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) 2012 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>
41 #include <sys/metaslab.h>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * See zil.h for more information about these fields.
72 zil_stats_t zil_stats = {
73 { "zil_commit_count", KSTAT_DATA_UINT64 },
74 { "zil_commit_writer_count", KSTAT_DATA_UINT64 },
75 { "zil_itx_count", KSTAT_DATA_UINT64 },
76 { "zil_itx_indirect_count", KSTAT_DATA_UINT64 },
77 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 },
78 { "zil_itx_copied_count", KSTAT_DATA_UINT64 },
79 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 },
80 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 },
81 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 },
82 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 },
83 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 },
84 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 },
85 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 },
88 static kstat_t *zil_ksp;
91 * This global ZIL switch affects all pools
93 int zil_replay_disable = 0; /* disable intent logging replay */
96 * Tunable parameter for debugging or performance analysis. Setting
97 * zfs_nocacheflush will cause corruption on power loss if a volatile
98 * out-of-order write cache is enabled.
100 int zfs_nocacheflush = 0;
102 static kmem_cache_t *zil_lwb_cache;
104 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
106 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
107 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
111 * ziltest is by and large an ugly hack, but very useful in
112 * checking replay without tedious work.
113 * When running ziltest we want to keep all itx's and so maintain
114 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
115 * We subtract TXG_CONCURRENT_STATES to allow for common code.
117 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
120 zil_bp_compare(const void *x1, const void *x2)
122 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
123 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
125 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
127 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
130 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
132 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
139 zil_bp_tree_init(zilog_t *zilog)
141 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
142 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
146 zil_bp_tree_fini(zilog_t *zilog)
148 avl_tree_t *t = &zilog->zl_bp_tree;
152 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
153 kmem_free(zn, sizeof (zil_bp_node_t));
159 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
161 avl_tree_t *t = &zilog->zl_bp_tree;
162 const dva_t *dva = BP_IDENTITY(bp);
166 if (avl_find(t, dva, &where) != NULL)
169 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_PUSHPAGE);
171 avl_insert(t, zn, where);
176 static zil_header_t *
177 zil_header_in_syncing_context(zilog_t *zilog)
179 return ((zil_header_t *)zilog->zl_header);
183 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
185 zio_cksum_t *zc = &bp->blk_cksum;
187 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
188 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
189 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
190 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
194 * Read a log block and make sure it's valid.
197 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
200 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
201 uint32_t aflags = ARC_WAIT;
202 arc_buf_t *abuf = NULL;
206 if (zilog->zl_header->zh_claim_txg == 0)
207 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
209 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
210 zio_flags |= ZIO_FLAG_SPECULATIVE;
212 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
213 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
215 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
216 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
219 zio_cksum_t cksum = bp->blk_cksum;
222 * Validate the checksummed log block.
224 * Sequence numbers should be... sequential. The checksum
225 * verifier for the next block should be bp's checksum plus 1.
227 * Also check the log chain linkage and size used.
229 cksum.zc_word[ZIL_ZC_SEQ]++;
231 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
232 zil_chain_t *zilc = abuf->b_data;
233 char *lr = (char *)(zilc + 1);
234 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
236 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
237 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
241 *end = (char *)dst + len;
242 *nbp = zilc->zc_next_blk;
245 char *lr = abuf->b_data;
246 uint64_t size = BP_GET_LSIZE(bp);
247 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
249 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
250 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
251 (zilc->zc_nused > (size - sizeof (*zilc)))) {
254 bcopy(lr, dst, zilc->zc_nused);
255 *end = (char *)dst + zilc->zc_nused;
256 *nbp = zilc->zc_next_blk;
260 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
267 * Read a TX_WRITE log data block.
270 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
272 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
273 const blkptr_t *bp = &lr->lr_blkptr;
274 uint32_t aflags = ARC_WAIT;
275 arc_buf_t *abuf = NULL;
279 if (BP_IS_HOLE(bp)) {
281 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
285 if (zilog->zl_header->zh_claim_txg == 0)
286 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
288 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
289 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
291 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
292 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
296 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
297 (void) arc_buf_remove_ref(abuf, &abuf);
304 * Parse the intent log, and call parse_func for each valid record within.
307 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
308 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
310 const zil_header_t *zh = zilog->zl_header;
311 boolean_t claimed = !!zh->zh_claim_txg;
312 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
313 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
314 uint64_t max_blk_seq = 0;
315 uint64_t max_lr_seq = 0;
316 uint64_t blk_count = 0;
317 uint64_t lr_count = 0;
318 blkptr_t blk, next_blk;
322 bzero(&next_blk, sizeof(blkptr_t));
325 * Old logs didn't record the maximum zh_claim_lr_seq.
327 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
328 claim_lr_seq = UINT64_MAX;
331 * Starting at the block pointed to by zh_log we read the log chain.
332 * For each block in the chain we strongly check that block to
333 * ensure its validity. We stop when an invalid block is found.
334 * For each block pointer in the chain we call parse_blk_func().
335 * For each record in each valid block we call parse_lr_func().
336 * If the log has been claimed, stop if we encounter a sequence
337 * number greater than the highest claimed sequence number.
339 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
340 zil_bp_tree_init(zilog);
342 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
343 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
347 if (blk_seq > claim_blk_seq)
349 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
351 ASSERT3U(max_blk_seq, <, blk_seq);
352 max_blk_seq = blk_seq;
355 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
358 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
362 for (lrp = lrbuf; lrp < end; lrp += reclen) {
363 lr_t *lr = (lr_t *)lrp;
364 reclen = lr->lrc_reclen;
365 ASSERT3U(reclen, >=, sizeof (lr_t));
366 if (lr->lrc_seq > claim_lr_seq)
368 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
370 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
371 max_lr_seq = lr->lrc_seq;
376 zilog->zl_parse_error = error;
377 zilog->zl_parse_blk_seq = max_blk_seq;
378 zilog->zl_parse_lr_seq = max_lr_seq;
379 zilog->zl_parse_blk_count = blk_count;
380 zilog->zl_parse_lr_count = lr_count;
382 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
383 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
385 zil_bp_tree_fini(zilog);
386 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
392 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
395 * Claim log block if not already committed and not already claimed.
