4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
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 * This global ZIL switch affects all pools
71 int zil_replay_disable = 0; /* disable intent logging replay */
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_notrim = B_TRUE;
87 TUNABLE_INT("vfs.zfs.trim_disable", &zfs_notrim);
88 SYSCTL_INT(_vfs_zfs, OID_AUTO, trim_disable, CTLFLAG_RDTUN, &zfs_notrim, 0,
91 static kmem_cache_t *zil_lwb_cache;
93 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
95 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
96 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
100 * ziltest is by and large an ugly hack, but very useful in
101 * checking replay without tedious work.
102 * When running ziltest we want to keep all itx's and so maintain
103 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
104 * We subtract TXG_CONCURRENT_STATES to allow for common code.
106 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
109 zil_bp_compare(const void *x1, const void *x2)
111 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
112 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
114 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
116 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
119 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
121 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
128 zil_bp_tree_init(zilog_t *zilog)
130 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
131 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
135 zil_bp_tree_fini(zilog_t *zilog)
137 avl_tree_t *t = &zilog->zl_bp_tree;
141 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
142 kmem_free(zn, sizeof (zil_bp_node_t));
148 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
150 avl_tree_t *t = &zilog->zl_bp_tree;
151 const dva_t *dva = BP_IDENTITY(bp);
155 if (avl_find(t, dva, &where) != NULL)
158 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
160 avl_insert(t, zn, where);
165 static zil_header_t *
166 zil_header_in_syncing_context(zilog_t *zilog)
168 return ((zil_header_t *)zilog->zl_header);
172 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
174 zio_cksum_t *zc = &bp->blk_cksum;
176 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
177 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
178 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
179 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
183 * Read a log block and make sure it's valid.
186 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
189 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
190 uint32_t aflags = ARC_WAIT;
191 arc_buf_t *abuf = NULL;
195 if (zilog->zl_header->zh_claim_txg == 0)
196 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
198 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
199 zio_flags |= ZIO_FLAG_SPECULATIVE;
201 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
202 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
204 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
205 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
208 zio_cksum_t cksum = bp->blk_cksum;
211 * Validate the checksummed log block.
213 * Sequence numbers should be... sequential. The checksum
214 * verifier for the next block should be bp's checksum plus 1.
216 * Also check the log chain linkage and size used.
218 cksum.zc_word[ZIL_ZC_SEQ]++;
220 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
221 zil_chain_t *zilc = abuf->b_data;
222 char *lr = (char *)(zilc + 1);
223 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
225 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
226 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
230 *end = (char *)dst + len;
231 *nbp = zilc->zc_next_blk;
234 char *lr = abuf->b_data;
235 uint64_t size = BP_GET_LSIZE(bp);
236 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
238 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
239 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
240 (zilc->zc_nused > (size - sizeof (*zilc)))) {
243 bcopy(lr, dst, zilc->zc_nused);
244 *end = (char *)dst + zilc->zc_nused;
245 *nbp = zilc->zc_next_blk;
249 VERIFY(arc_buf_remove_ref(abuf, &abuf));
256 * Read a TX_WRITE log data block.
259 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
261 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
262 const blkptr_t *bp = &lr->lr_blkptr;
263 uint32_t aflags = ARC_WAIT;
264 arc_buf_t *abuf = NULL;
268 if (BP_IS_HOLE(bp)) {
270 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
274 if (zilog->zl_header->zh_claim_txg == 0)
275 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
277 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
278 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
280 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
281 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
285 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
286 (void) arc_buf_remove_ref(abuf, &abuf);
293 * Parse the intent log, and call parse_func for each valid record within.
296 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
297 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
299 const zil_header_t *zh = zilog->zl_header;
300 boolean_t claimed = !!zh->zh_claim_txg;
301 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
302 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
303 uint64_t max_blk_seq = 0;
304 uint64_t max_lr_seq = 0;
305 uint64_t blk_count = 0;
306 uint64_t lr_count = 0;
307 blkptr_t blk, next_blk;
312 * Old logs didn't record the maximum zh_claim_lr_seq.
314 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
315 claim_lr_seq = UINT64_MAX;
318 * Starting at the block pointed to by zh_log we read the log chain.
319 * For each block in the chain we strongly check that block to
320 * ensure its validity. We stop when an invalid block is found.
321 * For each block pointer in the chain we call parse_blk_func().
322 * For each record in each valid block we call parse_lr_func().
323 * If the log has been claimed, stop if we encounter a sequence
324 * number greater than the highest claimed sequence number.
326 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
327 zil_bp_tree_init(zilog);
329 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
330 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
334 if (blk_seq > claim_blk_seq)
336 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
338 ASSERT3U(max_blk_seq, <, blk_seq);
339 max_blk_seq = blk_seq;
342 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
345 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
349 for (lrp = lrbuf; lrp < end; lrp += reclen) {
350 lr_t *lr = (lr_t *)lrp;
351 reclen = lr->lrc_reclen;
352 ASSERT3U(reclen, >=, sizeof (lr_t));
353 if (lr->lrc_seq > claim_lr_seq)
355 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
357 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
358 max_lr_seq = lr->lrc_seq;
363 zilog->zl_parse_error = error;
364 zilog->zl_parse_blk_seq = max_blk_seq;
365 zilog->zl_parse_lr_seq = max_lr_seq;
366 zilog->zl_parse_blk_count = blk_count;
367 zilog->zl_parse_lr_count = lr_count;
369 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
370 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
372 zil_bp_tree_fini(zilog);
373 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
379 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
382 * Claim log block if not already committed and not already claimed.
383 * If tx == NULL, just verify that the block is claimable.
