4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
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, txg, &blk, NULL,
669 ZIL_MIN_BLKSZ, &slog);
672 zil_init_log_chain(zilog, &blk);
676 * Allocate a log write block (lwb) for the first log block.
679 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
682 * If we just allocated the first log block, commit our transaction
683 * and wait for zil_sync() to stuff the block poiner into zh_log.
684 * (zh is part of the MOS, so we cannot modify it in open context.)
688 txg_wait_synced(zilog->zl_dmu_pool, txg);
691 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
697 * In one tx, free all log blocks and clear the log header. If keep_first
698 * is set, then we're replaying a log with no content. We want to keep the
699 * first block, however, so that the first synchronous transaction doesn't
700 * require a txg_wait_synced() in zil_create(). We don't need to
701 * txg_wait_synced() here either when keep_first is set, because both
702 * zil_create() and zil_destroy() will wait for any in-progress destroys
706 zil_destroy(zilog_t *zilog, boolean_t keep_first)
708 const zil_header_t *zh = zilog->zl_header;
714 * Wait for any previous destroy to complete.
716 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
718 zilog->zl_old_header = *zh; /* debugging aid */
720 if (BP_IS_HOLE(&zh->zh_log))
723 tx = dmu_tx_create(zilog->zl_os);
724 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
725 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
726 txg = dmu_tx_get_txg(tx);
728 mutex_enter(&zilog->zl_lock);
730 ASSERT3U(zilog->zl_destroy_txg, <, txg);
731 zilog->zl_destroy_txg = txg;
732 zilog->zl_keep_first = keep_first;
734 if (!list_is_empty(&zilog->zl_lwb_list)) {
735 ASSERT(zh->zh_claim_txg == 0);
737 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
738 list_remove(&zilog->zl_lwb_list, lwb);
739 if (lwb->lwb_buf != NULL)
740 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
741 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
742 zil_free_lwb(zilog, lwb);
744 } else if (!keep_first) {
745 zil_destroy_sync(zilog, tx);
747 mutex_exit(&zilog->zl_lock);
753 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
755 ASSERT(list_is_empty(&zilog->zl_lwb_list));
756 (void) zil_parse(zilog, zil_free_log_block,
757 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
761 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
763 dmu_tx_t *tx = txarg;
770 error = dmu_objset_own_obj(dp, ds->ds_object,
771 DMU_OST_ANY, B_FALSE, FTAG, &os);
774 * EBUSY indicates that the objset is inconsistent, in which
775 * case it can not have a ZIL.
777 if (error != EBUSY) {
778 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
779 (unsigned long long)ds->ds_object, error);
784 zilog = dmu_objset_zil(os);
785 zh = zil_header_in_syncing_context(zilog);
786 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
787 first_txg = spa_min_claim_txg(zilog->zl_spa);
790 * If the spa_log_state is not set to be cleared, check whether
791 * the current uberblock is a checkpoint one and if the current
792 * header has been claimed before moving on.
794 * If the current uberblock is a checkpointed uberblock then
795 * one of the following scenarios took place:
797 * 1] We are currently rewinding to the checkpoint of the pool.
798 * 2] We crashed in the middle of a checkpoint rewind but we
799 * did manage to write the checkpointed uberblock to the
800 * vdev labels, so when we tried to import the pool again
801 * the checkpointed uberblock was selected from the import
804 * In both cases we want to zero out all the ZIL blocks, except
805 * the ones that have been claimed at the time of the checkpoint
806 * (their zh_claim_txg != 0). The reason is that these blocks
807 * may be corrupted since we may have reused their locations on
808 * disk after we took the checkpoint.
810 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
811 * when we first figure out whether the current uberblock is
812 * checkpointed or not. Unfortunately, that would discard all
813 * the logs, including the ones that are claimed, and we would
816 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
817 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
818 zh->zh_claim_txg == 0)) {
819 if (!BP_IS_HOLE(&zh->zh_log)) {
820 (void) zil_parse(zilog, zil_clear_log_block,
821 zil_noop_log_record, tx, first_txg);
823 BP_ZERO(&zh->zh_log);
824 dsl_dataset_dirty(dmu_objset_ds(os), tx);
825 dmu_objset_disown(os, FTAG);
830 * If we are not rewinding and opening the pool normally, then
831 * the min_claim_txg should be equal to the first txg of the pool.
833 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
836 * Claim all log blocks if we haven't already done so, and remember
837 * the highest claimed sequence number. This ensures that if we can
838 * read only part of the log now (e.g. due to a missing device),
839 * but we can read the entire log later, we will not try to replay
840 * or destroy beyond the last block we successfully claimed.
842 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
843 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
844 (void) zil_parse(zilog, zil_claim_log_block,
845 zil_claim_log_record, tx, first_txg);
846 zh->zh_claim_txg = first_txg;
847 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
848 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
849 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
850 zh->zh_flags |= ZIL_REPLAY_NEEDED;
851 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
852 dsl_dataset_dirty(dmu_objset_ds(os), tx);
855 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
856 dmu_objset_disown(os, FTAG);
861 * Check the log by walking the log chain.
862 * Checksum errors are ok as they indicate the end of the chain.
863 * Any other error (no device or read failure) returns an error.
867 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
876 error = dmu_objset_from_ds(ds, &os);
878 cmn_err(CE_WARN, "can't open objset %llu, error %d",
879 (unsigned long long)ds->ds_object, error);
883 zilog = dmu_objset_zil(os);
884 bp = (blkptr_t *)&zilog->zl_header->zh_log;
886 if (!BP_IS_HOLE(bp)) {
888 boolean_t valid = B_TRUE;
891 * Check the first block and determine if it's on a log device
892 * which may have been removed or faulted prior to loading this
893 * pool. If so, there's no point in checking the rest of the
894 * log as its content should have already been synced to the
897 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
898 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
899 if (vd->vdev_islog && vdev_is_dead(vd))
900 valid = vdev_log_state_valid(vd);
901 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
907 * Check whether the current uberblock is checkpointed (e.g.
908 * we are rewinding) and whether the current header has been
909 * claimed or not. If it hasn't then skip verifying it. We
910 * do this because its ZIL blocks may be part of the pool's
911 * state before the rewind, which is no longer valid.
913 zil_header_t *zh = zil_header_in_syncing_context(zilog);
914 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
915 zh->zh_claim_txg == 0)
920 * Because tx == NULL, zil_claim_log_block() will not actually claim
921 * any blocks, but just determine whether it is possible to do so.
922 * In addition to checking the log chain, zil_claim_log_block()
923 * will invoke zio_claim() with a done func of spa_claim_notify(),
924 * which will update spa_max_claim_txg. See spa_load() for details.
926 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
927 zilog->zl_header->zh_claim_txg ? -1ULL :
928 spa_min_claim_txg(os->os_spa));
930 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
934 * When an itx is "skipped", this function is used to properly mark the
935 * waiter as "done, and signal any thread(s) waiting on it. An itx can
936 * be skipped (and not committed to an lwb) for a variety of reasons,
937 * one of them being that the itx was committed via spa_sync(), prior to
938 * it being committed to an lwb; this can happen if a thread calling
939 * zil_commit() is racing with spa_sync().
942 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
944 mutex_enter(&zcw->zcw_lock);
945 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
946 zcw->zcw_done = B_TRUE;
947 cv_broadcast(&zcw->zcw_cv);
948 mutex_exit(&zcw->zcw_lock);
952 * This function is used when the given waiter is to be linked into an
953 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
954 * At this point, the waiter will no longer be referenced by the itx,
955 * and instead, will be referenced by the lwb.
958 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
961 * The lwb_waiters field of the lwb is protected by the zilog's
962 * zl_lock, thus it must be held when calling this function.
964 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
966 mutex_enter(&zcw->zcw_lock);
967 ASSERT(!list_link_active(&zcw->zcw_node));
968 ASSERT3P(zcw->zcw_lwb, ==, NULL);
969 ASSERT3P(lwb, !=, NULL);
970 ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
971 lwb->lwb_state == LWB_STATE_ISSUED);
973 list_insert_tail(&lwb->lwb_waiters, zcw);
975 mutex_exit(&zcw->zcw_lock);
979 * This function is used when zio_alloc_zil() fails to allocate a ZIL
980 * block, and the given waiter must be linked to the "nolwb waiters"
981 * list inside of zil_process_commit_list().
984 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
986 mutex_enter(&zcw->zcw_lock);
987 ASSERT(!list_link_active(&zcw->zcw_node));
988 ASSERT3P(zcw->zcw_lwb, ==, NULL);
989 list_insert_tail(nolwb, zcw);
990 mutex_exit(&zcw->zcw_lock);
994 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
996 avl_tree_t *t = &lwb->lwb_vdev_tree;
998 zil_vdev_node_t *zv, zvsearch;
999 int ndvas = BP_GET_NDVAS(bp);
1002 if (zfs_nocacheflush)
1005 mutex_enter(&lwb->lwb_vdev_lock);
1006 for (i = 0; i < ndvas; i++) {
1007 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1008 if (avl_find(t, &zvsearch, &where) == NULL) {
1009 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1010 zv->zv_vdev = zvsearch.zv_vdev;
1011 avl_insert(t, zv, where);
1014 mutex_exit(&lwb->lwb_vdev_lock);
1018 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1020 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1024 * This function is a called after all VDEVs associated with a given lwb
1025 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1026 * as the lwb write completes, if "zfs_nocacheflush" is set.
