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, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
35 #include <sys/resource.h>
37 #include <sys/zil_impl.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/dmu_tx.h>
41 #include <sys/dsl_pool.h>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * Disable intent logging replay. This global ZIL switch affects all pools.
72 int zil_replay_disable = 0;
73 SYSCTL_DECL(_vfs_zfs);
74 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
75 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
76 &zil_replay_disable, 0, "Disable intent logging replay");
79 * Tunable parameter for debugging or performance analysis. Setting
80 * zfs_nocacheflush will cause corruption on power loss if a volatile
81 * out-of-order write cache is enabled.
83 boolean_t zfs_nocacheflush = B_FALSE;
84 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
85 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
86 &zfs_nocacheflush, 0, "Disable cache flush");
87 boolean_t zfs_trim_enabled = B_TRUE;
88 SYSCTL_DECL(_vfs_zfs_trim);
89 TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled);
90 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
94 * Limit SLOG write size per commit executed with synchronous priority.
95 * Any writes above that executed with lower (asynchronous) priority to
96 * limit potential SLOG device abuse by single active ZIL writer.
98 uint64_t zil_slog_limit = 768 * 1024;
99 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_limit, CTLFLAG_RWTUN,
100 &zil_slog_limit, 0, "Maximal SLOG commit size with sync priority");
102 static kmem_cache_t *zil_lwb_cache;
104 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
105 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
109 * ziltest is by and large an ugly hack, but very useful in
110 * checking replay without tedious work.
111 * When running ziltest we want to keep all itx's and so maintain
112 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
113 * We subtract TXG_CONCURRENT_STATES to allow for common code.
115 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
118 zil_bp_compare(const void *x1, const void *x2)
120 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
121 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
123 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
125 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
128 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
130 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
137 zil_bp_tree_init(zilog_t *zilog)
139 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
140 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
144 zil_bp_tree_fini(zilog_t *zilog)
146 avl_tree_t *t = &zilog->zl_bp_tree;
150 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
151 kmem_free(zn, sizeof (zil_bp_node_t));
157 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
159 avl_tree_t *t = &zilog->zl_bp_tree;
164 if (BP_IS_EMBEDDED(bp))
167 dva = BP_IDENTITY(bp);
169 if (avl_find(t, dva, &where) != NULL)
170 return (SET_ERROR(EEXIST));
172 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
174 avl_insert(t, zn, where);
179 static zil_header_t *
180 zil_header_in_syncing_context(zilog_t *zilog)
182 return ((zil_header_t *)zilog->zl_header);
186 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
188 zio_cksum_t *zc = &bp->blk_cksum;
190 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
191 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
192 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
193 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
197 * Read a log block and make sure it's valid.
200 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
203 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
204 arc_flags_t aflags = ARC_FLAG_WAIT;
205 arc_buf_t *abuf = NULL;
209 if (zilog->zl_header->zh_claim_txg == 0)
210 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
212 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
213 zio_flags |= ZIO_FLAG_SPECULATIVE;
215 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
216 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
218 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
219 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
222 zio_cksum_t cksum = bp->blk_cksum;
225 * Validate the checksummed log block.
227 * Sequence numbers should be... sequential. The checksum
228 * verifier for the next block should be bp's checksum plus 1.
230 * Also check the log chain linkage and size used.
232 cksum.zc_word[ZIL_ZC_SEQ]++;
234 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
235 zil_chain_t *zilc = abuf->b_data;
236 char *lr = (char *)(zilc + 1);
237 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
239 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
240 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
241 error = SET_ERROR(ECKSUM);
243 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
245 *end = (char *)dst + len;
246 *nbp = zilc->zc_next_blk;
249 char *lr = abuf->b_data;
250 uint64_t size = BP_GET_LSIZE(bp);
251 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
253 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
254 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
255 (zilc->zc_nused > (size - sizeof (*zilc)))) {
256 error = SET_ERROR(ECKSUM);
258 ASSERT3U(zilc->zc_nused, <=,
259 SPA_OLD_MAXBLOCKSIZE);
260 bcopy(lr, dst, zilc->zc_nused);
261 *end = (char *)dst + zilc->zc_nused;
262 *nbp = zilc->zc_next_blk;
266 arc_buf_destroy(abuf, &abuf);
273 * Read a TX_WRITE log data block.
276 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
278 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
279 const blkptr_t *bp = &lr->lr_blkptr;
280 arc_flags_t aflags = ARC_FLAG_WAIT;
281 arc_buf_t *abuf = NULL;
285 if (BP_IS_HOLE(bp)) {
287 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
291 if (zilog->zl_header->zh_claim_txg == 0)
292 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
294 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
295 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
297 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
298 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
302 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
303 arc_buf_destroy(abuf, &abuf);
310 * Parse the intent log, and call parse_func for each valid record within.
313 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
314 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
316 const zil_header_t *zh = zilog->zl_header;
317 boolean_t claimed = !!zh->zh_claim_txg;
318 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
319 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
320 uint64_t max_blk_seq = 0;
321 uint64_t max_lr_seq = 0;
322 uint64_t blk_count = 0;
323 uint64_t lr_count = 0;
324 blkptr_t blk, next_blk;
329 * Old logs didn't record the maximum zh_claim_lr_seq.
331 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
332 claim_lr_seq = UINT64_MAX;
335 * Starting at the block pointed to by zh_log we read the log chain.
336 * For each block in the chain we strongly check that block to
337 * ensure its validity. We stop when an invalid block is found.
338 * For each block pointer in the chain we call parse_blk_func().
339 * For each record in each valid block we call parse_lr_func().
340 * If the log has been claimed, stop if we encounter a sequence
341 * number greater than the highest claimed sequence number.
