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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 /* Portions Copyright 2010 Robert Milkowski */
30 #include <sys/zfs_context.h>
36 #include <sys/resource.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
45 * The zfs intent log (ZIL) saves transaction records of system calls
46 * that change the file system in memory with enough information
47 * to be able to replay them. These are stored in memory until
48 * either the DMU transaction group (txg) commits them to the stable pool
49 * and they can be discarded, or they are flushed to the stable log
50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
51 * requirement. In the event of a panic or power fail then those log
52 * records (transactions) are replayed.
54 * There is one ZIL per file system. Its on-disk (pool) format consists
61 * A log record holds a system call transaction. Log blocks can
62 * hold many log records and the blocks are chained together.
63 * Each ZIL block contains a block pointer (blkptr_t) to the next
64 * ZIL block in the chain. The ZIL header points to the first
65 * block in the chain. Note there is not a fixed place in the pool
66 * to hold blocks. They are dynamically allocated and freed as
67 * needed from the blocks available. Figure X shows the ZIL structure:
71 * Disable intent logging replay. This global ZIL switch affects all pools.
73 int zil_replay_disable = 0;
74 SYSCTL_DECL(_vfs_zfs);
75 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
76 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
77 &zil_replay_disable, 0, "Disable intent logging replay");
80 * Tunable parameter for debugging or performance analysis. Setting
81 * zfs_nocacheflush will cause corruption on power loss if a volatile
82 * out-of-order write cache is enabled.
84 boolean_t zfs_nocacheflush = B_FALSE;
85 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
86 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
87 &zfs_nocacheflush, 0, "Disable cache flush");
88 boolean_t zfs_trim_enabled = B_TRUE;
89 SYSCTL_DECL(_vfs_zfs_trim);
90 TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled);
91 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
94 static kmem_cache_t *zil_lwb_cache;
96 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
98 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
99 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
103 * ziltest is by and large an ugly hack, but very useful in
104 * checking replay without tedious work.
105 * When running ziltest we want to keep all itx's and so maintain
106 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
107 * We subtract TXG_CONCURRENT_STATES to allow for common code.
109 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
112 zil_bp_compare(const void *x1, const void *x2)
114 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
115 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
117 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
119 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
122 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
124 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
131 zil_bp_tree_init(zilog_t *zilog)
133 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
134 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
138 zil_bp_tree_fini(zilog_t *zilog)
140 avl_tree_t *t = &zilog->zl_bp_tree;
144 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
145 kmem_free(zn, sizeof (zil_bp_node_t));
151 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
153 avl_tree_t *t = &zilog->zl_bp_tree;
158 if (BP_IS_EMBEDDED(bp))
161 dva = BP_IDENTITY(bp);
163 if (avl_find(t, dva, &where) != NULL)
164 return (SET_ERROR(EEXIST));
166 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
168 avl_insert(t, zn, where);
173 static zil_header_t *
174 zil_header_in_syncing_context(zilog_t *zilog)
176 return ((zil_header_t *)zilog->zl_header);
180 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
182 zio_cksum_t *zc = &bp->blk_cksum;
184 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
185 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
186 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
187 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
191 * Read a log block and make sure it's valid.
194 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
197 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
198 arc_flags_t aflags = ARC_FLAG_WAIT;
199 arc_buf_t *abuf = NULL;
203 if (zilog->zl_header->zh_claim_txg == 0)
204 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
206 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
207 zio_flags |= ZIO_FLAG_SPECULATIVE;
209 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
210 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
212 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
213 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
216 zio_cksum_t cksum = bp->blk_cksum;
219 * Validate the checksummed log block.
221 * Sequence numbers should be... sequential. The checksum
222 * verifier for the next block should be bp's checksum plus 1.
224 * Also check the log chain linkage and size used.
226 cksum.zc_word[ZIL_ZC_SEQ]++;
228 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
229 zil_chain_t *zilc = abuf->b_data;
230 char *lr = (char *)(zilc + 1);
231 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
233 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
234 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
235 error = SET_ERROR(ECKSUM);
237 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
239 *end = (char *)dst + len;
240 *nbp = zilc->zc_next_blk;
243 char *lr = abuf->b_data;
244 uint64_t size = BP_GET_LSIZE(bp);
245 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
247 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
248 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
249 (zilc->zc_nused > (size - sizeof (*zilc)))) {
250 error = SET_ERROR(ECKSUM);
252 ASSERT3U(zilc->zc_nused, <=,
253 SPA_OLD_MAXBLOCKSIZE);
254 bcopy(lr, dst, zilc->zc_nused);
255 *end = (char *)dst + zilc->zc_nused;
256 *nbp = zilc->zc_next_blk;
260 arc_buf_destroy(abuf, &abuf);
267 * Read a TX_WRITE log data block.
270 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
272 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
273 const blkptr_t *bp = &lr->lr_blkptr;
274 arc_flags_t aflags = ARC_FLAG_WAIT;
275 arc_buf_t *abuf = NULL;
279 if (BP_IS_HOLE(bp)) {
281 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
285 if (zilog->zl_header->zh_claim_txg == 0)
286 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
288 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
289 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
291 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
292 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
296 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
297 arc_buf_destroy(abuf, &abuf);
304 * Parse the intent log, and call parse_func for each valid record within.
307 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
308 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
310 const zil_header_t *zh = zilog->zl_header;
311 boolean_t claimed = !!zh->zh_claim_txg;
312 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
313 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
314 uint64_t max_blk_seq = 0;
315 uint64_t max_lr_seq = 0;
316 uint64_t blk_count = 0;
317 uint64_t lr_count = 0;
318 blkptr_t blk, next_blk;
323 * Old logs didn't record the maximum zh_claim_lr_seq.
325 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
326 claim_lr_seq = UINT64_MAX;
329 * Starting at the block pointed to by zh_log we read the log chain.
