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 * Tunable parameter for debugging or performance analysis. Setting
102 * zfs_nocacheflush will cause corruption on power loss if a volatile
103 * out-of-order write cache is enabled.
105 boolean_t zfs_nocacheflush = B_FALSE;
106 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
107 &zfs_nocacheflush, 0, "Disable cache flush");
108 boolean_t zfs_trim_enabled = B_TRUE;
109 SYSCTL_DECL(_vfs_zfs_trim);
110 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
114 * Limit SLOG write size per commit executed with synchronous priority.
115 * Any writes above that will be executed with lower (asynchronous) priority
116 * to limit potential SLOG device abuse by single active ZIL writer.
118 uint64_t zil_slog_bulk = 768 * 1024;
119 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_bulk, CTLFLAG_RWTUN,
120 &zil_slog_bulk, 0, "Maximal SLOG commit size with sync priority");
122 static kmem_cache_t *zil_lwb_cache;
123 static kmem_cache_t *zil_zcw_cache;
125 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
126 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
129 zil_bp_compare(const void *x1, const void *x2)
131 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
132 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
134 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
136 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
139 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
141 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
148 zil_bp_tree_init(zilog_t *zilog)
150 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
151 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
155 zil_bp_tree_fini(zilog_t *zilog)
157 avl_tree_t *t = &zilog->zl_bp_tree;
161 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
162 kmem_free(zn, sizeof (zil_bp_node_t));
168 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
170 avl_tree_t *t = &zilog->zl_bp_tree;
175 if (BP_IS_EMBEDDED(bp))
178 dva = BP_IDENTITY(bp);
180 if (avl_find(t, dva, &where) != NULL)
181 return (SET_ERROR(EEXIST));
183 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
185 avl_insert(t, zn, where);
190 static zil_header_t *
191 zil_header_in_syncing_context(zilog_t *zilog)
193 return ((zil_header_t *)zilog->zl_header);
197 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
199 zio_cksum_t *zc = &bp->blk_cksum;
201 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
202 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
203 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
204 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
208 * Read a log block and make sure it's valid.
211 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
214 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
215 arc_flags_t aflags = ARC_FLAG_WAIT;
216 arc_buf_t *abuf = NULL;
220 if (zilog->zl_header->zh_claim_txg == 0)
221 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
223 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
224 zio_flags |= ZIO_FLAG_SPECULATIVE;
226 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
227 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
229 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
230 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
233 zio_cksum_t cksum = bp->blk_cksum;
236 * Validate the checksummed log block.
238 * Sequence numbers should be... sequential. The checksum
239 * verifier for the next block should be bp's checksum plus 1.
241 * Also check the log chain linkage and size used.
243 cksum.zc_word[ZIL_ZC_SEQ]++;
245 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
246 zil_chain_t *zilc = abuf->b_data;
247 char *lr = (char *)(zilc + 1);
248 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
250 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
251 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
252 error = SET_ERROR(ECKSUM);
254 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
256 *end = (char *)dst + len;
257 *nbp = zilc->zc_next_blk;
260 char *lr = abuf->b_data;
261 uint64_t size = BP_GET_LSIZE(bp);
262 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
264 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
265 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
266 (zilc->zc_nused > (size - sizeof (*zilc)))) {
267 error = SET_ERROR(ECKSUM);
269 ASSERT3U(zilc->zc_nused, <=,
270 SPA_OLD_MAXBLOCKSIZE);
271 bcopy(lr, dst, zilc->zc_nused);
272 *end = (char *)dst + zilc->zc_nused;
273 *nbp = zilc->zc_next_blk;
277 arc_buf_destroy(abuf, &abuf);
284 * Read a TX_WRITE log data block.
287 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
289 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
290 const blkptr_t *bp = &lr->lr_blkptr;
291 arc_flags_t aflags = ARC_FLAG_WAIT;
292 arc_buf_t *abuf = NULL;
296 if (BP_IS_HOLE(bp)) {
298 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
302 if (zilog->zl_header->zh_claim_txg == 0)
303 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
305 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
306 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
308 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
309 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
313 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
314 arc_buf_destroy(abuf, &abuf);
321 * Parse the intent log, and call parse_func for each valid record within.
324 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
325 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
327 const zil_header_t *zh = zilog->zl_header;
328 boolean_t claimed = !!zh->zh_claim_txg;
329 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
330 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
331 uint64_t max_blk_seq = 0;
332 uint64_t max_lr_seq = 0;
333 uint64_t blk_count = 0;
334 uint64_t lr_count = 0;
335 blkptr_t blk, next_blk;
340 * Old logs didn't record the maximum zh_claim_lr_seq.
342 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
343 claim_lr_seq = UINT64_MAX;
346 * Starting at the block pointed to by zh_log we read the log chain.
347 * For each block in the chain we strongly check that block to
348 * ensure its validity. We stop when an invalid block is found.
349 * For each block pointer in the chain we call parse_blk_func().
350 * For each record in each valid block we call parse_lr_func().
351 * If the log has been claimed, stop if we encounter a sequence
352 * number greater than the highest claimed sequence number.
354 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
355 zil_bp_tree_init(zilog);
357 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
358 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
362 if (blk_seq > claim_blk_seq)
364 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
366 ASSERT3U(max_blk_seq, <, blk_seq);
367 max_blk_seq = blk_seq;
370 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
373 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
377 for (lrp = lrbuf; lrp < end; lrp += reclen) {
378 lr_t *lr = (lr_t *)lrp;
379 reclen = lr->lrc_reclen;
380 ASSERT3U(reclen, >=, sizeof (lr_t));
381 if (lr->lrc_seq > claim_lr_seq)
383 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
385 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
386 max_lr_seq = lr->lrc_seq;
391 zilog->zl_parse_error = error;
392 zilog->zl_parse_blk_seq = max_blk_seq;
393 zilog->zl_parse_lr_seq = max_lr_seq;
394 zilog->zl_parse_blk_count = blk_count;
395 zilog->zl_parse_lr_count = lr_count;
397 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
398 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
400 zil_bp_tree_fini(zilog);
401 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
408 zil_clear_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
410 ASSERT(!BP_IS_HOLE(bp));
413 * As we call this function from the context of a rewind to a
414 * checkpoint, each ZIL block whose txg is later than the txg
415 * that we rewind to is invalid. Thus, we return -1 so
416 * zil_parse() doesn't attempt to read it.
418 if (bp->blk_birth >= first_txg)
421 if (zil_bp_tree_add(zilog, bp) != 0)
424 zio_free(zilog->zl_spa, first_txg, bp);
430 zil_noop_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
436 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
439 * Claim log block if not already committed and not already claimed.
440 * If tx == NULL, just verify that the block is claimable.
442 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
443 zil_bp_tree_add(zilog, bp) != 0)
446 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
447 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
448 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
452 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
454 lr_write_t *lr = (lr_write_t *)lrc;
457 if (lrc->lrc_txtype != TX_WRITE)
461 * If the block is not readable, don't claim it. This can happen
462 * in normal operation when a log block is written to disk before
463 * some of the dmu_sync() blocks it points to. In this case, the
464 * transaction cannot have been committed to anyone (we would have
465 * waited for all writes to be stable first), so it is semantically
466 * correct to declare this the end of the log.
468 if (lr->lr_blkptr.blk_birth >= first_txg &&
469 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
471 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
476 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
478 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
484 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
486 lr_write_t *lr = (lr_write_t *)lrc;
487 blkptr_t *bp = &lr->lr_blkptr;
490 * If we previously claimed it, we need to free it.
492 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
493 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
495 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
501 zil_lwb_vdev_compare(const void *x1, const void *x2)
503 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
504 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
515 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
519 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
520 lwb->lwb_zilog = zilog;
522 lwb->lwb_slog = slog;
523 lwb->lwb_state = LWB_STATE_CLOSED;
524 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
525 lwb->lwb_max_txg = txg;
526 lwb->lwb_write_zio = NULL;
527 lwb->lwb_root_zio = NULL;
529 lwb->lwb_issued_timestamp = 0;
530 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
531 lwb->lwb_nused = sizeof (zil_chain_t);
532 lwb->lwb_sz = BP_GET_LSIZE(bp);
535 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
538 mutex_enter(&zilog->zl_lock);
539 list_insert_tail(&zilog->zl_lwb_list, lwb);
540 mutex_exit(&zilog->zl_lock);
542 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
543 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
544 VERIFY(list_is_empty(&lwb->lwb_waiters));
550 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
552 ASSERT(MUTEX_HELD(&zilog->zl_lock));
553 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
554 VERIFY(list_is_empty(&lwb->lwb_waiters));
555 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
556 ASSERT3P(lwb->lwb_write_zio, ==, NULL);
557 ASSERT3P(lwb->lwb_root_zio, ==, NULL);
558 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
559 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
560 lwb->lwb_state == LWB_STATE_DONE);
563 * Clear the zilog's field to indicate this lwb is no longer
564 * valid, and prevent use-after-free errors.
566 if (zilog->zl_last_lwb_opened == lwb)
567 zilog->zl_last_lwb_opened = NULL;
569 kmem_cache_free(zil_lwb_cache, lwb);
573 * Called when we create in-memory log transactions so that we know
574 * to cleanup the itxs at the end of spa_sync().
577 zilog_dirty(zilog_t *zilog, uint64_t txg)
579 dsl_pool_t *dp = zilog->zl_dmu_pool;
580 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
582 ASSERT(spa_writeable(zilog->zl_spa));
584 if (ds->ds_is_snapshot)
585 panic("dirtying snapshot!");
587 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
588 /* up the hold count until we can be written out */
589 dmu_buf_add_ref(ds->ds_dbuf, zilog);
591 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
596 * Determine if the zil is dirty in the specified txg. Callers wanting to
597 * ensure that the dirty state does not change must hold the itxg_lock for
598 * the specified txg. Holding the lock will ensure that the zil cannot be
599 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
603 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
605 dsl_pool_t *dp = zilog->zl_dmu_pool;
607 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
613 * Determine if the zil is dirty. The zil is considered dirty if it has
614 * any pending itx records that have not been cleaned by zil_clean().
617 zilog_is_dirty(zilog_t *zilog)
619 dsl_pool_t *dp = zilog->zl_dmu_pool;
621 for (int t = 0; t < TXG_SIZE; t++) {
622 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
629 * Create an on-disk intent log.
