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) 2013 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);
650 cmn_err(CE_WARN, "can't open objset for %s", osname);
654 zilog = dmu_objset_zil(os);
655 zh = zil_header_in_syncing_context(zilog);
657 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
658 if (!BP_IS_HOLE(&zh->zh_log))
659 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
660 BP_ZERO(&zh->zh_log);
661 dsl_dataset_dirty(dmu_objset_ds(os), tx);
662 dmu_objset_disown(os, FTAG);
667 * Claim all log blocks if we haven't already done so, and remember
668 * the highest claimed sequence number. This ensures that if we can
669 * read only part of the log now (e.g. due to a missing device),
670 * but we can read the entire log later, we will not try to replay
671 * or destroy beyond the last block we successfully claimed.
673 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
674 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
675 (void) zil_parse(zilog, zil_claim_log_block,
676 zil_claim_log_record, tx, first_txg);
677 zh->zh_claim_txg = first_txg;
678 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
679 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
680 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
681 zh->zh_flags |= ZIL_REPLAY_NEEDED;
682 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
683 dsl_dataset_dirty(dmu_objset_ds(os), tx);
686 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
687 dmu_objset_disown(os, FTAG);
692 * Check the log by walking the log chain.
693 * Checksum errors are ok as they indicate the end of the chain.
694 * Any other error (no device or read failure) returns an error.
697 zil_check_log_chain(const char *osname, void *tx)
706 error = dmu_objset_hold(osname, FTAG, &os);
708 cmn_err(CE_WARN, "can't open objset for %s", osname);
712 zilog = dmu_objset_zil(os);
713 bp = (blkptr_t *)&zilog->zl_header->zh_log;
716 * Check the first block and determine if it's on a log device
717 * which may have been removed or faulted prior to loading this
718 * pool. If so, there's no point in checking the rest of the log
719 * as its content should have already been synced to the pool.
721 if (!BP_IS_HOLE(bp)) {
723 boolean_t valid = B_TRUE;
725 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
726 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
727 if (vd->vdev_islog && vdev_is_dead(vd))
728 valid = vdev_log_state_valid(vd);
729 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
732 dmu_objset_rele(os, FTAG);
738 * Because tx == NULL, zil_claim_log_block() will not actually claim
739 * any blocks, but just determine whether it is possible to do so.
740 * In addition to checking the log chain, zil_claim_log_block()
741 * will invoke zio_claim() with a done func of spa_claim_notify(),
742 * which will update spa_max_claim_txg. See spa_load() for details.
744 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
745 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
747 dmu_objset_rele(os, FTAG);
749 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
753 zil_vdev_compare(const void *x1, const void *x2)
755 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
756 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
767 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
769 avl_tree_t *t = &zilog->zl_vdev_tree;
771 zil_vdev_node_t *zv, zvsearch;
772 int ndvas = BP_GET_NDVAS(bp);
775 if (zfs_nocacheflush)
778 ASSERT(zilog->zl_writer);
781 * Even though we're zl_writer, we still need a lock because the
782 * zl_get_data() callbacks may have dmu_sync() done callbacks
783 * that will run concurrently.
785 mutex_enter(&zilog->zl_vdev_lock);
786 for (i = 0; i < ndvas; i++) {
787 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
788 if (avl_find(t, &zvsearch, &where) == NULL) {
789 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
790 zv->zv_vdev = zvsearch.zv_vdev;
791 avl_insert(t, zv, where);
794 mutex_exit(&zilog->zl_vdev_lock);
798 zil_flush_vdevs(zilog_t *zilog)
800 spa_t *spa = zilog->zl_spa;
801 avl_tree_t *t = &zilog->zl_vdev_tree;
806 ASSERT(zilog->zl_writer);
809 * We don't need zl_vdev_lock here because we're the zl_writer,
810 * and all zl_get_data() callbacks are done.
