
168 	cv_destroy(&tx->tx_sync_done_cv);
169 	cv_destroy(&tx->tx_quiesce_more_cv);
170 	cv_destroy(&tx->tx_quiesce_done_cv);
171 	cv_destroy(&tx->tx_exit_cv);
172 
173 	for (c = 0; c < max_ncpus; c++) {
174 		int i;
175 
176 		mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
177 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
178 		for (i = 0; i < TXG_SIZE; i++) {
179 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
180 			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
181 		}
182 	}
183 
184 	if (tx->tx_commit_cb_taskq != NULL)
185 		taskq_destroy(tx->tx_commit_cb_taskq);
186 
187 	vmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
188 
189 	memset(tx, 0, sizeof (tx_state_t));
190 }
191 
192 /*
193  * Start syncing transaction groups.
194  */
195 void
196 txg_sync_start(dsl_pool_t *dp)
197 {
198 	tx_state_t *tx = &dp->dp_tx;
199 
200 	mutex_enter(&tx->tx_sync_lock);
201 
202 	dprintf("pool %p\n", dp);
203 
204 	ASSERT0(tx->tx_threads);
205 
206 	tx->tx_threads = 2;
207 
208 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
209 	    dp, 0, &p0, TS_RUN, defclsyspri);
210 
211 	/*
212 	 * The sync thread can need a larger-than-default stack size on
213 	 * 32-bit x86.  This is due in part to nested pools and
214 	 * scrub_visitbp() recursion.
215 	 */
216 	tx->tx_sync_thread = thread_create(NULL, 0, txg_sync_thread,
217 	    dp, 0, &p0, TS_RUN, defclsyspri);
218 
219 	mutex_exit(&tx->tx_sync_lock);
220 }
221 
222 static void
223 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
224 {
225 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
226 	mutex_enter(&tx->tx_sync_lock);
227 }
228 
229 static void
230 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
231 {
232 	ASSERT(*tpp != NULL);
233 	*tpp = NULL;
234 	tx->tx_threads--;
235 	cv_broadcast(&tx->tx_exit_cv);
236 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
237 	thread_exit();
238 }
239 
240 static void
241 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
242 {
243 	CALLB_CPR_SAFE_BEGIN(cpr);
244 
245 	if (time) {
246 		(void) cv_timedwait_idle(cv, &tx->tx_sync_lock,
247 		    ddi_get_lbolt() + time);
248 	} else {
249 		cv_wait_idle(cv, &tx->tx_sync_lock);
250 	}
251 
252 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
253 }
254 
255 /*
256  * Stop syncing transaction groups.
257  */
258 void
259 txg_sync_stop(dsl_pool_t *dp)
260 {
261 	tx_state_t *tx = &dp->dp_tx;
262 
263 	dprintf("pool %p\n", dp);
264 	/*
265 	 * Finish off any work in progress.
266 	 */
267 	ASSERT3U(tx->tx_threads, ==, 2);
268 
269 	/*
270 	 * We need to ensure that we've vacated the deferred metaslab trees.
271 	 */
272 	txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
273 
274 	/*
275 	 * Wake all sync threads and wait for them to die.
276 	 */
277 	mutex_enter(&tx->tx_sync_lock);
278 
279 	ASSERT3U(tx->tx_threads, ==, 2);
280 
281 	tx->tx_exiting = 1;
282 
283 	cv_broadcast(&tx->tx_quiesce_more_cv);
284 	cv_broadcast(&tx->tx_quiesce_done_cv);
285 	cv_broadcast(&tx->tx_sync_more_cv);
286 
287 	while (tx->tx_threads != 0)
288 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
289 
290 	tx->tx_exiting = 0;
291 
292 	mutex_exit(&tx->tx_sync_lock);
293 }
294 
295 /*
296  * Get a handle on the currently open txg and keep it open.
297  *
298  * The txg is guaranteed to stay open until txg_rele_to_quiesce() is called for
299  * the handle. Once txg_rele_to_quiesce() has been called, the txg stays
300  * in quiescing state until txg_rele_to_sync() is called for the handle.