632 zil_create(zilog_t *zilog)
634 const zil_header_t *zh = zilog->zl_header;
640 boolean_t slog = FALSE;
643 * Wait for any previous destroy to complete.
645 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
647 ASSERT(zh->zh_claim_txg == 0);
648 ASSERT(zh->zh_replay_seq == 0);
653 * Allocate an initial log block if:
654 * - there isn't one already
655 * - the existing block is the wrong endianess
657 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
658 tx = dmu_tx_create(zilog->zl_os);
659 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
660 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
661 txg = dmu_tx_get_txg(tx);
663 if (!BP_IS_HOLE(&blk)) {
664 zio_free(zilog->zl_spa, txg, &blk);
668 error = zio_alloc_zil(zilog->zl_spa,
669 zilog->zl_os->os_dsl_dataset->ds_object, txg, &blk, NULL,
670 ZIL_MIN_BLKSZ, &slog);
673 zil_init_log_chain(zilog, &blk);
677 * Allocate a log write block (lwb) for the first log block.
680 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
683 * If we just allocated the first log block, commit our transaction
684 * and wait for zil_sync() to stuff the block poiner into zh_log.
685 * (zh is part of the MOS, so we cannot modify it in open context.)
689 txg_wait_synced(zilog->zl_dmu_pool, txg);
692 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
698 * In one tx, free all log blocks and clear the log header. If keep_first
699 * is set, then we're replaying a log with no content. We want to keep the
700 * first block, however, so that the first synchronous transaction doesn't
701 * require a txg_wait_synced() in zil_create(). We don't need to
702 * txg_wait_synced() here either when keep_first is set, because both
703 * zil_create() and zil_destroy() will wait for any in-progress destroys
707 zil_destroy(zilog_t *zilog, boolean_t keep_first)
709 const zil_header_t *zh = zilog->zl_header;
715 * Wait for any previous destroy to complete.
717 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
719 zilog->zl_old_header = *zh; /* debugging aid */
721 if (BP_IS_HOLE(&zh->zh_log))
724 tx = dmu_tx_create(zilog->zl_os);
725 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
726 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
727 txg = dmu_tx_get_txg(tx);
729 mutex_enter(&zilog->zl_lock);
731 ASSERT3U(zilog->zl_destroy_txg, <, txg);
732 zilog->zl_destroy_txg = txg;
733 zilog->zl_keep_first = keep_first;
735 if (!list_is_empty(&zilog->zl_lwb_list)) {
736 ASSERT(zh->zh_claim_txg == 0);
738 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
739 list_remove(&zilog->zl_lwb_list, lwb);
740 if (lwb->lwb_buf != NULL)
741 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
742 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
743 zil_free_lwb(zilog, lwb);
745 } else if (!keep_first) {
746 zil_destroy_sync(zilog, tx);
748 mutex_exit(&zilog->zl_lock);
754 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
756 ASSERT(list_is_empty(&zilog->zl_lwb_list));
757 (void) zil_parse(zilog, zil_free_log_block,
758 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
762 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
764 dmu_tx_t *tx = txarg;
771 error = dmu_objset_own_obj(dp, ds->ds_object,
772 DMU_OST_ANY, B_FALSE, FTAG, &os);
775 * EBUSY indicates that the objset is inconsistent, in which
776 * case it can not have a ZIL.
778 if (error != EBUSY) {
779 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
780 (unsigned long long)ds->ds_object, error);
785 zilog = dmu_objset_zil(os);
786 zh = zil_header_in_syncing_context(zilog);
787 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
788 first_txg = spa_min_claim_txg(zilog->zl_spa);
791 * If the spa_log_state is not set to be cleared, check whether
792 * the current uberblock is a checkpoint one and if the current
793 * header has been claimed before moving on.
795 * If the current uberblock is a checkpointed uberblock then
796 * one of the following scenarios took place:
798 * 1] We are currently rewinding to the checkpoint of the pool.
799 * 2] We crashed in the middle of a checkpoint rewind but we
800 * did manage to write the checkpointed uberblock to the
801 * vdev labels, so when we tried to import the pool again
802 * the checkpointed uberblock was selected from the import
805 * In both cases we want to zero out all the ZIL blocks, except
806 * the ones that have been claimed at the time of the checkpoint
807 * (their zh_claim_txg != 0). The reason is that these blocks
808 * may be corrupted since we may have reused their locations on
809 * disk after we took the checkpoint.
811 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
812 * when we first figure out whether the current uberblock is
813 * checkpointed or not. Unfortunately, that would discard all
814 * the logs, including the ones that are claimed, and we would
817 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
818 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
819 zh->zh_claim_txg == 0)) {
820 if (!BP_IS_HOLE(&zh->zh_log)) {
821 (void) zil_parse(zilog, zil_clear_log_block,
822 zil_noop_log_record, tx, first_txg);
824 BP_ZERO(&zh->zh_log);
825 dsl_dataset_dirty(dmu_objset_ds(os), tx);
826 dmu_objset_disown(os, FTAG);
831 * If we are not rewinding and opening the pool normally, then
832 * the min_claim_txg should be equal to the first txg of the pool.
834 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
837 * Claim all log blocks if we haven't already done so, and remember
838 * the highest claimed sequence number. This ensures that if we can
839 * read only part of the log now (e.g. due to a missing device),
840 * but we can read the entire log later, we will not try to replay
841 * or destroy beyond the last block we successfully claimed.
843 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
844 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
845 (void) zil_parse(zilog, zil_claim_log_block,
846 zil_claim_log_record, tx, first_txg);
847 zh->zh_claim_txg = first_txg;
848 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
849 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
850 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
851 zh->zh_flags |= ZIL_REPLAY_NEEDED;
852 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
853 dsl_dataset_dirty(dmu_objset_ds(os), tx);
856 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
857 dmu_objset_disown(os, FTAG);
862 * Check the log by walking the log chain.
863 * Checksum errors are ok as they indicate the end of the chain.
864 * Any other error (no device or read failure) returns an error.
868 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
877 error = dmu_objset_from_ds(ds, &os);
879 cmn_err(CE_WARN, "can't open objset %llu, error %d",
880 (unsigned long long)ds->ds_object, error);
884 zilog = dmu_objset_zil(os);
885 bp = (blkptr_t *)&zilog->zl_header->zh_log;
887 if (!BP_IS_HOLE(bp)) {
889 boolean_t valid = B_TRUE;
892 * Check the first block and determine if it's on a log device
893 * which may have been removed or faulted prior to loading this
894 * pool. If so, there's no point in checking the rest of the
895 * log as its content should have already been synced to the
898 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
899 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
900 if (vd->vdev_islog && vdev_is_dead(vd))
901 valid = vdev_log_state_valid(vd);
902 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
908 * Check whether the current uberblock is checkpointed (e.g.
909 * we are rewinding) and whether the current header has been
910 * claimed or not. If it hasn't then skip verifying it. We
911 * do this because its ZIL blocks may be part of the pool's
912 * state before the rewind, which is no longer valid.
914 zil_header_t *zh = zil_header_in_syncing_context(zilog);
915 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
916 zh->zh_claim_txg == 0)
921 * Because tx == NULL, zil_claim_log_block() will not actually claim
922 * any blocks, but just determine whether it is possible to do so.
923 * In addition to checking the log chain, zil_claim_log_block()
924 * will invoke zio_claim() with a done func of spa_claim_notify(),
925 * which will update spa_max_claim_txg. See spa_load() for details.
927 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
928 zilog->zl_header->zh_claim_txg ? -1ULL :
929 spa_min_claim_txg(os->os_spa));
931 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
935 * When an itx is "skipped", this function is used to properly mark the
936 * waiter as "done, and signal any thread(s) waiting on it. An itx can
937 * be skipped (and not committed to an lwb) for a variety of reasons,
938 * one of them being that the itx was committed via spa_sync(), prior to
939 * it being committed to an lwb; this can happen if a thread calling
940 * zil_commit() is racing with spa_sync().
943 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
945 mutex_enter(&zcw->zcw_lock);
946 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
947 zcw->zcw_done = B_TRUE;
948 cv_broadcast(&zcw->zcw_cv);
949 mutex_exit(&zcw->zcw_lock);
953 * This function is used when the given waiter is to be linked into an
954 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
955 * At this point, the waiter will no longer be referenced by the itx,
956 * and instead, will be referenced by the lwb.
959 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
962 * The lwb_waiters field of the lwb is protected by the zilog's
963 * zl_lock, thus it must be held when calling this function.
965 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
967 mutex_enter(&zcw->zcw_lock);
968 ASSERT(!list_link_active(&zcw->zcw_node));
969 ASSERT3P(zcw->zcw_lwb, ==, NULL);
970 ASSERT3P(lwb, !=, NULL);
971 ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
972 lwb->lwb_state == LWB_STATE_ISSUED);
974 list_insert_tail(&lwb->lwb_waiters, zcw);
976 mutex_exit(&zcw->zcw_lock);
980 * This function is used when zio_alloc_zil() fails to allocate a ZIL
981 * block, and the given waiter must be linked to the "nolwb waiters"
982 * list inside of zil_process_commit_list().
985 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
987 mutex_enter(&zcw->zcw_lock);
988 ASSERT(!list_link_active(&zcw->zcw_node));
989 ASSERT3P(zcw->zcw_lwb, ==, NULL);
990 list_insert_tail(nolwb, zcw);
991 mutex_exit(&zcw->zcw_lock);
995 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
997 avl_tree_t *t = &lwb->lwb_vdev_tree;
999 zil_vdev_node_t *zv, zvsearch;
1000 int ndvas = BP_GET_NDVAS(bp);
1003 if (zfs_nocacheflush)
1006 mutex_enter(&lwb->lwb_vdev_lock);
1007 for (i = 0; i < ndvas; i++) {
1008 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1009 if (avl_find(t, &zvsearch, &where) == NULL) {
1010 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1011 zv->zv_vdev = zvsearch.zv_vdev;
1012 avl_insert(t, zv, where);
1015 mutex_exit(&lwb->lwb_vdev_lock);
1019 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1021 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1025 * This function is a called after all VDEVs associated with a given lwb
1026 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1027 * as the lwb write completes, if "zfs_nocacheflush" is set.
1029 * The intention is for this function to be called as soon as the
1030 * contents of an lwb are considered "stable" on disk, and will survive
1031 * any sudden loss of power. At this point, any threads waiting for the
1032 * lwb to reach this state are signalled, and the "waiter" structures
1033 * are marked "done".
