xref: /titanic_51/usr/src/uts/common/fs/zfs/txg.c (revision bafd1f1462c49949e0251d74b4fbfa24d29bc79a)
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, 2014 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/callb.h>
34 
35 /*
36  * ZFS Transaction Groups
37  * ----------------------
38  *
39  * ZFS transaction groups are, as the name implies, groups of transactions
40  * that act on persistent state. ZFS asserts consistency at the granularity of
41  * these transaction groups. Each successive transaction group (txg) is
42  * assigned a 64-bit consecutive identifier. There are three active
43  * transaction group states: open, quiescing, or syncing. At any given time,
44  * there may be an active txg associated with each state; each active txg may
45  * either be processing, or blocked waiting to enter the next state. There may
46  * be up to three active txgs, and there is always a txg in the open state
47  * (though it may be blocked waiting to enter the quiescing state). In broad
48  * strokes, transactions -- operations that change in-memory structures -- are
49  * accepted into the txg in the open state, and are completed while the txg is
50  * in the open or quiescing states. The accumulated changes are written to
51  * disk in the syncing state.
52  *
53  * Open
54  *
55  * When a new txg becomes active, it first enters the open state. New
56  * transactions -- updates to in-memory structures -- are assigned to the
57  * currently open txg. There is always a txg in the open state so that ZFS can
58  * accept new changes (though the txg may refuse new changes if it has hit
59  * some limit). ZFS advances the open txg to the next state for a variety of
60  * reasons such as it hitting a time or size threshold, or the execution of an
61  * administrative action that must be completed in the syncing state.
62  *
63  * Quiescing
64  *
65  * After a txg exits the open state, it enters the quiescing state. The
66  * quiescing state is intended to provide a buffer between accepting new
67  * transactions in the open state and writing them out to stable storage in
68  * the syncing state. While quiescing, transactions can continue their
69  * operation without delaying either of the other states. Typically, a txg is
70  * in the quiescing state very briefly since the operations are bounded by
71  * software latencies rather than, say, slower I/O latencies. After all
72  * transactions complete, the txg is ready to enter the next state.
73  *
74  * Syncing
75  *
76  * In the syncing state, the in-memory state built up during the open and (to
77  * a lesser degree) the quiescing states is written to stable storage. The
78  * process of writing out modified data can, in turn modify more data. For
79  * example when we write new blocks, we need to allocate space for them; those
80  * allocations modify metadata (space maps)... which themselves must be
81  * written to stable storage. During the sync state, ZFS iterates, writing out
82  * data until it converges and all in-memory changes have been written out.
83  * The first such pass is the largest as it encompasses all the modified user
84  * data (as opposed to filesystem metadata). Subsequent passes typically have
85  * far less data to write as they consist exclusively of filesystem metadata.
86  *
87  * To ensure convergence, after a certain number of passes ZFS begins
88  * overwriting locations on stable storage that had been allocated earlier in
89  * the syncing state (and subsequently freed). ZFS usually allocates new
90  * blocks to optimize for large, continuous, writes. For the syncing state to
91  * converge however it must complete a pass where no new blocks are allocated
92  * since each allocation requires a modification of persistent metadata.
93  * Further, to hasten convergence, after a prescribed number of passes, ZFS
94  * also defers frees, and stops compressing.
95  *
96  * In addition to writing out user data, we must also execute synctasks during
97  * the syncing context. A synctask is the mechanism by which some
98  * administrative activities work such as creating and destroying snapshots or
99  * datasets. Note that when a synctask is initiated it enters the open txg,
100  * and ZFS then pushes that txg as quickly as possible to completion of the
101  * syncing state in order to reduce the latency of the administrative
102  * activity. To complete the syncing state, ZFS writes out a new uberblock,
103  * the root of the tree of blocks that comprise all state stored on the ZFS
104  * pool. Finally, if there is a quiesced txg waiting, we signal that it can
105  * now transition to the syncing state.
