4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
31 #include <sys/spa_impl.h>
36 #include <sys/resource.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
46 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
47 * calls that change the file system. Each itx has enough information to
48 * be able to replay them after a system crash, power loss, or
49 * equivalent failure mode. These are stored in memory until either:
51 * 1. they are committed to the pool by the DMU transaction group
52 * (txg), at which point they can be discarded; or
53 * 2. they are committed to the on-disk ZIL for the dataset being
54 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
57 * In the event of a crash or power loss, the itxs contained by each
58 * dataset's on-disk ZIL will be replayed when that dataset is first
59 * instantianted (e.g. if the dataset is a normal fileystem, when it is
62 * As hinted at above, there is one ZIL per dataset (both the in-memory
63 * representation, and the on-disk representation). The on-disk format
64 * consists of 3 parts:
66 * - a single, per-dataset, ZIL header; which points to a chain of
67 * - zero or more ZIL blocks; each of which contains
68 * - zero or more ZIL records
70 * A ZIL record holds the information necessary to replay a single
71 * system call transaction. A ZIL block can hold many ZIL records, and
72 * the blocks are chained together, similarly to a singly linked list.
74 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
75 * block in the chain, and the ZIL header points to the first block in
78 * Note, there is not a fixed place in the pool to hold these ZIL
79 * blocks; they are dynamically allocated and freed as needed from the
80 * blocks available on the pool, though they can be preferentially
81 * allocated from a dedicated "log" vdev.
85 * This controls the amount of time that a ZIL block (lwb) will remain
86 * "open" when it isn't "full", and it has a thread waiting for it to be
87 * committed to stable storage. Please refer to the zil_commit_waiter()
88 * function (and the comments within it) for more details.
90 int zfs_commit_timeout_pct = 5;
93 * Disable intent logging replay. This global ZIL switch affects all pools.
95 int zil_replay_disable = 0;
96 SYSCTL_DECL(_vfs_zfs);
97 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
98 &zil_replay_disable, 0, "Disable intent logging replay");
101 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
102 * the disk(s) by the ZIL after an LWB write has completed. Setting this
103 * will cause ZIL corruption on power loss if a volatile out-of-order
104 * write cache is enabled.
106 boolean_t zil_nocacheflush = B_FALSE;
107 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_nocacheflush, CTLFLAG_RWTUN,
108 &zil_nocacheflush, 0, "Disable ZIL cache flush");
110 boolean_t zfs_trim_enabled = B_TRUE;
111 SYSCTL_DECL(_vfs_zfs_trim);
112 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
116 * Limit SLOG write size per commit executed with synchronous priority.
117 * Any writes above that will be executed with lower (asynchronous) priority
118 * to limit potential SLOG device abuse by single active ZIL writer.
120 uint64_t zil_slog_bulk = 768 * 1024;
121 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_bulk, CTLFLAG_RWTUN,
122 &zil_slog_bulk, 0, "Maximal SLOG commit size with sync priority");
124 static kmem_cache_t *zil_lwb_cache;
125 static kmem_cache_t *zil_zcw_cache;
127 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
128 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
131 zil_bp_compare(const void *x1, const void *x2)
133 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
134 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
136 int cmp = AVL_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
140 return (AVL_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
144 zil_bp_tree_init(zilog_t *zilog)
146 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
147 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
151 zil_bp_tree_fini(zilog_t *zilog)
153 avl_tree_t *t = &zilog->zl_bp_tree;
157 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
158 kmem_free(zn, sizeof (zil_bp_node_t));
164 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
166 avl_tree_t *t = &zilog->zl_bp_tree;
171 if (BP_IS_EMBEDDED(bp))
174 dva = BP_IDENTITY(bp);
176 if (avl_find(t, dva, &where) != NULL)
177 return (SET_ERROR(EEXIST));
179 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
181 avl_insert(t, zn, where);
186 static zil_header_t *
187 zil_header_in_syncing_context(zilog_t *zilog)
189 return ((zil_header_t *)zilog->zl_header);
193 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
195 zio_cksum_t *zc = &bp->blk_cksum;
197 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
198 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
199 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
200 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
204 * Read a log block and make sure it's valid.
207 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
210 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
211 arc_flags_t aflags = ARC_FLAG_WAIT;
212 arc_buf_t *abuf = NULL;
216 if (zilog->zl_header->zh_claim_txg == 0)
217 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
219 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
220 zio_flags |= ZIO_FLAG_SPECULATIVE;
222 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
223 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
225 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
226 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
229 zio_cksum_t cksum = bp->blk_cksum;
232 * Validate the checksummed log block.
234 * Sequence numbers should be... sequential. The checksum
235 * verifier for the next block should be bp's checksum plus 1.
237 * Also check the log chain linkage and size used.
239 cksum.zc_word[ZIL_ZC_SEQ]++;
241 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
242 zil_chain_t *zilc = abuf->b_data;
243 char *lr = (char *)(zilc + 1);
244 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
246 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
247 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
248 error = SET_ERROR(ECKSUM);
250 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
252 *end = (char *)dst + len;
253 *nbp = zilc->zc_next_blk;
256 char *lr = abuf->b_data;
257 uint64_t size = BP_GET_LSIZE(bp);
258 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
260 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
261 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
262 (zilc->zc_nused > (size - sizeof (*zilc)))) {
263 error = SET_ERROR(ECKSUM);
265 ASSERT3U(zilc->zc_nused, <=,
266 SPA_OLD_MAXBLOCKSIZE);
267 bcopy(lr, dst, zilc->zc_nused);
268 *end = (char *)dst + zilc->zc_nused;
269 *nbp = zilc->zc_next_blk;
273 arc_buf_destroy(abuf, &abuf);
280 * Read a TX_WRITE log data block.
283 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
285 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
286 const blkptr_t *bp = &lr->lr_blkptr;
287 arc_flags_t aflags = ARC_FLAG_WAIT;
288 arc_buf_t *abuf = NULL;
292 if (BP_IS_HOLE(bp)) {
294 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
298 if (zilog->zl_header->zh_claim_txg == 0)
299 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
301 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
302 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
304 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
305 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
309 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
310 arc_buf_destroy(abuf, &abuf);
317 * Parse the intent log, and call parse_func for each valid record within.
320 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
321 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
323 const zil_header_t *zh = zilog->zl_header;
324 boolean_t claimed = !!zh->zh_claim_txg;
325 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
326 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
327 uint64_t max_blk_seq = 0;
328 uint64_t max_lr_seq = 0;
329 uint64_t blk_count = 0;
330 uint64_t lr_count = 0;
331 blkptr_t blk, next_blk;
336 * Old logs didn't record the maximum zh_claim_lr_seq.
338 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
339 claim_lr_seq = UINT64_MAX;
342 * Starting at the block pointed to by zh_log we read the log chain.
343 * For each block in the chain we strongly check that block to
344 * ensure its validity. We stop when an invalid block is found.
345 * For each block pointer in the chain we call parse_blk_func().
346 * For each record in each valid block we call parse_lr_func().
347 * If the log has been claimed, stop if we encounter a sequence
348 * number greater than the highest claimed sequence number.
350 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
351 zil_bp_tree_init(zilog);
353 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
354 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
358 if (blk_seq > claim_blk_seq)
360 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
362 ASSERT3U(max_blk_seq, <, blk_seq);
363 max_blk_seq = blk_seq;
366 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
369 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
373 for (lrp = lrbuf; lrp < end; lrp += reclen) {
374 lr_t *lr = (lr_t *)lrp;
375 reclen = lr->lrc_reclen;
376 ASSERT3U(reclen, >=, sizeof (lr_t));
377 if (lr->lrc_seq > claim_lr_seq)
379 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
381 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
382 max_lr_seq = lr->lrc_seq;
387 zilog->zl_parse_error = error;
388 zilog->zl_parse_blk_seq = max_blk_seq;
389 zilog->zl_parse_lr_seq = max_lr_seq;
390 zilog->zl_parse_blk_count = blk_count;
391 zilog->zl_parse_lr_count = lr_count;
393 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
394 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
396 zil_bp_tree_fini(zilog);
397 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
404 zil_clear_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
406 ASSERT(!BP_IS_HOLE(bp));
409 * As we call this function from the context of a rewind to a
410 * checkpoint, each ZIL block whose txg is later than the txg
411 * that we rewind to is invalid. Thus, we return -1 so
412 * zil_parse() doesn't attempt to read it.
414 if (bp->blk_birth >= first_txg)
417 if (zil_bp_tree_add(zilog, bp) != 0)
420 zio_free(zilog->zl_spa, first_txg, bp);
426 zil_noop_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
432 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
435 * Claim log block if not already committed and not already claimed.
436 * If tx == NULL, just verify that the block is claimable.
438 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
439 zil_bp_tree_add(zilog, bp) != 0)
442 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
443 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
444 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
448 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
450 lr_write_t *lr = (lr_write_t *)lrc;
453 if (lrc->lrc_txtype != TX_WRITE)
457 * If the block is not readable, don't claim it. This can happen
458 * in normal operation when a log block is written to disk before
459 * some of the dmu_sync() blocks it points to. In this case, the
460 * transaction cannot have been committed to anyone (we would have
461 * waited for all writes to be stable first), so it is semantically
462 * correct to declare this the end of the log.
464 if (lr->lr_blkptr.blk_birth >= first_txg &&
465 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
467 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
472 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
474 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
480 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
482 lr_write_t *lr = (lr_write_t *)lrc;
483 blkptr_t *bp = &lr->lr_blkptr;
486 * If we previously claimed it, we need to free it.
488 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
489 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
491 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
497 zil_lwb_vdev_compare(const void *x1, const void *x2)
499 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
500 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
502 return (AVL_CMP(v1, v2));
506 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
510 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
511 lwb->lwb_zilog = zilog;
513 lwb->lwb_slog = slog;
514 lwb->lwb_state = LWB_STATE_CLOSED;
515 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
516 lwb->lwb_max_txg = txg;
517 lwb->lwb_write_zio = NULL;
518 lwb->lwb_root_zio = NULL;
520 lwb->lwb_issued_timestamp = 0;
521 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
522 lwb->lwb_nused = sizeof (zil_chain_t);
523 lwb->lwb_sz = BP_GET_LSIZE(bp);
526 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
529 mutex_enter(&zilog->zl_lock);
530 list_insert_tail(&zilog->zl_lwb_list, lwb);
531 mutex_exit(&zilog->zl_lock);
533 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
534 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
535 VERIFY(list_is_empty(&lwb->lwb_waiters));
541 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
543 ASSERT(MUTEX_HELD(&zilog->zl_lock));
544 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
545 VERIFY(list_is_empty(&lwb->lwb_waiters));
546 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
547 ASSERT3P(lwb->lwb_write_zio, ==, NULL);
548 ASSERT3P(lwb->lwb_root_zio, ==, NULL);
549 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
550 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
551 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
554 * Clear the zilog's field to indicate this lwb is no longer
555 * valid, and prevent use-after-free errors.
557 if (zilog->zl_last_lwb_opened == lwb)
558 zilog->zl_last_lwb_opened = NULL;
560 kmem_cache_free(zil_lwb_cache, lwb);
564 * Called when we create in-memory log transactions so that we know
565 * to cleanup the itxs at the end of spa_sync().
