4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
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 * Create an on-disk intent log.
486 zil_create(zilog_t *zilog)
488 const zil_header_t *zh = zilog->zl_header;
494 boolean_t fastwrite = FALSE;
497 * Wait for any previous destroy to complete.
499 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
501 ASSERT(zh->zh_claim_txg == 0);
502 ASSERT(zh->zh_replay_seq == 0);
507 * Allocate an initial log block if:
508 * - there isn't one already
509 * - the existing block is the wrong endianess
511 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
512 tx = dmu_tx_create(zilog->zl_os);
513 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
514 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
515 txg = dmu_tx_get_txg(tx);
517 if (!BP_IS_HOLE(&blk)) {
518 zio_free_zil(zilog->zl_spa, txg, &blk);
522 error = zio_alloc_zil(zilog->zl_spa, txg, &blk,
523 ZIL_MIN_BLKSZ, B_TRUE);
527 zil_init_log_chain(zilog, &blk);
531 * Allocate a log write buffer (lwb) for the first log block.
534 lwb = zil_alloc_lwb(zilog, &blk, txg, fastwrite);
537 * If we just allocated the first log block, commit our transaction
538 * and wait for zil_sync() to stuff the block poiner into zh_log.
539 * (zh is part of the MOS, so we cannot modify it in open context.)
543 txg_wait_synced(zilog->zl_dmu_pool, txg);
546 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
552 * In one tx, free all log blocks and clear the log header.
553 * If keep_first is set, then we're replaying a log with no content.
554 * We want to keep the first block, however, so that the first
555 * synchronous transaction doesn't require a txg_wait_synced()
556 * in zil_create(). We don't need to txg_wait_synced() here either
557 * when keep_first is set, because both zil_create() and zil_destroy()
558 * will wait for any in-progress destroys to complete.
561 zil_destroy(zilog_t *zilog, boolean_t keep_first)
563 const zil_header_t *zh = zilog->zl_header;
569 * Wait for any previous destroy to complete.
571 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
573 zilog->zl_old_header = *zh; /* debugging aid */
575 if (BP_IS_HOLE(&zh->zh_log))
578 tx = dmu_tx_create(zilog->zl_os);
579 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
580 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
581 txg = dmu_tx_get_txg(tx);
583 mutex_enter(&zilog->zl_lock);
585 ASSERT3U(zilog->zl_destroy_txg, <, txg);
586 zilog->zl_destroy_txg = txg;
587 zilog->zl_keep_first = keep_first;
589 if (!list_is_empty(&zilog->zl_lwb_list)) {
590 ASSERT(zh->zh_claim_txg == 0);
592 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
593 ASSERT(lwb->lwb_zio == NULL);
594 if (lwb->lwb_fastwrite)
595 metaslab_fastwrite_unmark(zilog->zl_spa,
597 list_remove(&zilog->zl_lwb_list, lwb);
598 if (lwb->lwb_buf != NULL)
599 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
600 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
601 kmem_cache_free(zil_lwb_cache, lwb);
603 } else if (!keep_first) {
604 (void) zil_parse(zilog, zil_free_log_block,
605 zil_free_log_record, tx, zh->zh_claim_txg);
607 mutex_exit(&zilog->zl_lock);
613 zil_claim(const char *osname, void *txarg)
615 dmu_tx_t *tx = txarg;
616 uint64_t first_txg = dmu_tx_get_txg(tx);
622 error = dmu_objset_hold(osname, FTAG, &os);
624 cmn_err(CE_WARN, "can't open objset for %s", osname);
628 zilog = dmu_objset_zil(os);
629 zh = zil_header_in_syncing_context(zilog);
631 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
632 if (!BP_IS_HOLE(&zh->zh_log))
633 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
634 BP_ZERO(&zh->zh_log);
635 dsl_dataset_dirty(dmu_objset_ds(os), tx);
636 dmu_objset_rele(os, FTAG);
641 * Claim all log blocks if we haven't already done so, and remember
642 * the highest claimed sequence number. This ensures that if we can
643 * read only part of the log now (e.g. due to a missing device),
644 * but we can read the entire log later, we will not try to replay
645 * or destroy beyond the last block we successfully claimed.
647 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
648 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
649 (void) zil_parse(zilog, zil_claim_log_block,
650 zil_claim_log_record, tx, first_txg);
651 zh->zh_claim_txg = first_txg;
652 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
653 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
654 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
655 zh->zh_flags |= ZIL_REPLAY_NEEDED;
656 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
657 dsl_dataset_dirty(dmu_objset_ds(os), tx);
660 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
661 dmu_objset_rele(os, FTAG);
666 * Check the log by walking the log chain.
667 * Checksum errors are ok as they indicate the end of the chain.
668 * Any other error (no device or read failure) returns an error.