343 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
344 zil_bp_tree_init(zilog);
346 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
347 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
351 if (blk_seq > claim_blk_seq)
353 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
355 ASSERT3U(max_blk_seq, <, blk_seq);
356 max_blk_seq = blk_seq;
359 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
362 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
366 for (lrp = lrbuf; lrp < end; lrp += reclen) {
367 lr_t *lr = (lr_t *)lrp;
368 reclen = lr->lrc_reclen;
369 ASSERT3U(reclen, >=, sizeof (lr_t));
370 if (lr->lrc_seq > claim_lr_seq)
372 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
374 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
375 max_lr_seq = lr->lrc_seq;
380 zilog->zl_parse_error = error;
381 zilog->zl_parse_blk_seq = max_blk_seq;
382 zilog->zl_parse_lr_seq = max_lr_seq;
383 zilog->zl_parse_blk_count = blk_count;
384 zilog->zl_parse_lr_count = lr_count;
386 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
387 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
389 zil_bp_tree_fini(zilog);
390 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
396 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
399 * Claim log block if not already committed and not already claimed.
400 * If tx == NULL, just verify that the block is claimable.
402 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
403 zil_bp_tree_add(zilog, bp) != 0)
406 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
407 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
408 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
412 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
414 lr_write_t *lr = (lr_write_t *)lrc;
417 if (lrc->lrc_txtype != TX_WRITE)
421 * If the block is not readable, don't claim it. This can happen
422 * in normal operation when a log block is written to disk before
423 * some of the dmu_sync() blocks it points to. In this case, the
424 * transaction cannot have been committed to anyone (we would have
425 * waited for all writes to be stable first), so it is semantically
426 * correct to declare this the end of the log.
428 if (lr->lr_blkptr.blk_birth >= first_txg &&
429 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
431 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
436 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
438 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
444 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
446 lr_write_t *lr = (lr_write_t *)lrc;
447 blkptr_t *bp = &lr->lr_blkptr;
450 * If we previously claimed it, we need to free it.
452 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
453 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
455 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
461 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
465 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
466 lwb->lwb_zilog = zilog;
468 lwb->lwb_slog = slog;
469 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
470 lwb->lwb_max_txg = txg;
473 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
474 lwb->lwb_nused = sizeof (zil_chain_t);
475 lwb->lwb_sz = BP_GET_LSIZE(bp);
478 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
481 mutex_enter(&zilog->zl_lock);
482 list_insert_tail(&zilog->zl_lwb_list, lwb);
483 mutex_exit(&zilog->zl_lock);
489 * Called when we create in-memory log transactions so that we know
490 * to cleanup the itxs at the end of spa_sync().
493 zilog_dirty(zilog_t *zilog, uint64_t txg)
495 dsl_pool_t *dp = zilog->zl_dmu_pool;
496 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
498 if (ds->ds_is_snapshot)
499 panic("dirtying snapshot!");
501 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
502 /* up the hold count until we can be written out */
503 dmu_buf_add_ref(ds->ds_dbuf, zilog);
508 * Determine if the zil is dirty in the specified txg. Callers wanting to
509 * ensure that the dirty state does not change must hold the itxg_lock for
510 * the specified txg. Holding the lock will ensure that the zil cannot be
511 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
515 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
517 dsl_pool_t *dp = zilog->zl_dmu_pool;
519 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
525 * Determine if the zil is dirty. The zil is considered dirty if it has
526 * any pending itx records that have not been cleaned by zil_clean().
529 zilog_is_dirty(zilog_t *zilog)
531 dsl_pool_t *dp = zilog->zl_dmu_pool;
533 for (int t = 0; t < TXG_SIZE; t++) {
534 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
541 * Create an on-disk intent log.
544 zil_create(zilog_t *zilog)
546 const zil_header_t *zh = zilog->zl_header;
552 boolean_t slog = FALSE;
555 * Wait for any previous destroy to complete.
557 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
559 ASSERT(zh->zh_claim_txg == 0);
560 ASSERT(zh->zh_replay_seq == 0);
565 * Allocate an initial log block if:
566 * - there isn't one already
567 * - the existing block is the wrong endianess
569 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
570 tx = dmu_tx_create(zilog->zl_os);
571 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
572 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
573 txg = dmu_tx_get_txg(tx);
575 if (!BP_IS_HOLE(&blk)) {
576 zio_free_zil(zilog->zl_spa, txg, &blk);
580 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
581 ZIL_MIN_BLKSZ, &slog);
584 zil_init_log_chain(zilog, &blk);
588 * Allocate a log write buffer (lwb) for the first log block.
591 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
594 * If we just allocated the first log block, commit our transaction
595 * and wait for zil_sync() to stuff the block poiner into zh_log.
596 * (zh is part of the MOS, so we cannot modify it in open context.)
600 txg_wait_synced(zilog->zl_dmu_pool, txg);
603 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
609 * In one tx, free all log blocks and clear the log header.
610 * If keep_first is set, then we're replaying a log with no content.
611 * We want to keep the first block, however, so that the first
612 * synchronous transaction doesn't require a txg_wait_synced()
613 * in zil_create(). We don't need to txg_wait_synced() here either
614 * when keep_first is set, because both zil_create() and zil_destroy()
615 * will wait for any in-progress destroys to complete.
618 zil_destroy(zilog_t *zilog, boolean_t keep_first)
620 const zil_header_t *zh = zilog->zl_header;
626 * Wait for any previous destroy to complete.
628 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
630 zilog->zl_old_header = *zh; /* debugging aid */
632 if (BP_IS_HOLE(&zh->zh_log))
635 tx = dmu_tx_create(zilog->zl_os);
636 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
637 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
638 txg = dmu_tx_get_txg(tx);
640 mutex_enter(&zilog->zl_lock);
642 ASSERT3U(zilog->zl_destroy_txg, <, txg);
643 zilog->zl_destroy_txg = txg;
644 zilog->zl_keep_first = keep_first;
646 if (!list_is_empty(&zilog->zl_lwb_list)) {
647 ASSERT(zh->zh_claim_txg == 0);
649 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
650 list_remove(&zilog->zl_lwb_list, lwb);
651 if (lwb->lwb_buf != NULL)
652 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
653 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
654 kmem_cache_free(zil_lwb_cache, lwb);
656 } else if (!keep_first) {
657 zil_destroy_sync(zilog, tx);
659 mutex_exit(&zilog->zl_lock);
665 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
667 ASSERT(list_is_empty(&zilog->zl_lwb_list));
668 (void) zil_parse(zilog, zil_free_log_block,
669 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
673 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
675 dmu_tx_t *tx = txarg;
676 uint64_t first_txg = dmu_tx_get_txg(tx);
682 error = dmu_objset_own_obj(dp, ds->ds_object,
683 DMU_OST_ANY, B_FALSE, FTAG, &os);
686 * EBUSY indicates that the objset is inconsistent, in which
687 * case it can not have a ZIL.
