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.
25 /* Portions Copyright 2010 Robert Milkowski */
27 #include <sys/zfs_context.h>
33 #include <sys/resource.h>
35 #include <sys/zil_impl.h>
36 #include <sys/dsl_dataset.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dsl_pool.h>
42 * The zfs intent log (ZIL) saves transaction records of system calls
43 * that change the file system in memory with enough information
44 * to be able to replay them. These are stored in memory until
45 * either the DMU transaction group (txg) commits them to the stable pool
46 * and they can be discarded, or they are flushed to the stable log
47 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
48 * requirement. In the event of a panic or power fail then those log
49 * records (transactions) are replayed.
51 * There is one ZIL per file system. Its on-disk (pool) format consists
58 * A log record holds a system call transaction. Log blocks can
59 * hold many log records and the blocks are chained together.
60 * Each ZIL block contains a block pointer (blkptr_t) to the next
61 * ZIL block in the chain. The ZIL header points to the first
62 * block in the chain. Note there is not a fixed place in the pool
63 * to hold blocks. They are dynamically allocated and freed as
64 * needed from the blocks available. Figure X shows the ZIL structure:
68 * This global ZIL switch affects all pools
70 int zil_replay_disable = 0; /* disable intent logging replay */
73 * Tunable parameter for debugging or performance analysis. Setting
74 * zfs_nocacheflush will cause corruption on power loss if a volatile
75 * out-of-order write cache is enabled.
77 boolean_t zfs_nocacheflush = B_FALSE;
79 static kmem_cache_t *zil_lwb_cache;
81 static boolean_t zil_empty(zilog_t *zilog);
83 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
84 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
88 zil_bp_compare(const void *x1, const void *x2)
90 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
91 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
93 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
95 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
98 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
100 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
107 zil_bp_tree_init(zilog_t *zilog)
109 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
110 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
114 zil_bp_tree_fini(zilog_t *zilog)
116 avl_tree_t *t = &zilog->zl_bp_tree;
120 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
121 kmem_free(zn, sizeof (zil_bp_node_t));
127 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
129 avl_tree_t *t = &zilog->zl_bp_tree;
130 const dva_t *dva = BP_IDENTITY(bp);
134 if (avl_find(t, dva, &where) != NULL)
137 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
139 avl_insert(t, zn, where);
144 static zil_header_t *
145 zil_header_in_syncing_context(zilog_t *zilog)
147 return ((zil_header_t *)zilog->zl_header);
151 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
153 zio_cksum_t *zc = &bp->blk_cksum;
155 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
156 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
157 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
158 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
162 * Read a log block and make sure it's valid.
165 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
168 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
169 uint32_t aflags = ARC_WAIT;
170 arc_buf_t *abuf = NULL;
174 if (zilog->zl_header->zh_claim_txg == 0)
175 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
177 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
178 zio_flags |= ZIO_FLAG_SPECULATIVE;
180 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
181 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
183 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
184 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
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 (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
200 zil_chain_t *zilc = abuf->b_data;
201 char *lr = (char *)(zilc + 1);
202 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
204 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
205 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
209 *end = (char *)dst + len;
210 *nbp = zilc->zc_next_blk;
213 char *lr = abuf->b_data;
214 uint64_t size = BP_GET_LSIZE(bp);
215 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
217 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
218 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
219 (zilc->zc_nused > (size - sizeof (*zilc)))) {
222 bcopy(lr, dst, zilc->zc_nused);
223 *end = (char *)dst + zilc->zc_nused;
224 *nbp = zilc->zc_next_blk;
228 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
235 * Read a TX_WRITE log data block.
238 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
240 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
241 const blkptr_t *bp = &lr->lr_blkptr;
242 uint32_t aflags = ARC_WAIT;
243 arc_buf_t *abuf = NULL;
247 if (BP_IS_HOLE(bp)) {
249 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
253 if (zilog->zl_header->zh_claim_txg == 0)
254 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
256 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
257 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
259 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
260 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
264 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
265 (void) arc_buf_remove_ref(abuf, &abuf);
272 * Parse the intent log, and call parse_func for each valid record within.
275 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
276 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
278 const zil_header_t *zh = zilog->zl_header;
279 boolean_t claimed = !!zh->zh_claim_txg;
280 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
281 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
282 uint64_t max_blk_seq = 0;
283 uint64_t max_lr_seq = 0;
284 uint64_t blk_count = 0;
285 uint64_t lr_count = 0;
286 blkptr_t blk, next_blk;
291 * Old logs didn't record the maximum zh_claim_lr_seq.
293 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
294 claim_lr_seq = UINT64_MAX;
297 * Starting at the block pointed to by zh_log we read the log chain.
298 * For each block in the chain we strongly check that block to
299 * ensure its validity. We stop when an invalid block is found.
300 * For each block pointer in the chain we call parse_blk_func().
