4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 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_zil.h>
48 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
49 * calls that change the file system. Each itx has enough information to
50 * be able to replay them after a system crash, power loss, or
51 * equivalent failure mode. These are stored in memory until either:
53 * 1. they are committed to the pool by the DMU transaction group
54 * (txg), at which point they can be discarded; or
55 * 2. they are committed to the on-disk ZIL for the dataset being
56 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59 * In the event of a crash or power loss, the itxs contained by each
60 * dataset's on-disk ZIL will be replayed when that dataset is first
61 * instantiated (e.g. if the dataset is a normal filesystem, when it is
64 * As hinted at above, there is one ZIL per dataset (both the in-memory
65 * representation, and the on-disk representation). The on-disk format
66 * consists of 3 parts:
68 * - a single, per-dataset, ZIL header; which points to a chain of
69 * - zero or more ZIL blocks; each of which contains
70 * - zero or more ZIL records
72 * A ZIL record holds the information necessary to replay a single
73 * system call transaction. A ZIL block can hold many ZIL records, and
74 * the blocks are chained together, similarly to a singly linked list.
76 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
77 * block in the chain, and the ZIL header points to the first block in
80 * Note, there is not a fixed place in the pool to hold these ZIL
81 * blocks; they are dynamically allocated and freed as needed from the
82 * blocks available on the pool, though they can be preferentially
83 * allocated from a dedicated "log" vdev.
87 * This controls the amount of time that a ZIL block (lwb) will remain
88 * "open" when it isn't "full", and it has a thread waiting for it to be
89 * committed to stable storage. Please refer to the zil_commit_waiter()
90 * function (and the comments within it) for more details.
92 int zfs_commit_timeout_pct
= 5;
95 * See zil.h for more information about these fields.
97 zil_stats_t zil_stats
= {
98 { "zil_commit_count", KSTAT_DATA_UINT64
},
99 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
100 { "zil_itx_count", KSTAT_DATA_UINT64
},
101 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
102 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
103 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
104 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
105 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
106 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
107 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
108 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
109 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
110 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
113 static kstat_t
*zil_ksp
;
116 * Disable intent logging replay. This global ZIL switch affects all pools.
118 int zil_replay_disable
= 0;
121 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
122 * the disk(s) by the ZIL after an LWB write has completed. Setting this
123 * will cause ZIL corruption on power loss if a volatile out-of-order
124 * write cache is enabled.
126 int zil_nocacheflush
= 0;
129 * Limit SLOG write size per commit executed with synchronous priority.
130 * Any writes above that will be executed with lower (asynchronous) priority
131 * to limit potential SLOG device abuse by single active ZIL writer.
133 unsigned long zil_slog_bulk
= 768 * 1024;
135 static kmem_cache_t
*zil_lwb_cache
;
136 static kmem_cache_t
*zil_zcw_cache
;
138 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
140 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
141 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
144 zil_bp_compare(const void *x1
, const void *x2
)
146 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
147 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
149 int cmp
= AVL_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
153 return (AVL_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
157 zil_bp_tree_init(zilog_t
*zilog
)
159 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
160 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
164 zil_bp_tree_fini(zilog_t
*zilog
)
166 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
170 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
171 kmem_free(zn
, sizeof (zil_bp_node_t
));
177 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
179 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
184 if (BP_IS_EMBEDDED(bp
))
187 dva
= BP_IDENTITY(bp
);
189 if (avl_find(t
, dva
, &where
) != NULL
)
190 return (SET_ERROR(EEXIST
));
192 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
194 avl_insert(t
, zn
, where
);
199 static zil_header_t
*
200 zil_header_in_syncing_context(zilog_t
*zilog
)
202 return ((zil_header_t
*)zilog
->zl_header
);
206 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
208 zio_cksum_t
*zc
= &bp
->blk_cksum
;
210 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
211 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
212 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
213 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
217 * Read a log block and make sure it's valid.
220 zil_read_log_block(zilog_t
*zilog
, boolean_t decrypt
, const blkptr_t
*bp
,
221 blkptr_t
*nbp
, void *dst
, char **end
)
223 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
224 arc_flags_t aflags
= ARC_FLAG_WAIT
;
225 arc_buf_t
*abuf
= NULL
;
229 if (zilog
->zl_header
->zh_claim_txg
== 0)
230 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
232 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
233 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
236 zio_flags
|= ZIO_FLAG_RAW
;
238 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
239 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
241 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
,
242 &abuf
, ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
245 zio_cksum_t cksum
= bp
->blk_cksum
;
248 * Validate the checksummed log block.
250 * Sequence numbers should be... sequential. The checksum
251 * verifier for the next block should be bp's checksum plus 1.
253 * Also check the log chain linkage and size used.
255 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
257 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
258 zil_chain_t
*zilc
= abuf
->b_data
;
259 char *lr
= (char *)(zilc
+ 1);
260 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
262 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
263 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
264 error
= SET_ERROR(ECKSUM
);
266 ASSERT3U(len
, <=, SPA_OLD_MAXBLOCKSIZE
);
268 *end
= (char *)dst
+ len
;
269 *nbp
= zilc
->zc_next_blk
;
272 char *lr
= abuf
->b_data
;
273 uint64_t size
= BP_GET_LSIZE(bp
);
274 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
276 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
277 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
278 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
279 error
= SET_ERROR(ECKSUM
);
281 ASSERT3U(zilc
->zc_nused
, <=,
282 SPA_OLD_MAXBLOCKSIZE
);
283 bcopy(lr
, dst
, zilc
->zc_nused
);
284 *end
= (char *)dst
+ zilc
->zc_nused
;
285 *nbp
= zilc
->zc_next_blk
;
289 arc_buf_destroy(abuf
, &abuf
);
296 * Read a TX_WRITE log data block.
299 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
301 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
302 const blkptr_t
*bp
= &lr
->lr_blkptr
;
303 arc_flags_t aflags
= ARC_FLAG_WAIT
;
304 arc_buf_t
*abuf
= NULL
;
308 if (BP_IS_HOLE(bp
)) {
310 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
314 if (zilog
->zl_header
->zh_claim_txg
== 0)
315 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
318 * If we are not using the resulting data, we are just checking that
319 * it hasn't been corrupted so we don't need to waste CPU time
320 * decompressing and decrypting it.
323 zio_flags
|= ZIO_FLAG_RAW
;
325 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
326 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
328 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
329 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
333 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
334 arc_buf_destroy(abuf
, &abuf
);
341 * Parse the intent log, and call parse_func for each valid record within.
344 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
345 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
,
348 const zil_header_t
*zh
= zilog
->zl_header
;
349 boolean_t claimed
= !!zh
->zh_claim_txg
;
350 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
351 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
352 uint64_t max_blk_seq
= 0;
353 uint64_t max_lr_seq
= 0;
354 uint64_t blk_count
= 0;
355 uint64_t lr_count
= 0;
356 blkptr_t blk
, next_blk
;
360 bzero(&next_blk
, sizeof (blkptr_t
));
363 * Old logs didn't record the maximum zh_claim_lr_seq.
365 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
366 claim_lr_seq
= UINT64_MAX
;
369 * Starting at the block pointed to by zh_log we read the log chain.
370 * For each block in the chain we strongly check that block to
371 * ensure its validity. We stop when an invalid block is found.
372 * For each block pointer in the chain we call parse_blk_func().
373 * For each record in each valid block we call parse_lr_func().
374 * If the log has been claimed, stop if we encounter a sequence
375 * number greater than the highest claimed sequence number.
377 lrbuf
= zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE
);
378 zil_bp_tree_init(zilog
);
380 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
381 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
385 if (blk_seq
> claim_blk_seq
)
388 error
= parse_blk_func(zilog
, &blk
, arg
, txg
);
391 ASSERT3U(max_blk_seq
, <, blk_seq
);
392 max_blk_seq
= blk_seq
;
395 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
398 error
= zil_read_log_block(zilog
, decrypt
, &blk
, &next_blk
,
403 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
404 lr_t
*lr
= (lr_t
*)lrp
;
405 reclen
= lr
->lrc_reclen
;
406 ASSERT3U(reclen
, >=, sizeof (lr_t
));
407 if (lr
->lrc_seq
> claim_lr_seq
)
410 error
= parse_lr_func(zilog
, lr
, arg
, txg
);
413 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
414 max_lr_seq
= lr
->lrc_seq
;
419 zilog
->zl_parse_error
= error
;
420 zilog
->zl_parse_blk_seq
= max_blk_seq
;
421 zilog
->zl_parse_lr_seq
= max_lr_seq
;
422 zilog
->zl_parse_blk_count
= blk_count
;
423 zilog
->zl_parse_lr_count
= lr_count
;
425 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
426 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
) ||
427 (decrypt
&& error
== EIO
));
429 zil_bp_tree_fini(zilog
);
430 zio_buf_free(lrbuf
, SPA_OLD_MAXBLOCKSIZE
);
437 zil_clear_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
439 ASSERT(!BP_IS_HOLE(bp
));
442 * As we call this function from the context of a rewind to a
443 * checkpoint, each ZIL block whose txg is later than the txg
444 * that we rewind to is invalid. Thus, we return -1 so
445 * zil_parse() doesn't attempt to read it.
447 if (bp
->blk_birth
>= first_txg
)
450 if (zil_bp_tree_add(zilog
, bp
) != 0)
453 zio_free(zilog
->zl_spa
, first_txg
, bp
);
459 zil_noop_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
465 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
468 * Claim log block if not already committed and not already claimed.
469 * If tx == NULL, just verify that the block is claimable.
471 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
472 zil_bp_tree_add(zilog
, bp
) != 0)
475 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
476 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
477 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
481 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
483 lr_write_t
*lr
= (lr_write_t
*)lrc
;
486 if (lrc
->lrc_txtype
!= TX_WRITE
)
490 * If the block is not readable, don't claim it. This can happen
491 * in normal operation when a log block is written to disk before
492 * some of the dmu_sync() blocks it points to. In this case, the
493 * transaction cannot have been committed to anyone (we would have
494 * waited for all writes to be stable first), so it is semantically
495 * correct to declare this the end of the log.
497 if (lr
->lr_blkptr
.blk_birth
>= first_txg
) {
498 error
= zil_read_log_data(zilog
, lr
, NULL
);
503 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
508 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
510 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
516 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
518 lr_write_t
*lr
= (lr_write_t
*)lrc
;
519 blkptr_t
*bp
= &lr
->lr_blkptr
;
522 * If we previously claimed it, we need to free it.
524 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
525 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
527 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
533 zil_lwb_vdev_compare(const void *x1
, const void *x2
)
535 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
536 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
538 return (AVL_CMP(v1
, v2
));
542 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, boolean_t slog
, uint64_t txg
,
547 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
548 lwb
->lwb_zilog
= zilog
;
550 lwb
->lwb_fastwrite
= fastwrite
;
551 lwb
->lwb_slog
= slog
;
552 lwb
->lwb_state
= LWB_STATE_CLOSED
;
553 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
554 lwb
->lwb_max_txg
= txg
;
555 lwb
->lwb_write_zio
= NULL
;
556 lwb
->lwb_root_zio
= NULL
;
558 lwb
->lwb_issued_timestamp
= 0;
559 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
560 lwb
->lwb_nused
= sizeof (zil_chain_t
);
561 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
564 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
567 mutex_enter(&zilog
->zl_lock
);
568 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
569 mutex_exit(&zilog
->zl_lock
);
571 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
572 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
573 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
574 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
580 zil_free_lwb(zilog_t
*zilog
, lwb_t
*lwb
)
582 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
583 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
584 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
585 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
586 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
587 ASSERT3P(lwb
->lwb_write_zio
, ==, NULL
);
588 ASSERT3P(lwb
->lwb_root_zio
, ==, NULL
);
589 ASSERT3U(lwb
->lwb_max_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
590 ASSERT(lwb
->lwb_state
== LWB_STATE_CLOSED
||
591 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
594 * Clear the zilog's field to indicate this lwb is no longer
595 * valid, and prevent use-after-free errors.
597 if (zilog
->zl_last_lwb_opened
== lwb
)
598 zilog
->zl_last_lwb_opened
= NULL
;
600 kmem_cache_free(zil_lwb_cache
, lwb
);
604 * Called when we create in-memory log transactions so that we know
605 * to cleanup the itxs at the end of spa_sync().
