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, 2014 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
43 * The zfs intent log (ZIL) saves transaction records of system calls
44 * that change the file system in memory with enough information
45 * to be able to replay them. These are stored in memory until
46 * either the DMU transaction group (txg) commits them to the stable pool
47 * and they can be discarded, or they are flushed to the stable log
48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 * requirement. In the event of a panic or power fail then those log
50 * records (transactions) are replayed.
52 * There is one ZIL per file system. Its on-disk (pool) format consists
59 * A log record holds a system call transaction. Log blocks can
60 * hold many log records and the blocks are chained together.
61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 * ZIL block in the chain. The ZIL header points to the first
63 * block in the chain. Note there is not a fixed place in the pool
64 * to hold blocks. They are dynamically allocated and freed as
65 * needed from the blocks available. Figure X shows the ZIL structure:
69 * Disable intent logging replay. This global ZIL switch affects all pools.
71 int zil_replay_disable = 0;
72 SYSCTL_DECL(_vfs_zfs);
73 TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable);
74 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW,
75 &zil_replay_disable, 0, "Disable intent logging replay");
78 * Tunable parameter for debugging or performance analysis. Setting
79 * zfs_nocacheflush will cause corruption on power loss if a volatile
80 * out-of-order write cache is enabled.
82 boolean_t zfs_nocacheflush = B_FALSE;
83 TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush);
84 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
85 &zfs_nocacheflush, 0, "Disable cache flush");
86 boolean_t zfs_trim_enabled = B_TRUE;
87 SYSCTL_DECL(_vfs_zfs_trim);
88 TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled);
89 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
92 static kmem_cache_t *zil_lwb_cache;
94 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
96 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
97 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
101 * ziltest is by and large an ugly hack, but very useful in
102 * checking replay without tedious work.
103 * When running ziltest we want to keep all itx's and so maintain
104 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
105 * We subtract TXG_CONCURRENT_STATES to allow for common code.
107 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
110 zil_bp_compare(const void *x1, const void *x2)
112 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
113 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
115 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
117 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
120 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
122 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
129 zil_bp_tree_init(zilog_t *zilog)
131 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
132 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
136 zil_bp_tree_fini(zilog_t *zilog)
138 avl_tree_t *t = &zilog->zl_bp_tree;
142 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
143 kmem_free(zn, sizeof (zil_bp_node_t));
149 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
151 avl_tree_t *t = &zilog->zl_bp_tree;
156 if (BP_IS_EMBEDDED(bp))
159 dva = BP_IDENTITY(bp);
161 if (avl_find(t, dva, &where) != NULL)
162 return (SET_ERROR(EEXIST));
164 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
166 avl_insert(t, zn, where);
171 static zil_header_t *
172 zil_header_in_syncing_context(zilog_t *zilog)
174 return ((zil_header_t *)zilog->zl_header);
178 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
180 zio_cksum_t *zc = &bp->blk_cksum;
182 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
183 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
184 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
185 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
189 * Read a log block and make sure it's valid.
192 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
195 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
196 uint32_t aflags = ARC_WAIT;
197 arc_buf_t *abuf = NULL;
201 if (zilog->zl_header->zh_claim_txg == 0)
202 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
204 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
205 zio_flags |= ZIO_FLAG_SPECULATIVE;
207 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
208 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
210 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
211 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
214 zio_cksum_t cksum = bp->blk_cksum;
217 * Validate the checksummed log block.
219 * Sequence numbers should be... sequential. The checksum
220 * verifier for the next block should be bp's checksum plus 1.
222 * Also check the log chain linkage and size used.
224 cksum.zc_word[ZIL_ZC_SEQ]++;
226 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
227 zil_chain_t *zilc = abuf->b_data;
228 char *lr = (char *)(zilc + 1);
229 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
231 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
232 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
233 error = SET_ERROR(ECKSUM);
236 *end = (char *)dst + len;
237 *nbp = zilc->zc_next_blk;
240 char *lr = abuf->b_data;
241 uint64_t size = BP_GET_LSIZE(bp);
242 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
244 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
245 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
246 (zilc->zc_nused > (size - sizeof (*zilc)))) {
247 error = SET_ERROR(ECKSUM);
249 bcopy(lr, dst, zilc->zc_nused);
250 *end = (char *)dst + zilc->zc_nused;
251 *nbp = zilc->zc_next_blk;
255 VERIFY(arc_buf_remove_ref(abuf, &abuf));
262 * Read a TX_WRITE log data block.
265 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
267 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
268 const blkptr_t *bp = &lr->lr_blkptr;
269 uint32_t aflags = ARC_WAIT;
270 arc_buf_t *abuf = NULL;
274 if (BP_IS_HOLE(bp)) {
276 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
280 if (zilog->zl_header->zh_claim_txg == 0)
281 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
283 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
284 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
286 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
287 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
291 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
292 (void) arc_buf_remove_ref(abuf, &abuf);
299 * Parse the intent log, and call parse_func for each valid record within.
302 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
303 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
305 const zil_header_t *zh = zilog->zl_header;
306 boolean_t claimed = !!zh->zh_claim_txg;
307 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
308 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
309 uint64_t max_blk_seq = 0;
310 uint64_t max_lr_seq = 0;
311 uint64_t blk_count = 0;
312 uint64_t lr_count = 0;
313 blkptr_t blk, next_blk;
318 * Old logs didn't record the maximum zh_claim_lr_seq.
320 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
321 claim_lr_seq = UINT64_MAX;
324 * Starting at the block pointed to by zh_log we read the log chain.
325 * For each block in the chain we strongly check that block to
326 * ensure its validity. We stop when an invalid block is found.
327 * For each block pointer in the chain we call parse_blk_func().
328 * For each record in each valid block we call parse_lr_func().
329 * If the log has been claimed, stop if we encounter a sequence
330 * number greater than the highest claimed sequence number.