1028 * The intention is for this function to be called as soon as the
1029 * contents of an lwb are considered "stable" on disk, and will survive
1030 * any sudden loss of power. At this point, any threads waiting for the
1031 * lwb to reach this state are signalled, and the "waiter" structures
1032 * are marked "done".
1035 zil_lwb_flush_vdevs_done(zio_t *zio)
1037 lwb_t *lwb = zio->io_private;
1038 zilog_t *zilog = lwb->lwb_zilog;
1039 dmu_tx_t *tx = lwb->lwb_tx;
1040 zil_commit_waiter_t *zcw;
1042 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1044 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1046 mutex_enter(&zilog->zl_lock);
1049 * Ensure the lwb buffer pointer is cleared before releasing the
1050 * txg. If we have had an allocation failure and the txg is
1051 * waiting to sync then we want zil_sync() to remove the lwb so
1052 * that it's not picked up as the next new one in
1053 * zil_process_commit_list(). zil_sync() will only remove the
1054 * lwb if lwb_buf is null.
1056 lwb->lwb_buf = NULL;
1059 ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1060 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1062 lwb->lwb_root_zio = NULL;
1063 lwb->lwb_state = LWB_STATE_DONE;
1065 if (zilog->zl_last_lwb_opened == lwb) {
1067 * Remember the highest committed log sequence number
1068 * for ztest. We only update this value when all the log
1069 * writes succeeded, because ztest wants to ASSERT that
1070 * it got the whole log chain.
1072 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1075 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1076 mutex_enter(&zcw->zcw_lock);
1078 ASSERT(list_link_active(&zcw->zcw_node));
1079 list_remove(&lwb->lwb_waiters, zcw);
1081 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1082 zcw->zcw_lwb = NULL;
1084 zcw->zcw_zio_error = zio->io_error;
1086 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1087 zcw->zcw_done = B_TRUE;
1088 cv_broadcast(&zcw->zcw_cv);
1090 mutex_exit(&zcw->zcw_lock);
1093 mutex_exit(&zilog->zl_lock);
1096 * Now that we've written this log block, we have a stable pointer
1097 * to the next block in the chain, so it's OK to let the txg in
1098 * which we allocated the next block sync.
1104 * This is called when an lwb write completes. This means, this specific
1105 * lwb was written to disk, and all dependent lwb have also been
1108 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1109 * the VDEVs involved in writing out this specific lwb. The lwb will be
1110 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1111 * zio completion callback for the lwb's root zio.
1114 zil_lwb_write_done(zio_t *zio)
1116 lwb_t *lwb = zio->io_private;
1117 spa_t *spa = zio->io_spa;
1118 zilog_t *zilog = lwb->lwb_zilog;
1119 avl_tree_t *t = &lwb->lwb_vdev_tree;
1120 void *cookie = NULL;
1121 zil_vdev_node_t *zv;
1123 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1125 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1126 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1127 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1128 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1129 ASSERT(!BP_IS_GANG(zio->io_bp));
1130 ASSERT(!BP_IS_HOLE(zio->io_bp));
1131 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1133 abd_put(zio->io_abd);
1135 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1137 mutex_enter(&zilog->zl_lock);
1138 lwb->lwb_write_zio = NULL;
1139 mutex_exit(&zilog->zl_lock);
1141 if (avl_numnodes(t) == 0)
1145 * If there was an IO error, we're not going to call zio_flush()
1146 * on these vdevs, so we simply empty the tree and free the
1147 * nodes. We avoid calling zio_flush() since there isn't any
1148 * good reason for doing so, after the lwb block failed to be
1151 if (zio->io_error != 0) {
1152 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1153 kmem_free(zv, sizeof (*zv));
1157 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1158 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1160 zio_flush(lwb->lwb_root_zio, vd);
1161 kmem_free(zv, sizeof (*zv));
1166 * This function's purpose is to "open" an lwb such that it is ready to
1167 * accept new itxs being committed to it. To do this, the lwb's zio
1168 * structures are created, and linked to the lwb. This function is
1169 * idempotent; if the passed in lwb has already been opened, this
1170 * function is essentially a no-op.
1173 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1175 zbookmark_phys_t zb;
1176 zio_priority_t prio;
1178 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1179 ASSERT3P(lwb, !=, NULL);
1180 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1181 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1183 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1184 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1185 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1187 if (lwb->lwb_root_zio == NULL) {
1188 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1189 BP_GET_LSIZE(&lwb->lwb_blk));
1191 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1192 prio = ZIO_PRIORITY_SYNC_WRITE;
1194 prio = ZIO_PRIORITY_ASYNC_WRITE;
1196 lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1197 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1198 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1200 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1201 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1202 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1203 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1204 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1206 lwb->lwb_state = LWB_STATE_OPENED;
1208 mutex_enter(&zilog->zl_lock);
1211 * The zilog's "zl_last_lwb_opened" field is used to
1212 * build the lwb/zio dependency chain, which is used to
1213 * preserve the ordering of lwb completions that is
1214 * required by the semantics of the ZIL. Each new lwb
1215 * zio becomes a parent of the "previous" lwb zio, such
1216 * that the new lwb's zio cannot complete until the
1217 * "previous" lwb's zio completes.
1219 * This is required by the semantics of zil_commit();
1220 * the commit waiters attached to the lwbs will be woken
1221 * in the lwb zio's completion callback, so this zio
1222 * dependency graph ensures the waiters are woken in the
1223 * correct order (the same order the lwbs were created).
1225 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1226 if (last_lwb_opened != NULL &&
1227 last_lwb_opened->lwb_state != LWB_STATE_DONE) {
1228 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1229 last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1230 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1231 zio_add_child(lwb->lwb_root_zio,
1232 last_lwb_opened->lwb_root_zio);
1234 zilog->zl_last_lwb_opened = lwb;
1236 mutex_exit(&zilog->zl_lock);
1239 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1240 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1241 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1245 * Define a limited set of intent log block sizes.
1247 * These must be a multiple of 4KB. Note only the amount used (again
1248 * aligned to 4KB) actually gets written. However, we can't always just
1249 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1251 uint64_t zil_block_buckets[] = {
1252 4096, /* non TX_WRITE */
1253 8192+4096, /* data base */
1254 32*1024 + 4096, /* NFS writes */
1259 * Start a log block write and advance to the next log block.
1260 * Calls are serialized.
1263 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1267 spa_t *spa = zilog->zl_spa;
1271 uint64_t zil_blksz, wsz;
1275 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1276 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1277 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1278 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1280 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1281 zilc = (zil_chain_t *)lwb->lwb_buf;
1282 bp = &zilc->zc_next_blk;
1284 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1285 bp = &zilc->zc_next_blk;
1288 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1291 * Allocate the next block and save its address in this block
1292 * before writing it in order to establish the log chain.
1293 * Note that if the allocation of nlwb synced before we wrote
1294 * the block that points at it (lwb), we'd leak it if we crashed.
1295 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1296 * We dirty the dataset to ensure that zil_sync() will be called
1297 * to clean up in the event of allocation failure or I/O failure.
1300 tx = dmu_tx_create(zilog->zl_os);
1303 * Since we are not going to create any new dirty data, and we
1304 * can even help with clearing the existing dirty data, we
1305 * should not be subject to the dirty data based delays. We
1306 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1308 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1310 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1311 txg = dmu_tx_get_txg(tx);
1316 * Log blocks are pre-allocated. Here we select the size of the next
1317 * block, based on size used in the last block.
1318 * - first find the smallest bucket that will fit the block from a
1319 * limited set of block sizes. This is because it's faster to write
1320 * blocks allocated from the same metaslab as they are adjacent or
1322 * - next find the maximum from the new suggested size and an array of
1323 * previous sizes. This lessens a picket fence effect of wrongly
1324 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1327 * Note we only write what is used, but we can't just allocate
1328 * the maximum block size because we can exhaust the available
1331 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1332 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1334 zil_blksz = zil_block_buckets[i];
1335 if (zil_blksz == UINT64_MAX)
1336 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1337 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1338 for (i = 0; i < ZIL_PREV_BLKS; i++)
1339 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1340 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1344 /* pass the old blkptr in order to spread log blocks across devs */
1345 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1347 ASSERT3U(bp->blk_birth, ==, txg);
1348 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1349 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1352 * Allocate a new log write block (lwb).
