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, 2016 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>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * Disable intent logging replay. This global ZIL switch affects all pools.
72 int zil_replay_disable = 0;
73 SYSCTL_DECL(_vfs_zfs);
74 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
75 &zil_replay_disable, 0, "Disable intent logging replay");
78 * Tunable parameter for debugging or performance analysis. Setting
79 * zfs_nocacheflush will cause corruption on power loss if a volatile
80 * out-of-order write cache is enabled.
82 boolean_t zfs_nocacheflush = B_FALSE;
83 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
84 &zfs_nocacheflush, 0, "Disable cache flush");
85 boolean_t zfs_trim_enabled = B_TRUE;
86 SYSCTL_DECL(_vfs_zfs_trim);
87 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
91 * Limit SLOG write size per commit executed with synchronous priority.
92 * Any writes above that executed with lower (asynchronous) priority to
93 * limit potential SLOG device abuse by single active ZIL writer.
95 uint64_t zil_slog_limit = 768 * 1024;
96 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_limit, CTLFLAG_RWTUN,
97 &zil_slog_limit, 0, "Maximal SLOG commit size with sync priority");
99 static kmem_cache_t *zil_lwb_cache;
101 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
102 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
106 * ziltest is by and large an ugly hack, but very useful in
107 * checking replay without tedious work.
108 * When running ziltest we want to keep all itx's and so maintain
109 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
110 * We subtract TXG_CONCURRENT_STATES to allow for common code.
112 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
115 zil_bp_compare(const void *x1, const void *x2)
117 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
118 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
120 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
122 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
125 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
127 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
134 zil_bp_tree_init(zilog_t *zilog)
136 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
137 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
141 zil_bp_tree_fini(zilog_t *zilog)
143 avl_tree_t *t = &zilog->zl_bp_tree;
147 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
148 kmem_free(zn, sizeof (zil_bp_node_t));
154 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
156 avl_tree_t *t = &zilog->zl_bp_tree;
161 if (BP_IS_EMBEDDED(bp))
164 dva = BP_IDENTITY(bp);
166 if (avl_find(t, dva, &where) != NULL)
167 return (SET_ERROR(EEXIST));
169 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
171 avl_insert(t, zn, where);
176 static zil_header_t *
177 zil_header_in_syncing_context(zilog_t *zilog)
179 return ((zil_header_t *)zilog->zl_header);
183 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
185 zio_cksum_t *zc = &bp->blk_cksum;
187 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
188 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
189 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
190 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
194 * Read a log block and make sure it's valid.
197 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
200 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
201 arc_flags_t aflags = ARC_FLAG_WAIT;
202 arc_buf_t *abuf = NULL;
206 if (zilog->zl_header->zh_claim_txg == 0)
207 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
209 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
210 zio_flags |= ZIO_FLAG_SPECULATIVE;
212 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
213 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
215 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
216 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
219 zio_cksum_t cksum = bp->blk_cksum;
222 * Validate the checksummed log block.
224 * Sequence numbers should be... sequential. The checksum
225 * verifier for the next block should be bp's checksum plus 1.
227 * Also check the log chain linkage and size used.
229 cksum.zc_word[ZIL_ZC_SEQ]++;
231 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
232 zil_chain_t *zilc = abuf->b_data;
233 char *lr = (char *)(zilc + 1);
234 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
236 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
237 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
238 error = SET_ERROR(ECKSUM);
240 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
242 *end = (char *)dst + len;
243 *nbp = zilc->zc_next_blk;
246 char *lr = abuf->b_data;
247 uint64_t size = BP_GET_LSIZE(bp);
248 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
250 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
251 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
252 (zilc->zc_nused > (size - sizeof (*zilc)))) {
253 error = SET_ERROR(ECKSUM);
255 ASSERT3U(zilc->zc_nused, <=,
256 SPA_OLD_MAXBLOCKSIZE);
257 bcopy(lr, dst, zilc->zc_nused);
258 *end = (char *)dst + zilc->zc_nused;
259 *nbp = zilc->zc_next_blk;
263 arc_buf_destroy(abuf, &abuf);
270 * Read a TX_WRITE log data block.
273 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
275 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
276 const blkptr_t *bp = &lr->lr_blkptr;
277 arc_flags_t aflags = ARC_FLAG_WAIT;
278 arc_buf_t *abuf = NULL;
282 if (BP_IS_HOLE(bp)) {
284 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
288 if (zilog->zl_header->zh_claim_txg == 0)
289 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
291 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
292 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
294 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
295 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
299 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
300 arc_buf_destroy(abuf, &abuf);
307 * Parse the intent log, and call parse_func for each valid record within.
310 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
311 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
313 const zil_header_t *zh = zilog->zl_header;
314 boolean_t claimed = !!zh->zh_claim_txg;
315 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
316 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
317 uint64_t max_blk_seq = 0;
318 uint64_t max_lr_seq = 0;
319 uint64_t blk_count = 0;
320 uint64_t lr_count = 0;
321 blkptr_t blk, next_blk;
326 * Old logs didn't record the maximum zh_claim_lr_seq.
328 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
329 claim_lr_seq = UINT64_MAX;
332 * Starting at the block pointed to by zh_log we read the log chain.
333 * For each block in the chain we strongly check that block to
334 * ensure its validity. We stop when an invalid block is found.
335 * For each block pointer in the chain we call parse_blk_func().
336 * For each record in each valid block we call parse_lr_func().
337 * If the log has been claimed, stop if we encounter a sequence
338 * number greater than the highest claimed sequence number.
