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 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 * Copyright (c) 2018 Datto Inc.
28 /* Portions Copyright 2010 Robert Milkowski */
30 #include <sys/zfs_context.h>
32 #include <sys/spa_impl.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/metaslab.h>
44 #include <sys/trace_zfs.h>
48 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
49 * calls that change the file system. Each itx has enough information to
50 * be able to replay them after a system crash, power loss, or
51 * equivalent failure mode. These are stored in memory until either:
53 * 1. they are committed to the pool by the DMU transaction group
54 * (txg), at which point they can be discarded; or
55 * 2. they are committed to the on-disk ZIL for the dataset being
56 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59 * In the event of a crash or power loss, the itxs contained by each
60 * dataset's on-disk ZIL will be replayed when that dataset is first
61 * instantiated (e.g. if the dataset is a normal filesystem, when it is
64 * As hinted at above, there is one ZIL per dataset (both the in-memory
65 * representation, and the on-disk representation). The on-disk format
66 * consists of 3 parts:
68 * - a single, per-dataset, ZIL header; which points to a chain of
69 * - zero or more ZIL blocks; each of which contains
70 * - zero or more ZIL records
72 * A ZIL record holds the information necessary to replay a single
73 * system call transaction. A ZIL block can hold many ZIL records, and
74 * the blocks are chained together, similarly to a singly linked list.
76 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
77 * block in the chain, and the ZIL header points to the first block in
80 * Note, there is not a fixed place in the pool to hold these ZIL
81 * blocks; they are dynamically allocated and freed as needed from the
82 * blocks available on the pool, though they can be preferentially
83 * allocated from a dedicated "log" vdev.
87 * This controls the amount of time that a ZIL block (lwb) will remain
88 * "open" when it isn't "full", and it has a thread waiting for it to be
89 * committed to stable storage. Please refer to the zil_commit_waiter()
90 * function (and the comments within it) for more details.
92 int zfs_commit_timeout_pct = 5;
95 * See zil.h for more information about these fields.
97 zil_stats_t zil_stats = {
98 { "zil_commit_count", KSTAT_DATA_UINT64 },
99 { "zil_commit_writer_count", KSTAT_DATA_UINT64 },
100 { "zil_itx_count", KSTAT_DATA_UINT64 },
101 { "zil_itx_indirect_count", KSTAT_DATA_UINT64 },
102 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 },
103 { "zil_itx_copied_count", KSTAT_DATA_UINT64 },
104 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 },
105 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 },
106 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 },
107 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 },
108 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 },
109 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 },
110 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 },
113 static kstat_t *zil_ksp;
116 * Disable intent logging replay. This global ZIL switch affects all pools.
118 int zil_replay_disable = 0;
121 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
122 * the disk(s) by the ZIL after an LWB write has completed. Setting this
123 * will cause ZIL corruption on power loss if a volatile out-of-order
124 * write cache is enabled.
126 int zil_nocacheflush = 0;
129 * Limit SLOG write size per commit executed with synchronous priority.
130 * Any writes above that will be executed with lower (asynchronous) priority
131 * to limit potential SLOG device abuse by single active ZIL writer.
133 unsigned long zil_slog_bulk = 768 * 1024;
135 static kmem_cache_t *zil_lwb_cache;
136 static kmem_cache_t *zil_zcw_cache;
138 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
139 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
142 zil_bp_compare(const void *x1, const void *x2)
144 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
145 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
147 int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
151 return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
155 zil_bp_tree_init(zilog_t *zilog)
157 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
158 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
162 zil_bp_tree_fini(zilog_t *zilog)
164 avl_tree_t *t = &zilog->zl_bp_tree;
168 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
169 kmem_free(zn, sizeof (zil_bp_node_t));
175 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
177 avl_tree_t *t = &zilog->zl_bp_tree;
182 if (BP_IS_EMBEDDED(bp))
185 dva = BP_IDENTITY(bp);
187 if (avl_find(t, dva, &where) != NULL)
188 return (SET_ERROR(EEXIST));
190 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
192 avl_insert(t, zn, where);
197 static zil_header_t *
198 zil_header_in_syncing_context(zilog_t *zilog)
200 return ((zil_header_t *)zilog->zl_header);
204 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
206 zio_cksum_t *zc = &bp->blk_cksum;
208 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
209 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
210 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
211 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
215 * Read a log block and make sure it's valid.
218 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
219 blkptr_t *nbp, void *dst, char **end)
221 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
222 arc_flags_t aflags = ARC_FLAG_WAIT;
223 arc_buf_t *abuf = NULL;
227 if (zilog->zl_header->zh_claim_txg == 0)
228 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
230 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
231 zio_flags |= ZIO_FLAG_SPECULATIVE;
234 zio_flags |= ZIO_FLAG_RAW;
236 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
237 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
239 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
240 &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
243 zio_cksum_t cksum = bp->blk_cksum;
246 * Validate the checksummed log block.
248 * Sequence numbers should be... sequential. The checksum
249 * verifier for the next block should be bp's checksum plus 1.
251 * Also check the log chain linkage and size used.
253 cksum.zc_word[ZIL_ZC_SEQ]++;
255 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
256 zil_chain_t *zilc = abuf->b_data;
257 char *lr = (char *)(zilc + 1);
258 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
260 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
261 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
262 error = SET_ERROR(ECKSUM);
264 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
266 *end = (char *)dst + len;
267 *nbp = zilc->zc_next_blk;
270 char *lr = abuf->b_data;
271 uint64_t size = BP_GET_LSIZE(bp);
272 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
274 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
275 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
276 (zilc->zc_nused > (size - sizeof (*zilc)))) {
277 error = SET_ERROR(ECKSUM);
279 ASSERT3U(zilc->zc_nused, <=,
280 SPA_OLD_MAXBLOCKSIZE);
281 bcopy(lr, dst, zilc->zc_nused);
282 *end = (char *)dst + zilc->zc_nused;
283 *nbp = zilc->zc_next_blk;
287 arc_buf_destroy(abuf, &abuf);
294 * Read a TX_WRITE log data block.
297 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
299 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
300 const blkptr_t *bp = &lr->lr_blkptr;
301 arc_flags_t aflags = ARC_FLAG_WAIT;
302 arc_buf_t *abuf = NULL;
306 if (BP_IS_HOLE(bp)) {
308 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
312 if (zilog->zl_header->zh_claim_txg == 0)
313 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
316 * If we are not using the resulting data, we are just checking that
317 * it hasn't been corrupted so we don't need to waste CPU time
318 * decompressing and decrypting it.
321 zio_flags |= ZIO_FLAG_RAW;
323 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
324 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
326 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
327 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
331 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
332 arc_buf_destroy(abuf, &abuf);
339 * Parse the intent log, and call parse_func for each valid record within.
342 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
343 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
346 const zil_header_t *zh = zilog->zl_header;
347 boolean_t claimed = !!zh->zh_claim_txg;
348 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
349 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
350 uint64_t max_blk_seq = 0;
351 uint64_t max_lr_seq = 0;
352 uint64_t blk_count = 0;
353 uint64_t lr_count = 0;
354 blkptr_t blk, next_blk;
358 bzero(&next_blk, sizeof (blkptr_t));
361 * Old logs didn't record the maximum zh_claim_lr_seq.
363 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
364 claim_lr_seq = UINT64_MAX;
367 * Starting at the block pointed to by zh_log we read the log chain.
368 * For each block in the chain we strongly check that block to
369 * ensure its validity. We stop when an invalid block is found.
370 * For each block pointer in the chain we call parse_blk_func().
371 * For each record in each valid block we call parse_lr_func().
372 * If the log has been claimed, stop if we encounter a sequence
373 * number greater than the highest claimed sequence number.
375 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
376 zil_bp_tree_init(zilog);
378 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
379 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
383 if (blk_seq > claim_blk_seq)
386 error = parse_blk_func(zilog, &blk, arg, txg);
389 ASSERT3U(max_blk_seq, <, blk_seq);
390 max_blk_seq = blk_seq;
393 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
396 error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
401 for (lrp = lrbuf; lrp < end; lrp += reclen) {
402 lr_t *lr = (lr_t *)lrp;
403 reclen = lr->lrc_reclen;
404 ASSERT3U(reclen, >=, sizeof (lr_t));
405 if (lr->lrc_seq > claim_lr_seq)
408 error = parse_lr_func(zilog, lr, arg, txg);
411 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
412 max_lr_seq = lr->lrc_seq;
417 zilog->zl_parse_error = error;
418 zilog->zl_parse_blk_seq = max_blk_seq;
419 zilog->zl_parse_lr_seq = max_lr_seq;
420 zilog->zl_parse_blk_count = blk_count;
421 zilog->zl_parse_lr_count = lr_count;
423 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
424 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq) ||
425 (decrypt && error == EIO));
427 zil_bp_tree_fini(zilog);
428 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
435 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
438 ASSERT(!BP_IS_HOLE(bp));
441 * As we call this function from the context of a rewind to a
442 * checkpoint, each ZIL block whose txg is later than the txg
443 * that we rewind to is invalid. Thus, we return -1 so
444 * zil_parse() doesn't attempt to read it.
446 if (bp->blk_birth >= first_txg)
449 if (zil_bp_tree_add(zilog, bp) != 0)
452 zio_free(zilog->zl_spa, first_txg, bp);
458 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
465 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
469 * Claim log block if not already committed and not already claimed.
470 * If tx == NULL, just verify that the block is claimable.
472 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
473 zil_bp_tree_add(zilog, bp) != 0)
476 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
477 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
478 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
482 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
485 lr_write_t *lr = (lr_write_t *)lrc;
488 if (lrc->lrc_txtype != TX_WRITE)
492 * If the block is not readable, don't claim it. This can happen
493 * in normal operation when a log block is written to disk before
494 * some of the dmu_sync() blocks it points to. In this case, the
495 * transaction cannot have been committed to anyone (we would have
496 * waited for all writes to be stable first), so it is semantically
497 * correct to declare this the end of the log.
499 if (lr->lr_blkptr.blk_birth >= first_txg) {
500 error = zil_read_log_data(zilog, lr, NULL);
505 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
510 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
513 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
519 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
522 lr_write_t *lr = (lr_write_t *)lrc;
523 blkptr_t *bp = &lr->lr_blkptr;
526 * If we previously claimed it, we need to free it.
528 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
529 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
531 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
537 zil_lwb_vdev_compare(const void *x1, const void *x2)
539 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
540 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
542 return (TREE_CMP(v1, v2));
546 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
551 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
552 lwb->lwb_zilog = zilog;
554 lwb->lwb_fastwrite = fastwrite;
555 lwb->lwb_slog = slog;
556 lwb->lwb_state = LWB_STATE_CLOSED;
557 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
558 lwb->lwb_max_txg = txg;
559 lwb->lwb_write_zio = NULL;
560 lwb->lwb_root_zio = NULL;
562 lwb->lwb_issued_timestamp = 0;
563 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
564 lwb->lwb_nused = sizeof (zil_chain_t);
565 lwb->lwb_sz = BP_GET_LSIZE(bp);
568 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
571 mutex_enter(&zilog->zl_lock);
572 list_insert_tail(&zilog->zl_lwb_list, lwb);
573 mutex_exit(&zilog->zl_lock);
575 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
576 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
577 VERIFY(list_is_empty(&lwb->lwb_waiters));
578 VERIFY(list_is_empty(&lwb->lwb_itxs));
584 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
586 ASSERT(MUTEX_HELD(&zilog->zl_lock));
587 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
588 VERIFY(list_is_empty(&lwb->lwb_waiters));
589 VERIFY(list_is_empty(&lwb->lwb_itxs));
590 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
591 ASSERT3P(lwb->lwb_write_zio, ==, NULL);
592 ASSERT3P(lwb->lwb_root_zio, ==, NULL);
593 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
594 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
595 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
598 * Clear the zilog's field to indicate this lwb is no longer
599 * valid, and prevent use-after-free errors.