689 if (error != EBUSY) {
690 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
691 (unsigned long long)ds->ds_object, error);
696 zilog = dmu_objset_zil(os);
697 zh = zil_header_in_syncing_context(zilog);
699 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
700 if (!BP_IS_HOLE(&zh->zh_log))
701 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
702 BP_ZERO(&zh->zh_log);
703 dsl_dataset_dirty(dmu_objset_ds(os), tx);
704 dmu_objset_disown(os, FTAG);
709 * Claim all log blocks if we haven't already done so, and remember
710 * the highest claimed sequence number. This ensures that if we can
711 * read only part of the log now (e.g. due to a missing device),
712 * but we can read the entire log later, we will not try to replay
713 * or destroy beyond the last block we successfully claimed.
715 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
716 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
717 (void) zil_parse(zilog, zil_claim_log_block,
718 zil_claim_log_record, tx, first_txg);
719 zh->zh_claim_txg = first_txg;
720 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
721 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
722 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
723 zh->zh_flags |= ZIL_REPLAY_NEEDED;
724 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
725 dsl_dataset_dirty(dmu_objset_ds(os), tx);
728 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
729 dmu_objset_disown(os, FTAG);
734 * Check the log by walking the log chain.
735 * Checksum errors are ok as they indicate the end of the chain.
736 * Any other error (no device or read failure) returns an error.
740 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
749 error = dmu_objset_from_ds(ds, &os);
751 cmn_err(CE_WARN, "can't open objset %llu, error %d",
752 (unsigned long long)ds->ds_object, error);
756 zilog = dmu_objset_zil(os);
757 bp = (blkptr_t *)&zilog->zl_header->zh_log;
760 * Check the first block and determine if it's on a log device
761 * which may have been removed or faulted prior to loading this
762 * pool. If so, there's no point in checking the rest of the log
763 * as its content should have already been synced to the pool.
765 if (!BP_IS_HOLE(bp)) {
767 boolean_t valid = B_TRUE;
769 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
770 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
771 if (vd->vdev_islog && vdev_is_dead(vd))
772 valid = vdev_log_state_valid(vd);
773 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
780 * Because tx == NULL, zil_claim_log_block() will not actually claim
781 * any blocks, but just determine whether it is possible to do so.
782 * In addition to checking the log chain, zil_claim_log_block()
783 * will invoke zio_claim() with a done func of spa_claim_notify(),
784 * which will update spa_max_claim_txg. See spa_load() for details.
786 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
787 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
789 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
793 zil_vdev_compare(const void *x1, const void *x2)
795 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
796 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
807 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
809 avl_tree_t *t = &zilog->zl_vdev_tree;
811 zil_vdev_node_t *zv, zvsearch;
812 int ndvas = BP_GET_NDVAS(bp);
815 if (zfs_nocacheflush)
818 ASSERT(zilog->zl_writer);
821 * Even though we're zl_writer, we still need a lock because the
822 * zl_get_data() callbacks may have dmu_sync() done callbacks
823 * that will run concurrently.
825 mutex_enter(&zilog->zl_vdev_lock);
826 for (i = 0; i < ndvas; i++) {
827 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
828 if (avl_find(t, &zvsearch, &where) == NULL) {
829 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
830 zv->zv_vdev = zvsearch.zv_vdev;
831 avl_insert(t, zv, where);
834 mutex_exit(&zilog->zl_vdev_lock);
838 zil_flush_vdevs(zilog_t *zilog)
840 spa_t *spa = zilog->zl_spa;
841 avl_tree_t *t = &zilog->zl_vdev_tree;
846 ASSERT(zilog->zl_writer);
849 * We don't need zl_vdev_lock here because we're the zl_writer,
850 * and all zl_get_data() callbacks are done.
852 if (avl_numnodes(t) == 0)
855 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
857 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
858 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
859 if (vd != NULL && !vd->vdev_nowritecache) {
861 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
864 kmem_free(zv, sizeof (*zv));
868 * Wait for all the flushes to complete. Not all devices actually
869 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
872 (void) zio_wait(zio);
874 spa_config_exit(spa, SCL_STATE, FTAG);
878 * Function called when a log block write completes
881 zil_lwb_write_done(zio_t *zio)
883 lwb_t *lwb = zio->io_private;
884 zilog_t *zilog = lwb->lwb_zilog;
885 dmu_tx_t *tx = lwb->lwb_tx;
887 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
888 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
889 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
890 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
891 ASSERT(!BP_IS_GANG(zio->io_bp));
892 ASSERT(!BP_IS_HOLE(zio->io_bp));
893 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
896 * Ensure the lwb buffer pointer is cleared before releasing
897 * the txg. If we have had an allocation failure and
898 * the txg is waiting to sync then we want want zil_sync()
899 * to remove the lwb so that it's not picked up as the next new
900 * one in zil_commit_writer(). zil_sync() will only remove
901 * the lwb if lwb_buf is null.
903 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
904 mutex_enter(&zilog->zl_lock);
907 mutex_exit(&zilog->zl_lock);
910 * Now that we've written this log block, we have a stable pointer
911 * to the next block in the chain, so it's OK to let the txg in
912 * which we allocated the next block sync.
918 * Initialize the io for a log block.
921 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
926 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
927 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
928 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
930 if (zilog->zl_root_zio == NULL) {
931 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
934 if (lwb->lwb_zio == NULL) {
935 if (zilog->zl_cur_used <= zil_slog_limit || !lwb->lwb_slog)
936 prio = ZIO_PRIORITY_SYNC_WRITE;
938 prio = ZIO_PRIORITY_ASYNC_WRITE;
939 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
940 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
941 zil_lwb_write_done, lwb, prio,
942 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
947 * Define a limited set of intent log block sizes.