332 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
333 zil_bp_tree_init(zilog);
335 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
336 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
340 if (blk_seq > claim_blk_seq)
342 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
344 ASSERT3U(max_blk_seq, <, blk_seq);
345 max_blk_seq = blk_seq;
348 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
351 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
355 for (lrp = lrbuf; lrp < end; lrp += reclen) {
356 lr_t *lr = (lr_t *)lrp;
357 reclen = lr->lrc_reclen;
358 ASSERT3U(reclen, >=, sizeof (lr_t));
359 if (lr->lrc_seq > claim_lr_seq)
361 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
363 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
364 max_lr_seq = lr->lrc_seq;
369 zilog->zl_parse_error = error;
370 zilog->zl_parse_blk_seq = max_blk_seq;
371 zilog->zl_parse_lr_seq = max_lr_seq;
372 zilog->zl_parse_blk_count = blk_count;
373 zilog->zl_parse_lr_count = lr_count;
375 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
376 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
378 zil_bp_tree_fini(zilog);
379 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
385 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
388 * Claim log block if not already committed and not already claimed.
389 * If tx == NULL, just verify that the block is claimable.
391 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
392 zil_bp_tree_add(zilog, bp) != 0)
395 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
396 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
397 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
401 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
403 lr_write_t *lr = (lr_write_t *)lrc;
406 if (lrc->lrc_txtype != TX_WRITE)
410 * If the block is not readable, don't claim it. This can happen
411 * in normal operation when a log block is written to disk before
412 * some of the dmu_sync() blocks it points to. In this case, the
413 * transaction cannot have been committed to anyone (we would have
414 * waited for all writes to be stable first), so it is semantically
415 * correct to declare this the end of the log.
417 if (lr->lr_blkptr.blk_birth >= first_txg &&
418 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
420 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
425 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
427 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
433 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
435 lr_write_t *lr = (lr_write_t *)lrc;
436 blkptr_t *bp = &lr->lr_blkptr;
439 * If we previously claimed it, we need to free it.
441 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
442 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
444 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
450 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
454 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
455 lwb->lwb_zilog = zilog;
457 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
458 lwb->lwb_max_txg = txg;
461 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
462 lwb->lwb_nused = sizeof (zil_chain_t);
463 lwb->lwb_sz = BP_GET_LSIZE(bp);
466 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
469 mutex_enter(&zilog->zl_lock);
470 list_insert_tail(&zilog->zl_lwb_list, lwb);
471 mutex_exit(&zilog->zl_lock);
477 * Called when we create in-memory log transactions so that we know
478 * to cleanup the itxs at the end of spa_sync().
481 zilog_dirty(zilog_t *zilog, uint64_t txg)
483 dsl_pool_t *dp = zilog->zl_dmu_pool;
484 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
486 if (dsl_dataset_is_snapshot(ds))
487 panic("dirtying snapshot!");
489 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
490 /* up the hold count until we can be written out */
491 dmu_buf_add_ref(ds->ds_dbuf, zilog);
496 zilog_is_dirty(zilog_t *zilog)
498 dsl_pool_t *dp = zilog->zl_dmu_pool;
500 for (int t = 0; t < TXG_SIZE; t++) {
501 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
508 * Create an on-disk intent log.
511 zil_create(zilog_t *zilog)
513 const zil_header_t *zh = zilog->zl_header;
521 * Wait for any previous destroy to complete.
523 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
525 ASSERT(zh->zh_claim_txg == 0);
526 ASSERT(zh->zh_replay_seq == 0);
531 * Allocate an initial log block if:
532 * - there isn't one already
533 * - the existing block is the wrong endianess
535 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
536 tx = dmu_tx_create(zilog->zl_os);
537 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
538 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
539 txg = dmu_tx_get_txg(tx);
541 if (!BP_IS_HOLE(&blk)) {
542 zio_free_zil(zilog->zl_spa, txg, &blk);
546 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
547 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
550 zil_init_log_chain(zilog, &blk);
554 * Allocate a log write buffer (lwb) for the first log block.
557 lwb = zil_alloc_lwb(zilog, &blk, txg);
560 * If we just allocated the first log block, commit our transaction
561 * and wait for zil_sync() to stuff the block poiner into zh_log.
562 * (zh is part of the MOS, so we cannot modify it in open context.)
566 txg_wait_synced(zilog->zl_dmu_pool, txg);
569 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
575 * In one tx, free all log blocks and clear the log header.
576 * If keep_first is set, then we're replaying a log with no content.
577 * We want to keep the first block, however, so that the first
578 * synchronous transaction doesn't require a txg_wait_synced()
579 * in zil_create(). We don't need to txg_wait_synced() here either
580 * when keep_first is set, because both zil_create() and zil_destroy()
581 * will wait for any in-progress destroys to complete.
584 zil_destroy(zilog_t *zilog, boolean_t keep_first)
586 const zil_header_t *zh = zilog->zl_header;
592 * Wait for any previous destroy to complete.
594 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
596 zilog->zl_old_header = *zh; /* debugging aid */
598 if (BP_IS_HOLE(&zh->zh_log))
601 tx = dmu_tx_create(zilog->zl_os);
602 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
603 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
604 txg = dmu_tx_get_txg(tx);
606 mutex_enter(&zilog->zl_lock);
608 ASSERT3U(zilog->zl_destroy_txg, <, txg);
609 zilog->zl_destroy_txg = txg;
610 zilog->zl_keep_first = keep_first;
612 if (!list_is_empty(&zilog->zl_lwb_list)) {
613 ASSERT(zh->zh_claim_txg == 0);
615 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
616 list_remove(&zilog->zl_lwb_list, lwb);
617 if (lwb->lwb_buf != NULL)
618 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
619 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
620 kmem_cache_free(zil_lwb_cache, lwb);
622 } else if (!keep_first) {
623 zil_destroy_sync(zilog, tx);
625 mutex_exit(&zilog->zl_lock);
631 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
633 ASSERT(list_is_empty(&zilog->zl_lwb_list));
634 (void) zil_parse(zilog, zil_free_log_block,
635 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
639 zil_claim(const char *osname, void *txarg)
641 dmu_tx_t *tx = txarg;
642 uint64_t first_txg = dmu_tx_get_txg(tx);
648 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
651 * EBUSY indicates that the objset is inconsistent, in which
652 * case it can not have a ZIL.
