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 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #include <sys/zfs_context.h>
32 #include <sys/resource.h>
34 #include <sys/zil_impl.h>
35 #include <sys/dsl_dataset.h>
37 #include <sys/dmu_tx.h>
40 * The zfs intent log (ZIL) saves transaction records of system calls
41 * that change the file system in memory with enough information
42 * to be able to replay them. These are stored in memory until
43 * either the DMU transaction group (txg) commits them to the stable pool
44 * and they can be discarded, or they are flushed to the stable log
45 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
46 * requirement. In the event of a panic or power fail then those log
47 * records (transactions) are replayed.
49 * There is one ZIL per file system. Its on-disk (pool) format consists
56 * A log record holds a system call transaction. Log blocks can
57 * hold many log records and the blocks are chained together.
58 * Each ZIL block contains a block pointer (blkptr_t) to the next
59 * ZIL block in the chain. The ZIL header points to the first
60 * block in the chain. Note there is not a fixed place in the pool
61 * to hold blocks. They are dynamically allocated and freed as
62 * needed from the blocks available. Figure X shows the ZIL structure:
66 * This global ZIL switch affects all pools
68 int zil_disable = 0; /* disable intent logging */
71 * Tunable parameter for debugging or performance analysis. Setting
72 * zfs_nocacheflush will cause corruption on power loss if a volatile
73 * out-of-order write cache is enabled.
75 boolean_t zfs_nocacheflush = B_FALSE;
77 static kmem_cache_t *zil_lwb_cache;
80 zil_dva_compare(const void *x1, const void *x2)
82 const dva_t *dva1 = x1;
83 const dva_t *dva2 = x2;
85 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
87 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
90 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
92 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
99 zil_dva_tree_init(avl_tree_t *t)
101 avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t),
102 offsetof(zil_dva_node_t, zn_node));
106 zil_dva_tree_fini(avl_tree_t *t)
111 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
112 kmem_free(zn, sizeof (zil_dva_node_t));
118 zil_dva_tree_add(avl_tree_t *t, dva_t *dva)
123 if (avl_find(t, dva, &where) != NULL)
126 zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP);
128 avl_insert(t, zn, where);
133 static zil_header_t *
134 zil_header_in_syncing_context(zilog_t *zilog)
136 return ((zil_header_t *)zilog->zl_header);
140 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
142 zio_cksum_t *zc = &bp->blk_cksum;
144 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
145 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
146 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
147 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
151 * Read a log block, make sure it's valid, and byteswap it if necessary.
154 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp)
158 uint32_t aflags = ARC_WAIT;
161 zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET];
164 zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
169 * We shouldn't be doing any scrubbing while we're doing log
170 * replay, it's OK to not lock.
172 error = arc_read_nolock(NULL, zilog->zl_spa, &blk,
173 arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL |
174 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb);
177 char *data = (*abufpp)->b_data;
178 uint64_t blksz = BP_GET_LSIZE(bp);
179 zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1;
180 zio_cksum_t cksum = bp->blk_cksum;
183 * Validate the checksummed log block.
185 * Sequence numbers should be... sequential. The checksum
186 * verifier for the next block should be bp's checksum plus 1.
188 * Also check the log chain linkage and size used.
190 cksum.zc_word[ZIL_ZC_SEQ]++;
192 if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum,
193 sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) ||
194 (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) {
199 VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1);
204 dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid);
210 * Parse the intent log, and call parse_func for each valid record within.
211 * Return the highest sequence number.
214 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
215 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
217 const zil_header_t *zh = zilog->zl_header;
218 uint64_t claim_seq = zh->zh_claim_seq;
220 uint64_t max_seq = 0;
221 blkptr_t blk = zh->zh_log;
227 if (BP_IS_HOLE(&blk))
231 * Starting at the block pointed to by zh_log we read the log chain.
232 * For each block in the chain we strongly check that block to
233 * ensure its validity. We stop when an invalid block is found.
234 * For each block pointer in the chain we call parse_blk_func().
235 * For each record in each valid block we call parse_lr_func().
236 * If the log has been claimed, stop if we encounter a sequence
237 * number greater than the highest claimed sequence number.
239 zil_dva_tree_init(&zilog->zl_dva_tree);
241 seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
243 if (claim_seq != 0 && seq > claim_seq)
246 ASSERT(max_seq < seq);
249 error = zil_read_log_block(zilog, &blk, &abuf);
251 if (parse_blk_func != NULL)
252 parse_blk_func(zilog, &blk, arg, txg);
257 lrbuf = abuf->b_data;
258 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
259 blk = ztp->zit_next_blk;
261 if (parse_lr_func == NULL) {
262 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
266 for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) {
267 lr_t *lr = (lr_t *)lrp;
268 reclen = lr->lrc_reclen;
269 ASSERT3U(reclen, >=, sizeof (lr_t));
270 parse_lr_func(zilog, lr, arg, txg);
272 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
274 zil_dva_tree_fini(&zilog->zl_dva_tree);
281 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
283 spa_t *spa = zilog->zl_spa;
287 * Claim log block if not already committed and not already claimed.
