4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
31 #include <sys/spa_impl.h>
36 #include <sys/resource.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
46 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
47 * calls that change the file system. Each itx has enough information to
48 * be able to replay them after a system crash, power loss, or
49 * equivalent failure mode. These are stored in memory until either:
51 * 1. they are committed to the pool by the DMU transaction group
52 * (txg), at which point they can be discarded; or
53 * 2. they are committed to the on-disk ZIL for the dataset being
54 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
57 * In the event of a crash or power loss, the itxs contained by each
58 * dataset's on-disk ZIL will be replayed when that dataset is first
59 * instantianted (e.g. if the dataset is a normal fileystem, when it is
62 * As hinted at above, there is one ZIL per dataset (both the in-memory
63 * representation, and the on-disk representation). The on-disk format
64 * consists of 3 parts:
66 * - a single, per-dataset, ZIL header; which points to a chain of
67 * - zero or more ZIL blocks; each of which contains
68 * - zero or more ZIL records
70 * A ZIL record holds the information necessary to replay a single
71 * system call transaction. A ZIL block can hold many ZIL records, and
72 * the blocks are chained together, similarly to a singly linked list.
74 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
75 * block in the chain, and the ZIL header points to the first block in
78 * Note, there is not a fixed place in the pool to hold these ZIL
79 * blocks; they are dynamically allocated and freed as needed from the
80 * blocks available on the pool, though they can be preferentially
81 * allocated from a dedicated "log" vdev.
85 * This controls the amount of time that a ZIL block (lwb) will remain
86 * "open" when it isn't "full", and it has a thread waiting for it to be
87 * committed to stable storage. Please refer to the zil_commit_waiter()
88 * function (and the comments within it) for more details.
90 int zfs_commit_timeout_pct = 5;
93 * Disable intent logging replay. This global ZIL switch affects all pools.
95 int zil_replay_disable = 0;
96 SYSCTL_DECL(_vfs_zfs);
97 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
98 &zil_replay_disable, 0, "Disable intent logging replay");
101 * Tunable parameter for debugging or performance analysis. Setting
102 * zfs_nocacheflush will cause corruption on power loss if a volatile
103 * out-of-order write cache is enabled.
105 boolean_t zfs_nocacheflush = B_FALSE;
106 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
107 &zfs_nocacheflush, 0, "Disable cache flush");
108 boolean_t zfs_trim_enabled = B_TRUE;
109 SYSCTL_DECL(_vfs_zfs_trim);
110 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
114 * Limit SLOG write size per commit executed with synchronous priority.
115 * Any writes above that will be executed with lower (asynchronous) priority
116 * to limit potential SLOG device abuse by single active ZIL writer.
118 uint64_t zil_slog_bulk = 768 * 1024;
119 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_bulk, CTLFLAG_RWTUN,
120 &zil_slog_bulk, 0, "Maximal SLOG commit size with sync priority");
122 static kmem_cache_t *zil_lwb_cache;
123 static kmem_cache_t *zil_zcw_cache;
125 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
126 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
129 zil_bp_compare(const void *x1, const void *x2)
131 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
132 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
134 int cmp = AVL_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
138 return (AVL_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
142 zil_bp_tree_init(zilog_t *zilog)
144 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
145 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
149 zil_bp_tree_fini(zilog_t *zilog)
151 avl_tree_t *t = &zilog->zl_bp_tree;
155 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
156 kmem_free(zn, sizeof (zil_bp_node_t));
162 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
164 avl_tree_t *t = &zilog->zl_bp_tree;
169 if (BP_IS_EMBEDDED(bp))
172 dva = BP_IDENTITY(bp);
174 if (avl_find(t, dva, &where) != NULL)
175 return (SET_ERROR(EEXIST));
177 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
179 avl_insert(t, zn, where);
184 static zil_header_t *
185 zil_header_in_syncing_context(zilog_t *zilog)
187 return ((zil_header_t *)zilog->zl_header);
191 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
193 zio_cksum_t *zc = &bp->blk_cksum;
195 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
196 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
197 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
198 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
202 * Read a log block and make sure it's valid.
205 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
208 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
209 arc_flags_t aflags = ARC_FLAG_WAIT;
210 arc_buf_t *abuf = NULL;
214 if (zilog->zl_header->zh_claim_txg == 0)
215 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
217 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
218 zio_flags |= ZIO_FLAG_SPECULATIVE;
220 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
221 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
223 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
224 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
227 zio_cksum_t cksum = bp->blk_cksum;
230 * Validate the checksummed log block.
232 * Sequence numbers should be... sequential. The checksum
233 * verifier for the next block should be bp's checksum plus 1.
235 * Also check the log chain linkage and size used.
237 cksum.zc_word[ZIL_ZC_SEQ]++;
239 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
240 zil_chain_t *zilc = abuf->b_data;
241 char *lr = (char *)(zilc + 1);
242 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
244 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
245 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
246 error = SET_ERROR(ECKSUM);
248 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
250 *end = (char *)dst + len;
251 *nbp = zilc->zc_next_blk;
254 char *lr = abuf->b_data;
255 uint64_t size = BP_GET_LSIZE(bp);
256 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
258 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
259 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
260 (zilc->zc_nused > (size - sizeof (*zilc)))) {
261 error = SET_ERROR(ECKSUM);
263 ASSERT3U(zilc->zc_nused, <=,
264 SPA_OLD_MAXBLOCKSIZE);
265 bcopy(lr, dst, zilc->zc_nused);
266 *end = (char *)dst + zilc->zc_nused;
267 *nbp = zilc->zc_next_blk;
271 arc_buf_destroy(abuf, &abuf);
278 * Read a TX_WRITE log data block.
281 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
283 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
284 const blkptr_t *bp = &lr->lr_blkptr;
285 arc_flags_t aflags = ARC_FLAG_WAIT;
286 arc_buf_t *abuf = NULL;
290 if (BP_IS_HOLE(bp)) {
292 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
296 if (zilog->zl_header->zh_claim_txg == 0)
297 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
299 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
300 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
302 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
303 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
307 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
308 arc_buf_destroy(abuf, &abuf);
315 * Parse the intent log, and call parse_func for each valid record within.
318 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
319 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
321 const zil_header_t *zh = zilog->zl_header;
322 boolean_t claimed = !!zh->zh_claim_txg;
323 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
324 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
325 uint64_t max_blk_seq = 0;
326 uint64_t max_lr_seq = 0;
327 uint64_t blk_count = 0;
328 uint64_t lr_count = 0;
329 blkptr_t blk, next_blk;
334 * Old logs didn't record the maximum zh_claim_lr_seq.
336 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
337 claim_lr_seq = UINT64_MAX;
340 * Starting at the block pointed to by zh_log we read the log chain.
341 * For each block in the chain we strongly check that block to
342 * ensure its validity. We stop when an invalid block is found.
343 * For each block pointer in the chain we call parse_blk_func().
344 * For each record in each valid block we call parse_lr_func().
345 * If the log has been claimed, stop if we encounter a sequence
346 * number greater than the highest claimed sequence number.
348 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
349 zil_bp_tree_init(zilog);
351 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
352 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
356 if (blk_seq > claim_blk_seq)
358 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
360 ASSERT3U(max_blk_seq, <, blk_seq);
361 max_blk_seq = blk_seq;
364 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
367 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
371 for (lrp = lrbuf; lrp < end; lrp += reclen) {
372 lr_t *lr = (lr_t *)lrp;
373 reclen = lr->lrc_reclen;
374 ASSERT3U(reclen, >=, sizeof (lr_t));
375 if (lr->lrc_seq > claim_lr_seq)
377 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
379 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
380 max_lr_seq = lr->lrc_seq;
385 zilog->zl_parse_error = error;
386 zilog->zl_parse_blk_seq = max_blk_seq;
387 zilog->zl_parse_lr_seq = max_lr_seq;
388 zilog->zl_parse_blk_count = blk_count;
389 zilog->zl_parse_lr_count = lr_count;
391 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
392 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
394 zil_bp_tree_fini(zilog);
395 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
402 zil_clear_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
404 ASSERT(!BP_IS_HOLE(bp));
407 * As we call this function from the context of a rewind to a
408 * checkpoint, each ZIL block whose txg is later than the txg
409 * that we rewind to is invalid. Thus, we return -1 so
410 * zil_parse() doesn't attempt to read it.
412 if (bp->blk_birth >= first_txg)
415 if (zil_bp_tree_add(zilog, bp) != 0)
418 zio_free(zilog->zl_spa, first_txg, bp);
424 zil_noop_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
430 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
433 * Claim log block if not already committed and not already claimed.
434 * If tx == NULL, just verify that the block is claimable.
436 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
437 zil_bp_tree_add(zilog, bp) != 0)
440 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
441 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
442 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
446 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
448 lr_write_t *lr = (lr_write_t *)lrc;
451 if (lrc->lrc_txtype != TX_WRITE)
455 * If the block is not readable, don't claim it. This can happen
456 * in normal operation when a log block is written to disk before
457 * some of the dmu_sync() blocks it points to. In this case, the
458 * transaction cannot have been committed to anyone (we would have
459 * waited for all writes to be stable first), so it is semantically
460 * correct to declare this the end of the log.
462 if (lr->lr_blkptr.blk_birth >= first_txg &&
463 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
465 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
470 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
472 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
478 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
480 lr_write_t *lr = (lr_write_t *)lrc;
481 blkptr_t *bp = &lr->lr_blkptr;
484 * If we previously claimed it, we need to free it.
486 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
487 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
489 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
495 zil_lwb_vdev_compare(const void *x1, const void *x2)
497 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
498 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
500 return (AVL_CMP(v1, v2));
504 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
508 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
509 lwb->lwb_zilog = zilog;
511 lwb->lwb_slog = slog;
512 lwb->lwb_state = LWB_STATE_CLOSED;
513 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
514 lwb->lwb_max_txg = txg;
515 lwb->lwb_write_zio = NULL;
516 lwb->lwb_root_zio = NULL;
518 lwb->lwb_issued_timestamp = 0;
519 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
520 lwb->lwb_nused = sizeof (zil_chain_t);
521 lwb->lwb_sz = BP_GET_LSIZE(bp);
524 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
527 mutex_enter(&zilog->zl_lock);
528 list_insert_tail(&zilog->zl_lwb_list, lwb);
529 mutex_exit(&zilog->zl_lock);
531 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
532 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
533 VERIFY(list_is_empty(&lwb->lwb_waiters));
539 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
541 ASSERT(MUTEX_HELD(&zilog->zl_lock));
542 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
543 VERIFY(list_is_empty(&lwb->lwb_waiters));
544 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
545 ASSERT3P(lwb->lwb_write_zio, ==, NULL);
546 ASSERT3P(lwb->lwb_root_zio, ==, NULL);
547 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
548 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
549 lwb->lwb_state == LWB_STATE_DONE);
552 * Clear the zilog's field to indicate this lwb is no longer
553 * valid, and prevent use-after-free errors.
