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.
25 /* Portions Copyright 2010 Robert Milkowski */
27 #include <sys/zfs_context.h>
33 #include <sys/resource.h>
35 #include <sys/zil_impl.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dsl_pool.h>
42 * The zfs intent log (ZIL) saves transaction records of system calls
43 * that change the file system in memory with enough information
44 * to be able to replay them. These are stored in memory until
45 * either the DMU transaction group (txg) commits them to the stable pool
46 * and they can be discarded, or they are flushed to the stable log
47 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
48 * requirement. In the event of a panic or power fail then those log
49 * records (transactions) are replayed.
51 * There is one ZIL per file system. Its on-disk (pool) format consists
58 * A log record holds a system call transaction. Log blocks can
59 * hold many log records and the blocks are chained together.
60 * Each ZIL block contains a block pointer (blkptr_t) to the next
61 * ZIL block in the chain. The ZIL header points to the first
62 * block in the chain. Note there is not a fixed place in the pool
63 * to hold blocks. They are dynamically allocated and freed as
64 * needed from the blocks available. Figure X shows the ZIL structure:
68 * This global ZIL switch affects all pools
70 int zil_replay_disable = 0; /* disable intent logging replay */
73 * Tunable parameter for debugging or performance analysis. Setting
74 * zfs_nocacheflush will cause corruption on power loss if a volatile
75 * out-of-order write cache is enabled.
77 boolean_t zfs_nocacheflush = B_FALSE;
79 static kmem_cache_t *zil_lwb_cache;
81 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
83 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
84 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
88 * ziltest is by and large an ugly hack, but very useful in
89 * checking replay without tedious work.
90 * When running ziltest we want to keep all itx's and so maintain
91 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
92 * We subtract TXG_CONCURRENT_STATES to allow for common code.
94 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
97 zil_bp_compare(const void *x1, const void *x2)
99 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
100 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
102 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
104 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
107 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
109 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
116 zil_bp_tree_init(zilog_t *zilog)
118 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
119 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
123 zil_bp_tree_fini(zilog_t *zilog)
125 avl_tree_t *t = &zilog->zl_bp_tree;
129 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
130 kmem_free(zn, sizeof (zil_bp_node_t));
136 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
138 avl_tree_t *t = &zilog->zl_bp_tree;
139 const dva_t *dva = BP_IDENTITY(bp);
143 if (avl_find(t, dva, &where) != NULL)
146 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
148 avl_insert(t, zn, where);
153 static zil_header_t *
154 zil_header_in_syncing_context(zilog_t *zilog)
156 return ((zil_header_t *)zilog->zl_header);
160 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
162 zio_cksum_t *zc = &bp->blk_cksum;
164 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
165 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
166 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
167 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
171 * Read a log block and make sure it's valid.
174 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
177 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
178 uint32_t aflags = ARC_WAIT;
179 arc_buf_t *abuf = NULL;
183 if (zilog->zl_header->zh_claim_txg == 0)
184 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
186 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
187 zio_flags |= ZIO_FLAG_SPECULATIVE;
189 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
190 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
192 error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
193 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
196 zio_cksum_t cksum = bp->blk_cksum;
199 * Validate the checksummed log block.
201 * Sequence numbers should be... sequential. The checksum
202 * verifier for the next block should be bp's checksum plus 1.
204 * Also check the log chain linkage and size used.
206 cksum.zc_word[ZIL_ZC_SEQ]++;
208 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
209 zil_chain_t *zilc = abuf->b_data;
210 char *lr = (char *)(zilc + 1);
211 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
213 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
214 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
218 *end = (char *)dst + len;
219 *nbp = zilc->zc_next_blk;
222 char *lr = abuf->b_data;
223 uint64_t size = BP_GET_LSIZE(bp);
224 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
226 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
227 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
228 (zilc->zc_nused > (size - sizeof (*zilc)))) {
231 bcopy(lr, dst, zilc->zc_nused);
232 *end = (char *)dst + zilc->zc_nused;
233 *nbp = zilc->zc_next_blk;
237 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
244 * Read a TX_WRITE log data block.
247 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
249 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
250 const blkptr_t *bp = &lr->lr_blkptr;
251 uint32_t aflags = ARC_WAIT;
252 arc_buf_t *abuf = NULL;
256 if (BP_IS_HOLE(bp)) {
258 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
262 if (zilog->zl_header->zh_claim_txg == 0)
263 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
265 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
266 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
268 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
269 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
273 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
274 (void) arc_buf_remove_ref(abuf, &abuf);
281 * Parse the intent log, and call parse_func for each valid record within.
284 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
285 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
287 const zil_header_t *zh = zilog->zl_header;
288 boolean_t claimed = !!zh->zh_claim_txg;
289 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
290 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
291 uint64_t max_blk_seq = 0;
292 uint64_t max_lr_seq = 0;
293 uint64_t blk_count = 0;
294 uint64_t lr_count = 0;
295 blkptr_t blk, next_blk;
299 bzero(&next_blk, sizeof(blkptr_t));
302 * Old logs didn't record the maximum zh_claim_lr_seq.
304 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
305 claim_lr_seq = UINT64_MAX;
308 * Starting at the block pointed to by zh_log we read the log chain.
309 * For each block in the chain we strongly check that block to
310 * ensure its validity. We stop when an invalid block is found.
311 * For each block pointer in the chain we call parse_blk_func().
312 * For each record in each valid block we call parse_lr_func().