340 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
341 zil_bp_tree_init(zilog);
343 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
344 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
348 if (blk_seq > claim_blk_seq)
350 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
352 ASSERT3U(max_blk_seq, <, blk_seq);
353 max_blk_seq = blk_seq;
356 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
359 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
363 for (lrp = lrbuf; lrp < end; lrp += reclen) {
364 lr_t *lr = (lr_t *)lrp;
365 reclen = lr->lrc_reclen;
366 ASSERT3U(reclen, >=, sizeof (lr_t));
367 if (lr->lrc_seq > claim_lr_seq)
369 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
371 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
372 max_lr_seq = lr->lrc_seq;
377 zilog->zl_parse_error = error;
378 zilog->zl_parse_blk_seq = max_blk_seq;
379 zilog->zl_parse_lr_seq = max_lr_seq;
380 zilog->zl_parse_blk_count = blk_count;
381 zilog->zl_parse_lr_count = lr_count;
383 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
384 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
386 zil_bp_tree_fini(zilog);
387 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
393 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
396 * Claim log block if not already committed and not already claimed.
397 * If tx == NULL, just verify that the block is claimable.
399 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
400 zil_bp_tree_add(zilog, bp) != 0)
403 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
404 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
405 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
409 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
411 lr_write_t *lr = (lr_write_t *)lrc;
414 if (lrc->lrc_txtype != TX_WRITE)
418 * If the block is not readable, don't claim it. This can happen
419 * in normal operation when a log block is written to disk before
420 * some of the dmu_sync() blocks it points to. In this case, the
421 * transaction cannot have been committed to anyone (we would have
422 * waited for all writes to be stable first), so it is semantically
423 * correct to declare this the end of the log.
425 if (lr->lr_blkptr.blk_birth >= first_txg &&
426 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
428 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
433 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
435 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
441 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
443 lr_write_t *lr = (lr_write_t *)lrc;
444 blkptr_t *bp = &lr->lr_blkptr;
447 * If we previously claimed it, we need to free it.
449 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
450 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
452 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
458 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
462 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
463 lwb->lwb_zilog = zilog;
465 lwb->lwb_slog = slog;
466 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
467 lwb->lwb_max_txg = txg;
470 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
471 lwb->lwb_nused = sizeof (zil_chain_t);
472 lwb->lwb_sz = BP_GET_LSIZE(bp);
475 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
478 mutex_enter(&zilog->zl_lock);
479 list_insert_tail(&zilog->zl_lwb_list, lwb);
480 mutex_exit(&zilog->zl_lock);
486 * Called when we create in-memory log transactions so that we know
487 * to cleanup the itxs at the end of spa_sync().
490 zilog_dirty(zilog_t *zilog, uint64_t txg)
492 dsl_pool_t *dp = zilog->zl_dmu_pool;
493 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
495 if (ds->ds_is_snapshot)
496 panic("dirtying snapshot!");
498 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
499 /* up the hold count until we can be written out */
500 dmu_buf_add_ref(ds->ds_dbuf, zilog);
505 * Determine if the zil is dirty in the specified txg. Callers wanting to
506 * ensure that the dirty state does not change must hold the itxg_lock for
507 * the specified txg. Holding the lock will ensure that the zil cannot be
508 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
512 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
514 dsl_pool_t *dp = zilog->zl_dmu_pool;
516 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
522 * Determine if the zil is dirty. The zil is considered dirty if it has
523 * any pending itx records that have not been cleaned by zil_clean().
526 zilog_is_dirty(zilog_t *zilog)
528 dsl_pool_t *dp = zilog->zl_dmu_pool;
530 for (int t = 0; t < TXG_SIZE; t++) {
531 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
538 * Create an on-disk intent log.
541 zil_create(zilog_t *zilog)
543 const zil_header_t *zh = zilog->zl_header;
549 boolean_t slog = FALSE;
552 * Wait for any previous destroy to complete.
554 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
556 ASSERT(zh->zh_claim_txg == 0);
557 ASSERT(zh->zh_replay_seq == 0);
562 * Allocate an initial log block if:
563 * - there isn't one already
564 * - the existing block is the wrong endianess
566 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
567 tx = dmu_tx_create(zilog->zl_os);
568 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
569 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
570 txg = dmu_tx_get_txg(tx);
572 if (!BP_IS_HOLE(&blk)) {
573 zio_free_zil(zilog->zl_spa, txg, &blk);
577 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
578 ZIL_MIN_BLKSZ, &slog);
581 zil_init_log_chain(zilog, &blk);
585 * Allocate a log write buffer (lwb) for the first log block.
588 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
591 * If we just allocated the first log block, commit our transaction
592 * and wait for zil_sync() to stuff the block poiner into zh_log.
593 * (zh is part of the MOS, so we cannot modify it in open context.)
597 txg_wait_synced(zilog->zl_dmu_pool, txg);
600 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
606 * In one tx, free all log blocks and clear the log header.
607 * If keep_first is set, then we're replaying a log with no content.
608 * We want to keep the first block, however, so that the first
609 * synchronous transaction doesn't require a txg_wait_synced()
610 * in zil_create(). We don't need to txg_wait_synced() here either
611 * when keep_first is set, because both zil_create() and zil_destroy()
612 * will wait for any in-progress destroys to complete.
615 zil_destroy(zilog_t *zilog, boolean_t keep_first)
617 const zil_header_t *zh = zilog->zl_header;
623 * Wait for any previous destroy to complete.
625 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
627 zilog->zl_old_header = *zh; /* debugging aid */
629 if (BP_IS_HOLE(&zh->zh_log))
632 tx = dmu_tx_create(zilog->zl_os);
633 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
634 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
635 txg = dmu_tx_get_txg(tx);
637 mutex_enter(&zilog->zl_lock);
639 ASSERT3U(zilog->zl_destroy_txg, <, txg);
640 zilog->zl_destroy_txg = txg;
641 zilog->zl_keep_first = keep_first;
643 if (!list_is_empty(&zilog->zl_lwb_list)) {
644 ASSERT(zh->zh_claim_txg == 0);
646 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
647 list_remove(&zilog->zl_lwb_list, lwb);
648 if (lwb->lwb_buf != NULL)
649 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
650 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
651 kmem_cache_free(zil_lwb_cache, lwb);
653 } else if (!keep_first) {
654 zil_destroy_sync(zilog, tx);
656 mutex_exit(&zilog->zl_lock);
662 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
664 ASSERT(list_is_empty(&zilog->zl_lwb_list));
665 (void) zil_parse(zilog, zil_free_log_block,
666 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
670 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
672 dmu_tx_t *tx = txarg;
673 uint64_t first_txg = dmu_tx_get_txg(tx);
679 error = dmu_objset_own_obj(dp, ds->ds_object,
680 DMU_OST_ANY, B_FALSE, FTAG, &os);
683 * EBUSY indicates that the objset is inconsistent, in which
684 * case it can not have a ZIL.
