4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
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 */
74 * Tunable parameter for debugging or performance analysis. Setting
75 * zfs_nocacheflush will cause corruption on power loss if a volatile
76 * out-of-order write cache is enabled.
78 int zfs_nocacheflush = 0;
80 static kmem_cache_t *zil_lwb_cache;
82 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
84 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
85 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
89 * ziltest is by and large an ugly hack, but very useful in
90 * checking replay without tedious work.
91 * When running ziltest we want to keep all itx's and so maintain
92 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
93 * We subtract TXG_CONCURRENT_STATES to allow for common code.
95 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
98 zil_bp_compare(const void *x1, const void *x2)
100 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
101 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
103 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
105 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
108 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
110 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
117 zil_bp_tree_init(zilog_t *zilog)
119 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
120 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
124 zil_bp_tree_fini(zilog_t *zilog)
126 avl_tree_t *t = &zilog->zl_bp_tree;
130 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
131 kmem_free(zn, sizeof (zil_bp_node_t));
137 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
139 avl_tree_t *t = &zilog->zl_bp_tree;
140 const dva_t *dva = BP_IDENTITY(bp);
144 if (avl_find(t, dva, &where) != NULL)
147 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
149 avl_insert(t, zn, where);
154 static zil_header_t *
155 zil_header_in_syncing_context(zilog_t *zilog)
157 return ((zil_header_t *)zilog->zl_header);
161 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
163 zio_cksum_t *zc = &bp->blk_cksum;
165 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
166 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
167 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
168 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
172 * Read a log block and make sure it's valid.
175 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
178 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
179 uint32_t aflags = ARC_WAIT;
180 arc_buf_t *abuf = NULL;
184 if (zilog->zl_header->zh_claim_txg == 0)
185 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
187 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
188 zio_flags |= ZIO_FLAG_SPECULATIVE;
190 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
191 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
193 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
194 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
197 zio_cksum_t cksum = bp->blk_cksum;
200 * Validate the checksummed log block.
202 * Sequence numbers should be... sequential. The checksum
203 * verifier for the next block should be bp's checksum plus 1.
205 * Also check the log chain linkage and size used.
207 cksum.zc_word[ZIL_ZC_SEQ]++;
209 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
210 zil_chain_t *zilc = abuf->b_data;
211 char *lr = (char *)(zilc + 1);
212 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
214 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
215 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
219 *end = (char *)dst + len;
220 *nbp = zilc->zc_next_blk;
223 char *lr = abuf->b_data;
224 uint64_t size = BP_GET_LSIZE(bp);
225 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
227 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
228 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
229 (zilc->zc_nused > (size - sizeof (*zilc)))) {
232 bcopy(lr, dst, zilc->zc_nused);
233 *end = (char *)dst + zilc->zc_nused;
234 *nbp = zilc->zc_next_blk;
238 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
245 * Read a TX_WRITE log data block.
248 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
250 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
251 const blkptr_t *bp = &lr->lr_blkptr;
252 uint32_t aflags = ARC_WAIT;
253 arc_buf_t *abuf = NULL;
257 if (BP_IS_HOLE(bp)) {
259 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
263 if (zilog->zl_header->zh_claim_txg == 0)
264 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
266 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
267 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
269 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
270 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
274 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
275 (void) arc_buf_remove_ref(abuf, &abuf);
282 * Parse the intent log, and call parse_func for each valid record within.
285 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
286 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
288 const zil_header_t *zh = zilog->zl_header;
289 boolean_t claimed = !!zh->zh_claim_txg;
290 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
291 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
292 uint64_t max_blk_seq = 0;
293 uint64_t max_lr_seq = 0;
294 uint64_t blk_count = 0;
295 uint64_t lr_count = 0;
296 blkptr_t blk, next_blk;
300 bzero(&next_blk, sizeof(blkptr_t));
303 * Old logs didn't record the maximum zh_claim_lr_seq.
305 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
306 claim_lr_seq = UINT64_MAX;
309 * Starting at the block pointed to by zh_log we read the log chain.
310 * For each block in the chain we strongly check that block to
311 * ensure its validity. We stop when an invalid block is found.
312 * For each block pointer in the chain we call parse_blk_func().
313 * For each record in each valid block we call parse_lr_func().
314 * If the log has been claimed, stop if we encounter a sequence
315 * number greater than the highest claimed sequence number.
317 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
318 zil_bp_tree_init(zilog);
320 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
321 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
325 if (blk_seq > claim_blk_seq)
327 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
329 ASSERT3U(max_blk_seq, <, blk_seq);
330 max_blk_seq = blk_seq;
333 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
336 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
340 for (lrp = lrbuf; lrp < end; lrp += reclen) {
341 lr_t *lr = (lr_t *)lrp;
342 reclen = lr->lrc_reclen;
343 ASSERT3U(reclen, >=, sizeof (lr_t));
344 if (lr->lrc_seq > claim_lr_seq)
346 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
348 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
349 max_lr_seq = lr->lrc_seq;
354 zilog->zl_parse_error = error;
355 zilog->zl_parse_blk_seq = max_blk_seq;
356 zilog->zl_parse_lr_seq = max_lr_seq;
357 zilog->zl_parse_blk_count = blk_count;
358 zilog->zl_parse_lr_count = lr_count;
360 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
361 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
363 zil_bp_tree_fini(zilog);
364 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
370 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
373 * Claim log block if not already committed and not already claimed.