671 zil_check_log_chain(const char *osname, void *tx)
680 error = dmu_objset_hold(osname, FTAG, &os);
682 cmn_err(CE_WARN, "can't open objset for %s", osname);
686 zilog = dmu_objset_zil(os);
687 bp = (blkptr_t *)&zilog->zl_header->zh_log;
690 * Check the first block and determine if it's on a log device
691 * which may have been removed or faulted prior to loading this
692 * pool. If so, there's no point in checking the rest of the log
693 * as its content should have already been synced to the pool.
695 if (!BP_IS_HOLE(bp)) {
697 boolean_t valid = B_TRUE;
699 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
700 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
701 if (vd->vdev_islog && vdev_is_dead(vd))
702 valid = vdev_log_state_valid(vd);
703 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
706 dmu_objset_rele(os, FTAG);
712 * Because tx == NULL, zil_claim_log_block() will not actually claim
713 * any blocks, but just determine whether it is possible to do so.
714 * In addition to checking the log chain, zil_claim_log_block()
715 * will invoke zio_claim() with a done func of spa_claim_notify(),
716 * which will update spa_max_claim_txg. See spa_load() for details.
718 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
719 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
721 dmu_objset_rele(os, FTAG);
723 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
727 zil_vdev_compare(const void *x1, const void *x2)
729 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
730 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
741 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
743 avl_tree_t *t = &zilog->zl_vdev_tree;
745 zil_vdev_node_t *zv, zvsearch;
746 int ndvas = BP_GET_NDVAS(bp);
749 if (zfs_nocacheflush)
752 ASSERT(zilog->zl_writer);
755 * Even though we're zl_writer, we still need a lock because the
756 * zl_get_data() callbacks may have dmu_sync() done callbacks
757 * that will run concurrently.
759 mutex_enter(&zilog->zl_vdev_lock);
760 for (i = 0; i < ndvas; i++) {
761 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
762 if (avl_find(t, &zvsearch, &where) == NULL) {
763 zv = kmem_alloc(sizeof (*zv), KM_PUSHPAGE);
764 zv->zv_vdev = zvsearch.zv_vdev;
765 avl_insert(t, zv, where);
768 mutex_exit(&zilog->zl_vdev_lock);
772 zil_flush_vdevs(zilog_t *zilog)
774 spa_t *spa = zilog->zl_spa;
775 avl_tree_t *t = &zilog->zl_vdev_tree;
780 ASSERT(zilog->zl_writer);
783 * We don't need zl_vdev_lock here because we're the zl_writer,
784 * and all zl_get_data() callbacks are done.
786 if (avl_numnodes(t) == 0)
789 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
791 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
793 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
794 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
797 kmem_free(zv, sizeof (*zv));
801 * Wait for all the flushes to complete. Not all devices actually
802 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
804 (void) zio_wait(zio);
806 spa_config_exit(spa, SCL_STATE, FTAG);
810 * Function called when a log block write completes
813 zil_lwb_write_done(zio_t *zio)
815 lwb_t *lwb = zio->io_private;
816 zilog_t *zilog = lwb->lwb_zilog;
817 dmu_tx_t *tx = lwb->lwb_tx;
819 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
820 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
821 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
822 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
823 ASSERT(!BP_IS_GANG(zio->io_bp));
824 ASSERT(!BP_IS_HOLE(zio->io_bp));
825 ASSERT(zio->io_bp->blk_fill == 0);
828 * Ensure the lwb buffer pointer is cleared before releasing
829 * the txg. If we have had an allocation failure and
830 * the txg is waiting to sync then we want want zil_sync()
831 * to remove the lwb so that it's not picked up as the next new
832 * one in zil_commit_writer(). zil_sync() will only remove
833 * the lwb if lwb_buf is null.
835 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
836 mutex_enter(&zilog->zl_lock);
838 lwb->lwb_fastwrite = FALSE;
841 mutex_exit(&zilog->zl_lock);
844 * Now that we've written this log block, we have a stable pointer
845 * to the next block in the chain, so it's OK to let the txg in
846 * which we allocated the next block sync.
852 * Initialize the io for a log block.
855 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
859 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
860 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
861 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
863 if (zilog->zl_root_zio == NULL) {
864 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
868 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
869 mutex_enter(&zilog->zl_lock);
870 if (lwb->lwb_zio == NULL) {
871 if (!lwb->lwb_fastwrite) {
872 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
873 lwb->lwb_fastwrite = 1;
875 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
876 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
877 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
878 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
879 ZIO_FLAG_FASTWRITE, &zb);
881 mutex_exit(&zilog->zl_lock);
885 * Define a limited set of intent log block sizes.
