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
);
149 zil_bp_compare(const void *x1
, const void *x2
)
151 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
152 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
154 int cmp
= TREE_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
158 return (TREE_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
162 zil_bp_tree_init(zilog_t
*zilog
)
164 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
165 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
169 zil_bp_tree_fini(zilog_t
*zilog
)
171 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
175 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
176 kmem_free(zn
, sizeof (zil_bp_node_t
));
182 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
184 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
189 if (BP_IS_EMBEDDED(bp
))
192 dva
= BP_IDENTITY(bp
);
194 if (avl_find(t
, dva
, &where
) != NULL
)
195 return (SET_ERROR(EEXIST
));
197 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
199 avl_insert(t
, zn
, where
);
204 static zil_header_t
*
205 zil_header_in_syncing_context(zilog_t
*zilog
)
207 return ((zil_header_t
*)zilog
->zl_header
);
211 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
213 zio_cksum_t
*zc
= &bp
->blk_cksum
;
215 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_0
],
216 sizeof (zc
->zc_word
[ZIL_ZC_GUID_0
]));
217 (void) random_get_pseudo_bytes((void *)&zc
->zc_word
[ZIL_ZC_GUID_1
],
218 sizeof (zc
->zc_word
[ZIL_ZC_GUID_1
]));
219 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
220 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
224 zil_kstats_global_update(kstat_t
*ksp
, int rw
)
226 zil_kstat_values_t
*zs
= ksp
->ks_data
;
227 ASSERT3P(&zil_stats
, ==, zs
);
229 if (rw
== KSTAT_WRITE
) {
230 return (SET_ERROR(EACCES
));
233 zil_kstat_values_update(zs
, &zil_sums_global
);
239 * Read a log block and make sure it's valid.
242 zil_read_log_block(zilog_t
*zilog
, boolean_t decrypt
, const blkptr_t
*bp
,
243 blkptr_t
*nbp
, char **begin
, char **end
, arc_buf_t
**abuf
)
245 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
246 arc_flags_t aflags
= ARC_FLAG_WAIT
;
250 if (zilog
->zl_header
->zh_claim_txg
== 0)
251 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
253 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
254 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
257 zio_flags
|= ZIO_FLAG_RAW
;
259 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
260 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
262 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
,
263 abuf
, ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
266 zio_cksum_t cksum
= bp
->blk_cksum
;
269 * Validate the checksummed log block.
271 * Sequence numbers should be... sequential. The checksum
272 * verifier for the next block should be bp's checksum plus 1.
274 * Also check the log chain linkage and size used.
276 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
278 uint64_t size
= BP_GET_LSIZE(bp
);
279 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
280 zil_chain_t
*zilc
= (*abuf
)->b_data
;
281 char *lr
= (char *)(zilc
+ 1);
283 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
285 zilc
->zc_nused
< sizeof (*zilc
) ||
286 zilc
->zc_nused
> size
) {
287 error
= SET_ERROR(ECKSUM
);
290 *end
= lr
+ zilc
->zc_nused
- sizeof (*zilc
);
291 *nbp
= zilc
->zc_next_blk
;
294 char *lr
= (*abuf
)->b_data
;
295 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
297 if (memcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
299 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
300 error
= SET_ERROR(ECKSUM
);
303 *end
= lr
+ zilc
->zc_nused
;
304 *nbp
= zilc
->zc_next_blk
;
313 * Read a TX_WRITE log data block.
316 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
318 zio_flag_t zio_flags
= ZIO_FLAG_CANFAIL
;
319 const blkptr_t
*bp
= &lr
->lr_blkptr
;
320 arc_flags_t aflags
= ARC_FLAG_WAIT
;
321 arc_buf_t
*abuf
= NULL
;
325 if (BP_IS_HOLE(bp
)) {
327 memset(wbuf
, 0, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
331 if (zilog
->zl_header
->zh_claim_txg
== 0)
332 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
335 * If we are not using the resulting data, we are just checking that
336 * it hasn't been corrupted so we don't need to waste CPU time
337 * decompressing and decrypting it.
340 zio_flags
|= ZIO_FLAG_RAW
;
342 ASSERT3U(BP_GET_LSIZE(bp
), !=, 0);
343 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
344 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
346 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
347 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
351 memcpy(wbuf
, abuf
->b_data
, arc_buf_size(abuf
));
352 arc_buf_destroy(abuf
, &abuf
);
359 zil_sums_init(zil_sums_t
*zs
)
361 wmsum_init(&zs
->zil_commit_count
, 0);
362 wmsum_init(&zs
->zil_commit_writer_count
, 0);
363 wmsum_init(&zs
->zil_itx_count
, 0);
364 wmsum_init(&zs
->zil_itx_indirect_count
, 0);
365 wmsum_init(&zs
->zil_itx_indirect_bytes
, 0);
366 wmsum_init(&zs
->zil_itx_copied_count
, 0);
367 wmsum_init(&zs
->zil_itx_copied_bytes
, 0);
368 wmsum_init(&zs
->zil_itx_needcopy_count
, 0);
369 wmsum_init(&zs
->zil_itx_needcopy_bytes
, 0);
370 wmsum_init(&zs
->zil_itx_metaslab_normal_count
, 0);
371 wmsum_init(&zs
->zil_itx_metaslab_normal_bytes
, 0);
372 wmsum_init(&zs
->zil_itx_metaslab_normal_write
, 0);
373 wmsum_init(&zs
->zil_itx_metaslab_normal_alloc
, 0);
374 wmsum_init(&zs
->zil_itx_metaslab_slog_count
, 0);
375 wmsum_init(&zs
->zil_itx_metaslab_slog_bytes
, 0);
376 wmsum_init(&zs
->zil_itx_metaslab_slog_write
, 0);
377 wmsum_init(&zs
->zil_itx_metaslab_slog_alloc
, 0);
381 zil_sums_fini(zil_sums_t
*zs
)
383 wmsum_fini(&zs
->zil_commit_count
);
384 wmsum_fini(&zs
->zil_commit_writer_count
);
385 wmsum_fini(&zs
->zil_itx_count
);
386 wmsum_fini(&zs
->zil_itx_indirect_count
);
387 wmsum_fini(&zs
->zil_itx_indirect_bytes
);
388 wmsum_fini(&zs
->zil_itx_copied_count
);
389 wmsum_fini(&zs
->zil_itx_copied_bytes
);
390 wmsum_fini(&zs
->zil_itx_needcopy_count
);
391 wmsum_fini(&zs
->zil_itx_needcopy_bytes
);
392 wmsum_fini(&zs
->zil_itx_metaslab_normal_count
);
393 wmsum_fini(&zs
->zil_itx_metaslab_normal_bytes
);
394 wmsum_fini(&zs
->zil_itx_metaslab_normal_write
);
395 wmsum_fini(&zs
->zil_itx_metaslab_normal_alloc
);
396 wmsum_fini(&zs
->zil_itx_metaslab_slog_count
);
397 wmsum_fini(&zs
->zil_itx_metaslab_slog_bytes
);
398 wmsum_fini(&zs
->zil_itx_metaslab_slog_write
);
399 wmsum_fini(&zs
->zil_itx_metaslab_slog_alloc
);
403 zil_kstat_values_update(zil_kstat_values_t
*zs
, zil_sums_t
*zil_sums
)
405 zs
->zil_commit_count
.value
.ui64
=
406 wmsum_value(&zil_sums
->zil_commit_count
);
407 zs
->zil_commit_writer_count
.value
.ui64
=
408 wmsum_value(&zil_sums
->zil_commit_writer_count
);
409 zs
->zil_itx_count
.value
.ui64
=
410 wmsum_value(&zil_sums
->zil_itx_count
);
411 zs
->zil_itx_indirect_count
.value
.ui64
=
412 wmsum_value(&zil_sums
->zil_itx_indirect_count
);
413 zs
->zil_itx_indirect_bytes
.value
.ui64
=
414 wmsum_value(&zil_sums
->zil_itx_indirect_bytes
);
415 zs
->zil_itx_copied_count
.value
.ui64
=
416 wmsum_value(&zil_sums
->zil_itx_copied_count
);
417 zs
->zil_itx_copied_bytes
.value
.ui64
=
418 wmsum_value(&zil_sums
->zil_itx_copied_bytes
);
419 zs
->zil_itx_needcopy_count
.value
.ui64
=
420 wmsum_value(&zil_sums
->zil_itx_needcopy_count
);
421 zs
->zil_itx_needcopy_bytes
.value
.ui64
=
422 wmsum_value(&zil_sums
->zil_itx_needcopy_bytes
);
423 zs
->zil_itx_metaslab_normal_count
.value
.ui64
=
424 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_count
);
425 zs
->zil_itx_metaslab_normal_bytes
.value
.ui64
=
426 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_bytes
);
427 zs
->zil_itx_metaslab_normal_write
.value
.ui64
=
428 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_write
);
429 zs
->zil_itx_metaslab_normal_alloc
.value
.ui64
=
430 wmsum_value(&zil_sums
->zil_itx_metaslab_normal_alloc
);
431 zs
->zil_itx_metaslab_slog_count
.value
.ui64
=
432 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_count
);
433 zs
->zil_itx_metaslab_slog_bytes
.value
.ui64
=
434 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_bytes
);
435 zs
->zil_itx_metaslab_slog_write
.value
.ui64
=
436 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_write
);
437 zs
->zil_itx_metaslab_slog_alloc
.value
.ui64
=
438 wmsum_value(&zil_sums
->zil_itx_metaslab_slog_alloc
);
442 * Parse the intent log, and call parse_func for each valid record within.
445 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
446 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
,
449 const zil_header_t
*zh
= zilog
->zl_header
;
450 boolean_t claimed
= !!zh
->zh_claim_txg
;
451 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
452 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
453 uint64_t max_blk_seq
= 0;
454 uint64_t max_lr_seq
= 0;
455 uint64_t blk_count
= 0;
456 uint64_t lr_count
= 0;
457 blkptr_t blk
, next_blk
= {{{{0}}}};
461 * Old logs didn't record the maximum zh_claim_lr_seq.
463 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
464 claim_lr_seq
= UINT64_MAX
;
467 * Starting at the block pointed to by zh_log we read the log chain.
468 * For each block in the chain we strongly check that block to
469 * ensure its validity. We stop when an invalid block is found.
470 * For each block pointer in the chain we call parse_blk_func().
471 * For each record in each valid block we call parse_lr_func().
472 * If the log has been claimed, stop if we encounter a sequence
473 * number greater than the highest claimed sequence number.
475 zil_bp_tree_init(zilog
);
477 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
478 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
481 arc_buf_t
*abuf
= NULL
;
483 if (blk_seq
> claim_blk_seq
)
486 error
= parse_blk_func(zilog
, &blk
, arg
, txg
);
489 ASSERT3U(max_blk_seq
, <, blk_seq
);
490 max_blk_seq
= blk_seq
;
493 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
496 error
= zil_read_log_block(zilog
, decrypt
, &blk
, &next_blk
,
500 arc_buf_destroy(abuf
, &abuf
);
502 char name
[ZFS_MAX_DATASET_NAME_LEN
];
504 dmu_objset_name(zilog
->zl_os
, name
);
506 cmn_err(CE_WARN
, "ZFS read log block error %d, "
507 "dataset %s, seq 0x%llx\n", error
, name
,
508 (u_longlong_t
)blk_seq
);
513 for (; lrp
< end
; lrp
+= reclen
) {
514 lr_t
*lr
= (lr_t
*)lrp
;
515 reclen
= lr
->lrc_reclen
;
516 ASSERT3U(reclen
, >=, sizeof (lr_t
));
517 ASSERT3U(reclen
, <=, end
- lrp
);
518 if (lr
->lrc_seq
> claim_lr_seq
) {
519 arc_buf_destroy(abuf
, &abuf
);
523 error
= parse_lr_func(zilog
, lr
, arg
, txg
);
525 arc_buf_destroy(abuf
, &abuf
);
528 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
529 max_lr_seq
= lr
->lrc_seq
;
532 arc_buf_destroy(abuf
, &abuf
);
535 zilog
->zl_parse_error
= error
;
536 zilog
->zl_parse_blk_seq
= max_blk_seq
;
537 zilog
->zl_parse_lr_seq
= max_lr_seq
;
538 zilog
->zl_parse_blk_count
= blk_count
;
539 zilog
->zl_parse_lr_count
= lr_count
;
541 zil_bp_tree_fini(zilog
);
547 zil_clear_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
551 ASSERT(!BP_IS_HOLE(bp
));
554 * As we call this function from the context of a rewind to a
555 * checkpoint, each ZIL block whose txg is later than the txg
556 * that we rewind to is invalid. Thus, we return -1 so
557 * zil_parse() doesn't attempt to read it.
559 if (bp
->blk_birth
>= first_txg
)
562 if (zil_bp_tree_add(zilog
, bp
) != 0)
565 zio_free(zilog
->zl_spa
, first_txg
, bp
);
570 zil_noop_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
573 (void) zilog
, (void) lrc
, (void) tx
, (void) first_txg
;
578 zil_claim_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
582 * Claim log block if not already committed and not already claimed.
583 * If tx == NULL, just verify that the block is claimable.
585 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
586 zil_bp_tree_add(zilog
, bp
) != 0)
589 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
590 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
591 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
595 zil_claim_write(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
, uint64_t first_txg
)
597 lr_write_t
*lr
= (lr_write_t
*)lrc
;
600 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
603 * If the block is not readable, don't claim it. This can happen
604 * in normal operation when a log block is written to disk before
605 * some of the dmu_sync() blocks it points to. In this case, the
606 * transaction cannot have been committed to anyone (we would have
607 * waited for all writes to be stable first), so it is semantically
608 * correct to declare this the end of the log.
610 if (lr
->lr_blkptr
.blk_birth
>= first_txg
) {
611 error
= zil_read_log_data(zilog
, lr
, NULL
);
616 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
620 zil_claim_clone_range(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
623 const lr_clone_range_t
*lr
= (const lr_clone_range_t
*)lrc
;
625 spa_t
*spa
= zilog
->zl_spa
;
628 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
629 ASSERT3U(lrc
->lrc_reclen
, >=, offsetof(lr_clone_range_t
,
630 lr_bps
[lr
->lr_nbps
]));
637 * XXX: Do we need to byteswap lr?
640 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
641 bp
= &lr
->lr_bps
[ii
];
644 * When data is embedded into the BP there is no need to create
645 * BRT entry as there is no data block. Just copy the BP as it
648 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
652 * We can not handle block pointers from the future, since they
653 * are not yet allocated. It should not normally happen, but
654 * just in case lets be safe and just stop here now instead of
655 * corrupting the pool.
657 if (BP_PHYSICAL_BIRTH(bp
) >= first_txg
)
658 return (SET_ERROR(ENOENT
));
661 * Assert the block is really allocated before we reference it.
663 metaslab_check_free(spa
, bp
);
666 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
667 bp
= &lr
->lr_bps
[ii
];
668 if (!BP_IS_HOLE(bp
) && !BP_IS_EMBEDDED(bp
))
669 brt_pending_add(spa
, bp
, tx
);
676 zil_claim_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
680 switch (lrc
->lrc_txtype
) {
682 return (zil_claim_write(zilog
, lrc
, tx
, first_txg
));
684 return (zil_claim_clone_range(zilog
, lrc
, tx
, first_txg
));
691 zil_free_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, void *tx
,
696 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
702 zil_free_write(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
704 lr_write_t
*lr
= (lr_write_t
*)lrc
;
705 blkptr_t
*bp
= &lr
->lr_blkptr
;
707 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
710 * If we previously claimed it, we need to free it.