654 if (error != EBUSY) {
655 cmn_err(CE_WARN, "can't open objset for %s, error %u",
661 zilog = dmu_objset_zil(os);
662 zh = zil_header_in_syncing_context(zilog);
664 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
665 if (!BP_IS_HOLE(&zh->zh_log))
666 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
667 BP_ZERO(&zh->zh_log);
668 dsl_dataset_dirty(dmu_objset_ds(os), tx);
669 dmu_objset_disown(os, FTAG);
674 * Claim all log blocks if we haven't already done so, and remember
675 * the highest claimed sequence number. This ensures that if we can
676 * read only part of the log now (e.g. due to a missing device),
677 * but we can read the entire log later, we will not try to replay
678 * or destroy beyond the last block we successfully claimed.
680 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
681 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
682 (void) zil_parse(zilog, zil_claim_log_block,
683 zil_claim_log_record, tx, first_txg);
684 zh->zh_claim_txg = first_txg;
685 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
686 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
687 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
688 zh->zh_flags |= ZIL_REPLAY_NEEDED;
689 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
690 dsl_dataset_dirty(dmu_objset_ds(os), tx);
693 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
694 dmu_objset_disown(os, FTAG);
699 * Check the log by walking the log chain.
700 * Checksum errors are ok as they indicate the end of the chain.
701 * Any other error (no device or read failure) returns an error.
704 zil_check_log_chain(const char *osname, void *tx)
713 error = dmu_objset_hold(osname, FTAG, &os);
715 cmn_err(CE_WARN, "can't open objset for %s", osname);
719 zilog = dmu_objset_zil(os);
720 bp = (blkptr_t *)&zilog->zl_header->zh_log;
723 * Check the first block and determine if it's on a log device
724 * which may have been removed or faulted prior to loading this
725 * pool. If so, there's no point in checking the rest of the log
726 * as its content should have already been synced to the pool.
728 if (!BP_IS_HOLE(bp)) {
730 boolean_t valid = B_TRUE;
732 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
733 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
734 if (vd->vdev_islog && vdev_is_dead(vd))
735 valid = vdev_log_state_valid(vd);
736 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
739 dmu_objset_rele(os, FTAG);
745 * Because tx == NULL, zil_claim_log_block() will not actually claim
746 * any blocks, but just determine whether it is possible to do so.
747 * In addition to checking the log chain, zil_claim_log_block()
748 * will invoke zio_claim() with a done func of spa_claim_notify(),
749 * which will update spa_max_claim_txg. See spa_load() for details.
751 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
752 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
754 dmu_objset_rele(os, FTAG);
756 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
760 zil_vdev_compare(const void *x1, const void *x2)
762 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
763 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
774 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
776 avl_tree_t *t = &zilog->zl_vdev_tree;
778 zil_vdev_node_t *zv, zvsearch;
779 int ndvas = BP_GET_NDVAS(bp);
782 if (zfs_nocacheflush)
785 ASSERT(zilog->zl_writer);
788 * Even though we're zl_writer, we still need a lock because the
789 * zl_get_data() callbacks may have dmu_sync() done callbacks
790 * that will run concurrently.
792 mutex_enter(&zilog->zl_vdev_lock);
793 for (i = 0; i < ndvas; i++) {
794 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
795 if (avl_find(t, &zvsearch, &where) == NULL) {
796 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
797 zv->zv_vdev = zvsearch.zv_vdev;
798 avl_insert(t, zv, where);
801 mutex_exit(&zilog->zl_vdev_lock);
805 zil_flush_vdevs(zilog_t *zilog)
807 spa_t *spa = zilog->zl_spa;
808 avl_tree_t *t = &zilog->zl_vdev_tree;
813 ASSERT(zilog->zl_writer);
816 * We don't need zl_vdev_lock here because we're the zl_writer,
817 * and all zl_get_data() callbacks are done.
819 if (avl_numnodes(t) == 0)
822 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
824 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
826 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
827 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
830 kmem_free(zv, sizeof (*zv));
834 * Wait for all the flushes to complete. Not all devices actually
835 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
837 (void) zio_wait(zio);
839 spa_config_exit(spa, SCL_STATE, FTAG);
843 * Function called when a log block write completes
846 zil_lwb_write_done(zio_t *zio)
848 lwb_t *lwb = zio->io_private;
849 zilog_t *zilog = lwb->lwb_zilog;
850 dmu_tx_t *tx = lwb->lwb_tx;
852 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
853 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
854 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
855 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
856 ASSERT(!BP_IS_GANG(zio->io_bp));
857 ASSERT(!BP_IS_HOLE(zio->io_bp));
858 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
861 * Ensure the lwb buffer pointer is cleared before releasing
862 * the txg. If we have had an allocation failure and
863 * the txg is waiting to sync then we want want zil_sync()
864 * to remove the lwb so that it's not picked up as the next new
865 * one in zil_commit_writer(). zil_sync() will only remove
866 * the lwb if lwb_buf is null.
868 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
869 mutex_enter(&zilog->zl_lock);
872 mutex_exit(&zilog->zl_lock);
875 * Now that we've written this log block, we have a stable pointer
876 * to the next block in the chain, so it's OK to let the txg in
877 * which we allocated the next block sync.
883 * Initialize the io for a log block.