1036 zil_lwb_flush_vdevs_done(zio_t *zio)
1038 lwb_t *lwb = zio->io_private;
1039 zilog_t *zilog = lwb->lwb_zilog;
1040 dmu_tx_t *tx = lwb->lwb_tx;
1041 zil_commit_waiter_t *zcw;
1043 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1045 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1047 mutex_enter(&zilog->zl_lock);
1050 * Ensure the lwb buffer pointer is cleared before releasing the
1051 * txg. If we have had an allocation failure and the txg is
1052 * waiting to sync then we want zil_sync() to remove the lwb so
1053 * that it's not picked up as the next new one in
1054 * zil_process_commit_list(). zil_sync() will only remove the
1055 * lwb if lwb_buf is null.
1057 lwb->lwb_buf = NULL;
1060 ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1061 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1063 lwb->lwb_root_zio = NULL;
1064 lwb->lwb_state = LWB_STATE_DONE;
1066 if (zilog->zl_last_lwb_opened == lwb) {
1068 * Remember the highest committed log sequence number
1069 * for ztest. We only update this value when all the log
1070 * writes succeeded, because ztest wants to ASSERT that
1071 * it got the whole log chain.
1073 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1076 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1077 mutex_enter(&zcw->zcw_lock);
1079 ASSERT(list_link_active(&zcw->zcw_node));
1080 list_remove(&lwb->lwb_waiters, zcw);
1082 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1083 zcw->zcw_lwb = NULL;
1085 zcw->zcw_zio_error = zio->io_error;
1087 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1088 zcw->zcw_done = B_TRUE;
1089 cv_broadcast(&zcw->zcw_cv);
1091 mutex_exit(&zcw->zcw_lock);
1094 mutex_exit(&zilog->zl_lock);
1097 * Now that we've written this log block, we have a stable pointer
1098 * to the next block in the chain, so it's OK to let the txg in
1099 * which we allocated the next block sync.
1105 * This is called when an lwb write completes. This means, this specific
1106 * lwb was written to disk, and all dependent lwb have also been
1109 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1110 * the VDEVs involved in writing out this specific lwb. The lwb will be
1111 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1112 * zio completion callback for the lwb's root zio.
1115 zil_lwb_write_done(zio_t *zio)
1117 lwb_t *lwb = zio->io_private;
1118 spa_t *spa = zio->io_spa;
1119 zilog_t *zilog = lwb->lwb_zilog;
1120 avl_tree_t *t = &lwb->lwb_vdev_tree;
1121 void *cookie = NULL;
1122 zil_vdev_node_t *zv;
1124 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1126 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1127 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1128 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1129 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1130 ASSERT(!BP_IS_GANG(zio->io_bp));
1131 ASSERT(!BP_IS_HOLE(zio->io_bp));
1132 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1134 abd_put(zio->io_abd);
1136 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1138 mutex_enter(&zilog->zl_lock);
1139 lwb->lwb_write_zio = NULL;
1140 mutex_exit(&zilog->zl_lock);
1142 if (avl_numnodes(t) == 0)
1146 * If there was an IO error, we're not going to call zio_flush()
1147 * on these vdevs, so we simply empty the tree and free the
1148 * nodes. We avoid calling zio_flush() since there isn't any
1149 * good reason for doing so, after the lwb block failed to be
1152 if (zio->io_error != 0) {
1153 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1154 kmem_free(zv, sizeof (*zv));
1158 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1159 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1161 zio_flush(lwb->lwb_root_zio, vd);
1162 kmem_free(zv, sizeof (*zv));
1167 * This function's purpose is to "open" an lwb such that it is ready to
1168 * accept new itxs being committed to it. To do this, the lwb's zio
1169 * structures are created, and linked to the lwb. This function is
1170 * idempotent; if the passed in lwb has already been opened, this
1171 * function is essentially a no-op.
1174 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1176 zbookmark_phys_t zb;
1177 zio_priority_t prio;
1179 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1180 ASSERT3P(lwb, !=, NULL);
1181 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1182 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1184 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1185 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1186 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1188 if (lwb->lwb_root_zio == NULL) {
1189 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1190 BP_GET_LSIZE(&lwb->lwb_blk));
1192 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1193 prio = ZIO_PRIORITY_SYNC_WRITE;
1195 prio = ZIO_PRIORITY_ASYNC_WRITE;
1197 lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1198 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1199 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1201 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1202 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1203 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1204 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1205 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1207 lwb->lwb_state = LWB_STATE_OPENED;
1209 mutex_enter(&zilog->zl_lock);
1212 * The zilog's "zl_last_lwb_opened" field is used to
1213 * build the lwb/zio dependency chain, which is used to
1214 * preserve the ordering of lwb completions that is
1215 * required by the semantics of the ZIL. Each new lwb
1216 * zio becomes a parent of the "previous" lwb zio, such
1217 * that the new lwb's zio cannot complete until the
1218 * "previous" lwb's zio completes.
1220 * This is required by the semantics of zil_commit();
1221 * the commit waiters attached to the lwbs will be woken
1222 * in the lwb zio's completion callback, so this zio
1223 * dependency graph ensures the waiters are woken in the
1224 * correct order (the same order the lwbs were created).
1226 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1227 if (last_lwb_opened != NULL &&
1228 last_lwb_opened->lwb_state != LWB_STATE_DONE) {
1229 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1230 last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1231 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1232 zio_add_child(lwb->lwb_root_zio,
1233 last_lwb_opened->lwb_root_zio);
1235 zilog->zl_last_lwb_opened = lwb;
1237 mutex_exit(&zilog->zl_lock);
1240 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1241 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1242 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1246 * Define a limited set of intent log block sizes.
1248 * These must be a multiple of 4KB. Note only the amount used (again
1249 * aligned to 4KB) actually gets written. However, we can't always just
1250 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1252 uint64_t zil_block_buckets[] = {
1253 4096, /* non TX_WRITE */
1254 8192+4096, /* data base */
1255 32*1024 + 4096, /* NFS writes */
1260 * Start a log block write and advance to the next log block.
1261 * Calls are serialized.
1264 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1268 spa_t *spa = zilog->zl_spa;
1272 uint64_t zil_blksz, wsz;
1276 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1277 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1278 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1279 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1281 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1282 zilc = (zil_chain_t *)lwb->lwb_buf;
1283 bp = &zilc->zc_next_blk;
1285 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1286 bp = &zilc->zc_next_blk;
1289 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1292 * Allocate the next block and save its address in this block
1293 * before writing it in order to establish the log chain.
1294 * Note that if the allocation of nlwb synced before we wrote
1295 * the block that points at it (lwb), we'd leak it if we crashed.
1296 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1297 * We dirty the dataset to ensure that zil_sync() will be called
1298 * to clean up in the event of allocation failure or I/O failure.
1301 tx = dmu_tx_create(zilog->zl_os);
1304 * Since we are not going to create any new dirty data, and we
1305 * can even help with clearing the existing dirty data, we
1306 * should not be subject to the dirty data based delays. We
1307 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1309 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1311 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1312 txg = dmu_tx_get_txg(tx);
1317 * Log blocks are pre-allocated. Here we select the size of the next
1318 * block, based on size used in the last block.
1319 * - first find the smallest bucket that will fit the block from a
1320 * limited set of block sizes. This is because it's faster to write
1321 * blocks allocated from the same metaslab as they are adjacent or
1323 * - next find the maximum from the new suggested size and an array of
1324 * previous sizes. This lessens a picket fence effect of wrongly
1325 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1328 * Note we only write what is used, but we can't just allocate
1329 * the maximum block size because we can exhaust the available
1332 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1333 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1335 zil_blksz = zil_block_buckets[i];
1336 if (zil_blksz == UINT64_MAX)
1337 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1338 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1339 for (i = 0; i < ZIL_PREV_BLKS; i++)
1340 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1341 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1345 /* pass the old blkptr in order to spread log blocks across devs */
1346 error = zio_alloc_zil(spa, zilog->zl_os->os_dsl_dataset->ds_object,
1347 txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1349 ASSERT3U(bp->blk_birth, ==, txg);
1350 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1351 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1354 * Allocate a new log write block (lwb).
1356 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1359 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1360 /* For Slim ZIL only write what is used. */
1361 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1362 ASSERT3U(wsz, <=, lwb->lwb_sz);
1363 zio_shrink(lwb->lwb_write_zio, wsz);
1370 zilc->zc_nused = lwb->lwb_nused;
1371 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1374 * clear unused data for security
1376 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1378 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1380 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1381 lwb->lwb_issued_timestamp = gethrtime();
1382 lwb->lwb_state = LWB_STATE_ISSUED;
1384 zio_nowait(lwb->lwb_root_zio);
1385 zio_nowait(lwb->lwb_write_zio);
1388 * If there was an allocation failure then nlwb will be null which
1389 * forces a txg_wait_synced().
1395 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1398 lr_write_t *lrwb, *lrw;
1400 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1402 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1403 ASSERT3P(lwb, !=, NULL);
1404 ASSERT3P(lwb->lwb_buf, !=, NULL);
1406 zil_lwb_write_open(zilog, lwb);
1409 lrw = (lr_write_t *)lrc;
1412 * A commit itx doesn't represent any on-disk state; instead
1413 * it's simply used as a place holder on the commit list, and
1414 * provides a mechanism for attaching a "commit waiter" onto the
1415 * correct lwb (such that the waiter can be signalled upon
1416 * completion of that lwb). Thus, we don't process this itx's
1417 * log record if it's a commit itx (these itx's don't have log
1418 * records), and instead link the itx's waiter onto the lwb's
1421 * For more details, see the comment above zil_commit().
1423 if (lrc->lrc_txtype == TX_COMMIT) {
1424 mutex_enter(&zilog->zl_lock);
1425 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1426 itx->itx_private = NULL;
1427 mutex_exit(&zilog->zl_lock);
1431 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1432 dlen = P2ROUNDUP_TYPED(
1433 lrw->lr_length, sizeof (uint64_t), uint64_t);
1437 reclen = lrc->lrc_reclen;
1438 zilog->zl_cur_used += (reclen + dlen);
1441 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1445 * If this record won't fit in the current log block, start a new one.
1446 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1448 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1449 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1450 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1451 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1452 lwb = zil_lwb_write_issue(zilog, lwb);
1455 zil_lwb_write_open(zilog, lwb);
1456 ASSERT(LWB_EMPTY(lwb));
1457 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1458 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1461 dnow = MIN(dlen, lwb_sp - reclen);
1462 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1463 bcopy(lrc, lr_buf, reclen);
1464 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1465 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1468 * If it's a write, fetch the data or get its blkptr as appropriate.