712 if (bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
714 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
721 zil_free_clone_range(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
)
723 const lr_clone_range_t
*lr
= (const lr_clone_range_t
*)lrc
;
728 ASSERT3U(lrc
->lrc_reclen
, >=, sizeof (*lr
));
729 ASSERT3U(lrc
->lrc_reclen
, >=, offsetof(lr_clone_range_t
,
730 lr_bps
[lr
->lr_nbps
]));
738 for (ii
= 0; ii
< lr
->lr_nbps
; ii
++) {
739 bp
= &lr
->lr_bps
[ii
];
741 if (!BP_IS_HOLE(bp
)) {
742 zio_free(spa
, dmu_tx_get_txg(tx
), bp
);
750 zil_free_log_record(zilog_t
*zilog
, const lr_t
*lrc
, void *tx
,
754 if (claim_txg
== 0) {
758 switch (lrc
->lrc_txtype
) {
760 return (zil_free_write(zilog
, lrc
, tx
, claim_txg
));
762 return (zil_free_clone_range(zilog
, lrc
, tx
));
769 zil_lwb_vdev_compare(const void *x1
, const void *x2
)
771 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
772 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
774 return (TREE_CMP(v1
, v2
));
778 * Allocate a new lwb. We may already have a block pointer for it, in which
779 * case we get size and version from there. Or we may not yet, in which case
780 * we choose them here and later make the block allocation match.
783 zil_alloc_lwb(zilog_t
*zilog
, int sz
, blkptr_t
*bp
, boolean_t slog
,
784 uint64_t txg
, lwb_state_t state
)
788 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
789 lwb
->lwb_zilog
= zilog
;
792 lwb
->lwb_slim
= (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
);
793 sz
= BP_GET_LSIZE(bp
);
795 BP_ZERO(&lwb
->lwb_blk
);
796 lwb
->lwb_slim
= (spa_version(zilog
->zl_spa
) >=
797 SPA_VERSION_SLIM_ZIL
);
799 lwb
->lwb_slog
= slog
;
803 lwb
->lwb_nused
= lwb
->lwb_nfilled
= sizeof (zil_chain_t
);
805 lwb
->lwb_nmax
= sz
- sizeof (zil_chain_t
);
806 lwb
->lwb_nused
= lwb
->lwb_nfilled
= 0;
809 lwb
->lwb_state
= state
;
810 lwb
->lwb_buf
= zio_buf_alloc(sz
);
811 lwb
->lwb_child_zio
= NULL
;
812 lwb
->lwb_write_zio
= NULL
;
813 lwb
->lwb_root_zio
= NULL
;
814 lwb
->lwb_issued_timestamp
= 0;
815 lwb
->lwb_issued_txg
= 0;
816 lwb
->lwb_alloc_txg
= txg
;
817 lwb
->lwb_max_txg
= 0;
819 mutex_enter(&zilog
->zl_lock
);
820 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
821 if (state
!= LWB_STATE_NEW
)
822 zilog
->zl_last_lwb_opened
= lwb
;
823 mutex_exit(&zilog
->zl_lock
);
829 zil_free_lwb(zilog_t
*zilog
, lwb_t
*lwb
)
831 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
832 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
833 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
834 ASSERT3P(lwb
->lwb_child_zio
, ==, NULL
);
835 ASSERT3P(lwb
->lwb_write_zio
, ==, NULL
);
836 ASSERT3P(lwb
->lwb_root_zio
, ==, NULL
);
837 ASSERT3U(lwb
->lwb_alloc_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
838 ASSERT3U(lwb
->lwb_max_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
839 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
840 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
841 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
842 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
845 * Clear the zilog's field to indicate this lwb is no longer
846 * valid, and prevent use-after-free errors.
848 if (zilog
->zl_last_lwb_opened
== lwb
)
849 zilog
->zl_last_lwb_opened
= NULL
;
851 kmem_cache_free(zil_lwb_cache
, lwb
);
855 * Called when we create in-memory log transactions so that we know
856 * to cleanup the itxs at the end of spa_sync().
859 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
861 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
862 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
864 ASSERT(spa_writeable(zilog
->zl_spa
));
866 if (ds
->ds_is_snapshot
)
867 panic("dirtying snapshot!");
869 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
870 /* up the hold count until we can be written out */
871 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
873 zilog
->zl_dirty_max_txg
= MAX(txg
, zilog
->zl_dirty_max_txg
);
878 * Determine if the zil is dirty in the specified txg. Callers wanting to
879 * ensure that the dirty state does not change must hold the itxg_lock for
880 * the specified txg. Holding the lock will ensure that the zil cannot be
881 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
884 static boolean_t __maybe_unused
885 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
887 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
889 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
895 * Determine if the zil is dirty. The zil is considered dirty if it has
896 * any pending itx records that have not been cleaned by zil_clean().
899 zilog_is_dirty(zilog_t
*zilog
)
901 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
903 for (int t
= 0; t
< TXG_SIZE
; t
++) {
904 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
911 * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
912 * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
913 * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
917 zil_commit_activate_saxattr_feature(zilog_t
*zilog
)
919 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
923 if (spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
924 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
&&
925 !dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
)) {
926 tx
= dmu_tx_create(zilog
->zl_os
);
927 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
928 dsl_dataset_dirty(ds
, tx
);
929 txg
= dmu_tx_get_txg(tx
);
931 mutex_enter(&ds
->ds_lock
);
932 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
934 mutex_exit(&ds
->ds_lock
);
936 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
941 * Create an on-disk intent log.
944 zil_create(zilog_t
*zilog
)
946 const zil_header_t
*zh
= zilog
->zl_header
;
952 boolean_t slog
= FALSE
;
953 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
957 * Wait for any previous destroy to complete.
959 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
961 ASSERT(zh
->zh_claim_txg
== 0);
962 ASSERT(zh
->zh_replay_seq
== 0);
967 * Allocate an initial log block if:
968 * - there isn't one already
969 * - the existing block is the wrong endianness
971 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
972 tx
= dmu_tx_create(zilog
->zl_os
);
973 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
974 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
975 txg
= dmu_tx_get_txg(tx
);
977 if (!BP_IS_HOLE(&blk
)) {
978 zio_free(zilog
->zl_spa
, txg
, &blk
);
982 error
= zio_alloc_zil(zilog
->zl_spa
, zilog
->zl_os
, txg
, &blk
,
983 ZIL_MIN_BLKSZ
, &slog
);
985 zil_init_log_chain(zilog
, &blk
);
989 * Allocate a log write block (lwb) for the first log block.
992 lwb
= zil_alloc_lwb(zilog
, 0, &blk
, slog
, txg
, LWB_STATE_NEW
);
995 * If we just allocated the first log block, commit our transaction
996 * and wait for zil_sync() to stuff the block pointer into zh_log.
997 * (zh is part of the MOS, so we cannot modify it in open context.)
1001 * If "zilsaxattr" feature is enabled on zpool, then activate
1002 * it now when we're creating the ZIL chain. We can't wait with
1003 * this until we write the first xattr log record because we
1004 * need to wait for the feature activation to sync out.
1006 if (spa_feature_is_enabled(zilog
->zl_spa
,
1007 SPA_FEATURE_ZILSAXATTR
) && dmu_objset_type(zilog
->zl_os
) !=
1009 mutex_enter(&ds
->ds_lock
);
1010 ds
->ds_feature_activation
[SPA_FEATURE_ZILSAXATTR
] =
1012 mutex_exit(&ds
->ds_lock
);
1016 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1019 * This branch covers the case where we enable the feature on a
1020 * zpool that has existing ZIL headers.
1022 zil_commit_activate_saxattr_feature(zilog
);
1024 IMPLY(spa_feature_is_enabled(zilog
->zl_spa
, SPA_FEATURE_ZILSAXATTR
) &&
1025 dmu_objset_type(zilog
->zl_os
) != DMU_OST_ZVOL
,
1026 dsl_dataset_feature_is_active(ds
, SPA_FEATURE_ZILSAXATTR
));
1028 ASSERT(error
!= 0 || memcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
1029 IMPLY(error
== 0, lwb
!= NULL
);
1035 * In one tx, free all log blocks and clear the log header. If keep_first
1036 * is set, then we're replaying a log with no content. We want to keep the
1037 * first block, however, so that the first synchronous transaction doesn't
1038 * require a txg_wait_synced() in zil_create(). We don't need to
1039 * txg_wait_synced() here either when keep_first is set, because both
1040 * zil_create() and zil_destroy() will wait for any in-progress destroys
1042 * Return B_TRUE if there were any entries to replay.
1045 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
1047 const zil_header_t
*zh
= zilog
->zl_header
;
1053 * Wait for any previous destroy to complete.
1055 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
1057 zilog
->zl_old_header
= *zh
; /* debugging aid */
1059 if (BP_IS_HOLE(&zh
->zh_log
))
1062 tx
= dmu_tx_create(zilog
->zl_os
);
1063 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1064 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1065 txg
= dmu_tx_get_txg(tx
);
1067 mutex_enter(&zilog
->zl_lock
);
1069 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
1070 zilog
->zl_destroy_txg
= txg
;
1071 zilog
->zl_keep_first
= keep_first
;
1073 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
1074 ASSERT(zh
->zh_claim_txg
== 0);
1075 VERIFY(!keep_first
);
1076 while ((lwb
= list_remove_head(&zilog
->zl_lwb_list
)) != NULL
) {
1077 if (lwb
->lwb_buf
!= NULL
)
1078 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1079 if (!BP_IS_HOLE(&lwb
->lwb_blk
))
1080 zio_free(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
1081 zil_free_lwb(zilog
, lwb
);
1083 } else if (!keep_first
) {
1084 zil_destroy_sync(zilog
, tx
);
1086 mutex_exit(&zilog
->zl_lock
);
1094 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1096 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1097 (void) zil_parse(zilog
, zil_free_log_block
,
1098 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
, B_FALSE
);
1102 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
1104 dmu_tx_t
*tx
= txarg
;
1111 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
1112 DMU_OST_ANY
, B_FALSE
, B_FALSE
, FTAG
, &os
);
1115 * EBUSY indicates that the objset is inconsistent, in which
1116 * case it can not have a ZIL.
1118 if (error
!= EBUSY
) {
1119 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
1120 (unsigned long long)ds
->ds_object
, error
);
1126 zilog
= dmu_objset_zil(os
);
1127 zh
= zil_header_in_syncing_context(zilog
);
1128 ASSERT3U(tx
->tx_txg
, ==, spa_first_txg(zilog
->zl_spa
));
1129 first_txg
= spa_min_claim_txg(zilog
->zl_spa
);
1132 * If the spa_log_state is not set to be cleared, check whether
1133 * the current uberblock is a checkpoint one and if the current
1134 * header has been claimed before moving on.
1136 * If the current uberblock is a checkpointed uberblock then
1137 * one of the following scenarios took place:
1139 * 1] We are currently rewinding to the checkpoint of the pool.
1140 * 2] We crashed in the middle of a checkpoint rewind but we
1141 * did manage to write the checkpointed uberblock to the
1142 * vdev labels, so when we tried to import the pool again
1143 * the checkpointed uberblock was selected from the import
1146 * In both cases we want to zero out all the ZIL blocks, except
1147 * the ones that have been claimed at the time of the checkpoint
1148 * (their zh_claim_txg != 0). The reason is that these blocks
1149 * may be corrupted since we may have reused their locations on
1150 * disk after we took the checkpoint.
1152 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1153 * when we first figure out whether the current uberblock is
1154 * checkpointed or not. Unfortunately, that would discard all
1155 * the logs, including the ones that are claimed, and we would
1158 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
||
1159 (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1160 zh
->zh_claim_txg
== 0)) {
1161 if (!BP_IS_HOLE(&zh
->zh_log
)) {
1162 (void) zil_parse(zilog
, zil_clear_log_block
,
1163 zil_noop_log_record
, tx
, first_txg
, B_FALSE
);
1165 BP_ZERO(&zh
->zh_log
);
1166 if (os
->os_encrypted
)
1167 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1168 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1169 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1174 * If we are not rewinding and opening the pool normally, then
1175 * the min_claim_txg should be equal to the first txg of the pool.
1177 ASSERT3U(first_txg
, ==, spa_first_txg(zilog
->zl_spa
));
1180 * Claim all log blocks if we haven't already done so, and remember
1181 * the highest claimed sequence number. This ensures that if we can
1182 * read only part of the log now (e.g. due to a missing device),
1183 * but we can read the entire log later, we will not try to replay
1184 * or destroy beyond the last block we successfully claimed.
1186 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
1187 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
1188 (void) zil_parse(zilog
, zil_claim_log_block
,
1189 zil_claim_log_record
, tx
, first_txg
, B_FALSE
);
1190 zh
->zh_claim_txg
= first_txg
;
1191 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
1192 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
1193 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
1194 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
1195 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
1196 if (os
->os_encrypted
)
1197 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
1198 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
1201 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
1202 dmu_objset_disown(os
, B_FALSE
, FTAG
);
1207 * Check the log by walking the log chain.
1208 * Checksum errors are ok as they indicate the end of the chain.
1209 * Any other error (no device or read failure) returns an error.
1212 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
1222 error
= dmu_objset_from_ds(ds
, &os
);
1224 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
1225 (unsigned long long)ds
->ds_object
, error
);
1229 zilog
= dmu_objset_zil(os
);
1230 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
1232 if (!BP_IS_HOLE(bp
)) {
1234 boolean_t valid
= B_TRUE
;
1237 * Check the first block and determine if it's on a log device
1238 * which may have been removed or faulted prior to loading this
1239 * pool. If so, there's no point in checking the rest of the
1240 * log as its content should have already been synced to the
1243 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
1244 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
1245 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
1246 valid
= vdev_log_state_valid(vd
);
1247 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
1253 * Check whether the current uberblock is checkpointed (e.g.
1254 * we are rewinding) and whether the current header has been
1255 * claimed or not. If it hasn't then skip verifying it. We
1256 * do this because its ZIL blocks may be part of the pool's
1257 * state before the rewind, which is no longer valid.
1259 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1260 if (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
1261 zh
->zh_claim_txg
== 0)
1266 * Because tx == NULL, zil_claim_log_block() will not actually claim
1267 * any blocks, but just determine whether it is possible to do so.
1268 * In addition to checking the log chain, zil_claim_log_block()
1269 * will invoke zio_claim() with a done func of spa_claim_notify(),
1270 * which will update spa_max_claim_txg. See spa_load() for details.
1272 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
1273 zilog
->zl_header
->zh_claim_txg
? -1ULL :
1274 spa_min_claim_txg(os
->os_spa
), B_FALSE
);
1276 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
1280 * When an itx is "skipped", this function is used to properly mark the
1281 * waiter as "done, and signal any thread(s) waiting on it. An itx can
1282 * be skipped (and not committed to an lwb) for a variety of reasons,
1283 * one of them being that the itx was committed via spa_sync(), prior to
1284 * it being committed to an lwb; this can happen if a thread calling
1285 * zil_commit() is racing with spa_sync().
1288 zil_commit_waiter_skip(zil_commit_waiter_t
*zcw
)
1290 mutex_enter(&zcw
->zcw_lock
);
1291 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1292 zcw
->zcw_done
= B_TRUE
;
1293 cv_broadcast(&zcw
->zcw_cv
);
1294 mutex_exit(&zcw
->zcw_lock
);
1298 * This function is used when the given waiter is to be linked into an
1299 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1300 * At this point, the waiter will no longer be referenced by the itx,
1301 * and instead, will be referenced by the lwb.
1304 zil_commit_waiter_link_lwb(zil_commit_waiter_t
*zcw
, lwb_t
*lwb
)
1307 * The lwb_waiters field of the lwb is protected by the zilog's
1308 * zl_issuer_lock while the lwb is open and zl_lock otherwise.
1309 * zl_issuer_lock also protects leaving the open state.
1310 * zcw_lwb setting is protected by zl_issuer_lock and state !=
1311 * flush_done, which transition is protected by zl_lock.
1313 ASSERT(MUTEX_HELD(&lwb
->lwb_zilog
->zl_issuer_lock
));
1314 IMPLY(lwb
->lwb_state
!= LWB_STATE_OPENED
,
1315 MUTEX_HELD(&lwb
->lwb_zilog
->zl_lock
));
1316 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_NEW
);
1317 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1319 ASSERT(!list_link_active(&zcw
->zcw_node
));
1320 list_insert_tail(&lwb
->lwb_waiters
, zcw
);
1321 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1326 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1327 * block, and the given waiter must be linked to the "nolwb waiters"
1328 * list inside of zil_process_commit_list().
