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>
47 #include <sys/wmsum.h>
50 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
51 * calls that change the file system. Each itx has enough information to
52 * be able to replay them after a system crash, power loss, or
53 * equivalent failure mode. These are stored in memory until either:
55 * 1. they are committed to the pool by the DMU transaction group
56 * (txg), at which point they can be discarded; or
57 * 2. they are committed to the on-disk ZIL for the dataset being
58 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
61 * In the event of a crash or power loss, the itxs contained by each
62 * dataset's on-disk ZIL will be replayed when that dataset is first
63 * instantiated (e.g. if the dataset is a normal filesystem, when it is
66 * As hinted at above, there is one ZIL per dataset (both the in-memory
67 * representation, and the on-disk representation). The on-disk format
68 * consists of 3 parts:
70 * - a single, per-dataset, ZIL header; which points to a chain of
71 * - zero or more ZIL blocks; each of which contains
72 * - zero or more ZIL records
74 * A ZIL record holds the information necessary to replay a single
75 * system call transaction. A ZIL block can hold many ZIL records, and
76 * the blocks are chained together, similarly to a singly linked list.
78 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
79 * block in the chain, and the ZIL header points to the first block in
82 * Note, there is not a fixed place in the pool to hold these ZIL
83 * blocks; they are dynamically allocated and freed as needed from the
84 * blocks available on the pool, though they can be preferentially
85 * allocated from a dedicated "log" vdev.
89 * This controls the amount of time that a ZIL block (lwb) will remain
90 * "open" when it isn't "full", and it has a thread waiting for it to be
91 * committed to stable storage. Please refer to the zil_commit_waiter()
92 * function (and the comments within it) for more details.
94 static uint_t zfs_commit_timeout_pct
= 10;
97 * See zil.h for more information about these fields.
99 static zil_kstat_values_t zil_stats
= {
100 { "zil_commit_count", KSTAT_DATA_UINT64
},
101 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
102 { "zil_itx_count", KSTAT_DATA_UINT64
},
103 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
104 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
105 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
106 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
107 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
108 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
109 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
110 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
111 { "zil_itx_metaslab_normal_write", KSTAT_DATA_UINT64
},
112 { "zil_itx_metaslab_normal_alloc", KSTAT_DATA_UINT64
},
113 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
114 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
115 { "zil_itx_metaslab_slog_write", KSTAT_DATA_UINT64
},
116 { "zil_itx_metaslab_slog_alloc", KSTAT_DATA_UINT64
},
119 static zil_sums_t zil_sums_global
;
120 static kstat_t
*zil_kstats_global
;
123 * Disable intent logging replay. This global ZIL switch affects all pools.
125 int zil_replay_disable
= 0;
128 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
129 * the disk(s) by the ZIL after an LWB write has completed. Setting this
130 * will cause ZIL corruption on power loss if a volatile out-of-order
131 * write cache is enabled.
133 static int zil_nocacheflush
= 0;
136 * Limit SLOG write size per commit executed with synchronous priority.
137 * Any writes above that will be executed with lower (asynchronous) priority
138 * to limit potential SLOG device abuse by single active ZIL writer.
140 static uint64_t zil_slog_bulk
= 64 * 1024 * 1024;
142 static kmem_cache_t
*zil_lwb_cache
;
143 static kmem_cache_t
*zil_zcw_cache
;
145 static void zil_lwb_commit(zilog_t
*zilog
, lwb_t
*lwb
, itx_t
*itx
);
146 static itx_t
*zil_itx_clone(itx_t
*oitx
);
147 static uint64_t zil_max_waste_space(zilog_t
*zilog
);
150 zil_bp_compare(const void *x1
, const void *x2
)
152 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
153 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
155 int cmp
= TREE_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
159 return (TREE_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
163 zil_bp_tree_init(zilog_t
*zilog
)
165 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
166 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
170 zil_bp_tree_fini(zilog_t
*zilog
)
172 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
176 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
177 kmem_free(zn
, sizeof (zil_bp_node_t
));
183 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
185 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
190 if (BP_IS_EMBEDDED(bp
))
193 dva
= BP_IDENTITY(bp
);
195 if (avl_find(t
, dva
, &where
) != NULL
)
196 return (SET_ERROR(EEXIST
));
198 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
200 avl_insert(t
, zn
, where
);
205 static zil_header_t
*
206 zil_header_in_syncing_context(zilog_t
*zilog
)
208 return ((zil_header_t
*)zilog
->zl_header
);
212 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
214 zio_cksum_t
*zc
= &bp
->blk_cksum
;
216 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_0
],
217 sizeof (zc
->zc_word
[ZIL_ZC_GUID_0
]));
218 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_1
],
219 sizeof (zc
->zc_word
[ZIL_ZC_GUID_1
]));
220 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
221 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
225 zil_kstats_global_update(kstat_t
*ksp
, int rw
)
227 zil_kstat_values_t
*zs
= ksp
->ks_data
;
228 ASSERT3P(&zil_stats
, ==, zs
);
230 if (rw
== KSTAT_WRITE
) {
231 return (SET_ERROR(EACCES
));
234 zil_kstat_values_update(zs
, &zil_sums_global
);
240 * Read a log block and make sure it's valid.
243 zil_read_log_block(zilog_t
*zilog
, boolean_t decrypt
, const blkptr_t
*bp
,
244 blkptr_t
*nbp
, char **begin
, char **end
, arc_buf_t
**abuf
)
246 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
247 arc_flags_t aflags
= ARC_FLAG_WAIT
;
251 if (zilog
->zl_header
->zh_claim_txg
== 0)
252 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
254 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
255 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
258 zio_flags
|= ZIO_FLAG_RAW
;
260 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
261 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
263 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
,
264 abuf
, ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
267 zio_cksum_t cksum
= bp
->blk_cksum
;
270 * Validate the checksummed log block.
272 * Sequence numbers should be... sequential. The checksum
273 * verifier for the next block should be bp's checksum plus 1.
275 * Also check the log chain linkage and size used.
277 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
279 uint64_t size
= BP_GET_LSIZE(bp
);
280 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
281 zil_chain_t
*zilc
= (*abuf
)->b_data
;
282 char *lr
= (char *)(zilc
+ 1);
284 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
286 zilc
->zc_nused
< sizeof (*zilc
) ||
287 zilc
->zc_nused
> size
) {
288 error
= SET_ERROR(ECKSUM
);
291 *end
= lr
+ zilc
->zc_nused
- sizeof (*zilc
);
292 *nbp
= zilc
->zc_next_blk
;
295 char *lr
= (*abuf
)->b_data
;
296 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
298 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
300 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
301 error
= SET_ERROR(ECKSUM
);
304 *end
= lr
+ zilc
->zc_nused
;
305 *nbp
= zilc
->zc_next_blk
;
314 * Read a TX_WRITE log data block.
317 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
319 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
320 const blkptr_t
*bp
= &lr
->lr_blkptr
;
321 arc_flags_t aflags
= ARC_FLAG_WAIT
;
322 arc_buf_t
*abuf
= NULL
;
326 if (BP_IS_HOLE(bp
)) {
328 memset(wbuf
, 0, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
332 if (zilog
->zl_header
->zh_claim_txg
== 0)
333 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
336 * If we are not using the resulting data, we are just checking that
337 * it hasn't been corrupted so we don't need to waste CPU time
338 * decompressing and decrypting it.
341 zio_flags
|= ZIO_FLAG_RAW
;
343 ASSERT3U(BP_GET_LSIZE(bp
), !=, 0);
344 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
345 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
347 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
348 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
352 memcpy(wbuf
, abuf
->b_data
, arc_buf_size(abuf
));
353 arc_buf_destroy(abuf
, &abuf
);
360 zil_sums_init(zil_sums_t
*zs
)
362 wmsum_init(&zs
->zil_commit_count
, 0);
363 wmsum_init(&zs
->zil_commit_writer_count
, 0);
364 wmsum_init(&zs
->zil_itx_count
, 0);
365 wmsum_init(&zs
->zil_itx_indirect_count
, 0);
366 wmsum_init(&zs
->zil_itx_indirect_bytes
, 0);
367 wmsum_init(&zs
->zil_itx_copied_count
, 0);
368 wmsum_init(&zs
->zil_itx_copied_bytes
, 0);
369 wmsum_init(&zs
->zil_itx_needcopy_count
, 0);
370 wmsum_init(&zs
->zil_itx_needcopy_bytes
, 0);
371 wmsum_init(&zs
->zil_itx_metaslab_normal_count
, 0);
372 wmsum_init(&zs
->zil_itx_metaslab_normal_bytes
, 0);
373 wmsum_init(&zs
->zil_itx_metaslab_normal_write
, 0);
374 wmsum_init(&zs
->zil_itx_metaslab_normal_alloc
, 0);
375 wmsum_init(&zs
->zil_itx_metaslab_slog_count
, 0);
376 wmsum_init(&zs
->zil_itx_metaslab_slog_bytes
, 0);
377 wmsum_init(&zs
->zil_itx_metaslab_slog_write
, 0);
378 wmsum_init(&zs
->zil_itx_metaslab_slog_alloc
, 0);
382 zil_sums_fini(zil_sums_t
*zs
)
384 wmsum_fini(&zs
->zil_commit_count
);
385 wmsum_fini(&zs
->zil_commit_writer_count
);
386 wmsum_fini(&zs
->zil_itx_count
);
387 wmsum_fini(&zs
->zil_itx_indirect_count
);
388 wmsum_fini(&zs
->zil_itx_indirect_bytes
);
389 wmsum_fini(&zs
->zil_itx_copied_count
);
390 wmsum_fini(&zs
->zil_itx_copied_bytes
);
391 wmsum_fini(&zs
->zil_itx_needcopy_count
);
392 wmsum_fini(&zs
->zil_itx_needcopy_bytes
);
393 wmsum_fini(&zs
->zil_itx_metaslab_normal_count
);
394 wmsum_fini(&zs
->zil_itx_metaslab_normal_bytes
);
395 wmsum_fini(&zs
->zil_itx_metaslab_normal_write
);
396 wmsum_fini(&zs
->zil_itx_metaslab_normal_alloc
);
397 wmsum_fini(&zs
->zil_itx_metaslab_slog_count
);
398 wmsum_fini(&zs
->zil_itx_metaslab_slog_bytes
);
399 wmsum_fini(&zs
->zil_itx_metaslab_slog_write
);
400 wmsum_fini(&zs
->zil_itx_metaslab_slog_alloc
);
404 zil_kstat_values_update(zil_kstat_values_t
*zs
, zil_sums_t
*zil_sums
)
406 zs
->zil_commit_count
.value
.ui64
=
407 wmsum_value(&zil_sums
->zil_commit_count
);
408 zs
->zil_commit_writer_count
.value
.ui64
=
409 wmsum_value(&zil_sums
->zil_commit_writer_count
);
410 zs
->zil_itx_count
.value
.ui64
=
411 wmsum_value(&zil_sums
->zil_itx_count
);
412 zs
->zil_itx_indirect_count
.value
.ui64
=
413 wmsum_value(&zil_sums
->zil_itx_indirect_count
);
414 zs
->zil_itx_indirect_bytes
.value
.ui64
=
415 wmsum_value(&zil_sums
->zil_itx_indirect_bytes
);
416 zs
->zil_itx_copied_count
.value
.ui64
=
417 wmsum_value(&zil_sums
->zil_itx_copied_count
);
418 zs
->zil_itx_copied_bytes
.value
.ui64
=
419 wmsum_value(&zil_sums
->zil_itx_copied_bytes
);
420 zs
->zil_itx_needcopy_count
.value
.ui64
=
421 wmsum_value(&zil_sums
->zil_itx_needcopy_count
);
422 zs
->zil_itx_needcopy_bytes
.value
.ui64
=
423 wmsum_value(&zil_sums
->zil_itx_needcopy_bytes
);
424 zs
->zil_itx_metaslab_normal_count
.value
.ui64
=
425 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_count
);
426 zs
->zil_itx_metaslab_normal_bytes
.value
.ui64
=
427 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_bytes
);
428 zs
->zil_itx_metaslab_normal_write
.value
.ui64
=
429 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_write
);
430 zs
->zil_itx_metaslab_normal_alloc
.value
.ui64
=
431 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_alloc
);
432 zs
->zil_itx_metaslab_slog_count
.value
.ui64
=
433 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_count
);
434 zs
->zil_itx_metaslab_slog_bytes
.value
.ui64
=
435 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_bytes
);
436 zs
->zil_itx_metaslab_slog_write
.value
.ui64
=
437 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_write
);
438 zs
->zil_itx_metaslab_slog_alloc
.value
.ui64
=
439 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_alloc
);
443 * Parse the intent log, and call parse_func for each valid record within.
446 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
447 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
,
450 const zil_header_t
*zh
= zilog
->zl_header
;
451 boolean_t claimed
= !!zh
->zh_claim_txg
;
452 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
453 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
454 uint64_t max_blk_seq
= 0;
455 uint64_t max_lr_seq
= 0;
456 uint64_t blk_count
= 0;
457 uint64_t lr_count
= 0;
458 blkptr_t blk
, next_blk
= {{{{0}}}};
462 * Old logs didn't record the maximum zh_claim_lr_seq.
464 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
465 claim_lr_seq
= UINT64_MAX
;
468 * Starting at the block pointed to by zh_log we read the log chain.
469 * For each block in the chain we strongly check that block to
470 * ensure its validity. We stop when an invalid block is found.
471 * For each block pointer in the chain we call parse_blk_func().
472 * For each record in each valid block we call parse_lr_func().
473 * If the log has been claimed, stop if we encounter a sequence
474 * number greater than the highest claimed sequence number.
476 zil_bp_tree_init(zilog
);
478 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
479 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
482 arc_buf_t
*abuf
= NULL
;
484 if (blk_seq
> claim_blk_seq
)
487 error
= parse_blk_func(zilog
, &blk
, arg
, txg
);
490 ASSERT3U(max_blk_seq
, <, blk_seq
);
491 max_blk_seq
= blk_seq
;
494 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
497 error
= zil_read_log_block(zilog
, decrypt
, &blk
, &next_blk
,
501 arc_buf_destroy(abuf
, &abuf
);
503 char name
[ZFS_MAX_DATASET_NAME_LEN
];
505 dmu_objset_name(zilog
->zl_os
, name
);
507 cmn_err(CE_WARN
, "ZFS read log block error %d, "
508 "dataset %s, seq 0x%llx\n", error
, name
,
509 (u_longlong_t
)blk_seq
);
514 for (; lrp
< end
; lrp
+= reclen
) {
515 lr_t
*lr
= (lr_t
*)lrp
;
516 reclen
= lr
->lrc_reclen
;
517 ASSERT3U(reclen
, >=, sizeof (lr_t
));
518 ASSERT3U(reclen
, <=, end
- lrp
);
519 if (lr
->lrc_seq
> claim_lr_seq
) {
520 arc_buf_destroy(abuf
, &abuf
);
524 error
= parse_lr_func(zilog
, lr
, arg
, txg
);
526 arc_buf_destroy(abuf
, &abuf
);
529 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
530 max_lr_seq
= lr
->lrc_seq
;
533 arc_buf_destroy(abuf
, &abuf
);
536 zilog
->zl_parse_error
= error
;
537 zilog
->zl_parse_blk_seq
= max_blk_seq
;
538 zilog
->zl_parse_lr_seq
= max_lr_seq
;
539 zilog
->zl_parse_blk_count
= blk_count
;
540 zilog
->zl_parse_lr_count
= lr_count
;
542 zil_bp_tree_fini(zilog
);
548 zil_clear_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
552 ASSERT(!BP_IS_HOLE(bp
));
555 * As we call this function from the context of a rewind to a
556 * checkpoint, each ZIL block whose txg is later than the txg
557 * that we rewind to is invalid. Thus, we return -1 so
558 * zil_parse() doesn't attempt to read it.
560 if (bp
->blk_birth
>= first_txg
)
563 if (zil_bp_tree_add(zilog
, bp
) != 0)
566 zio_free(zilog
->zl_spa
, first_txg
, bp
);
571 zil_noop_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
574 (void) zilog
, (void) lrc
, (void) tx
, (void) first_txg
;
579 zil_claim_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
583 * Claim log block if not already committed and not already claimed.
584 * If tx == NULL, just verify that the block is claimable.
586 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
587 zil_bp_tree_add(zilog
, bp
) != 0)
590 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
591 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
592 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
596 zil_claim_write(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
, uint64_t first_txg
)
598 lr_write_t
*lr
= (lr_write_t
*)lrc
;
601 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
604 * If the block is not readable, don't claim it. This can happen
605 * in normal operation when a log block is written to disk before
606 * some of the dmu_sync() blocks it points to. In this case, the
607 * transaction cannot have been committed to anyone (we would have
608 * waited for all writes to be stable first), so it is semantically
609 * correct to declare this the end of the log.