601 if (zilog->zl_last_lwb_opened == lwb)
602 zilog->zl_last_lwb_opened = NULL;
604 kmem_cache_free(zil_lwb_cache, lwb);
608 * Called when we create in-memory log transactions so that we know
609 * to cleanup the itxs at the end of spa_sync().
612 zilog_dirty(zilog_t *zilog, uint64_t txg)
614 dsl_pool_t *dp = zilog->zl_dmu_pool;
615 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
617 ASSERT(spa_writeable(zilog->zl_spa));
619 if (ds->ds_is_snapshot)
620 panic("dirtying snapshot!");
622 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
623 /* up the hold count until we can be written out */
624 dmu_buf_add_ref(ds->ds_dbuf, zilog);
626 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
631 * Determine if the zil is dirty in the specified txg. Callers wanting to
632 * ensure that the dirty state does not change must hold the itxg_lock for
633 * the specified txg. Holding the lock will ensure that the zil cannot be
634 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
637 static boolean_t __maybe_unused
638 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
640 dsl_pool_t *dp = zilog->zl_dmu_pool;
642 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
648 * Determine if the zil is dirty. The zil is considered dirty if it has
649 * any pending itx records that have not been cleaned by zil_clean().
652 zilog_is_dirty(zilog_t *zilog)
654 dsl_pool_t *dp = zilog->zl_dmu_pool;
656 for (int t = 0; t < TXG_SIZE; t++) {
657 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
664 * Create an on-disk intent log.
667 zil_create(zilog_t *zilog)
669 const zil_header_t *zh = zilog->zl_header;
675 boolean_t fastwrite = FALSE;
676 boolean_t slog = FALSE;
679 * Wait for any previous destroy to complete.
681 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
683 ASSERT(zh->zh_claim_txg == 0);
684 ASSERT(zh->zh_replay_seq == 0);
689 * Allocate an initial log block if:
690 * - there isn't one already
691 * - the existing block is the wrong endianness
693 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
694 tx = dmu_tx_create(zilog->zl_os);
695 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
696 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
697 txg = dmu_tx_get_txg(tx);
699 if (!BP_IS_HOLE(&blk)) {
700 zio_free(zilog->zl_spa, txg, &blk);
704 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
705 ZIL_MIN_BLKSZ, &slog);
709 zil_init_log_chain(zilog, &blk);
713 * Allocate a log write block (lwb) for the first log block.
716 lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
719 * If we just allocated the first log block, commit our transaction
720 * and wait for zil_sync() to stuff the block pointer into zh_log.
721 * (zh is part of the MOS, so we cannot modify it in open context.)
725 txg_wait_synced(zilog->zl_dmu_pool, txg);
728 ASSERT(error != 0 || bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
729 IMPLY(error == 0, lwb != NULL);
735 * In one tx, free all log blocks and clear the log header. If keep_first
736 * is set, then we're replaying a log with no content. We want to keep the
737 * first block, however, so that the first synchronous transaction doesn't
738 * require a txg_wait_synced() in zil_create(). We don't need to
739 * txg_wait_synced() here either when keep_first is set, because both
740 * zil_create() and zil_destroy() will wait for any in-progress destroys
744 zil_destroy(zilog_t *zilog, boolean_t keep_first)
746 const zil_header_t *zh = zilog->zl_header;
752 * Wait for any previous destroy to complete.
754 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
756 zilog->zl_old_header = *zh; /* debugging aid */
758 if (BP_IS_HOLE(&zh->zh_log))
761 tx = dmu_tx_create(zilog->zl_os);
762 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
763 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
764 txg = dmu_tx_get_txg(tx);
766 mutex_enter(&zilog->zl_lock);
768 ASSERT3U(zilog->zl_destroy_txg, <, txg);
769 zilog->zl_destroy_txg = txg;
770 zilog->zl_keep_first = keep_first;
772 if (!list_is_empty(&zilog->zl_lwb_list)) {
773 ASSERT(zh->zh_claim_txg == 0);
775 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
776 if (lwb->lwb_fastwrite)
777 metaslab_fastwrite_unmark(zilog->zl_spa,
780 list_remove(&zilog->zl_lwb_list, lwb);
781 if (lwb->lwb_buf != NULL)
782 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
783 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
784 zil_free_lwb(zilog, lwb);
786 } else if (!keep_first) {
787 zil_destroy_sync(zilog, tx);
789 mutex_exit(&zilog->zl_lock);
795 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
797 ASSERT(list_is_empty(&zilog->zl_lwb_list));
798 (void) zil_parse(zilog, zil_free_log_block,
799 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
803 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
805 dmu_tx_t *tx = txarg;
812 error = dmu_objset_own_obj(dp, ds->ds_object,
813 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
816 * EBUSY indicates that the objset is inconsistent, in which
817 * case it can not have a ZIL.
819 if (error != EBUSY) {
820 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
821 (unsigned long long)ds->ds_object, error);
827 zilog = dmu_objset_zil(os);
828 zh = zil_header_in_syncing_context(zilog);
829 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
830 first_txg = spa_min_claim_txg(zilog->zl_spa);
833 * If the spa_log_state is not set to be cleared, check whether
834 * the current uberblock is a checkpoint one and if the current
835 * header has been claimed before moving on.
837 * If the current uberblock is a checkpointed uberblock then
838 * one of the following scenarios took place:
840 * 1] We are currently rewinding to the checkpoint of the pool.
841 * 2] We crashed in the middle of a checkpoint rewind but we
842 * did manage to write the checkpointed uberblock to the
843 * vdev labels, so when we tried to import the pool again
844 * the checkpointed uberblock was selected from the import
847 * In both cases we want to zero out all the ZIL blocks, except
848 * the ones that have been claimed at the time of the checkpoint
849 * (their zh_claim_txg != 0). The reason is that these blocks
850 * may be corrupted since we may have reused their locations on
851 * disk after we took the checkpoint.
853 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
854 * when we first figure out whether the current uberblock is
855 * checkpointed or not. Unfortunately, that would discard all
856 * the logs, including the ones that are claimed, and we would
859 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
860 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
861 zh->zh_claim_txg == 0)) {
862 if (!BP_IS_HOLE(&zh->zh_log)) {
863 (void) zil_parse(zilog, zil_clear_log_block,
864 zil_noop_log_record, tx, first_txg, B_FALSE);
866 BP_ZERO(&zh->zh_log);
867 if (os->os_encrypted)
868 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
869 dsl_dataset_dirty(dmu_objset_ds(os), tx);
870 dmu_objset_disown(os, B_FALSE, FTAG);
875 * If we are not rewinding and opening the pool normally, then
876 * the min_claim_txg should be equal to the first txg of the pool.
878 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
881 * Claim all log blocks if we haven't already done so, and remember
882 * the highest claimed sequence number. This ensures that if we can
883 * read only part of the log now (e.g. due to a missing device),
884 * but we can read the entire log later, we will not try to replay
885 * or destroy beyond the last block we successfully claimed.
887 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
888 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
889 (void) zil_parse(zilog, zil_claim_log_block,
890 zil_claim_log_record, tx, first_txg, B_FALSE);
891 zh->zh_claim_txg = first_txg;
892 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
893 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
894 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
895 zh->zh_flags |= ZIL_REPLAY_NEEDED;
896 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
897 if (os->os_encrypted)
898 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
899 dsl_dataset_dirty(dmu_objset_ds(os), tx);
902 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
903 dmu_objset_disown(os, B_FALSE, FTAG);
908 * Check the log by walking the log chain.
909 * Checksum errors are ok as they indicate the end of the chain.
910 * Any other error (no device or read failure) returns an error.
914 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
923 error = dmu_objset_from_ds(ds, &os);
925 cmn_err(CE_WARN, "can't open objset %llu, error %d",
926 (unsigned long long)ds->ds_object, error);
930 zilog = dmu_objset_zil(os);
931 bp = (blkptr_t *)&zilog->zl_header->zh_log;
933 if (!BP_IS_HOLE(bp)) {
935 boolean_t valid = B_TRUE;
938 * Check the first block and determine if it's on a log device
939 * which may have been removed or faulted prior to loading this
940 * pool. If so, there's no point in checking the rest of the
941 * log as its content should have already been synced to the
944 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
945 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
946 if (vd->vdev_islog && vdev_is_dead(vd))
947 valid = vdev_log_state_valid(vd);
948 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
954 * Check whether the current uberblock is checkpointed (e.g.
955 * we are rewinding) and whether the current header has been
956 * claimed or not. If it hasn't then skip verifying it. We
957 * do this because its ZIL blocks may be part of the pool's
958 * state before the rewind, which is no longer valid.
960 zil_header_t *zh = zil_header_in_syncing_context(zilog);
961 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
962 zh->zh_claim_txg == 0)
967 * Because tx == NULL, zil_claim_log_block() will not actually claim
968 * any blocks, but just determine whether it is possible to do so.
969 * In addition to checking the log chain, zil_claim_log_block()
970 * will invoke zio_claim() with a done func of spa_claim_notify(),
971 * which will update spa_max_claim_txg. See spa_load() for details.
973 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
974 zilog->zl_header->zh_claim_txg ? -1ULL :
975 spa_min_claim_txg(os->os_spa), B_FALSE);
977 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
981 * When an itx is "skipped", this function is used to properly mark the
982 * waiter as "done, and signal any thread(s) waiting on it. An itx can
983 * be skipped (and not committed to an lwb) for a variety of reasons,
984 * one of them being that the itx was committed via spa_sync(), prior to
985 * it being committed to an lwb; this can happen if a thread calling
986 * zil_commit() is racing with spa_sync().
989 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
991 mutex_enter(&zcw->zcw_lock);
992 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
993 zcw->zcw_done = B_TRUE;
994 cv_broadcast(&zcw->zcw_cv);
995 mutex_exit(&zcw->zcw_lock);
999 * This function is used when the given waiter is to be linked into an
1000 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1001 * At this point, the waiter will no longer be referenced by the itx,
1002 * and instead, will be referenced by the lwb.
1005 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1008 * The lwb_waiters field of the lwb is protected by the zilog's
1009 * zl_lock, thus it must be held when calling this function.
1011 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1013 mutex_enter(&zcw->zcw_lock);
1014 ASSERT(!list_link_active(&zcw->zcw_node));
1015 ASSERT3P(zcw->zcw_lwb, ==, NULL);
1016 ASSERT3P(lwb, !=, NULL);
1017 ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
1018 lwb->lwb_state == LWB_STATE_ISSUED ||
1019 lwb->lwb_state == LWB_STATE_WRITE_DONE);
1021 list_insert_tail(&lwb->lwb_waiters, zcw);
1023 mutex_exit(&zcw->zcw_lock);
1027 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1028 * block, and the given waiter must be linked to the "nolwb waiters"
1029 * list inside of zil_process_commit_list().
1032 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1034 mutex_enter(&zcw->zcw_lock);
1035 ASSERT(!list_link_active(&zcw->zcw_node));
1036 ASSERT3P(zcw->zcw_lwb, ==, NULL);
1037 list_insert_tail(nolwb, zcw);
1038 mutex_exit(&zcw->zcw_lock);
1042 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1044 avl_tree_t *t = &lwb->lwb_vdev_tree;
1046 zil_vdev_node_t *zv, zvsearch;
1047 int ndvas = BP_GET_NDVAS(bp);
1050 if (zil_nocacheflush)
1053 mutex_enter(&lwb->lwb_vdev_lock);
1054 for (i = 0; i < ndvas; i++) {
1055 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1056 if (avl_find(t, &zvsearch, &where) == NULL) {
1057 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1058 zv->zv_vdev = zvsearch.zv_vdev;
1059 avl_insert(t, zv, where);
1062 mutex_exit(&lwb->lwb_vdev_lock);
1066 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1068 avl_tree_t *src = &lwb->lwb_vdev_tree;
1069 avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1070 void *cookie = NULL;
1071 zil_vdev_node_t *zv;
1073 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1074 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1075 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1078 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1079 * not need the protection of lwb_vdev_lock (it will only be modified
1080 * while holding zilog->zl_lock) as its writes and those of its
1081 * children have all completed. The younger 'nlwb' may be waiting on
1082 * future writes to additional vdevs.
