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, 2017 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>
45 * The zfs intent log (ZIL) saves transaction records of system calls
46 * that change the file system in memory with enough information
47 * to be able to replay them. These are stored in memory until
48 * either the DMU transaction group (txg) commits them to the stable pool
49 * and they can be discarded, or they are flushed to the stable log
50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
51 * requirement. In the event of a panic or power fail then those log
52 * records (transactions) are replayed.
54 * There is one ZIL per file system. Its on-disk (pool) format consists
61 * A log record holds a system call transaction. Log blocks can
62 * hold many log records and the blocks are chained together.
63 * Each ZIL block contains a block pointer (blkptr_t) to the next
64 * ZIL block in the chain. The ZIL header points to the first
65 * block in the chain. Note there is not a fixed place in the pool
66 * to hold blocks. They are dynamically allocated and freed as
67 * needed from the blocks available. Figure X shows the ZIL structure:
71 * Disable intent logging replay. This global ZIL switch affects all pools.
73 int zil_replay_disable = 0;
74 SYSCTL_DECL(_vfs_zfs);
75 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
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 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
85 &zfs_nocacheflush, 0, "Disable cache flush");
86 boolean_t zfs_trim_enabled = B_TRUE;
87 SYSCTL_DECL(_vfs_zfs_trim);
88 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
92 * Limit SLOG write size per commit executed with synchronous priority.
93 * Any writes above that will be executed with lower (asynchronous) priority
94 * to limit potential SLOG device abuse by single active ZIL writer.
96 uint64_t zil_slog_bulk = 768 * 1024;
97 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_bulk, CTLFLAG_RWTUN,
98 &zil_slog_bulk, 0, "Maximal SLOG commit size with sync priority");
100 static kmem_cache_t *zil_lwb_cache;
102 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
103 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
106 zil_bp_compare(const void *x1, const void *x2)
108 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
109 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
111 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
113 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
116 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
118 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
125 zil_bp_tree_init(zilog_t *zilog)
127 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
128 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
132 zil_bp_tree_fini(zilog_t *zilog)
134 avl_tree_t *t = &zilog->zl_bp_tree;
138 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
139 kmem_free(zn, sizeof (zil_bp_node_t));
145 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
147 avl_tree_t *t = &zilog->zl_bp_tree;
152 if (BP_IS_EMBEDDED(bp))
155 dva = BP_IDENTITY(bp);
157 if (avl_find(t, dva, &where) != NULL)
158 return (SET_ERROR(EEXIST));
160 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
162 avl_insert(t, zn, where);
167 static zil_header_t *
168 zil_header_in_syncing_context(zilog_t *zilog)
170 return ((zil_header_t *)zilog->zl_header);
174 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
176 zio_cksum_t *zc = &bp->blk_cksum;
178 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
179 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
180 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
181 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
185 * Read a log block and make sure it's valid.
188 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
191 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
192 arc_flags_t aflags = ARC_FLAG_WAIT;
193 arc_buf_t *abuf = NULL;
197 if (zilog->zl_header->zh_claim_txg == 0)
198 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
200 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
201 zio_flags |= ZIO_FLAG_SPECULATIVE;
203 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
204 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
206 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
207 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
210 zio_cksum_t cksum = bp->blk_cksum;
213 * Validate the checksummed log block.
215 * Sequence numbers should be... sequential. The checksum
216 * verifier for the next block should be bp's checksum plus 1.
218 * Also check the log chain linkage and size used.
220 cksum.zc_word[ZIL_ZC_SEQ]++;
222 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
223 zil_chain_t *zilc = abuf->b_data;
224 char *lr = (char *)(zilc + 1);
225 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
227 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
228 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
229 error = SET_ERROR(ECKSUM);
231 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
233 *end = (char *)dst + len;
234 *nbp = zilc->zc_next_blk;
237 char *lr = abuf->b_data;
238 uint64_t size = BP_GET_LSIZE(bp);
239 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
241 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
242 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
243 (zilc->zc_nused > (size - sizeof (*zilc)))) {
244 error = SET_ERROR(ECKSUM);
246 ASSERT3U(zilc->zc_nused, <=,
247 SPA_OLD_MAXBLOCKSIZE);
248 bcopy(lr, dst, zilc->zc_nused);
249 *end = (char *)dst + zilc->zc_nused;
250 *nbp = zilc->zc_next_blk;
254 arc_buf_destroy(abuf, &abuf);
261 * Read a TX_WRITE log data block.
264 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
266 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
267 const blkptr_t *bp = &lr->lr_blkptr;
268 arc_flags_t aflags = ARC_FLAG_WAIT;
269 arc_buf_t *abuf = NULL;
273 if (BP_IS_HOLE(bp)) {
275 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
279 if (zilog->zl_header->zh_claim_txg == 0)
280 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
282 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
283 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
285 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
286 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
290 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
291 arc_buf_destroy(abuf, &abuf);
298 * Parse the intent log, and call parse_func for each valid record within.
301 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
302 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
304 const zil_header_t *zh = zilog->zl_header;
305 boolean_t claimed = !!zh->zh_claim_txg;
306 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
307 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
308 uint64_t max_blk_seq = 0;
309 uint64_t max_lr_seq = 0;
310 uint64_t blk_count = 0;
311 uint64_t lr_count = 0;
312 blkptr_t blk, next_blk;
317 * Old logs didn't record the maximum zh_claim_lr_seq.
319 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
320 claim_lr_seq = UINT64_MAX;
323 * Starting at the block pointed to by zh_log we read the log chain.
324 * For each block in the chain we strongly check that block to
325 * ensure its validity. We stop when an invalid block is found.
326 * For each block pointer in the chain we call parse_blk_func().
327 * For each record in each valid block we call parse_lr_func().
328 * If the log has been claimed, stop if we encounter a sequence
329 * number greater than the highest claimed sequence number.