611 if (lr
->lr_blkptr
.blk_birth
>= first_txg
) {
612 error
= zil_read_log_data(zilog
, lr
, NULL
);
617 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
621 zil_claim_clone_range(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
624 const lr_clone_range_t
*lr
= (const lr_clone_range_t
*)lrc
;
626 spa_t
*spa
= zilog
->zl_spa
;
629 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
630 ASSERT3U(lrc
->lrc_reclen
, >=, offsetof(lr_clone_range_t
,
631 lr_bps
[lr
->lr_nbps
]));
638 * XXX: Do we need to byteswap lr?
641 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
642 bp
= &lr
->lr_bps
[ii
];
645 * When data is embedded into the BP there is no need to create
646 * BRT entry as there is no data block. Just copy the BP as it
649 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
653 * We can not handle block pointers from the future, since they
654 * are not yet allocated. It should not normally happen, but
655 * just in case lets be safe and just stop here now instead of
656 * corrupting the pool.
658 if (BP_PHYSICAL_BIRTH(bp
) >= first_txg
)
659 return (SET_ERROR(ENOENT
));
662 * Assert the block is really allocated before we reference it.
664 metaslab_check_free(spa
, bp
);
667 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
668 bp
= &lr
->lr_bps
[ii
];
669 if (!BP_IS_HOLE(bp
) && !BP_IS_EMBEDDED(bp
))
670 brt_pending_add(spa
, bp
, tx
);
677 zil_claim_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
681 switch (lrc
->lrc_txtype
) {
683 return (zil_claim_write(zilog
, lrc
, tx
, first_txg
));
685 return (zil_claim_clone_range(zilog
, lrc
, tx
, first_txg
));
692 zil_free_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
697 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
703 zil_free_write(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
705 lr_write_t
*lr
= (lr_write_t
*)lrc
;
706 blkptr_t
*bp
= &lr
->lr_blkptr
;
708 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
711 * If we previously claimed it, we need to free it.
713 if (bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
715 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
722 zil_free_clone_range(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
)
724 const lr_clone_range_t
*lr
= (const lr_clone_range_t
*)lrc
;
729 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
730 ASSERT3U(lrc
->lrc_reclen
, >=, offsetof(lr_clone_range_t
,
731 lr_bps
[lr
->lr_nbps
]));
739 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
740 bp
= &lr
->lr_bps
[ii
];
742 if (!BP_IS_HOLE(bp
)) {
743 zio_free(spa
, dmu_tx_get_txg(tx
), bp
);
751 zil_free_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
755 if (claim_txg
== 0) {
759 switch (lrc
->lrc_txtype
) {
761 return (zil_free_write(zilog
, lrc
, tx
, claim_txg
));
763 return (zil_free_clone_range(zilog
, lrc
, tx
));
770 zil_lwb_vdev_compare(const void *x1
, const void *x2
)
772 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
773 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
775 return (TREE_CMP(v1
, v2
));
779 * Allocate a new lwb. We may already have a block pointer for it, in which
780 * case we get size and version from there. Or we may not yet, in which case
781 * we choose them here and later make the block allocation match.
784 zil_alloc_lwb(zilog_t
*zilog
, int sz
, blkptr_t
*bp
, boolean_t slog
,
785 uint64_t txg
, lwb_state_t state
)
789 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
790 lwb
->lwb_zilog
= zilog
;
793 lwb
->lwb_slim
= (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
);
794 sz
= BP_GET_LSIZE(bp
);
796 BP_ZERO(&lwb
->lwb_blk
);
797 lwb
->lwb_slim
= (spa_version(zilog
->zl_spa
) >=
798 SPA_VERSION_SLIM_ZIL
);
800 lwb
->lwb_slog
= slog
;
804 lwb
->lwb_nused
= lwb
->lwb_nfilled
= sizeof (zil_chain_t
);
806 lwb
->lwb_nmax
= sz
- sizeof (zil_chain_t
);
807 lwb
->lwb_nused
= lwb
->lwb_nfilled
= 0;
810 lwb
->lwb_state
= state
;
811 lwb
->lwb_buf
= zio_buf_alloc(sz
);
812 lwb
->lwb_child_zio
= NULL
;
813 lwb
->lwb_write_zio
= NULL
;
814 lwb
->lwb_root_zio
= NULL
;
815 lwb
->lwb_issued_timestamp
= 0;
816 lwb
->lwb_issued_txg
= 0;
817 lwb
->lwb_alloc_txg
= txg
;
818 lwb
->lwb_max_txg
= 0;
820 mutex_enter(&zilog
->zl_lock
);
821 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
822 if (state
!= LWB_STATE_NEW
)
823 zilog
->zl_last_lwb_opened
= lwb
;
824 mutex_exit(&zilog
->zl_lock
);
830 zil_free_lwb(zilog_t
*zilog
, lwb_t
*lwb
)
832 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
833 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
834 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
835 ASSERT3P(lwb
->lwb_child_zio
, ==, NULL
);
836 ASSERT3P(lwb
->lwb_write_zio
, ==, NULL
);
837 ASSERT3P(lwb
->lwb_root_zio
, ==, NULL
);
838 ASSERT3U(lwb
->lwb_alloc_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
839 ASSERT3U(lwb
->lwb_max_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
840 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
841 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
842 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
843 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
846 * Clear the zilog's field to indicate this lwb is no longer
847 * valid, and prevent use-after-free errors.
849 if (zilog
->zl_last_lwb_opened
== lwb
)
850 zilog
->zl_last_lwb_opened
= NULL
;
852 kmem_cache_free(zil_lwb_cache
, lwb
);
856 * Called when we create in-memory log transactions so that we know
857 * to cleanup the itxs at the end of spa_sync().
860 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
862 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
863 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
865 ASSERT(spa_writeable(zilog
->zl_spa
));
867 if (ds
->ds_is_snapshot
)
868 panic("dirtying snapshot!");
870 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
871 /* up the hold count until we can be written out */
872 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
874 zilog
->zl_dirty_max_txg
= MAX(txg
, zilog
->zl_dirty_max_txg
);
879 * Determine if the zil is dirty in the specified txg. Callers wanting to
880 * ensure that the dirty state does not change must hold the itxg_lock for
881 * the specified txg. Holding the lock will ensure that the zil cannot be
882 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
885 static boolean_t __maybe_unused
886 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
888 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
890 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
896 * Determine if the zil is dirty. The zil is considered dirty if it has
897 * any pending itx records that have not been cleaned by zil_clean().
900 zilog_is_dirty(zilog_t
*zilog
)
902 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
904 for (int t
= 0; t
< TXG_SIZE
; t
++) {
905 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
912 * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
913 * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
914 * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
918 zil_commit_activate_saxattr_feature(zilog_t
*zilog
)
920 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
924 if (spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
925 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
&&
926 !dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
)) {
927 tx
= dmu_tx_create(zilog
->zl_os
);
928 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
929 dsl_dataset_dirty(ds
, tx
);
930 txg
= dmu_tx_get_txg(tx
);
932 mutex_enter(&ds
->ds_lock
);
933 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
935 mutex_exit(&ds
->ds_lock
);
937 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
942 * Create an on-disk intent log.
945 zil_create(zilog_t
*zilog
)
947 const zil_header_t
*zh
= zilog
->zl_header
;
953 boolean_t slog
= FALSE
;
954 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
958 * Wait for any previous destroy to complete.
960 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
962 ASSERT(zh
->zh_claim_txg
== 0);
963 ASSERT(zh
->zh_replay_seq
== 0);
968 * Allocate an initial log block if:
969 * - there isn't one already
970 * - the existing block is the wrong endianness
972 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
973 tx
= dmu_tx_create(zilog
->zl_os
);
974 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
975 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
976 txg
= dmu_tx_get_txg(tx
);
978 if (!BP_IS_HOLE(&blk
)) {
979 zio_free(zilog
->zl_spa
, txg
, &blk
);
983 error
= zio_alloc_zil(zilog
->zl_spa
, zilog
->zl_os
, txg
, &blk
,
984 ZIL_MIN_BLKSZ
, &slog
);
986 zil_init_log_chain(zilog
, &blk
);
990 * Allocate a log write block (lwb) for the first log block.
993 lwb
= zil_alloc_lwb(zilog
, 0, &blk
, slog
, txg
, LWB_STATE_NEW
);
996 * If we just allocated the first log block, commit our transaction
997 * and wait for zil_sync() to stuff the block pointer into zh_log.
998 * (zh is part of the MOS, so we cannot modify it in open context.)
1002 * If "zilsaxattr" feature is enabled on zpool, then activate
1003 * it now when we're creating the ZIL chain. We can't wait with
1004 * this until we write the first xattr log record because we
1005 * need to wait for the feature activation to sync out.
1007 if (spa_feature_is_enabled(zilog
->zl_spa
,
1008 SPA_FEATURE_ZILSAXATTR
) && dmu_objset_type(zilog
->zl_os
) !=
1010 mutex_enter(&ds
->ds_lock
);
1011 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
1013 mutex_exit(&ds
->ds_lock
);
1017 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1020 * This branch covers the case where we enable the feature on a
1021 * zpool that has existing ZIL headers.
1023 zil_commit_activate_saxattr_feature(zilog
);
1025 IMPLY(spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
1026 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
,
1027 dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
));
1029 ASSERT(error
!= 0 || memcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
1030 IMPLY(error
== 0, lwb
!= NULL
);
1036 * In one tx, free all log blocks and clear the log header. If keep_first
1037 * is set, then we're replaying a log with no content. We want to keep the
1038 * first block, however, so that the first synchronous transaction doesn't
1039 * require a txg_wait_synced() in zil_create(). We don't need to
1040 * txg_wait_synced() here either when keep_first is set, because both
1041 * zil_create() and zil_destroy() will wait for any in-progress destroys
1043 * Return B_TRUE if there were any entries to replay.
1046 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
1048 const zil_header_t
*zh
= zilog
->zl_header
;
1054 * Wait for any previous destroy to complete.
1056 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
1058 zilog
->zl_old_header
= *zh
; /* debugging aid */
1060 if (BP_IS_HOLE(&zh
->zh_log
))
1063 tx
= dmu_tx_create(zilog
->zl_os
);
1064 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1065 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1066 txg
= dmu_tx_get_txg(tx
);
1068 mutex_enter(&zilog
->zl_lock
);
1070 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
1071 zilog
->zl_destroy_txg
= txg
;
1072 zilog
->zl_keep_first
= keep_first
;
1074 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
1075 ASSERT(zh
->zh_claim_txg
== 0);
1076 VERIFY(!keep_first
);
1077 while ((lwb
= list_remove_head(&zilog
->zl_lwb_list
)) != NULL
) {
1078 if (lwb
->lwb_buf
!= NULL
)
1079 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1080 if (!BP_IS_HOLE(&lwb
->lwb_blk
))
1081 zio_free(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
1082 zil_free_lwb(zilog
, lwb
);
1084 } else if (!keep_first
) {
1085 zil_destroy_sync(zilog
, tx
);
1087 mutex_exit(&zilog
->zl_lock
);
1095 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1097 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1098 (void) zil_parse(zilog
, zil_free_log_block
,
1099 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
, B_FALSE
);
1103 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
1105 dmu_tx_t
*tx
= txarg
;
1112 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
1113 DMU_OST_ANY
, B_FALSE
, B_FALSE
, FTAG
, &os
);
1116 * EBUSY indicates that the objset is inconsistent, in which
1117 * case it can not have a ZIL.
1119 if (error
!= EBUSY
) {
1120 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
1121 (unsigned long long)ds
->ds_object
, error
);
1127 zilog
= dmu_objset_zil(os
);
1128 zh
= zil_header_in_syncing_context(zilog
);
1129 ASSERT3U(tx
->tx_txg
, ==, spa_first_txg(zilog
->zl_spa
));
1130 first_txg
= spa_min_claim_txg(zilog
->zl_spa
);
1133 * If the spa_log_state is not set to be cleared, check whether
1134 * the current uberblock is a checkpoint one and if the current
1135 * header has been claimed before moving on.
1137 * If the current uberblock is a checkpointed uberblock then
1138 * one of the following scenarios took place:
1140 * 1] We are currently rewinding to the checkpoint of the pool.
1141 * 2] We crashed in the middle of a checkpoint rewind but we
1142 * did manage to write the checkpointed uberblock to the
1143 * vdev labels, so when we tried to import the pool again
1144 * the checkpointed uberblock was selected from the import
1147 * In both cases we want to zero out all the ZIL blocks, except
1148 * the ones that have been claimed at the time of the checkpoint
1149 * (their zh_claim_txg != 0). The reason is that these blocks
1150 * may be corrupted since we may have reused their locations on
1151 * disk after we took the checkpoint.
1153 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1154 * when we first figure out whether the current uberblock is
1155 * checkpointed or not. Unfortunately, that would discard all
1156 * the logs, including the ones that are claimed, and we would
1159 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
||
1160 (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1161 zh
->zh_claim_txg
== 0)) {
1162 if (!BP_IS_HOLE(&zh
->zh_log
)) {
1163 (void) zil_parse(zilog
, zil_clear_log_block
,
1164 zil_noop_log_record
, tx
, first_txg
, B_FALSE
);
1166 BP_ZERO(&zh
->zh_log
);
1167 if (os
->os_encrypted
)
1168 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1169 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1170 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1175 * If we are not rewinding and opening the pool normally, then
1176 * the min_claim_txg should be equal to the first txg of the pool.
1178 ASSERT3U(first_txg
, ==, spa_first_txg(zilog
->zl_spa
));
1181 * Claim all log blocks if we haven't already done so, and remember
1182 * the highest claimed sequence number. This ensures that if we can
1183 * read only part of the log now (e.g. due to a missing device),
1184 * but we can read the entire log later, we will not try to replay
1185 * or destroy beyond the last block we successfully claimed.
1187 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
1188 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
1189 (void) zil_parse(zilog
, zil_claim_log_block
,
1190 zil_claim_log_record
, tx
, first_txg
, B_FALSE
);
1191 zh
->zh_claim_txg
= first_txg
;
1192 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
1193 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
1194 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
1195 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
1196 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
1197 if (os
->os_encrypted
)
1198 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1199 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1202 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
1203 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1208 * Check the log by walking the log chain.
1209 * Checksum errors are ok as they indicate the end of the chain.
1210 * Any other error (no device or read failure) returns an error.
1213 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
1223 error
= dmu_objset_from_ds(ds
, &os
);
1225 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
1226 (unsigned long long)ds
->ds_object
, error
);
1230 zilog
= dmu_objset_zil(os
);
1231 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
1233 if (!BP_IS_HOLE(bp
)) {
1235 boolean_t valid
= B_TRUE
;
1238 * Check the first block and determine if it's on a log device
1239 * which may have been removed or faulted prior to loading this
1240 * pool. If so, there's no point in checking the rest of the
1241 * log as its content should have already been synced to the
1244 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
1245 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
1246 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
1247 valid
= vdev_log_state_valid(vd
);
1248 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
1254 * Check whether the current uberblock is checkpointed (e.g.
1255 * we are rewinding) and whether the current header has been
1256 * claimed or not. If it hasn't then skip verifying it. We
1257 * do this because its ZIL blocks may be part of the pool's
1258 * state before the rewind, which is no longer valid.
1260 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1261 if (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1262 zh
->zh_claim_txg
== 0)
1267 * Because tx == NULL, zil_claim_log_block() will not actually claim
1268 * any blocks, but just determine whether it is possible to do so.
1269 * In addition to checking the log chain, zil_claim_log_block()
1270 * will invoke zio_claim() with a done func of spa_claim_notify(),
1271 * which will update spa_max_claim_txg. See spa_load() for details.
1273 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
1274 zilog
->zl_header
->zh_claim_txg
? -1ULL :
1275 spa_min_claim_txg(os
->os_spa
), B_FALSE
);
1277 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
1281 * When an itx is "skipped", this function is used to properly mark the
1282 * waiter as "done, and signal any thread(s) waiting on it. An itx can
1283 * be skipped (and not committed to an lwb) for a variety of reasons,
1284 * one of them being that the itx was committed via spa_sync(), prior to
1285 * it being committed to an lwb; this can happen if a thread calling
1286 * zil_commit() is racing with spa_sync().
1289 zil_commit_waiter_skip(zil_commit_waiter_t
*zcw
)
1291 mutex_enter(&zcw
->zcw_lock
);
1292 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1293 zcw
->zcw_done
= B_TRUE
;
1294 cv_broadcast(&zcw
->zcw_cv
);
1295 mutex_exit(&zcw
->zcw_lock
);
1299 * This function is used when the given waiter is to be linked into an
1300 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1301 * At this point, the waiter will no longer be referenced by the itx,
1302 * and instead, will be referenced by the lwb.
1305 zil_commit_waiter_link_lwb(zil_commit_waiter_t
*zcw
, lwb_t
*lwb
)
1308 * The lwb_waiters field of the lwb is protected by the zilog's
1309 * zl_issuer_lock while the lwb is open and zl_lock otherwise.
1310 * zl_issuer_lock also protects leaving the open state.
1311 * zcw_lwb setting is protected by zl_issuer_lock and state !=
1312 * flush_done, which transition is protected by zl_lock.
