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;
152 const dva_t *dva = BP_IDENTITY(bp);
156 if (avl_find(t, dva, &where) != NULL)
157 return (SET_ERROR(EEXIST));
159 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
161 avl_insert(t, zn, where);
166 static zil_header_t *
167 zil_header_in_syncing_context(zilog_t *zilog)
169 return ((zil_header_t *)zilog->zl_header);
173 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
175 zio_cksum_t *zc = &bp->blk_cksum;
177 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
178 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
179 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
180 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
184 * Read a log block and make sure it's valid.
187 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
190 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
191 uint32_t aflags = ARC_WAIT;
192 arc_buf_t *abuf = NULL;
196 if (zilog->zl_header->zh_claim_txg == 0)
197 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
199 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
200 zio_flags |= ZIO_FLAG_SPECULATIVE;
202 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
203 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
205 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
206 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
209 zio_cksum_t cksum = bp->blk_cksum;
212 * Validate the checksummed log block.
214 * Sequence numbers should be... sequential. The checksum
215 * verifier for the next block should be bp's checksum plus 1.
217 * Also check the log chain linkage and size used.
219 cksum.zc_word[ZIL_ZC_SEQ]++;
221 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
222 zil_chain_t *zilc = abuf->b_data;
223 char *lr = (char *)(zilc + 1);
224 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
226 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
227 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
228 error = SET_ERROR(ECKSUM);
231 *end = (char *)dst + len;
232 *nbp = zilc->zc_next_blk;
235 char *lr = abuf->b_data;
236 uint64_t size = BP_GET_LSIZE(bp);
237 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
239 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
240 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
241 (zilc->zc_nused > (size - sizeof (*zilc)))) {
242 error = SET_ERROR(ECKSUM);
244 bcopy(lr, dst, zilc->zc_nused);
245 *end = (char *)dst + zilc->zc_nused;
246 *nbp = zilc->zc_next_blk;
250 VERIFY(arc_buf_remove_ref(abuf, &abuf));
257 * Read a TX_WRITE log data block.
260 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
262 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
263 const blkptr_t *bp = &lr->lr_blkptr;
264 uint32_t aflags = ARC_WAIT;
265 arc_buf_t *abuf = NULL;
269 if (BP_IS_HOLE(bp)) {
271 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
275 if (zilog->zl_header->zh_claim_txg == 0)
276 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
278 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
279 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
281 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
282 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
286 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
287 (void) arc_buf_remove_ref(abuf, &abuf);
294 * Parse the intent log, and call parse_func for each valid record within.
297 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
298 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
300 const zil_header_t *zh = zilog->zl_header;
301 boolean_t claimed = !!zh->zh_claim_txg;
302 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
303 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
304 uint64_t max_blk_seq = 0;
305 uint64_t max_lr_seq = 0;
306 uint64_t blk_count = 0;
307 uint64_t lr_count = 0;
308 blkptr_t blk, next_blk;
313 * Old logs didn't record the maximum zh_claim_lr_seq.
315 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
316 claim_lr_seq = UINT64_MAX;
319 * Starting at the block pointed to by zh_log we read the log chain.
320 * For each block in the chain we strongly check that block to
321 * ensure its validity. We stop when an invalid block is found.
322 * For each block pointer in the chain we call parse_blk_func().
323 * For each record in each valid block we call parse_lr_func().
324 * If the log has been claimed, stop if we encounter a sequence
325 * number greater than the highest claimed sequence number.
327 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
328 zil_bp_tree_init(zilog);
330 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
331 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
335 if (blk_seq > claim_blk_seq)
337 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
339 ASSERT3U(max_blk_seq, <, blk_seq);
340 max_blk_seq = blk_seq;
343 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
346 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
350 for (lrp = lrbuf; lrp < end; lrp += reclen) {
351 lr_t *lr = (lr_t *)lrp;
352 reclen = lr->lrc_reclen;
353 ASSERT3U(reclen, >=, sizeof (lr_t));
354 if (lr->lrc_seq > claim_lr_seq)
356 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
358 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
359 max_lr_seq = lr->lrc_seq;
364 zilog->zl_parse_error = error;
365 zilog->zl_parse_blk_seq = max_blk_seq;
366 zilog->zl_parse_lr_seq = max_lr_seq;
367 zilog->zl_parse_blk_count = blk_count;
368 zilog->zl_parse_lr_count = lr_count;
370 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
371 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
373 zil_bp_tree_fini(zilog);
374 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
380 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
383 * Claim log block if not already committed and not already claimed.
384 * If tx == NULL, just verify that the block is claimable.
386 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
387 zil_bp_tree_add(zilog, bp) != 0)
390 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
391 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
392 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
396 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
398 lr_write_t *lr = (lr_write_t *)lrc;
401 if (lrc->lrc_txtype != TX_WRITE)
405 * If the block is not readable, don't claim it. This can happen
406 * in normal operation when a log block is written to disk before
407 * some of the dmu_sync() blocks it points to. In this case, the
408 * transaction cannot have been committed to anyone (we would have
409 * waited for all writes to be stable first), so it is semantically
410 * correct to declare this the end of the log.
412 if (lr->lr_blkptr.blk_birth >= first_txg &&
413 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
415 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
420 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
422 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
428 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
430 lr_write_t *lr = (lr_write_t *)lrc;
431 blkptr_t *bp = &lr->lr_blkptr;
434 * If we previously claimed it, we need to free it.
