4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/metaslab.h>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * See zil.h for more information about these fields.
72 zil_stats_t zil_stats
= {
73 { "zil_commit_count", KSTAT_DATA_UINT64
},
74 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
75 { "zil_itx_count", KSTAT_DATA_UINT64
},
76 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
77 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
78 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
79 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
80 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
81 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
82 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
83 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
84 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
85 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
88 static kstat_t
*zil_ksp
;
91 * Disable intent logging replay. This global ZIL switch affects all pools.
93 int zil_replay_disable
= 0;
96 * Tunable parameter for debugging or performance analysis. Setting
97 * zfs_nocacheflush will cause corruption on power loss if a volatile
98 * out-of-order write cache is enabled.
100 int zfs_nocacheflush
= 0;
102 static kmem_cache_t
*zil_lwb_cache
;
104 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
106 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
107 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
111 * ziltest is by and large an ugly hack, but very useful in
112 * checking replay without tedious work.
113 * When running ziltest we want to keep all itx's and so maintain
114 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
115 * We subtract TXG_CONCURRENT_STATES to allow for common code.
117 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
120 zil_bp_compare(const void *x1
, const void *x2
)
122 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
123 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
125 if (DVA_GET_VDEV(dva1
) < DVA_GET_VDEV(dva2
))
127 if (DVA_GET_VDEV(dva1
) > DVA_GET_VDEV(dva2
))
130 if (DVA_GET_OFFSET(dva1
) < DVA_GET_OFFSET(dva2
))
132 if (DVA_GET_OFFSET(dva1
) > DVA_GET_OFFSET(dva2
))
139 zil_bp_tree_init(zilog_t
*zilog
)
141 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
142 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
146 zil_bp_tree_fini(zilog_t
*zilog
)
148 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
152 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
153 kmem_free(zn
, sizeof (zil_bp_node_t
));
159 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
161 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
162 const dva_t
*dva
= BP_IDENTITY(bp
);
166 if (avl_find(t
, dva
, &where
) != NULL
)
167 return (SET_ERROR(EEXIST
));
169 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_PUSHPAGE
);
171 avl_insert(t
, zn
, where
);
176 static zil_header_t
*
177 zil_header_in_syncing_context(zilog_t
*zilog
)
179 return ((zil_header_t
*)zilog
->zl_header
);
183 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
185 zio_cksum_t
*zc
= &bp
->blk_cksum
;
187 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
188 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
189 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
190 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
194 * Read a log block and make sure it's valid.
197 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
200 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
201 uint32_t aflags
= ARC_WAIT
;
202 arc_buf_t
*abuf
= NULL
;
206 if (zilog
->zl_header
->zh_claim_txg
== 0)
207 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
209 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
210 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
212 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
213 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
215 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
216 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
219 zio_cksum_t cksum
= bp
->blk_cksum
;
222 * Validate the checksummed log block.
224 * Sequence numbers should be... sequential. The checksum
225 * verifier for the next block should be bp's checksum plus 1.
227 * Also check the log chain linkage and size used.
229 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
231 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
232 zil_chain_t
*zilc
= abuf
->b_data
;
233 char *lr
= (char *)(zilc
+ 1);
234 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
236 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
237 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
238 error
= SET_ERROR(ECKSUM
);
241 *end
= (char *)dst
+ len
;
242 *nbp
= zilc
->zc_next_blk
;
245 char *lr
= abuf
->b_data
;
246 uint64_t size
= BP_GET_LSIZE(bp
);
247 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
249 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
250 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
251 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
252 error
= SET_ERROR(ECKSUM
);
254 bcopy(lr
, dst
, zilc
->zc_nused
);
255 *end
= (char *)dst
+ zilc
->zc_nused
;
256 *nbp
= zilc
->zc_next_blk
;
260 VERIFY(arc_buf_remove_ref(abuf
, &abuf
));
267 * Read a TX_WRITE log data block.
270 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
272 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
273 const blkptr_t
*bp
= &lr
->lr_blkptr
;
274 uint32_t aflags
= ARC_WAIT
;
275 arc_buf_t
*abuf
= NULL
;
279 if (BP_IS_HOLE(bp
)) {
281 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
285 if (zilog
->zl_header
->zh_claim_txg
== 0)
286 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
288 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
289 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
291 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
292 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
296 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
297 (void) arc_buf_remove_ref(abuf
, &abuf
);
304 * Parse the intent log, and call parse_func for each valid record within.
307 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
308 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
310 const zil_header_t
*zh
= zilog
->zl_header
;
311 boolean_t claimed
= !!zh
->zh_claim_txg
;
312 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
313 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
314 uint64_t max_blk_seq
= 0;
315 uint64_t max_lr_seq
= 0;
316 uint64_t blk_count
= 0;
317 uint64_t lr_count
= 0;
318 blkptr_t blk
, next_blk
;
322 bzero(&next_blk
, sizeof (blkptr_t
));
325 * Old logs didn't record the maximum zh_claim_lr_seq.
327 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
328 claim_lr_seq
= UINT64_MAX
;
331 * Starting at the block pointed to by zh_log we read the log chain.
332 * For each block in the chain we strongly check that block to
333 * ensure its validity. We stop when an invalid block is found.
334 * For each block pointer in the chain we call parse_blk_func().
335 * For each record in each valid block we call parse_lr_func().
336 * If the log has been claimed, stop if we encounter a sequence
337 * number greater than the highest claimed sequence number.
339 lrbuf
= zio_buf_alloc(SPA_MAXBLOCKSIZE
);
340 zil_bp_tree_init(zilog
);
342 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
343 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
347 if (blk_seq
> claim_blk_seq
)
349 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
351 ASSERT3U(max_blk_seq
, <, blk_seq
);
352 max_blk_seq
= blk_seq
;
355 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
358 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
362 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
363 lr_t
*lr
= (lr_t
*)lrp
;
364 reclen
= lr
->lrc_reclen
;
365 ASSERT3U(reclen
, >=, sizeof (lr_t
));
366 if (lr
->lrc_seq
> claim_lr_seq
)
368 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
370 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
371 max_lr_seq
= lr
->lrc_seq
;
376 zilog
->zl_parse_error
= error
;
377 zilog
->zl_parse_blk_seq
= max_blk_seq
;
378 zilog
->zl_parse_lr_seq
= max_lr_seq
;
379 zilog
->zl_parse_blk_count
= blk_count
;
380 zilog
->zl_parse_lr_count
= lr_count
;
382 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
383 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
385 zil_bp_tree_fini(zilog
);
386 zio_buf_free(lrbuf
, SPA_MAXBLOCKSIZE
);
392 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
395 * Claim log block if not already committed and not already claimed.