330 * For each block in the chain we strongly check that block to
331 * ensure its validity. We stop when an invalid block is found.
332 * For each block pointer in the chain we call parse_blk_func().
333 * For each record in each valid block we call parse_lr_func().
334 * If the log has been claimed, stop if we encounter a sequence
335 * number greater than the highest claimed sequence number.
337 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
338 zil_bp_tree_init(zilog);
340 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
341 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
345 if (blk_seq > claim_blk_seq)
347 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
349 ASSERT3U(max_blk_seq, <, blk_seq);
350 max_blk_seq = blk_seq;
353 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
356 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
360 for (lrp = lrbuf; lrp < end; lrp += reclen) {
361 lr_t *lr = (lr_t *)lrp;
362 reclen = lr->lrc_reclen;
363 ASSERT3U(reclen, >=, sizeof (lr_t));
364 if (lr->lrc_seq > claim_lr_seq)
366 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
368 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
369 max_lr_seq = lr->lrc_seq;
374 zilog->zl_parse_error = error;
375 zilog->zl_parse_blk_seq = max_blk_seq;
376 zilog->zl_parse_lr_seq = max_lr_seq;
377 zilog->zl_parse_blk_count = blk_count;
378 zilog->zl_parse_lr_count = lr_count;
380 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
381 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
383 zil_bp_tree_fini(zilog);
384 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
390 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
393 * Claim log block if not already committed and not already claimed.
394 * If tx == NULL, just verify that the block is claimable.
396 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
397 zil_bp_tree_add(zilog, bp) != 0)
400 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
401 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
402 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
406 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
408 lr_write_t *lr = (lr_write_t *)lrc;
411 if (lrc->lrc_txtype != TX_WRITE)
415 * If the block is not readable, don't claim it. This can happen
416 * in normal operation when a log block is written to disk before
417 * some of the dmu_sync() blocks it points to. In this case, the
418 * transaction cannot have been committed to anyone (we would have
419 * waited for all writes to be stable first), so it is semantically
420 * correct to declare this the end of the log.
422 if (lr->lr_blkptr.blk_birth >= first_txg &&
423 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
425 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
430 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
432 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
438 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
440 lr_write_t *lr = (lr_write_t *)lrc;
441 blkptr_t *bp = &lr->lr_blkptr;
444 * If we previously claimed it, we need to free it.
446 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
447 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
449 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
455 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
459 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
460 lwb->lwb_zilog = zilog;
462 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
463 lwb->lwb_max_txg = txg;
466 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
467 lwb->lwb_nused = sizeof (zil_chain_t);
468 lwb->lwb_sz = BP_GET_LSIZE(bp);
471 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
474 mutex_enter(&zilog->zl_lock);
475 list_insert_tail(&zilog->zl_lwb_list, lwb);
476 mutex_exit(&zilog->zl_lock);
482 * Called when we create in-memory log transactions so that we know
483 * to cleanup the itxs at the end of spa_sync().
486 zilog_dirty(zilog_t *zilog, uint64_t txg)
488 dsl_pool_t *dp = zilog->zl_dmu_pool;
489 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
491 if (ds->ds_is_snapshot)
492 panic("dirtying snapshot!");
494 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
495 /* up the hold count until we can be written out */
496 dmu_buf_add_ref(ds->ds_dbuf, zilog);
501 zilog_is_dirty(zilog_t *zilog)
503 dsl_pool_t *dp = zilog->zl_dmu_pool;
505 for (int t = 0; t < TXG_SIZE; t++) {
506 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
513 * Create an on-disk intent log.
516 zil_create(zilog_t *zilog)
518 const zil_header_t *zh = zilog->zl_header;
526 * Wait for any previous destroy to complete.
528 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
530 ASSERT(zh->zh_claim_txg == 0);
531 ASSERT(zh->zh_replay_seq == 0);
536 * Allocate an initial log block if:
537 * - there isn't one already
538 * - the existing block is the wrong endianess
540 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
541 tx = dmu_tx_create(zilog->zl_os);
542 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
543 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
544 txg = dmu_tx_get_txg(tx);
546 if (!BP_IS_HOLE(&blk)) {
547 zio_free_zil(zilog->zl_spa, txg, &blk);
551 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
552 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
555 zil_init_log_chain(zilog, &blk);
559 * Allocate a log write buffer (lwb) for the first log block.
562 lwb = zil_alloc_lwb(zilog, &blk, txg);
565 * If we just allocated the first log block, commit our transaction
566 * and wait for zil_sync() to stuff the block poiner into zh_log.
567 * (zh is part of the MOS, so we cannot modify it in open context.)
571 txg_wait_synced(zilog->zl_dmu_pool, txg);
574 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
580 * In one tx, free all log blocks and clear the log header.
581 * If keep_first is set, then we're replaying a log with no content.
582 * We want to keep the first block, however, so that the first
583 * synchronous transaction doesn't require a txg_wait_synced()
584 * in zil_create(). We don't need to txg_wait_synced() here either
585 * when keep_first is set, because both zil_create() and zil_destroy()
586 * will wait for any in-progress destroys to complete.
589 zil_destroy(zilog_t *zilog, boolean_t keep_first)
591 const zil_header_t *zh = zilog->zl_header;
597 * Wait for any previous destroy to complete.