949 * These must be a multiple of 4KB. Note only the amount used (again
950 * aligned to 4KB) actually gets written. However, we can't always just
951 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
953 uint64_t zil_block_buckets[] = {
954 4096, /* non TX_WRITE */
955 8192+4096, /* data base */
956 32*1024 + 4096, /* NFS writes */
961 * Start a log block write and advance to the next log block.
962 * Calls are serialized.
965 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last)
969 spa_t *spa = zilog->zl_spa;
973 uint64_t zil_blksz, wsz;
977 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
978 zilc = (zil_chain_t *)lwb->lwb_buf;
979 bp = &zilc->zc_next_blk;
981 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
982 bp = &zilc->zc_next_blk;
985 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
988 * Allocate the next block and save its address in this block
989 * before writing it in order to establish the log chain.
990 * Note that if the allocation of nlwb synced before we wrote
991 * the block that points at it (lwb), we'd leak it if we crashed.
992 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
993 * We dirty the dataset to ensure that zil_sync() will be called
994 * to clean up in the event of allocation failure or I/O failure.
996 tx = dmu_tx_create(zilog->zl_os);
997 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
998 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
999 txg = dmu_tx_get_txg(tx);
1004 * Log blocks are pre-allocated. Here we select the size of the next
1005 * block, based on size used in the last block.
1006 * - first find the smallest bucket that will fit the block from a
1007 * limited set of block sizes. This is because it's faster to write
1008 * blocks allocated from the same metaslab as they are adjacent or
1010 * - next find the maximum from the new suggested size and an array of
1011 * previous sizes. This lessens a picket fence effect of wrongly
1012 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1015 * Note we only write what is used, but we can't just allocate
1016 * the maximum block size because we can exhaust the available
1019 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1020 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1022 zil_blksz = zil_block_buckets[i];
1023 if (zil_blksz == UINT64_MAX)
1024 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1025 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1026 for (i = 0; i < ZIL_PREV_BLKS; i++)
1027 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1028 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1031 /* pass the old blkptr in order to spread log blocks across devs */
1032 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1034 ASSERT3U(bp->blk_birth, ==, txg);
1035 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1036 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1039 * Allocate a new log write buffer (lwb).
1041 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1043 /* Record the block for later vdev flushing */
1044 zil_add_block(zilog, &lwb->lwb_blk);
1047 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1048 /* For Slim ZIL only write what is used. */
1049 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1050 ASSERT3U(wsz, <=, lwb->lwb_sz);
1051 zio_shrink(lwb->lwb_zio, wsz);
1058 zilc->zc_nused = lwb->lwb_nused;
1059 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1062 * clear unused data for security
1064 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1067 lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC;
1068 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1071 * If there was an allocation failure then nlwb will be null which
1072 * forces a txg_wait_synced().
1078 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1080 lr_t *lrcb, *lrc = &itx->itx_lr; /* common log record */
1081 lr_write_t *lrwb, *lrw = (lr_write_t *)lrc;
1083 uint64_t txg = lrc->lrc_txg;
1084 uint64_t reclen = lrc->lrc_reclen;
1086 uint64_t dnow, lwb_sp;
1091 ASSERT(lwb->lwb_buf != NULL);
1093 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1094 dlen = P2ROUNDUP_TYPED(
1095 lrw->lr_length, sizeof (uint64_t), uint64_t);
1097 zilog->zl_cur_used += (reclen + dlen);
1099 zil_lwb_write_init(zilog, lwb);
1103 * If this record won't fit in the current log block, start a new one.
1104 * For WR_NEED_COPY optimize layout for minimal number of chunks, but
1105 * try to keep wasted space withing reasonable range (12%).
1107 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1108 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1109 lwb_sp < ZIL_MAX_LOG_DATA / 8 && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1110 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1111 lwb = zil_lwb_write_start(zilog, lwb, B_FALSE);
1114 zil_lwb_write_init(zilog, lwb);
1115 ASSERT(LWB_EMPTY(lwb));
1116 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1117 ASSERT3U(reclen + MIN(dlen, sizeof(uint64_t)), <=, lwb_sp);
1120 dnow = MIN(dlen, lwb_sp - reclen);
1121 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1122 bcopy(lrc, lr_buf, reclen);
1123 lrcb = (lr_t *)lr_buf;
1124 lrwb = (lr_write_t *)lrcb;
1127 * If it's a write, fetch the data or get its blkptr as appropriate.
1129 if (lrc->lrc_txtype == TX_WRITE) {
1130 if (txg > spa_freeze_txg(zilog->zl_spa))
1131 txg_wait_synced(zilog->zl_dmu_pool, txg);
1132 if (itx->itx_wr_state != WR_COPIED) {
1136 if (itx->itx_wr_state == WR_NEED_COPY) {
1137 dbuf = lr_buf + reclen;
1138 lrcb->lrc_reclen += dnow;
1139 if (lrwb->lr_length > dnow)
1140 lrwb->lr_length = dnow;
1141 lrw->lr_offset += dnow;
1142 lrw->lr_length -= dnow;
1144 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1147 error = zilog->zl_get_data(
1148 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1150 txg_wait_synced(zilog->zl_dmu_pool, txg);
1154 ASSERT(error == ENOENT || error == EEXIST ||
1162 * We're actually making an entry, so update lrc_seq to be the
1163 * log record sequence number. Note that this is generally not
1164 * equal to the itx sequence number because not all transactions
1165 * are synchronous, and sometimes spa_sync() gets there first.