331 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
332 zil_bp_tree_init(zilog);
334 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
335 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
339 if (blk_seq > claim_blk_seq)
341 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
343 ASSERT3U(max_blk_seq, <, blk_seq);
344 max_blk_seq = blk_seq;
347 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
350 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
354 for (lrp = lrbuf; lrp < end; lrp += reclen) {
355 lr_t *lr = (lr_t *)lrp;
356 reclen = lr->lrc_reclen;
357 ASSERT3U(reclen, >=, sizeof (lr_t));
358 if (lr->lrc_seq > claim_lr_seq)
360 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
362 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
363 max_lr_seq = lr->lrc_seq;
368 zilog->zl_parse_error = error;
369 zilog->zl_parse_blk_seq = max_blk_seq;
370 zilog->zl_parse_lr_seq = max_lr_seq;
371 zilog->zl_parse_blk_count = blk_count;
372 zilog->zl_parse_lr_count = lr_count;
374 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
375 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
377 zil_bp_tree_fini(zilog);
378 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
384 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
387 * Claim log block if not already committed and not already claimed.
388 * If tx == NULL, just verify that the block is claimable.
390 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
391 zil_bp_tree_add(zilog, bp) != 0)
394 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
395 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
396 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
400 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
402 lr_write_t *lr = (lr_write_t *)lrc;
405 if (lrc->lrc_txtype != TX_WRITE)
409 * If the block is not readable, don't claim it. This can happen
410 * in normal operation when a log block is written to disk before
411 * some of the dmu_sync() blocks it points to. In this case, the
412 * transaction cannot have been committed to anyone (we would have
413 * waited for all writes to be stable first), so it is semantically
414 * correct to declare this the end of the log.
416 if (lr->lr_blkptr.blk_birth >= first_txg &&
417 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
419 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
424 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
426 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
432 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
434 lr_write_t *lr = (lr_write_t *)lrc;
435 blkptr_t *bp = &lr->lr_blkptr;
438 * If we previously claimed it, we need to free it.
440 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
441 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
443 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
449 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
453 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
454 lwb->lwb_zilog = zilog;
456 lwb->lwb_slog = slog;
457 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
458 lwb->lwb_max_txg = txg;
461 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
462 lwb->lwb_nused = sizeof (zil_chain_t);
463 lwb->lwb_sz = BP_GET_LSIZE(bp);
466 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
469 mutex_enter(&zilog->zl_lock);
470 list_insert_tail(&zilog->zl_lwb_list, lwb);
471 mutex_exit(&zilog->zl_lock);
477 * Called when we create in-memory log transactions so that we know
478 * to cleanup the itxs at the end of spa_sync().
481 zilog_dirty(zilog_t *zilog, uint64_t txg)
483 dsl_pool_t *dp = zilog->zl_dmu_pool;
484 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
486 if (ds->ds_is_snapshot)
487 panic("dirtying snapshot!");
489 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
490 /* up the hold count until we can be written out */
491 dmu_buf_add_ref(ds->ds_dbuf, zilog);
496 * Determine if the zil is dirty in the specified txg. Callers wanting to
497 * ensure that the dirty state does not change must hold the itxg_lock for
498 * the specified txg. Holding the lock will ensure that the zil cannot be
499 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
503 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
505 dsl_pool_t *dp = zilog->zl_dmu_pool;
507 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
513 * Determine if the zil is dirty. The zil is considered dirty if it has
514 * any pending itx records that have not been cleaned by zil_clean().
517 zilog_is_dirty(zilog_t *zilog)
519 dsl_pool_t *dp = zilog->zl_dmu_pool;
521 for (int t = 0; t < TXG_SIZE; t++) {
522 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
529 * Create an on-disk intent log.
532 zil_create(zilog_t *zilog)
534 const zil_header_t *zh = zilog->zl_header;
540 boolean_t slog = FALSE;
543 * Wait for any previous destroy to complete.
545 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
547 ASSERT(zh->zh_claim_txg == 0);
548 ASSERT(zh->zh_replay_seq == 0);
553 * Allocate an initial log block if:
554 * - there isn't one already
555 * - the existing block is the wrong endianess
557 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
558 tx = dmu_tx_create(zilog->zl_os);
559 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
560 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
561 txg = dmu_tx_get_txg(tx);
563 if (!BP_IS_HOLE(&blk)) {
564 zio_free_zil(zilog->zl_spa, txg, &blk);
568 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
569 ZIL_MIN_BLKSZ, &slog);
572 zil_init_log_chain(zilog, &blk);
576 * Allocate a log write buffer (lwb) for the first log block.
579 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
582 * If we just allocated the first log block, commit our transaction
583 * and wait for zil_sync() to stuff the block poiner into zh_log.
584 * (zh is part of the MOS, so we cannot modify it in open context.)
588 txg_wait_synced(zilog->zl_dmu_pool, txg);
591 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
597 * In one tx, free all log blocks and clear the log header.
598 * If keep_first is set, then we're replaying a log with no content.
599 * We want to keep the first block, however, so that the first
600 * synchronous transaction doesn't require a txg_wait_synced()
601 * in zil_create(). We don't need to txg_wait_synced() here either
602 * when keep_first is set, because both zil_create() and zil_destroy()
603 * will wait for any in-progress destroys to complete.
606 zil_destroy(zilog_t *zilog, boolean_t keep_first)
608 const zil_header_t *zh = zilog->zl_header;
614 * Wait for any previous destroy to complete.