313 * If the log has been claimed, stop if we encounter a sequence
314 * number greater than the highest claimed sequence number.
316 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
317 zil_bp_tree_init(zilog);
319 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
320 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
324 if (blk_seq > claim_blk_seq)
326 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
328 ASSERT3U(max_blk_seq, <, blk_seq);
329 max_blk_seq = blk_seq;
332 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
335 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
339 for (lrp = lrbuf; lrp < end; lrp += reclen) {
340 lr_t *lr = (lr_t *)lrp;
341 reclen = lr->lrc_reclen;
342 ASSERT3U(reclen, >=, sizeof (lr_t));
343 if (lr->lrc_seq > claim_lr_seq)
345 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
347 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
348 max_lr_seq = lr->lrc_seq;
353 zilog->zl_parse_error = error;
354 zilog->zl_parse_blk_seq = max_blk_seq;
355 zilog->zl_parse_lr_seq = max_lr_seq;
356 zilog->zl_parse_blk_count = blk_count;
357 zilog->zl_parse_lr_count = lr_count;
359 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
360 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
362 zil_bp_tree_fini(zilog);
363 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
369 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
372 * Claim log block if not already committed and not already claimed.
373 * If tx == NULL, just verify that the block is claimable.
375 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
378 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
379 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
380 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
384 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
386 lr_write_t *lr = (lr_write_t *)lrc;
389 if (lrc->lrc_txtype != TX_WRITE)
393 * If the block is not readable, don't claim it. This can happen
394 * in normal operation when a log block is written to disk before
395 * some of the dmu_sync() blocks it points to. In this case, the
396 * transaction cannot have been committed to anyone (we would have
397 * waited for all writes to be stable first), so it is semantically
398 * correct to declare this the end of the log.
400 if (lr->lr_blkptr.blk_birth >= first_txg &&
401 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
403 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
408 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
410 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
416 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
418 lr_write_t *lr = (lr_write_t *)lrc;
419 blkptr_t *bp = &lr->lr_blkptr;
422 * If we previously claimed it, we need to free it.
424 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
425 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
426 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
432 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
436 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
437 lwb->lwb_zilog = zilog;
439 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
440 lwb->lwb_max_txg = txg;
443 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
444 lwb->lwb_nused = sizeof (zil_chain_t);
445 lwb->lwb_sz = BP_GET_LSIZE(bp);
448 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
451 mutex_enter(&zilog->zl_lock);
452 list_insert_tail(&zilog->zl_lwb_list, lwb);
453 mutex_exit(&zilog->zl_lock);
459 * Create an on-disk intent log.
462 zil_create(zilog_t *zilog)
464 const zil_header_t *zh = zilog->zl_header;
472 * Wait for any previous destroy to complete.
474 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
476 ASSERT(zh->zh_claim_txg == 0);
477 ASSERT(zh->zh_replay_seq == 0);
482 * Allocate an initial log block if:
483 * - there isn't one already
484 * - the existing block is the wrong endianess
486 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
487 tx = dmu_tx_create(zilog->zl_os);
488 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
489 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
490 txg = dmu_tx_get_txg(tx);
492 if (!BP_IS_HOLE(&blk)) {
493 zio_free_zil(zilog->zl_spa, txg, &blk);
497 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
498 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
501 zil_init_log_chain(zilog, &blk);
505 * Allocate a log write buffer (lwb) for the first log block.
508 lwb = zil_alloc_lwb(zilog, &blk, txg);
511 * If we just allocated the first log block, commit our transaction
512 * and wait for zil_sync() to stuff the block poiner into zh_log.
513 * (zh is part of the MOS, so we cannot modify it in open context.)
517 txg_wait_synced(zilog->zl_dmu_pool, txg);
520 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
526 * In one tx, free all log blocks and clear the log header.
527 * If keep_first is set, then we're replaying a log with no content.
528 * We want to keep the first block, however, so that the first
529 * synchronous transaction doesn't require a txg_wait_synced()
530 * in zil_create(). We don't need to txg_wait_synced() here either
531 * when keep_first is set, because both zil_create() and zil_destroy()
532 * will wait for any in-progress destroys to complete.
535 zil_destroy(zilog_t *zilog, boolean_t keep_first)
537 const zil_header_t *zh = zilog->zl_header;
543 * Wait for any previous destroy to complete.
545 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
547 zilog->zl_old_header = *zh; /* debugging aid */
549 if (BP_IS_HOLE(&zh->zh_log))
552 tx = dmu_tx_create(zilog->zl_os);
553 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
554 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
555 txg = dmu_tx_get_txg(tx);
557 mutex_enter(&zilog->zl_lock);
559 ASSERT3U(zilog->zl_destroy_txg, <, txg);
560 zilog->zl_destroy_txg = txg;
561 zilog->zl_keep_first = keep_first;
563 if (!list_is_empty(&zilog->zl_lwb_list)) {
564 ASSERT(zh->zh_claim_txg == 0);
566 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
567 list_remove(&zilog->zl_lwb_list, lwb);
568 if (lwb->lwb_buf != NULL)
569 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
570 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
571 kmem_cache_free(zil_lwb_cache, lwb);
573 } else if (!keep_first) {
574 (void) zil_parse(zilog, zil_free_log_block,
575 zil_free_log_record, tx, zh->zh_claim_txg);
577 mutex_exit(&zilog->zl_lock);
583 zil_claim(const char *osname, void *txarg)
585 dmu_tx_t *tx = txarg;
586 uint64_t first_txg = dmu_tx_get_txg(tx);
592 error = dmu_objset_hold(osname, FTAG, &os);
594 cmn_err(CE_WARN, "can't open objset for %s", osname);
598 zilog = dmu_objset_zil(os);
599 zh = zil_header_in_syncing_context(zilog);
601 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
602 if (!BP_IS_HOLE(&zh->zh_log))
603 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
604 BP_ZERO(&zh->zh_log);
605 dsl_dataset_dirty(dmu_objset_ds(os), tx);
606 dmu_objset_rele(os, FTAG);
611 * Claim all log blocks if we haven't already done so, and remember
612 * the highest claimed sequence number. This ensures that if we can
613 * read only part of the log now (e.g. due to a missing device),
614 * but we can read the entire log later, we will not try to replay
615 * or destroy beyond the last block we successfully claimed.
