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 */
69 SYSCTL_DECL(_vfs_zfs);
70 TUNABLE_INT("vfs.zfs.zil_disable", &zil_disable);
71 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_disable, CTLFLAG_RW, &zil_disable, 0,
72 "Disable ZFS Intent Log (ZIL)");
75 * Tunable parameter for debugging or performance analysis. Setting
76 * zfs_nocacheflush will cause corruption on power loss if a volatile
77 * out-of-order write cache is enabled.
79 boolean_t zfs_nocacheflush = B_FALSE;
80 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
81 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
82 &zfs_nocacheflush, 0, "Disable cache flush");
84 static kmem_cache_t *zil_lwb_cache;
87 zil_dva_compare(const void *x1, const void *x2)
89 const dva_t *dva1 = x1;
90 const dva_t *dva2 = x2;
92 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
94 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
97 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
99 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
106 zil_dva_tree_init(avl_tree_t *t)
108 avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t),
109 offsetof(zil_dva_node_t, zn_node));
113 zil_dva_tree_fini(avl_tree_t *t)
118 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
119 kmem_free(zn, sizeof (zil_dva_node_t));
125 zil_dva_tree_add(avl_tree_t *t, dva_t *dva)
130 if (avl_find(t, dva, &where) != NULL)
133 zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP);
135 avl_insert(t, zn, where);
140 static zil_header_t *
141 zil_header_in_syncing_context(zilog_t *zilog)
143 return ((zil_header_t *)zilog->zl_header);
147 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
149 zio_cksum_t *zc = &bp->blk_cksum;
151 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
152 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
153 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
154 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
158 * Read a log block, make sure it's valid, and byteswap it if necessary.
161 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp)
165 uint32_t aflags = ARC_WAIT;
168 zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET];
171 zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
176 * We shouldn't be doing any scrubbing while we're doing log
177 * replay, it's OK to not lock.
179 error = arc_read_nolock(NULL, zilog->zl_spa, &blk,
180 arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL |
181 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb);
184 char *data = (*abufpp)->b_data;
185 uint64_t blksz = BP_GET_LSIZE(bp);
186 zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1;
187 zio_cksum_t cksum = bp->blk_cksum;
190 * Validate the checksummed log block.
192 * Sequence numbers should be... sequential. The checksum
193 * verifier for the next block should be bp's checksum plus 1.
195 * Also check the log chain linkage and size used.
197 cksum.zc_word[ZIL_ZC_SEQ]++;
199 if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum,
200 sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) ||
201 (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) {
206 VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1);
211 dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid);
217 * Parse the intent log, and call parse_func for each valid record within.
218 * Return the highest sequence number.
221 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
222 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
224 const zil_header_t *zh = zilog->zl_header;
225 uint64_t claim_seq = zh->zh_claim_seq;
227 uint64_t max_seq = 0;
228 blkptr_t blk = zh->zh_log;
234 if (BP_IS_HOLE(&blk))
238 * Starting at the block pointed to by zh_log we read the log chain.
239 * For each block in the chain we strongly check that block to
240 * ensure its validity. We stop when an invalid block is found.
241 * For each block pointer in the chain we call parse_blk_func().
242 * For each record in each valid block we call parse_lr_func().
243 * If the log has been claimed, stop if we encounter a sequence
244 * number greater than the highest claimed sequence number.
246 zil_dva_tree_init(&zilog->zl_dva_tree);
248 seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
250 if (claim_seq != 0 && seq > claim_seq)
253 ASSERT(max_seq < seq);
256 error = zil_read_log_block(zilog, &blk, &abuf);
258 if (parse_blk_func != NULL)
259 parse_blk_func(zilog, &blk, arg, txg);
264 lrbuf = abuf->b_data;
265 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
266 blk = ztp->zit_next_blk;
268 if (parse_lr_func == NULL) {
269 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
273 for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) {
274 lr_t *lr = (lr_t *)lrp;
275 reclen = lr->lrc_reclen;
276 ASSERT3U(reclen, >=, sizeof (lr_t));
277 parse_lr_func(zilog, lr, arg, txg);
279 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
281 zil_dva_tree_fini(&zilog->zl_dva_tree);
288 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
290 spa_t *spa = zilog->zl_spa;
294 * Claim log block if not already committed and not already claimed.
296 if (bp->blk_birth >= first_txg &&
297 zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) {
298 err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL,
299 ZIO_FLAG_MUSTSUCCEED));
305 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
307 if (lrc->lrc_txtype == TX_WRITE) {
308 lr_write_t *lr = (lr_write_t *)lrc;
309 zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg);
315 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
317 zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx));
321 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
324 * If we previously claimed it, we need to free it.
326 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) {
327 lr_write_t *lr = (lr_write_t *)lrc;
328 blkptr_t *bp = &lr->lr_blkptr;
329 if (bp->blk_birth >= claim_txg &&
330 !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) {
331 (void) arc_free(NULL, zilog->zl_spa,
332 dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT);
338 * Create an on-disk intent log.