301 * For each record in each valid block we call parse_lr_func().
302 * If the log has been claimed, stop if we encounter a sequence
303 * number greater than the highest claimed sequence number.
305 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
306 zil_bp_tree_init(zilog);
308 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
309 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
313 if (blk_seq > claim_blk_seq)
315 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
317 ASSERT3U(max_blk_seq, <, blk_seq);
318 max_blk_seq = blk_seq;
321 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
324 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
328 for (lrp = lrbuf; lrp < end; lrp += reclen) {
329 lr_t *lr = (lr_t *)lrp;
330 reclen = lr->lrc_reclen;
331 ASSERT3U(reclen, >=, sizeof (lr_t));
332 if (lr->lrc_seq > claim_lr_seq)
334 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
336 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
337 max_lr_seq = lr->lrc_seq;
342 zilog->zl_parse_error = error;
343 zilog->zl_parse_blk_seq = max_blk_seq;
344 zilog->zl_parse_lr_seq = max_lr_seq;
345 zilog->zl_parse_blk_count = blk_count;
346 zilog->zl_parse_lr_count = lr_count;
348 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
349 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
351 zil_bp_tree_fini(zilog);
352 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
358 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
361 * Claim log block if not already committed and not already claimed.
362 * If tx == NULL, just verify that the block is claimable.
364 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
367 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
368 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
369 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
373 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
375 lr_write_t *lr = (lr_write_t *)lrc;
378 if (lrc->lrc_txtype != TX_WRITE)
382 * If the block is not readable, don't claim it. This can happen
383 * in normal operation when a log block is written to disk before
384 * some of the dmu_sync() blocks it points to. In this case, the
385 * transaction cannot have been committed to anyone (we would have
386 * waited for all writes to be stable first), so it is semantically
387 * correct to declare this the end of the log.
389 if (lr->lr_blkptr.blk_birth >= first_txg &&
390 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
392 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
397 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
399 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
405 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
407 lr_write_t *lr = (lr_write_t *)lrc;
408 blkptr_t *bp = &lr->lr_blkptr;
411 * If we previously claimed it, we need to free it.
413 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
414 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
415 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
421 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
425 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
426 lwb->lwb_zilog = zilog;
428 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
429 lwb->lwb_max_txg = txg;
432 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
433 lwb->lwb_nused = sizeof (zil_chain_t);
434 lwb->lwb_sz = BP_GET_LSIZE(bp);
437 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
440 mutex_enter(&zilog->zl_lock);
441 list_insert_tail(&zilog->zl_lwb_list, lwb);
442 mutex_exit(&zilog->zl_lock);
448 * Create an on-disk intent log.
451 zil_create(zilog_t *zilog)
453 const zil_header_t *zh = zilog->zl_header;
461 * Wait for any previous destroy to complete.
463 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
465 ASSERT(zh->zh_claim_txg == 0);
466 ASSERT(zh->zh_replay_seq == 0);
471 * Allocate an initial log block if:
472 * - there isn't one already
473 * - the existing block is the wrong endianess
475 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
476 tx = dmu_tx_create(zilog->zl_os);
477 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
478 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
479 txg = dmu_tx_get_txg(tx);
481 if (!BP_IS_HOLE(&blk)) {
482 zio_free_zil(zilog->zl_spa, txg, &blk);
486 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
487 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
490 zil_init_log_chain(zilog, &blk);
494 * Allocate a log write buffer (lwb) for the first log block.
497 lwb = zil_alloc_lwb(zilog, &blk, txg);
500 * If we just allocated the first log block, commit our transaction
501 * and wait for zil_sync() to stuff the block poiner into zh_log.
502 * (zh is part of the MOS, so we cannot modify it in open context.)
506 txg_wait_synced(zilog->zl_dmu_pool, txg);
509 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
515 * In one tx, free all log blocks and clear the log header.
516 * If keep_first is set, then we're replaying a log with no content.
517 * We want to keep the first block, however, so that the first
518 * synchronous transaction doesn't require a txg_wait_synced()
519 * in zil_create(). We don't need to txg_wait_synced() here either
520 * when keep_first is set, because both zil_create() and zil_destroy()
521 * will wait for any in-progress destroys to complete.
524 zil_destroy(zilog_t *zilog, boolean_t keep_first)
526 const zil_header_t *zh = zilog->zl_header;
532 * Wait for any previous destroy to complete.
