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, 2014 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
52 * There is one ZIL per file system. Its on-disk (pool) format consists
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
69 * Disable intent logging replay. This global ZIL switch affects all pools.
71 int zil_replay_disable = 0;
72 SYSCTL_DECL(_vfs_zfs);
73 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
74 &zil_replay_disable, 0, "Disable intent logging replay");
77 * Tunable parameter for debugging or performance analysis. Setting
78 * zfs_nocacheflush will cause corruption on power loss if a volatile
79 * out-of-order write cache is enabled.
81 boolean_t zfs_nocacheflush = B_FALSE;
82 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
83 &zfs_nocacheflush, 0, "Disable cache flush");
84 boolean_t zfs_trim_enabled = B_TRUE;
85 SYSCTL_DECL(_vfs_zfs_trim);
86 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
89 static kmem_cache_t *zil_lwb_cache;
91 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
93 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
94 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
98 * ziltest is by and large an ugly hack, but very useful in
99 * checking replay without tedious work.
100 * When running ziltest we want to keep all itx's and so maintain
101 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
102 * We subtract TXG_CONCURRENT_STATES to allow for common code.
104 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
107 zil_bp_compare(const void *x1, const void *x2)
109 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
110 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
112 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
114 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
117 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
119 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
126 zil_bp_tree_init(zilog_t *zilog)
128 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
129 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
133 zil_bp_tree_fini(zilog_t *zilog)
135 avl_tree_t *t = &zilog->zl_bp_tree;
139 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
140 kmem_free(zn, sizeof (zil_bp_node_t));
146 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
148 avl_tree_t *t = &zilog->zl_bp_tree;
153 if (BP_IS_EMBEDDED(bp))
156 dva = BP_IDENTITY(bp);
158 if (avl_find(t, dva, &where) != NULL)
159 return (SET_ERROR(EEXIST));
161 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
163 avl_insert(t, zn, where);
168 static zil_header_t *
169 zil_header_in_syncing_context(zilog_t *zilog)
171 return ((zil_header_t *)zilog->zl_header);
175 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
177 zio_cksum_t *zc = &bp->blk_cksum;
179 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
180 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
181 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
182 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
186 * Read a log block and make sure it's valid.
189 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
192 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
193 uint32_t aflags = ARC_WAIT;
194 arc_buf_t *abuf = NULL;
198 if (zilog->zl_header->zh_claim_txg == 0)
199 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
201 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
202 zio_flags |= ZIO_FLAG_SPECULATIVE;
204 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
205 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
207 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
208 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
211 zio_cksum_t cksum = bp->blk_cksum;
214 * Validate the checksummed log block.
216 * Sequence numbers should be... sequential. The checksum
217 * verifier for the next block should be bp's checksum plus 1.
219 * Also check the log chain linkage and size used.
221 cksum.zc_word[ZIL_ZC_SEQ]++;
223 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
224 zil_chain_t *zilc = abuf->b_data;
225 char *lr = (char *)(zilc + 1);
226 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
228 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
229 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
230 error = SET_ERROR(ECKSUM);
233 *end = (char *)dst + len;
234 *nbp = zilc->zc_next_blk;
237 char *lr = abuf->b_data;
238 uint64_t size = BP_GET_LSIZE(bp);
239 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
241 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
242 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
243 (zilc->zc_nused > (size - sizeof (*zilc)))) {
244 error = SET_ERROR(ECKSUM);
246 bcopy(lr, dst, zilc->zc_nused);
247 *end = (char *)dst + zilc->zc_nused;
248 *nbp = zilc->zc_next_blk;
252 VERIFY(arc_buf_remove_ref(abuf, &abuf));
259 * Read a TX_WRITE log data block.
262 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
264 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
265 const blkptr_t *bp = &lr->lr_blkptr;
266 uint32_t aflags = ARC_WAIT;
267 arc_buf_t *abuf = NULL;
271 if (BP_IS_HOLE(bp)) {
273 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
277 if (zilog->zl_header->zh_claim_txg == 0)
278 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
280 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
281 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
283 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
284 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
288 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
289 (void) arc_buf_remove_ref(abuf, &abuf);
296 * Parse the intent log, and call parse_func for each valid record within.
299 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
300 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
302 const zil_header_t *zh = zilog->zl_header;
303 boolean_t claimed = !!zh->zh_claim_txg;
304 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
305 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
306 uint64_t max_blk_seq = 0;
307 uint64_t max_lr_seq = 0;
308 uint64_t blk_count = 0;
309 uint64_t lr_count = 0;
310 blkptr_t blk, next_blk;
315 * Old logs didn't record the maximum zh_claim_lr_seq.
317 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
318 claim_lr_seq = UINT64_MAX;
321 * Starting at the block pointed to by zh_log we read the log chain.
322 * For each block in the chain we strongly check that block to
323 * ensure its validity. We stop when an invalid block is found.
324 * For each block pointer in the chain we call parse_blk_func().
325 * For each record in each valid block we call parse_lr_func().
326 * If the log has been claimed, stop if we encounter a sequence
327 * number greater than the highest claimed sequence number.
329 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
330 zil_bp_tree_init(zilog);
332 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
333 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
337 if (blk_seq > claim_blk_seq)
339 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
341 ASSERT3U(max_blk_seq, <, blk_seq);
342 max_blk_seq = blk_seq;
345 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
348 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
352 for (lrp = lrbuf; lrp < end; lrp += reclen) {
353 lr_t *lr = (lr_t *)lrp;
354 reclen = lr->lrc_reclen;
355 ASSERT3U(reclen, >=, sizeof (lr_t));
356 if (lr->lrc_seq > claim_lr_seq)
358 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
360 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
361 max_lr_seq = lr->lrc_seq;
366 zilog->zl_parse_error = error;
367 zilog->zl_parse_blk_seq = max_blk_seq;
368 zilog->zl_parse_lr_seq = max_lr_seq;
369 zilog->zl_parse_blk_count = blk_count;
370 zilog->zl_parse_lr_count = lr_count;
372 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
373 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
375 zil_bp_tree_fini(zilog);
376 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
382 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
385 * Claim log block if not already committed and not already claimed.