616 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
618 zilog->zl_old_header = *zh; /* debugging aid */
620 if (BP_IS_HOLE(&zh->zh_log))
623 tx = dmu_tx_create(zilog->zl_os);
624 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
625 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
626 txg = dmu_tx_get_txg(tx);
628 mutex_enter(&zilog->zl_lock);
630 ASSERT3U(zilog->zl_destroy_txg, <, txg);
631 zilog->zl_destroy_txg = txg;
632 zilog->zl_keep_first = keep_first;
634 if (!list_is_empty(&zilog->zl_lwb_list)) {
635 ASSERT(zh->zh_claim_txg == 0);
637 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
638 list_remove(&zilog->zl_lwb_list, lwb);
639 if (lwb->lwb_buf != NULL)
640 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
641 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
642 kmem_cache_free(zil_lwb_cache, lwb);
644 } else if (!keep_first) {
645 zil_destroy_sync(zilog, tx);
647 mutex_exit(&zilog->zl_lock);
653 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
655 ASSERT(list_is_empty(&zilog->zl_lwb_list));
656 (void) zil_parse(zilog, zil_free_log_block,
657 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
661 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
663 dmu_tx_t *tx = txarg;
664 uint64_t first_txg = dmu_tx_get_txg(tx);
670 error = dmu_objset_own_obj(dp, ds->ds_object,
671 DMU_OST_ANY, B_FALSE, FTAG, &os);
674 * EBUSY indicates that the objset is inconsistent, in which
675 * case it can not have a ZIL.
677 if (error != EBUSY) {
678 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
679 (unsigned long long)ds->ds_object, error);
684 zilog = dmu_objset_zil(os);
685 zh = zil_header_in_syncing_context(zilog);
687 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
688 if (!BP_IS_HOLE(&zh->zh_log))
689 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
690 BP_ZERO(&zh->zh_log);
691 dsl_dataset_dirty(dmu_objset_ds(os), tx);
692 dmu_objset_disown(os, FTAG);
697 * Claim all log blocks if we haven't already done so, and remember
698 * the highest claimed sequence number. This ensures that if we can
699 * read only part of the log now (e.g. due to a missing device),
700 * but we can read the entire log later, we will not try to replay
701 * or destroy beyond the last block we successfully claimed.
703 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
704 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
705 (void) zil_parse(zilog, zil_claim_log_block,
706 zil_claim_log_record, tx, first_txg);
707 zh->zh_claim_txg = first_txg;
708 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
709 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
710 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
711 zh->zh_flags |= ZIL_REPLAY_NEEDED;
712 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
713 dsl_dataset_dirty(dmu_objset_ds(os), tx);
716 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
717 dmu_objset_disown(os, FTAG);
722 * Check the log by walking the log chain.
723 * Checksum errors are ok as they indicate the end of the chain.
724 * Any other error (no device or read failure) returns an error.
728 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
737 error = dmu_objset_from_ds(ds, &os);
739 cmn_err(CE_WARN, "can't open objset %llu, error %d",
740 (unsigned long long)ds->ds_object, error);
744 zilog = dmu_objset_zil(os);
745 bp = (blkptr_t *)&zilog->zl_header->zh_log;
748 * Check the first block and determine if it's on a log device
749 * which may have been removed or faulted prior to loading this
750 * pool. If so, there's no point in checking the rest of the log
751 * as its content should have already been synced to the pool.
753 if (!BP_IS_HOLE(bp)) {
755 boolean_t valid = B_TRUE;
757 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
758 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
759 if (vd->vdev_islog && vdev_is_dead(vd))
760 valid = vdev_log_state_valid(vd);
761 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
768 * Because tx == NULL, zil_claim_log_block() will not actually claim
769 * any blocks, but just determine whether it is possible to do so.
770 * In addition to checking the log chain, zil_claim_log_block()
771 * will invoke zio_claim() with a done func of spa_claim_notify(),
772 * which will update spa_max_claim_txg. See spa_load() for details.
774 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
775 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
777 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
781 zil_vdev_compare(const void *x1, const void *x2)
783 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
784 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
795 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
797 avl_tree_t *t = &zilog->zl_vdev_tree;
799 zil_vdev_node_t *zv, zvsearch;
800 int ndvas = BP_GET_NDVAS(bp);
803 if (zfs_nocacheflush)
806 ASSERT(zilog->zl_writer);
809 * Even though we're zl_writer, we still need a lock because the
810 * zl_get_data() callbacks may have dmu_sync() done callbacks
811 * that will run concurrently.
813 mutex_enter(&zilog->zl_vdev_lock);
814 for (i = 0; i < ndvas; i++) {
815 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
816 if (avl_find(t, &zvsearch, &where) == NULL) {
817 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
818 zv->zv_vdev = zvsearch.zv_vdev;
819 avl_insert(t, zv, where);
822 mutex_exit(&zilog->zl_vdev_lock);
826 zil_flush_vdevs(zilog_t *zilog)
828 spa_t *spa = zilog->zl_spa;
829 avl_tree_t *t = &zilog->zl_vdev_tree;
834 ASSERT(zilog->zl_writer);
837 * We don't need zl_vdev_lock here because we're the zl_writer,
838 * and all zl_get_data() callbacks are done.
840 if (avl_numnodes(t) == 0)
843 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
845 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
846 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
847 if (vd != NULL && !vd->vdev_nowritecache) {
849 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
852 kmem_free(zv, sizeof (*zv));
856 * Wait for all the flushes to complete. Not all devices actually
857 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
860 (void) zio_wait(zio);
862 spa_config_exit(spa, SCL_STATE, FTAG);
866 * Function called when a log block write completes
869 zil_lwb_write_done(zio_t *zio)
871 lwb_t *lwb = zio->io_private;
872 zilog_t *zilog = lwb->lwb_zilog;
873 dmu_tx_t *tx = lwb->lwb_tx;
875 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
876 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
877 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
878 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
879 ASSERT(!BP_IS_GANG(zio->io_bp));
880 ASSERT(!BP_IS_HOLE(zio->io_bp));
881 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
884 * Ensure the lwb buffer pointer is cleared before releasing
885 * the txg. If we have had an allocation failure and
886 * the txg is waiting to sync then we want want zil_sync()
887 * to remove the lwb so that it's not picked up as the next new
888 * one in zil_commit_writer(). zil_sync() will only remove
889 * the lwb if lwb_buf is null.
891 abd_put(zio->io_abd);
892 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
893 mutex_enter(&zilog->zl_lock);
896 mutex_exit(&zilog->zl_lock);
899 * Now that we've written this log block, we have a stable pointer
900 * to the next block in the chain, so it's OK to let the txg in
901 * which we allocated the next block sync.
907 * Initialize the io for a log block.
