xref: /illumos-gate/usr/src/uts/common/fs/zfs/txg.c (revision d0cb1fb92629bc0283c88d4719df7285c1612700)
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  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Portions Copyright 2011 Martin Matuska
24  * Copyright (c) 2012, 2017 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/dmu_tx.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/zil.h>
34 #include <sys/callb.h>
35 
36 /*
37  * ZFS Transaction Groups
38  * ----------------------
39  *
40  * ZFS transaction groups are, as the name implies, groups of transactions
41  * that act on persistent state. ZFS asserts consistency at the granularity of
42  * these transaction groups. Each successive transaction group (txg) is
43  * assigned a 64-bit consecutive identifier. There are three active
44  * transaction group states: open, quiescing, or syncing. At any given time,
45  * there may be an active txg associated with each state; each active txg may
46  * either be processing, or blocked waiting to enter the next state. There may
47  * be up to three active txgs, and there is always a txg in the open state
48  * (though it may be blocked waiting to enter the quiescing state). In broad
49  * strokes, transactions -- operations that change in-memory structures -- are
50  * accepted into the txg in the open state, and are completed while the txg is
51  * in the open or quiescing states. The accumulated changes are written to
52  * disk in the syncing state.
53  *
54  * Open
55  *
56  * When a new txg becomes active, it first enters the open state. New
57  * transactions -- updates to in-memory structures -- are assigned to the
58  * currently open txg. There is always a txg in the open state so that ZFS can
59  * accept new changes (though the txg may refuse new changes if it has hit
60  * some limit). ZFS advances the open txg to the next state for a variety of
61  * reasons such as it hitting a time or size threshold, or the execution of an
62  * administrative action that must be completed in the syncing state.
63  *
64  * Quiescing
65  *
66  * After a txg exits the open state, it enters the quiescing state. The
67  * quiescing state is intended to provide a buffer between accepting new
68  * transactions in the open state and writing them out to stable storage in
69  * the syncing state. While quiescing, transactions can continue their
70  * operation without delaying either of the other states. Typically, a txg is
71  * in the quiescing state very briefly since the operations are bounded by
72  * software latencies rather than, say, slower I/O latencies. After all
73  * transactions complete, the txg is ready to enter the next state.
74  *
75  * Syncing
76  *
77  * In the syncing state, the in-memory state built up during the open and (to
78  * a lesser degree) the quiescing states is written to stable storage. The
79  * process of writing out modified data can, in turn modify more data. For
80  * example when we write new blocks, we need to allocate space for them; those
81  * allocations modify metadata (space maps)... which themselves must be
82  * written to stable storage. During the sync state, ZFS iterates, writing out
83  * data until it converges and all in-memory changes have been written out.
84  * The first such pass is the largest as it encompasses all the modified user
85  * data (as opposed to filesystem metadata). Subsequent passes typically have
86  * far less data to write as they consist exclusively of filesystem metadata.
87  *
88  * To ensure convergence, after a certain number of passes ZFS begins
89  * overwriting locations on stable storage that had been allocated earlier in
90  * the syncing state (and subsequently freed). ZFS usually allocates new
91  * blocks to optimize for large, continuous, writes. For the syncing state to
92  * converge however it must complete a pass where no new blocks are allocated
93  * since each allocation requires a modification of persistent metadata.
94  * Further, to hasten convergence, after a prescribed number of passes, ZFS
95  * also defers frees, and stops compressing.
96  *
97  * In addition to writing out user data, we must also execute synctasks during
98  * the syncing context. A synctask is the mechanism by which some
99  * administrative activities work such as creating and destroying snapshots or
100  * datasets. Note that when a synctask is initiated it enters the open txg,
101  * and ZFS then pushes that txg as quickly as possible to completion of the
102  * syncing state in order to reduce the latency of the administrative
103  * activity. To complete the syncing state, ZFS writes out a new uberblock,
104  * the root of the tree of blocks that comprise all state stored on the ZFS
105  * pool. Finally, if there is a quiesced txg waiting, we signal that it can
106  * now transition to the syncing state.
107  */
108 
109 static void txg_sync_thread(void *arg);
110 static void txg_quiesce_thread(void *arg);
111 
112 int zfs_txg_timeout = 5;	/* max seconds worth of delta per txg */
113 
114 /*
115  * Prepare the txg subsystem.