1354 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1357 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1358 /* For Slim ZIL only write what is used. */
1359 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1360 ASSERT3U(wsz, <=, lwb->lwb_sz);
1361 zio_shrink(lwb->lwb_write_zio, wsz);
1368 zilc->zc_nused = lwb->lwb_nused;
1369 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1372 * clear unused data for security
1374 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1376 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1378 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1379 lwb->lwb_issued_timestamp = gethrtime();
1380 lwb->lwb_state = LWB_STATE_ISSUED;
1382 zio_nowait(lwb->lwb_root_zio);
1383 zio_nowait(lwb->lwb_write_zio);
1386 * If there was an allocation failure then nlwb will be null which
1387 * forces a txg_wait_synced().
1393 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1396 lr_write_t *lrwb, *lrw;
1398 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1400 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1401 ASSERT3P(lwb, !=, NULL);
1402 ASSERT3P(lwb->lwb_buf, !=, NULL);
1404 zil_lwb_write_open(zilog, lwb);
1407 lrw = (lr_write_t *)lrc;
1410 * A commit itx doesn't represent any on-disk state; instead
1411 * it's simply used as a place holder on the commit list, and
1412 * provides a mechanism for attaching a "commit waiter" onto the
1413 * correct lwb (such that the waiter can be signalled upon
1414 * completion of that lwb). Thus, we don't process this itx's
1415 * log record if it's a commit itx (these itx's don't have log
1416 * records), and instead link the itx's waiter onto the lwb's
1419 * For more details, see the comment above zil_commit().
1421 if (lrc->lrc_txtype == TX_COMMIT) {
1422 mutex_enter(&zilog->zl_lock);
1423 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1424 itx->itx_private = NULL;
1425 mutex_exit(&zilog->zl_lock);
1429 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1430 dlen = P2ROUNDUP_TYPED(
1431 lrw->lr_length, sizeof (uint64_t), uint64_t);
1435 reclen = lrc->lrc_reclen;
1436 zilog->zl_cur_used += (reclen + dlen);
1439 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1443 * If this record won't fit in the current log block, start a new one.
1444 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1446 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1447 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1448 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1449 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1450 lwb = zil_lwb_write_issue(zilog, lwb);
1453 zil_lwb_write_open(zilog, lwb);
1454 ASSERT(LWB_EMPTY(lwb));
1455 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1456 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1459 dnow = MIN(dlen, lwb_sp - reclen);
1460 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1461 bcopy(lrc, lr_buf, reclen);
1462 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1463 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1466 * If it's a write, fetch the data or get its blkptr as appropriate.
1468 if (lrc->lrc_txtype == TX_WRITE) {
1469 if (txg > spa_freeze_txg(zilog->zl_spa))
1470 txg_wait_synced(zilog->zl_dmu_pool, txg);
1471 if (itx->itx_wr_state != WR_COPIED) {
1475 if (itx->itx_wr_state == WR_NEED_COPY) {
1476 dbuf = lr_buf + reclen;
1477 lrcb->lrc_reclen += dnow;
1478 if (lrwb->lr_length > dnow)
1479 lrwb->lr_length = dnow;
1480 lrw->lr_offset += dnow;
1481 lrw->lr_length -= dnow;
1483 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1488 * We pass in the "lwb_write_zio" rather than
1489 * "lwb_root_zio" so that the "lwb_write_zio"
1490 * becomes the parent of any zio's created by
1491 * the "zl_get_data" callback. The vdevs are
1492 * flushed after the "lwb_write_zio" completes,
1493 * so we want to make sure that completion
1494 * callback waits for these additional zio's,
1495 * such that the vdevs used by those zio's will
1496 * be included in the lwb's vdev tree, and those
1497 * vdevs will be properly flushed. If we passed
1498 * in "lwb_root_zio" here, then these additional
1499 * vdevs may not be flushed; e.g. if these zio's
1500 * completed after "lwb_write_zio" completed.
1502 error = zilog->zl_get_data(itx->itx_private,
1503 lrwb, dbuf, lwb, lwb->lwb_write_zio);
1506 txg_wait_synced(zilog->zl_dmu_pool, txg);
1510 ASSERT(error == ENOENT || error == EEXIST ||
1518 * We're actually making an entry, so update lrc_seq to be the
1519 * log record sequence number. Note that this is generally not
1520 * equal to the itx sequence number because not all transactions
1521 * are synchronous, and sometimes spa_sync() gets there first.
1523 lrcb->lrc_seq = ++zilog->zl_lr_seq;
1524 lwb->lwb_nused += reclen + dnow;
1526 zil_lwb_add_txg(lwb, txg);
1528 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1529 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1533 zilog->zl_cur_used += reclen;
1541 zil_itx_create(uint64_t txtype, size_t lrsize)
1545 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1547 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1548 itx->itx_lr.lrc_txtype = txtype;
1549 itx->itx_lr.lrc_reclen = lrsize;
1550 itx->itx_lr.lrc_seq = 0; /* defensive */
1551 itx->itx_sync = B_TRUE; /* default is synchronous */
1557 zil_itx_destroy(itx_t *itx)
1559 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1563 * Free up the sync and async itxs. The itxs_t has already been detached
1564 * so no locks are needed.
1567 zil_itxg_clean(itxs_t *itxs)
1573 itx_async_node_t *ian;
1575 list = &itxs->i_sync_list;
1576 while ((itx = list_head(list)) != NULL) {
1578 * In the general case, commit itxs will not be found
1579 * here, as they'll be committed to an lwb via
1580 * zil_lwb_commit(), and free'd in that function. Having
1581 * said that, it is still possible for commit itxs to be
1582 * found here, due to the following race:
1584 * - a thread calls zil_commit() which assigns the
1585 * commit itx to a per-txg i_sync_list
1586 * - zil_itxg_clean() is called (e.g. via spa_sync())
1587 * while the waiter is still on the i_sync_list
1589 * There's nothing to prevent syncing the txg while the
1590 * waiter is on the i_sync_list. This normally doesn't
1591 * happen because spa_sync() is slower than zil_commit(),
1592 * but if zil_commit() calls txg_wait_synced() (e.g.
1593 * because zil_create() or zil_commit_writer_stall() is
1594 * called) we will hit this case.
1596 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1597 zil_commit_waiter_skip(itx->itx_private);
1599 list_remove(list, itx);
1600 zil_itx_destroy(itx);
1604 t = &itxs->i_async_tree;
1605 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1606 list = &ian->ia_list;
1607 while ((itx = list_head(list)) != NULL) {
1608 list_remove(list, itx);
1609 /* commit itxs should never be on the async lists. */
1610 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1611 zil_itx_destroy(itx);
1614 kmem_free(ian, sizeof (itx_async_node_t));
1618 kmem_free(itxs, sizeof (itxs_t));
1622 zil_aitx_compare(const void *x1, const void *x2)
1624 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1625 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1636 * Remove all async itx with the given oid.
1639 zil_remove_async(zilog_t *zilog, uint64_t oid)
1642 itx_async_node_t *ian;
1649 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1651 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1654 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1656 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1657 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1659 mutex_enter(&itxg->itxg_lock);
1660 if (itxg->itxg_txg != txg) {
1661 mutex_exit(&itxg->itxg_lock);
1666 * Locate the object node and append its list.
1668 t = &itxg->itxg_itxs->i_async_tree;
1669 ian = avl_find(t, &oid, &where);
1671 list_move_tail(&clean_list, &ian->ia_list);
1672 mutex_exit(&itxg->itxg_lock);
1674 while ((itx = list_head(&clean_list)) != NULL) {
1675 list_remove(&clean_list, itx);
1676 /* commit itxs should never be on the async lists. */
1677 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1678 zil_itx_destroy(itx);
1680 list_destroy(&clean_list);
1684 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1688 itxs_t *itxs, *clean = NULL;
1691 * Object ids can be re-instantiated in the next txg so
1692 * remove any async transactions to avoid future leaks.
1693 * This can happen if a fsync occurs on the re-instantiated
1694 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1695 * the new file data and flushes a write record for the old object.
1697 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1698 zil_remove_async(zilog, itx->itx_oid);
1701 * Ensure the data of a renamed file is committed before the rename.
1703 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1704 zil_async_to_sync(zilog, itx->itx_oid);
1706 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1709 txg = dmu_tx_get_txg(tx);
1711 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1712 mutex_enter(&itxg->itxg_lock);
1713 itxs = itxg->itxg_itxs;
1714 if (itxg->itxg_txg != txg) {
1717 * The zil_clean callback hasn't got around to cleaning
1718 * this itxg. Save the itxs for release below.
1719 * This should be rare.
1721 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1722 "txg %llu", itxg->itxg_txg);
1723 clean = itxg->itxg_itxs;
1725 itxg->itxg_txg = txg;
1726 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1728 list_create(&itxs->i_sync_list, sizeof (itx_t),
1729 offsetof(itx_t, itx_node));
1730 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1731 sizeof (itx_async_node_t),
1732 offsetof(itx_async_node_t, ia_node));
1734 if (itx->itx_sync) {
1735 list_insert_tail(&itxs->i_sync_list, itx);
1737 avl_tree_t *t = &itxs->i_async_tree;
1738 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1739 itx_async_node_t *ian;
1742 ian = avl_find(t, &foid, &where);
1744 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1745 list_create(&ian->ia_list, sizeof (itx_t),
1746 offsetof(itx_t, itx_node));
1747 ian->ia_foid = foid;
1748 avl_insert(t, ian, where);
1750 list_insert_tail(&ian->ia_list, itx);
1753 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1756 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1757 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1758 * need to be careful to always dirty the ZIL using the "real"
1759 * TXG (not itxg_txg) even when the SPA is frozen.