686 if (error != EBUSY) {
687 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
688 (unsigned long long)ds->ds_object, error);
693 zilog = dmu_objset_zil(os);
694 zh = zil_header_in_syncing_context(zilog);
696 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
697 if (!BP_IS_HOLE(&zh->zh_log))
698 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
699 BP_ZERO(&zh->zh_log);
700 dsl_dataset_dirty(dmu_objset_ds(os), tx);
701 dmu_objset_disown(os, FTAG);
706 * Claim all log blocks if we haven't already done so, and remember
707 * the highest claimed sequence number. This ensures that if we can
708 * read only part of the log now (e.g. due to a missing device),
709 * but we can read the entire log later, we will not try to replay
710 * or destroy beyond the last block we successfully claimed.
712 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
713 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
714 (void) zil_parse(zilog, zil_claim_log_block,
715 zil_claim_log_record, tx, first_txg);
716 zh->zh_claim_txg = first_txg;
717 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
718 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
719 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
720 zh->zh_flags |= ZIL_REPLAY_NEEDED;
721 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
722 dsl_dataset_dirty(dmu_objset_ds(os), tx);
725 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
726 dmu_objset_disown(os, FTAG);
731 * Check the log by walking the log chain.
732 * Checksum errors are ok as they indicate the end of the chain.
733 * Any other error (no device or read failure) returns an error.
737 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
746 error = dmu_objset_from_ds(ds, &os);
748 cmn_err(CE_WARN, "can't open objset %llu, error %d",
749 (unsigned long long)ds->ds_object, error);
753 zilog = dmu_objset_zil(os);
754 bp = (blkptr_t *)&zilog->zl_header->zh_log;
757 * Check the first block and determine if it's on a log device
758 * which may have been removed or faulted prior to loading this
759 * pool. If so, there's no point in checking the rest of the log
760 * as its content should have already been synced to the pool.
762 if (!BP_IS_HOLE(bp)) {
764 boolean_t valid = B_TRUE;
766 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
767 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
768 if (vd->vdev_islog && vdev_is_dead(vd))
769 valid = vdev_log_state_valid(vd);
770 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
777 * Because tx == NULL, zil_claim_log_block() will not actually claim
778 * any blocks, but just determine whether it is possible to do so.
779 * In addition to checking the log chain, zil_claim_log_block()
780 * will invoke zio_claim() with a done func of spa_claim_notify(),
781 * which will update spa_max_claim_txg. See spa_load() for details.
783 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
784 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
786 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
790 zil_vdev_compare(const void *x1, const void *x2)
792 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
793 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
804 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
806 avl_tree_t *t = &zilog->zl_vdev_tree;
808 zil_vdev_node_t *zv, zvsearch;
809 int ndvas = BP_GET_NDVAS(bp);
812 if (zfs_nocacheflush)
815 ASSERT(zilog->zl_writer);
818 * Even though we're zl_writer, we still need a lock because the
819 * zl_get_data() callbacks may have dmu_sync() done callbacks
820 * that will run concurrently.
822 mutex_enter(&zilog->zl_vdev_lock);
823 for (i = 0; i < ndvas; i++) {
824 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
825 if (avl_find(t, &zvsearch, &where) == NULL) {
826 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
827 zv->zv_vdev = zvsearch.zv_vdev;
828 avl_insert(t, zv, where);
831 mutex_exit(&zilog->zl_vdev_lock);
835 zil_flush_vdevs(zilog_t *zilog)
837 spa_t *spa = zilog->zl_spa;
838 avl_tree_t *t = &zilog->zl_vdev_tree;
843 ASSERT(zilog->zl_writer);
846 * We don't need zl_vdev_lock here because we're the zl_writer,
847 * and all zl_get_data() callbacks are done.
849 if (avl_numnodes(t) == 0)
852 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
854 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
856 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
857 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
860 kmem_free(zv, sizeof (*zv));
864 * Wait for all the flushes to complete. Not all devices actually
865 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
867 (void) zio_wait(zio);
869 spa_config_exit(spa, SCL_STATE, FTAG);
873 * Function called when a log block write completes
876 zil_lwb_write_done(zio_t *zio)
878 lwb_t *lwb = zio->io_private;
879 zilog_t *zilog = lwb->lwb_zilog;
880 dmu_tx_t *tx = lwb->lwb_tx;
882 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
883 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
884 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
885 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
886 ASSERT(!BP_IS_GANG(zio->io_bp));
887 ASSERT(!BP_IS_HOLE(zio->io_bp));
888 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
891 * Ensure the lwb buffer pointer is cleared before releasing
892 * the txg. If we have had an allocation failure and
893 * the txg is waiting to sync then we want want zil_sync()
894 * to remove the lwb so that it's not picked up as the next new
895 * one in zil_commit_writer(). zil_sync() will only remove
896 * the lwb if lwb_buf is null.
898 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
899 mutex_enter(&zilog->zl_lock);
902 mutex_exit(&zilog->zl_lock);
905 * Now that we've written this log block, we have a stable pointer
906 * to the next block in the chain, so it's OK to let the txg in
907 * which we allocated the next block sync.
913 * Initialize the io for a log block.
916 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
921 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
922 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
923 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
925 if (zilog->zl_root_zio == NULL) {
926 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
929 if (lwb->lwb_zio == NULL) {
930 if (zilog->zl_cur_used <= zil_slog_limit || !lwb->lwb_slog)
931 prio = ZIO_PRIORITY_SYNC_WRITE;
933 prio = ZIO_PRIORITY_ASYNC_WRITE;
934 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
935 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
936 zil_lwb_write_done, lwb, prio,
937 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
942 * Define a limited set of intent log block sizes.
