4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
52 * There is one ZIL per file system. Its on-disk (pool) format consists
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
69 * This global ZIL switch affects all pools
71 int zil_replay_disable = 0; /* disable intent logging replay */
72 SYSCTL_DECL(_vfs_zfs);
73 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
74 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
75 &zil_replay_disable, 0, "Disable intent logging replay");
78 * Tunable parameter for debugging or performance analysis. Setting
79 * zfs_nocacheflush will cause corruption on power loss if a volatile
80 * out-of-order write cache is enabled.
82 boolean_t zfs_nocacheflush = B_FALSE;
83 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
84 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
85 &zfs_nocacheflush, 0, "Disable cache flush");
87 static kmem_cache_t *zil_lwb_cache;
89 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
91 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
92 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
96 * ziltest is by and large an ugly hack, but very useful in
97 * checking replay without tedious work.
98 * When running ziltest we want to keep all itx's and so maintain
99 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
100 * We subtract TXG_CONCURRENT_STATES to allow for common code.
102 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
105 zil_bp_compare(const void *x1, const void *x2)
107 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
108 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
110 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
112 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
115 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
117 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
124 zil_bp_tree_init(zilog_t *zilog)
126 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
127 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
131 zil_bp_tree_fini(zilog_t *zilog)
133 avl_tree_t *t = &zilog->zl_bp_tree;
137 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
138 kmem_free(zn, sizeof (zil_bp_node_t));
144 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
146 avl_tree_t *t = &zilog->zl_bp_tree;
147 const dva_t *dva = BP_IDENTITY(bp);
151 if (avl_find(t, dva, &where) != NULL)
154 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
156 avl_insert(t, zn, where);
161 static zil_header_t *
162 zil_header_in_syncing_context(zilog_t *zilog)
164 return ((zil_header_t *)zilog->zl_header);
168 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
170 zio_cksum_t *zc = &bp->blk_cksum;
172 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
173 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
174 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
175 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
179 * Read a log block and make sure it's valid.
182 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
185 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
186 uint32_t aflags = ARC_WAIT;
187 arc_buf_t *abuf = NULL;
191 if (zilog->zl_header->zh_claim_txg == 0)
192 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
194 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
195 zio_flags |= ZIO_FLAG_SPECULATIVE;
197 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
198 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
200 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
201 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
204 zio_cksum_t cksum = bp->blk_cksum;
207 * Validate the checksummed log block.
209 * Sequence numbers should be... sequential. The checksum
210 * verifier for the next block should be bp's checksum plus 1.
212 * Also check the log chain linkage and size used.
214 cksum.zc_word[ZIL_ZC_SEQ]++;
216 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
217 zil_chain_t *zilc = abuf->b_data;
218 char *lr = (char *)(zilc + 1);
219 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
221 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
222 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
226 *end = (char *)dst + len;
227 *nbp = zilc->zc_next_blk;
230 char *lr = abuf->b_data;
231 uint64_t size = BP_GET_LSIZE(bp);
232 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
234 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
235 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
236 (zilc->zc_nused > (size - sizeof (*zilc)))) {
239 bcopy(lr, dst, zilc->zc_nused);
240 *end = (char *)dst + zilc->zc_nused;
241 *nbp = zilc->zc_next_blk;
245 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
252 * Read a TX_WRITE log data block.
255 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
257 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
258 const blkptr_t *bp = &lr->lr_blkptr;
259 uint32_t aflags = ARC_WAIT;
260 arc_buf_t *abuf = NULL;
264 if (BP_IS_HOLE(bp)) {
266 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
270 if (zilog->zl_header->zh_claim_txg == 0)
271 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
273 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
274 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
276 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
277 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
281 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
282 (void) arc_buf_remove_ref(abuf, &abuf);
289 * Parse the intent log, and call parse_func for each valid record within.
292 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
293 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
295 const zil_header_t *zh = zilog->zl_header;
296 boolean_t claimed = !!zh->zh_claim_txg;
297 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
298 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
299 uint64_t max_blk_seq = 0;
300 uint64_t max_lr_seq = 0;
301 uint64_t blk_count = 0;
302 uint64_t lr_count = 0;
303 blkptr_t blk, next_blk;
308 * Old logs didn't record the maximum zh_claim_lr_seq.
310 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
311 claim_lr_seq = UINT64_MAX;
314 * Starting at the block pointed to by zh_log we read the log chain.
315 * For each block in the chain we strongly check that block to
316 * ensure its validity. We stop when an invalid block is found.
317 * For each block pointer in the chain we call parse_blk_func().
318 * For each record in each valid block we call parse_lr_func().
319 * If the log has been claimed, stop if we encounter a sequence
320 * number greater than the highest claimed sequence number.
322 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
323 zil_bp_tree_init(zilog);
325 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
326 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
330 if (blk_seq > claim_blk_seq)
332 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
334 ASSERT3U(max_blk_seq, <, blk_seq);
335 max_blk_seq = blk_seq;
338 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
341 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
345 for (lrp = lrbuf; lrp < end; lrp += reclen) {
346 lr_t *lr = (lr_t *)lrp;
347 reclen = lr->lrc_reclen;
348 ASSERT3U(reclen, >=, sizeof (lr_t));
349 if (lr->lrc_seq > claim_lr_seq)
351 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
353 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
354 max_lr_seq = lr->lrc_seq;
359 zilog->zl_parse_error = error;
360 zilog->zl_parse_blk_seq = max_blk_seq;
361 zilog->zl_parse_lr_seq = max_lr_seq;
362 zilog->zl_parse_blk_count = blk_count;
363 zilog->zl_parse_lr_count = lr_count;
365 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
366 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
368 zil_bp_tree_fini(zilog);
369 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
375 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
378 * Claim log block if not already committed and not already claimed.