116  */
117 void
118 txg_init(dsl_pool_t *dp, uint64_t txg)
119 {
120 	tx_state_t *tx = &dp->dp_tx;
121 	int c;
122 	bzero(tx, sizeof (tx_state_t));
123 
124 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
125 
126 	for (c = 0; c < max_ncpus; c++) {
127 		int i;
128 
129 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
130 		mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT,
131 		    NULL);
132 		for (i = 0; i < TXG_SIZE; i++) {
133 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
134 			    NULL);
135 			list_create(&tx->tx_cpu[c].tc_callbacks[i],
136 			    sizeof (dmu_tx_callback_t),
137 			    offsetof(dmu_tx_callback_t, dcb_node));
138 		}
139 	}
140 
141 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
142 
143 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
144 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
145 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
146 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
147 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
148 
149 	tx->tx_open_txg = txg;
150 }
151 
152 /*
153  * Close down the txg subsystem.
154  */
155 void
156 txg_fini(dsl_pool_t *dp)
157 {
158 	tx_state_t *tx = &dp->dp_tx;
159 	int c;
160 
161 	ASSERT0(tx->tx_threads);
162 
163 	mutex_destroy(&tx->tx_sync_lock);
164 
165 	cv_destroy(&tx->tx_sync_more_cv);
166 	cv_destroy(&tx->tx_sync_done_cv);
167 	cv_destroy(&tx->tx_quiesce_more_cv);
168 	cv_destroy(&tx->tx_quiesce_done_cv);
169 	cv_destroy(&tx->tx_exit_cv);
170 
171 	for (c = 0; c < max_ncpus; c++) {
172 		int i;
173 
174 		mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
175 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
176 		for (i = 0; i < TXG_SIZE; i++) {
177 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
178 			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
179 		}
180 	}
181 
182 	if (tx->tx_commit_cb_taskq != NULL)
183 		taskq_destroy(tx->tx_commit_cb_taskq);
184 
185 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
186 
187 	bzero(tx, sizeof (tx_state_t));
188 }
189 
190 /*
191  * Start syncing transaction groups.
192  */
193 void
194 txg_sync_start(dsl_pool_t *dp)
195 {
196 	tx_state_t *tx = &dp->dp_tx;
197 
198 	mutex_enter(&tx->tx_sync_lock);
199 
200 	dprintf("pool %p\n", dp);
201 
202 	ASSERT0(tx->tx_threads);
203 
204 	tx->tx_threads = 2;
205 
206 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
207 	    dp, 0, &p0, TS_RUN, minclsyspri);
208 
209 	/*
210 	 * The sync thread can need a larger-than-default stack size on
211 	 * 32-bit x86.  This is due in part to nested pools and
212 	 * scrub_visitbp() recursion.
213 	 */
214 	tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
215 	    dp, 0, &p0, TS_RUN, minclsyspri);
216 
217 	mutex_exit(&tx->tx_sync_lock);
218 }
219 
220 static void
221 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
222 {
223 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
224 	mutex_enter(&tx->tx_sync_lock);
225 }
226 
227 static void
228 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
229 {
230 	ASSERT(*tpp != NULL);
231 	*tpp = NULL;
232 	tx->tx_threads--;
233 	cv_broadcast(&tx->tx_exit_cv);
234 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
235 	thread_exit();
236 }
237 
238 static void
239 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
240 {
241 	CALLB_CPR_SAFE_BEGIN(cpr);
242 
243 	if (time)
244 		(void) cv_timedwait(cv, &tx->tx_sync_lock,
245 		    ddi_get_lbolt() + time);
246 	else
247 		cv_wait(cv, &tx->tx_sync_lock);
248 
249 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
250 }
251 
252 /*
253  * Stop syncing transaction groups.
254  */
255 void
256 txg_sync_stop(dsl_pool_t *dp)
257 {
258 	tx_state_t *tx = &dp->dp_tx;
259 
260 	dprintf("pool %p\n", dp);
261 	/*
262 	 * Finish off any work in progress.
263 	 */
264 	ASSERT3U(tx->tx_threads, ==, 2);
265 
266 	/*
267 	 * We need to ensure that we've vacated the deferred space_maps.
268 	 */
269 	txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
270 
271 	/*
272 	 * Wake all sync threads and wait for them to die.