534 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
536 zilog->zl_old_header = *zh; /* debugging aid */
538 if (BP_IS_HOLE(&zh->zh_log))
541 tx = dmu_tx_create(zilog->zl_os);
542 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
543 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
544 txg = dmu_tx_get_txg(tx);
546 mutex_enter(&zilog->zl_lock);
548 ASSERT3U(zilog->zl_destroy_txg, <, txg);
549 zilog->zl_destroy_txg = txg;
550 zilog->zl_keep_first = keep_first;
552 if (!list_is_empty(&zilog->zl_lwb_list)) {
553 ASSERT(zh->zh_claim_txg == 0);
555 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
556 list_remove(&zilog->zl_lwb_list, lwb);
557 if (lwb->lwb_buf != NULL)
558 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
559 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
560 kmem_cache_free(zil_lwb_cache, lwb);
562 } else if (!keep_first) {
563 (void) zil_parse(zilog, zil_free_log_block,
564 zil_free_log_record, tx, zh->zh_claim_txg);
566 mutex_exit(&zilog->zl_lock);
572 zil_claim(const char *osname, void *txarg)
574 dmu_tx_t *tx = txarg;
575 uint64_t first_txg = dmu_tx_get_txg(tx);
581 error = dmu_objset_hold(osname, FTAG, &os);
583 cmn_err(CE_WARN, "can't open objset for %s", osname);
587 zilog = dmu_objset_zil(os);
588 zh = zil_header_in_syncing_context(zilog);
590 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
591 if (!BP_IS_HOLE(&zh->zh_log))
592 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
593 BP_ZERO(&zh->zh_log);
594 dsl_dataset_dirty(dmu_objset_ds(os), tx);
595 dmu_objset_rele(os, FTAG);
600 * Claim all log blocks if we haven't already done so, and remember
601 * the highest claimed sequence number. This ensures that if we can
602 * read only part of the log now (e.g. due to a missing device),
603 * but we can read the entire log later, we will not try to replay
604 * or destroy beyond the last block we successfully claimed.
606 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
607 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
608 (void) zil_parse(zilog, zil_claim_log_block,
609 zil_claim_log_record, tx, first_txg);
610 zh->zh_claim_txg = first_txg;
611 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
612 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
613 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
614 zh->zh_flags |= ZIL_REPLAY_NEEDED;
615 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
616 dsl_dataset_dirty(dmu_objset_ds(os), tx);
619 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
620 dmu_objset_rele(os, FTAG);
625 * Check the log by walking the log chain.
626 * Checksum errors are ok as they indicate the end of the chain.
627 * Any other error (no device or read failure) returns an error.
630 zil_check_log_chain(const char *osname, void *tx)
638 error = dmu_objset_hold(osname, FTAG, &os);
640 cmn_err(CE_WARN, "can't open objset for %s", osname);
644 zilog = dmu_objset_zil(os);
647 * Because tx == NULL, zil_claim_log_block() will not actually claim
648 * any blocks, but just determine whether it is possible to do so.
649 * In addition to checking the log chain, zil_claim_log_block()
650 * will invoke zio_claim() with a done func of spa_claim_notify(),
651 * which will update spa_max_claim_txg. See spa_load() for details.
653 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
654 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
656 dmu_objset_rele(os, FTAG);
658 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
662 zil_vdev_compare(const void *x1, const void *x2)
664 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
665 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
676 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
678 avl_tree_t *t = &zilog->zl_vdev_tree;
680 zil_vdev_node_t *zv, zvsearch;
681 int ndvas = BP_GET_NDVAS(bp);
684 if (zfs_nocacheflush)
687 ASSERT(zilog->zl_writer);
690 * Even though we're zl_writer, we still need a lock because the
691 * zl_get_data() callbacks may have dmu_sync() done callbacks
692 * that will run concurrently.
694 mutex_enter(&zilog->zl_vdev_lock);
695 for (i = 0; i < ndvas; i++) {
696 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
697 if (avl_find(t, &zvsearch, &where) == NULL) {
698 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
699 zv->zv_vdev = zvsearch.zv_vdev;
700 avl_insert(t, zv, where);
703 mutex_exit(&zilog->zl_vdev_lock);
707 zil_flush_vdevs(zilog_t *zilog)
709 spa_t *spa = zilog->zl_spa;
710 avl_tree_t *t = &zilog->zl_vdev_tree;
715 ASSERT(zilog->zl_writer);
718 * We don't need zl_vdev_lock here because we're the zl_writer,
719 * and all zl_get_data() callbacks are done.
721 if (avl_numnodes(t) == 0)
724 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
726 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
728 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
729 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
732 kmem_free(zv, sizeof (*zv));
736 * Wait for all the flushes to complete. Not all devices actually
737 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
739 (void) zio_wait(zio);
741 spa_config_exit(spa, SCL_STATE, FTAG);
745 * Function called when a log block write completes
748 zil_lwb_write_done(zio_t *zio)
750 lwb_t *lwb = zio->io_private;
751 zilog_t *zilog = lwb->lwb_zilog;
752 dmu_tx_t *tx = lwb->lwb_tx;
754 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
755 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
756 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
757 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
758 ASSERT(!BP_IS_GANG(zio->io_bp));
759 ASSERT(!BP_IS_HOLE(zio->io_bp));
760 ASSERT(zio->io_bp->blk_fill == 0);
763 * Ensure the lwb buffer pointer is cleared before releasing
764 * the txg. If we have had an allocation failure and
765 * the txg is waiting to sync then we want want zil_sync()
766 * to remove the lwb so that it's not picked up as the next new
767 * one in zil_commit_writer(). zil_sync() will only remove
768 * the lwb if lwb_buf is null.