374 * If tx == NULL, just verify that the block is claimable.
376 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
379 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
380 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
381 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
385 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
387 lr_write_t *lr = (lr_write_t *)lrc;
390 if (lrc->lrc_txtype != TX_WRITE)
394 * If the block is not readable, don't claim it. This can happen
395 * in normal operation when a log block is written to disk before
396 * some of the dmu_sync() blocks it points to. In this case, the
397 * transaction cannot have been committed to anyone (we would have
398 * waited for all writes to be stable first), so it is semantically
399 * correct to declare this the end of the log.
401 if (lr->lr_blkptr.blk_birth >= first_txg &&
402 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
404 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
409 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
411 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
417 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
419 lr_write_t *lr = (lr_write_t *)lrc;
420 blkptr_t *bp = &lr->lr_blkptr;
423 * If we previously claimed it, we need to free it.
425 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
426 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
427 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
433 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
437 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
438 lwb->lwb_zilog = zilog;
440 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
441 lwb->lwb_max_txg = txg;
444 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
445 lwb->lwb_nused = sizeof (zil_chain_t);
446 lwb->lwb_sz = BP_GET_LSIZE(bp);
449 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
452 mutex_enter(&zilog->zl_lock);
453 list_insert_tail(&zilog->zl_lwb_list, lwb);
454 mutex_exit(&zilog->zl_lock);
460 * Create an on-disk intent log.
463 zil_create(zilog_t *zilog)
465 const zil_header_t *zh = zilog->zl_header;
473 * Wait for any previous destroy to complete.
475 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
477 ASSERT(zh->zh_claim_txg == 0);
478 ASSERT(zh->zh_replay_seq == 0);
483 * Allocate an initial log block if:
484 * - there isn't one already
485 * - the existing block is the wrong endianess
487 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
488 tx = dmu_tx_create(zilog->zl_os);
489 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
490 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
491 txg = dmu_tx_get_txg(tx);
493 if (!BP_IS_HOLE(&blk)) {
494 zio_free_zil(zilog->zl_spa, txg, &blk);
498 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
499 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
502 zil_init_log_chain(zilog, &blk);
506 * Allocate a log write buffer (lwb) for the first log block.
509 lwb = zil_alloc_lwb(zilog, &blk, txg);
512 * If we just allocated the first log block, commit our transaction
513 * and wait for zil_sync() to stuff the block poiner into zh_log.
514 * (zh is part of the MOS, so we cannot modify it in open context.)
518 txg_wait_synced(zilog->zl_dmu_pool, txg);
521 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
527 * In one tx, free all log blocks and clear the log header.
528 * If keep_first is set, then we're replaying a log with no content.
529 * We want to keep the first block, however, so that the first
530 * synchronous transaction doesn't require a txg_wait_synced()
531 * in zil_create(). We don't need to txg_wait_synced() here either
532 * when keep_first is set, because both zil_create() and zil_destroy()
533 * will wait for any in-progress destroys to complete.
536 zil_destroy(zilog_t *zilog, boolean_t keep_first)
538 const zil_header_t *zh = zilog->zl_header;
544 * Wait for any previous destroy to complete.
546 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
548 zilog->zl_old_header = *zh; /* debugging aid */
550 if (BP_IS_HOLE(&zh->zh_log))
553 tx = dmu_tx_create(zilog->zl_os);
554 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
555 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
556 txg = dmu_tx_get_txg(tx);
558 mutex_enter(&zilog->zl_lock);
560 ASSERT3U(zilog->zl_destroy_txg, <, txg);
561 zilog->zl_destroy_txg = txg;
562 zilog->zl_keep_first = keep_first;
564 if (!list_is_empty(&zilog->zl_lwb_list)) {
565 ASSERT(zh->zh_claim_txg == 0);
567 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
568 list_remove(&zilog->zl_lwb_list, lwb);
569 if (lwb->lwb_buf != NULL)
570 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
571 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
572 kmem_cache_free(zil_lwb_cache, lwb);
574 } else if (!keep_first) {
575 (void) zil_parse(zilog, zil_free_log_block,
576 zil_free_log_record, tx, zh->zh_claim_txg);
578 mutex_exit(&zilog->zl_lock);
584 zil_claim(const char *osname, void *txarg)
586 dmu_tx_t *tx = txarg;
587 uint64_t first_txg = dmu_tx_get_txg(tx);
593 error = dmu_objset_hold(osname, FTAG, &os);
595 cmn_err(CE_WARN, "can't open objset for %s", osname);
599 zilog = dmu_objset_zil(os);
600 zh = zil_header_in_syncing_context(zilog);
602 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
603 if (!BP_IS_HOLE(&zh->zh_log))
604 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
605 BP_ZERO(&zh->zh_log);
606 dsl_dataset_dirty(dmu_objset_ds(os), tx);
607 dmu_objset_rele(os, FTAG);
612 * Claim all log blocks if we haven't already done so, and remember
613 * the highest claimed sequence number. This ensures that if we can
614 * read only part of the log now (e.g. due to a missing device),
615 * but we can read the entire log later, we will not try to replay
616 * or destroy beyond the last block we successfully claimed.
