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 https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 * Copyright (c) 2018 Datto Inc.
28 /* Portions Copyright 2010 Robert Milkowski */
30 #include <sys/zfs_context.h>
32 #include <sys/spa_impl.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/metaslab.h>
44 #include <sys/trace_zfs.h>
46 #include <sys/wmsum.h>
49 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
50 * calls that change the file system. Each itx has enough information to
51 * be able to replay them after a system crash, power loss, or
52 * equivalent failure mode. These are stored in memory until either:
54 * 1. they are committed to the pool by the DMU transaction group
55 * (txg), at which point they can be discarded; or
56 * 2. they are committed to the on-disk ZIL for the dataset being
57 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
60 * In the event of a crash or power loss, the itxs contained by each
61 * dataset's on-disk ZIL will be replayed when that dataset is first
62 * instantiated (e.g. if the dataset is a normal filesystem, when it is
65 * As hinted at above, there is one ZIL per dataset (both the in-memory
66 * representation, and the on-disk representation). The on-disk format
67 * consists of 3 parts:
69 * - a single, per-dataset, ZIL header; which points to a chain of
70 * - zero or more ZIL blocks; each of which contains
71 * - zero or more ZIL records
73 * A ZIL record holds the information necessary to replay a single
74 * system call transaction. A ZIL block can hold many ZIL records, and
75 * the blocks are chained together, similarly to a singly linked list.
77 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
78 * block in the chain, and the ZIL header points to the first block in
81 * Note, there is not a fixed place in the pool to hold these ZIL
82 * blocks; they are dynamically allocated and freed as needed from the
83 * blocks available on the pool, though they can be preferentially
84 * allocated from a dedicated "log" vdev.
88 * This controls the amount of time that a ZIL block (lwb) will remain
89 * "open" when it isn't "full", and it has a thread waiting for it to be
90 * committed to stable storage. Please refer to the zil_commit_waiter()
91 * function (and the comments within it) for more details.
93 static uint_t zfs_commit_timeout_pct
= 5;
96 * See zil.h for more information about these fields.
98 static zil_kstat_values_t zil_stats
= {
99 { "zil_commit_count", KSTAT_DATA_UINT64
},
100 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
101 { "zil_itx_count", KSTAT_DATA_UINT64
},
102 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
103 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
104 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
105 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
106 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
107 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
108 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
109 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
110 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
111 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
114 static zil_sums_t zil_sums_global
;
115 static kstat_t
*zil_kstats_global
;
118 * Disable intent logging replay. This global ZIL switch affects all pools.
120 int zil_replay_disable
= 0;
123 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
124 * the disk(s) by the ZIL after an LWB write has completed. Setting this
125 * will cause ZIL corruption on power loss if a volatile out-of-order
126 * write cache is enabled.
128 static int zil_nocacheflush
= 0;
131 * Limit SLOG write size per commit executed with synchronous priority.
132 * Any writes above that will be executed with lower (asynchronous) priority
133 * to limit potential SLOG device abuse by single active ZIL writer.
135 static uint64_t zil_slog_bulk
= 768 * 1024;
137 static kmem_cache_t
*zil_lwb_cache
;
138 static kmem_cache_t
*zil_zcw_cache
;
140 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
141 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
144 zil_bp_compare(const void *x1
, const void *x2
)
146 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
147 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
149 int cmp
= TREE_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
153 return (TREE_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
157 zil_bp_tree_init(zilog_t
*zilog
)
159 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
160 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
164 zil_bp_tree_fini(zilog_t
*zilog
)
166 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
170 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
171 kmem_free(zn
, sizeof (zil_bp_node_t
));
177 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
179 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
184 if (BP_IS_EMBEDDED(bp
))
187 dva
= BP_IDENTITY(bp
);
189 if (avl_find(t
, dva
, &where
) != NULL
)
190 return (SET_ERROR(EEXIST
));
192 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
194 avl_insert(t
, zn
, where
);
199 static zil_header_t
*
200 zil_header_in_syncing_context(zilog_t
*zilog
)
202 return ((zil_header_t
*)zilog
->zl_header
);
206 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
208 zio_cksum_t
*zc
= &bp
->blk_cksum
;
210 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_0
],
211 sizeof (zc
->zc_word
[ZIL_ZC_GUID_0
]));
212 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_1
],
213 sizeof (zc
->zc_word
[ZIL_ZC_GUID_1
]));
214 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
215 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
219 zil_kstats_global_update(kstat_t
*ksp
, int rw
)
221 zil_kstat_values_t
*zs
= ksp
->ks_data
;
222 ASSERT3P(&zil_stats
, ==, zs
);
224 if (rw
== KSTAT_WRITE
) {
225 return (SET_ERROR(EACCES
));
228 zil_kstat_values_update(zs
, &zil_sums_global
);
234 * Read a log block and make sure it's valid.
237 zil_read_log_block(zilog_t
*zilog
, boolean_t decrypt
, const blkptr_t
*bp
,
238 blkptr_t
*nbp
, void *dst
, char **end
)
240 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
241 arc_flags_t aflags
= ARC_FLAG_WAIT
;
242 arc_buf_t
*abuf
= NULL
;
246 if (zilog
->zl_header
->zh_claim_txg
== 0)
247 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
249 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
250 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
253 zio_flags
|= ZIO_FLAG_RAW
;
255 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
256 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
258 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
,
259 &abuf
, ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
262 zio_cksum_t cksum
= bp
->blk_cksum
;
265 * Validate the checksummed log block.
267 * Sequence numbers should be... sequential. The checksum
268 * verifier for the next block should be bp's checksum plus 1.
270 * Also check the log chain linkage and size used.
272 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
274 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
275 zil_chain_t
*zilc
= abuf
->b_data
;
276 char *lr
= (char *)(zilc
+ 1);
277 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
279 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
280 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
281 error
= SET_ERROR(ECKSUM
);
283 ASSERT3U(len
, <=, SPA_OLD_MAXBLOCKSIZE
);
284 memcpy(dst
, lr
, len
);
285 *end
= (char *)dst
+ len
;
286 *nbp
= zilc
->zc_next_blk
;
289 char *lr
= abuf
->b_data
;
290 uint64_t size
= BP_GET_LSIZE(bp
);
291 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
293 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
294 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
295 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
296 error
= SET_ERROR(ECKSUM
);
298 ASSERT3U(zilc
->zc_nused
, <=,
299 SPA_OLD_MAXBLOCKSIZE
);
300 memcpy(dst
, lr
, zilc
->zc_nused
);
301 *end
= (char *)dst
+ zilc
->zc_nused
;
302 *nbp
= zilc
->zc_next_blk
;
306 arc_buf_destroy(abuf
, &abuf
);
313 * Read a TX_WRITE log data block.
316 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
318 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
319 const blkptr_t
*bp
= &lr
->lr_blkptr
;
320 arc_flags_t aflags
= ARC_FLAG_WAIT
;
321 arc_buf_t
*abuf
= NULL
;
325 if (BP_IS_HOLE(bp
)) {
327 memset(wbuf
, 0, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
331 if (zilog
->zl_header
->zh_claim_txg
== 0)
332 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
335 * If we are not using the resulting data, we are just checking that
336 * it hasn't been corrupted so we don't need to waste CPU time
337 * decompressing and decrypting it.
340 zio_flags
|= ZIO_FLAG_RAW
;
342 ASSERT3U(BP_GET_LSIZE(bp
), !=, 0);
343 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
344 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
346 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
347 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
351 memcpy(wbuf
, abuf
->b_data
, arc_buf_size(abuf
));
352 arc_buf_destroy(abuf
, &abuf
);
359 zil_sums_init(zil_sums_t
*zs
)
361 wmsum_init(&zs
->zil_commit_count
, 0);
362 wmsum_init(&zs
->zil_commit_writer_count
, 0);
363 wmsum_init(&zs
->zil_itx_count
, 0);
364 wmsum_init(&zs
->zil_itx_indirect_count
, 0);
365 wmsum_init(&zs
->zil_itx_indirect_bytes
, 0);
366 wmsum_init(&zs
->zil_itx_copied_count
, 0);
367 wmsum_init(&zs
->zil_itx_copied_bytes
, 0);
368 wmsum_init(&zs
->zil_itx_needcopy_count
, 0);
369 wmsum_init(&zs
->zil_itx_needcopy_bytes
, 0);
370 wmsum_init(&zs
->zil_itx_metaslab_normal_count
, 0);
371 wmsum_init(&zs
->zil_itx_metaslab_normal_bytes
, 0);
372 wmsum_init(&zs
->zil_itx_metaslab_slog_count
, 0);
373 wmsum_init(&zs
->zil_itx_metaslab_slog_bytes
, 0);
377 zil_sums_fini(zil_sums_t
*zs
)
379 wmsum_fini(&zs
->zil_commit_count
);
380 wmsum_fini(&zs
->zil_commit_writer_count
);
381 wmsum_fini(&zs
->zil_itx_count
);
382 wmsum_fini(&zs
->zil_itx_indirect_count
);
383 wmsum_fini(&zs
->zil_itx_indirect_bytes
);
384 wmsum_fini(&zs
->zil_itx_copied_count
);
385 wmsum_fini(&zs
->zil_itx_copied_bytes
);
386 wmsum_fini(&zs
->zil_itx_needcopy_count
);
387 wmsum_fini(&zs
->zil_itx_needcopy_bytes
);
388 wmsum_fini(&zs
->zil_itx_metaslab_normal_count
);
389 wmsum_fini(&zs
->zil_itx_metaslab_normal_bytes
);
390 wmsum_fini(&zs
->zil_itx_metaslab_slog_count
);
391 wmsum_fini(&zs
->zil_itx_metaslab_slog_bytes
);
395 zil_kstat_values_update(zil_kstat_values_t
*zs
, zil_sums_t
*zil_sums
)
397 zs
->zil_commit_count
.value
.ui64
=
398 wmsum_value(&zil_sums
->zil_commit_count
);
399 zs
->zil_commit_writer_count
.value
.ui64
=
400 wmsum_value(&zil_sums
->zil_commit_writer_count
);
401 zs
->zil_itx_count
.value
.ui64
=
402 wmsum_value(&zil_sums
->zil_itx_count
);
403 zs
->zil_itx_indirect_count
.value
.ui64
=
404 wmsum_value(&zil_sums
->zil_itx_indirect_count
);
405 zs
->zil_itx_indirect_bytes
.value
.ui64
=
406 wmsum_value(&zil_sums
->zil_itx_indirect_bytes
);
407 zs
->zil_itx_copied_count
.value
.ui64
=
408 wmsum_value(&zil_sums
->zil_itx_copied_count
);
409 zs
->zil_itx_copied_bytes
.value
.ui64
=
410 wmsum_value(&zil_sums
->zil_itx_copied_bytes
);
411 zs
->zil_itx_needcopy_count
.value
.ui64
=
412 wmsum_value(&zil_sums
->zil_itx_needcopy_count
);
413 zs
->zil_itx_needcopy_bytes
.value
.ui64
=
414 wmsum_value(&zil_sums
->zil_itx_needcopy_bytes
);
415 zs
->zil_itx_metaslab_normal_count
.value
.ui64
=
416 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_count
);
417 zs
->zil_itx_metaslab_normal_bytes
.value
.ui64
=
418 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_bytes
);
419 zs
->zil_itx_metaslab_slog_count
.value
.ui64
=
420 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_count
);
421 zs
->zil_itx_metaslab_slog_bytes
.value
.ui64
=
422 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_bytes
);
426 * Parse the intent log, and call parse_func for each valid record within.
429 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
430 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
,
433 const zil_header_t
*zh
= zilog
->zl_header
;
434 boolean_t claimed
= !!zh
->zh_claim_txg
;
435 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
436 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
437 uint64_t max_blk_seq
= 0;
438 uint64_t max_lr_seq
= 0;
439 uint64_t blk_count
= 0;
440 uint64_t lr_count
= 0;
441 blkptr_t blk
, next_blk
= {{{{0}}}};
446 * Old logs didn't record the maximum zh_claim_lr_seq.
448 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
449 claim_lr_seq
= UINT64_MAX
;
452 * Starting at the block pointed to by zh_log we read the log chain.
453 * For each block in the chain we strongly check that block to
454 * ensure its validity. We stop when an invalid block is found.
455 * For each block pointer in the chain we call parse_blk_func().
456 * For each record in each valid block we call parse_lr_func().
457 * If the log has been claimed, stop if we encounter a sequence
458 * number greater than the highest claimed sequence number.
460 lrbuf
= zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE
);
461 zil_bp_tree_init(zilog
);
463 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
464 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
468 if (blk_seq
> claim_blk_seq
)
471 error
= parse_blk_func(zilog
, &blk
, arg
, txg
);
474 ASSERT3U(max_blk_seq
, <, blk_seq
);
475 max_blk_seq
= blk_seq
;
478 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
481 error
= zil_read_log_block(zilog
, decrypt
, &blk
, &next_blk
,
486 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
487 lr_t
*lr
= (lr_t
*)lrp
;
488 reclen
= lr
->lrc_reclen
;
489 ASSERT3U(reclen
, >=, sizeof (lr_t
));
490 if (lr
->lrc_seq
> claim_lr_seq
)
493 error
= parse_lr_func(zilog
, lr
, arg
, txg
);
496 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
497 max_lr_seq
= lr
->lrc_seq
;
502 zilog
->zl_parse_error
= error
;
503 zilog
->zl_parse_blk_seq
= max_blk_seq
;
504 zilog
->zl_parse_lr_seq
= max_lr_seq
;
505 zilog
->zl_parse_blk_count
= blk_count
;
506 zilog
->zl_parse_lr_count
= lr_count
;
508 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
509 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
) ||
510 (decrypt
&& error
== EIO
));
512 zil_bp_tree_fini(zilog
);
513 zio_buf_free(lrbuf
, SPA_OLD_MAXBLOCKSIZE
);
519 zil_clear_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
523 ASSERT(!BP_IS_HOLE(bp
));
526 * As we call this function from the context of a rewind to a
527 * checkpoint, each ZIL block whose txg is later than the txg
528 * that we rewind to is invalid. Thus, we return -1 so
529 * zil_parse() doesn't attempt to read it.
531 if (bp
->blk_birth
>= first_txg
)
534 if (zil_bp_tree_add(zilog
, bp
) != 0)
537 zio_free(zilog
->zl_spa
, first_txg
, bp
);
542 zil_noop_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
545 (void) zilog
, (void) lrc
, (void) tx
, (void) first_txg
;
550 zil_claim_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
554 * Claim log block if not already committed and not already claimed.
555 * If tx == NULL, just verify that the block is claimable.