1761 zilog_dirty(zilog, dmu_tx_get_txg(tx));
1762 mutex_exit(&itxg->itxg_lock);
1764 /* Release the old itxs now we've dropped the lock */
1766 zil_itxg_clean(clean);
1770 * If there are any in-memory intent log transactions which have now been
1771 * synced then start up a taskq to free them. We should only do this after we
1772 * have written out the uberblocks (i.e. txg has been comitted) so that
1773 * don't inadvertently clean out in-memory log records that would be required
1777 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1779 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1782 ASSERT3U(synced_txg, <, ZILTEST_TXG);
1784 mutex_enter(&itxg->itxg_lock);
1785 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1786 mutex_exit(&itxg->itxg_lock);
1789 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1790 ASSERT3U(itxg->itxg_txg, !=, 0);
1791 clean_me = itxg->itxg_itxs;
1792 itxg->itxg_itxs = NULL;
1794 mutex_exit(&itxg->itxg_lock);
1796 * Preferably start a task queue to free up the old itxs but
1797 * if taskq_dispatch can't allocate resources to do that then
1798 * free it in-line. This should be rare. Note, using TQ_SLEEP
1799 * created a bad performance problem.
1801 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1802 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1803 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1804 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1805 zil_itxg_clean(clean_me);
1809 * This function will traverse the queue of itxs that need to be
1810 * committed, and move them onto the ZIL's zl_itx_commit_list.
1813 zil_get_commit_list(zilog_t *zilog)
1816 list_t *commit_list = &zilog->zl_itx_commit_list;
1818 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1820 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1823 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1826 * This is inherently racy, since there is nothing to prevent
1827 * the last synced txg from changing. That's okay since we'll
1828 * only commit things in the future.
1830 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1831 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1833 mutex_enter(&itxg->itxg_lock);
1834 if (itxg->itxg_txg != txg) {
1835 mutex_exit(&itxg->itxg_lock);
1840 * If we're adding itx records to the zl_itx_commit_list,
1841 * then the zil better be dirty in this "txg". We can assert
1842 * that here since we're holding the itxg_lock which will
1843 * prevent spa_sync from cleaning it. Once we add the itxs
1844 * to the zl_itx_commit_list we must commit it to disk even
1845 * if it's unnecessary (i.e. the txg was synced).
1847 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1848 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1849 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1851 mutex_exit(&itxg->itxg_lock);
1856 * Move the async itxs for a specified object to commit into sync lists.
1859 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1862 itx_async_node_t *ian;
1866 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1869 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1872 * This is inherently racy, since there is nothing to prevent
1873 * the last synced txg from changing.
1875 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1876 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1878 mutex_enter(&itxg->itxg_lock);
1879 if (itxg->itxg_txg != txg) {
1880 mutex_exit(&itxg->itxg_lock);
1885 * If a foid is specified then find that node and append its
1886 * list. Otherwise walk the tree appending all the lists
1887 * to the sync list. We add to the end rather than the
1888 * beginning to ensure the create has happened.
1890 t = &itxg->itxg_itxs->i_async_tree;
1892 ian = avl_find(t, &foid, &where);
1894 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1898 void *cookie = NULL;
1900 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1901 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1903 list_destroy(&ian->ia_list);
1904 kmem_free(ian, sizeof (itx_async_node_t));
1907 mutex_exit(&itxg->itxg_lock);
1912 * This function will prune commit itxs that are at the head of the
1913 * commit list (it won't prune past the first non-commit itx), and
1914 * either: a) attach them to the last lwb that's still pending
1915 * completion, or b) skip them altogether.
1917 * This is used as a performance optimization to prevent commit itxs
1918 * from generating new lwbs when it's unnecessary to do so.
1921 zil_prune_commit_list(zilog_t *zilog)
1925 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1927 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1928 lr_t *lrc = &itx->itx_lr;
1929 if (lrc->lrc_txtype != TX_COMMIT)
1932 mutex_enter(&zilog->zl_lock);
1934 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
1935 if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) {
1937 * All of the itxs this waiter was waiting on
1938 * must have already completed (or there were
1939 * never any itx's for it to wait on), so it's
1940 * safe to skip this waiter and mark it done.
1942 zil_commit_waiter_skip(itx->itx_private);
1944 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
1945 itx->itx_private = NULL;
1948 mutex_exit(&zilog->zl_lock);
1950 list_remove(&zilog->zl_itx_commit_list, itx);
1951 zil_itx_destroy(itx);
1954 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1958 zil_commit_writer_stall(zilog_t *zilog)
1961 * When zio_alloc_zil() fails to allocate the next lwb block on
1962 * disk, we must call txg_wait_synced() to ensure all of the
1963 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
1964 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
1965 * to zil_process_commit_list()) will have to call zil_create(),
1966 * and start a new ZIL chain.
1968 * Since zil_alloc_zil() failed, the lwb that was previously
1969 * issued does not have a pointer to the "next" lwb on disk.
1970 * Thus, if another ZIL writer thread was to allocate the "next"
1971 * on-disk lwb, that block could be leaked in the event of a
1972 * crash (because the previous lwb on-disk would not point to
1975 * We must hold the zilog's zl_issuer_lock while we do this, to
1976 * ensure no new threads enter zil_process_commit_list() until
1977 * all lwb's in the zl_lwb_list have been synced and freed
1978 * (which is achieved via the txg_wait_synced() call).
1980 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1981 txg_wait_synced(zilog->zl_dmu_pool, 0);
1982 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
1986 * This function will traverse the commit list, creating new lwbs as
1987 * needed, and committing the itxs from the commit list to these newly
1988 * created lwbs. Additionally, as a new lwb is created, the previous
1989 * lwb will be issued to the zio layer to be written to disk.
1992 zil_process_commit_list(zilog_t *zilog)
1994 spa_t *spa = zilog->zl_spa;
1995 list_t nolwb_waiters;
1999 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2002 * Return if there's nothing to commit before we dirty the fs by
2003 * calling zil_create().
2005 if (list_head(&zilog->zl_itx_commit_list) == NULL)
2008 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2009 offsetof(zil_commit_waiter_t, zcw_node));
2011 lwb = list_tail(&zilog->zl_lwb_list);
2013 lwb = zil_create(zilog);
2015 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2016 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2019 while (itx = list_head(&zilog->zl_itx_commit_list)) {
2020 lr_t *lrc = &itx->itx_lr;
2021 uint64_t txg = lrc->lrc_txg;
2023 ASSERT3U(txg, !=, 0);
2025 if (lrc->lrc_txtype == TX_COMMIT) {
2026 DTRACE_PROBE2(zil__process__commit__itx,
2027 zilog_t *, zilog, itx_t *, itx);
2029 DTRACE_PROBE2(zil__process__normal__itx,
2030 zilog_t *, zilog, itx_t *, itx);
2033 boolean_t synced = txg <= spa_last_synced_txg(spa);
2034 boolean_t frozen = txg > spa_freeze_txg(spa);
2037 * If the txg of this itx has already been synced out, then
2038 * we don't need to commit this itx to an lwb. This is
2039 * because the data of this itx will have already been
2040 * written to the main pool. This is inherently racy, and
2041 * it's still ok to commit an itx whose txg has already
2042 * been synced; this will result in a write that's
2043 * unnecessary, but will do no harm.
2045 * With that said, we always want to commit TX_COMMIT itxs
2046 * to an lwb, regardless of whether or not that itx's txg
2047 * has been synced out. We do this to ensure any OPENED lwb
2048 * will always have at least one zil_commit_waiter_t linked
2051 * As a counter-example, if we skipped TX_COMMIT itx's
2052 * whose txg had already been synced, the following
2053 * situation could occur if we happened to be racing with
2056 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
2057 * itx's txg is 10 and the last synced txg is 9.
2058 * 2. spa_sync finishes syncing out txg 10.
2059 * 3. we move to the next itx in the list, it's a TX_COMMIT
2060 * whose txg is 10, so we skip it rather than committing
2061 * it to the lwb used in (1).
2063 * If the itx that is skipped in (3) is the last TX_COMMIT
2064 * itx in the commit list, than it's possible for the lwb
2065 * used in (1) to remain in the OPENED state indefinitely.
2067 * To prevent the above scenario from occuring, ensuring
2068 * that once an lwb is OPENED it will transition to ISSUED
2069 * and eventually DONE, we always commit TX_COMMIT itx's to
2070 * an lwb here, even if that itx's txg has already been
2073 * Finally, if the pool is frozen, we _always_ commit the
2074 * itx. The point of freezing the pool is to prevent data
2075 * from being written to the main pool via spa_sync, and
2076 * instead rely solely on the ZIL to persistently store the
2077 * data; i.e. when the pool is frozen, the last synced txg
2078 * value can't be trusted.