608 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
610 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
611 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
613 ASSERT(spa_writeable(zilog
->zl_spa
));
615 if (ds
->ds_is_snapshot
)
616 panic("dirtying snapshot!");
618 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
619 /* up the hold count until we can be written out */
620 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
622 zilog
->zl_dirty_max_txg
= MAX(txg
, zilog
->zl_dirty_max_txg
);
627 * Determine if the zil is dirty in the specified txg. Callers wanting to
628 * ensure that the dirty state does not change must hold the itxg_lock for
629 * the specified txg. Holding the lock will ensure that the zil cannot be
630 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
634 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
636 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
638 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
644 * Determine if the zil is dirty. The zil is considered dirty if it has
645 * any pending itx records that have not been cleaned by zil_clean().
648 zilog_is_dirty(zilog_t
*zilog
)
650 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
652 for (int t
= 0; t
< TXG_SIZE
; t
++) {
653 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
660 * Create an on-disk intent log.
663 zil_create(zilog_t
*zilog
)
665 const zil_header_t
*zh
= zilog
->zl_header
;
671 boolean_t fastwrite
= FALSE
;
672 boolean_t slog
= FALSE
;
675 * Wait for any previous destroy to complete.
677 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
679 ASSERT(zh
->zh_claim_txg
== 0);
680 ASSERT(zh
->zh_replay_seq
== 0);
685 * Allocate an initial log block if:
686 * - there isn't one already
687 * - the existing block is the wrong endianness
689 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
690 tx
= dmu_tx_create(zilog
->zl_os
);
691 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
692 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
693 txg
= dmu_tx_get_txg(tx
);
695 if (!BP_IS_HOLE(&blk
)) {
696 zio_free(zilog
->zl_spa
, txg
, &blk
);
700 error
= zio_alloc_zil(zilog
->zl_spa
, zilog
->zl_os
, txg
, &blk
,
701 ZIL_MIN_BLKSZ
, &slog
);
705 zil_init_log_chain(zilog
, &blk
);
709 * Allocate a log write block (lwb) for the first log block.
712 lwb
= zil_alloc_lwb(zilog
, &blk
, slog
, txg
, fastwrite
);
715 * If we just allocated the first log block, commit our transaction
716 * and wait for zil_sync() to stuff the block pointer into zh_log.
717 * (zh is part of the MOS, so we cannot modify it in open context.)
721 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
724 ASSERT(error
!= 0 || bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
725 IMPLY(error
== 0, lwb
!= NULL
);
731 * In one tx, free all log blocks and clear the log header. If keep_first
732 * is set, then we're replaying a log with no content. We want to keep the
733 * first block, however, so that the first synchronous transaction doesn't
734 * require a txg_wait_synced() in zil_create(). We don't need to
735 * txg_wait_synced() here either when keep_first is set, because both
736 * zil_create() and zil_destroy() will wait for any in-progress destroys
740 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
742 const zil_header_t
*zh
= zilog
->zl_header
;
748 * Wait for any previous destroy to complete.
750 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
752 zilog
->zl_old_header
= *zh
; /* debugging aid */
754 if (BP_IS_HOLE(&zh
->zh_log
))
757 tx
= dmu_tx_create(zilog
->zl_os
);
758 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
759 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
760 txg
= dmu_tx_get_txg(tx
);
762 mutex_enter(&zilog
->zl_lock
);
764 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
765 zilog
->zl_destroy_txg
= txg
;
766 zilog
->zl_keep_first
= keep_first
;
768 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
769 ASSERT(zh
->zh_claim_txg
== 0);
771 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
772 if (lwb
->lwb_fastwrite
)
773 metaslab_fastwrite_unmark(zilog
->zl_spa
,
776 list_remove(&zilog
->zl_lwb_list
, lwb
);
777 if (lwb
->lwb_buf
!= NULL
)
778 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
779 zio_free(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
780 zil_free_lwb(zilog
, lwb
);
782 } else if (!keep_first
) {
783 zil_destroy_sync(zilog
, tx
);
785 mutex_exit(&zilog
->zl_lock
);
791 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
793 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
794 (void) zil_parse(zilog
, zil_free_log_block
,
795 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
, B_FALSE
);
799 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
801 dmu_tx_t
*tx
= txarg
;
808 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
809 DMU_OST_ANY
, B_FALSE
, B_FALSE
, FTAG
, &os
);
812 * EBUSY indicates that the objset is inconsistent, in which
813 * case it can not have a ZIL.
815 if (error
!= EBUSY
) {
816 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
817 (unsigned long long)ds
->ds_object
, error
);
823 zilog
= dmu_objset_zil(os
);
824 zh
= zil_header_in_syncing_context(zilog
);
825 ASSERT3U(tx
->tx_txg
, ==, spa_first_txg(zilog
->zl_spa
));
826 first_txg
= spa_min_claim_txg(zilog
->zl_spa
);
829 * If the spa_log_state is not set to be cleared, check whether
830 * the current uberblock is a checkpoint one and if the current
831 * header has been claimed before moving on.
833 * If the current uberblock is a checkpointed uberblock then
834 * one of the following scenarios took place:
836 * 1] We are currently rewinding to the checkpoint of the pool.
837 * 2] We crashed in the middle of a checkpoint rewind but we
838 * did manage to write the checkpointed uberblock to the
839 * vdev labels, so when we tried to import the pool again
840 * the checkpointed uberblock was selected from the import
843 * In both cases we want to zero out all the ZIL blocks, except
844 * the ones that have been claimed at the time of the checkpoint
845 * (their zh_claim_txg != 0). The reason is that these blocks
846 * may be corrupted since we may have reused their locations on
847 * disk after we took the checkpoint.
849 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
850 * when we first figure out whether the current uberblock is
851 * checkpointed or not. Unfortunately, that would discard all
852 * the logs, including the ones that are claimed, and we would
855 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
||
856 (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
857 zh
->zh_claim_txg
== 0)) {
858 if (!BP_IS_HOLE(&zh
->zh_log
)) {
859 (void) zil_parse(zilog
, zil_clear_log_block
,
860 zil_noop_log_record
, tx
, first_txg
, B_FALSE
);
862 BP_ZERO(&zh
->zh_log
);
863 if (os
->os_encrypted
)
864 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
865 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
866 dmu_objset_disown(os
, B_FALSE
, FTAG
);
871 * If we are not rewinding and opening the pool normally, then
872 * the min_claim_txg should be equal to the first txg of the pool.
874 ASSERT3U(first_txg
, ==, spa_first_txg(zilog
->zl_spa
));
877 * Claim all log blocks if we haven't already done so, and remember
878 * the highest claimed sequence number. This ensures that if we can
879 * read only part of the log now (e.g. due to a missing device),
880 * but we can read the entire log later, we will not try to replay
881 * or destroy beyond the last block we successfully claimed.
883 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
884 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
885 (void) zil_parse(zilog
, zil_claim_log_block
,
886 zil_claim_log_record
, tx
, first_txg
, B_FALSE
);
887 zh
->zh_claim_txg
= first_txg
;
888 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
889 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
890 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
891 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
892 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
893 if (os
->os_encrypted
)
894 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
895 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
898 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
899 dmu_objset_disown(os
, B_FALSE
, FTAG
);
904 * Check the log by walking the log chain.
905 * Checksum errors are ok as they indicate the end of the chain.
906 * Any other error (no device or read failure) returns an error.
910 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
919 error
= dmu_objset_from_ds(ds
, &os
);
921 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
922 (unsigned long long)ds
->ds_object
, error
);
926 zilog
= dmu_objset_zil(os
);
927 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
929 if (!BP_IS_HOLE(bp
)) {
931 boolean_t valid
= B_TRUE
;
934 * Check the first block and determine if it's on a log device
935 * which may have been removed or faulted prior to loading this
936 * pool. If so, there's no point in checking the rest of the
937 * log as its content should have already been synced to the
940 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
941 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
942 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
943 valid
= vdev_log_state_valid(vd
);
944 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
950 * Check whether the current uberblock is checkpointed (e.g.
951 * we are rewinding) and whether the current header has been
952 * claimed or not. If it hasn't then skip verifying it. We
953 * do this because its ZIL blocks may be part of the pool's
954 * state before the rewind, which is no longer valid.
956 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
957 if (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
958 zh
->zh_claim_txg
== 0)
963 * Because tx == NULL, zil_claim_log_block() will not actually claim
964 * any blocks, but just determine whether it is possible to do so.
965 * In addition to checking the log chain, zil_claim_log_block()
966 * will invoke zio_claim() with a done func of spa_claim_notify(),
967 * which will update spa_max_claim_txg. See spa_load() for details.
969 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
970 zilog
->zl_header
->zh_claim_txg
? -1ULL :
971 spa_min_claim_txg(os
->os_spa
), B_FALSE
);
973 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
977 * When an itx is "skipped", this function is used to properly mark the
978 * waiter as "done, and signal any thread(s) waiting on it. An itx can
979 * be skipped (and not committed to an lwb) for a variety of reasons,
980 * one of them being that the itx was committed via spa_sync(), prior to
981 * it being committed to an lwb; this can happen if a thread calling
982 * zil_commit() is racing with spa_sync().
985 zil_commit_waiter_skip(zil_commit_waiter_t
*zcw
)
987 mutex_enter(&zcw
->zcw_lock
);
988 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
989 zcw
->zcw_done
= B_TRUE
;
990 cv_broadcast(&zcw
->zcw_cv
);
991 mutex_exit(&zcw
->zcw_lock
);
995 * This function is used when the given waiter is to be linked into an
996 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
997 * At this point, the waiter will no longer be referenced by the itx,
998 * and instead, will be referenced by the lwb.
1001 zil_commit_waiter_link_lwb(zil_commit_waiter_t
*zcw
, lwb_t
*lwb
)
1004 * The lwb_waiters field of the lwb is protected by the zilog's
1005 * zl_lock, thus it must be held when calling this function.
1007 ASSERT(MUTEX_HELD(&lwb
->lwb_zilog
->zl_lock
));
1009 mutex_enter(&zcw
->zcw_lock
);
1010 ASSERT(!list_link_active(&zcw
->zcw_node
));
1011 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1012 ASSERT3P(lwb
, !=, NULL
);
1013 ASSERT(lwb
->lwb_state
== LWB_STATE_OPENED
||
1014 lwb
->lwb_state
== LWB_STATE_ISSUED
||
1015 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
);
1017 list_insert_tail(&lwb
->lwb_waiters
, zcw
);
1019 mutex_exit(&zcw
->zcw_lock
);
1023 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1024 * block, and the given waiter must be linked to the "nolwb waiters"
1025 * list inside of zil_process_commit_list().
1028 zil_commit_waiter_link_nolwb(zil_commit_waiter_t
*zcw
, list_t
*nolwb
)
1030 mutex_enter(&zcw
->zcw_lock
);
1031 ASSERT(!list_link_active(&zcw
->zcw_node
));
1032 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1033 list_insert_tail(nolwb
, zcw
);
1034 mutex_exit(&zcw
->zcw_lock
);
1038 zil_lwb_add_block(lwb_t
*lwb
, const blkptr_t
*bp
)
1040 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1042 zil_vdev_node_t
*zv
, zvsearch
;
1043 int ndvas
= BP_GET_NDVAS(bp
);
1046 if (zil_nocacheflush
)
1049 mutex_enter(&lwb
->lwb_vdev_lock
);
1050 for (i
= 0; i
< ndvas
; i
++) {
1051 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1052 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
1053 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
1054 zv
->zv_vdev
= zvsearch
.zv_vdev
;
1055 avl_insert(t
, zv
, where
);
1058 mutex_exit(&lwb
->lwb_vdev_lock
);
1062 zil_lwb_flush_defer(lwb_t
*lwb
, lwb_t
*nlwb
)
1064 avl_tree_t
*src
= &lwb
->lwb_vdev_tree
;
1065 avl_tree_t
*dst
= &nlwb
->lwb_vdev_tree
;
1066 void *cookie
= NULL
;
1067 zil_vdev_node_t
*zv
;
1069 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1070 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
1071 ASSERT3S(nlwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
1074 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1075 * not need the protection of lwb_vdev_lock (it will only be modified
1076 * while holding zilog->zl_lock) as its writes and those of its
1077 * children have all completed. The younger 'nlwb' may be waiting on
1078 * future writes to additional vdevs.