1470 if (lrc->lrc_txtype == TX_WRITE) {
1471 if (txg > spa_freeze_txg(zilog->zl_spa))
1472 txg_wait_synced(zilog->zl_dmu_pool, txg);
1473 if (itx->itx_wr_state != WR_COPIED) {
1477 if (itx->itx_wr_state == WR_NEED_COPY) {
1478 dbuf = lr_buf + reclen;
1479 lrcb->lrc_reclen += dnow;
1480 if (lrwb->lr_length > dnow)
1481 lrwb->lr_length = dnow;
1482 lrw->lr_offset += dnow;
1483 lrw->lr_length -= dnow;
1485 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1490 * We pass in the "lwb_write_zio" rather than
1491 * "lwb_root_zio" so that the "lwb_write_zio"
1492 * becomes the parent of any zio's created by
1493 * the "zl_get_data" callback. The vdevs are
1494 * flushed after the "lwb_write_zio" completes,
1495 * so we want to make sure that completion
1496 * callback waits for these additional zio's,
1497 * such that the vdevs used by those zio's will
1498 * be included in the lwb's vdev tree, and those
1499 * vdevs will be properly flushed. If we passed
1500 * in "lwb_root_zio" here, then these additional
1501 * vdevs may not be flushed; e.g. if these zio's
1502 * completed after "lwb_write_zio" completed.
1504 error = zilog->zl_get_data(itx->itx_private,
1505 lrwb, dbuf, lwb, lwb->lwb_write_zio);
1508 txg_wait_synced(zilog->zl_dmu_pool, txg);
1512 ASSERT(error == ENOENT || error == EEXIST ||
1520 * We're actually making an entry, so update lrc_seq to be the
1521 * log record sequence number. Note that this is generally not
1522 * equal to the itx sequence number because not all transactions
1523 * are synchronous, and sometimes spa_sync() gets there first.
1525 lrcb->lrc_seq = ++zilog->zl_lr_seq;
1526 lwb->lwb_nused += reclen + dnow;
1528 zil_lwb_add_txg(lwb, txg);
1530 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1531 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1535 zilog->zl_cur_used += reclen;
1543 zil_itx_create(uint64_t txtype, size_t lrsize)
1547 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1549 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1550 itx->itx_lr.lrc_txtype = txtype;
1551 itx->itx_lr.lrc_reclen = lrsize;
1552 itx->itx_lr.lrc_seq = 0; /* defensive */
1553 itx->itx_sync = B_TRUE; /* default is synchronous */
1559 zil_itx_destroy(itx_t *itx)
1561 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1565 * Free up the sync and async itxs. The itxs_t has already been detached
1566 * so no locks are needed.
1569 zil_itxg_clean(itxs_t *itxs)
1575 itx_async_node_t *ian;
1577 list = &itxs->i_sync_list;
1578 while ((itx = list_head(list)) != NULL) {
1580 * In the general case, commit itxs will not be found
1581 * here, as they'll be committed to an lwb via
1582 * zil_lwb_commit(), and free'd in that function. Having
1583 * said that, it is still possible for commit itxs to be
1584 * found here, due to the following race:
1586 * - a thread calls zil_commit() which assigns the
1587 * commit itx to a per-txg i_sync_list
1588 * - zil_itxg_clean() is called (e.g. via spa_sync())
1589 * while the waiter is still on the i_sync_list
1591 * There's nothing to prevent syncing the txg while the
1592 * waiter is on the i_sync_list. This normally doesn't
1593 * happen because spa_sync() is slower than zil_commit(),
1594 * but if zil_commit() calls txg_wait_synced() (e.g.
1595 * because zil_create() or zil_commit_writer_stall() is
1596 * called) we will hit this case.
1598 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1599 zil_commit_waiter_skip(itx->itx_private);
1601 list_remove(list, itx);
1602 zil_itx_destroy(itx);
1606 t = &itxs->i_async_tree;
1607 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1608 list = &ian->ia_list;
1609 while ((itx = list_head(list)) != NULL) {
1610 list_remove(list, itx);
1611 /* commit itxs should never be on the async lists. */
1612 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1613 zil_itx_destroy(itx);
1616 kmem_free(ian, sizeof (itx_async_node_t));
1620 kmem_free(itxs, sizeof (itxs_t));
1624 zil_aitx_compare(const void *x1, const void *x2)
1626 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1627 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1638 * Remove all async itx with the given oid.
1641 zil_remove_async(zilog_t *zilog, uint64_t oid)
1644 itx_async_node_t *ian;
1651 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1653 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1656 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1658 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1659 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1661 mutex_enter(&itxg->itxg_lock);
1662 if (itxg->itxg_txg != txg) {
1663 mutex_exit(&itxg->itxg_lock);
1668 * Locate the object node and append its list.
1670 t = &itxg->itxg_itxs->i_async_tree;
1671 ian = avl_find(t, &oid, &where);
1673 list_move_tail(&clean_list, &ian->ia_list);
1674 mutex_exit(&itxg->itxg_lock);
1676 while ((itx = list_head(&clean_list)) != NULL) {
1677 list_remove(&clean_list, itx);
1678 /* commit itxs should never be on the async lists. */
1679 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1680 zil_itx_destroy(itx);
1682 list_destroy(&clean_list);
1686 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1690 itxs_t *itxs, *clean = NULL;
1693 * Object ids can be re-instantiated in the next txg so
1694 * remove any async transactions to avoid future leaks.
1695 * This can happen if a fsync occurs on the re-instantiated
1696 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1697 * the new file data and flushes a write record for the old object.
1699 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1700 zil_remove_async(zilog, itx->itx_oid);
1703 * Ensure the data of a renamed file is committed before the rename.
1705 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1706 zil_async_to_sync(zilog, itx->itx_oid);
1708 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1711 txg = dmu_tx_get_txg(tx);
1713 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1714 mutex_enter(&itxg->itxg_lock);
1715 itxs = itxg->itxg_itxs;
1716 if (itxg->itxg_txg != txg) {
1719 * The zil_clean callback hasn't got around to cleaning
1720 * this itxg. Save the itxs for release below.
1721 * This should be rare.
1723 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1724 "txg %llu", itxg->itxg_txg);
1725 clean = itxg->itxg_itxs;
1727 itxg->itxg_txg = txg;
1728 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1730 list_create(&itxs->i_sync_list, sizeof (itx_t),
1731 offsetof(itx_t, itx_node));
1732 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1733 sizeof (itx_async_node_t),
1734 offsetof(itx_async_node_t, ia_node));
1736 if (itx->itx_sync) {
1737 list_insert_tail(&itxs->i_sync_list, itx);
1739 avl_tree_t *t = &itxs->i_async_tree;
1740 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1741 itx_async_node_t *ian;
1744 ian = avl_find(t, &foid, &where);
1746 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1747 list_create(&ian->ia_list, sizeof (itx_t),
1748 offsetof(itx_t, itx_node));
1749 ian->ia_foid = foid;
1750 avl_insert(t, ian, where);
1752 list_insert_tail(&ian->ia_list, itx);
1755 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1758 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1759 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1760 * need to be careful to always dirty the ZIL using the "real"
1761 * TXG (not itxg_txg) even when the SPA is frozen.
1763 zilog_dirty(zilog, dmu_tx_get_txg(tx));
1764 mutex_exit(&itxg->itxg_lock);
1766 /* Release the old itxs now we've dropped the lock */
1768 zil_itxg_clean(clean);
1772 * If there are any in-memory intent log transactions which have now been
1773 * synced then start up a taskq to free them. We should only do this after we
1774 * have written out the uberblocks (i.e. txg has been comitted) so that
1775 * don't inadvertently clean out in-memory log records that would be required
1779 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1781 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1784 ASSERT3U(synced_txg, <, ZILTEST_TXG);
1786 mutex_enter(&itxg->itxg_lock);
1787 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1788 mutex_exit(&itxg->itxg_lock);
1791 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1792 ASSERT3U(itxg->itxg_txg, !=, 0);
1793 clean_me = itxg->itxg_itxs;
1794 itxg->itxg_itxs = NULL;
1796 mutex_exit(&itxg->itxg_lock);
1798 * Preferably start a task queue to free up the old itxs but
1799 * if taskq_dispatch can't allocate resources to do that then
1800 * free it in-line. This should be rare. Note, using TQ_SLEEP
1801 * created a bad performance problem.
1803 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1804 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1805 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1806 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1807 zil_itxg_clean(clean_me);
1811 * This function will traverse the queue of itxs that need to be
1812 * committed, and move them onto the ZIL's zl_itx_commit_list.
1815 zil_get_commit_list(zilog_t *zilog)
1818 list_t *commit_list = &zilog->zl_itx_commit_list;
1820 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1822 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1825 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1828 * This is inherently racy, since there is nothing to prevent
1829 * the last synced txg from changing. That's okay since we'll
1830 * only commit things in the future.
1832 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1833 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1835 mutex_enter(&itxg->itxg_lock);
1836 if (itxg->itxg_txg != txg) {
1837 mutex_exit(&itxg->itxg_lock);
1842 * If we're adding itx records to the zl_itx_commit_list,
1843 * then the zil better be dirty in this "txg". We can assert
1844 * that here since we're holding the itxg_lock which will
1845 * prevent spa_sync from cleaning it. Once we add the itxs
1846 * to the zl_itx_commit_list we must commit it to disk even
1847 * if it's unnecessary (i.e. the txg was synced).
1849 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1850 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1851 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1853 mutex_exit(&itxg->itxg_lock);
1858 * Move the async itxs for a specified object to commit into sync lists.
1861 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1864 itx_async_node_t *ian;
1868 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1871 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1874 * This is inherently racy, since there is nothing to prevent
1875 * the last synced txg from changing.
1877 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1878 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1880 mutex_enter(&itxg->itxg_lock);
1881 if (itxg->itxg_txg != txg) {
1882 mutex_exit(&itxg->itxg_lock);
1887 * If a foid is specified then find that node and append its
1888 * list. Otherwise walk the tree appending all the lists
1889 * to the sync list. We add to the end rather than the
1890 * beginning to ensure the create has happened.