341 zil_create(zilog_t *zilog)
343 const zil_header_t *zh = zilog->zl_header;
351 * Wait for any previous destroy to complete.
353 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
355 ASSERT(zh->zh_claim_txg == 0);
356 ASSERT(zh->zh_replay_seq == 0);
361 * If we don't already have an initial log block, allocate one now.
363 if (BP_IS_HOLE(&blk)) {
364 tx = dmu_tx_create(zilog->zl_os);
365 (void) dmu_tx_assign(tx, TXG_WAIT);
366 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
367 txg = dmu_tx_get_txg(tx);
369 error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk,
373 zil_init_log_chain(zilog, &blk);
377 * Allocate a log write buffer (lwb) for the first log block.
380 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
381 lwb->lwb_zilog = zilog;
384 lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk);
385 lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz);
386 lwb->lwb_max_txg = txg;
389 mutex_enter(&zilog->zl_lock);
390 list_insert_tail(&zilog->zl_lwb_list, lwb);
391 mutex_exit(&zilog->zl_lock);
395 * If we just allocated the first log block, commit our transaction
396 * and wait for zil_sync() to stuff the block poiner into zh_log.
397 * (zh is part of the MOS, so we cannot modify it in open context.)
401 txg_wait_synced(zilog->zl_dmu_pool, txg);
404 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
408 * In one tx, free all log blocks and clear the log header.
409 * If keep_first is set, then we're replaying a log with no content.
410 * We want to keep the first block, however, so that the first
411 * synchronous transaction doesn't require a txg_wait_synced()
412 * in zil_create(). We don't need to txg_wait_synced() here either
413 * when keep_first is set, because both zil_create() and zil_destroy()
414 * will wait for any in-progress destroys to complete.
417 zil_destroy(zilog_t *zilog, boolean_t keep_first)
419 const zil_header_t *zh = zilog->zl_header;
425 * Wait for any previous destroy to complete.
427 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
429 if (BP_IS_HOLE(&zh->zh_log))
432 tx = dmu_tx_create(zilog->zl_os);
433 (void) dmu_tx_assign(tx, TXG_WAIT);
434 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
435 txg = dmu_tx_get_txg(tx);
437 mutex_enter(&zilog->zl_lock);
440 * It is possible for the ZIL to get the previously mounted zilog
441 * structure of the same dataset if quickly remounted and the dbuf
442 * eviction has not completed. In this case we can see a non
443 * empty lwb list and keep_first will be set. We fix this by
444 * clearing the keep_first. This will be slower but it's very rare.
446 if (!list_is_empty(&zilog->zl_lwb_list) && keep_first)
447 keep_first = B_FALSE;
449 ASSERT3U(zilog->zl_destroy_txg, <, txg);
450 zilog->zl_destroy_txg = txg;
451 zilog->zl_keep_first = keep_first;
453 if (!list_is_empty(&zilog->zl_lwb_list)) {
454 ASSERT(zh->zh_claim_txg == 0);
456 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
457 list_remove(&zilog->zl_lwb_list, lwb);
458 if (lwb->lwb_buf != NULL)
459 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
460 zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg);
461 kmem_cache_free(zil_lwb_cache, lwb);
465 (void) zil_parse(zilog, zil_free_log_block,
466 zil_free_log_record, tx, zh->zh_claim_txg);
469 mutex_exit(&zilog->zl_lock);
475 * zil_rollback_destroy() is only called by the rollback code.
476 * We already have a syncing tx. Rollback has exclusive access to the
477 * dataset, so we don't have to worry about concurrent zil access.
478 * The actual freeing of any log blocks occurs in zil_sync() later in
479 * this txg syncing phase.
482 zil_rollback_destroy(zilog_t *zilog, dmu_tx_t *tx)
484 const zil_header_t *zh = zilog->zl_header;
487 if (BP_IS_HOLE(&zh->zh_log))
490 txg = dmu_tx_get_txg(tx);
491 ASSERT3U(zilog->zl_destroy_txg, <, txg);
492 zilog->zl_destroy_txg = txg;
493 zilog->zl_keep_first = B_FALSE;
496 * Ensure there's no outstanding ZIL IO. No lwbs or just the
497 * unused one that allocated in advance is ok.
499 ASSERT(zilog->zl_lwb_list.list_head.list_next ==
500 zilog->zl_lwb_list.list_head.list_prev);
501 (void) zil_parse(zilog, zil_free_log_block, zil_free_log_record,
502 tx, zh->zh_claim_txg);
506 * return true if the initial log block is not valid
509 zil_empty(zilog_t *zilog)
511 const zil_header_t *zh = zilog->zl_header;
512 arc_buf_t *abuf = NULL;
514 if (BP_IS_HOLE(&zh->zh_log))
517 if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0)
520 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
525 zil_claim(char *osname, void *txarg)
527 dmu_tx_t *tx = txarg;
528 uint64_t first_txg = dmu_tx_get_txg(tx);
534 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
536 cmn_err(CE_WARN, "can't open objset for %s", osname);
540 zilog = dmu_objset_zil(os);
541 zh = zil_header_in_syncing_context(zilog);
544 * Record here whether the zil has any records to replay.
