4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
27 /* Portions Copyright 2010 Robert Milkowski */
29 #include <sys/zfs_context.h>
35 #include <sys/resource.h>
37 #include <sys/zil_impl.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/dmu_tx.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/metaslab.h>
43 #include <sys/trace_zil.h>
47 * The zfs intent log (ZIL) saves transaction records of system calls
48 * that change the file system in memory with enough information
49 * to be able to replay them. These are stored in memory until
50 * either the DMU transaction group (txg) commits them to the stable pool
51 * and they can be discarded, or they are flushed to the stable log
52 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
53 * requirement. In the event of a panic or power fail then those log
54 * records (transactions) are replayed.
56 * There is one ZIL per file system. Its on-disk (pool) format consists
63 * A log record holds a system call transaction. Log blocks can
64 * hold many log records and the blocks are chained together.
65 * Each ZIL block contains a block pointer (blkptr_t) to the next
66 * ZIL block in the chain. The ZIL header points to the first
67 * block in the chain. Note there is not a fixed place in the pool
68 * to hold blocks. They are dynamically allocated and freed as
69 * needed from the blocks available. Figure X shows the ZIL structure:
73 * See zil.h for more information about these fields.
75 zil_stats_t zil_stats = {
76 { "zil_commit_count", KSTAT_DATA_UINT64 },
77 { "zil_commit_writer_count", KSTAT_DATA_UINT64 },
78 { "zil_itx_count", KSTAT_DATA_UINT64 },
79 { "zil_itx_indirect_count", KSTAT_DATA_UINT64 },
80 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 },
81 { "zil_itx_copied_count", KSTAT_DATA_UINT64 },
82 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 },
83 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 },
84 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 },
85 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 },
86 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 },
87 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 },
88 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 },
91 static kstat_t *zil_ksp;
94 * Disable intent logging replay. This global ZIL switch affects all pools.
96 int zil_replay_disable = 0;
99 * Tunable parameter for debugging or performance analysis. Setting
100 * zfs_nocacheflush will cause corruption on power loss if a volatile
101 * out-of-order write cache is enabled.
103 int zfs_nocacheflush = 0;
106 * Limit SLOG write size per commit executed with synchronous priority.
107 * Any writes above that will be executed with lower (asynchronous) priority
108 * to limit potential SLOG device abuse by single active ZIL writer.
110 unsigned long zil_slog_bulk = 768 * 1024;
112 static kmem_cache_t *zil_lwb_cache;
114 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
116 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
117 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
120 zil_bp_compare(const void *x1, const void *x2)
122 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
123 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
125 int cmp = AVL_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
129 return (AVL_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
133 zil_bp_tree_init(zilog_t *zilog)
135 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
136 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
140 zil_bp_tree_fini(zilog_t *zilog)
142 avl_tree_t *t = &zilog->zl_bp_tree;
146 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
147 kmem_free(zn, sizeof (zil_bp_node_t));
153 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
155 avl_tree_t *t = &zilog->zl_bp_tree;
160 if (BP_IS_EMBEDDED(bp))
163 dva = BP_IDENTITY(bp);
165 if (avl_find(t, dva, &where) != NULL)
166 return (SET_ERROR(EEXIST));
168 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
170 avl_insert(t, zn, where);
175 static zil_header_t *
176 zil_header_in_syncing_context(zilog_t *zilog)
178 return ((zil_header_t *)zilog->zl_header);
182 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
184 zio_cksum_t *zc = &bp->blk_cksum;
186 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
187 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
188 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
189 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
193 * Read a log block and make sure it's valid.
196 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
197 blkptr_t *nbp, void *dst, char **end)
199 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
200 arc_flags_t aflags = ARC_FLAG_WAIT;
201 arc_buf_t *abuf = NULL;
205 if (zilog->zl_header->zh_claim_txg == 0)
206 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
208 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
209 zio_flags |= ZIO_FLAG_SPECULATIVE;
212 zio_flags |= ZIO_FLAG_RAW;
214 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
215 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
217 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
218 &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
221 zio_cksum_t cksum = bp->blk_cksum;
224 * Validate the checksummed log block.
226 * Sequence numbers should be... sequential. The checksum
227 * verifier for the next block should be bp's checksum plus 1.
229 * Also check the log chain linkage and size used.
231 cksum.zc_word[ZIL_ZC_SEQ]++;
233 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
234 zil_chain_t *zilc = abuf->b_data;
235 char *lr = (char *)(zilc + 1);
236 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
238 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
239 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
240 error = SET_ERROR(ECKSUM);
242 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
244 *end = (char *)dst + len;
245 *nbp = zilc->zc_next_blk;
248 char *lr = abuf->b_data;
249 uint64_t size = BP_GET_LSIZE(bp);
250 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
252 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
253 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
254 (zilc->zc_nused > (size - sizeof (*zilc)))) {
255 error = SET_ERROR(ECKSUM);
257 ASSERT3U(zilc->zc_nused, <=,
258 SPA_OLD_MAXBLOCKSIZE);
259 bcopy(lr, dst, zilc->zc_nused);
260 *end = (char *)dst + zilc->zc_nused;
261 *nbp = zilc->zc_next_blk;
265 arc_buf_destroy(abuf, &abuf);
272 * Read a TX_WRITE log data block.
275 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
277 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
278 const blkptr_t *bp = &lr->lr_blkptr;
279 arc_flags_t aflags = ARC_FLAG_WAIT;
280 arc_buf_t *abuf = NULL;
284 if (BP_IS_HOLE(bp)) {
286 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
290 if (zilog->zl_header->zh_claim_txg == 0)
291 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
294 * If we are not using the resulting data, we are just checking that
295 * it hasn't been corrupted so we don't need to waste CPU time
296 * decompressing and decrypting it.
299 zio_flags |= ZIO_FLAG_RAW;
301 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
302 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
304 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
305 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
309 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
310 arc_buf_destroy(abuf, &abuf);
317 * Parse the intent log, and call parse_func for each valid record within.
320 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
321 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
324 const zil_header_t *zh = zilog->zl_header;
325 boolean_t claimed = !!zh->zh_claim_txg;
326 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
327 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
328 uint64_t max_blk_seq = 0;
329 uint64_t max_lr_seq = 0;
330 uint64_t blk_count = 0;
331 uint64_t lr_count = 0;
332 blkptr_t blk, next_blk;
336 bzero(&next_blk, sizeof (blkptr_t));
339 * Old logs didn't record the maximum zh_claim_lr_seq.
341 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
342 claim_lr_seq = UINT64_MAX;
345 * Starting at the block pointed to by zh_log we read the log chain.
346 * For each block in the chain we strongly check that block to
347 * ensure its validity. We stop when an invalid block is found.
348 * For each block pointer in the chain we call parse_blk_func().
349 * For each record in each valid block we call parse_lr_func().
350 * If the log has been claimed, stop if we encounter a sequence
351 * number greater than the highest claimed sequence number.