555 if (zilog->zl_last_lwb_opened == lwb)
556 zilog->zl_last_lwb_opened = NULL;
558 kmem_cache_free(zil_lwb_cache, lwb);
562 * Called when we create in-memory log transactions so that we know
563 * to cleanup the itxs at the end of spa_sync().
566 zilog_dirty(zilog_t *zilog, uint64_t txg)
568 dsl_pool_t *dp = zilog->zl_dmu_pool;
569 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
571 ASSERT(spa_writeable(zilog->zl_spa));
573 if (ds->ds_is_snapshot)
574 panic("dirtying snapshot!");
576 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
577 /* up the hold count until we can be written out */
578 dmu_buf_add_ref(ds->ds_dbuf, zilog);
580 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
585 * Determine if the zil is dirty in the specified txg. Callers wanting to
586 * ensure that the dirty state does not change must hold the itxg_lock for
587 * the specified txg. Holding the lock will ensure that the zil cannot be
588 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
592 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
594 dsl_pool_t *dp = zilog->zl_dmu_pool;
596 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
602 * Determine if the zil is dirty. The zil is considered dirty if it has
603 * any pending itx records that have not been cleaned by zil_clean().
606 zilog_is_dirty(zilog_t *zilog)
608 dsl_pool_t *dp = zilog->zl_dmu_pool;
610 for (int t = 0; t < TXG_SIZE; t++) {
611 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
618 * Create an on-disk intent log.
621 zil_create(zilog_t *zilog)
623 const zil_header_t *zh = zilog->zl_header;
629 boolean_t slog = FALSE;
632 * Wait for any previous destroy to complete.
634 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
636 ASSERT(zh->zh_claim_txg == 0);
637 ASSERT(zh->zh_replay_seq == 0);
642 * Allocate an initial log block if:
643 * - there isn't one already
644 * - the existing block is the wrong endianess
646 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
647 tx = dmu_tx_create(zilog->zl_os);
648 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
649 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
650 txg = dmu_tx_get_txg(tx);
652 if (!BP_IS_HOLE(&blk)) {
653 zio_free(zilog->zl_spa, txg, &blk);
657 error = zio_alloc_zil(zilog->zl_spa,
658 zilog->zl_os->os_dsl_dataset->ds_object, txg, &blk, NULL,
659 ZIL_MIN_BLKSZ, &slog);
662 zil_init_log_chain(zilog, &blk);
666 * Allocate a log write block (lwb) for the first log block.
669 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
672 * If we just allocated the first log block, commit our transaction
673 * and wait for zil_sync() to stuff the block poiner into zh_log.
674 * (zh is part of the MOS, so we cannot modify it in open context.)
678 txg_wait_synced(zilog->zl_dmu_pool, txg);
681 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
687 * In one tx, free all log blocks and clear the log header. If keep_first
688 * is set, then we're replaying a log with no content. We want to keep the
689 * first block, however, so that the first synchronous transaction doesn't
690 * require a txg_wait_synced() in zil_create(). We don't need to
691 * txg_wait_synced() here either when keep_first is set, because both
692 * zil_create() and zil_destroy() will wait for any in-progress destroys
696 zil_destroy(zilog_t *zilog, boolean_t keep_first)
698 const zil_header_t *zh = zilog->zl_header;
704 * Wait for any previous destroy to complete.
706 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
708 zilog->zl_old_header = *zh; /* debugging aid */
710 if (BP_IS_HOLE(&zh->zh_log))
713 tx = dmu_tx_create(zilog->zl_os);
714 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
715 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
716 txg = dmu_tx_get_txg(tx);
718 mutex_enter(&zilog->zl_lock);
720 ASSERT3U(zilog->zl_destroy_txg, <, txg);
721 zilog->zl_destroy_txg = txg;
722 zilog->zl_keep_first = keep_first;
724 if (!list_is_empty(&zilog->zl_lwb_list)) {
725 ASSERT(zh->zh_claim_txg == 0);
727 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
728 list_remove(&zilog->zl_lwb_list, lwb);
729 if (lwb->lwb_buf != NULL)
730 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
731 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
732 zil_free_lwb(zilog, lwb);
734 } else if (!keep_first) {
735 zil_destroy_sync(zilog, tx);
737 mutex_exit(&zilog->zl_lock);
743 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
745 ASSERT(list_is_empty(&zilog->zl_lwb_list));
746 (void) zil_parse(zilog, zil_free_log_block,
747 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
751 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
753 dmu_tx_t *tx = txarg;
760 error = dmu_objset_own_obj(dp, ds->ds_object,
761 DMU_OST_ANY, B_FALSE, FTAG, &os);
764 * EBUSY indicates that the objset is inconsistent, in which
765 * case it can not have a ZIL.
767 if (error != EBUSY) {
768 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
769 (unsigned long long)ds->ds_object, error);
774 zilog = dmu_objset_zil(os);
775 zh = zil_header_in_syncing_context(zilog);
776 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
777 first_txg = spa_min_claim_txg(zilog->zl_spa);
780 * If the spa_log_state is not set to be cleared, check whether
781 * the current uberblock is a checkpoint one and if the current
782 * header has been claimed before moving on.
784 * If the current uberblock is a checkpointed uberblock then
785 * one of the following scenarios took place:
787 * 1] We are currently rewinding to the checkpoint of the pool.
788 * 2] We crashed in the middle of a checkpoint rewind but we
789 * did manage to write the checkpointed uberblock to the
790 * vdev labels, so when we tried to import the pool again
791 * the checkpointed uberblock was selected from the import
794 * In both cases we want to zero out all the ZIL blocks, except
795 * the ones that have been claimed at the time of the checkpoint
796 * (their zh_claim_txg != 0). The reason is that these blocks
797 * may be corrupted since we may have reused their locations on
798 * disk after we took the checkpoint.
800 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
801 * when we first figure out whether the current uberblock is
802 * checkpointed or not. Unfortunately, that would discard all
803 * the logs, including the ones that are claimed, and we would
806 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
807 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
808 zh->zh_claim_txg == 0)) {
809 if (!BP_IS_HOLE(&zh->zh_log)) {
810 (void) zil_parse(zilog, zil_clear_log_block,
811 zil_noop_log_record, tx, first_txg);
813 BP_ZERO(&zh->zh_log);
814 dsl_dataset_dirty(dmu_objset_ds(os), tx);
815 dmu_objset_disown(os, FTAG);
820 * If we are not rewinding and opening the pool normally, then
821 * the min_claim_txg should be equal to the first txg of the pool.
823 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
826 * Claim all log blocks if we haven't already done so, and remember
827 * the highest claimed sequence number. This ensures that if we can
828 * read only part of the log now (e.g. due to a missing device),
829 * but we can read the entire log later, we will not try to replay
830 * or destroy beyond the last block we successfully claimed.
832 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
833 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
834 (void) zil_parse(zilog, zil_claim_log_block,
835 zil_claim_log_record, tx, first_txg);
836 zh->zh_claim_txg = first_txg;
837 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
838 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
839 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
840 zh->zh_flags |= ZIL_REPLAY_NEEDED;
841 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
842 dsl_dataset_dirty(dmu_objset_ds(os), tx);
845 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
846 dmu_objset_disown(os, FTAG);
851 * Check the log by walking the log chain.
852 * Checksum errors are ok as they indicate the end of the chain.
853 * Any other error (no device or read failure) returns an error.
857 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
866 error = dmu_objset_from_ds(ds, &os);
868 cmn_err(CE_WARN, "can't open objset %llu, error %d",
869 (unsigned long long)ds->ds_object, error);
873 zilog = dmu_objset_zil(os);
874 bp = (blkptr_t *)&zilog->zl_header->zh_log;
876 if (!BP_IS_HOLE(bp)) {
878 boolean_t valid = B_TRUE;
881 * Check the first block and determine if it's on a log device
882 * which may have been removed or faulted prior to loading this
883 * pool. If so, there's no point in checking the rest of the
884 * log as its content should have already been synced to the
887 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
888 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
889 if (vd->vdev_islog && vdev_is_dead(vd))
890 valid = vdev_log_state_valid(vd);
891 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
897 * Check whether the current uberblock is checkpointed (e.g.
898 * we are rewinding) and whether the current header has been
899 * claimed or not. If it hasn't then skip verifying it. We
900 * do this because its ZIL blocks may be part of the pool's
901 * state before the rewind, which is no longer valid.
903 zil_header_t *zh = zil_header_in_syncing_context(zilog);
904 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
905 zh->zh_claim_txg == 0)
910 * Because tx == NULL, zil_claim_log_block() will not actually claim
911 * any blocks, but just determine whether it is possible to do so.
912 * In addition to checking the log chain, zil_claim_log_block()
913 * will invoke zio_claim() with a done func of spa_claim_notify(),
914 * which will update spa_max_claim_txg. See spa_load() for details.
916 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
917 zilog->zl_header->zh_claim_txg ? -1ULL :
918 spa_min_claim_txg(os->os_spa));
920 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
924 * When an itx is "skipped", this function is used to properly mark the
925 * waiter as "done, and signal any thread(s) waiting on it. An itx can
926 * be skipped (and not committed to an lwb) for a variety of reasons,
927 * one of them being that the itx was committed via spa_sync(), prior to
928 * it being committed to an lwb; this can happen if a thread calling
929 * zil_commit() is racing with spa_sync().
932 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
934 mutex_enter(&zcw->zcw_lock);
935 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
936 zcw->zcw_done = B_TRUE;
937 cv_broadcast(&zcw->zcw_cv);
938 mutex_exit(&zcw->zcw_lock);
942 * This function is used when the given waiter is to be linked into an
943 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
944 * At this point, the waiter will no longer be referenced by the itx,
945 * and instead, will be referenced by the lwb.
948 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
951 * The lwb_waiters field of the lwb is protected by the zilog's
952 * zl_lock, thus it must be held when calling this function.
954 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
956 mutex_enter(&zcw->zcw_lock);
957 ASSERT(!list_link_active(&zcw->zcw_node));
958 ASSERT3P(zcw->zcw_lwb, ==, NULL);
959 ASSERT3P(lwb, !=, NULL);
960 ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
961 lwb->lwb_state == LWB_STATE_ISSUED);
963 list_insert_tail(&lwb->lwb_waiters, zcw);
965 mutex_exit(&zcw->zcw_lock);
969 * This function is used when zio_alloc_zil() fails to allocate a ZIL
970 * block, and the given waiter must be linked to the "nolwb waiters"
971 * list inside of zil_process_commit_list().