944 * These must be a multiple of 4KB. Note only the amount used (again
945 * aligned to 4KB) actually gets written. However, we can't always just
946 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
948 uint64_t zil_block_buckets[] = {
949 4096, /* non TX_WRITE */
950 8192+4096, /* data base */
951 32*1024 + 4096, /* NFS writes */
956 * Start a log block write and advance to the next log block.
957 * Calls are serialized.
960 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
964 spa_t *spa = zilog->zl_spa;
968 uint64_t zil_blksz, wsz;
972 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
973 zilc = (zil_chain_t *)lwb->lwb_buf;
974 bp = &zilc->zc_next_blk;
976 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
977 bp = &zilc->zc_next_blk;
980 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
983 * Allocate the next block and save its address in this block
984 * before writing it in order to establish the log chain.
985 * Note that if the allocation of nlwb synced before we wrote
986 * the block that points at it (lwb), we'd leak it if we crashed.
987 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
988 * We dirty the dataset to ensure that zil_sync() will be called
989 * to clean up in the event of allocation failure or I/O failure.
991 tx = dmu_tx_create(zilog->zl_os);
992 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
993 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
994 txg = dmu_tx_get_txg(tx);
999 * Log blocks are pre-allocated. Here we select the size of the next
1000 * block, based on size used in the last block.
1001 * - first find the smallest bucket that will fit the block from a
1002 * limited set of block sizes. This is because it's faster to write
1003 * blocks allocated from the same metaslab as they are adjacent or
1005 * - next find the maximum from the new suggested size and an array of
1006 * previous sizes. This lessens a picket fence effect of wrongly
1007 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1010 * Note we only write what is used, but we can't just allocate
1011 * the maximum block size because we can exhaust the available
1014 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1015 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1017 zil_blksz = zil_block_buckets[i];
1018 if (zil_blksz == UINT64_MAX)
1019 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1020 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1021 for (i = 0; i < ZIL_PREV_BLKS; i++)
1022 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1023 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1026 /* pass the old blkptr in order to spread log blocks across devs */
1027 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1029 ASSERT3U(bp->blk_birth, ==, txg);
1030 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1031 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1034 * Allocate a new log write buffer (lwb).
1036 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1038 /* Record the block for later vdev flushing */
1039 zil_add_block(zilog, &lwb->lwb_blk);
1042 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1043 /* For Slim ZIL only write what is used. */
1044 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1045 ASSERT3U(wsz, <=, lwb->lwb_sz);
1046 zio_shrink(lwb->lwb_zio, wsz);
1053 zilc->zc_nused = lwb->lwb_nused;
1054 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1057 * clear unused data for security
1059 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1061 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1064 * If there was an allocation failure then nlwb will be null which
1065 * forces a txg_wait_synced().
1071 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1073 lr_t *lrcb, *lrc = &itx->itx_lr; /* common log record */
1074 lr_write_t *lrwb, *lrw = (lr_write_t *)lrc;
1076 uint64_t txg = lrc->lrc_txg;
1077 uint64_t reclen = lrc->lrc_reclen;
1079 uint64_t dnow, lwb_sp;
1084 ASSERT(lwb->lwb_buf != NULL);
1086 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1087 dlen = P2ROUNDUP_TYPED(
1088 lrw->lr_length, sizeof (uint64_t), uint64_t);
1090 zilog->zl_cur_used += (reclen + dlen);
1092 zil_lwb_write_init(zilog, lwb);
1096 * If this record won't fit in the current log block, start a new one.
1097 * For WR_NEED_COPY optimize layout for minimal number of chunks, but
1098 * try to keep wasted space withing reasonable range (12%).
1100 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1101 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1102 lwb_sp < ZIL_MAX_LOG_DATA / 8 && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1103 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1104 lwb = zil_lwb_write_start(zilog, lwb);
1107 zil_lwb_write_init(zilog, lwb);
1108 ASSERT(LWB_EMPTY(lwb));
1109 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1110 ASSERT3U(reclen + MIN(dlen, sizeof(uint64_t)), <=, lwb_sp);
1113 dnow = MIN(dlen, lwb_sp - reclen);
1114 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1115 bcopy(lrc, lr_buf, reclen);
1116 lrcb = (lr_t *)lr_buf;
1117 lrwb = (lr_write_t *)lrcb;
1120 * If it's a write, fetch the data or get its blkptr as appropriate.
1122 if (lrc->lrc_txtype == TX_WRITE) {
1123 if (txg > spa_freeze_txg(zilog->zl_spa))
1124 txg_wait_synced(zilog->zl_dmu_pool, txg);
1125 if (itx->itx_wr_state != WR_COPIED) {
1129 if (itx->itx_wr_state == WR_NEED_COPY) {
1130 dbuf = lr_buf + reclen;
1131 lrcb->lrc_reclen += dnow;
1132 if (lrwb->lr_length > dnow)
1133 lrwb->lr_length = dnow;
1134 lrw->lr_offset += dnow;
1135 lrw->lr_length -= dnow;
1137 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1140 error = zilog->zl_get_data(
1141 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1143 txg_wait_synced(zilog->zl_dmu_pool, txg);
1147 ASSERT(error == ENOENT || error == EEXIST ||
1155 * We're actually making an entry, so update lrc_seq to be the
1156 * log record sequence number. Note that this is generally not
1157 * equal to the itx sequence number because not all transactions
1158 * are synchronous, and sometimes spa_sync() gets there first.
1160 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1161 lwb->lwb_nused += reclen + dnow;
1162 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1163 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1164 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1168 zilog->zl_cur_used += reclen;
1176 zil_itx_create(uint64_t txtype, size_t lrsize)
1180 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1182 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1183 itx->itx_lr.lrc_txtype = txtype;
1184 itx->itx_lr.lrc_reclen = lrsize;
1185 itx->itx_lr.lrc_seq = 0; /* defensive */
1186 itx->itx_sync = B_TRUE; /* default is synchronous */
1192 zil_itx_destroy(itx_t *itx)
1194 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1198 * Free up the sync and async itxs. The itxs_t has already been detached
1199 * so no locks are needed.