812 if (avl_numnodes(t) == 0)
815 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
817 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
819 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
820 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
823 kmem_free(zv, sizeof (*zv));
827 * Wait for all the flushes to complete. Not all devices actually
828 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
830 (void) zio_wait(zio);
832 spa_config_exit(spa, SCL_STATE, FTAG);
836 * Function called when a log block write completes
839 zil_lwb_write_done(zio_t *zio)
841 lwb_t *lwb = zio->io_private;
842 zilog_t *zilog = lwb->lwb_zilog;
843 dmu_tx_t *tx = lwb->lwb_tx;
845 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
846 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
847 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
848 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
849 ASSERT(!BP_IS_GANG(zio->io_bp));
850 ASSERT(!BP_IS_HOLE(zio->io_bp));
851 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
854 * Ensure the lwb buffer pointer is cleared before releasing
855 * the txg. If we have had an allocation failure and
856 * the txg is waiting to sync then we want want zil_sync()
857 * to remove the lwb so that it's not picked up as the next new
858 * one in zil_commit_writer(). zil_sync() will only remove
859 * the lwb if lwb_buf is null.
861 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
862 mutex_enter(&zilog->zl_lock);
865 mutex_exit(&zilog->zl_lock);
868 * Now that we've written this log block, we have a stable pointer
869 * to the next block in the chain, so it's OK to let the txg in
870 * which we allocated the next block sync.
876 * Initialize the io for a log block.
879 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
883 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
884 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
885 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
887 if (zilog->zl_root_zio == NULL) {
888 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
891 if (lwb->lwb_zio == NULL) {
892 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
893 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
894 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
895 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
900 * Define a limited set of intent log block sizes.
902 * These must be a multiple of 4KB. Note only the amount used (again
903 * aligned to 4KB) actually gets written. However, we can't always just
904 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
906 uint64_t zil_block_buckets[] = {
907 4096, /* non TX_WRITE */
908 8192+4096, /* data base */
909 32*1024 + 4096, /* NFS writes */
914 * Use the slog as long as the logbias is 'latency' and the current commit size
915 * is less than the limit or the total list size is less than 2X the limit.
916 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
918 uint64_t zil_slog_limit = 1024 * 1024;
919 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
920 (((zilog)->zl_cur_used < zil_slog_limit) || \
921 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
924 * Start a log block write and advance to the next log block.
925 * Calls are serialized.
928 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
932 spa_t *spa = zilog->zl_spa;
936 uint64_t zil_blksz, wsz;
939 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
940 zilc = (zil_chain_t *)lwb->lwb_buf;
941 bp = &zilc->zc_next_blk;
943 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
944 bp = &zilc->zc_next_blk;
947 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
950 * Allocate the next block and save its address in this block
951 * before writing it in order to establish the log chain.
952 * Note that if the allocation of nlwb synced before we wrote
953 * the block that points at it (lwb), we'd leak it if we crashed.
954 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
955 * We dirty the dataset to ensure that zil_sync() will be called
956 * to clean up in the event of allocation failure or I/O failure.
958 tx = dmu_tx_create(zilog->zl_os);
959 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
960 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
961 txg = dmu_tx_get_txg(tx);
966 * Log blocks are pre-allocated. Here we select the size of the next
967 * block, based on size used in the last block.
968 * - first find the smallest bucket that will fit the block from a
969 * limited set of block sizes. This is because it's faster to write
970 * blocks allocated from the same metaslab as they are adjacent or
972 * - next find the maximum from the new suggested size and an array of
973 * previous sizes. This lessens a picket fence effect of wrongly
974 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
977 * Note we only write what is used, but we can't just allocate
978 * the maximum block size because we can exhaust the available
981 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
982 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
984 zil_blksz = zil_block_buckets[i];
985 if (zil_blksz == UINT64_MAX)
986 zil_blksz = SPA_MAXBLOCKSIZE;
987 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
988 for (i = 0; i < ZIL_PREV_BLKS; i++)
989 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
990 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
993 /* pass the old blkptr in order to spread log blocks across devs */
994 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
997 ASSERT3U(bp->blk_birth, ==, txg);
998 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
999 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1002 * Allocate a new log write buffer (lwb).
1004 nlwb = zil_alloc_lwb(zilog, bp, txg);
1006 /* Record the block for later vdev flushing */
1007 zil_add_block(zilog, &lwb->lwb_blk);
1010 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1011 /* For Slim ZIL only write what is used. */
1012 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1013 ASSERT3U(wsz, <=, lwb->lwb_sz);
1014 zio_shrink(lwb->lwb_zio, wsz);
1021 zilc->zc_nused = lwb->lwb_nused;
1022 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1025 * clear unused data for security
1027 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1029 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1032 * If there was an allocation failure then nlwb will be null which
1033 * forces a txg_wait_synced().
