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.
25 #include <sys/zfs_context.h>
26 #include <sys/txg_impl.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_scan.h>
31 #include <sys/callb.h>
34 * Pool-wide transaction groups.
37 static void txg_sync_thread(dsl_pool_t *dp);
38 static void txg_quiesce_thread(dsl_pool_t *dp);
40 int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
43 * Prepare the txg subsystem.
46 txg_init(dsl_pool_t *dp, uint64_t txg)
48 tx_state_t *tx = &dp->dp_tx;
50 bzero(tx, sizeof (tx_state_t));
52 tx->tx_cpu = vmem_zalloc(max_ncpus * sizeof (tx_cpu_t), KM_SLEEP);
54 for (c = 0; c < max_ncpus; c++) {
57 mutex_init(&tx->tx_cpu[c].tc_lock, NULL, MUTEX_DEFAULT, NULL);
58 for (i = 0; i < TXG_SIZE; i++) {
59 cv_init(&tx->tx_cpu[c].tc_cv[i], NULL, CV_DEFAULT,
61 list_create(&tx->tx_cpu[c].tc_callbacks[i],
62 sizeof (dmu_tx_callback_t),
63 offsetof(dmu_tx_callback_t, dcb_node));
67 mutex_init(&tx->tx_sync_lock, NULL, MUTEX_DEFAULT, NULL);
69 cv_init(&tx->tx_sync_more_cv, NULL, CV_DEFAULT, NULL);
70 cv_init(&tx->tx_sync_done_cv, NULL, CV_DEFAULT, NULL);
71 cv_init(&tx->tx_quiesce_more_cv, NULL, CV_DEFAULT, NULL);
72 cv_init(&tx->tx_quiesce_done_cv, NULL, CV_DEFAULT, NULL);
73 cv_init(&tx->tx_exit_cv, NULL, CV_DEFAULT, NULL);
75 tx->tx_open_txg = txg;
79 * Close down the txg subsystem.
82 txg_fini(dsl_pool_t *dp)
84 tx_state_t *tx = &dp->dp_tx;
87 ASSERT(tx->tx_threads == 0);
89 mutex_destroy(&tx->tx_sync_lock);
91 cv_destroy(&tx->tx_sync_more_cv);
92 cv_destroy(&tx->tx_sync_done_cv);
93 cv_destroy(&tx->tx_quiesce_more_cv);
94 cv_destroy(&tx->tx_quiesce_done_cv);
95 cv_destroy(&tx->tx_exit_cv);
97 for (c = 0; c < max_ncpus; c++) {
100 mutex_destroy(&tx->tx_cpu[c].tc_lock);
101 for (i = 0; i < TXG_SIZE; i++) {
102 cv_destroy(&tx->tx_cpu[c].tc_cv[i]);
103 list_destroy(&tx->tx_cpu[c].tc_callbacks[i]);
107 if (tx->tx_commit_cb_taskq != NULL)
108 taskq_destroy(tx->tx_commit_cb_taskq);
110 vmem_free(tx->tx_cpu, max_ncpus * sizeof (tx_cpu_t));
112 bzero(tx, sizeof (tx_state_t));
116 * Start syncing transaction groups.
119 txg_sync_start(dsl_pool_t *dp)
121 tx_state_t *tx = &dp->dp_tx;
123 mutex_enter(&tx->tx_sync_lock);
125 dprintf("pool %p\n", dp);
127 ASSERT(tx->tx_threads == 0);
131 tx->tx_quiesce_thread = thread_create(NULL, 0, txg_quiesce_thread,
132 dp, 0, &p0, TS_RUN, minclsyspri);
135 * The sync thread can need a larger-than-default stack size on
136 * 32-bit x86. This is due in part to nested pools and
137 * scrub_visitbp() recursion.