396 * If tx == NULL, just verify that the block is claimable.
398 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
401 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
402 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
403 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
407 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
409 lr_write_t *lr = (lr_write_t *)lrc;
412 if (lrc->lrc_txtype != TX_WRITE)
416 * If the block is not readable, don't claim it. This can happen
417 * in normal operation when a log block is written to disk before
418 * some of the dmu_sync() blocks it points to. In this case, the
419 * transaction cannot have been committed to anyone (we would have
420 * waited for all writes to be stable first), so it is semantically
421 * correct to declare this the end of the log.
423 if (lr->lr_blkptr.blk_birth >= first_txg &&
424 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
426 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
431 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
433 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
439 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
441 lr_write_t *lr = (lr_write_t *)lrc;
442 blkptr_t *bp = &lr->lr_blkptr;
445 * If we previously claimed it, we need to free it.
447 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
448 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
449 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
455 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg, boolean_t fastwrite)
459 lwb = kmem_cache_alloc(zil_lwb_cache, KM_PUSHPAGE);
460 lwb->lwb_zilog = zilog;
462 lwb->lwb_fastwrite = fastwrite;
463 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
464 lwb->lwb_max_txg = txg;
467 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
468 lwb->lwb_nused = sizeof (zil_chain_t);
469 lwb->lwb_sz = BP_GET_LSIZE(bp);
472 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
475 mutex_enter(&zilog->zl_lock);
476 list_insert_tail(&zilog->zl_lwb_list, lwb);
477 mutex_exit(&zilog->zl_lock);
483 * Called when we create in-memory log transactions so that we know
484 * to cleanup the itxs at the end of spa_sync().
487 zilog_dirty(zilog_t *zilog, uint64_t txg)
489 dsl_pool_t *dp = zilog->zl_dmu_pool;
490 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
492 if (dsl_dataset_is_snapshot(ds))
493 panic("dirtying snapshot!");
495 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg) == 0) {
496 /* up the hold count until we can be written out */
497 dmu_buf_add_ref(ds->ds_dbuf, zilog);
502 zilog_is_dirty(zilog_t *zilog)
504 dsl_pool_t *dp = zilog->zl_dmu_pool;
507 for (t = 0; t < TXG_SIZE; t++) {
508 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
515 * Create an on-disk intent log.
518 zil_create(zilog_t *zilog)
520 const zil_header_t *zh = zilog->zl_header;
526 boolean_t fastwrite = FALSE;
529 * Wait for any previous destroy to complete.
531 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
533 ASSERT(zh->zh_claim_txg == 0);
534 ASSERT(zh->zh_replay_seq == 0);
539 * Allocate an initial log block if:
540 * - there isn't one already
541 * - the existing block is the wrong endianess
543 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
544 tx = dmu_tx_create(zilog->zl_os);
545 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
546 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
547 txg = dmu_tx_get_txg(tx);
549 if (!BP_IS_HOLE(&blk)) {
550 zio_free_zil(zilog->zl_spa, txg, &blk);
554 error = zio_alloc_zil(zilog->zl_spa, txg, &blk,
555 ZIL_MIN_BLKSZ, B_TRUE);
559 zil_init_log_chain(zilog, &blk);
563 * Allocate a log write buffer (lwb) for the first log block.
566 lwb = zil_alloc_lwb(zilog, &blk, txg, fastwrite);
569 * If we just allocated the first log block, commit our transaction
570 * and wait for zil_sync() to stuff the block poiner into zh_log.
571 * (zh is part of the MOS, so we cannot modify it in open context.)
575 txg_wait_synced(zilog->zl_dmu_pool, txg);
578 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
584 * In one tx, free all log blocks and clear the log header.
585 * If keep_first is set, then we're replaying a log with no content.
586 * We want to keep the first block, however, so that the first
587 * synchronous transaction doesn't require a txg_wait_synced()
588 * in zil_create(). We don't need to txg_wait_synced() here either
589 * when keep_first is set, because both zil_create() and zil_destroy()
590 * will wait for any in-progress destroys to complete.
593 zil_destroy(zilog_t *zilog, boolean_t keep_first)
595 const zil_header_t *zh = zilog->zl_header;
601 * Wait for any previous destroy to complete.
603 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
605 zilog->zl_old_header = *zh; /* debugging aid */
607 if (BP_IS_HOLE(&zh->zh_log))
610 tx = dmu_tx_create(zilog->zl_os);
611 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
612 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
613 txg = dmu_tx_get_txg(tx);
615 mutex_enter(&zilog->zl_lock);
617 ASSERT3U(zilog->zl_destroy_txg, <, txg);
618 zilog->zl_destroy_txg = txg;
619 zilog->zl_keep_first = keep_first;
621 if (!list_is_empty(&zilog->zl_lwb_list)) {
622 ASSERT(zh->zh_claim_txg == 0);
624 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
625 ASSERT(lwb->lwb_zio == NULL);
626 if (lwb->lwb_fastwrite)
627 metaslab_fastwrite_unmark(zilog->zl_spa,
629 list_remove(&zilog->zl_lwb_list, lwb);
630 if (lwb->lwb_buf != NULL)
631 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
632 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
633 kmem_cache_free(zil_lwb_cache, lwb);
635 } else if (!keep_first) {
636 zil_destroy_sync(zilog, tx);
638 mutex_exit(&zilog->zl_lock);
644 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
646 ASSERT(list_is_empty(&zilog->zl_lwb_list));
647 (void) zil_parse(zilog, zil_free_log_block,
648 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
652 zil_claim(const char *osname, void *txarg)
654 dmu_tx_t *tx = txarg;
655 uint64_t first_txg = dmu_tx_get_txg(tx);
661 error = dmu_objset_hold(osname, FTAG, &os);
663 cmn_err(CE_WARN, "can't open objset for %s", osname);
667 zilog = dmu_objset_zil(os);
668 zh = zil_header_in_syncing_context(zilog);
670 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
671 if (!BP_IS_HOLE(&zh->zh_log))
672 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
673 BP_ZERO(&zh->zh_log);
674 dsl_dataset_dirty(dmu_objset_ds(os), tx);
675 dmu_objset_rele(os, FTAG);
680 * Claim all log blocks if we haven't already done so, and remember
681 * the highest claimed sequence number. This ensures that if we can
682 * read only part of the log now (e.g. due to a missing device),
683 * but we can read the entire log later, we will not try to replay
684 * or destroy beyond the last block we successfully claimed.