1314 ASSERT(MUTEX_HELD(&lwb
->lwb_zilog
->zl_issuer_lock
));
1315 IMPLY(lwb
->lwb_state
!= LWB_STATE_OPENED
,
1316 MUTEX_HELD(&lwb
->lwb_zilog
->zl_lock
));
1317 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_NEW
);
1318 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1320 ASSERT(!list_link_active(&zcw
->zcw_node
));
1321 list_insert_tail(&lwb
->lwb_waiters
, zcw
);
1322 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1327 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1328 * block, and the given waiter must be linked to the "nolwb waiters"
1329 * list inside of zil_process_commit_list().
1332 zil_commit_waiter_link_nolwb(zil_commit_waiter_t
*zcw
, list_t
*nolwb
)
1334 ASSERT(!list_link_active(&zcw
->zcw_node
));
1335 list_insert_tail(nolwb
, zcw
);
1336 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1340 zil_lwb_add_block(lwb_t
*lwb
, const blkptr_t
*bp
)
1342 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1344 zil_vdev_node_t
*zv
, zvsearch
;
1345 int ndvas
= BP_GET_NDVAS(bp
);
1348 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1349 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1351 if (zil_nocacheflush
)
1354 mutex_enter(&lwb
->lwb_vdev_lock
);
1355 for (i
= 0; i
< ndvas
; i
++) {
1356 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1357 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
1358 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
1359 zv
->zv_vdev
= zvsearch
.zv_vdev
;
1360 avl_insert(t
, zv
, where
);
1363 mutex_exit(&lwb
->lwb_vdev_lock
);
1367 zil_lwb_flush_defer(lwb_t
*lwb
, lwb_t
*nlwb
)
1369 avl_tree_t
*src
= &lwb
->lwb_vdev_tree
;
1370 avl_tree_t
*dst
= &nlwb
->lwb_vdev_tree
;
1371 void *cookie
= NULL
;
1372 zil_vdev_node_t
*zv
;
1374 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1375 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1376 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1379 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1380 * not need the protection of lwb_vdev_lock (it will only be modified
1381 * while holding zilog->zl_lock) as its writes and those of its
1382 * children have all completed. The younger 'nlwb' may be waiting on
1383 * future writes to additional vdevs.
1385 mutex_enter(&nlwb
->lwb_vdev_lock
);
1387 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1388 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1390 while ((zv
= avl_destroy_nodes(src
, &cookie
)) != NULL
) {
1393 if (avl_find(dst
, zv
, &where
) == NULL
) {
1394 avl_insert(dst
, zv
, where
);
1396 kmem_free(zv
, sizeof (*zv
));
1399 mutex_exit(&nlwb
->lwb_vdev_lock
);
1403 zil_lwb_add_txg(lwb_t
*lwb
, uint64_t txg
)
1405 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1409 * This function is a called after all vdevs associated with a given lwb
1410 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1411 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1412 * all "previous" lwb's will have completed before this function is
1413 * called; i.e. this function is called for all previous lwbs before
1414 * it's called for "this" lwb (enforced via zio the dependencies
1415 * configured in zil_lwb_set_zio_dependency()).
1417 * The intention is for this function to be called as soon as the
1418 * contents of an lwb are considered "stable" on disk, and will survive
1419 * any sudden loss of power. At this point, any threads waiting for the
1420 * lwb to reach this state are signalled, and the "waiter" structures
1421 * are marked "done".
1424 zil_lwb_flush_vdevs_done(zio_t
*zio
)
1426 lwb_t
*lwb
= zio
->io_private
;
1427 zilog_t
*zilog
= lwb
->lwb_zilog
;
1428 zil_commit_waiter_t
*zcw
;
1431 spa_config_exit(zilog
->zl_spa
, SCL_STATE
, lwb
);
1433 hrtime_t t
= gethrtime() - lwb
->lwb_issued_timestamp
;
1435 mutex_enter(&zilog
->zl_lock
);
1437 zilog
->zl_last_lwb_latency
= (zilog
->zl_last_lwb_latency
* 7 + t
) / 8;
1439 lwb
->lwb_root_zio
= NULL
;
1441 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1442 lwb
->lwb_state
= LWB_STATE_FLUSH_DONE
;
1444 if (zilog
->zl_last_lwb_opened
== lwb
) {
1446 * Remember the highest committed log sequence number
1447 * for ztest. We only update this value when all the log
1448 * writes succeeded, because ztest wants to ASSERT that
1449 * it got the whole log chain.
1451 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1454 while ((itx
= list_remove_head(&lwb
->lwb_itxs
)) != NULL
)
1455 zil_itx_destroy(itx
);
1457 while ((zcw
= list_remove_head(&lwb
->lwb_waiters
)) != NULL
) {
1458 mutex_enter(&zcw
->zcw_lock
);
1460 ASSERT3P(zcw
->zcw_lwb
, ==, lwb
);
1461 zcw
->zcw_lwb
= NULL
;
1463 * We expect any ZIO errors from child ZIOs to have been
1464 * propagated "up" to this specific LWB's root ZIO, in
1465 * order for this error handling to work correctly. This
1466 * includes ZIO errors from either this LWB's write or
1467 * flush, as well as any errors from other dependent LWBs
1468 * (e.g. a root LWB ZIO that might be a child of this LWB).
1470 * With that said, it's important to note that LWB flush
1471 * errors are not propagated up to the LWB root ZIO.
1472 * This is incorrect behavior, and results in VDEV flush
1473 * errors not being handled correctly here. See the
1474 * comment above the call to "zio_flush" for details.
1477 zcw
->zcw_zio_error
= zio
->io_error
;
1479 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1480 zcw
->zcw_done
= B_TRUE
;
1481 cv_broadcast(&zcw
->zcw_cv
);
1483 mutex_exit(&zcw
->zcw_lock
);
1486 uint64_t txg
= lwb
->lwb_issued_txg
;
1488 /* Once we drop the lock, lwb may be freed by zil_sync(). */
1489 mutex_exit(&zilog
->zl_lock
);
1491 mutex_enter(&zilog
->zl_lwb_io_lock
);
1492 ASSERT3U(zilog
->zl_lwb_inflight
[txg
& TXG_MASK
], >, 0);
1493 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]--;
1494 if (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] == 0)
1495 cv_broadcast(&zilog
->zl_lwb_io_cv
);
1496 mutex_exit(&zilog
->zl_lwb_io_lock
);
1500 * Wait for the completion of all issued write/flush of that txg provided.
1501 * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1504 zil_lwb_flush_wait_all(zilog_t
*zilog
, uint64_t txg
)
1506 ASSERT3U(txg
, ==, spa_syncing_txg(zilog
->zl_spa
));
1508 mutex_enter(&zilog
->zl_lwb_io_lock
);
1509 while (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] > 0)
1510 cv_wait(&zilog
->zl_lwb_io_cv
, &zilog
->zl_lwb_io_lock
);
1511 mutex_exit(&zilog
->zl_lwb_io_lock
);
1514 mutex_enter(&zilog
->zl_lock
);
1515 mutex_enter(&zilog
->zl_lwb_io_lock
);
1516 lwb_t
*lwb
= list_head(&zilog
->zl_lwb_list
);
1517 while (lwb
!= NULL
) {
1518 if (lwb
->lwb_issued_txg
<= txg
) {
1519 ASSERT(lwb
->lwb_state
!= LWB_STATE_ISSUED
);
1520 ASSERT(lwb
->lwb_state
!= LWB_STATE_WRITE_DONE
);
1521 IMPLY(lwb
->lwb_issued_txg
> 0,
1522 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
1524 IMPLY(lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
1525 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
,
1526 lwb
->lwb_buf
== NULL
);
1527 lwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1529 mutex_exit(&zilog
->zl_lwb_io_lock
);
1530 mutex_exit(&zilog
->zl_lock
);
1535 * This is called when an lwb's write zio completes. The callback's
1536 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1537 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1538 * in writing out this specific lwb's data, and in the case that cache
1539 * flushes have been deferred, vdevs involved in writing the data for
1540 * previous lwbs. The writes corresponding to all the vdevs in the
1541 * lwb_vdev_tree will have completed by the time this is called, due to
1542 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1543 * which takes deferred flushes into account. The lwb will be "done"
1544 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1545 * completion callback for the lwb's root zio.
1548 zil_lwb_write_done(zio_t
*zio
)
1550 lwb_t
*lwb
= zio
->io_private
;
1551 spa_t
*spa
= zio
->io_spa
;
1552 zilog_t
*zilog
= lwb
->lwb_zilog
;
1553 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1554 void *cookie
= NULL
;
1555 zil_vdev_node_t
*zv
;
1558 ASSERT3S(spa_config_held(spa
, SCL_STATE
, RW_READER
), !=, 0);
1560 abd_free(zio
->io_abd
);
1561 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1562 lwb
->lwb_buf
= NULL
;
1564 mutex_enter(&zilog
->zl_lock
);
1565 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_ISSUED
);
1566 lwb
->lwb_state
= LWB_STATE_WRITE_DONE
;
1567 lwb
->lwb_child_zio
= NULL
;
1568 lwb
->lwb_write_zio
= NULL
;
1571 * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not
1572 * called for it yet, and when it will be, it won't be able to make
1573 * its write ZIO a parent this ZIO. In such case we can not defer
1574 * our flushes or below may be a race between the done callbacks.
1576 nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1577 if (nlwb
&& nlwb
->lwb_state
!= LWB_STATE_ISSUED
)
1579 mutex_exit(&zilog
->zl_lock
);
1581 if (avl_numnodes(t
) == 0)
1585 * If there was an IO error, we're not going to call zio_flush()
1586 * on these vdevs, so we simply empty the tree and free the
1587 * nodes. We avoid calling zio_flush() since there isn't any
1588 * good reason for doing so, after the lwb block failed to be
1591 * Additionally, we don't perform any further error handling at
1592 * this point (e.g. setting "zcw_zio_error" appropriately), as
1593 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1594 * we expect any error seen here, to have been propagated to
1597 if (zio
->io_error
!= 0) {
1598 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
1599 kmem_free(zv
, sizeof (*zv
));
1604 * If this lwb does not have any threads waiting for it to
1605 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1606 * command to the vdevs written to by "this" lwb, and instead
1607 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1608 * command for those vdevs. Thus, we merge the vdev tree of
1609 * "this" lwb with the vdev tree of the "next" lwb in the list,
1610 * and assume the "next" lwb will handle flushing the vdevs (or
1611 * deferring the flush(s) again).
1613 * This is a useful performance optimization, especially for
1614 * workloads with lots of async write activity and few sync
1615 * write and/or fsync activity, as it has the potential to
1616 * coalesce multiple flush commands to a vdev into one.
1618 if (list_is_empty(&lwb
->lwb_waiters
) && nlwb
!= NULL
) {
1619 zil_lwb_flush_defer(lwb
, nlwb
);
1620 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
1624 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1625 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
1628 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1629 * always used within "zio_flush". This means,
1630 * any errors when flushing the vdev(s), will
1631 * (unfortunately) not be handled correctly,
1632 * since these "zio_flush" errors will not be
1633 * propagated up to "zil_lwb_flush_vdevs_done".
1635 zio_flush(lwb
->lwb_root_zio
, vd
);
1637 kmem_free(zv
, sizeof (*zv
));
1642 * Build the zio dependency chain, which is used to preserve the ordering of
1643 * lwb completions that is required by the semantics of the ZIL. Each new lwb
1644 * zio becomes a parent of the previous lwb zio, such that the new lwb's zio
1645 * cannot complete until the previous lwb's zio completes.
1647 * This is required by the semantics of zil_commit(): the commit waiters
1648 * attached to the lwbs will be woken in the lwb zio's completion callback,
1649 * so this zio dependency graph ensures the waiters are woken in the correct
1650 * order (the same order the lwbs were created).
1653 zil_lwb_set_zio_dependency(zilog_t
*zilog
, lwb_t
*lwb
)
1655 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
1657 lwb_t
*prev_lwb
= list_prev(&zilog
->zl_lwb_list
, lwb
);
1658 if (prev_lwb
== NULL
||
1659 prev_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
1663 * If the previous lwb's write hasn't already completed, we also want
1664 * to order the completion of the lwb write zios (above, we only order
1665 * the completion of the lwb root zios). This is required because of
1666 * how we can defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1668 * When the DKIOCFLUSHWRITECACHE commands are deferred, the previous
1669 * lwb will rely on this lwb to flush the vdevs written to by that
1670 * previous lwb. Thus, we need to ensure this lwb doesn't issue the
1671 * flush until after the previous lwb's write completes. We ensure
1672 * this ordering by setting the zio parent/child relationship here.
1674 * Without this relationship on the lwb's write zio, it's possible
1675 * for this lwb's write to complete prior to the previous lwb's write
1676 * completing; and thus, the vdevs for the previous lwb would be
1677 * flushed prior to that lwb's data being written to those vdevs (the
1678 * vdevs are flushed in the lwb write zio's completion handler,
1679 * zil_lwb_write_done()).
1681 if (prev_lwb
->lwb_state
== LWB_STATE_ISSUED
) {
1682 ASSERT3P(prev_lwb
->lwb_write_zio
, !=, NULL
);
1683 zio_add_child(lwb
->lwb_write_zio
, prev_lwb
->lwb_write_zio
);
1685 ASSERT3S(prev_lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1688 ASSERT3P(prev_lwb
->lwb_root_zio
, !=, NULL
);
1689 zio_add_child(lwb
->lwb_root_zio
, prev_lwb
->lwb_root_zio
);
1694 * This function's purpose is to "open" an lwb such that it is ready to
1695 * accept new itxs being committed to it. This function is idempotent; if
1696 * the passed in lwb has already been opened, it is essentially a no-op.
1699 zil_lwb_write_open(zilog_t
*zilog
, lwb_t
*lwb
)
1701 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1703 if (lwb
->lwb_state
!= LWB_STATE_NEW
) {
1704 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1708 mutex_enter(&zilog
->zl_lock
);
1709 lwb
->lwb_state
= LWB_STATE_OPENED
;
1710 zilog
->zl_last_lwb_opened
= lwb
;
1711 mutex_exit(&zilog
->zl_lock
);
1715 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1716 * initialized. Otherwise this should not be used directly; see
1717 * zl_max_block_size instead.
1719 static uint_t zil_maxblocksize
= SPA_OLD_MAXBLOCKSIZE
;
1722 * Plan splitting of the provided burst size between several blocks.
1725 zil_lwb_plan(zilog_t
*zilog
, uint64_t size
, uint_t
*minsize
)
1727 uint_t md
= zilog
->zl_max_block_size
- sizeof (zil_chain_t
);
1731 * Small bursts are written as-is in one block.
1735 } else if (size
> 8 * md
) {
1737 * Big bursts use maximum blocks. The first block size
1738 * is hard to predict, but it does not really matter.
1745 * Medium bursts try to divide evenly to better utilize several SLOG
1746 * VDEVs. The first block size we predict assuming the worst case of
1747 * maxing out others. Fall back to using maximum blocks if due to
1748 * large records or wasted space we can not predict anything better.
1751 uint_t n
= DIV_ROUND_UP(s
, md
- sizeof (lr_write_t
));
1752 uint_t chunk
= DIV_ROUND_UP(s
, n
);
1753 uint_t waste
= zil_max_waste_space(zilog
);
1754 waste
= MAX(waste
, zilog
->zl_cur_max
);
1755 if (chunk
<= md
- waste
) {
1756 *minsize
= MAX(s
- (md
- waste
) * (n
- 1), waste
);
1765 * Try to predict next block size based on previous history. Make prediction
1766 * sufficient for 7 of 8 previous bursts. Don't try to save if the saving is
1767 * less then 50%, extra writes may cost more, but we don't want single spike
1768 * to badly affect our predictions.
1771 zil_lwb_predict(zilog_t
*zilog
)
1775 /* If we are in the middle of a burst, take it into account also. */
1776 if (zilog
->zl_cur_size
> 0) {
1777 o
= zil_lwb_plan(zilog
, zilog
->zl_cur_size
, &m
);
1783 /* Find minimum optimal size. We don't need to go below that. */
1784 for (int i
= 0; i
< ZIL_BURSTS
; i
++)
1785 o
= MIN(o
, zilog
->zl_prev_opt
[i
]);
1787 /* Find two biggest minimal first block sizes above the optimal. */
1788 uint_t m1
= MAX(m
, o
), m2
= o
;
1789 for (int i
= 0; i
< ZIL_BURSTS
; i
++) {
1790 m
= zilog
->zl_prev_min
[i
];
1794 } else if (m
> m2
) {
1800 * If second minimum size gives 50% saving -- use it. It may cost us
1801 * one additional write later, but the space saving is just too big.
1803 return ((m1
< m2
* 2) ? m1
: m2
);
1807 * Close the log block for being issued and allocate the next one.
1808 * Has to be called under zl_issuer_lock to chain more lwbs.