1080 mutex_enter(&nlwb
->lwb_vdev_lock
);
1082 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1083 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1085 while ((zv
= avl_destroy_nodes(src
, &cookie
)) != NULL
) {
1088 if (avl_find(dst
, zv
, &where
) == NULL
) {
1089 avl_insert(dst
, zv
, where
);
1091 kmem_free(zv
, sizeof (*zv
));
1094 mutex_exit(&nlwb
->lwb_vdev_lock
);
1098 zil_lwb_add_txg(lwb_t
*lwb
, uint64_t txg
)
1100 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1104 * This function is a called after all vdevs associated with a given lwb
1105 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1106 * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1107 * all "previous" lwb's will have completed before this function is
1108 * called; i.e. this function is called for all previous lwbs before
1109 * it's called for "this" lwb (enforced via zio the dependencies
1110 * configured in zil_lwb_set_zio_dependency()).
1112 * The intention is for this function to be called as soon as the
1113 * contents of an lwb are considered "stable" on disk, and will survive
1114 * any sudden loss of power. At this point, any threads waiting for the
1115 * lwb to reach this state are signalled, and the "waiter" structures
1116 * are marked "done".
1119 zil_lwb_flush_vdevs_done(zio_t
*zio
)
1121 lwb_t
*lwb
= zio
->io_private
;
1122 zilog_t
*zilog
= lwb
->lwb_zilog
;
1123 dmu_tx_t
*tx
= lwb
->lwb_tx
;
1124 zil_commit_waiter_t
*zcw
;
1127 spa_config_exit(zilog
->zl_spa
, SCL_STATE
, lwb
);
1129 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1131 mutex_enter(&zilog
->zl_lock
);
1134 * Ensure the lwb buffer pointer is cleared before releasing the
1135 * txg. If we have had an allocation failure and the txg is
1136 * waiting to sync then we want zil_sync() to remove the lwb so
1137 * that it's not picked up as the next new one in
1138 * zil_process_commit_list(). zil_sync() will only remove the
1139 * lwb if lwb_buf is null.
1141 lwb
->lwb_buf
= NULL
;
1144 ASSERT3U(lwb
->lwb_issued_timestamp
, >, 0);
1145 zilog
->zl_last_lwb_latency
= gethrtime() - lwb
->lwb_issued_timestamp
;
1147 lwb
->lwb_root_zio
= NULL
;
1149 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_WRITE_DONE
);
1150 lwb
->lwb_state
= LWB_STATE_FLUSH_DONE
;
1152 if (zilog
->zl_last_lwb_opened
== lwb
) {
1154 * Remember the highest committed log sequence number
1155 * for ztest. We only update this value when all the log
1156 * writes succeeded, because ztest wants to ASSERT that
1157 * it got the whole log chain.
1159 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1162 while ((itx
= list_head(&lwb
->lwb_itxs
)) != NULL
) {
1163 list_remove(&lwb
->lwb_itxs
, itx
);
1164 zil_itx_destroy(itx
);
1167 while ((zcw
= list_head(&lwb
->lwb_waiters
)) != NULL
) {
1168 mutex_enter(&zcw
->zcw_lock
);
1170 ASSERT(list_link_active(&zcw
->zcw_node
));
1171 list_remove(&lwb
->lwb_waiters
, zcw
);
1173 ASSERT3P(zcw
->zcw_lwb
, ==, lwb
);
1174 zcw
->zcw_lwb
= NULL
;
1176 zcw
->zcw_zio_error
= zio
->io_error
;
1178 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1179 zcw
->zcw_done
= B_TRUE
;
1180 cv_broadcast(&zcw
->zcw_cv
);
1182 mutex_exit(&zcw
->zcw_lock
);
1185 mutex_exit(&zilog
->zl_lock
);
1188 * Now that we've written this log block, we have a stable pointer
1189 * to the next block in the chain, so it's OK to let the txg in
1190 * which we allocated the next block sync.
1196 * This is called when an lwb's write zio completes. The callback's
1197 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1198 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1199 * in writing out this specific lwb's data, and in the case that cache
1200 * flushes have been deferred, vdevs involved in writing the data for
1201 * previous lwbs. The writes corresponding to all the vdevs in the
1202 * lwb_vdev_tree will have completed by the time this is called, due to
1203 * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1204 * which takes deferred flushes into account. The lwb will be "done"
1205 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1206 * completion callback for the lwb's root zio.
1209 zil_lwb_write_done(zio_t
*zio
)
1211 lwb_t
*lwb
= zio
->io_private
;
1212 spa_t
*spa
= zio
->io_spa
;
1213 zilog_t
*zilog
= lwb
->lwb_zilog
;
1214 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1215 void *cookie
= NULL
;
1216 zil_vdev_node_t
*zv
;
1219 ASSERT3S(spa_config_held(spa
, SCL_STATE
, RW_READER
), !=, 0);
1221 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1222 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
1223 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
1224 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
1225 ASSERT(!BP_IS_GANG(zio
->io_bp
));
1226 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
1227 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
1229 abd_put(zio
->io_abd
);
1231 mutex_enter(&zilog
->zl_lock
);
1232 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_ISSUED
);
1233 lwb
->lwb_state
= LWB_STATE_WRITE_DONE
;
1234 lwb
->lwb_write_zio
= NULL
;
1235 lwb
->lwb_fastwrite
= FALSE
;
1236 nlwb
= list_next(&zilog
->zl_lwb_list
, lwb
);
1237 mutex_exit(&zilog
->zl_lock
);
1239 if (avl_numnodes(t
) == 0)
1243 * If there was an IO error, we're not going to call zio_flush()
1244 * on these vdevs, so we simply empty the tree and free the
1245 * nodes. We avoid calling zio_flush() since there isn't any
1246 * good reason for doing so, after the lwb block failed to be
1249 if (zio
->io_error
!= 0) {
1250 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
1251 kmem_free(zv
, sizeof (*zv
));
1256 * If this lwb does not have any threads waiting for it to
1257 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1258 * command to the vdevs written to by "this" lwb, and instead
1259 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1260 * command for those vdevs. Thus, we merge the vdev tree of
1261 * "this" lwb with the vdev tree of the "next" lwb in the list,
1262 * and assume the "next" lwb will handle flushing the vdevs (or
1263 * deferring the flush(s) again).
1265 * This is a useful performance optimization, especially for
1266 * workloads with lots of async write activity and few sync
1267 * write and/or fsync activity, as it has the potential to
1268 * coalesce multiple flush commands to a vdev into one.
1270 if (list_head(&lwb
->lwb_waiters
) == NULL
&& nlwb
!= NULL
) {
1271 zil_lwb_flush_defer(lwb
, nlwb
);
1272 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
1276 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1277 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
1279 zio_flush(lwb
->lwb_root_zio
, vd
);
1280 kmem_free(zv
, sizeof (*zv
));
1285 zil_lwb_set_zio_dependency(zilog_t
*zilog
, lwb_t
*lwb
)
1287 lwb_t
*last_lwb_opened
= zilog
->zl_last_lwb_opened
;
1289 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1290 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
1293 * The zilog's "zl_last_lwb_opened" field is used to build the
1294 * lwb/zio dependency chain, which is used to preserve the
1295 * ordering of lwb completions that is required by the semantics
1296 * of the ZIL. Each new lwb zio becomes a parent of the
1297 * "previous" lwb zio, such that the new lwb's zio cannot
1298 * complete until the "previous" lwb's zio completes.
1300 * This is required by the semantics of zil_commit(); the commit
1301 * waiters attached to the lwbs will be woken in the lwb zio's
1302 * completion callback, so this zio dependency graph ensures the
1303 * waiters are woken in the correct order (the same order the
1304 * lwbs were created).
1306 if (last_lwb_opened
!= NULL
&&
1307 last_lwb_opened
->lwb_state
!= LWB_STATE_FLUSH_DONE
) {
1308 ASSERT(last_lwb_opened
->lwb_state
== LWB_STATE_OPENED
||
1309 last_lwb_opened
->lwb_state
== LWB_STATE_ISSUED
||
1310 last_lwb_opened
->lwb_state
== LWB_STATE_WRITE_DONE
);
1312 ASSERT3P(last_lwb_opened
->lwb_root_zio
, !=, NULL
);
1313 zio_add_child(lwb
->lwb_root_zio
,
1314 last_lwb_opened
->lwb_root_zio
);
1317 * If the previous lwb's write hasn't already completed,
1318 * we also want to order the completion of the lwb write
1319 * zios (above, we only order the completion of the lwb
1320 * root zios). This is required because of how we can
1321 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1323 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1324 * the previous lwb will rely on this lwb to flush the
1325 * vdevs written to by that previous lwb. Thus, we need
1326 * to ensure this lwb doesn't issue the flush until
1327 * after the previous lwb's write completes. We ensure
1328 * this ordering by setting the zio parent/child
1329 * relationship here.
1331 * Without this relationship on the lwb's write zio,
1332 * it's possible for this lwb's write to complete prior
1333 * to the previous lwb's write completing; and thus, the
1334 * vdevs for the previous lwb would be flushed prior to
1335 * that lwb's data being written to those vdevs (the
1336 * vdevs are flushed in the lwb write zio's completion
1337 * handler, zil_lwb_write_done()).
1339 if (last_lwb_opened
->lwb_state
!= LWB_STATE_WRITE_DONE
) {
1340 ASSERT(last_lwb_opened
->lwb_state
== LWB_STATE_OPENED
||
1341 last_lwb_opened
->lwb_state
== LWB_STATE_ISSUED
);
1343 ASSERT3P(last_lwb_opened
->lwb_write_zio
, !=, NULL
);
1344 zio_add_child(lwb
->lwb_write_zio
,
1345 last_lwb_opened
->lwb_write_zio
);
1352 * This function's purpose is to "open" an lwb such that it is ready to
1353 * accept new itxs being committed to it. To do this, the lwb's zio
1354 * structures are created, and linked to the lwb. This function is
1355 * idempotent; if the passed in lwb has already been opened, this
1356 * function is essentially a no-op.
1359 zil_lwb_write_open(zilog_t
*zilog
, lwb_t
*lwb
)
1361 zbookmark_phys_t zb
;
1362 zio_priority_t prio
;
1364 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1365 ASSERT3P(lwb
, !=, NULL
);
1366 EQUIV(lwb
->lwb_root_zio
== NULL
, lwb
->lwb_state
== LWB_STATE_CLOSED
);
1367 EQUIV(lwb
->lwb_root_zio
!= NULL
, lwb
->lwb_state
== LWB_STATE_OPENED
);
1369 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1370 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
1371 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1373 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1374 mutex_enter(&zilog
->zl_lock
);
1375 if (lwb
->lwb_root_zio
== NULL
) {
1376 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
,
1377 BP_GET_LSIZE(&lwb
->lwb_blk
));
1379 if (!lwb
->lwb_fastwrite
) {
1380 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1381 lwb
->lwb_fastwrite
= 1;
1384 if (!lwb
->lwb_slog
|| zilog
->zl_cur_used
<= zil_slog_bulk
)
1385 prio
= ZIO_PRIORITY_SYNC_WRITE
;
1387 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
1389 lwb
->lwb_root_zio
= zio_root(zilog
->zl_spa
,
1390 zil_lwb_flush_vdevs_done
, lwb
, ZIO_FLAG_CANFAIL
);
1391 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1393 lwb
->lwb_write_zio
= zio_rewrite(lwb
->lwb_root_zio
,
1394 zilog
->zl_spa
, 0, &lwb
->lwb_blk
, lwb_abd
,
1395 BP_GET_LSIZE(&lwb
->lwb_blk
), zil_lwb_write_done
, lwb
,
1396 prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
1397 ZIO_FLAG_FASTWRITE
, &zb
);
1398 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1400 lwb
->lwb_state
= LWB_STATE_OPENED
;
1402 zil_lwb_set_zio_dependency(zilog
, lwb
);
1403 zilog
->zl_last_lwb_opened
= lwb
;
1405 mutex_exit(&zilog
->zl_lock
);
1407 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1408 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1409 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1413 * Define a limited set of intent log block sizes.
1415 * These must be a multiple of 4KB. Note only the amount used (again
1416 * aligned to 4KB) actually gets written. However, we can't always just
1417 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1419 uint64_t zil_block_buckets
[] = {
1420 4096, /* non TX_WRITE */
1421 8192+4096, /* data base */
1422 32*1024 + 4096, /* NFS writes */
1427 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1428 * initialized. Otherwise this should not be used directly; see
1429 * zl_max_block_size instead.
1431 int zil_maxblocksize
= SPA_OLD_MAXBLOCKSIZE
;
1434 * Start a log block write and advance to the next log block.
1435 * Calls are serialized.