396 * If tx == NULL, just verify that the block is claimable.
398 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
399 zil_bp_tree_add(zilog
, bp
) != 0)
402 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
403 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
404 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
408 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
410 lr_write_t
*lr
= (lr_write_t
*)lrc
;
413 if (lrc
->lrc_txtype
!= TX_WRITE
)
417 * If the block is not readable, don't claim it. This can happen
418 * in normal operation when a log block is written to disk before
419 * some of the dmu_sync() blocks it points to. In this case, the
420 * transaction cannot have been committed to anyone (we would have
421 * waited for all writes to be stable first), so it is semantically
422 * correct to declare this the end of the log.
424 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
425 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
427 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
432 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
434 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
440 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
442 lr_write_t
*lr
= (lr_write_t
*)lrc
;
443 blkptr_t
*bp
= &lr
->lr_blkptr
;
446 * If we previously claimed it, we need to free it.
448 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
449 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
451 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
457 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, uint64_t txg
, boolean_t fastwrite
)
461 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_PUSHPAGE
);
462 lwb
->lwb_zilog
= zilog
;
464 lwb
->lwb_fastwrite
= fastwrite
;
465 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
466 lwb
->lwb_max_txg
= txg
;
469 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
470 lwb
->lwb_nused
= sizeof (zil_chain_t
);
471 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
474 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
477 mutex_enter(&zilog
->zl_lock
);
478 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
479 mutex_exit(&zilog
->zl_lock
);
485 * Called when we create in-memory log transactions so that we know
486 * to cleanup the itxs at the end of spa_sync().
489 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
491 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
492 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
494 if (dsl_dataset_is_snapshot(ds
))
495 panic("dirtying snapshot!");
497 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
498 /* up the hold count until we can be written out */
499 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
504 zilog_is_dirty(zilog_t
*zilog
)
506 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
509 for (t
= 0; t
< TXG_SIZE
; t
++) {
510 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
517 * Create an on-disk intent log.
520 zil_create(zilog_t
*zilog
)
522 const zil_header_t
*zh
= zilog
->zl_header
;
528 boolean_t fastwrite
= FALSE
;
531 * Wait for any previous destroy to complete.
533 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
535 ASSERT(zh
->zh_claim_txg
== 0);
536 ASSERT(zh
->zh_replay_seq
== 0);
541 * Allocate an initial log block if:
542 * - there isn't one already
543 * - the existing block is the wrong endianess
545 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
546 tx
= dmu_tx_create(zilog
->zl_os
);
547 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
548 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
549 txg
= dmu_tx_get_txg(tx
);
551 if (!BP_IS_HOLE(&blk
)) {
552 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
556 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
,
557 ZIL_MIN_BLKSZ
, B_TRUE
);
561 zil_init_log_chain(zilog
, &blk
);
565 * Allocate a log write buffer (lwb) for the first log block.
568 lwb
= zil_alloc_lwb(zilog
, &blk
, txg
, fastwrite
);
571 * If we just allocated the first log block, commit our transaction
572 * and wait for zil_sync() to stuff the block poiner into zh_log.
573 * (zh is part of the MOS, so we cannot modify it in open context.)
577 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
580 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
586 * In one tx, free all log blocks and clear the log header.
587 * If keep_first is set, then we're replaying a log with no content.
588 * We want to keep the first block, however, so that the first
589 * synchronous transaction doesn't require a txg_wait_synced()
590 * in zil_create(). We don't need to txg_wait_synced() here either
591 * when keep_first is set, because both zil_create() and zil_destroy()
592 * will wait for any in-progress destroys to complete.
595 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
597 const zil_header_t
*zh
= zilog
->zl_header
;
603 * Wait for any previous destroy to complete.
605 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
607 zilog
->zl_old_header
= *zh
; /* debugging aid */
609 if (BP_IS_HOLE(&zh
->zh_log
))
612 tx
= dmu_tx_create(zilog
->zl_os
);
613 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
614 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
615 txg
= dmu_tx_get_txg(tx
);
617 mutex_enter(&zilog
->zl_lock
);
619 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
620 zilog
->zl_destroy_txg
= txg
;
621 zilog
->zl_keep_first
= keep_first
;
623 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
624 ASSERT(zh
->zh_claim_txg
== 0);
626 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
627 ASSERT(lwb
->lwb_zio
== NULL
);
628 if (lwb
->lwb_fastwrite
)
629 metaslab_fastwrite_unmark(zilog
->zl_spa
,
631 list_remove(&zilog
->zl_lwb_list
, lwb
);
632 if (lwb
->lwb_buf
!= NULL
)
633 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
634 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
635 kmem_cache_free(zil_lwb_cache
, lwb
);
637 } else if (!keep_first
) {
638 zil_destroy_sync(zilog
, tx
);
640 mutex_exit(&zilog
->zl_lock
);
646 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
648 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
649 (void) zil_parse(zilog
, zil_free_log_block
,
650 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
);
654 zil_claim(const char *osname
, void *txarg
)
656 dmu_tx_t
*tx
= txarg
;
657 uint64_t first_txg
= dmu_tx_get_txg(tx
);
663 error
= dmu_objset_own(osname
, DMU_OST_ANY
, B_FALSE
, FTAG
, &os
);
665 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
669 zilog
= dmu_objset_zil(os
);
670 zh
= zil_header_in_syncing_context(zilog
);
672 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
673 if (!BP_IS_HOLE(&zh
->zh_log
))
674 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
675 BP_ZERO(&zh
->zh_log
);
676 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
677 dmu_objset_disown(os
, FTAG
);
682 * Claim all log blocks if we haven't already done so, and remember
683 * the highest claimed sequence number. This ensures that if we can
684 * read only part of the log now (e.g. due to a missing device),
685 * but we can read the entire log later, we will not try to replay
686 * or destroy beyond the last block we successfully claimed.