910 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
915 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
916 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
917 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
919 if (zilog->zl_root_zio == NULL) {
920 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
923 if (lwb->lwb_zio == NULL) {
924 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
925 BP_GET_LSIZE(&lwb->lwb_blk));
926 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
927 prio = ZIO_PRIORITY_SYNC_WRITE;
929 prio = ZIO_PRIORITY_ASYNC_WRITE;
930 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
931 0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
932 zil_lwb_write_done, lwb, prio,
933 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
938 * Define a limited set of intent log block sizes.
940 * These must be a multiple of 4KB. Note only the amount used (again
941 * aligned to 4KB) actually gets written. However, we can't always just
942 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
944 uint64_t zil_block_buckets[] = {
945 4096, /* non TX_WRITE */
946 8192+4096, /* data base */
947 32*1024 + 4096, /* NFS writes */
952 * Start a log block write and advance to the next log block.
953 * Calls are serialized.
956 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last)
960 spa_t *spa = zilog->zl_spa;
964 uint64_t zil_blksz, wsz;
968 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
969 zilc = (zil_chain_t *)lwb->lwb_buf;
970 bp = &zilc->zc_next_blk;
972 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
973 bp = &zilc->zc_next_blk;
976 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
979 * Allocate the next block and save its address in this block
980 * before writing it in order to establish the log chain.
981 * Note that if the allocation of nlwb synced before we wrote
982 * the block that points at it (lwb), we'd leak it if we crashed.
983 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
984 * We dirty the dataset to ensure that zil_sync() will be called
985 * to clean up in the event of allocation failure or I/O failure.
987 tx = dmu_tx_create(zilog->zl_os);
990 * Since we are not going to create any new dirty data and we can even
991 * help with clearing the existing dirty data, we should not be subject
992 * to the dirty data based delays.
993 * We (ab)use TXG_WAITED to bypass the delay mechanism.
994 * One side effect from using TXG_WAITED is that dmu_tx_assign() can
995 * fail if the pool is suspended. Those are dramatic circumstances,
996 * so we return NULL to signal that the normal ZIL processing is not
997 * possible and txg_wait_synced() should be used to ensure that the data
1000 error = dmu_tx_assign(tx, TXG_WAITED);
1002 ASSERT3S(error, ==, EIO);
1006 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1007 txg = dmu_tx_get_txg(tx);
1012 * Log blocks are pre-allocated. Here we select the size of the next
1013 * block, based on size used in the last block.
1014 * - first find the smallest bucket that will fit the block from a
1015 * limited set of block sizes. This is because it's faster to write
1016 * blocks allocated from the same metaslab as they are adjacent or
1018 * - next find the maximum from the new suggested size and an array of
1019 * previous sizes. This lessens a picket fence effect of wrongly
1020 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1023 * Note we only write what is used, but we can't just allocate
1024 * the maximum block size because we can exhaust the available
1027 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1028 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1030 zil_blksz = zil_block_buckets[i];
1031 if (zil_blksz == UINT64_MAX)
1032 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1033 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1034 for (i = 0; i < ZIL_PREV_BLKS; i++)
1035 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1036 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1039 /* pass the old blkptr in order to spread log blocks across devs */
1040 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1042 ASSERT3U(bp->blk_birth, ==, txg);
1043 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1044 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1047 * Allocate a new log write buffer (lwb).
1049 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1051 /* Record the block for later vdev flushing */
1052 zil_add_block(zilog, &lwb->lwb_blk);
1055 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1056 /* For Slim ZIL only write what is used. */
1057 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1058 ASSERT3U(wsz, <=, lwb->lwb_sz);
1059 zio_shrink(lwb->lwb_zio, wsz);
1066 zilc->zc_nused = lwb->lwb_nused;
1067 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1070 * clear unused data for security
1072 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1075 lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC;
1076 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1079 * If there was an allocation failure then nlwb will be null which
1080 * forces a txg_wait_synced().
1086 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1089 lr_write_t *lrwb, *lrw;
1091 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1096 ASSERT(lwb->lwb_buf != NULL);
1098 lrc = &itx->itx_lr; /* Common log record inside itx. */
1099 lrw = (lr_write_t *)lrc; /* Write log record inside itx. */
1100 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1101 dlen = P2ROUNDUP_TYPED(
1102 lrw->lr_length, sizeof (uint64_t), uint64_t);
1106 reclen = lrc->lrc_reclen;
1107 zilog->zl_cur_used += (reclen + dlen);
1110 zil_lwb_write_init(zilog, lwb);
1114 * If this record won't fit in the current log block, start a new one.
1115 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1117 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1118 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1119 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1120 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1121 lwb = zil_lwb_write_start(zilog, lwb, B_FALSE);
1124 zil_lwb_write_init(zilog, lwb);
1125 ASSERT(LWB_EMPTY(lwb));
1126 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1127 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1130 dnow = MIN(dlen, lwb_sp - reclen);
1131 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1132 bcopy(lrc, lr_buf, reclen);
1133 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1134 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1137 * If it's a write, fetch the data or get its blkptr as appropriate.
1139 if (lrc->lrc_txtype == TX_WRITE) {
1140 if (txg > spa_freeze_txg(zilog->zl_spa))
1141 txg_wait_synced(zilog->zl_dmu_pool, txg);
1142 if (itx->itx_wr_state != WR_COPIED) {
1146 if (itx->itx_wr_state == WR_NEED_COPY) {
1147 dbuf = lr_buf + reclen;
1148 lrcb->lrc_reclen += dnow;
1149 if (lrwb->lr_length > dnow)
1150 lrwb->lr_length = dnow;
1151 lrw->lr_offset += dnow;
1152 lrw->lr_length -= dnow;
1154 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1157 error = zilog->zl_get_data(
1158 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1160 txg_wait_synced(zilog->zl_dmu_pool, txg);
1164 ASSERT(error == ENOENT || error == EEXIST ||
1172 * We're actually making an entry, so update lrc_seq to be the
1173 * log record sequence number. Note that this is generally not
1174 * equal to the itx sequence number because not all transactions
1175 * are synchronous, and sometimes spa_sync() gets there first.