618 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
619 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
620 (void) zil_parse(zilog, zil_claim_log_block,
621 zil_claim_log_record, tx, first_txg);
622 zh->zh_claim_txg = first_txg;
623 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
624 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
625 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
626 zh->zh_flags |= ZIL_REPLAY_NEEDED;
627 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
628 dsl_dataset_dirty(dmu_objset_ds(os), tx);
631 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
632 dmu_objset_rele(os, FTAG);
637 * Check the log by walking the log chain.
638 * Checksum errors are ok as they indicate the end of the chain.
639 * Any other error (no device or read failure) returns an error.
642 zil_check_log_chain(const char *osname, void *tx)
651 error = dmu_objset_hold(osname, FTAG, &os);
653 cmn_err(CE_WARN, "can't open objset for %s", osname);
657 zilog = dmu_objset_zil(os);
658 bp = (blkptr_t *)&zilog->zl_header->zh_log;
661 * Check the first block and determine if it's on a log device
662 * which may have been removed or faulted prior to loading this
663 * pool. If so, there's no point in checking the rest of the log
664 * as its content should have already been synced to the pool.
666 if (!BP_IS_HOLE(bp)) {
668 boolean_t valid = B_TRUE;
670 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
671 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
672 if (vd->vdev_islog && vdev_is_dead(vd))
673 valid = vdev_log_state_valid(vd);
674 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
677 dmu_objset_rele(os, FTAG);
683 * Because tx == NULL, zil_claim_log_block() will not actually claim
684 * any blocks, but just determine whether it is possible to do so.
685 * In addition to checking the log chain, zil_claim_log_block()
686 * will invoke zio_claim() with a done func of spa_claim_notify(),
687 * which will update spa_max_claim_txg. See spa_load() for details.
689 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
690 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
692 dmu_objset_rele(os, FTAG);
694 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
698 zil_vdev_compare(const void *x1, const void *x2)
700 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
701 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
712 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
714 avl_tree_t *t = &zilog->zl_vdev_tree;
716 zil_vdev_node_t *zv, zvsearch;
717 int ndvas = BP_GET_NDVAS(bp);
720 if (zfs_nocacheflush)
723 ASSERT(zilog->zl_writer);
726 * Even though we're zl_writer, we still need a lock because the
727 * zl_get_data() callbacks may have dmu_sync() done callbacks
728 * that will run concurrently.
730 mutex_enter(&zilog->zl_vdev_lock);
731 for (i = 0; i < ndvas; i++) {
732 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
733 if (avl_find(t, &zvsearch, &where) == NULL) {
734 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
735 zv->zv_vdev = zvsearch.zv_vdev;
736 avl_insert(t, zv, where);
739 mutex_exit(&zilog->zl_vdev_lock);
743 zil_flush_vdevs(zilog_t *zilog)
745 spa_t *spa = zilog->zl_spa;
746 avl_tree_t *t = &zilog->zl_vdev_tree;
751 ASSERT(zilog->zl_writer);
754 * We don't need zl_vdev_lock here because we're the zl_writer,
755 * and all zl_get_data() callbacks are done.
757 if (avl_numnodes(t) == 0)
760 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
762 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
764 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
765 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
768 kmem_free(zv, sizeof (*zv));
772 * Wait for all the flushes to complete. Not all devices actually
773 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
775 (void) zio_wait(zio);
777 spa_config_exit(spa, SCL_STATE, FTAG);
781 * Function called when a log block write completes
784 zil_lwb_write_done(zio_t *zio)
786 lwb_t *lwb = zio->io_private;
787 zilog_t *zilog = lwb->lwb_zilog;
788 dmu_tx_t *tx = lwb->lwb_tx;
790 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
791 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
792 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
793 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
794 ASSERT(!BP_IS_GANG(zio->io_bp));
795 ASSERT(!BP_IS_HOLE(zio->io_bp));
796 ASSERT(zio->io_bp->blk_fill == 0);
799 * Ensure the lwb buffer pointer is cleared before releasing
800 * the txg. If we have had an allocation failure and
801 * the txg is waiting to sync then we want want zil_sync()
802 * to remove the lwb so that it's not picked up as the next new
803 * one in zil_commit_writer(). zil_sync() will only remove
804 * the lwb if lwb_buf is null.
806 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
807 mutex_enter(&zilog->zl_lock);
810 mutex_exit(&zilog->zl_lock);
813 * Now that we've written this log block, we have a stable pointer
814 * to the next block in the chain, so it's OK to let the txg in
815 * which we allocated the next block sync.
821 * Initialize the io for a log block.
824 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
828 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
829 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
830 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
832 if (zilog->zl_root_zio == NULL) {
833 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
836 if (lwb->lwb_zio == NULL) {
837 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
838 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
839 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
840 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
845 * Define a limited set of intent log block sizes.
846 * These must be a multiple of 4KB. Note only the amount used (again
847 * aligned to 4KB) actually gets written. However, we can't always just
848 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
850 uint64_t zil_block_buckets[] = {
851 4096, /* non TX_WRITE */
852 8192+4096, /* data base */
853 32*1024 + 4096, /* NFS writes */
858 * Use the slog as long as the logbias is 'latency' and the current commit size
859 * is less than the limit or the total list size is less than 2X the limit.