688 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
689 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
690 (void) zil_parse(zilog
, zil_claim_log_block
,
691 zil_claim_log_record
, tx
, first_txg
);
692 zh
->zh_claim_txg
= first_txg
;
693 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
694 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
695 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
696 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
697 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
698 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
701 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
702 dmu_objset_disown(os
, FTAG
);
707 * Check the log by walking the log chain.
708 * Checksum errors are ok as they indicate the end of the chain.
709 * Any other error (no device or read failure) returns an error.
712 zil_check_log_chain(const char *osname
, void *tx
)
721 error
= dmu_objset_hold(osname
, FTAG
, &os
);
723 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
727 zilog
= dmu_objset_zil(os
);
728 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
731 * Check the first block and determine if it's on a log device
732 * which may have been removed or faulted prior to loading this
733 * pool. If so, there's no point in checking the rest of the log
734 * as its content should have already been synced to the pool.
736 if (!BP_IS_HOLE(bp
)) {
738 boolean_t valid
= B_TRUE
;
740 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
741 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
742 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
743 valid
= vdev_log_state_valid(vd
);
744 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
747 dmu_objset_rele(os
, FTAG
);
753 * Because tx == NULL, zil_claim_log_block() will not actually claim
754 * any blocks, but just determine whether it is possible to do so.
755 * In addition to checking the log chain, zil_claim_log_block()
756 * will invoke zio_claim() with a done func of spa_claim_notify(),
757 * which will update spa_max_claim_txg. See spa_load() for details.
759 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
760 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
762 dmu_objset_rele(os
, FTAG
);
764 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
768 zil_vdev_compare(const void *x1
, const void *x2
)
770 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
771 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
782 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
784 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
786 zil_vdev_node_t
*zv
, zvsearch
;
787 int ndvas
= BP_GET_NDVAS(bp
);
790 if (zfs_nocacheflush
)
793 ASSERT(zilog
->zl_writer
);
796 * Even though we're zl_writer, we still need a lock because the
797 * zl_get_data() callbacks may have dmu_sync() done callbacks
798 * that will run concurrently.
800 mutex_enter(&zilog
->zl_vdev_lock
);
801 for (i
= 0; i
< ndvas
; i
++) {
802 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
803 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
804 zv
= kmem_alloc(sizeof (*zv
), KM_PUSHPAGE
);
805 zv
->zv_vdev
= zvsearch
.zv_vdev
;
806 avl_insert(t
, zv
, where
);
809 mutex_exit(&zilog
->zl_vdev_lock
);
813 zil_flush_vdevs(zilog_t
*zilog
)
815 spa_t
*spa
= zilog
->zl_spa
;
816 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
821 ASSERT(zilog
->zl_writer
);
824 * We don't need zl_vdev_lock here because we're the zl_writer,
825 * and all zl_get_data() callbacks are done.
827 if (avl_numnodes(t
) == 0)
830 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
832 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
834 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
835 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
838 kmem_free(zv
, sizeof (*zv
));
842 * Wait for all the flushes to complete. Not all devices actually
843 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
845 (void) zio_wait(zio
);
847 spa_config_exit(spa
, SCL_STATE
, FTAG
);
851 * Function called when a log block write completes
854 zil_lwb_write_done(zio_t
*zio
)
856 lwb_t
*lwb
= zio
->io_private
;
857 zilog_t
*zilog
= lwb
->lwb_zilog
;
858 dmu_tx_t
*tx
= lwb
->lwb_tx
;
860 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
861 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
862 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
863 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
864 ASSERT(!BP_IS_GANG(zio
->io_bp
));
865 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
866 ASSERT(zio
->io_bp
->blk_fill
== 0);
869 * Ensure the lwb buffer pointer is cleared before releasing
870 * the txg. If we have had an allocation failure and
871 * the txg is waiting to sync then we want want zil_sync()
872 * to remove the lwb so that it's not picked up as the next new
873 * one in zil_commit_writer(). zil_sync() will only remove
874 * the lwb if lwb_buf is null.
876 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
877 mutex_enter(&zilog
->zl_lock
);
879 lwb
->lwb_fastwrite
= FALSE
;
882 mutex_exit(&zilog
->zl_lock
);
885 * Now that we've written this log block, we have a stable pointer
886 * to the next block in the chain, so it's OK to let the txg in
887 * which we allocated the next block sync.
893 * Initialize the io for a log block.
896 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
900 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
901 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
902 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
904 if (zilog
->zl_root_zio
== NULL
) {
905 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
909 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
910 mutex_enter(&zilog
->zl_lock
);
911 if (lwb
->lwb_zio
== NULL
) {
912 if (!lwb
->lwb_fastwrite
) {
913 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
914 lwb
->lwb_fastwrite
= 1;
916 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
917 0, &lwb
->lwb_blk
, lwb
->lwb_buf
, BP_GET_LSIZE(&lwb
->lwb_blk
),
918 zil_lwb_write_done
, lwb
, ZIO_PRIORITY_SYNC_WRITE
,
919 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
920 ZIO_FLAG_FASTWRITE
, &zb
);
922 mutex_exit(&zilog
->zl_lock
);
926 * Define a limited set of intent log block sizes.
928 * These must be a multiple of 4KB. Note only the amount used (again
929 * aligned to 4KB) actually gets written. However, we can't always just
930 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
932 uint64_t zil_block_buckets
[] = {
933 4096, /* non TX_WRITE */
934 8192+4096, /* data base */
935 32*1024 + 4096, /* NFS writes */
940 * Use the slog as long as the current commit size is less than the
941 * limit or the total list size is less than 2X the limit. Limit
942 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
944 unsigned long zil_slog_limit
= 1024 * 1024;
945 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
946 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
949 * Start a log block write and advance to the next log block.
950 * Calls are serialized.
953 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
957 spa_t
*spa
= zilog
->zl_spa
;
961 uint64_t zil_blksz
, wsz
;
965 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
966 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
967 bp
= &zilc
->zc_next_blk
;
969 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
970 bp
= &zilc
->zc_next_blk
;
973 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
976 * Allocate the next block and save its address in this block
977 * before writing it in order to establish the log chain.