1177 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1178 lwb->lwb_nused += reclen + dnow;
1179 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1180 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1181 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1185 zilog->zl_cur_used += reclen;
1193 zil_itx_create(uint64_t txtype, size_t lrsize)
1197 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1199 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1200 itx->itx_lr.lrc_txtype = txtype;
1201 itx->itx_lr.lrc_reclen = lrsize;
1202 itx->itx_lr.lrc_seq = 0; /* defensive */
1203 itx->itx_sync = B_TRUE; /* default is synchronous */
1209 zil_itx_destroy(itx_t *itx)
1211 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1215 * Free up the sync and async itxs. The itxs_t has already been detached
1216 * so no locks are needed.
1219 zil_itxg_clean(itxs_t *itxs)
1225 itx_async_node_t *ian;
1227 list = &itxs->i_sync_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);
1235 t = &itxs->i_async_tree;
1236 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1237 list = &ian->ia_list;
1238 while ((itx = list_head(list)) != NULL) {
1239 list_remove(list, itx);
1240 kmem_free(itx, offsetof(itx_t, itx_lr) +
1241 itx->itx_lr.lrc_reclen);
1244 kmem_free(ian, sizeof (itx_async_node_t));
1248 kmem_free(itxs, sizeof (itxs_t));
1252 zil_aitx_compare(const void *x1, const void *x2)
1254 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1255 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1266 * Remove all async itx with the given oid.
1269 zil_remove_async(zilog_t *zilog, uint64_t oid)
1272 itx_async_node_t *ian;
1279 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1281 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1284 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1286 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1287 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1289 mutex_enter(&itxg->itxg_lock);
1290 if (itxg->itxg_txg != txg) {
1291 mutex_exit(&itxg->itxg_lock);
1296 * Locate the object node and append its list.
1298 t = &itxg->itxg_itxs->i_async_tree;
1299 ian = avl_find(t, &oid, &where);
1301 list_move_tail(&clean_list, &ian->ia_list);
1302 mutex_exit(&itxg->itxg_lock);
1304 while ((itx = list_head(&clean_list)) != NULL) {
1305 list_remove(&clean_list, itx);
1306 kmem_free(itx, offsetof(itx_t, itx_lr) +
1307 itx->itx_lr.lrc_reclen);
1309 list_destroy(&clean_list);
1313 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1317 itxs_t *itxs, *clean = NULL;
1320 * Object ids can be re-instantiated in the next txg so
1321 * remove any async transactions to avoid future leaks.
1322 * This can happen if a fsync occurs on the re-instantiated
1323 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1324 * the new file data and flushes a write record for the old object.
1326 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1327 zil_remove_async(zilog, itx->itx_oid);
1330 * Ensure the data of a renamed file is committed before the rename.
1332 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1333 zil_async_to_sync(zilog, itx->itx_oid);
1335 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1338 txg = dmu_tx_get_txg(tx);
1340 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1341 mutex_enter(&itxg->itxg_lock);
1342 itxs = itxg->itxg_itxs;
1343 if (itxg->itxg_txg != txg) {
1346 * The zil_clean callback hasn't got around to cleaning
1347 * this itxg. Save the itxs for release below.
1348 * This should be rare.
1350 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1351 "txg %llu", itxg->itxg_txg);
1352 clean = itxg->itxg_itxs;
1354 itxg->itxg_txg = txg;
1355 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1357 list_create(&itxs->i_sync_list, sizeof (itx_t),
1358 offsetof(itx_t, itx_node));
1359 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1360 sizeof (itx_async_node_t),
1361 offsetof(itx_async_node_t, ia_node));
1363 if (itx->itx_sync) {
1364 list_insert_tail(&itxs->i_sync_list, itx);
1366 avl_tree_t *t = &itxs->i_async_tree;
1367 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1368 itx_async_node_t *ian;
1371 ian = avl_find(t, &foid, &where);
1373 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1374 list_create(&ian->ia_list, sizeof (itx_t),
1375 offsetof(itx_t, itx_node));
1376 ian->ia_foid = foid;
1377 avl_insert(t, ian, where);
1379 list_insert_tail(&ian->ia_list, itx);
1382 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1383 zilog_dirty(zilog, txg);
1384 mutex_exit(&itxg->itxg_lock);
1386 /* Release the old itxs now we've dropped the lock */
1388 zil_itxg_clean(clean);
1392 * If there are any in-memory intent log transactions which have now been
1393 * synced then start up a taskq to free them. We should only do this after we
1394 * have written out the uberblocks (i.e. txg has been comitted) so that
1395 * don't inadvertently clean out in-memory log records that would be required
1399 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1401 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1404 mutex_enter(&itxg->itxg_lock);
1405 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1406 mutex_exit(&itxg->itxg_lock);
1409 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1410 ASSERT3U(itxg->itxg_txg, !=, 0);
1411 clean_me = itxg->itxg_itxs;
1412 itxg->itxg_itxs = NULL;
1414 mutex_exit(&itxg->itxg_lock);
1416 * Preferably start a task queue to free up the old itxs but
1417 * if taskq_dispatch can't allocate resources to do that then
1418 * free it in-line. This should be rare. Note, using TQ_SLEEP
1419 * created a bad performance problem.
1421 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1422 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1423 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1424 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1425 zil_itxg_clean(clean_me);
1429 * Get the list of itxs to commit into zl_itx_commit_list.
1432 zil_get_commit_list(zilog_t *zilog)
1435 list_t *commit_list = &zilog->zl_itx_commit_list;
1437 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1440 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1443 * This is inherently racy, since there is nothing to prevent
1444 * the last synced txg from changing. That's okay since we'll
1445 * only commit things in the future.