436 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
437 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
439 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
445 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
449 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
450 lwb->lwb_zilog = zilog;
452 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
453 lwb->lwb_max_txg = txg;
456 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
457 lwb->lwb_nused = sizeof (zil_chain_t);
458 lwb->lwb_sz = BP_GET_LSIZE(bp);
461 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
464 mutex_enter(&zilog->zl_lock);
465 list_insert_tail(&zilog->zl_lwb_list, lwb);
466 mutex_exit(&zilog->zl_lock);
472 * Called when we create in-memory log transactions so that we know
473 * to cleanup the itxs at the end of spa_sync().
476 zilog_dirty(zilog_t *zilog, uint64_t txg)
478 dsl_pool_t *dp = zilog->zl_dmu_pool;
479 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
481 if (dsl_dataset_is_snapshot(ds))
482 panic("dirtying snapshot!");
484 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
485 /* up the hold count until we can be written out */
486 dmu_buf_add_ref(ds->ds_dbuf, zilog);
491 zilog_is_dirty(zilog_t *zilog)
493 dsl_pool_t *dp = zilog->zl_dmu_pool;
495 for (int t = 0; t < TXG_SIZE; t++) {
496 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
503 * Create an on-disk intent log.
506 zil_create(zilog_t *zilog)
508 const zil_header_t *zh = zilog->zl_header;
516 * Wait for any previous destroy to complete.
518 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
520 ASSERT(zh->zh_claim_txg == 0);
521 ASSERT(zh->zh_replay_seq == 0);
526 * Allocate an initial log block if:
527 * - there isn't one already
528 * - the existing block is the wrong endianess
530 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
531 tx = dmu_tx_create(zilog->zl_os);
532 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
533 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
534 txg = dmu_tx_get_txg(tx);
536 if (!BP_IS_HOLE(&blk)) {
537 zio_free_zil(zilog->zl_spa, txg, &blk);
541 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
542 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
545 zil_init_log_chain(zilog, &blk);
549 * Allocate a log write buffer (lwb) for the first log block.
552 lwb = zil_alloc_lwb(zilog, &blk, txg);
555 * If we just allocated the first log block, commit our transaction
556 * and wait for zil_sync() to stuff the block poiner into zh_log.
557 * (zh is part of the MOS, so we cannot modify it in open context.)
561 txg_wait_synced(zilog->zl_dmu_pool, txg);
564 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
570 * In one tx, free all log blocks and clear the log header.
571 * If keep_first is set, then we're replaying a log with no content.
572 * We want to keep the first block, however, so that the first
573 * synchronous transaction doesn't require a txg_wait_synced()
574 * in zil_create(). We don't need to txg_wait_synced() here either
575 * when keep_first is set, because both zil_create() and zil_destroy()
576 * will wait for any in-progress destroys to complete.
579 zil_destroy(zilog_t *zilog, boolean_t keep_first)
581 const zil_header_t *zh = zilog->zl_header;
587 * Wait for any previous destroy to complete.
589 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
591 zilog->zl_old_header = *zh; /* debugging aid */
593 if (BP_IS_HOLE(&zh->zh_log))
596 tx = dmu_tx_create(zilog->zl_os);
597 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
598 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
599 txg = dmu_tx_get_txg(tx);
601 mutex_enter(&zilog->zl_lock);
603 ASSERT3U(zilog->zl_destroy_txg, <, txg);
604 zilog->zl_destroy_txg = txg;
605 zilog->zl_keep_first = keep_first;
607 if (!list_is_empty(&zilog->zl_lwb_list)) {
608 ASSERT(zh->zh_claim_txg == 0);
610 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
611 list_remove(&zilog->zl_lwb_list, lwb);
612 if (lwb->lwb_buf != NULL)
613 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
614 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
615 kmem_cache_free(zil_lwb_cache, lwb);
617 } else if (!keep_first) {
618 zil_destroy_sync(zilog, tx);
620 mutex_exit(&zilog->zl_lock);
626 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
628 ASSERT(list_is_empty(&zilog->zl_lwb_list));
629 (void) zil_parse(zilog, zil_free_log_block,
630 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
634 zil_claim(const char *osname, void *txarg)
636 dmu_tx_t *tx = txarg;
637 uint64_t first_txg = dmu_tx_get_txg(tx);
643 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
645 cmn_err(CE_WARN, "can't open objset for %s", osname);
649 zilog = dmu_objset_zil(os);
650 zh = zil_header_in_syncing_context(zilog);
652 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
653 if (!BP_IS_HOLE(&zh->zh_log))
654 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
655 BP_ZERO(&zh->zh_log);
656 dsl_dataset_dirty(dmu_objset_ds(os), tx);
657 dmu_objset_disown(os, FTAG);
662 * Claim all log blocks if we haven't already done so, and remember
663 * the highest claimed sequence number. This ensures that if we can
664 * read only part of the log now (e.g. due to a missing device),
665 * but we can read the entire log later, we will not try to replay
666 * or destroy beyond the last block we successfully claimed.
668 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
669 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
670 (void) zil_parse(zilog, zil_claim_log_block,
671 zil_claim_log_record, tx, first_txg);
672 zh->zh_claim_txg = first_txg;
673 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
674 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
675 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
676 zh->zh_flags |= ZIL_REPLAY_NEEDED;
677 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
678 dsl_dataset_dirty(dmu_objset_ds(os), tx);
681 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
682 dmu_objset_disown(os, FTAG);
687 * Check the log by walking the log chain.
688 * Checksum errors are ok as they indicate the end of the chain.
689 * Any other error (no device or read failure) returns an error.