1331 zil_commit_waiter_link_nolwb(zil_commit_waiter_t
*zcw
, list_t
*nolwb
)
1333 ASSERT(!list_link_active(&zcw
->zcw_node
));
1334 list_insert_tail(nolwb
, zcw
);
1335 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1339 zil_lwb_add_block(lwb_t
*lwb
, const blkptr_t
*bp
)
1341 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1343 zil_vdev_node_t
*zv
, zvsearch
;
1344 int ndvas
= BP_GET_NDVAS(bp
);
1347 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1348 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1350 if (zil_nocacheflush
)
1353 mutex_enter(&lwb
->lwb_vdev_lock
);
1354 for (i
= 0; i
< ndvas
; i
++) {
1355 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1356 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
1357 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
1358 zv
->zv_vdev
= zvsearch
.zv_vdev
;
1359 avl_insert(t
, zv
, where
);
1362 mutex_exit(&lwb
->lwb_vdev_lock
);
1366 zil_lwb_flush_defer(lwb_t
*lwb
, lwb_t
*nlwb
)
1368 avl_tree_t
*src
= &lwb
->lwb_vdev_tree
;
1369 avl_tree_t
*dst
= &nlwb
->lwb_vdev_tree
;
1370 void *cookie
= NULL
;
1371 zil_vdev_node_t
*zv
;
1373 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1374 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1375 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1378 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1379 * not need the protection of lwb_vdev_lock (it will only be modified
1380 * while holding zilog->zl_lock) as its writes and those of its
1381 * children have all completed. The younger 'nlwb' may be waiting on
1382 * future writes to additional vdevs.
1384 mutex_enter(&nlwb
->lwb_vdev_lock
);
1386 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1387 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1389 while ((zv
= avl_destroy_nodes(src
, &cookie
)) != NULL
) {
1392 if (avl_find(dst
, zv
, &where
) == NULL
) {
1393 avl_insert(dst
, zv
, where
);
1395 kmem_free(zv
, sizeof (*zv
));
1398 mutex_exit(&nlwb
->lwb_vdev_lock
);
1402 zil_lwb_add_txg(lwb_t
*lwb
, uint64_t txg
)
1404 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1408 * This function is a called after all vdevs associated with a given lwb
1409 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1410 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1411 * all "previous" lwb's will have completed before this function is
1412 * called; i.e. this function is called for all previous lwbs before
1413 * it's called for "this" lwb (enforced via zio the dependencies
1414 * configured in zil_lwb_set_zio_dependency()).
1416 * The intention is for this function to be called as soon as the
1417 * contents of an lwb are considered "stable" on disk, and will survive
1418 * any sudden loss of power. At this point, any threads waiting for the
1419 * lwb to reach this state are signalled, and the "waiter" structures
1420 * are marked "done".
1423 zil_lwb_flush_vdevs_done(zio_t
*zio
)
1425 lwb_t
*lwb
= zio
->io_private
;
1426 zilog_t
*zilog
= lwb
->lwb_zilog
;
1427 zil_commit_waiter_t
*zcw
;
1430 spa_config_exit(zilog
->zl_spa
, SCL_STATE
, lwb
);
1432 hrtime_t t
= gethrtime() - lwb
->lwb_issued_timestamp
;
1434 mutex_enter(&zilog
->zl_lock
);
1436 zilog
->zl_last_lwb_latency
= (zilog
->zl_last_lwb_latency
* 7 + t
) / 8;
1438 lwb
->lwb_root_zio
= NULL
;
1440 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1441 lwb
->lwb_state
= LWB_STATE_FLUSH_DONE
;
1443 if (zilog
->zl_last_lwb_opened
== lwb
) {
1445 * Remember the highest committed log sequence number
1446 * for ztest. We only update this value when all the log
1447 * writes succeeded, because ztest wants to ASSERT that
1448 * it got the whole log chain.
1450 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1453 while ((itx
= list_remove_head(&lwb
->lwb_itxs
)) != NULL
)
1454 zil_itx_destroy(itx
);
1456 while ((zcw
= list_remove_head(&lwb
->lwb_waiters
)) != NULL
) {
1457 mutex_enter(&zcw
->zcw_lock
);
1459 ASSERT3P(zcw
->zcw_lwb
, ==, lwb
);
1460 zcw
->zcw_lwb
= NULL
;
1462 * We expect any ZIO errors from child ZIOs to have been
1463 * propagated "up" to this specific LWB's root ZIO, in
1464 * order for this error handling to work correctly. This
1465 * includes ZIO errors from either this LWB's write or
1466 * flush, as well as any errors from other dependent LWBs
1467 * (e.g. a root LWB ZIO that might be a child of this LWB).
1469 * With that said, it's important to note that LWB flush
1470 * errors are not propagated up to the LWB root ZIO.
1471 * This is incorrect behavior, and results in VDEV flush
1472 * errors not being handled correctly here. See the
1473 * comment above the call to "zio_flush" for details.
1476 zcw
->zcw_zio_error
= zio
->io_error
;
1478 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1479 zcw
->zcw_done
= B_TRUE
;
1480 cv_broadcast(&zcw
->zcw_cv
);
1482 mutex_exit(&zcw
->zcw_lock
);
1485 uint64_t txg
= lwb
->lwb_issued_txg
;
1487 /* Once we drop the lock, lwb may be freed by zil_sync(). */
1488 mutex_exit(&zilog
->zl_lock
);
1490 mutex_enter(&zilog
->zl_lwb_io_lock
);
1491 ASSERT3U(zilog
->zl_lwb_inflight
[txg
& TXG_MASK
], >, 0);
1492 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]--;
1493 if (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] == 0)
1494 cv_broadcast(&zilog
->zl_lwb_io_cv
);
1495 mutex_exit(&zilog
->zl_lwb_io_lock
);
1499 * Wait for the completion of all issued write/flush of that txg provided.
1500 * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1503 zil_lwb_flush_wait_all(zilog_t
*zilog
, uint64_t txg
)
1505 ASSERT3U(txg
, ==, spa_syncing_txg(zilog
->zl_spa
));
1507 mutex_enter(&zilog
->zl_lwb_io_lock
);
1508 while (zilog
->zl_lwb_inflight
[txg
& TXG_MASK
] > 0)
1509 cv_wait(&zilog
->zl_lwb_io_cv
, &zilog
->zl_lwb_io_lock
);
1510 mutex_exit(&zilog
->zl_lwb_io_lock
);
1513 mutex_enter(&zilog
->zl_lock
);
1514 mutex_enter(&zilog
->zl_lwb_io_lock
);
1515 lwb_t
*lwb
= list_head(&zilog
->zl_lwb_list
);
1516 while (lwb
!= NULL
) {
1517 if (lwb
->lwb_issued_txg
<= txg
) {
1518 ASSERT(lwb
->lwb_state
!= LWB_STATE_ISSUED
);
1519 ASSERT(lwb
->lwb_state
!= LWB_STATE_WRITE_DONE
);
1520 IMPLY(lwb
->lwb_issued_txg
> 0,
1521 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
1523 IMPLY(lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
1524 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
,
1525 lwb
->lwb_buf
== NULL
);
1526 lwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1528 mutex_exit(&zilog
->zl_lwb_io_lock
);
1529 mutex_exit(&zilog
->zl_lock
);
1534 * This is called when an lwb's write zio completes. The callback's
1535 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1536 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1537 * in writing out this specific lwb's data, and in the case that cache
1538 * flushes have been deferred, vdevs involved in writing the data for
1539 * previous lwbs. The writes corresponding to all the vdevs in the
1540 * lwb_vdev_tree will have completed by the time this is called, due to
1541 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1542 * which takes deferred flushes into account. The lwb will be "done"
1543 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1544 * completion callback for the lwb's root zio.
1547 zil_lwb_write_done(zio_t
*zio
)
1549 lwb_t
*lwb
= zio
->io_private
;
1550 spa_t
*spa
= zio
->io_spa
;
1551 zilog_t
*zilog
= lwb
->lwb_zilog
;
1552 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1553 void *cookie
= NULL
;
1554 zil_vdev_node_t
*zv
;
1557 ASSERT3S(spa_config_held(spa
, SCL_STATE
, RW_READER
), !=, 0);
1559 abd_free(zio
->io_abd
);
1560 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1561 lwb
->lwb_buf
= NULL
;
1563 mutex_enter(&zilog
->zl_lock
);
1564 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_ISSUED
);
1565 lwb
->lwb_state
= LWB_STATE_WRITE_DONE
;
1566 lwb
->lwb_child_zio
= NULL
;
1567 lwb
->lwb_write_zio
= NULL
;
1570 * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not
1571 * called for it yet, and when it will be, it won't be able to make
1572 * its write ZIO a parent this ZIO. In such case we can not defer
1573 * our flushes or below may be a race between the done callbacks.
1575 nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1576 if (nlwb
&& nlwb
->lwb_state
!= LWB_STATE_ISSUED
)
1578 mutex_exit(&zilog
->zl_lock
);
1580 if (avl_numnodes(t
) == 0)
1584 * If there was an IO error, we're not going to call zio_flush()
1585 * on these vdevs, so we simply empty the tree and free the
1586 * nodes. We avoid calling zio_flush() since there isn't any
1587 * good reason for doing so, after the lwb block failed to be
1590 * Additionally, we don't perform any further error handling at
1591 * this point (e.g. setting "zcw_zio_error" appropriately), as
1592 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1593 * we expect any error seen here, to have been propagated to
1596 if (zio
->io_error
!= 0) {
1597 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
1598 kmem_free(zv
, sizeof (*zv
));
1603 * If this lwb does not have any threads waiting for it to
1604 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1605 * command to the vdevs written to by "this" lwb, and instead
1606 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1607 * command for those vdevs. Thus, we merge the vdev tree of
1608 * "this" lwb with the vdev tree of the "next" lwb in the list,
1609 * and assume the "next" lwb will handle flushing the vdevs (or
1610 * deferring the flush(s) again).
1612 * This is a useful performance optimization, especially for
1613 * workloads with lots of async write activity and few sync
1614 * write and/or fsync activity, as it has the potential to
1615 * coalesce multiple flush commands to a vdev into one.
1617 if (list_is_empty(&lwb
->lwb_waiters
) && nlwb
!= NULL
) {
1618 zil_lwb_flush_defer(lwb
, nlwb
);
1619 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
1623 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1624 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
1627 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1628 * always used within "zio_flush". This means,
1629 * any errors when flushing the vdev(s), will
1630 * (unfortunately) not be handled correctly,
1631 * since these "zio_flush" errors will not be
1632 * propagated up to "zil_lwb_flush_vdevs_done".
1634 zio_flush(lwb
->lwb_root_zio
, vd
);
1636 kmem_free(zv
, sizeof (*zv
));
1641 * Build the zio dependency chain, which is used to preserve the ordering of
1642 * lwb completions that is required by the semantics of the ZIL. Each new lwb
1643 * zio becomes a parent of the previous lwb zio, such that the new lwb's zio
1644 * cannot complete until the previous lwb's zio completes.
1646 * This is required by the semantics of zil_commit(): the commit waiters
1647 * attached to the lwbs will be woken in the lwb zio's completion callback,
1648 * so this zio dependency graph ensures the waiters are woken in the correct
1649 * order (the same order the lwbs were created).
1652 zil_lwb_set_zio_dependency(zilog_t
*zilog
, lwb_t
*lwb
)
1654 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
1656 lwb_t
*prev_lwb
= list_prev(&zilog
->zl_lwb_list
, lwb
);
1657 if (prev_lwb
== NULL
||
1658 prev_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
1662 * If the previous lwb's write hasn't already completed, we also want
1663 * to order the completion of the lwb write zios (above, we only order
1664 * the completion of the lwb root zios). This is required because of
1665 * how we can defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1667 * When the DKIOCFLUSHWRITECACHE commands are deferred, the previous
1668 * lwb will rely on this lwb to flush the vdevs written to by that
1669 * previous lwb. Thus, we need to ensure this lwb doesn't issue the
1670 * flush until after the previous lwb's write completes. We ensure
1671 * this ordering by setting the zio parent/child relationship here.
1673 * Without this relationship on the lwb's write zio, it's possible
1674 * for this lwb's write to complete prior to the previous lwb's write
1675 * completing; and thus, the vdevs for the previous lwb would be
1676 * flushed prior to that lwb's data being written to those vdevs (the
1677 * vdevs are flushed in the lwb write zio's completion handler,
1678 * zil_lwb_write_done()).
1680 if (prev_lwb
->lwb_state
== LWB_STATE_ISSUED
) {
1681 ASSERT3P(prev_lwb
->lwb_write_zio
, !=, NULL
);
1682 zio_add_child(lwb
->lwb_write_zio
, prev_lwb
->lwb_write_zio
);
1684 ASSERT3S(prev_lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1687 ASSERT3P(prev_lwb
->lwb_root_zio
, !=, NULL
);
1688 zio_add_child(lwb
->lwb_root_zio
, prev_lwb
->lwb_root_zio
);
1693 * This function's purpose is to "open" an lwb such that it is ready to
1694 * accept new itxs being committed to it. This function is idempotent; if
1695 * the passed in lwb has already been opened, it is essentially a no-op.
1698 zil_lwb_write_open(zilog_t
*zilog
, lwb_t
*lwb
)
1700 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1702 if (lwb
->lwb_state
!= LWB_STATE_NEW
) {
1703 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1707 mutex_enter(&zilog
->zl_lock
);
1708 lwb
->lwb_state
= LWB_STATE_OPENED
;
1709 zilog
->zl_last_lwb_opened
= lwb
;
1710 mutex_exit(&zilog
->zl_lock
);
1714 * Define a limited set of intent log block sizes.
1716 * These must be a multiple of 4KB. Note only the amount used (again
1717 * aligned to 4KB) actually gets written. However, we can't always just
1718 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1720 static const struct {
1723 } zil_block_buckets
[] = {
1724 { 4096, 4096 }, /* non TX_WRITE */
1725 { 8192 + 4096, 8192 + 4096 }, /* database */
1726 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */
1727 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */
1728 { UINT64_MAX
, SPA_OLD_MAXBLOCKSIZE
}, /* > 128KB writes */
1732 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1733 * initialized. Otherwise this should not be used directly; see
1734 * zl_max_block_size instead.
1736 static uint_t zil_maxblocksize
= SPA_OLD_MAXBLOCKSIZE
;
1739 * Close the log block for being issued and allocate the next one.
1740 * Has to be called under zl_issuer_lock to chain more lwbs.
1743 zil_lwb_write_close(zilog_t
*zilog
, lwb_t
*lwb
, lwb_state_t state
)
1747 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1748 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1749 lwb
->lwb_state
= LWB_STATE_CLOSED
;
1752 * If there was an allocation failure then returned NULL will trigger
1753 * zil_commit_writer_stall() at the caller. This is inherently racy,
1754 * since allocation may not have happened yet.
1756 if (lwb
->lwb_error
!= 0)
1760 * Log blocks are pre-allocated. Here we select the size of the next
1761 * block, based on size used in the last block.
1762 * - first find the smallest bucket that will fit the block from a
1763 * limited set of block sizes. This is because it's faster to write
1764 * blocks allocated from the same metaslab as they are adjacent or
1766 * - next find the maximum from the new suggested size and an array of
1767 * previous sizes. This lessens a picket fence effect of wrongly
1768 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1771 * Note we only write what is used, but we can't just allocate
1772 * the maximum block size because we can exhaust the available
1775 uint64_t zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1776 for (i
= 0; zil_blksz
> zil_block_buckets
[i
].limit
; i
++)
1778 zil_blksz
= MIN(zil_block_buckets
[i
].blksz
, zilog
->zl_max_block_size
);
1779 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1780 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1781 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1782 DTRACE_PROBE3(zil__block__size
, zilog_t
*, zilog
,
1783 uint64_t, zil_blksz
,
1784 uint64_t, zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
]);
1785 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1787 return (zil_alloc_lwb(zilog
, zil_blksz
, NULL
, 0, 0, state
));
1791 * Finalize previously closed block and issue the write zio.