289 if (bp->blk_birth >= first_txg &&
290 zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) {
291 err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL,
292 ZIO_FLAG_MUSTSUCCEED));
298 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
300 if (lrc->lrc_txtype == TX_WRITE) {
301 lr_write_t *lr = (lr_write_t *)lrc;
302 zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg);
308 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
310 zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx));
314 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
317 * If we previously claimed it, we need to free it.
319 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) {
320 lr_write_t *lr = (lr_write_t *)lrc;
321 blkptr_t *bp = &lr->lr_blkptr;
322 if (bp->blk_birth >= claim_txg &&
323 !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) {
324 (void) arc_free(NULL, zilog->zl_spa,
325 dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT);
331 * Create an on-disk intent log.
334 zil_create(zilog_t *zilog)
336 const zil_header_t *zh = zilog->zl_header;
344 * Wait for any previous destroy to complete.
346 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
348 ASSERT(zh->zh_claim_txg == 0);
349 ASSERT(zh->zh_replay_seq == 0);
354 * If we don't already have an initial log block or we have one
355 * but it's the wrong endianness then allocate one.
357 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
358 tx = dmu_tx_create(zilog->zl_os);
359 (void) dmu_tx_assign(tx, TXG_WAIT);
360 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
361 txg = dmu_tx_get_txg(tx);
363 if (!BP_IS_HOLE(&blk)) {
364 zio_free_blk(zilog->zl_spa, &blk, txg);
368 error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk,
372 zil_init_log_chain(zilog, &blk);
376 * Allocate a log write buffer (lwb) for the first log block.
379 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
380 lwb->lwb_zilog = zilog;
383 lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk);
384 lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz);
385 lwb->lwb_max_txg = txg;
388 mutex_enter(&zilog->zl_lock);
389 list_insert_tail(&zilog->zl_lwb_list, lwb);
390 mutex_exit(&zilog->zl_lock);
394 * If we just allocated the first log block, commit our transaction
395 * and wait for zil_sync() to stuff the block poiner into zh_log.
396 * (zh is part of the MOS, so we cannot modify it in open context.)
400 txg_wait_synced(zilog->zl_dmu_pool, txg);
403 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
407 * In one tx, free all log blocks and clear the log header.
408 * If keep_first is set, then we're replaying a log with no content.
409 * We want to keep the first block, however, so that the first
410 * synchronous transaction doesn't require a txg_wait_synced()
411 * in zil_create(). We don't need to txg_wait_synced() here either
412 * when keep_first is set, because both zil_create() and zil_destroy()
413 * will wait for any in-progress destroys to complete.
416 zil_destroy(zilog_t *zilog, boolean_t keep_first)
418 const zil_header_t *zh = zilog->zl_header;
424 * Wait for any previous destroy to complete.
426 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
428 if (BP_IS_HOLE(&zh->zh_log))
431 tx = dmu_tx_create(zilog->zl_os);
432 (void) dmu_tx_assign(tx, TXG_WAIT);
433 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
434 txg = dmu_tx_get_txg(tx);
436 mutex_enter(&zilog->zl_lock);
439 * It is possible for the ZIL to get the previously mounted zilog
440 * structure of the same dataset if quickly remounted and the dbuf
441 * eviction has not completed. In this case we can see a non
442 * empty lwb list and keep_first will be set. We fix this by
443 * clearing the keep_first. This will be slower but it's very rare.
445 if (!list_is_empty(&zilog->zl_lwb_list) && keep_first)
446 keep_first = B_FALSE;
448 ASSERT3U(zilog->zl_destroy_txg, <, txg);
449 zilog->zl_destroy_txg = txg;
450 zilog->zl_keep_first = keep_first;
452 if (!list_is_empty(&zilog->zl_lwb_list)) {
453 ASSERT(zh->zh_claim_txg == 0);
455 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
456 list_remove(&zilog->zl_lwb_list, lwb);
457 if (lwb->lwb_buf != NULL)
458 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
459 zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg);
460 kmem_cache_free(zil_lwb_cache, lwb);
464 (void) zil_parse(zilog, zil_free_log_block,
465 zil_free_log_record, tx, zh->zh_claim_txg);
468 mutex_exit(&zilog->zl_lock);
474 * zil_rollback_destroy() is only called by the rollback code.
475 * We already have a syncing tx. Rollback has exclusive access to the
476 * dataset, so we don't have to worry about concurrent zil access.
