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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 /* Portions Copyright 2010 Robert Milkowski */
30 #include <sys/zfs_context.h>
36 #include <sys/resource.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
45 * The zfs intent log (ZIL) saves transaction records of system calls
46 * that change the file system in memory with enough information
47 * to be able to replay them. These are stored in memory until
48 * either the DMU transaction group (txg) commits them to the stable pool
49 * and they can be discarded, or they are flushed to the stable log
50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
51 * requirement. In the event of a panic or power fail then those log
52 * records (transactions) are replayed.
54 * There is one ZIL per file system. Its on-disk (pool) format consists
61 * A log record holds a system call transaction. Log blocks can
62 * hold many log records and the blocks are chained together.
63 * Each ZIL block contains a block pointer (blkptr_t) to the next
64 * ZIL block in the chain. The ZIL header points to the first
65 * block in the chain. Note there is not a fixed place in the pool
66 * to hold blocks. They are dynamically allocated and freed as
67 * needed from the blocks available. Figure X shows the ZIL structure:
71 * Disable intent logging replay. This global ZIL switch affects all pools.
73 int zil_replay_disable = 0;
74 SYSCTL_DECL(_vfs_zfs);
75 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
76 &zil_replay_disable, 0, "Disable intent logging replay");
79 * Tunable parameter for debugging or performance analysis. Setting
80 * zfs_nocacheflush will cause corruption on power loss if a volatile
81 * out-of-order write cache is enabled.
83 boolean_t zfs_nocacheflush = B_FALSE;
84 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
85 &zfs_nocacheflush, 0, "Disable cache flush");
86 boolean_t zfs_trim_enabled = B_TRUE;
87 SYSCTL_DECL(_vfs_zfs_trim);
88 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
91 static kmem_cache_t *zil_lwb_cache;
93 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
95 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
96 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
100 * ziltest is by and large an ugly hack, but very useful in
101 * checking replay without tedious work.
102 * When running ziltest we want to keep all itx's and so maintain
103 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
104 * We subtract TXG_CONCURRENT_STATES to allow for common code.
106 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
109 zil_bp_compare(const void *x1, const void *x2)
111 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
112 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
114 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
116 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
119 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
121 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
128 zil_bp_tree_init(zilog_t *zilog)
130 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
131 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
135 zil_bp_tree_fini(zilog_t *zilog)
137 avl_tree_t *t = &zilog->zl_bp_tree;
141 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
142 kmem_free(zn, sizeof (zil_bp_node_t));
148 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
150 avl_tree_t *t = &zilog->zl_bp_tree;
155 if (BP_IS_EMBEDDED(bp))
158 dva = BP_IDENTITY(bp);
160 if (avl_find(t, dva, &where) != NULL)
161 return (SET_ERROR(EEXIST));
163 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
165 avl_insert(t, zn, where);
170 static zil_header_t *
171 zil_header_in_syncing_context(zilog_t *zilog)
173 return ((zil_header_t *)zilog->zl_header);
177 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
179 zio_cksum_t *zc = &bp->blk_cksum;
181 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
182 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
183 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
184 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
188 * Read a log block and make sure it's valid.
191 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
194 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
195 arc_flags_t aflags = ARC_FLAG_WAIT;
196 arc_buf_t *abuf = NULL;
200 if (zilog->zl_header->zh_claim_txg == 0)
201 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
203 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
204 zio_flags |= ZIO_FLAG_SPECULATIVE;
206 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
207 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
209 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
210 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
213 zio_cksum_t cksum = bp->blk_cksum;
216 * Validate the checksummed log block.
218 * Sequence numbers should be... sequential. The checksum
219 * verifier for the next block should be bp's checksum plus 1.
221 * Also check the log chain linkage and size used.
223 cksum.zc_word[ZIL_ZC_SEQ]++;
225 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
226 zil_chain_t *zilc = abuf->b_data;
227 char *lr = (char *)(zilc + 1);
228 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
230 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
231 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
232 error = SET_ERROR(ECKSUM);
234 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
236 *end = (char *)dst + len;
237 *nbp = zilc->zc_next_blk;
240 char *lr = abuf->b_data;
241 uint64_t size = BP_GET_LSIZE(bp);
242 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
244 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
245 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
246 (zilc->zc_nused > (size - sizeof (*zilc)))) {
247 error = SET_ERROR(ECKSUM);
249 ASSERT3U(zilc->zc_nused, <=,
250 SPA_OLD_MAXBLOCKSIZE);
251 bcopy(lr, dst, zilc->zc_nused);
252 *end = (char *)dst + zilc->zc_nused;
253 *nbp = zilc->zc_next_blk;
257 arc_buf_destroy(abuf, &abuf);
264 * Read a TX_WRITE log data block.
267 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
269 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
270 const blkptr_t *bp = &lr->lr_blkptr;
271 arc_flags_t aflags = ARC_FLAG_WAIT;
272 arc_buf_t *abuf = NULL;
276 if (BP_IS_HOLE(bp)) {
278 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
282 if (zilog->zl_header->zh_claim_txg == 0)
283 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
285 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
286 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
288 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
289 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
293 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
294 arc_buf_destroy(abuf, &abuf);
301 * Parse the intent log, and call parse_func for each valid record within.
304 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
305 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
307 const zil_header_t *zh = zilog->zl_header;
308 boolean_t claimed = !!zh->zh_claim_txg;
309 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
310 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
311 uint64_t max_blk_seq = 0;
312 uint64_t max_lr_seq = 0;
313 uint64_t blk_count = 0;
314 uint64_t lr_count = 0;
315 blkptr_t blk, next_blk;
320 * Old logs didn't record the maximum zh_claim_lr_seq.
322 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
323 claim_lr_seq = UINT64_MAX;
326 * Starting at the block pointed to by zh_log we read the log chain.