686 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
687 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
688 (void) zil_parse(zilog, zil_claim_log_block,
689 zil_claim_log_record, tx, first_txg);
690 zh->zh_claim_txg = first_txg;
691 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
692 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
693 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
694 zh->zh_flags |= ZIL_REPLAY_NEEDED;
695 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
696 dsl_dataset_dirty(dmu_objset_ds(os), tx);
699 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
700 dmu_objset_rele(os, FTAG);
705 * Check the log by walking the log chain.
706 * Checksum errors are ok as they indicate the end of the chain.
707 * Any other error (no device or read failure) returns an error.
710 zil_check_log_chain(const char *osname, void *tx)
719 error = dmu_objset_hold(osname, FTAG, &os);
721 cmn_err(CE_WARN, "can't open objset for %s", osname);
725 zilog = dmu_objset_zil(os);
726 bp = (blkptr_t *)&zilog->zl_header->zh_log;
729 * Check the first block and determine if it's on a log device
730 * which may have been removed or faulted prior to loading this
731 * pool. If so, there's no point in checking the rest of the log
732 * as its content should have already been synced to the pool.
734 if (!BP_IS_HOLE(bp)) {
736 boolean_t valid = B_TRUE;
738 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
739 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
740 if (vd->vdev_islog && vdev_is_dead(vd))
741 valid = vdev_log_state_valid(vd);
742 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
745 dmu_objset_rele(os, FTAG);
751 * Because tx == NULL, zil_claim_log_block() will not actually claim
752 * any blocks, but just determine whether it is possible to do so.
753 * In addition to checking the log chain, zil_claim_log_block()
754 * will invoke zio_claim() with a done func of spa_claim_notify(),
755 * which will update spa_max_claim_txg. See spa_load() for details.
757 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
758 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
760 dmu_objset_rele(os, FTAG);
762 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
766 zil_vdev_compare(const void *x1, const void *x2)
768 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
769 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
780 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
782 avl_tree_t *t = &zilog->zl_vdev_tree;
784 zil_vdev_node_t *zv, zvsearch;
785 int ndvas = BP_GET_NDVAS(bp);
788 if (zfs_nocacheflush)
791 ASSERT(zilog->zl_writer);
794 * Even though we're zl_writer, we still need a lock because the
795 * zl_get_data() callbacks may have dmu_sync() done callbacks
796 * that will run concurrently.
798 mutex_enter(&zilog->zl_vdev_lock);
799 for (i = 0; i < ndvas; i++) {
800 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
801 if (avl_find(t, &zvsearch, &where) == NULL) {
802 zv = kmem_alloc(sizeof (*zv), KM_PUSHPAGE);
803 zv->zv_vdev = zvsearch.zv_vdev;
804 avl_insert(t, zv, where);
807 mutex_exit(&zilog->zl_vdev_lock);
811 zil_flush_vdevs(zilog_t *zilog)
813 spa_t *spa = zilog->zl_spa;
814 avl_tree_t *t = &zilog->zl_vdev_tree;
819 ASSERT(zilog->zl_writer);
822 * We don't need zl_vdev_lock here because we're the zl_writer,
823 * and all zl_get_data() callbacks are done.
825 if (avl_numnodes(t) == 0)
828 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
830 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
832 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
833 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
836 kmem_free(zv, sizeof (*zv));
840 * Wait for all the flushes to complete. Not all devices actually
841 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
843 (void) zio_wait(zio);
845 spa_config_exit(spa, SCL_STATE, FTAG);
849 * Function called when a log block write completes
852 zil_lwb_write_done(zio_t *zio)
854 lwb_t *lwb = zio->io_private;
855 zilog_t *zilog = lwb->lwb_zilog;
856 dmu_tx_t *tx = lwb->lwb_tx;
858 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
859 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
860 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
861 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
862 ASSERT(!BP_IS_GANG(zio->io_bp));
863 ASSERT(!BP_IS_HOLE(zio->io_bp));
864 ASSERT(zio->io_bp->blk_fill == 0);
867 * Ensure the lwb buffer pointer is cleared before releasing
868 * the txg. If we have had an allocation failure and
869 * the txg is waiting to sync then we want want zil_sync()
870 * to remove the lwb so that it's not picked up as the next new
871 * one in zil_commit_writer(). zil_sync() will only remove
872 * the lwb if lwb_buf is null.
874 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
875 mutex_enter(&zilog->zl_lock);
877 lwb->lwb_fastwrite = FALSE;
880 mutex_exit(&zilog->zl_lock);
883 * Now that we've written this log block, we have a stable pointer
884 * to the next block in the chain, so it's OK to let the txg in
885 * which we allocated the next block sync.
891 * Initialize the io for a log block.
894 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
898 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
899 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
900 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
902 if (zilog->zl_root_zio == NULL) {
903 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
907 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
908 mutex_enter(&zilog->zl_lock);
909 if (lwb->lwb_zio == NULL) {
910 if (!lwb->lwb_fastwrite) {
911 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
912 lwb->lwb_fastwrite = 1;
914 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
915 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
916 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
917 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
918 ZIO_FLAG_FASTWRITE, &zb);
920 mutex_exit(&zilog->zl_lock);
924 * Define a limited set of intent log block sizes.