1084 mutex_enter(&nlwb->lwb_vdev_lock);
1086 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1087 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1089 while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1092 if (avl_find(dst, zv, &where) == NULL) {
1093 avl_insert(dst, zv, where);
1095 kmem_free(zv, sizeof (*zv));
1098 mutex_exit(&nlwb->lwb_vdev_lock);
1102 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1104 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1108 * This function is a called after all vdevs associated with a given lwb
1109 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1110 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1111 * all "previous" lwb's will have completed before this function is
1112 * called; i.e. this function is called for all previous lwbs before
1113 * it's called for "this" lwb (enforced via zio the dependencies
1114 * configured in zil_lwb_set_zio_dependency()).
1116 * The intention is for this function to be called as soon as the
1117 * contents of an lwb are considered "stable" on disk, and will survive
1118 * any sudden loss of power. At this point, any threads waiting for the
1119 * lwb to reach this state are signalled, and the "waiter" structures
1120 * are marked "done".
1123 zil_lwb_flush_vdevs_done(zio_t *zio)
1125 lwb_t *lwb = zio->io_private;
1126 zilog_t *zilog = lwb->lwb_zilog;
1127 dmu_tx_t *tx = lwb->lwb_tx;
1128 zil_commit_waiter_t *zcw;
1131 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1133 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1135 mutex_enter(&zilog->zl_lock);
1138 * Ensure the lwb buffer pointer is cleared before releasing the
1139 * txg. If we have had an allocation failure and the txg is
1140 * waiting to sync then we want zil_sync() to remove the lwb so
1141 * that it's not picked up as the next new one in
1142 * zil_process_commit_list(). zil_sync() will only remove the
1143 * lwb if lwb_buf is null.
1145 lwb->lwb_buf = NULL;
1148 ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1149 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1151 lwb->lwb_root_zio = NULL;
1153 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1154 lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1156 if (zilog->zl_last_lwb_opened == lwb) {
1158 * Remember the highest committed log sequence number
1159 * for ztest. We only update this value when all the log
1160 * writes succeeded, because ztest wants to ASSERT that
1161 * it got the whole log chain.
1163 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1166 while ((itx = list_head(&lwb->lwb_itxs)) != NULL) {
1167 list_remove(&lwb->lwb_itxs, itx);
1168 zil_itx_destroy(itx);
1171 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1172 mutex_enter(&zcw->zcw_lock);
1174 ASSERT(list_link_active(&zcw->zcw_node));
1175 list_remove(&lwb->lwb_waiters, zcw);
1177 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1178 zcw->zcw_lwb = NULL;
1180 zcw->zcw_zio_error = zio->io_error;
1182 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1183 zcw->zcw_done = B_TRUE;
1184 cv_broadcast(&zcw->zcw_cv);
1186 mutex_exit(&zcw->zcw_lock);
1189 mutex_exit(&zilog->zl_lock);
1192 * Now that we've written this log block, we have a stable pointer
1193 * to the next block in the chain, so it's OK to let the txg in
1194 * which we allocated the next block sync.
1200 * This is called when an lwb's write zio completes. The callback's
1201 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1202 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1203 * in writing out this specific lwb's data, and in the case that cache
1204 * flushes have been deferred, vdevs involved in writing the data for
1205 * previous lwbs. The writes corresponding to all the vdevs in the
1206 * lwb_vdev_tree will have completed by the time this is called, due to
1207 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1208 * which takes deferred flushes into account. The lwb will be "done"
1209 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1210 * completion callback for the lwb's root zio.
1213 zil_lwb_write_done(zio_t *zio)
1215 lwb_t *lwb = zio->io_private;
1216 spa_t *spa = zio->io_spa;
1217 zilog_t *zilog = lwb->lwb_zilog;
1218 avl_tree_t *t = &lwb->lwb_vdev_tree;
1219 void *cookie = NULL;
1220 zil_vdev_node_t *zv;
1223 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1225 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1226 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1227 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1228 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1229 ASSERT(!BP_IS_GANG(zio->io_bp));
1230 ASSERT(!BP_IS_HOLE(zio->io_bp));
1231 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1233 abd_put(zio->io_abd);
1235 mutex_enter(&zilog->zl_lock);
1236 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1237 lwb->lwb_state = LWB_STATE_WRITE_DONE;
1238 lwb->lwb_write_zio = NULL;
1239 lwb->lwb_fastwrite = FALSE;
1240 nlwb = list_next(&zilog->zl_lwb_list, lwb);
1241 mutex_exit(&zilog->zl_lock);
1243 if (avl_numnodes(t) == 0)
1247 * If there was an IO error, we're not going to call zio_flush()
1248 * on these vdevs, so we simply empty the tree and free the
1249 * nodes. We avoid calling zio_flush() since there isn't any
1250 * good reason for doing so, after the lwb block failed to be
1253 if (zio->io_error != 0) {
1254 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1255 kmem_free(zv, sizeof (*zv));
1260 * If this lwb does not have any threads waiting for it to
1261 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1262 * command to the vdevs written to by "this" lwb, and instead
1263 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1264 * command for those vdevs. Thus, we merge the vdev tree of
1265 * "this" lwb with the vdev tree of the "next" lwb in the list,
1266 * and assume the "next" lwb will handle flushing the vdevs (or
1267 * deferring the flush(s) again).
1269 * This is a useful performance optimization, especially for
1270 * workloads with lots of async write activity and few sync
1271 * write and/or fsync activity, as it has the potential to
1272 * coalesce multiple flush commands to a vdev into one.
1274 if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) {
1275 zil_lwb_flush_defer(lwb, nlwb);
1276 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1280 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1281 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1283 zio_flush(lwb->lwb_root_zio, vd);
1284 kmem_free(zv, sizeof (*zv));
1289 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1291 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1293 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1294 ASSERT(MUTEX_HELD(&zilog->zl_lock));
1297 * The zilog's "zl_last_lwb_opened" field is used to build the
1298 * lwb/zio dependency chain, which is used to preserve the
1299 * ordering of lwb completions that is required by the semantics
1300 * of the ZIL. Each new lwb zio becomes a parent of the
1301 * "previous" lwb zio, such that the new lwb's zio cannot
1302 * complete until the "previous" lwb's zio completes.
1304 * This is required by the semantics of zil_commit(); the commit
1305 * waiters attached to the lwbs will be woken in the lwb zio's
1306 * completion callback, so this zio dependency graph ensures the
1307 * waiters are woken in the correct order (the same order the
1308 * lwbs were created).
1310 if (last_lwb_opened != NULL &&
1311 last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1312 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1313 last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1314 last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1316 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1317 zio_add_child(lwb->lwb_root_zio,
1318 last_lwb_opened->lwb_root_zio);
1321 * If the previous lwb's write hasn't already completed,
1322 * we also want to order the completion of the lwb write
1323 * zios (above, we only order the completion of the lwb
1324 * root zios). This is required because of how we can
1325 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1327 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1328 * the previous lwb will rely on this lwb to flush the
1329 * vdevs written to by that previous lwb. Thus, we need
1330 * to ensure this lwb doesn't issue the flush until
1331 * after the previous lwb's write completes. We ensure
1332 * this ordering by setting the zio parent/child
1333 * relationship here.
1335 * Without this relationship on the lwb's write zio,
1336 * it's possible for this lwb's write to complete prior
1337 * to the previous lwb's write completing; and thus, the
1338 * vdevs for the previous lwb would be flushed prior to
1339 * that lwb's data being written to those vdevs (the
1340 * vdevs are flushed in the lwb write zio's completion
1341 * handler, zil_lwb_write_done()).
1343 if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1344 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1345 last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1347 ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1348 zio_add_child(lwb->lwb_write_zio,
1349 last_lwb_opened->lwb_write_zio);
1356 * This function's purpose is to "open" an lwb such that it is ready to
1357 * accept new itxs being committed to it. To do this, the lwb's zio
1358 * structures are created, and linked to the lwb. This function is
1359 * idempotent; if the passed in lwb has already been opened, this
1360 * function is essentially a no-op.
1363 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1365 zbookmark_phys_t zb;
1366 zio_priority_t prio;
1368 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1369 ASSERT3P(lwb, !=, NULL);
1370 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1371 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1373 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1374 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1375 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1377 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1378 mutex_enter(&zilog->zl_lock);
1379 if (lwb->lwb_root_zio == NULL) {
1380 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1381 BP_GET_LSIZE(&lwb->lwb_blk));
1383 if (!lwb->lwb_fastwrite) {
1384 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
1385 lwb->lwb_fastwrite = 1;
1388 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1389 prio = ZIO_PRIORITY_SYNC_WRITE;
1391 prio = ZIO_PRIORITY_ASYNC_WRITE;
1393 lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1394 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1395 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1397 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1398 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1399 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1400 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
1401 ZIO_FLAG_FASTWRITE, &zb);
1402 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1404 lwb->lwb_state = LWB_STATE_OPENED;
1406 zil_lwb_set_zio_dependency(zilog, lwb);
1407 zilog->zl_last_lwb_opened = lwb;
1409 mutex_exit(&zilog->zl_lock);
1411 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1412 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1413 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1417 * Define a limited set of intent log block sizes.
1419 * These must be a multiple of 4KB. Note only the amount used (again
1420 * aligned to 4KB) actually gets written. However, we can't always just
1421 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1426 } zil_block_buckets[] = {
1427 { 4096, 4096 }, /* non TX_WRITE */
1428 { 8192 + 4096, 8192 + 4096 }, /* database */
1429 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */
1430 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */
1431 { 131072, 131072 }, /* < 128KB writes */
1432 { 131072 +4096, 65536 + 4096 }, /* 128KB writes */
1433 { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */
1437 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1438 * initialized. Otherwise this should not be used directly; see
1439 * zl_max_block_size instead.
1441 int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1444 * Start a log block write and advance to the next log block.
1445 * Calls are serialized.
1448 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1452 spa_t *spa = zilog->zl_spa;
1456 uint64_t zil_blksz, wsz;
1460 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1461 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1462 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1463 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1465 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1466 zilc = (zil_chain_t *)lwb->lwb_buf;
1467 bp = &zilc->zc_next_blk;
1469 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1470 bp = &zilc->zc_next_blk;
1473 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1476 * Allocate the next block and save its address in this block
1477 * before writing it in order to establish the log chain.
1478 * Note that if the allocation of nlwb synced before we wrote
1479 * the block that points at it (lwb), we'd leak it if we crashed.
1480 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1481 * We dirty the dataset to ensure that zil_sync() will be called
1482 * to clean up in the event of allocation failure or I/O failure.
1485 tx = dmu_tx_create(zilog->zl_os);
1488 * Since we are not going to create any new dirty data, and we
1489 * can even help with clearing the existing dirty data, we
1490 * should not be subject to the dirty data based delays. We
1491 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1493 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1495 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1496 txg = dmu_tx_get_txg(tx);
1501 * Log blocks are pre-allocated. Here we select the size of the next
1502 * block, based on size used in the last block.
1503 * - first find the smallest bucket that will fit the block from a
1504 * limited set of block sizes. This is because it's faster to write
1505 * blocks allocated from the same metaslab as they are adjacent or
1507 * - next find the maximum from the new suggested size and an array of
1508 * previous sizes. This lessens a picket fence effect of wrongly
1509 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1512 * Note we only write what is used, but we can't just allocate
1513 * the maximum block size because we can exhaust the available
1516 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1517 for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1519 zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1520 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1521 for (i = 0; i < ZIL_PREV_BLKS; i++)
1522 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1523 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1526 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1528 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
1529 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
1531 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
1532 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
1535 ASSERT3U(bp->blk_birth, ==, txg);
1536 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1537 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1540 * Allocate a new log write block (lwb).