886 * These must be a multiple of 4KB. Note only the amount used (again
887 * aligned to 4KB) actually gets written. However, we can't always just
888 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
890 uint64_t zil_block_buckets[] = {
891 4096, /* non TX_WRITE */
892 8192+4096, /* data base */
893 32*1024 + 4096, /* NFS writes */
898 * Use the slog as long as the current commit size is less than the
899 * limit or the total list size is less than 2X the limit. Limit
900 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
902 unsigned long zil_slog_limit = 1024 * 1024;
903 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
904 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
907 * Start a log block write and advance to the next log block.
908 * Calls are serialized.
911 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
915 spa_t *spa = zilog->zl_spa;
919 uint64_t zil_blksz, wsz;
923 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
924 zilc = (zil_chain_t *)lwb->lwb_buf;
925 bp = &zilc->zc_next_blk;
927 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
928 bp = &zilc->zc_next_blk;
931 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
934 * Allocate the next block and save its address in this block
935 * before writing it in order to establish the log chain.
936 * Note that if the allocation of nlwb synced before we wrote
937 * the block that points at it (lwb), we'd leak it if we crashed.
938 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
939 * We dirty the dataset to ensure that zil_sync() will be called
940 * to clean up in the event of allocation failure or I/O failure.
942 tx = dmu_tx_create(zilog->zl_os);
943 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
944 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
945 txg = dmu_tx_get_txg(tx);
950 * Log blocks are pre-allocated. Here we select the size of the next
951 * block, based on size used in the last block.
952 * - first find the smallest bucket that will fit the block from a
953 * limited set of block sizes. This is because it's faster to write
954 * blocks allocated from the same metaslab as they are adjacent or
956 * - next find the maximum from the new suggested size and an array of
957 * previous sizes. This lessens a picket fence effect of wrongly
958 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
961 * Note we only write what is used, but we can't just allocate
962 * the maximum block size because we can exhaust the available
965 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
966 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
968 zil_blksz = zil_block_buckets[i];
969 if (zil_blksz == UINT64_MAX)
970 zil_blksz = SPA_MAXBLOCKSIZE;
971 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
972 for (i = 0; i < ZIL_PREV_BLKS; i++)
973 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
974 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
977 use_slog = USE_SLOG(zilog);
978 error = zio_alloc_zil(spa, txg, bp, zil_blksz, USE_SLOG(zilog));
981 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count);
982 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused);
986 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count);
987 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused);
990 ASSERT3U(bp->blk_birth, ==, txg);
991 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
992 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
995 * Allocate a new log write buffer (lwb).
997 nlwb = zil_alloc_lwb(zilog, bp, txg, TRUE);
999 /* Record the block for later vdev flushing */
1000 zil_add_block(zilog, &lwb->lwb_blk);
1003 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1004 /* For Slim ZIL only write what is used. */
1005 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1006 ASSERT3U(wsz, <=, lwb->lwb_sz);
1007 zio_shrink(lwb->lwb_zio, wsz);
1014 zilc->zc_nused = lwb->lwb_nused;
1015 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1018 * clear unused data for security
1020 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1022 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1025 * If there was an allocation failure then nlwb will be null which
1026 * forces a txg_wait_synced().
1032 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1034 lr_t *lrc = &itx->itx_lr; /* common log record */
1035 lr_write_t *lrw = (lr_write_t *)lrc;
1037 uint64_t txg = lrc->lrc_txg;
1038 uint64_t reclen = lrc->lrc_reclen;
1044 ASSERT(lwb->lwb_buf != NULL);
1046 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1047 dlen = P2ROUNDUP_TYPED(
1048 lrw->lr_length, sizeof (uint64_t), uint64_t);
1050 zilog->zl_cur_used += (reclen + dlen);
1052 zil_lwb_write_init(zilog, lwb);
1055 * If this record won't fit in the current log block, start a new one.
1057 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1058 lwb = zil_lwb_write_start(zilog, lwb);
1061 zil_lwb_write_init(zilog, lwb);
1062 ASSERT(LWB_EMPTY(lwb));
1063 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1064 txg_wait_synced(zilog->zl_dmu_pool, txg);
1069 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1070 bcopy(lrc, lr_buf, reclen);
1071 lrc = (lr_t *)lr_buf;
1072 lrw = (lr_write_t *)lrc;
1074 ZIL_STAT_BUMP(zil_itx_count);
1077 * If it's a write, fetch the data or get its blkptr as appropriate.