770 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
771 mutex_enter(&zilog->zl_lock);
774 mutex_exit(&zilog->zl_lock);
777 * Now that we've written this log block, we have a stable pointer
778 * to the next block in the chain, so it's OK to let the txg in
779 * which we allocated the next block sync.
785 * Initialize the io for a log block.
788 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
792 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
793 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
794 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
796 if (zilog->zl_root_zio == NULL) {
797 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
800 if (lwb->lwb_zio == NULL) {
801 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
802 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
803 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
804 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
809 * Define a limited set of intent log block sizes.
810 * These must be a multiple of 4KB. Note only the amount used (again
811 * aligned to 4KB) actually gets written. However, we can't always just
812 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
814 uint64_t zil_block_buckets[] = {
815 4096, /* non TX_WRITE */
816 8192+4096, /* data base */
817 32*1024 + 4096, /* NFS writes */
822 * Use the slog as long as the logbias is 'latency' and the current commit size
823 * is less than the limit or the total list size is less than 2X the limit.
824 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
826 uint64_t zil_slog_limit = 1024 * 1024;
827 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
828 (((zilog)->zl_cur_used < zil_slog_limit) || \
829 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
832 * Start a log block write and advance to the next log block.
833 * Calls are serialized.
836 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
840 spa_t *spa = zilog->zl_spa;
844 uint64_t zil_blksz, wsz;
847 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
848 zilc = (zil_chain_t *)lwb->lwb_buf;
849 bp = &zilc->zc_next_blk;
851 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
852 bp = &zilc->zc_next_blk;
855 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
858 * Allocate the next block and save its address in this block
859 * before writing it in order to establish the log chain.
860 * Note that if the allocation of nlwb synced before we wrote
861 * the block that points at it (lwb), we'd leak it if we crashed.
862 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
863 * We dirty the dataset to ensure that zil_sync() will be called
864 * to clean up in the event of allocation failure or I/O failure.
866 tx = dmu_tx_create(zilog->zl_os);
867 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
868 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
869 txg = dmu_tx_get_txg(tx);
874 * Log blocks are pre-allocated. Here we select the size of the next
875 * block, based on size used in the last block.
876 * - first find the smallest bucket that will fit the block from a
877 * limited set of block sizes. This is because it's faster to write
878 * blocks allocated from the same metaslab as they are adjacent or
880 * - next find the maximum from the new suggested size and an array of
881 * previous sizes. This lessens a picket fence effect of wrongly
882 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
885 * Note we only write what is used, but we can't just allocate
886 * the maximum block size because we can exhaust the available
889 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
890 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
892 zil_blksz = zil_block_buckets[i];
893 if (zil_blksz == UINT64_MAX)
894 zil_blksz = SPA_MAXBLOCKSIZE;
895 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
896 for (i = 0; i < ZIL_PREV_BLKS; i++)
897 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
898 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
901 /* pass the old blkptr in order to spread log blocks across devs */
902 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
905 ASSERT3U(bp->blk_birth, ==, txg);
906 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
907 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
910 * Allocate a new log write buffer (lwb).
912 nlwb = zil_alloc_lwb(zilog, bp, txg);
914 /* Record the block for later vdev flushing */
915 zil_add_block(zilog, &lwb->lwb_blk);
918 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
919 /* For Slim ZIL only write what is used. */
920 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
921 ASSERT3U(wsz, <=, lwb->lwb_sz);
922 zio_shrink(lwb->lwb_zio, wsz);
929 zilc->zc_nused = lwb->lwb_nused;
930 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
933 * clear unused data for security
935 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
937 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
940 * If there was an allocation failure then nlwb will be null which
941 * forces a txg_wait_synced().
947 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
949 lr_t *lrc = &itx->itx_lr; /* common log record */
950 lr_write_t *lrw = (lr_write_t *)lrc;
952 uint64_t txg = lrc->lrc_txg;
953 uint64_t reclen = lrc->lrc_reclen;
959 ASSERT(lwb->lwb_buf != NULL);
961 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
962 dlen = P2ROUNDUP_TYPED(
963 lrw->lr_length, sizeof (uint64_t), uint64_t);
965 zilog->zl_cur_used += (reclen + dlen);
967 zil_lwb_write_init(zilog, lwb);
970 * If this record won't fit in the current log block, start a new one.
972 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
973 lwb = zil_lwb_write_start(zilog, lwb);
976 zil_lwb_write_init(zilog, lwb);
977 ASSERT(LWB_EMPTY(lwb));
978 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
979 txg_wait_synced(zilog->zl_dmu_pool, txg);
984 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
985 bcopy(lrc, lr_buf, reclen);
986 lrc = (lr_t *)lr_buf;
987 lrw = (lr_write_t *)lrc;
990 * If it's a write, fetch the data or get its blkptr as appropriate.
992 if (lrc->lrc_txtype == TX_WRITE) {
993 if (txg > spa_freeze_txg(zilog->zl_spa))
994 txg_wait_synced(zilog->zl_dmu_pool, txg);
995 if (itx->itx_wr_state != WR_COPIED) {
1000 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1001 dbuf = lr_buf + reclen;
1002 lrw->lr_common.lrc_reclen += dlen;
1004 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1007 error = zilog->zl_get_data(
1008 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1010 txg_wait_synced(zilog->zl_dmu_pool, txg);
1014 ASSERT(error == ENOENT || error == EEXIST ||
1022 * We're actually making an entry, so update lrc_seq to be the
1023 * log record sequence number. Note that this is generally not
1024 * equal to the itx sequence number because not all transactions
1025 * are synchronous, and sometimes spa_sync() gets there first.
1027 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1028 lwb->lwb_nused += reclen + dlen;
1029 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1030 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1031 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1037 zil_itx_create(uint64_t txtype, size_t lrsize)
1041 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1043 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1044 itx->itx_lr.lrc_txtype = txtype;
1045 itx->itx_lr.lrc_reclen = lrsize;
1046 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1047 itx->itx_lr.lrc_seq = 0; /* defensive */
1053 zil_itx_destroy(itx_t *itx)
1055 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1059 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1063 ASSERT(itx->itx_lr.lrc_seq == 0);
1064 ASSERT(!zilog->zl_replay);
1066 mutex_enter(&zilog->zl_lock);
1067 list_insert_tail(&zilog->zl_itx_list, itx);
1068 zilog->zl_itx_list_sz += itx->itx_sod;
1069 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1070 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
1071 mutex_exit(&zilog->zl_lock);
1077 * Free up all in-memory intent log transactions that have now been synced.
1080 zil_itx_clean(zilog_t *zilog)
1082 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
1083 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
1087 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1089 mutex_enter(&zilog->zl_lock);
1090 /* wait for a log writer to finish walking list */
1091 while (zilog->zl_writer) {
1092 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1096 * Move the sync'd log transactions to a separate list so we can call
1097 * kmem_free without holding the zl_lock.
1099 * There is no need to set zl_writer as we don't drop zl_lock here
1101 while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
1102 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
1103 list_remove(&zilog->zl_itx_list, itx);
1104 zilog->zl_itx_list_sz -= itx->itx_sod;
1105 list_insert_tail(&clean_list, itx);
1107 cv_broadcast(&zilog->zl_cv_writer);
1108 mutex_exit(&zilog->zl_lock);
1110 /* destroy sync'd log transactions */
1111 while ((itx = list_head(&clean_list)) != NULL) {
1112 list_remove(&clean_list, itx);
1113 zil_itx_destroy(itx);
1115 list_destroy(&clean_list);
1119 * If there are any in-memory intent log transactions which have now been
1120 * synced then start up a taskq to free them.
1123 zil_clean(zilog_t *zilog)
1127 mutex_enter(&zilog->zl_lock);
1128 itx = list_head(&zilog->zl_itx_list);
1129 if ((itx != NULL) &&
1130 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
1131 (void) taskq_dispatch(zilog->zl_clean_taskq,
1132 (task_func_t *)zil_itx_clean, zilog, TQ_NOSLEEP);
1134 mutex_exit(&zilog->zl_lock);
1138 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
1141 uint64_t commit_seq = 0;
1142 itx_t *itx, *itx_next;
1147 zilog->zl_writer = B_TRUE;
1148 ASSERT(zilog->zl_root_zio == NULL);
1149 spa = zilog->zl_spa;
1151 if (zilog->zl_suspend) {
1154 lwb = list_tail(&zilog->zl_lwb_list);
1157 * Return if there's nothing to flush before we
1158 * dirty the fs by calling zil_create()
1160 if (list_is_empty(&zilog->zl_itx_list)) {
1161 zilog->zl_writer = B_FALSE;
1164 mutex_exit(&zilog->zl_lock);
1165 lwb = zil_create(zilog);
1166 mutex_enter(&zilog->zl_lock);
1169 ASSERT(lwb == NULL || lwb->lwb_zio == NULL);
1171 /* Loop through in-memory log transactions filling log blocks. */
1172 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1174 for (itx = list_head(&zilog->zl_itx_list); itx; itx = itx_next) {
1176 * Save the next pointer. Even though we drop zl_lock below,
1177 * all threads that can remove itx list entries (other writers
1178 * and zil_itx_clean()) can't do so until they have zl_writer.