545 * If the header block pointer is null or the block points
546 * to the stubby then we know there are no valid log records.
547 * We use the header to store this state as the the zilog gets
548 * freed later in dmu_objset_close().
549 * The flags (and the rest of the header fields) are cleared in
550 * zil_sync() as a result of a zil_destroy(), after replaying the log.
552 * Note, the intent log can be empty but still need the
553 * stubby to be claimed.
555 if (!zil_empty(zilog))
556 zh->zh_flags |= ZIL_REPLAY_NEEDED;
559 * Claim all log blocks if we haven't already done so, and remember
560 * the highest claimed sequence number. This ensures that if we can
561 * read only part of the log now (e.g. due to a missing device),
562 * but we can read the entire log later, we will not try to replay
563 * or destroy beyond the last block we successfully claimed.
565 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
566 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
567 zh->zh_claim_txg = first_txg;
568 zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block,
569 zil_claim_log_record, tx, first_txg);
570 dsl_dataset_dirty(dmu_objset_ds(os), tx);
573 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
574 dmu_objset_close(os);
579 * Check the log by walking the log chain.
580 * Checksum errors are ok as they indicate the end of the chain.
581 * Any other error (no device or read failure) returns an error.
585 zil_check_log_chain(char *osname, void *txarg)
596 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
598 cmn_err(CE_WARN, "can't open objset for %s", osname);
602 zilog = dmu_objset_zil(os);
603 zh = zil_header_in_syncing_context(zilog);
605 if (BP_IS_HOLE(&blk)) {
606 dmu_objset_close(os);
607 return (0); /* no chain */
611 error = zil_read_log_block(zilog, &blk, &abuf);
614 lrbuf = abuf->b_data;
615 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1;
616 blk = ztp->zit_next_blk;
617 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
619 dmu_objset_close(os);
621 return (0); /* normal end of chain */
630 zil_clear_log_chain(char *osname, void *txarg)
638 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os);
640 cmn_err(CE_WARN, "can't open objset for %s", osname);
644 zilog = dmu_objset_zil(os);
645 tx = dmu_tx_create(zilog->zl_os);
646 (void) dmu_tx_assign(tx, TXG_WAIT);
647 zh = zil_header_in_syncing_context(zilog);
648 BP_ZERO(&zh->zh_log);
649 dsl_dataset_dirty(dmu_objset_ds(os), tx);
651 dmu_objset_close(os);
656 zil_vdev_compare(const void *x1, const void *x2)
658 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
659 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
670 zil_add_block(zilog_t *zilog, blkptr_t *bp)
672 avl_tree_t *t = &zilog->zl_vdev_tree;
674 zil_vdev_node_t *zv, zvsearch;
675 int ndvas = BP_GET_NDVAS(bp);
678 if (zfs_nocacheflush)
681 ASSERT(zilog->zl_writer);
684 * Even though we're zl_writer, we still need a lock because the
685 * zl_get_data() callbacks may have dmu_sync() done callbacks
686 * that will run concurrently.
688 mutex_enter(&zilog->zl_vdev_lock);
689 for (i = 0; i < ndvas; i++) {
690 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
691 if (avl_find(t, &zvsearch, &where) == NULL) {
692 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
693 zv->zv_vdev = zvsearch.zv_vdev;
694 avl_insert(t, zv, where);
697 mutex_exit(&zilog->zl_vdev_lock);
701 zil_flush_vdevs(zilog_t *zilog)
703 spa_t *spa = zilog->zl_spa;
704 avl_tree_t *t = &zilog->zl_vdev_tree;
709 ASSERT(zilog->zl_writer);
712 * We don't need zl_vdev_lock here because we're the zl_writer,
713 * and all zl_get_data() callbacks are done.
715 if (avl_numnodes(t) == 0)
718 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
720 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
722 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
723 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
726 kmem_free(zv, sizeof (*zv));
730 * Wait for all the flushes to complete. Not all devices actually
731 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
733 (void) zio_wait(zio);
735 spa_config_exit(spa, SCL_STATE, FTAG);
739 * Function called when a log block write completes
742 zil_lwb_write_done(zio_t *zio)
744 lwb_t *lwb = zio->io_private;
745 zilog_t *zilog = lwb->lwb_zilog;
747 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
748 ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG);
749 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
750 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
751 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
752 ASSERT(!BP_IS_GANG(zio->io_bp));
753 ASSERT(!BP_IS_HOLE(zio->io_bp));
754 ASSERT(zio->io_bp->blk_fill == 0);
757 * Now that we've written this log block, we have a stable pointer
758 * to the next block in the chain, so it's OK to let the txg in
759 * which we allocated the next block sync.