692 zil_check_log_chain(const char *osname, void *tx)
701 error = dmu_objset_hold(osname, FTAG, &os);
703 cmn_err(CE_WARN, "can't open objset for %s", osname);
707 zilog = dmu_objset_zil(os);
708 bp = (blkptr_t *)&zilog->zl_header->zh_log;
711 * Check the first block and determine if it's on a log device
712 * which may have been removed or faulted prior to loading this
713 * pool. If so, there's no point in checking the rest of the log
714 * as its content should have already been synced to the pool.
716 if (!BP_IS_HOLE(bp)) {
718 boolean_t valid = B_TRUE;
720 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
721 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
722 if (vd->vdev_islog && vdev_is_dead(vd))
723 valid = vdev_log_state_valid(vd);
724 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
727 dmu_objset_rele(os, FTAG);
733 * Because tx == NULL, zil_claim_log_block() will not actually claim
734 * any blocks, but just determine whether it is possible to do so.
735 * In addition to checking the log chain, zil_claim_log_block()
736 * will invoke zio_claim() with a done func of spa_claim_notify(),
737 * which will update spa_max_claim_txg. See spa_load() for details.
739 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
740 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
742 dmu_objset_rele(os, FTAG);
744 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
748 zil_vdev_compare(const void *x1, const void *x2)
750 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
751 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
762 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
764 avl_tree_t *t = &zilog->zl_vdev_tree;
766 zil_vdev_node_t *zv, zvsearch;
767 int ndvas = BP_GET_NDVAS(bp);
770 if (zfs_nocacheflush)
773 ASSERT(zilog->zl_writer);
776 * Even though we're zl_writer, we still need a lock because the
777 * zl_get_data() callbacks may have dmu_sync() done callbacks
778 * that will run concurrently.
780 mutex_enter(&zilog->zl_vdev_lock);
781 for (i = 0; i < ndvas; i++) {
782 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
783 if (avl_find(t, &zvsearch, &where) == NULL) {
784 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
785 zv->zv_vdev = zvsearch.zv_vdev;
786 avl_insert(t, zv, where);
789 mutex_exit(&zilog->zl_vdev_lock);
793 zil_flush_vdevs(zilog_t *zilog)
795 spa_t *spa = zilog->zl_spa;
796 avl_tree_t *t = &zilog->zl_vdev_tree;
801 ASSERT(zilog->zl_writer);
804 * We don't need zl_vdev_lock here because we're the zl_writer,
805 * and all zl_get_data() callbacks are done.
807 if (avl_numnodes(t) == 0)
810 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
812 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
814 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
815 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
818 kmem_free(zv, sizeof (*zv));
822 * Wait for all the flushes to complete. Not all devices actually
823 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
825 (void) zio_wait(zio);
827 spa_config_exit(spa, SCL_STATE, FTAG);
831 * Function called when a log block write completes
834 zil_lwb_write_done(zio_t *zio)
836 lwb_t *lwb = zio->io_private;
837 zilog_t *zilog = lwb->lwb_zilog;
838 dmu_tx_t *tx = lwb->lwb_tx;
840 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
841 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
842 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
843 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
844 ASSERT(!BP_IS_GANG(zio->io_bp));
845 ASSERT(!BP_IS_HOLE(zio->io_bp));
846 ASSERT(zio->io_bp->blk_fill == 0);
849 * Ensure the lwb buffer pointer is cleared before releasing
850 * the txg. If we have had an allocation failure and
851 * the txg is waiting to sync then we want want zil_sync()
852 * to remove the lwb so that it's not picked up as the next new
853 * one in zil_commit_writer(). zil_sync() will only remove
854 * the lwb if lwb_buf is null.
856 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
857 mutex_enter(&zilog->zl_lock);
860 mutex_exit(&zilog->zl_lock);
863 * Now that we've written this log block, we have a stable pointer
864 * to the next block in the chain, so it's OK to let the txg in
865 * which we allocated the next block sync.
871 * Initialize the io for a log block.
874 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
878 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
879 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
880 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
882 if (zilog->zl_root_zio == NULL) {
883 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
886 if (lwb->lwb_zio == NULL) {
887 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
888 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
889 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
890 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
895 * Define a limited set of intent log block sizes.
897 * These must be a multiple of 4KB. Note only the amount used (again
898 * aligned to 4KB) actually gets written. However, we can't always just
899 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
901 uint64_t zil_block_buckets[] = {
902 4096, /* non TX_WRITE */
903 8192+4096, /* data base */
904 32*1024 + 4096, /* NFS writes */
909 * Use the slog as long as the logbias is 'latency' and the current commit size
910 * is less than the limit or the total list size is less than 2X the limit.
911 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
913 uint64_t zil_slog_limit = 1024 * 1024;
914 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
915 (((zilog)->zl_cur_used < zil_slog_limit) || \
916 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
919 * Start a log block write and advance to the next log block.
920 * Calls are serialized.
923 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
927 spa_t *spa = zilog->zl_spa;
931 uint64_t zil_blksz, wsz;
934 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
935 zilc = (zil_chain_t *)lwb->lwb_buf;
936 bp = &zilc->zc_next_blk;
938 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
939 bp = &zilc->zc_next_blk;
942 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
945 * Allocate the next block and save its address in this block
946 * before writing it in order to establish the log chain.
947 * Note that if the allocation of nlwb synced before we wrote
948 * the block that points at it (lwb), we'd leak it if we crashed.
949 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
950 * We dirty the dataset to ensure that zil_sync() will be called
951 * to clean up in the event of allocation failure or I/O failure.