1079 if (lrc->lrc_txtype == TX_WRITE) {
1080 if (txg > spa_freeze_txg(zilog->zl_spa))
1081 txg_wait_synced(zilog->zl_dmu_pool, txg);
1082 if (itx->itx_wr_state == WR_COPIED) {
1083 ZIL_STAT_BUMP(zil_itx_copied_count);
1084 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length);
1090 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1091 dbuf = lr_buf + reclen;
1092 lrw->lr_common.lrc_reclen += dlen;
1093 ZIL_STAT_BUMP(zil_itx_needcopy_count);
1094 ZIL_STAT_INCR(zil_itx_needcopy_bytes, lrw->lr_length);
1096 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1098 ZIL_STAT_BUMP(zil_itx_indirect_count);
1099 ZIL_STAT_INCR(zil_itx_indirect_bytes, lrw->lr_length);
1101 error = zilog->zl_get_data(
1102 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1104 txg_wait_synced(zilog->zl_dmu_pool, txg);
1108 ASSERT(error == ENOENT || error == EEXIST ||
1116 * We're actually making an entry, so update lrc_seq to be the
1117 * log record sequence number. Note that this is generally not
1118 * equal to the itx sequence number because not all transactions
1119 * are synchronous, and sometimes spa_sync() gets there first.
1121 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1122 lwb->lwb_nused += reclen + dlen;
1123 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1124 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1125 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1131 zil_itx_create(uint64_t txtype, size_t lrsize)
1135 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1137 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize,
1138 KM_PUSHPAGE | KM_NODEBUG);
1139 itx->itx_lr.lrc_txtype = txtype;
1140 itx->itx_lr.lrc_reclen = lrsize;
1141 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1142 itx->itx_lr.lrc_seq = 0; /* defensive */
1143 itx->itx_sync = B_TRUE; /* default is synchronous */
1149 zil_itx_destroy(itx_t *itx)
1151 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1155 * Free up the sync and async itxs. The itxs_t has already been detached
1156 * so no locks are needed.
1159 zil_itxg_clean(itxs_t *itxs)
1165 itx_async_node_t *ian;
1167 list = &itxs->i_sync_list;
1168 while ((itx = list_head(list)) != NULL) {
1169 list_remove(list, itx);
1170 kmem_free(itx, offsetof(itx_t, itx_lr) +
1171 itx->itx_lr.lrc_reclen);
1175 t = &itxs->i_async_tree;
1176 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1177 list = &ian->ia_list;
1178 while ((itx = list_head(list)) != NULL) {
1179 list_remove(list, itx);
1180 kmem_free(itx, offsetof(itx_t, itx_lr) +
1181 itx->itx_lr.lrc_reclen);
1184 kmem_free(ian, sizeof (itx_async_node_t));
1188 kmem_free(itxs, sizeof (itxs_t));
1192 zil_aitx_compare(const void *x1, const void *x2)
1194 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1195 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1206 * Remove all async itx with the given oid.
1209 zil_remove_async(zilog_t *zilog, uint64_t oid)
1212 itx_async_node_t *ian;
1219 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1221 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1224 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1226 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1227 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1229 mutex_enter(&itxg->itxg_lock);
1230 if (itxg->itxg_txg != txg) {
1231 mutex_exit(&itxg->itxg_lock);
1236 * Locate the object node and append its list.
1238 t = &itxg->itxg_itxs->i_async_tree;
1239 ian = avl_find(t, &oid, &where);
1241 list_move_tail(&clean_list, &ian->ia_list);
1242 mutex_exit(&itxg->itxg_lock);
1244 while ((itx = list_head(&clean_list)) != NULL) {
1245 list_remove(&clean_list, itx);
1246 kmem_free(itx, offsetof(itx_t, itx_lr) +
1247 itx->itx_lr.lrc_reclen);
1249 list_destroy(&clean_list);
1253 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1257 itxs_t *itxs, *clean = NULL;
1260 * Object ids can be re-instantiated in the next txg so
1261 * remove any async transactions to avoid future leaks.
1262 * This can happen if a fsync occurs on the re-instantiated
1263 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1264 * the new file data and flushes a write record for the old object.
1266 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1267 zil_remove_async(zilog, itx->itx_oid);
1270 * Ensure the data of a renamed file is committed before the rename.
1272 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1273 zil_async_to_sync(zilog, itx->itx_oid);
1275 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1278 txg = dmu_tx_get_txg(tx);
1280 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1281 mutex_enter(&itxg->itxg_lock);
1282 itxs = itxg->itxg_itxs;
1283 if (itxg->itxg_txg != txg) {
1286 * The zil_clean callback hasn't got around to cleaning
1287 * this itxg. Save the itxs for release below.
1288 * This should be rare.
1290 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1292 clean = itxg->itxg_itxs;
1294 ASSERT(itxg->itxg_sod == 0);
1295 itxg->itxg_txg = txg;
1296 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_PUSHPAGE);
1298 list_create(&itxs->i_sync_list, sizeof (itx_t),
1299 offsetof(itx_t, itx_node));
1300 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1301 sizeof (itx_async_node_t),
1302 offsetof(itx_async_node_t, ia_node));
1304 if (itx->itx_sync) {
1305 list_insert_tail(&itxs->i_sync_list, itx);
1306 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1307 itxg->itxg_sod += itx->itx_sod;
1309 avl_tree_t *t = &itxs->i_async_tree;
1310 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1311 itx_async_node_t *ian;
1314 ian = avl_find(t, &foid, &where);
1316 ian = kmem_alloc(sizeof (itx_async_node_t), KM_PUSHPAGE);
1317 list_create(&ian->ia_list, sizeof (itx_t),
1318 offsetof(itx_t, itx_node));
1319 ian->ia_foid = foid;
1320 avl_insert(t, ian, where);
1322 list_insert_tail(&ian->ia_list, itx);
1325 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1326 mutex_exit(&itxg->itxg_lock);
1328 /* Release the old itxs now we've dropped the lock */
1330 zil_itxg_clean(clean);
1334 * If there are any in-memory intent log transactions which have now been
1335 * synced then start up a taskq to free them.
