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) 2013 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>
33 #include <sys/callb.h>
36 * ZFS Transaction Groups
37 * ----------------------
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.
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.
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.
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.
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.
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.
108 static void txg_sync_thread(void *arg);
109 static void txg_quiesce_thread(void *arg);
111 int zfs_txg_timeout = 5; /* max seconds worth of delta per txg */
113 SYSCTL_DECL(_vfs_zfs);
114 SYSCTL_NODE(_vfs_zfs, OID_AUTO, txg, CTLFLAG_RW, 0, "ZFS TXG");
115 TUNABLE_INT("vfs.zfs.txg.timeout", &zfs_txg_timeout);
116 SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, timeout, CTLFLAG_RW, &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 ASSERT(tx->tx_threads == 0);
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 ASSERT(tx->tx_threads == 0);
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 ASSERT(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 ASSERT(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.
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_sync_thread(void *arg)
456 dsl_pool_t *dp = arg;
457 spa_t *spa = dp->dp_spa;
458 tx_state_t *tx = &dp->dp_tx;
460 uint64_t start, delta;
462 txg_thread_enter(tx, &cpr);
466 uint64_t timeout = zfs_txg_timeout * hz;
471 * We sync when we're scanning, there's someone waiting
472 * on us, or the quiesce thread has handed off a txg to
473 * us, or we have reached our timeout.
475 timer = (delta >= timeout ? 0 : timeout - delta);
476 while (!dsl_scan_active(dp->dp_scan) &&
477 !tx->tx_exiting && timer > 0 &&
478 tx->tx_synced_txg >= tx->tx_sync_txg_waiting &&
479 tx->tx_quiesced_txg == 0 &&
480 dp->dp_dirty_total < zfs_dirty_data_sync) {
481 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
482 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
483 txg_thread_wait(tx, &cpr, &tx->tx_sync_more_cv, timer);
484 delta = ddi_get_lbolt() - start;
485 timer = (delta > timeout ? 0 : timeout - delta);
489 * Wait until the quiesce thread hands off a txg to us,
490 * prompting it to do so if necessary.
492 while (!tx->tx_exiting && tx->tx_quiesced_txg == 0) {
493 if (tx->tx_quiesce_txg_waiting < tx->tx_open_txg+1)
494 tx->tx_quiesce_txg_waiting = tx->tx_open_txg+1;
495 cv_broadcast(&tx->tx_quiesce_more_cv);
496 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_done_cv, 0);
500 txg_thread_exit(tx, &cpr, &tx->tx_sync_thread);
503 * Consume the quiesced txg which has been handed off to
504 * us. This may cause the quiescing thread to now be
505 * able to quiesce another txg, so we must signal it.
507 txg = tx->tx_quiesced_txg;
508 tx->tx_quiesced_txg = 0;
509 tx->tx_syncing_txg = txg;
510 DTRACE_PROBE2(txg__syncing, dsl_pool_t *, dp, uint64_t, txg);
511 cv_broadcast(&tx->tx_quiesce_more_cv);
513 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
514 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
515 mutex_exit(&tx->tx_sync_lock);
517 start = ddi_get_lbolt();
519 delta = ddi_get_lbolt() - start;
521 mutex_enter(&tx->tx_sync_lock);
522 tx->tx_synced_txg = txg;
523 tx->tx_syncing_txg = 0;
524 DTRACE_PROBE2(txg__synced, dsl_pool_t *, dp, uint64_t, txg);
525 cv_broadcast(&tx->tx_sync_done_cv);
528 * Dispatch commit callbacks to worker threads.
530 txg_dispatch_callbacks(dp, txg);
535 txg_quiesce_thread(void *arg)
537 dsl_pool_t *dp = arg;
538 tx_state_t *tx = &dp->dp_tx;
541 txg_thread_enter(tx, &cpr);
547 * We quiesce when there's someone waiting on us.
548 * However, we can only have one txg in "quiescing" or
549 * "quiesced, waiting to sync" state. So we wait until
550 * the "quiesced, waiting to sync" txg has been consumed
551 * by the sync thread.
553 while (!tx->tx_exiting &&
554 (tx->tx_open_txg >= tx->tx_quiesce_txg_waiting ||
555 tx->tx_quiesced_txg != 0))
556 txg_thread_wait(tx, &cpr, &tx->tx_quiesce_more_cv, 0);
559 txg_thread_exit(tx, &cpr, &tx->tx_quiesce_thread);
561 txg = tx->tx_open_txg;
562 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
563 txg, tx->tx_quiesce_txg_waiting,
564 tx->tx_sync_txg_waiting);
565 mutex_exit(&tx->tx_sync_lock);
566 txg_quiesce(dp, txg);
567 mutex_enter(&tx->tx_sync_lock);
570 * Hand this txg off to the sync thread.