1180 itx_next = list_next(&zilog->zl_itx_list, itx);
1183 * Determine whether to push this itx.
1184 * Push all transactions related to specified foid and
1185 * all other transactions except those that can be logged
1186 * out of order (TX_WRITE, TX_TRUNCATE, TX_SETATTR, TX_ACL)
1187 * for all other files.
1189 * If foid == 0 (meaning "push all foids") or
1190 * itx->itx_sync is set (meaning O_[D]SYNC), push regardless.
1192 if (foid != 0 && !itx->itx_sync &&
1193 TX_OOO(itx->itx_lr.lrc_txtype) &&
1194 ((lr_ooo_t *)&itx->itx_lr)->lr_foid != foid)
1195 continue; /* skip this record */
1197 if ((itx->itx_lr.lrc_seq > seq) &&
1198 ((lwb == NULL) || (LWB_EMPTY(lwb)) ||
1199 (lwb->lwb_nused + itx->itx_sod > lwb->lwb_sz)))
1202 list_remove(&zilog->zl_itx_list, itx);
1203 zilog->zl_itx_list_sz -= itx->itx_sod;
1205 mutex_exit(&zilog->zl_lock);
1207 txg = itx->itx_lr.lrc_txg;
1210 if (txg > spa_last_synced_txg(spa) ||
1211 txg > spa_freeze_txg(spa))
1212 lwb = zil_lwb_commit(zilog, itx, lwb);
1214 zil_itx_destroy(itx);
1216 mutex_enter(&zilog->zl_lock);
1218 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1219 /* determine commit sequence number */
1220 itx = list_head(&zilog->zl_itx_list);
1222 commit_seq = itx->itx_lr.lrc_seq - 1;
1224 commit_seq = zilog->zl_itx_seq;
1225 mutex_exit(&zilog->zl_lock);
1227 /* write the last block out */
1228 if (lwb != NULL && lwb->lwb_zio != NULL)
1229 lwb = zil_lwb_write_start(zilog, lwb);
1231 zilog->zl_prev_used = zilog->zl_cur_used;
1232 zilog->zl_cur_used = 0;
1235 * Wait if necessary for the log blocks to be on stable storage.
1237 if (zilog->zl_root_zio) {
1238 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1239 error = zio_wait(zilog->zl_root_zio);
1240 zilog->zl_root_zio = NULL;
1241 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1242 zil_flush_vdevs(zilog);
1245 if (error || lwb == NULL)
1246 txg_wait_synced(zilog->zl_dmu_pool, 0);
1248 mutex_enter(&zilog->zl_lock);
1249 zilog->zl_writer = B_FALSE;
1251 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1252 zilog->zl_commit_seq = commit_seq;
1255 * Remember the highest committed log sequence number for ztest.
1256 * We only update this value when all the log writes succeeded,
1257 * because ztest wants to ASSERT that it got the whole log chain.
1259 if (error == 0 && lwb != NULL)
1260 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1264 * Push zfs transactions to stable storage up to the supplied sequence number.
1265 * If foid is 0 push out all transactions, otherwise push only those
1266 * for that file or might have been used to create that file.
1269 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1271 if (zilog->zl_sync == ZFS_SYNC_DISABLED || seq == 0)
1274 mutex_enter(&zilog->zl_lock);
1276 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */
1278 while (zilog->zl_writer) {
1279 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1280 if (seq <= zilog->zl_commit_seq) {
1281 mutex_exit(&zilog->zl_lock);
1285 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1286 /* wake up others waiting on the commit */
1287 cv_broadcast(&zilog->zl_cv_writer);
1288 mutex_exit(&zilog->zl_lock);
1292 * Report whether all transactions are committed.
1295 zil_is_committed(zilog_t *zilog)
1298 boolean_t committed;
1300 mutex_enter(&zilog->zl_lock);
1302 while (zilog->zl_writer)
1303 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1305 if (!list_is_empty(&zilog->zl_itx_list))
1306 committed = B_FALSE; /* unpushed transactions */
1307 else if ((lwb = list_head(&zilog->zl_lwb_list)) == NULL)
1308 committed = B_TRUE; /* intent log never used */
1309 else if (list_next(&zilog->zl_lwb_list, lwb) != NULL)
1310 committed = B_FALSE; /* zil_sync() not done yet */
1312 committed = B_TRUE; /* everything synced */
1314 mutex_exit(&zilog->zl_lock);
1319 * Called in syncing context to free committed log blocks and update log header.
1322 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1324 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1325 uint64_t txg = dmu_tx_get_txg(tx);
1326 spa_t *spa = zilog->zl_spa;
1327 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1331 * We don't zero out zl_destroy_txg, so make sure we don't try
1332 * to destroy it twice.