1167 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1168 lwb->lwb_nused += reclen + dnow;
1169 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1170 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1171 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1175 zilog->zl_cur_used += reclen;
1183 zil_itx_create(uint64_t txtype, size_t lrsize)
1187 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1189 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1190 itx->itx_lr.lrc_txtype = txtype;
1191 itx->itx_lr.lrc_reclen = lrsize;
1192 itx->itx_lr.lrc_seq = 0; /* defensive */
1193 itx->itx_sync = B_TRUE; /* default is synchronous */
1199 zil_itx_destroy(itx_t *itx)
1201 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1205 * Free up the sync and async itxs. The itxs_t has already been detached
1206 * so no locks are needed.
1209 zil_itxg_clean(itxs_t *itxs)
1215 itx_async_node_t *ian;
1217 list = &itxs->i_sync_list;
1218 while ((itx = list_head(list)) != NULL) {
1219 list_remove(list, itx);
1220 kmem_free(itx, offsetof(itx_t, itx_lr) +
1221 itx->itx_lr.lrc_reclen);
1225 t = &itxs->i_async_tree;
1226 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1227 list = &ian->ia_list;
1228 while ((itx = list_head(list)) != NULL) {
1229 list_remove(list, itx);
1230 kmem_free(itx, offsetof(itx_t, itx_lr) +
1231 itx->itx_lr.lrc_reclen);
1234 kmem_free(ian, sizeof (itx_async_node_t));
1238 kmem_free(itxs, sizeof (itxs_t));
1242 zil_aitx_compare(const void *x1, const void *x2)
1244 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1245 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1256 * Remove all async itx with the given oid.
1259 zil_remove_async(zilog_t *zilog, uint64_t oid)
1262 itx_async_node_t *ian;
1269 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1271 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1274 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1276 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1277 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1279 mutex_enter(&itxg->itxg_lock);
1280 if (itxg->itxg_txg != txg) {
1281 mutex_exit(&itxg->itxg_lock);
1286 * Locate the object node and append its list.
1288 t = &itxg->itxg_itxs->i_async_tree;
1289 ian = avl_find(t, &oid, &where);
1291 list_move_tail(&clean_list, &ian->ia_list);
1292 mutex_exit(&itxg->itxg_lock);
1294 while ((itx = list_head(&clean_list)) != NULL) {
1295 list_remove(&clean_list, itx);
1296 kmem_free(itx, offsetof(itx_t, itx_lr) +
1297 itx->itx_lr.lrc_reclen);
1299 list_destroy(&clean_list);
1303 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1307 itxs_t *itxs, *clean = NULL;
1310 * Object ids can be re-instantiated in the next txg so
1311 * remove any async transactions to avoid future leaks.
1312 * This can happen if a fsync occurs on the re-instantiated
1313 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1314 * the new file data and flushes a write record for the old object.
1316 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1317 zil_remove_async(zilog, itx->itx_oid);
1320 * Ensure the data of a renamed file is committed before the rename.
1322 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1323 zil_async_to_sync(zilog, itx->itx_oid);
1325 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1328 txg = dmu_tx_get_txg(tx);
1330 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1331 mutex_enter(&itxg->itxg_lock);
1332 itxs = itxg->itxg_itxs;
1333 if (itxg->itxg_txg != txg) {
1336 * The zil_clean callback hasn't got around to cleaning
1337 * this itxg. Save the itxs for release below.
1338 * This should be rare.
1340 clean = itxg->itxg_itxs;
1342 itxg->itxg_txg = txg;
1343 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1345 list_create(&itxs->i_sync_list, sizeof (itx_t),
1346 offsetof(itx_t, itx_node));
1347 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1348 sizeof (itx_async_node_t),
1349 offsetof(itx_async_node_t, ia_node));
1351 if (itx->itx_sync) {
1352 list_insert_tail(&itxs->i_sync_list, itx);
1354 avl_tree_t *t = &itxs->i_async_tree;
1355 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1356 itx_async_node_t *ian;
1359 ian = avl_find(t, &foid, &where);
1361 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1362 list_create(&ian->ia_list, sizeof (itx_t),
1363 offsetof(itx_t, itx_node));
1364 ian->ia_foid = foid;
1365 avl_insert(t, ian, where);
1367 list_insert_tail(&ian->ia_list, itx);
1370 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1371 zilog_dirty(zilog, txg);
1372 mutex_exit(&itxg->itxg_lock);
1374 /* Release the old itxs now we've dropped the lock */
1376 zil_itxg_clean(clean);
1380 * If there are any in-memory intent log transactions which have now been
1381 * synced then start up a taskq to free them. We should only do this after we
1382 * have written out the uberblocks (i.e. txg has been comitted) so that
1383 * don't inadvertently clean out in-memory log records that would be required
1387 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1389 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1392 mutex_enter(&itxg->itxg_lock);
1393 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1394 mutex_exit(&itxg->itxg_lock);
1397 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1398 ASSERT(itxg->itxg_txg != 0);
1399 ASSERT(zilog->zl_clean_taskq != NULL);
1400 clean_me = itxg->itxg_itxs;
1401 itxg->itxg_itxs = NULL;
1403 mutex_exit(&itxg->itxg_lock);
1405 * Preferably start a task queue to free up the old itxs but
1406 * if taskq_dispatch can't allocate resources to do that then
1407 * free it in-line. This should be rare. Note, using TQ_SLEEP
1408 * created a bad performance problem.
1410 if (taskq_dispatch(zilog->zl_clean_taskq,
1411 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1412 zil_itxg_clean(clean_me);
1416 * Get the list of itxs to commit into zl_itx_commit_list.
1419 zil_get_commit_list(zilog_t *zilog)
1422 list_t *commit_list = &zilog->zl_itx_commit_list;
1424 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1427 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1430 * This is inherently racy, since there is nothing to prevent
1431 * the last synced txg from changing. That's okay since we'll
1432 * only commit things in the future.
1434 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1435 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1437 mutex_enter(&itxg->itxg_lock);
1438 if (itxg->itxg_txg != txg) {
1439 mutex_exit(&itxg->itxg_lock);
1444 * If we're adding itx records to the zl_itx_commit_list,
1445 * then the zil better be dirty in this "txg". We can assert
1446 * that here since we're holding the itxg_lock which will
1447 * prevent spa_sync from cleaning it. Once we add the itxs
1448 * to the zl_itx_commit_list we must commit it to disk even
1449 * if it's unnecessary (i.e. the txg was synced).