953 tx = dmu_tx_create(zilog->zl_os);
954 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
955 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
956 txg = dmu_tx_get_txg(tx);
961 * Log blocks are pre-allocated. Here we select the size of the next
962 * block, based on size used in the last block.
963 * - first find the smallest bucket that will fit the block from a
964 * limited set of block sizes. This is because it's faster to write
965 * blocks allocated from the same metaslab as they are adjacent or
967 * - next find the maximum from the new suggested size and an array of
968 * previous sizes. This lessens a picket fence effect of wrongly
969 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
972 * Note we only write what is used, but we can't just allocate
973 * the maximum block size because we can exhaust the available
976 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
977 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
979 zil_blksz = zil_block_buckets[i];
980 if (zil_blksz == UINT64_MAX)
981 zil_blksz = SPA_MAXBLOCKSIZE;
982 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
983 for (i = 0; i < ZIL_PREV_BLKS; i++)
984 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
985 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
988 /* pass the old blkptr in order to spread log blocks across devs */
989 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
992 ASSERT3U(bp->blk_birth, ==, txg);
993 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
994 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
997 * Allocate a new log write buffer (lwb).
999 nlwb = zil_alloc_lwb(zilog, bp, txg);
1001 /* Record the block for later vdev flushing */
1002 zil_add_block(zilog, &lwb->lwb_blk);
1005 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1006 /* For Slim ZIL only write what is used. */
1007 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1008 ASSERT3U(wsz, <=, lwb->lwb_sz);
1009 zio_shrink(lwb->lwb_zio, wsz);
1016 zilc->zc_nused = lwb->lwb_nused;
1017 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1020 * clear unused data for security
1022 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1024 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1027 * If there was an allocation failure then nlwb will be null which
1028 * forces a txg_wait_synced().
1034 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1036 lr_t *lrc = &itx->itx_lr; /* common log record */
1037 lr_write_t *lrw = (lr_write_t *)lrc;
1039 uint64_t txg = lrc->lrc_txg;
1040 uint64_t reclen = lrc->lrc_reclen;
1046 ASSERT(lwb->lwb_buf != NULL);
1047 ASSERT(zilog_is_dirty(zilog) ||
1048 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1050 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1051 dlen = P2ROUNDUP_TYPED(
1052 lrw->lr_length, sizeof (uint64_t), uint64_t);
1054 zilog->zl_cur_used += (reclen + dlen);
1056 zil_lwb_write_init(zilog, lwb);
1059 * If this record won't fit in the current log block, start a new one.
1061 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1062 lwb = zil_lwb_write_start(zilog, lwb);
1065 zil_lwb_write_init(zilog, lwb);
1066 ASSERT(LWB_EMPTY(lwb));
1067 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1068 txg_wait_synced(zilog->zl_dmu_pool, txg);
1073 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1074 bcopy(lrc, lr_buf, reclen);
1075 lrc = (lr_t *)lr_buf;
1076 lrw = (lr_write_t *)lrc;
1079 * If it's a write, fetch the data or get its blkptr as appropriate.
1081 if (lrc->lrc_txtype == TX_WRITE) {
1082 if (txg > spa_freeze_txg(zilog->zl_spa))
1083 txg_wait_synced(zilog->zl_dmu_pool, txg);
1084 if (itx->itx_wr_state != WR_COPIED) {
1089 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1090 dbuf = lr_buf + reclen;
1091 lrw->lr_common.lrc_reclen += dlen;
1093 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1096 error = zilog->zl_get_data(
1097 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1099 txg_wait_synced(zilog->zl_dmu_pool, txg);
1103 ASSERT(error == ENOENT || error == EEXIST ||
1111 * We're actually making an entry, so update lrc_seq to be the
1112 * log record sequence number. Note that this is generally not
1113 * equal to the itx sequence number because not all transactions
1114 * are synchronous, and sometimes spa_sync() gets there first.
1116 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1117 lwb->lwb_nused += reclen + dlen;
1118 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1119 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1120 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1126 zil_itx_create(uint64_t txtype, size_t lrsize)
1130 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1132 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1133 itx->itx_lr.lrc_txtype = txtype;
1134 itx->itx_lr.lrc_reclen = lrsize;
1135 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1136 itx->itx_lr.lrc_seq = 0; /* defensive */
1137 itx->itx_sync = B_TRUE; /* default is synchronous */
1143 zil_itx_destroy(itx_t *itx)
1145 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1149 * Free up the sync and async itxs. The itxs_t has already been detached
1150 * so no locks are needed.
1153 zil_itxg_clean(itxs_t *itxs)
1159 itx_async_node_t *ian;
1161 list = &itxs->i_sync_list;
1162 while ((itx = list_head(list)) != NULL) {
1163 list_remove(list, itx);
1164 kmem_free(itx, offsetof(itx_t, itx_lr) +
1165 itx->itx_lr.lrc_reclen);
1169 t = &itxs->i_async_tree;
1170 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1171 list = &ian->ia_list;
1172 while ((itx = list_head(list)) != NULL) {
1173 list_remove(list, itx);
1174 kmem_free(itx, offsetof(itx_t, itx_lr) +
1175 itx->itx_lr.lrc_reclen);
1178 kmem_free(ian, sizeof (itx_async_node_t));
1182 kmem_free(itxs, sizeof (itxs_t));
1186 zil_aitx_compare(const void *x1, const void *x2)
1188 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1189 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1200 * Remove all async itx with the given oid.