106  */
107 
108 static void txg_sync_thread(dsl_pool_t *dp);
109 static void txg_quiesce_thread(dsl_pool_t *dp);
110 
111 int zfs_txg_timeout = 5;	/* max seconds worth of delta per txg */
112 
113 /*
114  * Prepare the txg subsystem.
115  */
116 void
117 txg_init(dsl_pool_t *dp, uint64_t txg)
118 {
119 	tx_state_t *tx = &dp->dp_tx;
120 	int c;
121 	bzero(tx, sizeof (tx_state_t));
122 
123 	tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
124 
125 	for (c = 0; c < max_ncpus; c++) {
126 		int i;
127 
128 		mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
129 		mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT,
130 		    NULL);
131 		for (i = 0; i < TXG_SIZE; i++) {
132 			cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
133 			    NULL);
134 			list_create(&tx->tx_cpu[c].tc_callbacks[i],
135 			    sizeof (dmu_tx_callback_t),
136 			    offsetof(dmu_tx_callback_t, dcb_node));
137 		}
138 	}
139 
140 	mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
141 
142 	cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
143 	cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
144 	cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
145 	cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
146 	cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
147 
148 	tx->tx_open_txg = txg;
149 }
150 
151 /*
152  * Close down the txg subsystem.
153  */
154 void
155 txg_fini(dsl_pool_t *dp)
156 {
157 	tx_state_t *tx = &dp->dp_tx;
158 	int c;
159 
160 	ASSERT(tx->tx_threads == 0);
161 
162 	mutex_destroy(&tx->tx_sync_lock);
163 
164 	cv_destroy(&tx->tx_sync_more_cv);
165 	cv_destroy(&tx->tx_sync_done_cv);
166 	cv_destroy(&tx->tx_quiesce_more_cv);
167 	cv_destroy(&tx->tx_quiesce_done_cv);
168 	cv_destroy(&tx->tx_exit_cv);
169 
170 	for (c = 0; c < max_ncpus; c++) {
171 		int i;
172 
173 		mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
174 		mutex_destroy(&tx->tx_cpu[c].tc_lock);
175 		for (i = 0; i < TXG_SIZE; i++) {
176 			cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
177 			list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
178 		}
179 	}
180 
181 	if (tx->tx_commit_cb_taskq != NULL)
182 		taskq_destroy(tx->tx_commit_cb_taskq);
183 
184 	kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
185 
186 	bzero(tx, sizeof (tx_state_t));
187 }
188 
189 /*
190  * Start syncing transaction groups.
191  */
192 void
193 txg_sync_start(dsl_pool_t *dp)
194 {
195 	tx_state_t *tx = &dp->dp_tx;
196 
197 	mutex_enter(&tx->tx_sync_lock);
198 
199 	dprintf("pool %p\n", dp);
200 
201 	ASSERT(tx->tx_threads == 0);
202 
203 	tx->tx_threads = 2;
204 
205 	tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
206 	    dp, 0, &p0, TS_RUN, minclsyspri);
207 
208 	/*
209 	 * The sync thread can need a larger-than-default stack size on
210 	 * 32-bit x86.  This is due in part to nested pools and
211 	 * scrub_visitbp() recursion.
212 	 */
213 	tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
214 	    dp, 0, &p0, TS_RUN, minclsyspri);
215 
216 	mutex_exit(&tx->tx_sync_lock);
217 }
218 
219 static void
220 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
221 {
222 	CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
223 	mutex_enter(&tx->tx_sync_lock);
224 }
225 
226 static void
227 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
228 {
229 	ASSERT(*tpp != NULL);
230 	*tpp = NULL;
231 	tx->tx_threads--;
232 	cv_broadcast(&tx->tx_exit_cv);
233 	CALLB_CPR_EXIT(cpr);		/* drops &tx->tx_sync_lock */
234 	thread_exit();
235 }
236 
237 static void
238 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
239 {
240 	CALLB_CPR_SAFE_BEGIN(cpr);
241 
242 	if (time)
243 		(void) cv_timedwait(cv, &tx->tx_sync_lock,
244 		    ddi_get_lbolt() + time);
245 	else
246 		cv_wait(cv, &tx->tx_sync_lock);
247 
248 	CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
249 }
250 
251 /*
252  * Stop syncing transaction groups.
253  */
254 void
255 txg_sync_stop(dsl_pool_t *dp)
256 {
257 	tx_state_t *tx = &dp->dp_tx;
258 
259 	dprintf("pool %p\n", dp);
260 	/*
261 	 * Finish off any work in progress.
262 	 */
263 	ASSERT(tx->tx_threads == 2);
264 
265 	/*
266 	 * We need to ensure that we've vacated the deferred space_maps.
267 	 */
268 	txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
269 
270 	/*
271 	 * Wake all sync threads and wait for them to die.
272 	 */
273 	mutex_enter(&tx->tx_sync_lock);
274 
275 	ASSERT(tx->tx_threads == 2);
276 
277 	tx->tx_exiting = 1;
278 
279 	cv_broadcast(&tx->tx_quiesce_more_cv);
280 	cv_broadcast(&tx->tx_quiesce_done_cv);
281 	cv_broadcast(&tx->tx_sync_more_cv);
282 
283 	while (tx->tx_threads != 0)
284 		cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
285 
286 	tx->tx_exiting = 0;
287 
288 	mutex_exit(&tx->tx_sync_lock);
289 }
290 
291 uint64_t
292 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
293 {
294 	tx_state_t *tx = &dp->dp_tx;
295 	tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
296 	uint64_t txg;
297 
298 	mutex_enter(&tc->tc_open_lock);
299 	txg = tx->tx_open_txg;
300 
301 	mutex_enter(&tc->tc_lock);
302 	tc->tc_count[txg & TXG_MASK]++;
303 	mutex_exit(&tc->tc_lock);
304 
305 	th->th_cpu = tc;
306 	th->th_txg = txg;
307 
308 	return (txg);
309 }
310 
311 void
312 txg_rele_to_quiesce(txg_handle_t *th)
313 {
314 	tx_cpu_t *tc = th->th_cpu;
315 
316 	ASSERT(!MUTEX_HELD(&tc->tc_lock));
317 	mutex_exit(&tc->tc_open_lock);
318 }
319 
320 void
321 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
322 {
323 	tx_cpu_t *tc = th->th_cpu;
324 	int g = th->th_txg & TXG_MASK;
325 
326 	mutex_enter(&tc->tc_lock);
327 	list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
328 	mutex_exit(&tc->tc_lock);
329 }
330 
331 void
332 txg_rele_to_sync(txg_handle_t *th)
333 {
334 	tx_cpu_t *tc = th->th_cpu;
335 	int g = th->th_txg & TXG_MASK;
336 
337 	mutex_enter(&tc->tc_lock);
338 	ASSERT(tc->tc_count[g] != 0);
339 	if (--tc->tc_count[g] == 0)
340 		cv_broadcast(&tc->tc_cv[g]);
341 	mutex_exit(&tc->tc_lock);
342 
343 	th->th_cpu = NULL;	/* defensive */
344 }
345 
346 /*
347  * Blocks until all transactions in the group are committed.
348  *
349  * On return, the transaction group has reached a stable state in which it can
350  * then be passed off to the syncing context.
351  */
352 static void
353 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
354 {
355 	tx_state_t *tx = &dp->dp_tx;
356 	int g = txg & TXG_MASK;
357 	int c;
358 
359 	/*
360 	 * Grab all tc_open_locks so nobody else can get into this txg.
361 	 */
362 	for (c = 0; c < max_ncpus; c++)
363 		mutex_enter(&tx->tx_cpu[c].tc_open_lock);
364 
365 	ASSERT(txg == tx->tx_open_txg);
366 	tx->tx_open_txg++;
367 	tx->tx_open_time = gethrtime();
368 
369 	DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
370 	DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
371 
372 	/*
373 	 * Now that we've incremented tx_open_txg, we can let threads
374 	 * enter the next transaction group.
375 	 */
376 	for (c = 0; c < max_ncpus; c++)
377 		mutex_exit(&tx->tx_cpu[c].tc_open_lock);
378 
379 	/*
380 	 * Quiesce the transaction group by waiting for everyone to txg_exit().
381 	 */
382 	for (c = 0; c < max_ncpus; c++) {
383 		tx_cpu_t *tc = &tx->tx_cpu[c];
384 		mutex_enter(&tc->tc_lock);
385 		while (tc->tc_count[g] != 0)
386 			cv_wait(&tc->tc_cv[g], &tc->tc_lock);
387 		mutex_exit(&tc->tc_lock);
388 	}
389 }
390 
391 static void
392 txg_do_callbacks(list_t *cb_list)
393 {
394 	dmu_tx_do_callbacks(cb_list, 0);
395 
396 	list_destroy(cb_list);
397 
398 	kmem_free(cb_list, sizeof (list_t));
399 }
400 
401 /*
402  * Dispatch the commit callbacks registered on this txg to worker threads.
403  *
404  * If no callbacks are registered for a given TXG, nothing happens.
405  * This function creates a taskq for the associated pool, if needed.
406  */
407 static void
408 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
409 {
410 	int c;
411 	tx_state_t *tx = &dp->dp_tx;
412 	list_t *cb_list;
413 
414 	for (c = 0; c < max_ncpus; c++) {
415 		tx_cpu_t *tc = &tx->tx_cpu[c];
416 		/*
417 		 * No need to lock tx_cpu_t at this point, since this can
418 		 * only be called once a txg has been synced.
419 		 */
420 
421 		int g = txg & TXG_MASK;
422 
423 		if (list_is_empty(&tc->tc_callbacks[g]))
424 			continue;
425 
426 		if (tx->tx_commit_cb_taskq == NULL) {
427 			/*
428 			 * Commit callback taskq hasn't been created yet.
429 			 */
430 			tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
431 			    max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
432 			    TASKQ_PREPOPULATE);
433 		}
434 
435 		cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
436 		list_create(cb_list, sizeof (dmu_tx_callback_t),
437 		    offsetof(dmu_tx_callback_t, dcb_node));
438 
439 		list_move_tail(cb_list, &tc->tc_callbacks[g]);
440 
441 		(void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
442 		    txg_do_callbacks, cb_list, TQ_SLEEP);
443 	}
444 }
445 
446 static void
447 txg_sync_thread(dsl_pool_t *dp)
448 {
449 	spa_t *spa = dp->dp_spa;
450 	tx_state_t *tx = &dp->dp_tx;
451 	callb_cpr_t cpr;
452 	uint64_t start, delta;
453 
454 	txg_thread_enter(tx, &cpr);
455 
456 	start = delta = 0;
457 	for (;;) {
458 		uint64_t timeout = zfs_txg_timeout * hz;
459 		uint64_t timer;
460 		uint64_t txg;
461 
462 		/*
463 		 * We sync when we're scanning, there's someone waiting
464 		 * on us, or the quiesce thread has handed off a txg to
465 		 * us, or we have reached our timeout.
466 		 */
467 		timer = (delta >= timeout ? 0 : timeout - delta);
468 		while (!dsl_scan_active(dp->dp_scan) &&
469 		    !tx->tx_exiting && timer > 0 &&
470 		    tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
471 		    tx->tx_quiesced_txg == 0 &&
472 		    dp->dp_dirty_total < zfs_dirty_data_sync) {
473 			dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
474 			    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
475 			txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
476 			delta = ddi_get_lbolt() - start;
477 			timer = (delta > timeout ? 0 : timeout - delta);
478 		}
479 
480 		/*
481 		 * Wait until the quiesce thread hands off a txg to us,
482 		 * prompting it to do so if necessary.
483 		 */
484 		while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
485 			if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
486 				tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
487 			cv_broadcast(&tx->tx_quiesce_more_cv);
488 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
489 		}
490 
491 		if (tx->tx_exiting)
492 			txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
493 
494 		/*
495 		 * Consume the quiesced txg which has been handed off to
496 		 * us.  This may cause the quiescing thread to now be
497 		 * able to quiesce another txg, so we must signal it.
498 		 */
499 		txg = tx->tx_quiesced_txg;
500 		tx->tx_quiesced_txg = 0;
501 		tx->tx_syncing_txg = txg;
502 		DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
503 		cv_broadcast(&tx->tx_quiesce_more_cv);
504 
505 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
506 		    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
507 		mutex_exit(&tx->tx_sync_lock);
508 
509 		start = ddi_get_lbolt();
510 		spa_sync(spa, txg);
511 		delta = ddi_get_lbolt() - start;
512 
513 		mutex_enter(&tx->tx_sync_lock);
514 		tx->tx_synced_txg = txg;
515 		tx->tx_syncing_txg = 0;
516 		DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
517 		cv_broadcast(&tx->tx_sync_done_cv);
518 
519 		/*
520 		 * Dispatch commit callbacks to worker threads.
521 		 */
522 		txg_dispatch_callbacks(dp, txg);
523 	}
524 }
525 
526 static void
527 txg_quiesce_thread(dsl_pool_t *dp)
528 {
529 	tx_state_t *tx = &dp->dp_tx;
530 	callb_cpr_t cpr;
531 
532 	txg_thread_enter(tx, &cpr);
533 
534 	for (;;) {
535 		uint64_t txg;
536 
537 		/*
538 		 * We quiesce when there's someone waiting on us.
539 		 * However, we can only have one txg in "quiescing" or
540 		 * "quiesced, waiting to sync" state.  So we wait until
541 		 * the "quiesced, waiting to sync" txg has been consumed
542 		 * by the sync thread.
543 		 */
544 		while (!tx->tx_exiting &&
545 		    (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
546 		    tx->tx_quiesced_txg != 0))
547 			txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
548 
549 		if (tx->tx_exiting)
550 			txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
551 
552 		txg = tx->tx_open_txg;
553 		dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
554 		    txg, tx->tx_quiesce_txg_waiting,
555 		    tx->tx_sync_txg_waiting);
556 		mutex_exit(&tx->tx_sync_lock);
557 		txg_quiesce(dp, txg);
558 		mutex_enter(&tx->tx_sync_lock);
559 
560 		/*
561 		 * Hand this txg off to the sync thread.
562 		 */
563 		dprintf("quiesce done, handing off txg %llu\n", txg);
564 		tx->tx_quiesced_txg = txg;
565 		DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
566 		cv_broadcast(&tx->tx_sync_more_cv);
567 		cv_broadcast(&tx->tx_quiesce_done_cv);
568 	}
569 }
570 
571 /*
572  * Delay this thread by delay nanoseconds if we are still in the open
573  * transaction group and there is already a waiting txg quiescing or quiesced.
574  * Abort the delay if this txg stalls or enters the quiescing state.
575  */
576 void
577 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
578 {
579 	tx_state_t *tx = &dp->dp_tx;
580 	hrtime_t start = gethrtime();
581 
582 	/* don't delay if this txg could transition to quiescing immediately */
583 	if (tx->tx_open_txg > txg ||
584 	    tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
585 		return;
586 
587 	mutex_enter(&tx->tx_sync_lock);
588 	if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
589 		mutex_exit(&tx->tx_sync_lock);
590 		return;
591 	}
592 
593 	while (gethrtime() - start < delay &&
594 	    tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
595 		(void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
596 		    &tx->tx_sync_lock, delay, resolution, 0);
597 	}
598 
599 	mutex_exit(&tx->tx_sync_lock);
600 }
601 
602 void
603 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
604 {
605 	tx_state_t *tx = &dp->dp_tx;
606 
607 	ASSERT(!dsl_pool_config_held(dp));
608 
609 	mutex_enter(&tx->tx_sync_lock);
610 	ASSERT(tx->tx_threads == 2);
611 	if (txg == 0)
612 		txg = tx->tx_open_txg + TXG_DEFER_SIZE;
613 	if (tx->tx_sync_txg_waiting < txg)
614 		tx->tx_sync_txg_waiting = txg;
615 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
616 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
617 	while (tx->tx_synced_txg < txg) {
618 		dprintf("broadcasting sync more "
619 		    "tx_synced=%llu waiting=%llu dp=%p\n",
620 		    tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
621 		cv_broadcast(&tx->tx_sync_more_cv);
622 		cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
623 	}
624 	mutex_exit(&tx->tx_sync_lock);
625 }
626 
627 void
628 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
629 {
630 	tx_state_t *tx = &dp->dp_tx;
631 
632 	ASSERT(!dsl_pool_config_held(dp));
633 
634 	mutex_enter(&tx->tx_sync_lock);
635 	ASSERT(tx->tx_threads == 2);
636 	if (txg == 0)
637 		txg = tx->tx_open_txg + 1;
638 	if (tx->tx_quiesce_txg_waiting < txg)
639 		tx->tx_quiesce_txg_waiting = txg;
640 	dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
641 	    txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
642 	while (tx->tx_open_txg < txg) {
643 		cv_broadcast(&tx->tx_quiesce_more_cv);
644 		cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
645 	}
646 	mutex_exit(&tx->tx_sync_lock);
647 }
648 
649 /*
650  * If there isn't a txg syncing or in the pipeline, push another txg through
651  * the pipeline by queiscing the open txg.
652  */
653 void
654 txg_kick(dsl_pool_t *dp)
655 {
656 	tx_state_t *tx = &dp->dp_tx;
657 
658 	ASSERT(!dsl_pool_config_held(dp));
659 
660 	mutex_enter(&tx->tx_sync_lock);
661 	if (tx->tx_syncing_txg == 0 &&
662 	    tx->tx_quiesce_txg_waiting <= tx->tx_open_txg &&
663 	    tx->tx_sync_txg_waiting <= tx->tx_synced_txg &&
664 	    tx->tx_quiesced_txg <= tx->tx_synced_txg) {
665 		tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1;
666 		cv_broadcast(&tx->tx_quiesce_more_cv);
667 	}
668 	mutex_exit(&tx->tx_sync_lock);
669 }
670 
671 boolean_t
672 txg_stalled(dsl_pool_t *dp)
673 {
674 	tx_state_t *tx = &dp->dp_tx;
675 	return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
676 }
677 
678 boolean_t
679 txg_sync_waiting(dsl_pool_t *dp)
680 {
681 	tx_state_t *tx = &dp->dp_tx;
682 
683 	return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
684 	    tx->tx_quiesced_txg != 0);
685 }
686 
687 /*
688  * Per-txg object lists.
689  */
690 void
691 txg_list_create(txg_list_t *tl, size_t offset)
692 {
693 	int t;
694 
695 	mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
696 
697 	tl->tl_offset = offset;
698 
699 	for (t = 0; t < TXG_SIZE; t++)
700 		tl->tl_head[t] = NULL;
701 }
702 
703 void
704 txg_list_destroy(txg_list_t *tl)
705 {
706 	int t;
707 
708 	for (t = 0; t < TXG_SIZE; t++)
709 		ASSERT(txg_list_empty(tl, t));
710 
711 	mutex_destroy(&tl->tl_lock);
712 }
713 
714 boolean_t
715 txg_list_empty(txg_list_t *tl, uint64_t txg)
716 {
717 	return (tl->tl_head[txg & TXG_MASK] == NULL);
718 }
719 
720 /*
721  * Returns true if all txg lists are empty.
722  *
723  * Warning: this is inherently racy (an item could be added immediately after this
724  * function returns). We don't bother with the lock because it wouldn't change the
725  * semantics.
726  */
727 boolean_t
728 txg_all_lists_empty(txg_list_t *tl)
729 {
730 	for (int i = 0; i < TXG_SIZE; i++) {
731 		if (!txg_list_empty(tl, i)) {
732 			return (B_FALSE);
733 		}
734 	}
735 	return (B_TRUE);
736 }
737 
738 /*
739  * Add an entry to the list (unless it's already on the list).
740  * Returns B_TRUE if it was actually added.
741  */
742 boolean_t
743 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
744 {
745 	int t = txg & TXG_MASK;
746 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
747 	boolean_t add;
748 
749 	mutex_enter(&tl->tl_lock);
750 	add = (tn->tn_member[t] == 0);
751 	if (add) {
752 		tn->tn_member[t] = 1;
753 		tn->tn_next[t] = tl->tl_head[t];
754 		tl->tl_head[t] = tn;
755 	}
756 	mutex_exit(&tl->tl_lock);
757 
758 	return (add);
759 }
760 
761 /*
762  * Add an entry to the end of the list, unless it's already on the list.
763  * (walks list to find end)
764  * Returns B_TRUE if it was actually added.
765  */
766 boolean_t
767 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
768 {
769 	int t = txg & TXG_MASK;
770 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
771 	boolean_t add;
772 
773 	mutex_enter(&tl->tl_lock);
774 	add = (tn->tn_member[t] == 0);
775 	if (add) {
776 		txg_node_t **tp;
777 
778 		for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
779 			continue;
780 
781 		tn->tn_member[t] = 1;
782 		tn->tn_next[t] = NULL;
783 		*tp = tn;
784 	}
785 	mutex_exit(&tl->tl_lock);
786 
787 	return (add);
788 }
789 
790 /*
791  * Remove the head of the list and return it.
792  */
793 void *
794 txg_list_remove(txg_list_t *tl, uint64_t txg)
795 {
796 	int t = txg & TXG_MASK;
797 	txg_node_t *tn;
798 	void *p = NULL;
799 
800 	mutex_enter(&tl->tl_lock);
801 	if ((tn = tl->tl_head[t]) != NULL) {
802 		p = (char *)tn - tl->tl_offset;
803 		tl->tl_head[t] = tn->tn_next[t];
804 		tn->tn_next[t] = NULL;
805 		tn->tn_member[t] = 0;
806 	}
807 	mutex_exit(&tl->tl_lock);
808 
809 	return (p);
810 }
811 
812 /*
813  * Remove a specific item from the list and return it.
814  */
815 void *
816 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
817 {
818 	int t = txg & TXG_MASK;
819 	txg_node_t *tn, **tp;
820 
821 	mutex_enter(&tl->tl_lock);
822 
823 	for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
824 		if ((char *)tn - tl->tl_offset == p) {
825 			*tp = tn->tn_next[t];
826 			tn->tn_next[t] = NULL;
827 			tn->tn_member[t] = 0;
828 			mutex_exit(&tl->tl_lock);
829 			return (p);
830 		}
831 	}
832 
833 	mutex_exit(&tl->tl_lock);
834 
835 	return (NULL);
836 }
837 
838 boolean_t
839 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
840 {
841 	int t = txg & TXG_MASK;
842 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
843 
844 	return (tn->tn_member[t] != 0);
845 }
846 
847 /*
848  * Walk a txg list -- only safe if you know it's not changing.
849  */
850 void *
851 txg_list_head(txg_list_t *tl, uint64_t txg)
852 {
853 	int t = txg & TXG_MASK;
854 	txg_node_t *tn = tl->tl_head[t];
855 
856 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
857 }
858 
859 void *
860 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
861 {
862 	int t = txg & TXG_MASK;
863 	txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
864 
865 	tn = tn->tn_next[t];
866 
867 	return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
868 }
869