301  *
302  * It is guaranteed that subsequent calls return monotonically increasing
303  * txgs for the same dsl_pool_t. Of course this is not strong monotonicity,
304  * because the same txg can be returned multiple times in a row. This
305  * guarantee holds both for subsequent calls from one thread and for multiple
306  * threads. For example, it is impossible to observe the following sequence
307  * of events:
308  *
309  *           Thread 1                            Thread 2
310  *
311  *   1 <- txg_hold_open(P, ...)
312  *                                       2 <- txg_hold_open(P, ...)
313  *   1 <- txg_hold_open(P, ...)
314  *
315  */
316 uint64_t
317 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
318 {
319 	tx_state_t *tx = &dp->dp_tx;
320 	tx_cpu_t *tc;
321 	uint64_t txg;
322 
323 	/*
324 	 * It appears the processor id is simply used as a "random"
325 	 * number to index into the array, and there isn't any other
326 	 * significance to the chosen tx_cpu. Because.. Why not use
327 	 * the current cpu to index into the array?
328 	 */
329 	tc = &tx->tx_cpu[CPU_SEQID_UNSTABLE];
330 
331 	mutex_enter(&tc->tc_open_lock);
332 	txg = tx->tx_open_txg;
333 
334 	mutex_enter(&tc->tc_lock);
335 	tc->tc_count[txg & TXG_MASK]++;
336 	mutex_exit(&tc->tc_lock);
337 
338 	th->th_cpu = tc;
339 	th->th_txg = txg;
340 
341 	return (txg);
342 }
343 
344 void
345 txg_rele_to_quiesce(txg_handle_t *th)
346 {
347 	tx_cpu_t *tc = th->th_cpu;
348 
349 	ASSERT(!MUTEX_HELD(&tc->tc_lock));
350 	mutex_exit(&tc->tc_open_lock);
351 }
352 
353 void
354 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
355 {
356 	tx_cpu_t *tc = th->th_cpu;
357 	int g = th->th_txg & TXG_MASK;
358 
359 	mutex_enter(&tc->tc_lock);
360 	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
361 	mutex_exit(&tc->tc_lock);
362 }
363 
364 void
365 txg_rele_to_sync(txg_handle_t *th)
366 {
367 	tx_cpu_t *tc = th->th_cpu;
368 	int g = th->th_txg & TXG_MASK;
369 
370 	mutex_enter(&tc->tc_lock);
371 	ASSERT(tc->tc_count[g] != 0);
372 	if (--tc->tc_count[g] == 0)
373 		cv_broadcast(&tc->tc_cv[g]);
374 	mutex_exit(&tc->tc_lock);
375 
376 	th->th_cpu = NULL;	/* defensive */
377 }
378 
379 /*
380  * Blocks until all transactions in the group are committed.
381  *
382  * On return, the transaction group has reached a stable state in which it can
383  * then be passed off to the syncing context.
384  */
385 static void
386 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
387 {
388 	tx_state_t *tx = &dp->dp_tx;
389 	uint64_t tx_open_time;
390 	int g = txg & TXG_MASK;
391 	int c;
392 
393 	/*
394 	 * Grab all tc_open_locks so nobody else can get into this txg.
395 	 */
396 	for (c = 0; c < max_ncpus; c++)
397 		mutex_enter(&tx->tx_cpu[c].tc_open_lock);
398 
399 	ASSERT(txg == tx->tx_open_txg);
400 	tx->tx_open_txg++;
401 	tx->tx_open_time = tx_open_time = gethrtime();
402 
403 	DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
404 	DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
405 
406 	/*
407 	 * Now that we've incremented tx_open_txg, we can let threads
408 	 * enter the next transaction group.