273 	 */
274 	mutex_enter(&tx->tx_sync_lock);
275 
276 	ASSERT3U(tx->tx_threads, ==, 2);
277 
278 	tx->tx_exiting = 1;
279 
280 	cv_broadcast(&tx->tx_quiesce_more_cv);
281 	cv_broadcast(&tx->tx_quiesce_done_cv);
282 	cv_broadcast(&tx->tx_sync_more_cv);
283 
284 	while (tx->tx_threads != 0)
285 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
286 
287 	tx->tx_exiting = 0;
288 
289 	mutex_exit(&tx->tx_sync_lock);
290 }
291 
292 uint64_t
293 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
294 {
295 	tx_state_t *tx = &dp->dp_tx;
296 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
297 	uint64_t txg;
298 
299 	mutex_enter(&tc->tc_open_lock);
300 	txg = tx->tx_open_txg;
301 
302 	mutex_enter(&tc->tc_lock);
303 	tc->tc_count[txg & TXG_MASK]++;
304 	mutex_exit(&tc->tc_lock);
305 
306 	th->th_cpu = tc;
307 	th->th_txg = txg;
308 
309 	return (txg);
310 }
311 
312 void
313 txg_rele_to_quiesce(txg_handle_t *th)
314 {
315 	tx_cpu_t *tc = th->th_cpu;
316 
317 	ASSERT(!MUTEX_HELD(&tc->tc_lock));
318 	mutex_exit(&tc->tc_open_lock);
319 }
320 
321 void
322 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
323 {
324 	tx_cpu_t *tc = th->th_cpu;
325 	int g = th->th_txg & TXG_MASK;
326 
327 	mutex_enter(&tc->tc_lock);
328 	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
329 	mutex_exit(&tc->tc_lock);
330 }
331 
332 void
333 txg_rele_to_sync(txg_handle_t *th)
334 {
335 	tx_cpu_t *tc = th->th_cpu;
336 	int g = th->th_txg & TXG_MASK;
337 
338 	mutex_enter(&tc->tc_lock);
339 	ASSERT(tc->tc_count[g] != 0);
340 	if (--tc->tc_count[g] == 0)
341 		cv_broadcast(&tc->tc_cv[g]);
342 	mutex_exit(&tc->tc_lock);
343 
344 	th->th_cpu = NULL;	/* defensive */
345 }
346 
347 /*
348  * Blocks until all transactions in the group are committed.
349  *
350  * On return, the transaction group has reached a stable state in which it can
351  * then be passed off to the syncing context.
352  */
353 static void
354 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
355 {
356 	tx_state_t *tx = &dp->dp_tx;
357 	int g = txg & TXG_MASK;
358 	int c;
359 
360 	/*
361 	 * Grab all tc_open_locks so nobody else can get into this txg.
362 	 */
363 	for (c = 0; c < max_ncpus; c++)
364 		mutex_enter(&tx->tx_cpu[c].tc_open_lock);
365 
366 	ASSERT(txg == tx->tx_open_txg);
367 	tx->tx_open_txg++;
368 	tx->tx_open_time = gethrtime();
369 
370 	DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
371 	DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
372 
373 	/*
374 	 * Now that we've incremented tx_open_txg, we can let threads
375 	 * enter the next transaction group.
376 	 */
377 	for (c = 0; c < max_ncpus; c++)
378 		mutex_exit(&tx->tx_cpu[c].tc_open_lock);
379 
380 	/*
381 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
382 	 */
383 	for (c = 0; c < max_ncpus; c++) {
384 		tx_cpu_t *tc = &tx->tx_cpu[c];
385 		mutex_enter(&tc->tc_lock);
386 		while (tc->tc_count[g] != 0)
387 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
388 		mutex_exit(&tc->tc_lock);
389 	}
390 }
391 
392 static void
393 txg_do_callbacks(list_t *cb_list)
394 {
395 	dmu_tx_do_callbacks(cb_list, 0);
396 
397 	list_destroy(cb_list);
398 
399 	kmem_free(cb_list, sizeof (list_t));
400 }
401 
402 /*
403  * Dispatch the commit callbacks registered on this txg to worker threads.
404  *
405  * If no callbacks are registered for a given TXG, nothing happens.
406  * This function creates a taskq for the associated pool, if needed.
407  */
408 static void
409 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
410 {
411 	int c;
412 	tx_state_t *tx = &dp->dp_tx;
413 	list_t *cb_list;
414 
415 	for (c = 0; c < max_ncpus; c++) {
416 		tx_cpu_t *tc = &tx->tx_cpu[c];
417 		/*
418 		 * No need to lock tx_cpu_t at this point, since this can
419 		 * only be called once a txg has been synced.