139 tx->tx_sync_thread = thread_create(NULL, 32<<10, txg_sync_thread,
140 dp, 0, &p0, TS_RUN, minclsyspri);
142 mutex_exit(&tx->tx_sync_lock);
146 txg_thread_enter(tx_state_t *tx, callb_cpr_t *cpr)
148 CALLB_CPR_INIT(cpr, &tx->tx_sync_lock, callb_generic_cpr, FTAG);
149 mutex_enter(&tx->tx_sync_lock);
153 txg_thread_exit(tx_state_t *tx, callb_cpr_t *cpr, kthread_t **tpp)
155 ASSERT(*tpp != NULL);
158 cv_broadcast(&tx->tx_exit_cv);
159 CALLB_CPR_EXIT(cpr); /* drops &tx->tx_sync_lock */
164 txg_thread_wait(tx_state_t *tx, callb_cpr_t *cpr, kcondvar_t *cv, uint64_t time)
166 CALLB_CPR_SAFE_BEGIN(cpr);
169 (void) cv_timedwait(cv, &tx->tx_sync_lock,
170 ddi_get_lbolt() + time);
172 cv_wait(cv, &tx->tx_sync_lock);
174 CALLB_CPR_SAFE_END(cpr, &tx->tx_sync_lock);
178 * Stop syncing transaction groups.
181 txg_sync_stop(dsl_pool_t *dp)
183 tx_state_t *tx = &dp->dp_tx;
185 dprintf("pool %p\n", dp);
187 * Finish off any work in progress.
189 ASSERT(tx->tx_threads == 2);
192 * We need to ensure that we've vacated the deferred space_maps.
194 txg_wait_synced(dp, tx->tx_open_txg + TXG_DEFER_SIZE);
197 * Wake all sync threads and wait for them to die.
199 mutex_enter(&tx->tx_sync_lock);
201 ASSERT(tx->tx_threads == 2);
205 cv_broadcast(&tx->tx_quiesce_more_cv);
206 cv_broadcast(&tx->tx_quiesce_done_cv);
207 cv_broadcast(&tx->tx_sync_more_cv);
209 while (tx->tx_threads != 0)
210 cv_wait(&tx->tx_exit_cv, &tx->tx_sync_lock);
214 mutex_exit(&tx->tx_sync_lock);
218 txg_hold_open(dsl_pool_t *dp, txg_handle_t *th)
220 tx_state_t *tx = &dp->dp_tx;
221 tx_cpu_t *tc = &tx->tx_cpu[CPU_SEQID];
224 mutex_enter(&tc->tc_lock);
226 txg = tx->tx_open_txg;
227 tc->tc_count[txg & TXG_MASK]++;
236 txg_rele_to_quiesce(txg_handle_t *th)
238 tx_cpu_t *tc = th->th_cpu;
240 mutex_exit(&tc->tc_lock);
244 txg_register_callbacks(txg_handle_t *th, list_t *tx_callbacks)
246 tx_cpu_t *tc = th->th_cpu;
247 int g = th->th_txg & TXG_MASK;
249 mutex_enter(&tc->tc_lock);
250 list_move_tail(&tc->tc_callbacks[g], tx_callbacks);
251 mutex_exit(&tc->tc_lock);
255 txg_rele_to_sync(txg_handle_t *th)
257 tx_cpu_t *tc = th->th_cpu;
258 int g = th->th_txg & TXG_MASK;
260 mutex_enter(&tc->tc_lock);
261 ASSERT(tc->tc_count[g] != 0);
262 if (--tc->tc_count[g] == 0)
263 cv_broadcast(&tc->tc_cv[g]);
264 mutex_exit(&tc->tc_lock);
266 th->th_cpu = NULL; /* defensive */
270 txg_quiesce(dsl_pool_t *dp, uint64_t txg)
272 tx_state_t *tx = &dp->dp_tx;
273 int g = txg & TXG_MASK;
277 * Grab all tx_cpu locks so nobody else can get into this txg.
279 for (c = 0; c < max_ncpus; c++)
280 mutex_enter(&tx->tx_cpu[c].tc_lock);
282 ASSERT(txg == tx->tx_open_txg);
286 * Now that we've incremented tx_open_txg, we can let threads
287 * enter the next transaction group.