557 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
558 zil_bp_tree_add(zilog
, bp
) != 0)
561 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
562 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
563 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
567 zil_claim_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
570 lr_write_t
*lr
= (lr_write_t
*)lrc
;
573 if (lrc
->lrc_txtype
!= TX_WRITE
)
577 * If the block is not readable, don't claim it. This can happen
578 * in normal operation when a log block is written to disk before
579 * some of the dmu_sync() blocks it points to. In this case, the
580 * transaction cannot have been committed to anyone (we would have
581 * waited for all writes to be stable first), so it is semantically
582 * correct to declare this the end of the log.
584 if (lr
->lr_blkptr
.blk_birth
>= first_txg
) {
585 error
= zil_read_log_data(zilog
, lr
, NULL
);
590 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
594 zil_free_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
599 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
605 zil_free_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
608 lr_write_t
*lr
= (lr_write_t
*)lrc
;
609 blkptr_t
*bp
= &lr
->lr_blkptr
;
612 * If we previously claimed it, we need to free it.
614 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
615 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
617 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
623 zil_lwb_vdev_compare(const void *x1
, const void *x2
)
625 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
626 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
628 return (TREE_CMP(v1
, v2
));
632 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, boolean_t slog
, uint64_t txg
,
637 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
638 lwb
->lwb_zilog
= zilog
;
640 lwb
->lwb_fastwrite
= fastwrite
;
641 lwb
->lwb_slog
= slog
;
642 lwb
->lwb_state
= LWB_STATE_CLOSED
;
643 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
644 lwb
->lwb_max_txg
= txg
;
645 lwb
->lwb_write_zio
= NULL
;
646 lwb
->lwb_root_zio
= NULL
;
647 lwb
->lwb_issued_timestamp
= 0;
648 lwb
->lwb_issued_txg
= 0;
649 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
650 lwb
->lwb_nused
= sizeof (zil_chain_t
);
651 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
654 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
657 mutex_enter(&zilog
->zl_lock
);
658 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
659 mutex_exit(&zilog
->zl_lock
);
661 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
662 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
663 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
664 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
670 zil_free_lwb(zilog_t
*zilog
, lwb_t
*lwb
)
672 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
673 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
674 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
675 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
676 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
677 ASSERT3P(lwb
->lwb_write_zio
, ==, NULL
);
678 ASSERT3P(lwb
->lwb_root_zio
, ==, NULL
);
679 ASSERT3U(lwb
->lwb_max_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
680 ASSERT(lwb
->lwb_state
== LWB_STATE_CLOSED
||
681 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
684 * Clear the zilog's field to indicate this lwb is no longer
685 * valid, and prevent use-after-free errors.
687 if (zilog
->zl_last_lwb_opened
== lwb
)
688 zilog
->zl_last_lwb_opened
= NULL
;
690 kmem_cache_free(zil_lwb_cache
, lwb
);
694 * Called when we create in-memory log transactions so that we know
695 * to cleanup the itxs at the end of spa_sync().
698 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
700 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
701 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
703 ASSERT(spa_writeable(zilog
->zl_spa
));
705 if (ds
->ds_is_snapshot
)
706 panic("dirtying snapshot!");
708 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
709 /* up the hold count until we can be written out */
710 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
712 zilog
->zl_dirty_max_txg
= MAX(txg
, zilog
->zl_dirty_max_txg
);
717 * Determine if the zil is dirty in the specified txg. Callers wanting to
718 * ensure that the dirty state does not change must hold the itxg_lock for
719 * the specified txg. Holding the lock will ensure that the zil cannot be
720 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
723 static boolean_t __maybe_unused
724 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
726 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
728 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
734 * Determine if the zil is dirty. The zil is considered dirty if it has
735 * any pending itx records that have not been cleaned by zil_clean().
738 zilog_is_dirty(zilog_t
*zilog
)
740 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
742 for (int t
= 0; t
< TXG_SIZE
; t
++) {
743 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
750 * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
751 * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
752 * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
756 zil_commit_activate_saxattr_feature(zilog_t
*zilog
)
758 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
762 if (spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
763 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
&&
764 !dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
)) {
765 tx
= dmu_tx_create(zilog
->zl_os
);
766 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
767 dsl_dataset_dirty(ds
, tx
);
768 txg
= dmu_tx_get_txg(tx
);
770 mutex_enter(&ds
->ds_lock
);
771 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
773 mutex_exit(&ds
->ds_lock
);
775 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
780 * Create an on-disk intent log.
783 zil_create(zilog_t
*zilog
)
785 const zil_header_t
*zh
= zilog
->zl_header
;
791 boolean_t fastwrite
= FALSE
;
792 boolean_t slog
= FALSE
;
793 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
797 * Wait for any previous destroy to complete.
799 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
801 ASSERT(zh
->zh_claim_txg
== 0);
802 ASSERT(zh
->zh_replay_seq
== 0);
807 * Allocate an initial log block if:
808 * - there isn't one already
809 * - the existing block is the wrong endianness
811 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
812 tx
= dmu_tx_create(zilog
->zl_os
);
813 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
814 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
815 txg
= dmu_tx_get_txg(tx
);
817 if (!BP_IS_HOLE(&blk
)) {
818 zio_free(zilog
->zl_spa
, txg
, &blk
);
822 error
= zio_alloc_zil(zilog
->zl_spa
, zilog
->zl_os
, txg
, &blk
,
823 ZIL_MIN_BLKSZ
, &slog
);
827 zil_init_log_chain(zilog
, &blk
);
831 * Allocate a log write block (lwb) for the first log block.
834 lwb
= zil_alloc_lwb(zilog
, &blk
, slog
, txg
, fastwrite
);
837 * If we just allocated the first log block, commit our transaction
838 * and wait for zil_sync() to stuff the block pointer into zh_log.
839 * (zh is part of the MOS, so we cannot modify it in open context.)
843 * If "zilsaxattr" feature is enabled on zpool, then activate
844 * it now when we're creating the ZIL chain. We can't wait with
845 * this until we write the first xattr log record because we
846 * need to wait for the feature activation to sync out.
848 if (spa_feature_is_enabled(zilog
->zl_spa
,
849 SPA_FEATURE_ZILSAXATTR
) && dmu_objset_type(zilog
->zl_os
) !=
851 mutex_enter(&ds
->ds_lock
);
852 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
854 mutex_exit(&ds
->ds_lock
);
858 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
861 * This branch covers the case where we enable the feature on a
862 * zpool that has existing ZIL headers.
864 zil_commit_activate_saxattr_feature(zilog
);
866 IMPLY(spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
867 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
,
868 dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
));
870 ASSERT(error
!= 0 || memcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
871 IMPLY(error
== 0, lwb
!= NULL
);
877 * In one tx, free all log blocks and clear the log header. If keep_first
878 * is set, then we're replaying a log with no content. We want to keep the
879 * first block, however, so that the first synchronous transaction doesn't
880 * require a txg_wait_synced() in zil_create(). We don't need to
881 * txg_wait_synced() here either when keep_first is set, because both
882 * zil_create() and zil_destroy() will wait for any in-progress destroys
886 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
888 const zil_header_t
*zh
= zilog
->zl_header
;
894 * Wait for any previous destroy to complete.
896 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
898 zilog
->zl_old_header
= *zh
; /* debugging aid */
900 if (BP_IS_HOLE(&zh
->zh_log
))
903 tx
= dmu_tx_create(zilog
->zl_os
);
904 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
905 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
906 txg
= dmu_tx_get_txg(tx
);
908 mutex_enter(&zilog
->zl_lock
);
910 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
911 zilog
->zl_destroy_txg
= txg
;
912 zilog
->zl_keep_first
= keep_first
;
914 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
915 ASSERT(zh
->zh_claim_txg
== 0);
917 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
918 if (lwb
->lwb_fastwrite
)
919 metaslab_fastwrite_unmark(zilog
->zl_spa
,
922 list_remove(&zilog
->zl_lwb_list
, lwb
);
923 if (lwb
->lwb_buf
!= NULL
)
924 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
925 zio_free(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
926 zil_free_lwb(zilog
, lwb
);
928 } else if (!keep_first
) {
929 zil_destroy_sync(zilog
, tx
);
931 mutex_exit(&zilog
->zl_lock
);
937 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
939 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
940 (void) zil_parse(zilog
, zil_free_log_block
,
941 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
, B_FALSE
);
945 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
947 dmu_tx_t
*tx
= txarg
;
954 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
955 DMU_OST_ANY
, B_FALSE
, B_FALSE
, FTAG
, &os
);
958 * EBUSY indicates that the objset is inconsistent, in which
959 * case it can not have a ZIL.
961 if (error
!= EBUSY
) {
962 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
963 (unsigned long long)ds
->ds_object
, error
);
969 zilog
= dmu_objset_zil(os
);
970 zh
= zil_header_in_syncing_context(zilog
);
971 ASSERT3U(tx
->tx_txg
, ==, spa_first_txg(zilog
->zl_spa
));
972 first_txg
= spa_min_claim_txg(zilog
->zl_spa
);
975 * If the spa_log_state is not set to be cleared, check whether
976 * the current uberblock is a checkpoint one and if the current
977 * header has been claimed before moving on.
979 * If the current uberblock is a checkpointed uberblock then
980 * one of the following scenarios took place:
982 * 1] We are currently rewinding to the checkpoint of the pool.
983 * 2] We crashed in the middle of a checkpoint rewind but we
984 * did manage to write the checkpointed uberblock to the
985 * vdev labels, so when we tried to import the pool again
986 * the checkpointed uberblock was selected from the import
989 * In both cases we want to zero out all the ZIL blocks, except
990 * the ones that have been claimed at the time of the checkpoint
991 * (their zh_claim_txg != 0). The reason is that these blocks
992 * may be corrupted since we may have reused their locations on
993 * disk after we took the checkpoint.
995 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
996 * when we first figure out whether the current uberblock is
997 * checkpointed or not. Unfortunately, that would discard all
998 * the logs, including the ones that are claimed, and we would
1001 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
||
1002 (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1003 zh
->zh_claim_txg
== 0)) {
1004 if (!BP_IS_HOLE(&zh
->zh_log
)) {
1005 (void) zil_parse(zilog
, zil_clear_log_block
,
1006 zil_noop_log_record
, tx
, first_txg
, B_FALSE
);
1008 BP_ZERO(&zh
->zh_log
);
1009 if (os
->os_encrypted
)
1010 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1011 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1012 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1017 * If we are not rewinding and opening the pool normally, then
1018 * the min_claim_txg should be equal to the first txg of the pool.
1020 ASSERT3U(first_txg
, ==, spa_first_txg(zilog
->zl_spa
));
1023 * Claim all log blocks if we haven't already done so, and remember
1024 * the highest claimed sequence number. This ensures that if we can
1025 * read only part of the log now (e.g. due to a missing device),
1026 * but we can read the entire log later, we will not try to replay
1027 * or destroy beyond the last block we successfully claimed.
1029 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
1030 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
1031 (void) zil_parse(zilog
, zil_claim_log_block
,
1032 zil_claim_log_record
, tx
, first_txg
, B_FALSE
);
1033 zh
->zh_claim_txg
= first_txg
;
1034 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
1035 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
1036 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
1037 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
1038 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
1039 if (os
->os_encrypted
)
1040 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1041 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1044 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
1045 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1050 * Check the log by walking the log chain.
1051 * Checksum errors are ok as they indicate the end of the chain.
1052 * Any other error (no device or read failure) returns an error.
1055 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
1065 error
= dmu_objset_from_ds(ds
, &os
);
1067 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
1068 (unsigned long long)ds
->ds_object
, error
);
1072 zilog
= dmu_objset_zil(os
);
1073 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
1075 if (!BP_IS_HOLE(bp
)) {
1077 boolean_t valid
= B_TRUE
;
1080 * Check the first block and determine if it's on a log device
1081 * which may have been removed or faulted prior to loading this
1082 * pool. If so, there's no point in checking the rest of the
1083 * log as its content should have already been synced to the
1086 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
1087 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
1088 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
1089 valid
= vdev_log_state_valid(vd
);
1090 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
1096 * Check whether the current uberblock is checkpointed (e.g.
1097 * we are rewinding) and whether the current header has been
1098 * claimed or not. If it hasn't then skip verifying it. We
1099 * do this because its ZIL blocks may be part of the pool's
1100 * state before the rewind, which is no longer valid.
1102 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1103 if (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1104 zh
->zh_claim_txg
== 0)
1109 * Because tx == NULL, zil_claim_log_block() will not actually claim
1110 * any blocks, but just determine whether it is possible to do so.
1111 * In addition to checking the log chain, zil_claim_log_block()
1112 * will invoke zio_claim() with a done func of spa_claim_notify(),
1113 * which will update spa_max_claim_txg. See spa_load() for details.
1115 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
1116 zilog
->zl_header
->zh_claim_txg
? -1ULL :
1117 spa_min_claim_txg(os
->os_spa
), B_FALSE
);
1119 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
1123 * When an itx is "skipped", this function is used to properly mark the
1124 * waiter as "done, and signal any thread(s) waiting on it. An itx can
1125 * be skipped (and not committed to an lwb) for a variety of reasons,
1126 * one of them being that the itx was committed via spa_sync(), prior to
1127 * it being committed to an lwb; this can happen if a thread calling
1128 * zil_commit() is racing with spa_sync().
1131 zil_commit_waiter_skip(zil_commit_waiter_t
*zcw
)
1133 mutex_enter(&zcw
->zcw_lock
);
1134 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1135 zcw
->zcw_done
= B_TRUE
;
1136 cv_broadcast(&zcw
->zcw_cv
);
1137 mutex_exit(&zcw
->zcw_lock
);
1141 * This function is used when the given waiter is to be linked into an
1142 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1143 * At this point, the waiter will no longer be referenced by the itx,
1144 * and instead, will be referenced by the lwb.
1147 zil_commit_waiter_link_lwb(zil_commit_waiter_t
*zcw
, lwb_t
*lwb
)
1150 * The lwb_waiters field of the lwb is protected by the zilog's
1151 * zl_lock, thus it must be held when calling this function.
1153 ASSERT(MUTEX_HELD(&lwb
->lwb_zilog
->zl_lock
));
1155 mutex_enter(&zcw
->zcw_lock
);
1156 ASSERT(!list_link_active(&zcw
->zcw_node
));
1157 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1158 ASSERT3P(lwb
, !=, NULL
);
1159 ASSERT(lwb
->lwb_state
== LWB_STATE_OPENED
||
1160 lwb
->lwb_state
== LWB_STATE_ISSUED
||
1161 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
);
1163 list_insert_tail(&lwb
->lwb_waiters
, zcw
);
1165 mutex_exit(&zcw
->zcw_lock
);
1169 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1170 * block, and the given waiter must be linked to the "nolwb waiters"
1171 * list inside of zil_process_commit_list().