599 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
601 zilog->zl_old_header = *zh; /* debugging aid */
603 if (BP_IS_HOLE(&zh->zh_log))
606 tx = dmu_tx_create(zilog->zl_os);
607 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
608 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
609 txg = dmu_tx_get_txg(tx);
611 mutex_enter(&zilog->zl_lock);
613 ASSERT3U(zilog->zl_destroy_txg, <, txg);
614 zilog->zl_destroy_txg = txg;
615 zilog->zl_keep_first = keep_first;
617 if (!list_is_empty(&zilog->zl_lwb_list)) {
618 ASSERT(zh->zh_claim_txg == 0);
620 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
621 list_remove(&zilog->zl_lwb_list, lwb);
622 if (lwb->lwb_buf != NULL)
623 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
624 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
625 kmem_cache_free(zil_lwb_cache, lwb);
627 } else if (!keep_first) {
628 zil_destroy_sync(zilog, tx);
630 mutex_exit(&zilog->zl_lock);
636 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
638 ASSERT(list_is_empty(&zilog->zl_lwb_list));
639 (void) zil_parse(zilog, zil_free_log_block,
640 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
644 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
646 dmu_tx_t *tx = txarg;
647 uint64_t first_txg = dmu_tx_get_txg(tx);
653 error = dmu_objset_own_obj(dp, ds->ds_object,
654 DMU_OST_ANY, B_FALSE, FTAG, &os);
657 * EBUSY indicates that the objset is inconsistent, in which
658 * case it can not have a ZIL.
660 if (error != EBUSY) {
661 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
662 (unsigned long long)ds->ds_object, error);
667 zilog = dmu_objset_zil(os);
668 zh = zil_header_in_syncing_context(zilog);
670 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
671 if (!BP_IS_HOLE(&zh->zh_log))
672 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
673 BP_ZERO(&zh->zh_log);
674 dsl_dataset_dirty(dmu_objset_ds(os), tx);
675 dmu_objset_disown(os, FTAG);
680 * Claim all log blocks if we haven't already done so, and remember
681 * the highest claimed sequence number. This ensures that if we can
682 * read only part of the log now (e.g. due to a missing device),
683 * but we can read the entire log later, we will not try to replay
684 * or destroy beyond the last block we successfully claimed.
686 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
687 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
688 (void) zil_parse(zilog, zil_claim_log_block,
689 zil_claim_log_record, tx, first_txg);
690 zh->zh_claim_txg = first_txg;
691 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
692 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
693 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
694 zh->zh_flags |= ZIL_REPLAY_NEEDED;
695 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
696 dsl_dataset_dirty(dmu_objset_ds(os), tx);
699 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
700 dmu_objset_disown(os, FTAG);
705 * Check the log by walking the log chain.
706 * Checksum errors are ok as they indicate the end of the chain.
707 * Any other error (no device or read failure) returns an error.
711 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
720 error = dmu_objset_from_ds(ds, &os);
722 cmn_err(CE_WARN, "can't open objset %llu, error %d",
723 (unsigned long long)ds->ds_object, error);
727 zilog = dmu_objset_zil(os);
728 bp = (blkptr_t *)&zilog->zl_header->zh_log;
731 * Check the first block and determine if it's on a log device
732 * which may have been removed or faulted prior to loading this
733 * pool. If so, there's no point in checking the rest of the log
734 * as its content should have already been synced to the pool.
736 if (!BP_IS_HOLE(bp)) {
738 boolean_t valid = B_TRUE;
740 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
741 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
742 if (vd->vdev_islog && vdev_is_dead(vd))
743 valid = vdev_log_state_valid(vd);
744 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
751 * Because tx == NULL, zil_claim_log_block() will not actually claim
752 * any blocks, but just determine whether it is possible to do so.
753 * In addition to checking the log chain, zil_claim_log_block()
754 * will invoke zio_claim() with a done func of spa_claim_notify(),
755 * which will update spa_max_claim_txg. See spa_load() for details.
757 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
758 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
760 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
764 zil_vdev_compare(const void *x1, const void *x2)
766 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
767 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
778 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
780 avl_tree_t *t = &zilog->zl_vdev_tree;
782 zil_vdev_node_t *zv, zvsearch;
783 int ndvas = BP_GET_NDVAS(bp);
786 if (zfs_nocacheflush)
789 ASSERT(zilog->zl_writer);
792 * Even though we're zl_writer, we still need a lock because the
793 * zl_get_data() callbacks may have dmu_sync() done callbacks
794 * that will run concurrently.
796 mutex_enter(&zilog->zl_vdev_lock);
797 for (i = 0; i < ndvas; i++) {
798 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
799 if (avl_find(t, &zvsearch, &where) == NULL) {
800 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
801 zv->zv_vdev = zvsearch.zv_vdev;
802 avl_insert(t, zv, where);
805 mutex_exit(&zilog->zl_vdev_lock);
809 zil_flush_vdevs(zilog_t *zilog)
811 spa_t *spa = zilog->zl_spa;
812 avl_tree_t *t = &zilog->zl_vdev_tree;
817 ASSERT(zilog->zl_writer);
820 * We don't need zl_vdev_lock here because we're the zl_writer,
821 * and all zl_get_data() callbacks are done.
823 if (avl_numnodes(t) == 0)
826 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
828 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
830 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
831 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
834 kmem_free(zv, sizeof (*zv));
838 * Wait for all the flushes to complete. Not all devices actually
839 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
841 (void) zio_wait(zio);
843 spa_config_exit(spa, SCL_STATE, FTAG);
847 * Function called when a log block write completes
850 zil_lwb_write_done(zio_t *zio)
852 lwb_t *lwb = zio->io_private;
853 zilog_t *zilog = lwb->lwb_zilog;
854 dmu_tx_t *tx = lwb->lwb_tx;
856 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
857 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
858 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
859 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
860 ASSERT(!BP_IS_GANG(zio->io_bp));
861 ASSERT(!BP_IS_HOLE(zio->io_bp));
862 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
865 * Ensure the lwb buffer pointer is cleared before releasing
866 * the txg. If we have had an allocation failure and
867 * the txg is waiting to sync then we want want zil_sync()
868 * to remove the lwb so that it's not picked up as the next new
869 * one in zil_commit_writer(). zil_sync() will only remove
870 * the lwb if lwb_buf is null.
872 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
873 mutex_enter(&zilog->zl_lock);
876 mutex_exit(&zilog->zl_lock);
879 * Now that we've written this log block, we have a stable pointer
880 * to the next block in the chain, so it's OK to let the txg in
881 * which we allocated the next block sync.