353 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
354 zil_bp_tree_init(zilog);
356 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
357 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
361 if (blk_seq > claim_blk_seq)
364 error = parse_blk_func(zilog, &blk, arg, txg);
367 ASSERT3U(max_blk_seq, <, blk_seq);
368 max_blk_seq = blk_seq;
371 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
374 error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
379 for (lrp = lrbuf; lrp < end; lrp += reclen) {
380 lr_t *lr = (lr_t *)lrp;
381 reclen = lr->lrc_reclen;
382 ASSERT3U(reclen, >=, sizeof (lr_t));
383 if (lr->lrc_seq > claim_lr_seq)
386 error = parse_lr_func(zilog, lr, arg, txg);
389 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
390 max_lr_seq = lr->lrc_seq;
395 zilog->zl_parse_error = error;
396 zilog->zl_parse_blk_seq = max_blk_seq;
397 zilog->zl_parse_lr_seq = max_lr_seq;
398 zilog->zl_parse_blk_count = blk_count;
399 zilog->zl_parse_lr_count = lr_count;
401 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
402 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq) ||
403 (decrypt && error == EIO));
405 zil_bp_tree_fini(zilog);
406 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
412 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
415 * Claim log block if not already committed and not already claimed.
416 * If tx == NULL, just verify that the block is claimable.
418 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
419 zil_bp_tree_add(zilog, bp) != 0)
422 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
423 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
424 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
428 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
430 lr_write_t *lr = (lr_write_t *)lrc;
433 if (lrc->lrc_txtype != TX_WRITE)
437 * If the block is not readable, don't claim it. This can happen
438 * in normal operation when a log block is written to disk before
439 * some of the dmu_sync() blocks it points to. In this case, the
440 * transaction cannot have been committed to anyone (we would have
441 * waited for all writes to be stable first), so it is semantically
442 * correct to declare this the end of the log.
444 if (lr->lr_blkptr.blk_birth >= first_txg) {
445 error = zil_read_log_data(zilog, lr, NULL);
450 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
455 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
457 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
463 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
465 lr_write_t *lr = (lr_write_t *)lrc;
466 blkptr_t *bp = &lr->lr_blkptr;
469 * If we previously claimed it, we need to free it.
471 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
472 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
474 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
480 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
485 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
486 lwb->lwb_zilog = zilog;
488 lwb->lwb_fastwrite = fastwrite;
489 lwb->lwb_slog = slog;
490 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
491 lwb->lwb_max_txg = txg;
494 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
495 lwb->lwb_nused = sizeof (zil_chain_t);
496 lwb->lwb_sz = BP_GET_LSIZE(bp);
499 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
502 mutex_enter(&zilog->zl_lock);
503 list_insert_tail(&zilog->zl_lwb_list, lwb);
504 mutex_exit(&zilog->zl_lock);
510 * Called when we create in-memory log transactions so that we know
511 * to cleanup the itxs at the end of spa_sync().
514 zilog_dirty(zilog_t *zilog, uint64_t txg)
516 dsl_pool_t *dp = zilog->zl_dmu_pool;
517 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
519 if (ds->ds_is_snapshot)
520 panic("dirtying snapshot!");
522 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
523 /* up the hold count until we can be written out */
524 dmu_buf_add_ref(ds->ds_dbuf, zilog);
529 * Determine if the zil is dirty in the specified txg. Callers wanting to
530 * ensure that the dirty state does not change must hold the itxg_lock for
531 * the specified txg. Holding the lock will ensure that the zil cannot be
532 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
536 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
538 dsl_pool_t *dp = zilog->zl_dmu_pool;
540 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
546 * Determine if the zil is dirty. The zil is considered dirty if it has
547 * any pending itx records that have not been cleaned by zil_clean().
550 zilog_is_dirty(zilog_t *zilog)
552 dsl_pool_t *dp = zilog->zl_dmu_pool;
554 for (int t = 0; t < TXG_SIZE; t++) {
555 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
562 * Create an on-disk intent log.
565 zil_create(zilog_t *zilog)
567 const zil_header_t *zh = zilog->zl_header;
573 boolean_t fastwrite = FALSE;
574 boolean_t slog = FALSE;
577 * Wait for any previous destroy to complete.
579 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
581 ASSERT(zh->zh_claim_txg == 0);
582 ASSERT(zh->zh_replay_seq == 0);
587 * Allocate an initial log block if:
588 * - there isn't one already
589 * - the existing block is the wrong endianness
591 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
592 tx = dmu_tx_create(zilog->zl_os);
593 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
594 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
595 txg = dmu_tx_get_txg(tx);
597 if (!BP_IS_HOLE(&blk)) {
598 zio_free_zil(zilog->zl_spa, txg, &blk);
602 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
603 ZIL_MIN_BLKSZ, &slog);
607 zil_init_log_chain(zilog, &blk);
611 * Allocate a log write buffer (lwb) for the first log block.
614 lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
617 * If we just allocated the first log block, commit our transaction
618 * and wait for zil_sync() to stuff the block poiner into zh_log.
619 * (zh is part of the MOS, so we cannot modify it in open context.)
623 txg_wait_synced(zilog->zl_dmu_pool, txg);
626 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
632 * In one tx, free all log blocks and clear the log header.
633 * If keep_first is set, then we're replaying a log with no content.
634 * We want to keep the first block, however, so that the first
635 * synchronous transaction doesn't require a txg_wait_synced()
636 * in zil_create(). We don't need to txg_wait_synced() here either
637 * when keep_first is set, because both zil_create() and zil_destroy()
638 * will wait for any in-progress destroys to complete.
641 zil_destroy(zilog_t *zilog, boolean_t keep_first)
643 const zil_header_t *zh = zilog->zl_header;
649 * Wait for any previous destroy to complete.
651 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
653 zilog->zl_old_header = *zh; /* debugging aid */
655 if (BP_IS_HOLE(&zh->zh_log))
658 tx = dmu_tx_create(zilog->zl_os);
659 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
660 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
661 txg = dmu_tx_get_txg(tx);
663 mutex_enter(&zilog->zl_lock);
665 ASSERT3U(zilog->zl_destroy_txg, <, txg);
666 zilog->zl_destroy_txg = txg;
667 zilog->zl_keep_first = keep_first;
669 if (!list_is_empty(&zilog->zl_lwb_list)) {
670 ASSERT(zh->zh_claim_txg == 0);
672 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
673 ASSERT(lwb->lwb_zio == NULL);
674 if (lwb->lwb_fastwrite)
675 metaslab_fastwrite_unmark(zilog->zl_spa,
677 list_remove(&zilog->zl_lwb_list, lwb);
678 if (lwb->lwb_buf != NULL)
679 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
680 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
681 kmem_cache_free(zil_lwb_cache, lwb);
683 } else if (!keep_first) {
684 zil_destroy_sync(zilog, tx);
686 mutex_exit(&zilog->zl_lock);
692 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
694 ASSERT(list_is_empty(&zilog->zl_lwb_list));
695 (void) zil_parse(zilog, zil_free_log_block,
696 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
700 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
702 dmu_tx_t *tx = txarg;
703 uint64_t first_txg = dmu_tx_get_txg(tx);
709 error = dmu_objset_own_obj(dp, ds->ds_object,
710 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
713 * EBUSY indicates that the objset is inconsistent, in which
714 * case it can not have a ZIL.