568 zilog_dirty(zilog_t *zilog, uint64_t txg)
570 dsl_pool_t *dp = zilog->zl_dmu_pool;
571 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
573 ASSERT(spa_writeable(zilog->zl_spa));
575 if (ds->ds_is_snapshot)
576 panic("dirtying snapshot!");
578 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
579 /* up the hold count until we can be written out */
580 dmu_buf_add_ref(ds->ds_dbuf, zilog);
582 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
587 * Determine if the zil is dirty in the specified txg. Callers wanting to
588 * ensure that the dirty state does not change must hold the itxg_lock for
589 * the specified txg. Holding the lock will ensure that the zil cannot be
590 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
594 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
596 dsl_pool_t *dp = zilog->zl_dmu_pool;
598 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
604 * Determine if the zil is dirty. The zil is considered dirty if it has
605 * any pending itx records that have not been cleaned by zil_clean().
608 zilog_is_dirty(zilog_t *zilog)
610 dsl_pool_t *dp = zilog->zl_dmu_pool;
612 for (int t = 0; t < TXG_SIZE; t++) {
613 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
620 * Create an on-disk intent log.
623 zil_create(zilog_t *zilog)
625 const zil_header_t *zh = zilog->zl_header;
631 boolean_t slog = FALSE;
634 * Wait for any previous destroy to complete.
636 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
638 ASSERT(zh->zh_claim_txg == 0);
639 ASSERT(zh->zh_replay_seq == 0);
644 * Allocate an initial log block if:
645 * - there isn't one already
646 * - the existing block is the wrong endianess
648 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
649 tx = dmu_tx_create(zilog->zl_os);
650 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
651 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
652 txg = dmu_tx_get_txg(tx);
654 if (!BP_IS_HOLE(&blk)) {
655 zio_free(zilog->zl_spa, txg, &blk);
659 error = zio_alloc_zil(zilog->zl_spa,
660 zilog->zl_os->os_dsl_dataset->ds_object, txg, &blk, NULL,
661 ZIL_MIN_BLKSZ, &slog);
664 zil_init_log_chain(zilog, &blk);
668 * Allocate a log write block (lwb) for the first log block.
671 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
674 * If we just allocated the first log block, commit our transaction
675 * and wait for zil_sync() to stuff the block poiner into zh_log.
676 * (zh is part of the MOS, so we cannot modify it in open context.)
680 txg_wait_synced(zilog->zl_dmu_pool, txg);
683 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
689 * In one tx, free all log blocks and clear the log header. If keep_first
690 * is set, then we're replaying a log with no content. We want to keep the
691 * first block, however, so that the first synchronous transaction doesn't
692 * require a txg_wait_synced() in zil_create(). We don't need to
693 * txg_wait_synced() here either when keep_first is set, because both
694 * zil_create() and zil_destroy() will wait for any in-progress destroys
698 zil_destroy(zilog_t *zilog, boolean_t keep_first)
700 const zil_header_t *zh = zilog->zl_header;
706 * Wait for any previous destroy to complete.
708 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
710 zilog->zl_old_header = *zh; /* debugging aid */
712 if (BP_IS_HOLE(&zh->zh_log))
715 tx = dmu_tx_create(zilog->zl_os);
716 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
717 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
718 txg = dmu_tx_get_txg(tx);
720 mutex_enter(&zilog->zl_lock);
722 ASSERT3U(zilog->zl_destroy_txg, <, txg);
723 zilog->zl_destroy_txg = txg;
724 zilog->zl_keep_first = keep_first;
726 if (!list_is_empty(&zilog->zl_lwb_list)) {
727 ASSERT(zh->zh_claim_txg == 0);
729 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
730 list_remove(&zilog->zl_lwb_list, lwb);
731 if (lwb->lwb_buf != NULL)
732 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
733 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
734 zil_free_lwb(zilog, lwb);
736 } else if (!keep_first) {
737 zil_destroy_sync(zilog, tx);
739 mutex_exit(&zilog->zl_lock);
745 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
747 ASSERT(list_is_empty(&zilog->zl_lwb_list));
748 (void) zil_parse(zilog, zil_free_log_block,
749 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
753 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
755 dmu_tx_t *tx = txarg;
762 error = dmu_objset_own_obj(dp, ds->ds_object,
763 DMU_OST_ANY, B_FALSE, FTAG, &os);
766 * EBUSY indicates that the objset is inconsistent, in which
767 * case it can not have a ZIL.
769 if (error != EBUSY) {
770 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
771 (unsigned long long)ds->ds_object, error);
776 zilog = dmu_objset_zil(os);
777 zh = zil_header_in_syncing_context(zilog);
778 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
779 first_txg = spa_min_claim_txg(zilog->zl_spa);
782 * If the spa_log_state is not set to be cleared, check whether
783 * the current uberblock is a checkpoint one and if the current
784 * header has been claimed before moving on.
786 * If the current uberblock is a checkpointed uberblock then
787 * one of the following scenarios took place:
789 * 1] We are currently rewinding to the checkpoint of the pool.
790 * 2] We crashed in the middle of a checkpoint rewind but we
791 * did manage to write the checkpointed uberblock to the
792 * vdev labels, so when we tried to import the pool again
793 * the checkpointed uberblock was selected from the import
796 * In both cases we want to zero out all the ZIL blocks, except
797 * the ones that have been claimed at the time of the checkpoint
798 * (their zh_claim_txg != 0). The reason is that these blocks
799 * may be corrupted since we may have reused their locations on
800 * disk after we took the checkpoint.
802 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
803 * when we first figure out whether the current uberblock is
804 * checkpointed or not. Unfortunately, that would discard all
805 * the logs, including the ones that are claimed, and we would
808 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
809 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
810 zh->zh_claim_txg == 0)) {
811 if (!BP_IS_HOLE(&zh->zh_log)) {
812 (void) zil_parse(zilog, zil_clear_log_block,
813 zil_noop_log_record, tx, first_txg);
815 BP_ZERO(&zh->zh_log);
816 dsl_dataset_dirty(dmu_objset_ds(os), tx);
817 dmu_objset_disown(os, FTAG);
822 * If we are not rewinding and opening the pool normally, then
823 * the min_claim_txg should be equal to the first txg of the pool.
825 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
828 * Claim all log blocks if we haven't already done so, and remember
829 * the highest claimed sequence number. This ensures that if we can
830 * read only part of the log now (e.g. due to a missing device),
831 * but we can read the entire log later, we will not try to replay
832 * or destroy beyond the last block we successfully claimed.
834 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
835 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
836 (void) zil_parse(zilog, zil_claim_log_block,
837 zil_claim_log_record, tx, first_txg);
838 zh->zh_claim_txg = first_txg;
839 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
840 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
841 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
842 zh->zh_flags |= ZIL_REPLAY_NEEDED;
843 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
844 dsl_dataset_dirty(dmu_objset_ds(os), tx);
847 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
848 dmu_objset_disown(os, FTAG);
853 * Check the log by walking the log chain.
854 * Checksum errors are ok as they indicate the end of the chain.
855 * Any other error (no device or read failure) returns an error.
859 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
868 error = dmu_objset_from_ds(ds, &os);
870 cmn_err(CE_WARN, "can't open objset %llu, error %d",
871 (unsigned long long)ds->ds_object, error);
875 zilog = dmu_objset_zil(os);
876 bp = (blkptr_t *)&zilog->zl_header->zh_log;
878 if (!BP_IS_HOLE(bp)) {
880 boolean_t valid = B_TRUE;
883 * Check the first block and determine if it's on a log device
884 * which may have been removed or faulted prior to loading this
885 * pool. If so, there's no point in checking the rest of the
886 * log as its content should have already been synced to the
889 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
890 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
891 if (vd->vdev_islog && vdev_is_dead(vd))
892 valid = vdev_log_state_valid(vd);
893 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
899 * Check whether the current uberblock is checkpointed (e.g.
900 * we are rewinding) and whether the current header has been
901 * claimed or not. If it hasn't then skip verifying it. We
902 * do this because its ZIL blocks may be part of the pool's
903 * state before the rewind, which is no longer valid.
905 zil_header_t *zh = zil_header_in_syncing_context(zilog);
906 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
907 zh->zh_claim_txg == 0)
912 * Because tx == NULL, zil_claim_log_block() will not actually claim
913 * any blocks, but just determine whether it is possible to do so.
914 * In addition to checking the log chain, zil_claim_log_block()
915 * will invoke zio_claim() with a done func of spa_claim_notify(),
916 * which will update spa_max_claim_txg. See spa_load() for details.
918 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
919 zilog->zl_header->zh_claim_txg ? -1ULL :
920 spa_min_claim_txg(os->os_spa));
922 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
926 * When an itx is "skipped", this function is used to properly mark the
927 * waiter as "done, and signal any thread(s) waiting on it. An itx can
928 * be skipped (and not committed to an lwb) for a variety of reasons,
929 * one of them being that the itx was committed via spa_sync(), prior to
930 * it being committed to an lwb; this can happen if a thread calling
931 * zil_commit() is racing with spa_sync().
934 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
936 mutex_enter(&zcw->zcw_lock);
937 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
938 zcw->zcw_done = B_TRUE;
939 cv_broadcast(&zcw->zcw_cv);
940 mutex_exit(&zcw->zcw_lock);
944 * This function is used when the given waiter is to be linked into an
945 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
946 * At this point, the waiter will no longer be referenced by the itx,
947 * and instead, will be referenced by the lwb.
950 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
953 * The lwb_waiters field of the lwb is protected by the zilog's
954 * zl_lock, thus it must be held when calling this function.
956 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
958 mutex_enter(&zcw->zcw_lock);
959 ASSERT(!list_link_active(&zcw->zcw_node));
960 ASSERT3P(zcw->zcw_lwb, ==, NULL);
961 ASSERT3P(lwb, !=, NULL);
962 ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
963 lwb->lwb_state == LWB_STATE_ISSUED ||
964 lwb->lwb_state == LWB_STATE_WRITE_DONE);
966 list_insert_tail(&lwb->lwb_waiters, zcw);
968 mutex_exit(&zcw->zcw_lock);
972 * This function is used when zio_alloc_zil() fails to allocate a ZIL
973 * block, and the given waiter must be linked to the "nolwb waiters"
974 * list inside of zil_process_commit_list().
977 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
979 mutex_enter(&zcw->zcw_lock);
980 ASSERT(!list_link_active(&zcw->zcw_node));
981 ASSERT3P(zcw->zcw_lwb, ==, NULL);
982 list_insert_tail(nolwb, zcw);
983 mutex_exit(&zcw->zcw_lock);
987 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
989 avl_tree_t *t = &lwb->lwb_vdev_tree;
991 zil_vdev_node_t *zv, zvsearch;
992 int ndvas = BP_GET_NDVAS(bp);
995 if (zil_nocacheflush)
998 mutex_enter(&lwb->lwb_vdev_lock);
999 for (i = 0; i < ndvas; i++) {
1000 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1001 if (avl_find(t, &zvsearch, &where) == NULL) {
1002 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1003 zv->zv_vdev = zvsearch.zv_vdev;
1004 avl_insert(t, zv, where);
1007 mutex_exit(&lwb->lwb_vdev_lock);
1011 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1013 avl_tree_t *src = &lwb->lwb_vdev_tree;
1014 avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1015 void *cookie = NULL;
1016 zil_vdev_node_t *zv;
1018 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1019 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1020 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1023 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1024 * not need the protection of lwb_vdev_lock (it will only be modified
1025 * while holding zilog->zl_lock) as its writes and those of its
1026 * children have all completed. The younger 'nlwb' may be waiting on
1027 * future writes to additional vdevs.