1039 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1041 lr_t *lrc = &itx->itx_lr; /* common log record */
1042 lr_write_t *lrw = (lr_write_t *)lrc;
1044 uint64_t txg = lrc->lrc_txg;
1045 uint64_t reclen = lrc->lrc_reclen;
1051 ASSERT(lwb->lwb_buf != NULL);
1052 ASSERT(zilog_is_dirty(zilog) ||
1053 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1055 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1056 dlen = P2ROUNDUP_TYPED(
1057 lrw->lr_length, sizeof (uint64_t), uint64_t);
1059 zilog->zl_cur_used += (reclen + dlen);
1061 zil_lwb_write_init(zilog, lwb);
1064 * If this record won't fit in the current log block, start a new one.
1066 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1067 lwb = zil_lwb_write_start(zilog, lwb);
1070 zil_lwb_write_init(zilog, lwb);
1071 ASSERT(LWB_EMPTY(lwb));
1072 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1073 txg_wait_synced(zilog->zl_dmu_pool, txg);
1078 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1079 bcopy(lrc, lr_buf, reclen);
1080 lrc = (lr_t *)lr_buf;
1081 lrw = (lr_write_t *)lrc;
1084 * If it's a write, fetch the data or get its blkptr as appropriate.
1086 if (lrc->lrc_txtype == TX_WRITE) {
1087 if (txg > spa_freeze_txg(zilog->zl_spa))
1088 txg_wait_synced(zilog->zl_dmu_pool, txg);
1089 if (itx->itx_wr_state != WR_COPIED) {
1094 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1095 dbuf = lr_buf + reclen;
1096 lrw->lr_common.lrc_reclen += dlen;
1098 ASSERT(itx->itx_wr_state == WR_INDIRECT);
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 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
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, KM_SLEEP);
1138 itx->itx_lr.lrc_txtype = txtype;
1139 itx->itx_lr.lrc_reclen = lrsize;
1140 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1141 itx->itx_lr.lrc_seq = 0; /* defensive */
1142 itx->itx_sync = B_TRUE; /* default is synchronous */
1148 zil_itx_destroy(itx_t *itx)
1150 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1154 * Free up the sync and async itxs. The itxs_t has already been detached
1155 * so no locks are needed.
1158 zil_itxg_clean(itxs_t *itxs)
1164 itx_async_node_t *ian;
1166 list = &itxs->i_sync_list;
1167 while ((itx = list_head(list)) != NULL) {
1168 list_remove(list, itx);
1169 kmem_free(itx, offsetof(itx_t, itx_lr) +
1170 itx->itx_lr.lrc_reclen);
1174 t = &itxs->i_async_tree;
1175 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1176 list = &ian->ia_list;
1177 while ((itx = list_head(list)) != NULL) {
1178 list_remove(list, itx);
1179 kmem_free(itx, offsetof(itx_t, itx_lr) +
1180 itx->itx_lr.lrc_reclen);
1183 kmem_free(ian, sizeof (itx_async_node_t));
1187 kmem_free(itxs, sizeof (itxs_t));
1191 zil_aitx_compare(const void *x1, const void *x2)
1193 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1194 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1205 * Remove all async itx with the given oid.
1208 zil_remove_async(zilog_t *zilog, uint64_t oid)
1211 itx_async_node_t *ian;
1218 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1220 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1223 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1225 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1226 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1228 mutex_enter(&itxg->itxg_lock);
1229 if (itxg->itxg_txg != txg) {
1230 mutex_exit(&itxg->itxg_lock);
1235 * Locate the object node and append its list.
1237 t = &itxg->itxg_itxs->i_async_tree;
1238 ian = avl_find(t, &oid, &where);
1240 list_move_tail(&clean_list, &ian->ia_list);
1241 mutex_exit(&itxg->itxg_lock);
1243 while ((itx = list_head(&clean_list)) != NULL) {
1244 list_remove(&clean_list, itx);
1245 kmem_free(itx, offsetof(itx_t, itx_lr) +
1246 itx->itx_lr.lrc_reclen);
1248 list_destroy(&clean_list);
1252 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1256 itxs_t *itxs, *clean = NULL;
1259 * Object ids can be re-instantiated in the next txg so
1260 * remove any async transactions to avoid future leaks.