1447 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1448 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1450 mutex_enter(&itxg->itxg_lock);
1451 if (itxg->itxg_txg != txg) {
1452 mutex_exit(&itxg->itxg_lock);
1457 * If we're adding itx records to the zl_itx_commit_list,
1458 * then the zil better be dirty in this "txg". We can assert
1459 * that here since we're holding the itxg_lock which will
1460 * prevent spa_sync from cleaning it. Once we add the itxs
1461 * to the zl_itx_commit_list we must commit it to disk even
1462 * if it's unnecessary (i.e. the txg was synced).
1464 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1465 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1466 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1468 mutex_exit(&itxg->itxg_lock);
1473 * Move the async itxs for a specified object to commit into sync lists.
1476 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1479 itx_async_node_t *ian;
1483 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1486 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1489 * This is inherently racy, since there is nothing to prevent
1490 * the last synced txg from changing.
1492 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1493 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1495 mutex_enter(&itxg->itxg_lock);
1496 if (itxg->itxg_txg != txg) {
1497 mutex_exit(&itxg->itxg_lock);
1502 * If a foid is specified then find that node and append its
1503 * list. Otherwise walk the tree appending all the lists
1504 * to the sync list. We add to the end rather than the
1505 * beginning to ensure the create has happened.
1507 t = &itxg->itxg_itxs->i_async_tree;
1509 ian = avl_find(t, &foid, &where);
1511 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1515 void *cookie = NULL;
1517 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1518 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1520 list_destroy(&ian->ia_list);
1521 kmem_free(ian, sizeof (itx_async_node_t));
1524 mutex_exit(&itxg->itxg_lock);
1529 zil_commit_writer(zilog_t *zilog)
1534 spa_t *spa = zilog->zl_spa;
1537 ASSERT(zilog->zl_root_zio == NULL);
1539 mutex_exit(&zilog->zl_lock);
1541 zil_get_commit_list(zilog);
1544 * Return if there's nothing to commit before we dirty the fs by
1545 * calling zil_create().
1547 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1548 mutex_enter(&zilog->zl_lock);
1552 if (zilog->zl_suspend) {
1555 lwb = list_tail(&zilog->zl_lwb_list);
1557 lwb = zil_create(zilog);
1560 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1561 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1562 txg = itx->itx_lr.lrc_txg;
1563 ASSERT3U(txg, !=, 0);
1566 * This is inherently racy and may result in us writing
1567 * out a log block for a txg that was just synced. This is
1568 * ok since we'll end cleaning up that log block the next
1569 * time we call zil_sync().
1571 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1572 lwb = zil_lwb_commit(zilog, itx, lwb);
1573 list_remove(&zilog->zl_itx_commit_list, itx);
1574 kmem_free(itx, offsetof(itx_t, itx_lr)
1575 + itx->itx_lr.lrc_reclen);
1577 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1579 /* write the last block out */
1580 if (lwb != NULL && lwb->lwb_zio != NULL)
1581 lwb = zil_lwb_write_start(zilog, lwb, B_TRUE);
1583 zilog->zl_cur_used = 0;
1586 * Wait if necessary for the log blocks to be on stable storage.
1588 if (zilog->zl_root_zio) {
1589 error = zio_wait(zilog->zl_root_zio);
1590 zilog->zl_root_zio = NULL;
1591 zil_flush_vdevs(zilog);
1594 if (error || lwb == NULL)
1595 txg_wait_synced(zilog->zl_dmu_pool, 0);
1597 mutex_enter(&zilog->zl_lock);
1600 * Remember the highest committed log sequence number for ztest.
1601 * We only update this value when all the log writes succeeded,
1602 * because ztest wants to ASSERT that it got the whole log chain.
1604 if (error == 0 && lwb != NULL)
1605 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1609 * Commit zfs transactions to stable storage.
1610 * If foid is 0 push out all transactions, otherwise push only those
1611 * for that object or might reference that object.
1613 * itxs are committed in batches. In a heavily stressed zil there will be
1614 * a commit writer thread who is writing out a bunch of itxs to the log
1615 * for a set of committing threads (cthreads) in the same batch as the writer.
1616 * Those cthreads are all waiting on the same cv for that batch.
1618 * There will also be a different and growing batch of threads that are
1619 * waiting to commit (qthreads). When the committing batch completes
1620 * a transition occurs such that the cthreads exit and the qthreads become
1621 * cthreads. One of the new cthreads becomes the writer thread for the
1622 * batch. Any new threads arriving become new qthreads.
1624 * Only 2 condition variables are needed and there's no transition
1625 * between the two cvs needed. They just flip-flop between qthreads
1628 * Using this scheme we can efficiently wakeup up only those threads
1629 * that have been committed.
1632 zil_commit(zilog_t *zilog, uint64_t foid)
1636 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1639 /* move the async itxs for the foid to the sync queues */
1640 zil_async_to_sync(zilog, foid);
1642 mutex_enter(&zilog->zl_lock);
1643 mybatch = zilog->zl_next_batch;
1644 while (zilog->zl_writer) {
1645 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1646 if (mybatch <= zilog->zl_com_batch) {
1647 mutex_exit(&zilog->zl_lock);
1652 zilog->zl_next_batch++;
1653 zilog->zl_writer = B_TRUE;
1654 zil_commit_writer(zilog);
1655 zilog->zl_com_batch = mybatch;
1656 zilog->zl_writer = B_FALSE;
1657 mutex_exit(&zilog->zl_lock);
1659 /* wake up one thread to become the next writer */
1660 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1662 /* wake up all threads waiting for this batch to be committed */
1663 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1667 * Called in syncing context to free committed log blocks and update log header.
1670 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1672 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1673 uint64_t txg = dmu_tx_get_txg(tx);
1674 spa_t *spa = zilog->zl_spa;
1675 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1679 * We don't zero out zl_destroy_txg, so make sure we don't try
1680 * to destroy it twice.