477 * The actual freeing of any log blocks occurs in zil_sync() later in
478 * this txg syncing phase.
481 zil_rollback_destroy(zilog_t *zilog, dmu_tx_t *tx)
483 const zil_header_t *zh = zilog->zl_header;
486 if (BP_IS_HOLE(&zh->zh_log))
489 txg = dmu_tx_get_txg(tx);
490 ASSERT3U(zilog->zl_destroy_txg, <, txg);
491 zilog->zl_destroy_txg = txg;
492 zilog->zl_keep_first = B_FALSE;
495 * Ensure there's no outstanding ZIL IO. No lwbs or just the
496 * unused one that allocated in advance is ok.
498 ASSERT(zilog->zl_lwb_list.list_head.list_next ==
499 zilog->zl_lwb_list.list_head.list_prev);
500 (void) zil_parse(zilog, zil_free_log_block, zil_free_log_record,
501 tx, zh->zh_claim_txg);
505 zil_claim(char *osname, void *txarg)
507 dmu_tx_t *tx = txarg;
508 uint64_t first_txg = dmu_tx_get_txg(tx);
514 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
516 cmn_err(CE_WARN, "can't open objset for %s", osname);
520 zilog = dmu_objset_zil(os);
521 zh = zil_header_in_syncing_context(zilog);
524 * Claim all log blocks if we haven't already done so, and remember
525 * the highest claimed sequence number. This ensures that if we can
526 * read only part of the log now (e.g. due to a missing device),
527 * but we can read the entire log later, we will not try to replay
528 * or destroy beyond the last block we successfully claimed.
530 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
531 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
532 zh->zh_claim_txg = first_txg;
533 zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block,
534 zil_claim_log_record, tx, first_txg);
535 dsl_dataset_dirty(dmu_objset_ds(os), tx);
538 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
539 dmu_objset_close(os);
544 * Check the log by walking the log chain.
545 * Checksum errors are ok as they indicate the end of the chain.
546 * Any other error (no device or read failure) returns an error.
550 zil_check_log_chain(char *osname, void *txarg)
561 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
563 cmn_err(CE_WARN, "can't open objset for %s", osname);
567 zilog = dmu_objset_zil(os);
568 zh = zil_header_in_syncing_context(zilog);
570 if (BP_IS_HOLE(&blk)) {
571 dmu_objset_close(os);
572 return (0); /* no chain */
576 error = zil_read_log_block(zilog, &blk, &abuf);
579 lrbuf = abuf->b_data;
580 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
581 blk = ztp->zit_next_blk;
582 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
584 dmu_objset_close(os);
586 return (0); /* normal end of chain */
595 zil_clear_log_chain(char *osname, void *txarg)
603 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
605 cmn_err(CE_WARN, "can't open objset for %s", osname);
609 zilog = dmu_objset_zil(os);
610 tx = dmu_tx_create(zilog->zl_os);
611 (void) dmu_tx_assign(tx, TXG_WAIT);
612 zh = zil_header_in_syncing_context(zilog);
613 BP_ZERO(&zh->zh_log);
614 dsl_dataset_dirty(dmu_objset_ds(os), tx);
616 dmu_objset_close(os);
621 zil_vdev_compare(const void *x1, const void *x2)
623 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
624 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
635 zil_add_block(zilog_t *zilog, blkptr_t *bp)
637 avl_tree_t *t = &zilog->zl_vdev_tree;
639 zil_vdev_node_t *zv, zvsearch;
640 int ndvas = BP_GET_NDVAS(bp);
643 if (zfs_nocacheflush)
646 ASSERT(zilog->zl_writer);
649 * Even though we're zl_writer, we still need a lock because the
650 * zl_get_data() callbacks may have dmu_sync() done callbacks
651 * that will run concurrently.
653 mutex_enter(&zilog->zl_vdev_lock);
654 for (i = 0; i < ndvas; i++) {
655 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
656 if (avl_find(t, &zvsearch, &where) == NULL) {
657 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
658 zv->zv_vdev = zvsearch.zv_vdev;
659 avl_insert(t, zv, where);
662 mutex_exit(&zilog->zl_vdev_lock);
666 zil_flush_vdevs(zilog_t *zilog)
668 spa_t *spa = zilog->zl_spa;
669 avl_tree_t *t = &zilog->zl_vdev_tree;
674 ASSERT(zilog->zl_writer);
677 * We don't need zl_vdev_lock here because we're the zl_writer,
678 * and all zl_get_data() callbacks are done.
