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);
647 cmn_err(CE_WARN, "can't open objset for %s", osname);
651 zilog = dmu_objset_zil(os);
652 zh = zil_header_in_syncing_context(zilog);
654 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
655 if (!BP_IS_HOLE(&zh->zh_log))
656 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
657 BP_ZERO(&zh->zh_log);
658 dsl_dataset_dirty(dmu_objset_ds(os), tx);
659 dmu_objset_disown(os, FTAG);
664 * Claim all log blocks if we haven't already done so, and remember
665 * the highest claimed sequence number. This ensures that if we can
666 * read only part of the log now (e.g. due to a missing device),
667 * but we can read the entire log later, we will not try to replay
668 * or destroy beyond the last block we successfully claimed.
670 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
671 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
672 (void) zil_parse(zilog, zil_claim_log_block,
673 zil_claim_log_record, tx, first_txg);
674 zh->zh_claim_txg = first_txg;
675 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
676 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
677 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
678 zh->zh_flags |= ZIL_REPLAY_NEEDED;
679 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
680 dsl_dataset_dirty(dmu_objset_ds(os), tx);
683 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
684 dmu_objset_disown(os, FTAG);
689 * Check the log by walking the log chain.
690 * Checksum errors are ok as they indicate the end of the chain.
691 * Any other error (no device or read failure) returns an error.
694 zil_check_log_chain(const char *osname, void *tx)
703 error = dmu_objset_hold(osname, FTAG, &os);
705 cmn_err(CE_WARN, "can't open objset for %s", osname);
709 zilog = dmu_objset_zil(os);
710 bp = (blkptr_t *)&zilog->zl_header->zh_log;
713 * Check the first block and determine if it's on a log device
714 * which may have been removed or faulted prior to loading this
715 * pool. If so, there's no point in checking the rest of the log
716 * as its content should have already been synced to the pool.
718 if (!BP_IS_HOLE(bp)) {
720 boolean_t valid = B_TRUE;
722 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
723 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
724 if (vd->vdev_islog && vdev_is_dead(vd))
725 valid = vdev_log_state_valid(vd);
726 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
729 dmu_objset_rele(os, FTAG);
735 * Because tx == NULL, zil_claim_log_block() will not actually claim
736 * any blocks, but just determine whether it is possible to do so.
737 * In addition to checking the log chain, zil_claim_log_block()
738 * will invoke zio_claim() with a done func of spa_claim_notify(),
739 * which will update spa_max_claim_txg. See spa_load() for details.
741 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
742 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
744 dmu_objset_rele(os, FTAG);
746 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
750 zil_vdev_compare(const void *x1, const void *x2)
752 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
753 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
764 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
766 avl_tree_t *t = &zilog->zl_vdev_tree;
768 zil_vdev_node_t *zv, zvsearch;
769 int ndvas = BP_GET_NDVAS(bp);
772 if (zfs_nocacheflush)
775 ASSERT(zilog->zl_writer);
778 * Even though we're zl_writer, we still need a lock because the
779 * zl_get_data() callbacks may have dmu_sync() done callbacks
780 * that will run concurrently.
782 mutex_enter(&zilog->zl_vdev_lock);
783 for (i = 0; i < ndvas; i++) {
784 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
785 if (avl_find(t, &zvsearch, &where) == NULL) {
786 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
787 zv->zv_vdev = zvsearch.zv_vdev;
788 avl_insert(t, zv, where);
791 mutex_exit(&zilog->zl_vdev_lock);
795 zil_flush_vdevs(zilog_t *zilog)
797 spa_t *spa = zilog->zl_spa;
798 avl_tree_t *t = &zilog->zl_vdev_tree;
803 ASSERT(zilog->zl_writer);
806 * We don't need zl_vdev_lock here because we're the zl_writer,
807 * and all zl_get_data() callbacks are done.
