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;
555 for (t = 0; t < TXG_SIZE; t++) {
556 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
563 * Create an on-disk intent log.
566 zil_create(zilog_t *zilog)
568 const zil_header_t *zh = zilog->zl_header;
574 boolean_t fastwrite = FALSE;
575 boolean_t slog = FALSE;
578 * Wait for any previous destroy to complete.
580 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
582 ASSERT(zh->zh_claim_txg == 0);
583 ASSERT(zh->zh_replay_seq == 0);
588 * Allocate an initial log block if:
589 * - there isn't one already
590 * - the existing block is the wrong endianness
592 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
593 tx = dmu_tx_create(zilog->zl_os);
594 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
595 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
596 txg = dmu_tx_get_txg(tx);
598 if (!BP_IS_HOLE(&blk)) {
599 zio_free_zil(zilog->zl_spa, txg, &blk);
603 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
604 ZIL_MIN_BLKSZ, &slog);
608 zil_init_log_chain(zilog, &blk);
612 * Allocate a log write buffer (lwb) for the first log block.
615 lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
618 * If we just allocated the first log block, commit our transaction
619 * and wait for zil_sync() to stuff the block poiner into zh_log.
620 * (zh is part of the MOS, so we cannot modify it in open context.)
624 txg_wait_synced(zilog->zl_dmu_pool, txg);
627 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
633 * In one tx, free all log blocks and clear the log header.
634 * If keep_first is set, then we're replaying a log with no content.
635 * We want to keep the first block, however, so that the first
636 * synchronous transaction doesn't require a txg_wait_synced()
637 * in zil_create(). We don't need to txg_wait_synced() here either
638 * when keep_first is set, because both zil_create() and zil_destroy()
639 * will wait for any in-progress destroys to complete.
642 zil_destroy(zilog_t *zilog, boolean_t keep_first)
644 const zil_header_t *zh = zilog->zl_header;
650 * Wait for any previous destroy to complete.
652 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
654 zilog->zl_old_header = *zh; /* debugging aid */
656 if (BP_IS_HOLE(&zh->zh_log))
659 tx = dmu_tx_create(zilog->zl_os);
660 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
661 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
662 txg = dmu_tx_get_txg(tx);
664 mutex_enter(&zilog->zl_lock);
666 ASSERT3U(zilog->zl_destroy_txg, <, txg);
667 zilog->zl_destroy_txg = txg;
668 zilog->zl_keep_first = keep_first;
670 if (!list_is_empty(&zilog->zl_lwb_list)) {
671 ASSERT(zh->zh_claim_txg == 0);
673 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
674 ASSERT(lwb->lwb_zio == NULL);
675 if (lwb->lwb_fastwrite)
676 metaslab_fastwrite_unmark(zilog->zl_spa,
678 list_remove(&zilog->zl_lwb_list, lwb);
679 if (lwb->lwb_buf != NULL)
680 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
681 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
682 kmem_cache_free(zil_lwb_cache, lwb);
684 } else if (!keep_first) {
685 zil_destroy_sync(zilog, tx);
687 mutex_exit(&zilog->zl_lock);
693 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
695 ASSERT(list_is_empty(&zilog->zl_lwb_list));
696 (void) zil_parse(zilog, zil_free_log_block,
697 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
701 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
703 dmu_tx_t *tx = txarg;
704 uint64_t first_txg = dmu_tx_get_txg(tx);
710 error = dmu_objset_own_obj(dp, ds->ds_object,
711 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
714 * EBUSY indicates that the objset is inconsistent, in which
715 * case it can not have a ZIL.
717 if (error != EBUSY) {
718 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
719 (unsigned long long)ds->ds_object, error);
725 zilog = dmu_objset_zil(os);
726 zh = zil_header_in_syncing_context(zilog);
728 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
729 if (!BP_IS_HOLE(&zh->zh_log))
730 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
731 BP_ZERO(&zh->zh_log);
732 if (os->os_encrypted)
733 os->os_next_write_raw = B_TRUE;
734 dsl_dataset_dirty(dmu_objset_ds(os), tx);
735 dmu_objset_disown(os, B_FALSE, FTAG);
740 * Claim all log blocks if we haven't already done so, and remember
741 * the highest claimed sequence number. This ensures that if we can
742 * read only part of the log now (e.g. due to a missing device),
743 * but we can read the entire log later, we will not try to replay
744 * or destroy beyond the last block we successfully claimed.
746 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
747 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
748 (void) zil_parse(zilog, zil_claim_log_block,
749 zil_claim_log_record, tx, first_txg, B_FALSE);
750 zh->zh_claim_txg = first_txg;
751 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
752 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
753 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
754 zh->zh_flags |= ZIL_REPLAY_NEEDED;
755 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
756 dsl_dataset_dirty(dmu_objset_ds(os), tx);
759 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
760 dmu_objset_disown(os, B_FALSE, FTAG);
765 * Check the log by walking the log chain.
766 * Checksum errors are ok as they indicate the end of the chain.
767 * Any other error (no device or read failure) returns an error.
771 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
780 error = dmu_objset_from_ds(ds, &os);
782 cmn_err(CE_WARN, "can't open objset %llu, error %d",
783 (unsigned long long)ds->ds_object, error);
787 zilog = dmu_objset_zil(os);
788 bp = (blkptr_t *)&zilog->zl_header->zh_log;
791 * Check the first block and determine if it's on a log device
792 * which may have been removed or faulted prior to loading this
793 * pool. If so, there's no point in checking the rest of the log
794 * as its content should have already been synced to the pool.