289 for (c = 0; c < max_ncpus; c++)
290 mutex_exit(&tx->tx_cpu[c].tc_lock);
293 * Quiesce the transaction group by waiting for everyone to txg_exit().
295 for (c = 0; c < max_ncpus; c++) {
296 tx_cpu_t *tc = &tx->tx_cpu[c];
297 mutex_enter(&tc->tc_lock);
298 while (tc->tc_count[g] != 0)
299 cv_wait(&tc->tc_cv[g], &tc->tc_lock);
300 mutex_exit(&tc->tc_lock);
305 txg_do_callbacks(list_t *cb_list)
307 dmu_tx_do_callbacks(cb_list, 0);
309 list_destroy(cb_list);
311 kmem_free(cb_list, sizeof (list_t));
315 * Dispatch the commit callbacks registered on this txg to worker threads.
318 txg_dispatch_callbacks(dsl_pool_t *dp, uint64_t txg)
321 tx_state_t *tx = &dp->dp_tx;
324 for (c = 0; c < max_ncpus; c++) {
325 tx_cpu_t *tc = &tx->tx_cpu[c];
326 /* No need to lock tx_cpu_t at this point */
328 int g = txg & TXG_MASK;
330 if (list_is_empty(&tc->tc_callbacks[g]))
333 if (tx->tx_commit_cb_taskq == NULL) {
335 * Commit callback taskq hasn't been created yet.
337 tx->tx_commit_cb_taskq = taskq_create("tx_commit_cb",
338 100, minclsyspri, max_ncpus, INT_MAX,
339 TASKQ_THREADS_CPU_PCT | TASKQ_PREPOPULATE);
342 cb_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
343 list_create(cb_list, sizeof (dmu_tx_callback_t),
344 offsetof(dmu_tx_callback_t, dcb_node));
346 list_move_tail(cb_list, &tc->tc_callbacks[g]);
348 (void) taskq_dispatch(tx->tx_commit_cb_taskq, (task_func_t *)
349 txg_do_callbacks, cb_list, TQ_SLEEP);
354 * Wait for pending commit callbacks of already-synced transactions to finish
356 * Calling this function from within a commit callback will deadlock.
359 txg_wait_callbacks(dsl_pool_t *dp)
361 tx_state_t *tx = &dp->dp_tx;
363 if (tx->tx_commit_cb_taskq != NULL)
364 taskq_wait(tx->tx_commit_cb_taskq);
368 txg_sync_thread(dsl_pool_t *dp)
370 spa_t *spa = dp->dp_spa;
371 tx_state_t *tx = &dp->dp_tx;
373 uint64_t start, delta;
375 txg_thread_enter(tx, &cpr);
379 uint64_t timer, timeout = zfs_txg_timeout * hz;
383 * We sync when we're scanning, there's someone waiting
384 * on us, or the quiesce thread has handed off a txg to
385 * us, or we have reached our timeout.
387 timer = (delta >= timeout ? 0 : timeout - delta);
388 while (!dsl_scan_active(dp->dp_scan) &&
389 !tx->tx_exiting && timer > 0 &&
390 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
391 tx->tx_quiesced_txg == 0) {
392 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
393 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
394 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
395 delta = ddi_get_lbolt() - start;
396 timer = (delta > timeout ? 0 : timeout - delta);
400 * Wait until the quiesce thread hands off a txg to us,
401 * prompting it to do so if necessary.
403 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
404 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
405 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
406 cv_broadcast(&tx->tx_quiesce_more_cv);
407 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
411 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
414 * Consume the quiesced txg which has been handed off to
415 * us. This may cause the quiescing thread to now be
416 * able to quiesce another txg, so we must signal it.