1202 zil_itxg_clean(itxs_t *itxs)
1208 itx_async_node_t *ian;
1210 list = &itxs->i_sync_list;
1211 while ((itx = list_head(list)) != NULL) {
1212 list_remove(list, itx);
1213 kmem_free(itx, offsetof(itx_t, itx_lr) +
1214 itx->itx_lr.lrc_reclen);
1218 t = &itxs->i_async_tree;
1219 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1220 list = &ian->ia_list;
1221 while ((itx = list_head(list)) != NULL) {
1222 list_remove(list, itx);
1223 kmem_free(itx, offsetof(itx_t, itx_lr) +
1224 itx->itx_lr.lrc_reclen);
1227 kmem_free(ian, sizeof (itx_async_node_t));
1231 kmem_free(itxs, sizeof (itxs_t));
1235 zil_aitx_compare(const void *x1, const void *x2)
1237 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1238 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1249 * Remove all async itx with the given oid.
1252 zil_remove_async(zilog_t *zilog, uint64_t oid)
1255 itx_async_node_t *ian;
1262 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1264 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1267 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1269 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1270 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1272 mutex_enter(&itxg->itxg_lock);
1273 if (itxg->itxg_txg != txg) {
1274 mutex_exit(&itxg->itxg_lock);
1279 * Locate the object node and append its list.
1281 t = &itxg->itxg_itxs->i_async_tree;
1282 ian = avl_find(t, &oid, &where);
1284 list_move_tail(&clean_list, &ian->ia_list);
1285 mutex_exit(&itxg->itxg_lock);
1287 while ((itx = list_head(&clean_list)) != NULL) {
1288 list_remove(&clean_list, itx);
1289 kmem_free(itx, offsetof(itx_t, itx_lr) +
1290 itx->itx_lr.lrc_reclen);
1292 list_destroy(&clean_list);
1296 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1300 itxs_t *itxs, *clean = NULL;
1303 * Object ids can be re-instantiated in the next txg so
1304 * remove any async transactions to avoid future leaks.
1305 * This can happen if a fsync occurs on the re-instantiated
1306 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1307 * the new file data and flushes a write record for the old object.
1309 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1310 zil_remove_async(zilog, itx->itx_oid);
1313 * Ensure the data of a renamed file is committed before the rename.
1315 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1316 zil_async_to_sync(zilog, itx->itx_oid);
1318 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1321 txg = dmu_tx_get_txg(tx);
1323 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1324 mutex_enter(&itxg->itxg_lock);
1325 itxs = itxg->itxg_itxs;
1326 if (itxg->itxg_txg != txg) {
1329 * The zil_clean callback hasn't got around to cleaning
1330 * this itxg. Save the itxs for release below.
1331 * This should be rare.
1333 clean = itxg->itxg_itxs;
1335 itxg->itxg_txg = txg;
1336 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1338 list_create(&itxs->i_sync_list, sizeof (itx_t),
1339 offsetof(itx_t, itx_node));
1340 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1341 sizeof (itx_async_node_t),
1342 offsetof(itx_async_node_t, ia_node));
1344 if (itx->itx_sync) {
1345 list_insert_tail(&itxs->i_sync_list, itx);
1347 avl_tree_t *t = &itxs->i_async_tree;
1348 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1349 itx_async_node_t *ian;
1352 ian = avl_find(t, &foid, &where);
1354 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1355 list_create(&ian->ia_list, sizeof (itx_t),
1356 offsetof(itx_t, itx_node));
1357 ian->ia_foid = foid;
1358 avl_insert(t, ian, where);
1360 list_insert_tail(&ian->ia_list, itx);
1363 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1364 zilog_dirty(zilog, txg);
1365 mutex_exit(&itxg->itxg_lock);
1367 /* Release the old itxs now we've dropped the lock */
1369 zil_itxg_clean(clean);
1373 * If there are any in-memory intent log transactions which have now been
1374 * synced then start up a taskq to free them. We should only do this after we
1375 * have written out the uberblocks (i.e. txg has been comitted) so that
1376 * don't inadvertently clean out in-memory log records that would be required
1380 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1382 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1385 mutex_enter(&itxg->itxg_lock);
1386 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1387 mutex_exit(&itxg->itxg_lock);
1390 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1391 ASSERT(itxg->itxg_txg != 0);
1392 ASSERT(zilog->zl_clean_taskq != NULL);
1393 clean_me = itxg->itxg_itxs;
1394 itxg->itxg_itxs = NULL;
1396 mutex_exit(&itxg->itxg_lock);
1398 * Preferably start a task queue to free up the old itxs but
1399 * if taskq_dispatch can't allocate resources to do that then
1400 * free it in-line. This should be rare. Note, using TQ_SLEEP
1401 * created a bad performance problem.
1403 if (taskq_dispatch(zilog->zl_clean_taskq,
1404 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1405 zil_itxg_clean(clean_me);
1409 * Get the list of itxs to commit into zl_itx_commit_list.
1412 zil_get_commit_list(zilog_t *zilog)
1415 list_t *commit_list = &zilog->zl_itx_commit_list;
1417 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1420 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1423 * This is inherently racy, since there is nothing to prevent
1424 * the last synced txg from changing. That's okay since we'll
1425 * only commit things in the future.
1427 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1428 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1430 mutex_enter(&itxg->itxg_lock);
1431 if (itxg->itxg_txg != txg) {
1432 mutex_exit(&itxg->itxg_lock);
1437 * If we're adding itx records to the zl_itx_commit_list,
1438 * then the zil better be dirty in this "txg". We can assert
1439 * that here since we're holding the itxg_lock which will
1440 * prevent spa_sync from cleaning it. Once we add the itxs
1441 * to the zl_itx_commit_list we must commit it to disk even
1442 * if it's unnecessary (i.e. the txg was synced).