978 * Note that if the allocation of nlwb synced before we wrote
979 * the block that points at it (lwb), we'd leak it if we crashed.
980 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
981 * We dirty the dataset to ensure that zil_sync() will be called
982 * to clean up in the event of allocation failure or I/O failure.
984 tx
= dmu_tx_create(zilog
->zl_os
);
985 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
986 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
987 txg
= dmu_tx_get_txg(tx
);
992 * Log blocks are pre-allocated. Here we select the size of the next
993 * block, based on size used in the last block.
994 * - first find the smallest bucket that will fit the block from a
995 * limited set of block sizes. This is because it's faster to write
996 * blocks allocated from the same metaslab as they are adjacent or
998 * - next find the maximum from the new suggested size and an array of
999 * previous sizes. This lessens a picket fence effect of wrongly
1000 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1003 * Note we only write what is used, but we can't just allocate
1004 * the maximum block size because we can exhaust the available
1007 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1008 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1010 zil_blksz
= zil_block_buckets
[i
];
1011 if (zil_blksz
== UINT64_MAX
)
1012 zil_blksz
= SPA_MAXBLOCKSIZE
;
1013 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1014 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1015 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1016 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1019 use_slog
= USE_SLOG(zilog
);
1020 error
= zio_alloc_zil(spa
, txg
, bp
, zil_blksz
,
1023 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1024 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1026 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1027 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1030 ASSERT3U(bp
->blk_birth
, ==, txg
);
1031 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1032 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1035 * Allocate a new log write buffer (lwb).
1037 nlwb
= zil_alloc_lwb(zilog
, bp
, txg
, TRUE
);
1039 /* Record the block for later vdev flushing */
1040 zil_add_block(zilog
, &lwb
->lwb_blk
);
1043 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1044 /* For Slim ZIL only write what is used. */
1045 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1046 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1047 zio_shrink(lwb
->lwb_zio
, wsz
);
1054 zilc
->zc_nused
= lwb
->lwb_nused
;
1055 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1058 * clear unused data for security
1060 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1062 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
1065 * If there was an allocation failure then nlwb will be null which
1066 * forces a txg_wait_synced().
1072 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1074 lr_t
*lrc
= &itx
->itx_lr
; /* common log record */
1075 lr_write_t
*lrw
= (lr_write_t
*)lrc
;
1077 uint64_t txg
= lrc
->lrc_txg
;
1078 uint64_t reclen
= lrc
->lrc_reclen
;
1084 ASSERT(lwb
->lwb_buf
!= NULL
);
1085 ASSERT(zilog_is_dirty(zilog
) ||
1086 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1088 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
)
1089 dlen
= P2ROUNDUP_TYPED(
1090 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1092 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1094 zil_lwb_write_init(zilog
, lwb
);
1097 * If this record won't fit in the current log block, start a new one.
1099 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1100 lwb
= zil_lwb_write_start(zilog
, lwb
);
1103 zil_lwb_write_init(zilog
, lwb
);
1104 ASSERT(LWB_EMPTY(lwb
));
1105 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1106 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1111 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1112 bcopy(lrc
, lr_buf
, reclen
);
1113 lrc
= (lr_t
*)lr_buf
;
1114 lrw
= (lr_write_t
*)lrc
;
1116 ZIL_STAT_BUMP(zil_itx_count
);
1119 * If it's a write, fetch the data or get its blkptr as appropriate.
1121 if (lrc
->lrc_txtype
== TX_WRITE
) {
1122 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1123 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1124 if (itx
->itx_wr_state
== WR_COPIED
) {
1125 ZIL_STAT_BUMP(zil_itx_copied_count
);
1126 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1132 ASSERT(itx
->itx_wr_state
== WR_NEED_COPY
);
1133 dbuf
= lr_buf
+ reclen
;
1134 lrw
->lr_common
.lrc_reclen
+= dlen
;
1135 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1136 ZIL_STAT_INCR(zil_itx_needcopy_bytes
,
1139 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1141 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1142 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1145 error
= zilog
->zl_get_data(
1146 itx
->itx_private
, lrw
, dbuf
, lwb
->lwb_zio
);
1148 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1152 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1160 * We're actually making an entry, so update lrc_seq to be the
1161 * log record sequence number. Note that this is generally not
1162 * equal to the itx sequence number because not all transactions
1163 * are synchronous, and sometimes spa_sync() gets there first.
1165 lrc
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1166 lwb
->lwb_nused
+= reclen
+ dlen
;
1167 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1168 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1169 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1175 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1179 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1181 itx
= kmem_alloc(offsetof(itx_t
, itx_lr
) + lrsize
,
1182 KM_PUSHPAGE
| KM_NODEBUG
);
1183 itx
->itx_lr
.lrc_txtype
= txtype
;
1184 itx
->itx_lr
.lrc_reclen
= lrsize
;
1185 itx
->itx_sod
= lrsize
; /* if write & WR_NEED_COPY will be increased */
1186 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1187 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1188 itx
->itx_callback
= NULL
;
1189 itx
->itx_callback_data
= NULL
;
1195 zil_itx_destroy(itx_t
*itx
)
1197 kmem_free(itx
, offsetof(itx_t
, itx_lr
) + itx
->itx_lr
.lrc_reclen
);
1201 * Free up the sync and async itxs. The itxs_t has already been detached
1202 * so no locks are needed.
1205 zil_itxg_clean(itxs_t
*itxs
)
1211 itx_async_node_t
*ian
;
1213 list
= &itxs
->i_sync_list
;
1214 while ((itx
= list_head(list
)) != NULL
) {
1215 if (itx
->itx_callback
!= NULL
)
1216 itx
->itx_callback(itx
->itx_callback_data
);
1217 list_remove(list
, itx
);
1218 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1219 itx
->itx_lr
.lrc_reclen
);
1223 t
= &itxs
->i_async_tree
;
1224 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1225 list
= &ian
->ia_list
;
1226 while ((itx
= list_head(list
)) != NULL
) {
1227 if (itx
->itx_callback
!= NULL
)
1228 itx
->itx_callback(itx
->itx_callback_data
);
1229 list_remove(list
, itx
);
1230 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1231 itx
->itx_lr
.lrc_reclen
);
1234 kmem_free(ian
, sizeof (itx_async_node_t
));
1238 kmem_free(itxs
, sizeof (itxs_t
));
1242 zil_aitx_compare(const void *x1
, const void *x2
)
1244 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1245 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1256 * Remove all async itx with the given oid.