1542 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1545 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1546 /* For Slim ZIL only write what is used. */
1547 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1548 ASSERT3U(wsz, <=, lwb->lwb_sz);
1549 zio_shrink(lwb->lwb_write_zio, wsz);
1556 zilc->zc_nused = lwb->lwb_nused;
1557 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1560 * clear unused data for security
1562 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1564 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1566 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1567 lwb->lwb_issued_timestamp = gethrtime();
1568 lwb->lwb_state = LWB_STATE_ISSUED;
1570 zio_nowait(lwb->lwb_root_zio);
1571 zio_nowait(lwb->lwb_write_zio);
1574 * If there was an allocation failure then nlwb will be null which
1575 * forces a txg_wait_synced().
1581 * Maximum amount of write data that can be put into single log block.
1584 zil_max_log_data(zilog_t *zilog)
1586 return (zilog->zl_max_block_size -
1587 sizeof (zil_chain_t) - sizeof (lr_write_t));
1591 * Maximum amount of log space we agree to waste to reduce number of
1592 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1594 static inline uint64_t
1595 zil_max_waste_space(zilog_t *zilog)
1597 return (zil_max_log_data(zilog) / 8);
1601 * Maximum amount of write data for WR_COPIED. For correctness, consumers
1602 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1603 * maximum sized log block, because each WR_COPIED record must fit in a
1604 * single log block. For space efficiency, we want to fit two records into a
1605 * max-sized log block.
1608 zil_max_copied_data(zilog_t *zilog)
1610 return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1611 sizeof (lr_write_t));
1615 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1618 lr_write_t *lrwb, *lrw;
1620 uint64_t dlen, dnow, lwb_sp, reclen, txg, max_log_data;
1622 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1623 ASSERT3P(lwb, !=, NULL);
1624 ASSERT3P(lwb->lwb_buf, !=, NULL);
1626 zil_lwb_write_open(zilog, lwb);
1629 lrw = (lr_write_t *)lrc;
1632 * A commit itx doesn't represent any on-disk state; instead
1633 * it's simply used as a place holder on the commit list, and
1634 * provides a mechanism for attaching a "commit waiter" onto the
1635 * correct lwb (such that the waiter can be signalled upon
1636 * completion of that lwb). Thus, we don't process this itx's
1637 * log record if it's a commit itx (these itx's don't have log
1638 * records), and instead link the itx's waiter onto the lwb's
1641 * For more details, see the comment above zil_commit().
1643 if (lrc->lrc_txtype == TX_COMMIT) {
1644 mutex_enter(&zilog->zl_lock);
1645 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1646 itx->itx_private = NULL;
1647 mutex_exit(&zilog->zl_lock);
1651 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1652 dlen = P2ROUNDUP_TYPED(
1653 lrw->lr_length, sizeof (uint64_t), uint64_t);
1657 reclen = lrc->lrc_reclen;
1658 zilog->zl_cur_used += (reclen + dlen);
1661 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1665 * If this record won't fit in the current log block, start a new one.
1666 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1668 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1669 max_log_data = zil_max_log_data(zilog);
1670 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1671 lwb_sp < zil_max_waste_space(zilog) &&
1672 (dlen % max_log_data == 0 ||
1673 lwb_sp < reclen + dlen % max_log_data))) {
1674 lwb = zil_lwb_write_issue(zilog, lwb);
1677 zil_lwb_write_open(zilog, lwb);
1678 ASSERT(LWB_EMPTY(lwb));
1679 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1682 * There must be enough space in the new, empty log block to
1683 * hold reclen. For WR_COPIED, we need to fit the whole
1684 * record in one block, and reclen is the header size + the
1685 * data size. For WR_NEED_COPY, we can create multiple
1686 * records, splitting the data into multiple blocks, so we
1687 * only need to fit one word of data per block; in this case
1688 * reclen is just the header size (no data).
1690 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1693 dnow = MIN(dlen, lwb_sp - reclen);
1694 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1695 bcopy(lrc, lr_buf, reclen);
1696 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1697 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1699 ZIL_STAT_BUMP(zil_itx_count);
1702 * If it's a write, fetch the data or get its blkptr as appropriate.
1704 if (lrc->lrc_txtype == TX_WRITE) {
1705 if (txg > spa_freeze_txg(zilog->zl_spa))
1706 txg_wait_synced(zilog->zl_dmu_pool, txg);
1707 if (itx->itx_wr_state == WR_COPIED) {
1708 ZIL_STAT_BUMP(zil_itx_copied_count);
1709 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1714 if (itx->itx_wr_state == WR_NEED_COPY) {
1715 dbuf = lr_buf + reclen;
1716 lrcb->lrc_reclen += dnow;
1717 if (lrwb->lr_length > dnow)
1718 lrwb->lr_length = dnow;
1719 lrw->lr_offset += dnow;
1720 lrw->lr_length -= dnow;
1721 ZIL_STAT_BUMP(zil_itx_needcopy_count);
1722 ZIL_STAT_INCR(zil_itx_needcopy_bytes, dnow);
1724 ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
1726 ZIL_STAT_BUMP(zil_itx_indirect_count);
1727 ZIL_STAT_INCR(zil_itx_indirect_bytes,
1732 * We pass in the "lwb_write_zio" rather than
1733 * "lwb_root_zio" so that the "lwb_write_zio"
1734 * becomes the parent of any zio's created by
1735 * the "zl_get_data" callback. The vdevs are
1736 * flushed after the "lwb_write_zio" completes,
1737 * so we want to make sure that completion
1738 * callback waits for these additional zio's,
1739 * such that the vdevs used by those zio's will
1740 * be included in the lwb's vdev tree, and those
1741 * vdevs will be properly flushed. If we passed
1742 * in "lwb_root_zio" here, then these additional
1743 * vdevs may not be flushed; e.g. if these zio's
1744 * completed after "lwb_write_zio" completed.
1746 error = zilog->zl_get_data(itx->itx_private,
1747 lrwb, dbuf, lwb, lwb->lwb_write_zio);
1750 txg_wait_synced(zilog->zl_dmu_pool, txg);
1754 ASSERT(error == ENOENT || error == EEXIST ||
1762 * We're actually making an entry, so update lrc_seq to be the
1763 * log record sequence number. Note that this is generally not
1764 * equal to the itx sequence number because not all transactions
1765 * are synchronous, and sometimes spa_sync() gets there first.
1767 lrcb->lrc_seq = ++zilog->zl_lr_seq;
1768 lwb->lwb_nused += reclen + dnow;
1770 zil_lwb_add_txg(lwb, txg);
1772 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1773 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1777 zilog->zl_cur_used += reclen;
1785 zil_itx_create(uint64_t txtype, size_t lrsize)
1790 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1791 itxsize = offsetof(itx_t, itx_lr) + lrsize;
1793 itx = zio_data_buf_alloc(itxsize);
1794 itx->itx_lr.lrc_txtype = txtype;
1795 itx->itx_lr.lrc_reclen = lrsize;
1796 itx->itx_lr.lrc_seq = 0; /* defensive */
1797 itx->itx_sync = B_TRUE; /* default is synchronous */
1798 itx->itx_callback = NULL;
1799 itx->itx_callback_data = NULL;
1800 itx->itx_size = itxsize;
1806 zil_itx_destroy(itx_t *itx)
1808 IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
1809 IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1811 if (itx->itx_callback != NULL)
1812 itx->itx_callback(itx->itx_callback_data);
1814 zio_data_buf_free(itx, itx->itx_size);
1818 * Free up the sync and async itxs. The itxs_t has already been detached
1819 * so no locks are needed.
1822 zil_itxg_clean(itxs_t *itxs)
1828 itx_async_node_t *ian;
1830 list = &itxs->i_sync_list;
1831 while ((itx = list_head(list)) != NULL) {
1833 * In the general case, commit itxs will not be found
1834 * here, as they'll be committed to an lwb via
1835 * zil_lwb_commit(), and free'd in that function. Having
1836 * said that, it is still possible for commit itxs to be
1837 * found here, due to the following race:
1839 * - a thread calls zil_commit() which assigns the
1840 * commit itx to a per-txg i_sync_list
1841 * - zil_itxg_clean() is called (e.g. via spa_sync())
1842 * while the waiter is still on the i_sync_list
1844 * There's nothing to prevent syncing the txg while the
1845 * waiter is on the i_sync_list. This normally doesn't
1846 * happen because spa_sync() is slower than zil_commit(),
1847 * but if zil_commit() calls txg_wait_synced() (e.g.
1848 * because zil_create() or zil_commit_writer_stall() is
1849 * called) we will hit this case.
1851 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1852 zil_commit_waiter_skip(itx->itx_private);
1854 list_remove(list, itx);
1855 zil_itx_destroy(itx);
1859 t = &itxs->i_async_tree;
1860 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1861 list = &ian->ia_list;
1862 while ((itx = list_head(list)) != NULL) {
1863 list_remove(list, itx);
1864 /* commit itxs should never be on the async lists. */
1865 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1866 zil_itx_destroy(itx);
1869 kmem_free(ian, sizeof (itx_async_node_t));
1873 kmem_free(itxs, sizeof (itxs_t));
1877 zil_aitx_compare(const void *x1, const void *x2)
1879 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1880 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1882 return (TREE_CMP(o1, o2));
1886 * Remove all async itx with the given oid.
1889 zil_remove_async(zilog_t *zilog, uint64_t oid)
1892 itx_async_node_t *ian;
1899 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1901 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1904 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1906 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1907 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1909 mutex_enter(&itxg->itxg_lock);
1910 if (itxg->itxg_txg != txg) {
1911 mutex_exit(&itxg->itxg_lock);
1916 * Locate the object node and append its list.
1918 t = &itxg->itxg_itxs->i_async_tree;
1919 ian = avl_find(t, &oid, &where);
1921 list_move_tail(&clean_list, &ian->ia_list);
1922 mutex_exit(&itxg->itxg_lock);
1924 while ((itx = list_head(&clean_list)) != NULL) {
1925 list_remove(&clean_list, itx);
1926 /* commit itxs should never be on the async lists. */
1927 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1928 zil_itx_destroy(itx);
1930 list_destroy(&clean_list);
1934 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1938 itxs_t *itxs, *clean = NULL;
1941 * Ensure the data of a renamed file is committed before the rename.
1943 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1944 zil_async_to_sync(zilog, itx->itx_oid);
1946 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1949 txg = dmu_tx_get_txg(tx);
1951 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1952 mutex_enter(&itxg->itxg_lock);
1953 itxs = itxg->itxg_itxs;
1954 if (itxg->itxg_txg != txg) {
1957 * The zil_clean callback hasn't got around to cleaning
1958 * this itxg. Save the itxs for release below.
1959 * This should be rare.
1961 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1962 "txg %llu", itxg->itxg_txg);
1963 clean = itxg->itxg_itxs;
1965 itxg->itxg_txg = txg;
1966 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
1969 list_create(&itxs->i_sync_list, sizeof (itx_t),
1970 offsetof(itx_t, itx_node));
1971 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1972 sizeof (itx_async_node_t),
1973 offsetof(itx_async_node_t, ia_node));
1975 if (itx->itx_sync) {
1976 list_insert_tail(&itxs->i_sync_list, itx);
1978 avl_tree_t *t = &itxs->i_async_tree;
1980 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
1981 itx_async_node_t *ian;
1984 ian = avl_find(t, &foid, &where);
1986 ian = kmem_alloc(sizeof (itx_async_node_t),
1988 list_create(&ian->ia_list, sizeof (itx_t),
1989 offsetof(itx_t, itx_node));
1990 ian->ia_foid = foid;
1991 avl_insert(t, ian, where);
1993 list_insert_tail(&ian->ia_list, itx);
1996 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1999 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2000 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2001 * need to be careful to always dirty the ZIL using the "real"
2002 * TXG (not itxg_txg) even when the SPA is frozen.