1029 mutex_enter(&nlwb->lwb_vdev_lock);
1031 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1032 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1034 while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1037 if (avl_find(dst, zv, &where) == NULL) {
1038 avl_insert(dst, zv, where);
1040 kmem_free(zv, sizeof (*zv));
1043 mutex_exit(&nlwb->lwb_vdev_lock);
1047 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1049 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1053 * This function is a called after all vdevs associated with a given lwb
1054 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1055 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1056 * all "previous" lwb's will have completed before this function is
1057 * called; i.e. this function is called for all previous lwbs before
1058 * it's called for "this" lwb (enforced via zio the dependencies
1059 * configured in zil_lwb_set_zio_dependency()).
1061 * The intention is for this function to be called as soon as the
1062 * contents of an lwb are considered "stable" on disk, and will survive
1063 * any sudden loss of power. At this point, any threads waiting for the
1064 * lwb to reach this state are signalled, and the "waiter" structures
1065 * are marked "done".
1068 zil_lwb_flush_vdevs_done(zio_t *zio)
1070 lwb_t *lwb = zio->io_private;
1071 zilog_t *zilog = lwb->lwb_zilog;
1072 dmu_tx_t *tx = lwb->lwb_tx;
1073 zil_commit_waiter_t *zcw;
1075 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1077 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1079 mutex_enter(&zilog->zl_lock);
1082 * Ensure the lwb buffer pointer is cleared before releasing the
1083 * txg. If we have had an allocation failure and the txg is
1084 * waiting to sync then we want zil_sync() to remove the lwb so
1085 * that it's not picked up as the next new one in
1086 * zil_process_commit_list(). zil_sync() will only remove the
1087 * lwb if lwb_buf is null.
1089 lwb->lwb_buf = NULL;
1092 ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1093 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1095 lwb->lwb_root_zio = NULL;
1097 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1098 lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1100 if (zilog->zl_last_lwb_opened == lwb) {
1102 * Remember the highest committed log sequence number
1103 * for ztest. We only update this value when all the log
1104 * writes succeeded, because ztest wants to ASSERT that
1105 * it got the whole log chain.
1107 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1110 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1111 mutex_enter(&zcw->zcw_lock);
1113 ASSERT(list_link_active(&zcw->zcw_node));
1114 list_remove(&lwb->lwb_waiters, zcw);
1116 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1117 zcw->zcw_lwb = NULL;
1119 zcw->zcw_zio_error = zio->io_error;
1121 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1122 zcw->zcw_done = B_TRUE;
1123 cv_broadcast(&zcw->zcw_cv);
1125 mutex_exit(&zcw->zcw_lock);
1128 mutex_exit(&zilog->zl_lock);
1131 * Now that we've written this log block, we have a stable pointer
1132 * to the next block in the chain, so it's OK to let the txg in
1133 * which we allocated the next block sync.
1139 * This is called when an lwb's write zio completes. The callback's
1140 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1141 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1142 * in writing out this specific lwb's data, and in the case that cache
1143 * flushes have been deferred, vdevs involved in writing the data for
1144 * previous lwbs. The writes corresponding to all the vdevs in the
1145 * lwb_vdev_tree will have completed by the time this is called, due to
1146 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1147 * which takes deferred flushes into account. The lwb will be "done"
1148 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1149 * completion callback for the lwb's root zio.
1152 zil_lwb_write_done(zio_t *zio)
1154 lwb_t *lwb = zio->io_private;
1155 spa_t *spa = zio->io_spa;
1156 zilog_t *zilog = lwb->lwb_zilog;
1157 avl_tree_t *t = &lwb->lwb_vdev_tree;
1158 void *cookie = NULL;
1159 zil_vdev_node_t *zv;
1162 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1164 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1165 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1166 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1167 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1168 ASSERT(!BP_IS_GANG(zio->io_bp));
1169 ASSERT(!BP_IS_HOLE(zio->io_bp));
1170 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1172 abd_put(zio->io_abd);
1174 mutex_enter(&zilog->zl_lock);
1175 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1176 lwb->lwb_state = LWB_STATE_WRITE_DONE;
1177 lwb->lwb_write_zio = NULL;
1178 nlwb = list_next(&zilog->zl_lwb_list, lwb);
1179 mutex_exit(&zilog->zl_lock);
1181 if (avl_numnodes(t) == 0)
1185 * If there was an IO error, we're not going to call zio_flush()
1186 * on these vdevs, so we simply empty the tree and free the
1187 * nodes. We avoid calling zio_flush() since there isn't any
1188 * good reason for doing so, after the lwb block failed to be
1191 if (zio->io_error != 0) {
1192 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1193 kmem_free(zv, sizeof (*zv));
1198 * If this lwb does not have any threads waiting for it to
1199 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1200 * command to the vdevs written to by "this" lwb, and instead
1201 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1202 * command for those vdevs. Thus, we merge the vdev tree of
1203 * "this" lwb with the vdev tree of the "next" lwb in the list,
1204 * and assume the "next" lwb will handle flushing the vdevs (or
1205 * deferring the flush(s) again).
1207 * This is a useful performance optimization, especially for
1208 * workloads with lots of async write activity and few sync
1209 * write and/or fsync activity, as it has the potential to
1210 * coalesce multiple flush commands to a vdev into one.
1212 if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) {
1213 zil_lwb_flush_defer(lwb, nlwb);
1214 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1218 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1219 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1221 zio_flush(lwb->lwb_root_zio, vd);
1222 kmem_free(zv, sizeof (*zv));
1227 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1229 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1231 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1232 ASSERT(MUTEX_HELD(&zilog->zl_lock));
1235 * The zilog's "zl_last_lwb_opened" field is used to build the
1236 * lwb/zio dependency chain, which is used to preserve the
1237 * ordering of lwb completions that is required by the semantics
1238 * of the ZIL. Each new lwb zio becomes a parent of the
1239 * "previous" lwb zio, such that the new lwb's zio cannot
1240 * complete until the "previous" lwb's zio completes.
1242 * This is required by the semantics of zil_commit(); the commit
1243 * waiters attached to the lwbs will be woken in the lwb zio's
1244 * completion callback, so this zio dependency graph ensures the
1245 * waiters are woken in the correct order (the same order the
1246 * lwbs were created).
1248 if (last_lwb_opened != NULL &&
1249 last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1250 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1251 last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1252 last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1254 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1255 zio_add_child(lwb->lwb_root_zio,
1256 last_lwb_opened->lwb_root_zio);
1259 * If the previous lwb's write hasn't already completed,
1260 * we also want to order the completion of the lwb write
1261 * zios (above, we only order the completion of the lwb
1262 * root zios). This is required because of how we can
1263 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1265 * When the DKIOCFLUSHWRITECACHE commands are defered,
1266 * the previous lwb will rely on this lwb to flush the
1267 * vdevs written to by that previous lwb. Thus, we need
1268 * to ensure this lwb doesn't issue the flush until
1269 * after the previous lwb's write completes. We ensure
1270 * this ordering by setting the zio parent/child
1271 * relationship here.
1273 * Without this relationship on the lwb's write zio,
1274 * it's possible for this lwb's write to complete prior
1275 * to the previous lwb's write completing; and thus, the
1276 * vdevs for the previous lwb would be flushed prior to
1277 * that lwb's data being written to those vdevs (the
1278 * vdevs are flushed in the lwb write zio's completion
1279 * handler, zil_lwb_write_done()).
1281 if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1282 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1283 last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1285 ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1286 zio_add_child(lwb->lwb_write_zio,
1287 last_lwb_opened->lwb_write_zio);
1294 * This function's purpose is to "open" an lwb such that it is ready to
1295 * accept new itxs being committed to it. To do this, the lwb's zio
1296 * structures are created, and linked to the lwb. This function is
1297 * idempotent; if the passed in lwb has already been opened, this
1298 * function is essentially a no-op.
1301 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1303 zbookmark_phys_t zb;
1304 zio_priority_t prio;
1306 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1307 ASSERT3P(lwb, !=, NULL);
1308 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1309 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1311 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1312 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1313 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1315 if (lwb->lwb_root_zio == NULL) {
1316 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1317 BP_GET_LSIZE(&lwb->lwb_blk));
1319 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1320 prio = ZIO_PRIORITY_SYNC_WRITE;
1322 prio = ZIO_PRIORITY_ASYNC_WRITE;
1324 lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1325 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1326 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1328 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1329 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1330 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1331 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1332 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1334 lwb->lwb_state = LWB_STATE_OPENED;
1336 mutex_enter(&zilog->zl_lock);
1337 zil_lwb_set_zio_dependency(zilog, lwb);
1338 zilog->zl_last_lwb_opened = lwb;
1339 mutex_exit(&zilog->zl_lock);
1342 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1343 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1344 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1348 * Define a limited set of intent log block sizes.
1350 * These must be a multiple of 4KB. Note only the amount used (again
1351 * aligned to 4KB) actually gets written. However, we can't always just
1352 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1357 } zil_block_buckets[] = {
1358 { 4096, 4096 }, /* non TX_WRITE */
1359 { 8192 + 4096, 8192 + 4096 }, /* database */
1360 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */
1361 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */
1362 { 131072, 131072 }, /* < 128KB writes */
1363 { 131072 + 4096, 65536 + 4096 }, /* 128KB writes */
1364 { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */
1368 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1369 * initialized. Otherwise this should not be used directly; see
1370 * zl_max_block_size instead.
1372 int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1373 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_maxblocksize, CTLFLAG_RWTUN,
1374 &zil_maxblocksize, 0, "Limit in bytes of ZIL log block size");
1377 * Start a log block write and advance to the next log block.
1378 * Calls are serialized.
1381 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1385 spa_t *spa = zilog->zl_spa;
1389 uint64_t zil_blksz, wsz;
1393 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1394 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1395 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1396 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1398 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1399 zilc = (zil_chain_t *)lwb->lwb_buf;
1400 bp = &zilc->zc_next_blk;
1402 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1403 bp = &zilc->zc_next_blk;
1406 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1409 * Allocate the next block and save its address in this block
1410 * before writing it in order to establish the log chain.
1411 * Note that if the allocation of nlwb synced before we wrote
1412 * the block that points at it (lwb), we'd leak it if we crashed.
1413 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1414 * We dirty the dataset to ensure that zil_sync() will be called
1415 * to clean up in the event of allocation failure or I/O failure.
1418 tx = dmu_tx_create(zilog->zl_os);
1421 * Since we are not going to create any new dirty data, and we
1422 * can even help with clearing the existing dirty data, we
1423 * should not be subject to the dirty data based delays. We
1424 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1426 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1428 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1429 txg = dmu_tx_get_txg(tx);
1434 * Log blocks are pre-allocated. Here we select the size of the next
1435 * block, based on size used in the last block.