1794 zil_lwb_write_issue(zilog_t
*zilog
, lwb_t
*lwb
)
1796 spa_t
*spa
= zilog
->zl_spa
;
1799 zbookmark_phys_t zb
;
1800 zio_priority_t prio
;
1803 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_CLOSED
);
1805 /* Actually fill the lwb with the data. */
1806 for (itx_t
*itx
= list_head(&lwb
->lwb_itxs
); itx
;
1807 itx
= list_next(&lwb
->lwb_itxs
, itx
))
1808 zil_lwb_commit(zilog
, lwb
, itx
);
1809 lwb
->lwb_nused
= lwb
->lwb_nfilled
;
1810 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_nmax
);
1812 lwb
->lwb_root_zio
= zio_root(spa
, zil_lwb_flush_vdevs_done
, lwb
,
1816 * The lwb is now ready to be issued, but it can be only if it already
1817 * got its block pointer allocated or the allocation has failed.
1818 * Otherwise leave it as-is, relying on some other thread to issue it
1819 * after allocating its block pointer via calling zil_lwb_write_issue()
1820 * for the previous lwb(s) in the chain.
1822 mutex_enter(&zilog
->zl_lock
);
1823 lwb
->lwb_state
= LWB_STATE_READY
;
1824 if (BP_IS_HOLE(&lwb
->lwb_blk
) && lwb
->lwb_error
== 0) {
1825 mutex_exit(&zilog
->zl_lock
);
1828 mutex_exit(&zilog
->zl_lock
);
1832 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
1834 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_nmax
);
1835 int wsz
= lwb
->lwb_sz
;
1836 if (lwb
->lwb_error
== 0) {
1837 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
, lwb
->lwb_sz
);
1838 if (!lwb
->lwb_slog
|| zilog
->zl_cur_used
<= zil_slog_bulk
)
1839 prio
= ZIO_PRIORITY_SYNC_WRITE
;
1841 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
1842 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1843 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
1844 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1845 lwb
->lwb_write_zio
= zio_rewrite(lwb
->lwb_root_zio
, spa
, 0,
1846 &lwb
->lwb_blk
, lwb_abd
, lwb
->lwb_sz
, zil_lwb_write_done
,
1847 lwb
, prio
, ZIO_FLAG_CANFAIL
, &zb
);
1848 zil_lwb_add_block(lwb
, &lwb
->lwb_blk
);
1850 if (lwb
->lwb_slim
) {
1851 /* For Slim ZIL only write what is used. */
1852 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
,
1854 ASSERT3S(wsz
, <=, lwb
->lwb_sz
);
1855 zio_shrink(lwb
->lwb_write_zio
, wsz
);
1856 wsz
= lwb
->lwb_write_zio
->io_size
;
1858 memset(lwb
->lwb_buf
+ lwb
->lwb_nused
, 0, wsz
- lwb
->lwb_nused
);
1860 zilc
->zc_nused
= lwb
->lwb_nused
;
1861 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1864 * We can't write the lwb if there was an allocation failure,
1865 * so create a null zio instead just to maintain dependencies.
1867 lwb
->lwb_write_zio
= zio_null(lwb
->lwb_root_zio
, spa
, NULL
,
1868 zil_lwb_write_done
, lwb
, ZIO_FLAG_CANFAIL
);
1869 lwb
->lwb_write_zio
->io_error
= lwb
->lwb_error
;
1871 if (lwb
->lwb_child_zio
)
1872 zio_add_child(lwb
->lwb_write_zio
, lwb
->lwb_child_zio
);
1875 * Open transaction to allocate the next block pointer.
1877 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
1878 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
1879 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1880 uint64_t txg
= dmu_tx_get_txg(tx
);
1883 * Allocate next the block pointer unless we are already in error.
1885 lwb_t
*nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1886 blkptr_t
*bp
= &zilc
->zc_next_blk
;
1888 error
= lwb
->lwb_error
;
1890 error
= zio_alloc_zil(spa
, zilog
->zl_os
, txg
, bp
, nlwb
->lwb_sz
,
1894 ASSERT3U(bp
->blk_birth
, ==, txg
);
1895 BP_SET_CHECKSUM(bp
, nlwb
->lwb_slim
? ZIO_CHECKSUM_ZILOG2
:
1896 ZIO_CHECKSUM_ZILOG
);
1897 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1898 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1902 * Reduce TXG open time by incrementing inflight counter and committing
1903 * the transaciton. zil_sync() will wait for it to return to zero.
1905 mutex_enter(&zilog
->zl_lwb_io_lock
);
1906 lwb
->lwb_issued_txg
= txg
;
1907 zilog
->zl_lwb_inflight
[txg
& TXG_MASK
]++;
1908 zilog
->zl_lwb_max_issued_txg
= MAX(txg
, zilog
->zl_lwb_max_issued_txg
);
1909 mutex_exit(&zilog
->zl_lwb_io_lock
);
1912 spa_config_enter(spa
, SCL_STATE
, lwb
, RW_READER
);
1915 * We've completed all potentially blocking operations. Update the
1916 * nlwb and allow it proceed without possible lock order reversals.
1918 mutex_enter(&zilog
->zl_lock
);
1919 zil_lwb_set_zio_dependency(zilog
, lwb
);
1920 lwb
->lwb_state
= LWB_STATE_ISSUED
;
1923 nlwb
->lwb_blk
= *bp
;
1924 nlwb
->lwb_error
= error
;
1925 nlwb
->lwb_slog
= slog
;
1926 nlwb
->lwb_alloc_txg
= txg
;
1927 if (nlwb
->lwb_state
!= LWB_STATE_READY
)
1930 mutex_exit(&zilog
->zl_lock
);
1932 if (lwb
->lwb_slog
) {
1933 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_slog_count
);
1934 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_bytes
,
1936 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_write
,
1938 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_slog_alloc
,
1939 BP_GET_LSIZE(&lwb
->lwb_blk
));
1941 ZIL_STAT_BUMP(zilog
, zil_itx_metaslab_normal_count
);
1942 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_bytes
,
1944 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_write
,
1946 ZIL_STAT_INCR(zilog
, zil_itx_metaslab_normal_alloc
,
1947 BP_GET_LSIZE(&lwb
->lwb_blk
));
1949 lwb
->lwb_issued_timestamp
= gethrtime();
1950 if (lwb
->lwb_child_zio
)
1951 zio_nowait(lwb
->lwb_child_zio
);
1952 zio_nowait(lwb
->lwb_write_zio
);
1953 zio_nowait(lwb
->lwb_root_zio
);
1956 * If nlwb was ready when we gave it the block pointer,
1957 * it is on us to issue it and possibly following ones.
1965 * Maximum amount of data that can be put into single log block.
1968 zil_max_log_data(zilog_t
*zilog
, size_t hdrsize
)
1970 return (zilog
->zl_max_block_size
- sizeof (zil_chain_t
) - hdrsize
);
1974 * Maximum amount of log space we agree to waste to reduce number of
1975 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%).
1977 static inline uint64_t
1978 zil_max_waste_space(zilog_t
*zilog
)
1980 return (zil_max_log_data(zilog
, sizeof (lr_write_t
)) / 16);
1984 * Maximum amount of write data for WR_COPIED. For correctness, consumers
1985 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1986 * maximum sized log block, because each WR_COPIED record must fit in a
1987 * single log block. Below that it is a tradeoff of additional memory copy
1988 * and possibly worse log space efficiency vs additional range lock/unlock.
1990 static uint_t zil_maxcopied
= 7680;
1993 zil_max_copied_data(zilog_t
*zilog
)
1995 uint64_t max_data
= zil_max_log_data(zilog
, sizeof (lr_write_t
));
1996 return (MIN(max_data
, zil_maxcopied
));
2000 * Estimate space needed in the lwb for the itx. Allocate more lwbs or
2001 * split the itx as needed, but don't touch the actual transaction data.
2002 * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
2003 * to chain more lwbs.
2006 zil_lwb_assign(zilog_t
*zilog
, lwb_t
*lwb
, itx_t
*itx
, list_t
*ilwbs
)
2011 uint64_t dlen
, dnow
, lwb_sp
, reclen
, max_log_data
;
2013 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2014 ASSERT3P(lwb
, !=, NULL
);
2015 ASSERT3P(lwb
->lwb_buf
, !=, NULL
);
2017 zil_lwb_write_open(zilog
, lwb
);
2020 lrw
= (lr_write_t
*)lr
;
2023 * A commit itx doesn't represent any on-disk state; instead
2024 * it's simply used as a place holder on the commit list, and
2025 * provides a mechanism for attaching a "commit waiter" onto the
2026 * correct lwb (such that the waiter can be signalled upon
2027 * completion of that lwb). Thus, we don't process this itx's
2028 * log record if it's a commit itx (these itx's don't have log
2029 * records), and instead link the itx's waiter onto the lwb's
2032 * For more details, see the comment above zil_commit().
2034 if (lr
->lrc_txtype
== TX_COMMIT
) {
2035 zil_commit_waiter_link_lwb(itx
->itx_private
, lwb
);
2036 list_insert_tail(&lwb
->lwb_itxs
, itx
);
2040 reclen
= lr
->lrc_reclen
;
2041 if (lr
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
2042 ASSERT3U(reclen
, ==, sizeof (lr_write_t
));
2043 dlen
= P2ROUNDUP_TYPED(
2044 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
2046 ASSERT3U(reclen
, >=, sizeof (lr_t
));
2049 ASSERT3U(reclen
, <=, zil_max_log_data(zilog
, 0));
2050 zilog
->zl_cur_used
+= (reclen
+ dlen
);
2054 * If this record won't fit in the current log block, start a new one.
2055 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2057 lwb_sp
= lwb
->lwb_nmax
- lwb
->lwb_nused
;
2058 max_log_data
= zil_max_log_data(zilog
, sizeof (lr_write_t
));
2059 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
2060 lwb_sp
< zil_max_waste_space(zilog
) &&
2061 (dlen
% max_log_data
== 0 ||
2062 lwb_sp
< reclen
+ dlen
% max_log_data
))) {
2063 list_insert_tail(ilwbs
, lwb
);
2064 lwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_OPENED
);
2067 lwb_sp
= lwb
->lwb_nmax
- lwb
->lwb_nused
;
2071 * There must be enough space in the log block to hold reclen.
2072 * For WR_COPIED, we need to fit the whole record in one block,
2073 * and reclen is the write record header size + the data size.
2074 * For WR_NEED_COPY, we can create multiple records, splitting
2075 * the data into multiple blocks, so we only need to fit one
2076 * word of data per block; in this case reclen is just the header
2079 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
2081 dnow
= MIN(dlen
, lwb_sp
- reclen
);
2083 ASSERT3U(lr
->lrc_txtype
, ==, TX_WRITE
);
2084 ASSERT3U(itx
->itx_wr_state
, ==, WR_NEED_COPY
);
2085 citx
= zil_itx_clone(itx
);
2086 clr
= &citx
->itx_lr
;
2087 lr_write_t
*clrw
= (lr_write_t
*)clr
;
2088 clrw
->lr_length
= dnow
;
2089 lrw
->lr_offset
+= dnow
;
2090 lrw
->lr_length
-= dnow
;
2097 * We're actually making an entry, so update lrc_seq to be the
2098 * log record sequence number. Note that this is generally not
2099 * equal to the itx sequence number because not all transactions
2100 * are synchronous, and sometimes spa_sync() gets there first.
2102 clr
->lrc_seq
= ++zilog
->zl_lr_seq
;
2104 lwb
->lwb_nused
+= reclen
+ dnow
;
2105 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_nmax
);
2106 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
2108 zil_lwb_add_txg(lwb
, lr
->lrc_txg
);
2109 list_insert_tail(&lwb
->lwb_itxs
, citx
);
2113 zilog
->zl_cur_used
+= reclen
;
2117 if (lr
->lrc_txtype
== TX_WRITE
&&
2118 lr
->lrc_txg
> spa_freeze_txg(zilog
->zl_spa
))
2119 txg_wait_synced(zilog
->zl_dmu_pool
, lr
->lrc_txg
);
2125 * Fill the actual transaction data into the lwb, following zil_lwb_assign().
2126 * Does not require locking.
2129 zil_lwb_commit(zilog_t
*zilog
, lwb_t
*lwb
, itx_t
*itx
)
2132 lr_write_t
*lrw
, *lrwb
;
2134 uint64_t dlen
, reclen
;
2137 lrw
= (lr_write_t
*)lr
;
2139 if (lr
->lrc_txtype
== TX_COMMIT
)
2142 if (lr
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
2143 dlen
= P2ROUNDUP_TYPED(
2144 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
2148 reclen
= lr
->lrc_reclen
;
2149 ASSERT3U(reclen
+ dlen
, <=, lwb
->lwb_nused
- lwb
->lwb_nfilled
);
2151 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nfilled
;
2152 memcpy(lr_buf
, lr
, reclen
);
2153 lrb
= (lr_t
*)lr_buf
; /* Like lr, but inside lwb. */
2154 lrwb
= (lr_write_t
*)lrb
; /* Like lrw, but inside lwb. */
2156 ZIL_STAT_BUMP(zilog
, zil_itx_count
);
2159 * If it's a write, fetch the data or get its blkptr as appropriate.
2161 if (lr
->lrc_txtype
== TX_WRITE
) {
2162 if (itx
->itx_wr_state
== WR_COPIED
) {
2163 ZIL_STAT_BUMP(zilog
, zil_itx_copied_count
);
2164 ZIL_STAT_INCR(zilog
, zil_itx_copied_bytes
,
2170 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
2171 dbuf
= lr_buf
+ reclen
;
2172 lrb
->lrc_reclen
+= dlen
;
2173 ZIL_STAT_BUMP(zilog
, zil_itx_needcopy_count
);
2174 ZIL_STAT_INCR(zilog
, zil_itx_needcopy_bytes
,
2177 ASSERT3S(itx
->itx_wr_state
, ==, WR_INDIRECT
);
2179 ZIL_STAT_BUMP(zilog
, zil_itx_indirect_count
);
2180 ZIL_STAT_INCR(zilog
, zil_itx_indirect_bytes
,
2182 if (lwb
->lwb_child_zio
== NULL
) {
2183 lwb
->lwb_child_zio
= zio_null(NULL
,
2184 zilog
->zl_spa
, NULL
, NULL
, NULL
,
2190 * The "lwb_child_zio" we pass in will become a child of
2191 * "lwb_write_zio", when one is created, so one will be
2192 * a parent of any zio's created by the "zl_get_data".
2193 * This way "lwb_write_zio" will first wait for children
2194 * block pointers before own writing, and then for their
2195 * writing completion before the vdev cache flushing.
2197 error
= zilog
->zl_get_data(itx
->itx_private
,
2198 itx
->itx_gen
, lrwb
, dbuf
, lwb
,
2199 lwb
->lwb_child_zio
);
2200 if (dbuf
!= NULL
&& error
== 0) {
2201 /* Zero any padding bytes in the last block. */
2202 memset((char *)dbuf
+ lrwb
->lr_length
, 0,
2203 dlen
- lrwb
->lr_length
);
2207 * Typically, the only return values we should see from
2208 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2209 * EALREADY. However, it is also possible to see other
2210 * error values such as ENOSPC or EINVAL from
2211 * dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2212 * ENXIO as well as a multitude of others from the
2213 * block layer through dmu_buf_hold() -> dbuf_read()
2214 * -> zio_wait(), as well as through dmu_read() ->
2215 * dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2216 * zio_wait(). When these errors happen, we can assume
2217 * that neither an immediate write nor an indirect
2218 * write occurred, so we need to fall back to
2219 * txg_wait_synced(). This is unusual, so we print to
2220 * dmesg whenever one of these errors occurs.