809 if (avl_numnodes(t) == 0)
812 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
814 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
816 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
817 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
820 kmem_free(zv, sizeof (*zv));
824 * Wait for all the flushes to complete. Not all devices actually
825 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
827 (void) zio_wait(zio);
829 spa_config_exit(spa, SCL_STATE, FTAG);
833 * Function called when a log block write completes
836 zil_lwb_write_done(zio_t *zio)
838 lwb_t *lwb = zio->io_private;
839 zilog_t *zilog = lwb->lwb_zilog;
840 dmu_tx_t *tx = lwb->lwb_tx;
842 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
843 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
844 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
845 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
846 ASSERT(!BP_IS_GANG(zio->io_bp));
847 ASSERT(!BP_IS_HOLE(zio->io_bp));
848 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
851 * Ensure the lwb buffer pointer is cleared before releasing
852 * the txg. If we have had an allocation failure and
853 * the txg is waiting to sync then we want want zil_sync()
854 * to remove the lwb so that it's not picked up as the next new
855 * one in zil_commit_writer(). zil_sync() will only remove
856 * the lwb if lwb_buf is null.
858 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
859 mutex_enter(&zilog->zl_lock);
862 mutex_exit(&zilog->zl_lock);
865 * Now that we've written this log block, we have a stable pointer
866 * to the next block in the chain, so it's OK to let the txg in
867 * which we allocated the next block sync.
873 * Initialize the io for a log block.
876 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
880 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
881 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
882 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
884 if (zilog->zl_root_zio == NULL) {
885 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
888 if (lwb->lwb_zio == NULL) {
889 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
890 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
891 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
892 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
897 * Define a limited set of intent log block sizes.
899 * These must be a multiple of 4KB. Note only the amount used (again
900 * aligned to 4KB) actually gets written. However, we can't always just
901 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
903 uint64_t zil_block_buckets[] = {
904 4096, /* non TX_WRITE */
905 8192+4096, /* data base */
906 32*1024 + 4096, /* NFS writes */
911 * Use the slog as long as the logbias is 'latency' and the current commit size
912 * is less than the limit or the total list size is less than 2X the limit.
913 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
915 uint64_t zil_slog_limit = 1024 * 1024;
916 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
917 (((zilog)->zl_cur_used < zil_slog_limit) || \
918 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
921 * Start a log block write and advance to the next log block.
922 * Calls are serialized.
925 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
929 spa_t *spa = zilog->zl_spa;
933 uint64_t zil_blksz, wsz;
936 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
937 zilc = (zil_chain_t *)lwb->lwb_buf;
938 bp = &zilc->zc_next_blk;
940 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
941 bp = &zilc->zc_next_blk;
944 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
947 * Allocate the next block and save its address in this block
948 * before writing it in order to establish the log chain.
949 * Note that if the allocation of nlwb synced before we wrote
950 * the block that points at it (lwb), we'd leak it if we crashed.
951 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
952 * We dirty the dataset to ensure that zil_sync() will be called
953 * to clean up in the event of allocation failure or I/O failure.
955 tx = dmu_tx_create(zilog->zl_os);
956 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
957 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
958 txg = dmu_tx_get_txg(tx);
963 * Log blocks are pre-allocated. Here we select the size of the next
964 * block, based on size used in the last block.
965 * - first find the smallest bucket that will fit the block from a
966 * limited set of block sizes. This is because it's faster to write
967 * blocks allocated from the same metaslab as they are adjacent or
969 * - next find the maximum from the new suggested size and an array of
970 * previous sizes. This lessens a picket fence effect of wrongly
971 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
974 * Note we only write what is used, but we can't just allocate
975 * the maximum block size because we can exhaust the available
978 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
979 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
981 zil_blksz = zil_block_buckets[i];
982 if (zil_blksz == UINT64_MAX)
983 zil_blksz = SPA_MAXBLOCKSIZE;
984 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
985 for (i = 0; i < ZIL_PREV_BLKS; i++)
986 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
987 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
990 /* pass the old blkptr in order to spread log blocks across devs */
991 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
994 ASSERT3U(bp->blk_birth, ==, txg);
995 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
996 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
999 * Allocate a new log write buffer (lwb).
1001 nlwb = zil_alloc_lwb(zilog, bp, txg);
1003 /* Record the block for later vdev flushing */
1004 zil_add_block(zilog, &lwb->lwb_blk);
1007 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1008 /* For Slim ZIL only write what is used. */
1009 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1010 ASSERT3U(wsz, <=, lwb->lwb_sz);
1011 zio_shrink(lwb->lwb_zio, wsz);
1018 zilc->zc_nused = lwb->lwb_nused;
1019 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1022 * clear unused data for security
1024 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1026 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1029 * If there was an allocation failure then nlwb will be null which
1030 * forces a txg_wait_synced().