716 if (error != EBUSY) {
717 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
718 (unsigned long long)ds->ds_object, error);
724 zilog = dmu_objset_zil(os);
725 zh = zil_header_in_syncing_context(zilog);
727 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
728 if (!BP_IS_HOLE(&zh->zh_log))
729 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
730 BP_ZERO(&zh->zh_log);
731 if (os->os_encrypted)
732 os->os_next_write_raw = B_TRUE;
733 dsl_dataset_dirty(dmu_objset_ds(os), tx);
734 dmu_objset_disown(os, B_FALSE, FTAG);
739 * Claim all log blocks if we haven't already done so, and remember
740 * the highest claimed sequence number. This ensures that if we can
741 * read only part of the log now (e.g. due to a missing device),
742 * but we can read the entire log later, we will not try to replay
743 * or destroy beyond the last block we successfully claimed.
745 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
746 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
747 (void) zil_parse(zilog, zil_claim_log_block,
748 zil_claim_log_record, tx, first_txg, B_FALSE);
749 zh->zh_claim_txg = first_txg;
750 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
751 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
752 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
753 zh->zh_flags |= ZIL_REPLAY_NEEDED;
754 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
755 dsl_dataset_dirty(dmu_objset_ds(os), tx);
758 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
759 dmu_objset_disown(os, B_FALSE, FTAG);
764 * Check the log by walking the log chain.
765 * Checksum errors are ok as they indicate the end of the chain.
766 * Any other error (no device or read failure) returns an error.
770 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
779 error = dmu_objset_from_ds(ds, &os);
781 cmn_err(CE_WARN, "can't open objset %llu, error %d",
782 (unsigned long long)ds->ds_object, error);
786 zilog = dmu_objset_zil(os);
787 bp = (blkptr_t *)&zilog->zl_header->zh_log;
790 * Check the first block and determine if it's on a log device
791 * which may have been removed or faulted prior to loading this
792 * pool. If so, there's no point in checking the rest of the log
793 * as its content should have already been synced to the pool.
795 if (!BP_IS_HOLE(bp)) {
797 boolean_t valid = B_TRUE;
799 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
800 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
801 if (vd->vdev_islog && vdev_is_dead(vd))
802 valid = vdev_log_state_valid(vd);
803 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
810 * Because tx == NULL, zil_claim_log_block() will not actually claim
811 * any blocks, but just determine whether it is possible to do so.
812 * In addition to checking the log chain, zil_claim_log_block()
813 * will invoke zio_claim() with a done func of spa_claim_notify(),
814 * which will update spa_max_claim_txg. See spa_load() for details.
816 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
817 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa),
820 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
824 zil_vdev_compare(const void *x1, const void *x2)
826 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
827 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
829 return (AVL_CMP(v1, v2));
833 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
835 avl_tree_t *t = &zilog->zl_vdev_tree;
837 zil_vdev_node_t *zv, zvsearch;
838 int ndvas = BP_GET_NDVAS(bp);
841 if (zfs_nocacheflush)
844 ASSERT(zilog->zl_writer);
847 * Even though we're zl_writer, we still need a lock because the
848 * zl_get_data() callbacks may have dmu_sync() done callbacks
849 * that will run concurrently.
851 mutex_enter(&zilog->zl_vdev_lock);
852 for (i = 0; i < ndvas; i++) {
853 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
854 if (avl_find(t, &zvsearch, &where) == NULL) {
855 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
856 zv->zv_vdev = zvsearch.zv_vdev;
857 avl_insert(t, zv, where);
860 mutex_exit(&zilog->zl_vdev_lock);
864 zil_flush_vdevs(zilog_t *zilog)
866 spa_t *spa = zilog->zl_spa;
867 avl_tree_t *t = &zilog->zl_vdev_tree;
872 ASSERT(zilog->zl_writer);
875 * We don't need zl_vdev_lock here because we're the zl_writer,
876 * and all zl_get_data() callbacks are done.
878 if (avl_numnodes(t) == 0)
881 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
883 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
885 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
886 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
889 kmem_free(zv, sizeof (*zv));
893 * Wait for all the flushes to complete. Not all devices actually
894 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
896 (void) zio_wait(zio);
898 spa_config_exit(spa, SCL_STATE, FTAG);
902 * Function called when a log block write completes
905 zil_lwb_write_done(zio_t *zio)
907 lwb_t *lwb = zio->io_private;
908 zilog_t *zilog = lwb->lwb_zilog;
909 dmu_tx_t *tx = lwb->lwb_tx;
911 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
912 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
913 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
914 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
915 ASSERT(!BP_IS_GANG(zio->io_bp));
916 ASSERT(!BP_IS_HOLE(zio->io_bp));
917 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
920 * Ensure the lwb buffer pointer is cleared before releasing
921 * the txg. If we have had an allocation failure and
922 * the txg is waiting to sync then we want want zil_sync()
923 * to remove the lwb so that it's not picked up as the next new
924 * one in zil_commit_writer(). zil_sync() will only remove
925 * the lwb if lwb_buf is null.
927 abd_put(zio->io_abd);
928 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
929 mutex_enter(&zilog->zl_lock);
931 lwb->lwb_fastwrite = FALSE;
934 mutex_exit(&zilog->zl_lock);
937 * Now that we've written this log block, we have a stable pointer
938 * to the next block in the chain, so it's OK to let the txg in
939 * which we allocated the next block sync.
945 * Initialize the io for a log block.
948 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
953 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
954 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
955 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
957 if (zilog->zl_root_zio == NULL) {
958 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
962 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
963 mutex_enter(&zilog->zl_lock);
964 if (lwb->lwb_zio == NULL) {
965 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
966 BP_GET_LSIZE(&lwb->lwb_blk));
967 if (!lwb->lwb_fastwrite) {
968 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
969 lwb->lwb_fastwrite = 1;
971 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
972 prio = ZIO_PRIORITY_SYNC_WRITE;
974 prio = ZIO_PRIORITY_ASYNC_WRITE;
975 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
976 0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
977 zil_lwb_write_done, lwb, prio,
978 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
979 ZIO_FLAG_FASTWRITE, &zb);
981 mutex_exit(&zilog->zl_lock);
985 * Define a limited set of intent log block sizes.
987 * These must be a multiple of 4KB. Note only the amount used (again
988 * aligned to 4KB) actually gets written. However, we can't always just
989 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
991 uint64_t zil_block_buckets[] = {
992 4096, /* non TX_WRITE */
993 8192+4096, /* data base */
994 32*1024 + 4096, /* NFS writes */
999 * Start a log block write and advance to the next log block.