1892 t = &itxg->itxg_itxs->i_async_tree;
1894 ian = avl_find(t, &foid, &where);
1896 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1900 void *cookie = NULL;
1902 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1903 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1905 list_destroy(&ian->ia_list);
1906 kmem_free(ian, sizeof (itx_async_node_t));
1909 mutex_exit(&itxg->itxg_lock);
1914 * This function will prune commit itxs that are at the head of the
1915 * commit list (it won't prune past the first non-commit itx), and
1916 * either: a) attach them to the last lwb that's still pending
1917 * completion, or b) skip them altogether.
1919 * This is used as a performance optimization to prevent commit itxs
1920 * from generating new lwbs when it's unnecessary to do so.
1923 zil_prune_commit_list(zilog_t *zilog)
1927 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1929 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1930 lr_t *lrc = &itx->itx_lr;
1931 if (lrc->lrc_txtype != TX_COMMIT)
1934 mutex_enter(&zilog->zl_lock);
1936 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
1937 if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) {
1939 * All of the itxs this waiter was waiting on
1940 * must have already completed (or there were
1941 * never any itx's for it to wait on), so it's
1942 * safe to skip this waiter and mark it done.
1944 zil_commit_waiter_skip(itx->itx_private);
1946 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
1947 itx->itx_private = NULL;
1950 mutex_exit(&zilog->zl_lock);
1952 list_remove(&zilog->zl_itx_commit_list, itx);
1953 zil_itx_destroy(itx);
1956 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1960 zil_commit_writer_stall(zilog_t *zilog)
1963 * When zio_alloc_zil() fails to allocate the next lwb block on
1964 * disk, we must call txg_wait_synced() to ensure all of the
1965 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
1966 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
1967 * to zil_process_commit_list()) will have to call zil_create(),
1968 * and start a new ZIL chain.
1970 * Since zil_alloc_zil() failed, the lwb that was previously
1971 * issued does not have a pointer to the "next" lwb on disk.
1972 * Thus, if another ZIL writer thread was to allocate the "next"
1973 * on-disk lwb, that block could be leaked in the event of a
1974 * crash (because the previous lwb on-disk would not point to
1977 * We must hold the zilog's zl_issuer_lock while we do this, to
1978 * ensure no new threads enter zil_process_commit_list() until
1979 * all lwb's in the zl_lwb_list have been synced and freed
1980 * (which is achieved via the txg_wait_synced() call).
1982 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1983 txg_wait_synced(zilog->zl_dmu_pool, 0);
1984 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
1988 * This function will traverse the commit list, creating new lwbs as
1989 * needed, and committing the itxs from the commit list to these newly
1990 * created lwbs. Additionally, as a new lwb is created, the previous
1991 * lwb will be issued to the zio layer to be written to disk.
1994 zil_process_commit_list(zilog_t *zilog)
1996 spa_t *spa = zilog->zl_spa;
1997 list_t nolwb_waiters;
2001 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2004 * Return if there's nothing to commit before we dirty the fs by
2005 * calling zil_create().
2007 if (list_head(&zilog->zl_itx_commit_list) == NULL)
2010 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2011 offsetof(zil_commit_waiter_t, zcw_node));
2013 lwb = list_tail(&zilog->zl_lwb_list);
2015 lwb = zil_create(zilog);
2017 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2018 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2021 while (itx = list_head(&zilog->zl_itx_commit_list)) {
2022 lr_t *lrc = &itx->itx_lr;
2023 uint64_t txg = lrc->lrc_txg;
2025 ASSERT3U(txg, !=, 0);
2027 if (lrc->lrc_txtype == TX_COMMIT) {
2028 DTRACE_PROBE2(zil__process__commit__itx,
2029 zilog_t *, zilog, itx_t *, itx);
2031 DTRACE_PROBE2(zil__process__normal__itx,
2032 zilog_t *, zilog, itx_t *, itx);
2035 boolean_t synced = txg <= spa_last_synced_txg(spa);
2036 boolean_t frozen = txg > spa_freeze_txg(spa);
2039 * If the txg of this itx has already been synced out, then
2040 * we don't need to commit this itx to an lwb. This is
2041 * because the data of this itx will have already been
2042 * written to the main pool. This is inherently racy, and
2043 * it's still ok to commit an itx whose txg has already
2044 * been synced; this will result in a write that's
2045 * unnecessary, but will do no harm.
2047 * With that said, we always want to commit TX_COMMIT itxs
2048 * to an lwb, regardless of whether or not that itx's txg
2049 * has been synced out. We do this to ensure any OPENED lwb
2050 * will always have at least one zil_commit_waiter_t linked
2053 * As a counter-example, if we skipped TX_COMMIT itx's
2054 * whose txg had already been synced, the following
2055 * situation could occur if we happened to be racing with
2058 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
2059 * itx's txg is 10 and the last synced txg is 9.
2060 * 2. spa_sync finishes syncing out txg 10.
2061 * 3. we move to the next itx in the list, it's a TX_COMMIT
2062 * whose txg is 10, so we skip it rather than committing
2063 * it to the lwb used in (1).
2065 * If the itx that is skipped in (3) is the last TX_COMMIT
2066 * itx in the commit list, than it's possible for the lwb
2067 * used in (1) to remain in the OPENED state indefinitely.
2069 * To prevent the above scenario from occuring, ensuring
2070 * that once an lwb is OPENED it will transition to ISSUED
2071 * and eventually DONE, we always commit TX_COMMIT itx's to
2072 * an lwb here, even if that itx's txg has already been
2075 * Finally, if the pool is frozen, we _always_ commit the
2076 * itx. The point of freezing the pool is to prevent data
2077 * from being written to the main pool via spa_sync, and
2078 * instead rely solely on the ZIL to persistently store the
2079 * data; i.e. when the pool is frozen, the last synced txg
2080 * value can't be trusted.
2082 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2084 lwb = zil_lwb_commit(zilog, itx, lwb);
2085 } else if (lrc->lrc_txtype == TX_COMMIT) {
2086 ASSERT3P(lwb, ==, NULL);
2087 zil_commit_waiter_link_nolwb(
2088 itx->itx_private, &nolwb_waiters);
2092 list_remove(&zilog->zl_itx_commit_list, itx);
2093 zil_itx_destroy(itx);
2098 * This indicates zio_alloc_zil() failed to allocate the
2099 * "next" lwb on-disk. When this happens, we must stall
2100 * the ZIL write pipeline; see the comment within
2101 * zil_commit_writer_stall() for more details.
2103 zil_commit_writer_stall(zilog);
2106 * Additionally, we have to signal and mark the "nolwb"
2107 * waiters as "done" here, since without an lwb, we
2108 * can't do this via zil_lwb_flush_vdevs_done() like
2111 zil_commit_waiter_t *zcw;
2112 while (zcw = list_head(&nolwb_waiters)) {
2113 zil_commit_waiter_skip(zcw);
2114 list_remove(&nolwb_waiters, zcw);
2117 ASSERT(list_is_empty(&nolwb_waiters));
2118 ASSERT3P(lwb, !=, NULL);
2119 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2120 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2123 * At this point, the ZIL block pointed at by the "lwb"
2124 * variable is in one of the following states: "closed"
2127 * If its "closed", then no itxs have been committed to
2128 * it, so there's no point in issuing its zio (i.e.
2131 * If its "open" state, then it contains one or more
2132 * itxs that eventually need to be committed to stable
2133 * storage. In this case we intentionally do not issue
2134 * the lwb's zio to disk yet, and instead rely on one of
2135 * the following two mechanisms for issuing the zio:
2137 * 1. Ideally, there will be more ZIL activity occuring
2138 * on the system, such that this function will be
2139 * immediately called again (not necessarily by the same
2140 * thread) and this lwb's zio will be issued via
2141 * zil_lwb_commit(). This way, the lwb is guaranteed to
2142 * be "full" when it is issued to disk, and we'll make
2143 * use of the lwb's size the best we can.
2145 * 2. If there isn't sufficient ZIL activity occuring on
2146 * the system, such that this lwb's zio isn't issued via
2147 * zil_lwb_commit(), zil_commit_waiter() will issue the
2148 * lwb's zio. If this occurs, the lwb is not guaranteed
2149 * to be "full" by the time its zio is issued, and means
2150 * the size of the lwb was "too large" given the amount
2151 * of ZIL activity occuring on the system at that time.
2153 * We do this for a couple of reasons:
2155 * 1. To try and reduce the number of IOPs needed to
2156 * write the same number of itxs. If an lwb has space
2157 * available in it's buffer for more itxs, and more itxs
2158 * will be committed relatively soon (relative to the
2159 * latency of performing a write), then it's beneficial
2160 * to wait for these "next" itxs. This way, more itxs
2161 * can be committed to stable storage with fewer writes.
2163 * 2. To try and use the largest lwb block size that the
2164 * incoming rate of itxs can support. Again, this is to
2165 * try and pack as many itxs into as few lwbs as
2166 * possible, without significantly impacting the latency
2167 * of each individual itx.
2173 * This function is responsible for ensuring the passed in commit waiter
2174 * (and associated commit itx) is committed to an lwb. If the waiter is
2175 * not already committed to an lwb, all itxs in the zilog's queue of
2176 * itxs will be processed. The assumption is the passed in waiter's
2177 * commit itx will found in the queue just like the other non-commit
2178 * itxs, such that when the entire queue is processed, the waiter will
2179 * have been commited to an lwb.
2181 * The lwb associated with the passed in waiter is not guaranteed to
2182 * have been issued by the time this function completes. If the lwb is
2183 * not issued, we rely on future calls to zil_commit_writer() to issue
2184 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2187 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2189 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2190 ASSERT(spa_writeable(zilog->zl_spa));
2192 mutex_enter(&zilog->zl_issuer_lock);
2194 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2196 * It's possible that, while we were waiting to acquire
2197 * the "zl_issuer_lock", another thread committed this
2198 * waiter to an lwb. If that occurs, we bail out early,
2199 * without processing any of the zilog's queue of itxs.
2201 * On certain workloads and system configurations, the
2202 * "zl_issuer_lock" can become highly contended. In an
2203 * attempt to reduce this contention, we immediately drop
2204 * the lock if the waiter has already been processed.
2206 * We've measured this optimization to reduce CPU spent
2207 * contending on this lock by up to 5%, using a system
2208 * with 32 CPUs, low latency storage (~50 usec writes),
2209 * and 1024 threads performing sync writes.