1444 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1445 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1446 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1448 mutex_exit(&itxg->itxg_lock);
1453 * Move the async itxs for a specified object to commit into sync lists.
1456 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1459 itx_async_node_t *ian;
1463 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1466 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1469 * This is inherently racy, since there is nothing to prevent
1470 * the last synced txg from changing.
1472 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1473 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1475 mutex_enter(&itxg->itxg_lock);
1476 if (itxg->itxg_txg != txg) {
1477 mutex_exit(&itxg->itxg_lock);
1482 * If a foid is specified then find that node and append its
1483 * list. Otherwise walk the tree appending all the lists
1484 * to the sync list. We add to the end rather than the
1485 * beginning to ensure the create has happened.
1487 t = &itxg->itxg_itxs->i_async_tree;
1489 ian = avl_find(t, &foid, &where);
1491 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1495 void *cookie = NULL;
1497 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1498 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1500 list_destroy(&ian->ia_list);
1501 kmem_free(ian, sizeof (itx_async_node_t));
1504 mutex_exit(&itxg->itxg_lock);
1509 zil_commit_writer(zilog_t *zilog)
1514 spa_t *spa = zilog->zl_spa;
1517 ASSERT(zilog->zl_root_zio == NULL);
1519 mutex_exit(&zilog->zl_lock);
1521 zil_get_commit_list(zilog);
1524 * Return if there's nothing to commit before we dirty the fs by
1525 * calling zil_create().
1527 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1528 mutex_enter(&zilog->zl_lock);
1532 if (zilog->zl_suspend) {
1535 lwb = list_tail(&zilog->zl_lwb_list);
1537 lwb = zil_create(zilog);
1540 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1541 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1542 txg = itx->itx_lr.lrc_txg;
1543 ASSERT3U(txg, !=, 0);
1546 * This is inherently racy and may result in us writing
1547 * out a log block for a txg that was just synced. This is
1548 * ok since we'll end cleaning up that log block the next
1549 * time we call zil_sync().
1551 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1552 lwb = zil_lwb_commit(zilog, itx, lwb);
1553 list_remove(&zilog->zl_itx_commit_list, itx);
1554 kmem_free(itx, offsetof(itx_t, itx_lr)
1555 + itx->itx_lr.lrc_reclen);
1557 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1559 /* write the last block out */
1560 if (lwb != NULL && lwb->lwb_zio != NULL)
1561 lwb = zil_lwb_write_start(zilog, lwb);
1563 zilog->zl_cur_used = 0;
1566 * Wait if necessary for the log blocks to be on stable storage.
1568 if (zilog->zl_root_zio) {
1569 error = zio_wait(zilog->zl_root_zio);
1570 zilog->zl_root_zio = NULL;
1571 zil_flush_vdevs(zilog);
1574 if (error || lwb == NULL)
1575 txg_wait_synced(zilog->zl_dmu_pool, 0);
1577 mutex_enter(&zilog->zl_lock);
1580 * Remember the highest committed log sequence number for ztest.
1581 * We only update this value when all the log writes succeeded,
1582 * because ztest wants to ASSERT that it got the whole log chain.
1584 if (error == 0 && lwb != NULL)
1585 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1589 * Commit zfs transactions to stable storage.
1590 * If foid is 0 push out all transactions, otherwise push only those
1591 * for that object or might reference that object.
1593 * itxs are committed in batches. In a heavily stressed zil there will be
1594 * a commit writer thread who is writing out a bunch of itxs to the log
1595 * for a set of committing threads (cthreads) in the same batch as the writer.
1596 * Those cthreads are all waiting on the same cv for that batch.
1598 * There will also be a different and growing batch of threads that are
1599 * waiting to commit (qthreads). When the committing batch completes
1600 * a transition occurs such that the cthreads exit and the qthreads become
1601 * cthreads. One of the new cthreads becomes the writer thread for the
1602 * batch. Any new threads arriving become new qthreads.
1604 * Only 2 condition variables are needed and there's no transition
1605 * between the two cvs needed. They just flip-flop between qthreads
1608 * Using this scheme we can efficiently wakeup up only those threads
1609 * that have been committed.
1612 zil_commit(zilog_t *zilog, uint64_t foid)
1616 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1619 /* move the async itxs for the foid to the sync queues */
1620 zil_async_to_sync(zilog, foid);
1622 mutex_enter(&zilog->zl_lock);
1623 mybatch = zilog->zl_next_batch;
1624 while (zilog->zl_writer) {
1625 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1626 if (mybatch <= zilog->zl_com_batch) {
1627 mutex_exit(&zilog->zl_lock);
1632 zilog->zl_next_batch++;
1633 zilog->zl_writer = B_TRUE;
1634 zil_commit_writer(zilog);
1635 zilog->zl_com_batch = mybatch;
1636 zilog->zl_writer = B_FALSE;
1637 mutex_exit(&zilog->zl_lock);
1639 /* wake up one thread to become the next writer */
1640 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1642 /* wake up all threads waiting for this batch to be committed */
1643 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1647 * Called in syncing context to free committed log blocks and update log header.
1650 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1652 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1653 uint64_t txg = dmu_tx_get_txg(tx);
1654 spa_t *spa = zilog->zl_spa;
1655 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1659 * We don't zero out zl_destroy_txg, so make sure we don't try
1660 * to destroy it twice.
1662 if (spa_sync_pass(spa) != 1)
1665 mutex_enter(&zilog->zl_lock);
1667 ASSERT(zilog->zl_stop_sync == 0);
1669 if (*replayed_seq != 0) {
1670 ASSERT(zh->zh_replay_seq < *replayed_seq);
1671 zh->zh_replay_seq = *replayed_seq;
1675 if (zilog->zl_destroy_txg == txg) {
1676 blkptr_t blk = zh->zh_log;
1678 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1680 bzero(zh, sizeof (zil_header_t));
1681 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1683 if (zilog->zl_keep_first) {
1685 * If this block was part of log chain that couldn't
1686 * be claimed because a device was missing during
1687 * zil_claim(), but that device later returns,
1688 * then this block could erroneously appear valid.