1000 * Calls are serialized.
1003 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
1007 spa_t *spa = zilog->zl_spa;
1011 uint64_t zil_blksz, wsz;
1015 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1016 zilc = (zil_chain_t *)lwb->lwb_buf;
1017 bp = &zilc->zc_next_blk;
1019 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1020 bp = &zilc->zc_next_blk;
1023 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1026 * Allocate the next block and save its address in this block
1027 * before writing it in order to establish the log chain.
1028 * Note that if the allocation of nlwb synced before we wrote
1029 * the block that points at it (lwb), we'd leak it if we crashed.
1030 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1031 * We dirty the dataset to ensure that zil_sync() will be called
1032 * to clean up in the event of allocation failure or I/O failure.
1034 tx = dmu_tx_create(zilog->zl_os);
1037 * Since we are not going to create any new dirty data and we can even
1038 * help with clearing the existing dirty data, we should not be subject
1039 * to the dirty data based delays.
1040 * We (ab)use TXG_WAITED to bypass the delay mechanism.
1041 * One side effect from using TXG_WAITED is that dmu_tx_assign() can
1042 * fail if the pool is suspended. Those are dramatic circumstances,
1043 * so we return NULL to signal that the normal ZIL processing is not
1044 * possible and txg_wait_synced() should be used to ensure that the data
1047 error = dmu_tx_assign(tx, TXG_WAITED);
1049 ASSERT3S(error, ==, EIO);
1053 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1054 txg = dmu_tx_get_txg(tx);
1059 * Log blocks are pre-allocated. Here we select the size of the next
1060 * block, based on size used in the last block.
1061 * - first find the smallest bucket that will fit the block from a
1062 * limited set of block sizes. This is because it's faster to write
1063 * blocks allocated from the same metaslab as they are adjacent or
1065 * - next find the maximum from the new suggested size and an array of
1066 * previous sizes. This lessens a picket fence effect of wrongly
1067 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1070 * Note we only write what is used, but we can't just allocate
1071 * the maximum block size because we can exhaust the available
1074 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1075 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1077 zil_blksz = zil_block_buckets[i];
1078 if (zil_blksz == UINT64_MAX)
1079 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1080 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1081 for (i = 0; i < ZIL_PREV_BLKS; i++)
1082 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1083 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1086 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1088 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
1089 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
1091 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
1092 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
1095 ASSERT3U(bp->blk_birth, ==, txg);
1096 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1097 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1100 * Allocate a new log write buffer (lwb).
1102 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1104 /* Record the block for later vdev flushing */
1105 zil_add_block(zilog, &lwb->lwb_blk);
1108 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1109 /* For Slim ZIL only write what is used. */
1110 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1111 ASSERT3U(wsz, <=, lwb->lwb_sz);
1112 zio_shrink(lwb->lwb_zio, wsz);
1119 zilc->zc_nused = lwb->lwb_nused;
1120 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1123 * clear unused data for security
1125 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1127 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1130 * If there was an allocation failure then nlwb will be null which
1131 * forces a txg_wait_synced().
1137 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1140 lr_write_t *lrwb, *lrw;
1142 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1147 ASSERT(lwb->lwb_buf != NULL);
1149 lrc = &itx->itx_lr; /* Common log record inside itx. */
1150 lrw = (lr_write_t *)lrc; /* Write log record inside itx. */
1151 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1152 dlen = P2ROUNDUP_TYPED(
1153 lrw->lr_length, sizeof (uint64_t), uint64_t);
1157 reclen = lrc->lrc_reclen;
1158 zilog->zl_cur_used += (reclen + dlen);
1161 zil_lwb_write_init(zilog, lwb);
1165 * If this record won't fit in the current log block, start a new one.
1166 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1168 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1169 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1170 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1171 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1172 lwb = zil_lwb_write_start(zilog, lwb);
1175 zil_lwb_write_init(zilog, lwb);
1176 ASSERT(LWB_EMPTY(lwb));
1177 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1178 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1181 dnow = MIN(dlen, lwb_sp - reclen);
1182 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1183 bcopy(lrc, lr_buf, reclen);
1184 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1185 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1187 ZIL_STAT_BUMP(zil_itx_count);
1190 * If it's a write, fetch the data or get its blkptr as appropriate.
1192 if (lrc->lrc_txtype == TX_WRITE) {
1193 if (txg > spa_freeze_txg(zilog->zl_spa))
1194 txg_wait_synced(zilog->zl_dmu_pool, txg);
1195 if (itx->itx_wr_state == WR_COPIED) {
1196 ZIL_STAT_BUMP(zil_itx_copied_count);
1197 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1202 if (itx->itx_wr_state == WR_NEED_COPY) {
1203 dbuf = lr_buf + reclen;
1204 lrcb->lrc_reclen += dnow;
1205 if (lrwb->lr_length > dnow)
1206 lrwb->lr_length = dnow;
1207 lrw->lr_offset += dnow;
1208 lrw->lr_length -= dnow;
1209 ZIL_STAT_BUMP(zil_itx_needcopy_count);
1210 ZIL_STAT_INCR(zil_itx_needcopy_bytes,
1213 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1215 ZIL_STAT_BUMP(zil_itx_indirect_count);
1216 ZIL_STAT_INCR(zil_itx_indirect_bytes,
1219 error = zilog->zl_get_data(
1220 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1222 txg_wait_synced(zilog->zl_dmu_pool, txg);
1226 ASSERT(error == ENOENT || error == EEXIST ||
1234 * We're actually making an entry, so update lrc_seq to be the
1235 * log record sequence number. Note that this is generally not
1236 * equal to the itx sequence number because not all transactions
1237 * are synchronous, and sometimes spa_sync() gets there first.
1239 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1240 lwb->lwb_nused += reclen + dnow;
1241 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1242 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1243 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1247 zilog->zl_cur_used += reclen;
1255 zil_itx_create(uint64_t txtype, size_t lrsize)
1259 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1261 itx = zio_data_buf_alloc(offsetof(itx_t, itx_lr) + lrsize);
1262 itx->itx_lr.lrc_txtype = txtype;
1263 itx->itx_lr.lrc_reclen = lrsize;
1264 itx->itx_lr.lrc_seq = 0; /* defensive */
1265 itx->itx_sync = B_TRUE; /* default is synchronous */
1266 itx->itx_callback = NULL;
1267 itx->itx_callback_data = NULL;
1273 zil_itx_destroy(itx_t *itx)
1275 zio_data_buf_free(itx, offsetof(itx_t, itx_lr)+itx->itx_lr.lrc_reclen);
1279 * Free up the sync and async itxs. The itxs_t has already been detached
1280 * so no locks are needed.