2080 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2082 lwb = zil_lwb_commit(zilog, itx, lwb);
2083 } else if (lrc->lrc_txtype == TX_COMMIT) {
2084 ASSERT3P(lwb, ==, NULL);
2085 zil_commit_waiter_link_nolwb(
2086 itx->itx_private, &nolwb_waiters);
2090 list_remove(&zilog->zl_itx_commit_list, itx);
2091 zil_itx_destroy(itx);
2096 * This indicates zio_alloc_zil() failed to allocate the
2097 * "next" lwb on-disk. When this happens, we must stall
2098 * the ZIL write pipeline; see the comment within
2099 * zil_commit_writer_stall() for more details.
2101 zil_commit_writer_stall(zilog);
2104 * Additionally, we have to signal and mark the "nolwb"
2105 * waiters as "done" here, since without an lwb, we
2106 * can't do this via zil_lwb_flush_vdevs_done() like
2109 zil_commit_waiter_t *zcw;
2110 while (zcw = list_head(&nolwb_waiters)) {
2111 zil_commit_waiter_skip(zcw);
2112 list_remove(&nolwb_waiters, zcw);
2115 ASSERT(list_is_empty(&nolwb_waiters));
2116 ASSERT3P(lwb, !=, NULL);
2117 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2118 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2121 * At this point, the ZIL block pointed at by the "lwb"
2122 * variable is in one of the following states: "closed"
2125 * If its "closed", then no itxs have been committed to
2126 * it, so there's no point in issuing its zio (i.e.
2129 * If its "open" state, then it contains one or more
2130 * itxs that eventually need to be committed to stable
2131 * storage. In this case we intentionally do not issue
2132 * the lwb's zio to disk yet, and instead rely on one of
2133 * the following two mechanisms for issuing the zio:
2135 * 1. Ideally, there will be more ZIL activity occuring
2136 * on the system, such that this function will be
2137 * immediately called again (not necessarily by the same
2138 * thread) and this lwb's zio will be issued via
2139 * zil_lwb_commit(). This way, the lwb is guaranteed to
2140 * be "full" when it is issued to disk, and we'll make
2141 * use of the lwb's size the best we can.
2143 * 2. If there isn't sufficient ZIL activity occuring on
2144 * the system, such that this lwb's zio isn't issued via
2145 * zil_lwb_commit(), zil_commit_waiter() will issue the
2146 * lwb's zio. If this occurs, the lwb is not guaranteed
2147 * to be "full" by the time its zio is issued, and means
2148 * the size of the lwb was "too large" given the amount
2149 * of ZIL activity occuring on the system at that time.
2151 * We do this for a couple of reasons:
2153 * 1. To try and reduce the number of IOPs needed to
2154 * write the same number of itxs. If an lwb has space
2155 * available in it's buffer for more itxs, and more itxs
2156 * will be committed relatively soon (relative to the
2157 * latency of performing a write), then it's beneficial
2158 * to wait for these "next" itxs. This way, more itxs
2159 * can be committed to stable storage with fewer writes.
2161 * 2. To try and use the largest lwb block size that the
2162 * incoming rate of itxs can support. Again, this is to
2163 * try and pack as many itxs into as few lwbs as
2164 * possible, without significantly impacting the latency
2165 * of each individual itx.
2171 * This function is responsible for ensuring the passed in commit waiter
2172 * (and associated commit itx) is committed to an lwb. If the waiter is
2173 * not already committed to an lwb, all itxs in the zilog's queue of
2174 * itxs will be processed. The assumption is the passed in waiter's
2175 * commit itx will found in the queue just like the other non-commit
2176 * itxs, such that when the entire queue is processed, the waiter will
2177 * have been commited to an lwb.
2179 * The lwb associated with the passed in waiter is not guaranteed to
2180 * have been issued by the time this function completes. If the lwb is
2181 * not issued, we rely on future calls to zil_commit_writer() to issue
2182 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2185 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2187 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2188 ASSERT(spa_writeable(zilog->zl_spa));
2190 mutex_enter(&zilog->zl_issuer_lock);
2192 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2194 * It's possible that, while we were waiting to acquire
2195 * the "zl_issuer_lock", another thread committed this
2196 * waiter to an lwb. If that occurs, we bail out early,
2197 * without processing any of the zilog's queue of itxs.
2199 * On certain workloads and system configurations, the
2200 * "zl_issuer_lock" can become highly contended. In an
2201 * attempt to reduce this contention, we immediately drop
2202 * the lock if the waiter has already been processed.
2204 * We've measured this optimization to reduce CPU spent
2205 * contending on this lock by up to 5%, using a system
2206 * with 32 CPUs, low latency storage (~50 usec writes),
2207 * and 1024 threads performing sync writes.
2212 zil_get_commit_list(zilog);
2213 zil_prune_commit_list(zilog);
2214 zil_process_commit_list(zilog);
2217 mutex_exit(&zilog->zl_issuer_lock);
2221 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2223 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2224 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2225 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2227 lwb_t *lwb = zcw->zcw_lwb;
2228 ASSERT3P(lwb, !=, NULL);
2229 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2232 * If the lwb has already been issued by another thread, we can
2233 * immediately return since there's no work to be done (the
2234 * point of this function is to issue the lwb). Additionally, we
2235 * do this prior to acquiring the zl_issuer_lock, to avoid
2236 * acquiring it when it's not necessary to do so.
2238 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2239 lwb->lwb_state == LWB_STATE_DONE)
2243 * In order to call zil_lwb_write_issue() we must hold the
2244 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2245 * since we're already holding the commit waiter's "zcw_lock",
2246 * and those two locks are aquired in the opposite order
2249 mutex_exit(&zcw->zcw_lock);
2250 mutex_enter(&zilog->zl_issuer_lock);
2251 mutex_enter(&zcw->zcw_lock);
2254 * Since we just dropped and re-acquired the commit waiter's
2255 * lock, we have to re-check to see if the waiter was marked
2256 * "done" during that process. If the waiter was marked "done",
2257 * the "lwb" pointer is no longer valid (it can be free'd after
2258 * the waiter is marked "done"), so without this check we could
2259 * wind up with a use-after-free error below.
2264 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2267 * We've already checked this above, but since we hadn't acquired
2268 * the zilog's zl_issuer_lock, we have to perform this check a
2269 * second time while holding the lock.
2271 * We don't need to hold the zl_lock since the lwb cannot transition
2272 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2273 * _can_ transition from ISSUED to DONE, but it's OK to race with
2274 * that transition since we treat the lwb the same, whether it's in
2275 * the ISSUED or DONE states.
2277 * The important thing, is we treat the lwb differently depending on
2278 * if it's ISSUED or OPENED, and block any other threads that might
2279 * attempt to issue this lwb. For that reason we hold the
2280 * zl_issuer_lock when checking the lwb_state; we must not call
2281 * zil_lwb_write_issue() if the lwb had already been issued.
2283 * See the comment above the lwb_state_t structure definition for
2284 * more details on the lwb states, and locking requirements.
2286 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2287 lwb->lwb_state == LWB_STATE_DONE)
2290 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2293 * As described in the comments above zil_commit_waiter() and
2294 * zil_process_commit_list(), we need to issue this lwb's zio
2295 * since we've reached the commit waiter's timeout and it still
2296 * hasn't been issued.
2298 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2300 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2303 * Since the lwb's zio hadn't been issued by the time this thread
2304 * reached its timeout, we reset the zilog's "zl_cur_used" field
2305 * to influence the zil block size selection algorithm.
2307 * By having to issue the lwb's zio here, it means the size of the
2308 * lwb was too large, given the incoming throughput of itxs. By
2309 * setting "zl_cur_used" to zero, we communicate this fact to the
2310 * block size selection algorithm, so it can take this informaiton
2311 * into account, and potentially select a smaller size for the
2312 * next lwb block that is allocated.
2314 zilog->zl_cur_used = 0;
2318 * When zil_lwb_write_issue() returns NULL, this
2319 * indicates zio_alloc_zil() failed to allocate the
2320 * "next" lwb on-disk. When this occurs, the ZIL write
2321 * pipeline must be stalled; see the comment within the
2322 * zil_commit_writer_stall() function for more details.
2324 * We must drop the commit waiter's lock prior to
2325 * calling zil_commit_writer_stall() or else we can wind
2326 * up with the following deadlock:
2328 * - This thread is waiting for the txg to sync while
2329 * holding the waiter's lock; txg_wait_synced() is
2330 * used within txg_commit_writer_stall().
2332 * - The txg can't sync because it is waiting for this
2333 * lwb's zio callback to call dmu_tx_commit().
2335 * - The lwb's zio callback can't call dmu_tx_commit()
2336 * because it's blocked trying to acquire the waiter's
2337 * lock, which occurs prior to calling dmu_tx_commit()
2339 mutex_exit(&zcw->zcw_lock);
2340 zil_commit_writer_stall(zilog);
2341 mutex_enter(&zcw->zcw_lock);
2345 mutex_exit(&zilog->zl_issuer_lock);
2346 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2350 * This function is responsible for performing the following two tasks:
2352 * 1. its primary responsibility is to block until the given "commit
2353 * waiter" is considered "done".