385 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
388 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
389 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
390 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
394 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
396 lr_write_t *lr = (lr_write_t *)lrc;
399 if (lrc->lrc_txtype != TX_WRITE)
403 * If the block is not readable, don't claim it. This can happen
404 * in normal operation when a log block is written to disk before
405 * some of the dmu_sync() blocks it points to. In this case, the
406 * transaction cannot have been committed to anyone (we would have
407 * waited for all writes to be stable first), so it is semantically
408 * correct to declare this the end of the log.
410 if (lr->lr_blkptr.blk_birth >= first_txg &&
411 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
413 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
418 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
420 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
426 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
428 lr_write_t *lr = (lr_write_t *)lrc;
429 blkptr_t *bp = &lr->lr_blkptr;
432 * If we previously claimed it, we need to free it.
434 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
435 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
436 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
442 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
446 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
447 lwb->lwb_zilog = zilog;
449 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
450 lwb->lwb_max_txg = txg;
453 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
454 lwb->lwb_nused = sizeof (zil_chain_t);
455 lwb->lwb_sz = BP_GET_LSIZE(bp);
458 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
461 mutex_enter(&zilog->zl_lock);
462 list_insert_tail(&zilog->zl_lwb_list, lwb);
463 mutex_exit(&zilog->zl_lock);
469 * Called when we create in-memory log transactions so that we know
470 * to cleanup the itxs at the end of spa_sync().
473 zilog_dirty(zilog_t *zilog, uint64_t txg)
475 dsl_pool_t *dp = zilog->zl_dmu_pool;
476 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
478 if (dsl_dataset_is_snapshot(ds))
479 panic("dirtying snapshot!");
481 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
482 /* up the hold count until we can be written out */
483 dmu_buf_add_ref(ds->ds_dbuf, zilog);
488 zilog_is_dirty(zilog_t *zilog)
490 dsl_pool_t *dp = zilog->zl_dmu_pool;
492 for (int t = 0; t < TXG_SIZE; t++) {
493 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
500 * Create an on-disk intent log.
503 zil_create(zilog_t *zilog)
505 const zil_header_t *zh = zilog->zl_header;
513 * Wait for any previous destroy to complete.
515 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
517 ASSERT(zh->zh_claim_txg == 0);
518 ASSERT(zh->zh_replay_seq == 0);
523 * Allocate an initial log block if:
524 * - there isn't one already
525 * - the existing block is the wrong endianess
527 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
528 tx = dmu_tx_create(zilog->zl_os);
529 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
530 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
531 txg = dmu_tx_get_txg(tx);
533 if (!BP_IS_HOLE(&blk)) {
534 zio_free_zil(zilog->zl_spa, txg, &blk);
538 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
539 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
542 zil_init_log_chain(zilog, &blk);
546 * Allocate a log write buffer (lwb) for the first log block.
549 lwb = zil_alloc_lwb(zilog, &blk, txg);
552 * If we just allocated the first log block, commit our transaction
553 * and wait for zil_sync() to stuff the block poiner into zh_log.
554 * (zh is part of the MOS, so we cannot modify it in open context.)
558 txg_wait_synced(zilog->zl_dmu_pool, txg);
561 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
567 * In one tx, free all log blocks and clear the log header.
568 * If keep_first is set, then we're replaying a log with no content.
569 * We want to keep the first block, however, so that the first
570 * synchronous transaction doesn't require a txg_wait_synced()
571 * in zil_create(). We don't need to txg_wait_synced() here either
572 * when keep_first is set, because both zil_create() and zil_destroy()
573 * will wait for any in-progress destroys to complete.
576 zil_destroy(zilog_t *zilog, boolean_t keep_first)
578 const zil_header_t *zh = zilog->zl_header;
584 * Wait for any previous destroy to complete.
586 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
588 zilog->zl_old_header = *zh; /* debugging aid */
590 if (BP_IS_HOLE(&zh->zh_log))
593 tx = dmu_tx_create(zilog->zl_os);
594 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
595 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
596 txg = dmu_tx_get_txg(tx);
598 mutex_enter(&zilog->zl_lock);
600 ASSERT3U(zilog->zl_destroy_txg, <, txg);
601 zilog->zl_destroy_txg = txg;
602 zilog->zl_keep_first = keep_first;
604 if (!list_is_empty(&zilog->zl_lwb_list)) {
605 ASSERT(zh->zh_claim_txg == 0);
607 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
608 list_remove(&zilog->zl_lwb_list, lwb);
609 if (lwb->lwb_buf != NULL)
610 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
611 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
612 kmem_cache_free(zil_lwb_cache, lwb);
614 } else if (!keep_first) {
615 zil_destroy_sync(zilog, tx);
617 mutex_exit(&zilog->zl_lock);
623 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
625 ASSERT(list_is_empty(&zilog->zl_lwb_list));
626 (void) zil_parse(zilog, zil_free_log_block,
627 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
631 zil_claim(const char *osname, void *txarg)
633 dmu_tx_t *tx = txarg;
634 uint64_t first_txg = dmu_tx_get_txg(tx);
640 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
642 cmn_err(CE_WARN, "can't open objset for %s", osname);
646 zilog = dmu_objset_zil(os);
647 zh = zil_header_in_syncing_context(zilog);
649 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
650 if (!BP_IS_HOLE(&zh->zh_log))
651 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
652 BP_ZERO(&zh->zh_log);
653 dsl_dataset_dirty(dmu_objset_ds(os), tx);
654 dmu_objset_disown(os, FTAG);
659 * Claim all log blocks if we haven't already done so, and remember
660 * the highest claimed sequence number. This ensures that if we can
661 * read only part of the log now (e.g. due to a missing device),
662 * but we can read the entire log later, we will not try to replay
663 * or destroy beyond the last block we successfully claimed.