1203 zil_remove_async(zilog_t *zilog, uint64_t oid)
1206 itx_async_node_t *ian;
1213 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1215 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1218 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1220 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1221 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1223 mutex_enter(&itxg->itxg_lock);
1224 if (itxg->itxg_txg != txg) {
1225 mutex_exit(&itxg->itxg_lock);
1230 * Locate the object node and append its list.
1232 t = &itxg->itxg_itxs->i_async_tree;
1233 ian = avl_find(t, &oid, &where);
1235 list_move_tail(&clean_list, &ian->ia_list);
1236 mutex_exit(&itxg->itxg_lock);
1238 while ((itx = list_head(&clean_list)) != NULL) {
1239 list_remove(&clean_list, itx);
1240 kmem_free(itx, offsetof(itx_t, itx_lr) +
1241 itx->itx_lr.lrc_reclen);
1243 list_destroy(&clean_list);
1247 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1251 itxs_t *itxs, *clean = NULL;
1254 * Object ids can be re-instantiated in the next txg so
1255 * remove any async transactions to avoid future leaks.
1256 * This can happen if a fsync occurs on the re-instantiated
1257 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1258 * the new file data and flushes a write record for the old object.
1260 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1261 zil_remove_async(zilog, itx->itx_oid);
1264 * Ensure the data of a renamed file is committed before the rename.
1266 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1267 zil_async_to_sync(zilog, itx->itx_oid);
1269 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1272 txg = dmu_tx_get_txg(tx);
1274 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1275 mutex_enter(&itxg->itxg_lock);
1276 itxs = itxg->itxg_itxs;
1277 if (itxg->itxg_txg != txg) {
1280 * The zil_clean callback hasn't got around to cleaning
1281 * this itxg. Save the itxs for release below.
1282 * This should be rare.
1284 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1286 clean = itxg->itxg_itxs;
1288 ASSERT(itxg->itxg_sod == 0);
1289 itxg->itxg_txg = txg;
1290 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1292 list_create(&itxs->i_sync_list, sizeof (itx_t),
1293 offsetof(itx_t, itx_node));
1294 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1295 sizeof (itx_async_node_t),
1296 offsetof(itx_async_node_t, ia_node));
1298 if (itx->itx_sync) {
1299 list_insert_tail(&itxs->i_sync_list, itx);
1300 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1301 itxg->itxg_sod += itx->itx_sod;
1303 avl_tree_t *t = &itxs->i_async_tree;
1304 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1305 itx_async_node_t *ian;
1308 ian = avl_find(t, &foid, &where);
1310 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1311 list_create(&ian->ia_list, sizeof (itx_t),
1312 offsetof(itx_t, itx_node));
1313 ian->ia_foid = foid;
1314 avl_insert(t, ian, where);
1316 list_insert_tail(&ian->ia_list, itx);
1319 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1320 zilog_dirty(zilog, txg);
1321 mutex_exit(&itxg->itxg_lock);
1323 /* Release the old itxs now we've dropped the lock */
1325 zil_itxg_clean(clean);
1329 * If there are any in-memory intent log transactions which have now been
1330 * synced then start up a taskq to free them. We should only do this after we
1331 * have written out the uberblocks (i.e. txg has been comitted) so that
1332 * don't inadvertently clean out in-memory log records that would be required
1336 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1338 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1341 mutex_enter(&itxg->itxg_lock);
1342 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1343 mutex_exit(&itxg->itxg_lock);
1346 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1347 ASSERT(itxg->itxg_txg != 0);
1348 ASSERT(zilog->zl_clean_taskq != NULL);
1349 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1351 clean_me = itxg->itxg_itxs;
1352 itxg->itxg_itxs = NULL;
1354 mutex_exit(&itxg->itxg_lock);
1356 * Preferably start a task queue to free up the old itxs but
1357 * if taskq_dispatch can't allocate resources to do that then
1358 * free it in-line. This should be rare. Note, using TQ_SLEEP
1359 * created a bad performance problem.
1361 if (taskq_dispatch(zilog->zl_clean_taskq,
1362 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1363 zil_itxg_clean(clean_me);
1367 * Get the list of itxs to commit into zl_itx_commit_list.
1370 zil_get_commit_list(zilog_t *zilog)
1373 list_t *commit_list = &zilog->zl_itx_commit_list;
1374 uint64_t push_sod = 0;
1376 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1379 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1381 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1382 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1384 mutex_enter(&itxg->itxg_lock);
1385 if (itxg->itxg_txg != txg) {
1386 mutex_exit(&itxg->itxg_lock);
1390 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1391 push_sod += itxg->itxg_sod;
1394 mutex_exit(&itxg->itxg_lock);
1396 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1400 * Move the async itxs for a specified object to commit into sync lists.
1403 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1406 itx_async_node_t *ian;
1410 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1413 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1415 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1416 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1418 mutex_enter(&itxg->itxg_lock);
1419 if (itxg->itxg_txg != txg) {
1420 mutex_exit(&itxg->itxg_lock);
1425 * If a foid is specified then find that node and append its
1426 * list. Otherwise walk the tree appending all the lists
1427 * to the sync list. We add to the end rather than the
1428 * beginning to ensure the create has happened.
1430 t = &itxg->itxg_itxs->i_async_tree;
1432 ian = avl_find(t, &foid, &where);
1434 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1438 void *cookie = NULL;
1440 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1441 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1443 list_destroy(&ian->ia_list);
1444 kmem_free(ian, sizeof (itx_async_node_t));
1447 mutex_exit(&itxg->itxg_lock);
1452 zil_commit_writer(zilog_t *zilog)
1457 spa_t *spa = zilog->zl_spa;
1460 ASSERT(zilog->zl_root_zio == NULL);
1462 mutex_exit(&zilog->zl_lock);
1464 zil_get_commit_list(zilog);
1467 * Return if there's nothing to commit before we dirty the fs by
1468 * calling zil_create().
1470 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1471 mutex_enter(&zilog->zl_lock);
1475 if (zilog->zl_suspend) {
1478 lwb = list_tail(&zilog->zl_lwb_list);
1480 lwb = zil_create(zilog);
1483 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1484 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1485 txg = itx->itx_lr.lrc_txg;
1488 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1489 lwb = zil_lwb_commit(zilog, itx, lwb);
1490 list_remove(&zilog->zl_itx_commit_list, itx);
1491 kmem_free(itx, offsetof(itx_t, itx_lr)
1492 + itx->itx_lr.lrc_reclen);
1494 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1496 /* write the last block out */
1497 if (lwb != NULL && lwb->lwb_zio != NULL)
1498 lwb = zil_lwb_write_start(zilog, lwb);
1500 zilog->zl_cur_used = 0;
1503 * Wait if necessary for the log blocks to be on stable storage.
1505 if (zilog->zl_root_zio) {
1506 error = zio_wait(zilog->zl_root_zio);
1507 zilog->zl_root_zio = NULL;
1508 zil_flush_vdevs(zilog);
1511 if (error || lwb == NULL)
1512 txg_wait_synced(zilog->zl_dmu_pool, 0);
1514 mutex_enter(&zilog->zl_lock);
1517 * Remember the highest committed log sequence number for ztest.
1518 * We only update this value when all the log writes succeeded,
1519 * because ztest wants to ASSERT that it got the whole log chain.
1521 if (error == 0 && lwb != NULL)
1522 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1526 * Commit zfs transactions to stable storage.
1527 * If foid is 0 push out all transactions, otherwise push only those
1528 * for that object or might reference that object.
1530 * itxs are committed in batches. In a heavily stressed zil there will be
1531 * a commit writer thread who is writing out a bunch of itxs to the log
1532 * for a set of committing threads (cthreads) in the same batch as the writer.
1533 * Those cthreads are all waiting on the same cv for that batch.
1535 * There will also be a different and growing batch of threads that are
1536 * waiting to commit (qthreads). When the committing batch completes
1537 * a transition occurs such that the cthreads exit and the qthreads become
1538 * cthreads. One of the new cthreads becomes the writer thread for the
1539 * batch. Any new threads arriving become new qthreads.
1541 * Only 2 condition variables are needed and there's no transition
1542 * between the two cvs needed. They just flip-flop between qthreads
1545 * Using this scheme we can efficiently wakeup up only those threads
1546 * that have been committed.
1549 zil_commit(zilog_t *zilog, uint64_t foid)
1553 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1556 /* move the async itxs for the foid to the sync queues */
1557 zil_async_to_sync(zilog, foid);
1559 mutex_enter(&zilog->zl_lock);
1560 mybatch = zilog->zl_next_batch;
1561 while (zilog->zl_writer) {
1562 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1563 if (mybatch <= zilog->zl_com_batch) {
1564 mutex_exit(&zilog->zl_lock);
1569 zilog->zl_next_batch++;
1570 zilog->zl_writer = B_TRUE;
1571 zil_commit_writer(zilog);
1572 zilog->zl_com_batch = mybatch;
1573 zilog->zl_writer = B_FALSE;
1574 mutex_exit(&zilog->zl_lock);
1576 /* wake up one thread to become the next writer */
1577 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1579 /* wake up all threads waiting for this batch to be committed */
1580 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1584 * Called in syncing context to free committed log blocks and update log header.
1587 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1589 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1590 uint64_t txg = dmu_tx_get_txg(tx);
1591 spa_t *spa = zilog->zl_spa;
1592 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1596 * We don't zero out zl_destroy_txg, so make sure we don't try
1597 * to destroy it twice.
1599 if (spa_sync_pass(spa) != 1)
1602 mutex_enter(&zilog->zl_lock);
1604 ASSERT(zilog->zl_stop_sync == 0);
1606 if (*replayed_seq != 0) {
1607 ASSERT(zh->zh_replay_seq < *replayed_seq);
1608 zh->zh_replay_seq = *replayed_seq;
1612 if (zilog->zl_destroy_txg == txg) {
1613 blkptr_t blk = zh->zh_log;
1615 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1617 bzero(zh, sizeof (zil_header_t));
1618 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1620 if (zilog->zl_keep_first) {
1622 * If this block was part of log chain that couldn't
1623 * be claimed because a device was missing during
1624 * zil_claim(), but that device later returns,
1625 * then this block could erroneously appear valid.
1626 * To guard against this, assign a new GUID to the new
1627 * log chain so it doesn't matter what blk points to.
1629 zil_init_log_chain(zilog, &blk);
1634 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1635 zh->zh_log = lwb->lwb_blk;
1636 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1638 list_remove(&zilog->zl_lwb_list, lwb);
1639 zio_free_zil(spa, txg, &lwb->lwb_blk);
1640 kmem_cache_free(zil_lwb_cache, lwb);
1643 * If we don't have anything left in the lwb list then
1644 * we've had an allocation failure and we need to zero
1645 * out the zil_header blkptr so that we don't end
1646 * up freeing the same block twice.