925 * These must be a multiple of 4KB. Note only the amount used (again
926 * aligned to 4KB) actually gets written. However, we can't always just
927 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
929 uint64_t zil_block_buckets[] = {
930 4096, /* non TX_WRITE */
931 8192+4096, /* data base */
932 32*1024 + 4096, /* NFS writes */
937 * Use the slog as long as the current commit size is less than the
938 * limit or the total list size is less than 2X the limit. Limit
939 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
941 unsigned long zil_slog_limit = 1024 * 1024;
942 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
943 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
946 * Start a log block write and advance to the next log block.
947 * Calls are serialized.
950 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
954 spa_t *spa = zilog->zl_spa;
958 uint64_t zil_blksz, wsz;
962 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
963 zilc = (zil_chain_t *)lwb->lwb_buf;
964 bp = &zilc->zc_next_blk;
966 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
967 bp = &zilc->zc_next_blk;
970 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
973 * Allocate the next block and save its address in this block
974 * before writing it in order to establish the log chain.
975 * Note that if the allocation of nlwb synced before we wrote
976 * the block that points at it (lwb), we'd leak it if we crashed.
977 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
978 * We dirty the dataset to ensure that zil_sync() will be called
979 * to clean up in the event of allocation failure or I/O failure.
981 tx = dmu_tx_create(zilog->zl_os);
982 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
983 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
984 txg = dmu_tx_get_txg(tx);
989 * Log blocks are pre-allocated. Here we select the size of the next
990 * block, based on size used in the last block.
991 * - first find the smallest bucket that will fit the block from a
992 * limited set of block sizes. This is because it's faster to write
993 * blocks allocated from the same metaslab as they are adjacent or
995 * - next find the maximum from the new suggested size and an array of
996 * previous sizes. This lessens a picket fence effect of wrongly
997 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1000 * Note we only write what is used, but we can't just allocate
1001 * the maximum block size because we can exhaust the available
1004 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1005 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1007 zil_blksz = zil_block_buckets[i];
1008 if (zil_blksz == UINT64_MAX)
1009 zil_blksz = SPA_MAXBLOCKSIZE;
1010 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1011 for (i = 0; i < ZIL_PREV_BLKS; i++)
1012 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1013 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1016 use_slog = USE_SLOG(zilog);
1017 error = zio_alloc_zil(spa, txg, bp, zil_blksz, USE_SLOG(zilog));
1020 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
1021 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
1025 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
1026 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
1029 ASSERT3U(bp->blk_birth, ==, txg);
1030 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1031 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1034 * Allocate a new log write buffer (lwb).
1036 nlwb = zil_alloc_lwb(zilog, bp, txg, TRUE);
1038 /* Record the block for later vdev flushing */
1039 zil_add_block(zilog, &lwb->lwb_blk);
1042 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1043 /* For Slim ZIL only write what is used. */
1044 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1045 ASSERT3U(wsz, <=, lwb->lwb_sz);
1046 zio_shrink(lwb->lwb_zio, wsz);
1053 zilc->zc_nused = lwb->lwb_nused;
1054 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1057 * clear unused data for security
1059 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1061 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1064 * If there was an allocation failure then nlwb will be null which
1065 * forces a txg_wait_synced().
1071 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1073 lr_t *lrc = &itx->itx_lr; /* common log record */
1074 lr_write_t *lrw = (lr_write_t *)lrc;
1076 uint64_t txg = lrc->lrc_txg;
1077 uint64_t reclen = lrc->lrc_reclen;
1083 ASSERT(lwb->lwb_buf != NULL);
1084 ASSERT(zilog_is_dirty(zilog) ||
1085 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1087 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1088 dlen = P2ROUNDUP_TYPED(
1089 lrw->lr_length, sizeof (uint64_t), uint64_t);
1091 zilog->zl_cur_used += (reclen + dlen);
1093 zil_lwb_write_init(zilog, lwb);
1096 * If this record won't fit in the current log block, start a new one.
1098 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1099 lwb = zil_lwb_write_start(zilog, lwb);
1102 zil_lwb_write_init(zilog, lwb);
1103 ASSERT(LWB_EMPTY(lwb));
1104 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1105 txg_wait_synced(zilog->zl_dmu_pool, txg);
1110 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1111 bcopy(lrc, lr_buf, reclen);
1112 lrc = (lr_t *)lr_buf;
1113 lrw = (lr_write_t *)lrc;
1115 ZIL_STAT_BUMP(zil_itx_count);
1118 * If it's a write, fetch the data or get its blkptr as appropriate.
1120 if (lrc->lrc_txtype == TX_WRITE) {
1121 if (txg > spa_freeze_txg(zilog->zl_spa))
1122 txg_wait_synced(zilog->zl_dmu_pool, txg);
1123 if (itx->itx_wr_state == WR_COPIED) {
1124 ZIL_STAT_BUMP(zil_itx_copied_count);
1125 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1131 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1132 dbuf = lr_buf + reclen;
1133 lrw->lr_common.lrc_reclen += dlen;
1134 ZIL_STAT_BUMP(zil_itx_needcopy_count);
1135 ZIL_STAT_INCR(zil_itx_needcopy_bytes, lrw->lr_length);
1137 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1139 ZIL_STAT_BUMP(zil_itx_indirect_count);
1140 ZIL_STAT_INCR(zil_itx_indirect_bytes, lrw->lr_length);
1142 error = zilog->zl_get_data(
1143 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1145 txg_wait_synced(zilog->zl_dmu_pool, txg);
1149 ASSERT(error == ENOENT || error == EEXIST ||
1157 * We're actually making an entry, so update lrc_seq to be the
1158 * log record sequence number. Note that this is generally not
1159 * equal to the itx sequence number because not all transactions
1160 * are synchronous, and sometimes spa_sync() gets there first.