617 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
618 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
619 (void) zil_parse(zilog, zil_claim_log_block,
620 zil_claim_log_record, tx, first_txg);
621 zh->zh_claim_txg = first_txg;
622 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
623 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
624 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
625 zh->zh_flags |= ZIL_REPLAY_NEEDED;
626 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
627 dsl_dataset_dirty(dmu_objset_ds(os), tx);
630 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
631 dmu_objset_rele(os, FTAG);
636 * Check the log by walking the log chain.
637 * Checksum errors are ok as they indicate the end of the chain.
638 * Any other error (no device or read failure) returns an error.
641 zil_check_log_chain(const char *osname, void *tx)
650 error = dmu_objset_hold(osname, FTAG, &os);
652 cmn_err(CE_WARN, "can't open objset for %s", osname);
656 zilog = dmu_objset_zil(os);
657 bp = (blkptr_t *)&zilog->zl_header->zh_log;
660 * Check the first block and determine if it's on a log device
661 * which may have been removed or faulted prior to loading this
662 * pool. If so, there's no point in checking the rest of the log
663 * as its content should have already been synced to the pool.
665 if (!BP_IS_HOLE(bp)) {
667 boolean_t valid = B_TRUE;
669 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
670 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
671 if (vd->vdev_islog && vdev_is_dead(vd))
672 valid = vdev_log_state_valid(vd);
673 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
676 dmu_objset_rele(os, FTAG);
682 * Because tx == NULL, zil_claim_log_block() will not actually claim
683 * any blocks, but just determine whether it is possible to do so.
684 * In addition to checking the log chain, zil_claim_log_block()
685 * will invoke zio_claim() with a done func of spa_claim_notify(),
686 * which will update spa_max_claim_txg. See spa_load() for details.
688 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
689 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
691 dmu_objset_rele(os, FTAG);
693 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
697 zil_vdev_compare(const void *x1, const void *x2)
699 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
700 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
711 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
713 avl_tree_t *t = &zilog->zl_vdev_tree;
715 zil_vdev_node_t *zv, zvsearch;
716 int ndvas = BP_GET_NDVAS(bp);
719 if (zfs_nocacheflush)
722 ASSERT(zilog->zl_writer);
725 * Even though we're zl_writer, we still need a lock because the
726 * zl_get_data() callbacks may have dmu_sync() done callbacks
727 * that will run concurrently.
729 mutex_enter(&zilog->zl_vdev_lock);
730 for (i = 0; i < ndvas; i++) {
731 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
732 if (avl_find(t, &zvsearch, &where) == NULL) {
733 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
734 zv->zv_vdev = zvsearch.zv_vdev;
735 avl_insert(t, zv, where);
738 mutex_exit(&zilog->zl_vdev_lock);
742 zil_flush_vdevs(zilog_t *zilog)
744 spa_t *spa = zilog->zl_spa;
745 avl_tree_t *t = &zilog->zl_vdev_tree;
750 ASSERT(zilog->zl_writer);
753 * We don't need zl_vdev_lock here because we're the zl_writer,
754 * and all zl_get_data() callbacks are done.
756 if (avl_numnodes(t) == 0)
759 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
761 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
763 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
764 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
767 kmem_free(zv, sizeof (*zv));
771 * Wait for all the flushes to complete. Not all devices actually
772 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
774 (void) zio_wait(zio);
776 spa_config_exit(spa, SCL_STATE, FTAG);
780 * Function called when a log block write completes
783 zil_lwb_write_done(zio_t *zio)
785 lwb_t *lwb = zio->io_private;
786 zilog_t *zilog = lwb->lwb_zilog;
787 dmu_tx_t *tx = lwb->lwb_tx;
789 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
790 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
791 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
792 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
793 ASSERT(!BP_IS_GANG(zio->io_bp));
794 ASSERT(!BP_IS_HOLE(zio->io_bp));
795 ASSERT(zio->io_bp->blk_fill == 0);
798 * Ensure the lwb buffer pointer is cleared before releasing
799 * the txg. If we have had an allocation failure and
800 * the txg is waiting to sync then we want want zil_sync()
801 * to remove the lwb so that it's not picked up as the next new
802 * one in zil_commit_writer(). zil_sync() will only remove
803 * the lwb if lwb_buf is null.
805 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
806 mutex_enter(&zilog->zl_lock);
809 mutex_exit(&zilog->zl_lock);
812 * Now that we've written this log block, we have a stable pointer
813 * to the next block in the chain, so it's OK to let the txg in
814 * which we allocated the next block sync.