665 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
666 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
667 (void) zil_parse(zilog, zil_claim_log_block,
668 zil_claim_log_record, tx, first_txg);
669 zh->zh_claim_txg = first_txg;
670 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
671 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
672 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
673 zh->zh_flags |= ZIL_REPLAY_NEEDED;
674 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
675 dsl_dataset_dirty(dmu_objset_ds(os), tx);
678 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
679 dmu_objset_disown(os, FTAG);
684 * Check the log by walking the log chain.
685 * Checksum errors are ok as they indicate the end of the chain.
686 * Any other error (no device or read failure) returns an error.
689 zil_check_log_chain(const char *osname, void *tx)
698 error = dmu_objset_hold(osname, FTAG, &os);
700 cmn_err(CE_WARN, "can't open objset for %s", osname);
704 zilog = dmu_objset_zil(os);
705 bp = (blkptr_t *)&zilog->zl_header->zh_log;
708 * Check the first block and determine if it's on a log device
709 * which may have been removed or faulted prior to loading this
710 * pool. If so, there's no point in checking the rest of the log
711 * as its content should have already been synced to the pool.
713 if (!BP_IS_HOLE(bp)) {
715 boolean_t valid = B_TRUE;
717 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
718 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
719 if (vd->vdev_islog && vdev_is_dead(vd))
720 valid = vdev_log_state_valid(vd);
721 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
724 dmu_objset_rele(os, FTAG);
730 * Because tx == NULL, zil_claim_log_block() will not actually claim
731 * any blocks, but just determine whether it is possible to do so.
732 * In addition to checking the log chain, zil_claim_log_block()
733 * will invoke zio_claim() with a done func of spa_claim_notify(),
734 * which will update spa_max_claim_txg. See spa_load() for details.
736 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
737 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
739 dmu_objset_rele(os, FTAG);
741 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
745 zil_vdev_compare(const void *x1, const void *x2)
747 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
748 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
759 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
761 avl_tree_t *t = &zilog->zl_vdev_tree;
763 zil_vdev_node_t *zv, zvsearch;
764 int ndvas = BP_GET_NDVAS(bp);
767 if (zfs_nocacheflush)
770 ASSERT(zilog->zl_writer);
773 * Even though we're zl_writer, we still need a lock because the
774 * zl_get_data() callbacks may have dmu_sync() done callbacks
775 * that will run concurrently.
777 mutex_enter(&zilog->zl_vdev_lock);
778 for (i = 0; i < ndvas; i++) {
779 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
780 if (avl_find(t, &zvsearch, &where) == NULL) {
781 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
782 zv->zv_vdev = zvsearch.zv_vdev;
783 avl_insert(t, zv, where);
786 mutex_exit(&zilog->zl_vdev_lock);
790 zil_flush_vdevs(zilog_t *zilog)
792 spa_t *spa = zilog->zl_spa;
793 avl_tree_t *t = &zilog->zl_vdev_tree;
798 ASSERT(zilog->zl_writer);
801 * We don't need zl_vdev_lock here because we're the zl_writer,
802 * and all zl_get_data() callbacks are done.
804 if (avl_numnodes(t) == 0)
807 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
809 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
811 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
812 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
815 kmem_free(zv, sizeof (*zv));
819 * Wait for all the flushes to complete. Not all devices actually
820 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
822 (void) zio_wait(zio);
824 spa_config_exit(spa, SCL_STATE, FTAG);
828 * Function called when a log block write completes
831 zil_lwb_write_done(zio_t *zio)
833 lwb_t *lwb = zio->io_private;
834 zilog_t *zilog = lwb->lwb_zilog;
835 dmu_tx_t *tx = lwb->lwb_tx;
837 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
838 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
839 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
840 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
841 ASSERT(!BP_IS_GANG(zio->io_bp));
842 ASSERT(!BP_IS_HOLE(zio->io_bp));
843 ASSERT(zio->io_bp->blk_fill == 0);
846 * Ensure the lwb buffer pointer is cleared before releasing
847 * the txg. If we have had an allocation failure and
848 * the txg is waiting to sync then we want want zil_sync()
849 * to remove the lwb so that it's not picked up as the next new
850 * one in zil_commit_writer(). zil_sync() will only remove
851 * the lwb if lwb_buf is null.
853 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
854 mutex_enter(&zilog->zl_lock);
857 mutex_exit(&zilog->zl_lock);
860 * Now that we've written this log block, we have a stable pointer
861 * to the next block in the chain, so it's OK to let the txg in
862 * which we allocated the next block sync.
868 * Initialize the io for a log block.
871 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
875 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
876 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
877 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
879 if (zilog->zl_root_zio == NULL) {
880 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
883 if (lwb->lwb_zio == NULL) {
884 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
885 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
886 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
887 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
892 * Define a limited set of intent log block sizes.
893 * These must be a multiple of 4KB. Note only the amount used (again
894 * aligned to 4KB) actually gets written. However, we can't always just
895 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
897 uint64_t zil_block_buckets[] = {
898 4096, /* non TX_WRITE */
899 8192+4096, /* data base */
900 32*1024 + 4096, /* NFS writes */
905 * Use the slog as long as the logbias is 'latency' and the current commit size
906 * is less than the limit or the total list size is less than 2X the limit.
907 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
909 uint64_t zil_slog_limit = 1024 * 1024;
910 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
911 (((zilog)->zl_cur_used < zil_slog_limit) || \
912 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
915 * Start a log block write and advance to the next log block.