420 		 */
421 
422 		int g = txg & TXG_MASK;
423 
424 		if (list_is_empty(&tc->tc_callbacks[g]))
425 			continue;
426 
427 		if (tx->tx_commit_cb_taskq == NULL) {
428 			/*
429 			 * Commit callback taskq hasn't been created yet.
430 			 */
431 			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
432 			    max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
433 			    TASKQ_PREPOPULATE);
434 		}
435 
436 		cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
437 		list_create(cb_list, sizeof (dmu_tx_callback_t),
438 		    offsetof(dmu_tx_callback_t, dcb_node));
439 
440 		list_move_tail(cb_list, &tc->tc_callbacks[g]);
441 
442 		(void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
443 		    txg_do_callbacks, cb_list, TQ_SLEEP);
444 	}
445 }
446 
447 static boolean_t
448 txg_is_syncing(dsl_pool_t *dp)
449 {
450 	tx_state_t *tx = &dp->dp_tx;
451 	ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
452 	return (tx->tx_syncing_txg != 0);
453 }
454 
455 static boolean_t
456 txg_is_quiescing(dsl_pool_t *dp)
457 {
458 	tx_state_t *tx = &dp->dp_tx;
459 	ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
460 	return (tx->tx_quiescing_txg != 0);
461 }
462 
463 static boolean_t
464 txg_has_quiesced_to_sync(dsl_pool_t *dp)
465 {
466 	tx_state_t *tx = &dp->dp_tx;
467 	ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
468 	return (tx->tx_quiesced_txg != 0);
469 }
470 
471 static void
472 txg_sync_thread(void *arg)
473 {
474 	dsl_pool_t *dp = arg;
475 	spa_t *spa = dp->dp_spa;
476 	tx_state_t *tx = &dp->dp_tx;
477 	callb_cpr_t cpr;
478 	uint64_t start, delta;
479 
480 	txg_thread_enter(tx, &cpr);
481 
482 	start = delta = 0;
483 	for (;;) {
484 		uint64_t timeout = zfs_txg_timeout * hz;
485 		uint64_t timer;
486 		uint64_t txg;
487 		uint64_t dirty_min_bytes =
488 		    zfs_dirty_data_max * zfs_dirty_data_sync_pct / 100;
489 
490 		/*
491 		 * We sync when we're scanning, there's someone waiting
492 		 * on us, or the quiesce thread has handed off a txg to
493 		 * us, or we have reached our timeout.
494 		 */
495 		timer = (delta >= timeout ? 0 : timeout - delta);
496 		while (!dsl_scan_active(dp->dp_scan) &&
497 		    !tx->tx_exiting && timer > 0 &&
498 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
499 		    !txg_has_quiesced_to_sync(dp) &&
500 		    dp->dp_dirty_total < dirty_min_bytes) {
501 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
502 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
503 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
504 			delta = ddi_get_lbolt() - start;
505 			timer = (delta > timeout ? 0 : timeout - delta);
506 		}
507 
508 		/*
509 		 * Wait until the quiesce thread hands off a txg to us,
510 		 * prompting it to do so if necessary.
511 		 */
512 		while (!tx->tx_exiting && !txg_has_quiesced_to_sync(dp)) {
513 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
514 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
515 			cv_broadcast(&tx->tx_quiesce_more_cv);
516 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
517 		}
518 
519 		if (tx->tx_exiting)
520 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
521 
522 		/*
523 		 * Consume the quiesced txg which has been handed off to
524 		 * us.  This may cause the quiescing thread to now be
525 		 * able to quiesce another txg, so we must signal it.
526 		 */
527 		ASSERT(tx->tx_quiesced_txg != 0);
528 		txg = tx->tx_quiesced_txg;
529 		tx->tx_quiesced_txg = 0;
530 		tx->tx_syncing_txg = txg;
531 		DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
532 		cv_broadcast(&tx->tx_quiesce_more_cv);
533 
534 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
535 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
536 		mutex_exit(&tx->tx_sync_lock);
537 
538 		start = ddi_get_lbolt();
539 		spa_sync(spa, txg);
540 		delta = ddi_get_lbolt() - start;
541 
542 		mutex_enter(&tx->tx_sync_lock);
543 		tx->tx_synced_txg = txg;
544 		tx->tx_syncing_txg = 0;
545 		DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
546 		cv_broadcast(&tx->tx_sync_done_cv);
547 
548 		/*
549 		 * Dispatch commit callbacks to worker threads.