2004 zilog_dirty(zilog, dmu_tx_get_txg(tx));
2005 mutex_exit(&itxg->itxg_lock);
2007 /* Release the old itxs now we've dropped the lock */
2009 zil_itxg_clean(clean);
2013 * If there are any in-memory intent log transactions which have now been
2014 * synced then start up a taskq to free them. We should only do this after we
2015 * have written out the uberblocks (i.e. txg has been committed) so that
2016 * don't inadvertently clean out in-memory log records that would be required
2020 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2022 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2025 ASSERT3U(synced_txg, <, ZILTEST_TXG);
2027 mutex_enter(&itxg->itxg_lock);
2028 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2029 mutex_exit(&itxg->itxg_lock);
2032 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2033 ASSERT3U(itxg->itxg_txg, !=, 0);
2034 clean_me = itxg->itxg_itxs;
2035 itxg->itxg_itxs = NULL;
2037 mutex_exit(&itxg->itxg_lock);
2039 * Preferably start a task queue to free up the old itxs but
2040 * if taskq_dispatch can't allocate resources to do that then
2041 * free it in-line. This should be rare. Note, using TQ_SLEEP
2042 * created a bad performance problem.
2044 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2045 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2046 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2047 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP);
2048 if (id == TASKQID_INVALID)
2049 zil_itxg_clean(clean_me);
2053 * This function will traverse the queue of itxs that need to be
2054 * committed, and move them onto the ZIL's zl_itx_commit_list.
2057 zil_get_commit_list(zilog_t *zilog)
2060 list_t *commit_list = &zilog->zl_itx_commit_list;
2062 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2064 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2067 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2070 * This is inherently racy, since there is nothing to prevent
2071 * the last synced txg from changing. That's okay since we'll
2072 * only commit things in the future.
2074 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2075 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2077 mutex_enter(&itxg->itxg_lock);
2078 if (itxg->itxg_txg != txg) {
2079 mutex_exit(&itxg->itxg_lock);
2084 * If we're adding itx records to the zl_itx_commit_list,
2085 * then the zil better be dirty in this "txg". We can assert
2086 * that here since we're holding the itxg_lock which will
2087 * prevent spa_sync from cleaning it. Once we add the itxs
2088 * to the zl_itx_commit_list we must commit it to disk even
2089 * if it's unnecessary (i.e. the txg was synced).
2091 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2092 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2093 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
2095 mutex_exit(&itxg->itxg_lock);
2100 * Move the async itxs for a specified object to commit into sync lists.
2103 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2106 itx_async_node_t *ian;
2110 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2113 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2116 * This is inherently racy, since there is nothing to prevent
2117 * the last synced txg from changing.
2119 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2120 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2122 mutex_enter(&itxg->itxg_lock);
2123 if (itxg->itxg_txg != txg) {
2124 mutex_exit(&itxg->itxg_lock);
2129 * If a foid is specified then find that node and append its
2130 * list. Otherwise walk the tree appending all the lists
2131 * to the sync list. We add to the end rather than the
2132 * beginning to ensure the create has happened.
2134 t = &itxg->itxg_itxs->i_async_tree;
2136 ian = avl_find(t, &foid, &where);
2138 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2142 void *cookie = NULL;
2144 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2145 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2147 list_destroy(&ian->ia_list);
2148 kmem_free(ian, sizeof (itx_async_node_t));
2151 mutex_exit(&itxg->itxg_lock);
2156 * This function will prune commit itxs that are at the head of the
2157 * commit list (it won't prune past the first non-commit itx), and
2158 * either: a) attach them to the last lwb that's still pending
2159 * completion, or b) skip them altogether.
2161 * This is used as a performance optimization to prevent commit itxs
2162 * from generating new lwbs when it's unnecessary to do so.
2165 zil_prune_commit_list(zilog_t *zilog)
2169 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2171 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2172 lr_t *lrc = &itx->itx_lr;
2173 if (lrc->lrc_txtype != TX_COMMIT)
2176 mutex_enter(&zilog->zl_lock);
2178 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2179 if (last_lwb == NULL ||
2180 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2182 * All of the itxs this waiter was waiting on
2183 * must have already completed (or there were
2184 * never any itx's for it to wait on), so it's
2185 * safe to skip this waiter and mark it done.
2187 zil_commit_waiter_skip(itx->itx_private);
2189 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2190 itx->itx_private = NULL;
2193 mutex_exit(&zilog->zl_lock);
2195 list_remove(&zilog->zl_itx_commit_list, itx);
2196 zil_itx_destroy(itx);
2199 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2203 zil_commit_writer_stall(zilog_t *zilog)
2206 * When zio_alloc_zil() fails to allocate the next lwb block on
2207 * disk, we must call txg_wait_synced() to ensure all of the
2208 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2209 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2210 * to zil_process_commit_list()) will have to call zil_create(),
2211 * and start a new ZIL chain.
2213 * Since zil_alloc_zil() failed, the lwb that was previously
2214 * issued does not have a pointer to the "next" lwb on disk.
2215 * Thus, if another ZIL writer thread was to allocate the "next"
2216 * on-disk lwb, that block could be leaked in the event of a
2217 * crash (because the previous lwb on-disk would not point to
2220 * We must hold the zilog's zl_issuer_lock while we do this, to
2221 * ensure no new threads enter zil_process_commit_list() until
2222 * all lwb's in the zl_lwb_list have been synced and freed
2223 * (which is achieved via the txg_wait_synced() call).
2225 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2226 txg_wait_synced(zilog->zl_dmu_pool, 0);
2227 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2231 * This function will traverse the commit list, creating new lwbs as
2232 * needed, and committing the itxs from the commit list to these newly
2233 * created lwbs. Additionally, as a new lwb is created, the previous
2234 * lwb will be issued to the zio layer to be written to disk.
2237 zil_process_commit_list(zilog_t *zilog)
2239 spa_t *spa = zilog->zl_spa;
2241 list_t nolwb_waiters;
2245 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2248 * Return if there's nothing to commit before we dirty the fs by
2249 * calling zil_create().
2251 if (list_head(&zilog->zl_itx_commit_list) == NULL)
2254 list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2255 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2256 offsetof(zil_commit_waiter_t, zcw_node));
2258 lwb = list_tail(&zilog->zl_lwb_list);
2260 lwb = zil_create(zilog);
2262 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2263 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2264 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2267 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2268 lr_t *lrc = &itx->itx_lr;
2269 uint64_t txg = lrc->lrc_txg;
2271 ASSERT3U(txg, !=, 0);
2273 if (lrc->lrc_txtype == TX_COMMIT) {
2274 DTRACE_PROBE2(zil__process__commit__itx,
2275 zilog_t *, zilog, itx_t *, itx);
2277 DTRACE_PROBE2(zil__process__normal__itx,
2278 zilog_t *, zilog, itx_t *, itx);
2281 list_remove(&zilog->zl_itx_commit_list, itx);
2283 boolean_t synced = txg <= spa_last_synced_txg(spa);
2284 boolean_t frozen = txg > spa_freeze_txg(spa);
2287 * If the txg of this itx has already been synced out, then
2288 * we don't need to commit this itx to an lwb. This is
2289 * because the data of this itx will have already been
2290 * written to the main pool. This is inherently racy, and
2291 * it's still ok to commit an itx whose txg has already
2292 * been synced; this will result in a write that's
2293 * unnecessary, but will do no harm.
2295 * With that said, we always want to commit TX_COMMIT itxs
2296 * to an lwb, regardless of whether or not that itx's txg
2297 * has been synced out. We do this to ensure any OPENED lwb
2298 * will always have at least one zil_commit_waiter_t linked
2301 * As a counter-example, if we skipped TX_COMMIT itx's
2302 * whose txg had already been synced, the following
2303 * situation could occur if we happened to be racing with
2306 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2307 * itx's txg is 10 and the last synced txg is 9.
2308 * 2. spa_sync finishes syncing out txg 10.
2309 * 3. We move to the next itx in the list, it's a TX_COMMIT
2310 * whose txg is 10, so we skip it rather than committing
2311 * it to the lwb used in (1).
2313 * If the itx that is skipped in (3) is the last TX_COMMIT
2314 * itx in the commit list, than it's possible for the lwb
2315 * used in (1) to remain in the OPENED state indefinitely.
2317 * To prevent the above scenario from occurring, ensuring
2318 * that once an lwb is OPENED it will transition to ISSUED
2319 * and eventually DONE, we always commit TX_COMMIT itx's to
2320 * an lwb here, even if that itx's txg has already been
2323 * Finally, if the pool is frozen, we _always_ commit the
2324 * itx. The point of freezing the pool is to prevent data
2325 * from being written to the main pool via spa_sync, and
2326 * instead rely solely on the ZIL to persistently store the
2327 * data; i.e. when the pool is frozen, the last synced txg
2328 * value can't be trusted.
2330 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2332 lwb = zil_lwb_commit(zilog, itx, lwb);
2335 list_insert_tail(&nolwb_itxs, itx);
2337 list_insert_tail(&lwb->lwb_itxs, itx);
2339 if (lrc->lrc_txtype == TX_COMMIT) {
2340 zil_commit_waiter_link_nolwb(
2341 itx->itx_private, &nolwb_waiters);
2344 list_insert_tail(&nolwb_itxs, itx);
2347 ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2348 zil_itx_destroy(itx);
2354 * This indicates zio_alloc_zil() failed to allocate the
2355 * "next" lwb on-disk. When this happens, we must stall
2356 * the ZIL write pipeline; see the comment within
2357 * zil_commit_writer_stall() for more details.
2359 zil_commit_writer_stall(zilog);
2362 * Additionally, we have to signal and mark the "nolwb"
2363 * waiters as "done" here, since without an lwb, we
2364 * can't do this via zil_lwb_flush_vdevs_done() like
2367 zil_commit_waiter_t *zcw;
2368 while ((zcw = list_head(&nolwb_waiters)) != NULL) {
2369 zil_commit_waiter_skip(zcw);
2370 list_remove(&nolwb_waiters, zcw);
2374 * And finally, we have to destroy the itx's that
2375 * couldn't be committed to an lwb; this will also call
2376 * the itx's callback if one exists for the itx.
2378 while ((itx = list_head(&nolwb_itxs)) != NULL) {
2379 list_remove(&nolwb_itxs, itx);
2380 zil_itx_destroy(itx);
2383 ASSERT(list_is_empty(&nolwb_waiters));
2384 ASSERT3P(lwb, !=, NULL);
2385 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2386 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2387 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2390 * At this point, the ZIL block pointed at by the "lwb"
2391 * variable is in one of the following states: "closed"
2394 * If it's "closed", then no itxs have been committed to
2395 * it, so there's no point in issuing its zio (i.e. it's
2398 * If it's "open", then it contains one or more itxs that
2399 * eventually need to be committed to stable storage. In
2400 * this case we intentionally do not issue the lwb's zio
2401 * to disk yet, and instead rely on one of the following
2402 * two mechanisms for issuing the zio:
2404 * 1. Ideally, there will be more ZIL activity occurring
2405 * on the system, such that this function will be
2406 * immediately called again (not necessarily by the same
2407 * thread) and this lwb's zio will be issued via
2408 * zil_lwb_commit(). This way, the lwb is guaranteed to
2409 * be "full" when it is issued to disk, and we'll make
2410 * use of the lwb's size the best we can.
2412 * 2. If there isn't sufficient ZIL activity occurring on
2413 * the system, such that this lwb's zio isn't issued via
2414 * zil_lwb_commit(), zil_commit_waiter() will issue the
2415 * lwb's zio. If this occurs, the lwb is not guaranteed
2416 * to be "full" by the time its zio is issued, and means
2417 * the size of the lwb was "too large" given the amount
2418 * of ZIL activity occurring on the system at that time.