820 * Initialize the io for a log block.
823 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
827 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
828 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
829 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
831 if (zilog->zl_root_zio == NULL) {
832 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
835 if (lwb->lwb_zio == NULL) {
836 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
837 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
838 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
839 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
844 * Define a limited set of intent log block sizes.
845 * These must be a multiple of 4KB. Note only the amount used (again
846 * aligned to 4KB) actually gets written. However, we can't always just
847 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
849 uint64_t zil_block_buckets[] = {
850 4096, /* non TX_WRITE */
851 8192+4096, /* data base */
852 32*1024 + 4096, /* NFS writes */
857 * Use the slog as long as the logbias is 'latency' and the current commit size
858 * is less than the limit or the total list size is less than 2X the limit.
859 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
861 uint64_t zil_slog_limit = 1024 * 1024;
862 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
863 (((zilog)->zl_cur_used < zil_slog_limit) || \
864 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
867 * Start a log block write and advance to the next log block.
868 * Calls are serialized.
871 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
875 spa_t *spa = zilog->zl_spa;
879 uint64_t zil_blksz, wsz;
882 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
883 zilc = (zil_chain_t *)lwb->lwb_buf;
884 bp = &zilc->zc_next_blk;
886 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
887 bp = &zilc->zc_next_blk;
890 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
893 * Allocate the next block and save its address in this block
894 * before writing it in order to establish the log chain.
895 * Note that if the allocation of nlwb synced before we wrote
896 * the block that points at it (lwb), we'd leak it if we crashed.
897 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
898 * We dirty the dataset to ensure that zil_sync() will be called
899 * to clean up in the event of allocation failure or I/O failure.
901 tx = dmu_tx_create(zilog->zl_os);
902 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
903 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
904 txg = dmu_tx_get_txg(tx);
909 * Log blocks are pre-allocated. Here we select the size of the next
910 * block, based on size used in the last block.
911 * - first find the smallest bucket that will fit the block from a
912 * limited set of block sizes. This is because it's faster to write
913 * blocks allocated from the same metaslab as they are adjacent or
915 * - next find the maximum from the new suggested size and an array of
916 * previous sizes. This lessens a picket fence effect of wrongly
917 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
920 * Note we only write what is used, but we can't just allocate
921 * the maximum block size because we can exhaust the available
924 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
925 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
927 zil_blksz = zil_block_buckets[i];
928 if (zil_blksz == UINT64_MAX)
929 zil_blksz = SPA_MAXBLOCKSIZE;
930 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
931 for (i = 0; i < ZIL_PREV_BLKS; i++)
932 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
933 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
936 /* pass the old blkptr in order to spread log blocks across devs */
937 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
940 ASSERT3U(bp->blk_birth, ==, txg);
941 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
942 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
945 * Allocate a new log write buffer (lwb).
947 nlwb = zil_alloc_lwb(zilog, bp, txg);
949 /* Record the block for later vdev flushing */
950 zil_add_block(zilog, &lwb->lwb_blk);
953 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
954 /* For Slim ZIL only write what is used. */
955 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
956 ASSERT3U(wsz, <=, lwb->lwb_sz);
957 zio_shrink(lwb->lwb_zio, wsz);
964 zilc->zc_nused = lwb->lwb_nused;
965 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
968 * clear unused data for security
970 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
972 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
975 * If there was an allocation failure then nlwb will be null which
976 * forces a txg_wait_synced().
982 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
984 lr_t *lrc = &itx->itx_lr; /* common log record */
985 lr_write_t *lrw = (lr_write_t *)lrc;
987 uint64_t txg = lrc->lrc_txg;
988 uint64_t reclen = lrc->lrc_reclen;
994 ASSERT(lwb->lwb_buf != NULL);
996 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
997 dlen = P2ROUNDUP_TYPED(
998 lrw->lr_length, sizeof (uint64_t), uint64_t);
1000 zilog->zl_cur_used += (reclen + dlen);
1002 zil_lwb_write_init(zilog, lwb);
1005 * If this record won't fit in the current log block, start a new one.
1007 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1008 lwb = zil_lwb_write_start(zilog, lwb);
1011 zil_lwb_write_init(zilog, lwb);
1012 ASSERT(LWB_EMPTY(lwb));
1013 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1014 txg_wait_synced(zilog->zl_dmu_pool, txg);
1019 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1020 bcopy(lrc, lr_buf, reclen);
1021 lrc = (lr_t *)lr_buf;
1022 lrw = (lr_write_t *)lrc;
1025 * If it's a write, fetch the data or get its blkptr as appropriate.
1027 if (lrc->lrc_txtype == TX_WRITE) {
1028 if (txg > spa_freeze_txg(zilog->zl_spa))
1029 txg_wait_synced(zilog->zl_dmu_pool, txg);
1030 if (itx->itx_wr_state != WR_COPIED) {
1035 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1036 dbuf = lr_buf + reclen;
1037 lrw->lr_common.lrc_reclen += dlen;
1039 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1042 error = zilog->zl_get_data(
1043 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1045 txg_wait_synced(zilog->zl_dmu_pool, txg);
1049 ASSERT(error == ENOENT || error == EEXIST ||
1057 * We're actually making an entry, so update lrc_seq to be the
1058 * log record sequence number. Note that this is generally not
1059 * equal to the itx sequence number because not all transactions
1060 * are synchronous, and sometimes spa_sync() gets there first.
1062 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1063 lwb->lwb_nused += reclen + dlen;
1064 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1065 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1066 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1072 zil_itx_create(uint64_t txtype, size_t lrsize)
1076 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1078 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP|KM_NODEBUG);
1079 itx->itx_lr.lrc_txtype = txtype;
1080 itx->itx_lr.lrc_reclen = lrsize;
1081 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1082 itx->itx_lr.lrc_seq = 0; /* defensive */
1083 itx->itx_sync = B_TRUE; /* default is synchronous */
1089 zil_itx_destroy(itx_t *itx)
1091 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1095 * Free up the sync and async itxs. The itxs_t has already been detached
1096 * so no locks are needed.
1099 zil_itxg_clean(itxs_t *itxs)
1105 itx_async_node_t *ian;
1107 list = &itxs->i_sync_list;
1108 while ((itx = list_head(list)) != NULL) {
1109 list_remove(list, itx);
1110 kmem_free(itx, offsetof(itx_t, itx_lr) +
1111 itx->itx_lr.lrc_reclen);
1115 t = &itxs->i_async_tree;
1116 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1117 list = &ian->ia_list;
1118 while ((itx = list_head(list)) != NULL) {
1119 list_remove(list, itx);
1120 kmem_free(itx, offsetof(itx_t, itx_lr) +
1121 itx->itx_lr.lrc_reclen);
1124 kmem_free(ian, sizeof (itx_async_node_t));
1128 kmem_free(itxs, sizeof (itxs_t));
1132 zil_aitx_compare(const void *x1, const void *x2)
1134 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1135 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1146 * Remove all async itx with the given oid.
1149 zil_remove_async(zilog_t *zilog, uint64_t oid)
1152 itx_async_node_t *ian;
1159 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1161 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1164 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1166 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1167 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1169 mutex_enter(&itxg->itxg_lock);
1170 if (itxg->itxg_txg != txg) {
1171 mutex_exit(&itxg->itxg_lock);
1176 * Locate the object node and append its list.
1178 t = &itxg->itxg_itxs->i_async_tree;
1179 ian = avl_find(t, &oid, &where);
1181 list_move_tail(&clean_list, &ian->ia_list);
1182 mutex_exit(&itxg->itxg_lock);
1184 while ((itx = list_head(&clean_list)) != NULL) {
1185 list_remove(&clean_list, itx);
1186 kmem_free(itx, offsetof(itx_t, itx_lr) +
1187 itx->itx_lr.lrc_reclen);
1189 list_destroy(&clean_list);
1193 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1197 itxs_t *itxs, *clean = NULL;
1200 * Object ids can be re-instantiated in the next txg so
1201 * remove any async transactions to avoid future leaks.
1202 * This can happen if a fsync occurs on the re-instantiated
1203 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1204 * the new file data and flushes a write record for the old object.
1206 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1207 zil_remove_async(zilog, itx->itx_oid);
1210 * Ensure the data of a renamed file is committed before the rename.
1212 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1213 zil_async_to_sync(zilog, itx->itx_oid);
1215 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1218 txg = dmu_tx_get_txg(tx);
1220 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1221 mutex_enter(&itxg->itxg_lock);
1222 itxs = itxg->itxg_itxs;
1223 if (itxg->itxg_txg != txg) {
1226 * The zil_clean callback hasn't got around to cleaning
1227 * this itxg. Save the itxs for release below.
1228 * This should be rare.
1230 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1232 clean = itxg->itxg_itxs;
1234 ASSERT(itxg->itxg_sod == 0);
1235 itxg->itxg_txg = txg;
1236 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1238 list_create(&itxs->i_sync_list, sizeof (itx_t),
1239 offsetof(itx_t, itx_node));
1240 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1241 sizeof (itx_async_node_t),
1242 offsetof(itx_async_node_t, ia_node));
1244 if (itx->itx_sync) {
1245 list_insert_tail(&itxs->i_sync_list, itx);
1246 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1247 itxg->itxg_sod += itx->itx_sod;
1249 avl_tree_t *t = &itxs->i_async_tree;
1250 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1251 itx_async_node_t *ian;
1254 ian = avl_find(t, &foid, &where);
1256 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1257 list_create(&ian->ia_list, sizeof (itx_t),
1258 offsetof(itx_t, itx_node));
1259 ian->ia_foid = foid;
1260 avl_insert(t, ian, where);
1262 list_insert_tail(&ian->ia_list, itx);
1265 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1266 mutex_exit(&itxg->itxg_lock);
1268 /* Release the old itxs now we've dropped the lock */
1270 zil_itxg_clean(clean);
1274 * If there are any in-memory intent log transactions which have now been
1275 * synced then start up a taskq to free them.
1278 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1280 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1283 mutex_enter(&itxg->itxg_lock);
1284 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1285 mutex_exit(&itxg->itxg_lock);
1288 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1289 ASSERT(itxg->itxg_txg != 0);
1290 ASSERT(zilog->zl_clean_taskq != NULL);
1291 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1293 clean_me = itxg->itxg_itxs;
1294 itxg->itxg_itxs = NULL;
1296 mutex_exit(&itxg->itxg_lock);
1298 * Preferably start a task queue to free up the old itxs but
1299 * if taskq_dispatch can't allocate resources to do that then
1300 * free it in-line. This should be rare. Note, using TQ_SLEEP
1301 * created a bad performance problem.
1303 if (taskq_dispatch(zilog->zl_clean_taskq,
1304 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1305 zil_itxg_clean(clean_me);
1309 * Get the list of itxs to commit into zl_itx_commit_list.
1312 zil_get_commit_list(zilog_t *zilog)
1315 list_t *commit_list = &zilog->zl_itx_commit_list;
1316 uint64_t push_sod = 0;
1318 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1321 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1323 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1324 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1326 mutex_enter(&itxg->itxg_lock);
1327 if (itxg->itxg_txg != txg) {
1328 mutex_exit(&itxg->itxg_lock);
1332 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1333 push_sod += itxg->itxg_sod;
1336 mutex_exit(&itxg->itxg_lock);
1338 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1342 * Move the async itxs for a specified object to commit into sync lists.
1345 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1348 itx_async_node_t *ian;
1352 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1355 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1357 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1358 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1360 mutex_enter(&itxg->itxg_lock);
1361 if (itxg->itxg_txg != txg) {
1362 mutex_exit(&itxg->itxg_lock);
1367 * If a foid is specified then find that node and append its
1368 * list. Otherwise walk the tree appending all the lists
1369 * to the sync list. We add to the end rather than the
1370 * beginning to ensure the create has happened.
1372 t = &itxg->itxg_itxs->i_async_tree;
1374 ian = avl_find(t, &foid, &where);
1376 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1380 void *cookie = NULL;
1382 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1383 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1385 list_destroy(&ian->ia_list);
1386 kmem_free(ian, sizeof (itx_async_node_t));
1389 mutex_exit(&itxg->itxg_lock);
1394 zil_commit_writer(zilog_t *zilog)
1399 spa_t *spa = zilog->zl_spa;
1402 ASSERT(zilog->zl_root_zio == NULL);
1404 mutex_exit(&zilog->zl_lock);
1406 zil_get_commit_list(zilog);
1409 * Return if there's nothing to commit before we dirty the fs by
1410 * calling zil_create().
1412 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1413 mutex_enter(&zilog->zl_lock);
1417 if (zilog->zl_suspend) {
1420 lwb = list_tail(&zilog->zl_lwb_list);
1422 lwb = zil_create(zilog);
1425 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1426 while ((itx = list_head(&zilog->zl_itx_commit_list))) {
1427 txg = itx->itx_lr.lrc_txg;
1430 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1431 lwb = zil_lwb_commit(zilog, itx, lwb);
1432 list_remove(&zilog->zl_itx_commit_list, itx);
1433 kmem_free(itx, offsetof(itx_t, itx_lr)
1434 + itx->itx_lr.lrc_reclen);
1436 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1438 /* write the last block out */
1439 if (lwb != NULL && lwb->lwb_zio != NULL)
1440 lwb = zil_lwb_write_start(zilog, lwb);
1442 zilog->zl_cur_used = 0;
1445 * Wait if necessary for the log blocks to be on stable storage.
1447 if (zilog->zl_root_zio) {
1448 error = zio_wait(zilog->zl_root_zio);
1449 zilog->zl_root_zio = NULL;
1450 zil_flush_vdevs(zilog);
1453 if (error || lwb == NULL)
1454 txg_wait_synced(zilog->zl_dmu_pool, 0);
1456 mutex_enter(&zilog->zl_lock);
1459 * Remember the highest committed log sequence number for ztest.
1460 * We only update this value when all the log writes succeeded,
1461 * because ztest wants to ASSERT that it got the whole log chain.
1463 if (error == 0 && lwb != NULL)
1464 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1468 * Commit zfs transactions to stable storage.
1469 * If foid is 0 push out all transactions, otherwise push only those
1470 * for that object or might reference that object.
1472 * itxs are committed in batches. In a heavily stressed zil there will be
1473 * a commit writer thread who is writing out a bunch of itxs to the log
1474 * for a set of committing threads (cthreads) in the same batch as the writer.
1475 * Those cthreads are all waiting on the same cv for that batch.
1477 * There will also be a different and growing batch of threads that are
1478 * waiting to commit (qthreads). When the committing batch completes
1479 * a transition occurs such that the cthreads exit and the qthreads become
1480 * cthreads. One of the new cthreads becomes the writer thread for the
1481 * batch. Any new threads arriving become new qthreads.
1483 * Only 2 condition variables are needed and there's no transition
1484 * between the two cvs needed. They just flip-flop between qthreads
1487 * Using this scheme we can efficiently wakeup up only those threads
1488 * that have been committed.
1491 zil_commit(zilog_t *zilog, uint64_t foid)
1495 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1498 /* move the async itxs for the foid to the sync queues */
1499 zil_async_to_sync(zilog, foid);
1501 mutex_enter(&zilog->zl_lock);
1502 mybatch = zilog->zl_next_batch;
1503 while (zilog->zl_writer) {
1504 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1505 if (mybatch <= zilog->zl_com_batch) {
1506 mutex_exit(&zilog->zl_lock);
1511 zilog->zl_next_batch++;
1512 zilog->zl_writer = B_TRUE;
1513 zil_commit_writer(zilog);
1514 zilog->zl_com_batch = mybatch;
1515 zilog->zl_writer = B_FALSE;
1516 mutex_exit(&zilog->zl_lock);
1518 /* wake up one thread to become the next writer */
1519 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1521 /* wake up all threads waiting for this batch to be committed */
1522 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1526 * Called in syncing context to free committed log blocks and update log header.
1529 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1531 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1532 uint64_t txg = dmu_tx_get_txg(tx);
1533 spa_t *spa = zilog->zl_spa;
1534 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1538 * We don't zero out zl_destroy_txg, so make sure we don't try
1539 * to destroy it twice.
1541 if (spa_sync_pass(spa) != 1)
1544 mutex_enter(&zilog->zl_lock);
1546 ASSERT(zilog->zl_stop_sync == 0);
1548 if (*replayed_seq != 0) {
1549 ASSERT(zh->zh_replay_seq < *replayed_seq);
1550 zh->zh_replay_seq = *replayed_seq;
1554 if (zilog->zl_destroy_txg == txg) {
1555 blkptr_t blk = zh->zh_log;
1557 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1559 bzero(zh, sizeof (zil_header_t));
1560 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1562 if (zilog->zl_keep_first) {
1564 * If this block was part of log chain that couldn't
1565 * be claimed because a device was missing during
1566 * zil_claim(), but that device later returns,
1567 * then this block could erroneously appear valid.
1568 * To guard against this, assign a new GUID to the new
1569 * log chain so it doesn't matter what blk points to.
1571 zil_init_log_chain(zilog, &blk);
1576 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1577 zh->zh_log = lwb->lwb_blk;
1578 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1580 list_remove(&zilog->zl_lwb_list, lwb);
1581 zio_free_zil(spa, txg, &lwb->lwb_blk);
1582 kmem_cache_free(zil_lwb_cache, lwb);
1585 * If we don't have anything left in the lwb list then
1586 * we've had an allocation failure and we need to zero
1587 * out the zil_header blkptr so that we don't end
1588 * up freeing the same block twice.
1590 if (list_head(&zilog->zl_lwb_list) == NULL)
1591 BP_ZERO(&zh->zh_log);
1593 mutex_exit(&zilog->zl_lock);
1599 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1600 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1606 kmem_cache_destroy(zil_lwb_cache);
1610 zil_set_sync(zilog_t *zilog, uint64_t sync)
1612 zilog->zl_sync = sync;
1616 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1618 zilog->zl_logbias = logbias;
1622 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1627 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1629 zilog->zl_header = zh_phys;
1631 zilog->zl_spa = dmu_objset_spa(os);
1632 zilog->zl_dmu_pool = dmu_objset_pool(os);
1633 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1634 zilog->zl_logbias = dmu_objset_logbias(os);
1635 zilog->zl_sync = dmu_objset_syncprop(os);
1636 zilog->zl_next_batch = 1;
1638 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1640 for (i = 0; i < TXG_SIZE; i++) {
1641 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1642 MUTEX_DEFAULT, NULL);
1645 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1646 offsetof(lwb_t, lwb_node));
1648 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1649 offsetof(itx_t, itx_node));
1651 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1653 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1654 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1656 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1657 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1658 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1659 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1665 zil_free(zilog_t *zilog)
1670 zilog->zl_stop_sync = 1;
1673 * After zil_close() there should only be one lwb with a buffer.
1675 head_lwb = list_head(&zilog->zl_lwb_list);
1677 ASSERT(head_lwb == list_tail(&zilog->zl_lwb_list));
1678 list_remove(&zilog->zl_lwb_list, head_lwb);
1679 zio_buf_free(head_lwb->lwb_buf, head_lwb->lwb_sz);
1680 kmem_cache_free(zil_lwb_cache, head_lwb);
1682 list_destroy(&zilog->zl_lwb_list);
1684 avl_destroy(&zilog->zl_vdev_tree);
1685 mutex_destroy(&zilog->zl_vdev_lock);
1687 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1688 list_destroy(&zilog->zl_itx_commit_list);
1690 for (i = 0; i < TXG_SIZE; i++) {
1692 * It's possible for an itx to be generated that doesn't dirty
1693 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1694 * callback to remove the entry. We remove those here.
1696 * Also free up the ziltest itxs.
1698 if (zilog->zl_itxg[i].itxg_itxs)
1699 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1700 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1703 mutex_destroy(&zilog->zl_lock);
1705 cv_destroy(&zilog->zl_cv_writer);
1706 cv_destroy(&zilog->zl_cv_suspend);
1707 cv_destroy(&zilog->zl_cv_batch[0]);
1708 cv_destroy(&zilog->zl_cv_batch[1]);
1710 kmem_free(zilog, sizeof (zilog_t));
1714 * Open an intent log.
1717 zil_open(objset_t *os, zil_get_data_t *get_data)
1719 zilog_t *zilog = dmu_objset_zil(os);
1721 zilog->zl_get_data = get_data;
1722 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1723 2, 2, TASKQ_PREPOPULATE);
1729 * Close an intent log.
1732 zil_close(zilog_t *zilog)
1737 zil_commit(zilog, 0); /* commit all itx */
1740 * The lwb_max_txg for the stubby lwb will reflect the last activity
1741 * for the zil. After a txg_wait_synced() on the txg we know all the
1742 * callbacks have occurred that may clean the zil. Only then can we
1743 * destroy the zl_clean_taskq.
1745 mutex_enter(&zilog->zl_lock);
1746 tail_lwb = list_tail(&zilog->zl_lwb_list);
1747 if (tail_lwb != NULL)
1748 txg = tail_lwb->lwb_max_txg;
1749 mutex_exit(&zilog->zl_lock);
1751 txg_wait_synced(zilog->zl_dmu_pool, txg);
1753 taskq_destroy(zilog->zl_clean_taskq);
1754 zilog->zl_clean_taskq = NULL;
1755 zilog->zl_get_data = NULL;
1759 * Suspend an intent log. While in suspended mode, we still honor
1760 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1761 * We suspend the log briefly when taking a snapshot so that the snapshot
1762 * contains all the data it's supposed to, and has an empty intent log.