550 		 */
551 		txg_dispatch_callbacks(dp, txg);
552 	}
553 }
554 
555 static void
556 txg_quiesce_thread(void *arg)
557 {
558 	dsl_pool_t *dp = arg;
559 	tx_state_t *tx = &dp->dp_tx;
560 	callb_cpr_t cpr;
561 
562 	txg_thread_enter(tx, &cpr);
563 
564 	for (;;) {
565 		uint64_t txg;
566 
567 		/*
568 		 * We quiesce when there's someone waiting on us.
569 		 * However, we can only have one txg in "quiescing" or
570 		 * "quiesced, waiting to sync" state.  So we wait until
571 		 * the "quiesced, waiting to sync" txg has been consumed
572 		 * by the sync thread.
573 		 */
574 		while (!tx->tx_exiting &&
575 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
576 		    txg_has_quiesced_to_sync(dp)))
577 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
578 
579 		if (tx->tx_exiting)
580 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
581 
582 		txg = tx->tx_open_txg;
583 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
584 		    txg, tx->tx_quiesce_txg_waiting,
585 		    tx->tx_sync_txg_waiting);
586 		tx->tx_quiescing_txg = txg;
587 
588 		mutex_exit(&tx->tx_sync_lock);
589 		txg_quiesce(dp, txg);
590 		mutex_enter(&tx->tx_sync_lock);
591 
592 		/*
593 		 * Hand this txg off to the sync thread.
594 		 */
595 		dprintf("quiesce done, handing off txg %llu\n", txg);
596 		tx->tx_quiescing_txg = 0;
597 		tx->tx_quiesced_txg = txg;
598 		DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
599 		cv_broadcast(&tx->tx_sync_more_cv);
600 		cv_broadcast(&tx->tx_quiesce_done_cv);
601 	}
602 }
603 
604 /*
605  * Delay this thread by delay nanoseconds if we are still in the open
606  * transaction group and there is already a waiting txg quiescing or quiesced.
607  * Abort the delay if this txg stalls or enters the quiescing state.
608  */
609 void
610 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
611 {
612 	tx_state_t *tx = &dp->dp_tx;
613 	hrtime_t start = gethrtime();
614 
615 	/* don't delay if this txg could transition to quiescing immediately */
616 	if (tx->tx_open_txg > txg ||
617 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
618 		return;
619 
620 	mutex_enter(&tx->tx_sync_lock);
621 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
622 		mutex_exit(&tx->tx_sync_lock);
623 		return;
624 	}
625 
626 	while (gethrtime() - start < delay &&
627 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
628 		(void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
629 		    &tx->tx_sync_lock, delay, resolution, 0);
630 	}
631 
632 	mutex_exit(&tx->tx_sync_lock);
633 }
634 
635 static boolean_t
636 txg_wait_synced_impl(dsl_pool_t *dp, uint64_t txg, boolean_t wait_sig)
637 {
638 	tx_state_t *tx = &dp->dp_tx;
639 
640 	ASSERT(!dsl_pool_config_held(dp));
641 
642 	mutex_enter(&tx->tx_sync_lock);
643 	ASSERT3U(tx->tx_threads, ==, 2);
644 	if (txg == 0)
645 		txg = tx->tx_open_txg + TXG_DEFER_SIZE;
646 	if (tx->tx_sync_txg_waiting < txg)
647 		tx->tx_sync_txg_waiting = txg;
648 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
649 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
650 	while (tx->tx_synced_txg < txg) {
651 		dprintf("broadcasting sync more "
652 		    "tx_synced=%llu waiting=%llu dp=%p\n",
653 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
654 		cv_broadcast(&tx->tx_sync_more_cv);
655 		if (wait_sig) {
656 			/*
657 			 * Condition wait here but stop if the thread receives a
658 			 * signal. The caller may call txg_wait_synced*() again
659 			 * to resume waiting for this txg.
660 			 */
661 			if (cv_wait_sig(&tx->tx_sync_done_cv,
662 			    &tx->tx_sync_lock) == 0) {
663 				mutex_exit(&tx->tx_sync_lock);
664 				return (B_TRUE);
665 			}
666 		} else {
667 			cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
668 		}
669 	}
670 	mutex_exit(&tx->tx_sync_lock);
671 	return (B_FALSE);
672 }
673 
674 void
675 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
676 {
677 	VERIFY0(txg_wait_synced_impl(dp, txg, B_FALSE));
678 }
679 
680 /*
681  * Similar to a txg_wait_synced but it can be interrupted from a signal.