1811 zil_lwb_write_close(zilog_t
*zilog
, lwb_t
*lwb
, lwb_state_t state
)
1813 uint64_t blksz
, plan
, plan2
;
1815 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1816 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1817 lwb
->lwb_state
= LWB_STATE_CLOSED
;
1820 * If there was an allocation failure then returned NULL will trigger
1821 * zil_commit_writer_stall() at the caller. This is inherently racy,
1822 * since allocation may not have happened yet.
1824 if (lwb
->lwb_error
!= 0)
1828 * Log blocks are pre-allocated. Here we select the size of the next
1829 * block, based on what's left of this burst and the previous history.
1830 * While we try to only write used part of the block, we can't just
1831 * always allocate the maximum block size because we can exhaust all
1832 * available pool log space, so we try to be reasonable.
1834 if (zilog
->zl_cur_left
> 0) {
1836 * We are in the middle of a burst and know how much is left.
1837 * But if workload is multi-threaded there may be more soon.
1838 * Try to predict what can it be and plan for the worst case.
1841 plan
= zil_lwb_plan(zilog
, zilog
->zl_cur_left
, &m
);
1842 if (zilog
->zl_parallel
) {
1843 plan2
= zil_lwb_plan(zilog
, zilog
->zl_cur_left
+
1844 zil_lwb_predict(zilog
), &m
);
1850 * The previous burst is done and we can only predict what
1853 plan
= zil_lwb_predict(zilog
);
1855 blksz
= plan
+ sizeof (zil_chain_t
);
1856 blksz
= P2ROUNDUP_TYPED(blksz
, ZIL_MIN_BLKSZ
, uint64_t);
1857 blksz
= MIN(blksz
, zilog
->zl_max_block_size
);
1858 DTRACE_PROBE3(zil__block__size
, zilog_t
*, zilog
, uint64_t, blksz
,
1861 return (zil_alloc_lwb(zilog
, blksz
, NULL
, 0, 0, state
));
1865 * Finalize previously closed block and issue the write zio.
1868 zil_lwb_write_issue(zilog_t
*zilog
, lwb_t
*lwb
)
1870 spa_t
*spa
= zilog
->zl_spa
;
1873 zbookmark_phys_t zb
;
1874 zio_priority_t prio
;
1877 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_CLOSED
);
1879 /* Actually fill the lwb with the data. */
1880 for (itx_t
*itx
= list_head(&lwb
->lwb_itxs
); itx
;
1881 itx
= list_next(&lwb
->lwb_itxs
, itx
))
1882 zil_lwb_commit(zilog
, lwb
, itx
);
1883 lwb
->lwb_nused
= lwb
->lwb_nfilled
;
1884 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_nmax
);
1886 lwb
->lwb_root_zio
= zio_root(spa
, zil_lwb_flush_vdevs_done
, lwb
,
1890 * The lwb is now ready to be issued, but it can be only if it already
1891 * got its block pointer allocated or the allocation has failed.
1892 * Otherwise leave it as-is, relying on some other thread to issue it
1893 * after allocating its block pointer via calling zil_lwb_write_issue()
1894 * for the previous lwb(s) in the chain.
1896 mutex_enter(&zilog
->zl_lock
);
1897 lwb
->lwb_state
= LWB_STATE_READY
;
1898 if (BP_IS_HOLE(&lwb
->lwb_blk
) && lwb
->lwb_error
== 0) {
1899 mutex_exit(&zilog
->zl_lock
);
1902 mutex_exit(&zilog
->zl_lock
);
1906 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
1908 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_nmax
);
1909 int wsz
= lwb
->lwb_sz
;
1910 if (lwb
->lwb_error
== 0) {
1911 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
, lwb
->lwb_sz
);
1912 if (!lwb
->lwb_slog
|| zilog
->zl_cur_size
<= zil_slog_bulk
)
1913 prio
= ZIO_PRIORITY_SYNC_WRITE
;
1915 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
1916 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1917 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
1918 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1919 lwb
->lwb_write_zio
= zio_rewrite(lwb
->lwb_root_zio
, spa
, 0,
1920 &lwb
->lwb_blk
, lwb_abd
, lwb
->lwb_sz
, zil_lwb_write_done
,
1921 lwb
, prio
, ZIO_FLAG_CANFAIL
, &zb
);
1922 zil_lwb_add_block(lwb
, &lwb
->lwb_blk
);
1924 if (lwb
->lwb_slim
) {
1925 /* For Slim ZIL only write what is used. */
1926 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
,
1928 ASSERT3S(wsz
, <=, lwb
->lwb_sz
);
1929 zio_shrink(lwb
->lwb_write_zio
, wsz
);
1930 wsz
= lwb
->lwb_write_zio
->io_size
;
1932 memset(lwb
->lwb_buf
+ lwb
->lwb_nused
, 0, wsz
- lwb
->lwb_nused
);
1934 zilc
->zc_nused
= lwb
->lwb_nused
;
1935 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1938 * We can't write the lwb if there was an allocation failure,
1939 * so create a null zio instead just to maintain dependencies.
1941 lwb
->lwb_write_zio
= zio_null(lwb
->lwb_root_zio
, spa
, NULL
,
1942 zil_lwb_write_done
, lwb
, ZIO_FLAG_CANFAIL
);
1943 lwb
->lwb_write_zio
->io_error
= lwb
->lwb_error
;
1945 if (lwb
->lwb_child_zio
)
1946 zio_add_child(lwb
->lwb_write_zio
, lwb
->lwb_child_zio
);
1949 * Open transaction to allocate the next block pointer.
1951 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
1952 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
1953 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1954 uint64_t txg
= dmu_tx_get_txg(tx
);
1957 * Allocate next the block pointer unless we are already in error.
1959 lwb_t
*nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1960 blkptr_t
*bp
= &zilc
->zc_next_blk
;
1962 error
= lwb
->lwb_error
;
1964 error
= zio_alloc_zil(spa
, zilog
->zl_os
, txg
, bp
, nlwb
->lwb_sz
,
1968 ASSERT3U(bp
->blk_birth
, ==, txg
);
1969 BP_SET_CHECKSUM(bp
, nlwb
->lwb_slim
? ZIO_CHECKSUM_ZILOG2
:
1970 ZIO_CHECKSUM_ZILOG
);
1971 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1972 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1976 * Reduce TXG open time by incrementing inflight counter and committing
1977 * the transaciton. zil_sync() will wait for it to return to zero.
1979 mutex_enter(&zilog
->zl_lwb_io_lock
);
1980 lwb
->lwb_issued_txg
= txg
;
1981 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]++;
1982 zilog
->zl_lwb_max_issued_txg
= MAX(txg
, zilog
->zl_lwb_max_issued_txg
);
1983 mutex_exit(&zilog
->zl_lwb_io_lock
);
1986 spa_config_enter(spa
, SCL_STATE
, lwb
, RW_READER
);
1989 * We've completed all potentially blocking operations. Update the
1990 * nlwb and allow it proceed without possible lock order reversals.
1992 mutex_enter(&zilog
->zl_lock
);
1993 zil_lwb_set_zio_dependency(zilog
, lwb
);
1994 lwb
->lwb_state
= LWB_STATE_ISSUED
;
1997 nlwb
->lwb_blk
= *bp
;
1998 nlwb
->lwb_error
= error
;
1999 nlwb
->lwb_slog
= slog
;
2000 nlwb
->lwb_alloc_txg
= txg
;
2001 if (nlwb
->lwb_state
!= LWB_STATE_READY
)
2004 mutex_exit(&zilog
->zl_lock
);
2006 if (lwb
->lwb_slog
) {
2007 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_slog_count
);
2008 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_bytes
,
2010 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_write
,
2012 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_alloc
,
2013 BP_GET_LSIZE(&lwb
->lwb_blk
));
2015 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_normal_count
);
2016 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_bytes
,
2018 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_write
,
2020 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_alloc
,
2021 BP_GET_LSIZE(&lwb
->lwb_blk
));
2023 lwb
->lwb_issued_timestamp
= gethrtime();
2024 if (lwb
->lwb_child_zio
)
2025 zio_nowait(lwb
->lwb_child_zio
);
2026 zio_nowait(lwb
->lwb_write_zio
);
2027 zio_nowait(lwb
->lwb_root_zio
);
2030 * If nlwb was ready when we gave it the block pointer,
2031 * it is on us to issue it and possibly following ones.
2039 * Maximum amount of data that can be put into single log block.
2042 zil_max_log_data(zilog_t
*zilog
, size_t hdrsize
)
2044 return (zilog
->zl_max_block_size
- sizeof (zil_chain_t
) - hdrsize
);
2048 * Maximum amount of log space we agree to waste to reduce number of
2049 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%).
2051 static inline uint64_t
2052 zil_max_waste_space(zilog_t
*zilog
)
2054 return (zil_max_log_data(zilog
, sizeof (lr_write_t
)) / 16);
2058 * Maximum amount of write data for WR_COPIED. For correctness, consumers
2059 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
2060 * maximum sized log block, because each WR_COPIED record must fit in a
2061 * single log block. Below that it is a tradeoff of additional memory copy
2062 * and possibly worse log space efficiency vs additional range lock/unlock.
2064 static uint_t zil_maxcopied
= 7680;
2067 zil_max_copied_data(zilog_t
*zilog
)
2069 uint64_t max_data
= zil_max_log_data(zilog
, sizeof (lr_write_t
));
2070 return (MIN(max_data
, zil_maxcopied
));
2074 zil_itx_record_size(itx_t
*itx
)
2076 lr_t
*lr
= &itx
->itx_lr
;
2078 if (lr
->lrc_txtype
== TX_COMMIT
)
2080 ASSERT3U(lr
->lrc_reclen
, >=, sizeof (lr_t
));
2081 return (lr
->lrc_reclen
);
2085 zil_itx_data_size(itx_t
*itx
)
2087 lr_t
*lr
= &itx
->itx_lr
;
2088 lr_write_t
*lrw
= (lr_write_t
*)lr
;
2090 if (lr
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
2091 ASSERT3U(lr
->lrc_reclen
, ==, sizeof (lr_write_t
));
2092 return (P2ROUNDUP_TYPED(lrw
->lr_length
, sizeof (uint64_t),
2099 zil_itx_full_size(itx_t
*itx
)
2101 lr_t
*lr
= &itx
->itx_lr
;
2103 if (lr
->lrc_txtype
== TX_COMMIT
)
2105 ASSERT3U(lr
->lrc_reclen
, >=, sizeof (lr_t
));
2106 return (lr
->lrc_reclen
+ zil_itx_data_size(itx
));
2110 * Estimate space needed in the lwb for the itx. Allocate more lwbs or
2111 * split the itx as needed, but don't touch the actual transaction data.
2112 * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
2113 * to chain more lwbs.
2116 zil_lwb_assign(zilog_t
*zilog
, lwb_t
*lwb
, itx_t
*itx
, list_t
*ilwbs
)
2121 uint64_t dlen
, dnow
, lwb_sp
, reclen
, max_log_data
;
2123 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2124 ASSERT3P(lwb
, !=, NULL
);
2125 ASSERT3P(lwb
->lwb_buf
, !=, NULL
);
2127 zil_lwb_write_open(zilog
, lwb
);
2130 lrw
= (lr_write_t
*)lr
;
2133 * A commit itx doesn't represent any on-disk state; instead
2134 * it's simply used as a place holder on the commit list, and
2135 * provides a mechanism for attaching a "commit waiter" onto the
2136 * correct lwb (such that the waiter can be signalled upon
2137 * completion of that lwb). Thus, we don't process this itx's
2138 * log record if it's a commit itx (these itx's don't have log
2139 * records), and instead link the itx's waiter onto the lwb's
2142 * For more details, see the comment above zil_commit().
2144 if (lr
->lrc_txtype
== TX_COMMIT
) {
2145 zil_commit_waiter_link_lwb(itx
->itx_private
, lwb
);
2146 list_insert_tail(&lwb
->lwb_itxs
, itx
);
2150 reclen
= lr
->lrc_reclen
;
2151 ASSERT3U(reclen
, >=, sizeof (lr_t
));
2152 ASSERT3U(reclen
, <=, zil_max_log_data(zilog
, 0));
2153 dlen
= zil_itx_data_size(itx
);
2157 * If this record won't fit in the current log block, start a new one.
2158 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2160 lwb_sp
= lwb
->lwb_nmax
- lwb
->lwb_nused
;
2161 max_log_data
= zil_max_log_data(zilog
, sizeof (lr_write_t
));
2162 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
2163 lwb_sp
< zil_max_waste_space(zilog
) &&
2164 (dlen
% max_log_data
== 0 ||
2165 lwb_sp
< reclen
+ dlen
% max_log_data
))) {
2166 list_insert_tail(ilwbs
, lwb
);
2167 lwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_OPENED
);
2170 lwb_sp
= lwb
->lwb_nmax
- lwb
->lwb_nused
;
2174 * There must be enough space in the log block to hold reclen.
2175 * For WR_COPIED, we need to fit the whole record in one block,
2176 * and reclen is the write record header size + the data size.
2177 * For WR_NEED_COPY, we can create multiple records, splitting
2178 * the data into multiple blocks, so we only need to fit one
2179 * word of data per block; in this case reclen is just the header
2182 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
2184 dnow
= MIN(dlen
, lwb_sp
- reclen
);
2186 ASSERT3U(lr
->lrc_txtype
, ==, TX_WRITE
);
2187 ASSERT3U(itx
->itx_wr_state
, ==, WR_NEED_COPY
);
2188 citx
= zil_itx_clone(itx
);
2189 clr
= &citx
->itx_lr
;
2190 lr_write_t
*clrw
= (lr_write_t
*)clr
;
2191 clrw
->lr_length
= dnow
;
2192 lrw
->lr_offset
+= dnow
;
2193 lrw
->lr_length
-= dnow
;
2194 zilog
->zl_cur_left
-= dnow
;
2201 * We're actually making an entry, so update lrc_seq to be the
2202 * log record sequence number. Note that this is generally not
2203 * equal to the itx sequence number because not all transactions
2204 * are synchronous, and sometimes spa_sync() gets there first.
2206 clr
->lrc_seq
= ++zilog
->zl_lr_seq
;
2208 lwb
->lwb_nused
+= reclen
+ dnow
;
2209 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_nmax
);
2210 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
2212 zil_lwb_add_txg(lwb
, lr
->lrc_txg
);
2213 list_insert_tail(&lwb
->lwb_itxs
, citx
);
2219 if (lr
->lrc_txtype
== TX_WRITE
&&
2220 lr
->lrc_txg
> spa_freeze_txg(zilog
->zl_spa
))
2221 txg_wait_synced(zilog
->zl_dmu_pool
, lr
->lrc_txg
);
2227 * Fill the actual transaction data into the lwb, following zil_lwb_assign().
2228 * Does not require locking.
2231 zil_lwb_commit(zilog_t
*zilog
, lwb_t
*lwb
, itx_t
*itx
)
2234 lr_write_t
*lrw
, *lrwb
;
2236 uint64_t dlen
, reclen
;
2239 lrw
= (lr_write_t
*)lr
;
2241 if (lr
->lrc_txtype
== TX_COMMIT
)
2244 reclen
= lr
->lrc_reclen
;
2245 dlen
= zil_itx_data_size(itx
);
2246 ASSERT3U(reclen
+ dlen
, <=, lwb
->lwb_nused
- lwb
->lwb_nfilled
);
2248 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nfilled
;
2249 memcpy(lr_buf
, lr
, reclen
);
2250 lrb
= (lr_t
*)lr_buf
; /* Like lr, but inside lwb. */
2251 lrwb
= (lr_write_t
*)lrb
; /* Like lrw, but inside lwb. */
2253 ZIL_STAT_BUMP(zilog
, zil_itx_count
);
2256 * If it's a write, fetch the data or get its blkptr as appropriate.
2258 if (lr
->lrc_txtype
== TX_WRITE
) {
2259 if (itx
->itx_wr_state
== WR_COPIED
) {
2260 ZIL_STAT_BUMP(zilog
, zil_itx_copied_count
);
2261 ZIL_STAT_INCR(zilog
, zil_itx_copied_bytes
,
2267 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
2268 dbuf
= lr_buf
+ reclen
;
2269 lrb
->lrc_reclen
+= dlen
;
2270 ZIL_STAT_BUMP(zilog
, zil_itx_needcopy_count
);
2271 ZIL_STAT_INCR(zilog
, zil_itx_needcopy_bytes
,
2274 ASSERT3S(itx
->itx_wr_state
, ==, WR_INDIRECT
);
2276 ZIL_STAT_BUMP(zilog
, zil_itx_indirect_count
);
2277 ZIL_STAT_INCR(zilog
, zil_itx_indirect_bytes
,
2279 if (lwb
->lwb_child_zio
== NULL
) {
2280 lwb
->lwb_child_zio
= zio_null(NULL
,
2281 zilog
->zl_spa
, NULL
, NULL
, NULL
,
2287 * The "lwb_child_zio" we pass in will become a child of
2288 * "lwb_write_zio", when one is created, so one will be
2289 * a parent of any zio's created by the "zl_get_data".