1259 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1262 itx_async_node_t
*ian
;
1269 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1271 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1274 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1276 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1277 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1279 mutex_enter(&itxg
->itxg_lock
);
1280 if (itxg
->itxg_txg
!= txg
) {
1281 mutex_exit(&itxg
->itxg_lock
);
1286 * Locate the object node and append its list.
1288 t
= &itxg
->itxg_itxs
->i_async_tree
;
1289 ian
= avl_find(t
, &oid
, &where
);
1291 list_move_tail(&clean_list
, &ian
->ia_list
);
1292 mutex_exit(&itxg
->itxg_lock
);
1294 while ((itx
= list_head(&clean_list
)) != NULL
) {
1295 if (itx
->itx_callback
!= NULL
)
1296 itx
->itx_callback(itx
->itx_callback_data
);
1297 list_remove(&clean_list
, itx
);
1298 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1299 itx
->itx_lr
.lrc_reclen
);
1301 list_destroy(&clean_list
);
1305 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1309 itxs_t
*itxs
, *clean
= NULL
;
1312 * Object ids can be re-instantiated in the next txg so
1313 * remove any async transactions to avoid future leaks.
1314 * This can happen if a fsync occurs on the re-instantiated
1315 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1316 * the new file data and flushes a write record for the old object.
1318 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1319 zil_remove_async(zilog
, itx
->itx_oid
);
1322 * Ensure the data of a renamed file is committed before the rename.
1324 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1325 zil_async_to_sync(zilog
, itx
->itx_oid
);
1327 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1330 txg
= dmu_tx_get_txg(tx
);
1332 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1333 mutex_enter(&itxg
->itxg_lock
);
1334 itxs
= itxg
->itxg_itxs
;
1335 if (itxg
->itxg_txg
!= txg
) {
1338 * The zil_clean callback hasn't got around to cleaning
1339 * this itxg. Save the itxs for release below.
1340 * This should be rare.
1342 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1344 clean
= itxg
->itxg_itxs
;
1346 ASSERT(itxg
->itxg_sod
== 0);
1347 itxg
->itxg_txg
= txg
;
1348 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1351 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1352 offsetof(itx_t
, itx_node
));
1353 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1354 sizeof (itx_async_node_t
),
1355 offsetof(itx_async_node_t
, ia_node
));
1357 if (itx
->itx_sync
) {
1358 list_insert_tail(&itxs
->i_sync_list
, itx
);
1359 atomic_add_64(&zilog
->zl_itx_list_sz
, itx
->itx_sod
);
1360 itxg
->itxg_sod
+= itx
->itx_sod
;
1362 avl_tree_t
*t
= &itxs
->i_async_tree
;
1363 uint64_t foid
= ((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
;
1364 itx_async_node_t
*ian
;
1367 ian
= avl_find(t
, &foid
, &where
);
1369 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1371 list_create(&ian
->ia_list
, sizeof (itx_t
),
1372 offsetof(itx_t
, itx_node
));
1373 ian
->ia_foid
= foid
;
1374 avl_insert(t
, ian
, where
);
1376 list_insert_tail(&ian
->ia_list
, itx
);
1379 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1380 zilog_dirty(zilog
, txg
);
1381 mutex_exit(&itxg
->itxg_lock
);
1383 /* Release the old itxs now we've dropped the lock */
1385 zil_itxg_clean(clean
);
1389 * If there are any in-memory intent log transactions which have now been
1390 * synced then start up a taskq to free them. We should only do this after we
1391 * have written out the uberblocks (i.e. txg has been comitted) so that
1392 * don't inadvertently clean out in-memory log records that would be required
1396 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1398 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1401 mutex_enter(&itxg
->itxg_lock
);
1402 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1403 mutex_exit(&itxg
->itxg_lock
);
1406 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1407 ASSERT(itxg
->itxg_txg
!= 0);
1408 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1409 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1411 clean_me
= itxg
->itxg_itxs
;
1412 itxg
->itxg_itxs
= NULL
;
1414 mutex_exit(&itxg
->itxg_lock
);
1416 * Preferably start a task queue to free up the old itxs but
1417 * if taskq_dispatch can't allocate resources to do that then
1418 * free it in-line. This should be rare. Note, using TQ_SLEEP
1419 * created a bad performance problem.
1421 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1422 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1423 zil_itxg_clean(clean_me
);
1427 * Get the list of itxs to commit into zl_itx_commit_list.
1430 zil_get_commit_list(zilog_t
*zilog
)
1433 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1434 uint64_t push_sod
= 0;
1436 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1439 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1441 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1442 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1444 mutex_enter(&itxg
->itxg_lock
);
1445 if (itxg
->itxg_txg
!= txg
) {
1446 mutex_exit(&itxg
->itxg_lock
);
1450 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1451 push_sod
+= itxg
->itxg_sod
;
1454 mutex_exit(&itxg
->itxg_lock
);
1456 atomic_add_64(&zilog
->zl_itx_list_sz
, -push_sod
);
1460 * Move the async itxs for a specified object to commit into sync lists.
1463 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1466 itx_async_node_t
*ian
;
1470 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1473 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1475 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1476 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1478 mutex_enter(&itxg
->itxg_lock
);
1479 if (itxg
->itxg_txg
!= txg
) {
1480 mutex_exit(&itxg
->itxg_lock
);
1485 * If a foid is specified then find that node and append its
1486 * list. Otherwise walk the tree appending all the lists
1487 * to the sync list. We add to the end rather than the
1488 * beginning to ensure the create has happened.
1490 t
= &itxg
->itxg_itxs
->i_async_tree
;
1492 ian
= avl_find(t
, &foid
, &where
);
1494 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1498 void *cookie
= NULL
;
1500 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1501 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1503 list_destroy(&ian
->ia_list
);
1504 kmem_free(ian
, sizeof (itx_async_node_t
));
1507 mutex_exit(&itxg
->itxg_lock
);
1512 zil_commit_writer(zilog_t
*zilog
)
1517 spa_t
*spa
= zilog
->zl_spa
;
1520 ASSERT(zilog
->zl_root_zio
== NULL
);
1522 mutex_exit(&zilog
->zl_lock
);
1524 zil_get_commit_list(zilog
);
1527 * Return if there's nothing to commit before we dirty the fs by
1528 * calling zil_create().
1530 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1531 mutex_enter(&zilog
->zl_lock
);
1535 if (zilog
->zl_suspend
) {
1538 lwb
= list_tail(&zilog
->zl_lwb_list
);
1540 lwb
= zil_create(zilog
);
1543 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1544 for (itx
= list_head(&zilog
->zl_itx_commit_list
); itx
!= NULL
;
1545 itx
= list_next(&zilog
->zl_itx_commit_list
, itx
)) {
1546 txg
= itx
->itx_lr
.lrc_txg
;
1549 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1550 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1552 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1554 /* write the last block out */
1555 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1556 lwb
= zil_lwb_write_start(zilog
, lwb
);
1558 zilog
->zl_cur_used
= 0;
1561 * Wait if necessary for the log blocks to be on stable storage.
1563 if (zilog
->zl_root_zio
) {
1564 error
= zio_wait(zilog
->zl_root_zio
);
1565 zilog
->zl_root_zio
= NULL
;
1566 zil_flush_vdevs(zilog
);
1569 if (error
|| lwb
== NULL
)
1570 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1572 while ((itx
= list_head(&zilog
->zl_itx_commit_list
))) {
1573 txg
= itx
->itx_lr
.lrc_txg
;
1576 if (itx
->itx_callback
!= NULL
)
1577 itx
->itx_callback(itx
->itx_callback_data
);
1578 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1579 kmem_free(itx
, offsetof(itx_t
, itx_lr
)
1580 + itx
->itx_lr
.lrc_reclen
);
1583 mutex_enter(&zilog
->zl_lock
);
1586 * Remember the highest committed log sequence number for ztest.
1587 * We only update this value when all the log writes succeeded,
1588 * because ztest wants to ASSERT that it got the whole log chain.
1590 if (error
== 0 && lwb
!= NULL
)
1591 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1595 * Commit zfs transactions to stable storage.
1596 * If foid is 0 push out all transactions, otherwise push only those
1597 * for that object or might reference that object.
1599 * itxs are committed in batches. In a heavily stressed zil there will be
1600 * a commit writer thread who is writing out a bunch of itxs to the log
1601 * for a set of committing threads (cthreads) in the same batch as the writer.
1602 * Those cthreads are all waiting on the same cv for that batch.
1604 * There will also be a different and growing batch of threads that are
1605 * waiting to commit (qthreads). When the committing batch completes
1606 * a transition occurs such that the cthreads exit and the qthreads become
1607 * cthreads. One of the new cthreads becomes the writer thread for the
1608 * batch. Any new threads arriving become new qthreads.
1610 * Only 2 condition variables are needed and there's no transition
1611 * between the two cvs needed. They just flip-flop between qthreads
1614 * Using this scheme we can efficiently wakeup up only those threads
1615 * that have been committed.
1618 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1622 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1625 ZIL_STAT_BUMP(zil_commit_count
);
1627 /* move the async itxs for the foid to the sync queues */
1628 zil_async_to_sync(zilog
, foid
);
1630 mutex_enter(&zilog
->zl_lock
);
1631 mybatch
= zilog
->zl_next_batch
;
1632 while (zilog
->zl_writer
) {
1633 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1634 if (mybatch
<= zilog
->zl_com_batch
) {
1635 mutex_exit(&zilog
->zl_lock
);
1640 zilog
->zl_next_batch
++;
1641 zilog
->zl_writer
= B_TRUE
;
1642 ZIL_STAT_BUMP(zil_commit_writer_count
);
1643 zil_commit_writer(zilog
);
1644 zilog
->zl_com_batch
= mybatch
;
1645 zilog
->zl_writer
= B_FALSE
;
1647 /* wake up one thread to become the next writer */
1648 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1650 /* wake up all threads waiting for this batch to be committed */
1651 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1653 mutex_exit(&zilog
->zl_lock
);
1657 * Called in syncing context to free committed log blocks and update log header.
1660 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1662 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1663 uint64_t txg
= dmu_tx_get_txg(tx
);
1664 spa_t
*spa
= zilog
->zl_spa
;
1665 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1669 * We don't zero out zl_destroy_txg, so make sure we don't try
1670 * to destroy it twice.
1672 if (spa_sync_pass(spa
) != 1)
1675 mutex_enter(&zilog
->zl_lock
);
1677 ASSERT(zilog
->zl_stop_sync
== 0);
1679 if (*replayed_seq
!= 0) {
1680 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1681 zh
->zh_replay_seq
= *replayed_seq
;
1685 if (zilog
->zl_destroy_txg
== txg
) {
1686 blkptr_t blk
= zh
->zh_log
;
1688 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1690 bzero(zh
, sizeof (zil_header_t
));
1691 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1693 if (zilog
->zl_keep_first
) {
1695 * If this block was part of log chain that couldn't
1696 * be claimed because a device was missing during
1697 * zil_claim(), but that device later returns,
1698 * then this block could erroneously appear valid.
1699 * To guard against this, assign a new GUID to the new
1700 * log chain so it doesn't matter what blk points to.
1702 zil_init_log_chain(zilog
, &blk
);
1707 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1708 zh
->zh_log
= lwb
->lwb_blk
;
1709 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1712 ASSERT(lwb
->lwb_zio
== NULL
);
1714 list_remove(&zilog
->zl_lwb_list
, lwb
);
1715 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1716 kmem_cache_free(zil_lwb_cache
, lwb
);
1719 * If we don't have anything left in the lwb list then
1720 * we've had an allocation failure and we need to zero
1721 * out the zil_header blkptr so that we don't end
1722 * up freeing the same block twice.
1724 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1725 BP_ZERO(&zh
->zh_log
);
1729 * Remove fastwrite on any blocks that have been pre-allocated for
1730 * the next commit. This prevents fastwrite counter pollution by
1731 * unused, long-lived LWBs.