1436 * - first find the smallest bucket that will fit the block from a
1437 * limited set of block sizes. This is because it's faster to write
1438 * blocks allocated from the same metaslab as they are adjacent or
1440 * - next find the maximum from the new suggested size and an array of
1441 * previous sizes. This lessens a picket fence effect of wrongly
1442 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1445 * Note we only write what is used, but we can't just allocate
1446 * the maximum block size because we can exhaust the available
1449 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1450 for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1452 zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1453 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1454 for (i = 0; i < ZIL_PREV_BLKS; i++)
1455 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1456 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1460 /* pass the old blkptr in order to spread log blocks across devs */
1461 error = zio_alloc_zil(spa, zilog->zl_os->os_dsl_dataset->ds_object,
1462 txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1464 ASSERT3U(bp->blk_birth, ==, txg);
1465 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1466 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1469 * Allocate a new log write block (lwb).
1471 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1474 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1475 /* For Slim ZIL only write what is used. */
1476 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1477 ASSERT3U(wsz, <=, lwb->lwb_sz);
1478 zio_shrink(lwb->lwb_write_zio, wsz);
1485 zilc->zc_nused = lwb->lwb_nused;
1486 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1489 * clear unused data for security
1491 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1493 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1495 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1496 lwb->lwb_issued_timestamp = gethrtime();
1497 lwb->lwb_state = LWB_STATE_ISSUED;
1499 zio_nowait(lwb->lwb_root_zio);
1500 zio_nowait(lwb->lwb_write_zio);
1503 * If there was an allocation failure then nlwb will be null which
1504 * forces a txg_wait_synced().
1510 * Maximum amount of write data that can be put into single log block.
1513 zil_max_log_data(zilog_t *zilog)
1515 return (zilog->zl_max_block_size -
1516 sizeof (zil_chain_t) - sizeof (lr_write_t));
1520 * Maximum amount of log space we agree to waste to reduce number of
1521 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1523 static inline uint64_t
1524 zil_max_waste_space(zilog_t *zilog)
1526 return (zil_max_log_data(zilog) / 8);
1530 * Maximum amount of write data for WR_COPIED. For correctness, consumers
1531 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1532 * maximum sized log block, because each WR_COPIED record must fit in a
1533 * single log block. For space efficiency, we want to fit two records into a
1534 * max-sized log block.
1537 zil_max_copied_data(zilog_t *zilog)
1539 return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1540 sizeof (lr_write_t));
1544 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1547 lr_write_t *lrwb, *lrw;
1549 uint64_t dlen, dnow, lwb_sp, reclen, txg, max_log_data;
1551 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1552 ASSERT3P(lwb, !=, NULL);
1553 ASSERT3P(lwb->lwb_buf, !=, NULL);
1555 zil_lwb_write_open(zilog, lwb);
1558 lrw = (lr_write_t *)lrc;
1561 * A commit itx doesn't represent any on-disk state; instead
1562 * it's simply used as a place holder on the commit list, and
1563 * provides a mechanism for attaching a "commit waiter" onto the
1564 * correct lwb (such that the waiter can be signalled upon
1565 * completion of that lwb). Thus, we don't process this itx's
1566 * log record if it's a commit itx (these itx's don't have log
1567 * records), and instead link the itx's waiter onto the lwb's
1570 * For more details, see the comment above zil_commit().
1572 if (lrc->lrc_txtype == TX_COMMIT) {
1573 mutex_enter(&zilog->zl_lock);
1574 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1575 itx->itx_private = NULL;
1576 mutex_exit(&zilog->zl_lock);
1580 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1581 dlen = P2ROUNDUP_TYPED(
1582 lrw->lr_length, sizeof (uint64_t), uint64_t);
1586 reclen = lrc->lrc_reclen;
1587 zilog->zl_cur_used += (reclen + dlen);
1590 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1594 * If this record won't fit in the current log block, start a new one.
1595 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1597 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1598 max_log_data = zil_max_log_data(zilog);
1599 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1600 lwb_sp < zil_max_waste_space(zilog) &&
1601 (dlen % max_log_data == 0 ||
1602 lwb_sp < reclen + dlen % max_log_data))) {
1603 lwb = zil_lwb_write_issue(zilog, lwb);
1606 zil_lwb_write_open(zilog, lwb);
1607 ASSERT(LWB_EMPTY(lwb));
1608 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1611 * There must be enough space in the new, empty log block to
1612 * hold reclen. For WR_COPIED, we need to fit the whole
1613 * record in one block, and reclen is the header size + the
1614 * data size. For WR_NEED_COPY, we can create multiple
1615 * records, splitting the data into multiple blocks, so we
1616 * only need to fit one word of data per block; in this case
1617 * reclen is just the header size (no data).
1619 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1622 dnow = MIN(dlen, lwb_sp - reclen);
1623 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1624 bcopy(lrc, lr_buf, reclen);
1625 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1626 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1629 * If it's a write, fetch the data or get its blkptr as appropriate.
1631 if (lrc->lrc_txtype == TX_WRITE) {
1632 if (txg > spa_freeze_txg(zilog->zl_spa))
1633 txg_wait_synced(zilog->zl_dmu_pool, txg);
1634 if (itx->itx_wr_state != WR_COPIED) {
1638 if (itx->itx_wr_state == WR_NEED_COPY) {
1639 dbuf = lr_buf + reclen;
1640 lrcb->lrc_reclen += dnow;
1641 if (lrwb->lr_length > dnow)
1642 lrwb->lr_length = dnow;
1643 lrw->lr_offset += dnow;
1644 lrw->lr_length -= dnow;
1646 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1651 * We pass in the "lwb_write_zio" rather than
1652 * "lwb_root_zio" so that the "lwb_write_zio"
1653 * becomes the parent of any zio's created by
1654 * the "zl_get_data" callback. The vdevs are
1655 * flushed after the "lwb_write_zio" completes,
1656 * so we want to make sure that completion
1657 * callback waits for these additional zio's,
1658 * such that the vdevs used by those zio's will
1659 * be included in the lwb's vdev tree, and those
1660 * vdevs will be properly flushed. If we passed
1661 * in "lwb_root_zio" here, then these additional
1662 * vdevs may not be flushed; e.g. if these zio's
1663 * completed after "lwb_write_zio" completed.
1665 error = zilog->zl_get_data(itx->itx_private,
1666 lrwb, dbuf, lwb, lwb->lwb_write_zio);
1669 txg_wait_synced(zilog->zl_dmu_pool, txg);
1673 ASSERT(error == ENOENT || error == EEXIST ||
1681 * We're actually making an entry, so update lrc_seq to be the
1682 * log record sequence number. Note that this is generally not
1683 * equal to the itx sequence number because not all transactions
1684 * are synchronous, and sometimes spa_sync() gets there first.
1686 lrcb->lrc_seq = ++zilog->zl_lr_seq;
1687 lwb->lwb_nused += reclen + dnow;
1689 zil_lwb_add_txg(lwb, txg);
1691 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1692 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1696 zilog->zl_cur_used += reclen;
1704 zil_itx_create(uint64_t txtype, size_t lrsize)
1708 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1710 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1711 itx->itx_lr.lrc_txtype = txtype;
1712 itx->itx_lr.lrc_reclen = lrsize;
1713 itx->itx_lr.lrc_seq = 0; /* defensive */
1714 itx->itx_sync = B_TRUE; /* default is synchronous */
1720 zil_itx_destroy(itx_t *itx)
1722 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1726 * Free up the sync and async itxs. The itxs_t has already been detached
1727 * so no locks are needed.
1730 zil_itxg_clean(itxs_t *itxs)
1736 itx_async_node_t *ian;
1738 list = &itxs->i_sync_list;
1739 while ((itx = list_head(list)) != NULL) {
1741 * In the general case, commit itxs will not be found
1742 * here, as they'll be committed to an lwb via
1743 * zil_lwb_commit(), and free'd in that function. Having
1744 * said that, it is still possible for commit itxs to be
1745 * found here, due to the following race:
1747 * - a thread calls zil_commit() which assigns the
1748 * commit itx to a per-txg i_sync_list
1749 * - zil_itxg_clean() is called (e.g. via spa_sync())
1750 * while the waiter is still on the i_sync_list
1752 * There's nothing to prevent syncing the txg while the
1753 * waiter is on the i_sync_list. This normally doesn't
1754 * happen because spa_sync() is slower than zil_commit(),
1755 * but if zil_commit() calls txg_wait_synced() (e.g.
1756 * because zil_create() or zil_commit_writer_stall() is
1757 * called) we will hit this case.
1759 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1760 zil_commit_waiter_skip(itx->itx_private);
1762 list_remove(list, itx);
1763 zil_itx_destroy(itx);
1767 t = &itxs->i_async_tree;
1768 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1769 list = &ian->ia_list;
1770 while ((itx = list_head(list)) != NULL) {
1771 list_remove(list, itx);
1772 /* commit itxs should never be on the async lists. */
1773 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1774 zil_itx_destroy(itx);
1777 kmem_free(ian, sizeof (itx_async_node_t));
1781 kmem_free(itxs, sizeof (itxs_t));
1785 zil_aitx_compare(const void *x1, const void *x2)
1787 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1788 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1790 return (AVL_CMP(o1, o2));
1794 * Remove all async itx with the given oid.
1797 zil_remove_async(zilog_t *zilog, uint64_t oid)
1800 itx_async_node_t *ian;
1807 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1809 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1812 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1814 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1815 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1817 mutex_enter(&itxg->itxg_lock);
1818 if (itxg->itxg_txg != txg) {
1819 mutex_exit(&itxg->itxg_lock);
1824 * Locate the object node and append its list.
1826 t = &itxg->itxg_itxs->i_async_tree;
1827 ian = avl_find(t, &oid, &where);
1829 list_move_tail(&clean_list, &ian->ia_list);
1830 mutex_exit(&itxg->itxg_lock);
1832 while ((itx = list_head(&clean_list)) != NULL) {
1833 list_remove(&clean_list, itx);
1834 /* commit itxs should never be on the async lists. */
1835 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1836 zil_itx_destroy(itx);
1838 list_destroy(&clean_list);
1842 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1846 itxs_t *itxs, *clean = NULL;
1849 * Object ids can be re-instantiated in the next txg so
1850 * remove any async transactions to avoid future leaks.
1851 * This can happen if a fsync occurs on the re-instantiated
1852 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1853 * the new file data and flushes a write record for the old object.
1855 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1856 zil_remove_async(zilog, itx->itx_oid);
1859 * Ensure the data of a renamed file is committed before the rename.
1861 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1862 zil_async_to_sync(zilog, itx->itx_oid);
1864 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1867 txg = dmu_tx_get_txg(tx);
1869 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1870 mutex_enter(&itxg->itxg_lock);
1871 itxs = itxg->itxg_itxs;
1872 if (itxg->itxg_txg != txg) {
1875 * The zil_clean callback hasn't got around to cleaning
1876 * this itxg. Save the itxs for release below.
1877 * This should be rare.
1879 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1880 "txg %llu", itxg->itxg_txg);
1881 clean = itxg->itxg_itxs;
1883 itxg->itxg_txg = txg;
1884 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1886 list_create(&itxs->i_sync_list, sizeof (itx_t),
1887 offsetof(itx_t, itx_node));
1888 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1889 sizeof (itx_async_node_t),
1890 offsetof(itx_async_node_t, ia_node));
1892 if (itx->itx_sync) {
1893 list_insert_tail(&itxs->i_sync_list, itx);
1895 avl_tree_t *t = &itxs->i_async_tree;
1897 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
1898 itx_async_node_t *ian;
1901 ian = avl_find(t, &foid, &where);
1903 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1904 list_create(&ian->ia_list, sizeof (itx_t),
1905 offsetof(itx_t, itx_node));
1906 ian->ia_foid = foid;
1907 avl_insert(t, ian, where);
1909 list_insert_tail(&ian->ia_list, itx);
1912 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1915 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1916 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1917 * need to be careful to always dirty the ZIL using the "real"
1918 * TXG (not itxg_txg) even when the SPA is frozen.