1438 zil_lwb_write_issue(zilog_t
*zilog
, lwb_t
*lwb
)
1442 spa_t
*spa
= zilog
->zl_spa
;
1446 uint64_t zil_blksz
, wsz
;
1450 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1451 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1452 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1453 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1455 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1456 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
1457 bp
= &zilc
->zc_next_blk
;
1459 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
1460 bp
= &zilc
->zc_next_blk
;
1463 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
1466 * Allocate the next block and save its address in this block
1467 * before writing it in order to establish the log chain.
1468 * Note that if the allocation of nlwb synced before we wrote
1469 * the block that points at it (lwb), we'd leak it if we crashed.
1470 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1471 * We dirty the dataset to ensure that zil_sync() will be called
1472 * to clean up in the event of allocation failure or I/O failure.
1475 tx
= dmu_tx_create(zilog
->zl_os
);
1478 * Since we are not going to create any new dirty data, and we
1479 * can even help with clearing the existing dirty data, we
1480 * should not be subject to the dirty data based delays. We
1481 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1483 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
1485 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1486 txg
= dmu_tx_get_txg(tx
);
1491 * Log blocks are pre-allocated. Here we select the size of the next
1492 * block, based on size used in the last block.
1493 * - first find the smallest bucket that will fit the block from a
1494 * limited set of block sizes. This is because it's faster to write
1495 * blocks allocated from the same metaslab as they are adjacent or
1497 * - next find the maximum from the new suggested size and an array of
1498 * previous sizes. This lessens a picket fence effect of wrongly
1499 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1502 * Note we only write what is used, but we can't just allocate
1503 * the maximum block size because we can exhaust the available
1506 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1507 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1509 zil_blksz
= MIN(zil_block_buckets
[i
], zilog
->zl_max_block_size
);
1510 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1511 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1512 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1513 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1516 error
= zio_alloc_zil(spa
, zilog
->zl_os
, txg
, bp
, zil_blksz
, &slog
);
1518 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1519 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1521 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1522 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1525 ASSERT3U(bp
->blk_birth
, ==, txg
);
1526 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1527 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1530 * Allocate a new log write block (lwb).
1532 nlwb
= zil_alloc_lwb(zilog
, bp
, slog
, txg
, TRUE
);
1535 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1536 /* For Slim ZIL only write what is used. */
1537 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1538 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1539 zio_shrink(lwb
->lwb_write_zio
, wsz
);
1546 zilc
->zc_nused
= lwb
->lwb_nused
;
1547 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1550 * clear unused data for security
1552 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1554 spa_config_enter(zilog
->zl_spa
, SCL_STATE
, lwb
, RW_READER
);
1556 zil_lwb_add_block(lwb
, &lwb
->lwb_blk
);
1557 lwb
->lwb_issued_timestamp
= gethrtime();
1558 lwb
->lwb_state
= LWB_STATE_ISSUED
;
1560 zio_nowait(lwb
->lwb_root_zio
);
1561 zio_nowait(lwb
->lwb_write_zio
);
1564 * If there was an allocation failure then nlwb will be null which
1565 * forces a txg_wait_synced().
1571 * Maximum amount of write data that can be put into single log block.
1574 zil_max_log_data(zilog_t
*zilog
)
1576 return (zilog
->zl_max_block_size
-
1577 sizeof (zil_chain_t
) - sizeof (lr_write_t
));
1581 * Maximum amount of log space we agree to waste to reduce number of
1582 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1584 static inline uint64_t
1585 zil_max_waste_space(zilog_t
*zilog
)
1587 return (zil_max_log_data(zilog
) / 8);
1591 * Maximum amount of write data for WR_COPIED. For correctness, consumers
1592 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1593 * maximum sized log block, because each WR_COPIED record must fit in a
1594 * single log block. For space efficiency, we want to fit two records into a
1595 * max-sized log block.
1598 zil_max_copied_data(zilog_t
*zilog
)
1600 return ((zilog
->zl_max_block_size
- sizeof (zil_chain_t
)) / 2 -
1601 sizeof (lr_write_t
));
1605 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1608 lr_write_t
*lrwb
, *lrw
;
1610 uint64_t dlen
, dnow
, lwb_sp
, reclen
, txg
, max_log_data
;
1612 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1613 ASSERT3P(lwb
, !=, NULL
);
1614 ASSERT3P(lwb
->lwb_buf
, !=, NULL
);
1616 zil_lwb_write_open(zilog
, lwb
);
1619 lrw
= (lr_write_t
*)lrc
;
1622 * A commit itx doesn't represent any on-disk state; instead
1623 * it's simply used as a place holder on the commit list, and
1624 * provides a mechanism for attaching a "commit waiter" onto the
1625 * correct lwb (such that the waiter can be signalled upon
1626 * completion of that lwb). Thus, we don't process this itx's
1627 * log record if it's a commit itx (these itx's don't have log
1628 * records), and instead link the itx's waiter onto the lwb's
1631 * For more details, see the comment above zil_commit().
1633 if (lrc
->lrc_txtype
== TX_COMMIT
) {
1634 mutex_enter(&zilog
->zl_lock
);
1635 zil_commit_waiter_link_lwb(itx
->itx_private
, lwb
);
1636 itx
->itx_private
= NULL
;
1637 mutex_exit(&zilog
->zl_lock
);
1641 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
1642 dlen
= P2ROUNDUP_TYPED(
1643 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1647 reclen
= lrc
->lrc_reclen
;
1648 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1651 ASSERT3U(zilog
->zl_cur_used
, <, UINT64_MAX
- (reclen
+ dlen
));
1655 * If this record won't fit in the current log block, start a new one.
1656 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1658 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1659 max_log_data
= zil_max_log_data(zilog
);
1660 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
1661 lwb_sp
< zil_max_waste_space(zilog
) &&
1662 (dlen
% max_log_data
== 0 ||
1663 lwb_sp
< reclen
+ dlen
% max_log_data
))) {
1664 lwb
= zil_lwb_write_issue(zilog
, lwb
);
1667 zil_lwb_write_open(zilog
, lwb
);
1668 ASSERT(LWB_EMPTY(lwb
));
1669 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1672 * There must be enough space in the new, empty log block to
1673 * hold reclen. For WR_COPIED, we need to fit the whole
1674 * record in one block, and reclen is the header size + the
1675 * data size. For WR_NEED_COPY, we can create multiple
1676 * records, splitting the data into multiple blocks, so we
1677 * only need to fit one word of data per block; in this case
1678 * reclen is just the header size (no data).
1680 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
1683 dnow
= MIN(dlen
, lwb_sp
- reclen
);
1684 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1685 bcopy(lrc
, lr_buf
, reclen
);
1686 lrcb
= (lr_t
*)lr_buf
; /* Like lrc, but inside lwb. */
1687 lrwb
= (lr_write_t
*)lrcb
; /* Like lrw, but inside lwb. */
1689 ZIL_STAT_BUMP(zil_itx_count
);
1692 * If it's a write, fetch the data or get its blkptr as appropriate.
1694 if (lrc
->lrc_txtype
== TX_WRITE
) {
1695 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1696 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1697 if (itx
->itx_wr_state
== WR_COPIED
) {
1698 ZIL_STAT_BUMP(zil_itx_copied_count
);
1699 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1704 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
1705 dbuf
= lr_buf
+ reclen
;
1706 lrcb
->lrc_reclen
+= dnow
;
1707 if (lrwb
->lr_length
> dnow
)
1708 lrwb
->lr_length
= dnow
;
1709 lrw
->lr_offset
+= dnow
;
1710 lrw
->lr_length
-= dnow
;
1711 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1712 ZIL_STAT_INCR(zil_itx_needcopy_bytes
, dnow
);
1714 ASSERT3S(itx
->itx_wr_state
, ==, WR_INDIRECT
);
1716 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1717 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1722 * We pass in the "lwb_write_zio" rather than
1723 * "lwb_root_zio" so that the "lwb_write_zio"
1724 * becomes the parent of any zio's created by
1725 * the "zl_get_data" callback. The vdevs are
1726 * flushed after the "lwb_write_zio" completes,
1727 * so we want to make sure that completion
1728 * callback waits for these additional zio's,
1729 * such that the vdevs used by those zio's will
1730 * be included in the lwb's vdev tree, and those
1731 * vdevs will be properly flushed. If we passed
1732 * in "lwb_root_zio" here, then these additional
1733 * vdevs may not be flushed; e.g. if these zio's
1734 * completed after "lwb_write_zio" completed.
1736 error
= zilog
->zl_get_data(itx
->itx_private
,
1737 lrwb
, dbuf
, lwb
, lwb
->lwb_write_zio
);
1740 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1744 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1752 * We're actually making an entry, so update lrc_seq to be the
1753 * log record sequence number. Note that this is generally not
1754 * equal to the itx sequence number because not all transactions
1755 * are synchronous, and sometimes spa_sync() gets there first.
1757 lrcb
->lrc_seq
= ++zilog
->zl_lr_seq
;
1758 lwb
->lwb_nused
+= reclen
+ dnow
;
1760 zil_lwb_add_txg(lwb
, txg
);
1762 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1763 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1767 zilog
->zl_cur_used
+= reclen
;
1775 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1780 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1781 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
1783 itx
= zio_data_buf_alloc(itxsize
);
1784 itx
->itx_lr
.lrc_txtype
= txtype
;
1785 itx
->itx_lr
.lrc_reclen
= lrsize
;
1786 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1787 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1788 itx
->itx_callback
= NULL
;
1789 itx
->itx_callback_data
= NULL
;
1790 itx
->itx_size
= itxsize
;
1796 zil_itx_destroy(itx_t
*itx
)
1798 IMPLY(itx
->itx_lr
.lrc_txtype
== TX_COMMIT
, itx
->itx_callback
== NULL
);
1799 IMPLY(itx
->itx_callback
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
1801 if (itx
->itx_callback
!= NULL
)
1802 itx
->itx_callback(itx
->itx_callback_data
);
1804 zio_data_buf_free(itx
, itx
->itx_size
);
1808 * Free up the sync and async itxs. The itxs_t has already been detached
1809 * so no locks are needed.
1812 zil_itxg_clean(itxs_t
*itxs
)
1818 itx_async_node_t
*ian
;
1820 list
= &itxs
->i_sync_list
;
1821 while ((itx
= list_head(list
)) != NULL
) {
1823 * In the general case, commit itxs will not be found
1824 * here, as they'll be committed to an lwb via
1825 * zil_lwb_commit(), and free'd in that function. Having
1826 * said that, it is still possible for commit itxs to be
1827 * found here, due to the following race:
1829 * - a thread calls zil_commit() which assigns the
1830 * commit itx to a per-txg i_sync_list
1831 * - zil_itxg_clean() is called (e.g. via spa_sync())
1832 * while the waiter is still on the i_sync_list
1834 * There's nothing to prevent syncing the txg while the
1835 * waiter is on the i_sync_list. This normally doesn't
1836 * happen because spa_sync() is slower than zil_commit(),
1837 * but if zil_commit() calls txg_wait_synced() (e.g.
1838 * because zil_create() or zil_commit_writer_stall() is
1839 * called) we will hit this case.
1841 if (itx
->itx_lr
.lrc_txtype
== TX_COMMIT
)
1842 zil_commit_waiter_skip(itx
->itx_private
);
1844 list_remove(list
, itx
);
1845 zil_itx_destroy(itx
);
1849 t
= &itxs
->i_async_tree
;
1850 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1851 list
= &ian
->ia_list
;
1852 while ((itx
= list_head(list
)) != NULL
) {
1853 list_remove(list
, itx
);
1854 /* commit itxs should never be on the async lists. */
1855 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
1856 zil_itx_destroy(itx
);
1859 kmem_free(ian
, sizeof (itx_async_node_t
));
1863 kmem_free(itxs
, sizeof (itxs_t
));
1867 zil_aitx_compare(const void *x1
, const void *x2
)
1869 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1870 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1872 return (AVL_CMP(o1
, o2
));
1876 * Remove all async itx with the given oid.
1879 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1882 itx_async_node_t
*ian
;
1889 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1891 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1894 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1896 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1897 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1899 mutex_enter(&itxg
->itxg_lock
);
1900 if (itxg
->itxg_txg
!= txg
) {
1901 mutex_exit(&itxg
->itxg_lock
);
1906 * Locate the object node and append its list.