761 txg_rele_to_sync(&lwb->lwb_txgh);
763 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
764 mutex_enter(&zilog->zl_lock);
767 zilog->zl_log_error = B_TRUE;
768 mutex_exit(&zilog->zl_lock);
772 * Initialize the io for a log block.
775 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
779 zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET];
782 zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
784 if (zilog->zl_root_zio == NULL) {
785 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
788 if (lwb->lwb_zio == NULL) {
789 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
790 0, &lwb->lwb_blk, lwb->lwb_buf,
791 lwb->lwb_sz, zil_lwb_write_done, lwb,
792 ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb);
797 * Start a log block write and advance to the next log block.
798 * Calls are serialized.
801 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
804 zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1;
805 spa_t *spa = zilog->zl_spa;
806 blkptr_t *bp = &ztp->zit_next_blk;
811 ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb));
814 * Allocate the next block and save its address in this block
815 * before writing it in order to establish the log chain.
816 * Note that if the allocation of nlwb synced before we wrote
817 * the block that points at it (lwb), we'd leak it if we crashed.
818 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done().
820 txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh);
821 txg_rele_to_quiesce(&lwb->lwb_txgh);
824 * Pick a ZIL blocksize. We request a size that is the
825 * maximum of the previous used size, the current used size and
826 * the amount waiting in the queue.
828 zil_blksz = MAX(zilog->zl_prev_used,
829 zilog->zl_cur_used + sizeof (*ztp));
830 zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp));
831 zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t);
832 if (zil_blksz > ZIL_MAX_BLKSZ)
833 zil_blksz = ZIL_MAX_BLKSZ;
836 /* pass the old blkptr in order to spread log blocks across devs */
837 error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg);
839 dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg);
842 * We dirty the dataset to ensure that zil_sync() will
843 * be called to remove this lwb from our zl_lwb_list.
844 * Failing to do so, may leave an lwb with a NULL lwb_buf
845 * hanging around on the zl_lwb_list.
847 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
851 * Since we've just experienced an allocation failure so we
852 * terminate the current lwb and send it on its way.
855 ztp->zit_nused = lwb->lwb_nused;
856 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
857 zio_nowait(lwb->lwb_zio);
860 * By returning NULL the caller will call tx_wait_synced()
865 ASSERT3U(bp->blk_birth, ==, txg);
867 ztp->zit_nused = lwb->lwb_nused;
868 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum;
869 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
870 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
873 * Allocate a new log write buffer (lwb).
875 nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
877 nlwb->lwb_zilog = zilog;
880 nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk);
881 nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz);
882 nlwb->lwb_max_txg = txg;
883 nlwb->lwb_zio = NULL;
886 * Put new lwb at the end of the log chain
888 mutex_enter(&zilog->zl_lock);
889 list_insert_tail(&zilog->zl_lwb_list, nlwb);
890 mutex_exit(&zilog->zl_lock);
892 /* Record the block for later vdev flushing */
893 zil_add_block(zilog, &lwb->lwb_blk);
896 * kick off the write for the old log block
898 dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg);
899 ASSERT(lwb->lwb_zio);
900 zio_nowait(lwb->lwb_zio);
906 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
908 lr_t *lrc = &itx->itx_lr; /* common log record */
909 lr_write_t *lr = (lr_write_t *)lrc;
910 uint64_t txg = lrc->lrc_txg;
911 uint64_t reclen = lrc->lrc_reclen;
916 ASSERT(lwb->lwb_buf != NULL);
918 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
919 dlen = P2ROUNDUP_TYPED(
920 lr->lr_length, sizeof (uint64_t), uint64_t);
924 zilog->zl_cur_used += (reclen + dlen);
926 zil_lwb_write_init(zilog, lwb);
929 * If this record won't fit in the current log block, start a new one.
931 if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
932 lwb = zil_lwb_write_start(zilog, lwb);
935 zil_lwb_write_init(zilog, lwb);
936 ASSERT(lwb->lwb_nused == 0);
937 if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) {
938 txg_wait_synced(zilog->zl_dmu_pool, txg);
944 * Update the lrc_seq, to be log record sequence number. See zil.h
945 * Then copy the record to the log buffer.
947 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
948 bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen);
951 * If it's a write, fetch the data or get its blkptr as appropriate.