974 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
976 mutex_enter(&zcw->zcw_lock);
977 ASSERT(!list_link_active(&zcw->zcw_node));
978 ASSERT3P(zcw->zcw_lwb, ==, NULL);
979 list_insert_tail(nolwb, zcw);
980 mutex_exit(&zcw->zcw_lock);
984 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
986 avl_tree_t *t = &lwb->lwb_vdev_tree;
988 zil_vdev_node_t *zv, zvsearch;
989 int ndvas = BP_GET_NDVAS(bp);
992 if (zfs_nocacheflush)
995 mutex_enter(&lwb->lwb_vdev_lock);
996 for (i = 0; i < ndvas; i++) {
997 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
998 if (avl_find(t, &zvsearch, &where) == NULL) {
999 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1000 zv->zv_vdev = zvsearch.zv_vdev;
1001 avl_insert(t, zv, where);
1004 mutex_exit(&lwb->lwb_vdev_lock);
1008 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1010 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1014 * This function is a called after all VDEVs associated with a given lwb
1015 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1016 * as the lwb write completes, if "zfs_nocacheflush" is set.
1018 * The intention is for this function to be called as soon as the
1019 * contents of an lwb are considered "stable" on disk, and will survive
1020 * any sudden loss of power. At this point, any threads waiting for the
1021 * lwb to reach this state are signalled, and the "waiter" structures
1022 * are marked "done".
1025 zil_lwb_flush_vdevs_done(zio_t *zio)
1027 lwb_t *lwb = zio->io_private;
1028 zilog_t *zilog = lwb->lwb_zilog;
1029 dmu_tx_t *tx = lwb->lwb_tx;
1030 zil_commit_waiter_t *zcw;
1032 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1034 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1036 mutex_enter(&zilog->zl_lock);
1039 * Ensure the lwb buffer pointer is cleared before releasing the
1040 * txg. If we have had an allocation failure and the txg is
1041 * waiting to sync then we want zil_sync() to remove the lwb so
1042 * that it's not picked up as the next new one in
1043 * zil_process_commit_list(). zil_sync() will only remove the
1044 * lwb if lwb_buf is null.
1046 lwb->lwb_buf = NULL;
1049 ASSERT3U(lwb->lwb_issued_timestamp, >, 0);
1050 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp;
1052 lwb->lwb_root_zio = NULL;
1053 lwb->lwb_state = LWB_STATE_DONE;
1055 if (zilog->zl_last_lwb_opened == lwb) {
1057 * Remember the highest committed log sequence number
1058 * for ztest. We only update this value when all the log
1059 * writes succeeded, because ztest wants to ASSERT that
1060 * it got the whole log chain.
1062 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1065 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) {
1066 mutex_enter(&zcw->zcw_lock);
1068 ASSERT(list_link_active(&zcw->zcw_node));
1069 list_remove(&lwb->lwb_waiters, zcw);
1071 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1072 zcw->zcw_lwb = NULL;
1074 zcw->zcw_zio_error = zio->io_error;
1076 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1077 zcw->zcw_done = B_TRUE;
1078 cv_broadcast(&zcw->zcw_cv);
1080 mutex_exit(&zcw->zcw_lock);
1083 mutex_exit(&zilog->zl_lock);
1086 * Now that we've written this log block, we have a stable pointer
1087 * to the next block in the chain, so it's OK to let the txg in
1088 * which we allocated the next block sync.
1094 * This is called when an lwb write completes. This means, this specific
1095 * lwb was written to disk, and all dependent lwb have also been
1098 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1099 * the VDEVs involved in writing out this specific lwb. The lwb will be
1100 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1101 * zio completion callback for the lwb's root zio.
1104 zil_lwb_write_done(zio_t *zio)
1106 lwb_t *lwb = zio->io_private;
1107 spa_t *spa = zio->io_spa;
1108 zilog_t *zilog = lwb->lwb_zilog;
1109 avl_tree_t *t = &lwb->lwb_vdev_tree;
1110 void *cookie = NULL;
1111 zil_vdev_node_t *zv;
1113 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1115 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1116 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1117 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1118 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1119 ASSERT(!BP_IS_GANG(zio->io_bp));
1120 ASSERT(!BP_IS_HOLE(zio->io_bp));
1121 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1123 abd_put(zio->io_abd);
1125 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1127 mutex_enter(&zilog->zl_lock);
1128 lwb->lwb_write_zio = NULL;
1129 mutex_exit(&zilog->zl_lock);
1131 if (avl_numnodes(t) == 0)
1135 * If there was an IO error, we're not going to call zio_flush()
1136 * on these vdevs, so we simply empty the tree and free the
1137 * nodes. We avoid calling zio_flush() since there isn't any
1138 * good reason for doing so, after the lwb block failed to be
1141 if (zio->io_error != 0) {
1142 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1143 kmem_free(zv, sizeof (*zv));
1147 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1148 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1150 zio_flush(lwb->lwb_root_zio, vd);
1151 kmem_free(zv, sizeof (*zv));
1156 * This function's purpose is to "open" an lwb such that it is ready to
1157 * accept new itxs being committed to it. To do this, the lwb's zio
1158 * structures are created, and linked to the lwb. This function is
1159 * idempotent; if the passed in lwb has already been opened, this
1160 * function is essentially a no-op.
1163 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1165 zbookmark_phys_t zb;
1166 zio_priority_t prio;
1168 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1169 ASSERT3P(lwb, !=, NULL);
1170 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1171 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1173 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1174 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1175 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1177 if (lwb->lwb_root_zio == NULL) {
1178 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1179 BP_GET_LSIZE(&lwb->lwb_blk));
1181 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1182 prio = ZIO_PRIORITY_SYNC_WRITE;
1184 prio = ZIO_PRIORITY_ASYNC_WRITE;
1186 lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1187 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1188 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1190 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio,
1191 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd,
1192 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb,
1193 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
1194 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1196 lwb->lwb_state = LWB_STATE_OPENED;
1198 mutex_enter(&zilog->zl_lock);
1201 * The zilog's "zl_last_lwb_opened" field is used to
1202 * build the lwb/zio dependency chain, which is used to
1203 * preserve the ordering of lwb completions that is
1204 * required by the semantics of the ZIL. Each new lwb
1205 * zio becomes a parent of the "previous" lwb zio, such
1206 * that the new lwb's zio cannot complete until the
1207 * "previous" lwb's zio completes.
1209 * This is required by the semantics of zil_commit();
1210 * the commit waiters attached to the lwbs will be woken
1211 * in the lwb zio's completion callback, so this zio
1212 * dependency graph ensures the waiters are woken in the
1213 * correct order (the same order the lwbs were created).
1215 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1216 if (last_lwb_opened != NULL &&
1217 last_lwb_opened->lwb_state != LWB_STATE_DONE) {
1218 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1219 last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1220 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1221 zio_add_child(lwb->lwb_root_zio,
1222 last_lwb_opened->lwb_root_zio);
1224 zilog->zl_last_lwb_opened = lwb;
1226 mutex_exit(&zilog->zl_lock);
1229 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1230 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1231 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1235 * Define a limited set of intent log block sizes.
1237 * These must be a multiple of 4KB. Note only the amount used (again
1238 * aligned to 4KB) actually gets written. However, we can't always just
1239 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1241 uint64_t zil_block_buckets[] = {
1242 4096, /* non TX_WRITE */
1243 8192+4096, /* data base */
1244 32*1024 + 4096, /* NFS writes */
1249 * Start a log block write and advance to the next log block.
1250 * Calls are serialized.
1253 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1257 spa_t *spa = zilog->zl_spa;
1261 uint64_t zil_blksz, wsz;
1265 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1266 ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1267 ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1268 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1270 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1271 zilc = (zil_chain_t *)lwb->lwb_buf;
1272 bp = &zilc->zc_next_blk;
1274 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1275 bp = &zilc->zc_next_blk;
1278 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1281 * Allocate the next block and save its address in this block
1282 * before writing it in order to establish the log chain.
1283 * Note that if the allocation of nlwb synced before we wrote
1284 * the block that points at it (lwb), we'd leak it if we crashed.
1285 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1286 * We dirty the dataset to ensure that zil_sync() will be called
1287 * to clean up in the event of allocation failure or I/O failure.
1290 tx = dmu_tx_create(zilog->zl_os);
1293 * Since we are not going to create any new dirty data, and we
1294 * can even help with clearing the existing dirty data, we
1295 * should not be subject to the dirty data based delays. We
1296 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1298 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1300 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1301 txg = dmu_tx_get_txg(tx);
1306 * Log blocks are pre-allocated. Here we select the size of the next
1307 * block, based on size used in the last block.
1308 * - first find the smallest bucket that will fit the block from a
1309 * limited set of block sizes. This is because it's faster to write
1310 * blocks allocated from the same metaslab as they are adjacent or
1312 * - next find the maximum from the new suggested size and an array of
1313 * previous sizes. This lessens a picket fence effect of wrongly
1314 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1317 * Note we only write what is used, but we can't just allocate
1318 * the maximum block size because we can exhaust the available
1321 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1322 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1324 zil_blksz = zil_block_buckets[i];
1325 if (zil_blksz == UINT64_MAX)
1326 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1327 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1328 for (i = 0; i < ZIL_PREV_BLKS; i++)
1329 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1330 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1334 /* pass the old blkptr in order to spread log blocks across devs */
1335 error = zio_alloc_zil(spa, zilog->zl_os->os_dsl_dataset->ds_object,
1336 txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1338 ASSERT3U(bp->blk_birth, ==, txg);
1339 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1340 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1343 * Allocate a new log write block (lwb).
1345 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1348 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1349 /* For Slim ZIL only write what is used. */
1350 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1351 ASSERT3U(wsz, <=, lwb->lwb_sz);
1352 zio_shrink(lwb->lwb_write_zio, wsz);
1359 zilc->zc_nused = lwb->lwb_nused;
1360 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1363 * clear unused data for security
1365 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1367 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1369 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1370 lwb->lwb_issued_timestamp = gethrtime();
1371 lwb->lwb_state = LWB_STATE_ISSUED;
1373 zio_nowait(lwb->lwb_root_zio);
1374 zio_nowait(lwb->lwb_write_zio);
1377 * If there was an allocation failure then nlwb will be null which
1378 * forces a txg_wait_synced().
1384 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1387 lr_write_t *lrwb, *lrw;
1389 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1391 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1392 ASSERT3P(lwb, !=, NULL);
1393 ASSERT3P(lwb->lwb_buf, !=, NULL);
1395 zil_lwb_write_open(zilog, lwb);
1398 lrw = (lr_write_t *)lrc;
1401 * A commit itx doesn't represent any on-disk state; instead
1402 * it's simply used as a place holder on the commit list, and
1403 * provides a mechanism for attaching a "commit waiter" onto the
1404 * correct lwb (such that the waiter can be signalled upon
1405 * completion of that lwb). Thus, we don't process this itx's
1406 * log record if it's a commit itx (these itx's don't have log
1407 * records), and instead link the itx's waiter onto the lwb's
1410 * For more details, see the comment above zil_commit().