327 * For each block in the chain we strongly check that block to
328 * ensure its validity. We stop when an invalid block is found.
329 * For each block pointer in the chain we call parse_blk_func().
330 * For each record in each valid block we call parse_lr_func().
331 * If the log has been claimed, stop if we encounter a sequence
332 * number greater than the highest claimed sequence number.
334 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
335 zil_bp_tree_init(zilog);
337 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
338 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
342 if (blk_seq > claim_blk_seq)
344 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
346 ASSERT3U(max_blk_seq, <, blk_seq);
347 max_blk_seq = blk_seq;
350 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
353 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
357 for (lrp = lrbuf; lrp < end; lrp += reclen) {
358 lr_t *lr = (lr_t *)lrp;
359 reclen = lr->lrc_reclen;
360 ASSERT3U(reclen, >=, sizeof (lr_t));
361 if (lr->lrc_seq > claim_lr_seq)
363 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
365 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
366 max_lr_seq = lr->lrc_seq;
371 zilog->zl_parse_error = error;
372 zilog->zl_parse_blk_seq = max_blk_seq;
373 zilog->zl_parse_lr_seq = max_lr_seq;
374 zilog->zl_parse_blk_count = blk_count;
375 zilog->zl_parse_lr_count = lr_count;
377 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
378 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
380 zil_bp_tree_fini(zilog);
381 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
387 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
390 * Claim log block if not already committed and not already claimed.
391 * If tx == NULL, just verify that the block is claimable.
393 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
394 zil_bp_tree_add(zilog, bp) != 0)
397 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
398 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
399 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
403 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
405 lr_write_t *lr = (lr_write_t *)lrc;
408 if (lrc->lrc_txtype != TX_WRITE)
412 * If the block is not readable, don't claim it. This can happen
413 * in normal operation when a log block is written to disk before
414 * some of the dmu_sync() blocks it points to. In this case, the
415 * transaction cannot have been committed to anyone (we would have
416 * waited for all writes to be stable first), so it is semantically
417 * correct to declare this the end of the log.
419 if (lr->lr_blkptr.blk_birth >= first_txg &&
420 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
422 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
427 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
429 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
435 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
437 lr_write_t *lr = (lr_write_t *)lrc;
438 blkptr_t *bp = &lr->lr_blkptr;
441 * If we previously claimed it, we need to free it.
443 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
444 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
446 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
452 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
456 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
457 lwb->lwb_zilog = zilog;
459 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
460 lwb->lwb_max_txg = txg;
463 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
464 lwb->lwb_nused = sizeof (zil_chain_t);
465 lwb->lwb_sz = BP_GET_LSIZE(bp);
468 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
471 mutex_enter(&zilog->zl_lock);
472 list_insert_tail(&zilog->zl_lwb_list, lwb);
473 mutex_exit(&zilog->zl_lock);
479 * Called when we create in-memory log transactions so that we know
480 * to cleanup the itxs at the end of spa_sync().
483 zilog_dirty(zilog_t *zilog, uint64_t txg)
485 dsl_pool_t *dp = zilog->zl_dmu_pool;
486 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
488 if (ds->ds_is_snapshot)
489 panic("dirtying snapshot!");
491 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
492 /* up the hold count until we can be written out */
493 dmu_buf_add_ref(ds->ds_dbuf, zilog);
498 zilog_is_dirty(zilog_t *zilog)
500 dsl_pool_t *dp = zilog->zl_dmu_pool;
502 for (int t = 0; t < TXG_SIZE; t++) {
503 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
510 * Create an on-disk intent log.
513 zil_create(zilog_t *zilog)
515 const zil_header_t *zh = zilog->zl_header;
523 * Wait for any previous destroy to complete.
525 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
527 ASSERT(zh->zh_claim_txg == 0);
528 ASSERT(zh->zh_replay_seq == 0);
533 * Allocate an initial log block if:
534 * - there isn't one already
535 * - the existing block is the wrong endianess
537 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
538 tx = dmu_tx_create(zilog->zl_os);
539 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
540 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
541 txg = dmu_tx_get_txg(tx);
543 if (!BP_IS_HOLE(&blk)) {
544 zio_free_zil(zilog->zl_spa, txg, &blk);
548 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
549 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
552 zil_init_log_chain(zilog, &blk);
556 * Allocate a log write buffer (lwb) for the first log block.
559 lwb = zil_alloc_lwb(zilog, &blk, txg);
562 * If we just allocated the first log block, commit our transaction
563 * and wait for zil_sync() to stuff the block poiner into zh_log.
564 * (zh is part of the MOS, so we cannot modify it in open context.)
568 txg_wait_synced(zilog->zl_dmu_pool, txg);
571 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
577 * In one tx, free all log blocks and clear the log header.
578 * If keep_first is set, then we're replaying a log with no content.
579 * We want to keep the first block, however, so that the first
580 * synchronous transaction doesn't require a txg_wait_synced()
581 * in zil_create(). We don't need to txg_wait_synced() here either
582 * when keep_first is set, because both zil_create() and zil_destroy()
583 * will wait for any in-progress destroys to complete.
586 zil_destroy(zilog_t *zilog, boolean_t keep_first)
588 const zil_header_t *zh = zilog->zl_header;
594 * Wait for any previous destroy to complete.