2290 * This way "lwb_write_zio" will first wait for children
2291 * block pointers before own writing, and then for their
2292 * writing completion before the vdev cache flushing.
2294 error
= zilog
->zl_get_data(itx
->itx_private
,
2295 itx
->itx_gen
, lrwb
, dbuf
, lwb
,
2296 lwb
->lwb_child_zio
);
2297 if (dbuf
!= NULL
&& error
== 0) {
2298 /* Zero any padding bytes in the last block. */
2299 memset((char *)dbuf
+ lrwb
->lr_length
, 0,
2300 dlen
- lrwb
->lr_length
);
2304 * Typically, the only return values we should see from
2305 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2306 * EALREADY. However, it is also possible to see other
2307 * error values such as ENOSPC or EINVAL from
2308 * dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2309 * ENXIO as well as a multitude of others from the
2310 * block layer through dmu_buf_hold() -> dbuf_read()
2311 * -> zio_wait(), as well as through dmu_read() ->
2312 * dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2313 * zio_wait(). When these errors happen, we can assume
2314 * that neither an immediate write nor an indirect
2315 * write occurred, so we need to fall back to
2316 * txg_wait_synced(). This is unusual, so we print to
2317 * dmesg whenever one of these errors occurs.
2323 cmn_err(CE_WARN
, "zil_lwb_commit() received "
2324 "unexpected error %d from ->zl_get_data()"
2325 ". Falling back to txg_wait_synced().",
2329 txg_wait_synced(zilog
->zl_dmu_pool
,
2342 lwb
->lwb_nfilled
+= reclen
+ dlen
;
2343 ASSERT3S(lwb
->lwb_nfilled
, <=, lwb
->lwb_nused
);
2344 ASSERT0(P2PHASE(lwb
->lwb_nfilled
, sizeof (uint64_t)));
2348 zil_itx_create(uint64_t txtype
, size_t olrsize
)
2350 size_t itxsize
, lrsize
;
2353 ASSERT3U(olrsize
, >=, sizeof (lr_t
));
2354 lrsize
= P2ROUNDUP_TYPED(olrsize
, sizeof (uint64_t), size_t);
2355 ASSERT3U(lrsize
, >=, olrsize
);
2356 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
2358 itx
= zio_data_buf_alloc(itxsize
);
2359 itx
->itx_lr
.lrc_txtype
= txtype
;
2360 itx
->itx_lr
.lrc_reclen
= lrsize
;
2361 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
2362 memset((char *)&itx
->itx_lr
+ olrsize
, 0, lrsize
- olrsize
);
2363 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
2364 itx
->itx_callback
= NULL
;
2365 itx
->itx_callback_data
= NULL
;
2366 itx
->itx_size
= itxsize
;
2372 zil_itx_clone(itx_t
*oitx
)
2374 ASSERT3U(oitx
->itx_size
, >=, sizeof (itx_t
));
2375 ASSERT3U(oitx
->itx_size
, ==,
2376 offsetof(itx_t
, itx_lr
) + oitx
->itx_lr
.lrc_reclen
);
2378 itx_t
*itx
= zio_data_buf_alloc(oitx
->itx_size
);
2379 memcpy(itx
, oitx
, oitx
->itx_size
);
2380 itx
->itx_callback
= NULL
;
2381 itx
->itx_callback_data
= NULL
;
2386 zil_itx_destroy(itx_t
*itx
)
2388 ASSERT3U(itx
->itx_size
, >=, sizeof (itx_t
));
2389 ASSERT3U(itx
->itx_lr
.lrc_reclen
, ==,
2390 itx
->itx_size
- offsetof(itx_t
, itx_lr
));
2391 IMPLY(itx
->itx_lr
.lrc_txtype
== TX_COMMIT
, itx
->itx_callback
== NULL
);
2392 IMPLY(itx
->itx_callback
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2394 if (itx
->itx_callback
!= NULL
)
2395 itx
->itx_callback(itx
->itx_callback_data
);
2397 zio_data_buf_free(itx
, itx
->itx_size
);
2401 * Free up the sync and async itxs. The itxs_t has already been detached
2402 * so no locks are needed.
2405 zil_itxg_clean(void *arg
)
2412 itx_async_node_t
*ian
;
2414 list
= &itxs
->i_sync_list
;
2415 while ((itx
= list_remove_head(list
)) != NULL
) {
2417 * In the general case, commit itxs will not be found
2418 * here, as they'll be committed to an lwb via
2419 * zil_lwb_assign(), and free'd in that function. Having
2420 * said that, it is still possible for commit itxs to be
2421 * found here, due to the following race:
2423 * - a thread calls zil_commit() which assigns the
2424 * commit itx to a per-txg i_sync_list
2425 * - zil_itxg_clean() is called (e.g. via spa_sync())
2426 * while the waiter is still on the i_sync_list
2428 * There's nothing to prevent syncing the txg while the
2429 * waiter is on the i_sync_list. This normally doesn't
2430 * happen because spa_sync() is slower than zil_commit(),
2431 * but if zil_commit() calls txg_wait_synced() (e.g.
2432 * because zil_create() or zil_commit_writer_stall() is
2433 * called) we will hit this case.
2435 if (itx
->itx_lr
.lrc_txtype
== TX_COMMIT
)
2436 zil_commit_waiter_skip(itx
->itx_private
);
2438 zil_itx_destroy(itx
);
2442 t
= &itxs
->i_async_tree
;
2443 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2444 list
= &ian
->ia_list
;
2445 while ((itx
= list_remove_head(list
)) != NULL
) {
2446 /* commit itxs should never be on the async lists. */
2447 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2448 zil_itx_destroy(itx
);
2451 kmem_free(ian
, sizeof (itx_async_node_t
));
2455 kmem_free(itxs
, sizeof (itxs_t
));
2459 zil_aitx_compare(const void *x1
, const void *x2
)
2461 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
2462 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
2464 return (TREE_CMP(o1
, o2
));
2468 * Remove all async itx with the given oid.
2471 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
2474 itx_async_node_t
*ian
, ian_search
;
2481 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2483 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2486 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2488 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2489 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2491 mutex_enter(&itxg
->itxg_lock
);
2492 if (itxg
->itxg_txg
!= txg
) {
2493 mutex_exit(&itxg
->itxg_lock
);
2498 * Locate the object node and append its list.
2500 t
= &itxg
->itxg_itxs
->i_async_tree
;
2501 ian_search
.ia_foid
= oid
;
2502 ian
= avl_find(t
, &ian_search
, &where
);
2504 list_move_tail(&clean_list
, &ian
->ia_list
);
2505 mutex_exit(&itxg
->itxg_lock
);
2507 while ((itx
= list_remove_head(&clean_list
)) != NULL
) {
2508 /* commit itxs should never be on the async lists. */
2509 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2510 zil_itx_destroy(itx
);
2512 list_destroy(&clean_list
);
2516 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
2520 itxs_t
*itxs
, *clean
= NULL
;
2523 * Ensure the data of a renamed file is committed before the rename.
2525 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
2526 zil_async_to_sync(zilog
, itx
->itx_oid
);
2528 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
2531 txg
= dmu_tx_get_txg(tx
);
2533 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2534 mutex_enter(&itxg
->itxg_lock
);
2535 itxs
= itxg
->itxg_itxs
;
2536 if (itxg
->itxg_txg
!= txg
) {
2539 * The zil_clean callback hasn't got around to cleaning
2540 * this itxg. Save the itxs for release below.
2541 * This should be rare.
2543 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2544 "txg %llu", (u_longlong_t
)itxg
->itxg_txg
);
2545 clean
= itxg
->itxg_itxs
;
2547 itxg
->itxg_txg
= txg
;
2548 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
2551 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
2552 offsetof(itx_t
, itx_node
));
2553 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
2554 sizeof (itx_async_node_t
),
2555 offsetof(itx_async_node_t
, ia_node
));
2557 if (itx
->itx_sync
) {
2558 list_insert_tail(&itxs
->i_sync_list
, itx
);
2560 avl_tree_t
*t
= &itxs
->i_async_tree
;
2562 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
2563 itx_async_node_t
*ian
;
2566 ian
= avl_find(t
, &foid
, &where
);
2568 ian
= kmem_alloc(sizeof (itx_async_node_t
),
2570 list_create(&ian
->ia_list
, sizeof (itx_t
),
2571 offsetof(itx_t
, itx_node
));
2572 ian
->ia_foid
= foid
;
2573 avl_insert(t
, ian
, where
);
2575 list_insert_tail(&ian
->ia_list
, itx
);
2578 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
2581 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2582 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2583 * need to be careful to always dirty the ZIL using the "real"
2584 * TXG (not itxg_txg) even when the SPA is frozen.
2586 zilog_dirty(zilog
, dmu_tx_get_txg(tx
));
2587 mutex_exit(&itxg
->itxg_lock
);
2589 /* Release the old itxs now we've dropped the lock */
2591 zil_itxg_clean(clean
);
2595 * If there are any in-memory intent log transactions which have now been
2596 * synced then start up a taskq to free them. We should only do this after we
2597 * have written out the uberblocks (i.e. txg has been committed) so that
2598 * don't inadvertently clean out in-memory log records that would be required
2602 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
2604 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
2607 ASSERT3U(synced_txg
, <, ZILTEST_TXG
);
2609 mutex_enter(&itxg
->itxg_lock
);
2610 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
2611 mutex_exit(&itxg
->itxg_lock
);
2614 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
2615 ASSERT3U(itxg
->itxg_txg
, !=, 0);
2616 clean_me
= itxg
->itxg_itxs
;
2617 itxg
->itxg_itxs
= NULL
;
2619 mutex_exit(&itxg
->itxg_lock
);
2621 * Preferably start a task queue to free up the old itxs but
2622 * if taskq_dispatch can't allocate resources to do that then
2623 * free it in-line. This should be rare. Note, using TQ_SLEEP
2624 * created a bad performance problem.
2626 ASSERT3P(zilog
->zl_dmu_pool
, !=, NULL
);
2627 ASSERT3P(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
, !=, NULL
);
2628 taskqid_t id
= taskq_dispatch(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
,
2629 zil_itxg_clean
, clean_me
, TQ_NOSLEEP
);
2630 if (id
== TASKQID_INVALID
)
2631 zil_itxg_clean(clean_me
);
2635 * This function will traverse the queue of itxs that need to be
2636 * committed, and move them onto the ZIL's zl_itx_commit_list.
2639 zil_get_commit_list(zilog_t
*zilog
)
2641 uint64_t otxg
, txg
, wtxg
= 0;
2642 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
2644 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2646 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2649 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2652 * This is inherently racy, since there is nothing to prevent
2653 * the last synced txg from changing. That's okay since we'll
2654 * only commit things in the future.
2656 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2657 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2659 mutex_enter(&itxg
->itxg_lock
);
2660 if (itxg
->itxg_txg
!= txg
) {
2661 mutex_exit(&itxg
->itxg_lock
);
2666 * If we're adding itx records to the zl_itx_commit_list,
2667 * then the zil better be dirty in this "txg". We can assert
2668 * that here since we're holding the itxg_lock which will
2669 * prevent spa_sync from cleaning it. Once we add the itxs
2670 * to the zl_itx_commit_list we must commit it to disk even
2671 * if it's unnecessary (i.e. the txg was synced).
2673 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
2674 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
2675 list_t
*sync_list
= &itxg
->itxg_itxs
->i_sync_list
;
2677 if (unlikely(zilog
->zl_suspend
> 0)) {
2679 * ZIL was just suspended, but we lost the race.
2680 * Allow all earlier itxs to be committed, but ask
2681 * caller to do txg_wait_synced(txg) for any new.
2683 if (!list_is_empty(sync_list
))
2684 wtxg
= MAX(wtxg
, txg
);
2686 itx
= list_head(sync_list
);
2687 list_move_tail(commit_list
, sync_list
);
2690 mutex_exit(&itxg
->itxg_lock
);
2692 while (itx
!= NULL
) {
2693 uint64_t s
= zil_itx_full_size(itx
);
2694 zilog
->zl_cur_size
+= s
;
2695 zilog
->zl_cur_left
+= s
;
2696 s
= zil_itx_record_size(itx
);
2697 zilog
->zl_cur_max
= MAX(zilog
->zl_cur_max
, s
);
2698 itx
= list_next(commit_list
, itx
);
2705 * Move the async itxs for a specified object to commit into sync lists.
2708 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
2711 itx_async_node_t
*ian
, ian_search
;
2715 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2718 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2721 * This is inherently racy, since there is nothing to prevent
2722 * the last synced txg from changing.
2724 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2725 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2727 mutex_enter(&itxg
->itxg_lock
);
2728 if (itxg
->itxg_txg
!= txg
) {
2729 mutex_exit(&itxg
->itxg_lock
);
2734 * If a foid is specified then find that node and append its
2735 * list. Otherwise walk the tree appending all the lists
2736 * to the sync list. We add to the end rather than the
2737 * beginning to ensure the create has happened.
2739 t
= &itxg
->itxg_itxs
->i_async_tree
;
2741 ian_search
.ia_foid
= foid
;
2742 ian
= avl_find(t
, &ian_search
, &where
);
2744 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2748 void *cookie
= NULL
;
2750 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2751 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2753 list_destroy(&ian
->ia_list
);
2754 kmem_free(ian
, sizeof (itx_async_node_t
));
2757 mutex_exit(&itxg
->itxg_lock
);
2762 * This function will prune commit itxs that are at the head of the
2763 * commit list (it won't prune past the first non-commit itx), and
2764 * either: a) attach them to the last lwb that's still pending
2765 * completion, or b) skip them altogether.
2767 * This is used as a performance optimization to prevent commit itxs
2768 * from generating new lwbs when it's unnecessary to do so.
2771 zil_prune_commit_list(zilog_t
*zilog
)
2775 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2777 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2778 lr_t
*lrc
= &itx
->itx_lr
;
2779 if (lrc
->lrc_txtype
!= TX_COMMIT
)
2782 mutex_enter(&zilog
->zl_lock
);
2784 lwb_t
*last_lwb
= zilog
->zl_last_lwb_opened
;
2785 if (last_lwb
== NULL
||
2786 last_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
) {
2788 * All of the itxs this waiter was waiting on
2789 * must have already completed (or there were
2790 * never any itx's for it to wait on), so it's
2791 * safe to skip this waiter and mark it done.
2793 zil_commit_waiter_skip(itx
->itx_private
);
2795 zil_commit_waiter_link_lwb(itx
->itx_private
, last_lwb
);
2798 mutex_exit(&zilog
->zl_lock
);
2800 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2801 zil_itx_destroy(itx
);
2804 IMPLY(itx
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2808 zil_commit_writer_stall(zilog_t
*zilog
)
2811 * When zio_alloc_zil() fails to allocate the next lwb block on
2812 * disk, we must call txg_wait_synced() to ensure all of the
2813 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2814 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2815 * to zil_process_commit_list()) will have to call zil_create(),
2816 * and start a new ZIL chain.
2818 * Since zil_alloc_zil() failed, the lwb that was previously
2819 * issued does not have a pointer to the "next" lwb on disk.
2820 * Thus, if another ZIL writer thread was to allocate the "next"
2821 * on-disk lwb, that block could be leaked in the event of a
2822 * crash (because the previous lwb on-disk would not point to
2825 * We must hold the zilog's zl_issuer_lock while we do this, to
2826 * ensure no new threads enter zil_process_commit_list() until
2827 * all lwb's in the zl_lwb_list have been synced and freed
2828 * (which is achieved via the txg_wait_synced() call).
2830 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2831 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2832 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
2836 zil_burst_done(zilog_t
*zilog
)
2838 if (!list_is_empty(&zilog
->zl_itx_commit_list
) ||
2839 zilog
->zl_cur_size
== 0)
2842 if (zilog
->zl_parallel
)
2843 zilog
->zl_parallel
--;
2845 uint_t r
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_BURSTS
- 1);
2846 zilog
->zl_prev_rotor
= r
;
2847 zilog
->zl_prev_opt
[r
] = zil_lwb_plan(zilog
, zilog
->zl_cur_size
,
2848 &zilog
->zl_prev_min
[r
]);
2850 zilog
->zl_cur_size
= 0;
2851 zilog
->zl_cur_max
= 0;
2852 zilog
->zl_cur_left
= 0;
2856 * This function will traverse the commit list, creating new lwbs as
2857 * needed, and committing the itxs from the commit list to these newly
2858 * created lwbs. Additionally, as a new lwb is created, the previous
2859 * lwb will be issued to the zio layer to be written to disk.
2862 zil_process_commit_list(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
, list_t
*ilwbs
)
2864 spa_t
*spa
= zilog
->zl_spa
;
2866 list_t nolwb_waiters
;
2870 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2873 * Return if there's nothing to commit before we dirty the fs by
2874 * calling zil_create().