1338 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1340 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1343 mutex_enter(&itxg->itxg_lock);
1344 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1345 mutex_exit(&itxg->itxg_lock);
1348 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1349 ASSERT(itxg->itxg_txg != 0);
1350 ASSERT(zilog->zl_clean_taskq != NULL);
1351 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1353 clean_me = itxg->itxg_itxs;
1354 itxg->itxg_itxs = NULL;
1356 mutex_exit(&itxg->itxg_lock);
1358 * Preferably start a task queue to free up the old itxs but
1359 * if taskq_dispatch can't allocate resources to do that then
1360 * free it in-line. This should be rare. Note, using TQ_SLEEP
1361 * created a bad performance problem.
1363 if (taskq_dispatch(zilog->zl_clean_taskq,
1364 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1365 zil_itxg_clean(clean_me);
1369 * Get the list of itxs to commit into zl_itx_commit_list.
1372 zil_get_commit_list(zilog_t *zilog)
1375 list_t *commit_list = &zilog->zl_itx_commit_list;
1376 uint64_t push_sod = 0;
1378 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1381 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1383 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1384 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1386 mutex_enter(&itxg->itxg_lock);
1387 if (itxg->itxg_txg != txg) {
1388 mutex_exit(&itxg->itxg_lock);
1392 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1393 push_sod += itxg->itxg_sod;
1396 mutex_exit(&itxg->itxg_lock);
1398 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1402 * Move the async itxs for a specified object to commit into sync lists.
1405 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1408 itx_async_node_t *ian;
1412 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1415 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1417 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1418 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1420 mutex_enter(&itxg->itxg_lock);
1421 if (itxg->itxg_txg != txg) {
1422 mutex_exit(&itxg->itxg_lock);
1427 * If a foid is specified then find that node and append its
1428 * list. Otherwise walk the tree appending all the lists
1429 * to the sync list. We add to the end rather than the
1430 * beginning to ensure the create has happened.
1432 t = &itxg->itxg_itxs->i_async_tree;
1434 ian = avl_find(t, &foid, &where);
1436 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1440 void *cookie = NULL;
1442 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1443 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1445 list_destroy(&ian->ia_list);
1446 kmem_free(ian, sizeof (itx_async_node_t));
1449 mutex_exit(&itxg->itxg_lock);
1454 zil_commit_writer(zilog_t *zilog)
1459 spa_t *spa = zilog->zl_spa;
1462 ASSERT(zilog->zl_root_zio == NULL);
1464 mutex_exit(&zilog->zl_lock);
1466 zil_get_commit_list(zilog);
1469 * Return if there's nothing to commit before we dirty the fs by
1470 * calling zil_create().
1472 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1473 mutex_enter(&zilog->zl_lock);
1477 if (zilog->zl_suspend) {
1480 lwb = list_tail(&zilog->zl_lwb_list);
1482 lwb = zil_create(zilog);
1485 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1486 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1487 txg = itx->itx_lr.lrc_txg;
1490 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1491 lwb = zil_lwb_commit(zilog, itx, lwb);
1492 list_remove(&zilog->zl_itx_commit_list, itx);
1493 kmem_free(itx, offsetof(itx_t, itx_lr)
1494 + itx->itx_lr.lrc_reclen);
1496 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1498 /* write the last block out */
1499 if (lwb != NULL && lwb->lwb_zio != NULL)
1500 lwb = zil_lwb_write_start(zilog, lwb);
1502 zilog->zl_cur_used = 0;
1505 * Wait if necessary for the log blocks to be on stable storage.
1507 if (zilog->zl_root_zio) {
1508 error = zio_wait(zilog->zl_root_zio);
1509 zilog->zl_root_zio = NULL;
1510 zil_flush_vdevs(zilog);
1513 if (error || lwb == NULL)
1514 txg_wait_synced(zilog->zl_dmu_pool, 0);
1516 mutex_enter(&zilog->zl_lock);
1519 * Remember the highest committed log sequence number for ztest.
1520 * We only update this value when all the log writes succeeded,
1521 * because ztest wants to ASSERT that it got the whole log chain.
1523 if (error == 0 && lwb != NULL)
1524 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1528 * Commit zfs transactions to stable storage.