1451 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1452 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1453 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1455 mutex_exit(&itxg->itxg_lock);
1460 * Move the async itxs for a specified object to commit into sync lists.
1463 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1466 itx_async_node_t *ian;
1470 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1473 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1476 * This is inherently racy, since there is nothing to prevent
1477 * the last synced txg from changing.
1479 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1480 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1482 mutex_enter(&itxg->itxg_lock);
1483 if (itxg->itxg_txg != txg) {
1484 mutex_exit(&itxg->itxg_lock);
1489 * If a foid is specified then find that node and append its
1490 * list. Otherwise walk the tree appending all the lists
1491 * to the sync list. We add to the end rather than the
1492 * beginning to ensure the create has happened.
1494 t = &itxg->itxg_itxs->i_async_tree;
1496 ian = avl_find(t, &foid, &where);
1498 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1502 void *cookie = NULL;
1504 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1505 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1507 list_destroy(&ian->ia_list);
1508 kmem_free(ian, sizeof (itx_async_node_t));
1511 mutex_exit(&itxg->itxg_lock);
1516 zil_commit_writer(zilog_t *zilog)
1521 spa_t *spa = zilog->zl_spa;
1524 ASSERT(zilog->zl_root_zio == NULL);
1526 mutex_exit(&zilog->zl_lock);
1528 zil_get_commit_list(zilog);
1531 * Return if there's nothing to commit before we dirty the fs by
1532 * calling zil_create().
1534 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1535 mutex_enter(&zilog->zl_lock);
1539 if (zilog->zl_suspend) {
1542 lwb = list_tail(&zilog->zl_lwb_list);
1544 lwb = zil_create(zilog);
1547 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1548 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1549 txg = itx->itx_lr.lrc_txg;
1550 ASSERT3U(txg, !=, 0);
1553 * This is inherently racy and may result in us writing
1554 * out a log block for a txg that was just synced. This is
1555 * ok since we'll end cleaning up that log block the next
1556 * time we call zil_sync().
1558 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1559 lwb = zil_lwb_commit(zilog, itx, lwb);
1560 list_remove(&zilog->zl_itx_commit_list, itx);
1561 kmem_free(itx, offsetof(itx_t, itx_lr)
1562 + itx->itx_lr.lrc_reclen);
1564 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1566 /* write the last block out */
1567 if (lwb != NULL && lwb->lwb_zio != NULL)
1568 lwb = zil_lwb_write_start(zilog, lwb, B_TRUE);
1570 zilog->zl_cur_used = 0;
1573 * Wait if necessary for the log blocks to be on stable storage.
1575 if (zilog->zl_root_zio) {
1576 error = zio_wait(zilog->zl_root_zio);
1577 zilog->zl_root_zio = NULL;
1578 zil_flush_vdevs(zilog);
1581 if (error || lwb == NULL)
1582 txg_wait_synced(zilog->zl_dmu_pool, 0);
1584 mutex_enter(&zilog->zl_lock);
1587 * Remember the highest committed log sequence number for ztest.
1588 * We only update this value when all the log writes succeeded,
1589 * because ztest wants to ASSERT that it got the whole log chain.
1591 if (error == 0 && lwb != NULL)
1592 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1596 * Commit zfs transactions to stable storage.
1597 * If foid is 0 push out all transactions, otherwise push only those
1598 * for that object or might reference that object.
1600 * itxs are committed in batches. In a heavily stressed zil there will be
1601 * a commit writer thread who is writing out a bunch of itxs to the log
1602 * for a set of committing threads (cthreads) in the same batch as the writer.
1603 * Those cthreads are all waiting on the same cv for that batch.
1605 * There will also be a different and growing batch of threads that are
1606 * waiting to commit (qthreads). When the committing batch completes
1607 * a transition occurs such that the cthreads exit and the qthreads become
1608 * cthreads. One of the new cthreads becomes the writer thread for the
1609 * batch. Any new threads arriving become new qthreads.
1611 * Only 2 condition variables are needed and there's no transition
1612 * between the two cvs needed. They just flip-flop between qthreads
1615 * Using this scheme we can efficiently wakeup up only those threads
1616 * that have been committed.
1619 zil_commit(zilog_t *zilog, uint64_t foid)
1623 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1626 /* move the async itxs for the foid to the sync queues */
1627 zil_async_to_sync(zilog, foid);
1629 mutex_enter(&zilog->zl_lock);
1630 mybatch = zilog->zl_next_batch;
1631 while (zilog->zl_writer) {
1632 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1633 if (mybatch <= zilog->zl_com_batch) {
1634 mutex_exit(&zilog->zl_lock);
1639 zilog->zl_next_batch++;
1640 zilog->zl_writer = B_TRUE;
1641 zil_commit_writer(zilog);
1642 zilog->zl_com_batch = mybatch;
1643 zilog->zl_writer = B_FALSE;
1644 mutex_exit(&zilog->zl_lock);
1646 /* wake up one thread to become the next writer */
1647 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1649 /* wake up all threads waiting for this batch to be committed */
1650 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1654 * Called in syncing context to free committed log blocks and update log header.
1657 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1659 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1660 uint64_t txg = dmu_tx_get_txg(tx);
1661 spa_t *spa = zilog->zl_spa;
1662 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1666 * We don't zero out zl_destroy_txg, so make sure we don't try
1667 * to destroy it twice.
1669 if (spa_sync_pass(spa) != 1)
1672 mutex_enter(&zilog->zl_lock);
1674 ASSERT(zilog->zl_stop_sync == 0);
1676 if (*replayed_seq != 0) {
1677 ASSERT(zh->zh_replay_seq < *replayed_seq);
1678 zh->zh_replay_seq = *replayed_seq;
1682 if (zilog->zl_destroy_txg == txg) {
1683 blkptr_t blk = zh->zh_log;
1685 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1687 bzero(zh, sizeof (zil_header_t));
1688 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1690 if (zilog->zl_keep_first) {
1692 * If this block was part of log chain that couldn't
1693 * be claimed because a device was missing during
1694 * zil_claim(), but that device later returns,
1695 * then this block could erroneously appear valid.