1920 zilog_dirty(zilog, dmu_tx_get_txg(tx));
1921 mutex_exit(&itxg->itxg_lock);
1923 /* Release the old itxs now we've dropped the lock */
1925 zil_itxg_clean(clean);
1929 * If there are any in-memory intent log transactions which have now been
1930 * synced then start up a taskq to free them. We should only do this after we
1931 * have written out the uberblocks (i.e. txg has been comitted) so that
1932 * don't inadvertently clean out in-memory log records that would be required
1936 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1938 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1941 ASSERT3U(synced_txg, <, ZILTEST_TXG);
1943 mutex_enter(&itxg->itxg_lock);
1944 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1945 mutex_exit(&itxg->itxg_lock);
1948 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1949 ASSERT3U(itxg->itxg_txg, !=, 0);
1950 clean_me = itxg->itxg_itxs;
1951 itxg->itxg_itxs = NULL;
1953 mutex_exit(&itxg->itxg_lock);
1955 * Preferably start a task queue to free up the old itxs but
1956 * if taskq_dispatch can't allocate resources to do that then
1957 * free it in-line. This should be rare. Note, using TQ_SLEEP
1958 * created a bad performance problem.
1960 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1961 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1962 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1963 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1964 zil_itxg_clean(clean_me);
1968 * This function will traverse the queue of itxs that need to be
1969 * committed, and move them onto the ZIL's zl_itx_commit_list.
1972 zil_get_commit_list(zilog_t *zilog)
1975 list_t *commit_list = &zilog->zl_itx_commit_list;
1977 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1979 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1982 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1985 * This is inherently racy, since there is nothing to prevent
1986 * the last synced txg from changing. That's okay since we'll
1987 * only commit things in the future.
1989 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1990 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1992 mutex_enter(&itxg->itxg_lock);
1993 if (itxg->itxg_txg != txg) {
1994 mutex_exit(&itxg->itxg_lock);
1999 * If we're adding itx records to the zl_itx_commit_list,
2000 * then the zil better be dirty in this "txg". We can assert
2001 * that here since we're holding the itxg_lock which will
2002 * prevent spa_sync from cleaning it. Once we add the itxs
2003 * to the zl_itx_commit_list we must commit it to disk even
2004 * if it's unnecessary (i.e. the txg was synced).
2006 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2007 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2008 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
2010 mutex_exit(&itxg->itxg_lock);
2015 * Move the async itxs for a specified object to commit into sync lists.
2018 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2021 itx_async_node_t *ian;
2025 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2028 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2031 * This is inherently racy, since there is nothing to prevent
2032 * the last synced txg from changing.
2034 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2035 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2037 mutex_enter(&itxg->itxg_lock);
2038 if (itxg->itxg_txg != txg) {
2039 mutex_exit(&itxg->itxg_lock);
2044 * If a foid is specified then find that node and append its
2045 * list. Otherwise walk the tree appending all the lists
2046 * to the sync list. We add to the end rather than the
2047 * beginning to ensure the create has happened.
2049 t = &itxg->itxg_itxs->i_async_tree;
2051 ian = avl_find(t, &foid, &where);
2053 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2057 void *cookie = NULL;
2059 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2060 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2062 list_destroy(&ian->ia_list);
2063 kmem_free(ian, sizeof (itx_async_node_t));
2066 mutex_exit(&itxg->itxg_lock);
2071 * This function will prune commit itxs that are at the head of the
2072 * commit list (it won't prune past the first non-commit itx), and
2073 * either: a) attach them to the last lwb that's still pending
2074 * completion, or b) skip them altogether.
2076 * This is used as a performance optimization to prevent commit itxs
2077 * from generating new lwbs when it's unnecessary to do so.
2080 zil_prune_commit_list(zilog_t *zilog)
2084 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2086 while (itx = list_head(&zilog->zl_itx_commit_list)) {
2087 lr_t *lrc = &itx->itx_lr;
2088 if (lrc->lrc_txtype != TX_COMMIT)
2091 mutex_enter(&zilog->zl_lock);
2093 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2094 if (last_lwb == NULL ||
2095 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2097 * All of the itxs this waiter was waiting on
2098 * must have already completed (or there were
2099 * never any itx's for it to wait on), so it's
2100 * safe to skip this waiter and mark it done.
2102 zil_commit_waiter_skip(itx->itx_private);
2104 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2105 itx->itx_private = NULL;
2108 mutex_exit(&zilog->zl_lock);
2110 list_remove(&zilog->zl_itx_commit_list, itx);
2111 zil_itx_destroy(itx);
2114 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2118 zil_commit_writer_stall(zilog_t *zilog)
2121 * When zio_alloc_zil() fails to allocate the next lwb block on
2122 * disk, we must call txg_wait_synced() to ensure all of the
2123 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2124 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2125 * to zil_process_commit_list()) will have to call zil_create(),
2126 * and start a new ZIL chain.
2128 * Since zil_alloc_zil() failed, the lwb that was previously
2129 * issued does not have a pointer to the "next" lwb on disk.
2130 * Thus, if another ZIL writer thread was to allocate the "next"
2131 * on-disk lwb, that block could be leaked in the event of a
2132 * crash (because the previous lwb on-disk would not point to
2135 * We must hold the zilog's zl_issuer_lock while we do this, to
2136 * ensure no new threads enter zil_process_commit_list() until
2137 * all lwb's in the zl_lwb_list have been synced and freed
2138 * (which is achieved via the txg_wait_synced() call).
2140 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2141 txg_wait_synced(zilog->zl_dmu_pool, 0);
2142 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2146 * This function will traverse the commit list, creating new lwbs as
2147 * needed, and committing the itxs from the commit list to these newly
2148 * created lwbs. Additionally, as a new lwb is created, the previous
2149 * lwb will be issued to the zio layer to be written to disk.
2152 zil_process_commit_list(zilog_t *zilog)
2154 spa_t *spa = zilog->zl_spa;
2155 list_t nolwb_waiters;
2159 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2162 * Return if there's nothing to commit before we dirty the fs by
2163 * calling zil_create().
2165 if (list_head(&zilog->zl_itx_commit_list) == NULL)
2168 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2169 offsetof(zil_commit_waiter_t, zcw_node));
2171 lwb = list_tail(&zilog->zl_lwb_list);
2173 lwb = zil_create(zilog);
2175 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2176 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2177 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2180 while (itx = list_head(&zilog->zl_itx_commit_list)) {
2181 lr_t *lrc = &itx->itx_lr;
2182 uint64_t txg = lrc->lrc_txg;
2184 ASSERT3U(txg, !=, 0);
2186 if (lrc->lrc_txtype == TX_COMMIT) {
2187 DTRACE_PROBE2(zil__process__commit__itx,
2188 zilog_t *, zilog, itx_t *, itx);
2190 DTRACE_PROBE2(zil__process__normal__itx,
2191 zilog_t *, zilog, itx_t *, itx);
2194 boolean_t synced = txg <= spa_last_synced_txg(spa);
2195 boolean_t frozen = txg > spa_freeze_txg(spa);
2198 * If the txg of this itx has already been synced out, then
2199 * we don't need to commit this itx to an lwb. This is
2200 * because the data of this itx will have already been
2201 * written to the main pool. This is inherently racy, and
2202 * it's still ok to commit an itx whose txg has already
2203 * been synced; this will result in a write that's
2204 * unnecessary, but will do no harm.
2206 * With that said, we always want to commit TX_COMMIT itxs
2207 * to an lwb, regardless of whether or not that itx's txg
2208 * has been synced out. We do this to ensure any OPENED lwb
2209 * will always have at least one zil_commit_waiter_t linked
2212 * As a counter-example, if we skipped TX_COMMIT itx's
2213 * whose txg had already been synced, the following
2214 * situation could occur if we happened to be racing with
2217 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
2218 * itx's txg is 10 and the last synced txg is 9.
2219 * 2. spa_sync finishes syncing out txg 10.
2220 * 3. we move to the next itx in the list, it's a TX_COMMIT
2221 * whose txg is 10, so we skip it rather than committing
2222 * it to the lwb used in (1).
2224 * If the itx that is skipped in (3) is the last TX_COMMIT
2225 * itx in the commit list, than it's possible for the lwb
2226 * used in (1) to remain in the OPENED state indefinitely.
2228 * To prevent the above scenario from occuring, ensuring
2229 * that once an lwb is OPENED it will transition to ISSUED
2230 * and eventually DONE, we always commit TX_COMMIT itx's to
2231 * an lwb here, even if that itx's txg has already been
2234 * Finally, if the pool is frozen, we _always_ commit the
2235 * itx. The point of freezing the pool is to prevent data
2236 * from being written to the main pool via spa_sync, and
2237 * instead rely solely on the ZIL to persistently store the
2238 * data; i.e. when the pool is frozen, the last synced txg
2239 * value can't be trusted.
2241 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2243 lwb = zil_lwb_commit(zilog, itx, lwb);
2244 } else if (lrc->lrc_txtype == TX_COMMIT) {
2245 ASSERT3P(lwb, ==, NULL);
2246 zil_commit_waiter_link_nolwb(
2247 itx->itx_private, &nolwb_waiters);
2251 list_remove(&zilog->zl_itx_commit_list, itx);
2252 zil_itx_destroy(itx);
2257 * This indicates zio_alloc_zil() failed to allocate the
2258 * "next" lwb on-disk. When this happens, we must stall
2259 * the ZIL write pipeline; see the comment within
2260 * zil_commit_writer_stall() for more details.
2262 zil_commit_writer_stall(zilog);
2265 * Additionally, we have to signal and mark the "nolwb"
2266 * waiters as "done" here, since without an lwb, we
2267 * can't do this via zil_lwb_flush_vdevs_done() like
2270 zil_commit_waiter_t *zcw;
2271 while (zcw = list_head(&nolwb_waiters)) {
2272 zil_commit_waiter_skip(zcw);
2273 list_remove(&nolwb_waiters, zcw);
2276 ASSERT(list_is_empty(&nolwb_waiters));
2277 ASSERT3P(lwb, !=, NULL);
2278 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2279 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2280 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2283 * At this point, the ZIL block pointed at by the "lwb"
2284 * variable is in one of the following states: "closed"
2287 * If its "closed", then no itxs have been committed to
2288 * it, so there's no point in issuing its zio (i.e.
2291 * If its "open" state, then it contains one or more
2292 * itxs that eventually need to be committed to stable
2293 * storage. In this case we intentionally do not issue
2294 * the lwb's zio to disk yet, and instead rely on one of
2295 * the following two mechanisms for issuing the zio:
2297 * 1. Ideally, there will be more ZIL activity occuring
2298 * on the system, such that this function will be
2299 * immediately called again (not necessarily by the same
2300 * thread) and this lwb's zio will be issued via
2301 * zil_lwb_commit(). This way, the lwb is guaranteed to
2302 * be "full" when it is issued to disk, and we'll make
2303 * use of the lwb's size the best we can.