953 if (lrc->lrc_txtype == TX_WRITE) {
954 if (txg > spa_freeze_txg(zilog->zl_spa))
955 txg_wait_synced(zilog->zl_dmu_pool, txg);
956 if (itx->itx_wr_state != WR_COPIED) {
960 /* alignment is guaranteed */
961 lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused);
963 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
964 dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen;
965 lr->lr_common.lrc_reclen += dlen;
967 ASSERT(itx->itx_wr_state == WR_INDIRECT);
970 error = zilog->zl_get_data(
971 itx->itx_private, lr, dbuf, lwb->lwb_zio);
973 ASSERT(error == ENOENT || error == EEXIST ||
980 lwb->lwb_nused += reclen + dlen;
981 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
982 ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb));
983 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
989 zil_itx_create(uint64_t txtype, size_t lrsize)
993 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
995 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
996 itx->itx_lr.lrc_txtype = txtype;
997 itx->itx_lr.lrc_reclen = lrsize;
998 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
999 itx->itx_lr.lrc_seq = 0; /* defensive */
1005 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1009 ASSERT(itx->itx_lr.lrc_seq == 0);
1011 mutex_enter(&zilog->zl_lock);
1012 list_insert_tail(&zilog->zl_itx_list, itx);
1013 zilog->zl_itx_list_sz += itx->itx_sod;
1014 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1015 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
1016 mutex_exit(&zilog->zl_lock);
1022 * Free up all in-memory intent log transactions that have now been synced.
1025 zil_itx_clean(zilog_t *zilog)
1027 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
1028 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
1032 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1034 mutex_enter(&zilog->zl_lock);
1035 /* wait for a log writer to finish walking list */
1036 while (zilog->zl_writer) {
1037 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1041 * Move the sync'd log transactions to a separate list so we can call
1042 * kmem_free without holding the zl_lock.
1044 * There is no need to set zl_writer as we don't drop zl_lock here
1046 while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
1047 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
1048 list_remove(&zilog->zl_itx_list, itx);
1049 zilog->zl_itx_list_sz -= itx->itx_sod;
1050 list_insert_tail(&clean_list, itx);
1052 cv_broadcast(&zilog->zl_cv_writer);
1053 mutex_exit(&zilog->zl_lock);
1055 /* destroy sync'd log transactions */
1056 while ((itx = list_head(&clean_list)) != NULL) {
1057 list_remove(&clean_list, itx);
1058 kmem_free(itx, offsetof(itx_t, itx_lr)
1059 + itx->itx_lr.lrc_reclen);
1061 list_destroy(&clean_list);
1065 * If there are any in-memory intent log transactions which have now been
1066 * synced then start up a taskq to free them.
1069 zil_clean(zilog_t *zilog)
1073 mutex_enter(&zilog->zl_lock);
1074 itx = list_head(&zilog->zl_itx_list);
1075 if ((itx != NULL) &&
1076 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
1077 (void) taskq_dispatch(zilog->zl_clean_taskq,
1078 (task_func_t *)zil_itx_clean, zilog, TQ_SLEEP);
1080 mutex_exit(&zilog->zl_lock);
1084 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
1087 uint64_t commit_seq = 0;
1088 itx_t *itx, *itx_next = (itx_t *)-1;
1092 zilog->zl_writer = B_TRUE;
1093 ASSERT(zilog->zl_root_zio == NULL);
1094 spa = zilog->zl_spa;
1096 if (zilog->zl_suspend) {
1099 lwb = list_tail(&zilog->zl_lwb_list);
1102 * Return if there's nothing to flush before we
1103 * dirty the fs by calling zil_create()
1105 if (list_is_empty(&zilog->zl_itx_list)) {
1106 zilog->zl_writer = B_FALSE;
1109 mutex_exit(&zilog->zl_lock);
1111 mutex_enter(&zilog->zl_lock);
1112 lwb = list_tail(&zilog->zl_lwb_list);
1116 /* Loop through in-memory log transactions filling log blocks. */
1117 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1120 * Find the next itx to push:
1121 * Push all transactions related to specified foid and all
1122 * other transactions except TX_WRITE, TX_TRUNCATE,
1123 * TX_SETATTR and TX_ACL for all other files.
1125 if (itx_next != (itx_t *)-1)
1128 itx = list_head(&zilog->zl_itx_list);
1129 for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) {
1130 if (foid == 0) /* push all foids? */
1132 if (itx->itx_sync) /* push all O_[D]SYNC */
1134 switch (itx->itx_lr.lrc_txtype) {
1139 /* lr_foid is same offset for these records */
1140 if (((lr_write_t *)&itx->itx_lr)->lr_foid
1142 continue; /* skip this record */
1150 if ((itx->itx_lr.lrc_seq > seq) &&
1151 ((lwb == NULL) || (lwb->lwb_nused == 0) ||
1152 (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) {
1157 * Save the next pointer. Even though we soon drop
1158 * zl_lock all threads that may change the list
1159 * (another writer or zil_itx_clean) can't do so until
1160 * they have zl_writer.