1696 * To guard against this, assign a new GUID to the new
1697 * log chain so it doesn't matter what blk points to.
1699 zil_init_log_chain(zilog, &blk);
1704 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1705 zh->zh_log = lwb->lwb_blk;
1706 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1708 list_remove(&zilog->zl_lwb_list, lwb);
1709 zio_free_zil(spa, txg, &lwb->lwb_blk);
1710 kmem_cache_free(zil_lwb_cache, lwb);
1713 * If we don't have anything left in the lwb list then
1714 * we've had an allocation failure and we need to zero
1715 * out the zil_header blkptr so that we don't end
1716 * up freeing the same block twice.
1718 if (list_head(&zilog->zl_lwb_list) == NULL)
1719 BP_ZERO(&zh->zh_log);
1721 mutex_exit(&zilog->zl_lock);
1727 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1728 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1734 kmem_cache_destroy(zil_lwb_cache);
1738 zil_set_sync(zilog_t *zilog, uint64_t sync)
1740 zilog->zl_sync = sync;
1744 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1746 zilog->zl_logbias = logbias;
1750 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1754 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1756 zilog->zl_header = zh_phys;
1758 zilog->zl_spa = dmu_objset_spa(os);
1759 zilog->zl_dmu_pool = dmu_objset_pool(os);
1760 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1761 zilog->zl_logbias = dmu_objset_logbias(os);
1762 zilog->zl_sync = dmu_objset_syncprop(os);
1763 zilog->zl_next_batch = 1;
1765 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1767 for (int i = 0; i < TXG_SIZE; i++) {
1768 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1769 MUTEX_DEFAULT, NULL);
1772 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1773 offsetof(lwb_t, lwb_node));
1775 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1776 offsetof(itx_t, itx_node));
1778 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1780 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1781 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1783 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1784 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1785 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1786 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1792 zil_free(zilog_t *zilog)
1794 zilog->zl_stop_sync = 1;
1796 ASSERT0(zilog->zl_suspend);
1797 ASSERT0(zilog->zl_suspending);
1799 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1800 list_destroy(&zilog->zl_lwb_list);
1802 avl_destroy(&zilog->zl_vdev_tree);
1803 mutex_destroy(&zilog->zl_vdev_lock);
1805 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1806 list_destroy(&zilog->zl_itx_commit_list);
1808 for (int i = 0; i < TXG_SIZE; i++) {
1810 * It's possible for an itx to be generated that doesn't dirty
1811 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1812 * callback to remove the entry. We remove those here.
1814 * Also free up the ziltest itxs.
1816 if (zilog->zl_itxg[i].itxg_itxs)
1817 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1818 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1821 mutex_destroy(&zilog->zl_lock);
1823 cv_destroy(&zilog->zl_cv_writer);
1824 cv_destroy(&zilog->zl_cv_suspend);
1825 cv_destroy(&zilog->zl_cv_batch[0]);
1826 cv_destroy(&zilog->zl_cv_batch[1]);
1828 kmem_free(zilog, sizeof (zilog_t));
1832 * Open an intent log.
1835 zil_open(objset_t *os, zil_get_data_t *get_data)
1837 zilog_t *zilog = dmu_objset_zil(os);
1839 ASSERT(zilog->zl_clean_taskq == NULL);
1840 ASSERT(zilog->zl_get_data == NULL);
1841 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1843 zilog->zl_get_data = get_data;
1844 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1845 2, 2, TASKQ_PREPOPULATE);
1851 * Close an intent log.
1854 zil_close(zilog_t *zilog)
1859 zil_commit(zilog, 0); /* commit all itx */
1862 * The lwb_max_txg for the stubby lwb will reflect the last activity
1863 * for the zil. After a txg_wait_synced() on the txg we know all the
1864 * callbacks have occurred that may clean the zil. Only then can we
1865 * destroy the zl_clean_taskq.
1867 mutex_enter(&zilog->zl_lock);
1868 lwb = list_tail(&zilog->zl_lwb_list);
1870 txg = lwb->lwb_max_txg;
1871 mutex_exit(&zilog->zl_lock);
1873 txg_wait_synced(zilog->zl_dmu_pool, txg);
1875 if (zilog_is_dirty(zilog))
1876 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1877 VERIFY(!zilog_is_dirty(zilog));
1879 taskq_destroy(zilog->zl_clean_taskq);
1880 zilog->zl_clean_taskq = NULL;
1881 zilog->zl_get_data = NULL;
1884 * We should have only one LWB left on the list; remove it now.
1886 mutex_enter(&zilog->zl_lock);
1887 lwb = list_head(&zilog->zl_lwb_list);
1889 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1890 list_remove(&zilog->zl_lwb_list, lwb);
1891 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1892 kmem_cache_free(zil_lwb_cache, lwb);
1894 mutex_exit(&zilog->zl_lock);
1897 static char *suspend_tag = "zil suspending";
1900 * Suspend an intent log. While in suspended mode, we still honor
1901 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1902 * On old version pools, we suspend the log briefly when taking a
1903 * snapshot so that it will have an empty intent log.
1905 * Long holds are not really intended to be used the way we do here --
1906 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1907 * could fail. Therefore we take pains to only put a long hold if it is
1908 * actually necessary. Fortunately, it will only be necessary if the
1909 * objset is currently mounted (or the ZVOL equivalent). In that case it
1910 * will already have a long hold, so we are not really making things any worse.
1912 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1913 * zvol_state_t), and use their mechanism to prevent their hold from being
1914 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1917 * if cookiep == NULL, this does both the suspend & resume.
1918 * Otherwise, it returns with the dataset "long held", and the cookie
1919 * should be passed into zil_resume().