2420 * We do this for a couple of reasons:
2422 * 1. To try and reduce the number of IOPs needed to
2423 * write the same number of itxs. If an lwb has space
2424 * available in its buffer for more itxs, and more itxs
2425 * will be committed relatively soon (relative to the
2426 * latency of performing a write), then it's beneficial
2427 * to wait for these "next" itxs. This way, more itxs
2428 * can be committed to stable storage with fewer writes.
2430 * 2. To try and use the largest lwb block size that the
2431 * incoming rate of itxs can support. Again, this is to
2432 * try and pack as many itxs into as few lwbs as
2433 * possible, without significantly impacting the latency
2434 * of each individual itx.
2440 * This function is responsible for ensuring the passed in commit waiter
2441 * (and associated commit itx) is committed to an lwb. If the waiter is
2442 * not already committed to an lwb, all itxs in the zilog's queue of
2443 * itxs will be processed. The assumption is the passed in waiter's
2444 * commit itx will found in the queue just like the other non-commit
2445 * itxs, such that when the entire queue is processed, the waiter will
2446 * have been committed to an lwb.
2448 * The lwb associated with the passed in waiter is not guaranteed to
2449 * have been issued by the time this function completes. If the lwb is
2450 * not issued, we rely on future calls to zil_commit_writer() to issue
2451 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2454 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2456 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2457 ASSERT(spa_writeable(zilog->zl_spa));
2459 mutex_enter(&zilog->zl_issuer_lock);
2461 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2463 * It's possible that, while we were waiting to acquire
2464 * the "zl_issuer_lock", another thread committed this
2465 * waiter to an lwb. If that occurs, we bail out early,
2466 * without processing any of the zilog's queue of itxs.
2468 * On certain workloads and system configurations, the
2469 * "zl_issuer_lock" can become highly contended. In an
2470 * attempt to reduce this contention, we immediately drop
2471 * the lock if the waiter has already been processed.
2473 * We've measured this optimization to reduce CPU spent
2474 * contending on this lock by up to 5%, using a system
2475 * with 32 CPUs, low latency storage (~50 usec writes),
2476 * and 1024 threads performing sync writes.
2481 ZIL_STAT_BUMP(zil_commit_writer_count);
2483 zil_get_commit_list(zilog);
2484 zil_prune_commit_list(zilog);
2485 zil_process_commit_list(zilog);
2488 mutex_exit(&zilog->zl_issuer_lock);
2492 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2494 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2495 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2496 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2498 lwb_t *lwb = zcw->zcw_lwb;
2499 ASSERT3P(lwb, !=, NULL);
2500 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2503 * If the lwb has already been issued by another thread, we can
2504 * immediately return since there's no work to be done (the
2505 * point of this function is to issue the lwb). Additionally, we
2506 * do this prior to acquiring the zl_issuer_lock, to avoid
2507 * acquiring it when it's not necessary to do so.
2509 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2510 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2511 lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2515 * In order to call zil_lwb_write_issue() we must hold the
2516 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2517 * since we're already holding the commit waiter's "zcw_lock",
2518 * and those two locks are acquired in the opposite order
2521 mutex_exit(&zcw->zcw_lock);
2522 mutex_enter(&zilog->zl_issuer_lock);
2523 mutex_enter(&zcw->zcw_lock);
2526 * Since we just dropped and re-acquired the commit waiter's
2527 * lock, we have to re-check to see if the waiter was marked
2528 * "done" during that process. If the waiter was marked "done",
2529 * the "lwb" pointer is no longer valid (it can be free'd after
2530 * the waiter is marked "done"), so without this check we could
2531 * wind up with a use-after-free error below.
2536 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2539 * We've already checked this above, but since we hadn't acquired
2540 * the zilog's zl_issuer_lock, we have to perform this check a
2541 * second time while holding the lock.
2543 * We don't need to hold the zl_lock since the lwb cannot transition
2544 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2545 * _can_ transition from ISSUED to DONE, but it's OK to race with
2546 * that transition since we treat the lwb the same, whether it's in
2547 * the ISSUED or DONE states.
2549 * The important thing, is we treat the lwb differently depending on
2550 * if it's ISSUED or OPENED, and block any other threads that might
2551 * attempt to issue this lwb. For that reason we hold the
2552 * zl_issuer_lock when checking the lwb_state; we must not call
2553 * zil_lwb_write_issue() if the lwb had already been issued.
2555 * See the comment above the lwb_state_t structure definition for
2556 * more details on the lwb states, and locking requirements.
2558 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2559 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2560 lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2563 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2566 * As described in the comments above zil_commit_waiter() and
2567 * zil_process_commit_list(), we need to issue this lwb's zio
2568 * since we've reached the commit waiter's timeout and it still
2569 * hasn't been issued.
2571 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2573 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2576 * Since the lwb's zio hadn't been issued by the time this thread
2577 * reached its timeout, we reset the zilog's "zl_cur_used" field
2578 * to influence the zil block size selection algorithm.
2580 * By having to issue the lwb's zio here, it means the size of the
2581 * lwb was too large, given the incoming throughput of itxs. By
2582 * setting "zl_cur_used" to zero, we communicate this fact to the
2583 * block size selection algorithm, so it can take this information
2584 * into account, and potentially select a smaller size for the
2585 * next lwb block that is allocated.
2587 zilog->zl_cur_used = 0;
2591 * When zil_lwb_write_issue() returns NULL, this
2592 * indicates zio_alloc_zil() failed to allocate the
2593 * "next" lwb on-disk. When this occurs, the ZIL write
2594 * pipeline must be stalled; see the comment within the
2595 * zil_commit_writer_stall() function for more details.
2597 * We must drop the commit waiter's lock prior to
2598 * calling zil_commit_writer_stall() or else we can wind
2599 * up with the following deadlock:
2601 * - This thread is waiting for the txg to sync while
2602 * holding the waiter's lock; txg_wait_synced() is
2603 * used within txg_commit_writer_stall().
2605 * - The txg can't sync because it is waiting for this
2606 * lwb's zio callback to call dmu_tx_commit().
2608 * - The lwb's zio callback can't call dmu_tx_commit()
2609 * because it's blocked trying to acquire the waiter's
2610 * lock, which occurs prior to calling dmu_tx_commit()
2612 mutex_exit(&zcw->zcw_lock);
2613 zil_commit_writer_stall(zilog);
2614 mutex_enter(&zcw->zcw_lock);
2618 mutex_exit(&zilog->zl_issuer_lock);
2619 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2623 * This function is responsible for performing the following two tasks:
2625 * 1. its primary responsibility is to block until the given "commit
2626 * waiter" is considered "done".
2628 * 2. its secondary responsibility is to issue the zio for the lwb that
2629 * the given "commit waiter" is waiting on, if this function has
2630 * waited "long enough" and the lwb is still in the "open" state.
2632 * Given a sufficient amount of itxs being generated and written using
2633 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2634 * function. If this does not occur, this secondary responsibility will
2635 * ensure the lwb is issued even if there is not other synchronous
2636 * activity on the system.
2638 * For more details, see zil_process_commit_list(); more specifically,
2639 * the comment at the bottom of that function.
2642 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2644 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2645 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2646 ASSERT(spa_writeable(zilog->zl_spa));
2648 mutex_enter(&zcw->zcw_lock);
2651 * The timeout is scaled based on the lwb latency to avoid
2652 * significantly impacting the latency of each individual itx.
2653 * For more details, see the comment at the bottom of the
2654 * zil_process_commit_list() function.
2656 int pct = MAX(zfs_commit_timeout_pct, 1);
2657 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2658 hrtime_t wakeup = gethrtime() + sleep;
2659 boolean_t timedout = B_FALSE;
2661 while (!zcw->zcw_done) {
2662 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2664 lwb_t *lwb = zcw->zcw_lwb;
2667 * Usually, the waiter will have a non-NULL lwb field here,
2668 * but it's possible for it to be NULL as a result of
2669 * zil_commit() racing with spa_sync().
2671 * When zil_clean() is called, it's possible for the itxg
2672 * list (which may be cleaned via a taskq) to contain
2673 * commit itxs. When this occurs, the commit waiters linked
2674 * off of these commit itxs will not be committed to an
2675 * lwb. Additionally, these commit waiters will not be
2676 * marked done until zil_commit_waiter_skip() is called via
2679 * Thus, it's possible for this commit waiter (i.e. the
2680 * "zcw" variable) to be found in this "in between" state;
2681 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2682 * been skipped, so it's "zcw_done" field is still B_FALSE.
2684 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2686 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2687 ASSERT3B(timedout, ==, B_FALSE);
2690 * If the lwb hasn't been issued yet, then we
2691 * need to wait with a timeout, in case this
2692 * function needs to issue the lwb after the
2693 * timeout is reached; responsibility (2) from
2694 * the comment above this function.
2696 int rc = cv_timedwait_hires(&zcw->zcw_cv,
2697 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2698 CALLOUT_FLAG_ABSOLUTE);
2700 if (rc != -1 || zcw->zcw_done)
2704 zil_commit_waiter_timeout(zilog, zcw);
2706 if (!zcw->zcw_done) {
2708 * If the commit waiter has already been
2709 * marked "done", it's possible for the
2710 * waiter's lwb structure to have already
2711 * been freed. Thus, we can only reliably
2712 * make these assertions if the waiter
2715 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2716 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2720 * If the lwb isn't open, then it must have already
2721 * been issued. In that case, there's no need to
2722 * use a timeout when waiting for the lwb to
2725 * Additionally, if the lwb is NULL, the waiter
2726 * will soon be signaled and marked done via
2727 * zil_clean() and zil_itxg_clean(), so no timeout
2732 lwb->lwb_state == LWB_STATE_ISSUED ||
2733 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2734 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
2735 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2739 mutex_exit(&zcw->zcw_lock);
2742 static zil_commit_waiter_t *
2743 zil_alloc_commit_waiter(void)
2745 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2747 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2748 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2749 list_link_init(&zcw->zcw_node);
2750 zcw->zcw_lwb = NULL;
2751 zcw->zcw_done = B_FALSE;
2752 zcw->zcw_zio_error = 0;
2758 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2760 ASSERT(!list_link_active(&zcw->zcw_node));
2761 ASSERT3P(zcw->zcw_lwb, ==, NULL);
2762 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2763 mutex_destroy(&zcw->zcw_lock);
2764 cv_destroy(&zcw->zcw_cv);
2765 kmem_cache_free(zil_zcw_cache, zcw);
2769 * This function is used to create a TX_COMMIT itx and assign it. This
2770 * way, it will be linked into the ZIL's list of synchronous itxs, and
2771 * then later committed to an lwb (or skipped) when
2772 * zil_process_commit_list() is called.
2775 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2777 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2778 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2780 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2781 itx->itx_sync = B_TRUE;
2782 itx->itx_private = zcw;
2784 zil_itx_assign(zilog, itx, tx);
2790 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2792 * When writing ZIL transactions to the on-disk representation of the
2793 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2794 * itxs can be committed to a single lwb. Once a lwb is written and
2795 * committed to stable storage (i.e. the lwb is written, and vdevs have
2796 * been flushed), each itx that was committed to that lwb is also
2797 * considered to be committed to stable storage.
2799 * When an itx is committed to an lwb, the log record (lr_t) contained
2800 * by the itx is copied into the lwb's zio buffer, and once this buffer
2801 * is written to disk, it becomes an on-disk ZIL block.
2803 * As itxs are generated, they're inserted into the ZIL's queue of
2804 * uncommitted itxs. The semantics of zil_commit() are such that it will
2805 * block until all itxs that were in the queue when it was called, are
2806 * committed to stable storage.
2808 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2809 * itxs, for all objects in the dataset, will be committed to stable
2810 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2811 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2812 * that correspond to the foid passed in, will be committed to stable
2813 * storage prior to zil_commit() returning.
2815 * Generally speaking, when zil_commit() is called, the consumer doesn't
2816 * actually care about _all_ of the uncommitted itxs. Instead, they're
2817 * simply trying to waiting for a specific itx to be committed to disk,
2818 * but the interface(s) for interacting with the ZIL don't allow such
2819 * fine-grained communication. A better interface would allow a consumer
2820 * to create and assign an itx, and then pass a reference to this itx to
2821 * zil_commit(); such that zil_commit() would return as soon as that
2822 * specific itx was committed to disk (instead of waiting for _all_
2823 * itxs to be committed).