572 dprintf("quiesce done, handing off txg %llu\n", txg);
573 tx->tx_quiesced_txg = txg;
574 DTRACE_PROBE2(txg__quiesced, dsl_pool_t *, dp, uint64_t, txg);
575 cv_broadcast(&tx->tx_sync_more_cv);
576 cv_broadcast(&tx->tx_quiesce_done_cv);
581 * Delay this thread by delay nanoseconds if we are still in the open
582 * transaction group and there is already a waiting txg quiesing or quiesced.
583 * Abort the delay if this txg stalls or enters the quiesing state.
586 txg_delay(dsl_pool_t *dp, uint64_t txg, hrtime_t delay, hrtime_t resolution)
588 tx_state_t *tx = &dp->dp_tx;
589 hrtime_t start = gethrtime();
591 /* don't delay if this txg could transition to quiescing immediately */
592 if (tx->tx_open_txg > txg ||
593 tx->tx_syncing_txg == txg-1 || tx->tx_synced_txg == txg-1)
596 mutex_enter(&tx->tx_sync_lock);
597 if (tx->tx_open_txg > txg || tx->tx_synced_txg == txg-1) {
598 mutex_exit(&tx->tx_sync_lock);
602 while (gethrtime() - start < delay &&
603 tx->tx_syncing_txg < txg-1 && !txg_stalled(dp)) {
604 (void) cv_timedwait_hires(&tx->tx_quiesce_more_cv,
605 &tx->tx_sync_lock, delay, resolution, 0);
608 mutex_exit(&tx->tx_sync_lock);
612 txg_wait_synced(dsl_pool_t *dp, uint64_t txg)
614 tx_state_t *tx = &dp->dp_tx;
616 ASSERT(!dsl_pool_config_held(dp));
618 mutex_enter(&tx->tx_sync_lock);
619 ASSERT(tx->tx_threads == 2);
621 txg = tx->tx_open_txg + TXG_DEFER_SIZE;
622 if (tx->tx_sync_txg_waiting < txg)
623 tx->tx_sync_txg_waiting = txg;
624 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
625 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
626 while (tx->tx_synced_txg < txg) {
627 dprintf("broadcasting sync more "
628 "tx_synced=%llu waiting=%llu dp=%p\n",
629 tx->tx_synced_txg, tx->tx_sync_txg_waiting, dp);
630 cv_broadcast(&tx->tx_sync_more_cv);
631 cv_wait(&tx->tx_sync_done_cv, &tx->tx_sync_lock);
633 mutex_exit(&tx->tx_sync_lock);
637 txg_wait_open(dsl_pool_t *dp, uint64_t txg)
639 tx_state_t *tx = &dp->dp_tx;
641 ASSERT(!dsl_pool_config_held(dp));
643 mutex_enter(&tx->tx_sync_lock);
644 ASSERT(tx->tx_threads == 2);
646 txg = tx->tx_open_txg + 1;
647 if (tx->tx_quiesce_txg_waiting < txg)
648 tx->tx_quiesce_txg_waiting = txg;
649 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
650 txg, tx->tx_quiesce_txg_waiting, tx->tx_sync_txg_waiting);
651 while (tx->tx_open_txg < txg) {
652 cv_broadcast(&tx->tx_quiesce_more_cv);
653 cv_wait(&tx->tx_quiesce_done_cv, &tx->tx_sync_lock);
655 mutex_exit(&tx->tx_sync_lock);
659 * If there isn't a txg syncing or in the pipeline, push another txg through
660 * the pipeline by queiscing the open txg.