1529 * If foid is 0 push out all transactions, otherwise push only those
1530 * for that object or might reference that object.
1532 * itxs are committed in batches. In a heavily stressed zil there will be
1533 * a commit writer thread who is writing out a bunch of itxs to the log
1534 * for a set of committing threads (cthreads) in the same batch as the writer.
1535 * Those cthreads are all waiting on the same cv for that batch.
1537 * There will also be a different and growing batch of threads that are
1538 * waiting to commit (qthreads). When the committing batch completes
1539 * a transition occurs such that the cthreads exit and the qthreads become
1540 * cthreads. One of the new cthreads becomes the writer thread for the
1541 * batch. Any new threads arriving become new qthreads.
1543 * Only 2 condition variables are needed and there's no transition
1544 * between the two cvs needed. They just flip-flop between qthreads
1547 * Using this scheme we can efficiently wakeup up only those threads
1548 * that have been committed.
1551 zil_commit(zilog_t *zilog, uint64_t foid)
1555 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1558 ZIL_STAT_BUMP(zil_commit_count);
1560 /* move the async itxs for the foid to the sync queues */
1561 zil_async_to_sync(zilog, foid);
1563 mutex_enter(&zilog->zl_lock);
1564 mybatch = zilog->zl_next_batch;
1565 while (zilog->zl_writer) {
1566 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1567 if (mybatch <= zilog->zl_com_batch) {
1568 mutex_exit(&zilog->zl_lock);
1573 zilog->zl_next_batch++;
1574 zilog->zl_writer = B_TRUE;
1575 ZIL_STAT_BUMP(zil_commit_writer_count);
1576 zil_commit_writer(zilog);
1577 zilog->zl_com_batch = mybatch;
1578 zilog->zl_writer = B_FALSE;
1580 /* wake up one thread to become the next writer */
1581 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1583 /* wake up all threads waiting for this batch to be committed */
1584 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1586 mutex_exit(&zilog->zl_lock);
1590 * Called in syncing context to free committed log blocks and update log header.
1593 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1595 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1596 uint64_t txg = dmu_tx_get_txg(tx);
1597 spa_t *spa = zilog->zl_spa;
1598 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1602 * We don't zero out zl_destroy_txg, so make sure we don't try
1603 * to destroy it twice.
1605 if (spa_sync_pass(spa) != 1)
1608 mutex_enter(&zilog->zl_lock);
1610 ASSERT(zilog->zl_stop_sync == 0);
1612 if (*replayed_seq != 0) {
1613 ASSERT(zh->zh_replay_seq < *replayed_seq);
1614 zh->zh_replay_seq = *replayed_seq;
1618 if (zilog->zl_destroy_txg == txg) {
1619 blkptr_t blk = zh->zh_log;
1621 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1623 bzero(zh, sizeof (zil_header_t));
1624 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1626 if (zilog->zl_keep_first) {
1628 * If this block was part of log chain that couldn't
1629 * be claimed because a device was missing during
1630 * zil_claim(), but that device later returns,
1631 * then this block could erroneously appear valid.
1632 * To guard against this, assign a new GUID to the new
1633 * log chain so it doesn't matter what blk points to.
1635 zil_init_log_chain(zilog, &blk);
1640 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1641 zh->zh_log = lwb->lwb_blk;
1642 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1645 ASSERT(lwb->lwb_zio == NULL);
1647 list_remove(&zilog->zl_lwb_list, lwb);
1648 zio_free_zil(spa, txg, &lwb->lwb_blk);
1649 kmem_cache_free(zil_lwb_cache, lwb);
1652 * If we don't have anything left in the lwb list then
1653 * we've had an allocation failure and we need to zero
1654 * out the zil_header blkptr so that we don't end
1655 * up freeing the same block twice.
1657 if (list_head(&zilog->zl_lwb_list) == NULL)
1658 BP_ZERO(&zh->zh_log);
1662 * Remove fastwrite on any blocks that have been pre-allocated for
1663 * the next commit. This prevents fastwrite counter pollution by
1664 * unused, long-lived LWBs.