1261 * This can happen if a fsync occurs on the re-instantiated
1262 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1263 * the new file data and flushes a write record for the old object.
1265 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1266 zil_remove_async(zilog, itx->itx_oid);
1269 * Ensure the data of a renamed file is committed before the rename.
1271 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1272 zil_async_to_sync(zilog, itx->itx_oid);
1274 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1277 txg = dmu_tx_get_txg(tx);
1279 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1280 mutex_enter(&itxg->itxg_lock);
1281 itxs = itxg->itxg_itxs;
1282 if (itxg->itxg_txg != txg) {
1285 * The zil_clean callback hasn't got around to cleaning
1286 * this itxg. Save the itxs for release below.
1287 * This should be rare.
1289 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1291 clean = itxg->itxg_itxs;
1293 ASSERT(itxg->itxg_sod == 0);
1294 itxg->itxg_txg = txg;
1295 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1297 list_create(&itxs->i_sync_list, sizeof (itx_t),
1298 offsetof(itx_t, itx_node));
1299 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1300 sizeof (itx_async_node_t),
1301 offsetof(itx_async_node_t, ia_node));
1303 if (itx->itx_sync) {
1304 list_insert_tail(&itxs->i_sync_list, itx);
1305 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1306 itxg->itxg_sod += itx->itx_sod;
1308 avl_tree_t *t = &itxs->i_async_tree;
1309 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1310 itx_async_node_t *ian;
1313 ian = avl_find(t, &foid, &where);
1315 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1316 list_create(&ian->ia_list, sizeof (itx_t),
1317 offsetof(itx_t, itx_node));
1318 ian->ia_foid = foid;
1319 avl_insert(t, ian, where);
1321 list_insert_tail(&ian->ia_list, itx);
1324 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1325 zilog_dirty(zilog, txg);
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. We should only do this after we
1336 * have written out the uberblocks (i.e. txg has been comitted) so that
1337 * don't inadvertently clean out in-memory log records that would be required
1341 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1343 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1346 mutex_enter(&itxg->itxg_lock);
1347 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1348 mutex_exit(&itxg->itxg_lock);
1351 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1352 ASSERT(itxg->itxg_txg != 0);
1353 ASSERT(zilog->zl_clean_taskq != NULL);
1354 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1356 clean_me = itxg->itxg_itxs;
1357 itxg->itxg_itxs = NULL;
1359 mutex_exit(&itxg->itxg_lock);
1361 * Preferably start a task queue to free up the old itxs but
1362 * if taskq_dispatch can't allocate resources to do that then
1363 * free it in-line. This should be rare. Note, using TQ_SLEEP
1364 * created a bad performance problem.
1366 if (taskq_dispatch(zilog->zl_clean_taskq,
1367 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1368 zil_itxg_clean(clean_me);
1372 * Get the list of itxs to commit into zl_itx_commit_list.
1375 zil_get_commit_list(zilog_t *zilog)
1378 list_t *commit_list = &zilog->zl_itx_commit_list;
1379 uint64_t push_sod = 0;
1381 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1384 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1386 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1387 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1389 mutex_enter(&itxg->itxg_lock);
1390 if (itxg->itxg_txg != txg) {
1391 mutex_exit(&itxg->itxg_lock);
1395 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1396 push_sod += itxg->itxg_sod;
1399 mutex_exit(&itxg->itxg_lock);
1401 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1405 * Move the async itxs for a specified object to commit into sync lists.
1408 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1411 itx_async_node_t *ian;
1415 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1418 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1420 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1421 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1423 mutex_enter(&itxg->itxg_lock);
1424 if (itxg->itxg_txg != txg) {
1425 mutex_exit(&itxg->itxg_lock);
1430 * If a foid is specified then find that node and append its
1431 * list. Otherwise walk the tree appending all the lists
1432 * to the sync list. We add to the end rather than the
1433 * beginning to ensure the create has happened.