1162 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1163 lwb->lwb_nused += reclen + dlen;
1164 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1165 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1166 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1172 zil_itx_create(uint64_t txtype, size_t lrsize)
1176 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1178 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize,
1179 KM_PUSHPAGE | KM_NODEBUG);
1180 itx->itx_lr.lrc_txtype = txtype;
1181 itx->itx_lr.lrc_reclen = lrsize;
1182 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1183 itx->itx_lr.lrc_seq = 0; /* defensive */
1184 itx->itx_sync = B_TRUE; /* default is synchronous */
1190 zil_itx_destroy(itx_t *itx)
1192 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1196 * Free up the sync and async itxs. The itxs_t has already been detached
1197 * so no locks are needed.
1200 zil_itxg_clean(itxs_t *itxs)
1206 itx_async_node_t *ian;
1208 list = &itxs->i_sync_list;
1209 while ((itx = list_head(list)) != NULL) {
1210 list_remove(list, itx);
1211 kmem_free(itx, offsetof(itx_t, itx_lr) +
1212 itx->itx_lr.lrc_reclen);
1216 t = &itxs->i_async_tree;
1217 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1218 list = &ian->ia_list;
1219 while ((itx = list_head(list)) != NULL) {
1220 list_remove(list, itx);
1221 kmem_free(itx, offsetof(itx_t, itx_lr) +
1222 itx->itx_lr.lrc_reclen);
1225 kmem_free(ian, sizeof (itx_async_node_t));
1229 kmem_free(itxs, sizeof (itxs_t));
1233 zil_aitx_compare(const void *x1, const void *x2)
1235 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1236 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1247 * Remove all async itx with the given oid.
1250 zil_remove_async(zilog_t *zilog, uint64_t oid)
1253 itx_async_node_t *ian;
1260 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1262 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1265 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1267 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1268 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1270 mutex_enter(&itxg->itxg_lock);
1271 if (itxg->itxg_txg != txg) {
1272 mutex_exit(&itxg->itxg_lock);
1277 * Locate the object node and append its list.
1279 t = &itxg->itxg_itxs->i_async_tree;
1280 ian = avl_find(t, &oid, &where);
1282 list_move_tail(&clean_list, &ian->ia_list);
1283 mutex_exit(&itxg->itxg_lock);
1285 while ((itx = list_head(&clean_list)) != NULL) {
1286 list_remove(&clean_list, itx);
1287 kmem_free(itx, offsetof(itx_t, itx_lr) +
1288 itx->itx_lr.lrc_reclen);
1290 list_destroy(&clean_list);
1294 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1298 itxs_t *itxs, *clean = NULL;
1301 * Object ids can be re-instantiated in the next txg so
1302 * remove any async transactions to avoid future leaks.
1303 * This can happen if a fsync occurs on the re-instantiated
1304 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1305 * the new file data and flushes a write record for the old object.
1307 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1308 zil_remove_async(zilog, itx->itx_oid);
1311 * Ensure the data of a renamed file is committed before the rename.
1313 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1314 zil_async_to_sync(zilog, itx->itx_oid);
1316 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1319 txg = dmu_tx_get_txg(tx);
1321 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1322 mutex_enter(&itxg->itxg_lock);
1323 itxs = itxg->itxg_itxs;
1324 if (itxg->itxg_txg != txg) {
1327 * The zil_clean callback hasn't got around to cleaning
1328 * this itxg. Save the itxs for release below.
1329 * This should be rare.
1331 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1333 clean = itxg->itxg_itxs;
1335 ASSERT(itxg->itxg_sod == 0);
1336 itxg->itxg_txg = txg;
1337 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_PUSHPAGE);
1339 list_create(&itxs->i_sync_list, sizeof (itx_t),
1340 offsetof(itx_t, itx_node));
1341 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1342 sizeof (itx_async_node_t),
1343 offsetof(itx_async_node_t, ia_node));
1345 if (itx->itx_sync) {
1346 list_insert_tail(&itxs->i_sync_list, itx);
1347 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1348 itxg->itxg_sod += itx->itx_sod;
1350 avl_tree_t *t = &itxs->i_async_tree;
1351 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1352 itx_async_node_t *ian;
1355 ian = avl_find(t, &foid, &where);
1357 ian = kmem_alloc(sizeof (itx_async_node_t), KM_PUSHPAGE);
1358 list_create(&ian->ia_list, sizeof (itx_t),
1359 offsetof(itx_t, itx_node));
1360 ian->ia_foid = foid;
1361 avl_insert(t, ian, where);
1363 list_insert_tail(&ian->ia_list, itx);
1366 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1367 zilog_dirty(zilog, txg);
1368 mutex_exit(&itxg->itxg_lock);
1370 /* Release the old itxs now we've dropped the lock */
1372 zil_itxg_clean(clean);
1376 * If there are any in-memory intent log transactions which have now been
1377 * synced then start up a taskq to free them. We should only do this after we
1378 * have written out the uberblocks (i.e. txg has been comitted) so that
1379 * don't inadvertently clean out in-memory log records that would be required
1383 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1385 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1388 mutex_enter(&itxg->itxg_lock);
1389 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1390 mutex_exit(&itxg->itxg_lock);
1393 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1394 ASSERT(itxg->itxg_txg != 0);
1395 ASSERT(zilog->zl_clean_taskq != NULL);
1396 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1398 clean_me = itxg->itxg_itxs;
1399 itxg->itxg_itxs = NULL;
1401 mutex_exit(&itxg->itxg_lock);
1403 * Preferably start a task queue to free up the old itxs but
1404 * if taskq_dispatch can't allocate resources to do that then
1405 * free it in-line. This should be rare. Note, using TQ_SLEEP
1406 * created a bad performance problem.
1408 if (taskq_dispatch(zilog->zl_clean_taskq,
1409 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1410 zil_itxg_clean(clean_me);
1414 * Get the list of itxs to commit into zl_itx_commit_list.