916 * Calls are serialized.
919 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
923 spa_t *spa = zilog->zl_spa;
927 uint64_t zil_blksz, wsz;
930 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
931 zilc = (zil_chain_t *)lwb->lwb_buf;
932 bp = &zilc->zc_next_blk;
934 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
935 bp = &zilc->zc_next_blk;
938 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
941 * Allocate the next block and save its address in this block
942 * before writing it in order to establish the log chain.
943 * Note that if the allocation of nlwb synced before we wrote
944 * the block that points at it (lwb), we'd leak it if we crashed.
945 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
946 * We dirty the dataset to ensure that zil_sync() will be called
947 * to clean up in the event of allocation failure or I/O failure.
949 tx = dmu_tx_create(zilog->zl_os);
950 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
951 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
952 txg = dmu_tx_get_txg(tx);
957 * Log blocks are pre-allocated. Here we select the size of the next
958 * block, based on size used in the last block.
959 * - first find the smallest bucket that will fit the block from a
960 * limited set of block sizes. This is because it's faster to write
961 * blocks allocated from the same metaslab as they are adjacent or
963 * - next find the maximum from the new suggested size and an array of
964 * previous sizes. This lessens a picket fence effect of wrongly
965 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
968 * Note we only write what is used, but we can't just allocate
969 * the maximum block size because we can exhaust the available
972 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
973 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
975 zil_blksz = zil_block_buckets[i];
976 if (zil_blksz == UINT64_MAX)
977 zil_blksz = SPA_MAXBLOCKSIZE;
978 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
979 for (i = 0; i < ZIL_PREV_BLKS; i++)
980 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
981 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
984 /* pass the old blkptr in order to spread log blocks across devs */
985 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
988 ASSERT3U(bp->blk_birth, ==, txg);
989 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
990 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
993 * Allocate a new log write buffer (lwb).
995 nlwb = zil_alloc_lwb(zilog, bp, txg);
997 /* Record the block for later vdev flushing */
998 zil_add_block(zilog, &lwb->lwb_blk);
1001 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1002 /* For Slim ZIL only write what is used. */
1003 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1004 ASSERT3U(wsz, <=, lwb->lwb_sz);
1005 zio_shrink(lwb->lwb_zio, wsz);
1012 zilc->zc_nused = lwb->lwb_nused;
1013 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1016 * clear unused data for security
1018 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1020 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1023 * If there was an allocation failure then nlwb will be null which
1024 * forces a txg_wait_synced().
1030 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1032 lr_t *lrc = &itx->itx_lr; /* common log record */
1033 lr_write_t *lrw = (lr_write_t *)lrc;
1035 uint64_t txg = lrc->lrc_txg;
1036 uint64_t reclen = lrc->lrc_reclen;
1042 ASSERT(lwb->lwb_buf != NULL);
1043 ASSERT(zilog_is_dirty(zilog) ||
1044 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1046 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1047 dlen = P2ROUNDUP_TYPED(
1048 lrw->lr_length, sizeof (uint64_t), uint64_t);
1050 zilog->zl_cur_used += (reclen + dlen);
1052 zil_lwb_write_init(zilog, lwb);
1055 * If this record won't fit in the current log block, start a new one.
1057 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1058 lwb = zil_lwb_write_start(zilog, lwb);
1061 zil_lwb_write_init(zilog, lwb);
1062 ASSERT(LWB_EMPTY(lwb));
1063 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1064 txg_wait_synced(zilog->zl_dmu_pool, txg);
1069 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1070 bcopy(lrc, lr_buf, reclen);
1071 lrc = (lr_t *)lr_buf;
1072 lrw = (lr_write_t *)lrc;
1075 * If it's a write, fetch the data or get its blkptr as appropriate.
1077 if (lrc->lrc_txtype == TX_WRITE) {
1078 if (txg > spa_freeze_txg(zilog->zl_spa))
1079 txg_wait_synced(zilog->zl_dmu_pool, txg);
1080 if (itx->itx_wr_state != WR_COPIED) {
1085 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1086 dbuf = lr_buf + reclen;
1087 lrw->lr_common.lrc_reclen += dlen;
1089 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1092 error = zilog->zl_get_data(
1093 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1095 txg_wait_synced(zilog->zl_dmu_pool, txg);
1099 ASSERT(error == ENOENT || error == EEXIST ||
1107 * We're actually making an entry, so update lrc_seq to be the
1108 * log record sequence number. Note that this is generally not
1109 * equal to the itx sequence number because not all transactions
1110 * are synchronous, and sometimes spa_sync() gets there first.
1112 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1113 lwb->lwb_nused += reclen + dlen;
1114 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1115 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1116 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1122 zil_itx_create(uint64_t txtype, size_t lrsize)
1126 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1128 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1129 itx->itx_lr.lrc_txtype = txtype;
1130 itx->itx_lr.lrc_reclen = lrsize;
1131 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1132 itx->itx_lr.lrc_seq = 0; /* defensive */
1133 itx->itx_sync = B_TRUE; /* default is synchronous */
1139 zil_itx_destroy(itx_t *itx)
1141 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1145 * Free up the sync and async itxs. The itxs_t has already been detached
1146 * so no locks are needed.
1149 zil_itxg_clean(itxs_t *itxs)
1155 itx_async_node_t *ian;
1157 list = &itxs->i_sync_list;
1158 while ((itx = list_head(list)) != NULL) {
1159 list_remove(list, itx);
1160 kmem_free(itx, offsetof(itx_t, itx_lr) +
1161 itx->itx_lr.lrc_reclen);
1165 t = &itxs->i_async_tree;
1166 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1167 list = &ian->ia_list;
1168 while ((itx = list_head(list)) != NULL) {
1169 list_remove(list, itx);
1170 kmem_free(itx, offsetof(itx_t, itx_lr) +
1171 itx->itx_lr.lrc_reclen);
1174 kmem_free(ian, sizeof (itx_async_node_t));
1178 kmem_free(itxs, sizeof (itxs_t));
1182 zil_aitx_compare(const void *x1, const void *x2)
1184 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1185 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1196 * Remove all async itx with the given oid.
1199 zil_remove_async(zilog_t *zilog, uint64_t oid)
1202 itx_async_node_t *ian;
1209 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1211 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1214 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1216 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1217 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1219 mutex_enter(&itxg->itxg_lock);
1220 if (itxg->itxg_txg != txg) {
1221 mutex_exit(&itxg->itxg_lock);
1226 * Locate the object node and append its list.
1228 t = &itxg->itxg_itxs->i_async_tree;
1229 ian = avl_find(t, &oid, &where);
1231 list_move_tail(&clean_list, &ian->ia_list);
1232 mutex_exit(&itxg->itxg_lock);
1234 while ((itx = list_head(&clean_list)) != NULL) {
1235 list_remove(&clean_list, itx);
1236 kmem_free(itx, offsetof(itx_t, itx_lr) +
1237 itx->itx_lr.lrc_reclen);
1239 list_destroy(&clean_list);
1243 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1247 itxs_t *itxs, *clean = NULL;
1250 * Object ids can be re-instantiated in the next txg so
1251 * remove any async transactions to avoid future leaks.
1252 * This can happen if a fsync occurs on the re-instantiated
1253 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1254 * the new file data and flushes a write record for the old object.
1256 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1257 zil_remove_async(zilog, itx->itx_oid);
1260 * Ensure the data of a renamed file is committed before the rename.
1262 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1263 zil_async_to_sync(zilog, itx->itx_oid);
1265 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1268 txg = dmu_tx_get_txg(tx);
1270 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1271 mutex_enter(&itxg->itxg_lock);
1272 itxs = itxg->itxg_itxs;
1273 if (itxg->itxg_txg != txg) {
1276 * The zil_clean callback hasn't got around to cleaning
1277 * this itxg. Save the itxs for release below.
1278 * This should be rare.
1280 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1282 clean = itxg->itxg_itxs;
1284 ASSERT(itxg->itxg_sod == 0);
1285 itxg->itxg_txg = txg;
1286 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1288 list_create(&itxs->i_sync_list, sizeof (itx_t),
1289 offsetof(itx_t, itx_node));
1290 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1291 sizeof (itx_async_node_t),
1292 offsetof(itx_async_node_t, ia_node));
1294 if (itx->itx_sync) {
1295 list_insert_tail(&itxs->i_sync_list, itx);
1296 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1297 itxg->itxg_sod += itx->itx_sod;
1299 avl_tree_t *t = &itxs->i_async_tree;
1300 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1301 itx_async_node_t *ian;
1304 ian = avl_find(t, &foid, &where);
1306 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1307 list_create(&ian->ia_list, sizeof (itx_t),
1308 offsetof(itx_t, itx_node));
1309 ian->ia_foid = foid;
1310 avl_insert(t, ian, where);
1312 list_insert_tail(&ian->ia_list, itx);
1315 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1316 zilog_dirty(zilog, txg);
1317 mutex_exit(&itxg->itxg_lock);
1319 /* Release the old itxs now we've dropped the lock */
1321 zil_itxg_clean(clean);
1325 * If there are any in-memory intent log transactions which have now been
1326 * synced then start up a taskq to free them. We should only do this after we
1327 * have written out the uberblocks (i.e. txg has been comitted) so that
1328 * don't inadvertently clean out in-memory log records that would be required
1332 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1334 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1337 mutex_enter(&itxg->itxg_lock);
1338 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1339 mutex_exit(&itxg->itxg_lock);
1342 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1343 ASSERT(itxg->itxg_txg != 0);
1344 ASSERT(zilog->zl_clean_taskq != NULL);
1345 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1347 clean_me = itxg->itxg_itxs;
1348 itxg->itxg_itxs = NULL;
1350 mutex_exit(&itxg->itxg_lock);
1352 * Preferably start a task queue to free up the old itxs but
1353 * if taskq_dispatch can't allocate resources to do that then
1354 * free it in-line. This should be rare. Note, using TQ_SLEEP
1355 * created a bad performance problem.
1357 if (taskq_dispatch(zilog->zl_clean_taskq,
1358 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1359 zil_itxg_clean(clean_me);
1363 * Get the list of itxs to commit into zl_itx_commit_list.
1366 zil_get_commit_list(zilog_t *zilog)
1369 list_t *commit_list = &zilog->zl_itx_commit_list;
1370 uint64_t push_sod = 0;
1372 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1375 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1377 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1378 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1380 mutex_enter(&itxg->itxg_lock);
1381 if (itxg->itxg_txg != txg) {
1382 mutex_exit(&itxg->itxg_lock);
1386 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1387 push_sod += itxg->itxg_sod;
1390 mutex_exit(&itxg->itxg_lock);
1392 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1396 * Move the async itxs for a specified object to commit into sync lists.
1399 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1402 itx_async_node_t *ian;
1406 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1409 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1411 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1412 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1414 mutex_enter(&itxg->itxg_lock);
1415 if (itxg->itxg_txg != txg) {
1416 mutex_exit(&itxg->itxg_lock);
1421 * If a foid is specified then find that node and append its
1422 * list. Otherwise walk the tree appending all the lists
1423 * to the sync list. We add to the end rather than the
1424 * beginning to ensure the create has happened.