682  * Returns B_TRUE if the thread was signaled while waiting.
683  */
684 boolean_t
685 txg_wait_synced_sig(dsl_pool_t *dp, uint64_t txg)
686 {
687 	return (txg_wait_synced_impl(dp, txg, B_TRUE));
688 }
689 
690 void
691 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
692 {
693 	tx_state_t *tx = &dp->dp_tx;
694 
695 	ASSERT(!dsl_pool_config_held(dp));
696 
697 	mutex_enter(&tx->tx_sync_lock);
698 	ASSERT3U(tx->tx_threads, ==, 2);
699 	if (txg == 0)
700 		txg = tx->tx_open_txg + 1;
701 	if (tx->tx_quiesce_txg_waiting < txg)
702 		tx->tx_quiesce_txg_waiting = txg;
703 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
704 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
705 	while (tx->tx_open_txg < txg) {
706 		cv_broadcast(&tx->tx_quiesce_more_cv);
707 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
708 	}
709 	mutex_exit(&tx->tx_sync_lock);
710 }
711 
712 /*
713  * If there isn't a txg syncing or in the pipeline, push another txg through
714  * the pipeline by queiscing the open txg.
715  */
716 void
717 txg_kick(dsl_pool_t *dp)
718 {
719 	tx_state_t *tx = &dp->dp_tx;
720 
721 	ASSERT(!dsl_pool_config_held(dp));
722 
723 	mutex_enter(&tx->tx_sync_lock);
724 	if (!txg_is_syncing(dp) &&
725 	    !txg_is_quiescing(dp) &&
726 	    tx->tx_quiesce_txg_waiting <= tx->tx_open_txg &&
727 	    tx->tx_sync_txg_waiting <= tx->tx_synced_txg &&
728 	    tx->tx_quiesced_txg <= tx->tx_synced_txg) {
729 		tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1;
730 		cv_broadcast(&tx->tx_quiesce_more_cv);
731 	}
732 	mutex_exit(&tx->tx_sync_lock);
733 }
734 
735 boolean_t
736 txg_stalled(dsl_pool_t *dp)
737 {
738 	tx_state_t *tx = &dp->dp_tx;
739 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
740 }
741 
742 boolean_t
743 txg_sync_waiting(dsl_pool_t *dp)
744 {
745 	tx_state_t *tx = &dp->dp_tx;
746 
747 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
748 	    tx->tx_quiesced_txg != 0);
749 }
750 
751 /*
752  * Verify that this txg is active (open, quiescing, syncing).  Non-active
753  * txg's should not be manipulated.
754  */
755 void
756 txg_verify(spa_t *spa, uint64_t txg)
757 {
758 	dsl_pool_t *dp = spa_get_dsl(spa);
759 	if (txg <= TXG_INITIAL || txg == ZILTEST_TXG)
760 		return;
761 	ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
762 	ASSERT3U(txg, >=, dp->dp_tx.tx_synced_txg);
763 	ASSERT3U(txg, >=, dp->dp_tx.tx_open_txg - TXG_CONCURRENT_STATES);
764 }
765 
766 /*
767  * Per-txg object lists.
768  */
769 void
770 txg_list_create(txg_list_t *tl, spa_t *spa, size_t offset)
771 {
772 	int t;
773 
774 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
775 
776 	tl->tl_offset = offset;
777 	tl->tl_spa = spa;
778 
779 	for (t = 0; t < TXG_SIZE; t++)
780 		tl->tl_head[t] = NULL;
781 }
782 
783 void
784 txg_list_destroy(txg_list_t *tl)
785 {
786 	int t;
787 
788 	for (t = 0; t < TXG_SIZE; t++)
789 		ASSERT(txg_list_empty(tl, t));
790 
791 	mutex_destroy(&tl->tl_lock);
792 }
793 
794 boolean_t
795 txg_list_empty(txg_list_t *tl, uint64_t txg)
796 {
797 	txg_verify(tl->tl_spa, txg);
798 	return (tl->tl_head[txg & TXG_MASK] == NULL);
799 }
800 
801 /*
802  * Returns true if all txg lists are empty.
803  *
804  * Warning: this is inherently racy (an item could be added immediately
805  * after this function returns). We don't bother with the lock because
806  * it wouldn't change the semantics.