886 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
890 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
891 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
892 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
894 if (zilog->zl_root_zio == NULL) {
895 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
898 if (lwb->lwb_zio == NULL) {
899 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
900 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
901 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
902 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
907 * Define a limited set of intent log block sizes.
909 * These must be a multiple of 4KB. Note only the amount used (again
910 * aligned to 4KB) actually gets written. However, we can't always just
911 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
913 uint64_t zil_block_buckets[] = {
914 4096, /* non TX_WRITE */
915 8192+4096, /* data base */
916 32*1024 + 4096, /* NFS writes */
921 * Use the slog as long as the logbias is 'latency' and the current commit size
922 * is less than the limit or the total list size is less than 2X the limit.
923 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
925 uint64_t zil_slog_limit = 1024 * 1024;
926 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
927 (((zilog)->zl_cur_used < zil_slog_limit) || \
928 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
931 * Start a log block write and advance to the next log block.
932 * Calls are serialized.
935 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
939 spa_t *spa = zilog->zl_spa;
943 uint64_t zil_blksz, wsz;
946 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
947 zilc = (zil_chain_t *)lwb->lwb_buf;
948 bp = &zilc->zc_next_blk;
950 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
951 bp = &zilc->zc_next_blk;
954 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
957 * Allocate the next block and save its address in this block
958 * before writing it in order to establish the log chain.
959 * Note that if the allocation of nlwb synced before we wrote
960 * the block that points at it (lwb), we'd leak it if we crashed.
961 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
962 * We dirty the dataset to ensure that zil_sync() will be called
963 * to clean up in the event of allocation failure or I/O failure.
965 tx = dmu_tx_create(zilog->zl_os);
966 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
967 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
968 txg = dmu_tx_get_txg(tx);
973 * Log blocks are pre-allocated. Here we select the size of the next
974 * block, based on size used in the last block.
975 * - first find the smallest bucket that will fit the block from a
976 * limited set of block sizes. This is because it's faster to write
977 * blocks allocated from the same metaslab as they are adjacent or
979 * - next find the maximum from the new suggested size and an array of
980 * previous sizes. This lessens a picket fence effect of wrongly
981 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
984 * Note we only write what is used, but we can't just allocate
985 * the maximum block size because we can exhaust the available
988 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
989 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
991 zil_blksz = zil_block_buckets[i];
992 if (zil_blksz == UINT64_MAX)
993 zil_blksz = SPA_MAXBLOCKSIZE;
994 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
995 for (i = 0; i < ZIL_PREV_BLKS; i++)
996 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
997 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1000 /* pass the old blkptr in order to spread log blocks across devs */
1001 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
1004 ASSERT3U(bp->blk_birth, ==, txg);
1005 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1006 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1009 * Allocate a new log write buffer (lwb).
1011 nlwb = zil_alloc_lwb(zilog, bp, txg);
1013 /* Record the block for later vdev flushing */
1014 zil_add_block(zilog, &lwb->lwb_blk);
1017 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1018 /* For Slim ZIL only write what is used. */
1019 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1020 ASSERT3U(wsz, <=, lwb->lwb_sz);
1021 zio_shrink(lwb->lwb_zio, wsz);
1028 zilc->zc_nused = lwb->lwb_nused;
1029 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1032 * clear unused data for security
1034 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1036 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1039 * If there was an allocation failure then nlwb will be null which
1040 * forces a txg_wait_synced().
1046 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1048 lr_t *lrc = &itx->itx_lr; /* common log record */
1049 lr_write_t *lrw = (lr_write_t *)lrc;
1051 uint64_t txg = lrc->lrc_txg;
1052 uint64_t reclen = lrc->lrc_reclen;
1058 ASSERT(lwb->lwb_buf != NULL);
1059 ASSERT(zilog_is_dirty(zilog) ||
1060 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1062 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1063 dlen = P2ROUNDUP_TYPED(
1064 lrw->lr_length, sizeof (uint64_t), uint64_t);
1066 zilog->zl_cur_used += (reclen + dlen);
1068 zil_lwb_write_init(zilog, lwb);
1071 * If this record won't fit in the current log block, start a new one.
1073 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1074 lwb = zil_lwb_write_start(zilog, lwb);
1077 zil_lwb_write_init(zilog, lwb);
1078 ASSERT(LWB_EMPTY(lwb));
1079 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1080 txg_wait_synced(zilog->zl_dmu_pool, txg);
1085 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1086 bcopy(lrc, lr_buf, reclen);
1087 lrc = (lr_t *)lr_buf;
1088 lrw = (lr_write_t *)lrc;
1091 * If it's a write, fetch the data or get its blkptr as appropriate.
1093 if (lrc->lrc_txtype == TX_WRITE) {
1094 if (txg > spa_freeze_txg(zilog->zl_spa))
1095 txg_wait_synced(zilog->zl_dmu_pool, txg);
1096 if (itx->itx_wr_state != WR_COPIED) {
1101 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1102 dbuf = lr_buf + reclen;
1103 lrw->lr_common.lrc_reclen += dlen;
1105 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1108 error = zilog->zl_get_data(
1109 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1111 txg_wait_synced(zilog->zl_dmu_pool, txg);
1115 ASSERT(error == ENOENT || error == EEXIST ||
1123 * We're actually making an entry, so update lrc_seq to be the
1124 * log record sequence number. Note that this is generally not
1125 * equal to the itx sequence number because not all transactions
1126 * are synchronous, and sometimes spa_sync() gets there first.
1128 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1129 lwb->lwb_nused += reclen + dlen;
1130 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1131 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1132 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1138 zil_itx_create(uint64_t txtype, size_t lrsize)
1142 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1144 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1145 itx->itx_lr.lrc_txtype = txtype;
1146 itx->itx_lr.lrc_reclen = lrsize;
1147 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1148 itx->itx_lr.lrc_seq = 0; /* defensive */
1149 itx->itx_sync = B_TRUE; /* default is synchronous */
1155 zil_itx_destroy(itx_t *itx)
1157 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1161 * Free up the sync and async itxs. The itxs_t has already been detached
1162 * so no locks are needed.
1165 zil_itxg_clean(itxs_t *itxs)
1171 itx_async_node_t *ian;
1173 list = &itxs->i_sync_list;
1174 while ((itx = list_head(list)) != NULL) {
1175 list_remove(list, itx);
1176 kmem_free(itx, offsetof(itx_t, itx_lr) +
1177 itx->itx_lr.lrc_reclen);
1181 t = &itxs->i_async_tree;
1182 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1183 list = &ian->ia_list;
1184 while ((itx = list_head(list)) != NULL) {
1185 list_remove(list, itx);
1186 kmem_free(itx, offsetof(itx_t, itx_lr) +
1187 itx->itx_lr.lrc_reclen);
1190 kmem_free(ian, sizeof (itx_async_node_t));
1194 kmem_free(itxs, sizeof (itxs_t));
1198 zil_aitx_compare(const void *x1, const void *x2)
1200 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1201 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1212 * Remove all async itx with the given oid.
1215 zil_remove_async(zilog_t *zilog, uint64_t oid)
1218 itx_async_node_t *ian;
1225 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1227 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1230 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1232 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1233 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1235 mutex_enter(&itxg->itxg_lock);
1236 if (itxg->itxg_txg != txg) {
1237 mutex_exit(&itxg->itxg_lock);
1242 * Locate the object node and append its list.
1244 t = &itxg->itxg_itxs->i_async_tree;
1245 ian = avl_find(t, &oid, &where);
1247 list_move_tail(&clean_list, &ian->ia_list);
1248 mutex_exit(&itxg->itxg_lock);
1250 while ((itx = list_head(&clean_list)) != NULL) {
1251 list_remove(&clean_list, itx);
1252 kmem_free(itx, offsetof(itx_t, itx_lr) +
1253 itx->itx_lr.lrc_reclen);
1255 list_destroy(&clean_list);
1259 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1263 itxs_t *itxs, *clean = NULL;
1266 * Object ids can be re-instantiated in the next txg so
1267 * remove any async transactions to avoid future leaks.
1268 * This can happen if a fsync occurs on the re-instantiated
1269 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1270 * the new file data and flushes a write record for the old object.
1272 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1273 zil_remove_async(zilog, itx->itx_oid);
1276 * Ensure the data of a renamed file is committed before the rename.
1278 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1279 zil_async_to_sync(zilog, itx->itx_oid);
1281 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1284 txg = dmu_tx_get_txg(tx);
1286 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1287 mutex_enter(&itxg->itxg_lock);
1288 itxs = itxg->itxg_itxs;
1289 if (itxg->itxg_txg != txg) {
1292 * The zil_clean callback hasn't got around to cleaning
1293 * this itxg. Save the itxs for release below.
1294 * This should be rare.
1296 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1298 clean = itxg->itxg_itxs;
1300 ASSERT(itxg->itxg_sod == 0);
1301 itxg->itxg_txg = txg;
1302 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1304 list_create(&itxs->i_sync_list, sizeof (itx_t),
1305 offsetof(itx_t, itx_node));
1306 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1307 sizeof (itx_async_node_t),
1308 offsetof(itx_async_node_t, ia_node));
1310 if (itx->itx_sync) {
1311 list_insert_tail(&itxs->i_sync_list, itx);
1312 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1313 itxg->itxg_sod += itx->itx_sod;
1315 avl_tree_t *t = &itxs->i_async_tree;
1316 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1317 itx_async_node_t *ian;
1320 ian = avl_find(t, &foid, &where);
1322 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1323 list_create(&ian->ia_list, sizeof (itx_t),
1324 offsetof(itx_t, itx_node));
1325 ian->ia_foid = foid;
1326 avl_insert(t, ian, where);
1328 list_insert_tail(&ian->ia_list, itx);
1331 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1332 zilog_dirty(zilog, txg);
1333 mutex_exit(&itxg->itxg_lock);
1335 /* Release the old itxs now we've dropped the lock */
1337 zil_itxg_clean(clean);
1341 * If there are any in-memory intent log transactions which have now been
1342 * synced then start up a taskq to free them. We should only do this after we
1343 * have written out the uberblocks (i.e. txg has been comitted) so that
1344 * don't inadvertently clean out in-memory log records that would be required
1348 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1350 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1353 mutex_enter(&itxg->itxg_lock);
1354 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1355 mutex_exit(&itxg->itxg_lock);
1358 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1359 ASSERT(itxg->itxg_txg != 0);
1360 ASSERT(zilog->zl_clean_taskq != NULL);
1361 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1363 clean_me = itxg->itxg_itxs;
1364 itxg->itxg_itxs = NULL;
1366 mutex_exit(&itxg->itxg_lock);
1368 * Preferably start a task queue to free up the old itxs but
1369 * if taskq_dispatch can't allocate resources to do that then
1370 * free it in-line. This should be rare. Note, using TQ_SLEEP
1371 * created a bad performance problem.
1373 if (taskq_dispatch(zilog->zl_clean_taskq,
1374 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1375 zil_itxg_clean(clean_me);
1379 * Get the list of itxs to commit into zl_itx_commit_list.
1382 zil_get_commit_list(zilog_t *zilog)
1385 list_t *commit_list = &zilog->zl_itx_commit_list;
1386 uint64_t push_sod = 0;
1388 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1391 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1393 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1394 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1396 mutex_enter(&itxg->itxg_lock);
1397 if (itxg->itxg_txg != txg) {
1398 mutex_exit(&itxg->itxg_lock);
1402 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1403 push_sod += itxg->itxg_sod;
1406 mutex_exit(&itxg->itxg_lock);
1408 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1412 * Move the async itxs for a specified object to commit into sync lists.
1415 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1418 itx_async_node_t *ian;
1422 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1425 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1427 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1428 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1430 mutex_enter(&itxg->itxg_lock);
1431 if (itxg->itxg_txg != txg) {
1432 mutex_exit(&itxg->itxg_lock);
1437 * If a foid is specified then find that node and append its
1438 * list. Otherwise walk the tree appending all the lists
1439 * to the sync list. We add to the end rather than the
1440 * beginning to ensure the create has happened.
1442 t = &itxg->itxg_itxs->i_async_tree;
1444 ian = avl_find(t, &foid, &where);
1446 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1450 void *cookie = NULL;
1452 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1453 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1455 list_destroy(&ian->ia_list);
1456 kmem_free(ian, sizeof (itx_async_node_t));
1459 mutex_exit(&itxg->itxg_lock);
1464 zil_commit_writer(zilog_t *zilog)
1469 spa_t *spa = zilog->zl_spa;
1472 ASSERT(zilog->zl_root_zio == NULL);
1474 mutex_exit(&zilog->zl_lock);
1476 zil_get_commit_list(zilog);
1479 * Return if there's nothing to commit before we dirty the fs by
1480 * calling zil_create().
1482 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1483 mutex_enter(&zilog->zl_lock);
1487 if (zilog->zl_suspend) {
1490 lwb = list_tail(&zilog->zl_lwb_list);
1492 lwb = zil_create(zilog);
1495 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1496 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1497 txg = itx->itx_lr.lrc_txg;
1500 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1501 lwb = zil_lwb_commit(zilog, itx, lwb);
1502 list_remove(&zilog->zl_itx_commit_list, itx);
1503 kmem_free(itx, offsetof(itx_t, itx_lr)
1504 + itx->itx_lr.lrc_reclen);
1506 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1508 /* write the last block out */
1509 if (lwb != NULL && lwb->lwb_zio != NULL)
1510 lwb = zil_lwb_write_start(zilog, lwb);
1512 zilog->zl_cur_used = 0;
1515 * Wait if necessary for the log blocks to be on stable storage.
1517 if (zilog->zl_root_zio) {
1518 error = zio_wait(zilog->zl_root_zio);
1519 zilog->zl_root_zio = NULL;
1520 zil_flush_vdevs(zilog);
1523 if (error || lwb == NULL)
1524 txg_wait_synced(zilog->zl_dmu_pool, 0);
1526 mutex_enter(&zilog->zl_lock);
1529 * Remember the highest committed log sequence number for ztest.
1530 * We only update this value when all the log writes succeeded,
1531 * because ztest wants to ASSERT that it got the whole log chain.
1533 if (error == 0 && lwb != NULL)
1534 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1538 * Commit zfs transactions to stable storage.
1539 * If foid is 0 push out all transactions, otherwise push only those
1540 * for that object or might reference that object.
1542 * itxs are committed in batches. In a heavily stressed zil there will be
1543 * a commit writer thread who is writing out a bunch of itxs to the log
1544 * for a set of committing threads (cthreads) in the same batch as the writer.
1545 * Those cthreads are all waiting on the same cv for that batch.
1547 * There will also be a different and growing batch of threads that are
1548 * waiting to commit (qthreads). When the committing batch completes
1549 * a transition occurs such that the cthreads exit and the qthreads become
1550 * cthreads. One of the new cthreads becomes the writer thread for the
1551 * batch. Any new threads arriving become new qthreads.
1553 * Only 2 condition variables are needed and there's no transition
1554 * between the two cvs needed. They just flip-flop between qthreads
1557 * Using this scheme we can efficiently wakeup up only those threads
1558 * that have been committed.
1561 zil_commit(zilog_t *zilog, uint64_t foid)
1565 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1568 /* move the async itxs for the foid to the sync queues */
1569 zil_async_to_sync(zilog, foid);
1571 mutex_enter(&zilog->zl_lock);
1572 mybatch = zilog->zl_next_batch;
1573 while (zilog->zl_writer) {
1574 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1575 if (mybatch <= zilog->zl_com_batch) {
1576 mutex_exit(&zilog->zl_lock);
1581 zilog->zl_next_batch++;
1582 zilog->zl_writer = B_TRUE;
1583 zil_commit_writer(zilog);
1584 zilog->zl_com_batch = mybatch;
1585 zilog->zl_writer = B_FALSE;
1586 mutex_exit(&zilog->zl_lock);
1588 /* wake up one thread to become the next writer */
1589 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1591 /* wake up all threads waiting for this batch to be committed */
1592 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1596 * Called in syncing context to free committed log blocks and update log header.
1599 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1601 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1602 uint64_t txg = dmu_tx_get_txg(tx);
1603 spa_t *spa = zilog->zl_spa;
1604 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1608 * We don't zero out zl_destroy_txg, so make sure we don't try
1609 * to destroy it twice.
1611 if (spa_sync_pass(spa) != 1)
1614 mutex_enter(&zilog->zl_lock);
1616 ASSERT(zilog->zl_stop_sync == 0);
1618 if (*replayed_seq != 0) {
1619 ASSERT(zh->zh_replay_seq < *replayed_seq);
1620 zh->zh_replay_seq = *replayed_seq;
1624 if (zilog->zl_destroy_txg == txg) {
1625 blkptr_t blk = zh->zh_log;
1627 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1629 bzero(zh, sizeof (zil_header_t));
1630 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1632 if (zilog->zl_keep_first) {
1634 * If this block was part of log chain that couldn't
1635 * be claimed because a device was missing during
1636 * zil_claim(), but that device later returns,
1637 * then this block could erroneously appear valid.
1638 * To guard against this, assign a new GUID to the new
1639 * log chain so it doesn't matter what blk points to.