1435 t = &itxg->itxg_itxs->i_async_tree;
1437 ian = avl_find(t, &foid, &where);
1439 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1443 void *cookie = NULL;
1445 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1446 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1448 list_destroy(&ian->ia_list);
1449 kmem_free(ian, sizeof (itx_async_node_t));
1452 mutex_exit(&itxg->itxg_lock);
1457 zil_commit_writer(zilog_t *zilog)
1462 spa_t *spa = zilog->zl_spa;
1465 ASSERT(zilog->zl_root_zio == NULL);
1467 mutex_exit(&zilog->zl_lock);
1469 zil_get_commit_list(zilog);
1472 * Return if there's nothing to commit before we dirty the fs by
1473 * calling zil_create().
1475 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1476 mutex_enter(&zilog->zl_lock);
1480 if (zilog->zl_suspend) {
1483 lwb = list_tail(&zilog->zl_lwb_list);
1485 lwb = zil_create(zilog);
1488 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1489 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1490 txg = itx->itx_lr.lrc_txg;
1493 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1494 lwb = zil_lwb_commit(zilog, itx, lwb);
1495 list_remove(&zilog->zl_itx_commit_list, itx);
1496 kmem_free(itx, offsetof(itx_t, itx_lr)
1497 + itx->itx_lr.lrc_reclen);
1499 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1501 /* write the last block out */
1502 if (lwb != NULL && lwb->lwb_zio != NULL)
1503 lwb = zil_lwb_write_start(zilog, lwb);
1505 zilog->zl_cur_used = 0;
1508 * Wait if necessary for the log blocks to be on stable storage.
1510 if (zilog->zl_root_zio) {
1511 error = zio_wait(zilog->zl_root_zio);
1512 zilog->zl_root_zio = NULL;
1513 zil_flush_vdevs(zilog);
1516 if (error || lwb == NULL)
1517 txg_wait_synced(zilog->zl_dmu_pool, 0);
1519 mutex_enter(&zilog->zl_lock);
1522 * Remember the highest committed log sequence number for ztest.
1523 * We only update this value when all the log writes succeeded,
1524 * because ztest wants to ASSERT that it got the whole log chain.
1526 if (error == 0 && lwb != NULL)
1527 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1531 * Commit zfs transactions to stable storage.
1532 * If foid is 0 push out all transactions, otherwise push only those
1533 * for that object or might reference that object.
1535 * itxs are committed in batches. In a heavily stressed zil there will be
1536 * a commit writer thread who is writing out a bunch of itxs to the log
1537 * for a set of committing threads (cthreads) in the same batch as the writer.
1538 * Those cthreads are all waiting on the same cv for that batch.
1540 * There will also be a different and growing batch of threads that are
1541 * waiting to commit (qthreads). When the committing batch completes
1542 * a transition occurs such that the cthreads exit and the qthreads become
1543 * cthreads. One of the new cthreads becomes the writer thread for the
1544 * batch. Any new threads arriving become new qthreads.
1546 * Only 2 condition variables are needed and there's no transition
1547 * between the two cvs needed. They just flip-flop between qthreads
1550 * Using this scheme we can efficiently wakeup up only those threads
1551 * that have been committed.
1554 zil_commit(zilog_t *zilog, uint64_t foid)
1558 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1561 /* move the async itxs for the foid to the sync queues */
1562 zil_async_to_sync(zilog, foid);
1564 mutex_enter(&zilog->zl_lock);
1565 mybatch = zilog->zl_next_batch;
1566 while (zilog->zl_writer) {
1567 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1568 if (mybatch <= zilog->zl_com_batch) {
1569 mutex_exit(&zilog->zl_lock);
1574 zilog->zl_next_batch++;
1575 zilog->zl_writer = B_TRUE;
1576 zil_commit_writer(zilog);
1577 zilog->zl_com_batch = mybatch;
1578 zilog->zl_writer = B_FALSE;
1579 mutex_exit(&zilog->zl_lock);
1581 /* wake up one thread to become the next writer */
1582 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1584 /* wake up all threads waiting for this batch to be committed */
1585 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1589 * Called in syncing context to free committed log blocks and update log header.
1592 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1594 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1595 uint64_t txg = dmu_tx_get_txg(tx);
1596 spa_t *spa = zilog->zl_spa;
1597 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1601 * We don't zero out zl_destroy_txg, so make sure we don't try
1602 * to destroy it twice.