796 if (!BP_IS_HOLE(bp)) {
798 boolean_t valid = B_TRUE;
800 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
801 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
802 if (vd->vdev_islog && vdev_is_dead(vd))
803 valid = vdev_log_state_valid(vd);
804 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
811 * Because tx == NULL, zil_claim_log_block() will not actually claim
812 * any blocks, but just determine whether it is possible to do so.
813 * In addition to checking the log chain, zil_claim_log_block()
814 * will invoke zio_claim() with a done func of spa_claim_notify(),
815 * which will update spa_max_claim_txg. See spa_load() for details.
817 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
818 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa),
821 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
825 zil_vdev_compare(const void *x1, const void *x2)
827 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
828 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
830 return (AVL_CMP(v1, v2));
834 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
836 avl_tree_t *t = &zilog->zl_vdev_tree;
838 zil_vdev_node_t *zv, zvsearch;
839 int ndvas = BP_GET_NDVAS(bp);
842 if (zfs_nocacheflush)
845 ASSERT(zilog->zl_writer);
848 * Even though we're zl_writer, we still need a lock because the
849 * zl_get_data() callbacks may have dmu_sync() done callbacks
850 * that will run concurrently.
852 mutex_enter(&zilog->zl_vdev_lock);
853 for (i = 0; i < ndvas; i++) {
854 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
855 if (avl_find(t, &zvsearch, &where) == NULL) {
856 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
857 zv->zv_vdev = zvsearch.zv_vdev;
858 avl_insert(t, zv, where);
861 mutex_exit(&zilog->zl_vdev_lock);
865 zil_flush_vdevs(zilog_t *zilog)
867 spa_t *spa = zilog->zl_spa;
868 avl_tree_t *t = &zilog->zl_vdev_tree;
873 ASSERT(zilog->zl_writer);
876 * We don't need zl_vdev_lock here because we're the zl_writer,
877 * and all zl_get_data() callbacks are done.
879 if (avl_numnodes(t) == 0)
882 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
884 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
886 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
887 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
890 kmem_free(zv, sizeof (*zv));
894 * Wait for all the flushes to complete. Not all devices actually
895 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
897 (void) zio_wait(zio);
899 spa_config_exit(spa, SCL_STATE, FTAG);
903 * Function called when a log block write completes
906 zil_lwb_write_done(zio_t *zio)
908 lwb_t *lwb = zio->io_private;
909 zilog_t *zilog = lwb->lwb_zilog;
910 dmu_tx_t *tx = lwb->lwb_tx;
912 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
913 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
914 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
915 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
916 ASSERT(!BP_IS_GANG(zio->io_bp));
917 ASSERT(!BP_IS_HOLE(zio->io_bp));
918 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
921 * Ensure the lwb buffer pointer is cleared before releasing
922 * the txg. If we have had an allocation failure and
923 * the txg is waiting to sync then we want want zil_sync()
924 * to remove the lwb so that it's not picked up as the next new
925 * one in zil_commit_writer(). zil_sync() will only remove
926 * the lwb if lwb_buf is null.
928 abd_put(zio->io_abd);
929 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
930 mutex_enter(&zilog->zl_lock);
932 lwb->lwb_fastwrite = FALSE;
935 mutex_exit(&zilog->zl_lock);
938 * Now that we've written this log block, we have a stable pointer
939 * to the next block in the chain, so it's OK to let the txg in
940 * which we allocated the next block sync.
946 * Initialize the io for a log block.
949 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
954 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
955 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
956 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
958 if (zilog->zl_root_zio == NULL) {
959 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
963 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
964 mutex_enter(&zilog->zl_lock);
965 if (lwb->lwb_zio == NULL) {
966 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
967 BP_GET_LSIZE(&lwb->lwb_blk));
968 if (!lwb->lwb_fastwrite) {
969 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
970 lwb->lwb_fastwrite = 1;
972 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
973 prio = ZIO_PRIORITY_SYNC_WRITE;
975 prio = ZIO_PRIORITY_ASYNC_WRITE;
976 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
977 0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
978 zil_lwb_write_done, lwb, prio,
979 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
980 ZIO_FLAG_FASTWRITE, &zb);
982 mutex_exit(&zilog->zl_lock);
986 * Define a limited set of intent log block sizes.
988 * These must be a multiple of 4KB. Note only the amount used (again
989 * aligned to 4KB) actually gets written. However, we can't always just
990 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
992 uint64_t zil_block_buckets[] = {
993 4096, /* non TX_WRITE */
994 8192+4096, /* data base */
995 32*1024 + 4096, /* NFS writes */
1000 * Start a log block write and advance to the next log block.
1001 * Calls are serialized.
1004 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
1008 spa_t *spa = zilog->zl_spa;
1012 uint64_t zil_blksz, wsz;
1016 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1017 zilc = (zil_chain_t *)lwb->lwb_buf;
1018 bp = &zilc->zc_next_blk;
1020 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1021 bp = &zilc->zc_next_blk;
1024 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
1027 * Allocate the next block and save its address in this block
1028 * before writing it in order to establish the log chain.
1029 * Note that if the allocation of nlwb synced before we wrote
1030 * the block that points at it (lwb), we'd leak it if we crashed.
1031 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1032 * We dirty the dataset to ensure that zil_sync() will be called
1033 * to clean up in the event of allocation failure or I/O failure.
1035 tx = dmu_tx_create(zilog->zl_os);
1038 * Since we are not going to create any new dirty data and we can even
1039 * help with clearing the existing dirty data, we should not be subject
1040 * to the dirty data based delays.
1041 * We (ab)use TXG_WAITED to bypass the delay mechanism.