1908 t
= &itxg
->itxg_itxs
->i_async_tree
;
1909 ian
= avl_find(t
, &oid
, &where
);
1911 list_move_tail(&clean_list
, &ian
->ia_list
);
1912 mutex_exit(&itxg
->itxg_lock
);
1914 while ((itx
= list_head(&clean_list
)) != NULL
) {
1915 list_remove(&clean_list
, itx
);
1916 /* commit itxs should never be on the async lists. */
1917 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
1918 zil_itx_destroy(itx
);
1920 list_destroy(&clean_list
);
1924 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1928 itxs_t
*itxs
, *clean
= NULL
;
1931 * Ensure the data of a renamed file is committed before the rename.
1933 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1934 zil_async_to_sync(zilog
, itx
->itx_oid
);
1936 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1939 txg
= dmu_tx_get_txg(tx
);
1941 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1942 mutex_enter(&itxg
->itxg_lock
);
1943 itxs
= itxg
->itxg_itxs
;
1944 if (itxg
->itxg_txg
!= txg
) {
1947 * The zil_clean callback hasn't got around to cleaning
1948 * this itxg. Save the itxs for release below.
1949 * This should be rare.
1951 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1952 "txg %llu", itxg
->itxg_txg
);
1953 clean
= itxg
->itxg_itxs
;
1955 itxg
->itxg_txg
= txg
;
1956 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1959 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1960 offsetof(itx_t
, itx_node
));
1961 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1962 sizeof (itx_async_node_t
),
1963 offsetof(itx_async_node_t
, ia_node
));
1965 if (itx
->itx_sync
) {
1966 list_insert_tail(&itxs
->i_sync_list
, itx
);
1968 avl_tree_t
*t
= &itxs
->i_async_tree
;
1970 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
1971 itx_async_node_t
*ian
;
1974 ian
= avl_find(t
, &foid
, &where
);
1976 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1978 list_create(&ian
->ia_list
, sizeof (itx_t
),
1979 offsetof(itx_t
, itx_node
));
1980 ian
->ia_foid
= foid
;
1981 avl_insert(t
, ian
, where
);
1983 list_insert_tail(&ian
->ia_list
, itx
);
1986 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1989 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1990 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1991 * need to be careful to always dirty the ZIL using the "real"
1992 * TXG (not itxg_txg) even when the SPA is frozen.
1994 zilog_dirty(zilog
, dmu_tx_get_txg(tx
));
1995 mutex_exit(&itxg
->itxg_lock
);
1997 /* Release the old itxs now we've dropped the lock */
1999 zil_itxg_clean(clean
);
2003 * If there are any in-memory intent log transactions which have now been
2004 * synced then start up a taskq to free them. We should only do this after we
2005 * have written out the uberblocks (i.e. txg has been committed) so that
2006 * don't inadvertently clean out in-memory log records that would be required
2010 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
2012 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
2015 ASSERT3U(synced_txg
, <, ZILTEST_TXG
);
2017 mutex_enter(&itxg
->itxg_lock
);
2018 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
2019 mutex_exit(&itxg
->itxg_lock
);
2022 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
2023 ASSERT3U(itxg
->itxg_txg
, !=, 0);
2024 clean_me
= itxg
->itxg_itxs
;
2025 itxg
->itxg_itxs
= NULL
;
2027 mutex_exit(&itxg
->itxg_lock
);
2029 * Preferably start a task queue to free up the old itxs but
2030 * if taskq_dispatch can't allocate resources to do that then
2031 * free it in-line. This should be rare. Note, using TQ_SLEEP
2032 * created a bad performance problem.
2034 ASSERT3P(zilog
->zl_dmu_pool
, !=, NULL
);
2035 ASSERT3P(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
, !=, NULL
);
2036 taskqid_t id
= taskq_dispatch(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
,
2037 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
);
2038 if (id
== TASKQID_INVALID
)
2039 zil_itxg_clean(clean_me
);
2043 * This function will traverse the queue of itxs that need to be
2044 * committed, and move them onto the ZIL's zl_itx_commit_list.
2047 zil_get_commit_list(zilog_t
*zilog
)
2050 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
2052 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2054 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2057 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2060 * This is inherently racy, since there is nothing to prevent
2061 * the last synced txg from changing. That's okay since we'll
2062 * only commit things in the future.
2064 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2065 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2067 mutex_enter(&itxg
->itxg_lock
);
2068 if (itxg
->itxg_txg
!= txg
) {
2069 mutex_exit(&itxg
->itxg_lock
);
2074 * If we're adding itx records to the zl_itx_commit_list,
2075 * then the zil better be dirty in this "txg". We can assert
2076 * that here since we're holding the itxg_lock which will
2077 * prevent spa_sync from cleaning it. Once we add the itxs
2078 * to the zl_itx_commit_list we must commit it to disk even
2079 * if it's unnecessary (i.e. the txg was synced).
2081 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
2082 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
2083 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
2085 mutex_exit(&itxg
->itxg_lock
);
2090 * Move the async itxs for a specified object to commit into sync lists.
2093 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
2096 itx_async_node_t
*ian
;
2100 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
2103 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
2106 * This is inherently racy, since there is nothing to prevent
2107 * the last synced txg from changing.
2109 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
2110 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
2112 mutex_enter(&itxg
->itxg_lock
);
2113 if (itxg
->itxg_txg
!= txg
) {
2114 mutex_exit(&itxg
->itxg_lock
);
2119 * If a foid is specified then find that node and append its
2120 * list. Otherwise walk the tree appending all the lists
2121 * to the sync list. We add to the end rather than the
2122 * beginning to ensure the create has happened.
2124 t
= &itxg
->itxg_itxs
->i_async_tree
;
2126 ian
= avl_find(t
, &foid
, &where
);
2128 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2132 void *cookie
= NULL
;
2134 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
2135 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
2137 list_destroy(&ian
->ia_list
);
2138 kmem_free(ian
, sizeof (itx_async_node_t
));
2141 mutex_exit(&itxg
->itxg_lock
);
2146 * This function will prune commit itxs that are at the head of the
2147 * commit list (it won't prune past the first non-commit itx), and
2148 * either: a) attach them to the last lwb that's still pending
2149 * completion, or b) skip them altogether.
2151 * This is used as a performance optimization to prevent commit itxs
2152 * from generating new lwbs when it's unnecessary to do so.
2155 zil_prune_commit_list(zilog_t
*zilog
)
2159 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2161 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2162 lr_t
*lrc
= &itx
->itx_lr
;
2163 if (lrc
->lrc_txtype
!= TX_COMMIT
)
2166 mutex_enter(&zilog
->zl_lock
);
2168 lwb_t
*last_lwb
= zilog
->zl_last_lwb_opened
;
2169 if (last_lwb
== NULL
||
2170 last_lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
) {
2172 * All of the itxs this waiter was waiting on
2173 * must have already completed (or there were
2174 * never any itx's for it to wait on), so it's
2175 * safe to skip this waiter and mark it done.
2177 zil_commit_waiter_skip(itx
->itx_private
);
2179 zil_commit_waiter_link_lwb(itx
->itx_private
, last_lwb
);
2180 itx
->itx_private
= NULL
;
2183 mutex_exit(&zilog
->zl_lock
);
2185 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2186 zil_itx_destroy(itx
);
2189 IMPLY(itx
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2193 zil_commit_writer_stall(zilog_t
*zilog
)
2196 * When zio_alloc_zil() fails to allocate the next lwb block on
2197 * disk, we must call txg_wait_synced() to ensure all of the
2198 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2199 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2200 * to zil_process_commit_list()) will have to call zil_create(),
2201 * and start a new ZIL chain.
2203 * Since zil_alloc_zil() failed, the lwb that was previously
2204 * issued does not have a pointer to the "next" lwb on disk.
2205 * Thus, if another ZIL writer thread was to allocate the "next"
2206 * on-disk lwb, that block could be leaked in the event of a
2207 * crash (because the previous lwb on-disk would not point to
2210 * We must hold the zilog's zl_issuer_lock while we do this, to
2211 * ensure no new threads enter zil_process_commit_list() until
2212 * all lwb's in the zl_lwb_list have been synced and freed
2213 * (which is achieved via the txg_wait_synced() call).
2215 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2216 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2217 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
2221 * This function will traverse the commit list, creating new lwbs as
2222 * needed, and committing the itxs from the commit list to these newly
2223 * created lwbs. Additionally, as a new lwb is created, the previous
2224 * lwb will be issued to the zio layer to be written to disk.
2227 zil_process_commit_list(zilog_t
*zilog
)
2229 spa_t
*spa
= zilog
->zl_spa
;
2231 list_t nolwb_waiters
;
2235 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2238 * Return if there's nothing to commit before we dirty the fs by
2239 * calling zil_create().
2241 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
)
2244 list_create(&nolwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2245 list_create(&nolwb_waiters
, sizeof (zil_commit_waiter_t
),
2246 offsetof(zil_commit_waiter_t
, zcw_node
));
2248 lwb
= list_tail(&zilog
->zl_lwb_list
);
2250 lwb
= zil_create(zilog
);
2252 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2253 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
2254 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
2257 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2258 lr_t
*lrc
= &itx
->itx_lr
;
2259 uint64_t txg
= lrc
->lrc_txg
;
2261 ASSERT3U(txg
, !=, 0);
2263 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2264 DTRACE_PROBE2(zil__process__commit__itx
,
2265 zilog_t
*, zilog
, itx_t
*, itx
);
2267 DTRACE_PROBE2(zil__process__normal__itx
,
2268 zilog_t
*, zilog
, itx_t
*, itx
);
2271 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2273 boolean_t synced
= txg
<= spa_last_synced_txg(spa
);
2274 boolean_t frozen
= txg
> spa_freeze_txg(spa
);
2277 * If the txg of this itx has already been synced out, then
2278 * we don't need to commit this itx to an lwb. This is
2279 * because the data of this itx will have already been
2280 * written to the main pool. This is inherently racy, and
2281 * it's still ok to commit an itx whose txg has already
2282 * been synced; this will result in a write that's
2283 * unnecessary, but will do no harm.
2285 * With that said, we always want to commit TX_COMMIT itxs
2286 * to an lwb, regardless of whether or not that itx's txg
2287 * has been synced out. We do this to ensure any OPENED lwb
2288 * will always have at least one zil_commit_waiter_t linked
2291 * As a counter-example, if we skipped TX_COMMIT itx's
2292 * whose txg had already been synced, the following
2293 * situation could occur if we happened to be racing with
2296 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2297 * itx's txg is 10 and the last synced txg is 9.
2298 * 2. spa_sync finishes syncing out txg 10.
2299 * 3. We move to the next itx in the list, it's a TX_COMMIT
2300 * whose txg is 10, so we skip it rather than committing
2301 * it to the lwb used in (1).
2303 * If the itx that is skipped in (3) is the last TX_COMMIT
2304 * itx in the commit list, than it's possible for the lwb
2305 * used in (1) to remain in the OPENED state indefinitely.
2307 * To prevent the above scenario from occurring, ensuring
2308 * that once an lwb is OPENED it will transition to ISSUED
2309 * and eventually DONE, we always commit TX_COMMIT itx's to
2310 * an lwb here, even if that itx's txg has already been
2313 * Finally, if the pool is frozen, we _always_ commit the
2314 * itx. The point of freezing the pool is to prevent data
2315 * from being written to the main pool via spa_sync, and
2316 * instead rely solely on the ZIL to persistently store the
2317 * data; i.e. when the pool is frozen, the last synced txg
2318 * value can't be trusted.
2320 if (frozen
|| !synced
|| lrc
->lrc_txtype
== TX_COMMIT
) {
2322 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
2325 list_insert_tail(&nolwb_itxs
, itx
);
2327 list_insert_tail(&lwb
->lwb_itxs
, itx
);
2329 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2330 zil_commit_waiter_link_nolwb(
2331 itx
->itx_private
, &nolwb_waiters
);
2334 list_insert_tail(&nolwb_itxs
, itx
);
2337 ASSERT3S(lrc
->lrc_txtype
, !=, TX_COMMIT
);
2338 zil_itx_destroy(itx
);
2344 * This indicates zio_alloc_zil() failed to allocate the
2345 * "next" lwb on-disk. When this happens, we must stall
2346 * the ZIL write pipeline; see the comment within
2347 * zil_commit_writer_stall() for more details.
2349 zil_commit_writer_stall(zilog
);
2352 * Additionally, we have to signal and mark the "nolwb"
2353 * waiters as "done" here, since without an lwb, we
2354 * can't do this via zil_lwb_flush_vdevs_done() like
2357 zil_commit_waiter_t
*zcw
;
2358 while ((zcw
= list_head(&nolwb_waiters
)) != NULL
) {
2359 zil_commit_waiter_skip(zcw
);
2360 list_remove(&nolwb_waiters
, zcw
);
2364 * And finally, we have to destroy the itx's that
2365 * couldn't be committed to an lwb; this will also call
2366 * the itx's callback if one exists for the itx.