680 if (avl_numnodes(t) == 0)
683 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
685 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
687 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
688 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
691 kmem_free(zv, sizeof (*zv));
695 * Wait for all the flushes to complete. Not all devices actually
696 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
698 (void) zio_wait(zio);
700 spa_config_exit(spa, SCL_STATE, FTAG);
704 * Function called when a log block write completes
707 zil_lwb_write_done(zio_t *zio)
709 lwb_t *lwb = zio->io_private;
710 zilog_t *zilog = lwb->lwb_zilog;
712 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
713 ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG);
714 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
715 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
716 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
717 ASSERT(!BP_IS_GANG(zio->io_bp));
718 ASSERT(!BP_IS_HOLE(zio->io_bp));
719 ASSERT(zio->io_bp->blk_fill == 0);
722 * Now that we've written this log block, we have a stable pointer
723 * to the next block in the chain, so it's OK to let the txg in
724 * which we allocated the next block sync.
726 txg_rele_to_sync(&lwb->lwb_txgh);
728 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
729 mutex_enter(&zilog->zl_lock);
732 zilog->zl_log_error = B_TRUE;
733 mutex_exit(&zilog->zl_lock);
737 * Initialize the io for a log block.
740 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
744 zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET];
747 zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
749 if (zilog->zl_root_zio == NULL) {
750 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
753 if (lwb->lwb_zio == NULL) {
754 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
755 0, &lwb->lwb_blk, lwb->lwb_buf,
756 lwb->lwb_sz, zil_lwb_write_done, lwb,
757 ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb);
762 * Start a log block write and advance to the next log block.
763 * Calls are serialized.
766 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
769 zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1;
770 spa_t *spa = zilog->zl_spa;
771 blkptr_t *bp = &ztp->zit_next_blk;
776 ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb));
779 * Allocate the next block and save its address in this block
780 * before writing it in order to establish the log chain.
781 * Note that if the allocation of nlwb synced before we wrote
782 * the block that points at it (lwb), we'd leak it if we crashed.
783 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done().
785 txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh);
786 txg_rele_to_quiesce(&lwb->lwb_txgh);
789 * Pick a ZIL blocksize. We request a size that is the
790 * maximum of the previous used size, the current used size and
791 * the amount waiting in the queue.
793 zil_blksz = MAX(zilog->zl_prev_used,
794 zilog->zl_cur_used + sizeof (*ztp));
795 zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp));
796 zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t);
797 if (zil_blksz > ZIL_MAX_BLKSZ)
798 zil_blksz = ZIL_MAX_BLKSZ;
801 /* pass the old blkptr in order to spread log blocks across devs */
802 error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg);
804 dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg);
807 * We dirty the dataset to ensure that zil_sync() will
808 * be called to remove this lwb from our zl_lwb_list.
809 * Failing to do so, may leave an lwb with a NULL lwb_buf
810 * hanging around on the zl_lwb_list.
812 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
816 * Since we've just experienced an allocation failure so we
817 * terminate the current lwb and send it on its way.
820 ztp->zit_nused = lwb->lwb_nused;
821 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
822 zio_nowait(lwb->lwb_zio);
825 * By returning NULL the caller will call tx_wait_synced()
830 ASSERT3U(bp->blk_birth, ==, txg);
832 ztp->zit_nused = lwb->lwb_nused;
833 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
834 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
835 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
838 * Allocate a new log write buffer (lwb).
840 nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
842 nlwb->lwb_zilog = zilog;
845 nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk);
846 nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz);
847 nlwb->lwb_max_txg = txg;
848 nlwb->lwb_zio = NULL;
851 * Put new lwb at the end of the log chain
853 mutex_enter(&zilog->zl_lock);
854 list_insert_tail(&zilog->zl_lwb_list, nlwb);
855 mutex_exit(&zilog->zl_lock);
857 /* Record the block for later vdev flushing */
858 zil_add_block(zilog, &lwb->lwb_blk);
861 * kick off the write for the old log block
863 dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg);
864 ASSERT(lwb->lwb_zio);
865 zio_nowait(lwb->lwb_zio);
871 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
873 lr_t *lrc = &itx->itx_lr; /* common log record */
874 lr_write_t *lr = (lr_write_t *)lrc;
875 uint64_t txg = lrc->lrc_txg;
876 uint64_t reclen = lrc->lrc_reclen;
881 ASSERT(lwb->lwb_buf != NULL);
883 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
884 dlen = P2ROUNDUP_TYPED(
885 lr->lr_length, sizeof (uint64_t), uint64_t);
889 zilog->zl_cur_used += (reclen + dlen);
891 zil_lwb_write_init(zilog, lwb);
894 * If this record won't fit in the current log block, start a new one.
896 if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
897 lwb = zil_lwb_write_start(zilog, lwb);
900 zil_lwb_write_init(zilog, lwb);
901 ASSERT(lwb->lwb_nused == 0);
902 if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
903 txg_wait_synced(zilog->zl_dmu_pool, txg);
909 * Update the lrc_seq, to be log record sequence number. See zil.h
910 * Then copy the record to the log buffer.