1042 * One side effect from using TXG_WAITED is that dmu_tx_assign() can
1043 * fail if the pool is suspended. Those are dramatic circumstances,
1044 * so we return NULL to signal that the normal ZIL processing is not
1045 * possible and txg_wait_synced() should be used to ensure that the data
1048 error = dmu_tx_assign(tx, TXG_WAITED);
1050 ASSERT3S(error, ==, EIO);
1054 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1055 txg = dmu_tx_get_txg(tx);
1060 * Log blocks are pre-allocated. Here we select the size of the next
1061 * block, based on size used in the last block.
1062 * - first find the smallest bucket that will fit the block from a
1063 * limited set of block sizes. This is because it's faster to write
1064 * blocks allocated from the same metaslab as they are adjacent or
1066 * - next find the maximum from the new suggested size and an array of
1067 * previous sizes. This lessens a picket fence effect of wrongly
1068 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1071 * Note we only write what is used, but we can't just allocate
1072 * the maximum block size because we can exhaust the available
1075 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1076 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1078 zil_blksz = zil_block_buckets[i];
1079 if (zil_blksz == UINT64_MAX)
1080 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1081 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1082 for (i = 0; i < ZIL_PREV_BLKS; i++)
1083 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1084 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1087 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1089 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
1090 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
1092 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
1093 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
1096 ASSERT3U(bp->blk_birth, ==, txg);
1097 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1098 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1101 * Allocate a new log write buffer (lwb).
1103 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1105 /* Record the block for later vdev flushing */
1106 zil_add_block(zilog, &lwb->lwb_blk);
1109 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1110 /* For Slim ZIL only write what is used. */
1111 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1112 ASSERT3U(wsz, <=, lwb->lwb_sz);
1113 zio_shrink(lwb->lwb_zio, wsz);
1120 zilc->zc_nused = lwb->lwb_nused;
1121 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1124 * clear unused data for security
1126 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1128 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1131 * If there was an allocation failure then nlwb will be null which
1132 * forces a txg_wait_synced().
1138 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1141 lr_write_t *lrwb, *lrw;
1143 uint64_t dlen, dnow, lwb_sp, reclen, txg;
1148 ASSERT(lwb->lwb_buf != NULL);
1150 lrc = &itx->itx_lr; /* Common log record inside itx. */
1151 lrw = (lr_write_t *)lrc; /* Write log record inside itx. */
1152 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
1153 dlen = P2ROUNDUP_TYPED(
1154 lrw->lr_length, sizeof (uint64_t), uint64_t);
1158 reclen = lrc->lrc_reclen;
1159 zilog->zl_cur_used += (reclen + dlen);
1162 zil_lwb_write_init(zilog, lwb);
1166 * If this record won't fit in the current log block, start a new one.
1167 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1169 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1170 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1171 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1172 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1173 lwb = zil_lwb_write_start(zilog, lwb);
1176 zil_lwb_write_init(zilog, lwb);
1177 ASSERT(LWB_EMPTY(lwb));
1178 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1179 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
1182 dnow = MIN(dlen, lwb_sp - reclen);
1183 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1184 bcopy(lrc, lr_buf, reclen);
1185 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */
1186 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */
1188 ZIL_STAT_BUMP(zil_itx_count);
1191 * If it's a write, fetch the data or get its blkptr as appropriate.
1193 if (lrc->lrc_txtype == TX_WRITE) {
1194 if (txg > spa_freeze_txg(zilog->zl_spa))
1195 txg_wait_synced(zilog->zl_dmu_pool, txg);
1196 if (itx->itx_wr_state == WR_COPIED) {
1197 ZIL_STAT_BUMP(zil_itx_copied_count);
1198 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1203 if (itx->itx_wr_state == WR_NEED_COPY) {
1204 dbuf = lr_buf + reclen;
1205 lrcb->lrc_reclen += dnow;
1206 if (lrwb->lr_length > dnow)
1207 lrwb->lr_length = dnow;
1208 lrw->lr_offset += dnow;
1209 lrw->lr_length -= dnow;
1210 ZIL_STAT_BUMP(zil_itx_needcopy_count);
1211 ZIL_STAT_INCR(zil_itx_needcopy_bytes,
1214 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1216 ZIL_STAT_BUMP(zil_itx_indirect_count);
1217 ZIL_STAT_INCR(zil_itx_indirect_bytes,
1220 error = zilog->zl_get_data(
1221 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1223 txg_wait_synced(zilog->zl_dmu_pool, txg);
1227 ASSERT(error == ENOENT || error == EEXIST ||
1235 * We're actually making an entry, so update lrc_seq to be the
1236 * log record sequence number. Note that this is generally not
1237 * equal to the itx sequence number because not all transactions
1238 * are synchronous, and sometimes spa_sync() gets there first.
1240 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1241 lwb->lwb_nused += reclen + dnow;
1242 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1243 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1244 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1248 zilog->zl_cur_used += reclen;
1256 zil_itx_create(uint64_t txtype, size_t lrsize)
1260 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1262 itx = zio_data_buf_alloc(offsetof(itx_t, itx_lr) + lrsize);
1263 itx->itx_lr.lrc_txtype = txtype;
1264 itx->itx_lr.lrc_reclen = lrsize;
1265 itx->itx_lr.lrc_seq = 0; /* defensive */
1266 itx->itx_sync = B_TRUE; /* default is synchronous */
1267 itx->itx_callback = NULL;
1268 itx->itx_callback_data = NULL;
1274 zil_itx_destroy(itx_t *itx)
1276 zio_data_buf_free(itx, offsetof(itx_t, itx_lr)+itx->itx_lr.lrc_reclen);
1280 * Free up the sync and async itxs. The itxs_t has already been detached
1281 * so no locks are needed.