2368 while ((itx
= list_head(&nolwb_itxs
)) != NULL
) {
2369 list_remove(&nolwb_itxs
, itx
);
2370 zil_itx_destroy(itx
);
2373 ASSERT(list_is_empty(&nolwb_waiters
));
2374 ASSERT3P(lwb
, !=, NULL
);
2375 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2376 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_WRITE_DONE
);
2377 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_FLUSH_DONE
);
2380 * At this point, the ZIL block pointed at by the "lwb"
2381 * variable is in one of the following states: "closed"
2384 * If it's "closed", then no itxs have been committed to
2385 * it, so there's no point in issuing its zio (i.e. it's
2388 * If it's "open", then it contains one or more itxs that
2389 * eventually need to be committed to stable storage. In
2390 * this case we intentionally do not issue the lwb's zio
2391 * to disk yet, and instead rely on one of the following
2392 * two mechanisms for issuing the zio:
2394 * 1. Ideally, there will be more ZIL activity occurring
2395 * on the system, such that this function will be
2396 * immediately called again (not necessarily by the same
2397 * thread) and this lwb's zio will be issued via
2398 * zil_lwb_commit(). This way, the lwb is guaranteed to
2399 * be "full" when it is issued to disk, and we'll make
2400 * use of the lwb's size the best we can.
2402 * 2. If there isn't sufficient ZIL activity occurring on
2403 * the system, such that this lwb's zio isn't issued via
2404 * zil_lwb_commit(), zil_commit_waiter() will issue the
2405 * lwb's zio. If this occurs, the lwb is not guaranteed
2406 * to be "full" by the time its zio is issued, and means
2407 * the size of the lwb was "too large" given the amount
2408 * of ZIL activity occurring on the system at that time.
2410 * We do this for a couple of reasons:
2412 * 1. To try and reduce the number of IOPs needed to
2413 * write the same number of itxs. If an lwb has space
2414 * available in its buffer for more itxs, and more itxs
2415 * will be committed relatively soon (relative to the
2416 * latency of performing a write), then it's beneficial
2417 * to wait for these "next" itxs. This way, more itxs
2418 * can be committed to stable storage with fewer writes.
2420 * 2. To try and use the largest lwb block size that the
2421 * incoming rate of itxs can support. Again, this is to
2422 * try and pack as many itxs into as few lwbs as
2423 * possible, without significantly impacting the latency
2424 * of each individual itx.
2430 * This function is responsible for ensuring the passed in commit waiter
2431 * (and associated commit itx) is committed to an lwb. If the waiter is
2432 * not already committed to an lwb, all itxs in the zilog's queue of
2433 * itxs will be processed. The assumption is the passed in waiter's
2434 * commit itx will found in the queue just like the other non-commit
2435 * itxs, such that when the entire queue is processed, the waiter will
2436 * have been committed to an lwb.
2438 * The lwb associated with the passed in waiter is not guaranteed to
2439 * have been issued by the time this function completes. If the lwb is
2440 * not issued, we rely on future calls to zil_commit_writer() to issue
2441 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2444 zil_commit_writer(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2446 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2447 ASSERT(spa_writeable(zilog
->zl_spa
));
2449 mutex_enter(&zilog
->zl_issuer_lock
);
2451 if (zcw
->zcw_lwb
!= NULL
|| zcw
->zcw_done
) {
2453 * It's possible that, while we were waiting to acquire
2454 * the "zl_issuer_lock", another thread committed this
2455 * waiter to an lwb. If that occurs, we bail out early,
2456 * without processing any of the zilog's queue of itxs.
2458 * On certain workloads and system configurations, the
2459 * "zl_issuer_lock" can become highly contended. In an
2460 * attempt to reduce this contention, we immediately drop
2461 * the lock if the waiter has already been processed.
2463 * We've measured this optimization to reduce CPU spent
2464 * contending on this lock by up to 5%, using a system
2465 * with 32 CPUs, low latency storage (~50 usec writes),
2466 * and 1024 threads performing sync writes.
2471 ZIL_STAT_BUMP(zil_commit_writer_count
);
2473 zil_get_commit_list(zilog
);
2474 zil_prune_commit_list(zilog
);
2475 zil_process_commit_list(zilog
);
2478 mutex_exit(&zilog
->zl_issuer_lock
);
2482 zil_commit_waiter_timeout(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2484 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2485 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2486 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
2488 lwb_t
*lwb
= zcw
->zcw_lwb
;
2489 ASSERT3P(lwb
, !=, NULL
);
2490 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_CLOSED
);
2493 * If the lwb has already been issued by another thread, we can
2494 * immediately return since there's no work to be done (the
2495 * point of this function is to issue the lwb). Additionally, we
2496 * do this prior to acquiring the zl_issuer_lock, to avoid
2497 * acquiring it when it's not necessary to do so.
2499 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2500 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2501 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
2505 * In order to call zil_lwb_write_issue() we must hold the
2506 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2507 * since we're already holding the commit waiter's "zcw_lock",
2508 * and those two locks are acquired in the opposite order
2511 mutex_exit(&zcw
->zcw_lock
);
2512 mutex_enter(&zilog
->zl_issuer_lock
);
2513 mutex_enter(&zcw
->zcw_lock
);
2516 * Since we just dropped and re-acquired the commit waiter's
2517 * lock, we have to re-check to see if the waiter was marked
2518 * "done" during that process. If the waiter was marked "done",
2519 * the "lwb" pointer is no longer valid (it can be free'd after
2520 * the waiter is marked "done"), so without this check we could
2521 * wind up with a use-after-free error below.
2526 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2529 * We've already checked this above, but since we hadn't acquired
2530 * the zilog's zl_issuer_lock, we have to perform this check a
2531 * second time while holding the lock.
2533 * We don't need to hold the zl_lock since the lwb cannot transition
2534 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2535 * _can_ transition from ISSUED to DONE, but it's OK to race with
2536 * that transition since we treat the lwb the same, whether it's in
2537 * the ISSUED or DONE states.
2539 * The important thing, is we treat the lwb differently depending on
2540 * if it's ISSUED or OPENED, and block any other threads that might
2541 * attempt to issue this lwb. For that reason we hold the
2542 * zl_issuer_lock when checking the lwb_state; we must not call
2543 * zil_lwb_write_issue() if the lwb had already been issued.
2545 * See the comment above the lwb_state_t structure definition for
2546 * more details on the lwb states, and locking requirements.
2548 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2549 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2550 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
)
2553 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
2556 * As described in the comments above zil_commit_waiter() and
2557 * zil_process_commit_list(), we need to issue this lwb's zio
2558 * since we've reached the commit waiter's timeout and it still
2559 * hasn't been issued.
2561 lwb_t
*nlwb
= zil_lwb_write_issue(zilog
, lwb
);
2563 IMPLY(nlwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_OPENED
);
2566 * Since the lwb's zio hadn't been issued by the time this thread
2567 * reached its timeout, we reset the zilog's "zl_cur_used" field
2568 * to influence the zil block size selection algorithm.
2570 * By having to issue the lwb's zio here, it means the size of the
2571 * lwb was too large, given the incoming throughput of itxs. By
2572 * setting "zl_cur_used" to zero, we communicate this fact to the
2573 * block size selection algorithm, so it can take this information
2574 * into account, and potentially select a smaller size for the
2575 * next lwb block that is allocated.
2577 zilog
->zl_cur_used
= 0;
2581 * When zil_lwb_write_issue() returns NULL, this
2582 * indicates zio_alloc_zil() failed to allocate the
2583 * "next" lwb on-disk. When this occurs, the ZIL write
2584 * pipeline must be stalled; see the comment within the
2585 * zil_commit_writer_stall() function for more details.
2587 * We must drop the commit waiter's lock prior to
2588 * calling zil_commit_writer_stall() or else we can wind
2589 * up with the following deadlock:
2591 * - This thread is waiting for the txg to sync while
2592 * holding the waiter's lock; txg_wait_synced() is
2593 * used within txg_commit_writer_stall().
2595 * - The txg can't sync because it is waiting for this
2596 * lwb's zio callback to call dmu_tx_commit().
2598 * - The lwb's zio callback can't call dmu_tx_commit()
2599 * because it's blocked trying to acquire the waiter's
2600 * lock, which occurs prior to calling dmu_tx_commit()
2602 mutex_exit(&zcw
->zcw_lock
);
2603 zil_commit_writer_stall(zilog
);
2604 mutex_enter(&zcw
->zcw_lock
);
2608 mutex_exit(&zilog
->zl_issuer_lock
);
2609 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2613 * This function is responsible for performing the following two tasks:
2615 * 1. its primary responsibility is to block until the given "commit
2616 * waiter" is considered "done".
2618 * 2. its secondary responsibility is to issue the zio for the lwb that
2619 * the given "commit waiter" is waiting on, if this function has
2620 * waited "long enough" and the lwb is still in the "open" state.
2622 * Given a sufficient amount of itxs being generated and written using
2623 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2624 * function. If this does not occur, this secondary responsibility will
2625 * ensure the lwb is issued even if there is not other synchronous
2626 * activity on the system.
2628 * For more details, see zil_process_commit_list(); more specifically,
2629 * the comment at the bottom of that function.
2632 zil_commit_waiter(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2634 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2635 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2636 ASSERT(spa_writeable(zilog
->zl_spa
));
2638 mutex_enter(&zcw
->zcw_lock
);
2641 * The timeout is scaled based on the lwb latency to avoid
2642 * significantly impacting the latency of each individual itx.
2643 * For more details, see the comment at the bottom of the
2644 * zil_process_commit_list() function.
2646 int pct
= MAX(zfs_commit_timeout_pct
, 1);
2647 hrtime_t sleep
= (zilog
->zl_last_lwb_latency
* pct
) / 100;
2648 hrtime_t wakeup
= gethrtime() + sleep
;
2649 boolean_t timedout
= B_FALSE
;
2651 while (!zcw
->zcw_done
) {
2652 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2654 lwb_t
*lwb
= zcw
->zcw_lwb
;
2657 * Usually, the waiter will have a non-NULL lwb field here,
2658 * but it's possible for it to be NULL as a result of
2659 * zil_commit() racing with spa_sync().
2661 * When zil_clean() is called, it's possible for the itxg
2662 * list (which may be cleaned via a taskq) to contain
2663 * commit itxs. When this occurs, the commit waiters linked
2664 * off of these commit itxs will not be committed to an
2665 * lwb. Additionally, these commit waiters will not be
2666 * marked done until zil_commit_waiter_skip() is called via
2669 * Thus, it's possible for this commit waiter (i.e. the
2670 * "zcw" variable) to be found in this "in between" state;
2671 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2672 * been skipped, so it's "zcw_done" field is still B_FALSE.
2674 IMPLY(lwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_CLOSED
);
2676 if (lwb
!= NULL
&& lwb
->lwb_state
== LWB_STATE_OPENED
) {
2677 ASSERT3B(timedout
, ==, B_FALSE
);
2680 * If the lwb hasn't been issued yet, then we
2681 * need to wait with a timeout, in case this
2682 * function needs to issue the lwb after the
2683 * timeout is reached; responsibility (2) from
2684 * the comment above this function.