409 	 */
410 	for (c = 0; c < max_ncpus; c++)
411 		mutex_exit(&tx->tx_cpu[c].tc_open_lock);
412 
413 	spa_txg_history_set(dp->dp_spa, txg, TXG_STATE_OPEN, tx_open_time);
414 	spa_txg_history_add(dp->dp_spa, txg + 1, tx_open_time);
415 
416 	/*
417 	 * Quiesce the transaction group by waiting for everyone to
418 	 * call txg_rele_to_sync() for their open transaction handles.
419 	 */
420 	for (c = 0; c < max_ncpus; c++) {
421 		tx_cpu_t *tc = &tx->tx_cpu[c];
422 		mutex_enter(&tc->tc_lock);
423 		while (tc->tc_count[g] != 0)
424 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
425 		mutex_exit(&tc->tc_lock);
426 	}
427 
428 	spa_txg_history_set(dp->dp_spa, txg, TXG_STATE_QUIESCED, gethrtime());
429 }
430 
431 static void
432 txg_do_callbacks(void *cb_list)
433 {
434 	dmu_tx_do_callbacks(cb_list, 0);
435 
436 	list_destroy(cb_list);
437 
438 	kmem_free(cb_list, sizeof (list_t));
439 }
440 
441 /*
442  * Dispatch the commit callbacks registered on this txg to worker threads.
443  *
444  * If no callbacks are registered for a given TXG, nothing happens.
445  * This function creates a taskq for the associated pool, if needed.
446  */
447 static void
448 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
449 {
450 	int c;
451 	tx_state_t *tx = &dp->dp_tx;
452 	list_t *cb_list;
453 
454 	for (c = 0; c < max_ncpus; c++) {
455 		tx_cpu_t *tc = &tx->tx_cpu[c];
456 		/*
457 		 * No need to lock tx_cpu_t at this point, since this can
458 		 * only be called once a txg has been synced.
459 		 */
460 
461 		int g = txg & TXG_MASK;
462 
463 		if (list_is_empty(&tc->tc_callbacks[g]))
464 			continue;
465 
466 		if (tx->tx_commit_cb_taskq == NULL) {
467 			/*
468 			 * Commit callback taskq hasn't been created yet.
469 			 */
470 			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
471 			    100, defclsyspri, boot_ncpus, boot_ncpus * 2,
472 			    TASKQ_PREPOPULATE | TASKQ_DYNAMIC |
473 			    TASKQ_THREADS_CPU_PCT);
474 		}
475 
476 		cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
477 		list_create(cb_list, sizeof (dmu_tx_callback_t),
478 		    offsetof(dmu_tx_callback_t, dcb_node));
479 
480 		list_move_tail(cb_list, &tc->tc_callbacks[g]);
481 
482 		(void) taskq_dispatch(tx->tx_commit_cb_taskq,
483 		    txg_do_callbacks, cb_list, TQ_SLEEP);
484 	}
485 }
486 
487 /*
488  * Wait for pending commit callbacks of already-synced transactions to finish
489  * processing.
490  * Calling this function from within a commit callback will deadlock.
491  */
492 void
493 txg_wait_callbacks(dsl_pool_t *dp)
494 {
495 	tx_state_t *tx = &dp->dp_tx;
496 
497 	if (tx->tx_commit_cb_taskq != NULL)
498 		taskq_wait_outstanding(tx->tx_commit_cb_taskq, 0);
499 }
500 
501 static boolean_t
502 txg_is_quiescing(dsl_pool_t *dp)
503 {
504 	tx_state_t *tx = &dp->dp_tx;
505 	ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
506 	return (tx->tx_quiescing_txg != 0);
507 }
508 
509 static boolean_t
510 txg_has_quiesced_to_sync(dsl_pool_t *dp)
511 {
512 	tx_state_t *tx = &dp->dp_tx;
513 	ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
514 	return (tx->tx_quiesced_txg != 0);
515 }
516 
517 static __attribute__((noreturn)) void
518 txg_sync_thread(void *arg)
519 {
520 	dsl_pool_t *dp = arg;
521 	spa_t *spa = dp->dp_spa;
522 	tx_state_t *tx = &dp->dp_tx;
523 	callb_cpr_t cpr;
524 	clock_t start, delta;
525 
526 	(void) spl_fstrans_mark();
527 	txg_thread_enter(tx, &cpr);
528 
529 	start = delta = 0;
530 	for (;;) {
531 		clock_t timeout = zfs_txg_timeout * hz;
532 		clock_t timer;
533 		uint64_t txg;
534 
535 		/*
536 		 * We sync when we're scanning, there's someone waiting
537 		 * on us, or the quiesce thread has handed off a txg to
538 		 * us, or we have reached our timeout.