1174 zil_commit_waiter_link_nolwb(zil_commit_waiter_t
*zcw
, list_t
*nolwb
)
1176 mutex_enter(&zcw
->zcw_lock
);
1177 ASSERT(!list_link_active(&zcw
->zcw_node
));
1178 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1179 list_insert_tail(nolwb
, zcw
);
1180 mutex_exit(&zcw
->zcw_lock
);
1184 zil_lwb_add_block(lwb_t
*lwb
, const blkptr_t
*bp
)
1186 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1188 zil_vdev_node_t
*zv
, zvsearch
;
1189 int ndvas
= BP_GET_NDVAS(bp
);
1192 if (zil_nocacheflush
)
1195 mutex_enter(&lwb
->lwb_vdev_lock
);
1196 for (i
= 0; i
< ndvas
; i
++) {
1197 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1198 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
1199 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
1200 zv
->zv_vdev
= zvsearch
.zv_vdev
;
1201 avl_insert(t
, zv
, where
);
1204 mutex_exit(&lwb
->lwb_vdev_lock
);
1208 zil_lwb_flush_defer(lwb_t
*lwb
, lwb_t
*nlwb
)
1210 avl_tree_t
*src
= &lwb
->lwb_vdev_tree
;
1211 avl_tree_t
*dst
= &nlwb
->lwb_vdev_tree
;
1212 void *cookie
= NULL
;
1213 zil_vdev_node_t
*zv
;
1215 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1216 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1217 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1220 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1221 * not need the protection of lwb_vdev_lock (it will only be modified
1222 * while holding zilog->zl_lock) as its writes and those of its
1223 * children have all completed. The younger 'nlwb' may be waiting on
1224 * future writes to additional vdevs.
1226 mutex_enter(&nlwb
->lwb_vdev_lock
);
1228 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1229 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1231 while ((zv
= avl_destroy_nodes(src
, &cookie
)) != NULL
) {
1234 if (avl_find(dst
, zv
, &where
) == NULL
) {
1235 avl_insert(dst
, zv
, where
);
1237 kmem_free(zv
, sizeof (*zv
));
1240 mutex_exit(&nlwb
->lwb_vdev_lock
);
1244 zil_lwb_add_txg(lwb_t
*lwb
, uint64_t txg
)
1246 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1250 * This function is a called after all vdevs associated with a given lwb
1251 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1252 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1253 * all "previous" lwb's will have completed before this function is
1254 * called; i.e. this function is called for all previous lwbs before
1255 * it's called for "this" lwb (enforced via zio the dependencies
1256 * configured in zil_lwb_set_zio_dependency()).
1258 * The intention is for this function to be called as soon as the
1259 * contents of an lwb are considered "stable" on disk, and will survive
1260 * any sudden loss of power. At this point, any threads waiting for the
1261 * lwb to reach this state are signalled, and the "waiter" structures
1262 * are marked "done".
1265 zil_lwb_flush_vdevs_done(zio_t
*zio
)
1267 lwb_t
*lwb
= zio
->io_private
;
1268 zilog_t
*zilog
= lwb
->lwb_zilog
;
1269 zil_commit_waiter_t
*zcw
;
1273 spa_config_exit(zilog
->zl_spa
, SCL_STATE
, lwb
);
1275 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1277 mutex_enter(&zilog
->zl_lock
);
1280 * If we have had an allocation failure and the txg is
1281 * waiting to sync then we want zil_sync() to remove the lwb so
1282 * that it's not picked up as the next new one in
1283 * zil_process_commit_list(). zil_sync() will only remove the
1284 * lwb if lwb_buf is null.
1286 lwb
->lwb_buf
= NULL
;
1288 ASSERT3U(lwb
->lwb_issued_timestamp
, >, 0);
1289 zilog
->zl_last_lwb_latency
= gethrtime() - lwb
->lwb_issued_timestamp
;
1291 lwb
->lwb_root_zio
= NULL
;
1293 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1294 lwb
->lwb_state
= LWB_STATE_FLUSH_DONE
;
1296 if (zilog
->zl_last_lwb_opened
== lwb
) {
1298 * Remember the highest committed log sequence number
1299 * for ztest. We only update this value when all the log
1300 * writes succeeded, because ztest wants to ASSERT that
1301 * it got the whole log chain.
1303 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1306 while ((itx
= list_head(&lwb
->lwb_itxs
)) != NULL
) {
1307 list_remove(&lwb
->lwb_itxs
, itx
);
1308 zil_itx_destroy(itx
);
1311 while ((zcw
= list_head(&lwb
->lwb_waiters
)) != NULL
) {
1312 mutex_enter(&zcw
->zcw_lock
);
1314 ASSERT(list_link_active(&zcw
->zcw_node
));
1315 list_remove(&lwb
->lwb_waiters
, zcw
);
1317 ASSERT3P(zcw
->zcw_lwb
, ==, lwb
);
1318 zcw
->zcw_lwb
= NULL
;
1320 * We expect any ZIO errors from child ZIOs to have been
1321 * propagated "up" to this specific LWB's root ZIO, in
1322 * order for this error handling to work correctly. This
1323 * includes ZIO errors from either this LWB's write or
1324 * flush, as well as any errors from other dependent LWBs
1325 * (e.g. a root LWB ZIO that might be a child of this LWB).
1327 * With that said, it's important to note that LWB flush
1328 * errors are not propagated up to the LWB root ZIO.
1329 * This is incorrect behavior, and results in VDEV flush
1330 * errors not being handled correctly here. See the
1331 * comment above the call to "zio_flush" for details.
1334 zcw
->zcw_zio_error
= zio
->io_error
;
1336 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1337 zcw
->zcw_done
= B_TRUE
;
1338 cv_broadcast(&zcw
->zcw_cv
);
1340 mutex_exit(&zcw
->zcw_lock
);
1343 mutex_exit(&zilog
->zl_lock
);
1345 mutex_enter(&zilog
->zl_lwb_io_lock
);
1346 txg
= lwb
->lwb_issued_txg
;
1347 ASSERT3U(zilog
->zl_lwb_inflight
[txg
& TXG_MASK
], >, 0);
1348 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]--;
1349 if (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] == 0)
1350 cv_broadcast(&zilog
->zl_lwb_io_cv
);
1351 mutex_exit(&zilog
->zl_lwb_io_lock
);
1355 * Wait for the completion of all issued write/flush of that txg provided.
1356 * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1359 zil_lwb_flush_wait_all(zilog_t
*zilog
, uint64_t txg
)
1361 ASSERT3U(txg
, ==, spa_syncing_txg(zilog
->zl_spa
));
1363 mutex_enter(&zilog
->zl_lwb_io_lock
);
1364 while (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] > 0)
1365 cv_wait(&zilog
->zl_lwb_io_cv
, &zilog
->zl_lwb_io_lock
);
1366 mutex_exit(&zilog
->zl_lwb_io_lock
);
1369 mutex_enter(&zilog
->zl_lock
);
1370 mutex_enter(&zilog
->zl_lwb_io_lock
);
1371 lwb_t
*lwb
= list_head(&zilog
->zl_lwb_list
);
1372 while (lwb
!= NULL
&& lwb
->lwb_max_txg
<= txg
) {
1373 if (lwb
->lwb_issued_txg
<= txg
) {
1374 ASSERT(lwb
->lwb_state
!= LWB_STATE_ISSUED
);
1375 ASSERT(lwb
->lwb_state
!= LWB_STATE_WRITE_DONE
);
1376 IMPLY(lwb
->lwb_issued_txg
> 0,
1377 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
1379 IMPLY(lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
,
1380 lwb
->lwb_buf
== NULL
);
1381 lwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1383 mutex_exit(&zilog
->zl_lwb_io_lock
);
1384 mutex_exit(&zilog
->zl_lock
);
1389 * This is called when an lwb's write zio completes. The callback's
1390 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1391 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1392 * in writing out this specific lwb's data, and in the case that cache
1393 * flushes have been deferred, vdevs involved in writing the data for
1394 * previous lwbs. The writes corresponding to all the vdevs in the
1395 * lwb_vdev_tree will have completed by the time this is called, due to
1396 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1397 * which takes deferred flushes into account. The lwb will be "done"
1398 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1399 * completion callback for the lwb's root zio.
1402 zil_lwb_write_done(zio_t
*zio
)
1404 lwb_t
*lwb
= zio
->io_private
;
1405 spa_t
*spa
= zio
->io_spa
;
1406 zilog_t
*zilog
= lwb
->lwb_zilog
;
1407 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1408 void *cookie
= NULL
;
1409 zil_vdev_node_t
*zv
;
1412 ASSERT3S(spa_config_held(spa
, SCL_STATE
, RW_READER
), !=, 0);
1414 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1415 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
1416 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
1417 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
1418 ASSERT(!BP_IS_GANG(zio
->io_bp
));
1419 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
1420 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
1422 abd_free(zio
->io_abd
);
1424 mutex_enter(&zilog
->zl_lock
);
1425 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_ISSUED
);
1426 lwb
->lwb_state
= LWB_STATE_WRITE_DONE
;
1427 lwb
->lwb_write_zio
= NULL
;
1428 lwb
->lwb_fastwrite
= FALSE
;
1429 nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1430 mutex_exit(&zilog
->zl_lock
);
1432 if (avl_numnodes(t
) == 0)
1436 * If there was an IO error, we're not going to call zio_flush()
1437 * on these vdevs, so we simply empty the tree and free the
1438 * nodes. We avoid calling zio_flush() since there isn't any
1439 * good reason for doing so, after the lwb block failed to be
1442 * Additionally, we don't perform any further error handling at
1443 * this point (e.g. setting "zcw_zio_error" appropriately), as
1444 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1445 * we expect any error seen here, to have been propagated to
1448 if (zio
->io_error
!= 0) {
1449 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
1450 kmem_free(zv
, sizeof (*zv
));
1455 * If this lwb does not have any threads waiting for it to
1456 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1457 * command to the vdevs written to by "this" lwb, and instead
1458 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1459 * command for those vdevs. Thus, we merge the vdev tree of
1460 * "this" lwb with the vdev tree of the "next" lwb in the list,
1461 * and assume the "next" lwb will handle flushing the vdevs (or
1462 * deferring the flush(s) again).
1464 * This is a useful performance optimization, especially for
1465 * workloads with lots of async write activity and few sync
1466 * write and/or fsync activity, as it has the potential to
1467 * coalesce multiple flush commands to a vdev into one.
1469 if (list_head(&lwb
->lwb_waiters
) == NULL
&& nlwb
!= NULL
) {
1470 zil_lwb_flush_defer(lwb
, nlwb
);
1471 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
1475 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1476 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
1479 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1480 * always used within "zio_flush". This means,
1481 * any errors when flushing the vdev(s), will
1482 * (unfortunately) not be handled correctly,
1483 * since these "zio_flush" errors will not be
1484 * propagated up to "zil_lwb_flush_vdevs_done".
1486 zio_flush(lwb
->lwb_root_zio
, vd
);
1488 kmem_free(zv
, sizeof (*zv
));
1493 zil_lwb_set_zio_dependency(zilog_t
*zilog
, lwb_t
*lwb
)
1495 lwb_t
*last_lwb_opened
= zilog
->zl_last_lwb_opened
;
1497 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1498 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
1501 * The zilog's "zl_last_lwb_opened" field is used to build the
1502 * lwb/zio dependency chain, which is used to preserve the
1503 * ordering of lwb completions that is required by the semantics
1504 * of the ZIL. Each new lwb zio becomes a parent of the
1505 * "previous" lwb zio, such that the new lwb's zio cannot
1506 * complete until the "previous" lwb's zio completes.
1508 * This is required by the semantics of zil_commit(); the commit
1509 * waiters attached to the lwbs will be woken in the lwb zio's
1510 * completion callback, so this zio dependency graph ensures the
1511 * waiters are woken in the correct order (the same order the
1512 * lwbs were created).
1514 if (last_lwb_opened
!= NULL
&&
1515 last_lwb_opened
->lwb_state
!= LWB_STATE_FLUSH_DONE
) {
1516 ASSERT(last_lwb_opened
->lwb_state
== LWB_STATE_OPENED
||
1517 last_lwb_opened
->lwb_state
== LWB_STATE_ISSUED
||
1518 last_lwb_opened
->lwb_state
== LWB_STATE_WRITE_DONE
);
1520 ASSERT3P(last_lwb_opened
->lwb_root_zio
, !=, NULL
);
1521 zio_add_child(lwb
->lwb_root_zio
,
1522 last_lwb_opened
->lwb_root_zio
);
1525 * If the previous lwb's write hasn't already completed,
1526 * we also want to order the completion of the lwb write
1527 * zios (above, we only order the completion of the lwb
1528 * root zios). This is required because of how we can
1529 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1531 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1532 * the previous lwb will rely on this lwb to flush the
1533 * vdevs written to by that previous lwb. Thus, we need
1534 * to ensure this lwb doesn't issue the flush until
1535 * after the previous lwb's write completes. We ensure
1536 * this ordering by setting the zio parent/child
1537 * relationship here.
1539 * Without this relationship on the lwb's write zio,
1540 * it's possible for this lwb's write to complete prior
1541 * to the previous lwb's write completing; and thus, the
1542 * vdevs for the previous lwb would be flushed prior to
1543 * that lwb's data being written to those vdevs (the
1544 * vdevs are flushed in the lwb write zio's completion
1545 * handler, zil_lwb_write_done()).
1547 if (last_lwb_opened
->lwb_state
!= LWB_STATE_WRITE_DONE
) {
1548 ASSERT(last_lwb_opened
->lwb_state
== LWB_STATE_OPENED
||
1549 last_lwb_opened
->lwb_state
== LWB_STATE_ISSUED
);
1551 ASSERT3P(last_lwb_opened
->lwb_write_zio
, !=, NULL
);
1552 zio_add_child(lwb
->lwb_write_zio
,
1553 last_lwb_opened
->lwb_write_zio
);
1560 * This function's purpose is to "open" an lwb such that it is ready to
1561 * accept new itxs being committed to it. To do this, the lwb's zio
1562 * structures are created, and linked to the lwb. This function is
1563 * idempotent; if the passed in lwb has already been opened, this
1564 * function is essentially a no-op.
1567 zil_lwb_write_open(zilog_t
*zilog
, lwb_t
*lwb
)
1569 zbookmark_phys_t zb
;
1570 zio_priority_t prio
;
1572 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1573 ASSERT3P(lwb
, !=, NULL
);
1574 EQUIV(lwb
->lwb_root_zio
== NULL
, lwb
->lwb_state
== LWB_STATE_CLOSED
);
1575 EQUIV(lwb
->lwb_root_zio
!= NULL
, lwb
->lwb_state
== LWB_STATE_OPENED
);
1577 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1578 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
1579 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1581 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1582 mutex_enter(&zilog
->zl_lock
);
1583 if (lwb
->lwb_root_zio
== NULL
) {
1584 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
,
1585 BP_GET_LSIZE(&lwb
->lwb_blk
));
1587 if (!lwb
->lwb_fastwrite
) {
1588 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1589 lwb
->lwb_fastwrite
= 1;
1592 if (!lwb
->lwb_slog
|| zilog
->zl_cur_used
<= zil_slog_bulk
)
1593 prio
= ZIO_PRIORITY_SYNC_WRITE
;
1595 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
1597 lwb
->lwb_root_zio
= zio_root(zilog
->zl_spa
,
1598 zil_lwb_flush_vdevs_done
, lwb
, ZIO_FLAG_CANFAIL
);
1599 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1601 lwb
->lwb_write_zio
= zio_rewrite(lwb
->lwb_root_zio
,
1602 zilog
->zl_spa
, 0, &lwb
->lwb_blk
, lwb_abd
,
1603 BP_GET_LSIZE(&lwb
->lwb_blk
), zil_lwb_write_done
, lwb
,
1604 prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_FASTWRITE
, &zb
);
1605 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1607 lwb
->lwb_state
= LWB_STATE_OPENED
;
1609 zil_lwb_set_zio_dependency(zilog
, lwb
);
1610 zilog
->zl_last_lwb_opened
= lwb
;
1612 mutex_exit(&zilog
->zl_lock
);
1614 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1615 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1616 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1620 * Define a limited set of intent log block sizes.
1622 * These must be a multiple of 4KB. Note only the amount used (again
1623 * aligned to 4KB) actually gets written. However, we can't always just
1624 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1626 static const struct {
1629 } zil_block_buckets
[] = {
1630 { 4096, 4096 }, /* non TX_WRITE */
1631 { 8192 + 4096, 8192 + 4096 }, /* database */
1632 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */
1633 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */
1634 { 131072, 131072 }, /* < 128KB writes */
1635 { 131072 +4096, 65536 + 4096 }, /* 128KB writes */
1636 { UINT64_MAX
, SPA_OLD_MAXBLOCKSIZE
}, /* > 128KB writes */
1640 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1641 * initialized. Otherwise this should not be used directly; see
1642 * zl_max_block_size instead.
1644 static uint_t zil_maxblocksize
= SPA_OLD_MAXBLOCKSIZE
;
1647 * Start a log block write and advance to the next log block.
1648 * Calls are serialized.
1651 zil_lwb_write_issue(zilog_t
*zilog
, lwb_t
*lwb
)
1655 spa_t
*spa
= zilog
->zl_spa
;
1659 uint64_t zil_blksz
, wsz
;
1663 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1664 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1665 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1666 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1668 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1669 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
1670 bp
= &zilc
->zc_next_blk
;
1672 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
1673 bp
= &zilc
->zc_next_blk
;
1676 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
1679 * Allocate the next block and save its address in this block
1680 * before writing it in order to establish the log chain.
1683 tx
= dmu_tx_create(zilog
->zl_os
);
1686 * Since we are not going to create any new dirty data, and we
1687 * can even help with clearing the existing dirty data, we
1688 * should not be subject to the dirty data based delays. We
1689 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1691 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
1693 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1694 txg
= dmu_tx_get_txg(tx
);
1696 mutex_enter(&zilog
->zl_lwb_io_lock
);
1697 lwb
->lwb_issued_txg
= txg
;
1698 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]++;
1699 zilog
->zl_lwb_max_issued_txg
= MAX(txg
, zilog
->zl_lwb_max_issued_txg
);
1700 mutex_exit(&zilog
->zl_lwb_io_lock
);
1703 * Log blocks are pre-allocated. Here we select the size of the next
1704 * block, based on size used in the last block.
1705 * - first find the smallest bucket that will fit the block from a
1706 * limited set of block sizes. This is because it's faster to write
1707 * blocks allocated from the same metaslab as they are adjacent or
1709 * - next find the maximum from the new suggested size and an array of
1710 * previous sizes. This lessens a picket fence effect of wrongly
1711 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1714 * Note we only write what is used, but we can't just allocate
1715 * the maximum block size because we can exhaust the available
1718 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1719 for (i
= 0; zil_blksz
> zil_block_buckets
[i
].limit
; i
++)
1721 zil_blksz
= MIN(zil_block_buckets
[i
].blksz
, zilog
->zl_max_block_size
);
1722 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1723 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1724 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1725 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1728 error
= zio_alloc_zil(spa
, zilog
->zl_os
, txg
, bp
, zil_blksz
, &slog
);
1730 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_slog_count
);
1731 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_bytes
,
1734 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_normal_count
);
1735 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_bytes
,
1739 ASSERT3U(bp
->blk_birth
, ==, txg
);
1740 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1741 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1744 * Allocate a new log write block (lwb).
1746 nlwb
= zil_alloc_lwb(zilog
, bp
, slog
, txg
, TRUE
);
1749 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1750 /* For Slim ZIL only write what is used. */
1751 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1752 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1753 zio_shrink(lwb
->lwb_write_zio
, wsz
);
1760 zilc
->zc_nused
= lwb
->lwb_nused
;
1761 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1764 * clear unused data for security
1766 memset(lwb
->lwb_buf
+ lwb
->lwb_nused
, 0, wsz
- lwb
->lwb_nused
);
1768 spa_config_enter(zilog
->zl_spa
, SCL_STATE
, lwb
, RW_READER
);
1770 zil_lwb_add_block(lwb
, &lwb
->lwb_blk
);
1771 lwb
->lwb_issued_timestamp
= gethrtime();
1772 lwb
->lwb_state
= LWB_STATE_ISSUED
;
1774 zio_nowait(lwb
->lwb_root_zio
);
1775 zio_nowait(lwb
->lwb_write_zio
);
1780 * If there was an allocation failure then nlwb will be null which
1781 * forces a txg_wait_synced().
1787 * Maximum amount of write data that can be put into single log block.
1790 zil_max_log_data(zilog_t
*zilog
)
1792 return (zilog
->zl_max_block_size
-
1793 sizeof (zil_chain_t
) - sizeof (lr_write_t
));
1797 * Maximum amount of log space we agree to waste to reduce number of
1798 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1800 static inline uint64_t
1801 zil_max_waste_space(zilog_t
*zilog
)
1803 return (zil_max_log_data(zilog
) / 8);
1807 * Maximum amount of write data for WR_COPIED. For correctness, consumers
1808 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1809 * maximum sized log block, because each WR_COPIED record must fit in a
1810 * single log block. For space efficiency, we want to fit two records into a
1811 * max-sized log block.
1814 zil_max_copied_data(zilog_t
*zilog
)
1816 return ((zilog
->zl_max_block_size
- sizeof (zil_chain_t
)) / 2 -
1817 sizeof (lr_write_t
));
1821 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1824 lr_write_t
*lrwb
, *lrw
;
1826 uint64_t dlen
, dnow
, dpad
, lwb_sp
, reclen
, txg
, max_log_data
;
1828 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1829 ASSERT3P(lwb
, !=, NULL
);
1830 ASSERT3P(lwb
->lwb_buf
, !=, NULL
);
1832 zil_lwb_write_open(zilog
, lwb
);
1835 lrw
= (lr_write_t
*)lrc
;
1838 * A commit itx doesn't represent any on-disk state; instead
1839 * it's simply used as a place holder on the commit list, and
1840 * provides a mechanism for attaching a "commit waiter" onto the
1841 * correct lwb (such that the waiter can be signalled upon
1842 * completion of that lwb). Thus, we don't process this itx's
1843 * log record if it's a commit itx (these itx's don't have log
1844 * records), and instead link the itx's waiter onto the lwb's
1847 * For more details, see the comment above zil_commit().
1849 if (lrc
->lrc_txtype
== TX_COMMIT
) {
1850 mutex_enter(&zilog
->zl_lock
);
1851 zil_commit_waiter_link_lwb(itx
->itx_private
, lwb
);
1852 itx
->itx_private
= NULL
;
1853 mutex_exit(&zilog
->zl_lock
);
1857 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
1858 dlen
= P2ROUNDUP_TYPED(
1859 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1860 dpad
= dlen
- lrw
->lr_length
;
1864 reclen
= lrc
->lrc_reclen
;
1865 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1868 ASSERT3U(zilog
->zl_cur_used
, <, UINT64_MAX
- (reclen
+ dlen
));
1872 * If this record won't fit in the current log block, start a new one.
1873 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1875 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1876 max_log_data
= zil_max_log_data(zilog
);
1877 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
1878 lwb_sp
< zil_max_waste_space(zilog
) &&
1879 (dlen
% max_log_data
== 0 ||
1880 lwb_sp
< reclen
+ dlen
% max_log_data
))) {
1881 lwb
= zil_lwb_write_issue(zilog
, lwb
);
1884 zil_lwb_write_open(zilog
, lwb
);
1885 ASSERT(LWB_EMPTY(lwb
));
1886 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1889 * There must be enough space in the new, empty log block to
1890 * hold reclen. For WR_COPIED, we need to fit the whole
1891 * record in one block, and reclen is the header size + the
1892 * data size. For WR_NEED_COPY, we can create multiple
1893 * records, splitting the data into multiple blocks, so we
1894 * only need to fit one word of data per block; in this case
1895 * reclen is just the header size (no data).
1897 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
1900 dnow
= MIN(dlen
, lwb_sp
- reclen
);
1901 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1902 memcpy(lr_buf
, lrc
, reclen
);
1903 lrcb
= (lr_t
*)lr_buf
; /* Like lrc, but inside lwb. */
1904 lrwb
= (lr_write_t
*)lrcb
; /* Like lrw, but inside lwb. */
1906 ZIL_STAT_BUMP(zilog
, zil_itx_count
);
1909 * If it's a write, fetch the data or get its blkptr as appropriate.
1911 if (lrc
->lrc_txtype
== TX_WRITE
) {
1912 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1913 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1914 if (itx
->itx_wr_state
== WR_COPIED
) {
1915 ZIL_STAT_BUMP(zilog
, zil_itx_copied_count
);
1916 ZIL_STAT_INCR(zilog
, zil_itx_copied_bytes
,
1922 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
1923 dbuf
= lr_buf
+ reclen
;
1924 lrcb
->lrc_reclen
+= dnow
;
1925 if (lrwb
->lr_length
> dnow
)
1926 lrwb
->lr_length
= dnow
;
1927 lrw
->lr_offset
+= dnow
;
1928 lrw
->lr_length
-= dnow
;
1929 ZIL_STAT_BUMP(zilog
, zil_itx_needcopy_count
);
1930 ZIL_STAT_INCR(zilog
, zil_itx_needcopy_bytes
,
1933 ASSERT3S(itx
->itx_wr_state
, ==, WR_INDIRECT
);
1935 ZIL_STAT_BUMP(zilog
, zil_itx_indirect_count
);
1936 ZIL_STAT_INCR(zilog
, zil_itx_indirect_bytes
,
1941 * We pass in the "lwb_write_zio" rather than
1942 * "lwb_root_zio" so that the "lwb_write_zio"
1943 * becomes the parent of any zio's created by
1944 * the "zl_get_data" callback. The vdevs are
1945 * flushed after the "lwb_write_zio" completes,
1946 * so we want to make sure that completion
1947 * callback waits for these additional zio's,
1948 * such that the vdevs used by those zio's will
1949 * be included in the lwb's vdev tree, and those
1950 * vdevs will be properly flushed. If we passed
1951 * in "lwb_root_zio" here, then these additional
1952 * vdevs may not be flushed; e.g. if these zio's
1953 * completed after "lwb_write_zio" completed.
1955 error
= zilog
->zl_get_data(itx
->itx_private
,
1956 itx
->itx_gen
, lrwb
, dbuf
, lwb
,
1957 lwb
->lwb_write_zio
);
1958 if (dbuf
!= NULL
&& error
== 0 && dnow
== dlen
)
1959 /* Zero any padding bytes in the last block. */
1960 memset((char *)dbuf
+ lrwb
->lr_length
, 0, dpad
);
1963 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1967 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1975 * We're actually making an entry, so update lrc_seq to be the
1976 * log record sequence number. Note that this is generally not
1977 * equal to the itx sequence number because not all transactions
1978 * are synchronous, and sometimes spa_sync() gets there first.
1980 lrcb
->lrc_seq
= ++zilog
->zl_lr_seq
;
1981 lwb
->lwb_nused
+= reclen
+ dnow
;
1983 zil_lwb_add_txg(lwb
, txg
);
1985 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1986 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1990 zilog
->zl_cur_used
+= reclen
;
1998 zil_itx_create(uint64_t txtype
, size_t olrsize
)
2000 size_t itxsize
, lrsize
;
2003 lrsize
= P2ROUNDUP_TYPED(olrsize
, sizeof (uint64_t), size_t);
2004 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
2006 itx
= zio_data_buf_alloc(itxsize
);
2007 itx
->itx_lr
.lrc_txtype
= txtype
;
2008 itx
->itx_lr
.lrc_reclen
= lrsize
;
2009 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
2010 memset((char *)&itx
->itx_lr
+ olrsize
, 0, lrsize
- olrsize
);
2011 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
2012 itx
->itx_callback
= NULL
;
2013 itx
->itx_callback_data
= NULL
;
2014 itx
->itx_size
= itxsize
;
2020 zil_itx_destroy(itx_t
*itx
)
2022 IMPLY(itx
->itx_lr
.lrc_txtype
== TX_COMMIT
, itx
->itx_callback
== NULL
);
2023 IMPLY(itx
->itx_callback
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2025 if (itx
->itx_callback
!= NULL
)
2026 itx
->itx_callback(itx
->itx_callback_data
);
2028 zio_data_buf_free(itx
, itx
->itx_size
);
2032 * Free up the sync and async itxs. The itxs_t has already been detached
2033 * so no locks are needed.
2036 zil_itxg_clean(void *arg
)
2043 itx_async_node_t
*ian
;
2045 list
= &itxs
->i_sync_list
;
2046 while ((itx
= list_head(list
)) != NULL
) {
2048 * In the general case, commit itxs will not be found
2049 * here, as they'll be committed to an lwb via
2050 * zil_lwb_commit(), and free'd in that function. Having
2051 * said that, it is still possible for commit itxs to be
2052 * found here, due to the following race:
2054 * - a thread calls zil_commit() which assigns the
2055 * commit itx to a per-txg i_sync_list
2056 * - zil_itxg_clean() is called (e.g. via spa_sync())
2057 * while the waiter is still on the i_sync_list
2059 * There's nothing to prevent syncing the txg while the
2060 * waiter is on the i_sync_list. This normally doesn't
2061 * happen because spa_sync() is slower than zil_commit(),
2062 * but if zil_commit() calls txg_wait_synced() (e.g.
2063 * because zil_create() or zil_commit_writer_stall() is
2064 * called) we will hit this case.
2066 if (itx
->itx_lr
.lrc_txtype
== TX_COMMIT
)
2067 zil_commit_waiter_skip(itx
->itx_private
);
2069 list_remove(list
, itx
);
2070 zil_itx_destroy(itx
);
2074 t
= &itxs
->i_async_tree
;
2075 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2076 list
= &ian
->ia_list
;
2077 while ((itx
= list_head(list
)) != NULL
) {
2078 list_remove(list
, itx
);
2079 /* commit itxs should never be on the async lists. */
2080 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2081 zil_itx_destroy(itx
);
2084 kmem_free(ian
, sizeof (itx_async_node_t
));
2088 kmem_free(itxs
, sizeof (itxs_t
));
2092 zil_aitx_compare(const void *x1
, const void *x2
)
2094 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
2095 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
2097 return (TREE_CMP(o1
, o2
));
2101 * Remove all async itx with the given oid.
2104 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
2107 itx_async_node_t
*ian
;
2114 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2116 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2119 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2121 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2122 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2124 mutex_enter(&itxg
->itxg_lock
);
2125 if (itxg
->itxg_txg
!= txg
) {
2126 mutex_exit(&itxg
->itxg_lock
);
2131 * Locate the object node and append its list.
2133 t
= &itxg
->itxg_itxs
->i_async_tree
;
2134 ian
= avl_find(t
, &oid
, &where
);
2136 list_move_tail(&clean_list
, &ian
->ia_list
);
2137 mutex_exit(&itxg
->itxg_lock
);
2139 while ((itx
= list_head(&clean_list
)) != NULL
) {
2140 list_remove(&clean_list
, itx
);
2141 /* commit itxs should never be on the async lists. */
2142 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2143 zil_itx_destroy(itx
);
2145 list_destroy(&clean_list
);
2149 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
2153 itxs_t
*itxs
, *clean
= NULL
;
2156 * Ensure the data of a renamed file is committed before the rename.
2158 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
2159 zil_async_to_sync(zilog
, itx
->itx_oid
);
2161 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
2164 txg
= dmu_tx_get_txg(tx
);
2166 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2167 mutex_enter(&itxg
->itxg_lock
);
2168 itxs
= itxg
->itxg_itxs
;
2169 if (itxg
->itxg_txg
!= txg
) {
2172 * The zil_clean callback hasn't got around to cleaning
2173 * this itxg. Save the itxs for release below.
2174 * This should be rare.
2176 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2177 "txg %llu", (u_longlong_t
)itxg
->itxg_txg
);
2178 clean
= itxg
->itxg_itxs
;
2180 itxg
->itxg_txg
= txg
;
2181 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
2184 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
2185 offsetof(itx_t
, itx_node
));
2186 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
2187 sizeof (itx_async_node_t
),
2188 offsetof(itx_async_node_t
, ia_node
));
2190 if (itx
->itx_sync
) {
2191 list_insert_tail(&itxs
->i_sync_list
, itx
);
2193 avl_tree_t
*t
= &itxs
->i_async_tree
;
2195 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
2196 itx_async_node_t
*ian
;
2199 ian
= avl_find(t
, &foid
, &where
);
2201 ian
= kmem_alloc(sizeof (itx_async_node_t
),
2203 list_create(&ian
->ia_list
, sizeof (itx_t
),
2204 offsetof(itx_t
, itx_node
));
2205 ian
->ia_foid
= foid
;
2206 avl_insert(t
, ian
, where
);
2208 list_insert_tail(&ian
->ia_list
, itx
);
2211 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
2214 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2215 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2216 * need to be careful to always dirty the ZIL using the "real"
2217 * TXG (not itxg_txg) even when the SPA is frozen.
2219 zilog_dirty(zilog
, dmu_tx_get_txg(tx
));
2220 mutex_exit(&itxg
->itxg_lock
);
2222 /* Release the old itxs now we've dropped the lock */
2224 zil_itxg_clean(clean
);
2228 * If there are any in-memory intent log transactions which have now been
2229 * synced then start up a taskq to free them. We should only do this after we
2230 * have written out the uberblocks (i.e. txg has been committed) so that
2231 * don't inadvertently clean out in-memory log records that would be required
2235 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
2237 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
2240 ASSERT3U(synced_txg
, <, ZILTEST_TXG
);
2242 mutex_enter(&itxg
->itxg_lock
);
2243 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
2244 mutex_exit(&itxg
->itxg_lock
);
2247 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
2248 ASSERT3U(itxg
->itxg_txg
, !=, 0);
2249 clean_me
= itxg
->itxg_itxs
;
2250 itxg
->itxg_itxs
= NULL
;
2252 mutex_exit(&itxg
->itxg_lock
);
2254 * Preferably start a task queue to free up the old itxs but
2255 * if taskq_dispatch can't allocate resources to do that then
2256 * free it in-line. This should be rare. Note, using TQ_SLEEP
2257 * created a bad performance problem.
2259 ASSERT3P(zilog
->zl_dmu_pool
, !=, NULL
);
2260 ASSERT3P(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
, !=, NULL
);
2261 taskqid_t id
= taskq_dispatch(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
,
2262 zil_itxg_clean
, clean_me
, TQ_NOSLEEP
);
2263 if (id
== TASKQID_INVALID
)
2264 zil_itxg_clean(clean_me
);
2268 * This function will traverse the queue of itxs that need to be
2269 * committed, and move them onto the ZIL's zl_itx_commit_list.
2272 zil_get_commit_list(zilog_t
*zilog
)
2275 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
2277 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2279 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2282 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2285 * This is inherently racy, since there is nothing to prevent
2286 * the last synced txg from changing. That's okay since we'll
2287 * only commit things in the future.
2289 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2290 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2292 mutex_enter(&itxg
->itxg_lock
);
2293 if (itxg
->itxg_txg
!= txg
) {
2294 mutex_exit(&itxg
->itxg_lock
);
2299 * If we're adding itx records to the zl_itx_commit_list,
2300 * then the zil better be dirty in this "txg". We can assert
2301 * that here since we're holding the itxg_lock which will
2302 * prevent spa_sync from cleaning it. Once we add the itxs
2303 * to the zl_itx_commit_list we must commit it to disk even
2304 * if it's unnecessary (i.e. the txg was synced).
2306 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
2307 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
2308 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
2310 mutex_exit(&itxg
->itxg_lock
);
2315 * Move the async itxs for a specified object to commit into sync lists.
2318 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
2321 itx_async_node_t
*ian
;
2325 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2328 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2331 * This is inherently racy, since there is nothing to prevent
2332 * the last synced txg from changing.
2334 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2335 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2337 mutex_enter(&itxg
->itxg_lock
);
2338 if (itxg
->itxg_txg
!= txg
) {
2339 mutex_exit(&itxg
->itxg_lock
);
2344 * If a foid is specified then find that node and append its
2345 * list. Otherwise walk the tree appending all the lists
2346 * to the sync list. We add to the end rather than the
2347 * beginning to ensure the create has happened.
2349 t
= &itxg
->itxg_itxs
->i_async_tree
;
2351 ian
= avl_find(t
, &foid
, &where
);
2353 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2357 void *cookie
= NULL
;
2359 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2360 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2362 list_destroy(&ian
->ia_list
);
2363 kmem_free(ian
, sizeof (itx_async_node_t
));
2366 mutex_exit(&itxg
->itxg_lock
);
2371 * This function will prune commit itxs that are at the head of the
2372 * commit list (it won't prune past the first non-commit itx), and
2373 * either: a) attach them to the last lwb that's still pending
2374 * completion, or b) skip them altogether.
2376 * This is used as a performance optimization to prevent commit itxs
2377 * from generating new lwbs when it's unnecessary to do so.
2380 zil_prune_commit_list(zilog_t
*zilog
)
2384 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2386 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2387 lr_t
*lrc
= &itx
->itx_lr
;
2388 if (lrc
->lrc_txtype
!= TX_COMMIT
)
2391 mutex_enter(&zilog
->zl_lock
);
2393 lwb_t
*last_lwb
= zilog
->zl_last_lwb_opened
;
2394 if (last_lwb
== NULL
||
2395 last_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
) {
2397 * All of the itxs this waiter was waiting on
2398 * must have already completed (or there were
2399 * never any itx's for it to wait on), so it's
2400 * safe to skip this waiter and mark it done.
2402 zil_commit_waiter_skip(itx
->itx_private
);
2404 zil_commit_waiter_link_lwb(itx
->itx_private
, last_lwb
);
2405 itx
->itx_private
= NULL
;
2408 mutex_exit(&zilog
->zl_lock
);
2410 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2411 zil_itx_destroy(itx
);
2414 IMPLY(itx
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2418 zil_commit_writer_stall(zilog_t
*zilog
)
2421 * When zio_alloc_zil() fails to allocate the next lwb block on
2422 * disk, we must call txg_wait_synced() to ensure all of the
2423 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2424 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2425 * to zil_process_commit_list()) will have to call zil_create(),
2426 * and start a new ZIL chain.
2428 * Since zil_alloc_zil() failed, the lwb that was previously
2429 * issued does not have a pointer to the "next" lwb on disk.
2430 * Thus, if another ZIL writer thread was to allocate the "next"
2431 * on-disk lwb, that block could be leaked in the event of a
2432 * crash (because the previous lwb on-disk would not point to
2435 * We must hold the zilog's zl_issuer_lock while we do this, to
2436 * ensure no new threads enter zil_process_commit_list() until
2437 * all lwb's in the zl_lwb_list have been synced and freed
2438 * (which is achieved via the txg_wait_synced() call).
2440 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2441 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2442 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
2446 * This function will traverse the commit list, creating new lwbs as
2447 * needed, and committing the itxs from the commit list to these newly
2448 * created lwbs. Additionally, as a new lwb is created, the previous
2449 * lwb will be issued to the zio layer to be written to disk.
2452 zil_process_commit_list(zilog_t
*zilog
)
2454 spa_t
*spa
= zilog
->zl_spa
;
2456 list_t nolwb_waiters
;
2460 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2463 * Return if there's nothing to commit before we dirty the fs by
2464 * calling zil_create().
2466 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
)
2469 list_create(&nolwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2470 list_create(&nolwb_waiters
, sizeof (zil_commit_waiter_t
),
2471 offsetof(zil_commit_waiter_t
, zcw_node
));
2473 lwb
= list_tail(&zilog
->zl_lwb_list
);
2475 lwb
= zil_create(zilog
);
2478 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2479 * have already been created (zl_lwb_list not empty).
2481 zil_commit_activate_saxattr_feature(zilog
);
2482 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2483 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
2484 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
2487 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2488 lr_t
*lrc
= &itx
->itx_lr
;
2489 uint64_t txg
= lrc
->lrc_txg
;
2491 ASSERT3U(txg
, !=, 0);
2493 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2494 DTRACE_PROBE2(zil__process__commit__itx
,
2495 zilog_t
*, zilog
, itx_t
*, itx
);
2497 DTRACE_PROBE2(zil__process__normal__itx
,
2498 zilog_t
*, zilog
, itx_t
*, itx
);
2501 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2503 boolean_t synced
= txg
<= spa_last_synced_txg(spa
);
2504 boolean_t frozen
= txg
> spa_freeze_txg(spa
);
2507 * If the txg of this itx has already been synced out, then
2508 * we don't need to commit this itx to an lwb. This is
2509 * because the data of this itx will have already been
2510 * written to the main pool. This is inherently racy, and
2511 * it's still ok to commit an itx whose txg has already
2512 * been synced; this will result in a write that's
2513 * unnecessary, but will do no harm.
2515 * With that said, we always want to commit TX_COMMIT itxs
2516 * to an lwb, regardless of whether or not that itx's txg
2517 * has been synced out. We do this to ensure any OPENED lwb
2518 * will always have at least one zil_commit_waiter_t linked
2521 * As a counter-example, if we skipped TX_COMMIT itx's
2522 * whose txg had already been synced, the following
2523 * situation could occur if we happened to be racing with
2526 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2527 * itx's txg is 10 and the last synced txg is 9.
2528 * 2. spa_sync finishes syncing out txg 10.
2529 * 3. We move to the next itx in the list, it's a TX_COMMIT
2530 * whose txg is 10, so we skip it rather than committing
2531 * it to the lwb used in (1).
2533 * If the itx that is skipped in (3) is the last TX_COMMIT
2534 * itx in the commit list, than it's possible for the lwb
2535 * used in (1) to remain in the OPENED state indefinitely.
2537 * To prevent the above scenario from occurring, ensuring
2538 * that once an lwb is OPENED it will transition to ISSUED
2539 * and eventually DONE, we always commit TX_COMMIT itx's to
2540 * an lwb here, even if that itx's txg has already been
2543 * Finally, if the pool is frozen, we _always_ commit the
2544 * itx. The point of freezing the pool is to prevent data
2545 * from being written to the main pool via spa_sync, and
2546 * instead rely solely on the ZIL to persistently store the
2547 * data; i.e. when the pool is frozen, the last synced txg
2548 * value can't be trusted.
2550 if (frozen
|| !synced
|| lrc
->lrc_txtype
== TX_COMMIT
) {
2552 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
2555 list_insert_tail(&nolwb_itxs
, itx
);
2557 list_insert_tail(&lwb
->lwb_itxs
, itx
);
2559 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2560 zil_commit_waiter_link_nolwb(
2561 itx
->itx_private
, &nolwb_waiters
);
2564 list_insert_tail(&nolwb_itxs
, itx
);
2567 ASSERT3S(lrc
->lrc_txtype
, !=, TX_COMMIT
);
2568 zil_itx_destroy(itx
);
2574 * This indicates zio_alloc_zil() failed to allocate the
2575 * "next" lwb on-disk. When this happens, we must stall
2576 * the ZIL write pipeline; see the comment within
2577 * zil_commit_writer_stall() for more details.
2579 zil_commit_writer_stall(zilog
);
2582 * Additionally, we have to signal and mark the "nolwb"
2583 * waiters as "done" here, since without an lwb, we
2584 * can't do this via zil_lwb_flush_vdevs_done() like
2587 zil_commit_waiter_t
*zcw
;
2588 while ((zcw
= list_head(&nolwb_waiters
)) != NULL
) {
2589 zil_commit_waiter_skip(zcw
);
2590 list_remove(&nolwb_waiters
, zcw
);
2594 * And finally, we have to destroy the itx's that
2595 * couldn't be committed to an lwb; this will also call
2596 * the itx's callback if one exists for the itx.
2598 while ((itx
= list_head(&nolwb_itxs
)) != NULL
) {
2599 list_remove(&nolwb_itxs
, itx
);
2600 zil_itx_destroy(itx
);
2603 ASSERT(list_is_empty(&nolwb_waiters
));
2604 ASSERT3P(lwb
, !=, NULL
);
2605 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2606 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
2607 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
2610 * At this point, the ZIL block pointed at by the "lwb"
2611 * variable is in one of the following states: "closed"
2614 * If it's "closed", then no itxs have been committed to
2615 * it, so there's no point in issuing its zio (i.e. it's
2618 * If it's "open", then it contains one or more itxs that
2619 * eventually need to be committed to stable storage. In
2620 * this case we intentionally do not issue the lwb's zio
2621 * to disk yet, and instead rely on one of the following
2622 * two mechanisms for issuing the zio:
2624 * 1. Ideally, there will be more ZIL activity occurring
2625 * on the system, such that this function will be
2626 * immediately called again (not necessarily by the same
2627 * thread) and this lwb's zio will be issued via
2628 * zil_lwb_commit(). This way, the lwb is guaranteed to
2629 * be "full" when it is issued to disk, and we'll make
2630 * use of the lwb's size the best we can.
2632 * 2. If there isn't sufficient ZIL activity occurring on
2633 * the system, such that this lwb's zio isn't issued via
2634 * zil_lwb_commit(), zil_commit_waiter() will issue the
2635 * lwb's zio. If this occurs, the lwb is not guaranteed
2636 * to be "full" by the time its zio is issued, and means
2637 * the size of the lwb was "too large" given the amount
2638 * of ZIL activity occurring on the system at that time.
2640 * We do this for a couple of reasons:
2642 * 1. To try and reduce the number of IOPs needed to
2643 * write the same number of itxs. If an lwb has space
2644 * available in its buffer for more itxs, and more itxs
2645 * will be committed relatively soon (relative to the
2646 * latency of performing a write), then it's beneficial
2647 * to wait for these "next" itxs. This way, more itxs
2648 * can be committed to stable storage with fewer writes.
2650 * 2. To try and use the largest lwb block size that the
2651 * incoming rate of itxs can support. Again, this is to
2652 * try and pack as many itxs into as few lwbs as
2653 * possible, without significantly impacting the latency
2654 * of each individual itx.
2660 * This function is responsible for ensuring the passed in commit waiter
2661 * (and associated commit itx) is committed to an lwb. If the waiter is
2662 * not already committed to an lwb, all itxs in the zilog's queue of
2663 * itxs will be processed. The assumption is the passed in waiter's
2664 * commit itx will found in the queue just like the other non-commit
2665 * itxs, such that when the entire queue is processed, the waiter will
2666 * have been committed to an lwb.
2668 * The lwb associated with the passed in waiter is not guaranteed to
2669 * have been issued by the time this function completes. If the lwb is
2670 * not issued, we rely on future calls to zil_commit_writer() to issue
2671 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2674 zil_commit_writer(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2676 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2677 ASSERT(spa_writeable(zilog
->zl_spa
));
2679 mutex_enter(&zilog
->zl_issuer_lock
);
2681 if (zcw
->zcw_lwb
!= NULL
|| zcw
->zcw_done
) {
2683 * It's possible that, while we were waiting to acquire
2684 * the "zl_issuer_lock", another thread committed this
2685 * waiter to an lwb. If that occurs, we bail out early,
2686 * without processing any of the zilog's queue of itxs.
2688 * On certain workloads and system configurations, the
2689 * "zl_issuer_lock" can become highly contended. In an
2690 * attempt to reduce this contention, we immediately drop
2691 * the lock if the waiter has already been processed.
2693 * We've measured this optimization to reduce CPU spent
2694 * contending on this lock by up to 5%, using a system
2695 * with 32 CPUs, low latency storage (~50 usec writes),
2696 * and 1024 threads performing sync writes.
2701 ZIL_STAT_BUMP(zilog
, zil_commit_writer_count
);
2703 zil_get_commit_list(zilog
);
2704 zil_prune_commit_list(zilog
);
2705 zil_process_commit_list(zilog
);
2708 mutex_exit(&zilog
->zl_issuer_lock
);
2712 zil_commit_waiter_timeout(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2714 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2715 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2716 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
2718 lwb_t
*lwb
= zcw
->zcw_lwb
;
2719 ASSERT3P(lwb
, !=, NULL
);
2720 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_CLOSED
);
2723 * If the lwb has already been issued by another thread, we can
2724 * immediately return since there's no work to be done (the
2725 * point of this function is to issue the lwb). Additionally, we
2726 * do this prior to acquiring the zl_issuer_lock, to avoid
2727 * acquiring it when it's not necessary to do so.
2729 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2730 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2731 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
2735 * In order to call zil_lwb_write_issue() we must hold the
2736 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2737 * since we're already holding the commit waiter's "zcw_lock",
2738 * and those two locks are acquired in the opposite order
2741 mutex_exit(&zcw
->zcw_lock
);
2742 mutex_enter(&zilog
->zl_issuer_lock
);
2743 mutex_enter(&zcw
->zcw_lock
);
2746 * Since we just dropped and re-acquired the commit waiter's
2747 * lock, we have to re-check to see if the waiter was marked
2748 * "done" during that process. If the waiter was marked "done",
2749 * the "lwb" pointer is no longer valid (it can be free'd after
2750 * the waiter is marked "done"), so without this check we could
2751 * wind up with a use-after-free error below.
2756 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2759 * We've already checked this above, but since we hadn't acquired
2760 * the zilog's zl_issuer_lock, we have to perform this check a
2761 * second time while holding the lock.
2763 * We don't need to hold the zl_lock since the lwb cannot transition
2764 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2765 * _can_ transition from ISSUED to DONE, but it's OK to race with
2766 * that transition since we treat the lwb the same, whether it's in
2767 * the ISSUED or DONE states.
2769 * The important thing, is we treat the lwb differently depending on
2770 * if it's ISSUED or OPENED, and block any other threads that might
2771 * attempt to issue this lwb. For that reason we hold the
2772 * zl_issuer_lock when checking the lwb_state; we must not call
2773 * zil_lwb_write_issue() if the lwb had already been issued.
2775 * See the comment above the lwb_state_t structure definition for
2776 * more details on the lwb states, and locking requirements.
2778 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2779 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2780 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
2783 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
2786 * As described in the comments above zil_commit_waiter() and
2787 * zil_process_commit_list(), we need to issue this lwb's zio
2788 * since we've reached the commit waiter's timeout and it still
2789 * hasn't been issued.
2791 lwb_t
*nlwb
= zil_lwb_write_issue(zilog
, lwb
);
2793 IMPLY(nlwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_OPENED
);
2796 * Since the lwb's zio hadn't been issued by the time this thread
2797 * reached its timeout, we reset the zilog's "zl_cur_used" field
2798 * to influence the zil block size selection algorithm.
2800 * By having to issue the lwb's zio here, it means the size of the
2801 * lwb was too large, given the incoming throughput of itxs. By
2802 * setting "zl_cur_used" to zero, we communicate this fact to the
2803 * block size selection algorithm, so it can take this information
2804 * into account, and potentially select a smaller size for the
2805 * next lwb block that is allocated.
2807 zilog
->zl_cur_used
= 0;
2811 * When zil_lwb_write_issue() returns NULL, this
2812 * indicates zio_alloc_zil() failed to allocate the
2813 * "next" lwb on-disk. When this occurs, the ZIL write
2814 * pipeline must be stalled; see the comment within the
2815 * zil_commit_writer_stall() function for more details.
2817 * We must drop the commit waiter's lock prior to
2818 * calling zil_commit_writer_stall() or else we can wind
2819 * up with the following deadlock:
2821 * - This thread is waiting for the txg to sync while
2822 * holding the waiter's lock; txg_wait_synced() is
2823 * used within txg_commit_writer_stall().
2825 * - The txg can't sync because it is waiting for this
2826 * lwb's zio callback to call dmu_tx_commit().
2828 * - The lwb's zio callback can't call dmu_tx_commit()
2829 * because it's blocked trying to acquire the waiter's
2830 * lock, which occurs prior to calling dmu_tx_commit()
2832 mutex_exit(&zcw
->zcw_lock
);
2833 zil_commit_writer_stall(zilog
);
2834 mutex_enter(&zcw
->zcw_lock
);
2838 mutex_exit(&zilog
->zl_issuer_lock
);
2839 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2843 * This function is responsible for performing the following two tasks:
2845 * 1. its primary responsibility is to block until the given "commit
2846 * waiter" is considered "done".
2848 * 2. its secondary responsibility is to issue the zio for the lwb that
2849 * the given "commit waiter" is waiting on, if this function has
2850 * waited "long enough" and the lwb is still in the "open" state.
2852 * Given a sufficient amount of itxs being generated and written using
2853 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2854 * function. If this does not occur, this secondary responsibility will
2855 * ensure the lwb is issued even if there is not other synchronous
2856 * activity on the system.
2858 * For more details, see zil_process_commit_list(); more specifically,
2859 * the comment at the bottom of that function.
2862 zil_commit_waiter(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2864 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2865 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2866 ASSERT(spa_writeable(zilog
->zl_spa
));
2868 mutex_enter(&zcw
->zcw_lock
);
2871 * The timeout is scaled based on the lwb latency to avoid
2872 * significantly impacting the latency of each individual itx.
2873 * For more details, see the comment at the bottom of the
2874 * zil_process_commit_list() function.
2876 int pct
= MAX(zfs_commit_timeout_pct
, 1);
2877 hrtime_t sleep
= (zilog
->zl_last_lwb_latency
* pct
) / 100;
2878 hrtime_t wakeup
= gethrtime() + sleep
;
2879 boolean_t timedout
= B_FALSE
;
2881 while (!zcw
->zcw_done
) {
2882 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2884 lwb_t
*lwb
= zcw
->zcw_lwb
;
2887 * Usually, the waiter will have a non-NULL lwb field here,
2888 * but it's possible for it to be NULL as a result of
2889 * zil_commit() racing with spa_sync().
2891 * When zil_clean() is called, it's possible for the itxg
2892 * list (which may be cleaned via a taskq) to contain
2893 * commit itxs. When this occurs, the commit waiters linked
2894 * off of these commit itxs will not be committed to an
2895 * lwb. Additionally, these commit waiters will not be
2896 * marked done until zil_commit_waiter_skip() is called via
2899 * Thus, it's possible for this commit waiter (i.e. the
2900 * "zcw" variable) to be found in this "in between" state;
2901 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2902 * been skipped, so it's "zcw_done" field is still B_FALSE.
2904 IMPLY(lwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_CLOSED
);
2906 if (lwb
!= NULL
&& lwb
->lwb_state
== LWB_STATE_OPENED
) {
2907 ASSERT3B(timedout
, ==, B_FALSE
);
2910 * If the lwb hasn't been issued yet, then we
2911 * need to wait with a timeout, in case this
2912 * function needs to issue the lwb after the
2913 * timeout is reached; responsibility (2) from
2914 * the comment above this function.
2916 int rc
= cv_timedwait_hires(&zcw
->zcw_cv
,
2917 &zcw
->zcw_lock
, wakeup
, USEC2NSEC(1),
2918 CALLOUT_FLAG_ABSOLUTE
);
2920 if (rc
!= -1 || zcw
->zcw_done
)
2924 zil_commit_waiter_timeout(zilog
, zcw
);
2926 if (!zcw
->zcw_done
) {
2928 * If the commit waiter has already been
2929 * marked "done", it's possible for the
2930 * waiter's lwb structure to have already
2931 * been freed. Thus, we can only reliably
2932 * make these assertions if the waiter
2935 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2936 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_OPENED
);
2940 * If the lwb isn't open, then it must have already
2941 * been issued. In that case, there's no need to
2942 * use a timeout when waiting for the lwb to
2945 * Additionally, if the lwb is NULL, the waiter
2946 * will soon be signaled and marked done via
2947 * zil_clean() and zil_itxg_clean(), so no timeout
2952 lwb
->lwb_state
== LWB_STATE_ISSUED
||
2953 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2954 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
2955 cv_wait(&zcw
->zcw_cv
, &zcw
->zcw_lock
);
2959 mutex_exit(&zcw
->zcw_lock
);
2962 static zil_commit_waiter_t
*
2963 zil_alloc_commit_waiter(void)
2965 zil_commit_waiter_t
*zcw
= kmem_cache_alloc(zil_zcw_cache
, KM_SLEEP
);
2967 cv_init(&zcw
->zcw_cv
, NULL
, CV_DEFAULT
, NULL
);
2968 mutex_init(&zcw
->zcw_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2969 list_link_init(&zcw
->zcw_node
);
2970 zcw
->zcw_lwb
= NULL
;
2971 zcw
->zcw_done
= B_FALSE
;
2972 zcw
->zcw_zio_error
= 0;
2978 zil_free_commit_waiter(zil_commit_waiter_t
*zcw
)
2980 ASSERT(!list_link_active(&zcw
->zcw_node
));
2981 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
2982 ASSERT3B(zcw
->zcw_done
, ==, B_TRUE
);
2983 mutex_destroy(&zcw
->zcw_lock
);
2984 cv_destroy(&zcw
->zcw_cv
);
2985 kmem_cache_free(zil_zcw_cache
, zcw
);
2989 * This function is used to create a TX_COMMIT itx and assign it. This
2990 * way, it will be linked into the ZIL's list of synchronous itxs, and
2991 * then later committed to an lwb (or skipped) when
2992 * zil_process_commit_list() is called.
2995 zil_commit_itx_assign(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2997 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
2998 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
3000 itx_t
*itx
= zil_itx_create(TX_COMMIT
, sizeof (lr_t
));
3001 itx
->itx_sync
= B_TRUE
;
3002 itx
->itx_private
= zcw
;
3004 zil_itx_assign(zilog
, itx
, tx
);
3010 * Commit ZFS Intent Log transactions (itxs) to stable storage.
3012 * When writing ZIL transactions to the on-disk representation of the
3013 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3014 * itxs can be committed to a single lwb. Once a lwb is written and
3015 * committed to stable storage (i.e. the lwb is written, and vdevs have
3016 * been flushed), each itx that was committed to that lwb is also
3017 * considered to be committed to stable storage.
3019 * When an itx is committed to an lwb, the log record (lr_t) contained
3020 * by the itx is copied into the lwb's zio buffer, and once this buffer
3021 * is written to disk, it becomes an on-disk ZIL block.
3023 * As itxs are generated, they're inserted into the ZIL's queue of
3024 * uncommitted itxs. The semantics of zil_commit() are such that it will
3025 * block until all itxs that were in the queue when it was called, are
3026 * committed to stable storage.
3028 * If "foid" is zero, this means all "synchronous" and "asynchronous"
3029 * itxs, for all objects in the dataset, will be committed to stable
3030 * storage prior to zil_commit() returning. If "foid" is non-zero, all
3031 * "synchronous" itxs for all objects, but only "asynchronous" itxs
3032 * that correspond to the foid passed in, will be committed to stable
3033 * storage prior to zil_commit() returning.
3035 * Generally speaking, when zil_commit() is called, the consumer doesn't
3036 * actually care about _all_ of the uncommitted itxs. Instead, they're
3037 * simply trying to waiting for a specific itx to be committed to disk,
3038 * but the interface(s) for interacting with the ZIL don't allow such
3039 * fine-grained communication. A better interface would allow a consumer
3040 * to create and assign an itx, and then pass a reference to this itx to
3041 * zil_commit(); such that zil_commit() would return as soon as that
3042 * specific itx was committed to disk (instead of waiting for _all_
3043 * itxs to be committed).
3045 * When a thread calls zil_commit() a special "commit itx" will be
3046 * generated, along with a corresponding "waiter" for this commit itx.
3047 * zil_commit() will wait on this waiter's CV, such that when the waiter
3048 * is marked done, and signaled, zil_commit() will return.
3050 * This commit itx is inserted into the queue of uncommitted itxs. This
3051 * provides an easy mechanism for determining which itxs were in the
3052 * queue prior to zil_commit() having been called, and which itxs were
3053 * added after zil_commit() was called.
3055 * The commit itx is special; it doesn't have any on-disk representation.
3056 * When a commit itx is "committed" to an lwb, the waiter associated
3057 * with it is linked onto the lwb's list of waiters. Then, when that lwb
3058 * completes, each waiter on the lwb's list is marked done and signaled
3059 * -- allowing the thread waiting on the waiter to return from zil_commit().
3061 * It's important to point out a few critical factors that allow us
3062 * to make use of the commit itxs, commit waiters, per-lwb lists of
3063 * commit waiters, and zio completion callbacks like we're doing:
3065 * 1. The list of waiters for each lwb is traversed, and each commit
3066 * waiter is marked "done" and signaled, in the zio completion
3067 * callback of the lwb's zio[*].
3069 * * Actually, the waiters are signaled in the zio completion
3070 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3071 * that are sent to the vdevs upon completion of the lwb zio.
3073 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
3074 * itxs, the order in which they are inserted is preserved[*]; as
3075 * itxs are added to the queue, they are added to the tail of
3076 * in-memory linked lists.
3078 * When committing the itxs to lwbs (to be written to disk), they
3079 * are committed in the same order in which the itxs were added to
3080 * the uncommitted queue's linked list(s); i.e. the linked list of
3081 * itxs to commit is traversed from head to tail, and each itx is
3082 * committed to an lwb in that order.
3086 * - the order of "sync" itxs is preserved w.r.t. other
3087 * "sync" itxs, regardless of the corresponding objects.
3088 * - the order of "async" itxs is preserved w.r.t. other
3089 * "async" itxs corresponding to the same object.
3090 * - the order of "async" itxs is *not* preserved w.r.t. other
3091 * "async" itxs corresponding to different objects.
3092 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
3093 * versa) is *not* preserved, even for itxs that correspond
3094 * to the same object.
3096 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
3097 * zil_get_commit_list(), and zil_process_commit_list().
3099 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
3100 * lwb cannot be considered committed to stable storage, until its
3101 * "previous" lwb is also committed to stable storage. This fact,
3102 * coupled with the fact described above, means that itxs are
3103 * committed in (roughly) the order in which they were generated.
3104 * This is essential because itxs are dependent on prior itxs.
3105 * Thus, we *must not* deem an itx as being committed to stable
3106 * storage, until *all* prior itxs have also been committed to
3109 * To enforce this ordering of lwb zio's, while still leveraging as
3110 * much of the underlying storage performance as possible, we rely
3111 * on two fundamental concepts:
3113 * 1. The creation and issuance of lwb zio's is protected by
3114 * the zilog's "zl_issuer_lock", which ensures only a single
3115 * thread is creating and/or issuing lwb's at a time
3116 * 2. The "previous" lwb is a child of the "current" lwb
3117 * (leveraging the zio parent-child dependency graph)
3119 * By relying on this parent-child zio relationship, we can have
3120 * many lwb zio's concurrently issued to the underlying storage,
3121 * but the order in which they complete will be the same order in
3122 * which they were created.
3125 zil_commit(zilog_t
*zilog
, uint64_t foid
)
3128 * We should never attempt to call zil_commit on a snapshot for
3129 * a couple of reasons:
3131 * 1. A snapshot may never be modified, thus it cannot have any
3132 * in-flight itxs that would have modified the dataset.
3134 * 2. By design, when zil_commit() is called, a commit itx will
3135 * be assigned to this zilog; as a result, the zilog will be
3136 * dirtied. We must not dirty the zilog of a snapshot; there's
3137 * checks in the code that enforce this invariant, and will
3138 * cause a panic if it's not upheld.
3140 ASSERT3B(dmu_objset_is_snapshot(zilog
->zl_os
), ==, B_FALSE
);
3142 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3145 if (!spa_writeable(zilog
->zl_spa
)) {
3147 * If the SPA is not writable, there should never be any
3148 * pending itxs waiting to be committed to disk. If that
3149 * weren't true, we'd skip writing those itxs out, and
3150 * would break the semantics of zil_commit(); thus, we're
3151 * verifying that truth before we return to the caller.
3153 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3154 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3155 for (int i
= 0; i
< TXG_SIZE
; i
++)
3156 ASSERT3P(zilog
->zl_itxg
[i
].itxg_itxs
, ==, NULL
);
3161 * If the ZIL is suspended, we don't want to dirty it by calling
3162 * zil_commit_itx_assign() below, nor can we write out
3163 * lwbs like would be done in zil_commit_write(). Thus, we
3164 * simply rely on txg_wait_synced() to maintain the necessary
3165 * semantics, and avoid calling those functions altogether.
3167 if (zilog
->zl_suspend
> 0) {
3168 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3172 zil_commit_impl(zilog
, foid
);
3176 zil_commit_impl(zilog_t
*zilog
, uint64_t foid
)
3178 ZIL_STAT_BUMP(zilog
, zil_commit_count
);
3181 * Move the "async" itxs for the specified foid to the "sync"
3182 * queues, such that they will be later committed (or skipped)
3183 * to an lwb when zil_process_commit_list() is called.
3185 * Since these "async" itxs must be committed prior to this
3186 * call to zil_commit returning, we must perform this operation
3187 * before we call zil_commit_itx_assign().
3189 zil_async_to_sync(zilog
, foid
);
3192 * We allocate a new "waiter" structure which will initially be
3193 * linked to the commit itx using the itx's "itx_private" field.
3194 * Since the commit itx doesn't represent any on-disk state,
3195 * when it's committed to an lwb, rather than copying the its
3196 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3197 * added to the lwb's list of waiters. Then, when the lwb is
3198 * committed to stable storage, each waiter in the lwb's list of
3199 * waiters will be marked "done", and signalled.
3201 * We must create the waiter and assign the commit itx prior to
3202 * calling zil_commit_writer(), or else our specific commit itx
3203 * is not guaranteed to be committed to an lwb prior to calling
3204 * zil_commit_waiter().
3206 zil_commit_waiter_t
*zcw
= zil_alloc_commit_waiter();
3207 zil_commit_itx_assign(zilog
, zcw
);
3209 zil_commit_writer(zilog
, zcw
);
3210 zil_commit_waiter(zilog
, zcw
);
3212 if (zcw
->zcw_zio_error
!= 0) {
3214 * If there was an error writing out the ZIL blocks that
3215 * this thread is waiting on, then we fallback to
3216 * relying on spa_sync() to write out the data this
3217 * thread is waiting on. Obviously this has performance
3218 * implications, but the expectation is for this to be
3219 * an exceptional case, and shouldn't occur often.
3221 DTRACE_PROBE2(zil__commit__io__error
,
3222 zilog_t
*, zilog
, zil_commit_waiter_t
*, zcw
);
3223 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3226 zil_free_commit_waiter(zcw
);
3230 * Called in syncing context to free committed log blocks and update log header.
3233 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
3235 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
3236 uint64_t txg
= dmu_tx_get_txg(tx
);
3237 spa_t
*spa
= zilog
->zl_spa
;
3238 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
3242 * We don't zero out zl_destroy_txg, so make sure we don't try
3243 * to destroy it twice.
3245 if (spa_sync_pass(spa
) != 1)
3248 zil_lwb_flush_wait_all(zilog
, txg
);
3250 mutex_enter(&zilog
->zl_lock
);
3252 ASSERT(zilog
->zl_stop_sync
== 0);
3254 if (*replayed_seq
!= 0) {
3255 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
3256 zh
->zh_replay_seq
= *replayed_seq
;
3260 if (zilog
->zl_destroy_txg
== txg
) {
3261 blkptr_t blk
= zh
->zh_log
;
3262 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
3264 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
3266 memset(zh
, 0, sizeof (zil_header_t
));
3267 memset(zilog
->zl_replayed_seq
, 0,
3268 sizeof (zilog
->zl_replayed_seq
));
3270 if (zilog
->zl_keep_first
) {
3272 * If this block was part of log chain that couldn't
3273 * be claimed because a device was missing during
3274 * zil_claim(), but that device later returns,
3275 * then this block could erroneously appear valid.
3276 * To guard against this, assign a new GUID to the new
3277 * log chain so it doesn't matter what blk points to.
3279 zil_init_log_chain(zilog
, &blk
);
3283 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3284 * records. So, deactivate the feature for this dataset.
3285 * We activate it again when we start a new ZIL chain.
3287 if (dsl_dataset_feature_is_active(ds
,
3288 SPA_FEATURE_ZILSAXATTR
))
3289 dsl_dataset_deactivate_feature(ds
,
3290 SPA_FEATURE_ZILSAXATTR
, tx
);
3294 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
3295 zh
->zh_log
= lwb
->lwb_blk
;
3296 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
3298 list_remove(&zilog
->zl_lwb_list
, lwb
);
3299 zio_free(spa
, txg
, &lwb
->lwb_blk
);
3300 zil_free_lwb(zilog
, lwb
);
3303 * If we don't have anything left in the lwb list then
3304 * we've had an allocation failure and we need to zero
3305 * out the zil_header blkptr so that we don't end
3306 * up freeing the same block twice.
3308 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
3309 BP_ZERO(&zh
->zh_log
);
3313 * Remove fastwrite on any blocks that have been pre-allocated for
3314 * the next commit. This prevents fastwrite counter pollution by
3315 * unused, long-lived LWBs.
3317 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
3318 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_write_zio
) {
3319 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
3320 lwb
->lwb_fastwrite
= 0;
3324 mutex_exit(&zilog
->zl_lock
);
3328 zil_lwb_cons(void *vbuf
, void *unused
, int kmflag
)
3330 (void) unused
, (void) kmflag
;
3332 list_create(&lwb
->lwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
3333 list_create(&lwb
->lwb_waiters
, sizeof (zil_commit_waiter_t
),
3334 offsetof(zil_commit_waiter_t
, zcw_node
));
3335 avl_create(&lwb
->lwb_vdev_tree
, zil_lwb_vdev_compare
,
3336 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
3337 mutex_init(&lwb
->lwb_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3342 zil_lwb_dest(void *vbuf
, void *unused
)
3346 mutex_destroy(&lwb
->lwb_vdev_lock
);
3347 avl_destroy(&lwb
->lwb_vdev_tree
);
3348 list_destroy(&lwb
->lwb_waiters
);
3349 list_destroy(&lwb
->lwb_itxs
);
3355 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
3356 sizeof (lwb_t
), 0, zil_lwb_cons
, zil_lwb_dest
, NULL
, NULL
, NULL
, 0);
3358 zil_zcw_cache
= kmem_cache_create("zil_zcw_cache",
3359 sizeof (zil_commit_waiter_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
3361 zil_sums_init(&zil_sums_global
);
3362 zil_kstats_global
= kstat_create("zfs", 0, "zil", "misc",
3363 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
3364 KSTAT_FLAG_VIRTUAL
);
3366 if (zil_kstats_global
!= NULL
) {
3367 zil_kstats_global
->ks_data
= &zil_stats
;
3368 zil_kstats_global
->ks_update
= zil_kstats_global_update
;
3369 zil_kstats_global
->ks_private
= NULL
;
3370 kstat_install(zil_kstats_global
);
3377 kmem_cache_destroy(zil_zcw_cache
);
3378 kmem_cache_destroy(zil_lwb_cache
);
3380 if (zil_kstats_global
!= NULL
) {
3381 kstat_delete(zil_kstats_global
);
3382 zil_kstats_global
= NULL
;
3385 zil_sums_fini(&zil_sums_global
);
3389 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
3391 zilog
->zl_sync
= sync
;
3395 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
3397 zilog
->zl_logbias
= logbias
;
3401 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
3405 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
3407 zilog
->zl_header
= zh_phys
;
3409 zilog
->zl_spa
= dmu_objset_spa(os
);
3410 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
3411 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
3412 zilog
->zl_logbias
= dmu_objset_logbias(os
);
3413 zilog
->zl_sync
= dmu_objset_syncprop(os
);
3414 zilog
->zl_dirty_max_txg
= 0;
3415 zilog
->zl_last_lwb_opened
= NULL
;
3416 zilog
->zl_last_lwb_latency
= 0;
3417 zilog
->zl_max_block_size
= zil_maxblocksize
;
3419 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3420 mutex_init(&zilog
->zl_issuer_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3421 mutex_init(&zilog
->zl_lwb_io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3423 for (int i
= 0; i
< TXG_SIZE
; i
++) {
3424 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
3425 MUTEX_DEFAULT
, NULL
);
3428 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
3429 offsetof(lwb_t
, lwb_node
));
3431 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
3432 offsetof(itx_t
, itx_node
));
3434 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
3435 cv_init(&zilog
->zl_lwb_io_cv
, NULL
, CV_DEFAULT
, NULL
);
3441 zil_free(zilog_t
*zilog
)
3445 zilog
->zl_stop_sync
= 1;
3447 ASSERT0(zilog
->zl_suspend
);
3448 ASSERT0(zilog
->zl_suspending
);
3450 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3451 list_destroy(&zilog
->zl_lwb_list
);
3453 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
3454 list_destroy(&zilog
->zl_itx_commit_list
);
3456 for (i
= 0; i
< TXG_SIZE
; i
++) {
3458 * It's possible for an itx to be generated that doesn't dirty
3459 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3460 * callback to remove the entry. We remove those here.
3462 * Also free up the ziltest itxs.
3464 if (zilog
->zl_itxg
[i
].itxg_itxs
)
3465 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
3466 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
3469 mutex_destroy(&zilog
->zl_issuer_lock
);
3470 mutex_destroy(&zilog
->zl_lock
);
3471 mutex_destroy(&zilog
->zl_lwb_io_lock
);
3473 cv_destroy(&zilog
->zl_cv_suspend
);
3474 cv_destroy(&zilog
->zl_lwb_io_cv
);
3476 kmem_free(zilog
, sizeof (zilog_t
));
3480 * Open an intent log.
3483 zil_open(objset_t
*os
, zil_get_data_t
*get_data
, zil_sums_t
*zil_sums
)
3485 zilog_t
*zilog
= dmu_objset_zil(os
);
3487 ASSERT3P(zilog
->zl_get_data
, ==, NULL
);
3488 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3489 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3491 zilog
->zl_get_data
= get_data
;
3492 zilog
->zl_sums
= zil_sums
;
3498 * Close an intent log.
3501 zil_close(zilog_t
*zilog
)
3506 if (!dmu_objset_is_snapshot(zilog
->zl_os
)) {
3507 zil_commit(zilog
, 0);
3509 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
3510 ASSERT0(zilog
->zl_dirty_max_txg
);
3511 ASSERT3B(zilog_is_dirty(zilog
), ==, B_FALSE
);
3514 mutex_enter(&zilog
->zl_lock
);
3515 lwb
= list_tail(&zilog
->zl_lwb_list
);
3517 txg
= zilog
->zl_dirty_max_txg
;
3519 txg
= MAX(zilog
->zl_dirty_max_txg
, lwb
->lwb_max_txg
);
3520 mutex_exit(&zilog
->zl_lock
);
3523 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3524 * on the time when the dmu_tx transaction is assigned in
3525 * zil_lwb_write_issue().
3527 mutex_enter(&zilog
->zl_lwb_io_lock
);
3528 txg
= MAX(zilog
->zl_lwb_max_issued_txg
, txg
);
3529 mutex_exit(&zilog
->zl_lwb_io_lock
);
3532 * We need to use txg_wait_synced() to wait until that txg is synced.
3533 * zil_sync() will guarantee all lwbs up to that txg have been
3534 * written out, flushed, and cleaned.
3537 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
3539 if (zilog_is_dirty(zilog
))
3540 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog
,
3542 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
3543 VERIFY(!zilog_is_dirty(zilog
));
3545 zilog
->zl_get_data
= NULL
;
3548 * We should have only one lwb left on the list; remove it now.
3550 mutex_enter(&zilog
->zl_lock
);
3551 lwb
= list_head(&zilog
->zl_lwb_list
);
3553 ASSERT3P(lwb
, ==, list_tail(&zilog
->zl_lwb_list
));
3554 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
3556 if (lwb
->lwb_fastwrite
)
3557 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
3559 list_remove(&zilog
->zl_lwb_list
, lwb
);
3560 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
3561 zil_free_lwb(zilog
, lwb
);
3563 mutex_exit(&zilog
->zl_lock
);
3566 static const char *suspend_tag
= "zil suspending";
3569 * Suspend an intent log. While in suspended mode, we still honor
3570 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3571 * On old version pools, we suspend the log briefly when taking a
3572 * snapshot so that it will have an empty intent log.
3574 * Long holds are not really intended to be used the way we do here --
3575 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3576 * could fail. Therefore we take pains to only put a long hold if it is
3577 * actually necessary. Fortunately, it will only be necessary if the
3578 * objset is currently mounted (or the ZVOL equivalent). In that case it
3579 * will already have a long hold, so we are not really making things any worse.
3581 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3582 * zvol_state_t), and use their mechanism to prevent their hold from being
3583 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3586 * if cookiep == NULL, this does both the suspend & resume.
3587 * Otherwise, it returns with the dataset "long held", and the cookie
3588 * should be passed into zil_resume().
3591 zil_suspend(const char *osname
, void **cookiep
)
3595 const zil_header_t
*zh
;
3598 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
3601 zilog
= dmu_objset_zil(os
);
3603 mutex_enter(&zilog
->zl_lock
);
3604 zh
= zilog
->zl_header
;
3606 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
3607 mutex_exit(&zilog
->zl_lock
);
3608 dmu_objset_rele(os
, suspend_tag
);
3609 return (SET_ERROR(EBUSY
));
3613 * Don't put a long hold in the cases where we can avoid it. This
3614 * is when there is no cookie so we are doing a suspend & resume
3615 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3616 * for the suspend because it's already suspended, or there's no ZIL.
3618 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
3619 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
3620 mutex_exit(&zilog
->zl_lock
);
3621 dmu_objset_rele(os
, suspend_tag
);
3625 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
3626 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
3628 zilog
->zl_suspend
++;
3630 if (zilog
->zl_suspend
> 1) {
3632 * Someone else is already suspending it.
3633 * Just wait for them to finish.
3636 while (zilog
->zl_suspending
)
3637 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
3638 mutex_exit(&zilog
->zl_lock
);
3640 if (cookiep
== NULL
)
3648 * If there is no pointer to an on-disk block, this ZIL must not
3649 * be active (e.g. filesystem not mounted), so there's nothing
3652 if (BP_IS_HOLE(&zh
->zh_log
)) {
3653 ASSERT(cookiep
!= NULL
); /* fast path already handled */
3656 mutex_exit(&zilog
->zl_lock
);
3661 * The ZIL has work to do. Ensure that the associated encryption
3662 * key will remain mapped while we are committing the log by
3663 * grabbing a reference to it. If the key isn't loaded we have no
3664 * choice but to return an error until the wrapping key is loaded.
3666 if (os
->os_encrypted
&&
3667 dsl_dataset_create_key_mapping(dmu_objset_ds(os
)) != 0) {
3668 zilog
->zl_suspend
--;
3669 mutex_exit(&zilog
->zl_lock
);
3670 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3671 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3672 return (SET_ERROR(EACCES
));
3675 zilog
->zl_suspending
= B_TRUE
;
3676 mutex_exit(&zilog
->zl_lock
);
3679 * We need to use zil_commit_impl to ensure we wait for all
3680 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3681 * to disk before proceeding. If we used zil_commit instead, it
3682 * would just call txg_wait_synced(), because zl_suspend is set.
3683 * txg_wait_synced() doesn't wait for these lwb's to be
3684 * LWB_STATE_FLUSH_DONE before returning.
3686 zil_commit_impl(zilog
, 0);
3689 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3690 * use txg_wait_synced() to ensure the data from the zilog has
3691 * migrated to the main pool before calling zil_destroy().
3693 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3695 zil_destroy(zilog
, B_FALSE
);
3697 mutex_enter(&zilog
->zl_lock
);
3698 zilog
->zl_suspending
= B_FALSE
;
3699 cv_broadcast(&zilog
->zl_cv_suspend
);
3700 mutex_exit(&zilog
->zl_lock
);
3702 if (os
->os_encrypted
)
3703 dsl_dataset_remove_key_mapping(dmu_objset_ds(os
));
3705 if (cookiep
== NULL
)
3713 zil_resume(void *cookie
)
3715 objset_t
*os
= cookie
;
3716 zilog_t
*zilog
= dmu_objset_zil(os
);
3718 mutex_enter(&zilog
->zl_lock
);
3719 ASSERT(zilog
->zl_suspend
!= 0);
3720 zilog
->zl_suspend
--;
3721 mutex_exit(&zilog
->zl_lock
);
3722 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3723 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3726 typedef struct zil_replay_arg
{
3727 zil_replay_func_t
*const *zr_replay
;
3729 boolean_t zr_byteswap
;
3734 zil_replay_error(zilog_t
*zilog
, const lr_t
*lr
, int error
)
3736 char name
[ZFS_MAX_DATASET_NAME_LEN
];
3738 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
3740 dmu_objset_name(zilog
->zl_os
, name
);
3742 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
3743 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
3744 (u_longlong_t
)lr
->lrc_seq
,
3745 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
3746 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
3752 zil_replay_log_record(zilog_t
*zilog
, const lr_t
*lr
, void *zra
,
3755 zil_replay_arg_t
*zr
= zra
;
3756 const zil_header_t
*zh
= zilog
->zl_header
;
3757 uint64_t reclen
= lr
->lrc_reclen
;
3758 uint64_t txtype
= lr
->lrc_txtype
;
3761 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
3763 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
3766 if (lr
->lrc_txg
< claim_txg
) /* already committed */
3769 /* Strip case-insensitive bit, still present in log record */
3772 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
3773 return (zil_replay_error(zilog
, lr
, EINVAL
));
3776 * If this record type can be logged out of order, the object
3777 * (lr_foid) may no longer exist. That's legitimate, not an error.
3779 if (TX_OOO(txtype
)) {
3780 error
= dmu_object_info(zilog
->zl_os
,
3781 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
3782 if (error
== ENOENT
|| error
== EEXIST
)
3787 * Make a copy of the data so we can revise and extend it.
3789 memcpy(zr
->zr_lr
, lr
, reclen
);
3792 * If this is a TX_WRITE with a blkptr, suck in the data.
3794 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
3795 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
3796 zr
->zr_lr
+ reclen
);
3798 return (zil_replay_error(zilog
, lr
, error
));
3802 * The log block containing this lr may have been byteswapped
3803 * so that we can easily examine common fields like lrc_txtype.
3804 * However, the log is a mix of different record types, and only the
3805 * replay vectors know how to byteswap their records. Therefore, if
3806 * the lr was byteswapped, undo it before invoking the replay vector.
3808 if (zr
->zr_byteswap
)
3809 byteswap_uint64_array(zr
->zr_lr
, reclen
);
3812 * We must now do two things atomically: replay this log record,
3813 * and update the log header sequence number to reflect the fact that
3814 * we did so. At the end of each replay function the sequence number
3815 * is updated if we are in replay mode.
3817 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
3820 * The DMU's dnode layer doesn't see removes until the txg
3821 * commits, so a subsequent claim can spuriously fail with
3822 * EEXIST. So if we receive any error we try syncing out
3823 * any removes then retry the transaction. Note that we
3824 * specify B_FALSE for byteswap now, so we don't do it twice.
3826 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
3827 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
3829 return (zil_replay_error(zilog
, lr
, error
));
3835 zil_incr_blks(zilog_t
*zilog
, const blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
3837 (void) bp
, (void) arg
, (void) claim_txg
;
3839 zilog
->zl_replay_blks
++;
3845 * If this dataset has a non-empty intent log, replay it and destroy it.
3848 zil_replay(objset_t
*os
, void *arg
,
3849 zil_replay_func_t
*const replay_func
[TX_MAX_TYPE
])
3851 zilog_t
*zilog
= dmu_objset_zil(os
);
3852 const zil_header_t
*zh
= zilog
->zl_header
;
3853 zil_replay_arg_t zr
;
3855 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
3856 zil_destroy(zilog
, B_TRUE
);
3860 zr
.zr_replay
= replay_func
;
3862 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
3863 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
3866 * Wait for in-progress removes to sync before starting replay.
3868 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3870 zilog
->zl_replay
= B_TRUE
;
3871 zilog
->zl_replay_time
= ddi_get_lbolt();
3872 ASSERT(zilog
->zl_replay_blks
== 0);
3873 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
3874 zh
->zh_claim_txg
, B_TRUE
);
3875 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
3877 zil_destroy(zilog
, B_FALSE
);
3878 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
3879 zilog
->zl_replay
= B_FALSE
;
3883 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
3885 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3888 if (zilog
->zl_replay
) {
3889 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
3890 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
3891 zilog
->zl_replaying_seq
;
3899 zil_reset(const char *osname
, void *arg
)
3903 int error
= zil_suspend(osname
, NULL
);
3904 /* EACCES means crypto key not loaded */
3905 if ((error
== EACCES
) || (error
== EBUSY
))
3906 return (SET_ERROR(error
));
3908 return (SET_ERROR(EEXIST
));
3912 EXPORT_SYMBOL(zil_alloc
);
3913 EXPORT_SYMBOL(zil_free
);
3914 EXPORT_SYMBOL(zil_open
);
3915 EXPORT_SYMBOL(zil_close
);
3916 EXPORT_SYMBOL(zil_replay
);
3917 EXPORT_SYMBOL(zil_replaying
);
3918 EXPORT_SYMBOL(zil_destroy
);
3919 EXPORT_SYMBOL(zil_destroy_sync
);
3920 EXPORT_SYMBOL(zil_itx_create
);
3921 EXPORT_SYMBOL(zil_itx_destroy
);
3922 EXPORT_SYMBOL(zil_itx_assign
);
3923 EXPORT_SYMBOL(zil_commit
);
3924 EXPORT_SYMBOL(zil_claim
);
3925 EXPORT_SYMBOL(zil_check_log_chain
);
3926 EXPORT_SYMBOL(zil_sync
);
3927 EXPORT_SYMBOL(zil_clean
);
3928 EXPORT_SYMBOL(zil_suspend
);
3929 EXPORT_SYMBOL(zil_resume
);
3930 EXPORT_SYMBOL(zil_lwb_add_block
);
3931 EXPORT_SYMBOL(zil_bp_tree_add
);
3932 EXPORT_SYMBOL(zil_set_sync
);
3933 EXPORT_SYMBOL(zil_set_logbias
);
3934 EXPORT_SYMBOL(zil_sums_init
);
3935 EXPORT_SYMBOL(zil_sums_fini
);
3936 EXPORT_SYMBOL(zil_kstat_values_update
);
3938 ZFS_MODULE_PARAM(zfs
, zfs_
, commit_timeout_pct
, UINT
, ZMOD_RW
,
3939 "ZIL block open timeout percentage");
3941 ZFS_MODULE_PARAM(zfs_zil
, zil_
, replay_disable
, INT
, ZMOD_RW
,
3942 "Disable intent logging replay");
3944 ZFS_MODULE_PARAM(zfs_zil
, zil_
, nocacheflush
, INT
, ZMOD_RW
,
3945 "Disable ZIL cache flushes");
3947 ZFS_MODULE_PARAM(zfs_zil
, zil_
, slog_bulk
, U64
, ZMOD_RW
,
3948 "Limit in bytes slog sync writes per commit");
3950 ZFS_MODULE_PARAM(zfs_zil
, zil_
, maxblocksize
, UINT
, ZMOD_RW
,
3951 "Limit in bytes of ZIL log block size");