1284 zil_itxg_clean(itxs_t *itxs)
1290 itx_async_node_t *ian;
1292 list = &itxs->i_sync_list;
1293 while ((itx = list_head(list)) != NULL) {
1294 if (itx->itx_callback != NULL)
1295 itx->itx_callback(itx->itx_callback_data);
1296 list_remove(list, itx);
1297 zil_itx_destroy(itx);
1301 t = &itxs->i_async_tree;
1302 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1303 list = &ian->ia_list;
1304 while ((itx = list_head(list)) != NULL) {
1305 if (itx->itx_callback != NULL)
1306 itx->itx_callback(itx->itx_callback_data);
1307 list_remove(list, itx);
1308 zil_itx_destroy(itx);
1311 kmem_free(ian, sizeof (itx_async_node_t));
1315 kmem_free(itxs, sizeof (itxs_t));
1319 zil_aitx_compare(const void *x1, const void *x2)
1321 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1322 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1324 return (AVL_CMP(o1, o2));
1328 * Remove all async itx with the given oid.
1331 zil_remove_async(zilog_t *zilog, uint64_t oid)
1334 itx_async_node_t *ian;
1341 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1343 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1346 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1348 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1349 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1351 mutex_enter(&itxg->itxg_lock);
1352 if (itxg->itxg_txg != txg) {
1353 mutex_exit(&itxg->itxg_lock);
1358 * Locate the object node and append its list.
1360 t = &itxg->itxg_itxs->i_async_tree;
1361 ian = avl_find(t, &oid, &where);
1363 list_move_tail(&clean_list, &ian->ia_list);
1364 mutex_exit(&itxg->itxg_lock);
1366 while ((itx = list_head(&clean_list)) != NULL) {
1367 if (itx->itx_callback != NULL)
1368 itx->itx_callback(itx->itx_callback_data);
1369 list_remove(&clean_list, itx);
1370 zil_itx_destroy(itx);
1372 list_destroy(&clean_list);
1376 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1380 itxs_t *itxs, *clean = NULL;
1383 * Object ids can be re-instantiated in the next txg so
1384 * remove any async transactions to avoid future leaks.
1385 * This can happen if a fsync occurs on the re-instantiated
1386 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1387 * the new file data and flushes a write record for the old object.
1389 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1390 zil_remove_async(zilog, itx->itx_oid);
1393 * Ensure the data of a renamed file is committed before the rename.
1395 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1396 zil_async_to_sync(zilog, itx->itx_oid);
1398 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1401 txg = dmu_tx_get_txg(tx);
1403 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1404 mutex_enter(&itxg->itxg_lock);
1405 itxs = itxg->itxg_itxs;
1406 if (itxg->itxg_txg != txg) {
1409 * The zil_clean callback hasn't got around to cleaning
1410 * this itxg. Save the itxs for release below.
1411 * This should be rare.
1413 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1414 "txg %llu", itxg->itxg_txg);
1415 clean = itxg->itxg_itxs;
1417 itxg->itxg_txg = txg;
1418 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
1421 list_create(&itxs->i_sync_list, sizeof (itx_t),
1422 offsetof(itx_t, itx_node));
1423 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1424 sizeof (itx_async_node_t),
1425 offsetof(itx_async_node_t, ia_node));
1427 if (itx->itx_sync) {
1428 list_insert_tail(&itxs->i_sync_list, itx);
1430 avl_tree_t *t = &itxs->i_async_tree;
1432 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
1433 itx_async_node_t *ian;
1436 ian = avl_find(t, &foid, &where);
1438 ian = kmem_alloc(sizeof (itx_async_node_t),
1440 list_create(&ian->ia_list, sizeof (itx_t),
1441 offsetof(itx_t, itx_node));
1442 ian->ia_foid = foid;
1443 avl_insert(t, ian, where);
1445 list_insert_tail(&ian->ia_list, itx);
1448 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1449 zilog_dirty(zilog, txg);
1450 mutex_exit(&itxg->itxg_lock);
1452 /* Release the old itxs now we've dropped the lock */
1454 zil_itxg_clean(clean);
1458 * If there are any in-memory intent log transactions which have now been
1459 * synced then start up a taskq to free them. We should only do this after we
1460 * have written out the uberblocks (i.e. txg has been comitted) so that
1461 * don't inadvertently clean out in-memory log records that would be required
1465 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1467 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1470 mutex_enter(&itxg->itxg_lock);
1471 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1472 mutex_exit(&itxg->itxg_lock);
1475 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1476 ASSERT3U(itxg->itxg_txg, !=, 0);
1477 clean_me = itxg->itxg_itxs;
1478 itxg->itxg_itxs = NULL;
1480 mutex_exit(&itxg->itxg_lock);
1482 * Preferably start a task queue to free up the old itxs but
1483 * if taskq_dispatch can't allocate resources to do that then
1484 * free it in-line. This should be rare. Note, using TQ_SLEEP
1485 * created a bad performance problem.
1487 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
1488 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
1489 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
1490 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP);
1491 if (id == TASKQID_INVALID)
1492 zil_itxg_clean(clean_me);
1496 * Get the list of itxs to commit into zl_itx_commit_list.
1499 zil_get_commit_list(zilog_t *zilog)
1502 list_t *commit_list = &zilog->zl_itx_commit_list;
1504 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1507 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1510 * This is inherently racy, since there is nothing to prevent
1511 * the last synced txg from changing. That's okay since we'll
1512 * only commit things in the future.