2226 cmn_err(CE_WARN
, "zil_lwb_commit() received "
2227 "unexpected error %d from ->zl_get_data()"
2228 ". Falling back to txg_wait_synced().",
2232 txg_wait_synced(zilog
->zl_dmu_pool
,
2245 lwb
->lwb_nfilled
+= reclen
+ dlen
;
2246 ASSERT3S(lwb
->lwb_nfilled
, <=, lwb
->lwb_nused
);
2247 ASSERT0(P2PHASE(lwb
->lwb_nfilled
, sizeof (uint64_t)));
2251 zil_itx_create(uint64_t txtype
, size_t olrsize
)
2253 size_t itxsize
, lrsize
;
2256 ASSERT3U(olrsize
, >=, sizeof (lr_t
));
2257 lrsize
= P2ROUNDUP_TYPED(olrsize
, sizeof (uint64_t), size_t);
2258 ASSERT3U(lrsize
, >=, olrsize
);
2259 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
2261 itx
= zio_data_buf_alloc(itxsize
);
2262 itx
->itx_lr
.lrc_txtype
= txtype
;
2263 itx
->itx_lr
.lrc_reclen
= lrsize
;
2264 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
2265 memset((char *)&itx
->itx_lr
+ olrsize
, 0, lrsize
- olrsize
);
2266 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
2267 itx
->itx_callback
= NULL
;
2268 itx
->itx_callback_data
= NULL
;
2269 itx
->itx_size
= itxsize
;
2275 zil_itx_clone(itx_t
*oitx
)
2277 ASSERT3U(oitx
->itx_size
, >=, sizeof (itx_t
));
2278 ASSERT3U(oitx
->itx_size
, ==,
2279 offsetof(itx_t
, itx_lr
) + oitx
->itx_lr
.lrc_reclen
);
2281 itx_t
*itx
= zio_data_buf_alloc(oitx
->itx_size
);
2282 memcpy(itx
, oitx
, oitx
->itx_size
);
2283 itx
->itx_callback
= NULL
;
2284 itx
->itx_callback_data
= NULL
;
2289 zil_itx_destroy(itx_t
*itx
)
2291 ASSERT3U(itx
->itx_size
, >=, sizeof (itx_t
));
2292 ASSERT3U(itx
->itx_lr
.lrc_reclen
, ==,
2293 itx
->itx_size
- offsetof(itx_t
, itx_lr
));
2294 IMPLY(itx
->itx_lr
.lrc_txtype
== TX_COMMIT
, itx
->itx_callback
== NULL
);
2295 IMPLY(itx
->itx_callback
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2297 if (itx
->itx_callback
!= NULL
)
2298 itx
->itx_callback(itx
->itx_callback_data
);
2300 zio_data_buf_free(itx
, itx
->itx_size
);
2304 * Free up the sync and async itxs. The itxs_t has already been detached
2305 * so no locks are needed.
2308 zil_itxg_clean(void *arg
)
2315 itx_async_node_t
*ian
;
2317 list
= &itxs
->i_sync_list
;
2318 while ((itx
= list_remove_head(list
)) != NULL
) {
2320 * In the general case, commit itxs will not be found
2321 * here, as they'll be committed to an lwb via
2322 * zil_lwb_assign(), and free'd in that function. Having
2323 * said that, it is still possible for commit itxs to be
2324 * found here, due to the following race:
2326 * - a thread calls zil_commit() which assigns the
2327 * commit itx to a per-txg i_sync_list
2328 * - zil_itxg_clean() is called (e.g. via spa_sync())
2329 * while the waiter is still on the i_sync_list
2331 * There's nothing to prevent syncing the txg while the
2332 * waiter is on the i_sync_list. This normally doesn't
2333 * happen because spa_sync() is slower than zil_commit(),
2334 * but if zil_commit() calls txg_wait_synced() (e.g.
2335 * because zil_create() or zil_commit_writer_stall() is
2336 * called) we will hit this case.
2338 if (itx
->itx_lr
.lrc_txtype
== TX_COMMIT
)
2339 zil_commit_waiter_skip(itx
->itx_private
);
2341 zil_itx_destroy(itx
);
2345 t
= &itxs
->i_async_tree
;
2346 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2347 list
= &ian
->ia_list
;
2348 while ((itx
= list_remove_head(list
)) != NULL
) {
2349 /* commit itxs should never be on the async lists. */
2350 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2351 zil_itx_destroy(itx
);
2354 kmem_free(ian
, sizeof (itx_async_node_t
));
2358 kmem_free(itxs
, sizeof (itxs_t
));
2362 zil_aitx_compare(const void *x1
, const void *x2
)
2364 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
2365 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
2367 return (TREE_CMP(o1
, o2
));
2371 * Remove all async itx with the given oid.
2374 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
2377 itx_async_node_t
*ian
, ian_search
;
2384 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2386 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2389 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2391 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2392 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2394 mutex_enter(&itxg
->itxg_lock
);
2395 if (itxg
->itxg_txg
!= txg
) {
2396 mutex_exit(&itxg
->itxg_lock
);
2401 * Locate the object node and append its list.
2403 t
= &itxg
->itxg_itxs
->i_async_tree
;
2404 ian_search
.ia_foid
= oid
;
2405 ian
= avl_find(t
, &ian_search
, &where
);
2407 list_move_tail(&clean_list
, &ian
->ia_list
);
2408 mutex_exit(&itxg
->itxg_lock
);
2410 while ((itx
= list_remove_head(&clean_list
)) != NULL
) {
2411 /* commit itxs should never be on the async lists. */
2412 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
2413 zil_itx_destroy(itx
);
2415 list_destroy(&clean_list
);
2419 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
2423 itxs_t
*itxs
, *clean
= NULL
;
2426 * Ensure the data of a renamed file is committed before the rename.
2428 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
2429 zil_async_to_sync(zilog
, itx
->itx_oid
);
2431 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
2434 txg
= dmu_tx_get_txg(tx
);
2436 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2437 mutex_enter(&itxg
->itxg_lock
);
2438 itxs
= itxg
->itxg_itxs
;
2439 if (itxg
->itxg_txg
!= txg
) {
2442 * The zil_clean callback hasn't got around to cleaning
2443 * this itxg. Save the itxs for release below.
2444 * This should be rare.
2446 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2447 "txg %llu", (u_longlong_t
)itxg
->itxg_txg
);
2448 clean
= itxg
->itxg_itxs
;
2450 itxg
->itxg_txg
= txg
;
2451 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
2454 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
2455 offsetof(itx_t
, itx_node
));
2456 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
2457 sizeof (itx_async_node_t
),
2458 offsetof(itx_async_node_t
, ia_node
));
2460 if (itx
->itx_sync
) {
2461 list_insert_tail(&itxs
->i_sync_list
, itx
);
2463 avl_tree_t
*t
= &itxs
->i_async_tree
;
2465 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
2466 itx_async_node_t
*ian
;
2469 ian
= avl_find(t
, &foid
, &where
);
2471 ian
= kmem_alloc(sizeof (itx_async_node_t
),
2473 list_create(&ian
->ia_list
, sizeof (itx_t
),
2474 offsetof(itx_t
, itx_node
));
2475 ian
->ia_foid
= foid
;
2476 avl_insert(t
, ian
, where
);
2478 list_insert_tail(&ian
->ia_list
, itx
);
2481 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
2484 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2485 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2486 * need to be careful to always dirty the ZIL using the "real"
2487 * TXG (not itxg_txg) even when the SPA is frozen.
2489 zilog_dirty(zilog
, dmu_tx_get_txg(tx
));
2490 mutex_exit(&itxg
->itxg_lock
);
2492 /* Release the old itxs now we've dropped the lock */
2494 zil_itxg_clean(clean
);
2498 * If there are any in-memory intent log transactions which have now been
2499 * synced then start up a taskq to free them. We should only do this after we
2500 * have written out the uberblocks (i.e. txg has been committed) so that
2501 * don't inadvertently clean out in-memory log records that would be required
2505 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
2507 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
2510 ASSERT3U(synced_txg
, <, ZILTEST_TXG
);
2512 mutex_enter(&itxg
->itxg_lock
);
2513 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
2514 mutex_exit(&itxg
->itxg_lock
);
2517 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
2518 ASSERT3U(itxg
->itxg_txg
, !=, 0);
2519 clean_me
= itxg
->itxg_itxs
;
2520 itxg
->itxg_itxs
= NULL
;
2522 mutex_exit(&itxg
->itxg_lock
);
2524 * Preferably start a task queue to free up the old itxs but
2525 * if taskq_dispatch can't allocate resources to do that then
2526 * free it in-line. This should be rare. Note, using TQ_SLEEP
2527 * created a bad performance problem.
2529 ASSERT3P(zilog
->zl_dmu_pool
, !=, NULL
);
2530 ASSERT3P(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
, !=, NULL
);
2531 taskqid_t id
= taskq_dispatch(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
,
2532 zil_itxg_clean
, clean_me
, TQ_NOSLEEP
);
2533 if (id
== TASKQID_INVALID
)
2534 zil_itxg_clean(clean_me
);
2538 * This function will traverse the queue of itxs that need to be
2539 * committed, and move them onto the ZIL's zl_itx_commit_list.
2542 zil_get_commit_list(zilog_t
*zilog
)
2544 uint64_t otxg
, txg
, wtxg
= 0;
2545 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
2547 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2549 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2552 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2555 * This is inherently racy, since there is nothing to prevent
2556 * the last synced txg from changing. That's okay since we'll
2557 * only commit things in the future.
2559 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2560 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2562 mutex_enter(&itxg
->itxg_lock
);
2563 if (itxg
->itxg_txg
!= txg
) {
2564 mutex_exit(&itxg
->itxg_lock
);
2569 * If we're adding itx records to the zl_itx_commit_list,
2570 * then the zil better be dirty in this "txg". We can assert
2571 * that here since we're holding the itxg_lock which will
2572 * prevent spa_sync from cleaning it. Once we add the itxs
2573 * to the zl_itx_commit_list we must commit it to disk even
2574 * if it's unnecessary (i.e. the txg was synced).
2576 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
2577 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
2578 list_t
*sync_list
= &itxg
->itxg_itxs
->i_sync_list
;
2579 if (unlikely(zilog
->zl_suspend
> 0)) {
2581 * ZIL was just suspended, but we lost the race.
2582 * Allow all earlier itxs to be committed, but ask
2583 * caller to do txg_wait_synced(txg) for any new.
2585 if (!list_is_empty(sync_list
))
2586 wtxg
= MAX(wtxg
, txg
);
2588 list_move_tail(commit_list
, sync_list
);
2591 mutex_exit(&itxg
->itxg_lock
);
2597 * Move the async itxs for a specified object to commit into sync lists.
2600 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
2603 itx_async_node_t
*ian
, ian_search
;
2607 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2610 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2613 * This is inherently racy, since there is nothing to prevent
2614 * the last synced txg from changing.
2616 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2617 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2619 mutex_enter(&itxg
->itxg_lock
);
2620 if (itxg
->itxg_txg
!= txg
) {
2621 mutex_exit(&itxg
->itxg_lock
);
2626 * If a foid is specified then find that node and append its
2627 * list. Otherwise walk the tree appending all the lists
2628 * to the sync list. We add to the end rather than the
2629 * beginning to ensure the create has happened.
2631 t
= &itxg
->itxg_itxs
->i_async_tree
;
2633 ian_search
.ia_foid
= foid
;
2634 ian
= avl_find(t
, &ian_search
, &where
);
2636 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2640 void *cookie
= NULL
;
2642 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2643 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2645 list_destroy(&ian
->ia_list
);
2646 kmem_free(ian
, sizeof (itx_async_node_t
));
2649 mutex_exit(&itxg
->itxg_lock
);
2654 * This function will prune commit itxs that are at the head of the
2655 * commit list (it won't prune past the first non-commit itx), and
2656 * either: a) attach them to the last lwb that's still pending
2657 * completion, or b) skip them altogether.
2659 * This is used as a performance optimization to prevent commit itxs
2660 * from generating new lwbs when it's unnecessary to do so.
2663 zil_prune_commit_list(zilog_t
*zilog
)
2667 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2669 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2670 lr_t
*lrc
= &itx
->itx_lr
;
2671 if (lrc
->lrc_txtype
!= TX_COMMIT
)
2674 mutex_enter(&zilog
->zl_lock
);
2676 lwb_t
*last_lwb
= zilog
->zl_last_lwb_opened
;
2677 if (last_lwb
== NULL
||
2678 last_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
) {
2680 * All of the itxs this waiter was waiting on
2681 * must have already completed (or there were
2682 * never any itx's for it to wait on), so it's
2683 * safe to skip this waiter and mark it done.
2685 zil_commit_waiter_skip(itx
->itx_private
);
2687 zil_commit_waiter_link_lwb(itx
->itx_private
, last_lwb
);
2690 mutex_exit(&zilog
->zl_lock
);
2692 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2693 zil_itx_destroy(itx
);
2696 IMPLY(itx
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2700 zil_commit_writer_stall(zilog_t
*zilog
)
2703 * When zio_alloc_zil() fails to allocate the next lwb block on
2704 * disk, we must call txg_wait_synced() to ensure all of the
2705 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2706 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2707 * to zil_process_commit_list()) will have to call zil_create(),
2708 * and start a new ZIL chain.
2710 * Since zil_alloc_zil() failed, the lwb that was previously
2711 * issued does not have a pointer to the "next" lwb on disk.
2712 * Thus, if another ZIL writer thread was to allocate the "next"
2713 * on-disk lwb, that block could be leaked in the event of a
2714 * crash (because the previous lwb on-disk would not point to
2717 * We must hold the zilog's zl_issuer_lock while we do this, to
2718 * ensure no new threads enter zil_process_commit_list() until
2719 * all lwb's in the zl_lwb_list have been synced and freed
2720 * (which is achieved via the txg_wait_synced() call).
2722 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2723 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2724 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
2728 zil_burst_done(zilog_t
*zilog
)
2730 if (!list_is_empty(&zilog
->zl_itx_commit_list
) ||
2731 zilog
->zl_cur_used
== 0)
2734 if (zilog
->zl_parallel
)
2735 zilog
->zl_parallel
--;
2737 zilog
->zl_cur_used
= 0;
2741 * This function will traverse the commit list, creating new lwbs as
2742 * needed, and committing the itxs from the commit list to these newly
2743 * created lwbs. Additionally, as a new lwb is created, the previous
2744 * lwb will be issued to the zio layer to be written to disk.
2747 zil_process_commit_list(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
, list_t
*ilwbs
)
2749 spa_t
*spa
= zilog
->zl_spa
;
2751 list_t nolwb_waiters
;
2755 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2758 * Return if there's nothing to commit before we dirty the fs by
2759 * calling zil_create().
2761 if (list_is_empty(&zilog
->zl_itx_commit_list
))
2764 list_create(&nolwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2765 list_create(&nolwb_waiters
, sizeof (zil_commit_waiter_t
),
2766 offsetof(zil_commit_waiter_t
, zcw_node
));
2768 lwb
= list_tail(&zilog
->zl_lwb_list
);
2770 lwb
= zil_create(zilog
);
2773 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2774 * have already been created (zl_lwb_list not empty).
2776 zil_commit_activate_saxattr_feature(zilog
);
2777 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
2778 lwb
->lwb_state
== LWB_STATE_OPENED
);
2781 * If the lwb is still opened, it means the workload is really
2782 * multi-threaded and we won the chance of write aggregation.
2783 * If it is not opened yet, but previous lwb is still not
2784 * flushed, it still means the workload is multi-threaded, but
2785 * there was too much time between the commits to aggregate, so
2786 * we try aggregation next times, but without too much hopes.
2788 if (lwb
->lwb_state
== LWB_STATE_OPENED
) {
2789 zilog
->zl_parallel
= ZIL_BURSTS
;
2790 } else if ((plwb
= list_prev(&zilog
->zl_lwb_list
, lwb
))
2791 != NULL
&& plwb
->lwb_state
!= LWB_STATE_FLUSH_DONE
) {
2792 zilog
->zl_parallel
= MAX(zilog
->zl_parallel
,
2797 while ((itx
= list_remove_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2798 lr_t
*lrc
= &itx
->itx_lr
;
2799 uint64_t txg
= lrc
->lrc_txg
;
2801 ASSERT3U(txg
, !=, 0);
2803 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2804 DTRACE_PROBE2(zil__process__commit__itx
,
2805 zilog_t
*, zilog
, itx_t
*, itx
);
2807 DTRACE_PROBE2(zil__process__normal__itx
,
2808 zilog_t
*, zilog
, itx_t
*, itx
);
2811 boolean_t synced
= txg
<= spa_last_synced_txg(spa
);
2812 boolean_t frozen
= txg
> spa_freeze_txg(spa
);
2815 * If the txg of this itx has already been synced out, then
2816 * we don't need to commit this itx to an lwb. This is
2817 * because the data of this itx will have already been
2818 * written to the main pool. This is inherently racy, and
2819 * it's still ok to commit an itx whose txg has already
2820 * been synced; this will result in a write that's
2821 * unnecessary, but will do no harm.