2825 * When a thread calls zil_commit() a special "commit itx" will be
2826 * generated, along with a corresponding "waiter" for this commit itx.
2827 * zil_commit() will wait on this waiter's CV, such that when the waiter
2828 * is marked done, and signaled, zil_commit() will return.
2830 * This commit itx is inserted into the queue of uncommitted itxs. This
2831 * provides an easy mechanism for determining which itxs were in the
2832 * queue prior to zil_commit() having been called, and which itxs were
2833 * added after zil_commit() was called.
2835 * The commit it is special; it doesn't have any on-disk representation.
2836 * When a commit itx is "committed" to an lwb, the waiter associated
2837 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2838 * completes, each waiter on the lwb's list is marked done and signaled
2839 * -- allowing the thread waiting on the waiter to return from zil_commit().
2841 * It's important to point out a few critical factors that allow us
2842 * to make use of the commit itxs, commit waiters, per-lwb lists of
2843 * commit waiters, and zio completion callbacks like we're doing:
2845 * 1. The list of waiters for each lwb is traversed, and each commit
2846 * waiter is marked "done" and signaled, in the zio completion
2847 * callback of the lwb's zio[*].
2849 * * Actually, the waiters are signaled in the zio completion
2850 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2851 * that are sent to the vdevs upon completion of the lwb zio.
2853 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2854 * itxs, the order in which they are inserted is preserved[*]; as
2855 * itxs are added to the queue, they are added to the tail of
2856 * in-memory linked lists.
2858 * When committing the itxs to lwbs (to be written to disk), they
2859 * are committed in the same order in which the itxs were added to
2860 * the uncommitted queue's linked list(s); i.e. the linked list of
2861 * itxs to commit is traversed from head to tail, and each itx is
2862 * committed to an lwb in that order.
2866 * - the order of "sync" itxs is preserved w.r.t. other
2867 * "sync" itxs, regardless of the corresponding objects.
2868 * - the order of "async" itxs is preserved w.r.t. other
2869 * "async" itxs corresponding to the same object.
2870 * - the order of "async" itxs is *not* preserved w.r.t. other
2871 * "async" itxs corresponding to different objects.
2872 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2873 * versa) is *not* preserved, even for itxs that correspond
2874 * to the same object.
2876 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2877 * zil_get_commit_list(), and zil_process_commit_list().
2879 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2880 * lwb cannot be considered committed to stable storage, until its
2881 * "previous" lwb is also committed to stable storage. This fact,
2882 * coupled with the fact described above, means that itxs are
2883 * committed in (roughly) the order in which they were generated.
2884 * This is essential because itxs are dependent on prior itxs.
2885 * Thus, we *must not* deem an itx as being committed to stable
2886 * storage, until *all* prior itxs have also been committed to
2889 * To enforce this ordering of lwb zio's, while still leveraging as
2890 * much of the underlying storage performance as possible, we rely
2891 * on two fundamental concepts:
2893 * 1. The creation and issuance of lwb zio's is protected by
2894 * the zilog's "zl_issuer_lock", which ensures only a single
2895 * thread is creating and/or issuing lwb's at a time
2896 * 2. The "previous" lwb is a child of the "current" lwb
2897 * (leveraging the zio parent-child dependency graph)
2899 * By relying on this parent-child zio relationship, we can have
2900 * many lwb zio's concurrently issued to the underlying storage,
2901 * but the order in which they complete will be the same order in
2902 * which they were created.
2905 zil_commit(zilog_t *zilog, uint64_t foid)
2908 * We should never attempt to call zil_commit on a snapshot for
2909 * a couple of reasons:
2911 * 1. A snapshot may never be modified, thus it cannot have any
2912 * in-flight itxs that would have modified the dataset.
2914 * 2. By design, when zil_commit() is called, a commit itx will
2915 * be assigned to this zilog; as a result, the zilog will be
2916 * dirtied. We must not dirty the zilog of a snapshot; there's
2917 * checks in the code that enforce this invariant, and will
2918 * cause a panic if it's not upheld.
2920 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2922 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2925 if (!spa_writeable(zilog->zl_spa)) {
2927 * If the SPA is not writable, there should never be any
2928 * pending itxs waiting to be committed to disk. If that
2929 * weren't true, we'd skip writing those itxs out, and
2930 * would break the semantics of zil_commit(); thus, we're
2931 * verifying that truth before we return to the caller.
2933 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2934 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2935 for (int i = 0; i < TXG_SIZE; i++)
2936 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2941 * If the ZIL is suspended, we don't want to dirty it by calling
2942 * zil_commit_itx_assign() below, nor can we write out
2943 * lwbs like would be done in zil_commit_write(). Thus, we
2944 * simply rely on txg_wait_synced() to maintain the necessary
2945 * semantics, and avoid calling those functions altogether.
2947 if (zilog->zl_suspend > 0) {
2948 txg_wait_synced(zilog->zl_dmu_pool, 0);
2952 zil_commit_impl(zilog, foid);
2956 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2958 ZIL_STAT_BUMP(zil_commit_count);
2961 * Move the "async" itxs for the specified foid to the "sync"
2962 * queues, such that they will be later committed (or skipped)
2963 * to an lwb when zil_process_commit_list() is called.
2965 * Since these "async" itxs must be committed prior to this
2966 * call to zil_commit returning, we must perform this operation
2967 * before we call zil_commit_itx_assign().
2969 zil_async_to_sync(zilog, foid);
2972 * We allocate a new "waiter" structure which will initially be
2973 * linked to the commit itx using the itx's "itx_private" field.
2974 * Since the commit itx doesn't represent any on-disk state,
2975 * when it's committed to an lwb, rather than copying the its
2976 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2977 * added to the lwb's list of waiters. Then, when the lwb is
2978 * committed to stable storage, each waiter in the lwb's list of
2979 * waiters will be marked "done", and signalled.
2981 * We must create the waiter and assign the commit itx prior to
2982 * calling zil_commit_writer(), or else our specific commit itx
2983 * is not guaranteed to be committed to an lwb prior to calling
2984 * zil_commit_waiter().
2986 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2987 zil_commit_itx_assign(zilog, zcw);
2989 zil_commit_writer(zilog, zcw);
2990 zil_commit_waiter(zilog, zcw);
2992 if (zcw->zcw_zio_error != 0) {
2994 * If there was an error writing out the ZIL blocks that
2995 * this thread is waiting on, then we fallback to
2996 * relying on spa_sync() to write out the data this
2997 * thread is waiting on. Obviously this has performance
2998 * implications, but the expectation is for this to be
2999 * an exceptional case, and shouldn't occur often.
3001 DTRACE_PROBE2(zil__commit__io__error,
3002 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3003 txg_wait_synced(zilog->zl_dmu_pool, 0);
3006 zil_free_commit_waiter(zcw);
3010 * Called in syncing context to free committed log blocks and update log header.
3013 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3015 zil_header_t *zh = zil_header_in_syncing_context(zilog);
3016 uint64_t txg = dmu_tx_get_txg(tx);
3017 spa_t *spa = zilog->zl_spa;
3018 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3022 * We don't zero out zl_destroy_txg, so make sure we don't try
3023 * to destroy it twice.
3025 if (spa_sync_pass(spa) != 1)
3028 mutex_enter(&zilog->zl_lock);
3030 ASSERT(zilog->zl_stop_sync == 0);
3032 if (*replayed_seq != 0) {
3033 ASSERT(zh->zh_replay_seq < *replayed_seq);
3034 zh->zh_replay_seq = *replayed_seq;
3038 if (zilog->zl_destroy_txg == txg) {
3039 blkptr_t blk = zh->zh_log;
3041 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
3043 bzero(zh, sizeof (zil_header_t));
3044 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
3046 if (zilog->zl_keep_first) {
3048 * If this block was part of log chain that couldn't
3049 * be claimed because a device was missing during
3050 * zil_claim(), but that device later returns,
3051 * then this block could erroneously appear valid.
3052 * To guard against this, assign a new GUID to the new
3053 * log chain so it doesn't matter what blk points to.
3055 zil_init_log_chain(zilog, &blk);
3060 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3061 zh->zh_log = lwb->lwb_blk;
3062 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
3064 list_remove(&zilog->zl_lwb_list, lwb);
3065 zio_free(spa, txg, &lwb->lwb_blk);
3066 zil_free_lwb(zilog, lwb);
3069 * If we don't have anything left in the lwb list then
3070 * we've had an allocation failure and we need to zero
3071 * out the zil_header blkptr so that we don't end
3072 * up freeing the same block twice.
3074 if (list_head(&zilog->zl_lwb_list) == NULL)
3075 BP_ZERO(&zh->zh_log);
3079 * Remove fastwrite on any blocks that have been pre-allocated for
3080 * the next commit. This prevents fastwrite counter pollution by
3081 * unused, long-lived LWBs.
3083 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
3084 if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) {
3085 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3086 lwb->lwb_fastwrite = 0;
3090 mutex_exit(&zilog->zl_lock);
3095 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3098 list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3099 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3100 offsetof(zil_commit_waiter_t, zcw_node));
3101 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3102 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3103 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3109 zil_lwb_dest(void *vbuf, void *unused)
3112 mutex_destroy(&lwb->lwb_vdev_lock);
3113 avl_destroy(&lwb->lwb_vdev_tree);
3114 list_destroy(&lwb->lwb_waiters);
3115 list_destroy(&lwb->lwb_itxs);
3121 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3122 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3124 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3125 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3127 zil_ksp = kstat_create("zfs", 0, "zil", "misc",
3128 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3129 KSTAT_FLAG_VIRTUAL);
3131 if (zil_ksp != NULL) {
3132 zil_ksp->ks_data = &zil_stats;
3133 kstat_install(zil_ksp);
3140 kmem_cache_destroy(zil_zcw_cache);
3141 kmem_cache_destroy(zil_lwb_cache);
3143 if (zil_ksp != NULL) {
3144 kstat_delete(zil_ksp);
3150 zil_set_sync(zilog_t *zilog, uint64_t sync)
3152 zilog->zl_sync = sync;
3156 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3158 zilog->zl_logbias = logbias;
3162 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3166 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3168 zilog->zl_header = zh_phys;
3170 zilog->zl_spa = dmu_objset_spa(os);
3171 zilog->zl_dmu_pool = dmu_objset_pool(os);
3172 zilog->zl_destroy_txg = TXG_INITIAL - 1;
3173 zilog->zl_logbias = dmu_objset_logbias(os);
3174 zilog->zl_sync = dmu_objset_syncprop(os);
3175 zilog->zl_dirty_max_txg = 0;
3176 zilog->zl_last_lwb_opened = NULL;
3177 zilog->zl_last_lwb_latency = 0;
3178 zilog->zl_max_block_size = zil_maxblocksize;
3180 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3181 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3183 for (int i = 0; i < TXG_SIZE; i++) {
3184 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3185 MUTEX_DEFAULT, NULL);
3188 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3189 offsetof(lwb_t, lwb_node));
3191 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3192 offsetof(itx_t, itx_node));
3194 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3200 zil_free(zilog_t *zilog)
3204 zilog->zl_stop_sync = 1;
3206 ASSERT0(zilog->zl_suspend);
3207 ASSERT0(zilog->zl_suspending);
3209 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3210 list_destroy(&zilog->zl_lwb_list);
3212 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3213 list_destroy(&zilog->zl_itx_commit_list);
3215 for (i = 0; i < TXG_SIZE; i++) {
3217 * It's possible for an itx to be generated that doesn't dirty
3218 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3219 * callback to remove the entry. We remove those here.
3221 * Also free up the ziltest itxs.