887 * Initialize the io for a log block.
890 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
894 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
895 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
896 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
898 if (zilog->zl_root_zio == NULL) {
899 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
902 if (lwb->lwb_zio == NULL) {
903 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
904 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
905 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
906 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
911 * Define a limited set of intent log block sizes.
913 * These must be a multiple of 4KB. Note only the amount used (again
914 * aligned to 4KB) actually gets written. However, we can't always just
915 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
917 uint64_t zil_block_buckets[] = {
918 4096, /* non TX_WRITE */
919 8192+4096, /* data base */
920 32*1024 + 4096, /* NFS writes */
925 * Use the slog as long as the logbias is 'latency' and the current commit size
926 * is less than the limit or the total list size is less than 2X the limit.
927 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
929 uint64_t zil_slog_limit = 1024 * 1024;
930 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
931 (((zilog)->zl_cur_used < zil_slog_limit) || \
932 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
935 * Start a log block write and advance to the next log block.
936 * Calls are serialized.
939 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
943 spa_t *spa = zilog->zl_spa;
947 uint64_t zil_blksz, wsz;
950 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
951 zilc = (zil_chain_t *)lwb->lwb_buf;
952 bp = &zilc->zc_next_blk;
954 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
955 bp = &zilc->zc_next_blk;
958 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
961 * Allocate the next block and save its address in this block
962 * before writing it in order to establish the log chain.
963 * Note that if the allocation of nlwb synced before we wrote
964 * the block that points at it (lwb), we'd leak it if we crashed.
965 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
966 * We dirty the dataset to ensure that zil_sync() will be called
967 * to clean up in the event of allocation failure or I/O failure.
969 tx = dmu_tx_create(zilog->zl_os);
970 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
971 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
972 txg = dmu_tx_get_txg(tx);
977 * Log blocks are pre-allocated. Here we select the size of the next
978 * block, based on size used in the last block.
979 * - first find the smallest bucket that will fit the block from a
980 * limited set of block sizes. This is because it's faster to write
981 * blocks allocated from the same metaslab as they are adjacent or
983 * - next find the maximum from the new suggested size and an array of
984 * previous sizes. This lessens a picket fence effect of wrongly
985 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
988 * Note we only write what is used, but we can't just allocate
989 * the maximum block size because we can exhaust the available
992 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
993 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
995 zil_blksz = zil_block_buckets[i];
996 if (zil_blksz == UINT64_MAX)
997 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
998 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
999 for (i = 0; i < ZIL_PREV_BLKS; i++)
1000 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1001 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1004 /* pass the old blkptr in order to spread log blocks across devs */
1005 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1008 ASSERT3U(bp->blk_birth, ==, txg);
1009 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1010 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1013 * Allocate a new log write buffer (lwb).
1015 nlwb = zil_alloc_lwb(zilog, bp, txg);
1017 /* Record the block for later vdev flushing */
1018 zil_add_block(zilog, &lwb->lwb_blk);
1021 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1022 /* For Slim ZIL only write what is used. */
1023 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1024 ASSERT3U(wsz, <=, lwb->lwb_sz);
1025 zio_shrink(lwb->lwb_zio, wsz);
1032 zilc->zc_nused = lwb->lwb_nused;
1033 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1036 * clear unused data for security
1038 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1040 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1043 * If there was an allocation failure then nlwb will be null which
1044 * forces a txg_wait_synced().
1050 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1052 lr_t *lrc = &itx->itx_lr; /* common log record */
1053 lr_write_t *lrw = (lr_write_t *)lrc;
1055 uint64_t txg = lrc->lrc_txg;
1056 uint64_t reclen = lrc->lrc_reclen;
1062 ASSERT(lwb->lwb_buf != NULL);
1063 ASSERT(zilog_is_dirty(zilog) ||
1064 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1066 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1067 dlen = P2ROUNDUP_TYPED(
1068 lrw->lr_length, sizeof (uint64_t), uint64_t);
1070 zilog->zl_cur_used += (reclen + dlen);
1072 zil_lwb_write_init(zilog, lwb);
1075 * If this record won't fit in the current log block, start a new one.
1077 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1078 lwb = zil_lwb_write_start(zilog, lwb);
1081 zil_lwb_write_init(zilog, lwb);
1082 ASSERT(LWB_EMPTY(lwb));
1083 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1084 txg_wait_synced(zilog->zl_dmu_pool, txg);
1089 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1090 bcopy(lrc, lr_buf, reclen);
1091 lrc = (lr_t *)lr_buf;
1092 lrw = (lr_write_t *)lrc;
1095 * If it's a write, fetch the data or get its blkptr as appropriate.
1097 if (lrc->lrc_txtype == TX_WRITE) {
1098 if (txg > spa_freeze_txg(zilog->zl_spa))
1099 txg_wait_synced(zilog->zl_dmu_pool, txg);
1100 if (itx->itx_wr_state != WR_COPIED) {
1105 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1106 dbuf = lr_buf + reclen;
1107 lrw->lr_common.lrc_reclen += dlen;
1109 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1112 error = zilog->zl_get_data(
1113 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1115 txg_wait_synced(zilog->zl_dmu_pool, txg);
1119 ASSERT(error == ENOENT || error == EEXIST ||
1127 * We're actually making an entry, so update lrc_seq to be the
1128 * log record sequence number. Note that this is generally not
1129 * equal to the itx sequence number because not all transactions
1130 * are synchronous, and sometimes spa_sync() gets there first.