1514 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1515 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1517 mutex_enter(&itxg->itxg_lock);
1518 if (itxg->itxg_txg != txg) {
1519 mutex_exit(&itxg->itxg_lock);
1524 * If we're adding itx records to the zl_itx_commit_list,
1525 * then the zil better be dirty in this "txg". We can assert
1526 * that here since we're holding the itxg_lock which will
1527 * prevent spa_sync from cleaning it. Once we add the itxs
1528 * to the zl_itx_commit_list we must commit it to disk even
1529 * if it's unnecessary (i.e. the txg was synced).
1531 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1532 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1533 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1535 mutex_exit(&itxg->itxg_lock);
1540 * Move the async itxs for a specified object to commit into sync lists.
1543 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1546 itx_async_node_t *ian;
1550 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1553 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1556 * This is inherently racy, since there is nothing to prevent
1557 * the last synced txg from changing.
1559 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1560 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1562 mutex_enter(&itxg->itxg_lock);
1563 if (itxg->itxg_txg != txg) {
1564 mutex_exit(&itxg->itxg_lock);
1569 * If a foid is specified then find that node and append its
1570 * list. Otherwise walk the tree appending all the lists
1571 * to the sync list. We add to the end rather than the
1572 * beginning to ensure the create has happened.
1574 t = &itxg->itxg_itxs->i_async_tree;
1576 ian = avl_find(t, &foid, &where);
1578 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1582 void *cookie = NULL;
1584 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1585 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1587 list_destroy(&ian->ia_list);
1588 kmem_free(ian, sizeof (itx_async_node_t));
1591 mutex_exit(&itxg->itxg_lock);
1596 zil_commit_writer(zilog_t *zilog)
1601 spa_t *spa = zilog->zl_spa;
1604 ASSERT(zilog->zl_root_zio == NULL);
1606 mutex_exit(&zilog->zl_lock);
1608 zil_get_commit_list(zilog);
1611 * Return if there's nothing to commit before we dirty the fs by
1612 * calling zil_create().
1614 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1615 mutex_enter(&zilog->zl_lock);
1619 if (zilog->zl_suspend) {
1622 lwb = list_tail(&zilog->zl_lwb_list);
1624 lwb = zil_create(zilog);
1627 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1628 for (itx = list_head(&zilog->zl_itx_commit_list); itx != NULL;
1629 itx = list_next(&zilog->zl_itx_commit_list, itx)) {
1630 txg = itx->itx_lr.lrc_txg;
1631 ASSERT3U(txg, !=, 0);
1634 * This is inherently racy and may result in us writing
1635 * out a log block for a txg that was just synced. This is
1636 * ok since we'll end cleaning up that log block the next
1637 * time we call zil_sync().
1639 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1640 lwb = zil_lwb_commit(zilog, itx, lwb);
1642 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1644 /* write the last block out */
1645 if (lwb != NULL && lwb->lwb_zio != NULL)
1646 lwb = zil_lwb_write_start(zilog, lwb);
1648 zilog->zl_cur_used = 0;
1651 * Wait if necessary for the log blocks to be on stable storage.
1653 if (zilog->zl_root_zio) {
1654 error = zio_wait(zilog->zl_root_zio);
1655 zilog->zl_root_zio = NULL;
1656 zil_flush_vdevs(zilog);
1659 if (error || lwb == NULL)
1660 txg_wait_synced(zilog->zl_dmu_pool, 0);
1662 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1663 txg = itx->itx_lr.lrc_txg;
1666 if (itx->itx_callback != NULL)
1667 itx->itx_callback(itx->itx_callback_data);
1668 list_remove(&zilog->zl_itx_commit_list, itx);
1669 zil_itx_destroy(itx);
1672 mutex_enter(&zilog->zl_lock);
1675 * Remember the highest committed log sequence number for ztest.
1676 * We only update this value when all the log writes succeeded,
1677 * because ztest wants to ASSERT that it got the whole log chain.
1679 if (error == 0 && lwb != NULL)
1680 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1684 * Commit zfs transactions to stable storage.
1685 * If foid is 0 push out all transactions, otherwise push only those
1686 * for that object or might reference that object.
1688 * itxs are committed in batches. In a heavily stressed zil there will be
1689 * a commit writer thread who is writing out a bunch of itxs to the log
1690 * for a set of committing threads (cthreads) in the same batch as the writer.
1691 * Those cthreads are all waiting on the same cv for that batch.
1693 * There will also be a different and growing batch of threads that are
1694 * waiting to commit (qthreads). When the committing batch completes
1695 * a transition occurs such that the cthreads exit and the qthreads become
1696 * cthreads. One of the new cthreads becomes the writer thread for the
1697 * batch. Any new threads arriving become new qthreads.
1699 * Only 2 condition variables are needed and there's no transition
1700 * between the two cvs needed. They just flip-flop between qthreads
1703 * Using this scheme we can efficiently wakeup up only those threads
1704 * that have been committed.
1707 zil_commit(zilog_t *zilog, uint64_t foid)
1711 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1714 ZIL_STAT_BUMP(zil_commit_count);
1716 /* move the async itxs for the foid to the sync queues */
1717 zil_async_to_sync(zilog, foid);
1719 mutex_enter(&zilog->zl_lock);
1720 mybatch = zilog->zl_next_batch;
1721 while (zilog->zl_writer) {
1722 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1723 if (mybatch <= zilog->zl_com_batch) {
1724 mutex_exit(&zilog->zl_lock);
1729 zilog->zl_next_batch++;
1730 zilog->zl_writer = B_TRUE;
1731 ZIL_STAT_BUMP(zil_commit_writer_count);
1732 zil_commit_writer(zilog);
1733 zilog->zl_com_batch = mybatch;
1734 zilog->zl_writer = B_FALSE;
1736 /* wake up one thread to become the next writer */
1737 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1739 /* wake up all threads waiting for this batch to be committed */
1740 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1742 mutex_exit(&zilog->zl_lock);
1746 * Called in syncing context to free committed log blocks and update log header.