912 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
913 bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen);
916 * If it's a write, fetch the data or get its blkptr as appropriate.
918 if (lrc->lrc_txtype == TX_WRITE) {
919 if (txg > spa_freeze_txg(zilog->zl_spa))
920 txg_wait_synced(zilog->zl_dmu_pool, txg);
921 if (itx->itx_wr_state != WR_COPIED) {
925 /* alignment is guaranteed */
926 lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused);
928 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
929 dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen;
930 lr->lr_common.lrc_reclen += dlen;
932 ASSERT(itx->itx_wr_state == WR_INDIRECT);
935 error = zilog->zl_get_data(
936 itx->itx_private, lr, dbuf, lwb->lwb_zio);
938 ASSERT(error == ENOENT || error == EEXIST ||
945 lwb->lwb_nused += reclen + dlen;
946 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
947 ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb));
948 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
954 zil_itx_create(uint64_t txtype, size_t lrsize)
958 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
960 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
961 itx->itx_lr.lrc_txtype = txtype;
962 itx->itx_lr.lrc_reclen = lrsize;
963 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
964 itx->itx_lr.lrc_seq = 0; /* defensive */
970 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
974 ASSERT(itx->itx_lr.lrc_seq == 0);
976 mutex_enter(&zilog->zl_lock);
977 list_insert_tail(&zilog->zl_itx_list, itx);
978 zilog->zl_itx_list_sz += itx->itx_sod;
979 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
980 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
981 mutex_exit(&zilog->zl_lock);
987 * Free up all in-memory intent log transactions that have now been synced.
990 zil_itx_clean(zilog_t *zilog)
992 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
993 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
997 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
999 mutex_enter(&zilog->zl_lock);
1000 /* wait for a log writer to finish walking list */
1001 while (zilog->zl_writer) {
1002 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1006 * Move the sync'd log transactions to a separate list so we can call
1007 * kmem_free without holding the zl_lock.
1009 * There is no need to set zl_writer as we don't drop zl_lock here
1011 while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
1012 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
1013 list_remove(&zilog->zl_itx_list, itx);
1014 zilog->zl_itx_list_sz -= itx->itx_sod;
1015 list_insert_tail(&clean_list, itx);
1017 cv_broadcast(&zilog->zl_cv_writer);
1018 mutex_exit(&zilog->zl_lock);
1020 /* destroy sync'd log transactions */
1021 while ((itx = list_head(&clean_list)) != NULL) {
1022 list_remove(&clean_list, itx);
1023 kmem_free(itx, offsetof(itx_t, itx_lr)
1024 + itx->itx_lr.lrc_reclen);
1026 list_destroy(&clean_list);
1030 * If there are any in-memory intent log transactions which have now been
1031 * synced then start up a taskq to free them.
1034 zil_clean(zilog_t *zilog)
1038 mutex_enter(&zilog->zl_lock);
1039 itx = list_head(&zilog->zl_itx_list);
1040 if ((itx != NULL) &&
1041 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
1042 (void) taskq_dispatch(zilog->zl_clean_taskq,
1043 (void (*)(void *))zil_itx_clean, zilog, TQ_NOSLEEP);
1045 mutex_exit(&zilog->zl_lock);
1049 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
1052 uint64_t commit_seq = 0;
1053 itx_t *itx, *itx_next = (itx_t *)-1;
1057 zilog->zl_writer = B_TRUE;
1058 ASSERT(zilog->zl_root_zio == NULL);
1059 spa = zilog->zl_spa;
1061 if (zilog->zl_suspend) {
1064 lwb = list_tail(&zilog->zl_lwb_list);
1067 * Return if there's nothing to flush before we
1068 * dirty the fs by calling zil_create()
1070 if (list_is_empty(&zilog->zl_itx_list)) {
1071 zilog->zl_writer = B_FALSE;
1074 mutex_exit(&zilog->zl_lock);
1076 mutex_enter(&zilog->zl_lock);
1077 lwb = list_tail(&zilog->zl_lwb_list);
1081 /* Loop through in-memory log transactions filling log blocks. */
1082 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1085 * Find the next itx to push:
1086 * Push all transactions related to specified foid and all
1087 * other transactions except TX_WRITE, TX_TRUNCATE,
1088 * TX_SETATTR and TX_ACL for all other files.