1132 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1133 lwb->lwb_nused += reclen + dlen;
1134 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1135 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1136 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1142 zil_itx_create(uint64_t txtype, size_t lrsize)
1146 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1148 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1149 itx->itx_lr.lrc_txtype = txtype;
1150 itx->itx_lr.lrc_reclen = lrsize;
1151 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1152 itx->itx_lr.lrc_seq = 0; /* defensive */
1153 itx->itx_sync = B_TRUE; /* default is synchronous */
1159 zil_itx_destroy(itx_t *itx)
1161 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1165 * Free up the sync and async itxs. The itxs_t has already been detached
1166 * so no locks are needed.
1169 zil_itxg_clean(itxs_t *itxs)
1175 itx_async_node_t *ian;
1177 list = &itxs->i_sync_list;
1178 while ((itx = list_head(list)) != NULL) {
1179 list_remove(list, itx);
1180 kmem_free(itx, offsetof(itx_t, itx_lr) +
1181 itx->itx_lr.lrc_reclen);
1185 t = &itxs->i_async_tree;
1186 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1187 list = &ian->ia_list;
1188 while ((itx = list_head(list)) != NULL) {
1189 list_remove(list, itx);
1190 kmem_free(itx, offsetof(itx_t, itx_lr) +
1191 itx->itx_lr.lrc_reclen);
1194 kmem_free(ian, sizeof (itx_async_node_t));
1198 kmem_free(itxs, sizeof (itxs_t));
1202 zil_aitx_compare(const void *x1, const void *x2)
1204 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1205 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1216 * Remove all async itx with the given oid.
1219 zil_remove_async(zilog_t *zilog, uint64_t oid)
1222 itx_async_node_t *ian;
1229 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1231 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1234 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1236 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1237 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1239 mutex_enter(&itxg->itxg_lock);
1240 if (itxg->itxg_txg != txg) {
1241 mutex_exit(&itxg->itxg_lock);
1246 * Locate the object node and append its list.
1248 t = &itxg->itxg_itxs->i_async_tree;
1249 ian = avl_find(t, &oid, &where);
1251 list_move_tail(&clean_list, &ian->ia_list);
1252 mutex_exit(&itxg->itxg_lock);
1254 while ((itx = list_head(&clean_list)) != NULL) {
1255 list_remove(&clean_list, itx);
1256 kmem_free(itx, offsetof(itx_t, itx_lr) +
1257 itx->itx_lr.lrc_reclen);
1259 list_destroy(&clean_list);
1263 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1267 itxs_t *itxs, *clean = NULL;
1270 * Object ids can be re-instantiated in the next txg so
1271 * remove any async transactions to avoid future leaks.
1272 * This can happen if a fsync occurs on the re-instantiated
1273 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1274 * the new file data and flushes a write record for the old object.
1276 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1277 zil_remove_async(zilog, itx->itx_oid);
1280 * Ensure the data of a renamed file is committed before the rename.
1282 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1283 zil_async_to_sync(zilog, itx->itx_oid);
1285 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1288 txg = dmu_tx_get_txg(tx);
1290 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1291 mutex_enter(&itxg->itxg_lock);
1292 itxs = itxg->itxg_itxs;
1293 if (itxg->itxg_txg != txg) {
1296 * The zil_clean callback hasn't got around to cleaning
1297 * this itxg. Save the itxs for release below.
1298 * This should be rare.
1300 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1302 clean = itxg->itxg_itxs;
1304 ASSERT(itxg->itxg_sod == 0);
1305 itxg->itxg_txg = txg;
1306 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1308 list_create(&itxs->i_sync_list, sizeof (itx_t),
1309 offsetof(itx_t, itx_node));
1310 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1311 sizeof (itx_async_node_t),
1312 offsetof(itx_async_node_t, ia_node));
1314 if (itx->itx_sync) {
1315 list_insert_tail(&itxs->i_sync_list, itx);
1316 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1317 itxg->itxg_sod += itx->itx_sod;
1319 avl_tree_t *t = &itxs->i_async_tree;
1320 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1321 itx_async_node_t *ian;
1324 ian = avl_find(t, &foid, &where);
1326 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1327 list_create(&ian->ia_list, sizeof (itx_t),
1328 offsetof(itx_t, itx_node));
1329 ian->ia_foid = foid;
1330 avl_insert(t, ian, where);
1332 list_insert_tail(&ian->ia_list, itx);
1335 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1336 zilog_dirty(zilog, txg);
1337 mutex_exit(&itxg->itxg_lock);
1339 /* Release the old itxs now we've dropped the lock */
1341 zil_itxg_clean(clean);
1345 * If there are any in-memory intent log transactions which have now been
1346 * synced then start up a taskq to free them. We should only do this after we
1347 * have written out the uberblocks (i.e. txg has been comitted) so that
1348 * don't inadvertently clean out in-memory log records that would be required
1352 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1354 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1357 mutex_enter(&itxg->itxg_lock);
1358 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1359 mutex_exit(&itxg->itxg_lock);
1362 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1363 ASSERT(itxg->itxg_txg != 0);
1364 ASSERT(zilog->zl_clean_taskq != NULL);
1365 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1367 clean_me = itxg->itxg_itxs;
1368 itxg->itxg_itxs = NULL;
1370 mutex_exit(&itxg->itxg_lock);
1372 * Preferably start a task queue to free up the old itxs but
1373 * if taskq_dispatch can't allocate resources to do that then
1374 * free it in-line. This should be rare. Note, using TQ_SLEEP
1375 * created a bad performance problem.
1377 if (taskq_dispatch(zilog->zl_clean_taskq,
1378 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1379 zil_itxg_clean(clean_me);
1383 * Get the list of itxs to commit into zl_itx_commit_list.
1386 zil_get_commit_list(zilog_t *zilog)
1389 list_t *commit_list = &zilog->zl_itx_commit_list;
1390 uint64_t push_sod = 0;
1392 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1395 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1397 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1398 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1400 mutex_enter(&itxg->itxg_lock);
1401 if (itxg->itxg_txg != txg) {
1402 mutex_exit(&itxg->itxg_lock);
1406 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1407 push_sod += itxg->itxg_sod;
1410 mutex_exit(&itxg->itxg_lock);
1412 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1416 * Move the async itxs for a specified object to commit into sync lists.