1426 t = &itxg->itxg_itxs->i_async_tree;
1428 ian = avl_find(t, &foid, &where);
1430 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1434 void *cookie = NULL;
1436 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1437 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1439 list_destroy(&ian->ia_list);
1440 kmem_free(ian, sizeof (itx_async_node_t));
1443 mutex_exit(&itxg->itxg_lock);
1448 zil_commit_writer(zilog_t *zilog)
1453 spa_t *spa = zilog->zl_spa;
1456 ASSERT(zilog->zl_root_zio == NULL);
1458 mutex_exit(&zilog->zl_lock);
1460 zil_get_commit_list(zilog);
1463 * Return if there's nothing to commit before we dirty the fs by
1464 * calling zil_create().
1466 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1467 mutex_enter(&zilog->zl_lock);
1471 if (zilog->zl_suspend) {
1474 lwb = list_tail(&zilog->zl_lwb_list);
1476 lwb = zil_create(zilog);
1479 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1480 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1481 txg = itx->itx_lr.lrc_txg;
1484 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1485 lwb = zil_lwb_commit(zilog, itx, lwb);
1486 list_remove(&zilog->zl_itx_commit_list, itx);
1487 kmem_free(itx, offsetof(itx_t, itx_lr)
1488 + itx->itx_lr.lrc_reclen);
1490 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1492 /* write the last block out */
1493 if (lwb != NULL && lwb->lwb_zio != NULL)
1494 lwb = zil_lwb_write_start(zilog, lwb);
1496 zilog->zl_cur_used = 0;
1499 * Wait if necessary for the log blocks to be on stable storage.
1501 if (zilog->zl_root_zio) {
1502 error = zio_wait(zilog->zl_root_zio);
1503 zilog->zl_root_zio = NULL;
1504 zil_flush_vdevs(zilog);
1507 if (error || lwb == NULL)
1508 txg_wait_synced(zilog->zl_dmu_pool, 0);
1510 mutex_enter(&zilog->zl_lock);
1513 * Remember the highest committed log sequence number for ztest.
1514 * We only update this value when all the log writes succeeded,
1515 * because ztest wants to ASSERT that it got the whole log chain.
1517 if (error == 0 && lwb != NULL)
1518 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1522 * Commit zfs transactions to stable storage.
1523 * If foid is 0 push out all transactions, otherwise push only those
1524 * for that object or might reference that object.
1526 * itxs are committed in batches. In a heavily stressed zil there will be
1527 * a commit writer thread who is writing out a bunch of itxs to the log
1528 * for a set of committing threads (cthreads) in the same batch as the writer.
1529 * Those cthreads are all waiting on the same cv for that batch.
1531 * There will also be a different and growing batch of threads that are
1532 * waiting to commit (qthreads). When the committing batch completes
1533 * a transition occurs such that the cthreads exit and the qthreads become
1534 * cthreads. One of the new cthreads becomes the writer thread for the
1535 * batch. Any new threads arriving become new qthreads.
1537 * Only 2 condition variables are needed and there's no transition
1538 * between the two cvs needed. They just flip-flop between qthreads
1541 * Using this scheme we can efficiently wakeup up only those threads
1542 * that have been committed.
1545 zil_commit(zilog_t *zilog, uint64_t foid)
1549 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1552 /* move the async itxs for the foid to the sync queues */
1553 zil_async_to_sync(zilog, foid);
1555 mutex_enter(&zilog->zl_lock);
1556 mybatch = zilog->zl_next_batch;
1557 while (zilog->zl_writer) {
1558 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1559 if (mybatch <= zilog->zl_com_batch) {
1560 mutex_exit(&zilog->zl_lock);
1565 zilog->zl_next_batch++;
1566 zilog->zl_writer = B_TRUE;
1567 zil_commit_writer(zilog);
1568 zilog->zl_com_batch = mybatch;
1569 zilog->zl_writer = B_FALSE;
1570 mutex_exit(&zilog->zl_lock);
1572 /* wake up one thread to become the next writer */
1573 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1575 /* wake up all threads waiting for this batch to be committed */
1576 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1580 * Called in syncing context to free committed log blocks and update log header.
1583 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1585 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1586 uint64_t txg = dmu_tx_get_txg(tx);
1587 spa_t *spa = zilog->zl_spa;
1588 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1592 * We don't zero out zl_destroy_txg, so make sure we don't try
1593 * to destroy it twice.
1595 if (spa_sync_pass(spa) != 1)
1598 mutex_enter(&zilog->zl_lock);
1600 ASSERT(zilog->zl_stop_sync == 0);
1602 if (*replayed_seq != 0) {
1603 ASSERT(zh->zh_replay_seq < *replayed_seq);
1604 zh->zh_replay_seq = *replayed_seq;
1608 if (zilog->zl_destroy_txg == txg) {
1609 blkptr_t blk = zh->zh_log;
1611 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1613 bzero(zh, sizeof (zil_header_t));
1614 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1616 if (zilog->zl_keep_first) {
1618 * If this block was part of log chain that couldn't
1619 * be claimed because a device was missing during
1620 * zil_claim(), but that device later returns,
1621 * then this block could erroneously appear valid.
1622 * To guard against this, assign a new GUID to the new
1623 * log chain so it doesn't matter what blk points to.
1625 zil_init_log_chain(zilog, &blk);
1630 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1631 zh->zh_log = lwb->lwb_blk;
1632 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1634 list_remove(&zilog->zl_lwb_list, lwb);
1635 zio_free_zil(spa, txg, &lwb->lwb_blk);
1636 kmem_cache_free(zil_lwb_cache, lwb);
1639 * If we don't have anything left in the lwb list then
1640 * we've had an allocation failure and we need to zero
1641 * out the zil_header blkptr so that we don't end
1642 * up freeing the same block twice.
