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>
37 #include <sys/vdev_impl.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 */
71 SYSCTL_DECL(_vfs_zfs);
72 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
73 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
74 &zil_replay_disable, 0, "Disable intent logging replay");
77 * Tunable parameter for debugging or performance analysis. Setting
78 * zfs_nocacheflush will cause corruption on power loss if a volatile
79 * out-of-order write cache is enabled.
81 boolean_t zfs_nocacheflush = B_FALSE;
82 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
83 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
84 &zfs_nocacheflush, 0, "Disable cache flush");
86 static kmem_cache_t *zil_lwb_cache;
88 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
90 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
91 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
95 * ziltest is by and large an ugly hack, but very useful in
96 * checking replay without tedious work.
97 * When running ziltest we want to keep all itx's and so maintain
98 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
99 * We subtract TXG_CONCURRENT_STATES to allow for common code.
101 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
104 zil_bp_compare(const void *x1, const void *x2)
106 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
107 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
109 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
111 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
114 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
116 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
123 zil_bp_tree_init(zilog_t *zilog)
125 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
126 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
130 zil_bp_tree_fini(zilog_t *zilog)
132 avl_tree_t *t = &zilog->zl_bp_tree;
136 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
137 kmem_free(zn, sizeof (zil_bp_node_t));
143 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
145 avl_tree_t *t = &zilog->zl_bp_tree;
146 const dva_t *dva = BP_IDENTITY(bp);
150 if (avl_find(t, dva, &where) != NULL)
153 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
155 avl_insert(t, zn, where);
160 static zil_header_t *
161 zil_header_in_syncing_context(zilog_t *zilog)
163 return ((zil_header_t *)zilog->zl_header);
167 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
169 zio_cksum_t *zc = &bp->blk_cksum;
171 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
172 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
173 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
174 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
178 * Read a log block and make sure it's valid.
181 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
184 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
185 uint32_t aflags = ARC_WAIT;
186 arc_buf_t *abuf = NULL;
190 if (zilog->zl_header->zh_claim_txg == 0)
191 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
193 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
194 zio_flags |= ZIO_FLAG_SPECULATIVE;
196 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
197 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
199 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
200 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
203 zio_cksum_t cksum = bp->blk_cksum;
206 * Validate the checksummed log block.
208 * Sequence numbers should be... sequential. The checksum
209 * verifier for the next block should be bp's checksum plus 1.
211 * Also check the log chain linkage and size used.
213 cksum.zc_word[ZIL_ZC_SEQ]++;
215 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
216 zil_chain_t *zilc = abuf->b_data;
217 char *lr = (char *)(zilc + 1);
218 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
220 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
221 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
225 *end = (char *)dst + len;
226 *nbp = zilc->zc_next_blk;
229 char *lr = abuf->b_data;
230 uint64_t size = BP_GET_LSIZE(bp);
231 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
233 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
234 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
235 (zilc->zc_nused > (size - sizeof (*zilc)))) {
238 bcopy(lr, dst, zilc->zc_nused);
239 *end = (char *)dst + zilc->zc_nused;
240 *nbp = zilc->zc_next_blk;
244 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
251 * Read a TX_WRITE log data block.
254 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
256 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
257 const blkptr_t *bp = &lr->lr_blkptr;
258 uint32_t aflags = ARC_WAIT;
259 arc_buf_t *abuf = NULL;
263 if (BP_IS_HOLE(bp)) {
265 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
269 if (zilog->zl_header->zh_claim_txg == 0)
270 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
272 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
273 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
275 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
276 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
280 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
281 (void) arc_buf_remove_ref(abuf, &abuf);
288 * Parse the intent log, and call parse_func for each valid record within.
291 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
292 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
294 const zil_header_t *zh = zilog->zl_header;
295 boolean_t claimed = !!zh->zh_claim_txg;
296 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
297 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
298 uint64_t max_blk_seq = 0;
299 uint64_t max_lr_seq = 0;
300 uint64_t blk_count = 0;
301 uint64_t lr_count = 0;
302 blkptr_t blk, next_blk;
307 * Old logs didn't record the maximum zh_claim_lr_seq.
309 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
310 claim_lr_seq = UINT64_MAX;
313 * Starting at the block pointed to by zh_log we read the log chain.
314 * For each block in the chain we strongly check that block to
315 * ensure its validity. We stop when an invalid block is found.
316 * For each block pointer in the chain we call parse_blk_func().
317 * For each record in each valid block we call parse_lr_func().
318 * If the log has been claimed, stop if we encounter a sequence
319 * number greater than the highest claimed sequence number.
321 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
322 zil_bp_tree_init(zilog);
324 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
325 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
329 if (blk_seq > claim_blk_seq)
331 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
333 ASSERT3U(max_blk_seq, <, blk_seq);
334 max_blk_seq = blk_seq;
337 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
340 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
344 for (lrp = lrbuf; lrp < end; lrp += reclen) {
345 lr_t *lr = (lr_t *)lrp;
346 reclen = lr->lrc_reclen;
347 ASSERT3U(reclen, >=, sizeof (lr_t));
348 if (lr->lrc_seq > claim_lr_seq)
350 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
352 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
353 max_lr_seq = lr->lrc_seq;
358 zilog->zl_parse_error = error;
359 zilog->zl_parse_blk_seq = max_blk_seq;
360 zilog->zl_parse_lr_seq = max_lr_seq;
361 zilog->zl_parse_blk_count = blk_count;
362 zilog->zl_parse_lr_count = lr_count;
364 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
365 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
367 zil_bp_tree_fini(zilog);
368 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
374 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
377 * Claim log block if not already committed and not already claimed.
378 * If tx == NULL, just verify that the block is claimable.
380 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
383 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
384 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
385 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
389 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
391 lr_write_t *lr = (lr_write_t *)lrc;
394 if (lrc->lrc_txtype != TX_WRITE)
398 * If the block is not readable, don't claim it. This can happen
399 * in normal operation when a log block is written to disk before
400 * some of the dmu_sync() blocks it points to. In this case, the
401 * transaction cannot have been committed to anyone (we would have
402 * waited for all writes to be stable first), so it is semantically
403 * correct to declare this the end of the log.
405 if (lr->lr_blkptr.blk_birth >= first_txg &&
406 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
408 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
413 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
415 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
421 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
423 lr_write_t *lr = (lr_write_t *)lrc;
424 blkptr_t *bp = &lr->lr_blkptr;
427 * If we previously claimed it, we need to free it.
429 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
430 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
431 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
437 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
441 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
442 lwb->lwb_zilog = zilog;
444 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
445 lwb->lwb_max_txg = txg;
448 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
449 lwb->lwb_nused = sizeof (zil_chain_t);
450 lwb->lwb_sz = BP_GET_LSIZE(bp);
453 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
456 mutex_enter(&zilog->zl_lock);
457 list_insert_tail(&zilog->zl_lwb_list, lwb);
458 mutex_exit(&zilog->zl_lock);
464 * Create an on-disk intent log.
467 zil_create(zilog_t *zilog)
469 const zil_header_t *zh = zilog->zl_header;
477 * Wait for any previous destroy to complete.
479 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
481 ASSERT(zh->zh_claim_txg == 0);
482 ASSERT(zh->zh_replay_seq == 0);
487 * Allocate an initial log block if:
488 * - there isn't one already
489 * - the existing block is the wrong endianess
491 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
492 tx = dmu_tx_create(zilog->zl_os);
493 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
494 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
495 txg = dmu_tx_get_txg(tx);
497 if (!BP_IS_HOLE(&blk)) {
498 zio_free_zil(zilog->zl_spa, txg, &blk);
502 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
503 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
506 zil_init_log_chain(zilog, &blk);
510 * Allocate a log write buffer (lwb) for the first log block.
513 lwb = zil_alloc_lwb(zilog, &blk, txg);
516 * If we just allocated the first log block, commit our transaction
517 * and wait for zil_sync() to stuff the block poiner into zh_log.
518 * (zh is part of the MOS, so we cannot modify it in open context.)
522 txg_wait_synced(zilog->zl_dmu_pool, txg);
525 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
531 * In one tx, free all log blocks and clear the log header.
532 * If keep_first is set, then we're replaying a log with no content.
533 * We want to keep the first block, however, so that the first
534 * synchronous transaction doesn't require a txg_wait_synced()
535 * in zil_create(). We don't need to txg_wait_synced() here either
536 * when keep_first is set, because both zil_create() and zil_destroy()
537 * will wait for any in-progress destroys to complete.
540 zil_destroy(zilog_t *zilog, boolean_t keep_first)
542 const zil_header_t *zh = zilog->zl_header;
548 * Wait for any previous destroy to complete.
550 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
552 zilog->zl_old_header = *zh; /* debugging aid */
554 if (BP_IS_HOLE(&zh->zh_log))
557 tx = dmu_tx_create(zilog->zl_os);
558 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
559 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
560 txg = dmu_tx_get_txg(tx);
562 mutex_enter(&zilog->zl_lock);
564 ASSERT3U(zilog->zl_destroy_txg, <, txg);
565 zilog->zl_destroy_txg = txg;
566 zilog->zl_keep_first = keep_first;
568 if (!list_is_empty(&zilog->zl_lwb_list)) {
569 ASSERT(zh->zh_claim_txg == 0);
571 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
572 list_remove(&zilog->zl_lwb_list, lwb);
573 if (lwb->lwb_buf != NULL)
574 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
575 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
576 kmem_cache_free(zil_lwb_cache, lwb);
578 } else if (!keep_first) {
579 (void) zil_parse(zilog, zil_free_log_block,
580 zil_free_log_record, tx, zh->zh_claim_txg);
582 mutex_exit(&zilog->zl_lock);
588 zil_claim(const char *osname, void *txarg)
590 dmu_tx_t *tx = txarg;
591 uint64_t first_txg = dmu_tx_get_txg(tx);
597 error = dmu_objset_hold(osname, FTAG, &os);
599 cmn_err(CE_WARN, "can't open objset for %s", osname);
603 zilog = dmu_objset_zil(os);
604 zh = zil_header_in_syncing_context(zilog);
606 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
607 if (!BP_IS_HOLE(&zh->zh_log))
608 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
609 BP_ZERO(&zh->zh_log);
610 dsl_dataset_dirty(dmu_objset_ds(os), tx);
611 dmu_objset_rele(os, FTAG);
616 * Claim all log blocks if we haven't already done so, and remember
617 * the highest claimed sequence number. This ensures that if we can
618 * read only part of the log now (e.g. due to a missing device),
619 * but we can read the entire log later, we will not try to replay
620 * or destroy beyond the last block we successfully claimed.
622 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
623 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
624 (void) zil_parse(zilog, zil_claim_log_block,
625 zil_claim_log_record, tx, first_txg);
626 zh->zh_claim_txg = first_txg;
627 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
628 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
629 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
630 zh->zh_flags |= ZIL_REPLAY_NEEDED;
631 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
632 dsl_dataset_dirty(dmu_objset_ds(os), tx);
635 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
636 dmu_objset_rele(os, FTAG);
641 * Check the log by walking the log chain.
642 * Checksum errors are ok as they indicate the end of the chain.
643 * Any other error (no device or read failure) returns an error.
646 zil_check_log_chain(const char *osname, void *tx)
655 error = dmu_objset_hold(osname, FTAG, &os);
657 cmn_err(CE_WARN, "can't open objset for %s", osname);
661 zilog = dmu_objset_zil(os);
662 bp = (blkptr_t *)&zilog->zl_header->zh_log;
665 * Check the first block and determine if it's on a log device
666 * which may have been removed or faulted prior to loading this
667 * pool. If so, there's no point in checking the rest of the log
668 * as its content should have already been synced to the pool.
670 if (!BP_IS_HOLE(bp)) {
672 boolean_t valid = B_TRUE;
674 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
675 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
676 if (vd->vdev_islog && vdev_is_dead(vd))
677 valid = vdev_log_state_valid(vd);
678 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
681 dmu_objset_rele(os, FTAG);
687 * Because tx == NULL, zil_claim_log_block() will not actually claim
688 * any blocks, but just determine whether it is possible to do so.
689 * In addition to checking the log chain, zil_claim_log_block()
690 * will invoke zio_claim() with a done func of spa_claim_notify(),
691 * which will update spa_max_claim_txg. See spa_load() for details.
693 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
694 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
696 dmu_objset_rele(os, FTAG);
698 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
702 zil_vdev_compare(const void *x1, const void *x2)
704 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
705 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
716 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
718 avl_tree_t *t = &zilog->zl_vdev_tree;
720 zil_vdev_node_t *zv, zvsearch;
721 int ndvas = BP_GET_NDVAS(bp);
724 if (zfs_nocacheflush)
727 ASSERT(zilog->zl_writer);
730 * Even though we're zl_writer, we still need a lock because the
731 * zl_get_data() callbacks may have dmu_sync() done callbacks
732 * that will run concurrently.
734 mutex_enter(&zilog->zl_vdev_lock);
735 for (i = 0; i < ndvas; i++) {
736 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
737 if (avl_find(t, &zvsearch, &where) == NULL) {
738 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
739 zv->zv_vdev = zvsearch.zv_vdev;
740 avl_insert(t, zv, where);
743 mutex_exit(&zilog->zl_vdev_lock);
747 zil_flush_vdevs(zilog_t *zilog)
749 spa_t *spa = zilog->zl_spa;
750 avl_tree_t *t = &zilog->zl_vdev_tree;
755 ASSERT(zilog->zl_writer);
758 * We don't need zl_vdev_lock here because we're the zl_writer,
759 * and all zl_get_data() callbacks are done.
761 if (avl_numnodes(t) == 0)
764 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
766 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
768 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
769 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
772 kmem_free(zv, sizeof (*zv));
776 * Wait for all the flushes to complete. Not all devices actually
777 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
779 (void) zio_wait(zio);
781 spa_config_exit(spa, SCL_STATE, FTAG);
785 * Function called when a log block write completes
788 zil_lwb_write_done(zio_t *zio)
790 lwb_t *lwb = zio->io_private;
791 zilog_t *zilog = lwb->lwb_zilog;
792 dmu_tx_t *tx = lwb->lwb_tx;
794 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
795 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
796 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
797 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
798 ASSERT(!BP_IS_GANG(zio->io_bp));
799 ASSERT(!BP_IS_HOLE(zio->io_bp));
800 ASSERT(zio->io_bp->blk_fill == 0);
803 * Ensure the lwb buffer pointer is cleared before releasing
804 * the txg. If we have had an allocation failure and
805 * the txg is waiting to sync then we want want zil_sync()
806 * to remove the lwb so that it's not picked up as the next new
807 * one in zil_commit_writer(). zil_sync() will only remove
808 * the lwb if lwb_buf is null.
810 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
811 mutex_enter(&zilog->zl_lock);
814 mutex_exit(&zilog->zl_lock);
817 * Now that we've written this log block, we have a stable pointer
818 * to the next block in the chain, so it's OK to let the txg in
819 * which we allocated the next block sync.
825 * Initialize the io for a log block.
828 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
832 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
833 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
834 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
836 if (zilog->zl_root_zio == NULL) {
837 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
840 if (lwb->lwb_zio == NULL) {
841 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
842 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
843 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
844 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
849 * Define a limited set of intent log block sizes.
850 * These must be a multiple of 4KB. Note only the amount used (again
851 * aligned to 4KB) actually gets written. However, we can't always just
852 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
854 uint64_t zil_block_buckets[] = {
855 4096, /* non TX_WRITE */
856 8192+4096, /* data base */
857 32*1024 + 4096, /* NFS writes */
862 * Use the slog as long as the logbias is 'latency' and the current commit size
863 * is less than the limit or the total list size is less than 2X the limit.
864 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
866 uint64_t zil_slog_limit = 1024 * 1024;
867 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
868 (((zilog)->zl_cur_used < zil_slog_limit) || \
869 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
872 * Start a log block write and advance to the next log block.
873 * Calls are serialized.
876 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
880 spa_t *spa = zilog->zl_spa;
884 uint64_t zil_blksz, wsz;
887 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
888 zilc = (zil_chain_t *)lwb->lwb_buf;
889 bp = &zilc->zc_next_blk;
891 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
892 bp = &zilc->zc_next_blk;
895 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
898 * Allocate the next block and save its address in this block
899 * before writing it in order to establish the log chain.
900 * Note that if the allocation of nlwb synced before we wrote
901 * the block that points at it (lwb), we'd leak it if we crashed.
902 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
903 * We dirty the dataset to ensure that zil_sync() will be called
904 * to clean up in the event of allocation failure or I/O failure.
906 tx = dmu_tx_create(zilog->zl_os);
907 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
908 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
909 txg = dmu_tx_get_txg(tx);
914 * Log blocks are pre-allocated. Here we select the size of the next
915 * block, based on size used in the last block.
916 * - first find the smallest bucket that will fit the block from a
917 * limited set of block sizes. This is because it's faster to write
918 * blocks allocated from the same metaslab as they are adjacent or
920 * - next find the maximum from the new suggested size and an array of
921 * previous sizes. This lessens a picket fence effect of wrongly
922 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
925 * Note we only write what is used, but we can't just allocate
926 * the maximum block size because we can exhaust the available
929 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
930 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
932 zil_blksz = zil_block_buckets[i];
933 if (zil_blksz == UINT64_MAX)
934 zil_blksz = SPA_MAXBLOCKSIZE;
935 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
936 for (i = 0; i < ZIL_PREV_BLKS; i++)
937 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
938 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
941 /* pass the old blkptr in order to spread log blocks across devs */
942 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
945 ASSERT3U(bp->blk_birth, ==, txg);
946 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
947 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
950 * Allocate a new log write buffer (lwb).
952 nlwb = zil_alloc_lwb(zilog, bp, txg);
954 /* Record the block for later vdev flushing */
955 zil_add_block(zilog, &lwb->lwb_blk);
958 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
959 /* For Slim ZIL only write what is used. */
960 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
961 ASSERT3U(wsz, <=, lwb->lwb_sz);
962 zio_shrink(lwb->lwb_zio, wsz);
969 zilc->zc_nused = lwb->lwb_nused;
970 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
973 * clear unused data for security
975 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
977 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
980 * If there was an allocation failure then nlwb will be null which
981 * forces a txg_wait_synced().
987 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
989 lr_t *lrc = &itx->itx_lr; /* common log record */
990 lr_write_t *lrw = (lr_write_t *)lrc;
992 uint64_t txg = lrc->lrc_txg;
993 uint64_t reclen = lrc->lrc_reclen;
999 ASSERT(lwb->lwb_buf != NULL);
1001 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1002 dlen = P2ROUNDUP_TYPED(
1003 lrw->lr_length, sizeof (uint64_t), uint64_t);
1005 zilog->zl_cur_used += (reclen + dlen);
1007 zil_lwb_write_init(zilog, lwb);
1010 * If this record won't fit in the current log block, start a new one.
1012 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1013 lwb = zil_lwb_write_start(zilog, lwb);
1016 zil_lwb_write_init(zilog, lwb);
1017 ASSERT(LWB_EMPTY(lwb));
1018 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1019 txg_wait_synced(zilog->zl_dmu_pool, txg);
1024 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1025 bcopy(lrc, lr_buf, reclen);
1026 lrc = (lr_t *)lr_buf;
1027 lrw = (lr_write_t *)lrc;
1030 * If it's a write, fetch the data or get its blkptr as appropriate.
1032 if (lrc->lrc_txtype == TX_WRITE) {
1033 if (txg > spa_freeze_txg(zilog->zl_spa))
1034 txg_wait_synced(zilog->zl_dmu_pool, txg);
1035 if (itx->itx_wr_state != WR_COPIED) {
1040 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1041 dbuf = lr_buf + reclen;
1042 lrw->lr_common.lrc_reclen += dlen;
1044 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1047 error = zilog->zl_get_data(
1048 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1050 txg_wait_synced(zilog->zl_dmu_pool, txg);
1054 ASSERT(error == ENOENT || error == EEXIST ||
1062 * We're actually making an entry, so update lrc_seq to be the
1063 * log record sequence number. Note that this is generally not
1064 * equal to the itx sequence number because not all transactions
1065 * are synchronous, and sometimes spa_sync() gets there first.
1067 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1068 lwb->lwb_nused += reclen + dlen;
1069 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1070 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1071 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1077 zil_itx_create(uint64_t txtype, size_t lrsize)
1081 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1083 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1084 itx->itx_lr.lrc_txtype = txtype;
1085 itx->itx_lr.lrc_reclen = lrsize;
1086 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1087 itx->itx_lr.lrc_seq = 0; /* defensive */
1088 itx->itx_sync = B_TRUE; /* default is synchronous */
1094 zil_itx_destroy(itx_t *itx)
1096 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1100 * Free up the sync and async itxs. The itxs_t has already been detached
1101 * so no locks are needed.
1104 zil_itxg_clean(itxs_t *itxs)
1110 itx_async_node_t *ian;
1112 list = &itxs->i_sync_list;
1113 while ((itx = list_head(list)) != NULL) {
1114 list_remove(list, itx);
1115 kmem_free(itx, offsetof(itx_t, itx_lr) +
1116 itx->itx_lr.lrc_reclen);
1120 t = &itxs->i_async_tree;
1121 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1122 list = &ian->ia_list;
1123 while ((itx = list_head(list)) != NULL) {
1124 list_remove(list, itx);
1125 kmem_free(itx, offsetof(itx_t, itx_lr) +
1126 itx->itx_lr.lrc_reclen);
1129 kmem_free(ian, sizeof (itx_async_node_t));
1133 kmem_free(itxs, sizeof (itxs_t));
1137 zil_aitx_compare(const void *x1, const void *x2)
1139 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1140 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1151 * Remove all async itx with the given oid.
1154 zil_remove_async(zilog_t *zilog, uint64_t oid)
1157 itx_async_node_t *ian;
1164 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1166 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1169 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1171 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1172 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1174 mutex_enter(&itxg->itxg_lock);
1175 if (itxg->itxg_txg != txg) {
1176 mutex_exit(&itxg->itxg_lock);
1181 * Locate the object node and append its list.
1183 t = &itxg->itxg_itxs->i_async_tree;
1184 ian = avl_find(t, &oid, &where);
1186 list_move_tail(&clean_list, &ian->ia_list);
1187 mutex_exit(&itxg->itxg_lock);
1189 while ((itx = list_head(&clean_list)) != NULL) {
1190 list_remove(&clean_list, itx);
1191 kmem_free(itx, offsetof(itx_t, itx_lr) +
1192 itx->itx_lr.lrc_reclen);
1194 list_destroy(&clean_list);
1198 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1202 itxs_t *itxs, *clean = NULL;
1205 * Object ids can be re-instantiated in the next txg so
1206 * remove any async transactions to avoid future leaks.
1207 * This can happen if a fsync occurs on the re-instantiated
1208 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1209 * the new file data and flushes a write record for the old object.
1211 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1212 zil_remove_async(zilog, itx->itx_oid);
1215 * Ensure the data of a renamed file is committed before the rename.
1217 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1218 zil_async_to_sync(zilog, itx->itx_oid);
1220 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1223 txg = dmu_tx_get_txg(tx);
1225 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1226 mutex_enter(&itxg->itxg_lock);
1227 itxs = itxg->itxg_itxs;
1228 if (itxg->itxg_txg != txg) {
1231 * The zil_clean callback hasn't got around to cleaning
1232 * this itxg. Save the itxs for release below.
1233 * This should be rare.
1235 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1237 clean = itxg->itxg_itxs;
1239 ASSERT(itxg->itxg_sod == 0);
1240 itxg->itxg_txg = txg;
1241 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1243 list_create(&itxs->i_sync_list, sizeof (itx_t),
1244 offsetof(itx_t, itx_node));
1245 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1246 sizeof (itx_async_node_t),
1247 offsetof(itx_async_node_t, ia_node));
1249 if (itx->itx_sync) {
1250 list_insert_tail(&itxs->i_sync_list, itx);
1251 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1252 itxg->itxg_sod += itx->itx_sod;
1254 avl_tree_t *t = &itxs->i_async_tree;
1255 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1256 itx_async_node_t *ian;
1259 ian = avl_find(t, &foid, &where);
1261 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1262 list_create(&ian->ia_list, sizeof (itx_t),
1263 offsetof(itx_t, itx_node));
1264 ian->ia_foid = foid;
1265 avl_insert(t, ian, where);
1267 list_insert_tail(&ian->ia_list, itx);
1270 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1271 mutex_exit(&itxg->itxg_lock);
1273 /* Release the old itxs now we've dropped the lock */
1275 zil_itxg_clean(clean);
1279 * If there are any in-memory intent log transactions which have now been
1280 * synced then start up a taskq to free them.
1283 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1285 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1288 mutex_enter(&itxg->itxg_lock);
1289 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1290 mutex_exit(&itxg->itxg_lock);
1293 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1294 ASSERT(itxg->itxg_txg != 0);
1295 ASSERT(zilog->zl_clean_taskq != NULL);
1296 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1298 clean_me = itxg->itxg_itxs;
1299 itxg->itxg_itxs = NULL;
1301 mutex_exit(&itxg->itxg_lock);
1303 * Preferably start a task queue to free up the old itxs but
1304 * if taskq_dispatch can't allocate resources to do that then
1305 * free it in-line. This should be rare. Note, using TQ_SLEEP
1306 * created a bad performance problem.
1308 if (taskq_dispatch(zilog->zl_clean_taskq,
1309 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1310 zil_itxg_clean(clean_me);
1314 * Get the list of itxs to commit into zl_itx_commit_list.
1317 zil_get_commit_list(zilog_t *zilog)
1320 list_t *commit_list = &zilog->zl_itx_commit_list;
1321 uint64_t push_sod = 0;
1323 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1326 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1328 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1329 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1331 mutex_enter(&itxg->itxg_lock);
1332 if (itxg->itxg_txg != txg) {
1333 mutex_exit(&itxg->itxg_lock);
1337 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1338 push_sod += itxg->itxg_sod;
1341 mutex_exit(&itxg->itxg_lock);
1343 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1347 * Move the async itxs for a specified object to commit into sync lists.
1350 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1353 itx_async_node_t *ian;
1357 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1360 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1362 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1363 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1365 mutex_enter(&itxg->itxg_lock);
1366 if (itxg->itxg_txg != txg) {
1367 mutex_exit(&itxg->itxg_lock);
1372 * If a foid is specified then find that node and append its
1373 * list. Otherwise walk the tree appending all the lists
1374 * to the sync list. We add to the end rather than the
1375 * beginning to ensure the create has happened.
1377 t = &itxg->itxg_itxs->i_async_tree;
1379 ian = avl_find(t, &foid, &where);
1381 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1385 void *cookie = NULL;
1387 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1388 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1390 list_destroy(&ian->ia_list);
1391 kmem_free(ian, sizeof (itx_async_node_t));
1394 mutex_exit(&itxg->itxg_lock);
1399 zil_commit_writer(zilog_t *zilog)
1404 spa_t *spa = zilog->zl_spa;
1407 ASSERT(zilog->zl_root_zio == NULL);
1409 mutex_exit(&zilog->zl_lock);
1411 zil_get_commit_list(zilog);
1414 * Return if there's nothing to commit before we dirty the fs by
1415 * calling zil_create().
1417 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1418 mutex_enter(&zilog->zl_lock);
1422 if (zilog->zl_suspend) {
1425 lwb = list_tail(&zilog->zl_lwb_list);
1427 lwb = zil_create(zilog);
1430 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1431 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1432 txg = itx->itx_lr.lrc_txg;
1435 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1436 lwb = zil_lwb_commit(zilog, itx, lwb);
1437 list_remove(&zilog->zl_itx_commit_list, itx);
1438 kmem_free(itx, offsetof(itx_t, itx_lr)
1439 + itx->itx_lr.lrc_reclen);
1441 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1443 /* write the last block out */
1444 if (lwb != NULL && lwb->lwb_zio != NULL)
1445 lwb = zil_lwb_write_start(zilog, lwb);
1447 zilog->zl_cur_used = 0;
1450 * Wait if necessary for the log blocks to be on stable storage.
1452 if (zilog->zl_root_zio) {
1453 error = zio_wait(zilog->zl_root_zio);
1454 zilog->zl_root_zio = NULL;
1455 zil_flush_vdevs(zilog);
1458 if (error || lwb == NULL)
1459 txg_wait_synced(zilog->zl_dmu_pool, 0);
1461 mutex_enter(&zilog->zl_lock);
1464 * Remember the highest committed log sequence number for ztest.
1465 * We only update this value when all the log writes succeeded,
1466 * because ztest wants to ASSERT that it got the whole log chain.
1468 if (error == 0 && lwb != NULL)
1469 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1473 * Commit zfs transactions to stable storage.
1474 * If foid is 0 push out all transactions, otherwise push only those
1475 * for that object or might reference that object.
1477 * itxs are committed in batches. In a heavily stressed zil there will be
1478 * a commit writer thread who is writing out a bunch of itxs to the log
1479 * for a set of committing threads (cthreads) in the same batch as the writer.
1480 * Those cthreads are all waiting on the same cv for that batch.
1482 * There will also be a different and growing batch of threads that are
1483 * waiting to commit (qthreads). When the committing batch completes
1484 * a transition occurs such that the cthreads exit and the qthreads become
1485 * cthreads. One of the new cthreads becomes the writer thread for the
1486 * batch. Any new threads arriving become new qthreads.
1488 * Only 2 condition variables are needed and there's no transition
1489 * between the two cvs needed. They just flip-flop between qthreads
1492 * Using this scheme we can efficiently wakeup up only those threads
1493 * that have been committed.
1496 zil_commit(zilog_t *zilog, uint64_t foid)
1500 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1503 /* move the async itxs for the foid to the sync queues */
1504 zil_async_to_sync(zilog, foid);
1506 mutex_enter(&zilog->zl_lock);
1507 mybatch = zilog->zl_next_batch;
1508 while (zilog->zl_writer) {
1509 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1510 if (mybatch <= zilog->zl_com_batch) {
1511 mutex_exit(&zilog->zl_lock);
1516 zilog->zl_next_batch++;
1517 zilog->zl_writer = B_TRUE;
1518 zil_commit_writer(zilog);
1519 zilog->zl_com_batch = mybatch;
1520 zilog->zl_writer = B_FALSE;
1521 mutex_exit(&zilog->zl_lock);
1523 /* wake up one thread to become the next writer */
1524 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1526 /* wake up all threads waiting for this batch to be committed */
1527 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1531 * Called in syncing context to free committed log blocks and update log header.
1534 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1536 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1537 uint64_t txg = dmu_tx_get_txg(tx);
1538 spa_t *spa = zilog->zl_spa;
1539 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1543 * We don't zero out zl_destroy_txg, so make sure we don't try
1544 * to destroy it twice.
1546 if (spa_sync_pass(spa) != 1)
1549 mutex_enter(&zilog->zl_lock);
1551 ASSERT(zilog->zl_stop_sync == 0);
1553 if (*replayed_seq != 0) {
1554 ASSERT(zh->zh_replay_seq < *replayed_seq);
1555 zh->zh_replay_seq = *replayed_seq;
1559 if (zilog->zl_destroy_txg == txg) {
1560 blkptr_t blk = zh->zh_log;
1562 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1564 bzero(zh, sizeof (zil_header_t));
1565 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1567 if (zilog->zl_keep_first) {
1569 * If this block was part of log chain that couldn't
1570 * be claimed because a device was missing during
1571 * zil_claim(), but that device later returns,
1572 * then this block could erroneously appear valid.
1573 * To guard against this, assign a new GUID to the new
1574 * log chain so it doesn't matter what blk points to.
1576 zil_init_log_chain(zilog, &blk);
1581 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1582 zh->zh_log = lwb->lwb_blk;
1583 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1585 list_remove(&zilog->zl_lwb_list, lwb);
1586 zio_free_zil(spa, txg, &lwb->lwb_blk);
1587 kmem_cache_free(zil_lwb_cache, lwb);
1590 * If we don't have anything left in the lwb list then
1591 * we've had an allocation failure and we need to zero
1592 * out the zil_header blkptr so that we don't end
1593 * up freeing the same block twice.
1595 if (list_head(&zilog->zl_lwb_list) == NULL)
1596 BP_ZERO(&zh->zh_log);
1598 mutex_exit(&zilog->zl_lock);
1604 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1605 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1611 kmem_cache_destroy(zil_lwb_cache);
1615 zil_set_sync(zilog_t *zilog, uint64_t sync)
1617 zilog->zl_sync = sync;
1621 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1623 zilog->zl_logbias = logbias;
1627 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1631 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1633 zilog->zl_header = zh_phys;
1635 zilog->zl_spa = dmu_objset_spa(os);
1636 zilog->zl_dmu_pool = dmu_objset_pool(os);
1637 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1638 zilog->zl_logbias = dmu_objset_logbias(os);
1639 zilog->zl_sync = dmu_objset_syncprop(os);
1640 zilog->zl_next_batch = 1;
1642 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1644 for (int i = 0; i < TXG_SIZE; i++) {
1645 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1646 MUTEX_DEFAULT, NULL);
1649 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1650 offsetof(lwb_t, lwb_node));
1652 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1653 offsetof(itx_t, itx_node));
1655 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1657 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1658 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1660 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1661 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1662 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1663 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1669 zil_free(zilog_t *zilog)
1673 zilog->zl_stop_sync = 1;
1676 * After zil_close() there should only be one lwb with a buffer.
1678 head_lwb = list_head(&zilog->zl_lwb_list);
1680 ASSERT(head_lwb == list_tail(&zilog->zl_lwb_list));
1681 list_remove(&zilog->zl_lwb_list, head_lwb);
1682 zio_buf_free(head_lwb->lwb_buf, head_lwb->lwb_sz);
1683 kmem_cache_free(zil_lwb_cache, head_lwb);
1685 list_destroy(&zilog->zl_lwb_list);
1687 avl_destroy(&zilog->zl_vdev_tree);
1688 mutex_destroy(&zilog->zl_vdev_lock);
1690 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1691 list_destroy(&zilog->zl_itx_commit_list);
1693 for (int i = 0; i < TXG_SIZE; i++) {
1695 * It's possible for an itx to be generated that doesn't dirty
1696 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1697 * callback to remove the entry. We remove those here.
1699 * Also free up the ziltest itxs.
1701 if (zilog->zl_itxg[i].itxg_itxs)
1702 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1703 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1706 mutex_destroy(&zilog->zl_lock);
1708 cv_destroy(&zilog->zl_cv_writer);
1709 cv_destroy(&zilog->zl_cv_suspend);
1710 cv_destroy(&zilog->zl_cv_batch[0]);
1711 cv_destroy(&zilog->zl_cv_batch[1]);
1713 kmem_free(zilog, sizeof (zilog_t));
1717 * Open an intent log.
1720 zil_open(objset_t *os, zil_get_data_t *get_data)
1722 zilog_t *zilog = dmu_objset_zil(os);
1724 zilog->zl_get_data = get_data;
1725 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1726 2, 2, TASKQ_PREPOPULATE);
1732 * Close an intent log.
1735 zil_close(zilog_t *zilog)
1740 zil_commit(zilog, 0); /* commit all itx */
1743 * The lwb_max_txg for the stubby lwb will reflect the last activity
1744 * for the zil. After a txg_wait_synced() on the txg we know all the
1745 * callbacks have occurred that may clean the zil. Only then can we
1746 * destroy the zl_clean_taskq.
1748 mutex_enter(&zilog->zl_lock);
1749 tail_lwb = list_tail(&zilog->zl_lwb_list);
1750 if (tail_lwb != NULL)
1751 txg = tail_lwb->lwb_max_txg;
1752 mutex_exit(&zilog->zl_lock);
1754 txg_wait_synced(zilog->zl_dmu_pool, txg);
1756 taskq_destroy(zilog->zl_clean_taskq);
1757 zilog->zl_clean_taskq = NULL;
1758 zilog->zl_get_data = NULL;
1762 * Suspend an intent log. While in suspended mode, we still honor
1763 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1764 * We suspend the log briefly when taking a snapshot so that the snapshot
1765 * contains all the data it's supposed to, and has an empty intent log.
1768 zil_suspend(zilog_t *zilog)
1770 const zil_header_t *zh = zilog->zl_header;
1772 mutex_enter(&zilog->zl_lock);
1773 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1774 mutex_exit(&zilog->zl_lock);
1777 if (zilog->zl_suspend++ != 0) {
1779 * Someone else already began a suspend.
1780 * Just wait for them to finish.
1782 while (zilog->zl_suspending)
1783 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1784 mutex_exit(&zilog->zl_lock);
1787 zilog->zl_suspending = B_TRUE;
1788 mutex_exit(&zilog->zl_lock);
1790 zil_commit(zilog, 0);
1792 zil_destroy(zilog, B_FALSE);
1794 mutex_enter(&zilog->zl_lock);
1795 zilog->zl_suspending = B_FALSE;
1796 cv_broadcast(&zilog->zl_cv_suspend);
1797 mutex_exit(&zilog->zl_lock);
1803 zil_resume(zilog_t *zilog)
1805 mutex_enter(&zilog->zl_lock);
1806 ASSERT(zilog->zl_suspend != 0);
1807 zilog->zl_suspend--;
1808 mutex_exit(&zilog->zl_lock);
1811 typedef struct zil_replay_arg {
1812 zil_replay_func_t **zr_replay;
1814 boolean_t zr_byteswap;
1819 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1821 char name[MAXNAMELEN];
1823 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1825 dmu_objset_name(zilog->zl_os, name);
1827 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1828 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1829 (u_longlong_t)lr->lrc_seq,
1830 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1831 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1837 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1839 zil_replay_arg_t *zr = zra;
1840 const zil_header_t *zh = zilog->zl_header;
1841 uint64_t reclen = lr->lrc_reclen;
1842 uint64_t txtype = lr->lrc_txtype;
1845 zilog->zl_replaying_seq = lr->lrc_seq;
1847 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1850 if (lr->lrc_txg < claim_txg) /* already committed */
1853 /* Strip case-insensitive bit, still present in log record */
1856 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1857 return (zil_replay_error(zilog, lr, EINVAL));
1860 * If this record type can be logged out of order, the object
1861 * (lr_foid) may no longer exist. That's legitimate, not an error.
1863 if (TX_OOO(txtype)) {
1864 error = dmu_object_info(zilog->zl_os,
1865 ((lr_ooo_t *)lr)->lr_foid, NULL);
1866 if (error == ENOENT || error == EEXIST)
1871 * Make a copy of the data so we can revise and extend it.
1873 bcopy(lr, zr->zr_lr, reclen);
1876 * If this is a TX_WRITE with a blkptr, suck in the data.
1878 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1879 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1880 zr->zr_lr + reclen);
1882 return (zil_replay_error(zilog, lr, error));
1886 * The log block containing this lr may have been byteswapped
1887 * so that we can easily examine common fields like lrc_txtype.
1888 * However, the log is a mix of different record types, and only the
1889 * replay vectors know how to byteswap their records. Therefore, if
1890 * the lr was byteswapped, undo it before invoking the replay vector.
1892 if (zr->zr_byteswap)
1893 byteswap_uint64_array(zr->zr_lr, reclen);
1896 * We must now do two things atomically: replay this log record,
1897 * and update the log header sequence number to reflect the fact that
1898 * we did so. At the end of each replay function the sequence number
1899 * is updated if we are in replay mode.
1901 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1904 * The DMU's dnode layer doesn't see removes until the txg
1905 * commits, so a subsequent claim can spuriously fail with
1906 * EEXIST. So if we receive any error we try syncing out
1907 * any removes then retry the transaction. Note that we
1908 * specify B_FALSE for byteswap now, so we don't do it twice.
1910 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1911 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1913 return (zil_replay_error(zilog, lr, error));
1920 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1922 zilog->zl_replay_blks++;
1928 * If this dataset has a non-empty intent log, replay it and destroy it.
1931 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1933 zilog_t *zilog = dmu_objset_zil(os);
1934 const zil_header_t *zh = zilog->zl_header;
1935 zil_replay_arg_t zr;
1937 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1938 zil_destroy(zilog, B_TRUE);
1941 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
1943 zr.zr_replay = replay_func;
1945 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1946 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1949 * Wait for in-progress removes to sync before starting replay.
1951 txg_wait_synced(zilog->zl_dmu_pool, 0);
1953 zilog->zl_replay = B_TRUE;
1954 zilog->zl_replay_time = ddi_get_lbolt();
1955 ASSERT(zilog->zl_replay_blks == 0);
1956 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1958 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1960 zil_destroy(zilog, B_FALSE);
1961 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1962 zilog->zl_replay = B_FALSE;
1963 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
1967 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1969 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1972 if (zilog->zl_replay) {
1973 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1974 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1975 zilog->zl_replaying_seq;
1984 zil_vdev_offline(const char *osname, void *arg)
1990 error = dmu_objset_hold(osname, FTAG, &os);
1994 zilog = dmu_objset_zil(os);
1995 if (zil_suspend(zilog) != 0)
1999 dmu_objset_rele(os, FTAG);