1419 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1422 itx_async_node_t *ian;
1426 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1429 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1431 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1432 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1434 mutex_enter(&itxg->itxg_lock);
1435 if (itxg->itxg_txg != txg) {
1436 mutex_exit(&itxg->itxg_lock);
1441 * If a foid is specified then find that node and append its
1442 * list. Otherwise walk the tree appending all the lists
1443 * to the sync list. We add to the end rather than the
1444 * beginning to ensure the create has happened.
1446 t = &itxg->itxg_itxs->i_async_tree;
1448 ian = avl_find(t, &foid, &where);
1450 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1454 void *cookie = NULL;
1456 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1457 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1459 list_destroy(&ian->ia_list);
1460 kmem_free(ian, sizeof (itx_async_node_t));
1463 mutex_exit(&itxg->itxg_lock);
1468 zil_commit_writer(zilog_t *zilog)
1473 spa_t *spa = zilog->zl_spa;
1476 ASSERT(zilog->zl_root_zio == NULL);
1478 mutex_exit(&zilog->zl_lock);
1480 zil_get_commit_list(zilog);
1483 * Return if there's nothing to commit before we dirty the fs by
1484 * calling zil_create().
1486 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1487 mutex_enter(&zilog->zl_lock);
1491 if (zilog->zl_suspend) {
1494 lwb = list_tail(&zilog->zl_lwb_list);
1496 lwb = zil_create(zilog);
1499 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1500 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1501 txg = itx->itx_lr.lrc_txg;
1504 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1505 lwb = zil_lwb_commit(zilog, itx, lwb);
1506 list_remove(&zilog->zl_itx_commit_list, itx);
1507 kmem_free(itx, offsetof(itx_t, itx_lr)
1508 + itx->itx_lr.lrc_reclen);
1510 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1512 /* write the last block out */
1513 if (lwb != NULL && lwb->lwb_zio != NULL)
1514 lwb = zil_lwb_write_start(zilog, lwb);
1516 zilog->zl_cur_used = 0;
1519 * Wait if necessary for the log blocks to be on stable storage.
1521 if (zilog->zl_root_zio) {
1522 error = zio_wait(zilog->zl_root_zio);
1523 zilog->zl_root_zio = NULL;
1524 zil_flush_vdevs(zilog);
1527 if (error || lwb == NULL)
1528 txg_wait_synced(zilog->zl_dmu_pool, 0);
1530 mutex_enter(&zilog->zl_lock);
1533 * Remember the highest committed log sequence number for ztest.
1534 * We only update this value when all the log writes succeeded,
1535 * because ztest wants to ASSERT that it got the whole log chain.
1537 if (error == 0 && lwb != NULL)
1538 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1542 * Commit zfs transactions to stable storage.
1543 * If foid is 0 push out all transactions, otherwise push only those
1544 * for that object or might reference that object.
1546 * itxs are committed in batches. In a heavily stressed zil there will be
1547 * a commit writer thread who is writing out a bunch of itxs to the log
1548 * for a set of committing threads (cthreads) in the same batch as the writer.
1549 * Those cthreads are all waiting on the same cv for that batch.
1551 * There will also be a different and growing batch of threads that are
1552 * waiting to commit (qthreads). When the committing batch completes
1553 * a transition occurs such that the cthreads exit and the qthreads become
1554 * cthreads. One of the new cthreads becomes the writer thread for the
1555 * batch. Any new threads arriving become new qthreads.
1557 * Only 2 condition variables are needed and there's no transition
1558 * between the two cvs needed. They just flip-flop between qthreads
1561 * Using this scheme we can efficiently wakeup up only those threads
1562 * that have been committed.
1565 zil_commit(zilog_t *zilog, uint64_t foid)
1569 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1572 /* move the async itxs for the foid to the sync queues */
1573 zil_async_to_sync(zilog, foid);
1575 mutex_enter(&zilog->zl_lock);
1576 mybatch = zilog->zl_next_batch;
1577 while (zilog->zl_writer) {
1578 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1579 if (mybatch <= zilog->zl_com_batch) {
1580 mutex_exit(&zilog->zl_lock);
1585 zilog->zl_next_batch++;
1586 zilog->zl_writer = B_TRUE;
1587 zil_commit_writer(zilog);
1588 zilog->zl_com_batch = mybatch;
1589 zilog->zl_writer = B_FALSE;
1590 mutex_exit(&zilog->zl_lock);
1592 /* wake up one thread to become the next writer */
1593 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1595 /* wake up all threads waiting for this batch to be committed */
1596 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1600 * Called in syncing context to free committed log blocks and update log header.
1603 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1605 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1606 uint64_t txg = dmu_tx_get_txg(tx);
1607 spa_t *spa = zilog->zl_spa;
1608 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1612 * We don't zero out zl_destroy_txg, so make sure we don't try
1613 * to destroy it twice.
1615 if (spa_sync_pass(spa) != 1)
1618 mutex_enter(&zilog->zl_lock);
1620 ASSERT(zilog->zl_stop_sync == 0);
1622 if (*replayed_seq != 0) {
1623 ASSERT(zh->zh_replay_seq < *replayed_seq);
1624 zh->zh_replay_seq = *replayed_seq;
1628 if (zilog->zl_destroy_txg == txg) {
1629 blkptr_t blk = zh->zh_log;
1631 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1633 bzero(zh, sizeof (zil_header_t));
1634 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1636 if (zilog->zl_keep_first) {
1638 * If this block was part of log chain that couldn't
1639 * be claimed because a device was missing during
1640 * zil_claim(), but that device later returns,
1641 * then this block could erroneously appear valid.
1642 * To guard against this, assign a new GUID to the new
1643 * log chain so it doesn't matter what blk points to.