1090 if (itx_next != (itx_t *)-1)
1093 itx = list_head(&zilog->zl_itx_list);
1094 for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) {
1095 if (foid == 0) /* push all foids? */
1097 if (itx->itx_sync) /* push all O_[D]SYNC */
1099 switch (itx->itx_lr.lrc_txtype) {
1104 /* lr_foid is same offset for these records */
1105 if (((lr_write_t *)&itx->itx_lr)->lr_foid
1107 continue; /* skip this record */
1115 if ((itx->itx_lr.lrc_seq > seq) &&
1116 ((lwb == NULL) || (lwb->lwb_nused == 0) ||
1117 (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) {
1122 * Save the next pointer. Even though we soon drop
1123 * zl_lock all threads that may change the list
1124 * (another writer or zil_itx_clean) can't do so until
1125 * they have zl_writer.
1127 itx_next = list_next(&zilog->zl_itx_list, itx);
1128 list_remove(&zilog->zl_itx_list, itx);
1129 zilog->zl_itx_list_sz -= itx->itx_sod;
1130 mutex_exit(&zilog->zl_lock);
1131 txg = itx->itx_lr.lrc_txg;
1134 if (txg > spa_last_synced_txg(spa) ||
1135 txg > spa_freeze_txg(spa))
1136 lwb = zil_lwb_commit(zilog, itx, lwb);
1137 kmem_free(itx, offsetof(itx_t, itx_lr)
1138 + itx->itx_lr.lrc_reclen);
1139 mutex_enter(&zilog->zl_lock);
1141 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1142 /* determine commit sequence number */
1143 itx = list_head(&zilog->zl_itx_list);
1145 commit_seq = itx->itx_lr.lrc_seq;
1147 commit_seq = zilog->zl_itx_seq;
1148 mutex_exit(&zilog->zl_lock);
1150 /* write the last block out */
1151 if (lwb != NULL && lwb->lwb_zio != NULL)
1152 lwb = zil_lwb_write_start(zilog, lwb);
1154 zilog->zl_prev_used = zilog->zl_cur_used;
1155 zilog->zl_cur_used = 0;
1158 * Wait if necessary for the log blocks to be on stable storage.
1160 if (zilog->zl_root_zio) {
1161 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1162 (void) zio_wait(zilog->zl_root_zio);
1163 zilog->zl_root_zio = NULL;
1164 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1165 zil_flush_vdevs(zilog);
1168 if (zilog->zl_log_error || lwb == NULL) {
1169 zilog->zl_log_error = 0;
1170 txg_wait_synced(zilog->zl_dmu_pool, 0);
1173 mutex_enter(&zilog->zl_lock);
1174 zilog->zl_writer = B_FALSE;
1176 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1177 zilog->zl_commit_seq = commit_seq;
1181 * Push zfs transactions to stable storage up to the supplied sequence number.
1182 * If foid is 0 push out all transactions, otherwise push only those
1183 * for that file or might have been used to create that file.
1186 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1188 if (zilog == NULL || seq == 0)
1191 mutex_enter(&zilog->zl_lock);
1193 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */
1195 while (zilog->zl_writer) {
1196 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1197 if (seq < zilog->zl_commit_seq) {
1198 mutex_exit(&zilog->zl_lock);
1202 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1203 /* wake up others waiting on the commit */
1204 cv_broadcast(&zilog->zl_cv_writer);
1205 mutex_exit(&zilog->zl_lock);
1209 * Called in syncing context to free committed log blocks and update log header.
1212 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1214 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1215 uint64_t txg = dmu_tx_get_txg(tx);
1216 spa_t *spa = zilog->zl_spa;
1219 mutex_enter(&zilog->zl_lock);
1221 ASSERT(zilog->zl_stop_sync == 0);
1223 zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK];
1225 if (zilog->zl_destroy_txg == txg) {
1226 blkptr_t blk = zh->zh_log;
1228 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1229 ASSERT(spa_sync_pass(spa) == 1);
1231 bzero(zh, sizeof (zil_header_t));
1232 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1234 if (zilog->zl_keep_first) {
1236 * If this block was part of log chain that couldn't
1237 * be claimed because a device was missing during
1238 * zil_claim(), but that device later returns,
1239 * then this block could erroneously appear valid.
1240 * To guard against this, assign a new GUID to the new
1241 * log chain so it doesn't matter what blk points to.
1243 zil_init_log_chain(zilog, &blk);
1249 lwb = list_head(&zilog->zl_lwb_list);
1251 mutex_exit(&zilog->zl_lock);
1254 zh->zh_log = lwb->lwb_blk;
1255 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1257 list_remove(&zilog->zl_lwb_list, lwb);
1258 zio_free_blk(spa, &lwb->lwb_blk, txg);
1259 kmem_cache_free(zil_lwb_cache, lwb);
1262 * If we don't have anything left in the lwb list then
1263 * we've had an allocation failure and we need to zero
1264 * out the zil_header blkptr so that we don't end
1265 * up freeing the same block twice.
1267 if (list_head(&zilog->zl_lwb_list) == NULL)
1268 BP_ZERO(&zh->zh_log);
1270 mutex_exit(&zilog->zl_lock);
1276 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1277 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1283 kmem_cache_destroy(zil_lwb_cache);
1287 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1291 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1293 zilog->zl_header = zh_phys;
1295 zilog->zl_spa = dmu_objset_spa(os);
1296 zilog->zl_dmu_pool = dmu_objset_pool(os);
1297 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1299 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1301 list_create(&zilog->zl_itx_list, sizeof (itx_t),
1302 offsetof(itx_t, itx_node));
1304 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1305 offsetof(lwb_t, lwb_node));
1307 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1309 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1310 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1312 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1313 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1319 zil_free(zilog_t *zilog)
1323 zilog->zl_stop_sync = 1;
1325 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1326 list_remove(&zilog->zl_lwb_list, lwb);
1327 if (lwb->lwb_buf != NULL)
1328 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1329 kmem_cache_free(zil_lwb_cache, lwb);
1331 list_destroy(&zilog->zl_lwb_list);
1333 avl_destroy(&zilog->zl_vdev_tree);
1334 mutex_destroy(&zilog->zl_vdev_lock);
1336 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1337 list_destroy(&zilog->zl_itx_list);
1338 mutex_destroy(&zilog->zl_lock);
1340 cv_destroy(&zilog->zl_cv_writer);
1341 cv_destroy(&zilog->zl_cv_suspend);
1343 kmem_free(zilog, sizeof (zilog_t));
1347 * return true if the initial log block is not valid
1350 zil_empty(zilog_t *zilog)
1352 const zil_header_t *zh = zilog->zl_header;
1353 arc_buf_t *abuf = NULL;
1355 if (BP_IS_HOLE(&zh->zh_log))
1358 if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0)
1361 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1366 * Open an intent log.
1369 zil_open(objset_t *os, zil_get_data_t *get_data)
1371 zilog_t *zilog = dmu_objset_zil(os);
1373 zilog->zl_get_data = get_data;
1374 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1375 2, 2, TASKQ_PREPOPULATE);
1381 * Close an intent log.
1384 zil_close(zilog_t *zilog)
1387 * If the log isn't already committed, mark the objset dirty
1388 * (so zil_sync() will be called) and wait for that txg to sync.
1390 if (!zil_is_committed(zilog)) {
1392 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1393 (void) dmu_tx_assign(tx, TXG_WAIT);
1394 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1395 txg = dmu_tx_get_txg(tx);
1397 txg_wait_synced(zilog->zl_dmu_pool, txg);
1400 taskq_destroy(zilog->zl_clean_taskq);
1401 zilog->zl_clean_taskq = NULL;
1402 zilog->zl_get_data = NULL;
1404 zil_itx_clean(zilog);
1405 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1409 * Suspend an intent log. While in suspended mode, we still honor
1410 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1411 * We suspend the log briefly when taking a snapshot so that the snapshot
1412 * contains all the data it's supposed to, and has an empty intent log.
1415 zil_suspend(zilog_t *zilog)
1417 const zil_header_t *zh = zilog->zl_header;
1419 mutex_enter(&zilog->zl_lock);
1420 if (zh->zh_claim_txg != 0) { /* unplayed log */
1421 mutex_exit(&zilog->zl_lock);
1424 if (zilog->zl_suspend++ != 0) {
1426 * Someone else already began a suspend.
1427 * Just wait for them to finish.
1429 while (zilog->zl_suspending)
1430 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1431 mutex_exit(&zilog->zl_lock);
1434 zilog->zl_suspending = B_TRUE;
1435 mutex_exit(&zilog->zl_lock);
1437 zil_commit(zilog, UINT64_MAX, 0);
1440 * Wait for any in-flight log writes to complete.
1442 mutex_enter(&zilog->zl_lock);
1443 while (zilog->zl_writer)
1444 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1445 mutex_exit(&zilog->zl_lock);
1447 zil_destroy(zilog, B_FALSE);
1449 mutex_enter(&zilog->zl_lock);
1450 zilog->zl_suspending = B_FALSE;
1451 cv_broadcast(&zilog->zl_cv_suspend);
1452 mutex_exit(&zilog->zl_lock);
1458 zil_resume(zilog_t *zilog)
1460 mutex_enter(&zilog->zl_lock);
1461 ASSERT(zilog->zl_suspend != 0);
1462 zilog->zl_suspend--;
1463 mutex_exit(&zilog->zl_lock);
1466 typedef struct zil_replay_arg {
1468 zil_replay_func_t **zr_replay;
1470 boolean_t zr_byteswap;
1475 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1477 zil_replay_arg_t *zr = zra;
1478 const zil_header_t *zh = zilog->zl_header;
1479 uint64_t reclen = lr->lrc_reclen;
1480 uint64_t txtype = lr->lrc_txtype;
1484 if (!zilog->zl_replay) /* giving up */
1487 if (lr->lrc_txg < claim_txg) /* already committed */
1490 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1493 /* Strip case-insensitive bit, still present in log record */
1496 if (txtype == 0 || txtype >= TX_MAX_TYPE) {
1502 * Make a copy of the data so we can revise and extend it.