596 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
598 zilog->zl_old_header = *zh; /* debugging aid */
600 if (BP_IS_HOLE(&zh->zh_log))
603 tx = dmu_tx_create(zilog->zl_os);
604 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
605 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
606 txg = dmu_tx_get_txg(tx);
608 mutex_enter(&zilog->zl_lock);
610 ASSERT3U(zilog->zl_destroy_txg, <, txg);
611 zilog->zl_destroy_txg = txg;
612 zilog->zl_keep_first = keep_first;
614 if (!list_is_empty(&zilog->zl_lwb_list)) {
615 ASSERT(zh->zh_claim_txg == 0);
617 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
618 list_remove(&zilog->zl_lwb_list, lwb);
619 if (lwb->lwb_buf != NULL)
620 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
621 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
622 kmem_cache_free(zil_lwb_cache, lwb);
624 } else if (!keep_first) {
625 zil_destroy_sync(zilog, tx);
627 mutex_exit(&zilog->zl_lock);
633 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
635 ASSERT(list_is_empty(&zilog->zl_lwb_list));
636 (void) zil_parse(zilog, zil_free_log_block,
637 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
641 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
643 dmu_tx_t *tx = txarg;
644 uint64_t first_txg = dmu_tx_get_txg(tx);
650 error = dmu_objset_own_obj(dp, ds->ds_object,
651 DMU_OST_ANY, B_FALSE, FTAG, &os);
654 * EBUSY indicates that the objset is inconsistent, in which
655 * case it can not have a ZIL.
657 if (error != EBUSY) {
658 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
659 (unsigned long long)ds->ds_object, error);
664 zilog = dmu_objset_zil(os);
665 zh = zil_header_in_syncing_context(zilog);
667 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
668 if (!BP_IS_HOLE(&zh->zh_log))
669 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
670 BP_ZERO(&zh->zh_log);
671 dsl_dataset_dirty(dmu_objset_ds(os), tx);
672 dmu_objset_disown(os, FTAG);
677 * Claim all log blocks if we haven't already done so, and remember
678 * the highest claimed sequence number. This ensures that if we can
679 * read only part of the log now (e.g. due to a missing device),
680 * but we can read the entire log later, we will not try to replay
681 * or destroy beyond the last block we successfully claimed.
683 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
684 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
685 (void) zil_parse(zilog, zil_claim_log_block,
686 zil_claim_log_record, tx, first_txg);
687 zh->zh_claim_txg = first_txg;
688 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
689 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
690 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
691 zh->zh_flags |= ZIL_REPLAY_NEEDED;
692 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
693 dsl_dataset_dirty(dmu_objset_ds(os), tx);
696 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
697 dmu_objset_disown(os, FTAG);
702 * Check the log by walking the log chain.
703 * Checksum errors are ok as they indicate the end of the chain.
704 * Any other error (no device or read failure) returns an error.
708 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
717 error = dmu_objset_from_ds(ds, &os);
719 cmn_err(CE_WARN, "can't open objset %llu, error %d",
720 (unsigned long long)ds->ds_object, error);
724 zilog = dmu_objset_zil(os);
725 bp = (blkptr_t *)&zilog->zl_header->zh_log;
728 * Check the first block and determine if it's on a log device
729 * which may have been removed or faulted prior to loading this
730 * pool. If so, there's no point in checking the rest of the log
731 * as its content should have already been synced to the pool.
733 if (!BP_IS_HOLE(bp)) {
735 boolean_t valid = B_TRUE;
737 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
738 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
739 if (vd->vdev_islog && vdev_is_dead(vd))
740 valid = vdev_log_state_valid(vd);
741 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
748 * Because tx == NULL, zil_claim_log_block() will not actually claim
749 * any blocks, but just determine whether it is possible to do so.
750 * In addition to checking the log chain, zil_claim_log_block()
751 * will invoke zio_claim() with a done func of spa_claim_notify(),
752 * which will update spa_max_claim_txg. See spa_load() for details.
754 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
755 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
757 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
761 zil_vdev_compare(const void *x1, const void *x2)
763 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
764 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
775 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
777 avl_tree_t *t = &zilog->zl_vdev_tree;
779 zil_vdev_node_t *zv, zvsearch;
780 int ndvas = BP_GET_NDVAS(bp);
783 if (zfs_nocacheflush)
786 ASSERT(zilog->zl_writer);
789 * Even though we're zl_writer, we still need a lock because the
790 * zl_get_data() callbacks may have dmu_sync() done callbacks
791 * that will run concurrently.
793 mutex_enter(&zilog->zl_vdev_lock);
794 for (i = 0; i < ndvas; i++) {
795 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
796 if (avl_find(t, &zvsearch, &where) == NULL) {
797 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
798 zv->zv_vdev = zvsearch.zv_vdev;
799 avl_insert(t, zv, where);
802 mutex_exit(&zilog->zl_vdev_lock);
806 zil_flush_vdevs(zilog_t *zilog)
808 spa_t *spa = zilog->zl_spa;
809 avl_tree_t *t = &zilog->zl_vdev_tree;
814 ASSERT(zilog->zl_writer);
817 * We don't need zl_vdev_lock here because we're the zl_writer,
818 * and all zl_get_data() callbacks are done.
820 if (avl_numnodes(t) == 0)
823 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
825 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
827 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
828 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
831 kmem_free(zv, sizeof (*zv));
835 * Wait for all the flushes to complete. Not all devices actually
836 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
838 (void) zio_wait(zio);
840 spa_config_exit(spa, SCL_STATE, FTAG);
844 * Function called when a log block write completes
847 zil_lwb_write_done(zio_t *zio)
849 lwb_t *lwb = zio->io_private;
850 zilog_t *zilog = lwb->lwb_zilog;
851 dmu_tx_t *tx = lwb->lwb_tx;
853 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
854 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
855 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
856 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
857 ASSERT(!BP_IS_GANG(zio->io_bp));
858 ASSERT(!BP_IS_HOLE(zio->io_bp));
859 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
862 * Ensure the lwb buffer pointer is cleared before releasing
863 * the txg. If we have had an allocation failure and
864 * the txg is waiting to sync then we want want zil_sync()
865 * to remove the lwb so that it's not picked up as the next new
866 * one in zil_commit_writer(). zil_sync() will only remove
867 * the lwb if lwb_buf is null.
869 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
870 mutex_enter(&zilog->zl_lock);
873 mutex_exit(&zilog->zl_lock);
876 * Now that we've written this log block, we have a stable pointer
877 * to the next block in the chain, so it's OK to let the txg in
878 * which we allocated the next block sync.