1645 zil_init_log_chain(zilog, &blk);
1650 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1651 zh->zh_log = lwb->lwb_blk;
1652 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1654 list_remove(&zilog->zl_lwb_list, lwb);
1655 zio_free_zil(spa, txg, &lwb->lwb_blk);
1656 kmem_cache_free(zil_lwb_cache, lwb);
1659 * If we don't have anything left in the lwb list then
1660 * we've had an allocation failure and we need to zero
1661 * out the zil_header blkptr so that we don't end
1662 * up freeing the same block twice.
1664 if (list_head(&zilog->zl_lwb_list) == NULL)
1665 BP_ZERO(&zh->zh_log);
1667 mutex_exit(&zilog->zl_lock);
1673 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1674 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1680 kmem_cache_destroy(zil_lwb_cache);
1684 zil_set_sync(zilog_t *zilog, uint64_t sync)
1686 zilog->zl_sync = sync;
1690 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1692 zilog->zl_logbias = logbias;
1696 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1700 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1702 zilog->zl_header = zh_phys;
1704 zilog->zl_spa = dmu_objset_spa(os);
1705 zilog->zl_dmu_pool = dmu_objset_pool(os);
1706 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1707 zilog->zl_logbias = dmu_objset_logbias(os);
1708 zilog->zl_sync = dmu_objset_syncprop(os);
1709 zilog->zl_next_batch = 1;
1711 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1713 for (int i = 0; i < TXG_SIZE; i++) {
1714 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1715 MUTEX_DEFAULT, NULL);
1718 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1719 offsetof(lwb_t, lwb_node));
1721 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1722 offsetof(itx_t, itx_node));
1724 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1726 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1727 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1729 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1730 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1731 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1732 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1738 zil_free(zilog_t *zilog)
1740 zilog->zl_stop_sync = 1;
1742 ASSERT0(zilog->zl_suspend);
1743 ASSERT0(zilog->zl_suspending);
1745 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1746 list_destroy(&zilog->zl_lwb_list);
1748 avl_destroy(&zilog->zl_vdev_tree);
1749 mutex_destroy(&zilog->zl_vdev_lock);
1751 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1752 list_destroy(&zilog->zl_itx_commit_list);
1754 for (int i = 0; i < TXG_SIZE; i++) {
1756 * It's possible for an itx to be generated that doesn't dirty
1757 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1758 * callback to remove the entry. We remove those here.
1760 * Also free up the ziltest itxs.
1762 if (zilog->zl_itxg[i].itxg_itxs)
1763 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1764 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1767 mutex_destroy(&zilog->zl_lock);
1769 cv_destroy(&zilog->zl_cv_writer);
1770 cv_destroy(&zilog->zl_cv_suspend);
1771 cv_destroy(&zilog->zl_cv_batch[0]);
1772 cv_destroy(&zilog->zl_cv_batch[1]);
1774 kmem_free(zilog, sizeof (zilog_t));
1778 * Open an intent log.
1781 zil_open(objset_t *os, zil_get_data_t *get_data)
1783 zilog_t *zilog = dmu_objset_zil(os);
1785 ASSERT(zilog->zl_clean_taskq == NULL);
1786 ASSERT(zilog->zl_get_data == NULL);
1787 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1789 zilog->zl_get_data = get_data;
1790 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1791 2, 2, TASKQ_PREPOPULATE);
1797 * Close an intent log.
1800 zil_close(zilog_t *zilog)
1805 zil_commit(zilog, 0); /* commit all itx */
1808 * The lwb_max_txg for the stubby lwb will reflect the last activity
1809 * for the zil. After a txg_wait_synced() on the txg we know all the
1810 * callbacks have occurred that may clean the zil. Only then can we
1811 * destroy the zl_clean_taskq.
1813 mutex_enter(&zilog->zl_lock);
1814 lwb = list_tail(&zilog->zl_lwb_list);
1816 txg = lwb->lwb_max_txg;
1817 mutex_exit(&zilog->zl_lock);
1819 txg_wait_synced(zilog->zl_dmu_pool, txg);
1820 ASSERT(!zilog_is_dirty(zilog));
1822 taskq_destroy(zilog->zl_clean_taskq);
1823 zilog->zl_clean_taskq = NULL;
1824 zilog->zl_get_data = NULL;
1827 * We should have only one LWB left on the list; remove it now.
1829 mutex_enter(&zilog->zl_lock);
1830 lwb = list_head(&zilog->zl_lwb_list);
1832 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1833 list_remove(&zilog->zl_lwb_list, lwb);
1834 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1835 kmem_cache_free(zil_lwb_cache, lwb);
1837 mutex_exit(&zilog->zl_lock);
1840 static char *suspend_tag = "zil suspending";
1843 * Suspend an intent log. While in suspended mode, we still honor
1844 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1845 * On old version pools, we suspend the log briefly when taking a
1846 * snapshot so that it will have an empty intent log.
1848 * Long holds are not really intended to be used the way we do here --
1849 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1850 * could fail. Therefore we take pains to only put a long hold if it is
1851 * actually necessary. Fortunately, it will only be necessary if the
1852 * objset is currently mounted (or the ZVOL equivalent). In that case it
1853 * will already have a long hold, so we are not really making things any worse.
1855 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1856 * zvol_state_t), and use their mechanism to prevent their hold from being
1857 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1860 * if cookiep == NULL, this does both the suspend & resume.
1861 * Otherwise, it returns with the dataset "long held", and the cookie
1862 * should be passed into zil_resume().