379 * If tx == NULL, just verify that the block is claimable.
381 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
384 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
385 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
386 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
390 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
392 lr_write_t *lr = (lr_write_t *)lrc;
395 if (lrc->lrc_txtype != TX_WRITE)
399 * If the block is not readable, don't claim it. This can happen
400 * in normal operation when a log block is written to disk before
401 * some of the dmu_sync() blocks it points to. In this case, the
402 * transaction cannot have been committed to anyone (we would have
403 * waited for all writes to be stable first), so it is semantically
404 * correct to declare this the end of the log.
406 if (lr->lr_blkptr.blk_birth >= first_txg &&
407 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
409 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
414 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
416 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
422 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
424 lr_write_t *lr = (lr_write_t *)lrc;
425 blkptr_t *bp = &lr->lr_blkptr;
428 * If we previously claimed it, we need to free it.
430 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
431 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
432 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
438 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
442 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
443 lwb->lwb_zilog = zilog;
445 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
446 lwb->lwb_max_txg = txg;
449 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
450 lwb->lwb_nused = sizeof (zil_chain_t);
451 lwb->lwb_sz = BP_GET_LSIZE(bp);
454 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
457 mutex_enter(&zilog->zl_lock);
458 list_insert_tail(&zilog->zl_lwb_list, lwb);
459 mutex_exit(&zilog->zl_lock);
465 * Create an on-disk intent log.
468 zil_create(zilog_t *zilog)
470 const zil_header_t *zh = zilog->zl_header;
478 * Wait for any previous destroy to complete.
480 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
482 ASSERT(zh->zh_claim_txg == 0);
483 ASSERT(zh->zh_replay_seq == 0);
488 * Allocate an initial log block if:
489 * - there isn't one already
490 * - the existing block is the wrong endianess
492 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
493 tx = dmu_tx_create(zilog->zl_os);
494 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
495 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
496 txg = dmu_tx_get_txg(tx);
498 if (!BP_IS_HOLE(&blk)) {
499 zio_free_zil(zilog->zl_spa, txg, &blk);
503 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
504 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
507 zil_init_log_chain(zilog, &blk);
511 * Allocate a log write buffer (lwb) for the first log block.
514 lwb = zil_alloc_lwb(zilog, &blk, txg);
517 * If we just allocated the first log block, commit our transaction
518 * and wait for zil_sync() to stuff the block poiner into zh_log.
519 * (zh is part of the MOS, so we cannot modify it in open context.)
523 txg_wait_synced(zilog->zl_dmu_pool, txg);
526 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
532 * In one tx, free all log blocks and clear the log header.
533 * If keep_first is set, then we're replaying a log with no content.
534 * We want to keep the first block, however, so that the first
535 * synchronous transaction doesn't require a txg_wait_synced()
536 * in zil_create(). We don't need to txg_wait_synced() here either
537 * when keep_first is set, because both zil_create() and zil_destroy()
538 * will wait for any in-progress destroys to complete.
541 zil_destroy(zilog_t *zilog, boolean_t keep_first)
543 const zil_header_t *zh = zilog->zl_header;
549 * Wait for any previous destroy to complete.
551 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
553 zilog->zl_old_header = *zh; /* debugging aid */
555 if (BP_IS_HOLE(&zh->zh_log))
558 tx = dmu_tx_create(zilog->zl_os);
559 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
560 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
561 txg = dmu_tx_get_txg(tx);
563 mutex_enter(&zilog->zl_lock);
565 ASSERT3U(zilog->zl_destroy_txg, <, txg);
566 zilog->zl_destroy_txg = txg;
567 zilog->zl_keep_first = keep_first;
569 if (!list_is_empty(&zilog->zl_lwb_list)) {
570 ASSERT(zh->zh_claim_txg == 0);
572 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
573 list_remove(&zilog->zl_lwb_list, lwb);
574 if (lwb->lwb_buf != NULL)
575 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
576 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
577 kmem_cache_free(zil_lwb_cache, lwb);
579 } else if (!keep_first) {
580 (void) zil_parse(zilog, zil_free_log_block,
581 zil_free_log_record, tx, zh->zh_claim_txg);
583 mutex_exit(&zilog->zl_lock);
589 zil_claim(const char *osname, void *txarg)
591 dmu_tx_t *tx = txarg;
592 uint64_t first_txg = dmu_tx_get_txg(tx);
598 error = dmu_objset_hold(osname, FTAG, &os);
600 cmn_err(CE_WARN, "can't open objset for %s", osname);
604 zilog = dmu_objset_zil(os);
605 zh = zil_header_in_syncing_context(zilog);
607 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
608 if (!BP_IS_HOLE(&zh->zh_log))
609 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
610 BP_ZERO(&zh->zh_log);
611 dsl_dataset_dirty(dmu_objset_ds(os), tx);
612 dmu_objset_rele(os, FTAG);
617 * Claim all log blocks if we haven't already done so, and remember
618 * the highest claimed sequence number. This ensures that if we can
619 * read only part of the log now (e.g. due to a missing device),
620 * but we can read the entire log later, we will not try to replay
621 * or destroy beyond the last block we successfully claimed.
623 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
624 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
625 (void) zil_parse(zilog, zil_claim_log_block,
626 zil_claim_log_record, tx, first_txg);
627 zh->zh_claim_txg = first_txg;
628 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
629 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
630 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
631 zh->zh_flags |= ZIL_REPLAY_NEEDED;
632 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
633 dsl_dataset_dirty(dmu_objset_ds(os), tx);
636 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
637 dmu_objset_rele(os, FTAG);
642 * Check the log by walking the log chain.
643 * Checksum errors are ok as they indicate the end of the chain.
644 * Any other error (no device or read failure) returns an error.
647 zil_check_log_chain(const char *osname, void *tx)
656 error = dmu_objset_hold(osname, FTAG, &os);
658 cmn_err(CE_WARN, "can't open objset for %s", osname);
662 zilog = dmu_objset_zil(os);
663 bp = (blkptr_t *)&zilog->zl_header->zh_log;
666 * Check the first block and determine if it's on a log device
667 * which may have been removed or faulted prior to loading this
668 * pool. If so, there's no point in checking the rest of the log
669 * as its content should have already been synced to the pool.
671 if (!BP_IS_HOLE(bp)) {
673 boolean_t valid = B_TRUE;
675 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
676 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
677 if (vd->vdev_islog && vdev_is_dead(vd))
678 valid = vdev_log_state_valid(vd);
679 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
682 dmu_objset_rele(os, FTAG);
688 * Because tx == NULL, zil_claim_log_block() will not actually claim
689 * any blocks, but just determine whether it is possible to do so.
690 * In addition to checking the log chain, zil_claim_log_block()
691 * will invoke zio_claim() with a done func of spa_claim_notify(),
692 * which will update spa_max_claim_txg. See spa_load() for details.
694 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
695 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
697 dmu_objset_rele(os, FTAG);
699 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
703 zil_vdev_compare(const void *x1, const void *x2)
705 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
706 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
717 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
719 avl_tree_t *t = &zilog->zl_vdev_tree;
721 zil_vdev_node_t *zv, zvsearch;
722 int ndvas = BP_GET_NDVAS(bp);
725 if (zfs_nocacheflush)
728 ASSERT(zilog->zl_writer);
731 * Even though we're zl_writer, we still need a lock because the
732 * zl_get_data() callbacks may have dmu_sync() done callbacks
733 * that will run concurrently.
735 mutex_enter(&zilog->zl_vdev_lock);
736 for (i = 0; i < ndvas; i++) {
737 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
738 if (avl_find(t, &zvsearch, &where) == NULL) {
739 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
740 zv->zv_vdev = zvsearch.zv_vdev;
741 avl_insert(t, zv, where);
744 mutex_exit(&zilog->zl_vdev_lock);
748 zil_flush_vdevs(zilog_t *zilog)
750 spa_t *spa = zilog->zl_spa;
751 avl_tree_t *t = &zilog->zl_vdev_tree;
756 ASSERT(zilog->zl_writer);
759 * We don't need zl_vdev_lock here because we're the zl_writer,
760 * and all zl_get_data() callbacks are done.
762 if (avl_numnodes(t) == 0)
765 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
767 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
769 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
770 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
773 kmem_free(zv, sizeof (*zv));
777 * Wait for all the flushes to complete. Not all devices actually
778 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
780 (void) zio_wait(zio);
782 spa_config_exit(spa, SCL_STATE, FTAG);
786 * Function called when a log block write completes
789 zil_lwb_write_done(zio_t *zio)
791 lwb_t *lwb = zio->io_private;
792 zilog_t *zilog = lwb->lwb_zilog;
793 dmu_tx_t *tx = lwb->lwb_tx;
795 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
796 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
797 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
798 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
799 ASSERT(!BP_IS_GANG(zio->io_bp));
800 ASSERT(!BP_IS_HOLE(zio->io_bp));
801 ASSERT(zio->io_bp->blk_fill == 0);
804 * Ensure the lwb buffer pointer is cleared before releasing
805 * the txg. If we have had an allocation failure and
806 * the txg is waiting to sync then we want want zil_sync()
807 * to remove the lwb so that it's not picked up as the next new
808 * one in zil_commit_writer(). zil_sync() will only remove
809 * the lwb if lwb_buf is null.
811 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
812 mutex_enter(&zilog->zl_lock);
815 mutex_exit(&zilog->zl_lock);
818 * Now that we've written this log block, we have a stable pointer
819 * to the next block in the chain, so it's OK to let the txg in
820 * which we allocated the next block sync.
826 * Initialize the io for a log block.
829 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
833 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
834 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
835 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
837 if (zilog->zl_root_zio == NULL) {
838 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
841 if (lwb->lwb_zio == NULL) {
842 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
843 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
844 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
845 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
850 * Define a limited set of intent log block sizes.
851 * These must be a multiple of 4KB. Note only the amount used (again
852 * aligned to 4KB) actually gets written. However, we can't always just
853 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
855 uint64_t zil_block_buckets[] = {
856 4096, /* non TX_WRITE */
857 8192+4096, /* data base */
858 32*1024 + 4096, /* NFS writes */
863 * Use the slog as long as the logbias is 'latency' and the current commit size
864 * is less than the limit or the total list size is less than 2X the limit.
865 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
867 uint64_t zil_slog_limit = 1024 * 1024;
868 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
869 (((zilog)->zl_cur_used < zil_slog_limit) || \
870 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
873 * Start a log block write and advance to the next log block.
874 * Calls are serialized.
877 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
881 spa_t *spa = zilog->zl_spa;
885 uint64_t zil_blksz, wsz;
888 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
889 zilc = (zil_chain_t *)lwb->lwb_buf;
890 bp = &zilc->zc_next_blk;
892 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
893 bp = &zilc->zc_next_blk;
896 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
899 * Allocate the next block and save its address in this block
900 * before writing it in order to establish the log chain.
901 * Note that if the allocation of nlwb synced before we wrote
902 * the block that points at it (lwb), we'd leak it if we crashed.
903 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
904 * We dirty the dataset to ensure that zil_sync() will be called
905 * to clean up in the event of allocation failure or I/O failure.
907 tx = dmu_tx_create(zilog->zl_os);
908 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
909 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
910 txg = dmu_tx_get_txg(tx);
915 * Log blocks are pre-allocated. Here we select the size of the next
916 * block, based on size used in the last block.
917 * - first find the smallest bucket that will fit the block from a
918 * limited set of block sizes. This is because it's faster to write
919 * blocks allocated from the same metaslab as they are adjacent or
921 * - next find the maximum from the new suggested size and an array of
922 * previous sizes. This lessens a picket fence effect of wrongly
923 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
926 * Note we only write what is used, but we can't just allocate
927 * the maximum block size because we can exhaust the available
930 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
931 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
933 zil_blksz = zil_block_buckets[i];
934 if (zil_blksz == UINT64_MAX)
935 zil_blksz = SPA_MAXBLOCKSIZE;
936 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
937 for (i = 0; i < ZIL_PREV_BLKS; i++)
938 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
939 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
942 /* pass the old blkptr in order to spread log blocks across devs */
943 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
946 ASSERT3U(bp->blk_birth, ==, txg);
947 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
948 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
951 * Allocate a new log write buffer (lwb).
953 nlwb = zil_alloc_lwb(zilog, bp, txg);
955 /* Record the block for later vdev flushing */
956 zil_add_block(zilog, &lwb->lwb_blk);
959 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
960 /* For Slim ZIL only write what is used. */
961 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
962 ASSERT3U(wsz, <=, lwb->lwb_sz);
963 zio_shrink(lwb->lwb_zio, wsz);
970 zilc->zc_nused = lwb->lwb_nused;
971 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
974 * clear unused data for security
976 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
978 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
981 * If there was an allocation failure then nlwb will be null which
982 * forces a txg_wait_synced().
988 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
990 lr_t *lrc = &itx->itx_lr; /* common log record */
991 lr_write_t *lrw = (lr_write_t *)lrc;
993 uint64_t txg = lrc->lrc_txg;
994 uint64_t reclen = lrc->lrc_reclen;
1000 ASSERT(lwb->lwb_buf != NULL);
1002 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1003 dlen = P2ROUNDUP_TYPED(
1004 lrw->lr_length, sizeof (uint64_t), uint64_t);
1006 zilog->zl_cur_used += (reclen + dlen);
1008 zil_lwb_write_init(zilog, lwb);
1011 * If this record won't fit in the current log block, start a new one.
1013 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1014 lwb = zil_lwb_write_start(zilog, lwb);
1017 zil_lwb_write_init(zilog, lwb);
1018 ASSERT(LWB_EMPTY(lwb));
1019 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1020 txg_wait_synced(zilog->zl_dmu_pool, txg);
1025 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1026 bcopy(lrc, lr_buf, reclen);
1027 lrc = (lr_t *)lr_buf;
1028 lrw = (lr_write_t *)lrc;
1031 * If it's a write, fetch the data or get its blkptr as appropriate.
1033 if (lrc->lrc_txtype == TX_WRITE) {
1034 if (txg > spa_freeze_txg(zilog->zl_spa))
1035 txg_wait_synced(zilog->zl_dmu_pool, txg);
1036 if (itx->itx_wr_state != WR_COPIED) {
1041 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1042 dbuf = lr_buf + reclen;
1043 lrw->lr_common.lrc_reclen += dlen;
1045 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1048 error = zilog->zl_get_data(
1049 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1051 txg_wait_synced(zilog->zl_dmu_pool, txg);
1055 ASSERT(error == ENOENT || error == EEXIST ||
1063 * We're actually making an entry, so update lrc_seq to be the
1064 * log record sequence number. Note that this is generally not
1065 * equal to the itx sequence number because not all transactions
1066 * are synchronous, and sometimes spa_sync() gets there first.
1068 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1069 lwb->lwb_nused += reclen + dlen;
1070 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1071 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1072 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1078 zil_itx_create(uint64_t txtype, size_t lrsize)
1082 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1084 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1085 itx->itx_lr.lrc_txtype = txtype;
1086 itx->itx_lr.lrc_reclen = lrsize;
1087 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1088 itx->itx_lr.lrc_seq = 0; /* defensive */
1089 itx->itx_sync = B_TRUE; /* default is synchronous */
1095 zil_itx_destroy(itx_t *itx)
1097 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1101 * Free up the sync and async itxs. The itxs_t has already been detached
1102 * so no locks are needed.
1105 zil_itxg_clean(itxs_t *itxs)
1111 itx_async_node_t *ian;
1113 list = &itxs->i_sync_list;
1114 while ((itx = list_head(list)) != NULL) {
1115 list_remove(list, itx);
1116 kmem_free(itx, offsetof(itx_t, itx_lr) +
1117 itx->itx_lr.lrc_reclen);
1121 t = &itxs->i_async_tree;
1122 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1123 list = &ian->ia_list;
1124 while ((itx = list_head(list)) != NULL) {
1125 list_remove(list, itx);
1126 kmem_free(itx, offsetof(itx_t, itx_lr) +
1127 itx->itx_lr.lrc_reclen);
1130 kmem_free(ian, sizeof (itx_async_node_t));
1134 kmem_free(itxs, sizeof (itxs_t));
1138 zil_aitx_compare(const void *x1, const void *x2)
1140 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1141 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1152 * Remove all async itx with the given oid.
1155 zil_remove_async(zilog_t *zilog, uint64_t oid)
1158 itx_async_node_t *ian;
1165 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1167 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1170 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1172 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1173 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1175 mutex_enter(&itxg->itxg_lock);
1176 if (itxg->itxg_txg != txg) {
1177 mutex_exit(&itxg->itxg_lock);
1182 * Locate the object node and append its list.
1184 t = &itxg->itxg_itxs->i_async_tree;
1185 ian = avl_find(t, &oid, &where);
1187 list_move_tail(&clean_list, &ian->ia_list);
1188 mutex_exit(&itxg->itxg_lock);
1190 while ((itx = list_head(&clean_list)) != NULL) {
1191 list_remove(&clean_list, itx);
1192 kmem_free(itx, offsetof(itx_t, itx_lr) +
1193 itx->itx_lr.lrc_reclen);
1195 list_destroy(&clean_list);
1199 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1203 itxs_t *itxs, *clean = NULL;
1206 * Object ids can be re-instantiated in the next txg so
1207 * remove any async transactions to avoid future leaks.
1208 * This can happen if a fsync occurs on the re-instantiated
1209 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1210 * the new file data and flushes a write record for the old object.
1212 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1213 zil_remove_async(zilog, itx->itx_oid);
1216 * Ensure the data of a renamed file is committed before the rename.
1218 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1219 zil_async_to_sync(zilog, itx->itx_oid);
1221 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1224 txg = dmu_tx_get_txg(tx);
1226 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1227 mutex_enter(&itxg->itxg_lock);
1228 itxs = itxg->itxg_itxs;
1229 if (itxg->itxg_txg != txg) {
1232 * The zil_clean callback hasn't got around to cleaning
1233 * this itxg. Save the itxs for release below.
1234 * This should be rare.
1236 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1238 clean = itxg->itxg_itxs;
1240 ASSERT(itxg->itxg_sod == 0);
1241 itxg->itxg_txg = txg;
1242 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1244 list_create(&itxs->i_sync_list, sizeof (itx_t),
1245 offsetof(itx_t, itx_node));
1246 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1247 sizeof (itx_async_node_t),
1248 offsetof(itx_async_node_t, ia_node));
1250 if (itx->itx_sync) {
1251 list_insert_tail(&itxs->i_sync_list, itx);
1252 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1253 itxg->itxg_sod += itx->itx_sod;
1255 avl_tree_t *t = &itxs->i_async_tree;
1256 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1257 itx_async_node_t *ian;
1260 ian = avl_find(t, &foid, &where);
1262 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1263 list_create(&ian->ia_list, sizeof (itx_t),
1264 offsetof(itx_t, itx_node));
1265 ian->ia_foid = foid;
1266 avl_insert(t, ian, where);
1268 list_insert_tail(&ian->ia_list, itx);
1271 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1272 mutex_exit(&itxg->itxg_lock);
1274 /* Release the old itxs now we've dropped the lock */
1276 zil_itxg_clean(clean);
1280 * If there are any in-memory intent log transactions which have now been
1281 * synced then start up a taskq to free them.
1284 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1286 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1289 mutex_enter(&itxg->itxg_lock);
1290 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1291 mutex_exit(&itxg->itxg_lock);
1294 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1295 ASSERT(itxg->itxg_txg != 0);
1296 ASSERT(zilog->zl_clean_taskq != NULL);
1297 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1299 clean_me = itxg->itxg_itxs;
1300 itxg->itxg_itxs = NULL;
1302 mutex_exit(&itxg->itxg_lock);
1304 * Preferably start a task queue to free up the old itxs but
1305 * if taskq_dispatch can't allocate resources to do that then
1306 * free it in-line. This should be rare. Note, using TQ_SLEEP
1307 * created a bad performance problem.
1309 if (taskq_dispatch(zilog->zl_clean_taskq,
1310 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1311 zil_itxg_clean(clean_me);
1315 * Get the list of itxs to commit into zl_itx_commit_list.
1318 zil_get_commit_list(zilog_t *zilog)
1321 list_t *commit_list = &zilog->zl_itx_commit_list;
1322 uint64_t push_sod = 0;
1324 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1327 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1329 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1330 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1332 mutex_enter(&itxg->itxg_lock);
1333 if (itxg->itxg_txg != txg) {
1334 mutex_exit(&itxg->itxg_lock);
1338 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1339 push_sod += itxg->itxg_sod;
1342 mutex_exit(&itxg->itxg_lock);
1344 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1348 * Move the async itxs for a specified object to commit into sync lists.
1351 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1354 itx_async_node_t *ian;
1358 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1361 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1363 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1364 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1366 mutex_enter(&itxg->itxg_lock);
1367 if (itxg->itxg_txg != txg) {
1368 mutex_exit(&itxg->itxg_lock);
1373 * If a foid is specified then find that node and append its
1374 * list. Otherwise walk the tree appending all the lists
1375 * to the sync list. We add to the end rather than the
1376 * beginning to ensure the create has happened.
1378 t = &itxg->itxg_itxs->i_async_tree;
1380 ian = avl_find(t, &foid, &where);
1382 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1386 void *cookie = NULL;
1388 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1389 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1391 list_destroy(&ian->ia_list);
1392 kmem_free(ian, sizeof (itx_async_node_t));
1395 mutex_exit(&itxg->itxg_lock);
1400 zil_commit_writer(zilog_t *zilog)
1405 spa_t *spa = zilog->zl_spa;
1408 ASSERT(zilog->zl_root_zio == NULL);
1410 mutex_exit(&zilog->zl_lock);
1412 zil_get_commit_list(zilog);
1415 * Return if there's nothing to commit before we dirty the fs by
1416 * calling zil_create().
1418 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1419 mutex_enter(&zilog->zl_lock);
1423 if (zilog->zl_suspend) {
1426 lwb = list_tail(&zilog->zl_lwb_list);
1428 lwb = zil_create(zilog);
1431 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1432 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1433 txg = itx->itx_lr.lrc_txg;
1436 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1437 lwb = zil_lwb_commit(zilog, itx, lwb);
1438 list_remove(&zilog->zl_itx_commit_list, itx);
1439 kmem_free(itx, offsetof(itx_t, itx_lr)
1440 + itx->itx_lr.lrc_reclen);
1442 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1444 /* write the last block out */
1445 if (lwb != NULL && lwb->lwb_zio != NULL)
1446 lwb = zil_lwb_write_start(zilog, lwb);
1448 zilog->zl_cur_used = 0;
1451 * Wait if necessary for the log blocks to be on stable storage.
1453 if (zilog->zl_root_zio) {
1454 error = zio_wait(zilog->zl_root_zio);
1455 zilog->zl_root_zio = NULL;
1456 zil_flush_vdevs(zilog);
1459 if (error || lwb == NULL)
1460 txg_wait_synced(zilog->zl_dmu_pool, 0);
1462 mutex_enter(&zilog->zl_lock);
1465 * Remember the highest committed log sequence number for ztest.
1466 * We only update this value when all the log writes succeeded,
1467 * because ztest wants to ASSERT that it got the whole log chain.
1469 if (error == 0 && lwb != NULL)
1470 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1474 * Commit zfs transactions to stable storage.
1475 * If foid is 0 push out all transactions, otherwise push only those
1476 * for that object or might reference that object.
1478 * itxs are committed in batches. In a heavily stressed zil there will be
1479 * a commit writer thread who is writing out a bunch of itxs to the log
1480 * for a set of committing threads (cthreads) in the same batch as the writer.
1481 * Those cthreads are all waiting on the same cv for that batch.
1483 * There will also be a different and growing batch of threads that are
1484 * waiting to commit (qthreads). When the committing batch completes
1485 * a transition occurs such that the cthreads exit and the qthreads become
1486 * cthreads. One of the new cthreads becomes the writer thread for the
1487 * batch. Any new threads arriving become new qthreads.
1489 * Only 2 condition variables are needed and there's no transition
1490 * between the two cvs needed. They just flip-flop between qthreads
1493 * Using this scheme we can efficiently wakeup up only those threads
1494 * that have been committed.
1497 zil_commit(zilog_t *zilog, uint64_t foid)
1501 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1504 /* move the async itxs for the foid to the sync queues */
1505 zil_async_to_sync(zilog, foid);
1507 mutex_enter(&zilog->zl_lock);
1508 mybatch = zilog->zl_next_batch;
1509 while (zilog->zl_writer) {
1510 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1511 if (mybatch <= zilog->zl_com_batch) {
1512 mutex_exit(&zilog->zl_lock);
1517 zilog->zl_next_batch++;
1518 zilog->zl_writer = B_TRUE;
1519 zil_commit_writer(zilog);
1520 zilog->zl_com_batch = mybatch;
1521 zilog->zl_writer = B_FALSE;
1522 mutex_exit(&zilog->zl_lock);
1524 /* wake up one thread to become the next writer */
1525 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1527 /* wake up all threads waiting for this batch to be committed */
1528 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1532 * Called in syncing context to free committed log blocks and update log header.
1535 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1537 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1538 uint64_t txg = dmu_tx_get_txg(tx);
1539 spa_t *spa = zilog->zl_spa;
1540 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1544 * We don't zero out zl_destroy_txg, so make sure we don't try
1545 * to destroy it twice.
1547 if (spa_sync_pass(spa) != 1)
1550 mutex_enter(&zilog->zl_lock);
1552 ASSERT(zilog->zl_stop_sync == 0);
1554 if (*replayed_seq != 0) {
1555 ASSERT(zh->zh_replay_seq < *replayed_seq);
1556 zh->zh_replay_seq = *replayed_seq;
1560 if (zilog->zl_destroy_txg == txg) {
1561 blkptr_t blk = zh->zh_log;
1563 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1565 bzero(zh, sizeof (zil_header_t));
1566 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1568 if (zilog->zl_keep_first) {
1570 * If this block was part of log chain that couldn't
1571 * be claimed because a device was missing during
1572 * zil_claim(), but that device later returns,
1573 * then this block could erroneously appear valid.
1574 * To guard against this, assign a new GUID to the new
1575 * log chain so it doesn't matter what blk points to.
1577 zil_init_log_chain(zilog, &blk);
1582 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1583 zh->zh_log = lwb->lwb_blk;
1584 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1586 list_remove(&zilog->zl_lwb_list, lwb);
1587 zio_free_zil(spa, txg, &lwb->lwb_blk);
1588 kmem_cache_free(zil_lwb_cache, lwb);
1591 * If we don't have anything left in the lwb list then
1592 * we've had an allocation failure and we need to zero
1593 * out the zil_header blkptr so that we don't end
1594 * up freeing the same block twice.
1596 if (list_head(&zilog->zl_lwb_list) == NULL)
1597 BP_ZERO(&zh->zh_log);
1599 mutex_exit(&zilog->zl_lock);
1605 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1606 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1612 kmem_cache_destroy(zil_lwb_cache);
1616 zil_set_sync(zilog_t *zilog, uint64_t sync)
1618 zilog->zl_sync = sync;
1622 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1624 zilog->zl_logbias = logbias;
1628 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1632 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1634 zilog->zl_header = zh_phys;
1636 zilog->zl_spa = dmu_objset_spa(os);
1637 zilog->zl_dmu_pool = dmu_objset_pool(os);
1638 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1639 zilog->zl_logbias = dmu_objset_logbias(os);
1640 zilog->zl_sync = dmu_objset_syncprop(os);
1641 zilog->zl_next_batch = 1;
1643 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1645 for (int i = 0; i < TXG_SIZE; i++) {
1646 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1647 MUTEX_DEFAULT, NULL);
1650 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1651 offsetof(lwb_t, lwb_node));
1653 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1654 offsetof(itx_t, itx_node));
1656 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1658 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1659 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1661 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1662 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1663 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1664 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1670 zil_free(zilog_t *zilog)
1672 zilog->zl_stop_sync = 1;
1674 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1675 list_destroy(&zilog->zl_lwb_list);
1677 avl_destroy(&zilog->zl_vdev_tree);
1678 mutex_destroy(&zilog->zl_vdev_lock);
1680 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1681 list_destroy(&zilog->zl_itx_commit_list);
1683 for (int i = 0; i < TXG_SIZE; i++) {
1685 * It's possible for an itx to be generated that doesn't dirty
1686 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1687 * callback to remove the entry. We remove those here.
1689 * Also free up the ziltest itxs.
1691 if (zilog->zl_itxg[i].itxg_itxs)
1692 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1693 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1696 mutex_destroy(&zilog->zl_lock);
1698 cv_destroy(&zilog->zl_cv_writer);
1699 cv_destroy(&zilog->zl_cv_suspend);
1700 cv_destroy(&zilog->zl_cv_batch[0]);
1701 cv_destroy(&zilog->zl_cv_batch[1]);
1703 kmem_free(zilog, sizeof (zilog_t));
1707 * Open an intent log.
1710 zil_open(objset_t *os, zil_get_data_t *get_data)
1712 zilog_t *zilog = dmu_objset_zil(os);
1714 ASSERT(zilog->zl_clean_taskq == NULL);
1715 ASSERT(zilog->zl_get_data == NULL);
1716 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1718 zilog->zl_get_data = get_data;
1719 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1720 2, 2, TASKQ_PREPOPULATE);
1726 * Close an intent log.
1729 zil_close(zilog_t *zilog)
1734 zil_commit(zilog, 0); /* commit all itx */
1737 * The lwb_max_txg for the stubby lwb will reflect the last activity
1738 * for the zil. After a txg_wait_synced() on the txg we know all the
1739 * callbacks have occurred that may clean the zil. Only then can we
1740 * destroy the zl_clean_taskq.
1742 mutex_enter(&zilog->zl_lock);
1743 lwb = list_tail(&zilog->zl_lwb_list);
1745 txg = lwb->lwb_max_txg;
1746 mutex_exit(&zilog->zl_lock);
1748 txg_wait_synced(zilog->zl_dmu_pool, txg);
1750 taskq_destroy(zilog->zl_clean_taskq);
1751 zilog->zl_clean_taskq = NULL;
1752 zilog->zl_get_data = NULL;
1755 * We should have only one LWB left on the list; remove it now.
1757 mutex_enter(&zilog->zl_lock);
1758 lwb = list_head(&zilog->zl_lwb_list);
1760 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1761 list_remove(&zilog->zl_lwb_list, lwb);
1762 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1763 kmem_cache_free(zil_lwb_cache, lwb);
1765 mutex_exit(&zilog->zl_lock);
1769 * Suspend an intent log. While in suspended mode, we still honor
1770 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1771 * We suspend the log briefly when taking a snapshot so that the snapshot
1772 * contains all the data it's supposed to, and has an empty intent log.
1775 zil_suspend(zilog_t *zilog)
1777 const zil_header_t *zh = zilog->zl_header;
1779 mutex_enter(&zilog->zl_lock);
1780 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1781 mutex_exit(&zilog->zl_lock);
1784 if (zilog->zl_suspend++ != 0) {
1786 * Someone else already began a suspend.
1787 * Just wait for them to finish.
1789 while (zilog->zl_suspending)
1790 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1791 mutex_exit(&zilog->zl_lock);
1794 zilog->zl_suspending = B_TRUE;
1795 mutex_exit(&zilog->zl_lock);
1797 zil_commit(zilog, 0);
1799 zil_destroy(zilog, B_FALSE);
1801 mutex_enter(&zilog->zl_lock);
1802 zilog->zl_suspending = B_FALSE;
1803 cv_broadcast(&zilog->zl_cv_suspend);
1804 mutex_exit(&zilog->zl_lock);
1810 zil_resume(zilog_t *zilog)
1812 mutex_enter(&zilog->zl_lock);
1813 ASSERT(zilog->zl_suspend != 0);
1814 zilog->zl_suspend--;
1815 mutex_exit(&zilog->zl_lock);
1818 typedef struct zil_replay_arg {
1819 zil_replay_func_t **zr_replay;
1821 boolean_t zr_byteswap;
1826 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1828 char name[MAXNAMELEN];
1830 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1832 dmu_objset_name(zilog->zl_os, name);
1834 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1835 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1836 (u_longlong_t)lr->lrc_seq,
1837 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1838 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1844 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1846 zil_replay_arg_t *zr = zra;
1847 const zil_header_t *zh = zilog->zl_header;
1848 uint64_t reclen = lr->lrc_reclen;
1849 uint64_t txtype = lr->lrc_txtype;
1852 zilog->zl_replaying_seq = lr->lrc_seq;
1854 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1857 if (lr->lrc_txg < claim_txg) /* already committed */
1860 /* Strip case-insensitive bit, still present in log record */
1863 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1864 return (zil_replay_error(zilog, lr, EINVAL));
1867 * If this record type can be logged out of order, the object
1868 * (lr_foid) may no longer exist. That's legitimate, not an error.
1870 if (TX_OOO(txtype)) {
1871 error = dmu_object_info(zilog->zl_os,
1872 ((lr_ooo_t *)lr)->lr_foid, NULL);
1873 if (error == ENOENT || error == EEXIST)
1878 * Make a copy of the data so we can revise and extend it.
1880 bcopy(lr, zr->zr_lr, reclen);
1883 * If this is a TX_WRITE with a blkptr, suck in the data.
1885 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1886 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1887 zr->zr_lr + reclen);
1889 return (zil_replay_error(zilog, lr, error));
1893 * The log block containing this lr may have been byteswapped
1894 * so that we can easily examine common fields like lrc_txtype.
1895 * However, the log is a mix of different record types, and only the
1896 * replay vectors know how to byteswap their records. Therefore, if
1897 * the lr was byteswapped, undo it before invoking the replay vector.
1899 if (zr->zr_byteswap)
1900 byteswap_uint64_array(zr->zr_lr, reclen);
1903 * We must now do two things atomically: replay this log record,
1904 * and update the log header sequence number to reflect the fact that
1905 * we did so. At the end of each replay function the sequence number
1906 * is updated if we are in replay mode.
1908 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1911 * The DMU's dnode layer doesn't see removes until the txg
1912 * commits, so a subsequent claim can spuriously fail with
1913 * EEXIST. So if we receive any error we try syncing out
1914 * any removes then retry the transaction. Note that we
1915 * specify B_FALSE for byteswap now, so we don't do it twice.
1917 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1918 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1920 return (zil_replay_error(zilog, lr, error));
1927 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1929 zilog->zl_replay_blks++;
1935 * If this dataset has a non-empty intent log, replay it and destroy it.
1938 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1940 zilog_t *zilog = dmu_objset_zil(os);
1941 const zil_header_t *zh = zilog->zl_header;
1942 zil_replay_arg_t zr;
1944 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1945 zil_destroy(zilog, B_TRUE);
1948 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
1950 zr.zr_replay = replay_func;
1952 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1953 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1956 * Wait for in-progress removes to sync before starting replay.
1958 txg_wait_synced(zilog->zl_dmu_pool, 0);
1960 zilog->zl_replay = B_TRUE;
1961 zilog->zl_replay_time = ddi_get_lbolt();
1962 ASSERT(zilog->zl_replay_blks == 0);
1963 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1965 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1967 zil_destroy(zilog, B_FALSE);
1968 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1969 zilog->zl_replay = B_FALSE;
1970 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
1974 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1976 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1979 if (zilog->zl_replay) {
1980 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1981 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1982 zilog->zl_replaying_seq;
1991 zil_vdev_offline(const char *osname, void *arg)
1997 error = dmu_objset_hold(osname, FTAG, &os);
2001 zilog = dmu_objset_zil(os);
2002 if (zil_suspend(zilog) != 0)
2006 dmu_objset_rele(os, FTAG);