884 * Initialize the io for a log block.
887 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
891 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
892 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
893 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
895 if (zilog->zl_root_zio == NULL) {
896 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
899 if (lwb->lwb_zio == NULL) {
900 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
901 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
902 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
903 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
908 * Define a limited set of intent log block sizes.
910 * These must be a multiple of 4KB. Note only the amount used (again
911 * aligned to 4KB) actually gets written. However, we can't always just
912 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
914 uint64_t zil_block_buckets[] = {
915 4096, /* non TX_WRITE */
916 8192+4096, /* data base */
917 32*1024 + 4096, /* NFS writes */
922 * Use the slog as long as the logbias is 'latency' and the current commit size
923 * is less than the limit or the total list size is less than 2X the limit.
924 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
926 uint64_t zil_slog_limit = 1024 * 1024;
927 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
928 (((zilog)->zl_cur_used < zil_slog_limit) || \
929 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
932 * Start a log block write and advance to the next log block.
933 * Calls are serialized.
936 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
940 spa_t *spa = zilog->zl_spa;
944 uint64_t zil_blksz, wsz;
947 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
948 zilc = (zil_chain_t *)lwb->lwb_buf;
949 bp = &zilc->zc_next_blk;
951 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
952 bp = &zilc->zc_next_blk;
955 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
958 * Allocate the next block and save its address in this block
959 * before writing it in order to establish the log chain.
960 * Note that if the allocation of nlwb synced before we wrote
961 * the block that points at it (lwb), we'd leak it if we crashed.
962 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
963 * We dirty the dataset to ensure that zil_sync() will be called
964 * to clean up in the event of allocation failure or I/O failure.
966 tx = dmu_tx_create(zilog->zl_os);
967 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
968 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
969 txg = dmu_tx_get_txg(tx);
974 * Log blocks are pre-allocated. Here we select the size of the next
975 * block, based on size used in the last block.
976 * - first find the smallest bucket that will fit the block from a
977 * limited set of block sizes. This is because it's faster to write
978 * blocks allocated from the same metaslab as they are adjacent or
980 * - next find the maximum from the new suggested size and an array of
981 * previous sizes. This lessens a picket fence effect of wrongly
982 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
985 * Note we only write what is used, but we can't just allocate
986 * the maximum block size because we can exhaust the available
989 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
990 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
992 zil_blksz = zil_block_buckets[i];
993 if (zil_blksz == UINT64_MAX)
994 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
995 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
996 for (i = 0; i < ZIL_PREV_BLKS; i++)
997 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
998 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1001 /* pass the old blkptr in order to spread log blocks across devs */
1002 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1005 ASSERT3U(bp->blk_birth, ==, txg);
1006 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1007 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1010 * Allocate a new log write buffer (lwb).
1012 nlwb = zil_alloc_lwb(zilog, bp, txg);
1014 /* Record the block for later vdev flushing */
1015 zil_add_block(zilog, &lwb->lwb_blk);
1018 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1019 /* For Slim ZIL only write what is used. */
1020 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1021 ASSERT3U(wsz, <=, lwb->lwb_sz);
1022 zio_shrink(lwb->lwb_zio, wsz);
1029 zilc->zc_nused = lwb->lwb_nused;
1030 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1033 * clear unused data for security
1035 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1037 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1040 * If there was an allocation failure then nlwb will be null which
1041 * forces a txg_wait_synced().
1047 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1049 lr_t *lrc = &itx->itx_lr; /* common log record */
1050 lr_write_t *lrw = (lr_write_t *)lrc;
1052 uint64_t txg = lrc->lrc_txg;
1053 uint64_t reclen = lrc->lrc_reclen;
1059 ASSERT(lwb->lwb_buf != NULL);
1060 ASSERT(zilog_is_dirty(zilog) ||
1061 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1063 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1064 dlen = P2ROUNDUP_TYPED(
1065 lrw->lr_length, sizeof (uint64_t), uint64_t);
1067 zilog->zl_cur_used += (reclen + dlen);
1069 zil_lwb_write_init(zilog, lwb);
1072 * If this record won't fit in the current log block, start a new one.
1074 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1075 lwb = zil_lwb_write_start(zilog, lwb);
1078 zil_lwb_write_init(zilog, lwb);
1079 ASSERT(LWB_EMPTY(lwb));
1080 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1081 txg_wait_synced(zilog->zl_dmu_pool, txg);
1086 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1087 bcopy(lrc, lr_buf, reclen);
1088 lrc = (lr_t *)lr_buf;
1089 lrw = (lr_write_t *)lrc;
1092 * If it's a write, fetch the data or get its blkptr as appropriate.
1094 if (lrc->lrc_txtype == TX_WRITE) {
1095 if (txg > spa_freeze_txg(zilog->zl_spa))
1096 txg_wait_synced(zilog->zl_dmu_pool, txg);
1097 if (itx->itx_wr_state != WR_COPIED) {
1102 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1103 dbuf = lr_buf + reclen;
1104 lrw->lr_common.lrc_reclen += dlen;
1106 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1109 error = zilog->zl_get_data(
1110 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1112 txg_wait_synced(zilog->zl_dmu_pool, txg);
1116 ASSERT(error == ENOENT || error == EEXIST ||
1124 * We're actually making an entry, so update lrc_seq to be the
1125 * log record sequence number. Note that this is generally not
1126 * equal to the itx sequence number because not all transactions
1127 * are synchronous, and sometimes spa_sync() gets there first.