1036 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1038 lr_t *lrc = &itx->itx_lr; /* common log record */
1039 lr_write_t *lrw = (lr_write_t *)lrc;
1041 uint64_t txg = lrc->lrc_txg;
1042 uint64_t reclen = lrc->lrc_reclen;
1048 ASSERT(lwb->lwb_buf != NULL);
1049 ASSERT(zilog_is_dirty(zilog) ||
1050 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1052 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1053 dlen = P2ROUNDUP_TYPED(
1054 lrw->lr_length, sizeof (uint64_t), uint64_t);
1056 zilog->zl_cur_used += (reclen + dlen);
1058 zil_lwb_write_init(zilog, lwb);
1061 * If this record won't fit in the current log block, start a new one.
1063 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1064 lwb = zil_lwb_write_start(zilog, lwb);
1067 zil_lwb_write_init(zilog, lwb);
1068 ASSERT(LWB_EMPTY(lwb));
1069 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1070 txg_wait_synced(zilog->zl_dmu_pool, txg);
1075 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1076 bcopy(lrc, lr_buf, reclen);
1077 lrc = (lr_t *)lr_buf;
1078 lrw = (lr_write_t *)lrc;
1081 * If it's a write, fetch the data or get its blkptr as appropriate.
1083 if (lrc->lrc_txtype == TX_WRITE) {
1084 if (txg > spa_freeze_txg(zilog->zl_spa))
1085 txg_wait_synced(zilog->zl_dmu_pool, txg);
1086 if (itx->itx_wr_state != WR_COPIED) {
1091 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1092 dbuf = lr_buf + reclen;
1093 lrw->lr_common.lrc_reclen += dlen;
1095 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1098 error = zilog->zl_get_data(
1099 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1101 txg_wait_synced(zilog->zl_dmu_pool, txg);
1105 ASSERT(error == ENOENT || error == EEXIST ||
1113 * We're actually making an entry, so update lrc_seq to be the
1114 * log record sequence number. Note that this is generally not
1115 * equal to the itx sequence number because not all transactions
1116 * are synchronous, and sometimes spa_sync() gets there first.
1118 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1119 lwb->lwb_nused += reclen + dlen;
1120 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1121 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1122 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1128 zil_itx_create(uint64_t txtype, size_t lrsize)
1132 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1134 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1135 itx->itx_lr.lrc_txtype = txtype;
1136 itx->itx_lr.lrc_reclen = lrsize;
1137 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1138 itx->itx_lr.lrc_seq = 0; /* defensive */
1139 itx->itx_sync = B_TRUE; /* default is synchronous */
1145 zil_itx_destroy(itx_t *itx)
1147 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1151 * Free up the sync and async itxs. The itxs_t has already been detached
1152 * so no locks are needed.
1155 zil_itxg_clean(itxs_t *itxs)
1161 itx_async_node_t *ian;
1163 list = &itxs->i_sync_list;
1164 while ((itx = list_head(list)) != NULL) {
1165 list_remove(list, itx);
1166 kmem_free(itx, offsetof(itx_t, itx_lr) +
1167 itx->itx_lr.lrc_reclen);
1171 t = &itxs->i_async_tree;
1172 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1173 list = &ian->ia_list;
1174 while ((itx = list_head(list)) != NULL) {
1175 list_remove(list, itx);
1176 kmem_free(itx, offsetof(itx_t, itx_lr) +
1177 itx->itx_lr.lrc_reclen);
1180 kmem_free(ian, sizeof (itx_async_node_t));
1184 kmem_free(itxs, sizeof (itxs_t));
1188 zil_aitx_compare(const void *x1, const void *x2)
1190 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1191 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1202 * Remove all async itx with the given oid.
1205 zil_remove_async(zilog_t *zilog, uint64_t oid)
1208 itx_async_node_t *ian;
1215 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1217 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1220 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1222 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1223 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1225 mutex_enter(&itxg->itxg_lock);
1226 if (itxg->itxg_txg != txg) {
1227 mutex_exit(&itxg->itxg_lock);
1232 * Locate the object node and append its list.