1749 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1751 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1752 uint64_t txg = dmu_tx_get_txg(tx);
1753 spa_t *spa = zilog->zl_spa;
1754 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1758 * We don't zero out zl_destroy_txg, so make sure we don't try
1759 * to destroy it twice.
1761 if (spa_sync_pass(spa) != 1)
1764 mutex_enter(&zilog->zl_lock);
1766 ASSERT(zilog->zl_stop_sync == 0);
1768 if (*replayed_seq != 0) {
1769 ASSERT(zh->zh_replay_seq < *replayed_seq);
1770 zh->zh_replay_seq = *replayed_seq;
1774 if (zilog->zl_destroy_txg == txg) {
1775 blkptr_t blk = zh->zh_log;
1777 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1779 bzero(zh, sizeof (zil_header_t));
1780 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1782 if (zilog->zl_keep_first) {
1784 * If this block was part of log chain that couldn't
1785 * be claimed because a device was missing during
1786 * zil_claim(), but that device later returns,
1787 * then this block could erroneously appear valid.
1788 * To guard against this, assign a new GUID to the new
1789 * log chain so it doesn't matter what blk points to.
1791 zil_init_log_chain(zilog, &blk);
1796 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1797 zh->zh_log = lwb->lwb_blk;
1798 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1801 ASSERT(lwb->lwb_zio == NULL);
1803 list_remove(&zilog->zl_lwb_list, lwb);
1804 zio_free_zil(spa, txg, &lwb->lwb_blk);
1805 kmem_cache_free(zil_lwb_cache, lwb);
1808 * If we don't have anything left in the lwb list then
1809 * we've had an allocation failure and we need to zero
1810 * out the zil_header blkptr so that we don't end
1811 * up freeing the same block twice.
1813 if (list_head(&zilog->zl_lwb_list) == NULL)
1814 BP_ZERO(&zh->zh_log);
1818 * Remove fastwrite on any blocks that have been pre-allocated for
1819 * the next commit. This prevents fastwrite counter pollution by
1820 * unused, long-lived LWBs.
1822 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
1823 if (lwb->lwb_fastwrite && !lwb->lwb_zio) {
1824 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1825 lwb->lwb_fastwrite = 0;
1829 mutex_exit(&zilog->zl_lock);
1835 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1836 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1838 zil_ksp = kstat_create("zfs", 0, "zil", "misc",
1839 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
1840 KSTAT_FLAG_VIRTUAL);
1842 if (zil_ksp != NULL) {
1843 zil_ksp->ks_data = &zil_stats;
1844 kstat_install(zil_ksp);
1851 kmem_cache_destroy(zil_lwb_cache);
1853 if (zil_ksp != NULL) {
1854 kstat_delete(zil_ksp);
1860 zil_set_sync(zilog_t *zilog, uint64_t sync)
1862 zilog->zl_sync = sync;
1866 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1868 zilog->zl_logbias = logbias;
1872 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1877 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1879 zilog->zl_header = zh_phys;
1881 zilog->zl_spa = dmu_objset_spa(os);
1882 zilog->zl_dmu_pool = dmu_objset_pool(os);
1883 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1884 zilog->zl_logbias = dmu_objset_logbias(os);
1885 zilog->zl_sync = dmu_objset_syncprop(os);
1886 zilog->zl_next_batch = 1;
1888 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1890 for (i = 0; i < TXG_SIZE; i++) {
1891 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1892 MUTEX_DEFAULT, NULL);
1895 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1896 offsetof(lwb_t, lwb_node));
1898 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1899 offsetof(itx_t, itx_node));
1901 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1903 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1904 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1906 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1907 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1908 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1909 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1915 zil_free(zilog_t *zilog)
1919 zilog->zl_stop_sync = 1;
1921 ASSERT0(zilog->zl_suspend);
1922 ASSERT0(zilog->zl_suspending);
1924 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1925 list_destroy(&zilog->zl_lwb_list);
1927 avl_destroy(&zilog->zl_vdev_tree);
1928 mutex_destroy(&zilog->zl_vdev_lock);
1930 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1931 list_destroy(&zilog->zl_itx_commit_list);
1933 for (i = 0; i < TXG_SIZE; i++) {
1935 * It's possible for an itx to be generated that doesn't dirty
1936 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1937 * callback to remove the entry. We remove those here.
1939 * Also free up the ziltest itxs.
1941 if (zilog->zl_itxg[i].itxg_itxs)
1942 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1943 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1946 mutex_destroy(&zilog->zl_lock);
1948 cv_destroy(&zilog->zl_cv_writer);
1949 cv_destroy(&zilog->zl_cv_suspend);
1950 cv_destroy(&zilog->zl_cv_batch[0]);
1951 cv_destroy(&zilog->zl_cv_batch[1]);
1953 kmem_free(zilog, sizeof (zilog_t));
1957 * Open an intent log.
1960 zil_open(objset_t *os, zil_get_data_t *get_data)
1962 zilog_t *zilog = dmu_objset_zil(os);
1964 ASSERT(zilog->zl_get_data == NULL);
1965 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1967 zilog->zl_get_data = get_data;
1973 * Close an intent log.