1129 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1130 lwb->lwb_nused += reclen + dlen;
1131 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1132 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1133 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1139 zil_itx_create(uint64_t txtype, size_t lrsize)
1143 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1145 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1146 itx->itx_lr.lrc_txtype = txtype;
1147 itx->itx_lr.lrc_reclen = lrsize;
1148 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1149 itx->itx_lr.lrc_seq = 0; /* defensive */
1150 itx->itx_sync = B_TRUE; /* default is synchronous */
1156 zil_itx_destroy(itx_t *itx)
1158 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1162 * Free up the sync and async itxs. The itxs_t has already been detached
1163 * so no locks are needed.
1166 zil_itxg_clean(itxs_t *itxs)
1172 itx_async_node_t *ian;
1174 list = &itxs->i_sync_list;
1175 while ((itx = list_head(list)) != NULL) {
1176 list_remove(list, itx);
1177 kmem_free(itx, offsetof(itx_t, itx_lr) +
1178 itx->itx_lr.lrc_reclen);
1182 t = &itxs->i_async_tree;
1183 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1184 list = &ian->ia_list;
1185 while ((itx = list_head(list)) != NULL) {
1186 list_remove(list, itx);
1187 kmem_free(itx, offsetof(itx_t, itx_lr) +
1188 itx->itx_lr.lrc_reclen);
1191 kmem_free(ian, sizeof (itx_async_node_t));
1195 kmem_free(itxs, sizeof (itxs_t));
1199 zil_aitx_compare(const void *x1, const void *x2)
1201 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1202 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1213 * Remove all async itx with the given oid.
1216 zil_remove_async(zilog_t *zilog, uint64_t oid)
1219 itx_async_node_t *ian;
1226 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1228 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1231 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1233 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1234 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1236 mutex_enter(&itxg->itxg_lock);
1237 if (itxg->itxg_txg != txg) {
1238 mutex_exit(&itxg->itxg_lock);
1243 * Locate the object node and append its list.
1245 t = &itxg->itxg_itxs->i_async_tree;
1246 ian = avl_find(t, &oid, &where);
1248 list_move_tail(&clean_list, &ian->ia_list);
1249 mutex_exit(&itxg->itxg_lock);
1251 while ((itx = list_head(&clean_list)) != NULL) {
1252 list_remove(&clean_list, itx);
1253 kmem_free(itx, offsetof(itx_t, itx_lr) +
1254 itx->itx_lr.lrc_reclen);
1256 list_destroy(&clean_list);
1260 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1264 itxs_t *itxs, *clean = NULL;
1267 * Object ids can be re-instantiated in the next txg so
1268 * remove any async transactions to avoid future leaks.
1269 * This can happen if a fsync occurs on the re-instantiated
1270 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1271 * the new file data and flushes a write record for the old object.
1273 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1274 zil_remove_async(zilog, itx->itx_oid);
1277 * Ensure the data of a renamed file is committed before the rename.
1279 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1280 zil_async_to_sync(zilog, itx->itx_oid);
1282 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1285 txg = dmu_tx_get_txg(tx);
1287 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1288 mutex_enter(&itxg->itxg_lock);
1289 itxs = itxg->itxg_itxs;
1290 if (itxg->itxg_txg != txg) {
1293 * The zil_clean callback hasn't got around to cleaning
1294 * this itxg. Save the itxs for release below.
1295 * This should be rare.
1297 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1299 clean = itxg->itxg_itxs;
1301 ASSERT(itxg->itxg_sod == 0);
1302 itxg->itxg_txg = txg;
1303 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1305 list_create(&itxs->i_sync_list, sizeof (itx_t),
1306 offsetof(itx_t, itx_node));
1307 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1308 sizeof (itx_async_node_t),
1309 offsetof(itx_async_node_t, ia_node));
1311 if (itx->itx_sync) {
1312 list_insert_tail(&itxs->i_sync_list, itx);
1313 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1314 itxg->itxg_sod += itx->itx_sod;
1316 avl_tree_t *t = &itxs->i_async_tree;
1317 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1318 itx_async_node_t *ian;
1321 ian = avl_find(t, &foid, &where);
1323 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1324 list_create(&ian->ia_list, sizeof (itx_t),
1325 offsetof(itx_t, itx_node));
1326 ian->ia_foid = foid;
1327 avl_insert(t, ian, where);
1329 list_insert_tail(&ian->ia_list, itx);
1332 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1333 zilog_dirty(zilog, txg);
1334 mutex_exit(&itxg->itxg_lock);
1336 /* Release the old itxs now we've dropped the lock */
1338 zil_itxg_clean(clean);
1342 * If there are any in-memory intent log transactions which have now been
1343 * synced then start up a taskq to free them. We should only do this after we
1344 * have written out the uberblocks (i.e. txg has been comitted) so that
1345 * don't inadvertently clean out in-memory log records that would be required
1349 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1351 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1354 mutex_enter(&itxg->itxg_lock);
1355 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1356 mutex_exit(&itxg->itxg_lock);
1359 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1360 ASSERT(itxg->itxg_txg != 0);
1361 ASSERT(zilog->zl_clean_taskq != NULL);
1362 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1364 clean_me = itxg->itxg_itxs;
1365 itxg->itxg_itxs = NULL;
1367 mutex_exit(&itxg->itxg_lock);
1369 * Preferably start a task queue to free up the old itxs but
1370 * if taskq_dispatch can't allocate resources to do that then
1371 * free it in-line. This should be rare. Note, using TQ_SLEEP
1372 * created a bad performance problem.
1374 if (taskq_dispatch(zilog->zl_clean_taskq,
1375 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1376 zil_itxg_clean(clean_me);
1380 * Get the list of itxs to commit into zl_itx_commit_list.
1383 zil_get_commit_list(zilog_t *zilog)
1386 list_t *commit_list = &zilog->zl_itx_commit_list;
1387 uint64_t push_sod = 0;
1389 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1392 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1394 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1395 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1397 mutex_enter(&itxg->itxg_lock);
1398 if (itxg->itxg_txg != txg) {
1399 mutex_exit(&itxg->itxg_lock);
1403 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1404 push_sod += itxg->itxg_sod;
1407 mutex_exit(&itxg->itxg_lock);
1409 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1413 * Move the async itxs for a specified object to commit into sync lists.