1689 * To guard against this, assign a new GUID to the new
1690 * log chain so it doesn't matter what blk points to.
1692 zil_init_log_chain(zilog, &blk);
1697 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1698 zh->zh_log = lwb->lwb_blk;
1699 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1701 list_remove(&zilog->zl_lwb_list, lwb);
1702 zio_free_zil(spa, txg, &lwb->lwb_blk);
1703 kmem_cache_free(zil_lwb_cache, lwb);
1706 * If we don't have anything left in the lwb list then
1707 * we've had an allocation failure and we need to zero
1708 * out the zil_header blkptr so that we don't end
1709 * up freeing the same block twice.
1711 if (list_head(&zilog->zl_lwb_list) == NULL)
1712 BP_ZERO(&zh->zh_log);
1714 mutex_exit(&zilog->zl_lock);
1720 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1721 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1727 kmem_cache_destroy(zil_lwb_cache);
1731 zil_set_sync(zilog_t *zilog, uint64_t sync)
1733 zilog->zl_sync = sync;
1737 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1739 zilog->zl_logbias = logbias;
1743 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1747 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1749 zilog->zl_header = zh_phys;
1751 zilog->zl_spa = dmu_objset_spa(os);
1752 zilog->zl_dmu_pool = dmu_objset_pool(os);
1753 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1754 zilog->zl_logbias = dmu_objset_logbias(os);
1755 zilog->zl_sync = dmu_objset_syncprop(os);
1756 zilog->zl_next_batch = 1;
1758 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1760 for (int i = 0; i < TXG_SIZE; i++) {
1761 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1762 MUTEX_DEFAULT, NULL);
1765 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1766 offsetof(lwb_t, lwb_node));
1768 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1769 offsetof(itx_t, itx_node));
1771 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1773 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1774 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1776 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1777 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1778 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1779 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1785 zil_free(zilog_t *zilog)
1787 zilog->zl_stop_sync = 1;
1789 ASSERT0(zilog->zl_suspend);
1790 ASSERT0(zilog->zl_suspending);
1792 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1793 list_destroy(&zilog->zl_lwb_list);
1795 avl_destroy(&zilog->zl_vdev_tree);
1796 mutex_destroy(&zilog->zl_vdev_lock);
1798 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1799 list_destroy(&zilog->zl_itx_commit_list);
1801 for (int i = 0; i < TXG_SIZE; i++) {
1803 * It's possible for an itx to be generated that doesn't dirty
1804 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1805 * callback to remove the entry. We remove those here.
1807 * Also free up the ziltest itxs.
1809 if (zilog->zl_itxg[i].itxg_itxs)
1810 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1811 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1814 mutex_destroy(&zilog->zl_lock);
1816 cv_destroy(&zilog->zl_cv_writer);
1817 cv_destroy(&zilog->zl_cv_suspend);
1818 cv_destroy(&zilog->zl_cv_batch[0]);
1819 cv_destroy(&zilog->zl_cv_batch[1]);
1821 kmem_free(zilog, sizeof (zilog_t));
1825 * Open an intent log.
1828 zil_open(objset_t *os, zil_get_data_t *get_data)
1830 zilog_t *zilog = dmu_objset_zil(os);
1832 ASSERT(zilog->zl_clean_taskq == NULL);
1833 ASSERT(zilog->zl_get_data == NULL);
1834 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1836 zilog->zl_get_data = get_data;
1837 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1838 2, 2, TASKQ_PREPOPULATE);
1844 * Close an intent log.
1847 zil_close(zilog_t *zilog)
1852 zil_commit(zilog, 0); /* commit all itx */
1855 * The lwb_max_txg for the stubby lwb will reflect the last activity
1856 * for the zil. After a txg_wait_synced() on the txg we know all the
1857 * callbacks have occurred that may clean the zil. Only then can we
1858 * destroy the zl_clean_taskq.
1860 mutex_enter(&zilog->zl_lock);
1861 lwb = list_tail(&zilog->zl_lwb_list);
1863 txg = lwb->lwb_max_txg;
1864 mutex_exit(&zilog->zl_lock);
1866 txg_wait_synced(zilog->zl_dmu_pool, txg);
1868 if (zilog_is_dirty(zilog))
1869 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1870 VERIFY(!zilog_is_dirty(zilog));
1872 taskq_destroy(zilog->zl_clean_taskq);
1873 zilog->zl_clean_taskq = NULL;
1874 zilog->zl_get_data = NULL;
1877 * We should have only one LWB left on the list; remove it now.
1879 mutex_enter(&zilog->zl_lock);
1880 lwb = list_head(&zilog->zl_lwb_list);
1882 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1883 list_remove(&zilog->zl_lwb_list, lwb);
1884 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1885 kmem_cache_free(zil_lwb_cache, lwb);
1887 mutex_exit(&zilog->zl_lock);
1890 static char *suspend_tag = "zil suspending";
1893 * Suspend an intent log. While in suspended mode, we still honor
1894 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1895 * On old version pools, we suspend the log briefly when taking a
1896 * snapshot so that it will have an empty intent log.
1898 * Long holds are not really intended to be used the way we do here --
1899 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1900 * could fail. Therefore we take pains to only put a long hold if it is
1901 * actually necessary. Fortunately, it will only be necessary if the
1902 * objset is currently mounted (or the ZVOL equivalent). In that case it
1903 * will already have a long hold, so we are not really making things any worse.
1905 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1906 * zvol_state_t), and use their mechanism to prevent their hold from being
1907 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1910 * if cookiep == NULL, this does both the suspend & resume.
1911 * Otherwise, it returns with the dataset "long held", and the cookie
1912 * should be passed into zil_resume().