386 * If tx == NULL, just verify that the block is claimable.
388 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
389 zil_bp_tree_add(zilog, bp) != 0)
392 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
393 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
394 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
398 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
400 lr_write_t *lr = (lr_write_t *)lrc;
403 if (lrc->lrc_txtype != TX_WRITE)
407 * If the block is not readable, don't claim it. This can happen
408 * in normal operation when a log block is written to disk before
409 * some of the dmu_sync() blocks it points to. In this case, the
410 * transaction cannot have been committed to anyone (we would have
411 * waited for all writes to be stable first), so it is semantically
412 * correct to declare this the end of the log.
414 if (lr->lr_blkptr.blk_birth >= first_txg &&
415 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
417 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
422 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
424 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
430 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
432 lr_write_t *lr = (lr_write_t *)lrc;
433 blkptr_t *bp = &lr->lr_blkptr;
436 * If we previously claimed it, we need to free it.
438 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
439 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
441 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
447 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
451 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
452 lwb->lwb_zilog = zilog;
454 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
455 lwb->lwb_max_txg = txg;
458 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
459 lwb->lwb_nused = sizeof (zil_chain_t);
460 lwb->lwb_sz = BP_GET_LSIZE(bp);
463 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
466 mutex_enter(&zilog->zl_lock);
467 list_insert_tail(&zilog->zl_lwb_list, lwb);
468 mutex_exit(&zilog->zl_lock);
474 * Called when we create in-memory log transactions so that we know
475 * to cleanup the itxs at the end of spa_sync().
478 zilog_dirty(zilog_t *zilog, uint64_t txg)
480 dsl_pool_t *dp = zilog->zl_dmu_pool;
481 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
483 if (dsl_dataset_is_snapshot(ds))
484 panic("dirtying snapshot!");
486 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
487 /* up the hold count until we can be written out */
488 dmu_buf_add_ref(ds->ds_dbuf, zilog);
493 zilog_is_dirty(zilog_t *zilog)
495 dsl_pool_t *dp = zilog->zl_dmu_pool;
497 for (int t = 0; t < TXG_SIZE; t++) {
498 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
505 * Create an on-disk intent log.
508 zil_create(zilog_t *zilog)
510 const zil_header_t *zh = zilog->zl_header;
518 * Wait for any previous destroy to complete.
520 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
522 ASSERT(zh->zh_claim_txg == 0);
523 ASSERT(zh->zh_replay_seq == 0);
528 * Allocate an initial log block if:
529 * - there isn't one already
530 * - the existing block is the wrong endianess
532 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
533 tx = dmu_tx_create(zilog->zl_os);
534 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
535 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
536 txg = dmu_tx_get_txg(tx);
538 if (!BP_IS_HOLE(&blk)) {
539 zio_free_zil(zilog->zl_spa, txg, &blk);
543 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
544 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
547 zil_init_log_chain(zilog, &blk);
551 * Allocate a log write buffer (lwb) for the first log block.
554 lwb = zil_alloc_lwb(zilog, &blk, txg);
557 * If we just allocated the first log block, commit our transaction
558 * and wait for zil_sync() to stuff the block poiner into zh_log.
559 * (zh is part of the MOS, so we cannot modify it in open context.)
563 txg_wait_synced(zilog->zl_dmu_pool, txg);
566 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
572 * In one tx, free all log blocks and clear the log header.
573 * If keep_first is set, then we're replaying a log with no content.
574 * We want to keep the first block, however, so that the first
575 * synchronous transaction doesn't require a txg_wait_synced()
576 * in zil_create(). We don't need to txg_wait_synced() here either
577 * when keep_first is set, because both zil_create() and zil_destroy()
578 * will wait for any in-progress destroys to complete.
581 zil_destroy(zilog_t *zilog, boolean_t keep_first)
583 const zil_header_t *zh = zilog->zl_header;
589 * Wait for any previous destroy to complete.
591 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
593 zilog->zl_old_header = *zh; /* debugging aid */
595 if (BP_IS_HOLE(&zh->zh_log))
598 tx = dmu_tx_create(zilog->zl_os);
599 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
600 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
601 txg = dmu_tx_get_txg(tx);
603 mutex_enter(&zilog->zl_lock);
605 ASSERT3U(zilog->zl_destroy_txg, <, txg);
606 zilog->zl_destroy_txg = txg;
607 zilog->zl_keep_first = keep_first;
609 if (!list_is_empty(&zilog->zl_lwb_list)) {
610 ASSERT(zh->zh_claim_txg == 0);
612 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
613 list_remove(&zilog->zl_lwb_list, lwb);
614 if (lwb->lwb_buf != NULL)
615 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
616 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
617 kmem_cache_free(zil_lwb_cache, lwb);
619 } else if (!keep_first) {
620 zil_destroy_sync(zilog, tx);
622 mutex_exit(&zilog->zl_lock);
628 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
630 ASSERT(list_is_empty(&zilog->zl_lwb_list));
631 (void) zil_parse(zilog, zil_free_log_block,
632 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
636 zil_claim(const char *osname, void *txarg)
638 dmu_tx_t *tx = txarg;
639 uint64_t first_txg = dmu_tx_get_txg(tx);
645 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
648 * EBUSY indicates that the objset is inconsistent, in which
649 * case it can not have a ZIL.