2214 zil_get_commit_list(zilog);
2215 zil_prune_commit_list(zilog);
2216 zil_process_commit_list(zilog);
2219 mutex_exit(&zilog->zl_issuer_lock);
2223 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2225 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2226 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2227 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2229 lwb_t *lwb = zcw->zcw_lwb;
2230 ASSERT3P(lwb, !=, NULL);
2231 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2234 * If the lwb has already been issued by another thread, we can
2235 * immediately return since there's no work to be done (the
2236 * point of this function is to issue the lwb). Additionally, we
2237 * do this prior to acquiring the zl_issuer_lock, to avoid
2238 * acquiring it when it's not necessary to do so.
2240 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2241 lwb->lwb_state == LWB_STATE_DONE)
2245 * In order to call zil_lwb_write_issue() we must hold the
2246 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2247 * since we're already holding the commit waiter's "zcw_lock",
2248 * and those two locks are aquired in the opposite order
2251 mutex_exit(&zcw->zcw_lock);
2252 mutex_enter(&zilog->zl_issuer_lock);
2253 mutex_enter(&zcw->zcw_lock);
2256 * Since we just dropped and re-acquired the commit waiter's
2257 * lock, we have to re-check to see if the waiter was marked
2258 * "done" during that process. If the waiter was marked "done",
2259 * the "lwb" pointer is no longer valid (it can be free'd after
2260 * the waiter is marked "done"), so without this check we could
2261 * wind up with a use-after-free error below.
2266 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2269 * We've already checked this above, but since we hadn't acquired
2270 * the zilog's zl_issuer_lock, we have to perform this check a
2271 * second time while holding the lock.
2273 * We don't need to hold the zl_lock since the lwb cannot transition
2274 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2275 * _can_ transition from ISSUED to DONE, but it's OK to race with
2276 * that transition since we treat the lwb the same, whether it's in
2277 * the ISSUED or DONE states.
2279 * The important thing, is we treat the lwb differently depending on
2280 * if it's ISSUED or OPENED, and block any other threads that might
2281 * attempt to issue this lwb. For that reason we hold the
2282 * zl_issuer_lock when checking the lwb_state; we must not call
2283 * zil_lwb_write_issue() if the lwb had already been issued.
2285 * See the comment above the lwb_state_t structure definition for
2286 * more details on the lwb states, and locking requirements.
2288 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2289 lwb->lwb_state == LWB_STATE_DONE)
2292 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2295 * As described in the comments above zil_commit_waiter() and
2296 * zil_process_commit_list(), we need to issue this lwb's zio
2297 * since we've reached the commit waiter's timeout and it still
2298 * hasn't been issued.
2300 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2302 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2305 * Since the lwb's zio hadn't been issued by the time this thread
2306 * reached its timeout, we reset the zilog's "zl_cur_used" field
2307 * to influence the zil block size selection algorithm.
2309 * By having to issue the lwb's zio here, it means the size of the
2310 * lwb was too large, given the incoming throughput of itxs. By
2311 * setting "zl_cur_used" to zero, we communicate this fact to the
2312 * block size selection algorithm, so it can take this informaiton
2313 * into account, and potentially select a smaller size for the
2314 * next lwb block that is allocated.
2316 zilog->zl_cur_used = 0;
2320 * When zil_lwb_write_issue() returns NULL, this
2321 * indicates zio_alloc_zil() failed to allocate the
2322 * "next" lwb on-disk. When this occurs, the ZIL write
2323 * pipeline must be stalled; see the comment within the
2324 * zil_commit_writer_stall() function for more details.
2326 * We must drop the commit waiter's lock prior to
2327 * calling zil_commit_writer_stall() or else we can wind
2328 * up with the following deadlock:
2330 * - This thread is waiting for the txg to sync while
2331 * holding the waiter's lock; txg_wait_synced() is
2332 * used within txg_commit_writer_stall().
2334 * - The txg can't sync because it is waiting for this
2335 * lwb's zio callback to call dmu_tx_commit().
2337 * - The lwb's zio callback can't call dmu_tx_commit()
2338 * because it's blocked trying to acquire the waiter's
2339 * lock, which occurs prior to calling dmu_tx_commit()
2341 mutex_exit(&zcw->zcw_lock);
2342 zil_commit_writer_stall(zilog);
2343 mutex_enter(&zcw->zcw_lock);
2347 mutex_exit(&zilog->zl_issuer_lock);
2348 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2352 * This function is responsible for performing the following two tasks:
2354 * 1. its primary responsibility is to block until the given "commit
2355 * waiter" is considered "done".
2357 * 2. its secondary responsibility is to issue the zio for the lwb that
2358 * the given "commit waiter" is waiting on, if this function has
2359 * waited "long enough" and the lwb is still in the "open" state.
2361 * Given a sufficient amount of itxs being generated and written using
2362 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2363 * function. If this does not occur, this secondary responsibility will
2364 * ensure the lwb is issued even if there is not other synchronous
2365 * activity on the system.
2367 * For more details, see zil_process_commit_list(); more specifically,
2368 * the comment at the bottom of that function.
2371 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2373 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2374 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2375 ASSERT(spa_writeable(zilog->zl_spa));
2377 mutex_enter(&zcw->zcw_lock);
2380 * The timeout is scaled based on the lwb latency to avoid
2381 * significantly impacting the latency of each individual itx.
2382 * For more details, see the comment at the bottom of the
2383 * zil_process_commit_list() function.
2385 int pct = MAX(zfs_commit_timeout_pct, 1);
2386 #if defined(illumos) || !defined(_KERNEL)
2387 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2388 hrtime_t wakeup = gethrtime() + sleep;
2390 sbintime_t sleep = nstosbt((zilog->zl_last_lwb_latency * pct) / 100);
2391 sbintime_t wakeup = getsbinuptime() + sleep;
2393 boolean_t timedout = B_FALSE;
2395 while (!zcw->zcw_done) {
2396 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2398 lwb_t *lwb = zcw->zcw_lwb;
2401 * Usually, the waiter will have a non-NULL lwb field here,
2402 * but it's possible for it to be NULL as a result of
2403 * zil_commit() racing with spa_sync().
2405 * When zil_clean() is called, it's possible for the itxg
2406 * list (which may be cleaned via a taskq) to contain
2407 * commit itxs. When this occurs, the commit waiters linked
2408 * off of these commit itxs will not be committed to an
2409 * lwb. Additionally, these commit waiters will not be
2410 * marked done until zil_commit_waiter_skip() is called via
2413 * Thus, it's possible for this commit waiter (i.e. the
2414 * "zcw" variable) to be found in this "in between" state;
2415 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2416 * been skipped, so it's "zcw_done" field is still B_FALSE.
2418 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2420 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2421 ASSERT3B(timedout, ==, B_FALSE);
2424 * If the lwb hasn't been issued yet, then we
2425 * need to wait with a timeout, in case this
2426 * function needs to issue the lwb after the
2427 * timeout is reached; responsibility (2) from
2428 * the comment above this function.
2430 #if defined(illumos) || !defined(_KERNEL)
2431 clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2432 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2433 CALLOUT_FLAG_ABSOLUTE);
2435 if (timeleft >= 0 || zcw->zcw_done)
2438 int wait_err = cv_timedwait_sbt(&zcw->zcw_cv,
2439 &zcw->zcw_lock, wakeup, SBT_1NS, C_ABSOLUTE);
2440 if (wait_err != EWOULDBLOCK || zcw->zcw_done)
2445 zil_commit_waiter_timeout(zilog, zcw);
2447 if (!zcw->zcw_done) {
2449 * If the commit waiter has already been
2450 * marked "done", it's possible for the
2451 * waiter's lwb structure to have already
2452 * been freed. Thus, we can only reliably
2453 * make these assertions if the waiter
2456 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2457 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2461 * If the lwb isn't open, then it must have already
2462 * been issued. In that case, there's no need to
2463 * use a timeout when waiting for the lwb to
2466 * Additionally, if the lwb is NULL, the waiter
2467 * will soon be signalled and marked done via
2468 * zil_clean() and zil_itxg_clean(), so no timeout
2473 lwb->lwb_state == LWB_STATE_ISSUED ||
2474 lwb->lwb_state == LWB_STATE_DONE);
2475 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2479 mutex_exit(&zcw->zcw_lock);
2482 static zil_commit_waiter_t *
2483 zil_alloc_commit_waiter()
2485 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2487 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2488 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2489 list_link_init(&zcw->zcw_node);
2490 zcw->zcw_lwb = NULL;
2491 zcw->zcw_done = B_FALSE;
2492 zcw->zcw_zio_error = 0;
2498 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2500 ASSERT(!list_link_active(&zcw->zcw_node));
2501 ASSERT3P(zcw->zcw_lwb, ==, NULL);
2502 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2503 mutex_destroy(&zcw->zcw_lock);
2504 cv_destroy(&zcw->zcw_cv);
2505 kmem_cache_free(zil_zcw_cache, zcw);
2509 * This function is used to create a TX_COMMIT itx and assign it. This
2510 * way, it will be linked into the ZIL's list of synchronous itxs, and
2511 * then later committed to an lwb (or skipped) when
2512 * zil_process_commit_list() is called.
2515 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2517 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2518 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2520 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2521 itx->itx_sync = B_TRUE;
2522 itx->itx_private = zcw;
2524 zil_itx_assign(zilog, itx, tx);
2530 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2532 * When writing ZIL transactions to the on-disk representation of the
2533 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2534 * itxs can be committed to a single lwb. Once a lwb is written and
2535 * committed to stable storage (i.e. the lwb is written, and vdevs have
2536 * been flushed), each itx that was committed to that lwb is also
2537 * considered to be committed to stable storage.
2539 * When an itx is committed to an lwb, the log record (lr_t) contained
2540 * by the itx is copied into the lwb's zio buffer, and once this buffer
2541 * is written to disk, it becomes an on-disk ZIL block.
2543 * As itxs are generated, they're inserted into the ZIL's queue of
2544 * uncommitted itxs. The semantics of zil_commit() are such that it will
2545 * block until all itxs that were in the queue when it was called, are
2546 * committed to stable storage.
2548 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2549 * itxs, for all objects in the dataset, will be committed to stable
2550 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2551 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2552 * that correspond to the foid passed in, will be committed to stable
2553 * storage prior to zil_commit() returning.