1412 if (lrc->lrc_txtype == TX_COMMIT) {
1413 mutex_enter(&zilog->zl_lock);
1414 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
1415 itx->itx_private = NULL;
1416 mutex_exit(&zilog->zl_lock);
1420 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1421 dlen = P2ROUNDUP_TYPED(
1422 lrw->lr_length, sizeof (uint64_t), uint64_t);
1426 reclen = lrc->lrc_reclen;
1427 zilog->zl_cur_used += (reclen + dlen);
1430 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen));
1434 * If this record won't fit in the current log block, start a new one.
1435 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1437 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1438 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1439 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1440 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1441 lwb = zil_lwb_write_issue(zilog, lwb);
1444 zil_lwb_write_open(zilog, lwb);
1445 ASSERT(LWB_EMPTY(lwb));
1446 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1447 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1450 dnow = MIN(dlen, lwb_sp - reclen);
1451 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1452 bcopy(lrc, lr_buf, reclen);
1453 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1454 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1457 * If it's a write, fetch the data or get its blkptr as appropriate.
1459 if (lrc->lrc_txtype == TX_WRITE) {
1460 if (txg > spa_freeze_txg(zilog->zl_spa))
1461 txg_wait_synced(zilog->zl_dmu_pool, txg);
1462 if (itx->itx_wr_state != WR_COPIED) {
1466 if (itx->itx_wr_state == WR_NEED_COPY) {
1467 dbuf = lr_buf + reclen;
1468 lrcb->lrc_reclen += dnow;
1469 if (lrwb->lr_length > dnow)
1470 lrwb->lr_length = dnow;
1471 lrw->lr_offset += dnow;
1472 lrw->lr_length -= dnow;
1474 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1479 * We pass in the "lwb_write_zio" rather than
1480 * "lwb_root_zio" so that the "lwb_write_zio"
1481 * becomes the parent of any zio's created by
1482 * the "zl_get_data" callback. The vdevs are
1483 * flushed after the "lwb_write_zio" completes,
1484 * so we want to make sure that completion
1485 * callback waits for these additional zio's,
1486 * such that the vdevs used by those zio's will
1487 * be included in the lwb's vdev tree, and those
1488 * vdevs will be properly flushed. If we passed
1489 * in "lwb_root_zio" here, then these additional
1490 * vdevs may not be flushed; e.g. if these zio's
1491 * completed after "lwb_write_zio" completed.
1493 error = zilog->zl_get_data(itx->itx_private,
1494 lrwb, dbuf, lwb, lwb->lwb_write_zio);
1497 txg_wait_synced(zilog->zl_dmu_pool, txg);
1501 ASSERT(error == ENOENT || error == EEXIST ||
1509 * We're actually making an entry, so update lrc_seq to be the
1510 * log record sequence number. Note that this is generally not
1511 * equal to the itx sequence number because not all transactions
1512 * are synchronous, and sometimes spa_sync() gets there first.
1514 lrcb->lrc_seq = ++zilog->zl_lr_seq;
1515 lwb->lwb_nused += reclen + dnow;
1517 zil_lwb_add_txg(lwb, txg);
1519 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1520 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1524 zilog->zl_cur_used += reclen;
1532 zil_itx_create(uint64_t txtype, size_t lrsize)
1536 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1538 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1539 itx->itx_lr.lrc_txtype = txtype;
1540 itx->itx_lr.lrc_reclen = lrsize;
1541 itx->itx_lr.lrc_seq = 0; /* defensive */
1542 itx->itx_sync = B_TRUE; /* default is synchronous */
1548 zil_itx_destroy(itx_t *itx)
1550 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1554 * Free up the sync and async itxs. The itxs_t has already been detached
1555 * so no locks are needed.
1558 zil_itxg_clean(itxs_t *itxs)
1564 itx_async_node_t *ian;
1566 list = &itxs->i_sync_list;
1567 while ((itx = list_head(list)) != NULL) {
1569 * In the general case, commit itxs will not be found
1570 * here, as they'll be committed to an lwb via
1571 * zil_lwb_commit(), and free'd in that function. Having
1572 * said that, it is still possible for commit itxs to be
1573 * found here, due to the following race:
1575 * - a thread calls zil_commit() which assigns the
1576 * commit itx to a per-txg i_sync_list
1577 * - zil_itxg_clean() is called (e.g. via spa_sync())
1578 * while the waiter is still on the i_sync_list
1580 * There's nothing to prevent syncing the txg while the
1581 * waiter is on the i_sync_list. This normally doesn't
1582 * happen because spa_sync() is slower than zil_commit(),
1583 * but if zil_commit() calls txg_wait_synced() (e.g.
1584 * because zil_create() or zil_commit_writer_stall() is
1585 * called) we will hit this case.
1587 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
1588 zil_commit_waiter_skip(itx->itx_private);
1590 list_remove(list, itx);
1591 zil_itx_destroy(itx);
1595 t = &itxs->i_async_tree;
1596 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1597 list = &ian->ia_list;
1598 while ((itx = list_head(list)) != NULL) {
1599 list_remove(list, itx);
1600 /* commit itxs should never be on the async lists. */
1601 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1602 zil_itx_destroy(itx);
1605 kmem_free(ian, sizeof (itx_async_node_t));
1609 kmem_free(itxs, sizeof (itxs_t));
1613 zil_aitx_compare(const void *x1, const void *x2)
1615 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1616 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1618 return (AVL_CMP(o1, o2));
1622 * Remove all async itx with the given oid.
1625 zil_remove_async(zilog_t *zilog, uint64_t oid)
1628 itx_async_node_t *ian;
1635 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1637 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1640 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1642 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1643 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1645 mutex_enter(&itxg->itxg_lock);
1646 if (itxg->itxg_txg != txg) {
1647 mutex_exit(&itxg->itxg_lock);
1652 * Locate the object node and append its list.
1654 t = &itxg->itxg_itxs->i_async_tree;
1655 ian = avl_find(t, &oid, &where);
1657 list_move_tail(&clean_list, &ian->ia_list);
1658 mutex_exit(&itxg->itxg_lock);
1660 while ((itx = list_head(&clean_list)) != NULL) {
1661 list_remove(&clean_list, itx);
1662 /* commit itxs should never be on the async lists. */
1663 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
1664 zil_itx_destroy(itx);
1666 list_destroy(&clean_list);
1670 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1674 itxs_t *itxs, *clean = NULL;
1677 * Object ids can be re-instantiated in the next txg so
1678 * remove any async transactions to avoid future leaks.
1679 * This can happen if a fsync occurs on the re-instantiated
1680 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1681 * the new file data and flushes a write record for the old object.
1683 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1684 zil_remove_async(zilog, itx->itx_oid);
1687 * Ensure the data of a renamed file is committed before the rename.
1689 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1690 zil_async_to_sync(zilog, itx->itx_oid);
1692 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1695 txg = dmu_tx_get_txg(tx);
1697 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1698 mutex_enter(&itxg->itxg_lock);
1699 itxs = itxg->itxg_itxs;
1700 if (itxg->itxg_txg != txg) {
1703 * The zil_clean callback hasn't got around to cleaning
1704 * this itxg. Save the itxs for release below.
1705 * This should be rare.
1707 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1708 "txg %llu", itxg->itxg_txg);
1709 clean = itxg->itxg_itxs;
1711 itxg->itxg_txg = txg;
1712 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1714 list_create(&itxs->i_sync_list, sizeof (itx_t),
1715 offsetof(itx_t, itx_node));
1716 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1717 sizeof (itx_async_node_t),
1718 offsetof(itx_async_node_t, ia_node));
1720 if (itx->itx_sync) {
1721 list_insert_tail(&itxs->i_sync_list, itx);
1723 avl_tree_t *t = &itxs->i_async_tree;
1725 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
1726 itx_async_node_t *ian;
1729 ian = avl_find(t, &foid, &where);
1731 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1732 list_create(&ian->ia_list, sizeof (itx_t),
1733 offsetof(itx_t, itx_node));
1734 ian->ia_foid = foid;
1735 avl_insert(t, ian, where);
1737 list_insert_tail(&ian->ia_list, itx);
1740 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1743 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1744 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1745 * need to be careful to always dirty the ZIL using the "real"
1746 * TXG (not itxg_txg) even when the SPA is frozen.
1748 zilog_dirty(zilog, dmu_tx_get_txg(tx));
1749 mutex_exit(&itxg->itxg_lock);
1751 /* Release the old itxs now we've dropped the lock */
1753 zil_itxg_clean(clean);
1757 * If there are any in-memory intent log transactions which have now been
1758 * synced then start up a taskq to free them. We should only do this after we
1759 * have written out the uberblocks (i.e. txg has been comitted) so that
1760 * don't inadvertently clean out in-memory log records that would be required
1764 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1766 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1769 ASSERT3U(synced_txg, <, ZILTEST_TXG);
1771 mutex_enter(&itxg->itxg_lock);
1772 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1773 mutex_exit(&itxg->itxg_lock);
1776 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1777 ASSERT3U(itxg->itxg_txg, !=, 0);
1778 clean_me = itxg->itxg_itxs;
1779 itxg->itxg_itxs = NULL;
1781 mutex_exit(&itxg->itxg_lock);
1783 * Preferably start a task queue to free up the old itxs but
1784 * if taskq_dispatch can't allocate resources to do that then
1785 * free it in-line. This should be rare. Note, using TQ_SLEEP
1786 * created a bad performance problem.
1788 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1789 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1790 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1791 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1792 zil_itxg_clean(clean_me);
1796 * This function will traverse the queue of itxs that need to be
1797 * committed, and move them onto the ZIL's zl_itx_commit_list.
1800 zil_get_commit_list(zilog_t *zilog)
1803 list_t *commit_list = &zilog->zl_itx_commit_list;
1805 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1807 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1810 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1813 * This is inherently racy, since there is nothing to prevent
1814 * the last synced txg from changing. That's okay since we'll
1815 * only commit things in the future.
1817 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1818 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1820 mutex_enter(&itxg->itxg_lock);
1821 if (itxg->itxg_txg != txg) {
1822 mutex_exit(&itxg->itxg_lock);
1827 * If we're adding itx records to the zl_itx_commit_list,
1828 * then the zil better be dirty in this "txg". We can assert
1829 * that here since we're holding the itxg_lock which will
1830 * prevent spa_sync from cleaning it. Once we add the itxs
1831 * to the zl_itx_commit_list we must commit it to disk even
1832 * if it's unnecessary (i.e. the txg was synced).