1416 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1419 itx_async_node_t *ian;
1423 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1426 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1428 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1429 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1431 mutex_enter(&itxg->itxg_lock);
1432 if (itxg->itxg_txg != txg) {
1433 mutex_exit(&itxg->itxg_lock);
1438 * If a foid is specified then find that node and append its
1439 * list. Otherwise walk the tree appending all the lists
1440 * to the sync list. We add to the end rather than the
1441 * beginning to ensure the create has happened.
1443 t = &itxg->itxg_itxs->i_async_tree;
1445 ian = avl_find(t, &foid, &where);
1447 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1451 void *cookie = NULL;
1453 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1454 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1456 list_destroy(&ian->ia_list);
1457 kmem_free(ian, sizeof (itx_async_node_t));
1460 mutex_exit(&itxg->itxg_lock);
1465 zil_commit_writer(zilog_t *zilog)
1470 spa_t *spa = zilog->zl_spa;
1473 ASSERT(zilog->zl_root_zio == NULL);
1475 mutex_exit(&zilog->zl_lock);
1477 zil_get_commit_list(zilog);
1480 * Return if there's nothing to commit before we dirty the fs by
1481 * calling zil_create().
1483 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1484 mutex_enter(&zilog->zl_lock);
1488 if (zilog->zl_suspend) {
1491 lwb = list_tail(&zilog->zl_lwb_list);
1493 lwb = zil_create(zilog);
1496 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1497 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1498 txg = itx->itx_lr.lrc_txg;
1501 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1502 lwb = zil_lwb_commit(zilog, itx, lwb);
1503 list_remove(&zilog->zl_itx_commit_list, itx);
1504 kmem_free(itx, offsetof(itx_t, itx_lr)
1505 + itx->itx_lr.lrc_reclen);
1507 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1509 /* write the last block out */
1510 if (lwb != NULL && lwb->lwb_zio != NULL)
1511 lwb = zil_lwb_write_start(zilog, lwb);
1513 zilog->zl_cur_used = 0;
1516 * Wait if necessary for the log blocks to be on stable storage.
1518 if (zilog->zl_root_zio) {
1519 error = zio_wait(zilog->zl_root_zio);
1520 zilog->zl_root_zio = NULL;
1521 zil_flush_vdevs(zilog);
1524 if (error || lwb == NULL)
1525 txg_wait_synced(zilog->zl_dmu_pool, 0);
1527 mutex_enter(&zilog->zl_lock);
1530 * Remember the highest committed log sequence number for ztest.
1531 * We only update this value when all the log writes succeeded,
1532 * because ztest wants to ASSERT that it got the whole log chain.
1534 if (error == 0 && lwb != NULL)
1535 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1539 * Commit zfs transactions to stable storage.
1540 * If foid is 0 push out all transactions, otherwise push only those
1541 * for that object or might reference that object.
1543 * itxs are committed in batches. In a heavily stressed zil there will be
1544 * a commit writer thread who is writing out a bunch of itxs to the log
1545 * for a set of committing threads (cthreads) in the same batch as the writer.
1546 * Those cthreads are all waiting on the same cv for that batch.
1548 * There will also be a different and growing batch of threads that are
1549 * waiting to commit (qthreads). When the committing batch completes
1550 * a transition occurs such that the cthreads exit and the qthreads become
1551 * cthreads. One of the new cthreads becomes the writer thread for the
1552 * batch. Any new threads arriving become new qthreads.
1554 * Only 2 condition variables are needed and there's no transition
1555 * between the two cvs needed. They just flip-flop between qthreads
1558 * Using this scheme we can efficiently wakeup up only those threads
1559 * that have been committed.
1562 zil_commit(zilog_t *zilog, uint64_t foid)
1566 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1569 /* move the async itxs for the foid to the sync queues */
1570 zil_async_to_sync(zilog, foid);
1572 mutex_enter(&zilog->zl_lock);
1573 mybatch = zilog->zl_next_batch;
1574 while (zilog->zl_writer) {
1575 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1576 if (mybatch <= zilog->zl_com_batch) {
1577 mutex_exit(&zilog->zl_lock);
1582 zilog->zl_next_batch++;
1583 zilog->zl_writer = B_TRUE;
1584 zil_commit_writer(zilog);
1585 zilog->zl_com_batch = mybatch;
1586 zilog->zl_writer = B_FALSE;
1587 mutex_exit(&zilog->zl_lock);
1589 /* wake up one thread to become the next writer */
1590 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1592 /* wake up all threads waiting for this batch to be committed */
1593 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1597 * Called in syncing context to free committed log blocks and update log header.
1600 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1602 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1603 uint64_t txg = dmu_tx_get_txg(tx);
1604 spa_t *spa = zilog->zl_spa;
1605 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1609 * We don't zero out zl_destroy_txg, so make sure we don't try
1610 * to destroy it twice.
1612 if (spa_sync_pass(spa) != 1)
1615 mutex_enter(&zilog->zl_lock);
1617 ASSERT(zilog->zl_stop_sync == 0);
1619 if (*replayed_seq != 0) {
1620 ASSERT(zh->zh_replay_seq < *replayed_seq);
1621 zh->zh_replay_seq = *replayed_seq;
1625 if (zilog->zl_destroy_txg == txg) {
1626 blkptr_t blk = zh->zh_log;
1628 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1630 bzero(zh, sizeof (zil_header_t));
1631 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1633 if (zilog->zl_keep_first) {
1635 * If this block was part of log chain that couldn't
1636 * be claimed because a device was missing during
1637 * zil_claim(), but that device later returns,
1638 * then this block could erroneously appear valid.
1639 * To guard against this, assign a new GUID to the new
1640 * log chain so it doesn't matter what blk points to.