539 		 */
540 		timer = (delta >= timeout ? 0 : timeout - delta);
541 		while (!dsl_scan_active(dp->dp_scan) &&
542 		    !tx->tx_exiting && timer > 0 &&
543 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
544 		    !txg_has_quiesced_to_sync(dp)) {
545 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
546 			    (u_longlong_t)tx->tx_synced_txg,
547 			    (u_longlong_t)tx->tx_sync_txg_waiting, dp);
548 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
549 			delta = ddi_get_lbolt() - start;
550 			timer = (delta > timeout ? 0 : timeout - delta);
551 		}
552 
553 		/*
554 		 * When we're suspended, nothing should be changing and for
555 		 * MMP we don't want to bump anything that would make it
556 		 * harder to detect if another host is changing it when
557 		 * resuming after a MMP suspend.
558 		 */
559 		if (spa_suspended(spa))
560 			continue;
561 
562 		/*
563 		 * Wait until the quiesce thread hands off a txg to us,
564 		 * prompting it to do so if necessary.
565 		 */
566 		while (!tx->tx_exiting && !txg_has_quiesced_to_sync(dp)) {
567 			if (txg_is_quiescing(dp)) {
568 				txg_thread_wait(tx, &cpr,
569 				    &tx->tx_quiesce_done_cv, 0);
570 				continue;
571 			}
572 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
573 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
574 			cv_broadcast(&tx->tx_quiesce_more_cv);
575 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
576 		}
577 
578 		if (tx->tx_exiting)
579 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
580 
581 		/*
582 		 * Consume the quiesced txg which has been handed off to
583 		 * us.  This may cause the quiescing thread to now be
584 		 * able to quiesce another txg, so we must signal it.
585 		 */
586 		ASSERT(tx->tx_quiesced_txg != 0);
587 		txg = tx->tx_quiesced_txg;
588 		tx->tx_quiesced_txg = 0;
589 		tx->tx_syncing_txg = txg;
590 		DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
591 		cv_broadcast(&tx->tx_quiesce_more_cv);
592 
593 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
594 		    (u_longlong_t)txg, (u_longlong_t)tx->tx_quiesce_txg_waiting,
595 		    (u_longlong_t)tx->tx_sync_txg_waiting);
596 		mutex_exit(&tx->tx_sync_lock);
597 
598 		txg_stat_t *ts = spa_txg_history_init_io(spa, txg, dp);
599 		start = ddi_get_lbolt();
600 		spa_sync(spa, txg);
601 		delta = ddi_get_lbolt() - start;
602 		spa_txg_history_fini_io(spa, ts);
603 
604 		mutex_enter(&tx->tx_sync_lock);
605 		tx->tx_synced_txg = txg;
606 		tx->tx_syncing_txg = 0;
607 		DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
608 		cv_broadcast(&tx->tx_sync_done_cv);
609 
610 		/*
611 		 * Dispatch commit callbacks to worker threads.
612 		 */
613 		txg_dispatch_callbacks(dp, txg);
614 	}
615 }
616 
617 static __attribute__((noreturn)) void
618 txg_quiesce_thread(void *arg)
619 {
620 	dsl_pool_t *dp = arg;
621 	tx_state_t *tx = &dp->dp_tx;
622 	callb_cpr_t cpr;
623 
624 	txg_thread_enter(tx, &cpr);
625 
626 	for (;;) {
627 		uint64_t txg;
628 
629 		/*
630 		 * We quiesce when there's someone waiting on us.
631 		 * However, we can only have one txg in "quiescing" or
632 		 * "quiesced, waiting to sync" state.  So we wait until
633 		 * the "quiesced, waiting to sync" txg has been consumed
634 		 * by the sync thread.
635 		 */
636 		while (!tx->tx_exiting &&
637 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
638 		    txg_has_quiesced_to_sync(dp)))
639 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
640 
641 		if (tx->tx_exiting)
642 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
643 
644 		txg = tx->tx_open_txg;
645 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
646 		    (u_longlong_t)txg,
647 		    (u_longlong_t)tx->tx_quiesce_txg_waiting,
648 		    (u_longlong_t)tx->tx_sync_txg_waiting);
649 		tx->tx_quiescing_txg = txg;
650 
651 		mutex_exit(&tx->tx_sync_lock);
652 		txg_quiesce(dp, txg);
653 		mutex_enter(&tx->tx_sync_lock);
654 
655 		/*
656 		 * Hand this txg off to the sync thread.
657 		 */
658 		dprintf("quiesce done, handing off txg %llu\n",
659 		    (u_longlong_t)txg);
660 		tx->tx_quiescing_txg = 0;
661 		tx->tx_quiesced_txg = txg;
662 		DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
663 		cv_broadcast(&tx->tx_sync_more_cv);
664 		cv_broadcast(&tx->tx_quiesce_done_cv);
665 	}
666 }
667 
668 /*
669  * Delay this thread by delay nanoseconds if we are still in the open
670  * transaction group and there is already a waiting txg quiescing or quiesced.
671  * Abort the delay if this txg stalls or enters the quiescing state.
672  */
673 void
674 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
675 {
676 	tx_state_t *tx = &dp->dp_tx;
677 	hrtime_t start = gethrtime();
678 
679 	/* don't delay if this txg could transition to quiescing immediately */
680 	if (tx->tx_open_txg > txg ||
681 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
682 		return;
683 
684 	mutex_enter(&tx->tx_sync_lock);
685 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
686 		mutex_exit(&tx->tx_sync_lock);
687 		return;
688 	}
689 
690 	while (gethrtime() - start < delay &&
691 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
692 		(void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
693 		    &tx->tx_sync_lock, delay, resolution, 0);
694 	}
695 
696 	DMU_TX_STAT_BUMP(dmu_tx_delay);
697 
698 	mutex_exit(&tx->tx_sync_lock);
699 }
700 
701 static boolean_t
702 txg_wait_synced_impl(dsl_pool_t *dp, uint64_t txg, boolean_t wait_sig)
703 {
704 	tx_state_t *tx = &dp->dp_tx;
705 
706 	ASSERT(!dsl_pool_config_held(dp));
707 
708 	mutex_enter(&tx->tx_sync_lock);
709 	ASSERT3U(tx->tx_threads, ==, 2);
710 	if (txg == 0)
711 		txg = tx->tx_open_txg + TXG_DEFER_SIZE;
712 	if (tx->tx_sync_txg_waiting < txg)
713 		tx->tx_sync_txg_waiting = txg;
714 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
715 	    (u_longlong_t)txg, (u_longlong_t)tx->tx_quiesce_txg_waiting,
716 	    (u_longlong_t)tx->tx_sync_txg_waiting);
717 	while (tx->tx_synced_txg < txg) {
718 		dprintf("broadcasting sync more "
719 		    "tx_synced=%llu waiting=%llu dp=%px\n",
720 		    (u_longlong_t)tx->tx_synced_txg,
721 		    (u_longlong_t)tx->tx_sync_txg_waiting, dp);
722 		cv_broadcast(&tx->tx_sync_more_cv);
723 		if (wait_sig) {
724 			/*
725 			 * Condition wait here but stop if the thread receives a
726 			 * signal. The caller may call txg_wait_synced*() again
727 			 * to resume waiting for this txg.
728 			 */
729 			if (cv_wait_io_sig(&tx->tx_sync_done_cv,
730 			    &tx->tx_sync_lock) == 0) {
731 				mutex_exit(&tx->tx_sync_lock);
732 				return (B_TRUE);
733 			}
734 		} else {
735 			cv_wait_io(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
736 		}
737 	}
738 	mutex_exit(&tx->tx_sync_lock);
739 	return (B_FALSE);
740 }
741 
742 void
743 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
744 {
745 	VERIFY0(txg_wait_synced_impl(dp, txg, B_FALSE));
746 }
747 
748 /*
749  * Similar to a txg_wait_synced but it can be interrupted from a signal.
750  * Returns B_TRUE if the thread was signaled while waiting.
751  */
752 boolean_t
753 txg_wait_synced_sig(dsl_pool_t *dp, uint64_t txg)
754 {
755 	return (txg_wait_synced_impl(dp, txg, B_TRUE));
756 }
757 
758 /*
759  * Wait for the specified open transaction group.  Set should_quiesce
760  * when the current open txg should be quiesced immediately.
761  */
762 void
763 txg_wait_open(dsl_pool_t *dp, uint64_t txg, boolean_t should_quiesce)
764 {
765 	tx_state_t *tx = &dp->dp_tx;
766 
767 	ASSERT(!dsl_pool_config_held(dp));
768 
769 	mutex_enter(&tx->tx_sync_lock);
770 	ASSERT3U(tx->tx_threads, ==, 2);
771 	if (txg == 0)
772 		txg = tx->tx_open_txg + 1;
773 	if (tx->tx_quiesce_txg_waiting < txg && should_quiesce)
774 		tx->tx_quiesce_txg_waiting = txg;
775 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
776 	    (u_longlong_t)txg, (u_longlong_t)tx->tx_quiesce_txg_waiting,
777 	    (u_longlong_t)tx->tx_sync_txg_waiting);
778 	while (tx->tx_open_txg < txg) {
779 		cv_broadcast(&tx->tx_quiesce_more_cv);
780 		/*
781 		 * Callers setting should_quiesce will use cv_wait_io() and
782 		 * be accounted for as iowait time.  Otherwise, the caller is
783 		 * understood to be idle and cv_wait_sig() is used to prevent
784 		 * incorrectly inflating the system load average.
785 		 */
786 		if (should_quiesce == B_TRUE) {
787 			cv_wait_io(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
788 		} else {
789 			cv_wait_idle(&tx->tx_quiesce_done_cv,
790 			    &tx->tx_sync_lock);
791 		}
792 	}
793 	mutex_exit(&tx->tx_sync_lock);
794 }
795 
796 /*
797  * Pass in the txg number that should be synced.
798  */
799 void
800 txg_kick(dsl_pool_t *dp, uint64_t txg)
801 {
802 	tx_state_t *tx = &dp->dp_tx;
803 
804 	ASSERT(!dsl_pool_config_held(dp));
805 
806 	if (tx->tx_sync_txg_waiting >= txg)
807 		return;
808 
809 	mutex_enter(&tx->tx_sync_lock);
810 	if (tx->tx_sync_txg_waiting < txg) {
811 		tx->tx_sync_txg_waiting = txg;
812 		cv_broadcast(&tx->tx_sync_more_cv);
813 	}
814 	mutex_exit(&tx->tx_sync_lock);
815 }
816 
817 boolean_t
818 txg_stalled(dsl_pool_t *dp)
819 {
820 	tx_state_t *tx = &dp->dp_tx;
821 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
822 }
823 
824 boolean_t
825 txg_sync_waiting(dsl_pool_t *dp)
826 {
827 	tx_state_t *tx = &dp->dp_tx;
828 
829 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
830 	    tx->tx_quiesced_txg != 0);
831 }
832 
833 /*
834  * Verify that this txg is active (open, quiescing, syncing).  Non-active
835  * txg's should not be manipulated.