1334 if (spa_sync_pass(spa) != 1)
1337 mutex_enter(&zilog->zl_lock);
1339 ASSERT(zilog->zl_stop_sync == 0);
1341 if (*replayed_seq != 0) {
1342 ASSERT(zh->zh_replay_seq < *replayed_seq);
1343 zh->zh_replay_seq = *replayed_seq;
1347 if (zilog->zl_destroy_txg == txg) {
1348 blkptr_t blk = zh->zh_log;
1350 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1352 bzero(zh, sizeof (zil_header_t));
1353 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1355 if (zilog->zl_keep_first) {
1357 * If this block was part of log chain that couldn't
1358 * be claimed because a device was missing during
1359 * zil_claim(), but that device later returns,
1360 * then this block could erroneously appear valid.
1361 * To guard against this, assign a new GUID to the new
1362 * log chain so it doesn't matter what blk points to.
1364 zil_init_log_chain(zilog, &blk);
1369 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1370 zh->zh_log = lwb->lwb_blk;
1371 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1373 list_remove(&zilog->zl_lwb_list, lwb);
1374 zio_free_zil(spa, txg, &lwb->lwb_blk);
1375 kmem_cache_free(zil_lwb_cache, lwb);
1378 * If we don't have anything left in the lwb list then
1379 * we've had an allocation failure and we need to zero
1380 * out the zil_header blkptr so that we don't end
1381 * up freeing the same block twice.
1383 if (list_head(&zilog->zl_lwb_list) == NULL)
1384 BP_ZERO(&zh->zh_log);
1386 mutex_exit(&zilog->zl_lock);
1392 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1393 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1399 kmem_cache_destroy(zil_lwb_cache);
1403 zil_set_sync(zilog_t *zilog, uint64_t sync)
1405 zilog->zl_sync = sync;
1409 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1411 zilog->zl_logbias = logbias;
1415 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1419 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1421 zilog->zl_header = zh_phys;
1423 zilog->zl_spa = dmu_objset_spa(os);
1424 zilog->zl_dmu_pool = dmu_objset_pool(os);
1425 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1426 zilog->zl_logbias = dmu_objset_logbias(os);
1427 zilog->zl_sync = dmu_objset_syncprop(os);
1429 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1431 list_create(&zilog->zl_itx_list, sizeof (itx_t),
1432 offsetof(itx_t, itx_node));
1434 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1435 offsetof(lwb_t, lwb_node));
1437 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1439 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1440 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1442 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1443 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1449 zil_free(zilog_t *zilog)
1453 zilog->zl_stop_sync = 1;
1455 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1456 list_remove(&zilog->zl_lwb_list, lwb);
1457 if (lwb->lwb_buf != NULL)
1458 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1459 kmem_cache_free(zil_lwb_cache, lwb);
1461 list_destroy(&zilog->zl_lwb_list);
1463 avl_destroy(&zilog->zl_vdev_tree);
1464 mutex_destroy(&zilog->zl_vdev_lock);
1466 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1467 list_destroy(&zilog->zl_itx_list);
1468 mutex_destroy(&zilog->zl_lock);
1470 cv_destroy(&zilog->zl_cv_writer);
1471 cv_destroy(&zilog->zl_cv_suspend);
1473 kmem_free(zilog, sizeof (zilog_t));
1477 * Open an intent log.
1480 zil_open(objset_t *os, zil_get_data_t *get_data)
1482 zilog_t *zilog = dmu_objset_zil(os);
1484 zilog->zl_get_data = get_data;
1485 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1486 2, 2, TASKQ_PREPOPULATE);
1492 * Close an intent log.
1495 zil_close(zilog_t *zilog)
1498 * If the log isn't already committed, mark the objset dirty
1499 * (so zil_sync() will be called) and wait for that txg to sync.
1501 if (!zil_is_committed(zilog)) {
1503 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1504 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
1505 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1506 txg = dmu_tx_get_txg(tx);
1508 txg_wait_synced(zilog->zl_dmu_pool, txg);
1511 taskq_destroy(zilog->zl_clean_taskq);
1512 zilog->zl_clean_taskq = NULL;
1513 zilog->zl_get_data = NULL;
1515 zil_itx_clean(zilog);
1516 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1520 * Suspend an intent log. While in suspended mode, we still honor
1521 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1522 * We suspend the log briefly when taking a snapshot so that the snapshot
1523 * contains all the data it's supposed to, and has an empty intent log.
1526 zil_suspend(zilog_t *zilog)
1528 const zil_header_t *zh = zilog->zl_header;
1530 mutex_enter(&zilog->zl_lock);
1531 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1532 mutex_exit(&zilog->zl_lock);
1535 if (zilog->zl_suspend++ != 0) {
1537 * Someone else already began a suspend.
1538 * Just wait for them to finish.