1641 zil_init_log_chain(zilog, &blk);
1646 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1647 zh->zh_log = lwb->lwb_blk;
1648 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1650 list_remove(&zilog->zl_lwb_list, lwb);
1651 zio_free_zil(spa, txg, &lwb->lwb_blk);
1652 kmem_cache_free(zil_lwb_cache, lwb);
1655 * If we don't have anything left in the lwb list then
1656 * we've had an allocation failure and we need to zero
1657 * out the zil_header blkptr so that we don't end
1658 * up freeing the same block twice.
1660 if (list_head(&zilog->zl_lwb_list) == NULL)
1661 BP_ZERO(&zh->zh_log);
1663 mutex_exit(&zilog->zl_lock);
1669 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1670 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1676 kmem_cache_destroy(zil_lwb_cache);
1680 zil_set_sync(zilog_t *zilog, uint64_t sync)
1682 zilog->zl_sync = sync;
1686 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1688 zilog->zl_logbias = logbias;
1692 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1696 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1698 zilog->zl_header = zh_phys;
1700 zilog->zl_spa = dmu_objset_spa(os);
1701 zilog->zl_dmu_pool = dmu_objset_pool(os);
1702 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1703 zilog->zl_logbias = dmu_objset_logbias(os);
1704 zilog->zl_sync = dmu_objset_syncprop(os);
1705 zilog->zl_next_batch = 1;
1707 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1709 for (int i = 0; i < TXG_SIZE; i++) {
1710 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1711 MUTEX_DEFAULT, NULL);
1714 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1715 offsetof(lwb_t, lwb_node));
1717 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1718 offsetof(itx_t, itx_node));
1720 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1722 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1723 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1725 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1726 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1727 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1728 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1734 zil_free(zilog_t *zilog)
1736 zilog->zl_stop_sync = 1;
1738 ASSERT0(zilog->zl_suspend);
1739 ASSERT0(zilog->zl_suspending);
1741 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1742 list_destroy(&zilog->zl_lwb_list);
1744 avl_destroy(&zilog->zl_vdev_tree);
1745 mutex_destroy(&zilog->zl_vdev_lock);
1747 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1748 list_destroy(&zilog->zl_itx_commit_list);
1750 for (int i = 0; i < TXG_SIZE; i++) {
1752 * It's possible for an itx to be generated that doesn't dirty
1753 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1754 * callback to remove the entry. We remove those here.
1756 * Also free up the ziltest itxs.
1758 if (zilog->zl_itxg[i].itxg_itxs)
1759 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1760 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1763 mutex_destroy(&zilog->zl_lock);
1765 cv_destroy(&zilog->zl_cv_writer);
1766 cv_destroy(&zilog->zl_cv_suspend);
1767 cv_destroy(&zilog->zl_cv_batch[0]);
1768 cv_destroy(&zilog->zl_cv_batch[1]);
1770 kmem_free(zilog, sizeof (zilog_t));
1774 * Open an intent log.
1777 zil_open(objset_t *os, zil_get_data_t *get_data)
1779 zilog_t *zilog = dmu_objset_zil(os);
1781 ASSERT(zilog->zl_clean_taskq == NULL);
1782 ASSERT(zilog->zl_get_data == NULL);
1783 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1785 zilog->zl_get_data = get_data;
1786 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1787 2, 2, TASKQ_PREPOPULATE);
1793 * Close an intent log.
1796 zil_close(zilog_t *zilog)
1801 zil_commit(zilog, 0); /* commit all itx */
1804 * The lwb_max_txg for the stubby lwb will reflect the last activity
1805 * for the zil. After a txg_wait_synced() on the txg we know all the
1806 * callbacks have occurred that may clean the zil. Only then can we
1807 * destroy the zl_clean_taskq.
1809 mutex_enter(&zilog->zl_lock);
1810 lwb = list_tail(&zilog->zl_lwb_list);
1812 txg = lwb->lwb_max_txg;
1813 mutex_exit(&zilog->zl_lock);
1815 txg_wait_synced(zilog->zl_dmu_pool, txg);
1816 ASSERT(!zilog_is_dirty(zilog));
1818 taskq_destroy(zilog->zl_clean_taskq);
1819 zilog->zl_clean_taskq = NULL;
1820 zilog->zl_get_data = NULL;
1823 * We should have only one LWB left on the list; remove it now.
1825 mutex_enter(&zilog->zl_lock);
1826 lwb = list_head(&zilog->zl_lwb_list);
1828 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1829 list_remove(&zilog->zl_lwb_list, lwb);
1830 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1831 kmem_cache_free(zil_lwb_cache, lwb);
1833 mutex_exit(&zilog->zl_lock);
1836 static char *suspend_tag = "zil suspending";
1839 * Suspend an intent log. While in suspended mode, we still honor
1840 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1841 * On old version pools, we suspend the log briefly when taking a
1842 * snapshot so that it will have an empty intent log.
1844 * Long holds are not really intended to be used the way we do here --
1845 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1846 * could fail. Therefore we take pains to only put a long hold if it is
1847 * actually necessary. Fortunately, it will only be necessary if the
1848 * objset is currently mounted (or the ZVOL equivalent). In that case it
1849 * will already have a long hold, so we are not really making things any worse.
1851 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1852 * zvol_state_t), and use their mechanism to prevent their hold from being
1853 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1856 * if cookiep == NULL, this does both the suspend & resume.
1857 * Otherwise, it returns with the dataset "long held", and the cookie
1858 * should be passed into zil_resume().