1162 itx_next = list_next(&zilog->zl_itx_list, itx);
1163 list_remove(&zilog->zl_itx_list, itx);
1164 zilog->zl_itx_list_sz -= itx->itx_sod;
1165 mutex_exit(&zilog->zl_lock);
1166 txg = itx->itx_lr.lrc_txg;
1169 if (txg > spa_last_synced_txg(spa) ||
1170 txg > spa_freeze_txg(spa))
1171 lwb = zil_lwb_commit(zilog, itx, lwb);
1172 kmem_free(itx, offsetof(itx_t, itx_lr)
1173 + itx->itx_lr.lrc_reclen);
1174 mutex_enter(&zilog->zl_lock);
1176 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1177 /* determine commit sequence number */
1178 itx = list_head(&zilog->zl_itx_list);
1180 commit_seq = itx->itx_lr.lrc_seq;
1182 commit_seq = zilog->zl_itx_seq;
1183 mutex_exit(&zilog->zl_lock);
1185 /* write the last block out */
1186 if (lwb != NULL && lwb->lwb_zio != NULL)
1187 lwb = zil_lwb_write_start(zilog, lwb);
1189 zilog->zl_prev_used = zilog->zl_cur_used;
1190 zilog->zl_cur_used = 0;
1193 * Wait if necessary for the log blocks to be on stable storage.
1195 if (zilog->zl_root_zio) {
1196 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1197 (void) zio_wait(zilog->zl_root_zio);
1198 zilog->zl_root_zio = NULL;
1199 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1200 zil_flush_vdevs(zilog);
1203 if (zilog->zl_log_error || lwb == NULL) {
1204 zilog->zl_log_error = 0;
1205 txg_wait_synced(zilog->zl_dmu_pool, 0);
1208 mutex_enter(&zilog->zl_lock);
1209 zilog->zl_writer = B_FALSE;
1211 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1212 zilog->zl_commit_seq = commit_seq;
1216 * Push zfs transactions to stable storage up to the supplied sequence number.
1217 * If foid is 0 push out all transactions, otherwise push only those
1218 * for that file or might have been used to create that file.
1221 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1223 if (zilog == NULL || seq == 0)
1226 mutex_enter(&zilog->zl_lock);
1228 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */
1230 while (zilog->zl_writer) {
1231 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1232 if (seq < zilog->zl_commit_seq) {
1233 mutex_exit(&zilog->zl_lock);
1237 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1238 /* wake up others waiting on the commit */
1239 cv_broadcast(&zilog->zl_cv_writer);
1240 mutex_exit(&zilog->zl_lock);
1244 * Called in syncing context to free committed log blocks and update log header.
1247 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1249 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1250 uint64_t txg = dmu_tx_get_txg(tx);
1251 spa_t *spa = zilog->zl_spa;
1254 mutex_enter(&zilog->zl_lock);
1256 ASSERT(zilog->zl_stop_sync == 0);
1258 zh->zh_replay_seq = zilog->zl_replay_seq[txg & TXG_MASK];
1260 if (zilog->zl_destroy_txg == txg) {
1261 blkptr_t blk = zh->zh_log;
1263 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1264 ASSERT(spa_sync_pass(spa) == 1);
1266 bzero(zh, sizeof (zil_header_t));
1267 bzero(zilog->zl_replay_seq, sizeof (zilog->zl_replay_seq));
1269 if (zilog->zl_keep_first) {
1271 * If this block was part of log chain that couldn't
1272 * be claimed because a device was missing during
1273 * zil_claim(), but that device later returns,
1274 * then this block could erroneously appear valid.
1275 * To guard against this, assign a new GUID to the new
1276 * log chain so it doesn't matter what blk points to.
1278 zil_init_log_chain(zilog, &blk);
1284 lwb = list_head(&zilog->zl_lwb_list);
1286 mutex_exit(&zilog->zl_lock);
1289 zh->zh_log = lwb->lwb_blk;
1290 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1292 list_remove(&zilog->zl_lwb_list, lwb);
1293 zio_free_blk(spa, &lwb->lwb_blk, txg);
1294 kmem_cache_free(zil_lwb_cache, lwb);
1297 * If we don't have anything left in the lwb list then
1298 * we've had an allocation failure and we need to zero
1299 * out the zil_header blkptr so that we don't end
1300 * up freeing the same block twice.
1302 if (list_head(&zilog->zl_lwb_list) == NULL)
1303 BP_ZERO(&zh->zh_log);
1305 mutex_exit(&zilog->zl_lock);
1311 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1312 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1318 kmem_cache_destroy(zil_lwb_cache);
1322 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1326 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1328 zilog->zl_header = zh_phys;
1330 zilog->zl_spa = dmu_objset_spa(os);
1331 zilog->zl_dmu_pool = dmu_objset_pool(os);
1332 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1334 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1336 list_create(&zilog->zl_itx_list, sizeof (itx_t),
1337 offsetof(itx_t, itx_node));
1339 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1340 offsetof(lwb_t, lwb_node));
1342 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1344 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1345 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1347 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1348 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1354 zil_free(zilog_t *zilog)
1358 zilog->zl_stop_sync = 1;
1360 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1361 list_remove(&zilog->zl_lwb_list, lwb);
1362 if (lwb->lwb_buf != NULL)
1363 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1364 kmem_cache_free(zil_lwb_cache, lwb);
1366 list_destroy(&zilog->zl_lwb_list);
1368 avl_destroy(&zilog->zl_vdev_tree);
1369 mutex_destroy(&zilog->zl_vdev_lock);
1371 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1372 list_destroy(&zilog->zl_itx_list);
1373 mutex_destroy(&zilog->zl_lock);
1375 cv_destroy(&zilog->zl_cv_writer);
1376 cv_destroy(&zilog->zl_cv_suspend);
1378 kmem_free(zilog, sizeof (zilog_t));
1382 * Open an intent log.