1234 t = &itxg->itxg_itxs->i_async_tree;
1235 ian = avl_find(t, &oid, &where);
1237 list_move_tail(&clean_list, &ian->ia_list);
1238 mutex_exit(&itxg->itxg_lock);
1240 while ((itx = list_head(&clean_list)) != NULL) {
1241 list_remove(&clean_list, itx);
1242 kmem_free(itx, offsetof(itx_t, itx_lr) +
1243 itx->itx_lr.lrc_reclen);
1245 list_destroy(&clean_list);
1249 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1253 itxs_t *itxs, *clean = NULL;
1256 * Object ids can be re-instantiated in the next txg so
1257 * remove any async transactions to avoid future leaks.
1258 * This can happen if a fsync occurs on the re-instantiated
1259 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1260 * the new file data and flushes a write record for the old object.
1262 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1263 zil_remove_async(zilog, itx->itx_oid);
1266 * Ensure the data of a renamed file is committed before the rename.
1268 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1269 zil_async_to_sync(zilog, itx->itx_oid);
1271 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1274 txg = dmu_tx_get_txg(tx);
1276 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1277 mutex_enter(&itxg->itxg_lock);
1278 itxs = itxg->itxg_itxs;
1279 if (itxg->itxg_txg != txg) {
1282 * The zil_clean callback hasn't got around to cleaning
1283 * this itxg. Save the itxs for release below.
1284 * This should be rare.
1286 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1288 clean = itxg->itxg_itxs;
1290 ASSERT(itxg->itxg_sod == 0);
1291 itxg->itxg_txg = txg;
1292 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1294 list_create(&itxs->i_sync_list, sizeof (itx_t),
1295 offsetof(itx_t, itx_node));
1296 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1297 sizeof (itx_async_node_t),
1298 offsetof(itx_async_node_t, ia_node));
1300 if (itx->itx_sync) {
1301 list_insert_tail(&itxs->i_sync_list, itx);
1302 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1303 itxg->itxg_sod += itx->itx_sod;
1305 avl_tree_t *t = &itxs->i_async_tree;
1306 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1307 itx_async_node_t *ian;
1310 ian = avl_find(t, &foid, &where);
1312 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1313 list_create(&ian->ia_list, sizeof (itx_t),
1314 offsetof(itx_t, itx_node));
1315 ian->ia_foid = foid;
1316 avl_insert(t, ian, where);
1318 list_insert_tail(&ian->ia_list, itx);
1321 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1322 zilog_dirty(zilog, txg);
1323 mutex_exit(&itxg->itxg_lock);
1325 /* Release the old itxs now we've dropped the lock */
1327 zil_itxg_clean(clean);
1331 * If there are any in-memory intent log transactions which have now been
1332 * synced then start up a taskq to free them. We should only do this after we
1333 * have written out the uberblocks (i.e. txg has been comitted) so that
1334 * don't inadvertently clean out in-memory log records that would be required
1338 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1340 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1343 mutex_enter(&itxg->itxg_lock);
1344 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1345 mutex_exit(&itxg->itxg_lock);
1348 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1349 ASSERT(itxg->itxg_txg != 0);
1350 ASSERT(zilog->zl_clean_taskq != NULL);
1351 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1353 clean_me = itxg->itxg_itxs;
1354 itxg->itxg_itxs = NULL;
1356 mutex_exit(&itxg->itxg_lock);
1358 * Preferably start a task queue to free up the old itxs but
1359 * if taskq_dispatch can't allocate resources to do that then
1360 * free it in-line. This should be rare. Note, using TQ_SLEEP
1361 * created a bad performance problem.
1363 if (taskq_dispatch(zilog->zl_clean_taskq,
1364 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1365 zil_itxg_clean(clean_me);
1369 * Get the list of itxs to commit into zl_itx_commit_list.
1372 zil_get_commit_list(zilog_t *zilog)
1375 list_t *commit_list = &zilog->zl_itx_commit_list;
1376 uint64_t push_sod = 0;
1378 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1381 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1383 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1384 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1386 mutex_enter(&itxg->itxg_lock);
1387 if (itxg->itxg_txg != txg) {
1388 mutex_exit(&itxg->itxg_lock);
1392 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1393 push_sod += itxg->itxg_sod;
1396 mutex_exit(&itxg->itxg_lock);
1398 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1402 * Move the async itxs for a specified object to commit into sync lists.