1648 if (list_head(&zilog->zl_lwb_list) == NULL)
1649 BP_ZERO(&zh->zh_log);
1651 mutex_exit(&zilog->zl_lock);
1657 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1658 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1664 kmem_cache_destroy(zil_lwb_cache);
1668 zil_set_sync(zilog_t *zilog, uint64_t sync)
1670 zilog->zl_sync = sync;
1674 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1676 zilog->zl_logbias = logbias;
1680 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1684 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1686 zilog->zl_header = zh_phys;
1688 zilog->zl_spa = dmu_objset_spa(os);
1689 zilog->zl_dmu_pool = dmu_objset_pool(os);
1690 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1691 zilog->zl_logbias = dmu_objset_logbias(os);
1692 zilog->zl_sync = dmu_objset_syncprop(os);
1693 zilog->zl_next_batch = 1;
1695 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1697 for (int i = 0; i < TXG_SIZE; i++) {
1698 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1699 MUTEX_DEFAULT, NULL);
1702 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1703 offsetof(lwb_t, lwb_node));
1705 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1706 offsetof(itx_t, itx_node));
1708 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1710 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1711 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1713 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1714 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1715 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1716 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1722 zil_free(zilog_t *zilog)
1724 zilog->zl_stop_sync = 1;
1726 ASSERT0(zilog->zl_suspend);
1727 ASSERT0(zilog->zl_suspending);
1729 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1730 list_destroy(&zilog->zl_lwb_list);
1732 avl_destroy(&zilog->zl_vdev_tree);
1733 mutex_destroy(&zilog->zl_vdev_lock);
1735 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1736 list_destroy(&zilog->zl_itx_commit_list);
1738 for (int i = 0; i < TXG_SIZE; i++) {
1740 * It's possible for an itx to be generated that doesn't dirty
1741 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1742 * callback to remove the entry. We remove those here.
1744 * Also free up the ziltest itxs.
1746 if (zilog->zl_itxg[i].itxg_itxs)
1747 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1748 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1751 mutex_destroy(&zilog->zl_lock);
1753 cv_destroy(&zilog->zl_cv_writer);
1754 cv_destroy(&zilog->zl_cv_suspend);
1755 cv_destroy(&zilog->zl_cv_batch[0]);
1756 cv_destroy(&zilog->zl_cv_batch[1]);
1758 kmem_free(zilog, sizeof (zilog_t));
1762 * Open an intent log.
1765 zil_open(objset_t *os, zil_get_data_t *get_data)
1767 zilog_t *zilog = dmu_objset_zil(os);
1769 ASSERT(zilog->zl_clean_taskq == NULL);
1770 ASSERT(zilog->zl_get_data == NULL);
1771 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1773 zilog->zl_get_data = get_data;
1774 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1775 2, 2, TASKQ_PREPOPULATE);
1781 * Close an intent log.
1784 zil_close(zilog_t *zilog)
1789 zil_commit(zilog, 0); /* commit all itx */
1792 * The lwb_max_txg for the stubby lwb will reflect the last activity
1793 * for the zil. After a txg_wait_synced() on the txg we know all the
1794 * callbacks have occurred that may clean the zil. Only then can we
1795 * destroy the zl_clean_taskq.
1797 mutex_enter(&zilog->zl_lock);
1798 lwb = list_tail(&zilog->zl_lwb_list);
1800 txg = lwb->lwb_max_txg;
1801 mutex_exit(&zilog->zl_lock);
1803 txg_wait_synced(zilog->zl_dmu_pool, txg);
1804 ASSERT(!zilog_is_dirty(zilog));
1806 taskq_destroy(zilog->zl_clean_taskq);
1807 zilog->zl_clean_taskq = NULL;
1808 zilog->zl_get_data = NULL;
1811 * We should have only one LWB left on the list; remove it now.
1813 mutex_enter(&zilog->zl_lock);
1814 lwb = list_head(&zilog->zl_lwb_list);
1816 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1817 list_remove(&zilog->zl_lwb_list, lwb);
1818 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1819 kmem_cache_free(zil_lwb_cache, lwb);
1821 mutex_exit(&zilog->zl_lock);
1824 static char *suspend_tag = "zil suspending";
1827 * Suspend an intent log. While in suspended mode, we still honor
1828 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1829 * On old version pools, we suspend the log briefly when taking a
1830 * snapshot so that it will have an empty intent log.
1832 * Long holds are not really intended to be used the way we do here --
1833 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1834 * could fail. Therefore we take pains to only put a long hold if it is
1835 * actually necessary. Fortunately, it will only be necessary if the
1836 * objset is currently mounted (or the ZVOL equivalent). In that case it
1837 * will already have a long hold, so we are not really making things any worse.
1839 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1840 * zvol_state_t), and use their mechanism to prevent their hold from being
1841 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1844 * if cookiep == NULL, this does both the suspend & resume.
1845 * Otherwise, it returns with the dataset "long held", and the cookie
1846 * should be passed into zil_resume().
1849 zil_suspend(const char *osname, void **cookiep)
1853 const zil_header_t *zh;
1856 error = dmu_objset_hold(osname, suspend_tag, &os);
1859 zilog = dmu_objset_zil(os);
1861 mutex_enter(&zilog->zl_lock);
1862 zh = zilog->zl_header;
1864 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1865 mutex_exit(&zilog->zl_lock);
1866 dmu_objset_rele(os, suspend_tag);
1867 return (SET_ERROR(EBUSY));
1871 * Don't put a long hold in the cases where we can avoid it. This
1872 * is when there is no cookie so we are doing a suspend & resume
1873 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1874 * for the suspend because it's already suspended, or there's no ZIL.