2355 * 2. its secondary responsibility is to issue the zio for the lwb that
2356 * the given "commit waiter" is waiting on, if this function has
2357 * waited "long enough" and the lwb is still in the "open" state.
2359 * Given a sufficient amount of itxs being generated and written using
2360 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2361 * function. If this does not occur, this secondary responsibility will
2362 * ensure the lwb is issued even if there is not other synchronous
2363 * activity on the system.
2365 * For more details, see zil_process_commit_list(); more specifically,
2366 * the comment at the bottom of that function.
2369 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2371 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2372 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2373 ASSERT(spa_writeable(zilog->zl_spa));
2375 mutex_enter(&zcw->zcw_lock);
2378 * The timeout is scaled based on the lwb latency to avoid
2379 * significantly impacting the latency of each individual itx.
2380 * For more details, see the comment at the bottom of the
2381 * zil_process_commit_list() function.
2383 int pct = MAX(zfs_commit_timeout_pct, 1);
2384 #if defined(illumos) || !defined(_KERNEL)
2385 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2386 hrtime_t wakeup = gethrtime() + sleep;
2388 sbintime_t sleep = nstosbt((zilog->zl_last_lwb_latency * pct) / 100);
2389 sbintime_t wakeup = getsbinuptime() + sleep;
2391 boolean_t timedout = B_FALSE;
2393 while (!zcw->zcw_done) {
2394 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2396 lwb_t *lwb = zcw->zcw_lwb;
2399 * Usually, the waiter will have a non-NULL lwb field here,
2400 * but it's possible for it to be NULL as a result of
2401 * zil_commit() racing with spa_sync().
2403 * When zil_clean() is called, it's possible for the itxg
2404 * list (which may be cleaned via a taskq) to contain
2405 * commit itxs. When this occurs, the commit waiters linked
2406 * off of these commit itxs will not be committed to an
2407 * lwb. Additionally, these commit waiters will not be
2408 * marked done until zil_commit_waiter_skip() is called via
2411 * Thus, it's possible for this commit waiter (i.e. the
2412 * "zcw" variable) to be found in this "in between" state;
2413 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2414 * been skipped, so it's "zcw_done" field is still B_FALSE.
2416 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2418 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2419 ASSERT3B(timedout, ==, B_FALSE);
2422 * If the lwb hasn't been issued yet, then we
2423 * need to wait with a timeout, in case this
2424 * function needs to issue the lwb after the
2425 * timeout is reached; responsibility (2) from
2426 * the comment above this function.
2428 #if defined(illumos) || !defined(_KERNEL)
2429 clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2430 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2431 CALLOUT_FLAG_ABSOLUTE);
2433 if (timeleft >= 0 || zcw->zcw_done)
2436 int wait_err = cv_timedwait_sbt(&zcw->zcw_cv,
2437 &zcw->zcw_lock, wakeup, SBT_1NS, C_ABSOLUTE);
2438 if (wait_err != EWOULDBLOCK || zcw->zcw_done)
2443 zil_commit_waiter_timeout(zilog, zcw);
2445 if (!zcw->zcw_done) {
2447 * If the commit waiter has already been
2448 * marked "done", it's possible for the
2449 * waiter's lwb structure to have already
2450 * been freed. Thus, we can only reliably
2451 * make these assertions if the waiter
2454 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2455 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2459 * If the lwb isn't open, then it must have already
2460 * been issued. In that case, there's no need to
2461 * use a timeout when waiting for the lwb to
2464 * Additionally, if the lwb is NULL, the waiter
2465 * will soon be signalled and marked done via
2466 * zil_clean() and zil_itxg_clean(), so no timeout
2471 lwb->lwb_state == LWB_STATE_ISSUED ||
2472 lwb->lwb_state == LWB_STATE_DONE);
2473 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2477 mutex_exit(&zcw->zcw_lock);
2480 static zil_commit_waiter_t *
2481 zil_alloc_commit_waiter()
2483 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2485 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2486 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2487 list_link_init(&zcw->zcw_node);
2488 zcw->zcw_lwb = NULL;
2489 zcw->zcw_done = B_FALSE;
2490 zcw->zcw_zio_error = 0;
2496 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2498 ASSERT(!list_link_active(&zcw->zcw_node));
2499 ASSERT3P(zcw->zcw_lwb, ==, NULL);
2500 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2501 mutex_destroy(&zcw->zcw_lock);
2502 cv_destroy(&zcw->zcw_cv);
2503 kmem_cache_free(zil_zcw_cache, zcw);
2507 * This function is used to create a TX_COMMIT itx and assign it. This
2508 * way, it will be linked into the ZIL's list of synchronous itxs, and
2509 * then later committed to an lwb (or skipped) when
2510 * zil_process_commit_list() is called.
2513 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2515 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2516 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2518 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2519 itx->itx_sync = B_TRUE;
2520 itx->itx_private = zcw;
2522 zil_itx_assign(zilog, itx, tx);
2528 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2530 * When writing ZIL transactions to the on-disk representation of the
2531 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2532 * itxs can be committed to a single lwb. Once a lwb is written and
2533 * committed to stable storage (i.e. the lwb is written, and vdevs have
2534 * been flushed), each itx that was committed to that lwb is also
2535 * considered to be committed to stable storage.
2537 * When an itx is committed to an lwb, the log record (lr_t) contained
2538 * by the itx is copied into the lwb's zio buffer, and once this buffer
2539 * is written to disk, it becomes an on-disk ZIL block.
2541 * As itxs are generated, they're inserted into the ZIL's queue of
2542 * uncommitted itxs. The semantics of zil_commit() are such that it will
2543 * block until all itxs that were in the queue when it was called, are
2544 * committed to stable storage.
2546 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2547 * itxs, for all objects in the dataset, will be committed to stable
2548 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2549 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2550 * that correspond to the foid passed in, will be committed to stable
2551 * storage prior to zil_commit() returning.
2553 * Generally speaking, when zil_commit() is called, the consumer doesn't
2554 * actually care about _all_ of the uncommitted itxs. Instead, they're
2555 * simply trying to waiting for a specific itx to be committed to disk,
2556 * but the interface(s) for interacting with the ZIL don't allow such
2557 * fine-grained communication. A better interface would allow a consumer
2558 * to create and assign an itx, and then pass a reference to this itx to
2559 * zil_commit(); such that zil_commit() would return as soon as that
2560 * specific itx was committed to disk (instead of waiting for _all_
2561 * itxs to be committed).
2563 * When a thread calls zil_commit() a special "commit itx" will be
2564 * generated, along with a corresponding "waiter" for this commit itx.
2565 * zil_commit() will wait on this waiter's CV, such that when the waiter
2566 * is marked done, and signalled, zil_commit() will return.
2568 * This commit itx is inserted into the queue of uncommitted itxs. This
2569 * provides an easy mechanism for determining which itxs were in the
2570 * queue prior to zil_commit() having been called, and which itxs were
2571 * added after zil_commit() was called.
2573 * The commit it is special; it doesn't have any on-disk representation.
2574 * When a commit itx is "committed" to an lwb, the waiter associated
2575 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2576 * completes, each waiter on the lwb's list is marked done and signalled
2577 * -- allowing the thread waiting on the waiter to return from zil_commit().
2579 * It's important to point out a few critical factors that allow us
2580 * to make use of the commit itxs, commit waiters, per-lwb lists of
2581 * commit waiters, and zio completion callbacks like we're doing:
2583 * 1. The list of waiters for each lwb is traversed, and each commit
2584 * waiter is marked "done" and signalled, in the zio completion
2585 * callback of the lwb's zio[*].
2587 * * Actually, the waiters are signalled in the zio completion
2588 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2589 * that are sent to the vdevs upon completion of the lwb zio.
2591 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2592 * itxs, the order in which they are inserted is preserved[*]; as
2593 * itxs are added to the queue, they are added to the tail of
2594 * in-memory linked lists.
2596 * When committing the itxs to lwbs (to be written to disk), they
2597 * are committed in the same order in which the itxs were added to
2598 * the uncommitted queue's linked list(s); i.e. the linked list of
2599 * itxs to commit is traversed from head to tail, and each itx is
2600 * committed to an lwb in that order.
2604 * - the order of "sync" itxs is preserved w.r.t. other
2605 * "sync" itxs, regardless of the corresponding objects.
2606 * - the order of "async" itxs is preserved w.r.t. other
2607 * "async" itxs corresponding to the same object.
2608 * - the order of "async" itxs is *not* preserved w.r.t. other
2609 * "async" itxs corresponding to different objects.
2610 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2611 * versa) is *not* preserved, even for itxs that correspond
2612 * to the same object.
2614 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2615 * zil_get_commit_list(), and zil_process_commit_list().
2617 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2618 * lwb cannot be considered committed to stable storage, until its
2619 * "previous" lwb is also committed to stable storage. This fact,
2620 * coupled with the fact described above, means that itxs are
2621 * committed in (roughly) the order in which they were generated.
2622 * This is essential because itxs are dependent on prior itxs.
2623 * Thus, we *must not* deem an itx as being committed to stable
2624 * storage, until *all* prior itxs have also been committed to
2627 * To enforce this ordering of lwb zio's, while still leveraging as
2628 * much of the underlying storage performance as possible, we rely
2629 * on two fundamental concepts:
2631 * 1. The creation and issuance of lwb zio's is protected by
2632 * the zilog's "zl_issuer_lock", which ensures only a single
2633 * thread is creating and/or issuing lwb's at a time
2634 * 2. The "previous" lwb is a child of the "current" lwb
2635 * (leveraging the zio parent-child depenency graph)
2637 * By relying on this parent-child zio relationship, we can have
2638 * many lwb zio's concurrently issued to the underlying storage,
2639 * but the order in which they complete will be the same order in
2640 * which they were created.
2643 zil_commit(zilog_t *zilog, uint64_t foid)
2646 * We should never attempt to call zil_commit on a snapshot for
2647 * a couple of reasons:
2649 * 1. A snapshot may never be modified, thus it cannot have any
2650 * in-flight itxs that would have modified the dataset.
2652 * 2. By design, when zil_commit() is called, a commit itx will
2653 * be assigned to this zilog; as a result, the zilog will be
2654 * dirtied. We must not dirty the zilog of a snapshot; there's
2655 * checks in the code that enforce this invariant, and will
2656 * cause a panic if it's not upheld.
2658 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2660 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2663 if (!spa_writeable(zilog->zl_spa)) {
2665 * If the SPA is not writable, there should never be any
2666 * pending itxs waiting to be committed to disk. If that
2667 * weren't true, we'd skip writing those itxs out, and
2668 * would break the sematics of zil_commit(); thus, we're
2669 * verifying that truth before we return to the caller.
2671 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2672 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2673 for (int i = 0; i < TXG_SIZE; i++)
2674 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2679 * If the ZIL is suspended, we don't want to dirty it by calling
2680 * zil_commit_itx_assign() below, nor can we write out
2681 * lwbs like would be done in zil_commit_write(). Thus, we
2682 * simply rely on txg_wait_synced() to maintain the necessary
2683 * semantics, and avoid calling those functions altogether.
2685 if (zilog->zl_suspend > 0) {
2686 txg_wait_synced(zilog->zl_dmu_pool, 0);
2690 zil_commit_impl(zilog, foid);
2694 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2697 * Move the "async" itxs for the specified foid to the "sync"
2698 * queues, such that they will be later committed (or skipped)
2699 * to an lwb when zil_process_commit_list() is called.
2701 * Since these "async" itxs must be committed prior to this
2702 * call to zil_commit returning, we must perform this operation
2703 * before we call zil_commit_itx_assign().
2705 zil_async_to_sync(zilog, foid);
2708 * We allocate a new "waiter" structure which will initially be
2709 * linked to the commit itx using the itx's "itx_private" field.
2710 * Since the commit itx doesn't represent any on-disk state,
2711 * when it's committed to an lwb, rather than copying the its
2712 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2713 * added to the lwb's list of waiters. Then, when the lwb is
2714 * committed to stable storage, each waiter in the lwb's list of
2715 * waiters will be marked "done", and signalled.
2717 * We must create the waiter and assign the commit itx prior to
2718 * calling zil_commit_writer(), or else our specific commit itx
2719 * is not guaranteed to be committed to an lwb prior to calling
2720 * zil_commit_waiter().
2722 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2723 zil_commit_itx_assign(zilog, zcw);
2725 zil_commit_writer(zilog, zcw);
2726 zil_commit_waiter(zilog, zcw);
2728 if (zcw->zcw_zio_error != 0) {
2730 * If there was an error writing out the ZIL blocks that
2731 * this thread is waiting on, then we fallback to
2732 * relying on spa_sync() to write out the data this
2733 * thread is waiting on. Obviously this has performance
2734 * implications, but the expectation is for this to be
2735 * an exceptional case, and shouldn't occur often.
2737 DTRACE_PROBE2(zil__commit__io__error,
2738 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2739 txg_wait_synced(zilog->zl_dmu_pool, 0);
2742 zil_free_commit_waiter(zcw);
2746 * Called in syncing context to free committed log blocks and update log header.
2749 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2751 zil_header_t *zh = zil_header_in_syncing_context(zilog);
2752 uint64_t txg = dmu_tx_get_txg(tx);
2753 spa_t *spa = zilog->zl_spa;
2754 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2758 * We don't zero out zl_destroy_txg, so make sure we don't try
2759 * to destroy it twice.
2761 if (spa_sync_pass(spa) != 1)
2764 mutex_enter(&zilog->zl_lock);
2766 ASSERT(zilog->zl_stop_sync == 0);
2768 if (*replayed_seq != 0) {
2769 ASSERT(zh->zh_replay_seq < *replayed_seq);
2770 zh->zh_replay_seq = *replayed_seq;
2774 if (zilog->zl_destroy_txg == txg) {
2775 blkptr_t blk = zh->zh_log;
2777 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2779 bzero(zh, sizeof (zil_header_t));
2780 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2782 if (zilog->zl_keep_first) {
2784 * If this block was part of log chain that couldn't
2785 * be claimed because a device was missing during
2786 * zil_claim(), but that device later returns,
2787 * then this block could erroneously appear valid.
2788 * To guard against this, assign a new GUID to the new
2789 * log chain so it doesn't matter what blk points to.
2791 zil_init_log_chain(zilog, &blk);
2796 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2797 zh->zh_log = lwb->lwb_blk;
2798 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2800 list_remove(&zilog->zl_lwb_list, lwb);
2801 zio_free(spa, txg, &lwb->lwb_blk);
2802 zil_free_lwb(zilog, lwb);
2805 * If we don't have anything left in the lwb list then
2806 * we've had an allocation failure and we need to zero
2807 * out the zil_header blkptr so that we don't end
2808 * up freeing the same block twice.
2810 if (list_head(&zilog->zl_lwb_list) == NULL)
2811 BP_ZERO(&zh->zh_log);
2813 mutex_exit(&zilog->zl_lock);
2818 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2821 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2822 offsetof(zil_commit_waiter_t, zcw_node));
2823 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2824 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2825 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2831 zil_lwb_dest(void *vbuf, void *unused)
2834 mutex_destroy(&lwb->lwb_vdev_lock);
2835 avl_destroy(&lwb->lwb_vdev_tree);
2836 list_destroy(&lwb->lwb_waiters);
2842 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
2843 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
2845 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
2846 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2852 kmem_cache_destroy(zil_zcw_cache);
2853 kmem_cache_destroy(zil_lwb_cache);
2857 zil_set_sync(zilog_t *zilog, uint64_t sync)
2859 zilog->zl_sync = sync;
2863 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
2865 zilog->zl_logbias = logbias;
2869 zil_alloc(objset_t *os, zil_header_t *zh_phys)
2873 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
2875 zilog->zl_header = zh_phys;
2877 zilog->zl_spa = dmu_objset_spa(os);
2878 zilog->zl_dmu_pool = dmu_objset_pool(os);
2879 zilog->zl_destroy_txg = TXG_INITIAL - 1;
2880 zilog->zl_logbias = dmu_objset_logbias(os);
2881 zilog->zl_sync = dmu_objset_syncprop(os);
2882 zilog->zl_dirty_max_txg = 0;
2883 zilog->zl_last_lwb_opened = NULL;
2884 zilog->zl_last_lwb_latency = 0;
2886 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
2887 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
2889 for (int i = 0; i < TXG_SIZE; i++) {
2890 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
2891 MUTEX_DEFAULT, NULL);
2894 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
2895 offsetof(lwb_t, lwb_node));
2897 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
2898 offsetof(itx_t, itx_node));
2900 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
2906 zil_free(zilog_t *zilog)
2908 zilog->zl_stop_sync = 1;
2910 ASSERT0(zilog->zl_suspend);
2911 ASSERT0(zilog->zl_suspending);
2913 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2914 list_destroy(&zilog->zl_lwb_list);
2916 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
2917 list_destroy(&zilog->zl_itx_commit_list);
2919 for (int i = 0; i < TXG_SIZE; i++) {
2921 * It's possible for an itx to be generated that doesn't dirty
2922 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
2923 * callback to remove the entry. We remove those here.
2925 * Also free up the ziltest itxs.
2927 if (zilog->zl_itxg[i].itxg_itxs)
2928 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
2929 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
2932 mutex_destroy(&zilog->zl_issuer_lock);
2933 mutex_destroy(&zilog->zl_lock);
2935 cv_destroy(&zilog->zl_cv_suspend);
2937 kmem_free(zilog, sizeof (zilog_t));
2941 * Open an intent log.
2944 zil_open(objset_t *os, zil_get_data_t *get_data)
2946 zilog_t *zilog = dmu_objset_zil(os);
2948 ASSERT3P(zilog->zl_get_data, ==, NULL);
2949 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2950 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2952 zilog->zl_get_data = get_data;
2958 * Close an intent log.