807  */
808 boolean_t
809 txg_all_lists_empty(txg_list_t *tl)
810 {
811 	for (int i = 0; i < TXG_SIZE; i++) {
812 		if (!txg_list_empty(tl, i)) {
813 			return (B_FALSE);
814 		}
815 	}
816 	return (B_TRUE);
817 }
818 
819 /*
820  * Add an entry to the list (unless it's already on the list).
821  * Returns B_TRUE if it was actually added.
822  */
823 boolean_t
824 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
825 {
826 	int t = txg & TXG_MASK;
827 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
828 	boolean_t add;
829 
830 	txg_verify(tl->tl_spa, txg);
831 	mutex_enter(&tl->tl_lock);
832 	add = (tn->tn_member[t] == 0);
833 	if (add) {
834 		tn->tn_member[t] = 1;
835 		tn->tn_next[t] = tl->tl_head[t];
836 		tl->tl_head[t] = tn;
837 	}
838 	mutex_exit(&tl->tl_lock);
839 
840 	return (add);
841 }
842 
843 /*
844  * Add an entry to the end of the list, unless it's already on the list.
845  * (walks list to find end)
846  * Returns B_TRUE if it was actually added.
847  */
848 boolean_t
849 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
850 {
851 	int t = txg & TXG_MASK;
852 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
853 	boolean_t add;
854 
855 	txg_verify(tl->tl_spa, txg);
856 	mutex_enter(&tl->tl_lock);
857 	add = (tn->tn_member[t] == 0);
858 	if (add) {
859 		txg_node_t **tp;
860 
861 		for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
862 			continue;
863 
864 		tn->tn_member[t] = 1;
865 		tn->tn_next[t] = NULL;
866 		*tp = tn;
867 	}
868 	mutex_exit(&tl->tl_lock);
869 
870 	return (add);
871 }
872 
873 /*
874  * Remove the head of the list and return it.
875  */
876 void *
877 txg_list_remove(txg_list_t *tl, uint64_t txg)
878 {
879 	int t = txg & TXG_MASK;
880 	txg_node_t *tn;
881 	void *p = NULL;
882 
883 	txg_verify(tl->tl_spa, txg);
884 	mutex_enter(&tl->tl_lock);
885 	if ((tn = tl->tl_head[t]) != NULL) {
886 		ASSERT(tn->tn_member[t]);
887 		ASSERT(tn->tn_next[t] == NULL || tn->tn_next[t]->tn_member[t]);
888 		p = (char *)tn - tl->tl_offset;
889 		tl->tl_head[t] = tn->tn_next[t];
890 		tn->tn_next[t] = NULL;
891 		tn->tn_member[t] = 0;
892 	}
893 	mutex_exit(&tl->tl_lock);
894 
895 	return (p);
896 }
897 
898 /*
899  * Remove a specific item from the list and return it.
900  */
901 void *
902 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
903 {
904 	int t = txg & TXG_MASK;
905 	txg_node_t *tn, **tp;
906 
907 	txg_verify(tl->tl_spa, txg);
908 	mutex_enter(&tl->tl_lock);
909 
910 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
911 		if ((char *)tn - tl->tl_offset == p) {
912 			*tp = tn->tn_next[t];
913 			tn->tn_next[t] = NULL;
914 			tn->tn_member[t] = 0;
915 			mutex_exit(&tl->tl_lock);
916 			return (p);
917 		}
918 	}
919 
920 	mutex_exit(&tl->tl_lock);
921 
922 	return (NULL);
923 }
924 
925 boolean_t
926 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
927 {
928 	int t = txg & TXG_MASK;
929 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
930 
931 	txg_verify(tl->tl_spa, txg);
932 	return (tn->tn_member[t] != 0);
933 }
934 
935 /*
936  * Walk a txg list -- only safe if you know it's not changing.
937  */
938 void *
939 txg_list_head(txg_list_t *tl, uint64_t txg)
940 {
941 	int t = txg & TXG_MASK;
942 	txg_node_t *tn = tl->tl_head[t];
943 
944 	txg_verify(tl->tl_spa, txg);
945 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
946 }
947 
948 void *
949 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
950 {
951 	int t = txg & TXG_MASK;
952 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
953 
954 	txg_verify(tl->tl_spa, txg);
955 	tn = tn->tn_next[t];
956 
957 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
958 }
959