860 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
862 uint64_t zil_slog_limit = 1024 * 1024;
863 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
864 (((zilog)->zl_cur_used < zil_slog_limit) || \
865 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
868 * Start a log block write and advance to the next log block.
869 * Calls are serialized.
872 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
876 spa_t *spa = zilog->zl_spa;
880 uint64_t zil_blksz, wsz;
883 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
884 zilc = (zil_chain_t *)lwb->lwb_buf;
885 bp = &zilc->zc_next_blk;
887 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
888 bp = &zilc->zc_next_blk;
891 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
894 * Allocate the next block and save its address in this block
895 * before writing it in order to establish the log chain.
896 * Note that if the allocation of nlwb synced before we wrote
897 * the block that points at it (lwb), we'd leak it if we crashed.
898 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
899 * We dirty the dataset to ensure that zil_sync() will be called
900 * to clean up in the event of allocation failure or I/O failure.
902 tx = dmu_tx_create(zilog->zl_os);
903 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
904 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
905 txg = dmu_tx_get_txg(tx);
910 * Log blocks are pre-allocated. Here we select the size of the next
911 * block, based on size used in the last block.
912 * - first find the smallest bucket that will fit the block from a
913 * limited set of block sizes. This is because it's faster to write
914 * blocks allocated from the same metaslab as they are adjacent or
916 * - next find the maximum from the new suggested size and an array of
917 * previous sizes. This lessens a picket fence effect of wrongly
918 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
921 * Note we only write what is used, but we can't just allocate
922 * the maximum block size because we can exhaust the available
925 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
926 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
928 zil_blksz = zil_block_buckets[i];
929 if (zil_blksz == UINT64_MAX)
930 zil_blksz = SPA_MAXBLOCKSIZE;
931 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
932 for (i = 0; i < ZIL_PREV_BLKS; i++)
933 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
934 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
937 /* pass the old blkptr in order to spread log blocks across devs */
938 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
941 ASSERT3U(bp->blk_birth, ==, txg);
942 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
943 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
946 * Allocate a new log write buffer (lwb).
948 nlwb = zil_alloc_lwb(zilog, bp, txg);
950 /* Record the block for later vdev flushing */
951 zil_add_block(zilog, &lwb->lwb_blk);
954 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
955 /* For Slim ZIL only write what is used. */
956 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
957 ASSERT3U(wsz, <=, lwb->lwb_sz);
958 zio_shrink(lwb->lwb_zio, wsz);
965 zilc->zc_nused = lwb->lwb_nused;
966 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
969 * clear unused data for security
971 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
973 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
976 * If there was an allocation failure then nlwb will be null which
977 * forces a txg_wait_synced().
983 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
985 lr_t *lrc = &itx->itx_lr; /* common log record */
986 lr_write_t *lrw = (lr_write_t *)lrc;
988 uint64_t txg = lrc->lrc_txg;
989 uint64_t reclen = lrc->lrc_reclen;
995 ASSERT(lwb->lwb_buf != NULL);
997 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
998 dlen = P2ROUNDUP_TYPED(
999 lrw->lr_length, sizeof (uint64_t), uint64_t);
1001 zilog->zl_cur_used += (reclen + dlen);
1003 zil_lwb_write_init(zilog, lwb);
1006 * If this record won't fit in the current log block, start a new one.
1008 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1009 lwb = zil_lwb_write_start(zilog, lwb);
1012 zil_lwb_write_init(zilog, lwb);
1013 ASSERT(LWB_EMPTY(lwb));
1014 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1015 txg_wait_synced(zilog->zl_dmu_pool, txg);
1020 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1021 bcopy(lrc, lr_buf, reclen);
1022 lrc = (lr_t *)lr_buf;
1023 lrw = (lr_write_t *)lrc;
1026 * If it's a write, fetch the data or get its blkptr as appropriate.
1028 if (lrc->lrc_txtype == TX_WRITE) {
1029 if (txg > spa_freeze_txg(zilog->zl_spa))
1030 txg_wait_synced(zilog->zl_dmu_pool, txg);
1031 if (itx->itx_wr_state != WR_COPIED) {
1036 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1037 dbuf = lr_buf + reclen;
1038 lrw->lr_common.lrc_reclen += dlen;
1040 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1043 error = zilog->zl_get_data(
1044 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1046 txg_wait_synced(zilog->zl_dmu_pool, txg);
1050 ASSERT(error == ENOENT || error == EEXIST ||
1058 * We're actually making an entry, so update lrc_seq to be the
1059 * log record sequence number. Note that this is generally not
1060 * equal to the itx sequence number because not all transactions
1061 * are synchronous, and sometimes spa_sync() gets there first.
1063 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1064 lwb->lwb_nused += reclen + dlen;
1065 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1066 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1067 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1073 zil_itx_create(uint64_t txtype, size_t lrsize)
1077 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1079 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize,
1080 KM_PUSHPAGE | KM_NODEBUG);
1081 itx->itx_lr.lrc_txtype = txtype;
1082 itx->itx_lr.lrc_reclen = lrsize;
1083 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1084 itx->itx_lr.lrc_seq = 0; /* defensive */
1085 itx->itx_sync = B_TRUE; /* default is synchronous */
1091 zil_itx_destroy(itx_t *itx)
1093 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1097 * Free up the sync and async itxs. The itxs_t has already been detached
1098 * so no locks are needed.