1765 zil_suspend(zilog_t *zilog)
1767 const zil_header_t *zh = zilog->zl_header;
1769 mutex_enter(&zilog->zl_lock);
1770 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1771 mutex_exit(&zilog->zl_lock);
1774 if (zilog->zl_suspend++ != 0) {
1776 * Someone else already began a suspend.
1777 * Just wait for them to finish.
1779 while (zilog->zl_suspending)
1780 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1781 mutex_exit(&zilog->zl_lock);
1784 zilog->zl_suspending = B_TRUE;
1785 mutex_exit(&zilog->zl_lock);
1787 zil_commit(zilog, 0);
1789 zil_destroy(zilog, B_FALSE);
1791 mutex_enter(&zilog->zl_lock);
1792 zilog->zl_suspending = B_FALSE;
1793 cv_broadcast(&zilog->zl_cv_suspend);
1794 mutex_exit(&zilog->zl_lock);
1800 zil_resume(zilog_t *zilog)
1802 mutex_enter(&zilog->zl_lock);
1803 ASSERT(zilog->zl_suspend != 0);
1804 zilog->zl_suspend--;
1805 mutex_exit(&zilog->zl_lock);
1808 typedef struct zil_replay_arg {
1809 zil_replay_func_t **zr_replay;
1811 boolean_t zr_byteswap;
1816 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1818 char name[MAXNAMELEN];
1820 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1822 dmu_objset_name(zilog->zl_os, name);
1824 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1825 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1826 (u_longlong_t)lr->lrc_seq,
1827 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1828 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1834 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1836 zil_replay_arg_t *zr = zra;
1837 const zil_header_t *zh = zilog->zl_header;
1838 uint64_t reclen = lr->lrc_reclen;
1839 uint64_t txtype = lr->lrc_txtype;
1842 zilog->zl_replaying_seq = lr->lrc_seq;
1844 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1847 if (lr->lrc_txg < claim_txg) /* already committed */
1850 /* Strip case-insensitive bit, still present in log record */
1853 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1854 return (zil_replay_error(zilog, lr, EINVAL));
1857 * If this record type can be logged out of order, the object
1858 * (lr_foid) may no longer exist. That's legitimate, not an error.
1860 if (TX_OOO(txtype)) {
1861 error = dmu_object_info(zilog->zl_os,
1862 ((lr_ooo_t *)lr)->lr_foid, NULL);
1863 if (error == ENOENT || error == EEXIST)
1868 * Make a copy of the data so we can revise and extend it.
1870 bcopy(lr, zr->zr_lr, reclen);
1873 * If this is a TX_WRITE with a blkptr, suck in the data.
1875 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1876 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1877 zr->zr_lr + reclen);
1879 return (zil_replay_error(zilog, lr, error));
1883 * The log block containing this lr may have been byteswapped
1884 * so that we can easily examine common fields like lrc_txtype.
1885 * However, the log is a mix of different record types, and only the
1886 * replay vectors know how to byteswap their records. Therefore, if
1887 * the lr was byteswapped, undo it before invoking the replay vector.
1889 if (zr->zr_byteswap)
1890 byteswap_uint64_array(zr->zr_lr, reclen);
1893 * We must now do two things atomically: replay this log record,
1894 * and update the log header sequence number to reflect the fact that
1895 * we did so. At the end of each replay function the sequence number
1896 * is updated if we are in replay mode.
1898 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1901 * The DMU's dnode layer doesn't see removes until the txg
1902 * commits, so a subsequent claim can spuriously fail with
1903 * EEXIST. So if we receive any error we try syncing out
1904 * any removes then retry the transaction. Note that we
1905 * specify B_FALSE for byteswap now, so we don't do it twice.
1907 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1908 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1910 return (zil_replay_error(zilog, lr, error));
1917 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1919 zilog->zl_replay_blks++;
1925 * If this dataset has a non-empty intent log, replay it and destroy it.
1928 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1930 zilog_t *zilog = dmu_objset_zil(os);
1931 const zil_header_t *zh = zilog->zl_header;
1932 zil_replay_arg_t zr;
1934 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1935 zil_destroy(zilog, B_TRUE);
1939 zr.zr_replay = replay_func;
1941 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1942 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1945 * Wait for in-progress removes to sync before starting replay.
1947 txg_wait_synced(zilog->zl_dmu_pool, 0);
1949 zilog->zl_replay = B_TRUE;
1950 zilog->zl_replay_time = ddi_get_lbolt();
1951 ASSERT(zilog->zl_replay_blks == 0);
1952 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1954 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1956 zil_destroy(zilog, B_FALSE);
1957 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1958 zilog->zl_replay = B_FALSE;
1962 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1964 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1967 if (zilog->zl_replay) {
1968 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1969 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1970 zilog->zl_replaying_seq;
1979 zil_vdev_offline(const char *osname, void *arg)
1985 error = dmu_objset_hold(osname, FTAG, &os);
1989 zilog = dmu_objset_zil(os);
1990 if (zil_suspend(zilog) != 0)
1994 dmu_objset_rele(os, FTAG);