836  */
837 #ifdef ZFS_DEBUG
838 void
839 txg_verify(spa_t *spa, uint64_t txg)
840 {
841 	dsl_pool_t *dp __maybe_unused = spa_get_dsl(spa);
842 	if (txg <= TXG_INITIAL || txg == ZILTEST_TXG)
843 		return;
844 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
845 	ASSERT3U(txg, >=, dp->dp_tx.tx_synced_txg);
846 	ASSERT3U(txg, >=, dp->dp_tx.tx_open_txg - TXG_CONCURRENT_STATES);
847 }
848 #endif
849 
850 /*
851  * Per-txg object lists.
852  */
853 void
854 txg_list_create(txg_list_t *tl, spa_t *spa, size_t offset)
855 {
856 	int t;
857 
858 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
859 
860 	tl->tl_offset = offset;
861 	tl->tl_spa = spa;
862 
863 	for (t = 0; t < TXG_SIZE; t++)
864 		tl->tl_head[t] = NULL;
865 }
866 
867 static boolean_t
868 txg_list_empty_impl(txg_list_t *tl, uint64_t txg)
869 {
870 	ASSERT(MUTEX_HELD(&tl->tl_lock));
871 	TXG_VERIFY(tl->tl_spa, txg);
872 	return (tl->tl_head[txg & TXG_MASK] == NULL);
873 }
874 
875 boolean_t
876 txg_list_empty(txg_list_t *tl, uint64_t txg)
877 {
878 	mutex_enter(&tl->tl_lock);
879 	boolean_t ret = txg_list_empty_impl(tl, txg);
880 	mutex_exit(&tl->tl_lock);
881 
882 	return (ret);
883 }
884 
885 void
886 txg_list_destroy(txg_list_t *tl)
887 {
888 	int t;
889 
890 	mutex_enter(&tl->tl_lock);
891 	for (t = 0; t < TXG_SIZE; t++)
892 		ASSERT(txg_list_empty_impl(tl, t));
893 	mutex_exit(&tl->tl_lock);
894 
895 	mutex_destroy(&tl->tl_lock);
896 }
897 
898 /*
899  * Returns true if all txg lists are empty.
900  *
901  * Warning: this is inherently racy (an item could be added immediately
902  * after this function returns).
903  */
904 boolean_t
905 txg_all_lists_empty(txg_list_t *tl)
906 {
907 	boolean_t res = B_TRUE;
908 	for (int i = 0; i < TXG_SIZE; i++)
909 		res &= (tl->tl_head[i] == NULL);
910 	return (res);
911 }
912 
913 /*
914  * Add an entry to the list (unless it's already on the list).
915  * Returns B_TRUE if it was actually added.
916  */
917 boolean_t
918 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
919 {
920 	int t = txg & TXG_MASK;
921 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
922 	boolean_t add;
923 
924 	TXG_VERIFY(tl->tl_spa, txg);
925 	mutex_enter(&tl->tl_lock);
926 	add = (tn->tn_member[t] == 0);
927 	if (add) {
928 		tn->tn_member[t] = 1;
929 		tn->tn_next[t] = tl->tl_head[t];
930 		tl->tl_head[t] = tn;
931 	}
932 	mutex_exit(&tl->tl_lock);
933 
934 	return (add);
935 }
936 
937 /*
938  * Add an entry to the end of the list, unless it's already on the list.
939  * (walks list to find end)
940  * Returns B_TRUE if it was actually added.
941  */
942 boolean_t
943 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
944 {
945 	int t = txg & TXG_MASK;
946 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
947 	boolean_t add;
948 
949 	TXG_VERIFY(tl->tl_spa, txg);
950 	mutex_enter(&tl->tl_lock);
951 	add = (tn->tn_member[t] == 0);
952 	if (add) {
953 		txg_node_t **tp;
954 
955 		for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
956 			continue;
957 
958 		tn->tn_member[t] = 1;
959 		tn->tn_next[t] = NULL;
960 		*tp = tn;
961 	}
962 	mutex_exit(&tl->tl_lock);
963 
964 	return (add);
965 }
966 
967 /*
968  * Remove the head of the list and return it.