1642 zil_init_log_chain(zilog, &blk);
1647 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1648 zh->zh_log = lwb->lwb_blk;
1649 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1651 list_remove(&zilog->zl_lwb_list, lwb);
1652 zio_free_zil(spa, txg, &lwb->lwb_blk);
1653 kmem_cache_free(zil_lwb_cache, lwb);
1656 * If we don't have anything left in the lwb list then
1657 * we've had an allocation failure and we need to zero
1658 * out the zil_header blkptr so that we don't end
1659 * up freeing the same block twice.
1661 if (list_head(&zilog->zl_lwb_list) == NULL)
1662 BP_ZERO(&zh->zh_log);
1664 mutex_exit(&zilog->zl_lock);
1670 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1671 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1677 kmem_cache_destroy(zil_lwb_cache);
1681 zil_set_sync(zilog_t *zilog, uint64_t sync)
1683 zilog->zl_sync = sync;
1687 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1689 zilog->zl_logbias = logbias;
1693 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1697 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1699 zilog->zl_header = zh_phys;
1701 zilog->zl_spa = dmu_objset_spa(os);
1702 zilog->zl_dmu_pool = dmu_objset_pool(os);
1703 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1704 zilog->zl_logbias = dmu_objset_logbias(os);
1705 zilog->zl_sync = dmu_objset_syncprop(os);
1706 zilog->zl_next_batch = 1;
1708 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1710 for (int i = 0; i < TXG_SIZE; i++) {
1711 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1712 MUTEX_DEFAULT, NULL);
1715 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1716 offsetof(lwb_t, lwb_node));
1718 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1719 offsetof(itx_t, itx_node));
1721 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1723 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1724 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1726 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1727 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1728 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1729 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1735 zil_free(zilog_t *zilog)
1737 zilog->zl_stop_sync = 1;
1739 ASSERT0(zilog->zl_suspend);
1740 ASSERT0(zilog->zl_suspending);
1742 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1743 list_destroy(&zilog->zl_lwb_list);
1745 avl_destroy(&zilog->zl_vdev_tree);
1746 mutex_destroy(&zilog->zl_vdev_lock);
1748 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1749 list_destroy(&zilog->zl_itx_commit_list);
1751 for (int i = 0; i < TXG_SIZE; i++) {
1753 * It's possible for an itx to be generated that doesn't dirty
1754 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1755 * callback to remove the entry. We remove those here.
1757 * Also free up the ziltest itxs.
1759 if (zilog->zl_itxg[i].itxg_itxs)
1760 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1761 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1764 mutex_destroy(&zilog->zl_lock);
1766 cv_destroy(&zilog->zl_cv_writer);
1767 cv_destroy(&zilog->zl_cv_suspend);
1768 cv_destroy(&zilog->zl_cv_batch[0]);
1769 cv_destroy(&zilog->zl_cv_batch[1]);
1771 kmem_free(zilog, sizeof (zilog_t));
1775 * Open an intent log.
1778 zil_open(objset_t *os, zil_get_data_t *get_data)
1780 zilog_t *zilog = dmu_objset_zil(os);
1782 ASSERT(zilog->zl_clean_taskq == NULL);
1783 ASSERT(zilog->zl_get_data == NULL);
1784 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1786 zilog->zl_get_data = get_data;
1787 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1788 2, 2, TASKQ_PREPOPULATE);
1794 * Close an intent log.
1797 zil_close(zilog_t *zilog)
1802 zil_commit(zilog, 0); /* commit all itx */
1805 * The lwb_max_txg for the stubby lwb will reflect the last activity
1806 * for the zil. After a txg_wait_synced() on the txg we know all the
1807 * callbacks have occurred that may clean the zil. Only then can we
1808 * destroy the zl_clean_taskq.
1810 mutex_enter(&zilog->zl_lock);
1811 lwb = list_tail(&zilog->zl_lwb_list);
1813 txg = lwb->lwb_max_txg;
1814 mutex_exit(&zilog->zl_lock);
1816 txg_wait_synced(zilog->zl_dmu_pool, txg);
1817 ASSERT(!zilog_is_dirty(zilog));
1819 taskq_destroy(zilog->zl_clean_taskq);
1820 zilog->zl_clean_taskq = NULL;
1821 zilog->zl_get_data = NULL;
1824 * We should have only one LWB left on the list; remove it now.
1826 mutex_enter(&zilog->zl_lock);
1827 lwb = list_head(&zilog->zl_lwb_list);
1829 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1830 list_remove(&zilog->zl_lwb_list, lwb);
1831 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1832 kmem_cache_free(zil_lwb_cache, lwb);
1834 mutex_exit(&zilog->zl_lock);
1837 static char *suspend_tag = "zil suspending";
1840 * Suspend an intent log. While in suspended mode, we still honor
1841 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1842 * On old version pools, we suspend the log briefly when taking a
1843 * snapshot so that it will have an empty intent log.
1845 * Long holds are not really intended to be used the way we do here --
1846 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1847 * could fail. Therefore we take pains to only put a long hold if it is
1848 * actually necessary. Fortunately, it will only be necessary if the
1849 * objset is currently mounted (or the ZVOL equivalent). In that case it
1850 * will already have a long hold, so we are not really making things any worse.
1852 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1853 * zvol_state_t), and use their mechanism to prevent their hold from being
1854 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1857 * if cookiep == NULL, this does both the suspend & resume.
1858 * Otherwise, it returns with the dataset "long held", and the cookie
1859 * should be passed into zil_resume().