1682 if (spa_sync_pass(spa) != 1)
1685 mutex_enter(&zilog->zl_lock);
1687 ASSERT(zilog->zl_stop_sync == 0);
1689 if (*replayed_seq != 0) {
1690 ASSERT(zh->zh_replay_seq < *replayed_seq);
1691 zh->zh_replay_seq = *replayed_seq;
1695 if (zilog->zl_destroy_txg == txg) {
1696 blkptr_t blk = zh->zh_log;
1698 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1700 bzero(zh, sizeof (zil_header_t));
1701 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1703 if (zilog->zl_keep_first) {
1705 * If this block was part of log chain that couldn't
1706 * be claimed because a device was missing during
1707 * zil_claim(), but that device later returns,
1708 * then this block could erroneously appear valid.
1709 * To guard against this, assign a new GUID to the new
1710 * log chain so it doesn't matter what blk points to.
1712 zil_init_log_chain(zilog, &blk);
1717 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1718 zh->zh_log = lwb->lwb_blk;
1719 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1721 list_remove(&zilog->zl_lwb_list, lwb);
1722 zio_free_zil(spa, txg, &lwb->lwb_blk);
1723 kmem_cache_free(zil_lwb_cache, lwb);
1726 * If we don't have anything left in the lwb list then
1727 * we've had an allocation failure and we need to zero
1728 * out the zil_header blkptr so that we don't end
1729 * up freeing the same block twice.
1731 if (list_head(&zilog->zl_lwb_list) == NULL)
1732 BP_ZERO(&zh->zh_log);
1734 mutex_exit(&zilog->zl_lock);
1740 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1741 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1747 kmem_cache_destroy(zil_lwb_cache);
1751 zil_set_sync(zilog_t *zilog, uint64_t sync)
1753 zilog->zl_sync = sync;
1757 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1759 zilog->zl_logbias = logbias;
1763 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1767 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1769 zilog->zl_header = zh_phys;
1771 zilog->zl_spa = dmu_objset_spa(os);
1772 zilog->zl_dmu_pool = dmu_objset_pool(os);
1773 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1774 zilog->zl_logbias = dmu_objset_logbias(os);
1775 zilog->zl_sync = dmu_objset_syncprop(os);
1776 zilog->zl_next_batch = 1;
1778 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1780 for (int i = 0; i < TXG_SIZE; i++) {
1781 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1782 MUTEX_DEFAULT, NULL);
1785 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1786 offsetof(lwb_t, lwb_node));
1788 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1789 offsetof(itx_t, itx_node));
1791 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1793 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1794 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1796 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1797 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1798 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1799 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1805 zil_free(zilog_t *zilog)
1807 zilog->zl_stop_sync = 1;
1809 ASSERT0(zilog->zl_suspend);
1810 ASSERT0(zilog->zl_suspending);
1812 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1813 list_destroy(&zilog->zl_lwb_list);
1815 avl_destroy(&zilog->zl_vdev_tree);
1816 mutex_destroy(&zilog->zl_vdev_lock);
1818 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1819 list_destroy(&zilog->zl_itx_commit_list);
1821 for (int i = 0; i < TXG_SIZE; i++) {
1823 * It's possible for an itx to be generated that doesn't dirty
1824 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1825 * callback to remove the entry. We remove those here.
1827 * Also free up the ziltest itxs.
1829 if (zilog->zl_itxg[i].itxg_itxs)
1830 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1831 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1834 mutex_destroy(&zilog->zl_lock);
1836 cv_destroy(&zilog->zl_cv_writer);
1837 cv_destroy(&zilog->zl_cv_suspend);
1838 cv_destroy(&zilog->zl_cv_batch[0]);
1839 cv_destroy(&zilog->zl_cv_batch[1]);
1841 kmem_free(zilog, sizeof (zilog_t));
1845 * Open an intent log.
1848 zil_open(objset_t *os, zil_get_data_t *get_data)
1850 zilog_t *zilog = dmu_objset_zil(os);
1852 ASSERT(zilog->zl_get_data == NULL);
1853 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1855 zilog->zl_get_data = get_data;
1861 * Close an intent log.
1864 zil_close(zilog_t *zilog)
1869 zil_commit(zilog, 0); /* commit all itx */
1872 * The lwb_max_txg for the stubby lwb will reflect the last activity
1873 * for the zil. After a txg_wait_synced() on the txg we know all the
1874 * callbacks have occurred that may clean the zil. Only then can we
1875 * destroy the zl_clean_taskq.
1877 mutex_enter(&zilog->zl_lock);
1878 lwb = list_tail(&zilog->zl_lwb_list);
1880 txg = lwb->lwb_max_txg;
1881 mutex_exit(&zilog->zl_lock);
1883 txg_wait_synced(zilog->zl_dmu_pool, txg);
1885 if (zilog_is_dirty(zilog))
1886 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1887 VERIFY(!zilog_is_dirty(zilog));
1889 zilog->zl_get_data = NULL;
1892 * We should have only one LWB left on the list; remove it now.
1894 mutex_enter(&zilog->zl_lock);
1895 lwb = list_head(&zilog->zl_lwb_list);
1897 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1898 list_remove(&zilog->zl_lwb_list, lwb);
1899 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1900 kmem_cache_free(zil_lwb_cache, lwb);
1902 mutex_exit(&zilog->zl_lock);
1905 static char *suspend_tag = "zil suspending";
1908 * Suspend an intent log. While in suspended mode, we still honor
1909 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1910 * On old version pools, we suspend the log briefly when taking a
1911 * snapshot so that it will have an empty intent log.
1913 * Long holds are not really intended to be used the way we do here --
1914 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1915 * could fail. Therefore we take pains to only put a long hold if it is
1916 * actually necessary. Fortunately, it will only be necessary if the
1917 * objset is currently mounted (or the ZVOL equivalent). In that case it
1918 * will already have a long hold, so we are not really making things any worse.
1920 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1921 * zvol_state_t), and use their mechanism to prevent their hold from being
1922 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1925 * if cookiep == NULL, this does both the suspend & resume.
1926 * Otherwise, it returns with the dataset "long held", and the cookie
1927 * should be passed into zil_resume().