1865 zil_suspend(const char *osname, void **cookiep)
1869 const zil_header_t *zh;
1872 error = dmu_objset_hold(osname, suspend_tag, &os);
1875 zilog = dmu_objset_zil(os);
1877 mutex_enter(&zilog->zl_lock);
1878 zh = zilog->zl_header;
1880 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1881 mutex_exit(&zilog->zl_lock);
1882 dmu_objset_rele(os, suspend_tag);
1883 return (SET_ERROR(EBUSY));
1887 * Don't put a long hold in the cases where we can avoid it. This
1888 * is when there is no cookie so we are doing a suspend & resume
1889 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1890 * for the suspend because it's already suspended, or there's no ZIL.
1892 if (cookiep == NULL && !zilog->zl_suspending &&
1893 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1894 mutex_exit(&zilog->zl_lock);
1895 dmu_objset_rele(os, suspend_tag);
1899 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1900 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1902 zilog->zl_suspend++;
1904 if (zilog->zl_suspend > 1) {
1906 * Someone else is already suspending it.
1907 * Just wait for them to finish.
1910 while (zilog->zl_suspending)
1911 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1912 mutex_exit(&zilog->zl_lock);
1914 if (cookiep == NULL)
1922 * If there is no pointer to an on-disk block, this ZIL must not
1923 * be active (e.g. filesystem not mounted), so there's nothing
1926 if (BP_IS_HOLE(&zh->zh_log)) {
1927 ASSERT(cookiep != NULL); /* fast path already handled */
1930 mutex_exit(&zilog->zl_lock);
1934 zilog->zl_suspending = B_TRUE;
1935 mutex_exit(&zilog->zl_lock);
1937 zil_commit(zilog, 0);
1939 zil_destroy(zilog, B_FALSE);
1941 mutex_enter(&zilog->zl_lock);
1942 zilog->zl_suspending = B_FALSE;
1943 cv_broadcast(&zilog->zl_cv_suspend);
1944 mutex_exit(&zilog->zl_lock);
1946 if (cookiep == NULL)
1954 zil_resume(void *cookie)
1956 objset_t *os = cookie;
1957 zilog_t *zilog = dmu_objset_zil(os);
1959 mutex_enter(&zilog->zl_lock);
1960 ASSERT(zilog->zl_suspend != 0);
1961 zilog->zl_suspend--;
1962 mutex_exit(&zilog->zl_lock);
1963 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1964 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1967 typedef struct zil_replay_arg {
1968 zil_replay_func_t **zr_replay;
1970 boolean_t zr_byteswap;
1975 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1977 char name[ZFS_MAX_DATASET_NAME_LEN];
1979 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1981 dmu_objset_name(zilog->zl_os, name);
1983 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1984 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1985 (u_longlong_t)lr->lrc_seq,
1986 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1987 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1993 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1995 zil_replay_arg_t *zr = zra;
1996 const zil_header_t *zh = zilog->zl_header;
1997 uint64_t reclen = lr->lrc_reclen;
1998 uint64_t txtype = lr->lrc_txtype;
2001 zilog->zl_replaying_seq = lr->lrc_seq;
2003 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2006 if (lr->lrc_txg < claim_txg) /* already committed */
2009 /* Strip case-insensitive bit, still present in log record */
2012 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2013 return (zil_replay_error(zilog, lr, EINVAL));
2016 * If this record type can be logged out of order, the object
2017 * (lr_foid) may no longer exist. That's legitimate, not an error.
2019 if (TX_OOO(txtype)) {
2020 error = dmu_object_info(zilog->zl_os,
2021 ((lr_ooo_t *)lr)->lr_foid, NULL);
2022 if (error == ENOENT || error == EEXIST)
2027 * Make a copy of the data so we can revise and extend it.
2029 bcopy(lr, zr->zr_lr, reclen);
2032 * If this is a TX_WRITE with a blkptr, suck in the data.
2034 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2035 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2036 zr->zr_lr + reclen);
2038 return (zil_replay_error(zilog, lr, error));
2042 * The log block containing this lr may have been byteswapped
2043 * so that we can easily examine common fields like lrc_txtype.
2044 * However, the log is a mix of different record types, and only the
2045 * replay vectors know how to byteswap their records. Therefore, if
2046 * the lr was byteswapped, undo it before invoking the replay vector.
2048 if (zr->zr_byteswap)
2049 byteswap_uint64_array(zr->zr_lr, reclen);
2052 * We must now do two things atomically: replay this log record,
2053 * and update the log header sequence number to reflect the fact that
2054 * we did so. At the end of each replay function the sequence number
2055 * is updated if we are in replay mode.
2057 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2060 * The DMU's dnode layer doesn't see removes until the txg
2061 * commits, so a subsequent claim can spuriously fail with
2062 * EEXIST. So if we receive any error we try syncing out
2063 * any removes then retry the transaction. Note that we
2064 * specify B_FALSE for byteswap now, so we don't do it twice.
2066 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2067 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2069 return (zil_replay_error(zilog, lr, error));
2076 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2078 zilog->zl_replay_blks++;
2084 * If this dataset has a non-empty intent log, replay it and destroy it.
2087 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2089 zilog_t *zilog = dmu_objset_zil(os);
2090 const zil_header_t *zh = zilog->zl_header;
2091 zil_replay_arg_t zr;
2093 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2094 zil_destroy(zilog, B_TRUE);
2098 zr.zr_replay = replay_func;
2100 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2101 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2104 * Wait for in-progress removes to sync before starting replay.
2106 txg_wait_synced(zilog->zl_dmu_pool, 0);
2108 zilog->zl_replay = B_TRUE;
2109 zilog->zl_replay_time = ddi_get_lbolt();
2110 ASSERT(zilog->zl_replay_blks == 0);
2111 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2113 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2115 zil_destroy(zilog, B_FALSE);
2116 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2117 zilog->zl_replay = B_FALSE;
2121 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2123 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2126 if (zilog->zl_replay) {
2127 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2128 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2129 zilog->zl_replaying_seq;
2138 zil_vdev_offline(const char *osname, void *arg)
2142 error = zil_suspend(osname, NULL);
2144 return (SET_ERROR(EEXIST));