1876 if (cookiep == NULL && !zilog->zl_suspending &&
1877 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1878 mutex_exit(&zilog->zl_lock);
1879 dmu_objset_rele(os, suspend_tag);
1883 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1884 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1886 zilog->zl_suspend++;
1888 if (zilog->zl_suspend > 1) {
1890 * Someone else is already suspending it.
1891 * Just wait for them to finish.
1894 while (zilog->zl_suspending)
1895 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1896 mutex_exit(&zilog->zl_lock);
1898 if (cookiep == NULL)
1906 * If there is no pointer to an on-disk block, this ZIL must not
1907 * be active (e.g. filesystem not mounted), so there's nothing
1910 if (BP_IS_HOLE(&zh->zh_log)) {
1911 ASSERT(cookiep != NULL); /* fast path already handled */
1914 mutex_exit(&zilog->zl_lock);
1918 zilog->zl_suspending = B_TRUE;
1919 mutex_exit(&zilog->zl_lock);
1921 zil_commit(zilog, 0);
1923 zil_destroy(zilog, B_FALSE);
1925 mutex_enter(&zilog->zl_lock);
1926 zilog->zl_suspending = B_FALSE;
1927 cv_broadcast(&zilog->zl_cv_suspend);
1928 mutex_exit(&zilog->zl_lock);
1930 if (cookiep == NULL)
1938 zil_resume(void *cookie)
1940 objset_t *os = cookie;
1941 zilog_t *zilog = dmu_objset_zil(os);
1943 mutex_enter(&zilog->zl_lock);
1944 ASSERT(zilog->zl_suspend != 0);
1945 zilog->zl_suspend--;
1946 mutex_exit(&zilog->zl_lock);
1947 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1948 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1951 typedef struct zil_replay_arg {
1952 zil_replay_func_t **zr_replay;
1954 boolean_t zr_byteswap;
1959 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1961 char name[MAXNAMELEN];
1963 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1965 dmu_objset_name(zilog->zl_os, name);
1967 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1968 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1969 (u_longlong_t)lr->lrc_seq,
1970 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1971 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1977 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1979 zil_replay_arg_t *zr = zra;
1980 const zil_header_t *zh = zilog->zl_header;
1981 uint64_t reclen = lr->lrc_reclen;
1982 uint64_t txtype = lr->lrc_txtype;
1985 zilog->zl_replaying_seq = lr->lrc_seq;
1987 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1990 if (lr->lrc_txg < claim_txg) /* already committed */
1993 /* Strip case-insensitive bit, still present in log record */
1996 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1997 return (zil_replay_error(zilog, lr, EINVAL));
2000 * If this record type can be logged out of order, the object
2001 * (lr_foid) may no longer exist. That's legitimate, not an error.
2003 if (TX_OOO(txtype)) {
2004 error = dmu_object_info(zilog->zl_os,
2005 ((lr_ooo_t *)lr)->lr_foid, NULL);
2006 if (error == ENOENT || error == EEXIST)
2011 * Make a copy of the data so we can revise and extend it.
2013 bcopy(lr, zr->zr_lr, reclen);
2016 * If this is a TX_WRITE with a blkptr, suck in the data.
2018 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2019 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2020 zr->zr_lr + reclen);
2022 return (zil_replay_error(zilog, lr, error));
2026 * The log block containing this lr may have been byteswapped
2027 * so that we can easily examine common fields like lrc_txtype.
2028 * However, the log is a mix of different record types, and only the
2029 * replay vectors know how to byteswap their records. Therefore, if
2030 * the lr was byteswapped, undo it before invoking the replay vector.
2032 if (zr->zr_byteswap)
2033 byteswap_uint64_array(zr->zr_lr, reclen);
2036 * We must now do two things atomically: replay this log record,
2037 * and update the log header sequence number to reflect the fact that
2038 * we did so. At the end of each replay function the sequence number
2039 * is updated if we are in replay mode.
2041 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2044 * The DMU's dnode layer doesn't see removes until the txg
2045 * commits, so a subsequent claim can spuriously fail with
2046 * EEXIST. So if we receive any error we try syncing out
2047 * any removes then retry the transaction. Note that we
2048 * specify B_FALSE for byteswap now, so we don't do it twice.
2050 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2051 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2053 return (zil_replay_error(zilog, lr, error));
2060 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2062 zilog->zl_replay_blks++;
2068 * If this dataset has a non-empty intent log, replay it and destroy it.
2071 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2073 zilog_t *zilog = dmu_objset_zil(os);
2074 const zil_header_t *zh = zilog->zl_header;
2075 zil_replay_arg_t zr;
2077 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2078 zil_destroy(zilog, B_TRUE);
2081 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2083 zr.zr_replay = replay_func;
2085 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2086 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2089 * Wait for in-progress removes to sync before starting replay.
2091 txg_wait_synced(zilog->zl_dmu_pool, 0);
2093 zilog->zl_replay = B_TRUE;
2094 zilog->zl_replay_time = ddi_get_lbolt();
2095 ASSERT(zilog->zl_replay_blks == 0);
2096 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2098 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2100 zil_destroy(zilog, B_FALSE);
2101 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2102 zilog->zl_replay = B_FALSE;
2103 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2107 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2109 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2112 if (zilog->zl_replay) {
2113 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2114 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2115 zilog->zl_replaying_seq;
2124 zil_vdev_offline(const char *osname, void *arg)
2128 error = zil_suspend(osname, NULL);
2130 return (SET_ERROR(EEXIST));