1861 zil_suspend(const char *osname, void **cookiep)
1865 const zil_header_t *zh;
1868 error = dmu_objset_hold(osname, suspend_tag, &os);
1871 zilog = dmu_objset_zil(os);
1873 mutex_enter(&zilog->zl_lock);
1874 zh = zilog->zl_header;
1876 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1877 mutex_exit(&zilog->zl_lock);
1878 dmu_objset_rele(os, suspend_tag);
1879 return (SET_ERROR(EBUSY));
1883 * Don't put a long hold in the cases where we can avoid it. This
1884 * is when there is no cookie so we are doing a suspend & resume
1885 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1886 * for the suspend because it's already suspended, or there's no ZIL.
1888 if (cookiep == NULL && !zilog->zl_suspending &&
1889 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1890 mutex_exit(&zilog->zl_lock);
1891 dmu_objset_rele(os, suspend_tag);
1895 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1896 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1898 zilog->zl_suspend++;
1900 if (zilog->zl_suspend > 1) {
1902 * Someone else is already suspending it.
1903 * Just wait for them to finish.
1906 while (zilog->zl_suspending)
1907 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1908 mutex_exit(&zilog->zl_lock);
1910 if (cookiep == NULL)
1918 * If there is no pointer to an on-disk block, this ZIL must not
1919 * be active (e.g. filesystem not mounted), so there's nothing
1922 if (BP_IS_HOLE(&zh->zh_log)) {
1923 ASSERT(cookiep != NULL); /* fast path already handled */
1926 mutex_exit(&zilog->zl_lock);
1930 zilog->zl_suspending = B_TRUE;
1931 mutex_exit(&zilog->zl_lock);
1933 zil_commit(zilog, 0);
1935 zil_destroy(zilog, B_FALSE);
1937 mutex_enter(&zilog->zl_lock);
1938 zilog->zl_suspending = B_FALSE;
1939 cv_broadcast(&zilog->zl_cv_suspend);
1940 mutex_exit(&zilog->zl_lock);
1942 if (cookiep == NULL)
1950 zil_resume(void *cookie)
1952 objset_t *os = cookie;
1953 zilog_t *zilog = dmu_objset_zil(os);
1955 mutex_enter(&zilog->zl_lock);
1956 ASSERT(zilog->zl_suspend != 0);
1957 zilog->zl_suspend--;
1958 mutex_exit(&zilog->zl_lock);
1959 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1960 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1963 typedef struct zil_replay_arg {
1964 zil_replay_func_t **zr_replay;
1966 boolean_t zr_byteswap;
1971 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1973 char name[MAXNAMELEN];
1975 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1977 dmu_objset_name(zilog->zl_os, name);
1979 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1980 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1981 (u_longlong_t)lr->lrc_seq,
1982 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1983 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1989 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1991 zil_replay_arg_t *zr = zra;
1992 const zil_header_t *zh = zilog->zl_header;
1993 uint64_t reclen = lr->lrc_reclen;
1994 uint64_t txtype = lr->lrc_txtype;
1997 zilog->zl_replaying_seq = lr->lrc_seq;
1999 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2002 if (lr->lrc_txg < claim_txg) /* already committed */
2005 /* Strip case-insensitive bit, still present in log record */
2008 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2009 return (zil_replay_error(zilog, lr, EINVAL));
2012 * If this record type can be logged out of order, the object
2013 * (lr_foid) may no longer exist. That's legitimate, not an error.
2015 if (TX_OOO(txtype)) {
2016 error = dmu_object_info(zilog->zl_os,
2017 ((lr_ooo_t *)lr)->lr_foid, NULL);
2018 if (error == ENOENT || error == EEXIST)
2023 * Make a copy of the data so we can revise and extend it.
2025 bcopy(lr, zr->zr_lr, reclen);
2028 * If this is a TX_WRITE with a blkptr, suck in the data.
2030 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2031 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2032 zr->zr_lr + reclen);
2034 return (zil_replay_error(zilog, lr, error));
2038 * The log block containing this lr may have been byteswapped
2039 * so that we can easily examine common fields like lrc_txtype.
2040 * However, the log is a mix of different record types, and only the
2041 * replay vectors know how to byteswap their records. Therefore, if
2042 * the lr was byteswapped, undo it before invoking the replay vector.
2044 if (zr->zr_byteswap)
2045 byteswap_uint64_array(zr->zr_lr, reclen);
2048 * We must now do two things atomically: replay this log record,
2049 * and update the log header sequence number to reflect the fact that
2050 * we did so. At the end of each replay function the sequence number
2051 * is updated if we are in replay mode.
2053 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2056 * The DMU's dnode layer doesn't see removes until the txg
2057 * commits, so a subsequent claim can spuriously fail with
2058 * EEXIST. So if we receive any error we try syncing out
2059 * any removes then retry the transaction. Note that we
2060 * specify B_FALSE for byteswap now, so we don't do it twice.
2062 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2063 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2065 return (zil_replay_error(zilog, lr, error));
2072 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2074 zilog->zl_replay_blks++;
2080 * If this dataset has a non-empty intent log, replay it and destroy it.
2083 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2085 zilog_t *zilog = dmu_objset_zil(os);
2086 const zil_header_t *zh = zilog->zl_header;
2087 zil_replay_arg_t zr;
2089 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2090 zil_destroy(zilog, B_TRUE);
2093 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2095 zr.zr_replay = replay_func;
2097 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2098 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2101 * Wait for in-progress removes to sync before starting replay.
2103 txg_wait_synced(zilog->zl_dmu_pool, 0);
2105 zilog->zl_replay = B_TRUE;
2106 zilog->zl_replay_time = ddi_get_lbolt();
2107 ASSERT(zilog->zl_replay_blks == 0);
2108 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2110 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2112 zil_destroy(zilog, B_FALSE);
2113 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2114 zilog->zl_replay = B_FALSE;
2115 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2119 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2121 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2124 if (zilog->zl_replay) {
2125 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2126 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2127 zilog->zl_replaying_seq;
2136 zil_vdev_offline(const char *osname, void *arg)
2140 error = zil_suspend(osname, NULL);
2142 return (SET_ERROR(EEXIST));