1834 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1835 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1836 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1838 mutex_exit(&itxg->itxg_lock);
1843 * Move the async itxs for a specified object to commit into sync lists.
1846 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1849 itx_async_node_t *ian;
1853 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1856 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1859 * This is inherently racy, since there is nothing to prevent
1860 * the last synced txg from changing.
1862 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1863 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1865 mutex_enter(&itxg->itxg_lock);
1866 if (itxg->itxg_txg != txg) {
1867 mutex_exit(&itxg->itxg_lock);
1872 * If a foid is specified then find that node and append its
1873 * list. Otherwise walk the tree appending all the lists
1874 * to the sync list. We add to the end rather than the
1875 * beginning to ensure the create has happened.
1877 t = &itxg->itxg_itxs->i_async_tree;
1879 ian = avl_find(t, &foid, &where);
1881 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1885 void *cookie = NULL;
1887 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1888 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1890 list_destroy(&ian->ia_list);
1891 kmem_free(ian, sizeof (itx_async_node_t));
1894 mutex_exit(&itxg->itxg_lock);
1899 * This function will prune commit itxs that are at the head of the
1900 * commit list (it won't prune past the first non-commit itx), and
1901 * either: a) attach them to the last lwb that's still pending
1902 * completion, or b) skip them altogether.
1904 * This is used as a performance optimization to prevent commit itxs
1905 * from generating new lwbs when it's unnecessary to do so.
1908 zil_prune_commit_list(zilog_t *zilog)
1912 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1914 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1915 lr_t *lrc = &itx->itx_lr;
1916 if (lrc->lrc_txtype != TX_COMMIT)
1919 mutex_enter(&zilog->zl_lock);
1921 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
1922 if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) {
1924 * All of the itxs this waiter was waiting on
1925 * must have already completed (or there were
1926 * never any itx's for it to wait on), so it's
1927 * safe to skip this waiter and mark it done.
1929 zil_commit_waiter_skip(itx->itx_private);
1931 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
1932 itx->itx_private = NULL;
1935 mutex_exit(&zilog->zl_lock);
1937 list_remove(&zilog->zl_itx_commit_list, itx);
1938 zil_itx_destroy(itx);
1941 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
1945 zil_commit_writer_stall(zilog_t *zilog)
1948 * When zio_alloc_zil() fails to allocate the next lwb block on
1949 * disk, we must call txg_wait_synced() to ensure all of the
1950 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
1951 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
1952 * to zil_process_commit_list()) will have to call zil_create(),
1953 * and start a new ZIL chain.
1955 * Since zil_alloc_zil() failed, the lwb that was previously
1956 * issued does not have a pointer to the "next" lwb on disk.
1957 * Thus, if another ZIL writer thread was to allocate the "next"
1958 * on-disk lwb, that block could be leaked in the event of a
1959 * crash (because the previous lwb on-disk would not point to
1962 * We must hold the zilog's zl_issuer_lock while we do this, to
1963 * ensure no new threads enter zil_process_commit_list() until
1964 * all lwb's in the zl_lwb_list have been synced and freed
1965 * (which is achieved via the txg_wait_synced() call).
1967 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1968 txg_wait_synced(zilog->zl_dmu_pool, 0);
1969 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
1973 * This function will traverse the commit list, creating new lwbs as
1974 * needed, and committing the itxs from the commit list to these newly
1975 * created lwbs. Additionally, as a new lwb is created, the previous
1976 * lwb will be issued to the zio layer to be written to disk.
1979 zil_process_commit_list(zilog_t *zilog)
1981 spa_t *spa = zilog->zl_spa;
1982 list_t nolwb_waiters;
1986 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1989 * Return if there's nothing to commit before we dirty the fs by
1990 * calling zil_create().
1992 if (list_head(&zilog->zl_itx_commit_list) == NULL)
1995 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
1996 offsetof(zil_commit_waiter_t, zcw_node));
1998 lwb = list_tail(&zilog->zl_lwb_list);
2000 lwb = zil_create(zilog);
2002 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2003 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2006 while (itx = list_head(&zilog->zl_itx_commit_list)) {
2007 lr_t *lrc = &itx->itx_lr;
2008 uint64_t txg = lrc->lrc_txg;
2010 ASSERT3U(txg, !=, 0);
2012 if (lrc->lrc_txtype == TX_COMMIT) {
2013 DTRACE_PROBE2(zil__process__commit__itx,
2014 zilog_t *, zilog, itx_t *, itx);
2016 DTRACE_PROBE2(zil__process__normal__itx,
2017 zilog_t *, zilog, itx_t *, itx);
2020 boolean_t synced = txg <= spa_last_synced_txg(spa);
2021 boolean_t frozen = txg > spa_freeze_txg(spa);
2024 * If the txg of this itx has already been synced out, then
2025 * we don't need to commit this itx to an lwb. This is
2026 * because the data of this itx will have already been
2027 * written to the main pool. This is inherently racy, and
2028 * it's still ok to commit an itx whose txg has already
2029 * been synced; this will result in a write that's
2030 * unnecessary, but will do no harm.
2032 * With that said, we always want to commit TX_COMMIT itxs
2033 * to an lwb, regardless of whether or not that itx's txg
2034 * has been synced out. We do this to ensure any OPENED lwb
2035 * will always have at least one zil_commit_waiter_t linked
2038 * As a counter-example, if we skipped TX_COMMIT itx's
2039 * whose txg had already been synced, the following
2040 * situation could occur if we happened to be racing with
2043 * 1. we commit a non-TX_COMMIT itx to an lwb, where the
2044 * itx's txg is 10 and the last synced txg is 9.
2045 * 2. spa_sync finishes syncing out txg 10.
2046 * 3. we move to the next itx in the list, it's a TX_COMMIT
2047 * whose txg is 10, so we skip it rather than committing
2048 * it to the lwb used in (1).
2050 * If the itx that is skipped in (3) is the last TX_COMMIT
2051 * itx in the commit list, than it's possible for the lwb
2052 * used in (1) to remain in the OPENED state indefinitely.
2054 * To prevent the above scenario from occuring, ensuring
2055 * that once an lwb is OPENED it will transition to ISSUED
2056 * and eventually DONE, we always commit TX_COMMIT itx's to
2057 * an lwb here, even if that itx's txg has already been
2060 * Finally, if the pool is frozen, we _always_ commit the
2061 * itx. The point of freezing the pool is to prevent data
2062 * from being written to the main pool via spa_sync, and
2063 * instead rely solely on the ZIL to persistently store the
2064 * data; i.e. when the pool is frozen, the last synced txg
2065 * value can't be trusted.
2067 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2069 lwb = zil_lwb_commit(zilog, itx, lwb);
2070 } else if (lrc->lrc_txtype == TX_COMMIT) {
2071 ASSERT3P(lwb, ==, NULL);
2072 zil_commit_waiter_link_nolwb(
2073 itx->itx_private, &nolwb_waiters);
2077 list_remove(&zilog->zl_itx_commit_list, itx);
2078 zil_itx_destroy(itx);
2083 * This indicates zio_alloc_zil() failed to allocate the
2084 * "next" lwb on-disk. When this happens, we must stall
2085 * the ZIL write pipeline; see the comment within
2086 * zil_commit_writer_stall() for more details.
2088 zil_commit_writer_stall(zilog);
2091 * Additionally, we have to signal and mark the "nolwb"
2092 * waiters as "done" here, since without an lwb, we
2093 * can't do this via zil_lwb_flush_vdevs_done() like
2096 zil_commit_waiter_t *zcw;
2097 while (zcw = list_head(&nolwb_waiters)) {
2098 zil_commit_waiter_skip(zcw);
2099 list_remove(&nolwb_waiters, zcw);
2102 ASSERT(list_is_empty(&nolwb_waiters));
2103 ASSERT3P(lwb, !=, NULL);
2104 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2105 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE);
2108 * At this point, the ZIL block pointed at by the "lwb"
2109 * variable is in one of the following states: "closed"
2112 * If its "closed", then no itxs have been committed to
2113 * it, so there's no point in issuing its zio (i.e.
2116 * If its "open" state, then it contains one or more
2117 * itxs that eventually need to be committed to stable
2118 * storage. In this case we intentionally do not issue
2119 * the lwb's zio to disk yet, and instead rely on one of
2120 * the following two mechanisms for issuing the zio:
2122 * 1. Ideally, there will be more ZIL activity occuring
2123 * on the system, such that this function will be
2124 * immediately called again (not necessarily by the same
2125 * thread) and this lwb's zio will be issued via
2126 * zil_lwb_commit(). This way, the lwb is guaranteed to
2127 * be "full" when it is issued to disk, and we'll make
2128 * use of the lwb's size the best we can.
2130 * 2. If there isn't sufficient ZIL activity occuring on
2131 * the system, such that this lwb's zio isn't issued via
2132 * zil_lwb_commit(), zil_commit_waiter() will issue the
2133 * lwb's zio. If this occurs, the lwb is not guaranteed
2134 * to be "full" by the time its zio is issued, and means
2135 * the size of the lwb was "too large" given the amount
2136 * of ZIL activity occuring on the system at that time.
2138 * We do this for a couple of reasons:
2140 * 1. To try and reduce the number of IOPs needed to
2141 * write the same number of itxs. If an lwb has space
2142 * available in it's buffer for more itxs, and more itxs
2143 * will be committed relatively soon (relative to the
2144 * latency of performing a write), then it's beneficial
2145 * to wait for these "next" itxs. This way, more itxs
2146 * can be committed to stable storage with fewer writes.
2148 * 2. To try and use the largest lwb block size that the
2149 * incoming rate of itxs can support. Again, this is to
2150 * try and pack as many itxs into as few lwbs as
2151 * possible, without significantly impacting the latency
2152 * of each individual itx.
2158 * This function is responsible for ensuring the passed in commit waiter
2159 * (and associated commit itx) is committed to an lwb. If the waiter is
2160 * not already committed to an lwb, all itxs in the zilog's queue of
2161 * itxs will be processed. The assumption is the passed in waiter's
2162 * commit itx will found in the queue just like the other non-commit
2163 * itxs, such that when the entire queue is processed, the waiter will
2164 * have been commited to an lwb.
2166 * The lwb associated with the passed in waiter is not guaranteed to
2167 * have been issued by the time this function completes. If the lwb is
2168 * not issued, we rely on future calls to zil_commit_writer() to issue
2169 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2172 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2174 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2175 ASSERT(spa_writeable(zilog->zl_spa));
2177 mutex_enter(&zilog->zl_issuer_lock);
2179 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2181 * It's possible that, while we were waiting to acquire
2182 * the "zl_issuer_lock", another thread committed this
2183 * waiter to an lwb. If that occurs, we bail out early,
2184 * without processing any of the zilog's queue of itxs.
2186 * On certain workloads and system configurations, the
2187 * "zl_issuer_lock" can become highly contended. In an
2188 * attempt to reduce this contention, we immediately drop
2189 * the lock if the waiter has already been processed.
2191 * We've measured this optimization to reduce CPU spent
2192 * contending on this lock by up to 5%, using a system
2193 * with 32 CPUs, low latency storage (~50 usec writes),
2194 * and 1024 threads performing sync writes.
2199 zil_get_commit_list(zilog);
2200 zil_prune_commit_list(zilog);
2201 zil_process_commit_list(zilog);
2204 mutex_exit(&zilog->zl_issuer_lock);
2208 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2210 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2211 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2212 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
2214 lwb_t *lwb = zcw->zcw_lwb;
2215 ASSERT3P(lwb, !=, NULL);
2216 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
2219 * If the lwb has already been issued by another thread, we can
2220 * immediately return since there's no work to be done (the
2221 * point of this function is to issue the lwb). Additionally, we
2222 * do this prior to acquiring the zl_issuer_lock, to avoid
2223 * acquiring it when it's not necessary to do so.
2225 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2226 lwb->lwb_state == LWB_STATE_DONE)
2230 * In order to call zil_lwb_write_issue() we must hold the
2231 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2232 * since we're already holding the commit waiter's "zcw_lock",
2233 * and those two locks are aquired in the opposite order
2236 mutex_exit(&zcw->zcw_lock);
2237 mutex_enter(&zilog->zl_issuer_lock);
2238 mutex_enter(&zcw->zcw_lock);
2241 * Since we just dropped and re-acquired the commit waiter's
2242 * lock, we have to re-check to see if the waiter was marked
2243 * "done" during that process. If the waiter was marked "done",
2244 * the "lwb" pointer is no longer valid (it can be free'd after
2245 * the waiter is marked "done"), so without this check we could
2246 * wind up with a use-after-free error below.
2251 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2254 * We've already checked this above, but since we hadn't acquired
2255 * the zilog's zl_issuer_lock, we have to perform this check a
2256 * second time while holding the lock.
2258 * We don't need to hold the zl_lock since the lwb cannot transition
2259 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2260 * _can_ transition from ISSUED to DONE, but it's OK to race with
2261 * that transition since we treat the lwb the same, whether it's in
2262 * the ISSUED or DONE states.
2264 * The important thing, is we treat the lwb differently depending on
2265 * if it's ISSUED or OPENED, and block any other threads that might
2266 * attempt to issue this lwb. For that reason we hold the
2267 * zl_issuer_lock when checking the lwb_state; we must not call
2268 * zil_lwb_write_issue() if the lwb had already been issued.
2270 * See the comment above the lwb_state_t structure definition for
2271 * more details on the lwb states, and locking requirements.
2273 if (lwb->lwb_state == LWB_STATE_ISSUED ||
2274 lwb->lwb_state == LWB_STATE_DONE)
2277 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
2280 * As described in the comments above zil_commit_waiter() and
2281 * zil_process_commit_list(), we need to issue this lwb's zio
2282 * since we've reached the commit waiter's timeout and it still
2283 * hasn't been issued.
2285 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb);
2287 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED);
2290 * Since the lwb's zio hadn't been issued by the time this thread
2291 * reached its timeout, we reset the zilog's "zl_cur_used" field
2292 * to influence the zil block size selection algorithm.
2294 * By having to issue the lwb's zio here, it means the size of the
2295 * lwb was too large, given the incoming throughput of itxs. By
2296 * setting "zl_cur_used" to zero, we communicate this fact to the
2297 * block size selection algorithm, so it can take this informaiton
2298 * into account, and potentially select a smaller size for the
2299 * next lwb block that is allocated.
2301 zilog->zl_cur_used = 0;
2305 * When zil_lwb_write_issue() returns NULL, this
2306 * indicates zio_alloc_zil() failed to allocate the
2307 * "next" lwb on-disk. When this occurs, the ZIL write
2308 * pipeline must be stalled; see the comment within the
2309 * zil_commit_writer_stall() function for more details.
2311 * We must drop the commit waiter's lock prior to
2312 * calling zil_commit_writer_stall() or else we can wind
2313 * up with the following deadlock:
2315 * - This thread is waiting for the txg to sync while
2316 * holding the waiter's lock; txg_wait_synced() is
2317 * used within txg_commit_writer_stall().
2319 * - The txg can't sync because it is waiting for this
2320 * lwb's zio callback to call dmu_tx_commit().
2322 * - The lwb's zio callback can't call dmu_tx_commit()
2323 * because it's blocked trying to acquire the waiter's
2324 * lock, which occurs prior to calling dmu_tx_commit()
2326 mutex_exit(&zcw->zcw_lock);
2327 zil_commit_writer_stall(zilog);
2328 mutex_enter(&zcw->zcw_lock);
2332 mutex_exit(&zilog->zl_issuer_lock);
2333 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2337 * This function is responsible for performing the following two tasks:
2339 * 1. its primary responsibility is to block until the given "commit
2340 * waiter" is considered "done".
2342 * 2. its secondary responsibility is to issue the zio for the lwb that
2343 * the given "commit waiter" is waiting on, if this function has
2344 * waited "long enough" and the lwb is still in the "open" state.
2346 * Given a sufficient amount of itxs being generated and written using
2347 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2348 * function. If this does not occur, this secondary responsibility will
2349 * ensure the lwb is issued even if there is not other synchronous
2350 * activity on the system.
2352 * For more details, see zil_process_commit_list(); more specifically,
2353 * the comment at the bottom of that function.
2356 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
2358 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2359 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
2360 ASSERT(spa_writeable(zilog->zl_spa));
2362 mutex_enter(&zcw->zcw_lock);
2365 * The timeout is scaled based on the lwb latency to avoid
2366 * significantly impacting the latency of each individual itx.
2367 * For more details, see the comment at the bottom of the
2368 * zil_process_commit_list() function.
2370 int pct = MAX(zfs_commit_timeout_pct, 1);
2371 #if defined(illumos) || !defined(_KERNEL)
2372 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
2373 hrtime_t wakeup = gethrtime() + sleep;
2375 sbintime_t sleep = nstosbt((zilog->zl_last_lwb_latency * pct) / 100);
2376 sbintime_t wakeup = getsbinuptime() + sleep;
2378 boolean_t timedout = B_FALSE;
2380 while (!zcw->zcw_done) {
2381 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
2383 lwb_t *lwb = zcw->zcw_lwb;
2386 * Usually, the waiter will have a non-NULL lwb field here,
2387 * but it's possible for it to be NULL as a result of
2388 * zil_commit() racing with spa_sync().
2390 * When zil_clean() is called, it's possible for the itxg
2391 * list (which may be cleaned via a taskq) to contain
2392 * commit itxs. When this occurs, the commit waiters linked
2393 * off of these commit itxs will not be committed to an
2394 * lwb. Additionally, these commit waiters will not be
2395 * marked done until zil_commit_waiter_skip() is called via
2398 * Thus, it's possible for this commit waiter (i.e. the
2399 * "zcw" variable) to be found in this "in between" state;
2400 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2401 * been skipped, so it's "zcw_done" field is still B_FALSE.
2403 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
2405 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
2406 ASSERT3B(timedout, ==, B_FALSE);
2409 * If the lwb hasn't been issued yet, then we
2410 * need to wait with a timeout, in case this
2411 * function needs to issue the lwb after the
2412 * timeout is reached; responsibility (2) from
2413 * the comment above this function.
2415 #if defined(illumos) || !defined(_KERNEL)
2416 clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv,
2417 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
2418 CALLOUT_FLAG_ABSOLUTE);
2420 if (timeleft >= 0 || zcw->zcw_done)
2423 int wait_err = cv_timedwait_sbt(&zcw->zcw_cv,
2424 &zcw->zcw_lock, wakeup, SBT_1NS, C_ABSOLUTE);
2425 if (wait_err != EWOULDBLOCK || zcw->zcw_done)
2430 zil_commit_waiter_timeout(zilog, zcw);
2432 if (!zcw->zcw_done) {
2434 * If the commit waiter has already been
2435 * marked "done", it's possible for the
2436 * waiter's lwb structure to have already
2437 * been freed. Thus, we can only reliably
2438 * make these assertions if the waiter
2441 ASSERT3P(lwb, ==, zcw->zcw_lwb);
2442 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
2446 * If the lwb isn't open, then it must have already
2447 * been issued. In that case, there's no need to
2448 * use a timeout when waiting for the lwb to
2451 * Additionally, if the lwb is NULL, the waiter
2452 * will soon be signalled and marked done via
2453 * zil_clean() and zil_itxg_clean(), so no timeout
2458 lwb->lwb_state == LWB_STATE_ISSUED ||
2459 lwb->lwb_state == LWB_STATE_DONE);
2460 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
2464 mutex_exit(&zcw->zcw_lock);
2467 static zil_commit_waiter_t *
2468 zil_alloc_commit_waiter()
2470 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
2472 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
2473 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
2474 list_link_init(&zcw->zcw_node);
2475 zcw->zcw_lwb = NULL;
2476 zcw->zcw_done = B_FALSE;
2477 zcw->zcw_zio_error = 0;
2483 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
2485 ASSERT(!list_link_active(&zcw->zcw_node));
2486 ASSERT3P(zcw->zcw_lwb, ==, NULL);
2487 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
2488 mutex_destroy(&zcw->zcw_lock);
2489 cv_destroy(&zcw->zcw_cv);
2490 kmem_cache_free(zil_zcw_cache, zcw);
2494 * This function is used to create a TX_COMMIT itx and assign it. This
2495 * way, it will be linked into the ZIL's list of synchronous itxs, and
2496 * then later committed to an lwb (or skipped) when
2497 * zil_process_commit_list() is called.
2500 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
2502 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
2503 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
2505 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
2506 itx->itx_sync = B_TRUE;
2507 itx->itx_private = zcw;
2509 zil_itx_assign(zilog, itx, tx);
2515 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2517 * When writing ZIL transactions to the on-disk representation of the
2518 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2519 * itxs can be committed to a single lwb. Once a lwb is written and
2520 * committed to stable storage (i.e. the lwb is written, and vdevs have
2521 * been flushed), each itx that was committed to that lwb is also
2522 * considered to be committed to stable storage.
2524 * When an itx is committed to an lwb, the log record (lr_t) contained
2525 * by the itx is copied into the lwb's zio buffer, and once this buffer
2526 * is written to disk, it becomes an on-disk ZIL block.
2528 * As itxs are generated, they're inserted into the ZIL's queue of
2529 * uncommitted itxs. The semantics of zil_commit() are such that it will
2530 * block until all itxs that were in the queue when it was called, are
2531 * committed to stable storage.
2533 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2534 * itxs, for all objects in the dataset, will be committed to stable
2535 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2536 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2537 * that correspond to the foid passed in, will be committed to stable
2538 * storage prior to zil_commit() returning.
2540 * Generally speaking, when zil_commit() is called, the consumer doesn't
2541 * actually care about _all_ of the uncommitted itxs. Instead, they're
2542 * simply trying to waiting for a specific itx to be committed to disk,
2543 * but the interface(s) for interacting with the ZIL don't allow such
2544 * fine-grained communication. A better interface would allow a consumer
2545 * to create and assign an itx, and then pass a reference to this itx to
2546 * zil_commit(); such that zil_commit() would return as soon as that
2547 * specific itx was committed to disk (instead of waiting for _all_
2548 * itxs to be committed).
2550 * When a thread calls zil_commit() a special "commit itx" will be
2551 * generated, along with a corresponding "waiter" for this commit itx.
2552 * zil_commit() will wait on this waiter's CV, such that when the waiter
2553 * is marked done, and signalled, zil_commit() will return.
2555 * This commit itx is inserted into the queue of uncommitted itxs. This
2556 * provides an easy mechanism for determining which itxs were in the
2557 * queue prior to zil_commit() having been called, and which itxs were
2558 * added after zil_commit() was called.
2560 * The commit it is special; it doesn't have any on-disk representation.
2561 * When a commit itx is "committed" to an lwb, the waiter associated
2562 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2563 * completes, each waiter on the lwb's list is marked done and signalled
2564 * -- allowing the thread waiting on the waiter to return from zil_commit().
2566 * It's important to point out a few critical factors that allow us
2567 * to make use of the commit itxs, commit waiters, per-lwb lists of
2568 * commit waiters, and zio completion callbacks like we're doing:
2570 * 1. The list of waiters for each lwb is traversed, and each commit
2571 * waiter is marked "done" and signalled, in the zio completion
2572 * callback of the lwb's zio[*].
2574 * * Actually, the waiters are signalled in the zio completion
2575 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2576 * that are sent to the vdevs upon completion of the lwb zio.
2578 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2579 * itxs, the order in which they are inserted is preserved[*]; as
2580 * itxs are added to the queue, they are added to the tail of
2581 * in-memory linked lists.
2583 * When committing the itxs to lwbs (to be written to disk), they
2584 * are committed in the same order in which the itxs were added to
2585 * the uncommitted queue's linked list(s); i.e. the linked list of
2586 * itxs to commit is traversed from head to tail, and each itx is
2587 * committed to an lwb in that order.
2591 * - the order of "sync" itxs is preserved w.r.t. other
2592 * "sync" itxs, regardless of the corresponding objects.
2593 * - the order of "async" itxs is preserved w.r.t. other
2594 * "async" itxs corresponding to the same object.
2595 * - the order of "async" itxs is *not* preserved w.r.t. other
2596 * "async" itxs corresponding to different objects.
2597 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2598 * versa) is *not* preserved, even for itxs that correspond
2599 * to the same object.
2601 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2602 * zil_get_commit_list(), and zil_process_commit_list().
2604 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2605 * lwb cannot be considered committed to stable storage, until its
2606 * "previous" lwb is also committed to stable storage. This fact,
2607 * coupled with the fact described above, means that itxs are
2608 * committed in (roughly) the order in which they were generated.
2609 * This is essential because itxs are dependent on prior itxs.
2610 * Thus, we *must not* deem an itx as being committed to stable
2611 * storage, until *all* prior itxs have also been committed to
2614 * To enforce this ordering of lwb zio's, while still leveraging as
2615 * much of the underlying storage performance as possible, we rely
2616 * on two fundamental concepts:
2618 * 1. The creation and issuance of lwb zio's is protected by
2619 * the zilog's "zl_issuer_lock", which ensures only a single
2620 * thread is creating and/or issuing lwb's at a time
2621 * 2. The "previous" lwb is a child of the "current" lwb
2622 * (leveraging the zio parent-child depenency graph)
2624 * By relying on this parent-child zio relationship, we can have
2625 * many lwb zio's concurrently issued to the underlying storage,
2626 * but the order in which they complete will be the same order in
2627 * which they were created.
2630 zil_commit(zilog_t *zilog, uint64_t foid)
2633 * We should never attempt to call zil_commit on a snapshot for
2634 * a couple of reasons:
2636 * 1. A snapshot may never be modified, thus it cannot have any
2637 * in-flight itxs that would have modified the dataset.
2639 * 2. By design, when zil_commit() is called, a commit itx will
2640 * be assigned to this zilog; as a result, the zilog will be
2641 * dirtied. We must not dirty the zilog of a snapshot; there's
2642 * checks in the code that enforce this invariant, and will
2643 * cause a panic if it's not upheld.
2645 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
2647 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2650 if (!spa_writeable(zilog->zl_spa)) {
2652 * If the SPA is not writable, there should never be any
2653 * pending itxs waiting to be committed to disk. If that
2654 * weren't true, we'd skip writing those itxs out, and
2655 * would break the sematics of zil_commit(); thus, we're
2656 * verifying that truth before we return to the caller.
2658 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2659 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2660 for (int i = 0; i < TXG_SIZE; i++)
2661 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
2666 * If the ZIL is suspended, we don't want to dirty it by calling
2667 * zil_commit_itx_assign() below, nor can we write out
2668 * lwbs like would be done in zil_commit_write(). Thus, we
2669 * simply rely on txg_wait_synced() to maintain the necessary
2670 * semantics, and avoid calling those functions altogether.
2672 if (zilog->zl_suspend > 0) {
2673 txg_wait_synced(zilog->zl_dmu_pool, 0);
2677 zil_commit_impl(zilog, foid);
2681 zil_commit_impl(zilog_t *zilog, uint64_t foid)
2684 * Move the "async" itxs for the specified foid to the "sync"
2685 * queues, such that they will be later committed (or skipped)
2686 * to an lwb when zil_process_commit_list() is called.
2688 * Since these "async" itxs must be committed prior to this
2689 * call to zil_commit returning, we must perform this operation
2690 * before we call zil_commit_itx_assign().
2692 zil_async_to_sync(zilog, foid);
2695 * We allocate a new "waiter" structure which will initially be
2696 * linked to the commit itx using the itx's "itx_private" field.
2697 * Since the commit itx doesn't represent any on-disk state,
2698 * when it's committed to an lwb, rather than copying the its
2699 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2700 * added to the lwb's list of waiters. Then, when the lwb is
2701 * committed to stable storage, each waiter in the lwb's list of
2702 * waiters will be marked "done", and signalled.
2704 * We must create the waiter and assign the commit itx prior to
2705 * calling zil_commit_writer(), or else our specific commit itx
2706 * is not guaranteed to be committed to an lwb prior to calling
2707 * zil_commit_waiter().
2709 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
2710 zil_commit_itx_assign(zilog, zcw);
2712 zil_commit_writer(zilog, zcw);
2713 zil_commit_waiter(zilog, zcw);
2715 if (zcw->zcw_zio_error != 0) {
2717 * If there was an error writing out the ZIL blocks that
2718 * this thread is waiting on, then we fallback to
2719 * relying on spa_sync() to write out the data this
2720 * thread is waiting on. Obviously this has performance
2721 * implications, but the expectation is for this to be
2722 * an exceptional case, and shouldn't occur often.
2724 DTRACE_PROBE2(zil__commit__io__error,
2725 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
2726 txg_wait_synced(zilog->zl_dmu_pool, 0);
2729 zil_free_commit_waiter(zcw);
2733 * Called in syncing context to free committed log blocks and update log header.
2736 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
2738 zil_header_t *zh = zil_header_in_syncing_context(zilog);
2739 uint64_t txg = dmu_tx_get_txg(tx);
2740 spa_t *spa = zilog->zl_spa;
2741 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
2745 * We don't zero out zl_destroy_txg, so make sure we don't try
2746 * to destroy it twice.
2748 if (spa_sync_pass(spa) != 1)
2751 mutex_enter(&zilog->zl_lock);
2753 ASSERT(zilog->zl_stop_sync == 0);
2755 if (*replayed_seq != 0) {
2756 ASSERT(zh->zh_replay_seq < *replayed_seq);
2757 zh->zh_replay_seq = *replayed_seq;
2761 if (zilog->zl_destroy_txg == txg) {
2762 blkptr_t blk = zh->zh_log;
2764 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
2766 bzero(zh, sizeof (zil_header_t));
2767 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
2769 if (zilog->zl_keep_first) {
2771 * If this block was part of log chain that couldn't
2772 * be claimed because a device was missing during
2773 * zil_claim(), but that device later returns,
2774 * then this block could erroneously appear valid.
2775 * To guard against this, assign a new GUID to the new
2776 * log chain so it doesn't matter what blk points to.
2778 zil_init_log_chain(zilog, &blk);
2783 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
2784 zh->zh_log = lwb->lwb_blk;
2785 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
2787 list_remove(&zilog->zl_lwb_list, lwb);
2788 zio_free(spa, txg, &lwb->lwb_blk);
2789 zil_free_lwb(zilog, lwb);
2792 * If we don't have anything left in the lwb list then
2793 * we've had an allocation failure and we need to zero
2794 * out the zil_header blkptr so that we don't end
2795 * up freeing the same block twice.
2797 if (list_head(&zilog->zl_lwb_list) == NULL)
2798 BP_ZERO(&zh->zh_log);
2800 mutex_exit(&zilog->zl_lock);
2805 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
2808 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
2809 offsetof(zil_commit_waiter_t, zcw_node));
2810 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
2811 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
2812 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
2818 zil_lwb_dest(void *vbuf, void *unused)
2821 mutex_destroy(&lwb->lwb_vdev_lock);
2822 avl_destroy(&lwb->lwb_vdev_tree);
2823 list_destroy(&lwb->lwb_waiters);
2829 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
2830 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
2832 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
2833 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2839 kmem_cache_destroy(zil_zcw_cache);
2840 kmem_cache_destroy(zil_lwb_cache);
2844 zil_set_sync(zilog_t *zilog, uint64_t sync)
2846 zilog->zl_sync = sync;
2850 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
2852 zilog->zl_logbias = logbias;
2856 zil_alloc(objset_t *os, zil_header_t *zh_phys)
2860 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
2862 zilog->zl_header = zh_phys;
2864 zilog->zl_spa = dmu_objset_spa(os);
2865 zilog->zl_dmu_pool = dmu_objset_pool(os);
2866 zilog->zl_destroy_txg = TXG_INITIAL - 1;
2867 zilog->zl_logbias = dmu_objset_logbias(os);
2868 zilog->zl_sync = dmu_objset_syncprop(os);
2869 zilog->zl_dirty_max_txg = 0;
2870 zilog->zl_last_lwb_opened = NULL;
2871 zilog->zl_last_lwb_latency = 0;
2873 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
2874 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
2876 for (int i = 0; i < TXG_SIZE; i++) {
2877 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
2878 MUTEX_DEFAULT, NULL);
2881 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
2882 offsetof(lwb_t, lwb_node));
2884 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
2885 offsetof(itx_t, itx_node));
2887 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
2893 zil_free(zilog_t *zilog)
2895 zilog->zl_stop_sync = 1;
2897 ASSERT0(zilog->zl_suspend);
2898 ASSERT0(zilog->zl_suspending);
2900 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2901 list_destroy(&zilog->zl_lwb_list);
2903 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
2904 list_destroy(&zilog->zl_itx_commit_list);
2906 for (int i = 0; i < TXG_SIZE; i++) {
2908 * It's possible for an itx to be generated that doesn't dirty
2909 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
2910 * callback to remove the entry. We remove those here.
2912 * Also free up the ziltest itxs.
2914 if (zilog->zl_itxg[i].itxg_itxs)
2915 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
2916 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
2919 mutex_destroy(&zilog->zl_issuer_lock);
2920 mutex_destroy(&zilog->zl_lock);
2922 cv_destroy(&zilog->zl_cv_suspend);
2924 kmem_free(zilog, sizeof (zilog_t));
2928 * Open an intent log.
2931 zil_open(objset_t *os, zil_get_data_t *get_data)
2933 zilog_t *zilog = dmu_objset_zil(os);
2935 ASSERT3P(zilog->zl_get_data, ==, NULL);
2936 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
2937 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2939 zilog->zl_get_data = get_data;
2945 * Close an intent log.
2948 zil_close(zilog_t *zilog)
2953 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
2954 zil_commit(zilog, 0);
2956 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL);
2957 ASSERT0(zilog->zl_dirty_max_txg);
2958 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
2961 mutex_enter(&zilog->zl_lock);
2962 lwb = list_tail(&zilog->zl_lwb_list);
2964 txg = zilog->zl_dirty_max_txg;
2966 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
2967 mutex_exit(&zilog->zl_lock);
2970 * We need to use txg_wait_synced() to wait long enough for the
2971 * ZIL to be clean, and to wait for all pending lwbs to be
2975 txg_wait_synced(zilog->zl_dmu_pool, txg);
2977 if (txg < spa_freeze_txg(zilog->zl_spa))
2978 ASSERT(!zilog_is_dirty(zilog));
2980 zilog->zl_get_data = NULL;
2983 * We should have only one lwb left on the list; remove it now.
2985 mutex_enter(&zilog->zl_lock);
2986 lwb = list_head(&zilog->zl_lwb_list);
2988 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list));
2989 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2990 list_remove(&zilog->zl_lwb_list, lwb);
2991 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2992 zil_free_lwb(zilog, lwb);
2994 mutex_exit(&zilog->zl_lock);
2997 static char *suspend_tag = "zil suspending";
3000 * Suspend an intent log. While in suspended mode, we still honor
3001 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3002 * On old version pools, we suspend the log briefly when taking a
3003 * snapshot so that it will have an empty intent log.
3005 * Long holds are not really intended to be used the way we do here --
3006 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3007 * could fail. Therefore we take pains to only put a long hold if it is
3008 * actually necessary. Fortunately, it will only be necessary if the
3009 * objset is currently mounted (or the ZVOL equivalent). In that case it
3010 * will already have a long hold, so we are not really making things any worse.
3012 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3013 * zvol_state_t), and use their mechanism to prevent their hold from being
3014 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3017 * if cookiep == NULL, this does both the suspend & resume.
3018 * Otherwise, it returns with the dataset "long held", and the cookie
3019 * should be passed into zil_resume().
3022 zil_suspend(const char *osname, void **cookiep)
3026 const zil_header_t *zh;
3029 error = dmu_objset_hold(osname, suspend_tag, &os);
3032 zilog = dmu_objset_zil(os);
3034 mutex_enter(&zilog->zl_lock);
3035 zh = zilog->zl_header;
3037 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
3038 mutex_exit(&zilog->zl_lock);
3039 dmu_objset_rele(os, suspend_tag);
3040 return (SET_ERROR(EBUSY));
3044 * Don't put a long hold in the cases where we can avoid it. This
3045 * is when there is no cookie so we are doing a suspend & resume
3046 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3047 * for the suspend because it's already suspended, or there's no ZIL.
3049 if (cookiep == NULL && !zilog->zl_suspending &&
3050 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3051 mutex_exit(&zilog->zl_lock);
3052 dmu_objset_rele(os, suspend_tag);
3056 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3057 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3059 zilog->zl_suspend++;
3061 if (zilog->zl_suspend > 1) {
3063 * Someone else is already suspending it.
3064 * Just wait for them to finish.
3067 while (zilog->zl_suspending)
3068 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3069 mutex_exit(&zilog->zl_lock);
3071 if (cookiep == NULL)
3079 * If there is no pointer to an on-disk block, this ZIL must not
3080 * be active (e.g. filesystem not mounted), so there's nothing
3083 if (BP_IS_HOLE(&zh->zh_log)) {
3084 ASSERT(cookiep != NULL); /* fast path already handled */
3087 mutex_exit(&zilog->zl_lock);
3091 zilog->zl_suspending = B_TRUE;
3092 mutex_exit(&zilog->zl_lock);
3095 * We need to use zil_commit_impl to ensure we wait for all
3096 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed
3097 * to disk before proceeding. If we used zil_commit instead, it
3098 * would just call txg_wait_synced(), because zl_suspend is set.
3099 * txg_wait_synced() doesn't wait for these lwb's to be
3100 * LWB_STATE_DONE before returning.
3102 zil_commit_impl(zilog, 0);
3105 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
3106 * txg_wait_synced() to ensure the data from the zilog has
3107 * migrated to the main pool before calling zil_destroy().
3109 txg_wait_synced(zilog->zl_dmu_pool, 0);
3111 zil_destroy(zilog, B_FALSE);
3113 mutex_enter(&zilog->zl_lock);
3114 zilog->zl_suspending = B_FALSE;
3115 cv_broadcast(&zilog->zl_cv_suspend);
3116 mutex_exit(&zilog->zl_lock);
3118 if (cookiep == NULL)
3126 zil_resume(void *cookie)
3128 objset_t *os = cookie;
3129 zilog_t *zilog = dmu_objset_zil(os);
3131 mutex_enter(&zilog->zl_lock);
3132 ASSERT(zilog->zl_suspend != 0);
3133 zilog->zl_suspend--;
3134 mutex_exit(&zilog->zl_lock);
3135 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3136 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3139 typedef struct zil_replay_arg {
3140 zil_replay_func_t **zr_replay;
3142 boolean_t zr_byteswap;
3147 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
3149 char name[ZFS_MAX_DATASET_NAME_LEN];
3151 zilog->zl_replaying_seq--; /* didn't actually replay this one */
3153 dmu_objset_name(zilog->zl_os, name);
3155 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
3156 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
3157 (u_longlong_t)lr->lrc_seq,
3158 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
3159 (lr->lrc_txtype & TX_CI) ? "CI" : "");
3165 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
3167 zil_replay_arg_t *zr = zra;
3168 const zil_header_t *zh = zilog->zl_header;
3169 uint64_t reclen = lr->lrc_reclen;
3170 uint64_t txtype = lr->lrc_txtype;
3173 zilog->zl_replaying_seq = lr->lrc_seq;
3175 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
3178 if (lr->lrc_txg < claim_txg) /* already committed */
3181 /* Strip case-insensitive bit, still present in log record */
3184 if (txtype == 0 || txtype >= TX_MAX_TYPE)
3185 return (zil_replay_error(zilog, lr, EINVAL));
3188 * If this record type can be logged out of order, the object
3189 * (lr_foid) may no longer exist. That's legitimate, not an error.
3191 if (TX_OOO(txtype)) {
3192 error = dmu_object_info(zilog->zl_os,
3193 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
3194 if (error == ENOENT || error == EEXIST)
3199 * Make a copy of the data so we can revise and extend it.
3201 bcopy(lr, zr->zr_lr, reclen);
3204 * If this is a TX_WRITE with a blkptr, suck in the data.
3206 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
3207 error = zil_read_log_data(zilog, (lr_write_t *)lr,
3208 zr->zr_lr + reclen);
3210 return (zil_replay_error(zilog, lr, error));
3214 * The log block containing this lr may have been byteswapped
3215 * so that we can easily examine common fields like lrc_txtype.
3216 * However, the log is a mix of different record types, and only the
3217 * replay vectors know how to byteswap their records. Therefore, if
3218 * the lr was byteswapped, undo it before invoking the replay vector.
3220 if (zr->zr_byteswap)
3221 byteswap_uint64_array(zr->zr_lr, reclen);
3224 * We must now do two things atomically: replay this log record,
3225 * and update the log header sequence number to reflect the fact that
3226 * we did so. At the end of each replay function the sequence number
3227 * is updated if we are in replay mode.
3229 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
3232 * The DMU's dnode layer doesn't see removes until the txg
3233 * commits, so a subsequent claim can spuriously fail with
3234 * EEXIST. So if we receive any error we try syncing out
3235 * any removes then retry the transaction. Note that we
3236 * specify B_FALSE for byteswap now, so we don't do it twice.
3238 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
3239 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
3241 return (zil_replay_error(zilog, lr, error));
3248 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
3250 zilog->zl_replay_blks++;
3256 * If this dataset has a non-empty intent log, replay it and destroy it.
3259 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
3261 zilog_t *zilog = dmu_objset_zil(os);
3262 const zil_header_t *zh = zilog->zl_header;
3263 zil_replay_arg_t zr;
3265 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
3266 zil_destroy(zilog, B_TRUE);
3270 zr.zr_replay = replay_func;
3272 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
3273 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
3276 * Wait for in-progress removes to sync before starting replay.
3278 txg_wait_synced(zilog->zl_dmu_pool, 0);
3280 zilog->zl_replay = B_TRUE;
3281 zilog->zl_replay_time = ddi_get_lbolt();
3282 ASSERT(zilog->zl_replay_blks == 0);
3283 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
3285 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
3287 zil_destroy(zilog, B_FALSE);
3288 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
3289 zilog->zl_replay = B_FALSE;
3293 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
3295 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3298 if (zilog->zl_replay) {
3299 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
3300 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
3301 zilog->zl_replaying_seq;
3310 zil_reset(const char *osname, void *arg)
3314 error = zil_suspend(osname, NULL);
3316 return (SET_ERROR(EEXIST));