2876 if (list_is_empty(&zilog
->zl_itx_commit_list
))
2879 list_create(&nolwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2880 list_create(&nolwb_waiters
, sizeof (zil_commit_waiter_t
),
2881 offsetof(zil_commit_waiter_t
, zcw_node
));
2883 lwb
= list_tail(&zilog
->zl_lwb_list
);
2885 lwb
= zil_create(zilog
);
2888 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2889 * have already been created (zl_lwb_list not empty).
2891 zil_commit_activate_saxattr_feature(zilog
);
2892 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
2893 lwb
->lwb_state
== LWB_STATE_OPENED
);
2896 * If the lwb is still opened, it means the workload is really
2897 * multi-threaded and we won the chance of write aggregation.
2898 * If it is not opened yet, but previous lwb is still not
2899 * flushed, it still means the workload is multi-threaded, but
2900 * there was too much time between the commits to aggregate, so
2901 * we try aggregation next times, but without too much hopes.
2903 if (lwb
->lwb_state
== LWB_STATE_OPENED
) {
2904 zilog
->zl_parallel
= ZIL_BURSTS
;
2905 } else if ((plwb
= list_prev(&zilog
->zl_lwb_list
, lwb
))
2906 != NULL
&& plwb
->lwb_state
!= LWB_STATE_FLUSH_DONE
) {
2907 zilog
->zl_parallel
= MAX(zilog
->zl_parallel
,
2912 while ((itx
= list_remove_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2913 lr_t
*lrc
= &itx
->itx_lr
;
2914 uint64_t txg
= lrc
->lrc_txg
;
2916 ASSERT3U(txg
, !=, 0);
2918 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2919 DTRACE_PROBE2(zil__process__commit__itx
,
2920 zilog_t
*, zilog
, itx_t
*, itx
);
2922 DTRACE_PROBE2(zil__process__normal__itx
,
2923 zilog_t
*, zilog
, itx_t
*, itx
);
2926 boolean_t synced
= txg
<= spa_last_synced_txg(spa
);
2927 boolean_t frozen
= txg
> spa_freeze_txg(spa
);
2930 * If the txg of this itx has already been synced out, then
2931 * we don't need to commit this itx to an lwb. This is
2932 * because the data of this itx will have already been
2933 * written to the main pool. This is inherently racy, and
2934 * it's still ok to commit an itx whose txg has already
2935 * been synced; this will result in a write that's
2936 * unnecessary, but will do no harm.
2938 * With that said, we always want to commit TX_COMMIT itxs
2939 * to an lwb, regardless of whether or not that itx's txg
2940 * has been synced out. We do this to ensure any OPENED lwb
2941 * will always have at least one zil_commit_waiter_t linked
2944 * As a counter-example, if we skipped TX_COMMIT itx's
2945 * whose txg had already been synced, the following
2946 * situation could occur if we happened to be racing with
2949 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2950 * itx's txg is 10 and the last synced txg is 9.
2951 * 2. spa_sync finishes syncing out txg 10.
2952 * 3. We move to the next itx in the list, it's a TX_COMMIT
2953 * whose txg is 10, so we skip it rather than committing
2954 * it to the lwb used in (1).
2956 * If the itx that is skipped in (3) is the last TX_COMMIT
2957 * itx in the commit list, than it's possible for the lwb
2958 * used in (1) to remain in the OPENED state indefinitely.
2960 * To prevent the above scenario from occurring, ensuring
2961 * that once an lwb is OPENED it will transition to ISSUED
2962 * and eventually DONE, we always commit TX_COMMIT itx's to
2963 * an lwb here, even if that itx's txg has already been
2966 * Finally, if the pool is frozen, we _always_ commit the
2967 * itx. The point of freezing the pool is to prevent data
2968 * from being written to the main pool via spa_sync, and
2969 * instead rely solely on the ZIL to persistently store the
2970 * data; i.e. when the pool is frozen, the last synced txg
2971 * value can't be trusted.
2973 if (frozen
|| !synced
|| lrc
->lrc_txtype
== TX_COMMIT
) {
2975 lwb
= zil_lwb_assign(zilog
, lwb
, itx
, ilwbs
);
2977 list_insert_tail(&nolwb_itxs
, itx
);
2978 } else if ((zcw
->zcw_lwb
!= NULL
&&
2979 zcw
->zcw_lwb
!= lwb
) || zcw
->zcw_done
) {
2981 * Our lwb is done, leave the rest of
2982 * itx list to somebody else who care.
2984 zilog
->zl_parallel
= ZIL_BURSTS
;
2985 zilog
->zl_cur_left
-=
2986 zil_itx_full_size(itx
);
2990 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2991 zil_commit_waiter_link_nolwb(
2992 itx
->itx_private
, &nolwb_waiters
);
2994 list_insert_tail(&nolwb_itxs
, itx
);
2996 zilog
->zl_cur_left
-= zil_itx_full_size(itx
);
2998 ASSERT3S(lrc
->lrc_txtype
, !=, TX_COMMIT
);
2999 zilog
->zl_cur_left
-= zil_itx_full_size(itx
);
3000 zil_itx_destroy(itx
);
3006 * This indicates zio_alloc_zil() failed to allocate the
3007 * "next" lwb on-disk. When this happens, we must stall
3008 * the ZIL write pipeline; see the comment within
3009 * zil_commit_writer_stall() for more details.
3011 while ((lwb
= list_remove_head(ilwbs
)) != NULL
)
3012 zil_lwb_write_issue(zilog
, lwb
);
3013 zil_commit_writer_stall(zilog
);
3016 * Additionally, we have to signal and mark the "nolwb"
3017 * waiters as "done" here, since without an lwb, we
3018 * can't do this via zil_lwb_flush_vdevs_done() like
3021 zil_commit_waiter_t
*zcw
;
3022 while ((zcw
= list_remove_head(&nolwb_waiters
)) != NULL
)
3023 zil_commit_waiter_skip(zcw
);
3026 * And finally, we have to destroy the itx's that
3027 * couldn't be committed to an lwb; this will also call
3028 * the itx's callback if one exists for the itx.
3030 while ((itx
= list_remove_head(&nolwb_itxs
)) != NULL
)
3031 zil_itx_destroy(itx
);
3033 ASSERT(list_is_empty(&nolwb_waiters
));
3034 ASSERT3P(lwb
, !=, NULL
);
3035 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
3036 lwb
->lwb_state
== LWB_STATE_OPENED
);
3039 * At this point, the ZIL block pointed at by the "lwb"
3040 * variable is in "new" or "opened" state.
3042 * If it's "new", then no itxs have been committed to it, so
3043 * there's no point in issuing its zio (i.e. it's "empty").
3045 * If it's "opened", then it contains one or more itxs that
3046 * eventually need to be committed to stable storage. In
3047 * this case we intentionally do not issue the lwb's zio
3048 * to disk yet, and instead rely on one of the following
3049 * two mechanisms for issuing the zio:
3051 * 1. Ideally, there will be more ZIL activity occurring on
3052 * the system, such that this function will be immediately
3053 * called again by different thread and this lwb will be
3054 * closed by zil_lwb_assign(). This way, the lwb will be
3055 * "full" when it is issued to disk, and we'll make use of
3056 * the lwb's size the best we can.
3058 * 2. If there isn't sufficient ZIL activity occurring on
3059 * the system, zil_commit_waiter() will close it and issue
3060 * the zio. If this occurs, the lwb is not guaranteed
3061 * to be "full" by the time its zio is issued, and means
3062 * the size of the lwb was "too large" given the amount
3063 * of ZIL activity occurring on the system at that time.
3065 * We do this for a couple of reasons:
3067 * 1. To try and reduce the number of IOPs needed to
3068 * write the same number of itxs. If an lwb has space
3069 * available in its buffer for more itxs, and more itxs
3070 * will be committed relatively soon (relative to the
3071 * latency of performing a write), then it's beneficial
3072 * to wait for these "next" itxs. This way, more itxs
3073 * can be committed to stable storage with fewer writes.
3075 * 2. To try and use the largest lwb block size that the
3076 * incoming rate of itxs can support. Again, this is to
3077 * try and pack as many itxs into as few lwbs as
3078 * possible, without significantly impacting the latency
3079 * of each individual itx.
3081 if (lwb
->lwb_state
== LWB_STATE_OPENED
&& !zilog
->zl_parallel
) {
3082 zil_burst_done(zilog
);
3083 list_insert_tail(ilwbs
, lwb
);
3084 lwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_NEW
);
3086 while ((lwb
= list_remove_head(ilwbs
)) != NULL
)
3087 zil_lwb_write_issue(zilog
, lwb
);
3088 zil_commit_writer_stall(zilog
);
3095 * This function is responsible for ensuring the passed in commit waiter
3096 * (and associated commit itx) is committed to an lwb. If the waiter is
3097 * not already committed to an lwb, all itxs in the zilog's queue of
3098 * itxs will be processed. The assumption is the passed in waiter's
3099 * commit itx will found in the queue just like the other non-commit
3100 * itxs, such that when the entire queue is processed, the waiter will
3101 * have been committed to an lwb.
3103 * The lwb associated with the passed in waiter is not guaranteed to
3104 * have been issued by the time this function completes. If the lwb is
3105 * not issued, we rely on future calls to zil_commit_writer() to issue
3106 * the lwb, or the timeout mechanism found in zil_commit_waiter().
3109 zil_commit_writer(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3115 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
3116 ASSERT(spa_writeable(zilog
->zl_spa
));
3118 list_create(&ilwbs
, sizeof (lwb_t
), offsetof(lwb_t
, lwb_issue_node
));
3119 mutex_enter(&zilog
->zl_issuer_lock
);
3121 if (zcw
->zcw_lwb
!= NULL
|| zcw
->zcw_done
) {
3123 * It's possible that, while we were waiting to acquire
3124 * the "zl_issuer_lock", another thread committed this
3125 * waiter to an lwb. If that occurs, we bail out early,
3126 * without processing any of the zilog's queue of itxs.
3128 * On certain workloads and system configurations, the
3129 * "zl_issuer_lock" can become highly contended. In an
3130 * attempt to reduce this contention, we immediately drop
3131 * the lock if the waiter has already been processed.
3133 * We've measured this optimization to reduce CPU spent
3134 * contending on this lock by up to 5%, using a system
3135 * with 32 CPUs, low latency storage (~50 usec writes),
3136 * and 1024 threads performing sync writes.
3141 ZIL_STAT_BUMP(zilog
, zil_commit_writer_count
);
3143 wtxg
= zil_get_commit_list(zilog
);
3144 zil_prune_commit_list(zilog
);
3145 zil_process_commit_list(zilog
, zcw
, &ilwbs
);
3148 mutex_exit(&zilog
->zl_issuer_lock
);
3149 while ((lwb
= list_remove_head(&ilwbs
)) != NULL
)
3150 zil_lwb_write_issue(zilog
, lwb
);
3151 list_destroy(&ilwbs
);
3156 zil_commit_waiter_timeout(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3158 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
3159 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
3160 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
3162 lwb_t
*lwb
= zcw
->zcw_lwb
;
3163 ASSERT3P(lwb
, !=, NULL
);
3164 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_NEW
);
3167 * If the lwb has already been issued by another thread, we can
3168 * immediately return since there's no work to be done (the
3169 * point of this function is to issue the lwb). Additionally, we
3170 * do this prior to acquiring the zl_issuer_lock, to avoid
3171 * acquiring it when it's not necessary to do so.
3173 if (lwb
->lwb_state
!= LWB_STATE_OPENED
)
3177 * In order to call zil_lwb_write_close() we must hold the
3178 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3179 * since we're already holding the commit waiter's "zcw_lock",
3180 * and those two locks are acquired in the opposite order
3183 mutex_exit(&zcw
->zcw_lock
);
3184 mutex_enter(&zilog
->zl_issuer_lock
);
3185 mutex_enter(&zcw
->zcw_lock
);
3188 * Since we just dropped and re-acquired the commit waiter's
3189 * lock, we have to re-check to see if the waiter was marked
3190 * "done" during that process. If the waiter was marked "done",
3191 * the "lwb" pointer is no longer valid (it can be free'd after
3192 * the waiter is marked "done"), so without this check we could
3193 * wind up with a use-after-free error below.
3195 if (zcw
->zcw_done
) {
3196 mutex_exit(&zilog
->zl_issuer_lock
);
3200 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
3203 * We've already checked this above, but since we hadn't acquired
3204 * the zilog's zl_issuer_lock, we have to perform this check a
3205 * second time while holding the lock.
3207 * We don't need to hold the zl_lock since the lwb cannot transition
3208 * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb
3209 * _can_ transition from CLOSED to DONE, but it's OK to race with
3210 * that transition since we treat the lwb the same, whether it's in
3211 * the CLOSED, ISSUED or DONE states.
3213 * The important thing, is we treat the lwb differently depending on
3214 * if it's OPENED or CLOSED, and block any other threads that might
3215 * attempt to close/issue this lwb. For that reason we hold the
3216 * zl_issuer_lock when checking the lwb_state; we must not call
3217 * zil_lwb_write_close() if the lwb had already been closed/issued.
3219 * See the comment above the lwb_state_t structure definition for
3220 * more details on the lwb states, and locking requirements.
3222 if (lwb
->lwb_state
!= LWB_STATE_OPENED
) {
3223 mutex_exit(&zilog
->zl_issuer_lock
);
3228 * We do not need zcw_lock once we hold zl_issuer_lock and know lwb
3229 * is still open. But we have to drop it to avoid a deadlock in case
3230 * callback of zio issued by zil_lwb_write_issue() try to get it,
3231 * while zil_lwb_write_issue() is blocked on attempt to issue next
3232 * lwb it found in LWB_STATE_READY state.
3234 mutex_exit(&zcw
->zcw_lock
);
3237 * As described in the comments above zil_commit_waiter() and
3238 * zil_process_commit_list(), we need to issue this lwb's zio
3239 * since we've reached the commit waiter's timeout and it still
3240 * hasn't been issued.
3242 zil_burst_done(zilog
);
3243 lwb_t
*nlwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_NEW
);
3245 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_CLOSED
);
3249 * When zil_lwb_write_close() returns NULL, this
3250 * indicates zio_alloc_zil() failed to allocate the
3251 * "next" lwb on-disk. When this occurs, the ZIL write
3252 * pipeline must be stalled; see the comment within the
3253 * zil_commit_writer_stall() function for more details.
3255 zil_lwb_write_issue(zilog
, lwb
);
3256 zil_commit_writer_stall(zilog
);
3257 mutex_exit(&zilog
->zl_issuer_lock
);
3259 mutex_exit(&zilog
->zl_issuer_lock
);
3260 zil_lwb_write_issue(zilog
, lwb
);
3262 mutex_enter(&zcw
->zcw_lock
);
3266 * This function is responsible for performing the following two tasks:
3268 * 1. its primary responsibility is to block until the given "commit
3269 * waiter" is considered "done".
3271 * 2. its secondary responsibility is to issue the zio for the lwb that
3272 * the given "commit waiter" is waiting on, if this function has
3273 * waited "long enough" and the lwb is still in the "open" state.
3275 * Given a sufficient amount of itxs being generated and written using
3276 * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3277 * function. If this does not occur, this secondary responsibility will
3278 * ensure the lwb is issued even if there is not other synchronous
3279 * activity on the system.
3281 * For more details, see zil_process_commit_list(); more specifically,
3282 * the comment at the bottom of that function.
3285 zil_commit_waiter(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3287 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
3288 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
3289 ASSERT(spa_writeable(zilog
->zl_spa
));
3291 mutex_enter(&zcw
->zcw_lock
);
3294 * The timeout is scaled based on the lwb latency to avoid
3295 * significantly impacting the latency of each individual itx.
3296 * For more details, see the comment at the bottom of the
3297 * zil_process_commit_list() function.
3299 int pct
= MAX(zfs_commit_timeout_pct
, 1);
3300 hrtime_t sleep
= (zilog
->zl_last_lwb_latency
* pct
) / 100;
3301 hrtime_t wakeup
= gethrtime() + sleep
;
3302 boolean_t timedout
= B_FALSE
;
3304 while (!zcw
->zcw_done
) {
3305 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
3307 lwb_t
*lwb
= zcw
->zcw_lwb
;
3310 * Usually, the waiter will have a non-NULL lwb field here,
3311 * but it's possible for it to be NULL as a result of
3312 * zil_commit() racing with spa_sync().
3314 * When zil_clean() is called, it's possible for the itxg
3315 * list (which may be cleaned via a taskq) to contain
3316 * commit itxs. When this occurs, the commit waiters linked
3317 * off of these commit itxs will not be committed to an
3318 * lwb. Additionally, these commit waiters will not be
3319 * marked done until zil_commit_waiter_skip() is called via
3322 * Thus, it's possible for this commit waiter (i.e. the
3323 * "zcw" variable) to be found in this "in between" state;
3324 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3325 * been skipped, so it's "zcw_done" field is still B_FALSE.
3327 IMPLY(lwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_NEW
);
3329 if (lwb
!= NULL
&& lwb
->lwb_state
== LWB_STATE_OPENED
) {
3330 ASSERT3B(timedout
, ==, B_FALSE
);
3333 * If the lwb hasn't been issued yet, then we
3334 * need to wait with a timeout, in case this
3335 * function needs to issue the lwb after the
3336 * timeout is reached; responsibility (2) from
3337 * the comment above this function.