1101 zil_itxg_clean(itxs_t *itxs)
1107 itx_async_node_t *ian;
1109 list = &itxs->i_sync_list;
1110 while ((itx = list_head(list)) != NULL) {
1111 list_remove(list, itx);
1112 kmem_free(itx, offsetof(itx_t, itx_lr) +
1113 itx->itx_lr.lrc_reclen);
1117 t = &itxs->i_async_tree;
1118 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1119 list = &ian->ia_list;
1120 while ((itx = list_head(list)) != NULL) {
1121 list_remove(list, itx);
1122 kmem_free(itx, offsetof(itx_t, itx_lr) +
1123 itx->itx_lr.lrc_reclen);
1126 kmem_free(ian, sizeof (itx_async_node_t));
1130 kmem_free(itxs, sizeof (itxs_t));
1134 zil_aitx_compare(const void *x1, const void *x2)
1136 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1137 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1148 * Remove all async itx with the given oid.
1151 zil_remove_async(zilog_t *zilog, uint64_t oid)
1154 itx_async_node_t *ian;
1161 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1163 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1166 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1168 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1169 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1171 mutex_enter(&itxg->itxg_lock);
1172 if (itxg->itxg_txg != txg) {
1173 mutex_exit(&itxg->itxg_lock);
1178 * Locate the object node and append its list.
1180 t = &itxg->itxg_itxs->i_async_tree;
1181 ian = avl_find(t, &oid, &where);
1183 list_move_tail(&clean_list, &ian->ia_list);
1184 mutex_exit(&itxg->itxg_lock);
1186 while ((itx = list_head(&clean_list)) != NULL) {
1187 list_remove(&clean_list, itx);
1188 kmem_free(itx, offsetof(itx_t, itx_lr) +
1189 itx->itx_lr.lrc_reclen);
1191 list_destroy(&clean_list);
1195 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1199 itxs_t *itxs, *clean = NULL;
1202 * Object ids can be re-instantiated in the next txg so
1203 * remove any async transactions to avoid future leaks.
1204 * This can happen if a fsync occurs on the re-instantiated
1205 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1206 * the new file data and flushes a write record for the old object.
1208 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1209 zil_remove_async(zilog, itx->itx_oid);
1212 * Ensure the data of a renamed file is committed before the rename.
1214 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1215 zil_async_to_sync(zilog, itx->itx_oid);
1217 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1220 txg = dmu_tx_get_txg(tx);
1222 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1223 mutex_enter(&itxg->itxg_lock);
1224 itxs = itxg->itxg_itxs;
1225 if (itxg->itxg_txg != txg) {
1228 * The zil_clean callback hasn't got around to cleaning
1229 * this itxg. Save the itxs for release below.
1230 * This should be rare.
1232 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1234 clean = itxg->itxg_itxs;
1236 ASSERT(itxg->itxg_sod == 0);
1237 itxg->itxg_txg = txg;
1238 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1240 list_create(&itxs->i_sync_list, sizeof (itx_t),
1241 offsetof(itx_t, itx_node));
1242 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1243 sizeof (itx_async_node_t),
1244 offsetof(itx_async_node_t, ia_node));
1246 if (itx->itx_sync) {
1247 list_insert_tail(&itxs->i_sync_list, itx);
1248 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1249 itxg->itxg_sod += itx->itx_sod;
1251 avl_tree_t *t = &itxs->i_async_tree;
1252 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1253 itx_async_node_t *ian;
1256 ian = avl_find(t, &foid, &where);
1258 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1259 list_create(&ian->ia_list, sizeof (itx_t),
1260 offsetof(itx_t, itx_node));
1261 ian->ia_foid = foid;
1262 avl_insert(t, ian, where);
1264 list_insert_tail(&ian->ia_list, itx);
1267 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1268 mutex_exit(&itxg->itxg_lock);
1270 /* Release the old itxs now we've dropped the lock */
1272 zil_itxg_clean(clean);
1276 * If there are any in-memory intent log transactions which have now been
1277 * synced then start up a taskq to free them.
1280 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1282 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1285 mutex_enter(&itxg->itxg_lock);
1286 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1287 mutex_exit(&itxg->itxg_lock);
1290 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1291 ASSERT(itxg->itxg_txg != 0);
1292 ASSERT(zilog->zl_clean_taskq != NULL);
1293 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1295 clean_me = itxg->itxg_itxs;
1296 itxg->itxg_itxs = NULL;
1298 mutex_exit(&itxg->itxg_lock);
1300 * Preferably start a task queue to free up the old itxs but
1301 * if taskq_dispatch can't allocate resources to do that then
1302 * free it in-line. This should be rare. Note, using TQ_SLEEP
1303 * created a bad performance problem.
1305 if (taskq_dispatch(zilog->zl_clean_taskq,
1306 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1307 zil_itxg_clean(clean_me);
1311 * Get the list of itxs to commit into zl_itx_commit_list.
1314 zil_get_commit_list(zilog_t *zilog)
1317 list_t *commit_list = &zilog->zl_itx_commit_list;
1318 uint64_t push_sod = 0;
1320 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1323 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1325 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1326 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1328 mutex_enter(&itxg->itxg_lock);
1329 if (itxg->itxg_txg != txg) {
1330 mutex_exit(&itxg->itxg_lock);
1334 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1335 push_sod += itxg->itxg_sod;
1338 mutex_exit(&itxg->itxg_lock);
1340 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1344 * Move the async itxs for a specified object to commit into sync lists.