1976 zil_close(zilog_t *zilog)
1981 zil_commit(zilog, 0); /* commit all itx */
1984 * The lwb_max_txg for the stubby lwb will reflect the last activity
1985 * for the zil. After a txg_wait_synced() on the txg we know all the
1986 * callbacks have occurred that may clean the zil. Only then can we
1987 * destroy the zl_clean_taskq.
1989 mutex_enter(&zilog->zl_lock);
1990 lwb = list_tail(&zilog->zl_lwb_list);
1992 txg = lwb->lwb_max_txg;
1993 mutex_exit(&zilog->zl_lock);
1995 txg_wait_synced(zilog->zl_dmu_pool, txg);
1997 if (zilog_is_dirty(zilog))
1998 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1999 if (txg < spa_freeze_txg(zilog->zl_spa))
2000 VERIFY(!zilog_is_dirty(zilog));
2002 zilog->zl_get_data = NULL;
2005 * We should have only one LWB left on the list; remove it now.
2007 mutex_enter(&zilog->zl_lock);
2008 lwb = list_head(&zilog->zl_lwb_list);
2010 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
2011 ASSERT(lwb->lwb_zio == NULL);
2012 if (lwb->lwb_fastwrite)
2013 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
2014 list_remove(&zilog->zl_lwb_list, lwb);
2015 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
2016 kmem_cache_free(zil_lwb_cache, lwb);
2018 mutex_exit(&zilog->zl_lock);
2021 static char *suspend_tag = "zil suspending";
2024 * Suspend an intent log. While in suspended mode, we still honor
2025 * synchronous semantics, but we rely on txg_wait_synced() to do it.
2026 * On old version pools, we suspend the log briefly when taking a
2027 * snapshot so that it will have an empty intent log.
2029 * Long holds are not really intended to be used the way we do here --
2030 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
2031 * could fail. Therefore we take pains to only put a long hold if it is
2032 * actually necessary. Fortunately, it will only be necessary if the
2033 * objset is currently mounted (or the ZVOL equivalent). In that case it
2034 * will already have a long hold, so we are not really making things any worse.
2036 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
2037 * zvol_state_t), and use their mechanism to prevent their hold from being
2038 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
2041 * if cookiep == NULL, this does both the suspend & resume.
2042 * Otherwise, it returns with the dataset "long held", and the cookie
2043 * should be passed into zil_resume().
2046 zil_suspend(const char *osname, void **cookiep)
2050 const zil_header_t *zh;
2053 error = dmu_objset_hold(osname, suspend_tag, &os);
2056 zilog = dmu_objset_zil(os);
2058 mutex_enter(&zilog->zl_lock);
2059 zh = zilog->zl_header;
2061 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
2062 mutex_exit(&zilog->zl_lock);
2063 dmu_objset_rele(os, suspend_tag);
2064 return (SET_ERROR(EBUSY));
2068 * Don't put a long hold in the cases where we can avoid it. This
2069 * is when there is no cookie so we are doing a suspend & resume
2070 * (i.e. called from zil_vdev_offline()), and there's nothing to do
2071 * for the suspend because it's already suspended, or there's no ZIL.
2073 if (cookiep == NULL && !zilog->zl_suspending &&
2074 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
2075 mutex_exit(&zilog->zl_lock);
2076 dmu_objset_rele(os, suspend_tag);
2080 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
2081 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
2083 zilog->zl_suspend++;
2085 if (zilog->zl_suspend > 1) {
2087 * Someone else is already suspending it.
2088 * Just wait for them to finish.
2091 while (zilog->zl_suspending)
2092 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
2093 mutex_exit(&zilog->zl_lock);
2095 if (cookiep == NULL)
2103 * If there is no pointer to an on-disk block, this ZIL must not
2104 * be active (e.g. filesystem not mounted), so there's nothing
2107 if (BP_IS_HOLE(&zh->zh_log)) {
2108 ASSERT(cookiep != NULL); /* fast path already handled */
2111 mutex_exit(&zilog->zl_lock);
2116 * The ZIL has work to do. Ensure that the associated encryption
2117 * key will remain mapped while we are committing the log by
2118 * grabbing a reference to it. If the key isn't loaded we have no
2119 * choice but to return an error until the wrapping key is loaded.
2121 if (os->os_encrypted && spa_keystore_create_mapping(os->os_spa,
2122 dmu_objset_ds(os), FTAG) != 0) {
2123 zilog->zl_suspend--;
2124 mutex_exit(&zilog->zl_lock);
2125 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2126 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2127 return (SET_ERROR(EBUSY));
2130 zilog->zl_suspending = B_TRUE;
2131 mutex_exit(&zilog->zl_lock);
2133 zil_commit(zilog, 0);
2135 zil_destroy(zilog, B_FALSE);
2137 mutex_enter(&zilog->zl_lock);
2138 zilog->zl_suspending = B_FALSE;
2139 cv_broadcast(&zilog->zl_cv_suspend);
2140 mutex_exit(&zilog->zl_lock);
2142 if (os->os_encrypted) {
2144 * Encrypted datasets need to wait for all data to be
2145 * synced out before removing the mapping.
2147 * XXX: Depending on the number of datasets with
2148 * outstanding ZIL data on a given log device, this
2149 * might cause spa_offline_log() to take a long time.