1733 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
1734 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_zio
) {
1735 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1736 lwb
->lwb_fastwrite
= 0;
1740 mutex_exit(&zilog
->zl_lock
);
1746 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1747 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1749 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
1750 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
1751 KSTAT_FLAG_VIRTUAL
);
1753 if (zil_ksp
!= NULL
) {
1754 zil_ksp
->ks_data
= &zil_stats
;
1755 kstat_install(zil_ksp
);
1762 kmem_cache_destroy(zil_lwb_cache
);
1764 if (zil_ksp
!= NULL
) {
1765 kstat_delete(zil_ksp
);
1771 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1773 zilog
->zl_sync
= sync
;
1777 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1779 zilog
->zl_logbias
= logbias
;
1783 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1788 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_PUSHPAGE
);
1790 zilog
->zl_header
= zh_phys
;
1792 zilog
->zl_spa
= dmu_objset_spa(os
);
1793 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1794 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1795 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1796 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1797 zilog
->zl_next_batch
= 1;
1799 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1801 for (i
= 0; i
< TXG_SIZE
; i
++) {
1802 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1803 MUTEX_DEFAULT
, NULL
);
1806 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1807 offsetof(lwb_t
, lwb_node
));
1809 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1810 offsetof(itx_t
, itx_node
));
1812 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1814 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1815 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1817 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1818 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1819 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1820 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1826 zil_free(zilog_t
*zilog
)
1830 zilog
->zl_stop_sync
= 1;
1832 ASSERT0(zilog
->zl_suspend
);
1833 ASSERT0(zilog
->zl_suspending
);
1835 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1836 list_destroy(&zilog
->zl_lwb_list
);
1838 avl_destroy(&zilog
->zl_vdev_tree
);
1839 mutex_destroy(&zilog
->zl_vdev_lock
);
1841 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1842 list_destroy(&zilog
->zl_itx_commit_list
);
1844 for (i
= 0; i
< TXG_SIZE
; i
++) {
1846 * It's possible for an itx to be generated that doesn't dirty
1847 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1848 * callback to remove the entry. We remove those here.
1850 * Also free up the ziltest itxs.
1852 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1853 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1854 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1857 mutex_destroy(&zilog
->zl_lock
);
1859 cv_destroy(&zilog
->zl_cv_writer
);
1860 cv_destroy(&zilog
->zl_cv_suspend
);
1861 cv_destroy(&zilog
->zl_cv_batch
[0]);
1862 cv_destroy(&zilog
->zl_cv_batch
[1]);
1864 kmem_free(zilog
, sizeof (zilog_t
));
1868 * Open an intent log.
1871 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1873 zilog_t
*zilog
= dmu_objset_zil(os
);
1875 ASSERT(zilog
->zl_clean_taskq
== NULL
);
1876 ASSERT(zilog
->zl_get_data
== NULL
);
1877 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1879 zilog
->zl_get_data
= get_data
;
1880 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, minclsyspri
,
1881 2, 2, TASKQ_PREPOPULATE
);
1887 * Close an intent log.
1890 zil_close(zilog_t
*zilog
)
1895 zil_commit(zilog
, 0); /* commit all itx */
1898 * The lwb_max_txg for the stubby lwb will reflect the last activity
1899 * for the zil. After a txg_wait_synced() on the txg we know all the
1900 * callbacks have occurred that may clean the zil. Only then can we
1901 * destroy the zl_clean_taskq.
1903 mutex_enter(&zilog
->zl_lock
);
1904 lwb
= list_tail(&zilog
->zl_lwb_list
);
1906 txg
= lwb
->lwb_max_txg
;
1907 mutex_exit(&zilog
->zl_lock
);
1909 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1910 ASSERT(!zilog_is_dirty(zilog
));
1912 taskq_destroy(zilog
->zl_clean_taskq
);
1913 zilog
->zl_clean_taskq
= NULL
;
1914 zilog
->zl_get_data
= NULL
;
1917 * We should have only one LWB left on the list; remove it now.
1919 mutex_enter(&zilog
->zl_lock
);
1920 lwb
= list_head(&zilog
->zl_lwb_list
);
1922 ASSERT(lwb
== list_tail(&zilog
->zl_lwb_list
));
1923 ASSERT(lwb
->lwb_zio
== NULL
);
1924 if (lwb
->lwb_fastwrite
)
1925 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1926 list_remove(&zilog
->zl_lwb_list
, lwb
);
1927 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1928 kmem_cache_free(zil_lwb_cache
, lwb
);
1930 mutex_exit(&zilog
->zl_lock
);
1933 static char *suspend_tag
= "zil suspending";
1936 * Suspend an intent log. While in suspended mode, we still honor
1937 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1938 * On old version pools, we suspend the log briefly when taking a
1939 * snapshot so that it will have an empty intent log.
1941 * Long holds are not really intended to be used the way we do here --
1942 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1943 * could fail. Therefore we take pains to only put a long hold if it is
1944 * actually necessary. Fortunately, it will only be necessary if the
1945 * objset is currently mounted (or the ZVOL equivalent). In that case it
1946 * will already have a long hold, so we are not really making things any worse.
1948 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1949 * zvol_state_t), and use their mechanism to prevent their hold from being
1950 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1953 * if cookiep == NULL, this does both the suspend & resume.
1954 * Otherwise, it returns with the dataset "long held", and the cookie
1955 * should be passed into zil_resume().
1958 zil_suspend(const char *osname
, void **cookiep
)
1962 const zil_header_t
*zh
;
1965 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
1968 zilog
= dmu_objset_zil(os
);
1970 mutex_enter(&zilog
->zl_lock
);
1971 zh
= zilog
->zl_header
;
1973 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
1974 mutex_exit(&zilog
->zl_lock
);
1975 dmu_objset_rele(os
, suspend_tag
);
1976 return (SET_ERROR(EBUSY
));
1980 * Don't put a long hold in the cases where we can avoid it. This
1981 * is when there is no cookie so we are doing a suspend & resume
1982 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1983 * for the suspend because it's already suspended, or there's no ZIL.