1347 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1350 itx_async_node_t *ian;
1354 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1357 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1359 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1360 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1362 mutex_enter(&itxg->itxg_lock);
1363 if (itxg->itxg_txg != txg) {
1364 mutex_exit(&itxg->itxg_lock);
1369 * If a foid is specified then find that node and append its
1370 * list. Otherwise walk the tree appending all the lists
1371 * to the sync list. We add to the end rather than the
1372 * beginning to ensure the create has happened.
1374 t = &itxg->itxg_itxs->i_async_tree;
1376 ian = avl_find(t, &foid, &where);
1378 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1382 void *cookie = NULL;
1384 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1385 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1387 list_destroy(&ian->ia_list);
1388 kmem_free(ian, sizeof (itx_async_node_t));
1391 mutex_exit(&itxg->itxg_lock);
1396 zil_commit_writer(zilog_t *zilog)
1401 spa_t *spa = zilog->zl_spa;
1404 ASSERT(zilog->zl_root_zio == NULL);
1406 mutex_exit(&zilog->zl_lock);
1408 zil_get_commit_list(zilog);
1411 * Return if there's nothing to commit before we dirty the fs by
1412 * calling zil_create().
1414 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1415 mutex_enter(&zilog->zl_lock);
1419 if (zilog->zl_suspend) {
1422 lwb = list_tail(&zilog->zl_lwb_list);
1424 lwb = zil_create(zilog);
1427 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1428 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1429 txg = itx->itx_lr.lrc_txg;
1432 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1433 lwb = zil_lwb_commit(zilog, itx, lwb);
1434 list_remove(&zilog->zl_itx_commit_list, itx);
1435 kmem_free(itx, offsetof(itx_t, itx_lr)
1436 + itx->itx_lr.lrc_reclen);
1438 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1440 /* write the last block out */
1441 if (lwb != NULL && lwb->lwb_zio != NULL)
1442 lwb = zil_lwb_write_start(zilog, lwb);
1444 zilog->zl_cur_used = 0;
1447 * Wait if necessary for the log blocks to be on stable storage.
1449 if (zilog->zl_root_zio) {
1450 error = zio_wait(zilog->zl_root_zio);
1451 zilog->zl_root_zio = NULL;
1452 zil_flush_vdevs(zilog);
1455 if (error || lwb == NULL)
1456 txg_wait_synced(zilog->zl_dmu_pool, 0);
1458 mutex_enter(&zilog->zl_lock);
1461 * Remember the highest committed log sequence number for ztest.
1462 * We only update this value when all the log writes succeeded,
1463 * because ztest wants to ASSERT that it got the whole log chain.
1465 if (error == 0 && lwb != NULL)
1466 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1470 * Commit zfs transactions to stable storage.
1471 * If foid is 0 push out all transactions, otherwise push only those
1472 * for that object or might reference that object.
1474 * itxs are committed in batches. In a heavily stressed zil there will be
1475 * a commit writer thread who is writing out a bunch of itxs to the log
1476 * for a set of committing threads (cthreads) in the same batch as the writer.
1477 * Those cthreads are all waiting on the same cv for that batch.
1479 * There will also be a different and growing batch of threads that are
1480 * waiting to commit (qthreads). When the committing batch completes
1481 * a transition occurs such that the cthreads exit and the qthreads become
1482 * cthreads. One of the new cthreads becomes the writer thread for the
1483 * batch. Any new threads arriving become new qthreads.
1485 * Only 2 condition variables are needed and there's no transition
1486 * between the two cvs needed. They just flip-flop between qthreads
1489 * Using this scheme we can efficiently wakeup up only those threads
1490 * that have been committed.
1493 zil_commit(zilog_t *zilog, uint64_t foid)
1497 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1500 /* move the async itxs for the foid to the sync queues */
1501 zil_async_to_sync(zilog, foid);
1503 mutex_enter(&zilog->zl_lock);
1504 mybatch = zilog->zl_next_batch;
1505 while (zilog->zl_writer) {
1506 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1507 if (mybatch <= zilog->zl_com_batch) {
1508 mutex_exit(&zilog->zl_lock);
1513 zilog->zl_next_batch++;
1514 zilog->zl_writer = B_TRUE;
1515 zil_commit_writer(zilog);
1516 zilog->zl_com_batch = mybatch;
1517 zilog->zl_writer = B_FALSE;
1518 mutex_exit(&zilog->zl_lock);
1520 /* wake up one thread to become the next writer */
1521 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1523 /* wake up all threads waiting for this batch to be committed */
1524 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1528 * Called in syncing context to free committed log blocks and update log header.
1531 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1533 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1534 uint64_t txg = dmu_tx_get_txg(tx);
1535 spa_t *spa = zilog->zl_spa;
1536 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1540 * We don't zero out zl_destroy_txg, so make sure we don't try
1541 * to destroy it twice.