1540 while (zilog->zl_suspending)
1541 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1542 mutex_exit(&zilog->zl_lock);
1545 zilog->zl_suspending = B_TRUE;
1546 mutex_exit(&zilog->zl_lock);
1548 zil_commit(zilog, UINT64_MAX, 0);
1551 * Wait for any in-flight log writes to complete.
1553 mutex_enter(&zilog->zl_lock);
1554 while (zilog->zl_writer)
1555 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1556 mutex_exit(&zilog->zl_lock);
1558 zil_destroy(zilog, B_FALSE);
1560 mutex_enter(&zilog->zl_lock);
1561 zilog->zl_suspending = B_FALSE;
1562 cv_broadcast(&zilog->zl_cv_suspend);
1563 mutex_exit(&zilog->zl_lock);
1569 zil_resume(zilog_t *zilog)
1571 mutex_enter(&zilog->zl_lock);
1572 ASSERT(zilog->zl_suspend != 0);
1573 zilog->zl_suspend--;
1574 mutex_exit(&zilog->zl_lock);
1577 typedef struct zil_replay_arg {
1578 zil_replay_func_t **zr_replay;
1580 boolean_t zr_byteswap;
1585 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1587 char name[MAXNAMELEN];
1589 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1591 dmu_objset_name(zilog->zl_os, name);
1593 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1594 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1595 (u_longlong_t)lr->lrc_seq,
1596 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1597 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1603 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1605 zil_replay_arg_t *zr = zra;
1606 const zil_header_t *zh = zilog->zl_header;
1607 uint64_t reclen = lr->lrc_reclen;
1608 uint64_t txtype = lr->lrc_txtype;
1611 zilog->zl_replaying_seq = lr->lrc_seq;
1613 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1616 if (lr->lrc_txg < claim_txg) /* already committed */
1619 /* Strip case-insensitive bit, still present in log record */
1622 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1623 return (zil_replay_error(zilog, lr, EINVAL));
1626 * If this record type can be logged out of order, the object
1627 * (lr_foid) may no longer exist. That's legitimate, not an error.
1629 if (TX_OOO(txtype)) {
1630 error = dmu_object_info(zilog->zl_os,
1631 ((lr_ooo_t *)lr)->lr_foid, NULL);
1632 if (error == ENOENT || error == EEXIST)
1637 * Make a copy of the data so we can revise and extend it.
1639 bcopy(lr, zr->zr_lr, reclen);
1642 * If this is a TX_WRITE with a blkptr, suck in the data.
1644 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1645 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1646 zr->zr_lr + reclen);
1648 return (zil_replay_error(zilog, lr, error));
1652 * The log block containing this lr may have been byteswapped
1653 * so that we can easily examine common fields like lrc_txtype.
1654 * However, the log is a mix of different record types, and only the
1655 * replay vectors know how to byteswap their records. Therefore, if
1656 * the lr was byteswapped, undo it before invoking the replay vector.
1658 if (zr->zr_byteswap)
1659 byteswap_uint64_array(zr->zr_lr, reclen);
1662 * We must now do two things atomically: replay this log record,
1663 * and update the log header sequence number to reflect the fact that
1664 * we did so. At the end of each replay function the sequence number
1665 * is updated if we are in replay mode.
1667 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1670 * The DMU's dnode layer doesn't see removes until the txg
1671 * commits, so a subsequent claim can spuriously fail with
1672 * EEXIST. So if we receive any error we try syncing out
1673 * any removes then retry the transaction. Note that we
1674 * specify B_FALSE for byteswap now, so we don't do it twice.
1676 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1677 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1679 return (zil_replay_error(zilog, lr, error));
1686 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1688 zilog->zl_replay_blks++;
1694 * If this dataset has a non-empty intent log, replay it and destroy it.
1697 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1699 zilog_t *zilog = dmu_objset_zil(os);
1700 const zil_header_t *zh = zilog->zl_header;
1701 zil_replay_arg_t zr;
1703 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1704 zil_destroy(zilog, B_TRUE);
1708 zr.zr_replay = replay_func;
1710 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1711 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1714 * Wait for in-progress removes to sync before starting replay.
1716 txg_wait_synced(zilog->zl_dmu_pool, 0);
1718 zilog->zl_replay = B_TRUE;
1719 zilog->zl_replay_time = ddi_get_lbolt();
1720 ASSERT(zilog->zl_replay_blks == 0);
1721 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1723 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1725 zil_destroy(zilog, B_FALSE);
1726 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1727 zilog->zl_replay = B_FALSE;
1731 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1733 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1736 if (zilog->zl_replay) {
1737 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1738 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1739 zilog->zl_replaying_seq;
1748 zil_vdev_offline(const char *osname, void *arg)
1754 error = dmu_objset_hold(osname, FTAG, &os);
1758 zilog = dmu_objset_zil(os);
1759 if (zil_suspend(zilog) != 0)
1763 dmu_objset_rele(os, FTAG);