663 txg_kick(dsl_pool_t *dp)
665 tx_state_t *tx = &dp->dp_tx;
667 ASSERT(!dsl_pool_config_held(dp));
669 mutex_enter(&tx->tx_sync_lock);
670 if (tx->tx_syncing_txg == 0 &&
671 tx->tx_quiesce_txg_waiting <= tx->tx_open_txg &&
672 tx->tx_sync_txg_waiting <= tx->tx_synced_txg &&
673 tx->tx_quiesced_txg <= tx->tx_synced_txg) {
674 tx->tx_quiesce_txg_waiting = tx->tx_open_txg + 1;
675 cv_broadcast(&tx->tx_quiesce_more_cv);
677 mutex_exit(&tx->tx_sync_lock);
681 txg_stalled(dsl_pool_t *dp)
683 tx_state_t *tx = &dp->dp_tx;
684 return (tx->tx_quiesce_txg_waiting > tx->tx_open_txg);
688 txg_sync_waiting(dsl_pool_t *dp)
690 tx_state_t *tx = &dp->dp_tx;
692 return (tx->tx_syncing_txg <= tx->tx_sync_txg_waiting ||
693 tx->tx_quiesced_txg != 0);
697 * Per-txg object lists.
700 txg_list_create(txg_list_t *tl, size_t offset)
704 mutex_init(&tl->tl_lock, NULL, MUTEX_DEFAULT, NULL);
706 tl->tl_offset = offset;
708 for (t = 0; t < TXG_SIZE; t++)
709 tl->tl_head[t] = NULL;
713 txg_list_destroy(txg_list_t *tl)
717 for (t = 0; t < TXG_SIZE; t++)
718 ASSERT(txg_list_empty(tl, t));
720 mutex_destroy(&tl->tl_lock);
724 txg_list_empty(txg_list_t *tl, uint64_t txg)
726 return (tl->tl_head[txg & TXG_MASK] == NULL);
730 * Add an entry to the list (unless it's already on the list).
731 * Returns B_TRUE if it was actually added.
734 txg_list_add(txg_list_t *tl, void *p, uint64_t txg)
736 int t = txg & TXG_MASK;
737 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
740 mutex_enter(&tl->tl_lock);
741 add = (tn->tn_member[t] == 0);
743 tn->tn_member[t] = 1;
744 tn->tn_next[t] = tl->tl_head[t];
747 mutex_exit(&tl->tl_lock);
753 * Add an entry to the end of the list, unless it's already on the list.
754 * (walks list to find end)
755 * Returns B_TRUE if it was actually added.
758 txg_list_add_tail(txg_list_t *tl, void *p, uint64_t txg)
760 int t = txg & TXG_MASK;
761 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
764 mutex_enter(&tl->tl_lock);
765 add = (tn->tn_member[t] == 0);
769 for (tp = &tl->tl_head[t]; *tp != NULL; tp = &(*tp)->tn_next[t])
772 tn->tn_member[t] = 1;
773 tn->tn_next[t] = NULL;
776 mutex_exit(&tl->tl_lock);
782 * Remove the head of the list and return it.
785 txg_list_remove(txg_list_t *tl, uint64_t txg)
787 int t = txg & TXG_MASK;
791 mutex_enter(&tl->tl_lock);
792 if ((tn = tl->tl_head[t]) != NULL) {
793 p = (char *)tn - tl->tl_offset;
794 tl->tl_head[t] = tn->tn_next[t];
795 tn->tn_next[t] = NULL;
796 tn->tn_member[t] = 0;
798 mutex_exit(&tl->tl_lock);
804 * Remove a specific item from the list and return it.
807 txg_list_remove_this(txg_list_t *tl, void *p, uint64_t txg)
809 int t = txg & TXG_MASK;
810 txg_node_t *tn, **tp;
812 mutex_enter(&tl->tl_lock);
814 for (tp = &tl->tl_head[t]; (tn = *tp) != NULL; tp = &tn->tn_next[t]) {
815 if ((char *)tn - tl->tl_offset == p) {
816 *tp = tn->tn_next[t];
817 tn->tn_next[t] = NULL;
818 tn->tn_member[t] = 0;
819 mutex_exit(&tl->tl_lock);
824 mutex_exit(&tl->tl_lock);
830 txg_list_member(txg_list_t *tl, void *p, uint64_t txg)
832 int t = txg & TXG_MASK;
833 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
835 return (tn->tn_member[t] != 0);
839 * Walk a txg list -- only safe if you know it's not changing.
842 txg_list_head(txg_list_t *tl, uint64_t txg)
844 int t = txg & TXG_MASK;
845 txg_node_t *tn = tl->tl_head[t];
847 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);
851 txg_list_next(txg_list_t *tl, void *p, uint64_t txg)
853 int t = txg & TXG_MASK;
854 txg_node_t *tn = (txg_node_t *)((char *)p + tl->tl_offset);
858 return (tn == NULL ? NULL : (char *)tn - tl->tl_offset);