1543 if (spa_sync_pass(spa) != 1)
1546 mutex_enter(&zilog->zl_lock);
1548 ASSERT(zilog->zl_stop_sync == 0);
1550 if (*replayed_seq != 0) {
1551 ASSERT(zh->zh_replay_seq < *replayed_seq);
1552 zh->zh_replay_seq = *replayed_seq;
1556 if (zilog->zl_destroy_txg == txg) {
1557 blkptr_t blk = zh->zh_log;
1559 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1561 bzero(zh, sizeof (zil_header_t));
1562 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1564 if (zilog->zl_keep_first) {
1566 * If this block was part of log chain that couldn't
1567 * be claimed because a device was missing during
1568 * zil_claim(), but that device later returns,
1569 * then this block could erroneously appear valid.
1570 * To guard against this, assign a new GUID to the new
1571 * log chain so it doesn't matter what blk points to.
1573 zil_init_log_chain(zilog, &blk);
1578 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1579 zh->zh_log = lwb->lwb_blk;
1580 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1582 list_remove(&zilog->zl_lwb_list, lwb);
1583 zio_free_zil(spa, txg, &lwb->lwb_blk);
1584 kmem_cache_free(zil_lwb_cache, lwb);
1587 * If we don't have anything left in the lwb list then
1588 * we've had an allocation failure and we need to zero
1589 * out the zil_header blkptr so that we don't end
1590 * up freeing the same block twice.
1592 if (list_head(&zilog->zl_lwb_list) == NULL)
1593 BP_ZERO(&zh->zh_log);
1595 mutex_exit(&zilog->zl_lock);
1601 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1602 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1608 kmem_cache_destroy(zil_lwb_cache);
1612 zil_set_sync(zilog_t *zilog, uint64_t sync)
1614 zilog->zl_sync = sync;
1618 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1620 zilog->zl_logbias = logbias;
1624 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1629 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1631 zilog->zl_header = zh_phys;
1633 zilog->zl_spa = dmu_objset_spa(os);
1634 zilog->zl_dmu_pool = dmu_objset_pool(os);
1635 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1636 zilog->zl_logbias = dmu_objset_logbias(os);
1637 zilog->zl_sync = dmu_objset_syncprop(os);
1638 zilog->zl_next_batch = 1;
1640 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1642 for (i = 0; i < TXG_SIZE; i++) {
1643 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1644 MUTEX_DEFAULT, NULL);
1647 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1648 offsetof(lwb_t, lwb_node));
1650 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1651 offsetof(itx_t, itx_node));
1653 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1655 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1656 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1658 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1659 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1660 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1661 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1667 zil_free(zilog_t *zilog)
1671 zilog->zl_stop_sync = 1;
1673 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1674 list_destroy(&zilog->zl_lwb_list);
1676 avl_destroy(&zilog->zl_vdev_tree);
1677 mutex_destroy(&zilog->zl_vdev_lock);
1679 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1680 list_destroy(&zilog->zl_itx_commit_list);
1682 for (i = 0; i < TXG_SIZE; i++) {
1684 * It's possible for an itx to be generated that doesn't dirty
1685 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1686 * callback to remove the entry. We remove those here.
1688 * Also free up the ziltest itxs.
1690 if (zilog->zl_itxg[i].itxg_itxs)
1691 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1692 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1695 mutex_destroy(&zilog->zl_lock);
1697 cv_destroy(&zilog->zl_cv_writer);
1698 cv_destroy(&zilog->zl_cv_suspend);
1699 cv_destroy(&zilog->zl_cv_batch[0]);
1700 cv_destroy(&zilog->zl_cv_batch[1]);
1702 kmem_free(zilog, sizeof (zilog_t));
1706 * Open an intent log.
1709 zil_open(objset_t *os, zil_get_data_t *get_data)
1711 zilog_t *zilog = dmu_objset_zil(os);
1713 ASSERT(zilog->zl_clean_taskq == NULL);
1714 ASSERT(zilog->zl_get_data == NULL);
1715 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1717 zilog->zl_get_data = get_data;
1718 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1719 2, 2, TASKQ_PREPOPULATE);
1725 * Close an intent log.
1728 zil_close(zilog_t *zilog)
1733 zil_commit(zilog, 0); /* commit all itx */
1736 * The lwb_max_txg for the stubby lwb will reflect the last activity
1737 * for the zil. After a txg_wait_synced() on the txg we know all the
1738 * callbacks have occurred that may clean the zil. Only then can we
1739 * destroy the zl_clean_taskq.
1741 mutex_enter(&zilog->zl_lock);
1742 lwb = list_tail(&zilog->zl_lwb_list);
1744 txg = lwb->lwb_max_txg;
1745 mutex_exit(&zilog->zl_lock);
1747 txg_wait_synced(zilog->zl_dmu_pool, txg);
1749 taskq_destroy(zilog->zl_clean_taskq);
1750 zilog->zl_clean_taskq = NULL;
1751 zilog->zl_get_data = NULL;
1754 * We should have only one LWB left on the list; remove it now.
1756 mutex_enter(&zilog->zl_lock);
1757 lwb = list_head(&zilog->zl_lwb_list);
1759 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1760 list_remove(&zilog->zl_lwb_list, lwb);
1761 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1762 kmem_cache_free(zil_lwb_cache, lwb);
1764 mutex_exit(&zilog->zl_lock);
1768 * Suspend an intent log. While in suspended mode, we still honor
1769 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1770 * We suspend the log briefly when taking a snapshot so that the snapshot
1771 * contains all the data it's supposed to, and has an empty intent log.