2823 * With that said, we always want to commit TX_COMMIT itxs
2824 * to an lwb, regardless of whether or not that itx's txg
2825 * has been synced out. We do this to ensure any OPENED lwb
2826 * will always have at least one zil_commit_waiter_t linked
2829 * As a counter-example, if we skipped TX_COMMIT itx's
2830 * whose txg had already been synced, the following
2831 * situation could occur if we happened to be racing with
2834 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2835 * itx's txg is 10 and the last synced txg is 9.
2836 * 2. spa_sync finishes syncing out txg 10.
2837 * 3. We move to the next itx in the list, it's a TX_COMMIT
2838 * whose txg is 10, so we skip it rather than committing
2839 * it to the lwb used in (1).
2841 * If the itx that is skipped in (3) is the last TX_COMMIT
2842 * itx in the commit list, than it's possible for the lwb
2843 * used in (1) to remain in the OPENED state indefinitely.
2845 * To prevent the above scenario from occurring, ensuring
2846 * that once an lwb is OPENED it will transition to ISSUED
2847 * and eventually DONE, we always commit TX_COMMIT itx's to
2848 * an lwb here, even if that itx's txg has already been
2851 * Finally, if the pool is frozen, we _always_ commit the
2852 * itx. The point of freezing the pool is to prevent data
2853 * from being written to the main pool via spa_sync, and
2854 * instead rely solely on the ZIL to persistently store the
2855 * data; i.e. when the pool is frozen, the last synced txg
2856 * value can't be trusted.
2858 if (frozen
|| !synced
|| lrc
->lrc_txtype
== TX_COMMIT
) {
2860 lwb
= zil_lwb_assign(zilog
, lwb
, itx
, ilwbs
);
2862 list_insert_tail(&nolwb_itxs
, itx
);
2863 } else if ((zcw
->zcw_lwb
!= NULL
&&
2864 zcw
->zcw_lwb
!= lwb
) || zcw
->zcw_done
) {
2866 * Our lwb is done, leave the rest of
2867 * itx list to somebody else who care.
2869 zilog
->zl_parallel
= ZIL_BURSTS
;
2873 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2874 zil_commit_waiter_link_nolwb(
2875 itx
->itx_private
, &nolwb_waiters
);
2877 list_insert_tail(&nolwb_itxs
, itx
);
2880 ASSERT3S(lrc
->lrc_txtype
, !=, TX_COMMIT
);
2881 zil_itx_destroy(itx
);
2887 * This indicates zio_alloc_zil() failed to allocate the
2888 * "next" lwb on-disk. When this happens, we must stall
2889 * the ZIL write pipeline; see the comment within
2890 * zil_commit_writer_stall() for more details.
2892 while ((lwb
= list_remove_head(ilwbs
)) != NULL
)
2893 zil_lwb_write_issue(zilog
, lwb
);
2894 zil_commit_writer_stall(zilog
);
2897 * Additionally, we have to signal and mark the "nolwb"
2898 * waiters as "done" here, since without an lwb, we
2899 * can't do this via zil_lwb_flush_vdevs_done() like
2902 zil_commit_waiter_t
*zcw
;
2903 while ((zcw
= list_remove_head(&nolwb_waiters
)) != NULL
)
2904 zil_commit_waiter_skip(zcw
);
2907 * And finally, we have to destroy the itx's that
2908 * couldn't be committed to an lwb; this will also call
2909 * the itx's callback if one exists for the itx.
2911 while ((itx
= list_remove_head(&nolwb_itxs
)) != NULL
)
2912 zil_itx_destroy(itx
);
2914 ASSERT(list_is_empty(&nolwb_waiters
));
2915 ASSERT3P(lwb
, !=, NULL
);
2916 ASSERT(lwb
->lwb_state
== LWB_STATE_NEW
||
2917 lwb
->lwb_state
== LWB_STATE_OPENED
);
2920 * At this point, the ZIL block pointed at by the "lwb"
2921 * variable is in "new" or "opened" state.
2923 * If it's "new", then no itxs have been committed to it, so
2924 * there's no point in issuing its zio (i.e. it's "empty").
2926 * If it's "opened", then it contains one or more itxs that
2927 * eventually need to be committed to stable storage. In
2928 * this case we intentionally do not issue the lwb's zio
2929 * to disk yet, and instead rely on one of the following
2930 * two mechanisms for issuing the zio:
2932 * 1. Ideally, there will be more ZIL activity occurring on
2933 * the system, such that this function will be immediately
2934 * called again by different thread and this lwb will be
2935 * closed by zil_lwb_assign(). This way, the lwb will be
2936 * "full" when it is issued to disk, and we'll make use of
2937 * the lwb's size the best we can.
2939 * 2. If there isn't sufficient ZIL activity occurring on
2940 * the system, zil_commit_waiter() will close it and issue
2941 * the zio. If this occurs, the lwb is not guaranteed
2942 * to be "full" by the time its zio is issued, and means
2943 * the size of the lwb was "too large" given the amount
2944 * of ZIL activity occurring on the system at that time.
2946 * We do this for a couple of reasons:
2948 * 1. To try and reduce the number of IOPs needed to
2949 * write the same number of itxs. If an lwb has space
2950 * available in its buffer for more itxs, and more itxs
2951 * will be committed relatively soon (relative to the
2952 * latency of performing a write), then it's beneficial
2953 * to wait for these "next" itxs. This way, more itxs
2954 * can be committed to stable storage with fewer writes.
2956 * 2. To try and use the largest lwb block size that the
2957 * incoming rate of itxs can support. Again, this is to
2958 * try and pack as many itxs into as few lwbs as
2959 * possible, without significantly impacting the latency
2960 * of each individual itx.
2962 if (lwb
->lwb_state
== LWB_STATE_OPENED
&& !zilog
->zl_parallel
) {
2963 list_insert_tail(ilwbs
, lwb
);
2964 lwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_NEW
);
2965 zil_burst_done(zilog
);
2967 while ((lwb
= list_remove_head(ilwbs
)) != NULL
)
2968 zil_lwb_write_issue(zilog
, lwb
);
2969 zil_commit_writer_stall(zilog
);
2976 * This function is responsible for ensuring the passed in commit waiter
2977 * (and associated commit itx) is committed to an lwb. If the waiter is
2978 * not already committed to an lwb, all itxs in the zilog's queue of
2979 * itxs will be processed. The assumption is the passed in waiter's
2980 * commit itx will found in the queue just like the other non-commit
2981 * itxs, such that when the entire queue is processed, the waiter will
2982 * have been committed to an lwb.
2984 * The lwb associated with the passed in waiter is not guaranteed to
2985 * have been issued by the time this function completes. If the lwb is
2986 * not issued, we rely on future calls to zil_commit_writer() to issue
2987 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2990 zil_commit_writer(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2996 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2997 ASSERT(spa_writeable(zilog
->zl_spa
));
2999 list_create(&ilwbs
, sizeof (lwb_t
), offsetof(lwb_t
, lwb_issue_node
));
3000 mutex_enter(&zilog
->zl_issuer_lock
);
3002 if (zcw
->zcw_lwb
!= NULL
|| zcw
->zcw_done
) {
3004 * It's possible that, while we were waiting to acquire
3005 * the "zl_issuer_lock", another thread committed this
3006 * waiter to an lwb. If that occurs, we bail out early,
3007 * without processing any of the zilog's queue of itxs.
3009 * On certain workloads and system configurations, the
3010 * "zl_issuer_lock" can become highly contended. In an
3011 * attempt to reduce this contention, we immediately drop
3012 * the lock if the waiter has already been processed.
3014 * We've measured this optimization to reduce CPU spent
3015 * contending on this lock by up to 5%, using a system
3016 * with 32 CPUs, low latency storage (~50 usec writes),
3017 * and 1024 threads performing sync writes.
3022 ZIL_STAT_BUMP(zilog
, zil_commit_writer_count
);
3024 wtxg
= zil_get_commit_list(zilog
);
3025 zil_prune_commit_list(zilog
);
3026 zil_process_commit_list(zilog
, zcw
, &ilwbs
);
3029 mutex_exit(&zilog
->zl_issuer_lock
);
3030 while ((lwb
= list_remove_head(&ilwbs
)) != NULL
)
3031 zil_lwb_write_issue(zilog
, lwb
);
3032 list_destroy(&ilwbs
);
3037 zil_commit_waiter_timeout(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3039 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
3040 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
3041 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
3043 lwb_t
*lwb
= zcw
->zcw_lwb
;
3044 ASSERT3P(lwb
, !=, NULL
);
3045 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_NEW
);
3048 * If the lwb has already been issued by another thread, we can
3049 * immediately return since there's no work to be done (the
3050 * point of this function is to issue the lwb). Additionally, we
3051 * do this prior to acquiring the zl_issuer_lock, to avoid
3052 * acquiring it when it's not necessary to do so.
3054 if (lwb
->lwb_state
!= LWB_STATE_OPENED
)
3058 * In order to call zil_lwb_write_close() we must hold the
3059 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3060 * since we're already holding the commit waiter's "zcw_lock",
3061 * and those two locks are acquired in the opposite order
3064 mutex_exit(&zcw
->zcw_lock
);
3065 mutex_enter(&zilog
->zl_issuer_lock
);
3066 mutex_enter(&zcw
->zcw_lock
);
3069 * Since we just dropped and re-acquired the commit waiter's
3070 * lock, we have to re-check to see if the waiter was marked
3071 * "done" during that process. If the waiter was marked "done",
3072 * the "lwb" pointer is no longer valid (it can be free'd after
3073 * the waiter is marked "done"), so without this check we could
3074 * wind up with a use-after-free error below.
3076 if (zcw
->zcw_done
) {
3077 mutex_exit(&zilog
->zl_issuer_lock
);
3081 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
3084 * We've already checked this above, but since we hadn't acquired
3085 * the zilog's zl_issuer_lock, we have to perform this check a
3086 * second time while holding the lock.
3088 * We don't need to hold the zl_lock since the lwb cannot transition
3089 * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb
3090 * _can_ transition from CLOSED to DONE, but it's OK to race with
3091 * that transition since we treat the lwb the same, whether it's in
3092 * the CLOSED, ISSUED or DONE states.
3094 * The important thing, is we treat the lwb differently depending on
3095 * if it's OPENED or CLOSED, and block any other threads that might
3096 * attempt to close/issue this lwb. For that reason we hold the
3097 * zl_issuer_lock when checking the lwb_state; we must not call
3098 * zil_lwb_write_close() if the lwb had already been closed/issued.
3100 * See the comment above the lwb_state_t structure definition for
3101 * more details on the lwb states, and locking requirements.
3103 if (lwb
->lwb_state
!= LWB_STATE_OPENED
) {
3104 mutex_exit(&zilog
->zl_issuer_lock
);
3109 * We do not need zcw_lock once we hold zl_issuer_lock and know lwb
3110 * is still open. But we have to drop it to avoid a deadlock in case
3111 * callback of zio issued by zil_lwb_write_issue() try to get it,
3112 * while zil_lwb_write_issue() is blocked on attempt to issue next
3113 * lwb it found in LWB_STATE_READY state.
3115 mutex_exit(&zcw
->zcw_lock
);
3118 * As described in the comments above zil_commit_waiter() and
3119 * zil_process_commit_list(), we need to issue this lwb's zio
3120 * since we've reached the commit waiter's timeout and it still
3121 * hasn't been issued.
3123 lwb_t
*nlwb
= zil_lwb_write_close(zilog
, lwb
, LWB_STATE_NEW
);
3125 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_CLOSED
);
3127 zil_burst_done(zilog
);
3131 * When zil_lwb_write_close() returns NULL, this
3132 * indicates zio_alloc_zil() failed to allocate the
3133 * "next" lwb on-disk. When this occurs, the ZIL write
3134 * pipeline must be stalled; see the comment within the
3135 * zil_commit_writer_stall() function for more details.
3137 zil_lwb_write_issue(zilog
, lwb
);
3138 zil_commit_writer_stall(zilog
);
3139 mutex_exit(&zilog
->zl_issuer_lock
);
3141 mutex_exit(&zilog
->zl_issuer_lock
);
3142 zil_lwb_write_issue(zilog
, lwb
);
3144 mutex_enter(&zcw
->zcw_lock
);
3148 * This function is responsible for performing the following two tasks:
3150 * 1. its primary responsibility is to block until the given "commit
3151 * waiter" is considered "done".
3153 * 2. its secondary responsibility is to issue the zio for the lwb that
3154 * the given "commit waiter" is waiting on, if this function has
3155 * waited "long enough" and the lwb is still in the "open" state.
3157 * Given a sufficient amount of itxs being generated and written using
3158 * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3159 * function. If this does not occur, this secondary responsibility will
3160 * ensure the lwb is issued even if there is not other synchronous
3161 * activity on the system.
3163 * For more details, see zil_process_commit_list(); more specifically,
3164 * the comment at the bottom of that function.
3167 zil_commit_waiter(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3169 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
3170 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
3171 ASSERT(spa_writeable(zilog
->zl_spa
));
3173 mutex_enter(&zcw
->zcw_lock
);
3176 * The timeout is scaled based on the lwb latency to avoid
3177 * significantly impacting the latency of each individual itx.
3178 * For more details, see the comment at the bottom of the
3179 * zil_process_commit_list() function.
3181 int pct
= MAX(zfs_commit_timeout_pct
, 1);
3182 hrtime_t sleep
= (zilog
->zl_last_lwb_latency
* pct
) / 100;
3183 hrtime_t wakeup
= gethrtime() + sleep
;
3184 boolean_t timedout
= B_FALSE
;
3186 while (!zcw
->zcw_done
) {
3187 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
3189 lwb_t
*lwb
= zcw
->zcw_lwb
;
3192 * Usually, the waiter will have a non-NULL lwb field here,
3193 * but it's possible for it to be NULL as a result of
3194 * zil_commit() racing with spa_sync().
3196 * When zil_clean() is called, it's possible for the itxg
3197 * list (which may be cleaned via a taskq) to contain
3198 * commit itxs. When this occurs, the commit waiters linked
3199 * off of these commit itxs will not be committed to an
3200 * lwb. Additionally, these commit waiters will not be
3201 * marked done until zil_commit_waiter_skip() is called via
3204 * Thus, it's possible for this commit waiter (i.e. the
3205 * "zcw" variable) to be found in this "in between" state;
3206 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3207 * been skipped, so it's "zcw_done" field is still B_FALSE.
3209 IMPLY(lwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_NEW
);
3211 if (lwb
!= NULL
&& lwb
->lwb_state
== LWB_STATE_OPENED
) {
3212 ASSERT3B(timedout
, ==, B_FALSE
);
3215 * If the lwb hasn't been issued yet, then we
3216 * need to wait with a timeout, in case this
3217 * function needs to issue the lwb after the
3218 * timeout is reached; responsibility (2) from
3219 * the comment above this function.