651 if (error != EBUSY) {
652 cmn_err(CE_WARN, "can't open objset for %s, error %u",
658 zilog = dmu_objset_zil(os);
659 zh = zil_header_in_syncing_context(zilog);
661 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
662 if (!BP_IS_HOLE(&zh->zh_log))
663 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
664 BP_ZERO(&zh->zh_log);
665 dsl_dataset_dirty(dmu_objset_ds(os), tx);
666 dmu_objset_disown(os, FTAG);
671 * Claim all log blocks if we haven't already done so, and remember
672 * the highest claimed sequence number. This ensures that if we can
673 * read only part of the log now (e.g. due to a missing device),
674 * but we can read the entire log later, we will not try to replay
675 * or destroy beyond the last block we successfully claimed.
677 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
678 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
679 (void) zil_parse(zilog, zil_claim_log_block,
680 zil_claim_log_record, tx, first_txg);
681 zh->zh_claim_txg = first_txg;
682 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
683 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
684 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
685 zh->zh_flags |= ZIL_REPLAY_NEEDED;
686 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
687 dsl_dataset_dirty(dmu_objset_ds(os), tx);
690 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
691 dmu_objset_disown(os, FTAG);
696 * Check the log by walking the log chain.
697 * Checksum errors are ok as they indicate the end of the chain.
698 * Any other error (no device or read failure) returns an error.
701 zil_check_log_chain(const char *osname, void *tx)
710 error = dmu_objset_hold(osname, FTAG, &os);
712 cmn_err(CE_WARN, "can't open objset for %s", osname);
716 zilog = dmu_objset_zil(os);
717 bp = (blkptr_t *)&zilog->zl_header->zh_log;
720 * Check the first block and determine if it's on a log device
721 * which may have been removed or faulted prior to loading this
722 * pool. If so, there's no point in checking the rest of the log
723 * as its content should have already been synced to the pool.
725 if (!BP_IS_HOLE(bp)) {
727 boolean_t valid = B_TRUE;
729 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
730 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
731 if (vd->vdev_islog && vdev_is_dead(vd))
732 valid = vdev_log_state_valid(vd);
733 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
736 dmu_objset_rele(os, FTAG);
742 * Because tx == NULL, zil_claim_log_block() will not actually claim
743 * any blocks, but just determine whether it is possible to do so.
744 * In addition to checking the log chain, zil_claim_log_block()
745 * will invoke zio_claim() with a done func of spa_claim_notify(),
746 * which will update spa_max_claim_txg. See spa_load() for details.
748 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
749 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
751 dmu_objset_rele(os, FTAG);
753 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
757 zil_vdev_compare(const void *x1, const void *x2)
759 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
760 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
771 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
773 avl_tree_t *t = &zilog->zl_vdev_tree;
775 zil_vdev_node_t *zv, zvsearch;
776 int ndvas = BP_GET_NDVAS(bp);
779 if (zfs_nocacheflush)
782 ASSERT(zilog->zl_writer);
785 * Even though we're zl_writer, we still need a lock because the
786 * zl_get_data() callbacks may have dmu_sync() done callbacks
787 * that will run concurrently.
789 mutex_enter(&zilog->zl_vdev_lock);
790 for (i = 0; i < ndvas; i++) {
791 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
792 if (avl_find(t, &zvsearch, &where) == NULL) {
793 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
794 zv->zv_vdev = zvsearch.zv_vdev;
795 avl_insert(t, zv, where);
798 mutex_exit(&zilog->zl_vdev_lock);
802 zil_flush_vdevs(zilog_t *zilog)
804 spa_t *spa = zilog->zl_spa;
805 avl_tree_t *t = &zilog->zl_vdev_tree;
810 ASSERT(zilog->zl_writer);
813 * We don't need zl_vdev_lock here because we're the zl_writer,
814 * and all zl_get_data() callbacks are done.
816 if (avl_numnodes(t) == 0)
819 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
821 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
823 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
824 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
827 kmem_free(zv, sizeof (*zv));
831 * Wait for all the flushes to complete. Not all devices actually
832 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
834 (void) zio_wait(zio);
836 spa_config_exit(spa, SCL_STATE, FTAG);
840 * Function called when a log block write completes
843 zil_lwb_write_done(zio_t *zio)
845 lwb_t *lwb = zio->io_private;
846 zilog_t *zilog = lwb->lwb_zilog;
847 dmu_tx_t *tx = lwb->lwb_tx;
849 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
850 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
851 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
852 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
853 ASSERT(!BP_IS_GANG(zio->io_bp));
854 ASSERT(!BP_IS_HOLE(zio->io_bp));
855 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
858 * Ensure the lwb buffer pointer is cleared before releasing
859 * the txg. If we have had an allocation failure and
860 * the txg is waiting to sync then we want want zil_sync()
861 * to remove the lwb so that it's not picked up as the next new
862 * one in zil_commit_writer(). zil_sync() will only remove
863 * the lwb if lwb_buf is null.
865 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
866 mutex_enter(&zilog->zl_lock);
869 mutex_exit(&zilog->zl_lock);
872 * Now that we've written this log block, we have a stable pointer
873 * to the next block in the chain, so it's OK to let the txg in
874 * which we allocated the next block sync.
880 * Initialize the io for a log block.
883 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
887 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
888 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
889 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
891 if (zilog->zl_root_zio == NULL) {
892 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
895 if (lwb->lwb_zio == NULL) {
896 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
897 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
898 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
899 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
904 * Define a limited set of intent log block sizes.
906 * These must be a multiple of 4KB. Note only the amount used (again
907 * aligned to 4KB) actually gets written. However, we can't always just
908 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
910 uint64_t zil_block_buckets[] = {
911 4096, /* non TX_WRITE */
912 8192+4096, /* data base */
913 32*1024 + 4096, /* NFS writes */
918 * Use the slog as long as the logbias is 'latency' and the current commit size
919 * is less than the limit or the total list size is less than 2X the limit.