418 txg = tx->tx_quiesced_txg;
419 tx->tx_quiesced_txg = 0;
420 tx->tx_syncing_txg = txg;
421 cv_broadcast(&tx->tx_quiesce_more_cv);
423 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
424 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
425 mutex_exit(&tx->tx_sync_lock);
427 start = ddi_get_lbolt();
429 delta = ddi_get_lbolt() - start;
431 mutex_enter(&tx->tx_sync_lock);
432 tx->tx_synced_txg = txg;
433 tx->tx_syncing_txg = 0;
434 cv_broadcast(&tx->tx_sync_done_cv);
437 * Dispatch commit callbacks to worker threads.
439 txg_dispatch_callbacks(dp, txg);
444 txg_quiesce_thread(dsl_pool_t *dp)
446 tx_state_t *tx = &dp->dp_tx;
449 txg_thread_enter(tx, &cpr);
455 * We quiesce when there's someone waiting on us.
456 * However, we can only have one txg in "quiescing" or
457 * "quiesced, waiting to sync" state. So we wait until
458 * the "quiesced, waiting to sync" txg has been consumed
459 * by the sync thread.
461 while (!tx->tx_exiting &&
462 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
463 tx->tx_quiesced_txg != 0))
464 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
467 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
469 txg = tx->tx_open_txg;
470 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
471 txg, tx->tx_quiesce_txg_waiting,
472 tx->tx_sync_txg_waiting);
473 mutex_exit(&tx->tx_sync_lock);
474 txg_quiesce(dp, txg);
475 mutex_enter(&tx->tx_sync_lock);
478 * Hand this txg off to the sync thread.
480 dprintf("quiesce done, handing off txg %llu\n", txg);
481 tx->tx_quiesced_txg = txg;
482 cv_broadcast(&tx->tx_sync_more_cv);
483 cv_broadcast(&tx->tx_quiesce_done_cv);
488 * Delay this thread by 'ticks' if we are still in the open transaction
489 * group and there is already a waiting txg quiesing or quiesced. Abort
490 * the delay if this txg stalls or enters the quiesing state.
493 txg_delay(dsl_pool_t *dp, uint64_t txg, int ticks)
495 tx_state_t *tx = &dp->dp_tx;
496 int timeout = ddi_get_lbolt() + ticks;
498 /* don't delay if this txg could transition to quiesing immediately */
499 if (tx->tx_open_txg > txg ||
500 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
503 mutex_enter(&tx->tx_sync_lock);
504 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
505 mutex_exit(&tx->tx_sync_lock);
509 while (ddi_get_lbolt() < timeout &&
510 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
511 (void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
514 mutex_exit(&tx->tx_sync_lock);
518 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
520 tx_state_t *tx = &dp->dp_tx;
522 mutex_enter(&tx->tx_sync_lock);
523 ASSERT(tx->tx_threads == 2);
525 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
526 if (tx->tx_sync_txg_waiting < txg)
527 tx->tx_sync_txg_waiting = txg;
528 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
529 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
530 while (tx->tx_synced_txg < txg) {
531 dprintf("broadcasting sync more "
532 "tx_synced=%llu waiting=%llu dp=%p\n",
533 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
534 cv_broadcast(&tx->tx_sync_more_cv);
535 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
537 mutex_exit(&tx->tx_sync_lock);
541 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
543 tx_state_t *tx = &dp->dp_tx;
545 mutex_enter(&tx->tx_sync_lock);
546 ASSERT(tx->tx_threads == 2);
548 txg = tx->tx_open_txg + 1;
549 if (tx->tx_quiesce_txg_waiting < txg)
550 tx->tx_quiesce_txg_waiting = txg;
551 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
552 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
553 while (tx->tx_open_txg < txg) {
554 cv_broadcast(&tx->tx_quiesce_more_cv);
555 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
557 mutex_exit(&tx->tx_sync_lock);
561 txg_stalled(dsl_pool_t *dp)
563 tx_state_t *tx = &dp->dp_tx;
564 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
568 txg_sync_waiting(dsl_pool_t *dp)
570 tx_state_t *tx = &dp->dp_tx;
572 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
573 tx->tx_quiesced_txg != 0);
577 * Per-txg object lists.