2686 clock_t timeleft
= cv_timedwait_hires(&zcw
->zcw_cv
,
2687 &zcw
->zcw_lock
, wakeup
, USEC2NSEC(1),
2688 CALLOUT_FLAG_ABSOLUTE
);
2690 if (timeleft
>= 0 || zcw
->zcw_done
)
2694 zil_commit_waiter_timeout(zilog
, zcw
);
2696 if (!zcw
->zcw_done
) {
2698 * If the commit waiter has already been
2699 * marked "done", it's possible for the
2700 * waiter's lwb structure to have already
2701 * been freed. Thus, we can only reliably
2702 * make these assertions if the waiter
2705 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2706 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_OPENED
);
2710 * If the lwb isn't open, then it must have already
2711 * been issued. In that case, there's no need to
2712 * use a timeout when waiting for the lwb to
2715 * Additionally, if the lwb is NULL, the waiter
2716 * will soon be signaled and marked done via
2717 * zil_clean() and zil_itxg_clean(), so no timeout
2722 lwb
->lwb_state
== LWB_STATE_ISSUED
||
2723 lwb
->lwb_state
== LWB_STATE_WRITE_DONE
||
2724 lwb
->lwb_state
== LWB_STATE_FLUSH_DONE
);
2725 cv_wait(&zcw
->zcw_cv
, &zcw
->zcw_lock
);
2729 mutex_exit(&zcw
->zcw_lock
);
2732 static zil_commit_waiter_t
*
2733 zil_alloc_commit_waiter(void)
2735 zil_commit_waiter_t
*zcw
= kmem_cache_alloc(zil_zcw_cache
, KM_SLEEP
);
2737 cv_init(&zcw
->zcw_cv
, NULL
, CV_DEFAULT
, NULL
);
2738 mutex_init(&zcw
->zcw_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2739 list_link_init(&zcw
->zcw_node
);
2740 zcw
->zcw_lwb
= NULL
;
2741 zcw
->zcw_done
= B_FALSE
;
2742 zcw
->zcw_zio_error
= 0;
2748 zil_free_commit_waiter(zil_commit_waiter_t
*zcw
)
2750 ASSERT(!list_link_active(&zcw
->zcw_node
));
2751 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
2752 ASSERT3B(zcw
->zcw_done
, ==, B_TRUE
);
2753 mutex_destroy(&zcw
->zcw_lock
);
2754 cv_destroy(&zcw
->zcw_cv
);
2755 kmem_cache_free(zil_zcw_cache
, zcw
);
2759 * This function is used to create a TX_COMMIT itx and assign it. This
2760 * way, it will be linked into the ZIL's list of synchronous itxs, and
2761 * then later committed to an lwb (or skipped) when
2762 * zil_process_commit_list() is called.
2765 zil_commit_itx_assign(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2767 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
2768 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
2770 itx_t
*itx
= zil_itx_create(TX_COMMIT
, sizeof (lr_t
));
2771 itx
->itx_sync
= B_TRUE
;
2772 itx
->itx_private
= zcw
;
2774 zil_itx_assign(zilog
, itx
, tx
);
2780 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2782 * When writing ZIL transactions to the on-disk representation of the
2783 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2784 * itxs can be committed to a single lwb. Once a lwb is written and
2785 * committed to stable storage (i.e. the lwb is written, and vdevs have
2786 * been flushed), each itx that was committed to that lwb is also
2787 * considered to be committed to stable storage.
2789 * When an itx is committed to an lwb, the log record (lr_t) contained
2790 * by the itx is copied into the lwb's zio buffer, and once this buffer
2791 * is written to disk, it becomes an on-disk ZIL block.
2793 * As itxs are generated, they're inserted into the ZIL's queue of
2794 * uncommitted itxs. The semantics of zil_commit() are such that it will
2795 * block until all itxs that were in the queue when it was called, are
2796 * committed to stable storage.
2798 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2799 * itxs, for all objects in the dataset, will be committed to stable
2800 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2801 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2802 * that correspond to the foid passed in, will be committed to stable
2803 * storage prior to zil_commit() returning.
2805 * Generally speaking, when zil_commit() is called, the consumer doesn't
2806 * actually care about _all_ of the uncommitted itxs. Instead, they're
2807 * simply trying to waiting for a specific itx to be committed to disk,
2808 * but the interface(s) for interacting with the ZIL don't allow such
2809 * fine-grained communication. A better interface would allow a consumer
2810 * to create and assign an itx, and then pass a reference to this itx to
2811 * zil_commit(); such that zil_commit() would return as soon as that
2812 * specific itx was committed to disk (instead of waiting for _all_
2813 * itxs to be committed).
2815 * When a thread calls zil_commit() a special "commit itx" will be
2816 * generated, along with a corresponding "waiter" for this commit itx.
2817 * zil_commit() will wait on this waiter's CV, such that when the waiter
2818 * is marked done, and signaled, zil_commit() will return.
2820 * This commit itx is inserted into the queue of uncommitted itxs. This
2821 * provides an easy mechanism for determining which itxs were in the
2822 * queue prior to zil_commit() having been called, and which itxs were
2823 * added after zil_commit() was called.
2825 * The commit it is special; it doesn't have any on-disk representation.
2826 * When a commit itx is "committed" to an lwb, the waiter associated
2827 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2828 * completes, each waiter on the lwb's list is marked done and signaled
2829 * -- allowing the thread waiting on the waiter to return from zil_commit().
2831 * It's important to point out a few critical factors that allow us
2832 * to make use of the commit itxs, commit waiters, per-lwb lists of
2833 * commit waiters, and zio completion callbacks like we're doing:
2835 * 1. The list of waiters for each lwb is traversed, and each commit
2836 * waiter is marked "done" and signaled, in the zio completion
2837 * callback of the lwb's zio[*].
2839 * * Actually, the waiters are signaled in the zio completion
2840 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2841 * that are sent to the vdevs upon completion of the lwb zio.
2843 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2844 * itxs, the order in which they are inserted is preserved[*]; as
2845 * itxs are added to the queue, they are added to the tail of
2846 * in-memory linked lists.
2848 * When committing the itxs to lwbs (to be written to disk), they
2849 * are committed in the same order in which the itxs were added to
2850 * the uncommitted queue's linked list(s); i.e. the linked list of
2851 * itxs to commit is traversed from head to tail, and each itx is
2852 * committed to an lwb in that order.
2856 * - the order of "sync" itxs is preserved w.r.t. other
2857 * "sync" itxs, regardless of the corresponding objects.
2858 * - the order of "async" itxs is preserved w.r.t. other
2859 * "async" itxs corresponding to the same object.
2860 * - the order of "async" itxs is *not* preserved w.r.t. other
2861 * "async" itxs corresponding to different objects.
2862 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2863 * versa) is *not* preserved, even for itxs that correspond
2864 * to the same object.
2866 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2867 * zil_get_commit_list(), and zil_process_commit_list().
2869 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2870 * lwb cannot be considered committed to stable storage, until its
2871 * "previous" lwb is also committed to stable storage. This fact,
2872 * coupled with the fact described above, means that itxs are
2873 * committed in (roughly) the order in which they were generated.
2874 * This is essential because itxs are dependent on prior itxs.
2875 * Thus, we *must not* deem an itx as being committed to stable
2876 * storage, until *all* prior itxs have also been committed to
2879 * To enforce this ordering of lwb zio's, while still leveraging as
2880 * much of the underlying storage performance as possible, we rely
2881 * on two fundamental concepts:
2883 * 1. The creation and issuance of lwb zio's is protected by
2884 * the zilog's "zl_issuer_lock", which ensures only a single
2885 * thread is creating and/or issuing lwb's at a time
2886 * 2. The "previous" lwb is a child of the "current" lwb
2887 * (leveraging the zio parent-child dependency graph)
2889 * By relying on this parent-child zio relationship, we can have
2890 * many lwb zio's concurrently issued to the underlying storage,
2891 * but the order in which they complete will be the same order in
2892 * which they were created.
2895 zil_commit(zilog_t
*zilog
, uint64_t foid
)
2898 * We should never attempt to call zil_commit on a snapshot for
2899 * a couple of reasons:
2901 * 1. A snapshot may never be modified, thus it cannot have any
2902 * in-flight itxs that would have modified the dataset.
2904 * 2. By design, when zil_commit() is called, a commit itx will
2905 * be assigned to this zilog; as a result, the zilog will be
2906 * dirtied. We must not dirty the zilog of a snapshot; there's
2907 * checks in the code that enforce this invariant, and will
2908 * cause a panic if it's not upheld.
2910 ASSERT3B(dmu_objset_is_snapshot(zilog
->zl_os
), ==, B_FALSE
);
2912 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2915 if (!spa_writeable(zilog
->zl_spa
)) {
2917 * If the SPA is not writable, there should never be any
2918 * pending itxs waiting to be committed to disk. If that
2919 * weren't true, we'd skip writing those itxs out, and
2920 * would break the semantics of zil_commit(); thus, we're
2921 * verifying that truth before we return to the caller.
2923 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
2924 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
2925 for (int i
= 0; i
< TXG_SIZE
; i
++)
2926 ASSERT3P(zilog
->zl_itxg
[i
].itxg_itxs
, ==, NULL
);
2931 * If the ZIL is suspended, we don't want to dirty it by calling
2932 * zil_commit_itx_assign() below, nor can we write out
2933 * lwbs like would be done in zil_commit_write(). Thus, we
2934 * simply rely on txg_wait_synced() to maintain the necessary
2935 * semantics, and avoid calling those functions altogether.
2937 if (zilog
->zl_suspend
> 0) {
2938 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2942 zil_commit_impl(zilog
, foid
);
2946 zil_commit_impl(zilog_t
*zilog
, uint64_t foid
)
2948 ZIL_STAT_BUMP(zil_commit_count
);
2951 * Move the "async" itxs for the specified foid to the "sync"
2952 * queues, such that they will be later committed (or skipped)
2953 * to an lwb when zil_process_commit_list() is called.
2955 * Since these "async" itxs must be committed prior to this
2956 * call to zil_commit returning, we must perform this operation
2957 * before we call zil_commit_itx_assign().
2959 zil_async_to_sync(zilog
, foid
);
2962 * We allocate a new "waiter" structure which will initially be
2963 * linked to the commit itx using the itx's "itx_private" field.
2964 * Since the commit itx doesn't represent any on-disk state,
2965 * when it's committed to an lwb, rather than copying the its
2966 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2967 * added to the lwb's list of waiters. Then, when the lwb is
2968 * committed to stable storage, each waiter in the lwb's list of
2969 * waiters will be marked "done", and signalled.
2971 * We must create the waiter and assign the commit itx prior to
2972 * calling zil_commit_writer(), or else our specific commit itx
2973 * is not guaranteed to be committed to an lwb prior to calling
2974 * zil_commit_waiter().
2976 zil_commit_waiter_t
*zcw
= zil_alloc_commit_waiter();
2977 zil_commit_itx_assign(zilog
, zcw
);
2979 zil_commit_writer(zilog
, zcw
);
2980 zil_commit_waiter(zilog
, zcw
);
2982 if (zcw
->zcw_zio_error
!= 0) {
2984 * If there was an error writing out the ZIL blocks that
2985 * this thread is waiting on, then we fallback to
2986 * relying on spa_sync() to write out the data this
2987 * thread is waiting on. Obviously this has performance
2988 * implications, but the expectation is for this to be
2989 * an exceptional case, and shouldn't occur often.
2991 DTRACE_PROBE2(zil__commit__io__error
,
2992 zilog_t
*, zilog
, zil_commit_waiter_t
*, zcw
);
2993 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2996 zil_free_commit_waiter(zcw
);
3000 * Called in syncing context to free committed log blocks and update log header.
3003 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
3005 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
3006 uint64_t txg
= dmu_tx_get_txg(tx
);
3007 spa_t
*spa
= zilog
->zl_spa
;
3008 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
3012 * We don't zero out zl_destroy_txg, so make sure we don't try
3013 * to destroy it twice.
3015 if (spa_sync_pass(spa
) != 1)
3018 mutex_enter(&zilog
->zl_lock
);
3020 ASSERT(zilog
->zl_stop_sync
== 0);
3022 if (*replayed_seq
!= 0) {
3023 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
3024 zh
->zh_replay_seq
= *replayed_seq
;
3028 if (zilog
->zl_destroy_txg
== txg
) {
3029 blkptr_t blk
= zh
->zh_log
;
3031 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
3033 bzero(zh
, sizeof (zil_header_t
));
3034 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
3036 if (zilog
->zl_keep_first
) {
3038 * If this block was part of log chain that couldn't
3039 * be claimed because a device was missing during
3040 * zil_claim(), but that device later returns,
3041 * then this block could erroneously appear valid.
3042 * To guard against this, assign a new GUID to the new
3043 * log chain so it doesn't matter what blk points to.
3045 zil_init_log_chain(zilog
, &blk
);
3050 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
3051 zh
->zh_log
= lwb
->lwb_blk
;
3052 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
3054 list_remove(&zilog
->zl_lwb_list
, lwb
);
3055 zio_free(spa
, txg
, &lwb
->lwb_blk
);
3056 zil_free_lwb(zilog
, lwb
);
3059 * If we don't have anything left in the lwb list then
3060 * we've had an allocation failure and we need to zero
3061 * out the zil_header blkptr so that we don't end
3062 * up freeing the same block twice.
3064 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
3065 BP_ZERO(&zh
->zh_log
);
3069 * Remove fastwrite on any blocks that have been pre-allocated for
3070 * the next commit. This prevents fastwrite counter pollution by
3071 * unused, long-lived LWBs.