2151 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2152 VERIFY0(spa_keystore_remove_mapping(os->os_spa,
2153 dmu_objset_id(os), FTAG));
2156 if (cookiep == NULL)
2164 zil_resume(void *cookie)
2166 objset_t *os = cookie;
2167 zilog_t *zilog = dmu_objset_zil(os);
2169 mutex_enter(&zilog->zl_lock);
2170 ASSERT(zilog->zl_suspend != 0);
2171 zilog->zl_suspend--;
2172 mutex_exit(&zilog->zl_lock);
2173 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2174 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2177 typedef struct zil_replay_arg {
2178 zil_replay_func_t **zr_replay;
2180 boolean_t zr_byteswap;
2185 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2187 char name[ZFS_MAX_DATASET_NAME_LEN];
2189 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2191 dmu_objset_name(zilog->zl_os, name);
2193 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2194 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2195 (u_longlong_t)lr->lrc_seq,
2196 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2197 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2203 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2205 zil_replay_arg_t *zr = zra;
2206 const zil_header_t *zh = zilog->zl_header;
2207 uint64_t reclen = lr->lrc_reclen;
2208 uint64_t txtype = lr->lrc_txtype;
2211 zilog->zl_replaying_seq = lr->lrc_seq;
2213 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2216 if (lr->lrc_txg < claim_txg) /* already committed */
2219 /* Strip case-insensitive bit, still present in log record */
2222 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2223 return (zil_replay_error(zilog, lr, EINVAL));
2226 * If this record type can be logged out of order, the object
2227 * (lr_foid) may no longer exist. That's legitimate, not an error.
2229 if (TX_OOO(txtype)) {
2230 error = dmu_object_info(zilog->zl_os,
2231 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
2232 if (error == ENOENT || error == EEXIST)
2237 * Make a copy of the data so we can revise and extend it.
2239 bcopy(lr, zr->zr_lr, reclen);
2242 * If this is a TX_WRITE with a blkptr, suck in the data.
2244 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2245 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2246 zr->zr_lr + reclen);
2248 return (zil_replay_error(zilog, lr, error));
2252 * The log block containing this lr may have been byteswapped
2253 * so that we can easily examine common fields like lrc_txtype.
2254 * However, the log is a mix of different record types, and only the
2255 * replay vectors know how to byteswap their records. Therefore, if
2256 * the lr was byteswapped, undo it before invoking the replay vector.
2258 if (zr->zr_byteswap)
2259 byteswap_uint64_array(zr->zr_lr, reclen);
2262 * We must now do two things atomically: replay this log record,
2263 * and update the log header sequence number to reflect the fact that
2264 * we did so. At the end of each replay function the sequence number
2265 * is updated if we are in replay mode.
2267 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2270 * The DMU's dnode layer doesn't see removes until the txg
2271 * commits, so a subsequent claim can spuriously fail with
2272 * EEXIST. So if we receive any error we try syncing out
2273 * any removes then retry the transaction. Note that we
2274 * specify B_FALSE for byteswap now, so we don't do it twice.
2276 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2277 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2279 return (zil_replay_error(zilog, lr, error));
2286 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2288 zilog->zl_replay_blks++;
2294 * If this dataset has a non-empty intent log, replay it and destroy it.
2297 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2299 zilog_t *zilog = dmu_objset_zil(os);
2300 const zil_header_t *zh = zilog->zl_header;
2301 zil_replay_arg_t zr;
2303 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2304 zil_destroy(zilog, B_TRUE);
2308 zr.zr_replay = replay_func;
2310 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2311 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2314 * Wait for in-progress removes to sync before starting replay.
2316 txg_wait_synced(zilog->zl_dmu_pool, 0);
2318 zilog->zl_replay = B_TRUE;
2319 zilog->zl_replay_time = ddi_get_lbolt();
2320 ASSERT(zilog->zl_replay_blks == 0);
2321 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2322 zh->zh_claim_txg, B_TRUE);
2323 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2325 zil_destroy(zilog, B_FALSE);
2326 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2327 zilog->zl_replay = B_FALSE;
2331 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2333 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2336 if (zilog->zl_replay) {
2337 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2338 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2339 zilog->zl_replaying_seq;
2348 zil_vdev_offline(const char *osname, void *arg)
2352 error = zil_suspend(osname, NULL);
2354 return (SET_ERROR(EEXIST));
2358 #if defined(_KERNEL) && defined(HAVE_SPL)
2359 EXPORT_SYMBOL(zil_alloc);
2360 EXPORT_SYMBOL(zil_free);
2361 EXPORT_SYMBOL(zil_open);
2362 EXPORT_SYMBOL(zil_close);
2363 EXPORT_SYMBOL(zil_replay);
2364 EXPORT_SYMBOL(zil_replaying);
2365 EXPORT_SYMBOL(zil_destroy);
2366 EXPORT_SYMBOL(zil_destroy_sync);
2367 EXPORT_SYMBOL(zil_itx_create);
2368 EXPORT_SYMBOL(zil_itx_destroy);
2369 EXPORT_SYMBOL(zil_itx_assign);
2370 EXPORT_SYMBOL(zil_commit);
2371 EXPORT_SYMBOL(zil_vdev_offline);
2372 EXPORT_SYMBOL(zil_claim);
2373 EXPORT_SYMBOL(zil_check_log_chain);
2374 EXPORT_SYMBOL(zil_sync);
2375 EXPORT_SYMBOL(zil_clean);
2376 EXPORT_SYMBOL(zil_suspend);
2377 EXPORT_SYMBOL(zil_resume);
2378 EXPORT_SYMBOL(zil_add_block);
2379 EXPORT_SYMBOL(zil_bp_tree_add);
2380 EXPORT_SYMBOL(zil_set_sync);
2381 EXPORT_SYMBOL(zil_set_logbias);
2384 module_param(zil_replay_disable, int, 0644);
2385 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
2387 module_param(zfs_nocacheflush, int, 0644);
2388 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");
2390 module_param(zil_slog_bulk, ulong, 0644);
2391 MODULE_PARM_DESC(zil_slog_bulk, "Limit in bytes slog sync writes per commit");