1644 if (list_head(&zilog->zl_lwb_list) == NULL)
1645 BP_ZERO(&zh->zh_log);
1647 mutex_exit(&zilog->zl_lock);
1653 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1654 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1660 kmem_cache_destroy(zil_lwb_cache);
1664 zil_set_sync(zilog_t *zilog, uint64_t sync)
1666 zilog->zl_sync = sync;
1670 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1672 zilog->zl_logbias = logbias;
1676 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1680 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1682 zilog->zl_header = zh_phys;
1684 zilog->zl_spa = dmu_objset_spa(os);
1685 zilog->zl_dmu_pool = dmu_objset_pool(os);
1686 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1687 zilog->zl_logbias = dmu_objset_logbias(os);
1688 zilog->zl_sync = dmu_objset_syncprop(os);
1689 zilog->zl_next_batch = 1;
1691 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1693 for (int i = 0; i < TXG_SIZE; i++) {
1694 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1695 MUTEX_DEFAULT, NULL);
1698 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1699 offsetof(lwb_t, lwb_node));
1701 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1702 offsetof(itx_t, itx_node));
1704 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1706 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1707 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1709 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1710 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1711 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1712 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1718 zil_free(zilog_t *zilog)
1720 zilog->zl_stop_sync = 1;
1722 ASSERT0(zilog->zl_suspend);
1723 ASSERT0(zilog->zl_suspending);
1725 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1726 list_destroy(&zilog->zl_lwb_list);
1728 avl_destroy(&zilog->zl_vdev_tree);
1729 mutex_destroy(&zilog->zl_vdev_lock);
1731 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1732 list_destroy(&zilog->zl_itx_commit_list);
1734 for (int i = 0; i < TXG_SIZE; i++) {
1736 * It's possible for an itx to be generated that doesn't dirty
1737 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1738 * callback to remove the entry. We remove those here.
1740 * Also free up the ziltest itxs.
1742 if (zilog->zl_itxg[i].itxg_itxs)
1743 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1744 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1747 mutex_destroy(&zilog->zl_lock);
1749 cv_destroy(&zilog->zl_cv_writer);
1750 cv_destroy(&zilog->zl_cv_suspend);
1751 cv_destroy(&zilog->zl_cv_batch[0]);
1752 cv_destroy(&zilog->zl_cv_batch[1]);
1754 kmem_free(zilog, sizeof (zilog_t));
1758 * Open an intent log.
1761 zil_open(objset_t *os, zil_get_data_t *get_data)
1763 zilog_t *zilog = dmu_objset_zil(os);
1765 ASSERT(zilog->zl_clean_taskq == NULL);
1766 ASSERT(zilog->zl_get_data == NULL);
1767 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1769 zilog->zl_get_data = get_data;
1770 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1771 2, 2, TASKQ_PREPOPULATE);
1777 * Close an intent log.
1780 zil_close(zilog_t *zilog)
1785 zil_commit(zilog, 0); /* commit all itx */
1788 * The lwb_max_txg for the stubby lwb will reflect the last activity
1789 * for the zil. After a txg_wait_synced() on the txg we know all the
1790 * callbacks have occurred that may clean the zil. Only then can we
1791 * destroy the zl_clean_taskq.
1793 mutex_enter(&zilog->zl_lock);
1794 lwb = list_tail(&zilog->zl_lwb_list);
1796 txg = lwb->lwb_max_txg;
1797 mutex_exit(&zilog->zl_lock);
1799 txg_wait_synced(zilog->zl_dmu_pool, txg);
1800 ASSERT(!zilog_is_dirty(zilog));
1802 taskq_destroy(zilog->zl_clean_taskq);
1803 zilog->zl_clean_taskq = NULL;
1804 zilog->zl_get_data = NULL;
1807 * We should have only one LWB left on the list; remove it now.
1809 mutex_enter(&zilog->zl_lock);
1810 lwb = list_head(&zilog->zl_lwb_list);
1812 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1813 list_remove(&zilog->zl_lwb_list, lwb);
1814 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1815 kmem_cache_free(zil_lwb_cache, lwb);
1817 mutex_exit(&zilog->zl_lock);
1820 static char *suspend_tag = "zil suspending";
1823 * Suspend an intent log. While in suspended mode, we still honor
1824 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1825 * On old version pools, we suspend the log briefly when taking a
1826 * snapshot so that it will have an empty intent log.
1828 * Long holds are not really intended to be used the way we do here --
1829 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1830 * could fail. Therefore we take pains to only put a long hold if it is
1831 * actually necessary. Fortunately, it will only be necessary if the
1832 * objset is currently mounted (or the ZVOL equivalent). In that case it
1833 * will already have a long hold, so we are not really making things any worse.
1835 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1836 * zvol_state_t), and use their mechanism to prevent their hold from being
1837 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1840 * if cookiep == NULL, this does both the suspend & resume.
1841 * Otherwise, it returns with the dataset "long held", and the cookie
1842 * should be passed into zil_resume().
1845 zil_suspend(const char *osname, void **cookiep)
1849 const zil_header_t *zh;
1852 error = dmu_objset_hold(osname, suspend_tag, &os);
1855 zilog = dmu_objset_zil(os);
1857 mutex_enter(&zilog->zl_lock);
1858 zh = zilog->zl_header;
1860 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1861 mutex_exit(&zilog->zl_lock);
1862 dmu_objset_rele(os, suspend_tag);
1867 * Don't put a long hold in the cases where we can avoid it. This
1868 * is when there is no cookie so we are doing a suspend & resume
1869 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1870 * for the suspend because it's already suspended, or there's no ZIL.