1604 if (spa_sync_pass(spa) != 1)
1607 mutex_enter(&zilog->zl_lock);
1609 ASSERT(zilog->zl_stop_sync == 0);
1611 if (*replayed_seq != 0) {
1612 ASSERT(zh->zh_replay_seq < *replayed_seq);
1613 zh->zh_replay_seq = *replayed_seq;
1617 if (zilog->zl_destroy_txg == txg) {
1618 blkptr_t blk = zh->zh_log;
1620 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1622 bzero(zh, sizeof (zil_header_t));
1623 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1625 if (zilog->zl_keep_first) {
1627 * If this block was part of log chain that couldn't
1628 * be claimed because a device was missing during
1629 * zil_claim(), but that device later returns,
1630 * then this block could erroneously appear valid.
1631 * To guard against this, assign a new GUID to the new
1632 * log chain so it doesn't matter what blk points to.
1634 zil_init_log_chain(zilog, &blk);
1639 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1640 zh->zh_log = lwb->lwb_blk;
1641 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1643 list_remove(&zilog->zl_lwb_list, lwb);
1644 zio_free_zil(spa, txg, &lwb->lwb_blk);
1645 kmem_cache_free(zil_lwb_cache, lwb);
1648 * If we don't have anything left in the lwb list then
1649 * we've had an allocation failure and we need to zero
1650 * out the zil_header blkptr so that we don't end
1651 * up freeing the same block twice.
1653 if (list_head(&zilog->zl_lwb_list) == NULL)
1654 BP_ZERO(&zh->zh_log);
1656 mutex_exit(&zilog->zl_lock);
1662 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1663 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1669 kmem_cache_destroy(zil_lwb_cache);
1673 zil_set_sync(zilog_t *zilog, uint64_t sync)
1675 zilog->zl_sync = sync;
1679 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1681 zilog->zl_logbias = logbias;
1685 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1689 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1691 zilog->zl_header = zh_phys;
1693 zilog->zl_spa = dmu_objset_spa(os);
1694 zilog->zl_dmu_pool = dmu_objset_pool(os);
1695 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1696 zilog->zl_logbias = dmu_objset_logbias(os);
1697 zilog->zl_sync = dmu_objset_syncprop(os);
1698 zilog->zl_next_batch = 1;
1700 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1702 for (int i = 0; i < TXG_SIZE; i++) {
1703 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1704 MUTEX_DEFAULT, NULL);
1707 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1708 offsetof(lwb_t, lwb_node));
1710 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1711 offsetof(itx_t, itx_node));
1713 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1715 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1716 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1718 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1719 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1720 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1721 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1727 zil_free(zilog_t *zilog)
1729 zilog->zl_stop_sync = 1;
1731 ASSERT0(zilog->zl_suspend);
1732 ASSERT0(zilog->zl_suspending);
1734 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1735 list_destroy(&zilog->zl_lwb_list);
1737 avl_destroy(&zilog->zl_vdev_tree);
1738 mutex_destroy(&zilog->zl_vdev_lock);
1740 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1741 list_destroy(&zilog->zl_itx_commit_list);
1743 for (int i = 0; i < TXG_SIZE; i++) {
1745 * It's possible for an itx to be generated that doesn't dirty
1746 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1747 * callback to remove the entry. We remove those here.
1749 * Also free up the ziltest itxs.
1751 if (zilog->zl_itxg[i].itxg_itxs)
1752 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1753 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1756 mutex_destroy(&zilog->zl_lock);
1758 cv_destroy(&zilog->zl_cv_writer);
1759 cv_destroy(&zilog->zl_cv_suspend);
1760 cv_destroy(&zilog->zl_cv_batch[0]);
1761 cv_destroy(&zilog->zl_cv_batch[1]);
1763 kmem_free(zilog, sizeof (zilog_t));
1767 * Open an intent log.
1770 zil_open(objset_t *os, zil_get_data_t *get_data)
1772 zilog_t *zilog = dmu_objset_zil(os);
1774 ASSERT(zilog->zl_clean_taskq == NULL);
1775 ASSERT(zilog->zl_get_data == NULL);
1776 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1778 zilog->zl_get_data = get_data;
1779 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1780 2, 2, TASKQ_PREPOPULATE);
1786 * Close an intent log.