1417 zil_get_commit_list(zilog_t *zilog)
1420 list_t *commit_list = &zilog->zl_itx_commit_list;
1421 uint64_t push_sod = 0;
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);
1437 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1438 push_sod += itxg->itxg_sod;
1441 mutex_exit(&itxg->itxg_lock);
1443 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1447 * Move the async itxs for a specified object to commit into sync lists.
1450 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1453 itx_async_node_t *ian;
1457 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1460 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1462 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1463 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1465 mutex_enter(&itxg->itxg_lock);
1466 if (itxg->itxg_txg != txg) {
1467 mutex_exit(&itxg->itxg_lock);
1472 * If a foid is specified then find that node and append its
1473 * list. Otherwise walk the tree appending all the lists
1474 * to the sync list. We add to the end rather than the
1475 * beginning to ensure the create has happened.
1477 t = &itxg->itxg_itxs->i_async_tree;
1479 ian = avl_find(t, &foid, &where);
1481 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1485 void *cookie = NULL;
1487 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1488 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1490 list_destroy(&ian->ia_list);
1491 kmem_free(ian, sizeof (itx_async_node_t));
1494 mutex_exit(&itxg->itxg_lock);
1499 zil_commit_writer(zilog_t *zilog)
1504 spa_t *spa = zilog->zl_spa;
1507 ASSERT(zilog->zl_root_zio == NULL);
1509 mutex_exit(&zilog->zl_lock);
1511 zil_get_commit_list(zilog);
1514 * Return if there's nothing to commit before we dirty the fs by
1515 * calling zil_create().
1517 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1518 mutex_enter(&zilog->zl_lock);
1522 if (zilog->zl_suspend) {
1525 lwb = list_tail(&zilog->zl_lwb_list);
1527 lwb = zil_create(zilog);
1530 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1531 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1532 txg = itx->itx_lr.lrc_txg;
1535 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1536 lwb = zil_lwb_commit(zilog, itx, lwb);
1537 list_remove(&zilog->zl_itx_commit_list, itx);
1538 kmem_free(itx, offsetof(itx_t, itx_lr)
1539 + itx->itx_lr.lrc_reclen);
1541 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1543 /* write the last block out */
1544 if (lwb != NULL && lwb->lwb_zio != NULL)
1545 lwb = zil_lwb_write_start(zilog, lwb);
1547 zilog->zl_cur_used = 0;
1550 * Wait if necessary for the log blocks to be on stable storage.
1552 if (zilog->zl_root_zio) {
1553 error = zio_wait(zilog->zl_root_zio);
1554 zilog->zl_root_zio = NULL;
1555 zil_flush_vdevs(zilog);
1558 if (error || lwb == NULL)
1559 txg_wait_synced(zilog->zl_dmu_pool, 0);
1561 mutex_enter(&zilog->zl_lock);
1564 * Remember the highest committed log sequence number for ztest.
1565 * We only update this value when all the log writes succeeded,
1566 * because ztest wants to ASSERT that it got the whole log chain.
1568 if (error == 0 && lwb != NULL)
1569 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1573 * Commit zfs transactions to stable storage.
1574 * If foid is 0 push out all transactions, otherwise push only those
1575 * for that object or might reference that object.
1577 * itxs are committed in batches. In a heavily stressed zil there will be
1578 * a commit writer thread who is writing out a bunch of itxs to the log
1579 * for a set of committing threads (cthreads) in the same batch as the writer.
1580 * Those cthreads are all waiting on the same cv for that batch.
1582 * There will also be a different and growing batch of threads that are
1583 * waiting to commit (qthreads). When the committing batch completes
1584 * a transition occurs such that the cthreads exit and the qthreads become
1585 * cthreads. One of the new cthreads becomes the writer thread for the
1586 * batch. Any new threads arriving become new qthreads.
1588 * Only 2 condition variables are needed and there's no transition
1589 * between the two cvs needed. They just flip-flop between qthreads
1592 * Using this scheme we can efficiently wakeup up only those threads
1593 * that have been committed.
1596 zil_commit(zilog_t *zilog, uint64_t foid)
1600 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1603 ZIL_STAT_BUMP(zil_commit_count);
1605 /* move the async itxs for the foid to the sync queues */
1606 zil_async_to_sync(zilog, foid);
1608 mutex_enter(&zilog->zl_lock);
1609 mybatch = zilog->zl_next_batch;
1610 while (zilog->zl_writer) {
1611 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1612 if (mybatch <= zilog->zl_com_batch) {
1613 mutex_exit(&zilog->zl_lock);
1618 zilog->zl_next_batch++;
1619 zilog->zl_writer = B_TRUE;
1620 ZIL_STAT_BUMP(zil_commit_writer_count);
1621 zil_commit_writer(zilog);
1622 zilog->zl_com_batch = mybatch;
1623 zilog->zl_writer = B_FALSE;
1625 /* wake up one thread to become the next writer */
1626 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1628 /* wake up all threads waiting for this batch to be committed */
1629 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1631 mutex_exit(&zilog->zl_lock);
1635 * Called in syncing context to free committed log blocks and update log header.
1638 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1640 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1641 uint64_t txg = dmu_tx_get_txg(tx);
1642 spa_t *spa = zilog->zl_spa;
1643 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1647 * We don't zero out zl_destroy_txg, so make sure we don't try
1648 * to destroy it twice.
1650 if (spa_sync_pass(spa) != 1)
1653 mutex_enter(&zilog->zl_lock);
1655 ASSERT(zilog->zl_stop_sync == 0);
1657 if (*replayed_seq != 0) {
1658 ASSERT(zh->zh_replay_seq < *replayed_seq);
1659 zh->zh_replay_seq = *replayed_seq;
1663 if (zilog->zl_destroy_txg == txg) {
1664 blkptr_t blk = zh->zh_log;
1666 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1668 bzero(zh, sizeof (zil_header_t));
1669 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1671 if (zilog->zl_keep_first) {
1673 * If this block was part of log chain that couldn't
1674 * be claimed because a device was missing during
1675 * zil_claim(), but that device later returns,
1676 * then this block could erroneously appear valid.