2305 * 2. If there isn't sufficient ZIL activity occuring on
2306 * the system, such that this lwb's zio isn't issued via
2307 * zil_lwb_commit(), zil_commit_waiter() will issue the
2308 * lwb's zio. If this occurs, the lwb is not guaranteed
2309 * to be "full" by the time its zio is issued, and means
2310 * the size of the lwb was "too large" given the amount
2311 * of ZIL activity occuring on the system at that time.
2313 * We do this for a couple of reasons:
2315 * 1. To try and reduce the number of IOPs needed to
2316 * write the same number of itxs. If an lwb has space
2317 * available in it's buffer for more itxs, and more itxs
2318 * will be committed relatively soon (relative to the
2319 * latency of performing a write), then it's beneficial
2320 * to wait for these "next" itxs. This way, more itxs
2321 * can be committed to stable storage with fewer writes.
2323 * 2. To try and use the largest lwb block size that the
2324 * incoming rate of itxs can support. Again, this is to
2325 * try and pack as many itxs into as few lwbs as
2326 * possible, without significantly impacting the latency
2327 * of each individual itx.
2333 * This function is responsible for ensuring the passed in commit waiter
2334 * (and associated commit itx) is committed to an lwb. If the waiter is
2335 * not already committed to an lwb, all itxs in the zilog's queue of
2336 * itxs will be processed. The assumption is the passed in waiter's
2337 * commit itx will found in the queue just like the other non-commit
2338 * itxs, such that when the entire queue is processed, the waiter will
2339 * have been commited to an lwb.
2341 * The lwb associated with the passed in waiter is not guaranteed to
2342 * have been issued by the time this function completes. If the lwb is
2343 * not issued, we rely on future calls to zil_commit_writer() to issue
2344 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2347 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2349 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2350 ASSERT(spa_writeable(zilog->zl_spa));
2352 mutex_enter(&zilog->zl_issuer_lock);
2354 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2356 * It's possible that, while we were waiting to acquire
2357 * the "zl_issuer_lock", another thread committed this
2358 * waiter to an lwb. If that occurs, we bail out early,
2359 * without processing any of the zilog's queue of itxs.
2361 * On certain workloads and system configurations, the
2362 * "zl_issuer_lock" can become highly contended. In an
2363 * attempt to reduce this contention, we immediately drop
2364 * the lock if the waiter has already been processed.
2366 * We've measured this optimization to reduce CPU spent
2367 * contending on this lock by up to 5%, using a system
2368 * with 32 CPUs, low latency storage (~50 usec writes),
2369 * and 1024 threads performing sync writes.
2374 zil_get_commit_list(zilog);
2375 zil_prune_commit_list(zilog);
2376 zil_process_commit_list(zilog);
2379 mutex_exit(&zilog->zl_issuer_lock);
2383 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2385 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2386 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2387 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2389 lwb_t *lwb = zcw->zcw_lwb;
2390 ASSERT3P(lwb, !=, NULL);
2391 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2394 * If the lwb has already been issued by another thread, we can
2395 * immediately return since there's no work to be done (the
2396 * point of this function is to issue the lwb). Additionally, we
2397 * do this prior to acquiring the zl_issuer_lock, to avoid
2398 * acquiring it when it's not necessary to do so.
2400 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2401 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2402 lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2406 * In order to call zil_lwb_write_issue() we must hold the
2407 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2408 * since we're already holding the commit waiter's "zcw_lock",
2409 * and those two locks are aquired in the opposite order
2412 mutex_exit(&zcw->zcw_lock);
2413 mutex_enter(&zilog->zl_issuer_lock);
2414 mutex_enter(&zcw->zcw_lock);
2417 * Since we just dropped and re-acquired the commit waiter's
2418 * lock, we have to re-check to see if the waiter was marked
2419 * "done" during that process. If the waiter was marked "done",
2420 * the "lwb" pointer is no longer valid (it can be free'd after
2421 * the waiter is marked "done"), so without this check we could
2422 * wind up with a use-after-free error below.
2427 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2430 * We've already checked this above, but since we hadn't acquired
2431 * the zilog's zl_issuer_lock, we have to perform this check a
2432 * second time while holding the lock.
2434 * We don't need to hold the zl_lock since the lwb cannot transition
2435 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2436 * _can_ transition from ISSUED to DONE, but it's OK to race with
2437 * that transition since we treat the lwb the same, whether it's in
2438 * the ISSUED or DONE states.
2440 * The important thing, is we treat the lwb differently depending on
2441 * if it's ISSUED or OPENED, and block any other threads that might
2442 * attempt to issue this lwb. For that reason we hold the
2443 * zl_issuer_lock when checking the lwb_state; we must not call
2444 * zil_lwb_write_issue() if the lwb had already been issued.
2446 * See the comment above the lwb_state_t structure definition for
2447 * more details on the lwb states, and locking requirements.
2449 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2450 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2451 lwb->lwb_state == LWB_STATE_FLUSH_DONE)
2454 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2457 * As described in the comments above zil_commit_waiter() and
2458 * zil_process_commit_list(), we need to issue this lwb's zio
2459 * since we've reached the commit waiter's timeout and it still
2460 * hasn't been issued.
2462 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2464 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2467 * Since the lwb's zio hadn't been issued by the time this thread
2468 * reached its timeout, we reset the zilog's "zl_cur_used" field
2469 * to influence the zil block size selection algorithm.
2471 * By having to issue the lwb's zio here, it means the size of the
2472 * lwb was too large, given the incoming throughput of itxs. By
2473 * setting "zl_cur_used" to zero, we communicate this fact to the
2474 * block size selection algorithm, so it can take this informaiton
2475 * into account, and potentially select a smaller size for the
2476 * next lwb block that is allocated.
2478 zilog->zl_cur_used = 0;
2482 * When zil_lwb_write_issue() returns NULL, this
2483 * indicates zio_alloc_zil() failed to allocate the
2484 * "next" lwb on-disk. When this occurs, the ZIL write
2485 * pipeline must be stalled; see the comment within the
2486 * zil_commit_writer_stall() function for more details.
2488 * We must drop the commit waiter's lock prior to
2489 * calling zil_commit_writer_stall() or else we can wind
2490 * up with the following deadlock:
2492 * - This thread is waiting for the txg to sync while
2493 * holding the waiter's lock; txg_wait_synced() is
2494 * used within txg_commit_writer_stall().
2496 * - The txg can't sync because it is waiting for this
2497 * lwb's zio callback to call dmu_tx_commit().
2499 * - The lwb's zio callback can't call dmu_tx_commit()
2500 * because it's blocked trying to acquire the waiter's
2501 * lock, which occurs prior to calling dmu_tx_commit()
2503 mutex_exit(&zcw->zcw_lock);
2504 zil_commit_writer_stall(zilog);
2505 mutex_enter(&zcw->zcw_lock);
2509 mutex_exit(&zilog->zl_issuer_lock);
2510 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2514 * This function is responsible for performing the following two tasks:
2516 * 1. its primary responsibility is to block until the given "commit
2517 * waiter" is considered "done".
2519 * 2. its secondary responsibility is to issue the zio for the lwb that
2520 * the given "commit waiter" is waiting on, if this function has
2521 * waited "long enough" and the lwb is still in the "open" state.
2523 * Given a sufficient amount of itxs being generated and written using
2524 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2525 * function. If this does not occur, this secondary responsibility will
2526 * ensure the lwb is issued even if there is not other synchronous
2527 * activity on the system.
2529 * For more details, see zil_process_commit_list(); more specifically,
2530 * the comment at the bottom of that function.
2533 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2535 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2536 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2537 ASSERT(spa_writeable(zilog->zl_spa));
2539 mutex_enter(&zcw->zcw_lock);
2542 * The timeout is scaled based on the lwb latency to avoid
2543 * significantly impacting the latency of each individual itx.
2544 * For more details, see the comment at the bottom of the
2545 * zil_process_commit_list() function.
2547 int pct = MAX(zfs_commit_timeout_pct, 1);
2548 #if defined(illumos) || !defined(_KERNEL)
2549 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2550 hrtime_t wakeup = gethrtime() + sleep;
2552 sbintime_t sleep = nstosbt((zilog->zl_last_lwb_latency * pct) / 100);
2553 sbintime_t wakeup = getsbinuptime() + sleep;
2555 boolean_t timedout = B_FALSE;
2557 while (!zcw->zcw_done) {
2558 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2560 lwb_t *lwb = zcw->zcw_lwb;
2563 * Usually, the waiter will have a non-NULL lwb field here,
2564 * but it's possible for it to be NULL as a result of
2565 * zil_commit() racing with spa_sync().
2567 * When zil_clean() is called, it's possible for the itxg
2568 * list (which may be cleaned via a taskq) to contain
2569 * commit itxs. When this occurs, the commit waiters linked
2570 * off of these commit itxs will not be committed to an
2571 * lwb. Additionally, these commit waiters will not be
2572 * marked done until zil_commit_waiter_skip() is called via
2575 * Thus, it's possible for this commit waiter (i.e. the
2576 * "zcw" variable) to be found in this "in between" state;
2577 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2578 * been skipped, so it's "zcw_done" field is still B_FALSE.
2580 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2582 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2583 ASSERT3B(timedout, ==, B_FALSE);
2586 * If the lwb hasn't been issued yet, then we
2587 * need to wait with a timeout, in case this
2588 * function needs to issue the lwb after the
2589 * timeout is reached; responsibility (2) from
2590 * the comment above this function.
2592 #if defined(illumos) || !defined(_KERNEL)
2593 clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2594 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2595 CALLOUT_FLAG_ABSOLUTE);
2597 if (timeleft >= 0 || zcw->zcw_done)
2600 int wait_err = cv_timedwait_sbt(&zcw->zcw_cv,
2601 &zcw->zcw_lock, wakeup, SBT_1NS, C_ABSOLUTE);
2602 if (wait_err != EWOULDBLOCK || zcw->zcw_done)
2607 zil_commit_waiter_timeout(zilog, zcw);
2609 if (!zcw->zcw_done) {
2611 * If the commit waiter has already been
2612 * marked "done", it's possible for the
2613 * waiter's lwb structure to have already
2614 * been freed. Thus, we can only reliably
2615 * make these assertions if the waiter
2618 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2619 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2623 * If the lwb isn't open, then it must have already
2624 * been issued. In that case, there's no need to
2625 * use a timeout when waiting for the lwb to
2628 * Additionally, if the lwb is NULL, the waiter
2629 * will soon be signalled and marked done via
2630 * zil_clean() and zil_itxg_clean(), so no timeout
2635 lwb->lwb_state == LWB_STATE_ISSUED ||
2636 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
2637 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
2638 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2642 mutex_exit(&zcw->zcw_lock);
2645 static zil_commit_waiter_t *
2646 zil_alloc_commit_waiter()
2648 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2650 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2651 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2652 list_link_init(&zcw->zcw_node);
2653 zcw->zcw_lwb = NULL;
2654 zcw->zcw_done = B_FALSE;
2655 zcw->zcw_zio_error = 0;
2661 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2663 ASSERT(!list_link_active(&zcw->zcw_node));
2664 ASSERT3P(zcw->zcw_lwb, ==, NULL);
2665 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2666 mutex_destroy(&zcw->zcw_lock);
2667 cv_destroy(&zcw->zcw_cv);
2668 kmem_cache_free(zil_zcw_cache, zcw);
2672 * This function is used to create a TX_COMMIT itx and assign it. This
2673 * way, it will be linked into the ZIL's list of synchronous itxs, and
2674 * then later committed to an lwb (or skipped) when
2675 * zil_process_commit_list() is called.
2678 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2680 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2681 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2683 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2684 itx->itx_sync = B_TRUE;
2685 itx->itx_private = zcw;
2687 zil_itx_assign(zilog, itx, tx);
2693 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2695 * When writing ZIL transactions to the on-disk representation of the
2696 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2697 * itxs can be committed to a single lwb. Once a lwb is written and
2698 * committed to stable storage (i.e. the lwb is written, and vdevs have
2699 * been flushed), each itx that was committed to that lwb is also
2700 * considered to be committed to stable storage.
2702 * When an itx is committed to an lwb, the log record (lr_t) contained
2703 * by the itx is copied into the lwb's zio buffer, and once this buffer
2704 * is written to disk, it becomes an on-disk ZIL block.
2706 * As itxs are generated, they're inserted into the ZIL's queue of
2707 * uncommitted itxs. The semantics of zil_commit() are such that it will
2708 * block until all itxs that were in the queue when it was called, are
2709 * committed to stable storage.
2711 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2712 * itxs, for all objects in the dataset, will be committed to stable
2713 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2714 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2715 * that correspond to the foid passed in, will be committed to stable
2716 * storage prior to zil_commit() returning.
2718 * Generally speaking, when zil_commit() is called, the consumer doesn't
2719 * actually care about _all_ of the uncommitted itxs. Instead, they're
2720 * simply trying to waiting for a specific itx to be committed to disk,
2721 * but the interface(s) for interacting with the ZIL don't allow such
2722 * fine-grained communication. A better interface would allow a consumer
2723 * to create and assign an itx, and then pass a reference to this itx to
2724 * zil_commit(); such that zil_commit() would return as soon as that
2725 * specific itx was committed to disk (instead of waiting for _all_
2726 * itxs to be committed).
2728 * When a thread calls zil_commit() a special "commit itx" will be
2729 * generated, along with a corresponding "waiter" for this commit itx.
2730 * zil_commit() will wait on this waiter's CV, such that when the waiter
2731 * is marked done, and signalled, zil_commit() will return.
2733 * This commit itx is inserted into the queue of uncommitted itxs. This
2734 * provides an easy mechanism for determining which itxs were in the
2735 * queue prior to zil_commit() having been called, and which itxs were
2736 * added after zil_commit() was called.
2738 * The commit it is special; it doesn't have any on-disk representation.
2739 * When a commit itx is "committed" to an lwb, the waiter associated
2740 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2741 * completes, each waiter on the lwb's list is marked done and signalled
2742 * -- allowing the thread waiting on the waiter to return from zil_commit().
2744 * It's important to point out a few critical factors that allow us
2745 * to make use of the commit itxs, commit waiters, per-lwb lists of
2746 * commit waiters, and zio completion callbacks like we're doing:
2748 * 1. The list of waiters for each lwb is traversed, and each commit
2749 * waiter is marked "done" and signalled, in the zio completion
2750 * callback of the lwb's zio[*].
2752 * * Actually, the waiters are signalled in the zio completion
2753 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2754 * that are sent to the vdevs upon completion of the lwb zio.
2756 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2757 * itxs, the order in which they are inserted is preserved[*]; as
2758 * itxs are added to the queue, they are added to the tail of
2759 * in-memory linked lists.
2761 * When committing the itxs to lwbs (to be written to disk), they
2762 * are committed in the same order in which the itxs were added to
2763 * the uncommitted queue's linked list(s); i.e. the linked list of
2764 * itxs to commit is traversed from head to tail, and each itx is
2765 * committed to an lwb in that order.
2769 * - the order of "sync" itxs is preserved w.r.t. other
2770 * "sync" itxs, regardless of the corresponding objects.
2771 * - the order of "async" itxs is preserved w.r.t. other
2772 * "async" itxs corresponding to the same object.
2773 * - the order of "async" itxs is *not* preserved w.r.t. other
2774 * "async" itxs corresponding to different objects.
2775 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2776 * versa) is *not* preserved, even for itxs that correspond
2777 * to the same object.
2779 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2780 * zil_get_commit_list(), and zil_process_commit_list().
2782 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2783 * lwb cannot be considered committed to stable storage, until its
2784 * "previous" lwb is also committed to stable storage. This fact,
2785 * coupled with the fact described above, means that itxs are
2786 * committed in (roughly) the order in which they were generated.
2787 * This is essential because itxs are dependent on prior itxs.
2788 * Thus, we *must not* deem an itx as being committed to stable
2789 * storage, until *all* prior itxs have also been committed to
2792 * To enforce this ordering of lwb zio's, while still leveraging as
2793 * much of the underlying storage performance as possible, we rely
2794 * on two fundamental concepts:
2796 * 1. The creation and issuance of lwb zio's is protected by
2797 * the zilog's "zl_issuer_lock", which ensures only a single
2798 * thread is creating and/or issuing lwb's at a time
2799 * 2. The "previous" lwb is a child of the "current" lwb
2800 * (leveraging the zio parent-child depenency graph)
2802 * By relying on this parent-child zio relationship, we can have
2803 * many lwb zio's concurrently issued to the underlying storage,
2804 * but the order in which they complete will be the same order in
2805 * which they were created.
2808 zil_commit(zilog_t *zilog, uint64_t foid)
2811 * We should never attempt to call zil_commit on a snapshot for
2812 * a couple of reasons:
2814 * 1. A snapshot may never be modified, thus it cannot have any
2815 * in-flight itxs that would have modified the dataset.
2817 * 2. By design, when zil_commit() is called, a commit itx will
2818 * be assigned to this zilog; as a result, the zilog will be
2819 * dirtied. We must not dirty the zilog of a snapshot; there's
2820 * checks in the code that enforce this invariant, and will
2821 * cause a panic if it's not upheld.
2823 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2825 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2828 if (!spa_writeable(zilog->zl_spa)) {
2830 * If the SPA is not writable, there should never be any
2831 * pending itxs waiting to be committed to disk. If that
2832 * weren't true, we'd skip writing those itxs out, and
2833 * would break the sematics of zil_commit(); thus, we're
2834 * verifying that truth before we return to the caller.
2836 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2837 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2838 for (int i = 0; i < TXG_SIZE; i++)
2839 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2844 * If the ZIL is suspended, we don't want to dirty it by calling
2845 * zil_commit_itx_assign() below, nor can we write out
2846 * lwbs like would be done in zil_commit_write(). Thus, we
2847 * simply rely on txg_wait_synced() to maintain the necessary
2848 * semantics, and avoid calling those functions altogether.
2850 if (zilog->zl_suspend > 0) {
2851 txg_wait_synced(zilog->zl_dmu_pool, 0);
2855 zil_commit_impl(zilog, foid);
2859 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2862 * Move the "async" itxs for the specified foid to the "sync"
2863 * queues, such that they will be later committed (or skipped)
2864 * to an lwb when zil_process_commit_list() is called.
2866 * Since these "async" itxs must be committed prior to this
2867 * call to zil_commit returning, we must perform this operation
2868 * before we call zil_commit_itx_assign().
2870 zil_async_to_sync(zilog, foid);
2873 * We allocate a new "waiter" structure which will initially be
2874 * linked to the commit itx using the itx's "itx_private" field.
2875 * Since the commit itx doesn't represent any on-disk state,
2876 * when it's committed to an lwb, rather than copying the its
2877 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2878 * added to the lwb's list of waiters. Then, when the lwb is
2879 * committed to stable storage, each waiter in the lwb's list of
2880 * waiters will be marked "done", and signalled.
2882 * We must create the waiter and assign the commit itx prior to
2883 * calling zil_commit_writer(), or else our specific commit itx
2884 * is not guaranteed to be committed to an lwb prior to calling
2885 * zil_commit_waiter().
2887 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2888 zil_commit_itx_assign(zilog, zcw);
2890 zil_commit_writer(zilog, zcw);
2891 zil_commit_waiter(zilog, zcw);
2893 if (zcw->zcw_zio_error != 0) {
2895 * If there was an error writing out the ZIL blocks that
2896 * this thread is waiting on, then we fallback to
2897 * relying on spa_sync() to write out the data this
2898 * thread is waiting on. Obviously this has performance
2899 * implications, but the expectation is for this to be
2900 * an exceptional case, and shouldn't occur often.
2902 DTRACE_PROBE2(zil__commit__io__error,
2903 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2904 txg_wait_synced(zilog->zl_dmu_pool, 0);
2907 zil_free_commit_waiter(zcw);
2911 * Called in syncing context to free committed log blocks and update log header.
2914 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2916 zil_header_t *zh = zil_header_in_syncing_context(zilog);
2917 uint64_t txg = dmu_tx_get_txg(tx);
2918 spa_t *spa = zilog->zl_spa;
2919 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2923 * We don't zero out zl_destroy_txg, so make sure we don't try
2924 * to destroy it twice.
2926 if (spa_sync_pass(spa) != 1)
2929 mutex_enter(&zilog->zl_lock);
2931 ASSERT(zilog->zl_stop_sync == 0);
2933 if (*replayed_seq != 0) {
2934 ASSERT(zh->zh_replay_seq < *replayed_seq);
2935 zh->zh_replay_seq = *replayed_seq;
2939 if (zilog->zl_destroy_txg == txg) {
2940 blkptr_t blk = zh->zh_log;
2942 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2944 bzero(zh, sizeof (zil_header_t));
2945 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2947 if (zilog->zl_keep_first) {
2949 * If this block was part of log chain that couldn't
2950 * be claimed because a device was missing during
2951 * zil_claim(), but that device later returns,
2952 * then this block could erroneously appear valid.
2953 * To guard against this, assign a new GUID to the new
2954 * log chain so it doesn't matter what blk points to.
2956 zil_init_log_chain(zilog, &blk);
2961 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2962 zh->zh_log = lwb->lwb_blk;
2963 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2965 list_remove(&zilog->zl_lwb_list, lwb);
2966 zio_free(spa, txg, &lwb->lwb_blk);
2967 zil_free_lwb(zilog, lwb);
2970 * If we don't have anything left in the lwb list then
2971 * we've had an allocation failure and we need to zero
2972 * out the zil_header blkptr so that we don't end
2973 * up freeing the same block twice.
2975 if (list_head(&zilog->zl_lwb_list) == NULL)
2976 BP_ZERO(&zh->zh_log);
2978 mutex_exit(&zilog->zl_lock);
2983 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2986 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2987 offsetof(zil_commit_waiter_t, zcw_node));
2988 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2989 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2990 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2996 zil_lwb_dest(void *vbuf, void *unused)
2999 mutex_destroy(&lwb->lwb_vdev_lock);
3000 avl_destroy(&lwb->lwb_vdev_tree);
3001 list_destroy(&lwb->lwb_waiters);
3007 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3008 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3010 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3011 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3017 kmem_cache_destroy(zil_zcw_cache);
3018 kmem_cache_destroy(zil_lwb_cache);
3022 zil_set_sync(zilog_t *zilog, uint64_t sync)
3024 zilog->zl_sync = sync;
3028 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3030 zilog->zl_logbias = logbias;
3034 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3038 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3040 zilog->zl_header = zh_phys;
3042 zilog->zl_spa = dmu_objset_spa(os);
3043 zilog->zl_dmu_pool = dmu_objset_pool(os);
3044 zilog->zl_destroy_txg = TXG_INITIAL - 1;
3045 zilog->zl_logbias = dmu_objset_logbias(os);
3046 zilog->zl_sync = dmu_objset_syncprop(os);
3047 zilog->zl_dirty_max_txg = 0;
3048 zilog->zl_last_lwb_opened = NULL;
3049 zilog->zl_last_lwb_latency = 0;
3050 zilog->zl_max_block_size = zil_maxblocksize;
3052 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3053 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3055 for (int i = 0; i < TXG_SIZE; i++) {
3056 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3057 MUTEX_DEFAULT, NULL);
3060 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3061 offsetof(lwb_t, lwb_node));
3063 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3064 offsetof(itx_t, itx_node));
3066 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3072 zil_free(zilog_t *zilog)
3074 zilog->zl_stop_sync = 1;
3076 ASSERT0(zilog->zl_suspend);
3077 ASSERT0(zilog->zl_suspending);
3079 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3080 list_destroy(&zilog->zl_lwb_list);
3082 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3083 list_destroy(&zilog->zl_itx_commit_list);
3085 for (int i = 0; i < TXG_SIZE; i++) {
3087 * It's possible for an itx to be generated that doesn't dirty
3088 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3089 * callback to remove the entry. We remove those here.
3091 * Also free up the ziltest itxs.
3093 if (zilog->zl_itxg[i].itxg_itxs)
3094 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3095 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3098 mutex_destroy(&zilog->zl_issuer_lock);
3099 mutex_destroy(&zilog->zl_lock);
3101 cv_destroy(&zilog->zl_cv_suspend);
3103 kmem_free(zilog, sizeof (zilog_t));
3107 * Open an intent log.
3110 zil_open(objset_t *os, zil_get_data_t *get_data)
3112 zilog_t *zilog = dmu_objset_zil(os);
3114 ASSERT3P(zilog->zl_get_data, ==, NULL);
3115 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3116 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3118 zilog->zl_get_data = get_data;
3124 * Close an intent log.
3127 zil_close(zilog_t *zilog)
3132 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3133 zil_commit(zilog, 0);
3135 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
3136 ASSERT0(zilog->zl_dirty_max_txg);
3137 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3140 mutex_enter(&zilog->zl_lock);
3141 lwb = list_tail(&zilog->zl_lwb_list);
3143 txg = zilog->zl_dirty_max_txg;
3145 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3146 mutex_exit(&zilog->zl_lock);
3149 * We need to use txg_wait_synced() to wait long enough for the
3150 * ZIL to be clean, and to wait for all pending lwbs to be
3154 txg_wait_synced(zilog->zl_dmu_pool, txg);
3156 if (zilog_is_dirty(zilog))
3157 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
3158 if (txg < spa_freeze_txg(zilog->zl_spa))
3159 VERIFY(!zilog_is_dirty(zilog));
3161 zilog->zl_get_data = NULL;
3164 * We should have only one lwb left on the list; remove it now.
3166 mutex_enter(&zilog->zl_lock);
3167 lwb = list_head(&zilog->zl_lwb_list);
3169 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3170 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3171 list_remove(&zilog->zl_lwb_list, lwb);
3172 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3173 zil_free_lwb(zilog, lwb);
3175 mutex_exit(&zilog->zl_lock);
3178 static char *suspend_tag = "zil suspending";
3181 * Suspend an intent log. While in suspended mode, we still honor
3182 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3183 * On old version pools, we suspend the log briefly when taking a
3184 * snapshot so that it will have an empty intent log.
3186 * Long holds are not really intended to be used the way we do here --
3187 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3188 * could fail. Therefore we take pains to only put a long hold if it is
3189 * actually necessary. Fortunately, it will only be necessary if the
3190 * objset is currently mounted (or the ZVOL equivalent). In that case it
3191 * will already have a long hold, so we are not really making things any worse.
3193 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3194 * zvol_state_t), and use their mechanism to prevent their hold from being
3195 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3198 * if cookiep == NULL, this does both the suspend & resume.
3199 * Otherwise, it returns with the dataset "long held", and the cookie
3200 * should be passed into zil_resume().
3203 zil_suspend(const char *osname, void **cookiep)
3207 const zil_header_t *zh;
3210 error = dmu_objset_hold(osname, suspend_tag, &os);
3213 zilog = dmu_objset_zil(os);
3215 mutex_enter(&zilog->zl_lock);
3216 zh = zilog->zl_header;
3218 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
3219 mutex_exit(&zilog->zl_lock);
3220 dmu_objset_rele(os, suspend_tag);
3221 return (SET_ERROR(EBUSY));
3225 * Don't put a long hold in the cases where we can avoid it. This
3226 * is when there is no cookie so we are doing a suspend & resume
3227 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3228 * for the suspend because it's already suspended, or there's no ZIL.
3230 if (cookiep == NULL && !zilog->zl_suspending &&
3231 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3232 mutex_exit(&zilog->zl_lock);
3233 dmu_objset_rele(os, suspend_tag);
3237 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3238 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3240 zilog->zl_suspend++;
3242 if (zilog->zl_suspend > 1) {
3244 * Someone else is already suspending it.
3245 * Just wait for them to finish.
3248 while (zilog->zl_suspending)
3249 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3250 mutex_exit(&zilog->zl_lock);
3252 if (cookiep == NULL)
3260 * If there is no pointer to an on-disk block, this ZIL must not
3261 * be active (e.g. filesystem not mounted), so there's nothing
3264 if (BP_IS_HOLE(&zh->zh_log)) {
3265 ASSERT(cookiep != NULL); /* fast path already handled */
3268 mutex_exit(&zilog->zl_lock);
3272 zilog->zl_suspending = B_TRUE;
3273 mutex_exit(&zilog->zl_lock);
3276 * We need to use zil_commit_impl to ensure we wait for all
3277 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3278 * to disk before proceeding. If we used zil_commit instead, it
3279 * would just call txg_wait_synced(), because zl_suspend is set.
3280 * txg_wait_synced() doesn't wait for these lwb's to be
3281 * LWB_STATE_FLUSH_DONE before returning.
3283 zil_commit_impl(zilog, 0);
3286 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3287 * use txg_wait_synced() to ensure the data from the zilog has
3288 * migrated to the main pool before calling zil_destroy().
3290 txg_wait_synced(zilog->zl_dmu_pool, 0);
3292 zil_destroy(zilog, B_FALSE);
3294 mutex_enter(&zilog->zl_lock);
3295 zilog->zl_suspending = B_FALSE;
3296 cv_broadcast(&zilog->zl_cv_suspend);
3297 mutex_exit(&zilog->zl_lock);
3299 if (cookiep == NULL)
3307 zil_resume(void *cookie)
3309 objset_t *os = cookie;
3310 zilog_t *zilog = dmu_objset_zil(os);
3312 mutex_enter(&zilog->zl_lock);
3313 ASSERT(zilog->zl_suspend != 0);
3314 zilog->zl_suspend--;
3315 mutex_exit(&zilog->zl_lock);
3316 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3317 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3320 typedef struct zil_replay_arg {
3321 zil_replay_func_t **zr_replay;
3323 boolean_t zr_byteswap;
3328 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3330 char name[ZFS_MAX_DATASET_NAME_LEN];
3332 zilog->zl_replaying_seq--; /* didn't actually replay this one */
3334 dmu_objset_name(zilog->zl_os, name);
3336 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3337 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3338 (u_longlong_t)lr->lrc_seq,
3339 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3340 (lr->lrc_txtype & TX_CI) ? "CI" : "");
3346 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3348 zil_replay_arg_t *zr = zra;
3349 const zil_header_t *zh = zilog->zl_header;
3350 uint64_t reclen = lr->lrc_reclen;
3351 uint64_t txtype = lr->lrc_txtype;
3354 zilog->zl_replaying_seq = lr->lrc_seq;
3356 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
3359 if (lr->lrc_txg < claim_txg) /* already committed */
3362 /* Strip case-insensitive bit, still present in log record */
3365 if (txtype == 0 || txtype >= TX_MAX_TYPE)
3366 return (zil_replay_error(zilog, lr, EINVAL));
3369 * If this record type can be logged out of order, the object
3370 * (lr_foid) may no longer exist. That's legitimate, not an error.
3372 if (TX_OOO(txtype)) {
3373 error = dmu_object_info(zilog->zl_os,
3374 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3375 if (error == ENOENT || error == EEXIST)
3380 * Make a copy of the data so we can revise and extend it.
3382 bcopy(lr, zr->zr_lr, reclen);
3385 * If this is a TX_WRITE with a blkptr, suck in the data.
3387 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3388 error = zil_read_log_data(zilog, (lr_write_t *)lr,
3389 zr->zr_lr + reclen);
3391 return (zil_replay_error(zilog, lr, error));
3395 * The log block containing this lr may have been byteswapped
3396 * so that we can easily examine common fields like lrc_txtype.
3397 * However, the log is a mix of different record types, and only the
3398 * replay vectors know how to byteswap their records. Therefore, if
3399 * the lr was byteswapped, undo it before invoking the replay vector.
3401 if (zr->zr_byteswap)
3402 byteswap_uint64_array(zr->zr_lr, reclen);
3405 * We must now do two things atomically: replay this log record,
3406 * and update the log header sequence number to reflect the fact that
3407 * we did so. At the end of each replay function the sequence number
3408 * is updated if we are in replay mode.
3410 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3413 * The DMU's dnode layer doesn't see removes until the txg
3414 * commits, so a subsequent claim can spuriously fail with
3415 * EEXIST. So if we receive any error we try syncing out
3416 * any removes then retry the transaction. Note that we
3417 * specify B_FALSE for byteswap now, so we don't do it twice.
3419 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3420 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3422 return (zil_replay_error(zilog, lr, error));
3429 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3431 zilog->zl_replay_blks++;
3437 * If this dataset has a non-empty intent log, replay it and destroy it.
3440 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3442 zilog_t *zilog = dmu_objset_zil(os);
3443 const zil_header_t *zh = zilog->zl_header;
3444 zil_replay_arg_t zr;
3446 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3447 zil_destroy(zilog, B_TRUE);
3451 zr.zr_replay = replay_func;
3453 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3454 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3457 * Wait for in-progress removes to sync before starting replay.
3459 txg_wait_synced(zilog->zl_dmu_pool, 0);
3461 zilog->zl_replay = B_TRUE;
3462 zilog->zl_replay_time = ddi_get_lbolt();
3463 ASSERT(zilog->zl_replay_blks == 0);
3464 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3466 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3468 zil_destroy(zilog, B_FALSE);
3469 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3470 zilog->zl_replay = B_FALSE;
3474 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3476 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3479 if (zilog->zl_replay) {
3480 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3481 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3482 zilog->zl_replaying_seq;
3491 zil_reset(const char *osname, void *arg)
3495 error = zil_suspend(osname, NULL);
3497 return (SET_ERROR(EEXIST));