1789 zil_close(zilog_t *zilog)
1794 zil_commit(zilog, 0); /* commit all itx */
1797 * The lwb_max_txg for the stubby lwb will reflect the last activity
1798 * for the zil. After a txg_wait_synced() on the txg we know all the
1799 * callbacks have occurred that may clean the zil. Only then can we
1800 * destroy the zl_clean_taskq.
1802 mutex_enter(&zilog->zl_lock);
1803 lwb = list_tail(&zilog->zl_lwb_list);
1805 txg = lwb->lwb_max_txg;
1806 mutex_exit(&zilog->zl_lock);
1808 txg_wait_synced(zilog->zl_dmu_pool, txg);
1809 ASSERT(!zilog_is_dirty(zilog));
1811 taskq_destroy(zilog->zl_clean_taskq);
1812 zilog->zl_clean_taskq = NULL;
1813 zilog->zl_get_data = NULL;
1816 * We should have only one LWB left on the list; remove it now.
1818 mutex_enter(&zilog->zl_lock);
1819 lwb = list_head(&zilog->zl_lwb_list);
1821 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1822 list_remove(&zilog->zl_lwb_list, lwb);
1823 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1824 kmem_cache_free(zil_lwb_cache, lwb);
1826 mutex_exit(&zilog->zl_lock);
1829 static char *suspend_tag = "zil suspending";
1832 * Suspend an intent log. While in suspended mode, we still honor
1833 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1834 * On old version pools, we suspend the log briefly when taking a
1835 * snapshot so that it will have an empty intent log.
1837 * Long holds are not really intended to be used the way we do here --
1838 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1839 * could fail. Therefore we take pains to only put a long hold if it is
1840 * actually necessary. Fortunately, it will only be necessary if the
1841 * objset is currently mounted (or the ZVOL equivalent). In that case it
1842 * will already have a long hold, so we are not really making things any worse.
1844 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1845 * zvol_state_t), and use their mechanism to prevent their hold from being
1846 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1849 * if cookiep == NULL, this does both the suspend & resume.
1850 * Otherwise, it returns with the dataset "long held", and the cookie
1851 * should be passed into zil_resume().
1854 zil_suspend(const char *osname, void **cookiep)
1858 const zil_header_t *zh;
1861 error = dmu_objset_hold(osname, suspend_tag, &os);
1864 zilog = dmu_objset_zil(os);
1866 mutex_enter(&zilog->zl_lock);
1867 zh = zilog->zl_header;
1869 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1870 mutex_exit(&zilog->zl_lock);
1871 dmu_objset_rele(os, suspend_tag);
1872 return (SET_ERROR(EBUSY));
1876 * Don't put a long hold in the cases where we can avoid it. This
1877 * is when there is no cookie so we are doing a suspend & resume
1878 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1879 * for the suspend because it's already suspended, or there's no ZIL.
1881 if (cookiep == NULL && !zilog->zl_suspending &&
1882 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1883 mutex_exit(&zilog->zl_lock);
1884 dmu_objset_rele(os, suspend_tag);
1888 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1889 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1891 zilog->zl_suspend++;
1893 if (zilog->zl_suspend > 1) {
1895 * Someone else is already suspending it.
1896 * Just wait for them to finish.