1774 zil_suspend(zilog_t *zilog)
1776 const zil_header_t *zh = zilog->zl_header;
1778 mutex_enter(&zilog->zl_lock);
1779 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1780 mutex_exit(&zilog->zl_lock);
1783 if (zilog->zl_suspend++ != 0) {
1785 * Someone else already began a suspend.
1786 * Just wait for them to finish.
1788 while (zilog->zl_suspending)
1789 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1790 mutex_exit(&zilog->zl_lock);
1793 zilog->zl_suspending = B_TRUE;
1794 mutex_exit(&zilog->zl_lock);
1796 zil_commit(zilog, 0);
1798 zil_destroy(zilog, B_FALSE);
1800 mutex_enter(&zilog->zl_lock);
1801 zilog->zl_suspending = B_FALSE;
1802 cv_broadcast(&zilog->zl_cv_suspend);
1803 mutex_exit(&zilog->zl_lock);
1809 zil_resume(zilog_t *zilog)
1811 mutex_enter(&zilog->zl_lock);
1812 ASSERT(zilog->zl_suspend != 0);
1813 zilog->zl_suspend--;
1814 mutex_exit(&zilog->zl_lock);
1817 typedef struct zil_replay_arg {
1818 zil_replay_func_t **zr_replay;
1820 boolean_t zr_byteswap;
1825 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1827 char name[MAXNAMELEN];
1829 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1831 dmu_objset_name(zilog->zl_os, name);
1833 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1834 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1835 (u_longlong_t)lr->lrc_seq,
1836 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1837 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1843 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1845 zil_replay_arg_t *zr = zra;
1846 const zil_header_t *zh = zilog->zl_header;
1847 uint64_t reclen = lr->lrc_reclen;
1848 uint64_t txtype = lr->lrc_txtype;
1851 zilog->zl_replaying_seq = lr->lrc_seq;
1853 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1856 if (lr->lrc_txg < claim_txg) /* already committed */
1859 /* Strip case-insensitive bit, still present in log record */
1862 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1863 return (zil_replay_error(zilog, lr, EINVAL));
1866 * If this record type can be logged out of order, the object
1867 * (lr_foid) may no longer exist. That's legitimate, not an error.
1869 if (TX_OOO(txtype)) {
1870 error = dmu_object_info(zilog->zl_os,
1871 ((lr_ooo_t *)lr)->lr_foid, NULL);
1872 if (error == ENOENT || error == EEXIST)
1877 * Make a copy of the data so we can revise and extend it.
1879 bcopy(lr, zr->zr_lr, reclen);
1882 * If this is a TX_WRITE with a blkptr, suck in the data.
1884 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1885 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1886 zr->zr_lr + reclen);
1888 return (zil_replay_error(zilog, lr, error));
1892 * The log block containing this lr may have been byteswapped
1893 * so that we can easily examine common fields like lrc_txtype.
1894 * However, the log is a mix of different record types, and only the
1895 * replay vectors know how to byteswap their records. Therefore, if
1896 * the lr was byteswapped, undo it before invoking the replay vector.
1898 if (zr->zr_byteswap)
1899 byteswap_uint64_array(zr->zr_lr, reclen);
1902 * We must now do two things atomically: replay this log record,
1903 * and update the log header sequence number to reflect the fact that
1904 * we did so. At the end of each replay function the sequence number
1905 * is updated if we are in replay mode.
1907 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1910 * The DMU's dnode layer doesn't see removes until the txg
1911 * commits, so a subsequent claim can spuriously fail with
1912 * EEXIST. So if we receive any error we try syncing out
1913 * any removes then retry the transaction. Note that we
1914 * specify B_FALSE for byteswap now, so we don't do it twice.
1916 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1917 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1919 return (zil_replay_error(zilog, lr, error));
1926 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1928 zilog->zl_replay_blks++;
1934 * If this dataset has a non-empty intent log, replay it and destroy it.
1937 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1939 zilog_t *zilog = dmu_objset_zil(os);
1940 const zil_header_t *zh = zilog->zl_header;
1941 zil_replay_arg_t zr;
1943 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1944 zil_destroy(zilog, B_TRUE);
1948 zr.zr_replay = replay_func;
1950 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1951 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1954 * Wait for in-progress removes to sync before starting replay.
1956 txg_wait_synced(zilog->zl_dmu_pool, 0);
1958 zilog->zl_replay = B_TRUE;
1959 zilog->zl_replay_time = ddi_get_lbolt();
1960 ASSERT(zilog->zl_replay_blks == 0);
1961 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1963 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1965 zil_destroy(zilog, B_FALSE);
1966 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1967 zilog->zl_replay = B_FALSE;
1971 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1973 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1976 if (zilog->zl_replay) {
1977 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1978 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1979 zilog->zl_replaying_seq;
1988 zil_vdev_offline(const char *osname, void *arg)
1994 error = dmu_objset_hold(osname, FTAG, &os);
1998 zilog = dmu_objset_zil(os);
1999 if (zil_suspend(zilog) != 0)
2003 dmu_objset_rele(os, FTAG);
2007 #if defined(_KERNEL) && defined(HAVE_SPL)
2008 module_param(zil_replay_disable, int, 0644);
2009 MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay");
2011 module_param(zfs_nocacheflush, int, 0644);
2012 MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes");