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.
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.
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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Portions Copyright 2011 Martin Matuska <mm@FreeBSD.org>
24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
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>
34 #include <sys/callb.h>
37 * ZFS Transaction Groups
38 * ----------------------
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.
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.
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.
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.
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.
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.
109 static void txg_sync_thread(void *arg);
110 static void txg_quiesce_thread(void *arg);
112 int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
114 SYSCTL_DECL(_vfs_zfs);
115 SYSCTL_NODE(_vfs_zfs, OID_AUTO, txg, CTLFLAG_RW, 0, "ZFS TXG");
116 SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, timeout, CTLFLAG_RWTUN, &zfs_txg_timeout, 0,
117 "Maximum seconds worth of delta per txg");
120 * Prepare the txg subsystem.
123 txg_init(dsl_pool_t *dp, uint64_t txg)
125 tx_state_t *tx = &dp->dp_tx;
127 bzero(tx, sizeof (tx_state_t));
129 tx->tx_cpu = kmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
131 for (c = 0; c < max_ncpus; c++) {
134 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
135 mutex_init(&tx->tx_cpu[c].tc_open_lock, NULL, MUTEX_DEFAULT,
137 for (i = 0; i < TXG_SIZE; i++) {
138 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
140 list_create(&tx->tx_cpu[c].tc_callbacks[i],
141 sizeof (dmu_tx_callback_t),
142 offsetof(dmu_tx_callback_t, dcb_node));
146 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
148 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
149 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
150 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
151 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
152 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
154 tx->tx_open_txg = txg;
158 * Close down the txg subsystem.
161 txg_fini(dsl_pool_t *dp)
163 tx_state_t *tx = &dp->dp_tx;
166 ASSERT0(tx->tx_threads);
168 mutex_destroy(&tx->tx_sync_lock);
170 cv_destroy(&tx->tx_sync_more_cv);
171 cv_destroy(&tx->tx_sync_done_cv);
172 cv_destroy(&tx->tx_quiesce_more_cv);
173 cv_destroy(&tx->tx_quiesce_done_cv);
174 cv_destroy(&tx->tx_exit_cv);
176 for (c = 0; c < max_ncpus; c++) {
179 mutex_destroy(&tx->tx_cpu[c].tc_open_lock);
180 mutex_destroy(&tx->tx_cpu[c].tc_lock);
181 for (i = 0; i < TXG_SIZE; i++) {
182 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
183 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
187 if (tx->tx_commit_cb_taskq != NULL)
188 taskq_destroy(tx->tx_commit_cb_taskq);
190 kmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
192 bzero(tx, sizeof (tx_state_t));
196 * Start syncing transaction groups.
199 txg_sync_start(dsl_pool_t *dp)
201 tx_state_t *tx = &dp->dp_tx;
203 mutex_enter(&tx->tx_sync_lock);
205 dprintf("pool %p\n", dp);
207 ASSERT0(tx->tx_threads);
211 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
212 dp, 0, &p0, TS_RUN, minclsyspri);
215 * The sync thread can need a larger-than-default stack size on
216 * 32-bit x86. This is due in part to nested pools and
217 * scrub_visitbp() recursion.
219 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
220 dp, 0, &p0, TS_RUN, minclsyspri);
222 mutex_exit(&tx->tx_sync_lock);
226 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
228 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
229 mutex_enter(&tx->tx_sync_lock);
233 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
235 ASSERT(*tpp != NULL);
238 cv_broadcast(&tx->tx_exit_cv);
239 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
244 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, clock_t time)
246 CALLB_CPR_SAFE_BEGIN(cpr);
249 (void) cv_timedwait(cv, &tx->tx_sync_lock, time);
251 cv_wait(cv, &tx->tx_sync_lock);
253 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
257 * Stop syncing transaction groups.
260 txg_sync_stop(dsl_pool_t *dp)
262 tx_state_t *tx = &dp->dp_tx;
264 dprintf("pool %p\n", dp);
266 * Finish off any work in progress.
268 ASSERT3U(tx->tx_threads, ==, 2);
271 * We need to ensure that we've vacated the deferred space_maps.
273 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
276 * Wake all sync threads and wait for them to die.
278 mutex_enter(&tx->tx_sync_lock);
280 ASSERT3U(tx->tx_threads, ==, 2);
284 cv_broadcast(&tx->tx_quiesce_more_cv);
285 cv_broadcast(&tx->tx_quiesce_done_cv);
286 cv_broadcast(&tx->tx_sync_more_cv);
288 while (tx->tx_threads != 0)
289 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
293 mutex_exit(&tx->tx_sync_lock);
297 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
299 tx_state_t *tx = &dp->dp_tx;
300 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
303 mutex_enter(&tc->tc_open_lock);
304 txg = tx->tx_open_txg;
306 mutex_enter(&tc->tc_lock);
307 tc->tc_count[txg & TXG_MASK]++;
308 mutex_exit(&tc->tc_lock);
317 txg_rele_to_quiesce(txg_handle_t *th)
319 tx_cpu_t *tc = th->th_cpu;
321 ASSERT(!MUTEX_HELD(&tc->tc_lock));
322 mutex_exit(&tc->tc_open_lock);
326 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
328 tx_cpu_t *tc = th->th_cpu;
329 int g = th->th_txg & TXG_MASK;
331 mutex_enter(&tc->tc_lock);
332 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
333 mutex_exit(&tc->tc_lock);
337 txg_rele_to_sync(txg_handle_t *th)
339 tx_cpu_t *tc = th->th_cpu;
340 int g = th->th_txg & TXG_MASK;
342 mutex_enter(&tc->tc_lock);
343 ASSERT(tc->tc_count[g] != 0);
344 if (--tc->tc_count[g] == 0)
345 cv_broadcast(&tc->tc_cv[g]);
346 mutex_exit(&tc->tc_lock);
348 th->th_cpu = NULL; /* defensive */
352 * Blocks until all transactions in the group are committed.
354 * On return, the transaction group has reached a stable state in which it can
355 * then be passed off to the syncing context.
357 static __noinline void
358 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
360 tx_state_t *tx = &dp->dp_tx;
361 int g = txg & TXG_MASK;
365 * Grab all tc_open_locks so nobody else can get into this txg.
367 for (c = 0; c < max_ncpus; c++)
368 mutex_enter(&tx->tx_cpu[c].tc_open_lock);
370 ASSERT(txg == tx->tx_open_txg);
372 tx->tx_open_time = gethrtime();
374 DTRACE_PROBE2(txg__quiescing, dsl_pool_t *, dp, uint64_t, txg);
375 DTRACE_PROBE2(txg__opened, dsl_pool_t *, dp, uint64_t, tx->tx_open_txg);
378 * Now that we've incremented tx_open_txg, we can let threads
379 * enter the next transaction group.
381 for (c = 0; c < max_ncpus; c++)
382 mutex_exit(&tx->tx_cpu[c].tc_open_lock);
385 * Quiesce the transaction group by waiting for everyone to txg_exit().
387 for (c = 0; c < max_ncpus; c++) {
388 tx_cpu_t *tc = &tx->tx_cpu[c];
389 mutex_enter(&tc->tc_lock);
390 while (tc->tc_count[g] != 0)
391 cv_wait(&tc->tc_cv[g], &tc->tc_lock);
392 mutex_exit(&tc->tc_lock);
397 txg_do_callbacks(void *arg)
399 list_t *cb_list = arg;
401 dmu_tx_do_callbacks(cb_list, 0);
403 list_destroy(cb_list);
405 kmem_free(cb_list, sizeof (list_t));
409 * Dispatch the commit callbacks registered on this txg to worker threads.
411 * If no callbacks are registered for a given TXG, nothing happens.
412 * This function creates a taskq for the associated pool, if needed.
415 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
418 tx_state_t *tx = &dp->dp_tx;
421 for (c = 0; c < max_ncpus; c++) {
422 tx_cpu_t *tc = &tx->tx_cpu[c];
424 * No need to lock tx_cpu_t at this point, since this can
425 * only be called once a txg has been synced.
428 int g = txg & TXG_MASK;
430 if (list_is_empty(&tc->tc_callbacks[g]))
433 if (tx->tx_commit_cb_taskq == NULL) {
435 * Commit callback taskq hasn't been created yet.
437 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
438 max_ncpus, minclsyspri, max_ncpus, max_ncpus * 2,
442 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
443 list_create(cb_list, sizeof (dmu_tx_callback_t),
444 offsetof(dmu_tx_callback_t, dcb_node));
446 list_move_tail(cb_list, &tc->tc_callbacks[g]);
448 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
449 txg_do_callbacks, cb_list, TQ_SLEEP);
454 txg_is_syncing(dsl_pool_t *dp)
456 tx_state_t *tx = &dp->dp_tx;
457 ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
458 return (tx->tx_syncing_txg != 0);
462 txg_is_quiescing(dsl_pool_t *dp)
464 tx_state_t *tx = &dp->dp_tx;
465 ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
466 return (tx->tx_quiescing_txg != 0);
470 txg_has_quiesced_to_sync(dsl_pool_t *dp)
472 tx_state_t *tx = &dp->dp_tx;
473 ASSERT(MUTEX_HELD(&tx->tx_sync_lock));
474 return (tx->tx_quiesced_txg != 0);
478 txg_sync_thread(void *arg)
480 dsl_pool_t *dp = arg;
481 spa_t *spa = dp->dp_spa;
482 tx_state_t *tx = &dp->dp_tx;
484 uint64_t start, delta;
486 txg_thread_enter(tx, &cpr);
490 uint64_t timeout = zfs_txg_timeout * hz;
495 * We sync when we're scanning, there's someone waiting
496 * on us, or the quiesce thread has handed off a txg to
497 * us, or we have reached our timeout.
499 timer = (delta >= timeout ? 0 : timeout - delta);
500 while (!dsl_scan_active(dp->dp_scan) &&
501 !tx->tx_exiting && timer > 0 &&
502 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
503 !txg_has_quiesced_to_sync(dp) &&
504 dp->dp_dirty_total < zfs_dirty_data_sync) {
505 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
506 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
507 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
508 delta = ddi_get_lbolt() - start;
509 timer = (delta > timeout ? 0 : timeout - delta);
513 * Wait until the quiesce thread hands off a txg to us,
514 * prompting it to do so if necessary.
516 while (!tx->tx_exiting && !txg_has_quiesced_to_sync(dp)) {
517 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
518 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
519 cv_broadcast(&tx->tx_quiesce_more_cv);
520 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
524 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
527 * Consume the quiesced txg which has been handed off to
528 * us. This may cause the quiescing thread to now be
529 * able to quiesce another txg, so we must signal it.
531 ASSERT(tx->tx_quiesced_txg != 0);
532 txg = tx->tx_quiesced_txg;
533 tx->tx_quiesced_txg = 0;
534 tx->tx_syncing_txg = txg;
535 DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
536 cv_broadcast(&tx->tx_quiesce_more_cv);
538 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
539 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
540 mutex_exit(&tx->tx_sync_lock);
542 start = ddi_get_lbolt();
544 delta = ddi_get_lbolt() - start;
546 mutex_enter(&tx->tx_sync_lock);
547 tx->tx_synced_txg = txg;
548 tx->tx_syncing_txg = 0;
549 DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
550 cv_broadcast(&tx->tx_sync_done_cv);
553 * Dispatch commit callbacks to worker threads.
555 txg_dispatch_callbacks(dp, txg);
560 txg_quiesce_thread(void *arg)
562 dsl_pool_t *dp = arg;
563 tx_state_t *tx = &dp->dp_tx;
566 txg_thread_enter(tx, &cpr);
572 * We quiesce when there's someone waiting on us.
573 * However, we can only have one txg in "quiescing" or
574 * "quiesced, waiting to sync" state. So we wait until
575 * the "quiesced, waiting to sync" txg has been consumed
576 * by the sync thread.
578 while (!tx->tx_exiting &&
579 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
580 txg_has_quiesced_to_sync(dp)))
581 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
584 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
586 txg = tx->tx_open_txg;
587 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
588 txg, tx->tx_quiesce_txg_waiting,
589 tx->tx_sync_txg_waiting);
590 tx->tx_quiescing_txg = txg;
592 mutex_exit(&tx->tx_sync_lock);
593 txg_quiesce(dp, txg);
594 mutex_enter(&tx->tx_sync_lock);
597 * Hand this txg off to the sync thread.
599 dprintf("quiesce done, handing off txg %llu\n", txg);
600 tx->tx_quiescing_txg = 0;
601 tx->tx_quiesced_txg = txg;
602 DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
603 cv_broadcast(&tx->tx_sync_more_cv);
604 cv_broadcast(&tx->tx_quiesce_done_cv);
609 * Delay this thread by delay nanoseconds if we are still in the open
610 * transaction group and there is already a waiting txg quiesing or quiesced.
611 * Abort the delay if this txg stalls or enters the quiesing state.
614 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
616 tx_state_t *tx = &dp->dp_tx;
617 hrtime_t start = gethrtime();
619 /* don't delay if this txg could transition to quiescing immediately */
620 if (tx->tx_open_txg > txg ||
621 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
624 mutex_enter(&tx->tx_sync_lock);
625 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
626 mutex_exit(&tx->tx_sync_lock);
630 while (gethrtime() - start < delay &&
631 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
632 (void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
633 &tx->tx_sync_lock, delay, resolution, 0);
636 mutex_exit(&tx->tx_sync_lock);
640 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
642 tx_state_t *tx = &dp->dp_tx;
644 ASSERT(!dsl_pool_config_held(dp));
646 mutex_enter(&tx->tx_sync_lock);
647 ASSERT3U(tx->tx_threads, ==, 2);
649 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
650 if (tx->tx_sync_txg_waiting < txg)
651 tx->tx_sync_txg_waiting = txg;
652 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
653 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
654 while (tx->tx_synced_txg < txg) {
655 dprintf("broadcasting sync more "
656 "tx_synced=%llu waiting=%llu dp=%p\n",
657 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
658 cv_broadcast(&tx->tx_sync_more_cv);
659 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
661 mutex_exit(&tx->tx_sync_lock);
665 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
667 tx_state_t *tx = &dp->dp_tx;
669 ASSERT(!dsl_pool_config_held(dp));
671 mutex_enter(&tx->tx_sync_lock);
672 ASSERT3U(tx->tx_threads, ==, 2);
674 txg = tx->tx_open_txg + 1;
675 if (tx->tx_quiesce_txg_waiting < txg)
676 tx->tx_quiesce_txg_waiting = txg;
677 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
678 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
679 while (tx->tx_open_txg < txg) {
680 cv_broadcast(&tx->tx_quiesce_more_cv);
681 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
683 mutex_exit(&tx->tx_sync_lock);
687 * If there isn't a txg syncing or in the pipeline, push another txg through
688 * the pipeline by queiscing the open txg.
691 txg_kick(dsl_pool_t *dp)
693 tx_state_t *tx = &dp->dp_tx;
695 ASSERT(!dsl_pool_config_held(dp));
697 mutex_enter(&tx->tx_sync_lock);
698 if (!txg_is_syncing(dp) &&
699 !txg_is_quiescing(dp) &&
700 tx->tx_quiesce_txg_waiting <= tx->tx_open_txg &&
701 tx->tx_sync_txg_waiting <= tx->tx_synced_txg &&
702 tx->tx_quiesced_txg <= tx->tx_synced_txg) {
703 tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1;
704 cv_broadcast(&tx->tx_quiesce_more_cv);
706 mutex_exit(&tx->tx_sync_lock);
710 txg_stalled(dsl_pool_t *dp)
712 tx_state_t *tx = &dp->dp_tx;
713 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
717 txg_sync_waiting(dsl_pool_t *dp)
719 tx_state_t *tx = &dp->dp_tx;
721 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
722 tx->tx_quiesced_txg != 0);
726 * Verify that this txg is active (open, quiescing, syncing). Non-active
727 * txg's should not be manipulated.
730 txg_verify(spa_t *spa, uint64_t txg)
732 dsl_pool_t *dp = spa_get_dsl(spa);
733 if (txg <= TXG_INITIAL || txg == ZILTEST_TXG)
735 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
736 ASSERT3U(txg, >=, dp->dp_tx.tx_synced_txg);
737 ASSERT3U(txg, >=, dp->dp_tx.tx_open_txg - TXG_CONCURRENT_STATES);
741 * Per-txg object lists.
744 txg_list_create(txg_list_t *tl, spa_t *spa, size_t offset)
748 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
750 tl->tl_offset = offset;
753 for (t = 0; t < TXG_SIZE; t++)
754 tl->tl_head[t] = NULL;
758 txg_list_destroy(txg_list_t *tl)
762 for (t = 0; t < TXG_SIZE; t++)
763 ASSERT(txg_list_empty(tl, t));
765 mutex_destroy(&tl->tl_lock);
769 txg_list_empty(txg_list_t *tl, uint64_t txg)
771 txg_verify(tl->tl_spa, txg);
772 return (tl->tl_head[txg & TXG_MASK] == NULL);
776 * Returns true if all txg lists are empty.
778 * Warning: this is inherently racy (an item could be added immediately
779 * after this function returns). We don't bother with the lock because
780 * it wouldn't change the semantics.
783 txg_all_lists_empty(txg_list_t *tl)
785 for (int i = 0; i < TXG_SIZE; i++) {
786 if (!txg_list_empty(tl, i)) {
794 * Add an entry to the list (unless it's already on the list).
795 * Returns B_TRUE if it was actually added.
798 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
800 int t = txg & TXG_MASK;
801 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
804 txg_verify(tl->tl_spa, txg);
805 mutex_enter(&tl->tl_lock);
806 add = (tn->tn_member[t] == 0);
808 tn->tn_member[t] = 1;
809 tn->tn_next[t] = tl->tl_head[t];
812 mutex_exit(&tl->tl_lock);
818 * Add an entry to the end of the list, unless it's already on the list.
819 * (walks list to find end)
820 * Returns B_TRUE if it was actually added.
823 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
825 int t = txg & TXG_MASK;
826 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
829 txg_verify(tl->tl_spa, txg);
830 mutex_enter(&tl->tl_lock);
831 add = (tn->tn_member[t] == 0);
835 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
838 tn->tn_member[t] = 1;
839 tn->tn_next[t] = NULL;
842 mutex_exit(&tl->tl_lock);
848 * Remove the head of the list and return it.
851 txg_list_remove(txg_list_t *tl, uint64_t txg)
853 int t = txg & TXG_MASK;
857 txg_verify(tl->tl_spa, txg);
858 mutex_enter(&tl->tl_lock);
859 if ((tn = tl->tl_head[t]) != NULL) {
860 ASSERT(tn->tn_member[t]);
861 ASSERT(tn->tn_next[t] == NULL || tn->tn_next[t]->tn_member[t]);
862 p = (char *)tn - tl->tl_offset;
863 tl->tl_head[t] = tn->tn_next[t];
864 tn->tn_next[t] = NULL;
865 tn->tn_member[t] = 0;
867 mutex_exit(&tl->tl_lock);
873 * Remove a specific item from the list and return it.
876 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
878 int t = txg & TXG_MASK;
879 txg_node_t *tn, **tp;
881 txg_verify(tl->tl_spa, txg);
882 mutex_enter(&tl->tl_lock);
884 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
885 if ((char *)tn - tl->tl_offset == p) {
886 *tp = tn->tn_next[t];
887 tn->tn_next[t] = NULL;
888 tn->tn_member[t] = 0;
889 mutex_exit(&tl->tl_lock);
894 mutex_exit(&tl->tl_lock);
900 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
902 int t = txg & TXG_MASK;
903 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
905 txg_verify(tl->tl_spa, txg);
906 return (tn->tn_member[t] != 0);
910 * Walk a txg list -- only safe if you know it's not changing.
913 txg_list_head(txg_list_t *tl, uint64_t txg)
915 int t = txg & TXG_MASK;
916 txg_node_t *tn = tl->tl_head[t];
918 txg_verify(tl->tl_spa, txg);
919 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
923 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
925 int t = txg & TXG_MASK;
926 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
928 txg_verify(tl->tl_spa, txg);
931 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);