1930 zil_suspend(const char *osname, void **cookiep)
1934 const zil_header_t *zh;
1937 error = dmu_objset_hold(osname, suspend_tag, &os);
1940 zilog = dmu_objset_zil(os);
1942 mutex_enter(&zilog->zl_lock);
1943 zh = zilog->zl_header;
1945 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1946 mutex_exit(&zilog->zl_lock);
1947 dmu_objset_rele(os, suspend_tag);
1948 return (SET_ERROR(EBUSY));
1952 * Don't put a long hold in the cases where we can avoid it. This
1953 * is when there is no cookie so we are doing a suspend & resume
1954 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1955 * for the suspend because it's already suspended, or there's no ZIL.
1957 if (cookiep == NULL && !zilog->zl_suspending &&
1958 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1959 mutex_exit(&zilog->zl_lock);
1960 dmu_objset_rele(os, suspend_tag);
1964 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1965 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1967 zilog->zl_suspend++;
1969 if (zilog->zl_suspend > 1) {
1971 * Someone else is already suspending it.
1972 * Just wait for them to finish.
1975 while (zilog->zl_suspending)
1976 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1977 mutex_exit(&zilog->zl_lock);
1979 if (cookiep == NULL)
1987 * If there is no pointer to an on-disk block, this ZIL must not
1988 * be active (e.g. filesystem not mounted), so there's nothing
1991 if (BP_IS_HOLE(&zh->zh_log)) {
1992 ASSERT(cookiep != NULL); /* fast path already handled */
1995 mutex_exit(&zilog->zl_lock);
1999 zilog->zl_suspending = B_TRUE;
2000 mutex_exit(&zilog->zl_lock);
2002 zil_commit(zilog, 0);
2004 zil_destroy(zilog, B_FALSE);
2006 mutex_enter(&zilog->zl_lock);
2007 zilog->zl_suspending = B_FALSE;
2008 cv_broadcast(&zilog->zl_cv_suspend);
2009 mutex_exit(&zilog->zl_lock);
2011 if (cookiep == NULL)
2019 zil_resume(void *cookie)
2021 objset_t *os = cookie;
2022 zilog_t *zilog = dmu_objset_zil(os);
2024 mutex_enter(&zilog->zl_lock);
2025 ASSERT(zilog->zl_suspend != 0);
2026 zilog->zl_suspend--;
2027 mutex_exit(&zilog->zl_lock);
2028 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2029 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2032 typedef struct zil_replay_arg {
2033 zil_replay_func_t **zr_replay;
2035 boolean_t zr_byteswap;
2040 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2042 char name[ZFS_MAX_DATASET_NAME_LEN];
2044 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2046 dmu_objset_name(zilog->zl_os, name);
2048 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2049 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2050 (u_longlong_t)lr->lrc_seq,
2051 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2052 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2058 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2060 zil_replay_arg_t *zr = zra;
2061 const zil_header_t *zh = zilog->zl_header;
2062 uint64_t reclen = lr->lrc_reclen;
2063 uint64_t txtype = lr->lrc_txtype;
2066 zilog->zl_replaying_seq = lr->lrc_seq;
2068 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2071 if (lr->lrc_txg < claim_txg) /* already committed */
2074 /* Strip case-insensitive bit, still present in log record */
2077 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2078 return (zil_replay_error(zilog, lr, EINVAL));
2081 * If this record type can be logged out of order, the object
2082 * (lr_foid) may no longer exist. That's legitimate, not an error.
2084 if (TX_OOO(txtype)) {
2085 error = dmu_object_info(zilog->zl_os,
2086 ((lr_ooo_t *)lr)->lr_foid, NULL);
2087 if (error == ENOENT || error == EEXIST)
2092 * Make a copy of the data so we can revise and extend it.
2094 bcopy(lr, zr->zr_lr, reclen);
2097 * If this is a TX_WRITE with a blkptr, suck in the data.
2099 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2100 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2101 zr->zr_lr + reclen);
2103 return (zil_replay_error(zilog, lr, error));
2107 * The log block containing this lr may have been byteswapped
2108 * so that we can easily examine common fields like lrc_txtype.
2109 * However, the log is a mix of different record types, and only the
2110 * replay vectors know how to byteswap their records. Therefore, if
2111 * the lr was byteswapped, undo it before invoking the replay vector.
2113 if (zr->zr_byteswap)
2114 byteswap_uint64_array(zr->zr_lr, reclen);
2117 * We must now do two things atomically: replay this log record,
2118 * and update the log header sequence number to reflect the fact that
2119 * we did so. At the end of each replay function the sequence number
2120 * is updated if we are in replay mode.
2122 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2125 * The DMU's dnode layer doesn't see removes until the txg
2126 * commits, so a subsequent claim can spuriously fail with
2127 * EEXIST. So if we receive any error we try syncing out
2128 * any removes then retry the transaction. Note that we
2129 * specify B_FALSE for byteswap now, so we don't do it twice.
2131 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2132 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2134 return (zil_replay_error(zilog, lr, error));
2141 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2143 zilog->zl_replay_blks++;
2149 * If this dataset has a non-empty intent log, replay it and destroy it.
2152 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2154 zilog_t *zilog = dmu_objset_zil(os);
2155 const zil_header_t *zh = zilog->zl_header;
2156 zil_replay_arg_t zr;
2158 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2159 zil_destroy(zilog, B_TRUE);
2163 zr.zr_replay = replay_func;
2165 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2166 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2169 * Wait for in-progress removes to sync before starting replay.
2171 txg_wait_synced(zilog->zl_dmu_pool, 0);
2173 zilog->zl_replay = B_TRUE;
2174 zilog->zl_replay_time = ddi_get_lbolt();
2175 ASSERT(zilog->zl_replay_blks == 0);
2176 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2178 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2180 zil_destroy(zilog, B_FALSE);
2181 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2182 zilog->zl_replay = B_FALSE;
2186 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2188 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2191 if (zilog->zl_replay) {
2192 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2193 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2194 zilog->zl_replaying_seq;
2203 zil_vdev_offline(const char *osname, void *arg)
2207 error = zil_suspend(osname, NULL);
2209 return (SET_ERROR(EEXIST));