580 txg_list_create(txg_list_t *tl, size_t offset)
584 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
586 tl->tl_offset = offset;
588 for (t = 0; t < TXG_SIZE; t++)
589 tl->tl_head[t] = NULL;
593 txg_list_destroy(txg_list_t *tl)
597 for (t = 0; t < TXG_SIZE; t++)
598 ASSERT(txg_list_empty(tl, t));
600 mutex_destroy(&tl->tl_lock);
604 txg_list_empty(txg_list_t *tl, uint64_t txg)
606 return (tl->tl_head[txg & TXG_MASK] == NULL);
610 * Add an entry to the list.
611 * Returns 0 if it's a new entry, 1 if it's already there.
614 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
616 int t = txg & TXG_MASK;
617 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
620 mutex_enter(&tl->tl_lock);
621 already_on_list = tn->tn_member[t];
622 if (!already_on_list) {
623 tn->tn_member[t] = 1;
624 tn->tn_next[t] = tl->tl_head[t];
627 mutex_exit(&tl->tl_lock);
629 return (already_on_list);
633 * Add an entry to the end of the list (walks list to find end).
634 * Returns 0 if it's a new entry, 1 if it's already there.
637 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
639 int t = txg & TXG_MASK;
640 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
643 mutex_enter(&tl->tl_lock);
644 already_on_list = tn->tn_member[t];
645 if (!already_on_list) {
648 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
651 tn->tn_member[t] = 1;
652 tn->tn_next[t] = NULL;
655 mutex_exit(&tl->tl_lock);
657 return (already_on_list);
661 * Remove the head of the list and return it.
664 txg_list_remove(txg_list_t *tl, uint64_t txg)
666 int t = txg & TXG_MASK;
670 mutex_enter(&tl->tl_lock);
671 if ((tn = tl->tl_head[t]) != NULL) {
672 p = (char *)tn - tl->tl_offset;
673 tl->tl_head[t] = tn->tn_next[t];
674 tn->tn_next[t] = NULL;
675 tn->tn_member[t] = 0;
677 mutex_exit(&tl->tl_lock);
683 * Remove a specific item from the list and return it.
686 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
688 int t = txg & TXG_MASK;
689 txg_node_t *tn, **tp;
691 mutex_enter(&tl->tl_lock);
693 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
694 if ((char *)tn - tl->tl_offset == p) {
695 *tp = tn->tn_next[t];
696 tn->tn_next[t] = NULL;
697 tn->tn_member[t] = 0;
698 mutex_exit(&tl->tl_lock);
703 mutex_exit(&tl->tl_lock);
709 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
711 int t = txg & TXG_MASK;
712 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
714 return (tn->tn_member[t]);
718 * Walk a txg list -- only safe if you know it's not changing.
721 txg_list_head(txg_list_t *tl, uint64_t txg)
723 int t = txg & TXG_MASK;
724 txg_node_t *tn = tl->tl_head[t];
726 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
730 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
732 int t = txg & TXG_MASK;
733 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
737 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
740 #if defined(_KERNEL) && defined(HAVE_SPL)
741 EXPORT_SYMBOL(txg_init);
742 EXPORT_SYMBOL(txg_fini);
743 EXPORT_SYMBOL(txg_sync_start);
744 EXPORT_SYMBOL(txg_sync_stop);
745 EXPORT_SYMBOL(txg_hold_open);
746 EXPORT_SYMBOL(txg_rele_to_quiesce);
747 EXPORT_SYMBOL(txg_rele_to_sync);
748 EXPORT_SYMBOL(txg_register_callbacks);
749 EXPORT_SYMBOL(txg_delay);
750 EXPORT_SYMBOL(txg_wait_synced);
751 EXPORT_SYMBOL(txg_wait_open);
752 EXPORT_SYMBOL(txg_wait_callbacks);
753 EXPORT_SYMBOL(txg_stalled);
754 EXPORT_SYMBOL(txg_sync_waiting);