1985 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
1986 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
1987 mutex_exit(&zilog
->zl_lock
);
1988 dmu_objset_rele(os
, suspend_tag
);
1992 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
1993 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
1995 zilog
->zl_suspend
++;
1997 if (zilog
->zl_suspend
> 1) {
1999 * Someone else is already suspending it.
2000 * Just wait for them to finish.
2003 while (zilog
->zl_suspending
)
2004 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
2005 mutex_exit(&zilog
->zl_lock
);
2007 if (cookiep
== NULL
)
2015 * If there is no pointer to an on-disk block, this ZIL must not
2016 * be active (e.g. filesystem not mounted), so there's nothing
2019 if (BP_IS_HOLE(&zh
->zh_log
)) {
2020 ASSERT(cookiep
!= NULL
); /* fast path already handled */
2023 mutex_exit(&zilog
->zl_lock
);
2027 zilog
->zl_suspending
= B_TRUE
;
2028 mutex_exit(&zilog
->zl_lock
);
2030 zil_commit(zilog
, 0);
2032 zil_destroy(zilog
, B_FALSE
);
2034 mutex_enter(&zilog
->zl_lock
);
2035 zilog
->zl_suspending
= B_FALSE
;
2036 cv_broadcast(&zilog
->zl_cv_suspend
);
2037 mutex_exit(&zilog
->zl_lock
);
2039 if (cookiep
== NULL
)
2047 zil_resume(void *cookie
)
2049 objset_t
*os
= cookie
;
2050 zilog_t
*zilog
= dmu_objset_zil(os
);
2052 mutex_enter(&zilog
->zl_lock
);
2053 ASSERT(zilog
->zl_suspend
!= 0);
2054 zilog
->zl_suspend
--;
2055 mutex_exit(&zilog
->zl_lock
);
2056 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
2057 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
2060 typedef struct zil_replay_arg
{
2061 zil_replay_func_t
*zr_replay
;
2063 boolean_t zr_byteswap
;
2068 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
2070 char name
[MAXNAMELEN
];
2072 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
2074 dmu_objset_name(zilog
->zl_os
, name
);
2076 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
2077 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
2078 (u_longlong_t
)lr
->lrc_seq
,
2079 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
2080 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
2086 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
2088 zil_replay_arg_t
*zr
= zra
;
2089 const zil_header_t
*zh
= zilog
->zl_header
;
2090 uint64_t reclen
= lr
->lrc_reclen
;
2091 uint64_t txtype
= lr
->lrc_txtype
;
2094 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
2096 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
2099 if (lr
->lrc_txg
< claim_txg
) /* already committed */
2102 /* Strip case-insensitive bit, still present in log record */
2105 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
2106 return (zil_replay_error(zilog
, lr
, EINVAL
));
2109 * If this record type can be logged out of order, the object
2110 * (lr_foid) may no longer exist. That's legitimate, not an error.
2112 if (TX_OOO(txtype
)) {
2113 error
= dmu_object_info(zilog
->zl_os
,
2114 ((lr_ooo_t
*)lr
)->lr_foid
, NULL
);
2115 if (error
== ENOENT
|| error
== EEXIST
)
2120 * Make a copy of the data so we can revise and extend it.
2122 bcopy(lr
, zr
->zr_lr
, reclen
);
2125 * If this is a TX_WRITE with a blkptr, suck in the data.
2127 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
2128 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
2129 zr
->zr_lr
+ reclen
);
2131 return (zil_replay_error(zilog
, lr
, error
));
2135 * The log block containing this lr may have been byteswapped
2136 * so that we can easily examine common fields like lrc_txtype.
2137 * However, the log is a mix of different record types, and only the
2138 * replay vectors know how to byteswap their records. Therefore, if
2139 * the lr was byteswapped, undo it before invoking the replay vector.
2141 if (zr
->zr_byteswap
)
2142 byteswap_uint64_array(zr
->zr_lr
, reclen
);
2145 * We must now do two things atomically: replay this log record,
2146 * and update the log header sequence number to reflect the fact that
2147 * we did so. At the end of each replay function the sequence number
2148 * is updated if we are in replay mode.
2150 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
2153 * The DMU's dnode layer doesn't see removes until the txg
2154 * commits, so a subsequent claim can spuriously fail with
2155 * EEXIST. So if we receive any error we try syncing out
2156 * any removes then retry the transaction. Note that we
2157 * specify B_FALSE for byteswap now, so we don't do it twice.
2159 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
2160 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
2162 return (zil_replay_error(zilog
, lr
, error
));
2169 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
2171 zilog
->zl_replay_blks
++;
2177 * If this dataset has a non-empty intent log, replay it and destroy it.
2180 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t replay_func
[TX_MAX_TYPE
])
2182 zilog_t
*zilog
= dmu_objset_zil(os
);
2183 const zil_header_t
*zh
= zilog
->zl_header
;
2184 zil_replay_arg_t zr
;
2186 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
2187 zil_destroy(zilog
, B_TRUE
);
2191 zr
.zr_replay
= replay_func
;
2193 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
2194 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_PUSHPAGE
);
2197 * Wait for in-progress removes to sync before starting replay.
2199 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2201 zilog
->zl_replay
= B_TRUE
;
2202 zilog
->zl_replay_time
= ddi_get_lbolt();
2203 ASSERT(zilog
->zl_replay_blks
== 0);
2204 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
2206 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
2208 zil_destroy(zilog
, B_FALSE
);
2209 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
2210 zilog
->zl_replay
= B_FALSE
;
2214 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
2216 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2219 if (zilog
->zl_replay
) {
2220 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
2221 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
2222 zilog
->zl_replaying_seq
;
2231 zil_vdev_offline(const char *osname
, void *arg
)
2235 error
= zil_suspend(osname
, NULL
);
2237 return (SET_ERROR(EEXIST
));
2241 #if defined(_KERNEL) && defined(HAVE_SPL)
2242 module_param(zil_replay_disable
, int, 0644);
2243 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
2245 module_param(zfs_nocacheflush
, int, 0644);
2246 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
2248 module_param(zil_slog_limit
, ulong
, 0644);
2249 MODULE_PARM_DESC(zil_slog_limit
, "Max commit bytes to separate log device");