1862 zil_suspend(const char *osname, void **cookiep)
1866 const zil_header_t *zh;
1869 error = dmu_objset_hold(osname, suspend_tag, &os);
1872 zilog = dmu_objset_zil(os);
1874 mutex_enter(&zilog->zl_lock);
1875 zh = zilog->zl_header;
1877 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1878 mutex_exit(&zilog->zl_lock);
1879 dmu_objset_rele(os, suspend_tag);
1880 return (SET_ERROR(EBUSY));
1884 * Don't put a long hold in the cases where we can avoid it. This
1885 * is when there is no cookie so we are doing a suspend & resume
1886 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1887 * for the suspend because it's already suspended, or there's no ZIL.
1889 if (cookiep == NULL && !zilog->zl_suspending &&
1890 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1891 mutex_exit(&zilog->zl_lock);
1892 dmu_objset_rele(os, suspend_tag);
1896 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1897 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1899 zilog->zl_suspend++;
1901 if (zilog->zl_suspend > 1) {
1903 * Someone else is already suspending it.
1904 * Just wait for them to finish.
1907 while (zilog->zl_suspending)
1908 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1909 mutex_exit(&zilog->zl_lock);
1911 if (cookiep == NULL)
1919 * If there is no pointer to an on-disk block, this ZIL must not
1920 * be active (e.g. filesystem not mounted), so there's nothing
1923 if (BP_IS_HOLE(&zh->zh_log)) {
1924 ASSERT(cookiep != NULL); /* fast path already handled */
1927 mutex_exit(&zilog->zl_lock);
1931 zilog->zl_suspending = B_TRUE;
1932 mutex_exit(&zilog->zl_lock);
1934 zil_commit(zilog, 0);
1936 zil_destroy(zilog, B_FALSE);
1938 mutex_enter(&zilog->zl_lock);
1939 zilog->zl_suspending = B_FALSE;
1940 cv_broadcast(&zilog->zl_cv_suspend);
1941 mutex_exit(&zilog->zl_lock);
1943 if (cookiep == NULL)
1951 zil_resume(void *cookie)
1953 objset_t *os = cookie;
1954 zilog_t *zilog = dmu_objset_zil(os);
1956 mutex_enter(&zilog->zl_lock);
1957 ASSERT(zilog->zl_suspend != 0);
1958 zilog->zl_suspend--;
1959 mutex_exit(&zilog->zl_lock);
1960 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1961 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1964 typedef struct zil_replay_arg {
1965 zil_replay_func_t **zr_replay;
1967 boolean_t zr_byteswap;
1972 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1974 char name[ZFS_MAX_DATASET_NAME_LEN];
1976 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1978 dmu_objset_name(zilog->zl_os, name);
1980 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1981 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1982 (u_longlong_t)lr->lrc_seq,
1983 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1984 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1990 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1992 zil_replay_arg_t *zr = zra;
1993 const zil_header_t *zh = zilog->zl_header;
1994 uint64_t reclen = lr->lrc_reclen;
1995 uint64_t txtype = lr->lrc_txtype;
1998 zilog->zl_replaying_seq = lr->lrc_seq;
2000 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2003 if (lr->lrc_txg < claim_txg) /* already committed */
2006 /* Strip case-insensitive bit, still present in log record */
2009 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2010 return (zil_replay_error(zilog, lr, EINVAL));
2013 * If this record type can be logged out of order, the object
2014 * (lr_foid) may no longer exist. That's legitimate, not an error.
2016 if (TX_OOO(txtype)) {
2017 error = dmu_object_info(zilog->zl_os,
2018 ((lr_ooo_t *)lr)->lr_foid, NULL);
2019 if (error == ENOENT || error == EEXIST)
2024 * Make a copy of the data so we can revise and extend it.
2026 bcopy(lr, zr->zr_lr, reclen);
2029 * If this is a TX_WRITE with a blkptr, suck in the data.
2031 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2032 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2033 zr->zr_lr + reclen);
2035 return (zil_replay_error(zilog, lr, error));
2039 * The log block containing this lr may have been byteswapped
2040 * so that we can easily examine common fields like lrc_txtype.
2041 * However, the log is a mix of different record types, and only the
2042 * replay vectors know how to byteswap their records. Therefore, if
2043 * the lr was byteswapped, undo it before invoking the replay vector.
2045 if (zr->zr_byteswap)
2046 byteswap_uint64_array(zr->zr_lr, reclen);
2049 * We must now do two things atomically: replay this log record,
2050 * and update the log header sequence number to reflect the fact that
2051 * we did so. At the end of each replay function the sequence number
2052 * is updated if we are in replay mode.
2054 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2057 * The DMU's dnode layer doesn't see removes until the txg
2058 * commits, so a subsequent claim can spuriously fail with
2059 * EEXIST. So if we receive any error we try syncing out
2060 * any removes then retry the transaction. Note that we
2061 * specify B_FALSE for byteswap now, so we don't do it twice.
2063 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2064 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2066 return (zil_replay_error(zilog, lr, error));
2073 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2075 zilog->zl_replay_blks++;
2081 * If this dataset has a non-empty intent log, replay it and destroy it.
2084 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2086 zilog_t *zilog = dmu_objset_zil(os);
2087 const zil_header_t *zh = zilog->zl_header;
2088 zil_replay_arg_t zr;
2090 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2091 zil_destroy(zilog, B_TRUE);
2095 zr.zr_replay = replay_func;
2097 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2098 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2101 * Wait for in-progress removes to sync before starting replay.
2103 txg_wait_synced(zilog->zl_dmu_pool, 0);
2105 zilog->zl_replay = B_TRUE;
2106 zilog->zl_replay_time = ddi_get_lbolt();
2107 ASSERT(zilog->zl_replay_blks == 0);
2108 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2110 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2112 zil_destroy(zilog, B_FALSE);
2113 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2114 zilog->zl_replay = B_FALSE;
2118 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2120 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2123 if (zilog->zl_replay) {
2124 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2125 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2126 zilog->zl_replaying_seq;
2135 zil_vdev_offline(const char *osname, void *arg)
2139 error = zil_suspend(osname, NULL);
2141 return (SET_ERROR(EEXIST));