1677 * To guard against this, assign a new GUID to the new
1678 * log chain so it doesn't matter what blk points to.
1680 zil_init_log_chain(zilog, &blk);
1685 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1686 zh->zh_log = lwb->lwb_blk;
1687 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1690 ASSERT(lwb->lwb_zio == NULL);
1692 list_remove(&zilog->zl_lwb_list, lwb);
1693 zio_free_zil(spa, txg, &lwb->lwb_blk);
1694 kmem_cache_free(zil_lwb_cache, lwb);
1697 * If we don't have anything left in the lwb list then
1698 * we've had an allocation failure and we need to zero
1699 * out the zil_header blkptr so that we don't end
1700 * up freeing the same block twice.
1702 if (list_head(&zilog->zl_lwb_list) == NULL)
1703 BP_ZERO(&zh->zh_log);
1707 * Remove fastwrite on any blocks that have been pre-allocated for
1708 * the next commit. This prevents fastwrite counter pollution by
1709 * unused, long-lived LWBs.
1711 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
1712 if (lwb->lwb_fastwrite && !lwb->lwb_zio) {
1713 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1714 lwb->lwb_fastwrite = 0;
1718 mutex_exit(&zilog->zl_lock);
1724 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1725 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1727 zil_ksp = kstat_create("zfs", 0, "zil", "misc",
1728 KSTAT_TYPE_NAMED, sizeof(zil_stats) / sizeof(kstat_named_t),
1729 KSTAT_FLAG_VIRTUAL);
1731 if (zil_ksp != NULL) {
1732 zil_ksp->ks_data = &zil_stats;
1733 kstat_install(zil_ksp);
1740 kmem_cache_destroy(zil_lwb_cache);
1742 if (zil_ksp != NULL) {
1743 kstat_delete(zil_ksp);
1749 zil_set_sync(zilog_t *zilog, uint64_t sync)
1751 zilog->zl_sync = sync;
1755 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1757 zilog->zl_logbias = logbias;
1761 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1766 zilog = kmem_zalloc(sizeof (zilog_t), KM_PUSHPAGE);
1768 zilog->zl_header = zh_phys;
1770 zilog->zl_spa = dmu_objset_spa(os);
1771 zilog->zl_dmu_pool = dmu_objset_pool(os);
1772 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1773 zilog->zl_logbias = dmu_objset_logbias(os);
1774 zilog->zl_sync = dmu_objset_syncprop(os);
1775 zilog->zl_next_batch = 1;
1777 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1779 for (i = 0; i < TXG_SIZE; i++) {
1780 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1781 MUTEX_DEFAULT, NULL);
1784 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1785 offsetof(lwb_t, lwb_node));
1787 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1788 offsetof(itx_t, itx_node));
1790 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1792 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1793 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1795 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1796 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1797 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1798 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1804 zil_free(zilog_t *zilog)
1808 zilog->zl_stop_sync = 1;
1810 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1811 list_destroy(&zilog->zl_lwb_list);
1813 avl_destroy(&zilog->zl_vdev_tree);
1814 mutex_destroy(&zilog->zl_vdev_lock);
1816 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1817 list_destroy(&zilog->zl_itx_commit_list);
1819 for (i = 0; i < TXG_SIZE; i++) {
1821 * It's possible for an itx to be generated that doesn't dirty
1822 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1823 * callback to remove the entry. We remove those here.
1825 * Also free up the ziltest itxs.
1827 if (zilog->zl_itxg[i].itxg_itxs)
1828 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1829 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1832 mutex_destroy(&zilog->zl_lock);
1834 cv_destroy(&zilog->zl_cv_writer);
1835 cv_destroy(&zilog->zl_cv_suspend);
1836 cv_destroy(&zilog->zl_cv_batch[0]);
1837 cv_destroy(&zilog->zl_cv_batch[1]);
1839 kmem_free(zilog, sizeof (zilog_t));
1843 * Open an intent log.
1846 zil_open(objset_t *os, zil_get_data_t *get_data)
1848 zilog_t *zilog = dmu_objset_zil(os);
1850 ASSERT(zilog->zl_clean_taskq == NULL);
1851 ASSERT(zilog->zl_get_data == NULL);
1852 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1854 zilog->zl_get_data = get_data;
1855 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1856 2, 2, TASKQ_PREPOPULATE);
1862 * Close an intent log.
1865 zil_close(zilog_t *zilog)
1870 zil_commit(zilog, 0); /* commit all itx */
1873 * The lwb_max_txg for the stubby lwb will reflect the last activity
1874 * for the zil. After a txg_wait_synced() on the txg we know all the
1875 * callbacks have occurred that may clean the zil. Only then can we
1876 * destroy the zl_clean_taskq.
1878 mutex_enter(&zilog->zl_lock);
1879 lwb = list_tail(&zilog->zl_lwb_list);
1881 txg = lwb->lwb_max_txg;
1882 mutex_exit(&zilog->zl_lock);
1884 txg_wait_synced(zilog->zl_dmu_pool, txg);
1885 ASSERT(!zilog_is_dirty(zilog));
1887 taskq_destroy(zilog->zl_clean_taskq);
1888 zilog->zl_clean_taskq = NULL;
1889 zilog->zl_get_data = NULL;
1892 * We should have only one LWB left on the list; remove it now.
1894 mutex_enter(&zilog->zl_lock);
1895 lwb = list_head(&zilog->zl_lwb_list);
1897 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1898 ASSERT(lwb->lwb_zio == NULL);
1899 if (lwb->lwb_fastwrite)
1900 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1901 list_remove(&zilog->zl_lwb_list, lwb);
1902 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1903 kmem_cache_free(zil_lwb_cache, lwb);
1905 mutex_exit(&zilog->zl_lock);
1909 * Suspend an intent log. While in suspended mode, we still honor
1910 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1911 * We suspend the log briefly when taking a snapshot so that the snapshot
1912 * contains all the data it's supposed to, and has an empty intent log.
1915 zil_suspend(zilog_t *zilog)
1917 const zil_header_t *zh = zilog->zl_header;
1919 mutex_enter(&zilog->zl_lock);
1920 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1921 mutex_exit(&zilog->zl_lock);
1924 if (zilog->zl_suspend++ != 0) {
1926 * Someone else already began a suspend.
1927 * Just wait for them to finish.
1929 while (zilog->zl_suspending)
1930 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1931 mutex_exit(&zilog->zl_lock);
1934 zilog->zl_suspending = B_TRUE;
1935 mutex_exit(&zilog->zl_lock);
1937 zil_commit(zilog, 0);
1939 zil_destroy(zilog, B_FALSE);
1941 mutex_enter(&zilog->zl_lock);
1942 zilog->zl_suspending = B_FALSE;
1943 cv_broadcast(&zilog->zl_cv_suspend);
1944 mutex_exit(&zilog->zl_lock);
1950 zil_resume(zilog_t *zilog)
1952 mutex_enter(&zilog->zl_lock);
1953 ASSERT(zilog->zl_suspend != 0);
1954 zilog->zl_suspend--;
1955 mutex_exit(&zilog->zl_lock);
1958 typedef struct zil_replay_arg {
1959 zil_replay_func_t *zr_replay;
1961 boolean_t zr_byteswap;
1966 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1968 char name[MAXNAMELEN];
1970 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1972 dmu_objset_name(zilog->zl_os, name);
1974 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1975 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1976 (u_longlong_t)lr->lrc_seq,
1977 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1978 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1984 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1986 zil_replay_arg_t *zr = zra;
1987 const zil_header_t *zh = zilog->zl_header;
1988 uint64_t reclen = lr->lrc_reclen;
1989 uint64_t txtype = lr->lrc_txtype;
1992 zilog->zl_replaying_seq = lr->lrc_seq;
1994 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1997 if (lr->lrc_txg < claim_txg) /* already committed */
2000 /* Strip case-insensitive bit, still present in log record */
2003 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2004 return (zil_replay_error(zilog, lr, EINVAL));
2007 * If this record type can be logged out of order, the object
2008 * (lr_foid) may no longer exist. That's legitimate, not an error.
2010 if (TX_OOO(txtype)) {
2011 error = dmu_object_info(zilog->zl_os,
2012 ((lr_ooo_t *)lr)->lr_foid, NULL);
2013 if (error == ENOENT || error == EEXIST)
2018 * Make a copy of the data so we can revise and extend it.
2020 bcopy(lr, zr->zr_lr, reclen);
2023 * If this is a TX_WRITE with a blkptr, suck in the data.
2025 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2026 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2027 zr->zr_lr + reclen);
2029 return (zil_replay_error(zilog, lr, error));
2033 * The log block containing this lr may have been byteswapped
2034 * so that we can easily examine common fields like lrc_txtype.
2035 * However, the log is a mix of different record types, and only the
2036 * replay vectors know how to byteswap their records. Therefore, if
2037 * the lr was byteswapped, undo it before invoking the replay vector.
2039 if (zr->zr_byteswap)
2040 byteswap_uint64_array(zr->zr_lr, reclen);
2043 * We must now do two things atomically: replay this log record,
2044 * and update the log header sequence number to reflect the fact that
2045 * we did so. At the end of each replay function the sequence number
2046 * is updated if we are in replay mode.
2048 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2051 * The DMU's dnode layer doesn't see removes until the txg
2052 * commits, so a subsequent claim can spuriously fail with
2053 * EEXIST. So if we receive any error we try syncing out
2054 * any removes then retry the transaction. Note that we
2055 * specify B_FALSE for byteswap now, so we don't do it twice.
2057 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2058 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2060 return (zil_replay_error(zilog, lr, error));
2067 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2069 zilog->zl_replay_blks++;
2075 * If this dataset has a non-empty intent log, replay it and destroy it.
2078 zil_replay(objset_t *os, void *arg, zil_replay_func_t replay_func[TX_MAX_TYPE])
2080 zilog_t *zilog = dmu_objset_zil(os);
2081 const zil_header_t *zh = zilog->zl_header;
2082 zil_replay_arg_t zr;
2084 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2085 zil_destroy(zilog, B_TRUE);
2089 zr.zr_replay = replay_func;
2091 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2092 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_PUSHPAGE);
2095 * Wait for in-progress removes to sync before starting replay.
2097 txg_wait_synced(zilog->zl_dmu_pool, 0);
2099 zilog->zl_replay = B_TRUE;
2100 zilog->zl_replay_time = ddi_get_lbolt();
2101 ASSERT(zilog->zl_replay_blks == 0);
2102 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2104 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2106 zil_destroy(zilog, B_FALSE);
2107 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2108 zilog->zl_replay = B_FALSE;
2112 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2114 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2117 if (zilog->zl_replay) {
2118 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2119 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2120 zilog->zl_replaying_seq;
2129 zil_vdev_offline(const char *osname, void *arg)
2135 error = dmu_objset_hold(osname, FTAG, &os);
2139 zilog = dmu_objset_zil(os);
2140 if (zil_suspend(zilog) != 0)
2144 dmu_objset_rele(os, FTAG);
2148 #if defined(_KERNEL) && defined(HAVE_SPL)
2149 module_param(zil_replay_disable, int, 0644);
2150 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
2152 module_param(zfs_nocacheflush, int, 0644);
2153 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");
2155 module_param(zil_slog_limit, ulong, 0644);
2156 MODULE_PARM_DESC(zil_slog_limit, "Max commit bytes to separate log device");