1385 zil_open(objset_t *os, zil_get_data_t *get_data)
1387 zilog_t *zilog = dmu_objset_zil(os);
1389 zilog->zl_get_data = get_data;
1390 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1391 2, 2, TASKQ_PREPOPULATE);
1397 * Close an intent log.
1400 zil_close(zilog_t *zilog)
1403 * If the log isn't already committed, mark the objset dirty
1404 * (so zil_sync() will be called) and wait for that txg to sync.
1406 if (!zil_is_committed(zilog)) {
1408 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1409 (void) dmu_tx_assign(tx, TXG_WAIT);
1410 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1411 txg = dmu_tx_get_txg(tx);
1413 txg_wait_synced(zilog->zl_dmu_pool, txg);
1416 taskq_destroy(zilog->zl_clean_taskq);
1417 zilog->zl_clean_taskq = NULL;
1418 zilog->zl_get_data = NULL;
1420 zil_itx_clean(zilog);
1421 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1425 * Suspend an intent log. While in suspended mode, we still honor
1426 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1427 * We suspend the log briefly when taking a snapshot so that the snapshot
1428 * contains all the data it's supposed to, and has an empty intent log.
1431 zil_suspend(zilog_t *zilog)
1433 const zil_header_t *zh = zilog->zl_header;
1435 mutex_enter(&zilog->zl_lock);
1436 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1437 mutex_exit(&zilog->zl_lock);
1440 if (zilog->zl_suspend++ != 0) {
1442 * Someone else already began a suspend.
1443 * Just wait for them to finish.
1445 while (zilog->zl_suspending)
1446 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1447 mutex_exit(&zilog->zl_lock);
1450 zilog->zl_suspending = B_TRUE;
1451 mutex_exit(&zilog->zl_lock);
1453 zil_commit(zilog, UINT64_MAX, 0);
1456 * Wait for any in-flight log writes to complete.
1458 mutex_enter(&zilog->zl_lock);
1459 while (zilog->zl_writer)
1460 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1461 mutex_exit(&zilog->zl_lock);
1463 zil_destroy(zilog, B_FALSE);
1465 mutex_enter(&zilog->zl_lock);
1466 zilog->zl_suspending = B_FALSE;
1467 cv_broadcast(&zilog->zl_cv_suspend);
1468 mutex_exit(&zilog->zl_lock);
1474 zil_resume(zilog_t *zilog)
1476 mutex_enter(&zilog->zl_lock);
1477 ASSERT(zilog->zl_suspend != 0);
1478 zilog->zl_suspend--;
1479 mutex_exit(&zilog->zl_lock);
1482 typedef struct zil_replay_arg {
1484 zil_replay_func_t **zr_replay;
1485 zil_replay_cleaner_t *zr_replay_cleaner;
1488 boolean_t zr_byteswap;
1493 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1495 zil_replay_arg_t *zr = zra;
1496 const zil_header_t *zh = zilog->zl_header;
1497 uint64_t reclen = lr->lrc_reclen;
1498 uint64_t txtype = lr->lrc_txtype;
1500 int pass, error, sunk;
1502 if (zilog->zl_stop_replay)
1505 if (lr->lrc_txg < claim_txg) /* already committed */
1508 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1511 /* Strip case-insensitive bit, still present in log record */
1515 * Make a copy of the data so we can revise and extend it.
1517 bcopy(lr, zr->zr_lrbuf, reclen);
1520 * The log block containing this lr may have been byteswapped
1521 * so that we can easily examine common fields like lrc_txtype.
1522 * However, the log is a mix of different data types, and only the
1523 * replay vectors know how to byteswap their records. Therefore, if
1524 * the lr was byteswapped, undo it before invoking the replay vector.
1526 if (zr->zr_byteswap)
1527 byteswap_uint64_array(zr->zr_lrbuf, reclen);
1530 * If this is a TX_WRITE with a blkptr, suck in the data.
1532 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1533 lr_write_t *lrw = (lr_write_t *)lr;
1534 blkptr_t *wbp = &lrw->lr_blkptr;
1535 uint64_t wlen = lrw->lr_length;
1536 char *wbuf = zr->zr_lrbuf + reclen;
1538 if (BP_IS_HOLE(wbp)) { /* compressed to a hole */
1542 * A subsequent write may have overwritten this block,
1543 * in which case wbp may have been been freed and
1544 * reallocated, and our read of wbp may fail with a
1545 * checksum error. We can safely ignore this because
1546 * the later write will provide the correct data.