1872 if (cookiep == NULL && !zilog->zl_suspending &&
1873 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1874 mutex_exit(&zilog->zl_lock);
1875 dmu_objset_rele(os, suspend_tag);
1879 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1880 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1882 zilog->zl_suspend++;
1884 if (zilog->zl_suspend > 1) {
1886 * Someone else is already suspending it.
1887 * Just wait for them to finish.
1890 while (zilog->zl_suspending)
1891 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1892 mutex_exit(&zilog->zl_lock);
1894 if (cookiep == NULL)
1902 * If there is no pointer to an on-disk block, this ZIL must not
1903 * be active (e.g. filesystem not mounted), so there's nothing
1906 if (BP_IS_HOLE(&zh->zh_log)) {
1907 ASSERT(cookiep != NULL); /* fast path already handled */
1910 mutex_exit(&zilog->zl_lock);
1914 zilog->zl_suspending = B_TRUE;
1915 mutex_exit(&zilog->zl_lock);
1917 zil_commit(zilog, 0);
1919 zil_destroy(zilog, B_FALSE);
1921 mutex_enter(&zilog->zl_lock);
1922 zilog->zl_suspending = B_FALSE;
1923 cv_broadcast(&zilog->zl_cv_suspend);
1924 mutex_exit(&zilog->zl_lock);
1926 if (cookiep == NULL)
1934 zil_resume(void *cookie)
1936 objset_t *os = cookie;
1937 zilog_t *zilog = dmu_objset_zil(os);
1939 mutex_enter(&zilog->zl_lock);
1940 ASSERT(zilog->zl_suspend != 0);
1941 zilog->zl_suspend--;
1942 mutex_exit(&zilog->zl_lock);
1943 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1944 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1947 typedef struct zil_replay_arg {
1948 zil_replay_func_t **zr_replay;
1950 boolean_t zr_byteswap;
1955 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1957 char name[MAXNAMELEN];
1959 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1961 dmu_objset_name(zilog->zl_os, name);
1963 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1964 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1965 (u_longlong_t)lr->lrc_seq,
1966 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1967 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1973 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1975 zil_replay_arg_t *zr = zra;
1976 const zil_header_t *zh = zilog->zl_header;
1977 uint64_t reclen = lr->lrc_reclen;
1978 uint64_t txtype = lr->lrc_txtype;
1981 zilog->zl_replaying_seq = lr->lrc_seq;
1983 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1986 if (lr->lrc_txg < claim_txg) /* already committed */
1989 /* Strip case-insensitive bit, still present in log record */
1992 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1993 return (zil_replay_error(zilog, lr, EINVAL));
1996 * If this record type can be logged out of order, the object
1997 * (lr_foid) may no longer exist. That's legitimate, not an error.
1999 if (TX_OOO(txtype)) {
2000 error = dmu_object_info(zilog->zl_os,
2001 ((lr_ooo_t *)lr)->lr_foid, NULL);
2002 if (error == ENOENT || error == EEXIST)
2007 * Make a copy of the data so we can revise and extend it.
2009 bcopy(lr, zr->zr_lr, reclen);
2012 * If this is a TX_WRITE with a blkptr, suck in the data.
2014 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2015 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2016 zr->zr_lr + reclen);
2018 return (zil_replay_error(zilog, lr, error));
2022 * The log block containing this lr may have been byteswapped
2023 * so that we can easily examine common fields like lrc_txtype.
2024 * However, the log is a mix of different record types, and only the
2025 * replay vectors know how to byteswap their records. Therefore, if
2026 * the lr was byteswapped, undo it before invoking the replay vector.
2028 if (zr->zr_byteswap)
2029 byteswap_uint64_array(zr->zr_lr, reclen);
2032 * We must now do two things atomically: replay this log record,
2033 * and update the log header sequence number to reflect the fact that
2034 * we did so. At the end of each replay function the sequence number
2035 * is updated if we are in replay mode.
2037 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2040 * The DMU's dnode layer doesn't see removes until the txg
2041 * commits, so a subsequent claim can spuriously fail with
2042 * EEXIST. So if we receive any error we try syncing out
2043 * any removes then retry the transaction. Note that we
2044 * specify B_FALSE for byteswap now, so we don't do it twice.
2046 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2047 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2049 return (zil_replay_error(zilog, lr, error));
2056 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2058 zilog->zl_replay_blks++;
2064 * If this dataset has a non-empty intent log, replay it and destroy it.
2067 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2069 zilog_t *zilog = dmu_objset_zil(os);
2070 const zil_header_t *zh = zilog->zl_header;
2071 zil_replay_arg_t zr;
2073 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2074 zil_destroy(zilog, B_TRUE);
2077 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name);
2079 zr.zr_replay = replay_func;
2081 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2082 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2085 * Wait for in-progress removes to sync before starting replay.
2087 txg_wait_synced(zilog->zl_dmu_pool, 0);
2089 zilog->zl_replay = B_TRUE;
2090 zilog->zl_replay_time = ddi_get_lbolt();
2091 ASSERT(zilog->zl_replay_blks == 0);
2092 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2094 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2096 zil_destroy(zilog, B_FALSE);
2097 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2098 zilog->zl_replay = B_FALSE;
2099 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name);
2103 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2105 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2108 if (zilog->zl_replay) {
2109 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2110 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2111 zilog->zl_replaying_seq;
2120 zil_vdev_offline(const char *osname, void *arg)
2124 error = zil_suspend(osname, NULL);