2555 * Generally speaking, when zil_commit() is called, the consumer doesn't
2556 * actually care about _all_ of the uncommitted itxs. Instead, they're
2557 * simply trying to waiting for a specific itx to be committed to disk,
2558 * but the interface(s) for interacting with the ZIL don't allow such
2559 * fine-grained communication. A better interface would allow a consumer
2560 * to create and assign an itx, and then pass a reference to this itx to
2561 * zil_commit(); such that zil_commit() would return as soon as that
2562 * specific itx was committed to disk (instead of waiting for _all_
2563 * itxs to be committed).
2565 * When a thread calls zil_commit() a special "commit itx" will be
2566 * generated, along with a corresponding "waiter" for this commit itx.
2567 * zil_commit() will wait on this waiter's CV, such that when the waiter
2568 * is marked done, and signalled, zil_commit() will return.
2570 * This commit itx is inserted into the queue of uncommitted itxs. This
2571 * provides an easy mechanism for determining which itxs were in the
2572 * queue prior to zil_commit() having been called, and which itxs were
2573 * added after zil_commit() was called.
2575 * The commit it is special; it doesn't have any on-disk representation.
2576 * When a commit itx is "committed" to an lwb, the waiter associated
2577 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2578 * completes, each waiter on the lwb's list is marked done and signalled
2579 * -- allowing the thread waiting on the waiter to return from zil_commit().
2581 * It's important to point out a few critical factors that allow us
2582 * to make use of the commit itxs, commit waiters, per-lwb lists of
2583 * commit waiters, and zio completion callbacks like we're doing:
2585 * 1. The list of waiters for each lwb is traversed, and each commit
2586 * waiter is marked "done" and signalled, in the zio completion
2587 * callback of the lwb's zio[*].
2589 * * Actually, the waiters are signalled in the zio completion
2590 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2591 * that are sent to the vdevs upon completion of the lwb zio.
2593 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2594 * itxs, the order in which they are inserted is preserved[*]; as
2595 * itxs are added to the queue, they are added to the tail of
2596 * in-memory linked lists.
2598 * When committing the itxs to lwbs (to be written to disk), they
2599 * are committed in the same order in which the itxs were added to
2600 * the uncommitted queue's linked list(s); i.e. the linked list of
2601 * itxs to commit is traversed from head to tail, and each itx is
2602 * committed to an lwb in that order.
2606 * - the order of "sync" itxs is preserved w.r.t. other
2607 * "sync" itxs, regardless of the corresponding objects.
2608 * - the order of "async" itxs is preserved w.r.t. other
2609 * "async" itxs corresponding to the same object.
2610 * - the order of "async" itxs is *not* preserved w.r.t. other
2611 * "async" itxs corresponding to different objects.
2612 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2613 * versa) is *not* preserved, even for itxs that correspond
2614 * to the same object.
2616 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2617 * zil_get_commit_list(), and zil_process_commit_list().
2619 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2620 * lwb cannot be considered committed to stable storage, until its
2621 * "previous" lwb is also committed to stable storage. This fact,
2622 * coupled with the fact described above, means that itxs are
2623 * committed in (roughly) the order in which they were generated.
2624 * This is essential because itxs are dependent on prior itxs.
2625 * Thus, we *must not* deem an itx as being committed to stable
2626 * storage, until *all* prior itxs have also been committed to
2629 * To enforce this ordering of lwb zio's, while still leveraging as
2630 * much of the underlying storage performance as possible, we rely
2631 * on two fundamental concepts:
2633 * 1. The creation and issuance of lwb zio's is protected by
2634 * the zilog's "zl_issuer_lock", which ensures only a single
2635 * thread is creating and/or issuing lwb's at a time
2636 * 2. The "previous" lwb is a child of the "current" lwb
2637 * (leveraging the zio parent-child depenency graph)
2639 * By relying on this parent-child zio relationship, we can have
2640 * many lwb zio's concurrently issued to the underlying storage,
2641 * but the order in which they complete will be the same order in
2642 * which they were created.
2645 zil_commit(zilog_t *zilog, uint64_t foid)
2648 * We should never attempt to call zil_commit on a snapshot for
2649 * a couple of reasons:
2651 * 1. A snapshot may never be modified, thus it cannot have any
2652 * in-flight itxs that would have modified the dataset.
2654 * 2. By design, when zil_commit() is called, a commit itx will
2655 * be assigned to this zilog; as a result, the zilog will be
2656 * dirtied. We must not dirty the zilog of a snapshot; there's
2657 * checks in the code that enforce this invariant, and will
2658 * cause a panic if it's not upheld.
2660 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2662 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2665 if (!spa_writeable(zilog->zl_spa)) {
2667 * If the SPA is not writable, there should never be any
2668 * pending itxs waiting to be committed to disk. If that
2669 * weren't true, we'd skip writing those itxs out, and
2670 * would break the sematics of zil_commit(); thus, we're
2671 * verifying that truth before we return to the caller.
2673 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2674 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2675 for (int i = 0; i < TXG_SIZE; i++)
2676 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2681 * If the ZIL is suspended, we don't want to dirty it by calling
2682 * zil_commit_itx_assign() below, nor can we write out
2683 * lwbs like would be done in zil_commit_write(). Thus, we
2684 * simply rely on txg_wait_synced() to maintain the necessary
2685 * semantics, and avoid calling those functions altogether.
2687 if (zilog->zl_suspend > 0) {
2688 txg_wait_synced(zilog->zl_dmu_pool, 0);
2692 zil_commit_impl(zilog, foid);
2696 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2699 * Move the "async" itxs for the specified foid to the "sync"
2700 * queues, such that they will be later committed (or skipped)
2701 * to an lwb when zil_process_commit_list() is called.
2703 * Since these "async" itxs must be committed prior to this
2704 * call to zil_commit returning, we must perform this operation
2705 * before we call zil_commit_itx_assign().
2707 zil_async_to_sync(zilog, foid);
2710 * We allocate a new "waiter" structure which will initially be
2711 * linked to the commit itx using the itx's "itx_private" field.
2712 * Since the commit itx doesn't represent any on-disk state,
2713 * when it's committed to an lwb, rather than copying the its
2714 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2715 * added to the lwb's list of waiters. Then, when the lwb is
2716 * committed to stable storage, each waiter in the lwb's list of
2717 * waiters will be marked "done", and signalled.
2719 * We must create the waiter and assign the commit itx prior to
2720 * calling zil_commit_writer(), or else our specific commit itx
2721 * is not guaranteed to be committed to an lwb prior to calling
2722 * zil_commit_waiter().
2724 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2725 zil_commit_itx_assign(zilog, zcw);
2727 zil_commit_writer(zilog, zcw);
2728 zil_commit_waiter(zilog, zcw);
2730 if (zcw->zcw_zio_error != 0) {
2732 * If there was an error writing out the ZIL blocks that
2733 * this thread is waiting on, then we fallback to
2734 * relying on spa_sync() to write out the data this
2735 * thread is waiting on. Obviously this has performance
2736 * implications, but the expectation is for this to be
2737 * an exceptional case, and shouldn't occur often.
2739 DTRACE_PROBE2(zil__commit__io__error,
2740 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2741 txg_wait_synced(zilog->zl_dmu_pool, 0);
2744 zil_free_commit_waiter(zcw);
2748 * Called in syncing context to free committed log blocks and update log header.
2751 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2753 zil_header_t *zh = zil_header_in_syncing_context(zilog);
2754 uint64_t txg = dmu_tx_get_txg(tx);
2755 spa_t *spa = zilog->zl_spa;
2756 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2760 * We don't zero out zl_destroy_txg, so make sure we don't try
2761 * to destroy it twice.
2763 if (spa_sync_pass(spa) != 1)
2766 mutex_enter(&zilog->zl_lock);
2768 ASSERT(zilog->zl_stop_sync == 0);
2770 if (*replayed_seq != 0) {
2771 ASSERT(zh->zh_replay_seq < *replayed_seq);
2772 zh->zh_replay_seq = *replayed_seq;
2776 if (zilog->zl_destroy_txg == txg) {
2777 blkptr_t blk = zh->zh_log;
2779 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2781 bzero(zh, sizeof (zil_header_t));
2782 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2784 if (zilog->zl_keep_first) {
2786 * If this block was part of log chain that couldn't
2787 * be claimed because a device was missing during
2788 * zil_claim(), but that device later returns,
2789 * then this block could erroneously appear valid.
2790 * To guard against this, assign a new GUID to the new
2791 * log chain so it doesn't matter what blk points to.
2793 zil_init_log_chain(zilog, &blk);
2798 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2799 zh->zh_log = lwb->lwb_blk;
2800 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2802 list_remove(&zilog->zl_lwb_list, lwb);
2803 zio_free(spa, txg, &lwb->lwb_blk);
2804 zil_free_lwb(zilog, lwb);
2807 * If we don't have anything left in the lwb list then
2808 * we've had an allocation failure and we need to zero
2809 * out the zil_header blkptr so that we don't end
2810 * up freeing the same block twice.
2812 if (list_head(&zilog->zl_lwb_list) == NULL)
2813 BP_ZERO(&zh->zh_log);
2815 mutex_exit(&zilog->zl_lock);
2820 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2823 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2824 offsetof(zil_commit_waiter_t, zcw_node));
2825 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2826 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2827 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2833 zil_lwb_dest(void *vbuf, void *unused)
2836 mutex_destroy(&lwb->lwb_vdev_lock);
2837 avl_destroy(&lwb->lwb_vdev_tree);
2838 list_destroy(&lwb->lwb_waiters);
2844 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
2845 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
2847 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
2848 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2854 kmem_cache_destroy(zil_zcw_cache);
2855 kmem_cache_destroy(zil_lwb_cache);
2859 zil_set_sync(zilog_t *zilog, uint64_t sync)
2861 zilog->zl_sync = sync;
2865 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
2867 zilog->zl_logbias = logbias;
2871 zil_alloc(objset_t *os, zil_header_t *zh_phys)
2875 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
2877 zilog->zl_header = zh_phys;
2879 zilog->zl_spa = dmu_objset_spa(os);
2880 zilog->zl_dmu_pool = dmu_objset_pool(os);
2881 zilog->zl_destroy_txg = TXG_INITIAL - 1;
2882 zilog->zl_logbias = dmu_objset_logbias(os);
2883 zilog->zl_sync = dmu_objset_syncprop(os);
2884 zilog->zl_dirty_max_txg = 0;
2885 zilog->zl_last_lwb_opened = NULL;
2886 zilog->zl_last_lwb_latency = 0;
2888 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
2889 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
2891 for (int i = 0; i < TXG_SIZE; i++) {
2892 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
2893 MUTEX_DEFAULT, NULL);
2896 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
2897 offsetof(lwb_t, lwb_node));
2899 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
2900 offsetof(itx_t, itx_node));
2902 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
2908 zil_free(zilog_t *zilog)
2910 zilog->zl_stop_sync = 1;
2912 ASSERT0(zilog->zl_suspend);
2913 ASSERT0(zilog->zl_suspending);
2915 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2916 list_destroy(&zilog->zl_lwb_list);
2918 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
2919 list_destroy(&zilog->zl_itx_commit_list);
2921 for (int i = 0; i < TXG_SIZE; i++) {
2923 * It's possible for an itx to be generated that doesn't dirty
2924 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
2925 * callback to remove the entry. We remove those here.