1550 zb.zb_objset = dmu_objset_id(zilog->zl_os);
1551 zb.zb_object = lrw->lr_foid;
1553 zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp);
1555 (void) zio_wait(zio_read(NULL, zilog->zl_spa,
1556 wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL,
1557 ZIO_PRIORITY_SYNC_READ,
1558 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb));
1559 (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen);
1564 * We must now do two things atomically: replay this log record,
1565 * and update the log header to reflect the fact that we did so.
1566 * We use the DMU's ability to assign into a specific txg to do this.
1568 for (pass = 1, sunk = B_FALSE; /* CONSTANTCONDITION */; pass++) {
1569 uint64_t replay_txg;
1570 dmu_tx_t *replay_tx;
1572 replay_tx = dmu_tx_create(zr->zr_os);
1573 error = dmu_tx_assign(replay_tx, TXG_WAIT);
1575 dmu_tx_abort(replay_tx);
1579 replay_txg = dmu_tx_get_txg(replay_tx);
1581 if (txtype == 0 || txtype >= TX_MAX_TYPE) {
1585 * On the first pass, arrange for the replay vector
1586 * to fail its dmu_tx_assign(). That's the only way
1587 * to ensure that those code paths remain well tested.
1589 * Only byteswap (if needed) on the 1st pass.
1591 *zr->zr_txgp = replay_txg - (pass == 1);
1592 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
1593 zr->zr_byteswap && pass == 1);
1594 *zr->zr_txgp = TXG_NOWAIT;
1598 dsl_dataset_dirty(dmu_objset_ds(zr->zr_os), replay_tx);
1599 zilog->zl_replay_seq[replay_txg & TXG_MASK] =
1603 dmu_tx_commit(replay_tx);
1609 * The DMU's dnode layer doesn't see removes until the txg
1610 * commits, so a subsequent claim can spuriously fail with
1611 * EEXIST. So if we receive any error other than ERESTART
1612 * we try syncing out any removes then retrying the
1615 if (error != ERESTART && !sunk) {
1616 if (zr->zr_replay_cleaner)
1617 zr->zr_replay_cleaner(zr->zr_arg);
1618 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1620 continue; /* retry */
1623 if (error != ERESTART)
1627 txg_wait_open(spa_get_dsl(zilog->zl_spa),
1630 dprintf("pass %d, retrying\n", pass);
1633 ASSERT(error && error != ERESTART);
1634 name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
1635 dmu_objset_name(zr->zr_os, name);
1636 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1637 "dataset %s, seq 0x%llx, txtype %llu %s\n",
1638 error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype,
1639 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1640 zilog->zl_stop_replay = 1;
1641 kmem_free(name, MAXNAMELEN);
1646 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1648 zilog->zl_replay_blks++;
1652 * If this dataset has a non-empty intent log, replay it and destroy it.
1655 zil_replay(objset_t *os, void *arg, uint64_t *txgp,
1656 zil_replay_func_t *replay_func[TX_MAX_TYPE],
1657 zil_replay_cleaner_t *replay_cleaner)
1659 zilog_t *zilog = dmu_objset_zil(os);
1660 const zil_header_t *zh = zilog->zl_header;
1661 zil_replay_arg_t zr;
1663 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1664 zil_destroy(zilog, B_TRUE);
1667 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
1670 zr.zr_replay = replay_func;
1671 zr.zr_replay_cleaner = replay_cleaner;
1674 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1675 zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1678 * Wait for in-progress removes to sync before starting replay.
1680 txg_wait_synced(zilog->zl_dmu_pool, 0);
1682 zilog->zl_stop_replay = 0;
1683 zilog->zl_replay_time = LBOLT;
1684 ASSERT(zilog->zl_replay_blks == 0);
1685 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1687 kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE);
1689 zil_destroy(zilog, B_FALSE);
1690 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1691 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
1695 * Report whether all transactions are committed
1698 zil_is_committed(zilog_t *zilog)
1703 mutex_enter(&zilog->zl_lock);
1704 while (zilog->zl_writer)
1705 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1707 /* recent unpushed intent log transactions? */
1708 if (!list_is_empty(&zilog->zl_itx_list)) {
1713 /* intent log never used? */
1714 lwb = list_head(&zilog->zl_lwb_list);
1721 * more than 1 log buffer means zil_sync() hasn't yet freed
1722 * entries after a txg has committed
1724 if (list_next(&zilog->zl_lwb_list, lwb)) {
1729 ASSERT(zil_empty(zilog));
1732 cv_broadcast(&zilog->zl_cv_writer);
1733 mutex_exit(&zilog->zl_lock);