3221 int rc
= cv_timedwait_hires(&zcw
->zcw_cv
,
3222 &zcw
->zcw_lock
, wakeup
, USEC2NSEC(1),
3223 CALLOUT_FLAG_ABSOLUTE
);
3225 if (rc
!= -1 || zcw
->zcw_done
)
3229 zil_commit_waiter_timeout(zilog
, zcw
);
3231 if (!zcw
->zcw_done
) {
3233 * If the commit waiter has already been
3234 * marked "done", it's possible for the
3235 * waiter's lwb structure to have already
3236 * been freed. Thus, we can only reliably
3237 * make these assertions if the waiter
3240 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
3241 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_OPENED
);
3245 * If the lwb isn't open, then it must have already
3246 * been issued. In that case, there's no need to
3247 * use a timeout when waiting for the lwb to
3250 * Additionally, if the lwb is NULL, the waiter
3251 * will soon be signaled and marked done via
3252 * zil_clean() and zil_itxg_clean(), so no timeout
3257 lwb
->lwb_state
== LWB_STATE_CLOSED
||
3258 lwb
->lwb_state
== LWB_STATE_READY
||
3259 lwb
->lwb_state
== LWB_STATE_ISSUED
||
3260 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
3261 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
3262 cv_wait(&zcw
->zcw_cv
, &zcw
->zcw_lock
);
3266 mutex_exit(&zcw
->zcw_lock
);
3269 static zil_commit_waiter_t
*
3270 zil_alloc_commit_waiter(void)
3272 zil_commit_waiter_t
*zcw
= kmem_cache_alloc(zil_zcw_cache
, KM_SLEEP
);
3274 cv_init(&zcw
->zcw_cv
, NULL
, CV_DEFAULT
, NULL
);
3275 mutex_init(&zcw
->zcw_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3276 list_link_init(&zcw
->zcw_node
);
3277 zcw
->zcw_lwb
= NULL
;
3278 zcw
->zcw_done
= B_FALSE
;
3279 zcw
->zcw_zio_error
= 0;
3285 zil_free_commit_waiter(zil_commit_waiter_t
*zcw
)
3287 ASSERT(!list_link_active(&zcw
->zcw_node
));
3288 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
3289 ASSERT3B(zcw
->zcw_done
, ==, B_TRUE
);
3290 mutex_destroy(&zcw
->zcw_lock
);
3291 cv_destroy(&zcw
->zcw_cv
);
3292 kmem_cache_free(zil_zcw_cache
, zcw
);
3296 * This function is used to create a TX_COMMIT itx and assign it. This
3297 * way, it will be linked into the ZIL's list of synchronous itxs, and
3298 * then later committed to an lwb (or skipped) when
3299 * zil_process_commit_list() is called.
3302 zil_commit_itx_assign(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
3304 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
3307 * Since we are not going to create any new dirty data, and we
3308 * can even help with clearing the existing dirty data, we
3309 * should not be subject to the dirty data based delays. We
3310 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3312 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
3314 itx_t
*itx
= zil_itx_create(TX_COMMIT
, sizeof (lr_t
));
3315 itx
->itx_sync
= B_TRUE
;
3316 itx
->itx_private
= zcw
;
3318 zil_itx_assign(zilog
, itx
, tx
);
3324 * Commit ZFS Intent Log transactions (itxs) to stable storage.
3326 * When writing ZIL transactions to the on-disk representation of the
3327 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3328 * itxs can be committed to a single lwb. Once a lwb is written and
3329 * committed to stable storage (i.e. the lwb is written, and vdevs have
3330 * been flushed), each itx that was committed to that lwb is also
3331 * considered to be committed to stable storage.
3333 * When an itx is committed to an lwb, the log record (lr_t) contained
3334 * by the itx is copied into the lwb's zio buffer, and once this buffer
3335 * is written to disk, it becomes an on-disk ZIL block.
3337 * As itxs are generated, they're inserted into the ZIL's queue of
3338 * uncommitted itxs. The semantics of zil_commit() are such that it will
3339 * block until all itxs that were in the queue when it was called, are
3340 * committed to stable storage.
3342 * If "foid" is zero, this means all "synchronous" and "asynchronous"
3343 * itxs, for all objects in the dataset, will be committed to stable
3344 * storage prior to zil_commit() returning. If "foid" is non-zero, all
3345 * "synchronous" itxs for all objects, but only "asynchronous" itxs
3346 * that correspond to the foid passed in, will be committed to stable
3347 * storage prior to zil_commit() returning.
3349 * Generally speaking, when zil_commit() is called, the consumer doesn't
3350 * actually care about _all_ of the uncommitted itxs. Instead, they're
3351 * simply trying to waiting for a specific itx to be committed to disk,
3352 * but the interface(s) for interacting with the ZIL don't allow such
3353 * fine-grained communication. A better interface would allow a consumer
3354 * to create and assign an itx, and then pass a reference to this itx to
3355 * zil_commit(); such that zil_commit() would return as soon as that
3356 * specific itx was committed to disk (instead of waiting for _all_
3357 * itxs to be committed).
3359 * When a thread calls zil_commit() a special "commit itx" will be
3360 * generated, along with a corresponding "waiter" for this commit itx.
3361 * zil_commit() will wait on this waiter's CV, such that when the waiter
3362 * is marked done, and signaled, zil_commit() will return.
3364 * This commit itx is inserted into the queue of uncommitted itxs. This
3365 * provides an easy mechanism for determining which itxs were in the
3366 * queue prior to zil_commit() having been called, and which itxs were
3367 * added after zil_commit() was called.
3369 * The commit itx is special; it doesn't have any on-disk representation.
3370 * When a commit itx is "committed" to an lwb, the waiter associated
3371 * with it is linked onto the lwb's list of waiters. Then, when that lwb
3372 * completes, each waiter on the lwb's list is marked done and signaled
3373 * -- allowing the thread waiting on the waiter to return from zil_commit().
3375 * It's important to point out a few critical factors that allow us
3376 * to make use of the commit itxs, commit waiters, per-lwb lists of
3377 * commit waiters, and zio completion callbacks like we're doing:
3379 * 1. The list of waiters for each lwb is traversed, and each commit
3380 * waiter is marked "done" and signaled, in the zio completion
3381 * callback of the lwb's zio[*].
3383 * * Actually, the waiters are signaled in the zio completion
3384 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3385 * that are sent to the vdevs upon completion of the lwb zio.
3387 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
3388 * itxs, the order in which they are inserted is preserved[*]; as
3389 * itxs are added to the queue, they are added to the tail of
3390 * in-memory linked lists.
3392 * When committing the itxs to lwbs (to be written to disk), they
3393 * are committed in the same order in which the itxs were added to
3394 * the uncommitted queue's linked list(s); i.e. the linked list of
3395 * itxs to commit is traversed from head to tail, and each itx is
3396 * committed to an lwb in that order.
3400 * - the order of "sync" itxs is preserved w.r.t. other
3401 * "sync" itxs, regardless of the corresponding objects.
3402 * - the order of "async" itxs is preserved w.r.t. other
3403 * "async" itxs corresponding to the same object.
3404 * - the order of "async" itxs is *not* preserved w.r.t. other
3405 * "async" itxs corresponding to different objects.
3406 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
3407 * versa) is *not* preserved, even for itxs that correspond
3408 * to the same object.
3410 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
3411 * zil_get_commit_list(), and zil_process_commit_list().
3413 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
3414 * lwb cannot be considered committed to stable storage, until its
3415 * "previous" lwb is also committed to stable storage. This fact,
3416 * coupled with the fact described above, means that itxs are
3417 * committed in (roughly) the order in which they were generated.
3418 * This is essential because itxs are dependent on prior itxs.
3419 * Thus, we *must not* deem an itx as being committed to stable
3420 * storage, until *all* prior itxs have also been committed to
3423 * To enforce this ordering of lwb zio's, while still leveraging as
3424 * much of the underlying storage performance as possible, we rely
3425 * on two fundamental concepts:
3427 * 1. The creation and issuance of lwb zio's is protected by
3428 * the zilog's "zl_issuer_lock", which ensures only a single
3429 * thread is creating and/or issuing lwb's at a time
3430 * 2. The "previous" lwb is a child of the "current" lwb
3431 * (leveraging the zio parent-child dependency graph)
3433 * By relying on this parent-child zio relationship, we can have
3434 * many lwb zio's concurrently issued to the underlying storage,
3435 * but the order in which they complete will be the same order in
3436 * which they were created.
3439 zil_commit(zilog_t
*zilog
, uint64_t foid
)
3442 * We should never attempt to call zil_commit on a snapshot for
3443 * a couple of reasons:
3445 * 1. A snapshot may never be modified, thus it cannot have any
3446 * in-flight itxs that would have modified the dataset.
3448 * 2. By design, when zil_commit() is called, a commit itx will
3449 * be assigned to this zilog; as a result, the zilog will be
3450 * dirtied. We must not dirty the zilog of a snapshot; there's
3451 * checks in the code that enforce this invariant, and will
3452 * cause a panic if it's not upheld.
3454 ASSERT3B(dmu_objset_is_snapshot(zilog
->zl_os
), ==, B_FALSE
);
3456 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3459 if (!spa_writeable(zilog
->zl_spa
)) {
3461 * If the SPA is not writable, there should never be any
3462 * pending itxs waiting to be committed to disk. If that
3463 * weren't true, we'd skip writing those itxs out, and
3464 * would break the semantics of zil_commit(); thus, we're
3465 * verifying that truth before we return to the caller.
3467 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3468 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3469 for (int i
= 0; i
< TXG_SIZE
; i
++)
3470 ASSERT3P(zilog
->zl_itxg
[i
].itxg_itxs
, ==, NULL
);
3475 * If the ZIL is suspended, we don't want to dirty it by calling
3476 * zil_commit_itx_assign() below, nor can we write out
3477 * lwbs like would be done in zil_commit_write(). Thus, we
3478 * simply rely on txg_wait_synced() to maintain the necessary
3479 * semantics, and avoid calling those functions altogether.
3481 if (zilog
->zl_suspend
> 0) {
3482 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3486 zil_commit_impl(zilog
, foid
);
3490 zil_commit_impl(zilog_t
*zilog
, uint64_t foid
)
3492 ZIL_STAT_BUMP(zilog
, zil_commit_count
);
3495 * Move the "async" itxs for the specified foid to the "sync"
3496 * queues, such that they will be later committed (or skipped)
3497 * to an lwb when zil_process_commit_list() is called.
3499 * Since these "async" itxs must be committed prior to this
3500 * call to zil_commit returning, we must perform this operation
3501 * before we call zil_commit_itx_assign().
3503 zil_async_to_sync(zilog
, foid
);
3506 * We allocate a new "waiter" structure which will initially be
3507 * linked to the commit itx using the itx's "itx_private" field.
3508 * Since the commit itx doesn't represent any on-disk state,
3509 * when it's committed to an lwb, rather than copying the its
3510 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3511 * added to the lwb's list of waiters. Then, when the lwb is
3512 * committed to stable storage, each waiter in the lwb's list of
3513 * waiters will be marked "done", and signalled.
3515 * We must create the waiter and assign the commit itx prior to
3516 * calling zil_commit_writer(), or else our specific commit itx
3517 * is not guaranteed to be committed to an lwb prior to calling
3518 * zil_commit_waiter().
3520 zil_commit_waiter_t
*zcw
= zil_alloc_commit_waiter();
3521 zil_commit_itx_assign(zilog
, zcw
);
3523 uint64_t wtxg
= zil_commit_writer(zilog
, zcw
);
3524 zil_commit_waiter(zilog
, zcw
);
3526 if (zcw
->zcw_zio_error
!= 0) {
3528 * If there was an error writing out the ZIL blocks that
3529 * this thread is waiting on, then we fallback to
3530 * relying on spa_sync() to write out the data this
3531 * thread is waiting on. Obviously this has performance
3532 * implications, but the expectation is for this to be
3533 * an exceptional case, and shouldn't occur often.
3535 DTRACE_PROBE2(zil__commit__io__error
,
3536 zilog_t
*, zilog
, zil_commit_waiter_t
*, zcw
);
3537 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3538 } else if (wtxg
!= 0) {
3539 txg_wait_synced(zilog
->zl_dmu_pool
, wtxg
);
3542 zil_free_commit_waiter(zcw
);
3546 * Called in syncing context to free committed log blocks and update log header.
3549 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
3551 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
3552 uint64_t txg
= dmu_tx_get_txg(tx
);
3553 spa_t
*spa
= zilog
->zl_spa
;
3554 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
3558 * We don't zero out zl_destroy_txg, so make sure we don't try
3559 * to destroy it twice.
3561 if (spa_sync_pass(spa
) != 1)
3564 zil_lwb_flush_wait_all(zilog
, txg
);
3566 mutex_enter(&zilog
->zl_lock
);
3568 ASSERT(zilog
->zl_stop_sync
== 0);
3570 if (*replayed_seq
!= 0) {
3571 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
3572 zh
->zh_replay_seq
= *replayed_seq
;
3576 if (zilog
->zl_destroy_txg
== txg
) {
3577 blkptr_t blk
= zh
->zh_log
;
3578 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
3580 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3582 memset(zh
, 0, sizeof (zil_header_t
));
3583 memset(zilog
->zl_replayed_seq
, 0,
3584 sizeof (zilog
->zl_replayed_seq
));
3586 if (zilog
->zl_keep_first
) {
3588 * If this block was part of log chain that couldn't
3589 * be claimed because a device was missing during
3590 * zil_claim(), but that device later returns,
3591 * then this block could erroneously appear valid.
3592 * To guard against this, assign a new GUID to the new
3593 * log chain so it doesn't matter what blk points to.
3595 zil_init_log_chain(zilog
, &blk
);
3599 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3600 * records. So, deactivate the feature for this dataset.
3601 * We activate it again when we start a new ZIL chain.
3603 if (dsl_dataset_feature_is_active(ds
,
3604 SPA_FEATURE_ZILSAXATTR
))
3605 dsl_dataset_deactivate_feature(ds
,
3606 SPA_FEATURE_ZILSAXATTR
, tx
);
3610 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
3611 zh
->zh_log
= lwb
->lwb_blk
;
3612 if (lwb
->lwb_state
!= LWB_STATE_FLUSH_DONE
||
3613 lwb
->lwb_alloc_txg
> txg
|| lwb
->lwb_max_txg
> txg
)
3615 list_remove(&zilog
->zl_lwb_list
, lwb
);
3616 if (!BP_IS_HOLE(&lwb
->lwb_blk
))
3617 zio_free(spa
, txg
, &lwb
->lwb_blk
);
3618 zil_free_lwb(zilog
, lwb
);
3621 * If we don't have anything left in the lwb list then
3622 * we've had an allocation failure and we need to zero
3623 * out the zil_header blkptr so that we don't end
3624 * up freeing the same block twice.
3626 if (list_is_empty(&zilog
->zl_lwb_list
))
3627 BP_ZERO(&zh
->zh_log
);
3630 mutex_exit(&zilog
->zl_lock
);
3634 zil_lwb_cons(void *vbuf
, void *unused
, int kmflag
)
3636 (void) unused
, (void) kmflag
;
3638 list_create(&lwb
->lwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
3639 list_create(&lwb
->lwb_waiters
, sizeof (zil_commit_waiter_t
),
3640 offsetof(zil_commit_waiter_t
, zcw_node
));
3641 avl_create(&lwb
->lwb_vdev_tree
, zil_lwb_vdev_compare
,
3642 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
3643 mutex_init(&lwb
->lwb_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3648 zil_lwb_dest(void *vbuf
, void *unused
)
3652 mutex_destroy(&lwb
->lwb_vdev_lock
);
3653 avl_destroy(&lwb
->lwb_vdev_tree
);
3654 list_destroy(&lwb
->lwb_waiters
);
3655 list_destroy(&lwb
->lwb_itxs
);
3661 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
3662 sizeof (lwb_t
), 0, zil_lwb_cons
, zil_lwb_dest
, NULL
, NULL
, NULL
, 0);
3664 zil_zcw_cache
= kmem_cache_create("zil_zcw_cache",
3665 sizeof (zil_commit_waiter_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
3667 zil_sums_init(&zil_sums_global
);
3668 zil_kstats_global
= kstat_create("zfs", 0, "zil", "misc",
3669 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
3670 KSTAT_FLAG_VIRTUAL
);
3672 if (zil_kstats_global
!= NULL
) {
3673 zil_kstats_global
->ks_data
= &zil_stats
;
3674 zil_kstats_global
->ks_update
= zil_kstats_global_update
;
3675 zil_kstats_global
->ks_private
= NULL
;
3676 kstat_install(zil_kstats_global
);
3683 kmem_cache_destroy(zil_zcw_cache
);
3684 kmem_cache_destroy(zil_lwb_cache
);
3686 if (zil_kstats_global
!= NULL
) {
3687 kstat_delete(zil_kstats_global
);
3688 zil_kstats_global
= NULL
;
3691 zil_sums_fini(&zil_sums_global
);
3695 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
3697 zilog
->zl_sync
= sync
;
3701 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
3703 zilog
->zl_logbias
= logbias
;
3707 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
3711 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
3713 zilog
->zl_header
= zh_phys
;
3715 zilog
->zl_spa
= dmu_objset_spa(os
);
3716 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
3717 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
3718 zilog
->zl_logbias
= dmu_objset_logbias(os
);
3719 zilog
->zl_sync
= dmu_objset_syncprop(os
);
3720 zilog
->zl_dirty_max_txg
= 0;
3721 zilog
->zl_last_lwb_opened
= NULL
;
3722 zilog
->zl_last_lwb_latency
= 0;
3723 zilog
->zl_max_block_size
= zil_maxblocksize
;
3725 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3726 mutex_init(&zilog
->zl_issuer_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3727 mutex_init(&zilog
->zl_lwb_io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3729 for (int i
= 0; i
< TXG_SIZE
; i
++) {
3730 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
3731 MUTEX_DEFAULT
, NULL
);
3734 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
3735 offsetof(lwb_t
, lwb_node
));
3737 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
3738 offsetof(itx_t
, itx_node
));
3740 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
3741 cv_init(&zilog
->zl_lwb_io_cv
, NULL
, CV_DEFAULT
, NULL
);
3747 zil_free(zilog_t
*zilog
)
3751 zilog
->zl_stop_sync
= 1;
3753 ASSERT0(zilog
->zl_suspend
);
3754 ASSERT0(zilog
->zl_suspending
);
3756 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3757 list_destroy(&zilog
->zl_lwb_list
);
3759 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
3760 list_destroy(&zilog
->zl_itx_commit_list
);
3762 for (i
= 0; i
< TXG_SIZE
; i
++) {
3764 * It's possible for an itx to be generated that doesn't dirty
3765 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3766 * callback to remove the entry. We remove those here.