920 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
922 uint64_t zil_slog_limit = 1024 * 1024;
923 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
924 (((zilog)->zl_cur_used < zil_slog_limit) || \
925 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
928 * Start a log block write and advance to the next log block.
929 * Calls are serialized.
932 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
936 spa_t *spa = zilog->zl_spa;
940 uint64_t zil_blksz, wsz;
943 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
944 zilc = (zil_chain_t *)lwb->lwb_buf;
945 bp = &zilc->zc_next_blk;
947 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
948 bp = &zilc->zc_next_blk;
951 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
954 * Allocate the next block and save its address in this block
955 * before writing it in order to establish the log chain.
956 * Note that if the allocation of nlwb synced before we wrote
957 * the block that points at it (lwb), we'd leak it if we crashed.
958 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
959 * We dirty the dataset to ensure that zil_sync() will be called
960 * to clean up in the event of allocation failure or I/O failure.
962 tx = dmu_tx_create(zilog->zl_os);
963 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
964 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
965 txg = dmu_tx_get_txg(tx);
970 * Log blocks are pre-allocated. Here we select the size of the next
971 * block, based on size used in the last block.
972 * - first find the smallest bucket that will fit the block from a
973 * limited set of block sizes. This is because it's faster to write
974 * blocks allocated from the same metaslab as they are adjacent or
976 * - next find the maximum from the new suggested size and an array of
977 * previous sizes. This lessens a picket fence effect of wrongly
978 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
981 * Note we only write what is used, but we can't just allocate
982 * the maximum block size because we can exhaust the available
985 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
986 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
988 zil_blksz = zil_block_buckets[i];
989 if (zil_blksz == UINT64_MAX)
990 zil_blksz = SPA_MAXBLOCKSIZE;
991 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
992 for (i = 0; i < ZIL_PREV_BLKS; i++)
993 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
994 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
997 /* pass the old blkptr in order to spread log blocks across devs */
998 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1001 ASSERT3U(bp->blk_birth, ==, txg);
1002 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1003 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1006 * Allocate a new log write buffer (lwb).
1008 nlwb = zil_alloc_lwb(zilog, bp, txg);
1010 /* Record the block for later vdev flushing */
1011 zil_add_block(zilog, &lwb->lwb_blk);
1014 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1015 /* For Slim ZIL only write what is used. */
1016 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1017 ASSERT3U(wsz, <=, lwb->lwb_sz);
1018 zio_shrink(lwb->lwb_zio, wsz);
1025 zilc->zc_nused = lwb->lwb_nused;
1026 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1029 * clear unused data for security
1031 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1033 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1036 * If there was an allocation failure then nlwb will be null which
1037 * forces a txg_wait_synced().
1043 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1045 lr_t *lrc = &itx->itx_lr; /* common log record */
1046 lr_write_t *lrw = (lr_write_t *)lrc;
1048 uint64_t txg = lrc->lrc_txg;
1049 uint64_t reclen = lrc->lrc_reclen;
1055 ASSERT(lwb->lwb_buf != NULL);
1056 ASSERT(zilog_is_dirty(zilog) ||
1057 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1059 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1060 dlen = P2ROUNDUP_TYPED(
1061 lrw->lr_length, sizeof (uint64_t), uint64_t);
1063 zilog->zl_cur_used += (reclen + dlen);
1065 zil_lwb_write_init(zilog, lwb);
1068 * If this record won't fit in the current log block, start a new one.
1070 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1071 lwb = zil_lwb_write_start(zilog, lwb);
1074 zil_lwb_write_init(zilog, lwb);
1075 ASSERT(LWB_EMPTY(lwb));
1076 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1077 txg_wait_synced(zilog->zl_dmu_pool, txg);
1082 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1083 bcopy(lrc, lr_buf, reclen);
1084 lrc = (lr_t *)lr_buf;
1085 lrw = (lr_write_t *)lrc;
1088 * If it's a write, fetch the data or get its blkptr as appropriate.
1090 if (lrc->lrc_txtype == TX_WRITE) {
1091 if (txg > spa_freeze_txg(zilog->zl_spa))
1092 txg_wait_synced(zilog->zl_dmu_pool, txg);
1093 if (itx->itx_wr_state != WR_COPIED) {
1098 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1099 dbuf = lr_buf + reclen;
1100 lrw->lr_common.lrc_reclen += dlen;
1102 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1105 error = zilog->zl_get_data(
1106 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1108 txg_wait_synced(zilog->zl_dmu_pool, txg);
1112 ASSERT(error == ENOENT || error == EEXIST ||
1120 * We're actually making an entry, so update lrc_seq to be the
1121 * log record sequence number. Note that this is generally not
1122 * equal to the itx sequence number because not all transactions
1123 * are synchronous, and sometimes spa_sync() gets there first.
1125 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1126 lwb->lwb_nused += reclen + dlen;
1127 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1128 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1129 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1135 zil_itx_create(uint64_t txtype, size_t lrsize)
1139 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1141 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1142 itx->itx_lr.lrc_txtype = txtype;
1143 itx->itx_lr.lrc_reclen = lrsize;
1144 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1145 itx->itx_lr.lrc_seq = 0; /* defensive */
1146 itx->itx_sync = B_TRUE; /* default is synchronous */
1152 zil_itx_destroy(itx_t *itx)
1154 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1158 * Free up the sync and async itxs. The itxs_t has already been detached
1159 * so no locks are needed.
1162 zil_itxg_clean(itxs_t *itxs)
1168 itx_async_node_t *ian;
1170 list = &itxs->i_sync_list;
1171 while ((itx = list_head(list)) != NULL) {
1172 list_remove(list, itx);
1173 kmem_free(itx, offsetof(itx_t, itx_lr) +
1174 itx->itx_lr.lrc_reclen);
1178 t = &itxs->i_async_tree;
1179 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1180 list = &ian->ia_list;
1181 while ((itx = list_head(list)) != NULL) {
1182 list_remove(list, itx);
1183 kmem_free(itx, offsetof(itx_t, itx_lr) +
1184 itx->itx_lr.lrc_reclen);
1187 kmem_free(ian, sizeof (itx_async_node_t));
1191 kmem_free(itxs, sizeof (itxs_t));
1195 zil_aitx_compare(const void *x1, const void *x2)
1197 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1198 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1209 * Remove all async itx with the given oid.
1212 zil_remove_async(zilog_t *zilog, uint64_t oid)
1215 itx_async_node_t *ian;
1222 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1224 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1227 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1229 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1230 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1232 mutex_enter(&itxg->itxg_lock);
1233 if (itxg->itxg_txg != txg) {
1234 mutex_exit(&itxg->itxg_lock);
1239 * Locate the object node and append its list.
1241 t = &itxg->itxg_itxs->i_async_tree;
1242 ian = avl_find(t, &oid, &where);
1244 list_move_tail(&clean_list, &ian->ia_list);
1245 mutex_exit(&itxg->itxg_lock);
1247 while ((itx = list_head(&clean_list)) != NULL) {
1248 list_remove(&clean_list, itx);
1249 kmem_free(itx, offsetof(itx_t, itx_lr) +
1250 itx->itx_lr.lrc_reclen);
1252 list_destroy(&clean_list);
1256 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1260 itxs_t *itxs, *clean = NULL;
1263 * Object ids can be re-instantiated in the next txg so
1264 * remove any async transactions to avoid future leaks.
1265 * This can happen if a fsync occurs on the re-instantiated
1266 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1267 * the new file data and flushes a write record for the old object.
1269 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1270 zil_remove_async(zilog, itx->itx_oid);
1273 * Ensure the data of a renamed file is committed before the rename.
1275 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1276 zil_async_to_sync(zilog, itx->itx_oid);
1278 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1281 txg = dmu_tx_get_txg(tx);
1283 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1284 mutex_enter(&itxg->itxg_lock);
1285 itxs = itxg->itxg_itxs;
1286 if (itxg->itxg_txg != txg) {
1289 * The zil_clean callback hasn't got around to cleaning
1290 * this itxg. Save the itxs for release below.
1291 * This should be rare.
1293 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1295 clean = itxg->itxg_itxs;
1297 ASSERT(itxg->itxg_sod == 0);
1298 itxg->itxg_txg = txg;
1299 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1301 list_create(&itxs->i_sync_list, sizeof (itx_t),
1302 offsetof(itx_t, itx_node));
1303 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1304 sizeof (itx_async_node_t),
1305 offsetof(itx_async_node_t, ia_node));
1307 if (itx->itx_sync) {
1308 list_insert_tail(&itxs->i_sync_list, itx);
1309 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1310 itxg->itxg_sod += itx->itx_sod;
1312 avl_tree_t *t = &itxs->i_async_tree;
1313 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1314 itx_async_node_t *ian;
1317 ian = avl_find(t, &foid, &where);
1319 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1320 list_create(&ian->ia_list, sizeof (itx_t),
1321 offsetof(itx_t, itx_node));
1322 ian->ia_foid = foid;
1323 avl_insert(t, ian, where);
1325 list_insert_tail(&ian->ia_list, itx);
1328 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1329 zilog_dirty(zilog, txg);
1330 mutex_exit(&itxg->itxg_lock);
1332 /* Release the old itxs now we've dropped the lock */
1334 zil_itxg_clean(clean);
1338 * If there are any in-memory intent log transactions which have now been
1339 * synced then start up a taskq to free them. We should only do this after we
1340 * have written out the uberblocks (i.e. txg has been comitted) so that
1341 * don't inadvertently clean out in-memory log records that would be required
1345 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1347 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1350 mutex_enter(&itxg->itxg_lock);
1351 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1352 mutex_exit(&itxg->itxg_lock);
1355 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1356 ASSERT(itxg->itxg_txg != 0);
1357 ASSERT(zilog->zl_clean_taskq != NULL);
1358 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1360 clean_me = itxg->itxg_itxs;
1361 itxg->itxg_itxs = NULL;
1363 mutex_exit(&itxg->itxg_lock);
1365 * Preferably start a task queue to free up the old itxs but
1366 * if taskq_dispatch can't allocate resources to do that then
1367 * free it in-line. This should be rare. Note, using TQ_SLEEP
1368 * created a bad performance problem.
1370 if (taskq_dispatch(zilog->zl_clean_taskq,
1371 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1372 zil_itxg_clean(clean_me);
1376 * Get the list of itxs to commit into zl_itx_commit_list.
1379 zil_get_commit_list(zilog_t *zilog)
1382 list_t *commit_list = &zilog->zl_itx_commit_list;
1383 uint64_t push_sod = 0;
1385 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1388 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1390 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1391 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1393 mutex_enter(&itxg->itxg_lock);
1394 if (itxg->itxg_txg != txg) {
1395 mutex_exit(&itxg->itxg_lock);
1399 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1400 push_sod += itxg->itxg_sod;
1403 mutex_exit(&itxg->itxg_lock);
1405 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1409 * Move the async itxs for a specified object to commit into sync lists.
1412 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1415 itx_async_node_t *ian;
1419 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1422 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1424 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1425 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1427 mutex_enter(&itxg->itxg_lock);
1428 if (itxg->itxg_txg != txg) {
1429 mutex_exit(&itxg->itxg_lock);
1434 * If a foid is specified then find that node and append its
1435 * list. Otherwise walk the tree appending all the lists
1436 * to the sync list. We add to the end rather than the
1437 * beginning to ensure the create has happened.
1439 t = &itxg->itxg_itxs->i_async_tree;
1441 ian = avl_find(t, &foid, &where);
1443 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1447 void *cookie = NULL;
1449 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1450 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1452 list_destroy(&ian->ia_list);
1453 kmem_free(ian, sizeof (itx_async_node_t));
1456 mutex_exit(&itxg->itxg_lock);
1461 zil_commit_writer(zilog_t *zilog)
1466 spa_t *spa = zilog->zl_spa;
1469 ASSERT(zilog->zl_root_zio == NULL);
1471 mutex_exit(&zilog->zl_lock);
1473 zil_get_commit_list(zilog);
1476 * Return if there's nothing to commit before we dirty the fs by
1477 * calling zil_create().
1479 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1480 mutex_enter(&zilog->zl_lock);
1484 if (zilog->zl_suspend) {
1487 lwb = list_tail(&zilog->zl_lwb_list);
1489 lwb = zil_create(zilog);
1492 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1493 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1494 txg = itx->itx_lr.lrc_txg;
1497 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1498 lwb = zil_lwb_commit(zilog, itx, lwb);
1499 list_remove(&zilog->zl_itx_commit_list, itx);
1500 kmem_free(itx, offsetof(itx_t, itx_lr)
1501 + itx->itx_lr.lrc_reclen);
1503 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1505 /* write the last block out */
1506 if (lwb != NULL && lwb->lwb_zio != NULL)
1507 lwb = zil_lwb_write_start(zilog, lwb);
1509 zilog->zl_cur_used = 0;
1512 * Wait if necessary for the log blocks to be on stable storage.
1514 if (zilog->zl_root_zio) {
1515 error = zio_wait(zilog->zl_root_zio);
1516 zilog->zl_root_zio = NULL;
1517 zil_flush_vdevs(zilog);
1520 if (error || lwb == NULL)
1521 txg_wait_synced(zilog->zl_dmu_pool, 0);
1523 mutex_enter(&zilog->zl_lock);
1526 * Remember the highest committed log sequence number for ztest.
1527 * We only update this value when all the log writes succeeded,
1528 * because ztest wants to ASSERT that it got the whole log chain.
1530 if (error == 0 && lwb != NULL)
1531 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1535 * Commit zfs transactions to stable storage.
1536 * If foid is 0 push out all transactions, otherwise push only those
1537 * for that object or might reference that object.
1539 * itxs are committed in batches. In a heavily stressed zil there will be
1540 * a commit writer thread who is writing out a bunch of itxs to the log
1541 * for a set of committing threads (cthreads) in the same batch as the writer.
1542 * Those cthreads are all waiting on the same cv for that batch.
1544 * There will also be a different and growing batch of threads that are
1545 * waiting to commit (qthreads). When the committing batch completes
1546 * a transition occurs such that the cthreads exit and the qthreads become
1547 * cthreads. One of the new cthreads becomes the writer thread for the
1548 * batch. Any new threads arriving become new qthreads.
1550 * Only 2 condition variables are needed and there's no transition
1551 * between the two cvs needed. They just flip-flop between qthreads
1554 * Using this scheme we can efficiently wakeup up only those threads
1555 * that have been committed.
1558 zil_commit(zilog_t *zilog, uint64_t foid)
1562 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1565 /* move the async itxs for the foid to the sync queues */
1566 zil_async_to_sync(zilog, foid);
1568 mutex_enter(&zilog->zl_lock);
1569 mybatch = zilog->zl_next_batch;
1570 while (zilog->zl_writer) {
1571 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1572 if (mybatch <= zilog->zl_com_batch) {
1573 mutex_exit(&zilog->zl_lock);
1578 zilog->zl_next_batch++;
1579 zilog->zl_writer = B_TRUE;
1580 zil_commit_writer(zilog);
1581 zilog->zl_com_batch = mybatch;
1582 zilog->zl_writer = B_FALSE;
1583 mutex_exit(&zilog->zl_lock);
1585 /* wake up one thread to become the next writer */
1586 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1588 /* wake up all threads waiting for this batch to be committed */
1589 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1593 * Called in syncing context to free committed log blocks and update log header.
1596 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1598 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1599 uint64_t txg = dmu_tx_get_txg(tx);
1600 spa_t *spa = zilog->zl_spa;
1601 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1605 * We don't zero out zl_destroy_txg, so make sure we don't try
1606 * to destroy it twice.
1608 if (spa_sync_pass(spa) != 1)
1611 mutex_enter(&zilog->zl_lock);
1613 ASSERT(zilog->zl_stop_sync == 0);
1615 if (*replayed_seq != 0) {
1616 ASSERT(zh->zh_replay_seq < *replayed_seq);
1617 zh->zh_replay_seq = *replayed_seq;
1621 if (zilog->zl_destroy_txg == txg) {
1622 blkptr_t blk = zh->zh_log;
1624 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1626 bzero(zh, sizeof (zil_header_t));
1627 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1629 if (zilog->zl_keep_first) {
1631 * If this block was part of log chain that couldn't
1632 * be claimed because a device was missing during
1633 * zil_claim(), but that device later returns,
1634 * then this block could erroneously appear valid.
1635 * To guard against this, assign a new GUID to the new
1636 * log chain so it doesn't matter what blk points to.
1638 zil_init_log_chain(zilog, &blk);
1643 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1644 zh->zh_log = lwb->lwb_blk;
1645 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1647 list_remove(&zilog->zl_lwb_list, lwb);
1648 zio_free_zil(spa, txg, &lwb->lwb_blk);
1649 kmem_cache_free(zil_lwb_cache, lwb);
1652 * If we don't have anything left in the lwb list then
1653 * we've had an allocation failure and we need to zero
1654 * out the zil_header blkptr so that we don't end
1655 * up freeing the same block twice.
1657 if (list_head(&zilog->zl_lwb_list) == NULL)
1658 BP_ZERO(&zh->zh_log);
1660 mutex_exit(&zilog->zl_lock);
1666 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1667 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1673 kmem_cache_destroy(zil_lwb_cache);
1677 zil_set_sync(zilog_t *zilog, uint64_t sync)
1679 zilog->zl_sync = sync;
1683 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1685 zilog->zl_logbias = logbias;
1689 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1693 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1695 zilog->zl_header = zh_phys;
1697 zilog->zl_spa = dmu_objset_spa(os);
1698 zilog->zl_dmu_pool = dmu_objset_pool(os);
1699 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1700 zilog->zl_logbias = dmu_objset_logbias(os);
1701 zilog->zl_sync = dmu_objset_syncprop(os);
1702 zilog->zl_next_batch = 1;
1704 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1706 for (int i = 0; i < TXG_SIZE; i++) {
1707 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1708 MUTEX_DEFAULT, NULL);
1711 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1712 offsetof(lwb_t, lwb_node));
1714 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1715 offsetof(itx_t, itx_node));
1717 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1719 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1720 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1722 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1723 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1724 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1725 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1731 zil_free(zilog_t *zilog)
1733 zilog->zl_stop_sync = 1;
1735 ASSERT0(zilog->zl_suspend);
1736 ASSERT0(zilog->zl_suspending);
1738 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1739 list_destroy(&zilog->zl_lwb_list);
1741 avl_destroy(&zilog->zl_vdev_tree);
1742 mutex_destroy(&zilog->zl_vdev_lock);
1744 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1745 list_destroy(&zilog->zl_itx_commit_list);
1747 for (int i = 0; i < TXG_SIZE; i++) {
1749 * It's possible for an itx to be generated that doesn't dirty
1750 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1751 * callback to remove the entry. We remove those here.
1753 * Also free up the ziltest itxs.
1755 if (zilog->zl_itxg[i].itxg_itxs)
1756 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1757 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1760 mutex_destroy(&zilog->zl_lock);
1762 cv_destroy(&zilog->zl_cv_writer);
1763 cv_destroy(&zilog->zl_cv_suspend);
1764 cv_destroy(&zilog->zl_cv_batch[0]);
1765 cv_destroy(&zilog->zl_cv_batch[1]);
1767 kmem_free(zilog, sizeof (zilog_t));
1771 * Open an intent log.
1774 zil_open(objset_t *os, zil_get_data_t *get_data)
1776 zilog_t *zilog = dmu_objset_zil(os);
1778 ASSERT(zilog->zl_clean_taskq == NULL);
1779 ASSERT(zilog->zl_get_data == NULL);
1780 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1782 zilog->zl_get_data = get_data;
1783 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1784 2, 2, TASKQ_PREPOPULATE);
1790 * Close an intent log.
1793 zil_close(zilog_t *zilog)
1798 zil_commit(zilog, 0); /* commit all itx */
1801 * The lwb_max_txg for the stubby lwb will reflect the last activity
1802 * for the zil. After a txg_wait_synced() on the txg we know all the
1803 * callbacks have occurred that may clean the zil. Only then can we
1804 * destroy the zl_clean_taskq.
1806 mutex_enter(&zilog->zl_lock);
1807 lwb = list_tail(&zilog->zl_lwb_list);
1809 txg = lwb->lwb_max_txg;
1810 mutex_exit(&zilog->zl_lock);
1812 txg_wait_synced(zilog->zl_dmu_pool, txg);
1813 ASSERT(!zilog_is_dirty(zilog));
1815 taskq_destroy(zilog->zl_clean_taskq);
1816 zilog->zl_clean_taskq = NULL;
1817 zilog->zl_get_data = NULL;
1820 * We should have only one LWB left on the list; remove it now.
1822 mutex_enter(&zilog->zl_lock);
1823 lwb = list_head(&zilog->zl_lwb_list);
1825 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1826 list_remove(&zilog->zl_lwb_list, lwb);
1827 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1828 kmem_cache_free(zil_lwb_cache, lwb);
1830 mutex_exit(&zilog->zl_lock);
1833 static char *suspend_tag = "zil suspending";
1836 * Suspend an intent log. While in suspended mode, we still honor
1837 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1838 * On old version pools, we suspend the log briefly when taking a
1839 * snapshot so that it will have an empty intent log.
1841 * Long holds are not really intended to be used the way we do here --
1842 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1843 * could fail. Therefore we take pains to only put a long hold if it is
1844 * actually necessary. Fortunately, it will only be necessary if the
1845 * objset is currently mounted (or the ZVOL equivalent). In that case it
1846 * will already have a long hold, so we are not really making things any worse.
1848 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1849 * zvol_state_t), and use their mechanism to prevent their hold from being
1850 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1853 * if cookiep == NULL, this does both the suspend & resume.
1854 * Otherwise, it returns with the dataset "long held", and the cookie
1855 * should be passed into zil_resume().
1858 zil_suspend(const char *osname, void **cookiep)
1862 const zil_header_t *zh;
1865 error = dmu_objset_hold(osname, suspend_tag, &os);
1868 zilog = dmu_objset_zil(os);
1870 mutex_enter(&zilog->zl_lock);
1871 zh = zilog->zl_header;
1873 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1874 mutex_exit(&zilog->zl_lock);
1875 dmu_objset_rele(os, suspend_tag);
1876 return (SET_ERROR(EBUSY));
1880 * Don't put a long hold in the cases where we can avoid it. This
1881 * is when there is no cookie so we are doing a suspend & resume
1882 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1883 * for the suspend because it's already suspended, or there's no ZIL.
1885 if (cookiep == NULL && !zilog->zl_suspending &&
1886 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1887 mutex_exit(&zilog->zl_lock);
1888 dmu_objset_rele(os, suspend_tag);
1892 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1893 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1895 zilog->zl_suspend++;
1897 if (zilog->zl_suspend > 1) {
1899 * Someone else is already suspending it.
1900 * Just wait for them to finish.
1903 while (zilog->zl_suspending)
1904 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1905 mutex_exit(&zilog->zl_lock);
1907 if (cookiep == NULL)
1915 * If there is no pointer to an on-disk block, this ZIL must not
1916 * be active (e.g. filesystem not mounted), so there's nothing
1919 if (BP_IS_HOLE(&zh->zh_log)) {
1920 ASSERT(cookiep != NULL); /* fast path already handled */
1923 mutex_exit(&zilog->zl_lock);
1927 zilog->zl_suspending = B_TRUE;
1928 mutex_exit(&zilog->zl_lock);
1930 zil_commit(zilog, 0);
1932 zil_destroy(zilog, B_FALSE);
1934 mutex_enter(&zilog->zl_lock);
1935 zilog->zl_suspending = B_FALSE;
1936 cv_broadcast(&zilog->zl_cv_suspend);
1937 mutex_exit(&zilog->zl_lock);
1939 if (cookiep == NULL)
1947 zil_resume(void *cookie)
1949 objset_t *os = cookie;
1950 zilog_t *zilog = dmu_objset_zil(os);
1952 mutex_enter(&zilog->zl_lock);
1953 ASSERT(zilog->zl_suspend != 0);
1954 zilog->zl_suspend--;
1955 mutex_exit(&zilog->zl_lock);
1956 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1957 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1960 typedef struct zil_replay_arg {
1961 zil_replay_func_t **zr_replay;
1963 boolean_t zr_byteswap;
1968 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1970 char name[MAXNAMELEN];
1972 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1974 dmu_objset_name(zilog->zl_os, name);
1976 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1977 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1978 (u_longlong_t)lr->lrc_seq,
1979 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1980 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1986 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1988 zil_replay_arg_t *zr = zra;
1989 const zil_header_t *zh = zilog->zl_header;
1990 uint64_t reclen = lr->lrc_reclen;
1991 uint64_t txtype = lr->lrc_txtype;
1994 zilog->zl_replaying_seq = lr->lrc_seq;
1996 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1999 if (lr->lrc_txg < claim_txg) /* already committed */
2002 /* Strip case-insensitive bit, still present in log record */
2005 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2006 return (zil_replay_error(zilog, lr, EINVAL));
2009 * If this record type can be logged out of order, the object
2010 * (lr_foid) may no longer exist. That's legitimate, not an error.
2012 if (TX_OOO(txtype)) {
2013 error = dmu_object_info(zilog->zl_os,
2014 ((lr_ooo_t *)lr)->lr_foid, NULL);
2015 if (error == ENOENT || error == EEXIST)
2020 * Make a copy of the data so we can revise and extend it.
2022 bcopy(lr, zr->zr_lr, reclen);
2025 * If this is a TX_WRITE with a blkptr, suck in the data.
2027 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2028 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2029 zr->zr_lr + reclen);
2031 return (zil_replay_error(zilog, lr, error));
2035 * The log block containing this lr may have been byteswapped
2036 * so that we can easily examine common fields like lrc_txtype.
2037 * However, the log is a mix of different record types, and only the
2038 * replay vectors know how to byteswap their records. Therefore, if
2039 * the lr was byteswapped, undo it before invoking the replay vector.
2041 if (zr->zr_byteswap)
2042 byteswap_uint64_array(zr->zr_lr, reclen);
2045 * We must now do two things atomically: replay this log record,
2046 * and update the log header sequence number to reflect the fact that
2047 * we did so. At the end of each replay function the sequence number
2048 * is updated if we are in replay mode.
2050 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2053 * The DMU's dnode layer doesn't see removes until the txg
2054 * commits, so a subsequent claim can spuriously fail with
2055 * EEXIST. So if we receive any error we try syncing out
2056 * any removes then retry the transaction. Note that we
2057 * specify B_FALSE for byteswap now, so we don't do it twice.
2059 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2060 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2062 return (zil_replay_error(zilog, lr, error));
2069 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2071 zilog->zl_replay_blks++;
2077 * If this dataset has a non-empty intent log, replay it and destroy it.
2080 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2082 zilog_t *zilog = dmu_objset_zil(os);
2083 const zil_header_t *zh = zilog->zl_header;
2084 zil_replay_arg_t zr;
2086 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2087 zil_destroy(zilog, B_TRUE);
2090 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2092 zr.zr_replay = replay_func;
2094 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2095 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2098 * Wait for in-progress removes to sync before starting replay.
2100 txg_wait_synced(zilog->zl_dmu_pool, 0);
2102 zilog->zl_replay = B_TRUE;
2103 zilog->zl_replay_time = ddi_get_lbolt();
2104 ASSERT(zilog->zl_replay_blks == 0);
2105 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2107 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2109 zil_destroy(zilog, B_FALSE);
2110 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2111 zilog->zl_replay = B_FALSE;
2112 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2116 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2118 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2121 if (zilog->zl_replay) {
2122 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2123 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2124 zilog->zl_replaying_seq;
2133 zil_vdev_offline(const char *osname, void *arg)
2137 error = zil_suspend(osname, NULL);
2139 return (SET_ERROR(EEXIST));