1666 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
1667 if (lwb->lwb_fastwrite && !lwb->lwb_zio) {
1668 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1669 lwb->lwb_fastwrite = 0;
1673 mutex_exit(&zilog->zl_lock);
1679 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1680 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1682 zil_ksp = kstat_create("zfs", 0, "zil", "misc",
1683 KSTAT_TYPE_NAMED, sizeof(zil_stats) / sizeof(kstat_named_t),
1684 KSTAT_FLAG_VIRTUAL);
1686 if (zil_ksp != NULL) {
1687 zil_ksp->ks_data = &zil_stats;
1688 kstat_install(zil_ksp);
1695 kmem_cache_destroy(zil_lwb_cache);
1697 if (zil_ksp != NULL) {
1698 kstat_delete(zil_ksp);
1704 zil_set_sync(zilog_t *zilog, uint64_t sync)
1706 zilog->zl_sync = sync;
1710 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1712 zilog->zl_logbias = logbias;
1716 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1721 zilog = kmem_zalloc(sizeof (zilog_t), KM_PUSHPAGE);
1723 zilog->zl_header = zh_phys;
1725 zilog->zl_spa = dmu_objset_spa(os);
1726 zilog->zl_dmu_pool = dmu_objset_pool(os);
1727 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1728 zilog->zl_logbias = dmu_objset_logbias(os);
1729 zilog->zl_sync = dmu_objset_syncprop(os);
1730 zilog->zl_next_batch = 1;
1732 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1734 for (i = 0; i < TXG_SIZE; i++) {
1735 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1736 MUTEX_DEFAULT, NULL);
1739 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1740 offsetof(lwb_t, lwb_node));
1742 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1743 offsetof(itx_t, itx_node));
1745 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1747 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1748 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1750 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1751 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1752 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1753 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1759 zil_free(zilog_t *zilog)
1763 zilog->zl_stop_sync = 1;
1765 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1766 list_destroy(&zilog->zl_lwb_list);
1768 avl_destroy(&zilog->zl_vdev_tree);
1769 mutex_destroy(&zilog->zl_vdev_lock);
1771 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1772 list_destroy(&zilog->zl_itx_commit_list);
1774 for (i = 0; i < TXG_SIZE; i++) {
1776 * It's possible for an itx to be generated that doesn't dirty
1777 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1778 * callback to remove the entry. We remove those here.
1780 * Also free up the ziltest itxs.
1782 if (zilog->zl_itxg[i].itxg_itxs)
1783 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1784 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1787 mutex_destroy(&zilog->zl_lock);
1789 cv_destroy(&zilog->zl_cv_writer);
1790 cv_destroy(&zilog->zl_cv_suspend);
1791 cv_destroy(&zilog->zl_cv_batch[0]);
1792 cv_destroy(&zilog->zl_cv_batch[1]);
1794 kmem_free(zilog, sizeof (zilog_t));
1798 * Open an intent log.
1801 zil_open(objset_t *os, zil_get_data_t *get_data)
1803 zilog_t *zilog = dmu_objset_zil(os);
1805 ASSERT(zilog->zl_clean_taskq == NULL);
1806 ASSERT(zilog->zl_get_data == NULL);
1807 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1809 zilog->zl_get_data = get_data;
1810 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1811 2, 2, TASKQ_PREPOPULATE);
1817 * Close an intent log.
1820 zil_close(zilog_t *zilog)
1825 zil_commit(zilog, 0); /* commit all itx */
1828 * The lwb_max_txg for the stubby lwb will reflect the last activity
1829 * for the zil. After a txg_wait_synced() on the txg we know all the
1830 * callbacks have occurred that may clean the zil. Only then can we
1831 * destroy the zl_clean_taskq.
1833 mutex_enter(&zilog->zl_lock);
1834 lwb = list_tail(&zilog->zl_lwb_list);
1836 txg = lwb->lwb_max_txg;
1837 mutex_exit(&zilog->zl_lock);
1839 txg_wait_synced(zilog->zl_dmu_pool, txg);
1841 taskq_destroy(zilog->zl_clean_taskq);
1842 zilog->zl_clean_taskq = NULL;
1843 zilog->zl_get_data = NULL;
1846 * We should have only one LWB left on the list; remove it now.
1848 mutex_enter(&zilog->zl_lock);
1849 lwb = list_head(&zilog->zl_lwb_list);
1851 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1852 ASSERT(lwb->lwb_zio == NULL);
1853 if (lwb->lwb_fastwrite)
1854 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
1855 list_remove(&zilog->zl_lwb_list, lwb);
1856 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1857 kmem_cache_free(zil_lwb_cache, lwb);
1859 mutex_exit(&zilog->zl_lock);
1863 * Suspend an intent log. While in suspended mode, we still honor
1864 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1865 * We suspend the log briefly when taking a snapshot so that the snapshot
1866 * contains all the data it's supposed to, and has an empty intent log.
1869 zil_suspend(zilog_t *zilog)
1871 const zil_header_t *zh = zilog->zl_header;
1873 mutex_enter(&zilog->zl_lock);
1874 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1875 mutex_exit(&zilog->zl_lock);
1878 if (zilog->zl_suspend++ != 0) {
1880 * Someone else already began a suspend.
1881 * Just wait for them to finish.
1883 while (zilog->zl_suspending)
1884 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1885 mutex_exit(&zilog->zl_lock);
1888 zilog->zl_suspending = B_TRUE;
1889 mutex_exit(&zilog->zl_lock);
1891 zil_commit(zilog, 0);
1893 zil_destroy(zilog, B_FALSE);
1895 mutex_enter(&zilog->zl_lock);
1896 zilog->zl_suspending = B_FALSE;
1897 cv_broadcast(&zilog->zl_cv_suspend);
1898 mutex_exit(&zilog->zl_lock);
1904 zil_resume(zilog_t *zilog)
1906 mutex_enter(&zilog->zl_lock);
1907 ASSERT(zilog->zl_suspend != 0);
1908 zilog->zl_suspend--;
1909 mutex_exit(&zilog->zl_lock);
1912 typedef struct zil_replay_arg {
1913 zil_replay_func_t **zr_replay;
1915 boolean_t zr_byteswap;
1920 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1922 char name[MAXNAMELEN];
1924 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1926 dmu_objset_name(zilog->zl_os, name);
1928 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1929 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1930 (u_longlong_t)lr->lrc_seq,
1931 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1932 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1938 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1940 zil_replay_arg_t *zr = zra;
1941 const zil_header_t *zh = zilog->zl_header;
1942 uint64_t reclen = lr->lrc_reclen;
1943 uint64_t txtype = lr->lrc_txtype;
1946 zilog->zl_replaying_seq = lr->lrc_seq;
1948 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1951 if (lr->lrc_txg < claim_txg) /* already committed */
1954 /* Strip case-insensitive bit, still present in log record */
1957 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1958 return (zil_replay_error(zilog, lr, EINVAL));
1961 * If this record type can be logged out of order, the object
1962 * (lr_foid) may no longer exist. That's legitimate, not an error.
1964 if (TX_OOO(txtype)) {
1965 error = dmu_object_info(zilog->zl_os,
1966 ((lr_ooo_t *)lr)->lr_foid, NULL);
1967 if (error == ENOENT || error == EEXIST)
1972 * Make a copy of the data so we can revise and extend it.
1974 bcopy(lr, zr->zr_lr, reclen);
1977 * If this is a TX_WRITE with a blkptr, suck in the data.
1979 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1980 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1981 zr->zr_lr + reclen);
1983 return (zil_replay_error(zilog, lr, error));
1987 * The log block containing this lr may have been byteswapped
1988 * so that we can easily examine common fields like lrc_txtype.
1989 * However, the log is a mix of different record types, and only the
1990 * replay vectors know how to byteswap their records. Therefore, if
1991 * the lr was byteswapped, undo it before invoking the replay vector.
1993 if (zr->zr_byteswap)
1994 byteswap_uint64_array(zr->zr_lr, reclen);
1997 * We must now do two things atomically: replay this log record,
1998 * and update the log header sequence number to reflect the fact that
1999 * we did so. At the end of each replay function the sequence number
2000 * is updated if we are in replay mode.
2002 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2005 * The DMU's dnode layer doesn't see removes until the txg
2006 * commits, so a subsequent claim can spuriously fail with
2007 * EEXIST. So if we receive any error we try syncing out
2008 * any removes then retry the transaction. Note that we
2009 * specify B_FALSE for byteswap now, so we don't do it twice.
2011 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2012 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2014 return (zil_replay_error(zilog, lr, error));
2021 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2023 zilog->zl_replay_blks++;
2029 * If this dataset has a non-empty intent log, replay it and destroy it.
2032 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2034 zilog_t *zilog = dmu_objset_zil(os);
2035 const zil_header_t *zh = zilog->zl_header;
2036 zil_replay_arg_t zr;
2038 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2039 zil_destroy(zilog, B_TRUE);
2043 zr.zr_replay = replay_func;
2045 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2046 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_PUSHPAGE);
2049 * Wait for in-progress removes to sync before starting replay.
2051 txg_wait_synced(zilog->zl_dmu_pool, 0);
2053 zilog->zl_replay = B_TRUE;
2054 zilog->zl_replay_time = ddi_get_lbolt();
2055 ASSERT(zilog->zl_replay_blks == 0);
2056 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2058 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2060 zil_destroy(zilog, B_FALSE);
2061 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2062 zilog->zl_replay = B_FALSE;
2066 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2068 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2071 if (zilog->zl_replay) {
2072 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2073 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2074 zilog->zl_replaying_seq;
2083 zil_vdev_offline(const char *osname, void *arg)
2089 error = dmu_objset_hold(osname, FTAG, &os);
2093 zilog = dmu_objset_zil(os);
2094 if (zil_suspend(zilog) != 0)
2098 dmu_objset_rele(os, FTAG);
2102 #if defined(_KERNEL) && defined(HAVE_SPL)
2103 module_param(zil_replay_disable, int, 0644);
2104 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
2106 module_param(zfs_nocacheflush, int, 0644);
2107 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");
2109 module_param(zil_slog_limit, ulong, 0644);
2110 MODULE_PARM_DESC(zil_slog_limit, "Max commit bytes to separate log device");