3768 * Also free up the ziltest itxs.
3770 if (zilog
->zl_itxg
[i
].itxg_itxs
)
3771 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
3772 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
3775 mutex_destroy(&zilog
->zl_issuer_lock
);
3776 mutex_destroy(&zilog
->zl_lock
);
3777 mutex_destroy(&zilog
->zl_lwb_io_lock
);
3779 cv_destroy(&zilog
->zl_cv_suspend
);
3780 cv_destroy(&zilog
->zl_lwb_io_cv
);
3782 kmem_free(zilog
, sizeof (zilog_t
));
3786 * Open an intent log.
3789 zil_open(objset_t
*os
, zil_get_data_t
*get_data
, zil_sums_t
*zil_sums
)
3791 zilog_t
*zilog
= dmu_objset_zil(os
);
3793 ASSERT3P(zilog
->zl_get_data
, ==, NULL
);
3794 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3795 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3797 zilog
->zl_get_data
= get_data
;
3798 zilog
->zl_sums
= zil_sums
;
3804 * Close an intent log.
3807 zil_close(zilog_t
*zilog
)
3812 if (!dmu_objset_is_snapshot(zilog
->zl_os
)) {
3813 zil_commit(zilog
, 0);
3815 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3816 ASSERT0(zilog
->zl_dirty_max_txg
);
3817 ASSERT3B(zilog_is_dirty(zilog
), ==, B_FALSE
);
3820 mutex_enter(&zilog
->zl_lock
);
3821 txg
= zilog
->zl_dirty_max_txg
;
3822 lwb
= list_tail(&zilog
->zl_lwb_list
);
3824 txg
= MAX(txg
, lwb
->lwb_alloc_txg
);
3825 txg
= MAX(txg
, lwb
->lwb_max_txg
);
3827 mutex_exit(&zilog
->zl_lock
);
3830 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3831 * on the time when the dmu_tx transaction is assigned in
3832 * zil_lwb_write_issue().
3834 mutex_enter(&zilog
->zl_lwb_io_lock
);
3835 txg
= MAX(zilog
->zl_lwb_max_issued_txg
, txg
);
3836 mutex_exit(&zilog
->zl_lwb_io_lock
);
3839 * We need to use txg_wait_synced() to wait until that txg is synced.
3840 * zil_sync() will guarantee all lwbs up to that txg have been
3841 * written out, flushed, and cleaned.
3844 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
3846 if (zilog_is_dirty(zilog
))
3847 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog
,
3849 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
3850 VERIFY(!zilog_is_dirty(zilog
));
3852 zilog
->zl_get_data
= NULL
;
3855 * We should have only one lwb left on the list; remove it now.
3857 mutex_enter(&zilog
->zl_lock
);
3858 lwb
= list_remove_head(&zilog
->zl_lwb_list
);
3860 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3861 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_NEW
);
3862 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
3863 zil_free_lwb(zilog
, lwb
);
3865 mutex_exit(&zilog
->zl_lock
);
3868 static const char *suspend_tag
= "zil suspending";
3871 * Suspend an intent log. While in suspended mode, we still honor
3872 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3873 * On old version pools, we suspend the log briefly when taking a
3874 * snapshot so that it will have an empty intent log.
3876 * Long holds are not really intended to be used the way we do here --
3877 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3878 * could fail. Therefore we take pains to only put a long hold if it is
3879 * actually necessary. Fortunately, it will only be necessary if the
3880 * objset is currently mounted (or the ZVOL equivalent). In that case it
3881 * will already have a long hold, so we are not really making things any worse.
3883 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3884 * zvol_state_t), and use their mechanism to prevent their hold from being
3885 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3888 * if cookiep == NULL, this does both the suspend & resume.
3889 * Otherwise, it returns with the dataset "long held", and the cookie
3890 * should be passed into zil_resume().
3893 zil_suspend(const char *osname
, void **cookiep
)
3897 const zil_header_t
*zh
;
3900 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
3903 zilog
= dmu_objset_zil(os
);
3905 mutex_enter(&zilog
->zl_lock
);
3906 zh
= zilog
->zl_header
;
3908 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
3909 mutex_exit(&zilog
->zl_lock
);
3910 dmu_objset_rele(os
, suspend_tag
);
3911 return (SET_ERROR(EBUSY
));
3915 * Don't put a long hold in the cases where we can avoid it. This
3916 * is when there is no cookie so we are doing a suspend & resume
3917 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3918 * for the suspend because it's already suspended, or there's no ZIL.
3920 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
3921 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
3922 mutex_exit(&zilog
->zl_lock
);
3923 dmu_objset_rele(os
, suspend_tag
);
3927 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
3928 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
3930 zilog
->zl_suspend
++;
3932 if (zilog
->zl_suspend
> 1) {
3934 * Someone else is already suspending it.
3935 * Just wait for them to finish.
3938 while (zilog
->zl_suspending
)
3939 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
3940 mutex_exit(&zilog
->zl_lock
);
3942 if (cookiep
== NULL
)
3950 * If there is no pointer to an on-disk block, this ZIL must not
3951 * be active (e.g. filesystem not mounted), so there's nothing
3954 if (BP_IS_HOLE(&zh
->zh_log
)) {
3955 ASSERT(cookiep
!= NULL
); /* fast path already handled */
3958 mutex_exit(&zilog
->zl_lock
);
3963 * The ZIL has work to do. Ensure that the associated encryption
3964 * key will remain mapped while we are committing the log by
3965 * grabbing a reference to it. If the key isn't loaded we have no
3966 * choice but to return an error until the wrapping key is loaded.
3968 if (os
->os_encrypted
&&
3969 dsl_dataset_create_key_mapping(dmu_objset_ds(os
)) != 0) {
3970 zilog
->zl_suspend
--;
3971 mutex_exit(&zilog
->zl_lock
);
3972 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3973 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3974 return (SET_ERROR(EACCES
));
3977 zilog
->zl_suspending
= B_TRUE
;
3978 mutex_exit(&zilog
->zl_lock
);
3981 * We need to use zil_commit_impl to ensure we wait for all
3982 * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed
3983 * to disk before proceeding. If we used zil_commit instead, it
3984 * would just call txg_wait_synced(), because zl_suspend is set.
3985 * txg_wait_synced() doesn't wait for these lwb's to be
3986 * LWB_STATE_FLUSH_DONE before returning.
3988 zil_commit_impl(zilog
, 0);
3991 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3992 * use txg_wait_synced() to ensure the data from the zilog has
3993 * migrated to the main pool before calling zil_destroy().
3995 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3997 zil_destroy(zilog
, B_FALSE
);
3999 mutex_enter(&zilog
->zl_lock
);
4000 zilog
->zl_suspending
= B_FALSE
;
4001 cv_broadcast(&zilog
->zl_cv_suspend
);
4002 mutex_exit(&zilog
->zl_lock
);
4004 if (os
->os_encrypted
)
4005 dsl_dataset_remove_key_mapping(dmu_objset_ds(os
));
4007 if (cookiep
== NULL
)
4015 zil_resume(void *cookie
)
4017 objset_t
*os
= cookie
;
4018 zilog_t
*zilog
= dmu_objset_zil(os
);
4020 mutex_enter(&zilog
->zl_lock
);
4021 ASSERT(zilog
->zl_suspend
!= 0);
4022 zilog
->zl_suspend
--;
4023 mutex_exit(&zilog
->zl_lock
);
4024 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
4025 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
4028 typedef struct zil_replay_arg
{
4029 zil_replay_func_t
*const *zr_replay
;
4031 boolean_t zr_byteswap
;
4036 zil_replay_error(zilog_t
*zilog
, const lr_t
*lr
, int error
)
4038 char name
[ZFS_MAX_DATASET_NAME_LEN
];
4040 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
4042 dmu_objset_name(zilog
->zl_os
, name
);
4044 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
4045 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
4046 (u_longlong_t
)lr
->lrc_seq
,
4047 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
4048 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
4054 zil_replay_log_record(zilog_t
*zilog
, const lr_t
*lr
, void *zra
,
4057 zil_replay_arg_t
*zr
= zra
;
4058 const zil_header_t
*zh
= zilog
->zl_header
;
4059 uint64_t reclen
= lr
->lrc_reclen
;
4060 uint64_t txtype
= lr
->lrc_txtype
;
4063 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
4065 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
4068 if (lr
->lrc_txg
< claim_txg
) /* already committed */
4071 /* Strip case-insensitive bit, still present in log record */
4074 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
4075 return (zil_replay_error(zilog
, lr
, EINVAL
));
4078 * If this record type can be logged out of order, the object
4079 * (lr_foid) may no longer exist. That's legitimate, not an error.
4081 if (TX_OOO(txtype
)) {
4082 error
= dmu_object_info(zilog
->zl_os
,
4083 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
4084 if (error
== ENOENT
|| error
== EEXIST
)
4089 * Make a copy of the data so we can revise and extend it.
4091 memcpy(zr
->zr_lr
, lr
, reclen
);
4094 * If this is a TX_WRITE with a blkptr, suck in the data.
4096 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
4097 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
4098 zr
->zr_lr
+ reclen
);
4100 return (zil_replay_error(zilog
, lr
, error
));
4104 * The log block containing this lr may have been byteswapped
4105 * so that we can easily examine common fields like lrc_txtype.
4106 * However, the log is a mix of different record types, and only the
4107 * replay vectors know how to byteswap their records. Therefore, if
4108 * the lr was byteswapped, undo it before invoking the replay vector.
4110 if (zr
->zr_byteswap
)
4111 byteswap_uint64_array(zr
->zr_lr
, reclen
);
4114 * We must now do two things atomically: replay this log record,
4115 * and update the log header sequence number to reflect the fact that
4116 * we did so. At the end of each replay function the sequence number
4117 * is updated if we are in replay mode.
4119 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
4122 * The DMU's dnode layer doesn't see removes until the txg
4123 * commits, so a subsequent claim can spuriously fail with
4124 * EEXIST. So if we receive any error we try syncing out
4125 * any removes then retry the transaction. Note that we
4126 * specify B_FALSE for byteswap now, so we don't do it twice.
4128 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
4129 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
4131 return (zil_replay_error(zilog
, lr
, error
));
4137 zil_incr_blks(zilog_t
*zilog
, const blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
4139 (void) bp
, (void) arg
, (void) claim_txg
;
4141 zilog
->zl_replay_blks
++;
4147 * If this dataset has a non-empty intent log, replay it and destroy it.
4148 * Return B_TRUE if there were any entries to replay.
4151 zil_replay(objset_t
*os
, void *arg
,
4152 zil_replay_func_t
*const replay_func
[TX_MAX_TYPE
])
4154 zilog_t
*zilog
= dmu_objset_zil(os
);
4155 const zil_header_t
*zh
= zilog
->zl_header
;
4156 zil_replay_arg_t zr
;
4158 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
4159 return (zil_destroy(zilog
, B_TRUE
));
4162 zr
.zr_replay
= replay_func
;
4164 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
4165 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
4168 * Wait for in-progress removes to sync before starting replay.
4170 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
4172 zilog
->zl_replay
= B_TRUE
;
4173 zilog
->zl_replay_time
= ddi_get_lbolt();
4174 ASSERT(zilog
->zl_replay_blks
== 0);
4175 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
4176 zh
->zh_claim_txg
, B_TRUE
);
4177 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
4179 zil_destroy(zilog
, B_FALSE
);
4180 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
4181 zilog
->zl_replay
= B_FALSE
;
4187 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
4189 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
4192 if (zilog
->zl_replay
) {
4193 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
4194 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
4195 zilog
->zl_replaying_seq
;
4203 zil_reset(const char *osname
, void *arg
)
4207 int error
= zil_suspend(osname
, NULL
);
4208 /* EACCES means crypto key not loaded */
4209 if ((error
== EACCES
) || (error
== EBUSY
))
4210 return (SET_ERROR(error
));
4212 return (SET_ERROR(EEXIST
));
4216 EXPORT_SYMBOL(zil_alloc
);
4217 EXPORT_SYMBOL(zil_free
);
4218 EXPORT_SYMBOL(zil_open
);
4219 EXPORT_SYMBOL(zil_close
);
4220 EXPORT_SYMBOL(zil_replay
);
4221 EXPORT_SYMBOL(zil_replaying
);
4222 EXPORT_SYMBOL(zil_destroy
);
4223 EXPORT_SYMBOL(zil_destroy_sync
);
4224 EXPORT_SYMBOL(zil_itx_create
);
4225 EXPORT_SYMBOL(zil_itx_destroy
);
4226 EXPORT_SYMBOL(zil_itx_assign
);
4227 EXPORT_SYMBOL(zil_commit
);
4228 EXPORT_SYMBOL(zil_claim
);
4229 EXPORT_SYMBOL(zil_check_log_chain
);
4230 EXPORT_SYMBOL(zil_sync
);
4231 EXPORT_SYMBOL(zil_clean
);
4232 EXPORT_SYMBOL(zil_suspend
);
4233 EXPORT_SYMBOL(zil_resume
);
4234 EXPORT_SYMBOL(zil_lwb_add_block
);
4235 EXPORT_SYMBOL(zil_bp_tree_add
);
4236 EXPORT_SYMBOL(zil_set_sync
);
4237 EXPORT_SYMBOL(zil_set_logbias
);
4238 EXPORT_SYMBOL(zil_sums_init
);
4239 EXPORT_SYMBOL(zil_sums_fini
);
4240 EXPORT_SYMBOL(zil_kstat_values_update
);
4242 ZFS_MODULE_PARAM(zfs
, zfs_
, commit_timeout_pct
, UINT
, ZMOD_RW
,
4243 "ZIL block open timeout percentage");
4245 ZFS_MODULE_PARAM(zfs_zil
, zil_
, replay_disable
, INT
, ZMOD_RW
,
4246 "Disable intent logging replay");
4248 ZFS_MODULE_PARAM(zfs_zil
, zil_
, nocacheflush
, INT
, ZMOD_RW
,
4249 "Disable ZIL cache flushes");
4251 ZFS_MODULE_PARAM(zfs_zil
, zil_
, slog_bulk
, U64
, ZMOD_RW
,
4252 "Limit in bytes slog sync writes per commit");
4254 ZFS_MODULE_PARAM(zfs_zil
, zil_
, maxblocksize
, UINT
, ZMOD_RW
,
4255 "Limit in bytes of ZIL log block size");
4257 ZFS_MODULE_PARAM(zfs_zil
, zil_
, maxcopied
, UINT
, ZMOD_RW
,
4258 "Limit in bytes WR_COPIED size");