1283 zil_itxg_clean(itxs_t *itxs)
1289 itx_async_node_t *ian;
1291 list = &itxs->i_sync_list;
1292 while ((itx = list_head(list)) != NULL) {
1293 if (itx->itx_callback != NULL)
1294 itx->itx_callback(itx->itx_callback_data);
1295 list_remove(list, itx);
1296 zil_itx_destroy(itx);
1300 t = &itxs->i_async_tree;
1301 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1302 list = &ian->ia_list;
1303 while ((itx = list_head(list)) != NULL) {
1304 if (itx->itx_callback != NULL)
1305 itx->itx_callback(itx->itx_callback_data);
1306 list_remove(list, itx);
1307 zil_itx_destroy(itx);
1310 kmem_free(ian, sizeof (itx_async_node_t));
1314 kmem_free(itxs, sizeof (itxs_t));
1318 zil_aitx_compare(const void *x1, const void *x2)
1320 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1321 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1323 return (AVL_CMP(o1, o2));
1327 * Remove all async itx with the given oid.
1330 zil_remove_async(zilog_t *zilog, uint64_t oid)
1333 itx_async_node_t *ian;
1340 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1342 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1345 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1347 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1348 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1350 mutex_enter(&itxg->itxg_lock);
1351 if (itxg->itxg_txg != txg) {
1352 mutex_exit(&itxg->itxg_lock);
1357 * Locate the object node and append its list.
1359 t = &itxg->itxg_itxs->i_async_tree;
1360 ian = avl_find(t, &oid, &where);
1362 list_move_tail(&clean_list, &ian->ia_list);
1363 mutex_exit(&itxg->itxg_lock);
1365 while ((itx = list_head(&clean_list)) != NULL) {
1366 if (itx->itx_callback != NULL)
1367 itx->itx_callback(itx->itx_callback_data);
1368 list_remove(&clean_list, itx);
1369 zil_itx_destroy(itx);
1371 list_destroy(&clean_list);
1375 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1379 itxs_t *itxs, *clean = NULL;
1382 * Object ids can be re-instantiated in the next txg so
1383 * remove any async transactions to avoid future leaks.
1384 * This can happen if a fsync occurs on the re-instantiated
1385 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1386 * the new file data and flushes a write record for the old object.
1388 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1389 zil_remove_async(zilog, itx->itx_oid);
1392 * Ensure the data of a renamed file is committed before the rename.
1394 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1395 zil_async_to_sync(zilog, itx->itx_oid);
1397 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1400 txg = dmu_tx_get_txg(tx);
1402 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1403 mutex_enter(&itxg->itxg_lock);
1404 itxs = itxg->itxg_itxs;
1405 if (itxg->itxg_txg != txg) {
1408 * The zil_clean callback hasn't got around to cleaning
1409 * this itxg. Save the itxs for release below.
1410 * This should be rare.
1412 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1413 "txg %llu", itxg->itxg_txg);
1414 clean = itxg->itxg_itxs;
1416 itxg->itxg_txg = txg;
1417 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
1420 list_create(&itxs->i_sync_list, sizeof (itx_t),
1421 offsetof(itx_t, itx_node));
1422 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1423 sizeof (itx_async_node_t),
1424 offsetof(itx_async_node_t, ia_node));
1426 if (itx->itx_sync) {
1427 list_insert_tail(&itxs->i_sync_list, itx);
1429 avl_tree_t *t = &itxs->i_async_tree;
1431 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
1432 itx_async_node_t *ian;
1435 ian = avl_find(t, &foid, &where);
1437 ian = kmem_alloc(sizeof (itx_async_node_t),
1439 list_create(&ian->ia_list, sizeof (itx_t),
1440 offsetof(itx_t, itx_node));
1441 ian->ia_foid = foid;
1442 avl_insert(t, ian, where);
1444 list_insert_tail(&ian->ia_list, itx);
1447 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1448 zilog_dirty(zilog, txg);
1449 mutex_exit(&itxg->itxg_lock);
1451 /* Release the old itxs now we've dropped the lock */
1453 zil_itxg_clean(clean);
1457 * If there are any in-memory intent log transactions which have now been
1458 * synced then start up a taskq to free them. We should only do this after we
1459 * have written out the uberblocks (i.e. txg has been comitted) so that
1460 * don't inadvertently clean out in-memory log records that would be required
1464 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1466 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1469 mutex_enter(&itxg->itxg_lock);
1470 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1471 mutex_exit(&itxg->itxg_lock);
1474 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1475 ASSERT3U(itxg->itxg_txg, !=, 0);
1476 clean_me = itxg->itxg_itxs;
1477 itxg->itxg_itxs = NULL;
1479 mutex_exit(&itxg->itxg_lock);
1481 * Preferably start a task queue to free up the old itxs but
1482 * if taskq_dispatch can't allocate resources to do that then
1483 * free it in-line. This should be rare. Note, using TQ_SLEEP
1484 * created a bad performance problem.
1486 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1487 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1488 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1489 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP);
1490 if (id == TASKQID_INVALID)
1491 zil_itxg_clean(clean_me);
1495 * Get the list of itxs to commit into zl_itx_commit_list.
1498 zil_get_commit_list(zilog_t *zilog)
1501 list_t *commit_list = &zilog->zl_itx_commit_list;
1503 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1506 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1509 * This is inherently racy, since there is nothing to prevent
1510 * the last synced txg from changing. That's okay since we'll
1511 * only commit things in the future.
1513 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1514 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1516 mutex_enter(&itxg->itxg_lock);
1517 if (itxg->itxg_txg != txg) {
1518 mutex_exit(&itxg->itxg_lock);
1523 * If we're adding itx records to the zl_itx_commit_list,
1524 * then the zil better be dirty in this "txg". We can assert
1525 * that here since we're holding the itxg_lock which will
1526 * prevent spa_sync from cleaning it. Once we add the itxs
1527 * to the zl_itx_commit_list we must commit it to disk even
1528 * if it's unnecessary (i.e. the txg was synced).
1530 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1531 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1532 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1534 mutex_exit(&itxg->itxg_lock);
1539 * Move the async itxs for a specified object to commit into sync lists.
1542 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1545 itx_async_node_t *ian;
1549 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1552 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1555 * This is inherently racy, since there is nothing to prevent
1556 * the last synced txg from changing.
1558 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1559 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1561 mutex_enter(&itxg->itxg_lock);
1562 if (itxg->itxg_txg != txg) {
1563 mutex_exit(&itxg->itxg_lock);
1568 * If a foid is specified then find that node and append its
1569 * list. Otherwise walk the tree appending all the lists
1570 * to the sync list. We add to the end rather than the
1571 * beginning to ensure the create has happened.
1573 t = &itxg->itxg_itxs->i_async_tree;
1575 ian = avl_find(t, &foid, &where);
1577 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1581 void *cookie = NULL;
1583 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1584 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1586 list_destroy(&ian->ia_list);
1587 kmem_free(ian, sizeof (itx_async_node_t));
1590 mutex_exit(&itxg->itxg_lock);
1595 zil_commit_writer(zilog_t *zilog)
1600 spa_t *spa = zilog->zl_spa;
1603 ASSERT(zilog->zl_root_zio == NULL);
1605 mutex_exit(&zilog->zl_lock);
1607 zil_get_commit_list(zilog);
1610 * Return if there's nothing to commit before we dirty the fs by
1611 * calling zil_create().