3339 int rc
= cv_timedwait_hires(&zcw
->zcw_cv
,
3340 &zcw
->zcw_lock
, wakeup
, USEC2NSEC(1),
3341 CALLOUT_FLAG_ABSOLUTE
);
3343 if (rc
!= -1 || zcw
->zcw_done
)
3347 zil_commit_waiter_timeout(zilog
, zcw
);
3349 if (!zcw
->zcw_done
) {
3351 * If the commit waiter has already been
3352 * marked "done", it's possible for the
3353 * waiter's lwb structure to have already
3354 * been freed. Thus, we can only reliably
3355 * make these assertions if the waiter
3358 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
3359 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_OPENED
);
3363 * If the lwb isn't open, then it must have already
3364 * been issued. In that case, there's no need to
3365 * use a timeout when waiting for the lwb to
3368 * Additionally, if the lwb is NULL, the waiter
3369 * will soon be signaled and marked done via
3370 * zil_clean() and zil_itxg_clean(), so no timeout
3375 lwb
->lwb_state
== LWB_STATE_CLOSED
||
3376 lwb
->lwb_state
== LWB_STATE_READY
||
3377 lwb
->lwb_state
== LWB_STATE_ISSUED
||
3378 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
3379 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
3380 cv_wait(&zcw
->zcw_cv
, &zcw
->zcw_lock
);
3384 mutex_exit(&zcw
->zcw_lock
);
3387 static zil_commit_waiter_t
*
3388 zil_alloc_commit_waiter(void)
3390 zil_commit_waiter_t
*zcw
= kmem_cache_alloc(zil_zcw_cache
, KM_SLEEP
);
3392 cv_init(&zcw
->zcw_cv
, NULL
, CV_DEFAULT
, NULL
);
3393 mutex_init(&zcw
->zcw_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3394 list_link_init(&zcw
->zcw_node
);
3395 zcw
->zcw_lwb
= NULL
;
3396 zcw
->zcw_done
= B_FALSE
;
3397 zcw
->zcw_zio_error
= 0;
3403 zil_free_commit_waiter(zil_commit_waiter_t
*zcw
)
3405 ASSERT(!list_link_active(&zcw
->zcw_node
));
3406 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
3407 ASSERT3B(zcw
->zcw_done
, ==, B_TRUE
);
3408 mutex_destroy(&zcw
->zcw_lock
);
3409 cv_destroy(&zcw
->zcw_cv
);
3410 kmem_cache_free(zil_zcw_cache
, zcw
);
3414 * This function is used to create a TX_COMMIT itx and assign it. This
3415 * way, it will be linked into the ZIL's list of synchronous itxs, and
3416 * then later committed to an lwb (or skipped) when
3417 * zil_process_commit_list() is called.
3420 zil_commit_itx_assign(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3422 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
3425 * Since we are not going to create any new dirty data, and we
3426 * can even help with clearing the existing dirty data, we
3427 * should not be subject to the dirty data based delays. We
3428 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3430 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
3432 itx_t
*itx
= zil_itx_create(TX_COMMIT
, sizeof (lr_t
));
3433 itx
->itx_sync
= B_TRUE
;
3434 itx
->itx_private
= zcw
;
3436 zil_itx_assign(zilog
, itx
, tx
);
3442 * Commit ZFS Intent Log transactions (itxs) to stable storage.
3444 * When writing ZIL transactions to the on-disk representation of the
3445 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3446 * itxs can be committed to a single lwb. Once a lwb is written and
3447 * committed to stable storage (i.e. the lwb is written, and vdevs have
3448 * been flushed), each itx that was committed to that lwb is also
3449 * considered to be committed to stable storage.
3451 * When an itx is committed to an lwb, the log record (lr_t) contained
3452 * by the itx is copied into the lwb's zio buffer, and once this buffer
3453 * is written to disk, it becomes an on-disk ZIL block.
3455 * As itxs are generated, they're inserted into the ZIL's queue of
3456 * uncommitted itxs. The semantics of zil_commit() are such that it will
3457 * block until all itxs that were in the queue when it was called, are
3458 * committed to stable storage.
3460 * If "foid" is zero, this means all "synchronous" and "asynchronous"
3461 * itxs, for all objects in the dataset, will be committed to stable
3462 * storage prior to zil_commit() returning. If "foid" is non-zero, all
3463 * "synchronous" itxs for all objects, but only "asynchronous" itxs
3464 * that correspond to the foid passed in, will be committed to stable
3465 * storage prior to zil_commit() returning.
3467 * Generally speaking, when zil_commit() is called, the consumer doesn't
3468 * actually care about _all_ of the uncommitted itxs. Instead, they're
3469 * simply trying to waiting for a specific itx to be committed to disk,
3470 * but the interface(s) for interacting with the ZIL don't allow such
3471 * fine-grained communication. A better interface would allow a consumer
3472 * to create and assign an itx, and then pass a reference to this itx to
3473 * zil_commit(); such that zil_commit() would return as soon as that
3474 * specific itx was committed to disk (instead of waiting for _all_
3475 * itxs to be committed).
3477 * When a thread calls zil_commit() a special "commit itx" will be
3478 * generated, along with a corresponding "waiter" for this commit itx.
3479 * zil_commit() will wait on this waiter's CV, such that when the waiter
3480 * is marked done, and signaled, zil_commit() will return.
3482 * This commit itx is inserted into the queue of uncommitted itxs. This
3483 * provides an easy mechanism for determining which itxs were in the
3484 * queue prior to zil_commit() having been called, and which itxs were
3485 * added after zil_commit() was called.
3487 * The commit itx is special; it doesn't have any on-disk representation.
3488 * When a commit itx is "committed" to an lwb, the waiter associated
3489 * with it is linked onto the lwb's list of waiters. Then, when that lwb
3490 * completes, each waiter on the lwb's list is marked done and signaled
3491 * -- allowing the thread waiting on the waiter to return from zil_commit().
3493 * It's important to point out a few critical factors that allow us
3494 * to make use of the commit itxs, commit waiters, per-lwb lists of
3495 * commit waiters, and zio completion callbacks like we're doing:
3497 * 1. The list of waiters for each lwb is traversed, and each commit
3498 * waiter is marked "done" and signaled, in the zio completion
3499 * callback of the lwb's zio[*].
3501 * * Actually, the waiters are signaled in the zio completion
3502 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3503 * that are sent to the vdevs upon completion of the lwb zio.
3505 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
3506 * itxs, the order in which they are inserted is preserved[*]; as
3507 * itxs are added to the queue, they are added to the tail of
3508 * in-memory linked lists.
3510 * When committing the itxs to lwbs (to be written to disk), they
3511 * are committed in the same order in which the itxs were added to
3512 * the uncommitted queue's linked list(s); i.e. the linked list of
3513 * itxs to commit is traversed from head to tail, and each itx is
3514 * committed to an lwb in that order.
3518 * - the order of "sync" itxs is preserved w.r.t. other
3519 * "sync" itxs, regardless of the corresponding objects.
3520 * - the order of "async" itxs is preserved w.r.t. other
3521 * "async" itxs corresponding to the same object.
3522 * - the order of "async" itxs is *not* preserved w.r.t. other
3523 * "async" itxs corresponding to different objects.
3524 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
3525 * versa) is *not* preserved, even for itxs that correspond
3526 * to the same object.
3528 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
3529 * zil_get_commit_list(), and zil_process_commit_list().
3531 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
3532 * lwb cannot be considered committed to stable storage, until its
3533 * "previous" lwb is also committed to stable storage. This fact,
3534 * coupled with the fact described above, means that itxs are
3535 * committed in (roughly) the order in which they were generated.
3536 * This is essential because itxs are dependent on prior itxs.
3537 * Thus, we *must not* deem an itx as being committed to stable
3538 * storage, until *all* prior itxs have also been committed to
3541 * To enforce this ordering of lwb zio's, while still leveraging as
3542 * much of the underlying storage performance as possible, we rely
3543 * on two fundamental concepts:
3545 * 1. The creation and issuance of lwb zio's is protected by
3546 * the zilog's "zl_issuer_lock", which ensures only a single
3547 * thread is creating and/or issuing lwb's at a time
3548 * 2. The "previous" lwb is a child of the "current" lwb
3549 * (leveraging the zio parent-child dependency graph)
3551 * By relying on this parent-child zio relationship, we can have
3552 * many lwb zio's concurrently issued to the underlying storage,
3553 * but the order in which they complete will be the same order in
3554 * which they were created.
3557 zil_commit(zilog_t
*zilog
, uint64_t foid
)
3560 * We should never attempt to call zil_commit on a snapshot for
3561 * a couple of reasons:
3563 * 1. A snapshot may never be modified, thus it cannot have any
3564 * in-flight itxs that would have modified the dataset.
3566 * 2. By design, when zil_commit() is called, a commit itx will
3567 * be assigned to this zilog; as a result, the zilog will be
3568 * dirtied. We must not dirty the zilog of a snapshot; there's
3569 * checks in the code that enforce this invariant, and will
3570 * cause a panic if it's not upheld.
3572 ASSERT3B(dmu_objset_is_snapshot(zilog
->zl_os
), ==, B_FALSE
);
3574 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3577 if (!spa_writeable(zilog
->zl_spa
)) {
3579 * If the SPA is not writable, there should never be any
3580 * pending itxs waiting to be committed to disk. If that
3581 * weren't true, we'd skip writing those itxs out, and
3582 * would break the semantics of zil_commit(); thus, we're
3583 * verifying that truth before we return to the caller.
3585 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3586 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3587 for (int i
= 0; i
< TXG_SIZE
; i
++)
3588 ASSERT3P(zilog
->zl_itxg
[i
].itxg_itxs
, ==, NULL
);
3593 * If the ZIL is suspended, we don't want to dirty it by calling
3594 * zil_commit_itx_assign() below, nor can we write out
3595 * lwbs like would be done in zil_commit_write(). Thus, we
3596 * simply rely on txg_wait_synced() to maintain the necessary
3597 * semantics, and avoid calling those functions altogether.
3599 if (zilog
->zl_suspend
> 0) {
3600 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3604 zil_commit_impl(zilog
, foid
);
3608 zil_commit_impl(zilog_t
*zilog
, uint64_t foid
)
3610 ZIL_STAT_BUMP(zilog
, zil_commit_count
);
3613 * Move the "async" itxs for the specified foid to the "sync"
3614 * queues, such that they will be later committed (or skipped)
3615 * to an lwb when zil_process_commit_list() is called.
3617 * Since these "async" itxs must be committed prior to this
3618 * call to zil_commit returning, we must perform this operation
3619 * before we call zil_commit_itx_assign().
3621 zil_async_to_sync(zilog
, foid
);
3624 * We allocate a new "waiter" structure which will initially be
3625 * linked to the commit itx using the itx's "itx_private" field.
3626 * Since the commit itx doesn't represent any on-disk state,
3627 * when it's committed to an lwb, rather than copying the its
3628 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3629 * added to the lwb's list of waiters. Then, when the lwb is
3630 * committed to stable storage, each waiter in the lwb's list of
3631 * waiters will be marked "done", and signalled.
3633 * We must create the waiter and assign the commit itx prior to
3634 * calling zil_commit_writer(), or else our specific commit itx
3635 * is not guaranteed to be committed to an lwb prior to calling
3636 * zil_commit_waiter().
3638 zil_commit_waiter_t
*zcw
= zil_alloc_commit_waiter();
3639 zil_commit_itx_assign(zilog
, zcw
);
3641 uint64_t wtxg
= zil_commit_writer(zilog
, zcw
);
3642 zil_commit_waiter(zilog
, zcw
);
3644 if (zcw
->zcw_zio_error
!= 0) {
3646 * If there was an error writing out the ZIL blocks that
3647 * this thread is waiting on, then we fallback to
3648 * relying on spa_sync() to write out the data this
3649 * thread is waiting on. Obviously this has performance
3650 * implications, but the expectation is for this to be
3651 * an exceptional case, and shouldn't occur often.
3653 DTRACE_PROBE2(zil__commit__io__error
,
3654 zilog_t
*, zilog
, zil_commit_waiter_t
*, zcw
);
3655 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3656 } else if (wtxg
!= 0) {
3657 txg_wait_synced(zilog
->zl_dmu_pool
, wtxg
);
3660 zil_free_commit_waiter(zcw
);
3664 * Called in syncing context to free committed log blocks and update log header.
3667 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
3669 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
3670 uint64_t txg
= dmu_tx_get_txg(tx
);
3671 spa_t
*spa
= zilog
->zl_spa
;
3672 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
3676 * We don't zero out zl_destroy_txg, so make sure we don't try
3677 * to destroy it twice.
3679 if (spa_sync_pass(spa
) != 1)
3682 zil_lwb_flush_wait_all(zilog
, txg
);
3684 mutex_enter(&zilog
->zl_lock
);
3686 ASSERT(zilog
->zl_stop_sync
== 0);
3688 if (*replayed_seq
!= 0) {
3689 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
3690 zh
->zh_replay_seq
= *replayed_seq
;
3694 if (zilog
->zl_destroy_txg
== txg
) {
3695 blkptr_t blk
= zh
->zh_log
;
3696 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
3698 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3700 memset(zh
, 0, sizeof (zil_header_t
));
3701 memset(zilog
->zl_replayed_seq
, 0,
3702 sizeof (zilog
->zl_replayed_seq
));
3704 if (zilog
->zl_keep_first
) {
3706 * If this block was part of log chain that couldn't
3707 * be claimed because a device was missing during
3708 * zil_claim(), but that device later returns,
3709 * then this block could erroneously appear valid.
3710 * To guard against this, assign a new GUID to the new
3711 * log chain so it doesn't matter what blk points to.
3713 zil_init_log_chain(zilog
, &blk
);
3717 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3718 * records. So, deactivate the feature for this dataset.
3719 * We activate it again when we start a new ZIL chain.
3721 if (dsl_dataset_feature_is_active(ds
,
3722 SPA_FEATURE_ZILSAXATTR
))
3723 dsl_dataset_deactivate_feature(ds
,
3724 SPA_FEATURE_ZILSAXATTR
, tx
);
3728 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
3729 zh
->zh_log
= lwb
->lwb_blk
;
3730 if (lwb
->lwb_state
!= LWB_STATE_FLUSH_DONE
||
3731 lwb
->lwb_alloc_txg
> txg
|| lwb
->lwb_max_txg
> txg
)
3733 list_remove(&zilog
->zl_lwb_list
, lwb
);
3734 if (!BP_IS_HOLE(&lwb
->lwb_blk
))
3735 zio_free(spa
, txg
, &lwb
->lwb_blk
);
3736 zil_free_lwb(zilog
, lwb
);
3739 * If we don't have anything left in the lwb list then
3740 * we've had an allocation failure and we need to zero
3741 * out the zil_header blkptr so that we don't end
3742 * up freeing the same block twice.