2927 * Also free up the ziltest itxs.
2929 if (zilog->zl_itxg[i].itxg_itxs)
2930 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
2931 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
2934 mutex_destroy(&zilog->zl_issuer_lock);
2935 mutex_destroy(&zilog->zl_lock);
2937 cv_destroy(&zilog->zl_cv_suspend);
2939 kmem_free(zilog, sizeof (zilog_t));
2943 * Open an intent log.
2946 zil_open(objset_t *os, zil_get_data_t *get_data)
2948 zilog_t *zilog = dmu_objset_zil(os);
2950 ASSERT3P(zilog->zl_get_data, ==, NULL);
2951 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2952 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2954 zilog->zl_get_data = get_data;
2960 * Close an intent log.
2963 zil_close(zilog_t *zilog)
2968 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
2969 zil_commit(zilog, 0);
2971 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2972 ASSERT0(zilog->zl_dirty_max_txg);
2973 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
2976 mutex_enter(&zilog->zl_lock);
2977 lwb = list_tail(&zilog->zl_lwb_list);
2979 txg = zilog->zl_dirty_max_txg;
2981 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
2982 mutex_exit(&zilog->zl_lock);
2985 * We need to use txg_wait_synced() to wait long enough for the
2986 * ZIL to be clean, and to wait for all pending lwbs to be
2990 txg_wait_synced(zilog->zl_dmu_pool, txg);
2992 if (zilog_is_dirty(zilog))
2993 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
2994 VERIFY(!zilog_is_dirty(zilog));
2996 zilog->zl_get_data = NULL;
2999 * We should have only one lwb left on the list; remove it now.
3001 mutex_enter(&zilog->zl_lock);
3002 lwb = list_head(&zilog->zl_lwb_list);
3004 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3005 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3006 list_remove(&zilog->zl_lwb_list, lwb);
3007 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3008 zil_free_lwb(zilog, lwb);
3010 mutex_exit(&zilog->zl_lock);
3013 static char *suspend_tag = "zil suspending";
3016 * Suspend an intent log. While in suspended mode, we still honor
3017 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3018 * On old version pools, we suspend the log briefly when taking a
3019 * snapshot so that it will have an empty intent log.
3021 * Long holds are not really intended to be used the way we do here --
3022 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3023 * could fail. Therefore we take pains to only put a long hold if it is
3024 * actually necessary. Fortunately, it will only be necessary if the
3025 * objset is currently mounted (or the ZVOL equivalent). In that case it
3026 * will already have a long hold, so we are not really making things any worse.
3028 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3029 * zvol_state_t), and use their mechanism to prevent their hold from being
3030 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3033 * if cookiep == NULL, this does both the suspend & resume.
3034 * Otherwise, it returns with the dataset "long held", and the cookie
3035 * should be passed into zil_resume().
3038 zil_suspend(const char *osname, void **cookiep)
3042 const zil_header_t *zh;
3045 error = dmu_objset_hold(osname, suspend_tag, &os);
3048 zilog = dmu_objset_zil(os);
3050 mutex_enter(&zilog->zl_lock);
3051 zh = zilog->zl_header;
3053 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
3054 mutex_exit(&zilog->zl_lock);
3055 dmu_objset_rele(os, suspend_tag);
3056 return (SET_ERROR(EBUSY));
3060 * Don't put a long hold in the cases where we can avoid it. This
3061 * is when there is no cookie so we are doing a suspend & resume
3062 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3063 * for the suspend because it's already suspended, or there's no ZIL.
3065 if (cookiep == NULL && !zilog->zl_suspending &&
3066 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3067 mutex_exit(&zilog->zl_lock);
3068 dmu_objset_rele(os, suspend_tag);
3072 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3073 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3075 zilog->zl_suspend++;
3077 if (zilog->zl_suspend > 1) {
3079 * Someone else is already suspending it.
3080 * Just wait for them to finish.
3083 while (zilog->zl_suspending)
3084 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3085 mutex_exit(&zilog->zl_lock);
3087 if (cookiep == NULL)
3095 * If there is no pointer to an on-disk block, this ZIL must not
3096 * be active (e.g. filesystem not mounted), so there's nothing
3099 if (BP_IS_HOLE(&zh->zh_log)) {
3100 ASSERT(cookiep != NULL); /* fast path already handled */
3103 mutex_exit(&zilog->zl_lock);
3107 zilog->zl_suspending = B_TRUE;
3108 mutex_exit(&zilog->zl_lock);
3111 * We need to use zil_commit_impl to ensure we wait for all
3112 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3113 * to disk before proceeding. If we used zil_commit instead, it
3114 * would just call txg_wait_synced(), because zl_suspend is set.
3115 * txg_wait_synced() doesn't wait for these lwb's to be
3116 * LWB_STATE_DONE before returning.
3118 zil_commit_impl(zilog, 0);
3121 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
3122 * txg_wait_synced() to ensure the data from the zilog has
3123 * migrated to the main pool before calling zil_destroy().
3125 txg_wait_synced(zilog->zl_dmu_pool, 0);
3127 zil_destroy(zilog, B_FALSE);
3129 mutex_enter(&zilog->zl_lock);
3130 zilog->zl_suspending = B_FALSE;
3131 cv_broadcast(&zilog->zl_cv_suspend);
3132 mutex_exit(&zilog->zl_lock);
3134 if (cookiep == NULL)
3142 zil_resume(void *cookie)
3144 objset_t *os = cookie;
3145 zilog_t *zilog = dmu_objset_zil(os);
3147 mutex_enter(&zilog->zl_lock);
3148 ASSERT(zilog->zl_suspend != 0);
3149 zilog->zl_suspend--;
3150 mutex_exit(&zilog->zl_lock);
3151 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3152 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3155 typedef struct zil_replay_arg {
3156 zil_replay_func_t **zr_replay;
3158 boolean_t zr_byteswap;
3163 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3165 char name[ZFS_MAX_DATASET_NAME_LEN];
3167 zilog->zl_replaying_seq--; /* didn't actually replay this one */
3169 dmu_objset_name(zilog->zl_os, name);
3171 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3172 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3173 (u_longlong_t)lr->lrc_seq,
3174 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3175 (lr->lrc_txtype & TX_CI) ? "CI" : "");
3181 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3183 zil_replay_arg_t *zr = zra;
3184 const zil_header_t *zh = zilog->zl_header;
3185 uint64_t reclen = lr->lrc_reclen;
3186 uint64_t txtype = lr->lrc_txtype;
3189 zilog->zl_replaying_seq = lr->lrc_seq;
3191 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
3194 if (lr->lrc_txg < claim_txg) /* already committed */
3197 /* Strip case-insensitive bit, still present in log record */
3200 if (txtype == 0 || txtype >= TX_MAX_TYPE)
3201 return (zil_replay_error(zilog, lr, EINVAL));
3204 * If this record type can be logged out of order, the object
3205 * (lr_foid) may no longer exist. That's legitimate, not an error.
3207 if (TX_OOO(txtype)) {
3208 error = dmu_object_info(zilog->zl_os,
3209 ((lr_ooo_t *)lr)->lr_foid, NULL);
3210 if (error == ENOENT || error == EEXIST)
3215 * Make a copy of the data so we can revise and extend it.
3217 bcopy(lr, zr->zr_lr, reclen);
3220 * If this is a TX_WRITE with a blkptr, suck in the data.
3222 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3223 error = zil_read_log_data(zilog, (lr_write_t *)lr,
3224 zr->zr_lr + reclen);
3226 return (zil_replay_error(zilog, lr, error));
3230 * The log block containing this lr may have been byteswapped
3231 * so that we can easily examine common fields like lrc_txtype.
3232 * However, the log is a mix of different record types, and only the
3233 * replay vectors know how to byteswap their records. Therefore, if
3234 * the lr was byteswapped, undo it before invoking the replay vector.
3236 if (zr->zr_byteswap)
3237 byteswap_uint64_array(zr->zr_lr, reclen);
3240 * We must now do two things atomically: replay this log record,
3241 * and update the log header sequence number to reflect the fact that
3242 * we did so. At the end of each replay function the sequence number
3243 * is updated if we are in replay mode.
3245 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3248 * The DMU's dnode layer doesn't see removes until the txg
3249 * commits, so a subsequent claim can spuriously fail with
3250 * EEXIST. So if we receive any error we try syncing out
3251 * any removes then retry the transaction. Note that we
3252 * specify B_FALSE for byteswap now, so we don't do it twice.
3254 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3255 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3257 return (zil_replay_error(zilog, lr, error));
3264 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3266 zilog->zl_replay_blks++;
3272 * If this dataset has a non-empty intent log, replay it and destroy it.
3275 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3277 zilog_t *zilog = dmu_objset_zil(os);
3278 const zil_header_t *zh = zilog->zl_header;
3279 zil_replay_arg_t zr;
3281 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3282 zil_destroy(zilog, B_TRUE);
3286 zr.zr_replay = replay_func;
3288 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3289 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3292 * Wait for in-progress removes to sync before starting replay.
3294 txg_wait_synced(zilog->zl_dmu_pool, 0);
3296 zilog->zl_replay = B_TRUE;
3297 zilog->zl_replay_time = ddi_get_lbolt();
3298 ASSERT(zilog->zl_replay_blks == 0);
3299 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3301 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3303 zil_destroy(zilog, B_FALSE);
3304 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3305 zilog->zl_replay = B_FALSE;
3309 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3311 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3314 if (zilog->zl_replay) {
3315 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3316 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3317 zilog->zl_replaying_seq;
3326 zil_reset(const char *osname, void *arg)
3330 error = zil_suspend(osname, NULL);
3332 return (SET_ERROR(EEXIST));