1504 bcopy(lr, zr->zr_lrbuf, reclen);
1507 * The log block containing this lr may have been byteswapped
1508 * so that we can easily examine common fields like lrc_txtype.
1509 * However, the log is a mix of different data types, and only the
1510 * replay vectors know how to byteswap their records. Therefore, if
1511 * the lr was byteswapped, undo it before invoking the replay vector.
1513 if (zr->zr_byteswap)
1514 byteswap_uint64_array(zr->zr_lrbuf, reclen);
1517 * If this is a TX_WRITE with a blkptr, suck in the data.
1519 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1520 lr_write_t *lrw = (lr_write_t *)lr;
1521 blkptr_t *wbp = &lrw->lr_blkptr;
1522 uint64_t wlen = lrw->lr_length;
1523 char *wbuf = zr->zr_lrbuf + reclen;
1525 if (BP_IS_HOLE(wbp)) { /* compressed to a hole */
1529 * A subsequent write may have overwritten this block,
1530 * in which case wbp may have been been freed and
1531 * reallocated, and our read of wbp may fail with a
1532 * checksum error. We can safely ignore this because
1533 * the later write will provide the correct data.
1537 zb.zb_objset = dmu_objset_id(zilog->zl_os);
1538 zb.zb_object = lrw->lr_foid;
1540 zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp);
1542 (void) zio_wait(zio_read(NULL, zilog->zl_spa,
1543 wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL,
1544 ZIO_PRIORITY_SYNC_READ,
1545 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb));
1546 (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen);
1551 * We must now do two things atomically: replay this log record,
1552 * and update the log header sequence number to reflect the fact that
1553 * we did so. At the end of each replay function the sequence number
1554 * is updated if we are in replay mode.
1556 for (pass = 1; pass <= 2; pass++) {
1557 zilog->zl_replaying_seq = lr->lrc_seq;
1558 /* Only byteswap (if needed) on the 1st pass. */
1559 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
1560 zr->zr_byteswap && pass == 1);
1566 * The DMU's dnode layer doesn't see removes until the txg
1567 * commits, so a subsequent claim can spuriously fail with
1568 * EEXIST. So if we receive any error we try syncing out
1569 * any removes then retry the transaction.
1572 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1577 name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1578 dmu_objset_name(zr->zr_os, name);
1579 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1580 "dataset %s, seq 0x%llx, txtype %llu %s\n",
1581 error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype,
1582 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1583 zilog->zl_replay = B_FALSE;
1584 kmem_free(name, MAXNAMELEN);
1589 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1591 zilog->zl_replay_blks++;
1595 * If this dataset has a non-empty intent log, replay it and destroy it.
1598 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1600 zilog_t *zilog = dmu_objset_zil(os);
1601 const zil_header_t *zh = zilog->zl_header;
1602 zil_replay_arg_t zr;
1604 if (zil_empty(zilog)) {
1605 zil_destroy(zilog, B_TRUE);
1610 zr.zr_replay = replay_func;
1612 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1613 zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1616 * Wait for in-progress removes to sync before starting replay.
1618 txg_wait_synced(zilog->zl_dmu_pool, 0);
1620 zilog->zl_replay = B_TRUE;
1621 zilog->zl_replay_time = lbolt;
1622 ASSERT(zilog->zl_replay_blks == 0);
1623 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1625 kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE);
1627 zil_destroy(zilog, B_FALSE);
1628 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1629 zilog->zl_replay = B_FALSE;
1633 * Report whether all transactions are committed
1636 zil_is_committed(zilog_t *zilog)
1641 mutex_enter(&zilog->zl_lock);
1642 while (zilog->zl_writer)
1643 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1645 /* recent unpushed intent log transactions? */
1646 if (!list_is_empty(&zilog->zl_itx_list)) {
1651 /* intent log never used? */
1652 lwb = list_head(&zilog->zl_lwb_list);
1659 * more than 1 log buffer means zil_sync() hasn't yet freed
1660 * entries after a txg has committed
1662 if (list_next(&zilog->zl_lwb_list, lwb)) {
1667 ASSERT(zil_empty(zilog));
1670 cv_broadcast(&zilog->zl_cv_writer);
1671 mutex_exit(&zilog->zl_lock);