2961 zil_close(zilog_t *zilog)
2966 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
2967 zil_commit(zilog, 0);
2969 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2970 ASSERT0(zilog->zl_dirty_max_txg);
2971 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
2974 mutex_enter(&zilog->zl_lock);
2975 lwb = list_tail(&zilog->zl_lwb_list);
2977 txg = zilog->zl_dirty_max_txg;
2979 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
2980 mutex_exit(&zilog->zl_lock);
2983 * We need to use txg_wait_synced() to wait long enough for the
2984 * ZIL to be clean, and to wait for all pending lwbs to be
2988 txg_wait_synced(zilog->zl_dmu_pool, txg);
2990 if (zilog_is_dirty(zilog))
2991 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
2992 VERIFY(!zilog_is_dirty(zilog));
2994 zilog->zl_get_data = NULL;
2997 * We should have only one lwb left on the list; remove it now.
2999 mutex_enter(&zilog->zl_lock);
3000 lwb = list_head(&zilog->zl_lwb_list);
3002 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
3003 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3004 list_remove(&zilog->zl_lwb_list, lwb);
3005 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3006 zil_free_lwb(zilog, lwb);
3008 mutex_exit(&zilog->zl_lock);
3011 static char *suspend_tag = "zil suspending";
3014 * Suspend an intent log. While in suspended mode, we still honor
3015 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3016 * On old version pools, we suspend the log briefly when taking a
3017 * snapshot so that it will have an empty intent log.
3019 * Long holds are not really intended to be used the way we do here --
3020 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3021 * could fail. Therefore we take pains to only put a long hold if it is
3022 * actually necessary. Fortunately, it will only be necessary if the
3023 * objset is currently mounted (or the ZVOL equivalent). In that case it
3024 * will already have a long hold, so we are not really making things any worse.
3026 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3027 * zvol_state_t), and use their mechanism to prevent their hold from being
3028 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3031 * if cookiep == NULL, this does both the suspend & resume.
3032 * Otherwise, it returns with the dataset "long held", and the cookie
3033 * should be passed into zil_resume().
3036 zil_suspend(const char *osname, void **cookiep)
3040 const zil_header_t *zh;
3043 error = dmu_objset_hold(osname, suspend_tag, &os);
3046 zilog = dmu_objset_zil(os);
3048 mutex_enter(&zilog->zl_lock);
3049 zh = zilog->zl_header;
3051 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
3052 mutex_exit(&zilog->zl_lock);
3053 dmu_objset_rele(os, suspend_tag);
3054 return (SET_ERROR(EBUSY));
3058 * Don't put a long hold in the cases where we can avoid it. This
3059 * is when there is no cookie so we are doing a suspend & resume
3060 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3061 * for the suspend because it's already suspended, or there's no ZIL.
3063 if (cookiep == NULL && !zilog->zl_suspending &&
3064 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3065 mutex_exit(&zilog->zl_lock);
3066 dmu_objset_rele(os, suspend_tag);
3070 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3071 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3073 zilog->zl_suspend++;
3075 if (zilog->zl_suspend > 1) {
3077 * Someone else is already suspending it.
3078 * Just wait for them to finish.
3081 while (zilog->zl_suspending)
3082 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3083 mutex_exit(&zilog->zl_lock);
3085 if (cookiep == NULL)
3093 * If there is no pointer to an on-disk block, this ZIL must not
3094 * be active (e.g. filesystem not mounted), so there's nothing
3097 if (BP_IS_HOLE(&zh->zh_log)) {
3098 ASSERT(cookiep != NULL); /* fast path already handled */
3101 mutex_exit(&zilog->zl_lock);
3105 zilog->zl_suspending = B_TRUE;
3106 mutex_exit(&zilog->zl_lock);
3109 * We need to use zil_commit_impl to ensure we wait for all
3110 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3111 * to disk before proceeding. If we used zil_commit instead, it
3112 * would just call txg_wait_synced(), because zl_suspend is set.
3113 * txg_wait_synced() doesn't wait for these lwb's to be
3114 * LWB_STATE_DONE before returning.
3116 zil_commit_impl(zilog, 0);
3119 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
3120 * txg_wait_synced() to ensure the data from the zilog has
3121 * migrated to the main pool before calling zil_destroy().
3123 txg_wait_synced(zilog->zl_dmu_pool, 0);
3125 zil_destroy(zilog, B_FALSE);
3127 mutex_enter(&zilog->zl_lock);
3128 zilog->zl_suspending = B_FALSE;
3129 cv_broadcast(&zilog->zl_cv_suspend);
3130 mutex_exit(&zilog->zl_lock);
3132 if (cookiep == NULL)
3140 zil_resume(void *cookie)
3142 objset_t *os = cookie;
3143 zilog_t *zilog = dmu_objset_zil(os);
3145 mutex_enter(&zilog->zl_lock);
3146 ASSERT(zilog->zl_suspend != 0);
3147 zilog->zl_suspend--;
3148 mutex_exit(&zilog->zl_lock);
3149 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3150 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3153 typedef struct zil_replay_arg {
3154 zil_replay_func_t **zr_replay;
3156 boolean_t zr_byteswap;
3161 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3163 char name[ZFS_MAX_DATASET_NAME_LEN];
3165 zilog->zl_replaying_seq--; /* didn't actually replay this one */
3167 dmu_objset_name(zilog->zl_os, name);
3169 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3170 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3171 (u_longlong_t)lr->lrc_seq,
3172 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3173 (lr->lrc_txtype & TX_CI) ? "CI" : "");
3179 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3181 zil_replay_arg_t *zr = zra;
3182 const zil_header_t *zh = zilog->zl_header;
3183 uint64_t reclen = lr->lrc_reclen;
3184 uint64_t txtype = lr->lrc_txtype;
3187 zilog->zl_replaying_seq = lr->lrc_seq;
3189 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
3192 if (lr->lrc_txg < claim_txg) /* already committed */
3195 /* Strip case-insensitive bit, still present in log record */
3198 if (txtype == 0 || txtype >= TX_MAX_TYPE)
3199 return (zil_replay_error(zilog, lr, EINVAL));
3202 * If this record type can be logged out of order, the object
3203 * (lr_foid) may no longer exist. That's legitimate, not an error.
3205 if (TX_OOO(txtype)) {
3206 error = dmu_object_info(zilog->zl_os,
3207 ((lr_ooo_t *)lr)->lr_foid, NULL);
3208 if (error == ENOENT || error == EEXIST)
3213 * Make a copy of the data so we can revise and extend it.
3215 bcopy(lr, zr->zr_lr, reclen);
3218 * If this is a TX_WRITE with a blkptr, suck in the data.
3220 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3221 error = zil_read_log_data(zilog, (lr_write_t *)lr,
3222 zr->zr_lr + reclen);
3224 return (zil_replay_error(zilog, lr, error));
3228 * The log block containing this lr may have been byteswapped
3229 * so that we can easily examine common fields like lrc_txtype.
3230 * However, the log is a mix of different record types, and only the
3231 * replay vectors know how to byteswap their records. Therefore, if
3232 * the lr was byteswapped, undo it before invoking the replay vector.
3234 if (zr->zr_byteswap)
3235 byteswap_uint64_array(zr->zr_lr, reclen);
3238 * We must now do two things atomically: replay this log record,
3239 * and update the log header sequence number to reflect the fact that
3240 * we did so. At the end of each replay function the sequence number
3241 * is updated if we are in replay mode.
3243 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3246 * The DMU's dnode layer doesn't see removes until the txg
3247 * commits, so a subsequent claim can spuriously fail with
3248 * EEXIST. So if we receive any error we try syncing out
3249 * any removes then retry the transaction. Note that we
3250 * specify B_FALSE for byteswap now, so we don't do it twice.
3252 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3253 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3255 return (zil_replay_error(zilog, lr, error));
3262 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3264 zilog->zl_replay_blks++;
3270 * If this dataset has a non-empty intent log, replay it and destroy it.
3273 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3275 zilog_t *zilog = dmu_objset_zil(os);
3276 const zil_header_t *zh = zilog->zl_header;
3277 zil_replay_arg_t zr;
3279 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3280 zil_destroy(zilog, B_TRUE);
3284 zr.zr_replay = replay_func;
3286 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3287 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3290 * Wait for in-progress removes to sync before starting replay.
3292 txg_wait_synced(zilog->zl_dmu_pool, 0);
3294 zilog->zl_replay = B_TRUE;
3295 zilog->zl_replay_time = ddi_get_lbolt();
3296 ASSERT(zilog->zl_replay_blks == 0);
3297 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3299 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3301 zil_destroy(zilog, B_FALSE);
3302 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3303 zilog->zl_replay = B_FALSE;
3307 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3309 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3312 if (zilog->zl_replay) {
3313 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3314 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3315 zilog->zl_replaying_seq;
3324 zil_reset(const char *osname, void *arg)
3328 error = zil_suspend(osname, NULL);
3330 return (SET_ERROR(EEXIST));