1405 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1408 itx_async_node_t *ian;
1412 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1415 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1417 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1418 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1420 mutex_enter(&itxg->itxg_lock);
1421 if (itxg->itxg_txg != txg) {
1422 mutex_exit(&itxg->itxg_lock);
1427 * If a foid is specified then find that node and append its
1428 * list. Otherwise walk the tree appending all the lists
1429 * to the sync list. We add to the end rather than the
1430 * beginning to ensure the create has happened.
1432 t = &itxg->itxg_itxs->i_async_tree;
1434 ian = avl_find(t, &foid, &where);
1436 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1440 void *cookie = NULL;
1442 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1443 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1445 list_destroy(&ian->ia_list);
1446 kmem_free(ian, sizeof (itx_async_node_t));
1449 mutex_exit(&itxg->itxg_lock);
1454 zil_commit_writer(zilog_t *zilog)
1459 spa_t *spa = zilog->zl_spa;
1462 ASSERT(zilog->zl_root_zio == NULL);
1464 mutex_exit(&zilog->zl_lock);
1466 zil_get_commit_list(zilog);
1469 * Return if there's nothing to commit before we dirty the fs by
1470 * calling zil_create().
1472 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1473 mutex_enter(&zilog->zl_lock);
1477 if (zilog->zl_suspend) {
1480 lwb = list_tail(&zilog->zl_lwb_list);
1482 lwb = zil_create(zilog);
1485 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1486 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1487 txg = itx->itx_lr.lrc_txg;
1490 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1491 lwb = zil_lwb_commit(zilog, itx, lwb);
1492 list_remove(&zilog->zl_itx_commit_list, itx);
1493 kmem_free(itx, offsetof(itx_t, itx_lr)
1494 + itx->itx_lr.lrc_reclen);
1496 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1498 /* write the last block out */
1499 if (lwb != NULL && lwb->lwb_zio != NULL)
1500 lwb = zil_lwb_write_start(zilog, lwb);
1502 zilog->zl_cur_used = 0;
1505 * Wait if necessary for the log blocks to be on stable storage.
1507 if (zilog->zl_root_zio) {
1508 error = zio_wait(zilog->zl_root_zio);
1509 zilog->zl_root_zio = NULL;
1510 zil_flush_vdevs(zilog);
1513 if (error || lwb == NULL)
1514 txg_wait_synced(zilog->zl_dmu_pool, 0);
1516 mutex_enter(&zilog->zl_lock);
1519 * Remember the highest committed log sequence number for ztest.
1520 * We only update this value when all the log writes succeeded,
1521 * because ztest wants to ASSERT that it got the whole log chain.
1523 if (error == 0 && lwb != NULL)
1524 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1528 * Commit zfs transactions to stable storage.
1529 * If foid is 0 push out all transactions, otherwise push only those
1530 * for that object or might reference that object.
1532 * itxs are committed in batches. In a heavily stressed zil there will be
1533 * a commit writer thread who is writing out a bunch of itxs to the log
1534 * for a set of committing threads (cthreads) in the same batch as the writer.
1535 * Those cthreads are all waiting on the same cv for that batch.
1537 * There will also be a different and growing batch of threads that are
1538 * waiting to commit (qthreads). When the committing batch completes
1539 * a transition occurs such that the cthreads exit and the qthreads become
1540 * cthreads. One of the new cthreads becomes the writer thread for the
1541 * batch. Any new threads arriving become new qthreads.
1543 * Only 2 condition variables are needed and there's no transition
1544 * between the two cvs needed. They just flip-flop between qthreads
1547 * Using this scheme we can efficiently wakeup up only those threads
1548 * that have been committed.
1551 zil_commit(zilog_t *zilog, uint64_t foid)
1555 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1558 /* move the async itxs for the foid to the sync queues */
1559 zil_async_to_sync(zilog, foid);
1561 mutex_enter(&zilog->zl_lock);
1562 mybatch = zilog->zl_next_batch;
1563 while (zilog->zl_writer) {
1564 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1565 if (mybatch <= zilog->zl_com_batch) {
1566 mutex_exit(&zilog->zl_lock);
1571 zilog->zl_next_batch++;
1572 zilog->zl_writer = B_TRUE;
1573 zil_commit_writer(zilog);
1574 zilog->zl_com_batch = mybatch;
1575 zilog->zl_writer = B_FALSE;
1576 mutex_exit(&zilog->zl_lock);
1578 /* wake up one thread to become the next writer */
1579 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1581 /* wake up all threads waiting for this batch to be committed */
1582 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1586 * Called in syncing context to free committed log blocks and update log header.
1589 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1591 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1592 uint64_t txg = dmu_tx_get_txg(tx);
1593 spa_t *spa = zilog->zl_spa;
1594 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1598 * We don't zero out zl_destroy_txg, so make sure we don't try
1599 * to destroy it twice.
1601 if (spa_sync_pass(spa) != 1)
1604 mutex_enter(&zilog->zl_lock);
1606 ASSERT(zilog->zl_stop_sync == 0);
1608 if (*replayed_seq != 0) {
1609 ASSERT(zh->zh_replay_seq < *replayed_seq);
1610 zh->zh_replay_seq = *replayed_seq;
1614 if (zilog->zl_destroy_txg == txg) {
1615 blkptr_t blk = zh->zh_log;
1617 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1619 bzero(zh, sizeof (zil_header_t));
1620 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1622 if (zilog->zl_keep_first) {
1624 * If this block was part of log chain that couldn't
1625 * be claimed because a device was missing during
1626 * zil_claim(), but that device later returns,
1627 * then this block could erroneously appear valid.
1628 * To guard against this, assign a new GUID to the new
1629 * log chain so it doesn't matter what blk points to.
1631 zil_init_log_chain(zilog, &blk);
1636 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1637 zh->zh_log = lwb->lwb_blk;
1638 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1640 list_remove(&zilog->zl_lwb_list, lwb);
1641 zio_free_zil(spa, txg, &lwb->lwb_blk);
1642 kmem_cache_free(zil_lwb_cache, lwb);
1645 * If we don't have anything left in the lwb list then
1646 * we've had an allocation failure and we need to zero
1647 * out the zil_header blkptr so that we don't end
1648 * up freeing the same block twice.
1650 if (list_head(&zilog->zl_lwb_list) == NULL)
1651 BP_ZERO(&zh->zh_log);
1653 mutex_exit(&zilog->zl_lock);
1659 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1660 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1666 kmem_cache_destroy(zil_lwb_cache);
1670 zil_set_sync(zilog_t *zilog, uint64_t sync)
1672 zilog->zl_sync = sync;
1676 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1678 zilog->zl_logbias = logbias;
1682 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1686 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1688 zilog->zl_header = zh_phys;
1690 zilog->zl_spa = dmu_objset_spa(os);
1691 zilog->zl_dmu_pool = dmu_objset_pool(os);
1692 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1693 zilog->zl_logbias = dmu_objset_logbias(os);
1694 zilog->zl_sync = dmu_objset_syncprop(os);
1695 zilog->zl_next_batch = 1;
1697 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1699 for (int i = 0; i < TXG_SIZE; i++) {
1700 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1701 MUTEX_DEFAULT, NULL);
1704 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1705 offsetof(lwb_t, lwb_node));
1707 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1708 offsetof(itx_t, itx_node));
1710 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1712 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1713 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1715 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1716 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1717 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1718 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1724 zil_free(zilog_t *zilog)
1726 zilog->zl_stop_sync = 1;
1728 ASSERT0(zilog->zl_suspend);
1729 ASSERT0(zilog->zl_suspending);
1731 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1732 list_destroy(&zilog->zl_lwb_list);
1734 avl_destroy(&zilog->zl_vdev_tree);
1735 mutex_destroy(&zilog->zl_vdev_lock);
1737 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1738 list_destroy(&zilog->zl_itx_commit_list);
1740 for (int i = 0; i < TXG_SIZE; i++) {
1742 * It's possible for an itx to be generated that doesn't dirty
1743 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1744 * callback to remove the entry. We remove those here.
1746 * Also free up the ziltest itxs.
1748 if (zilog->zl_itxg[i].itxg_itxs)
1749 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1750 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1753 mutex_destroy(&zilog->zl_lock);
1755 cv_destroy(&zilog->zl_cv_writer);
1756 cv_destroy(&zilog->zl_cv_suspend);
1757 cv_destroy(&zilog->zl_cv_batch[0]);
1758 cv_destroy(&zilog->zl_cv_batch[1]);
1760 kmem_free(zilog, sizeof (zilog_t));
1764 * Open an intent log.
1767 zil_open(objset_t *os, zil_get_data_t *get_data)
1769 zilog_t *zilog = dmu_objset_zil(os);
1771 ASSERT(zilog->zl_clean_taskq == NULL);
1772 ASSERT(zilog->zl_get_data == NULL);
1773 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1775 zilog->zl_get_data = get_data;
1776 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1777 2, 2, TASKQ_PREPOPULATE);
1783 * Close an intent log.
1786 zil_close(zilog_t *zilog)
1791 zil_commit(zilog, 0); /* commit all itx */
1794 * The lwb_max_txg for the stubby lwb will reflect the last activity
1795 * for the zil. After a txg_wait_synced() on the txg we know all the
1796 * callbacks have occurred that may clean the zil. Only then can we
1797 * destroy the zl_clean_taskq.
1799 mutex_enter(&zilog->zl_lock);
1800 lwb = list_tail(&zilog->zl_lwb_list);
1802 txg = lwb->lwb_max_txg;
1803 mutex_exit(&zilog->zl_lock);
1805 txg_wait_synced(zilog->zl_dmu_pool, txg);
1806 ASSERT(!zilog_is_dirty(zilog));
1808 taskq_destroy(zilog->zl_clean_taskq);
1809 zilog->zl_clean_taskq = NULL;
1810 zilog->zl_get_data = NULL;
1813 * We should have only one LWB left on the list; remove it now.
1815 mutex_enter(&zilog->zl_lock);
1816 lwb = list_head(&zilog->zl_lwb_list);
1818 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1819 list_remove(&zilog->zl_lwb_list, lwb);
1820 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1821 kmem_cache_free(zil_lwb_cache, lwb);
1823 mutex_exit(&zilog->zl_lock);
1826 static char *suspend_tag = "zil suspending";
1829 * Suspend an intent log. While in suspended mode, we still honor
1830 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1831 * On old version pools, we suspend the log briefly when taking a
1832 * snapshot so that it will have an empty intent log.
1834 * Long holds are not really intended to be used the way we do here --
1835 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1836 * could fail. Therefore we take pains to only put a long hold if it is
1837 * actually necessary. Fortunately, it will only be necessary if the
1838 * objset is currently mounted (or the ZVOL equivalent). In that case it
1839 * will already have a long hold, so we are not really making things any worse.
1841 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1842 * zvol_state_t), and use their mechanism to prevent their hold from being
1843 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1846 * if cookiep == NULL, this does both the suspend & resume.
1847 * Otherwise, it returns with the dataset "long held", and the cookie
1848 * should be passed into zil_resume().
1851 zil_suspend(const char *osname, void **cookiep)
1855 const zil_header_t *zh;
1858 error = dmu_objset_hold(osname, suspend_tag, &os);
1861 zilog = dmu_objset_zil(os);
1863 mutex_enter(&zilog->zl_lock);
1864 zh = zilog->zl_header;
1866 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1867 mutex_exit(&zilog->zl_lock);
1868 dmu_objset_rele(os, suspend_tag);
1869 return (SET_ERROR(EBUSY));
1873 * Don't put a long hold in the cases where we can avoid it. This
1874 * is when there is no cookie so we are doing a suspend & resume
1875 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1876 * for the suspend because it's already suspended, or there's no ZIL.
1878 if (cookiep == NULL && !zilog->zl_suspending &&
1879 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1880 mutex_exit(&zilog->zl_lock);
1881 dmu_objset_rele(os, suspend_tag);
1885 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1886 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1888 zilog->zl_suspend++;
1890 if (zilog->zl_suspend > 1) {
1892 * Someone else is already suspending it.
1893 * Just wait for them to finish.
1896 while (zilog->zl_suspending)
1897 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1898 mutex_exit(&zilog->zl_lock);
1900 if (cookiep == NULL)
1908 * If there is no pointer to an on-disk block, this ZIL must not
1909 * be active (e.g. filesystem not mounted), so there's nothing
1912 if (BP_IS_HOLE(&zh->zh_log)) {
1913 ASSERT(cookiep != NULL); /* fast path already handled */
1916 mutex_exit(&zilog->zl_lock);
1920 zilog->zl_suspending = B_TRUE;
1921 mutex_exit(&zilog->zl_lock);
1923 zil_commit(zilog, 0);
1925 zil_destroy(zilog, B_FALSE);
1927 mutex_enter(&zilog->zl_lock);
1928 zilog->zl_suspending = B_FALSE;
1929 cv_broadcast(&zilog->zl_cv_suspend);
1930 mutex_exit(&zilog->zl_lock);
1932 if (cookiep == NULL)
1940 zil_resume(void *cookie)
1942 objset_t *os = cookie;
1943 zilog_t *zilog = dmu_objset_zil(os);
1945 mutex_enter(&zilog->zl_lock);
1946 ASSERT(zilog->zl_suspend != 0);
1947 zilog->zl_suspend--;
1948 mutex_exit(&zilog->zl_lock);
1949 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1950 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1953 typedef struct zil_replay_arg {
1954 zil_replay_func_t **zr_replay;
1956 boolean_t zr_byteswap;
1961 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1963 char name[MAXNAMELEN];
1965 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1967 dmu_objset_name(zilog->zl_os, name);
1969 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1970 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1971 (u_longlong_t)lr->lrc_seq,
1972 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1973 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1979 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1981 zil_replay_arg_t *zr = zra;
1982 const zil_header_t *zh = zilog->zl_header;
1983 uint64_t reclen = lr->lrc_reclen;
1984 uint64_t txtype = lr->lrc_txtype;
1987 zilog->zl_replaying_seq = lr->lrc_seq;
1989 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1992 if (lr->lrc_txg < claim_txg) /* already committed */
1995 /* Strip case-insensitive bit, still present in log record */
1998 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1999 return (zil_replay_error(zilog, lr, EINVAL));
2002 * If this record type can be logged out of order, the object
2003 * (lr_foid) may no longer exist. That's legitimate, not an error.
2005 if (TX_OOO(txtype)) {
2006 error = dmu_object_info(zilog->zl_os,
2007 ((lr_ooo_t *)lr)->lr_foid, NULL);
2008 if (error == ENOENT || error == EEXIST)
2013 * Make a copy of the data so we can revise and extend it.
2015 bcopy(lr, zr->zr_lr, reclen);
2018 * If this is a TX_WRITE with a blkptr, suck in the data.
2020 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2021 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2022 zr->zr_lr + reclen);
2024 return (zil_replay_error(zilog, lr, error));
2028 * The log block containing this lr may have been byteswapped
2029 * so that we can easily examine common fields like lrc_txtype.
2030 * However, the log is a mix of different record types, and only the
2031 * replay vectors know how to byteswap their records. Therefore, if
2032 * the lr was byteswapped, undo it before invoking the replay vector.
2034 if (zr->zr_byteswap)
2035 byteswap_uint64_array(zr->zr_lr, reclen);
2038 * We must now do two things atomically: replay this log record,
2039 * and update the log header sequence number to reflect the fact that
2040 * we did so. At the end of each replay function the sequence number
2041 * is updated if we are in replay mode.
2043 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2046 * The DMU's dnode layer doesn't see removes until the txg
2047 * commits, so a subsequent claim can spuriously fail with
2048 * EEXIST. So if we receive any error we try syncing out
2049 * any removes then retry the transaction. Note that we
2050 * specify B_FALSE for byteswap now, so we don't do it twice.
2052 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2053 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2055 return (zil_replay_error(zilog, lr, error));
2062 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2064 zilog->zl_replay_blks++;
2070 * If this dataset has a non-empty intent log, replay it and destroy it.
2073 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2075 zilog_t *zilog = dmu_objset_zil(os);
2076 const zil_header_t *zh = zilog->zl_header;
2077 zil_replay_arg_t zr;
2079 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2080 zil_destroy(zilog, B_TRUE);
2083 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2085 zr.zr_replay = replay_func;
2087 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2088 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2091 * Wait for in-progress removes to sync before starting replay.
2093 txg_wait_synced(zilog->zl_dmu_pool, 0);
2095 zilog->zl_replay = B_TRUE;
2096 zilog->zl_replay_time = ddi_get_lbolt();
2097 ASSERT(zilog->zl_replay_blks == 0);
2098 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2100 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2102 zil_destroy(zilog, B_FALSE);
2103 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2104 zilog->zl_replay = B_FALSE;
2105 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2109 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2111 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2114 if (zilog->zl_replay) {
2115 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2116 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2117 zilog->zl_replaying_seq;
2126 zil_vdev_offline(const char *osname, void *arg)
2130 error = zil_suspend(osname, NULL);
2132 return (SET_ERROR(EEXIST));