3744 if (list_is_empty(&zilog
->zl_lwb_list
))
3745 BP_ZERO(&zh
->zh_log
);
3748 mutex_exit(&zilog
->zl_lock
);
3752 zil_lwb_cons(void *vbuf
, void *unused
, int kmflag
)
3754 (void) unused
, (void) kmflag
;
3756 list_create(&lwb
->lwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
3757 list_create(&lwb
->lwb_waiters
, sizeof (zil_commit_waiter_t
),
3758 offsetof(zil_commit_waiter_t
, zcw_node
));
3759 avl_create(&lwb
->lwb_vdev_tree
, zil_lwb_vdev_compare
,
3760 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
3761 mutex_init(&lwb
->lwb_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3766 zil_lwb_dest(void *vbuf
, void *unused
)
3770 mutex_destroy(&lwb
->lwb_vdev_lock
);
3771 avl_destroy(&lwb
->lwb_vdev_tree
);
3772 list_destroy(&lwb
->lwb_waiters
);
3773 list_destroy(&lwb
->lwb_itxs
);
3779 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
3780 sizeof (lwb_t
), 0, zil_lwb_cons
, zil_lwb_dest
, NULL
, NULL
, NULL
, 0);
3782 zil_zcw_cache
= kmem_cache_create("zil_zcw_cache",
3783 sizeof (zil_commit_waiter_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
3785 zil_sums_init(&zil_sums_global
);
3786 zil_kstats_global
= kstat_create("zfs", 0, "zil", "misc",
3787 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
3788 KSTAT_FLAG_VIRTUAL
);
3790 if (zil_kstats_global
!= NULL
) {
3791 zil_kstats_global
->ks_data
= &zil_stats
;
3792 zil_kstats_global
->ks_update
= zil_kstats_global_update
;
3793 zil_kstats_global
->ks_private
= NULL
;
3794 kstat_install(zil_kstats_global
);
3801 kmem_cache_destroy(zil_zcw_cache
);
3802 kmem_cache_destroy(zil_lwb_cache
);
3804 if (zil_kstats_global
!= NULL
) {
3805 kstat_delete(zil_kstats_global
);
3806 zil_kstats_global
= NULL
;
3809 zil_sums_fini(&zil_sums_global
);
3813 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
3815 zilog
->zl_sync
= sync
;
3819 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
3821 zilog
->zl_logbias
= logbias
;
3825 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
3829 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
3831 zilog
->zl_header
= zh_phys
;
3833 zilog
->zl_spa
= dmu_objset_spa(os
);
3834 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
3835 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
3836 zilog
->zl_logbias
= dmu_objset_logbias(os
);
3837 zilog
->zl_sync
= dmu_objset_syncprop(os
);
3838 zilog
->zl_dirty_max_txg
= 0;
3839 zilog
->zl_last_lwb_opened
= NULL
;
3840 zilog
->zl_last_lwb_latency
= 0;
3841 zilog
->zl_max_block_size
= MIN(MAX(P2ALIGN_TYPED(zil_maxblocksize
,
3842 ZIL_MIN_BLKSZ
, uint64_t), ZIL_MIN_BLKSZ
),
3843 spa_maxblocksize(dmu_objset_spa(os
)));
3845 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3846 mutex_init(&zilog
->zl_issuer_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3847 mutex_init(&zilog
->zl_lwb_io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3849 for (int i
= 0; i
< TXG_SIZE
; i
++) {
3850 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
3851 MUTEX_DEFAULT
, NULL
);
3854 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
3855 offsetof(lwb_t
, lwb_node
));
3857 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
3858 offsetof(itx_t
, itx_node
));
3860 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
3861 cv_init(&zilog
->zl_lwb_io_cv
, NULL
, CV_DEFAULT
, NULL
);
3863 for (int i
= 0; i
< ZIL_BURSTS
; i
++) {
3864 zilog
->zl_prev_opt
[i
] = zilog
->zl_max_block_size
-
3865 sizeof (zil_chain_t
);
3872 zil_free(zilog_t
*zilog
)
3876 zilog
->zl_stop_sync
= 1;
3878 ASSERT0(zilog
->zl_suspend
);
3879 ASSERT0(zilog
->zl_suspending
);
3881 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3882 list_destroy(&zilog
->zl_lwb_list
);
3884 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
3885 list_destroy(&zilog
->zl_itx_commit_list
);
3887 for (i
= 0; i
< TXG_SIZE
; i
++) {
3889 * It's possible for an itx to be generated that doesn't dirty
3890 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3891 * callback to remove the entry. We remove those here.
3893 * Also free up the ziltest itxs.
3895 if (zilog
->zl_itxg
[i
].itxg_itxs
)
3896 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
3897 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
3900 mutex_destroy(&zilog
->zl_issuer_lock
);
3901 mutex_destroy(&zilog
->zl_lock
);
3902 mutex_destroy(&zilog
->zl_lwb_io_lock
);
3904 cv_destroy(&zilog
->zl_cv_suspend
);
3905 cv_destroy(&zilog
->zl_lwb_io_cv
);
3907 kmem_free(zilog
, sizeof (zilog_t
));
3911 * Open an intent log.
3914 zil_open(objset_t
*os
, zil_get_data_t
*get_data
, zil_sums_t
*zil_sums
)
3916 zilog_t
*zilog
= dmu_objset_zil(os
);
3918 ASSERT3P(zilog
->zl_get_data
, ==, NULL
);
3919 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3920 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3922 zilog
->zl_get_data
= get_data
;
3923 zilog
->zl_sums
= zil_sums
;
3929 * Close an intent log.
3932 zil_close(zilog_t
*zilog
)
3937 if (!dmu_objset_is_snapshot(zilog
->zl_os
)) {
3938 zil_commit(zilog
, 0);
3940 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3941 ASSERT0(zilog
->zl_dirty_max_txg
);
3942 ASSERT3B(zilog_is_dirty(zilog
), ==, B_FALSE
);
3945 mutex_enter(&zilog
->zl_lock
);
3946 txg
= zilog
->zl_dirty_max_txg
;
3947 lwb
= list_tail(&zilog
->zl_lwb_list
);
3949 txg
= MAX(txg
, lwb
->lwb_alloc_txg
);
3950 txg
= MAX(txg
, lwb
->lwb_max_txg
);
3952 mutex_exit(&zilog
->zl_lock
);
3955 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3956 * on the time when the dmu_tx transaction is assigned in
3957 * zil_lwb_write_issue().
3959 mutex_enter(&zilog
->zl_lwb_io_lock
);
3960 txg
= MAX(zilog
->zl_lwb_max_issued_txg
, txg
);
3961 mutex_exit(&zilog
->zl_lwb_io_lock
);
3964 * We need to use txg_wait_synced() to wait until that txg is synced.
3965 * zil_sync() will guarantee all lwbs up to that txg have been
3966 * written out, flushed, and cleaned.
3969 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
3971 if (zilog_is_dirty(zilog
))
3972 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog
,
3974 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
3975 VERIFY(!zilog_is_dirty(zilog
));
3977 zilog
->zl_get_data
= NULL
;
3980 * We should have only one lwb left on the list; remove it now.
3982 mutex_enter(&zilog
->zl_lock
);
3983 lwb
= list_remove_head(&zilog
->zl_lwb_list
);
3985 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3986 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_NEW
);
3987 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
3988 zil_free_lwb(zilog
, lwb
);
3990 mutex_exit(&zilog
->zl_lock
);
3993 static const char *suspend_tag
= "zil suspending";
3996 * Suspend an intent log. While in suspended mode, we still honor
3997 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3998 * On old version pools, we suspend the log briefly when taking a
3999 * snapshot so that it will have an empty intent log.
4001 * Long holds are not really intended to be used the way we do here --
4002 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
4003 * could fail. Therefore we take pains to only put a long hold if it is
4004 * actually necessary. Fortunately, it will only be necessary if the
4005 * objset is currently mounted (or the ZVOL equivalent). In that case it
4006 * will already have a long hold, so we are not really making things any worse.
4008 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
4009 * zvol_state_t), and use their mechanism to prevent their hold from being
4010 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
4013 * if cookiep == NULL, this does both the suspend & resume.
4014 * Otherwise, it returns with the dataset "long held", and the cookie
4015 * should be passed into zil_resume().
4018 zil_suspend(const char *osname
, void **cookiep
)
4022 const zil_header_t
*zh
;
4025 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
4028 zilog
= dmu_objset_zil(os
);
4030 mutex_enter(&zilog
->zl_lock
);
4031 zh
= zilog
->zl_header
;
4033 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
4034 mutex_exit(&zilog
->zl_lock
);
4035 dmu_objset_rele(os
, suspend_tag
);
4036 return (SET_ERROR(EBUSY
));
4040 * Don't put a long hold in the cases where we can avoid it. This
4041 * is when there is no cookie so we are doing a suspend & resume
4042 * (i.e. called from zil_vdev_offline()), and there's nothing to do
4043 * for the suspend because it's already suspended, or there's no ZIL.
4045 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
4046 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
4047 mutex_exit(&zilog
->zl_lock
);
4048 dmu_objset_rele(os
, suspend_tag
);
4052 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
4053 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
4055 zilog
->zl_suspend
++;
4057 if (zilog
->zl_suspend
> 1) {
4059 * Someone else is already suspending it.
4060 * Just wait for them to finish.
4063 while (zilog
->zl_suspending
)
4064 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
4065 mutex_exit(&zilog
->zl_lock
);
4067 if (cookiep
== NULL
)
4075 * If there is no pointer to an on-disk block, this ZIL must not
4076 * be active (e.g. filesystem not mounted), so there's nothing
4079 if (BP_IS_HOLE(&zh
->zh_log
)) {
4080 ASSERT(cookiep
!= NULL
); /* fast path already handled */
4083 mutex_exit(&zilog
->zl_lock
);
4088 * The ZIL has work to do. Ensure that the associated encryption
4089 * key will remain mapped while we are committing the log by
4090 * grabbing a reference to it. If the key isn't loaded we have no
4091 * choice but to return an error until the wrapping key is loaded.
4093 if (os
->os_encrypted
&&
4094 dsl_dataset_create_key_mapping(dmu_objset_ds(os
)) != 0) {
4095 zilog
->zl_suspend
--;
4096 mutex_exit(&zilog
->zl_lock
);
4097 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
4098 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
4099 return (SET_ERROR(EACCES
));
4102 zilog
->zl_suspending
= B_TRUE
;
4103 mutex_exit(&zilog
->zl_lock
);
4106 * We need to use zil_commit_impl to ensure we wait for all
4107 * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed
4108 * to disk before proceeding. If we used zil_commit instead, it
4109 * would just call txg_wait_synced(), because zl_suspend is set.
4110 * txg_wait_synced() doesn't wait for these lwb's to be
4111 * LWB_STATE_FLUSH_DONE before returning.
4113 zil_commit_impl(zilog
, 0);
4116 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
4117 * use txg_wait_synced() to ensure the data from the zilog has
4118 * migrated to the main pool before calling zil_destroy().
4120 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
4122 zil_destroy(zilog
, B_FALSE
);
4124 mutex_enter(&zilog
->zl_lock
);
4125 zilog
->zl_suspending
= B_FALSE
;
4126 cv_broadcast(&zilog
->zl_cv_suspend
);
4127 mutex_exit(&zilog
->zl_lock
);
4129 if (os
->os_encrypted
)
4130 dsl_dataset_remove_key_mapping(dmu_objset_ds(os
));
4132 if (cookiep
== NULL
)
4140 zil_resume(void *cookie
)
4142 objset_t
*os
= cookie
;
4143 zilog_t
*zilog
= dmu_objset_zil(os
);
4145 mutex_enter(&zilog
->zl_lock
);
4146 ASSERT(zilog
->zl_suspend
!= 0);
4147 zilog
->zl_suspend
--;
4148 mutex_exit(&zilog
->zl_lock
);
4149 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
4150 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
4153 typedef struct zil_replay_arg
{
4154 zil_replay_func_t
*const *zr_replay
;
4156 boolean_t zr_byteswap
;
4161 zil_replay_error(zilog_t
*zilog
, const lr_t
*lr
, int error
)
4163 char name
[ZFS_MAX_DATASET_NAME_LEN
];
4165 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
4167 dmu_objset_name(zilog
->zl_os
, name
);
4169 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
4170 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
4171 (u_longlong_t
)lr
->lrc_seq
,
4172 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
4173 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
4179 zil_replay_log_record(zilog_t
*zilog
, const lr_t
*lr
, void *zra
,
4182 zil_replay_arg_t
*zr
= zra
;
4183 const zil_header_t
*zh
= zilog
->zl_header
;
4184 uint64_t reclen
= lr
->lrc_reclen
;
4185 uint64_t txtype
= lr
->lrc_txtype
;
4188 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
4190 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
4193 if (lr
->lrc_txg
< claim_txg
) /* already committed */
4196 /* Strip case-insensitive bit, still present in log record */
4199 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
4200 return (zil_replay_error(zilog
, lr
, EINVAL
));
4203 * If this record type can be logged out of order, the object
4204 * (lr_foid) may no longer exist. That's legitimate, not an error.
4206 if (TX_OOO(txtype
)) {
4207 error
= dmu_object_info(zilog
->zl_os
,
4208 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
4209 if (error
== ENOENT
|| error
== EEXIST
)
4214 * Make a copy of the data so we can revise and extend it.
4216 memcpy(zr
->zr_lr
, lr
, reclen
);
4219 * If this is a TX_WRITE with a blkptr, suck in the data.
4221 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
4222 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
4223 zr
->zr_lr
+ reclen
);
4225 return (zil_replay_error(zilog
, lr
, error
));
4229 * The log block containing this lr may have been byteswapped
4230 * so that we can easily examine common fields like lrc_txtype.
4231 * However, the log is a mix of different record types, and only the
4232 * replay vectors know how to byteswap their records. Therefore, if
4233 * the lr was byteswapped, undo it before invoking the replay vector.
4235 if (zr
->zr_byteswap
)
4236 byteswap_uint64_array(zr
->zr_lr
, reclen
);
4239 * We must now do two things atomically: replay this log record,
4240 * and update the log header sequence number to reflect the fact that
4241 * we did so. At the end of each replay function the sequence number
4242 * is updated if we are in replay mode.
4244 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
4247 * The DMU's dnode layer doesn't see removes until the txg
4248 * commits, so a subsequent claim can spuriously fail with
4249 * EEXIST. So if we receive any error we try syncing out
4250 * any removes then retry the transaction. Note that we
4251 * specify B_FALSE for byteswap now, so we don't do it twice.
4253 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
4254 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
4256 return (zil_replay_error(zilog
, lr
, error
));
4262 zil_incr_blks(zilog_t
*zilog
, const blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
4264 (void) bp
, (void) arg
, (void) claim_txg
;
4266 zilog
->zl_replay_blks
++;
4272 * If this dataset has a non-empty intent log, replay it and destroy it.
4273 * Return B_TRUE if there were any entries to replay.
4276 zil_replay(objset_t
*os
, void *arg
,
4277 zil_replay_func_t
*const replay_func
[TX_MAX_TYPE
])
4279 zilog_t
*zilog
= dmu_objset_zil(os
);
4280 const zil_header_t
*zh
= zilog
->zl_header
;
4281 zil_replay_arg_t zr
;
4283 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
4284 return (zil_destroy(zilog
, B_TRUE
));
4287 zr
.zr_replay
= replay_func
;
4289 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
4290 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
4293 * Wait for in-progress removes to sync before starting replay.
4295 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
4297 zilog
->zl_replay
= B_TRUE
;
4298 zilog
->zl_replay_time
= ddi_get_lbolt();
4299 ASSERT(zilog
->zl_replay_blks
== 0);
4300 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
4301 zh
->zh_claim_txg
, B_TRUE
);
4302 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
4304 zil_destroy(zilog
, B_FALSE
);
4305 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
4306 zilog
->zl_replay
= B_FALSE
;
4312 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
4314 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
4317 if (zilog
->zl_replay
) {
4318 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
4319 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
4320 zilog
->zl_replaying_seq
;
4328 zil_reset(const char *osname
, void *arg
)
4332 int error
= zil_suspend(osname
, NULL
);
4333 /* EACCES means crypto key not loaded */
4334 if ((error
== EACCES
) || (error
== EBUSY
))
4335 return (SET_ERROR(error
));
4337 return (SET_ERROR(EEXIST
));
4341 EXPORT_SYMBOL(zil_alloc
);
4342 EXPORT_SYMBOL(zil_free
);
4343 EXPORT_SYMBOL(zil_open
);
4344 EXPORT_SYMBOL(zil_close
);
4345 EXPORT_SYMBOL(zil_replay
);
4346 EXPORT_SYMBOL(zil_replaying
);
4347 EXPORT_SYMBOL(zil_destroy
);
4348 EXPORT_SYMBOL(zil_destroy_sync
);
4349 EXPORT_SYMBOL(zil_itx_create
);
4350 EXPORT_SYMBOL(zil_itx_destroy
);
4351 EXPORT_SYMBOL(zil_itx_assign
);
4352 EXPORT_SYMBOL(zil_commit
);
4353 EXPORT_SYMBOL(zil_claim
);
4354 EXPORT_SYMBOL(zil_check_log_chain
);
4355 EXPORT_SYMBOL(zil_sync
);
4356 EXPORT_SYMBOL(zil_clean
);
4357 EXPORT_SYMBOL(zil_suspend
);
4358 EXPORT_SYMBOL(zil_resume
);
4359 EXPORT_SYMBOL(zil_lwb_add_block
);
4360 EXPORT_SYMBOL(zil_bp_tree_add
);
4361 EXPORT_SYMBOL(zil_set_sync
);
4362 EXPORT_SYMBOL(zil_set_logbias
);
4363 EXPORT_SYMBOL(zil_sums_init
);
4364 EXPORT_SYMBOL(zil_sums_fini
);
4365 EXPORT_SYMBOL(zil_kstat_values_update
);
4367 ZFS_MODULE_PARAM(zfs
, zfs_
, commit_timeout_pct
, UINT
, ZMOD_RW
,
4368 "ZIL block open timeout percentage");
4370 ZFS_MODULE_PARAM(zfs_zil
, zil_
, replay_disable
, INT
, ZMOD_RW
,
4371 "Disable intent logging replay");
4373 ZFS_MODULE_PARAM(zfs_zil
, zil_
, nocacheflush
, INT
, ZMOD_RW
,
4374 "Disable ZIL cache flushes");
4376 ZFS_MODULE_PARAM(zfs_zil
, zil_
, slog_bulk
, U64
, ZMOD_RW
,
4377 "Limit in bytes slog sync writes per commit");
4379 ZFS_MODULE_PARAM(zfs_zil
, zil_
, maxblocksize
, UINT
, ZMOD_RW
,
4380 "Limit in bytes of ZIL log block size");
4382 ZFS_MODULE_PARAM(zfs_zil
, zil_
, maxcopied
, UINT
, ZMOD_RW
,
4383 "Limit in bytes WR_COPIED size");