1613 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1614 mutex_enter(&zilog->zl_lock);
1618 if (zilog->zl_suspend) {
1621 lwb = list_tail(&zilog->zl_lwb_list);
1623 lwb = zil_create(zilog);
1626 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1627 for (itx = list_head(&zilog->zl_itx_commit_list); itx != NULL;
1628 itx = list_next(&zilog->zl_itx_commit_list, itx)) {
1629 txg = itx->itx_lr.lrc_txg;
1630 ASSERT3U(txg, !=, 0);
1633 * This is inherently racy and may result in us writing
1634 * out a log block for a txg that was just synced. This is
1635 * ok since we'll end cleaning up that log block the next
1636 * time we call zil_sync().
1638 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1639 lwb = zil_lwb_commit(zilog, itx, lwb);
1641 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1643 /* write the last block out */
1644 if (lwb != NULL && lwb->lwb_zio != NULL)
1645 lwb = zil_lwb_write_start(zilog, lwb);
1647 zilog->zl_cur_used = 0;
1650 * Wait if necessary for the log blocks to be on stable storage.
1652 if (zilog->zl_root_zio) {
1653 error = zio_wait(zilog->zl_root_zio);
1654 zilog->zl_root_zio = NULL;
1655 zil_flush_vdevs(zilog);
1658 if (error || lwb == NULL)
1659 txg_wait_synced(zilog->zl_dmu_pool, 0);
1661 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1662 txg = itx->itx_lr.lrc_txg;
1665 if (itx->itx_callback != NULL)
1666 itx->itx_callback(itx->itx_callback_data);
1667 list_remove(&zilog->zl_itx_commit_list, itx);
1668 zil_itx_destroy(itx);
1671 mutex_enter(&zilog->zl_lock);
1674 * Remember the highest committed log sequence number for ztest.
1675 * We only update this value when all the log writes succeeded,
1676 * because ztest wants to ASSERT that it got the whole log chain.
1678 if (error == 0 && lwb != NULL)
1679 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1683 * Commit zfs transactions to stable storage.
1684 * If foid is 0 push out all transactions, otherwise push only those
1685 * for that object or might reference that object.
1687 * itxs are committed in batches. In a heavily stressed zil there will be
1688 * a commit writer thread who is writing out a bunch of itxs to the log
1689 * for a set of committing threads (cthreads) in the same batch as the writer.
1690 * Those cthreads are all waiting on the same cv for that batch.
1692 * There will also be a different and growing batch of threads that are
1693 * waiting to commit (qthreads). When the committing batch completes
1694 * a transition occurs such that the cthreads exit and the qthreads become
1695 * cthreads. One of the new cthreads becomes the writer thread for the
1696 * batch. Any new threads arriving become new qthreads.
1698 * Only 2 condition variables are needed and there's no transition
1699 * between the two cvs needed. They just flip-flop between qthreads
1702 * Using this scheme we can efficiently wakeup up only those threads
1703 * that have been committed.
1706 zil_commit(zilog_t *zilog, uint64_t foid)
1710 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1713 ZIL_STAT_BUMP(zil_commit_count);
1715 /* move the async itxs for the foid to the sync queues */
1716 zil_async_to_sync(zilog, foid);
1718 mutex_enter(&zilog->zl_lock);
1719 mybatch = zilog->zl_next_batch;
1720 while (zilog->zl_writer) {
1721 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1722 if (mybatch <= zilog->zl_com_batch) {
1723 mutex_exit(&zilog->zl_lock);
1728 zilog->zl_next_batch++;
1729 zilog->zl_writer = B_TRUE;
1730 ZIL_STAT_BUMP(zil_commit_writer_count);
1731 zil_commit_writer(zilog);
1732 zilog->zl_com_batch = mybatch;
1733 zilog->zl_writer = B_FALSE;
1735 /* wake up one thread to become the next writer */
1736 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1738 /* wake up all threads waiting for this batch to be committed */
1739 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1741 mutex_exit(&zilog->zl_lock);
1745 * Called in syncing context to free committed log blocks and update log header.
1748 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1750 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1751 uint64_t txg = dmu_tx_get_txg(tx);
1752 spa_t *spa = zilog->zl_spa;
1753 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1757 * We don't zero out zl_destroy_txg, so make sure we don't try
1758 * to destroy it twice.
1760 if (spa_sync_pass(spa) != 1)
1763 mutex_enter(&zilog->zl_lock);
1765 ASSERT(zilog->zl_stop_sync == 0);
1767 if (*replayed_seq != 0) {
1768 ASSERT(zh->zh_replay_seq < *replayed_seq);
1769 zh->zh_replay_seq = *replayed_seq;
1773 if (zilog->zl_destroy_txg == txg) {
1774 blkptr_t blk = zh->zh_log;
1776 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1778 bzero(zh, sizeof (zil_header_t));
1779 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1781 if (zilog->zl_keep_first) {
1783 * If this block was part of log chain that couldn't
1784 * be claimed because a device was missing during
1785 * zil_claim(), but that device later returns,
1786 * then this block could erroneously appear valid.
1787 * To guard against this, assign a new GUID to the new
1788 * log chain so it doesn't matter what blk points to.
1790 zil_init_log_chain(zilog, &blk);
1795 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1796 zh->zh_log = lwb->lwb_blk;
1797 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1800 ASSERT(lwb->lwb_zio == NULL);
1802 list_remove(&zilog->zl_lwb_list, lwb);
1803 zio_free_zil(spa, txg, &lwb->lwb_blk);
1804 kmem_cache_free(zil_lwb_cache, lwb);
1807 * If we don't have anything left in the lwb list then
1808 * we've had an allocation failure and we need to zero
1809 * out the zil_header blkptr so that we don't end
1810 * up freeing the same block twice.
1812 if (list_head(&zilog->zl_lwb_list) == NULL)
1813 BP_ZERO(&zh->zh_log);
1817 * Remove fastwrite on any blocks that have been pre-allocated for
1818 * the next commit. This prevents fastwrite counter pollution by
1819 * unused, long-lived LWBs.