1915 zil_suspend(const char *osname, void **cookiep)
1919 const zil_header_t *zh;
1922 error = dmu_objset_hold(osname, suspend_tag, &os);
1925 zilog = dmu_objset_zil(os);
1927 mutex_enter(&zilog->zl_lock);
1928 zh = zilog->zl_header;
1930 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1931 mutex_exit(&zilog->zl_lock);
1932 dmu_objset_rele(os, suspend_tag);
1933 return (SET_ERROR(EBUSY));
1937 * Don't put a long hold in the cases where we can avoid it. This
1938 * is when there is no cookie so we are doing a suspend & resume
1939 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1940 * for the suspend because it's already suspended, or there's no ZIL.
1942 if (cookiep == NULL && !zilog->zl_suspending &&
1943 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1944 mutex_exit(&zilog->zl_lock);
1945 dmu_objset_rele(os, suspend_tag);
1949 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1950 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1952 zilog->zl_suspend++;
1954 if (zilog->zl_suspend > 1) {
1956 * Someone else is already suspending it.
1957 * Just wait for them to finish.
1960 while (zilog->zl_suspending)
1961 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1962 mutex_exit(&zilog->zl_lock);
1964 if (cookiep == NULL)
1972 * If there is no pointer to an on-disk block, this ZIL must not
1973 * be active (e.g. filesystem not mounted), so there's nothing
1976 if (BP_IS_HOLE(&zh->zh_log)) {
1977 ASSERT(cookiep != NULL); /* fast path already handled */
1980 mutex_exit(&zilog->zl_lock);
1984 zilog->zl_suspending = B_TRUE;
1985 mutex_exit(&zilog->zl_lock);
1987 zil_commit(zilog, 0);
1989 zil_destroy(zilog, B_FALSE);
1991 mutex_enter(&zilog->zl_lock);
1992 zilog->zl_suspending = B_FALSE;
1993 cv_broadcast(&zilog->zl_cv_suspend);
1994 mutex_exit(&zilog->zl_lock);
1996 if (cookiep == NULL)
2004 zil_resume(void *cookie)
2006 objset_t *os = cookie;
2007 zilog_t *zilog = dmu_objset_zil(os);
2009 mutex_enter(&zilog->zl_lock);
2010 ASSERT(zilog->zl_suspend != 0);
2011 zilog->zl_suspend--;
2012 mutex_exit(&zilog->zl_lock);
2013 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2014 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2017 typedef struct zil_replay_arg {
2018 zil_replay_func_t **zr_replay;
2020 boolean_t zr_byteswap;
2025 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2027 char name[ZFS_MAX_DATASET_NAME_LEN];
2029 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2031 dmu_objset_name(zilog->zl_os, name);
2033 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2034 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2035 (u_longlong_t)lr->lrc_seq,
2036 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2037 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2043 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2045 zil_replay_arg_t *zr = zra;
2046 const zil_header_t *zh = zilog->zl_header;
2047 uint64_t reclen = lr->lrc_reclen;
2048 uint64_t txtype = lr->lrc_txtype;
2051 zilog->zl_replaying_seq = lr->lrc_seq;
2053 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2056 if (lr->lrc_txg < claim_txg) /* already committed */
2059 /* Strip case-insensitive bit, still present in log record */
2062 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2063 return (zil_replay_error(zilog, lr, EINVAL));
2066 * If this record type can be logged out of order, the object
2067 * (lr_foid) may no longer exist. That's legitimate, not an error.
2069 if (TX_OOO(txtype)) {
2070 error = dmu_object_info(zilog->zl_os,
2071 ((lr_ooo_t *)lr)->lr_foid, NULL);
2072 if (error == ENOENT || error == EEXIST)
2077 * Make a copy of the data so we can revise and extend it.
2079 bcopy(lr, zr->zr_lr, reclen);
2082 * If this is a TX_WRITE with a blkptr, suck in the data.
2084 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2085 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2086 zr->zr_lr + reclen);
2088 return (zil_replay_error(zilog, lr, error));
2092 * The log block containing this lr may have been byteswapped
2093 * so that we can easily examine common fields like lrc_txtype.
2094 * However, the log is a mix of different record types, and only the
2095 * replay vectors know how to byteswap their records. Therefore, if
2096 * the lr was byteswapped, undo it before invoking the replay vector.
2098 if (zr->zr_byteswap)
2099 byteswap_uint64_array(zr->zr_lr, reclen);
2102 * We must now do two things atomically: replay this log record,
2103 * and update the log header sequence number to reflect the fact that
2104 * we did so. At the end of each replay function the sequence number
2105 * is updated if we are in replay mode.
2107 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2110 * The DMU's dnode layer doesn't see removes until the txg
2111 * commits, so a subsequent claim can spuriously fail with
2112 * EEXIST. So if we receive any error we try syncing out
2113 * any removes then retry the transaction. Note that we
2114 * specify B_FALSE for byteswap now, so we don't do it twice.
2116 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2117 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2119 return (zil_replay_error(zilog, lr, error));
2126 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2128 zilog->zl_replay_blks++;
2134 * If this dataset has a non-empty intent log, replay it and destroy it.
2137 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2139 zilog_t *zilog = dmu_objset_zil(os);
2140 const zil_header_t *zh = zilog->zl_header;
2141 zil_replay_arg_t zr;
2143 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2144 zil_destroy(zilog, B_TRUE);
2148 zr.zr_replay = replay_func;
2150 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2151 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2154 * Wait for in-progress removes to sync before starting replay.
2156 txg_wait_synced(zilog->zl_dmu_pool, 0);
2158 zilog->zl_replay = B_TRUE;
2159 zilog->zl_replay_time = ddi_get_lbolt();
2160 ASSERT(zilog->zl_replay_blks == 0);
2161 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2163 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2165 zil_destroy(zilog, B_FALSE);
2166 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2167 zilog->zl_replay = B_FALSE;
2171 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2173 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2176 if (zilog->zl_replay) {
2177 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2178 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2179 zilog->zl_replaying_seq;
2188 zil_vdev_offline(const char *osname, void *arg)
2192 error = zil_suspend(osname, NULL);
2194 return (SET_ERROR(EEXIST));