1899 while (zilog->zl_suspending)
1900 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1901 mutex_exit(&zilog->zl_lock);
1903 if (cookiep == NULL)
1911 * If there is no pointer to an on-disk block, this ZIL must not
1912 * be active (e.g. filesystem not mounted), so there's nothing
1915 if (BP_IS_HOLE(&zh->zh_log)) {
1916 ASSERT(cookiep != NULL); /* fast path already handled */
1919 mutex_exit(&zilog->zl_lock);
1923 zilog->zl_suspending = B_TRUE;
1924 mutex_exit(&zilog->zl_lock);
1926 zil_commit(zilog, 0);
1928 zil_destroy(zilog, B_FALSE);
1930 mutex_enter(&zilog->zl_lock);
1931 zilog->zl_suspending = B_FALSE;
1932 cv_broadcast(&zilog->zl_cv_suspend);
1933 mutex_exit(&zilog->zl_lock);
1935 if (cookiep == NULL)
1943 zil_resume(void *cookie)
1945 objset_t *os = cookie;
1946 zilog_t *zilog = dmu_objset_zil(os);
1948 mutex_enter(&zilog->zl_lock);
1949 ASSERT(zilog->zl_suspend != 0);
1950 zilog->zl_suspend--;
1951 mutex_exit(&zilog->zl_lock);
1952 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1953 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1956 typedef struct zil_replay_arg {
1957 zil_replay_func_t **zr_replay;
1959 boolean_t zr_byteswap;
1964 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1966 char name[MAXNAMELEN];
1968 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1970 dmu_objset_name(zilog->zl_os, name);
1972 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1973 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1974 (u_longlong_t)lr->lrc_seq,
1975 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1976 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1982 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1984 zil_replay_arg_t *zr = zra;
1985 const zil_header_t *zh = zilog->zl_header;
1986 uint64_t reclen = lr->lrc_reclen;
1987 uint64_t txtype = lr->lrc_txtype;
1990 zilog->zl_replaying_seq = lr->lrc_seq;
1992 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1995 if (lr->lrc_txg < claim_txg) /* already committed */
1998 /* Strip case-insensitive bit, still present in log record */
2001 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2002 return (zil_replay_error(zilog, lr, EINVAL));
2005 * If this record type can be logged out of order, the object
2006 * (lr_foid) may no longer exist. That's legitimate, not an error.
2008 if (TX_OOO(txtype)) {
2009 error = dmu_object_info(zilog->zl_os,
2010 ((lr_ooo_t *)lr)->lr_foid, NULL);
2011 if (error == ENOENT || error == EEXIST)
2016 * Make a copy of the data so we can revise and extend it.
2018 bcopy(lr, zr->zr_lr, reclen);
2021 * If this is a TX_WRITE with a blkptr, suck in the data.
2023 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2024 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2025 zr->zr_lr + reclen);
2027 return (zil_replay_error(zilog, lr, error));
2031 * The log block containing this lr may have been byteswapped
2032 * so that we can easily examine common fields like lrc_txtype.
2033 * However, the log is a mix of different record types, and only the
2034 * replay vectors know how to byteswap their records. Therefore, if
2035 * the lr was byteswapped, undo it before invoking the replay vector.
2037 if (zr->zr_byteswap)
2038 byteswap_uint64_array(zr->zr_lr, reclen);
2041 * We must now do two things atomically: replay this log record,
2042 * and update the log header sequence number to reflect the fact that
2043 * we did so. At the end of each replay function the sequence number
2044 * is updated if we are in replay mode.
2046 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2049 * The DMU's dnode layer doesn't see removes until the txg
2050 * commits, so a subsequent claim can spuriously fail with
2051 * EEXIST. So if we receive any error we try syncing out
2052 * any removes then retry the transaction. Note that we
2053 * specify B_FALSE for byteswap now, so we don't do it twice.
2055 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2056 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2058 return (zil_replay_error(zilog, lr, error));
2065 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2067 zilog->zl_replay_blks++;
2073 * If this dataset has a non-empty intent log, replay it and destroy it.
2076 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2078 zilog_t *zilog = dmu_objset_zil(os);
2079 const zil_header_t *zh = zilog->zl_header;
2080 zil_replay_arg_t zr;
2082 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2083 zil_destroy(zilog, B_TRUE);
2086 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2088 zr.zr_replay = replay_func;
2090 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2091 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2094 * Wait for in-progress removes to sync before starting replay.
2096 txg_wait_synced(zilog->zl_dmu_pool, 0);
2098 zilog->zl_replay = B_TRUE;
2099 zilog->zl_replay_time = ddi_get_lbolt();
2100 ASSERT(zilog->zl_replay_blks == 0);
2101 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2103 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2105 zil_destroy(zilog, B_FALSE);
2106 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2107 zilog->zl_replay = B_FALSE;
2108 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2112 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2114 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2117 if (zilog->zl_replay) {
2118 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2119 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2120 zilog->zl_replaying_seq;
2129 zil_vdev_offline(const char *osname, void *arg)
2133 error = zil_suspend(osname, NULL);
2135 return (SET_ERROR(EEXIST));