3223 if (zilog->zl_itxg[i].itxg_itxs)
3224 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3225 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3228 mutex_destroy(&zilog->zl_issuer_lock);
3229 mutex_destroy(&zilog->zl_lock);
3231 cv_destroy(&zilog->zl_cv_suspend);
3233 kmem_free(zilog, sizeof (zilog_t));
3237 * Open an intent log.
3240 zil_open(objset_t *os, zil_get_data_t *get_data)
3242 zilog_t *zilog = dmu_objset_zil(os);
3244 ASSERT3P(zilog->zl_get_data, ==, NULL);
3245 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3246 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3248 zilog->zl_get_data = get_data;
3254 * Close an intent log.
3257 zil_close(zilog_t *zilog)
3262 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3263 zil_commit(zilog, 0);
3265 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
3266 ASSERT0(zilog->zl_dirty_max_txg);
3267 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3270 mutex_enter(&zilog->zl_lock);
3271 lwb = list_tail(&zilog->zl_lwb_list);
3273 txg = zilog->zl_dirty_max_txg;
3275 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3276 mutex_exit(&zilog->zl_lock);
3279 * We need to use txg_wait_synced() to wait long enough for the
3280 * ZIL to be clean, and to wait for all pending lwbs to be
3284 txg_wait_synced(zilog->zl_dmu_pool, txg);
3286 if (zilog_is_dirty(zilog))
3287 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog, txg);
3288 if (txg < spa_freeze_txg(zilog->zl_spa))
3289 VERIFY(!zilog_is_dirty(zilog));
3291 zilog->zl_get_data = NULL;
3294 * We should have only one lwb left on the list; remove it now.
3296 mutex_enter(&zilog->zl_lock);
3297 lwb = list_head(&zilog->zl_lwb_list);
3299 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3300 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3302 if (lwb->lwb_fastwrite)
3303 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3305 list_remove(&zilog->zl_lwb_list, lwb);
3306 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3307 zil_free_lwb(zilog, lwb);
3309 mutex_exit(&zilog->zl_lock);
3312 static char *suspend_tag = "zil suspending";
3315 * Suspend an intent log. While in suspended mode, we still honor
3316 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3317 * On old version pools, we suspend the log briefly when taking a
3318 * snapshot so that it will have an empty intent log.
3320 * Long holds are not really intended to be used the way we do here --
3321 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3322 * could fail. Therefore we take pains to only put a long hold if it is
3323 * actually necessary. Fortunately, it will only be necessary if the
3324 * objset is currently mounted (or the ZVOL equivalent). In that case it
3325 * will already have a long hold, so we are not really making things any worse.
3327 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3328 * zvol_state_t), and use their mechanism to prevent their hold from being
3329 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3332 * if cookiep == NULL, this does both the suspend & resume.
3333 * Otherwise, it returns with the dataset "long held", and the cookie
3334 * should be passed into zil_resume().
3337 zil_suspend(const char *osname, void **cookiep)
3341 const zil_header_t *zh;
3344 error = dmu_objset_hold(osname, suspend_tag, &os);
3347 zilog = dmu_objset_zil(os);
3349 mutex_enter(&zilog->zl_lock);
3350 zh = zilog->zl_header;
3352 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
3353 mutex_exit(&zilog->zl_lock);
3354 dmu_objset_rele(os, suspend_tag);
3355 return (SET_ERROR(EBUSY));
3359 * Don't put a long hold in the cases where we can avoid it. This
3360 * is when there is no cookie so we are doing a suspend & resume
3361 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3362 * for the suspend because it's already suspended, or there's no ZIL.
3364 if (cookiep == NULL && !zilog->zl_suspending &&
3365 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3366 mutex_exit(&zilog->zl_lock);
3367 dmu_objset_rele(os, suspend_tag);
3371 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3372 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3374 zilog->zl_suspend++;
3376 if (zilog->zl_suspend > 1) {
3378 * Someone else is already suspending it.
3379 * Just wait for them to finish.
3382 while (zilog->zl_suspending)
3383 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3384 mutex_exit(&zilog->zl_lock);
3386 if (cookiep == NULL)
3394 * If there is no pointer to an on-disk block, this ZIL must not
3395 * be active (e.g. filesystem not mounted), so there's nothing
3398 if (BP_IS_HOLE(&zh->zh_log)) {
3399 ASSERT(cookiep != NULL); /* fast path already handled */
3402 mutex_exit(&zilog->zl_lock);
3407 * The ZIL has work to do. Ensure that the associated encryption
3408 * key will remain mapped while we are committing the log by
3409 * grabbing a reference to it. If the key isn't loaded we have no
3410 * choice but to return an error until the wrapping key is loaded.
3412 if (os->os_encrypted &&
3413 dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3414 zilog->zl_suspend--;
3415 mutex_exit(&zilog->zl_lock);
3416 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3417 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3418 return (SET_ERROR(EACCES));
3421 zilog->zl_suspending = B_TRUE;
3422 mutex_exit(&zilog->zl_lock);
3425 * We need to use zil_commit_impl to ensure we wait for all
3426 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3427 * to disk before proceeding. If we used zil_commit instead, it
3428 * would just call txg_wait_synced(), because zl_suspend is set.
3429 * txg_wait_synced() doesn't wait for these lwb's to be
3430 * LWB_STATE_FLUSH_DONE before returning.
3432 zil_commit_impl(zilog, 0);
3435 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3436 * use txg_wait_synced() to ensure the data from the zilog has
3437 * migrated to the main pool before calling zil_destroy().
3439 txg_wait_synced(zilog->zl_dmu_pool, 0);
3441 zil_destroy(zilog, B_FALSE);
3443 mutex_enter(&zilog->zl_lock);
3444 zilog->zl_suspending = B_FALSE;
3445 cv_broadcast(&zilog->zl_cv_suspend);
3446 mutex_exit(&zilog->zl_lock);
3448 if (os->os_encrypted)
3449 dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
3451 if (cookiep == NULL)
3459 zil_resume(void *cookie)
3461 objset_t *os = cookie;
3462 zilog_t *zilog = dmu_objset_zil(os);
3464 mutex_enter(&zilog->zl_lock);
3465 ASSERT(zilog->zl_suspend != 0);
3466 zilog->zl_suspend--;
3467 mutex_exit(&zilog->zl_lock);
3468 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3469 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3472 typedef struct zil_replay_arg {
3473 zil_replay_func_t **zr_replay;
3475 boolean_t zr_byteswap;
3480 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
3482 char name[ZFS_MAX_DATASET_NAME_LEN];
3484 zilog->zl_replaying_seq--; /* didn't actually replay this one */
3486 dmu_objset_name(zilog->zl_os, name);
3488 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3489 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3490 (u_longlong_t)lr->lrc_seq,
3491 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3492 (lr->lrc_txtype & TX_CI) ? "CI" : "");
3498 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
3501 zil_replay_arg_t *zr = zra;
3502 const zil_header_t *zh = zilog->zl_header;
3503 uint64_t reclen = lr->lrc_reclen;
3504 uint64_t txtype = lr->lrc_txtype;
3507 zilog->zl_replaying_seq = lr->lrc_seq;
3509 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
3512 if (lr->lrc_txg < claim_txg) /* already committed */
3515 /* Strip case-insensitive bit, still present in log record */
3518 if (txtype == 0 || txtype >= TX_MAX_TYPE)
3519 return (zil_replay_error(zilog, lr, EINVAL));
3522 * If this record type can be logged out of order, the object
3523 * (lr_foid) may no longer exist. That's legitimate, not an error.
3525 if (TX_OOO(txtype)) {
3526 error = dmu_object_info(zilog->zl_os,
3527 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3528 if (error == ENOENT || error == EEXIST)
3533 * Make a copy of the data so we can revise and extend it.
3535 bcopy(lr, zr->zr_lr, reclen);
3538 * If this is a TX_WRITE with a blkptr, suck in the data.
3540 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3541 error = zil_read_log_data(zilog, (lr_write_t *)lr,
3542 zr->zr_lr + reclen);
3544 return (zil_replay_error(zilog, lr, error));
3548 * The log block containing this lr may have been byteswapped
3549 * so that we can easily examine common fields like lrc_txtype.
3550 * However, the log is a mix of different record types, and only the
3551 * replay vectors know how to byteswap their records. Therefore, if
3552 * the lr was byteswapped, undo it before invoking the replay vector.
3554 if (zr->zr_byteswap)
3555 byteswap_uint64_array(zr->zr_lr, reclen);
3558 * We must now do two things atomically: replay this log record,
3559 * and update the log header sequence number to reflect the fact that
3560 * we did so. At the end of each replay function the sequence number
3561 * is updated if we are in replay mode.
3563 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3566 * The DMU's dnode layer doesn't see removes until the txg
3567 * commits, so a subsequent claim can spuriously fail with
3568 * EEXIST. So if we receive any error we try syncing out
3569 * any removes then retry the transaction. Note that we
3570 * specify B_FALSE for byteswap now, so we don't do it twice.
3572 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3573 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3575 return (zil_replay_error(zilog, lr, error));
3582 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
3584 zilog->zl_replay_blks++;
3590 * If this dataset has a non-empty intent log, replay it and destroy it.
3593 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3595 zilog_t *zilog = dmu_objset_zil(os);
3596 const zil_header_t *zh = zilog->zl_header;
3597 zil_replay_arg_t zr;
3599 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3600 zil_destroy(zilog, B_TRUE);
3604 zr.zr_replay = replay_func;
3606 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3607 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3610 * Wait for in-progress removes to sync before starting replay.
3612 txg_wait_synced(zilog->zl_dmu_pool, 0);
3614 zilog->zl_replay = B_TRUE;
3615 zilog->zl_replay_time = ddi_get_lbolt();
3616 ASSERT(zilog->zl_replay_blks == 0);
3617 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3618 zh->zh_claim_txg, B_TRUE);
3619 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3621 zil_destroy(zilog, B_FALSE);
3622 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3623 zilog->zl_replay = B_FALSE;
3627 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3629 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3632 if (zilog->zl_replay) {
3633 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3634 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3635 zilog->zl_replaying_seq;
3644 zil_reset(const char *osname, void *arg)
3648 error = zil_suspend(osname, NULL);
3649 /* EACCES means crypto key not loaded */
3650 if ((error == EACCES) || (error == EBUSY))
3651 return (SET_ERROR(error));
3653 return (SET_ERROR(EEXIST));
3657 EXPORT_SYMBOL(zil_alloc);
3658 EXPORT_SYMBOL(zil_free);
3659 EXPORT_SYMBOL(zil_open);
3660 EXPORT_SYMBOL(zil_close);
3661 EXPORT_SYMBOL(zil_replay);
3662 EXPORT_SYMBOL(zil_replaying);
3663 EXPORT_SYMBOL(zil_destroy);
3664 EXPORT_SYMBOL(zil_destroy_sync);
3665 EXPORT_SYMBOL(zil_itx_create);
3666 EXPORT_SYMBOL(zil_itx_destroy);
3667 EXPORT_SYMBOL(zil_itx_assign);
3668 EXPORT_SYMBOL(zil_commit);
3669 EXPORT_SYMBOL(zil_claim);
3670 EXPORT_SYMBOL(zil_check_log_chain);
3671 EXPORT_SYMBOL(zil_sync);
3672 EXPORT_SYMBOL(zil_clean);
3673 EXPORT_SYMBOL(zil_suspend);
3674 EXPORT_SYMBOL(zil_resume);
3675 EXPORT_SYMBOL(zil_lwb_add_block);
3676 EXPORT_SYMBOL(zil_bp_tree_add);
3677 EXPORT_SYMBOL(zil_set_sync);
3678 EXPORT_SYMBOL(zil_set_logbias);
3681 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW,
3682 "ZIL block open timeout percentage");
3684 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
3685 "Disable intent logging replay");
3687 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
3688 "Disable ZIL cache flushes");
3690 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW,
3691 "Limit in bytes slog sync writes per commit");
3693 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW,
3694 "Limit in bytes of ZIL log block size");