1922 zil_suspend(const char *osname, void **cookiep)
1926 const zil_header_t *zh;
1929 error = dmu_objset_hold(osname, suspend_tag, &os);
1932 zilog = dmu_objset_zil(os);
1934 mutex_enter(&zilog->zl_lock);
1935 zh = zilog->zl_header;
1937 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1938 mutex_exit(&zilog->zl_lock);
1939 dmu_objset_rele(os, suspend_tag);
1940 return (SET_ERROR(EBUSY));
1944 * Don't put a long hold in the cases where we can avoid it. This
1945 * is when there is no cookie so we are doing a suspend & resume
1946 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1947 * for the suspend because it's already suspended, or there's no ZIL.
1949 if (cookiep == NULL && !zilog->zl_suspending &&
1950 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1951 mutex_exit(&zilog->zl_lock);
1952 dmu_objset_rele(os, suspend_tag);
1956 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1957 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1959 zilog->zl_suspend++;
1961 if (zilog->zl_suspend > 1) {
1963 * Someone else is already suspending it.
1964 * Just wait for them to finish.
1967 while (zilog->zl_suspending)
1968 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1969 mutex_exit(&zilog->zl_lock);
1971 if (cookiep == NULL)
1979 * If there is no pointer to an on-disk block, this ZIL must not
1980 * be active (e.g. filesystem not mounted), so there's nothing
1983 if (BP_IS_HOLE(&zh->zh_log)) {
1984 ASSERT(cookiep != NULL); /* fast path already handled */
1987 mutex_exit(&zilog->zl_lock);
1991 zilog->zl_suspending = B_TRUE;
1992 mutex_exit(&zilog->zl_lock);
1994 zil_commit(zilog, 0);
1996 zil_destroy(zilog, B_FALSE);
1998 mutex_enter(&zilog->zl_lock);
1999 zilog->zl_suspending = B_FALSE;
2000 cv_broadcast(&zilog->zl_cv_suspend);
2001 mutex_exit(&zilog->zl_lock);
2003 if (cookiep == NULL)
2011 zil_resume(void *cookie)
2013 objset_t *os = cookie;
2014 zilog_t *zilog = dmu_objset_zil(os);
2016 mutex_enter(&zilog->zl_lock);
2017 ASSERT(zilog->zl_suspend != 0);
2018 zilog->zl_suspend--;
2019 mutex_exit(&zilog->zl_lock);
2020 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2021 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2024 typedef struct zil_replay_arg {
2025 zil_replay_func_t **zr_replay;
2027 boolean_t zr_byteswap;
2032 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2034 char name[ZFS_MAX_DATASET_NAME_LEN];
2036 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2038 dmu_objset_name(zilog->zl_os, name);
2040 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2041 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2042 (u_longlong_t)lr->lrc_seq,
2043 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2044 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2050 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2052 zil_replay_arg_t *zr = zra;
2053 const zil_header_t *zh = zilog->zl_header;
2054 uint64_t reclen = lr->lrc_reclen;
2055 uint64_t txtype = lr->lrc_txtype;
2058 zilog->zl_replaying_seq = lr->lrc_seq;
2060 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2063 if (lr->lrc_txg < claim_txg) /* already committed */
2066 /* Strip case-insensitive bit, still present in log record */
2069 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2070 return (zil_replay_error(zilog, lr, EINVAL));
2073 * If this record type can be logged out of order, the object
2074 * (lr_foid) may no longer exist. That's legitimate, not an error.
2076 if (TX_OOO(txtype)) {
2077 error = dmu_object_info(zilog->zl_os,
2078 ((lr_ooo_t *)lr)->lr_foid, NULL);
2079 if (error == ENOENT || error == EEXIST)
2084 * Make a copy of the data so we can revise and extend it.
2086 bcopy(lr, zr->zr_lr, reclen);
2089 * If this is a TX_WRITE with a blkptr, suck in the data.
2091 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2092 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2093 zr->zr_lr + reclen);
2095 return (zil_replay_error(zilog, lr, error));
2099 * The log block containing this lr may have been byteswapped
2100 * so that we can easily examine common fields like lrc_txtype.
2101 * However, the log is a mix of different record types, and only the
2102 * replay vectors know how to byteswap their records. Therefore, if
2103 * the lr was byteswapped, undo it before invoking the replay vector.
2105 if (zr->zr_byteswap)
2106 byteswap_uint64_array(zr->zr_lr, reclen);
2109 * We must now do two things atomically: replay this log record,
2110 * and update the log header sequence number to reflect the fact that
2111 * we did so. At the end of each replay function the sequence number
2112 * is updated if we are in replay mode.
2114 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2117 * The DMU's dnode layer doesn't see removes until the txg
2118 * commits, so a subsequent claim can spuriously fail with
2119 * EEXIST. So if we receive any error we try syncing out
2120 * any removes then retry the transaction. Note that we
2121 * specify B_FALSE for byteswap now, so we don't do it twice.
2123 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2124 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2126 return (zil_replay_error(zilog, lr, error));
2133 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2135 zilog->zl_replay_blks++;
2141 * If this dataset has a non-empty intent log, replay it and destroy it.
2144 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2146 zilog_t *zilog = dmu_objset_zil(os);
2147 const zil_header_t *zh = zilog->zl_header;
2148 zil_replay_arg_t zr;
2150 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2151 zil_destroy(zilog, B_TRUE);
2155 zr.zr_replay = replay_func;
2157 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2158 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2161 * Wait for in-progress removes to sync before starting replay.
2163 txg_wait_synced(zilog->zl_dmu_pool, 0);
2165 zilog->zl_replay = B_TRUE;
2166 zilog->zl_replay_time = ddi_get_lbolt();
2167 ASSERT(zilog->zl_replay_blks == 0);
2168 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2170 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2172 zil_destroy(zilog, B_FALSE);
2173 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2174 zilog->zl_replay = B_FALSE;
2178 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2180 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2183 if (zilog->zl_replay) {
2184 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2185 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2186 zilog->zl_replaying_seq;
2195 zil_vdev_offline(const char *osname, void *arg)
2199 error = zil_suspend(osname, NULL);
2201 return (SET_ERROR(EEXIST));