969  */
970 void *
971 txg_list_remove(txg_list_t *tl, uint64_t txg)
972 {
973 	int t = txg & TXG_MASK;
974 	txg_node_t *tn;
975 	void *p = NULL;
976 
977 	TXG_VERIFY(tl->tl_spa, txg);
978 	mutex_enter(&tl->tl_lock);
979 	if ((tn = tl->tl_head[t]) != NULL) {
980 		ASSERT(tn->tn_member[t]);
981 		ASSERT(tn->tn_next[t] == NULL || tn->tn_next[t]->tn_member[t]);
982 		p = (char *)tn - tl->tl_offset;
983 		tl->tl_head[t] = tn->tn_next[t];
984 		tn->tn_next[t] = NULL;
985 		tn->tn_member[t] = 0;
986 	}
987 	mutex_exit(&tl->tl_lock);
988 
989 	return (p);
990 }
991 
992 /*
993  * Remove a specific item from the list and return it.
994  */
995 void *
996 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
997 {
998 	int t = txg & TXG_MASK;
999 	txg_node_t *tn, **tp;
1000 
1001 	TXG_VERIFY(tl->tl_spa, txg);
1002 	mutex_enter(&tl->tl_lock);
1003 
1004 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
1005 		if ((char *)tn - tl->tl_offset == p) {
1006 			*tp = tn->tn_next[t];
1007 			tn->tn_next[t] = NULL;
1008 			tn->tn_member[t] = 0;
1009 			mutex_exit(&tl->tl_lock);
1010 			return (p);
1011 		}
1012 	}
1013 
1014 	mutex_exit(&tl->tl_lock);
1015 
1016 	return (NULL);
1017 }
1018 
1019 boolean_t
1020 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
1021 {
1022 	int t = txg & TXG_MASK;
1023 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
1024 
1025 	TXG_VERIFY(tl->tl_spa, txg);
1026 	return (tn->tn_member[t] != 0);
1027 }
1028 
1029 /*
1030  * Walk a txg list
1031  */
1032 void *
1033 txg_list_head(txg_list_t *tl, uint64_t txg)
1034 {
1035 	int t = txg & TXG_MASK;
1036 	txg_node_t *tn;
1037 
1038 	mutex_enter(&tl->tl_lock);
1039 	tn = tl->tl_head[t];
1040 	mutex_exit(&tl->tl_lock);
1041 
1042 	TXG_VERIFY(tl->tl_spa, txg);
1043 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
1044 }
1045 
1046 void *
1047 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
1048 {
1049 	int t = txg & TXG_MASK;
1050 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
1051 
1052 	TXG_VERIFY(tl->tl_spa, txg);
1053 
1054 	mutex_enter(&tl->tl_lock);
1055 	tn = tn->tn_next[t];
1056 	mutex_exit(&tl->tl_lock);
1057 
1058 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
1059 }
1060 
1061 EXPORT_SYMBOL(txg_init);
1062 EXPORT_SYMBOL(txg_fini);
1063 EXPORT_SYMBOL(txg_sync_start);
1064 EXPORT_SYMBOL(txg_sync_stop);
1065 EXPORT_SYMBOL(txg_hold_open);
1066 EXPORT_SYMBOL(txg_rele_to_quiesce);
1067 EXPORT_SYMBOL(txg_rele_to_sync);
1068 EXPORT_SYMBOL(txg_register_callbacks);
1069 EXPORT_SYMBOL(txg_delay);
1070 EXPORT_SYMBOL(txg_wait_synced);
1071 EXPORT_SYMBOL(txg_wait_open);
1072 EXPORT_SYMBOL(txg_wait_callbacks);
1073 EXPORT_SYMBOL(txg_stalled);
1074 EXPORT_SYMBOL(txg_sync_waiting);
1075 
1076 ZFS_MODULE_PARAM(zfs_txg, zfs_txg_, timeout, UINT, ZMOD_RW,
1077 	"Max seconds worth of delta per txg");
1078