1821 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
1822 if (lwb->lwb_fastwrite && !lwb->lwb_zio) {
1823 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1824 lwb->lwb_fastwrite = 0;
1828 mutex_exit(&zilog->zl_lock);
1834 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1835 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1837 zil_ksp = kstat_create("zfs", 0, "zil", "misc",
1838 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
1839 KSTAT_FLAG_VIRTUAL);
1841 if (zil_ksp != NULL) {
1842 zil_ksp->ks_data = &zil_stats;
1843 kstat_install(zil_ksp);
1850 kmem_cache_destroy(zil_lwb_cache);
1852 if (zil_ksp != NULL) {
1853 kstat_delete(zil_ksp);
1859 zil_set_sync(zilog_t *zilog, uint64_t sync)
1861 zilog->zl_sync = sync;
1865 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1867 zilog->zl_logbias = logbias;
1871 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1875 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1877 zilog->zl_header = zh_phys;
1879 zilog->zl_spa = dmu_objset_spa(os);
1880 zilog->zl_dmu_pool = dmu_objset_pool(os);
1881 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1882 zilog->zl_logbias = dmu_objset_logbias(os);
1883 zilog->zl_sync = dmu_objset_syncprop(os);
1884 zilog->zl_next_batch = 1;
1886 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1888 for (int i = 0; i < TXG_SIZE; i++) {
1889 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1890 MUTEX_DEFAULT, NULL);
1893 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1894 offsetof(lwb_t, lwb_node));
1896 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1897 offsetof(itx_t, itx_node));
1899 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1901 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1902 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1904 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1905 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1906 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1907 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1913 zil_free(zilog_t *zilog)
1917 zilog->zl_stop_sync = 1;
1919 ASSERT0(zilog->zl_suspend);
1920 ASSERT0(zilog->zl_suspending);
1922 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1923 list_destroy(&zilog->zl_lwb_list);
1925 avl_destroy(&zilog->zl_vdev_tree);
1926 mutex_destroy(&zilog->zl_vdev_lock);
1928 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1929 list_destroy(&zilog->zl_itx_commit_list);
1931 for (i = 0; i < TXG_SIZE; i++) {
1933 * It's possible for an itx to be generated that doesn't dirty
1934 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1935 * callback to remove the entry. We remove those here.
1937 * Also free up the ziltest itxs.
1939 if (zilog->zl_itxg[i].itxg_itxs)
1940 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1941 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1944 mutex_destroy(&zilog->zl_lock);
1946 cv_destroy(&zilog->zl_cv_writer);
1947 cv_destroy(&zilog->zl_cv_suspend);
1948 cv_destroy(&zilog->zl_cv_batch[0]);
1949 cv_destroy(&zilog->zl_cv_batch[1]);
1951 kmem_free(zilog, sizeof (zilog_t));
1955 * Open an intent log.
1958 zil_open(objset_t *os, zil_get_data_t *get_data)
1960 zilog_t *zilog = dmu_objset_zil(os);
1962 ASSERT(zilog->zl_get_data == NULL);
1963 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1965 zilog->zl_get_data = get_data;
1971 * Close an intent log.
1974 zil_close(zilog_t *zilog)
1979 zil_commit(zilog, 0); /* commit all itx */
1982 * The lwb_max_txg for the stubby lwb will reflect the last activity
1983 * for the zil. After a txg_wait_synced() on the txg we know all the
1984 * callbacks have occurred that may clean the zil. Only then can we
1985 * destroy the zl_clean_taskq.
1987 mutex_enter(&zilog->zl_lock);
1988 lwb = list_tail(&zilog->zl_lwb_list);
1990 txg = lwb->lwb_max_txg;
1991 mutex_exit(&zilog->zl_lock);
1993 txg_wait_synced(zilog->zl_dmu_pool, txg);
1995 if (zilog_is_dirty(zilog))
1996 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1997 if (txg < spa_freeze_txg(zilog->zl_spa))
1998 VERIFY(!zilog_is_dirty(zilog));
2000 zilog->zl_get_data = NULL;
2003 * We should have only one LWB left on the list; remove it now.
2005 mutex_enter(&zilog->zl_lock);
2006 lwb = list_head(&zilog->zl_lwb_list);
2008 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
2009 ASSERT(lwb->lwb_zio == NULL);
2010 if (lwb->lwb_fastwrite)
2011 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
2012 list_remove(&zilog->zl_lwb_list, lwb);
2013 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2014 kmem_cache_free(zil_lwb_cache, lwb);
2016 mutex_exit(&zilog->zl_lock);
2019 static char *suspend_tag = "zil suspending";
2022 * Suspend an intent log. While in suspended mode, we still honor
2023 * synchronous semantics, but we rely on txg_wait_synced() to do it.
2024 * On old version pools, we suspend the log briefly when taking a
2025 * snapshot so that it will have an empty intent log.
2027 * Long holds are not really intended to be used the way we do here --
2028 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
2029 * could fail. Therefore we take pains to only put a long hold if it is
2030 * actually necessary. Fortunately, it will only be necessary if the
2031 * objset is currently mounted (or the ZVOL equivalent). In that case it
2032 * will already have a long hold, so we are not really making things any worse.
2034 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
2035 * zvol_state_t), and use their mechanism to prevent their hold from being
2036 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
2039 * if cookiep == NULL, this does both the suspend & resume.
2040 * Otherwise, it returns with the dataset "long held", and the cookie
2041 * should be passed into zil_resume().
2044 zil_suspend(const char *osname, void **cookiep)
2048 const zil_header_t *zh;
2051 error = dmu_objset_hold(osname, suspend_tag, &os);
2054 zilog = dmu_objset_zil(os);
2056 mutex_enter(&zilog->zl_lock);
2057 zh = zilog->zl_header;
2059 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
2060 mutex_exit(&zilog->zl_lock);
2061 dmu_objset_rele(os, suspend_tag);
2062 return (SET_ERROR(EBUSY));
2066 * Don't put a long hold in the cases where we can avoid it. This
2067 * is when there is no cookie so we are doing a suspend & resume
2068 * (i.e. called from zil_vdev_offline()), and there's nothing to do
2069 * for the suspend because it's already suspended, or there's no ZIL.
2071 if (cookiep == NULL && !zilog->zl_suspending &&
2072 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
2073 mutex_exit(&zilog->zl_lock);
2074 dmu_objset_rele(os, suspend_tag);
2078 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
2079 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
2081 zilog->zl_suspend++;
2083 if (zilog->zl_suspend > 1) {
2085 * Someone else is already suspending it.
2086 * Just wait for them to finish.
2089 while (zilog->zl_suspending)
2090 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
2091 mutex_exit(&zilog->zl_lock);
2093 if (cookiep == NULL)
2101 * If there is no pointer to an on-disk block, this ZIL must not
2102 * be active (e.g. filesystem not mounted), so there's nothing
2105 if (BP_IS_HOLE(&zh->zh_log)) {
2106 ASSERT(cookiep != NULL); /* fast path already handled */
2109 mutex_exit(&zilog->zl_lock);
2114 * The ZIL has work to do. Ensure that the associated encryption
2115 * key will remain mapped while we are committing the log by
2116 * grabbing a reference to it. If the key isn't loaded we have no
2117 * choice but to return an error until the wrapping key is loaded.
2119 if (os->os_encrypted && spa_keystore_create_mapping(os->os_spa,
2120 dmu_objset_ds(os), FTAG) != 0) {
2121 zilog->zl_suspend--;
2122 mutex_exit(&zilog->zl_lock);
2123 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2124 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2125 return (SET_ERROR(EBUSY));
2128 zilog->zl_suspending = B_TRUE;
2129 mutex_exit(&zilog->zl_lock);
2131 zil_commit(zilog, 0);
2133 zil_destroy(zilog, B_FALSE);
2135 mutex_enter(&zilog->zl_lock);
2136 zilog->zl_suspending = B_FALSE;
2137 cv_broadcast(&zilog->zl_cv_suspend);
2138 mutex_exit(&zilog->zl_lock);
2140 if (os->os_encrypted) {
2142 * Encrypted datasets need to wait for all data to be
2143 * synced out before removing the mapping.
2145 * XXX: Depending on the number of datasets with
2146 * outstanding ZIL data on a given log device, this
2147 * might cause spa_offline_log() to take a long time.
2149 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2150 VERIFY0(spa_keystore_remove_mapping(os->os_spa,
2151 dmu_objset_id(os), FTAG));
2154 if (cookiep == NULL)
2162 zil_resume(void *cookie)
2164 objset_t *os = cookie;
2165 zilog_t *zilog = dmu_objset_zil(os);
2167 mutex_enter(&zilog->zl_lock);
2168 ASSERT(zilog->zl_suspend != 0);
2169 zilog->zl_suspend--;
2170 mutex_exit(&zilog->zl_lock);
2171 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2172 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2175 typedef struct zil_replay_arg {
2176 zil_replay_func_t **zr_replay;
2178 boolean_t zr_byteswap;
2183 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2185 char name[ZFS_MAX_DATASET_NAME_LEN];
2187 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2189 dmu_objset_name(zilog->zl_os, name);
2191 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2192 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2193 (u_longlong_t)lr->lrc_seq,
2194 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2195 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2201 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2203 zil_replay_arg_t *zr = zra;
2204 const zil_header_t *zh = zilog->zl_header;
2205 uint64_t reclen = lr->lrc_reclen;
2206 uint64_t txtype = lr->lrc_txtype;
2209 zilog->zl_replaying_seq = lr->lrc_seq;
2211 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2214 if (lr->lrc_txg < claim_txg) /* already committed */
2217 /* Strip case-insensitive bit, still present in log record */
2220 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2221 return (zil_replay_error(zilog, lr, EINVAL));
2224 * If this record type can be logged out of order, the object
2225 * (lr_foid) may no longer exist. That's legitimate, not an error.
2227 if (TX_OOO(txtype)) {
2228 error = dmu_object_info(zilog->zl_os,
2229 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
2230 if (error == ENOENT || error == EEXIST)
2235 * Make a copy of the data so we can revise and extend it.
2237 bcopy(lr, zr->zr_lr, reclen);
2240 * If this is a TX_WRITE with a blkptr, suck in the data.
2242 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2243 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2244 zr->zr_lr + reclen);
2246 return (zil_replay_error(zilog, lr, error));
2250 * The log block containing this lr may have been byteswapped
2251 * so that we can easily examine common fields like lrc_txtype.
2252 * However, the log is a mix of different record types, and only the
2253 * replay vectors know how to byteswap their records. Therefore, if
2254 * the lr was byteswapped, undo it before invoking the replay vector.
2256 if (zr->zr_byteswap)
2257 byteswap_uint64_array(zr->zr_lr, reclen);
2260 * We must now do two things atomically: replay this log record,
2261 * and update the log header sequence number to reflect the fact that
2262 * we did so. At the end of each replay function the sequence number
2263 * is updated if we are in replay mode.
2265 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2268 * The DMU's dnode layer doesn't see removes until the txg
2269 * commits, so a subsequent claim can spuriously fail with
2270 * EEXIST. So if we receive any error we try syncing out
2271 * any removes then retry the transaction. Note that we
2272 * specify B_FALSE for byteswap now, so we don't do it twice.
2274 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2275 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2277 return (zil_replay_error(zilog, lr, error));
2284 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2286 zilog->zl_replay_blks++;
2292 * If this dataset has a non-empty intent log, replay it and destroy it.
2295 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2297 zilog_t *zilog = dmu_objset_zil(os);
2298 const zil_header_t *zh = zilog->zl_header;
2299 zil_replay_arg_t zr;
2301 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2302 zil_destroy(zilog, B_TRUE);
2306 zr.zr_replay = replay_func;
2308 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2309 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2312 * Wait for in-progress removes to sync before starting replay.
2314 txg_wait_synced(zilog->zl_dmu_pool, 0);
2316 zilog->zl_replay = B_TRUE;
2317 zilog->zl_replay_time = ddi_get_lbolt();
2318 ASSERT(zilog->zl_replay_blks == 0);
2319 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2320 zh->zh_claim_txg, B_TRUE);
2321 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2323 zil_destroy(zilog, B_FALSE);
2324 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2325 zilog->zl_replay = B_FALSE;
2329 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2331 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2334 if (zilog->zl_replay) {
2335 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2336 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2337 zilog->zl_replaying_seq;
2346 zil_vdev_offline(const char *osname, void *arg)
2350 error = zil_suspend(osname, NULL);
2352 return (SET_ERROR(EEXIST));
2356 #if defined(_KERNEL) && defined(HAVE_SPL)
2357 EXPORT_SYMBOL(zil_alloc);
2358 EXPORT_SYMBOL(zil_free);
2359 EXPORT_SYMBOL(zil_open);
2360 EXPORT_SYMBOL(zil_close);
2361 EXPORT_SYMBOL(zil_replay);
2362 EXPORT_SYMBOL(zil_replaying);
2363 EXPORT_SYMBOL(zil_destroy);
2364 EXPORT_SYMBOL(zil_destroy_sync);
2365 EXPORT_SYMBOL(zil_itx_create);
2366 EXPORT_SYMBOL(zil_itx_destroy);
2367 EXPORT_SYMBOL(zil_itx_assign);
2368 EXPORT_SYMBOL(zil_commit);
2369 EXPORT_SYMBOL(zil_vdev_offline);
2370 EXPORT_SYMBOL(zil_claim);
2371 EXPORT_SYMBOL(zil_check_log_chain);
2372 EXPORT_SYMBOL(zil_sync);
2373 EXPORT_SYMBOL(zil_clean);
2374 EXPORT_SYMBOL(zil_suspend);
2375 EXPORT_SYMBOL(zil_resume);
2376 EXPORT_SYMBOL(zil_add_block);
2377 EXPORT_SYMBOL(zil_bp_tree_add);
2378 EXPORT_SYMBOL(zil_set_sync);
2379 EXPORT_SYMBOL(zil_set_logbias);
2382 module_param(zil_replay_disable, int, 0644);
2383 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
2385 module_param(zfs_nocacheflush, int, 0644);
2386 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");
2388 module_param(zil_slog_bulk, ulong, 0644);
2389 MODULE_PARM_DESC(zil_slog_bulk, "Limit in bytes slog sync writes per commit");