3073 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
3074 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_write_zio
) {
3075 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
3076 lwb
->lwb_fastwrite
= 0;
3080 mutex_exit(&zilog
->zl_lock
);
3085 zil_lwb_cons(void *vbuf
, void *unused
, int kmflag
)
3088 list_create(&lwb
->lwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
3089 list_create(&lwb
->lwb_waiters
, sizeof (zil_commit_waiter_t
),
3090 offsetof(zil_commit_waiter_t
, zcw_node
));
3091 avl_create(&lwb
->lwb_vdev_tree
, zil_lwb_vdev_compare
,
3092 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
3093 mutex_init(&lwb
->lwb_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3099 zil_lwb_dest(void *vbuf
, void *unused
)
3102 mutex_destroy(&lwb
->lwb_vdev_lock
);
3103 avl_destroy(&lwb
->lwb_vdev_tree
);
3104 list_destroy(&lwb
->lwb_waiters
);
3105 list_destroy(&lwb
->lwb_itxs
);
3111 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
3112 sizeof (lwb_t
), 0, zil_lwb_cons
, zil_lwb_dest
, NULL
, NULL
, NULL
, 0);
3114 zil_zcw_cache
= kmem_cache_create("zil_zcw_cache",
3115 sizeof (zil_commit_waiter_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
3117 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
3118 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
3119 KSTAT_FLAG_VIRTUAL
);
3121 if (zil_ksp
!= NULL
) {
3122 zil_ksp
->ks_data
= &zil_stats
;
3123 kstat_install(zil_ksp
);
3130 kmem_cache_destroy(zil_zcw_cache
);
3131 kmem_cache_destroy(zil_lwb_cache
);
3133 if (zil_ksp
!= NULL
) {
3134 kstat_delete(zil_ksp
);
3140 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
3142 zilog
->zl_sync
= sync
;
3146 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
3148 zilog
->zl_logbias
= logbias
;
3152 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
3156 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
3158 zilog
->zl_header
= zh_phys
;
3160 zilog
->zl_spa
= dmu_objset_spa(os
);
3161 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
3162 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
3163 zilog
->zl_logbias
= dmu_objset_logbias(os
);
3164 zilog
->zl_sync
= dmu_objset_syncprop(os
);
3165 zilog
->zl_dirty_max_txg
= 0;
3166 zilog
->zl_last_lwb_opened
= NULL
;
3167 zilog
->zl_last_lwb_latency
= 0;
3168 zilog
->zl_max_block_size
= zil_maxblocksize
;
3170 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3171 mutex_init(&zilog
->zl_issuer_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3173 for (int i
= 0; i
< TXG_SIZE
; i
++) {
3174 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
3175 MUTEX_DEFAULT
, NULL
);
3178 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
3179 offsetof(lwb_t
, lwb_node
));
3181 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
3182 offsetof(itx_t
, itx_node
));
3184 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
3190 zil_free(zilog_t
*zilog
)
3194 zilog
->zl_stop_sync
= 1;
3196 ASSERT0(zilog
->zl_suspend
);
3197 ASSERT0(zilog
->zl_suspending
);
3199 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3200 list_destroy(&zilog
->zl_lwb_list
);
3202 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
3203 list_destroy(&zilog
->zl_itx_commit_list
);
3205 for (i
= 0; i
< TXG_SIZE
; i
++) {
3207 * It's possible for an itx to be generated that doesn't dirty
3208 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3209 * callback to remove the entry. We remove those here.
3211 * Also free up the ziltest itxs.
3213 if (zilog
->zl_itxg
[i
].itxg_itxs
)
3214 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
3215 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
3218 mutex_destroy(&zilog
->zl_issuer_lock
);
3219 mutex_destroy(&zilog
->zl_lock
);
3221 cv_destroy(&zilog
->zl_cv_suspend
);
3223 kmem_free(zilog
, sizeof (zilog_t
));
3227 * Open an intent log.
3230 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
3232 zilog_t
*zilog
= dmu_objset_zil(os
);
3234 ASSERT3P(zilog
->zl_get_data
, ==, NULL
);
3235 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3236 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3238 zilog
->zl_get_data
= get_data
;
3244 * Close an intent log.
3247 zil_close(zilog_t
*zilog
)
3252 if (!dmu_objset_is_snapshot(zilog
->zl_os
)) {
3253 zil_commit(zilog
, 0);
3255 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
3256 ASSERT0(zilog
->zl_dirty_max_txg
);
3257 ASSERT3B(zilog_is_dirty(zilog
), ==, B_FALSE
);
3260 mutex_enter(&zilog
->zl_lock
);
3261 lwb
= list_tail(&zilog
->zl_lwb_list
);
3263 txg
= zilog
->zl_dirty_max_txg
;
3265 txg
= MAX(zilog
->zl_dirty_max_txg
, lwb
->lwb_max_txg
);
3266 mutex_exit(&zilog
->zl_lock
);
3269 * We need to use txg_wait_synced() to wait long enough for the
3270 * ZIL to be clean, and to wait for all pending lwbs to be
3274 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
3276 if (zilog_is_dirty(zilog
))
3277 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog
, txg
);
3278 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
3279 VERIFY(!zilog_is_dirty(zilog
));
3281 zilog
->zl_get_data
= NULL
;
3284 * We should have only one lwb left on the list; remove it now.
3286 mutex_enter(&zilog
->zl_lock
);
3287 lwb
= list_head(&zilog
->zl_lwb_list
);
3289 ASSERT3P(lwb
, ==, list_tail(&zilog
->zl_lwb_list
));
3290 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
3292 if (lwb
->lwb_fastwrite
)
3293 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
3295 list_remove(&zilog
->zl_lwb_list
, lwb
);
3296 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
3297 zil_free_lwb(zilog
, lwb
);
3299 mutex_exit(&zilog
->zl_lock
);
3302 static char *suspend_tag
= "zil suspending";
3305 * Suspend an intent log. While in suspended mode, we still honor
3306 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3307 * On old version pools, we suspend the log briefly when taking a
3308 * snapshot so that it will have an empty intent log.
3310 * Long holds are not really intended to be used the way we do here --
3311 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3312 * could fail. Therefore we take pains to only put a long hold if it is
3313 * actually necessary. Fortunately, it will only be necessary if the
3314 * objset is currently mounted (or the ZVOL equivalent). In that case it
3315 * will already have a long hold, so we are not really making things any worse.
3317 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3318 * zvol_state_t), and use their mechanism to prevent their hold from being
3319 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3322 * if cookiep == NULL, this does both the suspend & resume.
3323 * Otherwise, it returns with the dataset "long held", and the cookie
3324 * should be passed into zil_resume().
3327 zil_suspend(const char *osname
, void **cookiep
)
3331 const zil_header_t
*zh
;
3334 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
3337 zilog
= dmu_objset_zil(os
);
3339 mutex_enter(&zilog
->zl_lock
);
3340 zh
= zilog
->zl_header
;
3342 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
3343 mutex_exit(&zilog
->zl_lock
);
3344 dmu_objset_rele(os
, suspend_tag
);
3345 return (SET_ERROR(EBUSY
));
3349 * Don't put a long hold in the cases where we can avoid it. This
3350 * is when there is no cookie so we are doing a suspend & resume
3351 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3352 * for the suspend because it's already suspended, or there's no ZIL.
3354 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
3355 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
3356 mutex_exit(&zilog
->zl_lock
);
3357 dmu_objset_rele(os
, suspend_tag
);
3361 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
3362 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
3364 zilog
->zl_suspend
++;
3366 if (zilog
->zl_suspend
> 1) {
3368 * Someone else is already suspending it.
3369 * Just wait for them to finish.
3372 while (zilog
->zl_suspending
)
3373 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
3374 mutex_exit(&zilog
->zl_lock
);
3376 if (cookiep
== NULL
)
3384 * If there is no pointer to an on-disk block, this ZIL must not
3385 * be active (e.g. filesystem not mounted), so there's nothing
3388 if (BP_IS_HOLE(&zh
->zh_log
)) {
3389 ASSERT(cookiep
!= NULL
); /* fast path already handled */
3392 mutex_exit(&zilog
->zl_lock
);
3397 * The ZIL has work to do. Ensure that the associated encryption
3398 * key will remain mapped while we are committing the log by
3399 * grabbing a reference to it. If the key isn't loaded we have no
3400 * choice but to return an error until the wrapping key is loaded.
3402 if (os
->os_encrypted
&&
3403 dsl_dataset_create_key_mapping(dmu_objset_ds(os
)) != 0) {
3404 zilog
->zl_suspend
--;
3405 mutex_exit(&zilog
->zl_lock
);
3406 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3407 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3408 return (SET_ERROR(EACCES
));
3411 zilog
->zl_suspending
= B_TRUE
;
3412 mutex_exit(&zilog
->zl_lock
);
3415 * We need to use zil_commit_impl to ensure we wait for all
3416 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3417 * to disk before proceeding. If we used zil_commit instead, it
3418 * would just call txg_wait_synced(), because zl_suspend is set.
3419 * txg_wait_synced() doesn't wait for these lwb's to be
3420 * LWB_STATE_FLUSH_DONE before returning.
3422 zil_commit_impl(zilog
, 0);
3425 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3426 * use txg_wait_synced() to ensure the data from the zilog has
3427 * migrated to the main pool before calling zil_destroy().
3429 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3431 zil_destroy(zilog
, B_FALSE
);
3433 mutex_enter(&zilog
->zl_lock
);
3434 zilog
->zl_suspending
= B_FALSE
;
3435 cv_broadcast(&zilog
->zl_cv_suspend
);
3436 mutex_exit(&zilog
->zl_lock
);
3438 if (os
->os_encrypted
)
3439 dsl_dataset_remove_key_mapping(dmu_objset_ds(os
));
3441 if (cookiep
== NULL
)
3449 zil_resume(void *cookie
)
3451 objset_t
*os
= cookie
;
3452 zilog_t
*zilog
= dmu_objset_zil(os
);
3454 mutex_enter(&zilog
->zl_lock
);
3455 ASSERT(zilog
->zl_suspend
!= 0);
3456 zilog
->zl_suspend
--;
3457 mutex_exit(&zilog
->zl_lock
);
3458 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3459 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3462 typedef struct zil_replay_arg
{
3463 zil_replay_func_t
**zr_replay
;
3465 boolean_t zr_byteswap
;
3470 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
3472 char name
[ZFS_MAX_DATASET_NAME_LEN
];
3474 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
3476 dmu_objset_name(zilog
->zl_os
, name
);
3478 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
3479 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
3480 (u_longlong_t
)lr
->lrc_seq
,
3481 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
3482 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
3488 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
3490 zil_replay_arg_t
*zr
= zra
;
3491 const zil_header_t
*zh
= zilog
->zl_header
;
3492 uint64_t reclen
= lr
->lrc_reclen
;
3493 uint64_t txtype
= lr
->lrc_txtype
;
3496 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
3498 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
3501 if (lr
->lrc_txg
< claim_txg
) /* already committed */
3504 /* Strip case-insensitive bit, still present in log record */
3507 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
3508 return (zil_replay_error(zilog
, lr
, EINVAL
));
3511 * If this record type can be logged out of order, the object
3512 * (lr_foid) may no longer exist. That's legitimate, not an error.
3514 if (TX_OOO(txtype
)) {
3515 error
= dmu_object_info(zilog
->zl_os
,
3516 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
3517 if (error
== ENOENT
|| error
== EEXIST
)
3522 * Make a copy of the data so we can revise and extend it.
3524 bcopy(lr
, zr
->zr_lr
, reclen
);
3527 * If this is a TX_WRITE with a blkptr, suck in the data.
3529 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
3530 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
3531 zr
->zr_lr
+ reclen
);
3533 return (zil_replay_error(zilog
, lr
, error
));
3537 * The log block containing this lr may have been byteswapped
3538 * so that we can easily examine common fields like lrc_txtype.
3539 * However, the log is a mix of different record types, and only the
3540 * replay vectors know how to byteswap their records. Therefore, if
3541 * the lr was byteswapped, undo it before invoking the replay vector.
3543 if (zr
->zr_byteswap
)
3544 byteswap_uint64_array(zr
->zr_lr
, reclen
);
3547 * We must now do two things atomically: replay this log record,
3548 * and update the log header sequence number to reflect the fact that
3549 * we did so. At the end of each replay function the sequence number
3550 * is updated if we are in replay mode.
3552 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
3555 * The DMU's dnode layer doesn't see removes until the txg
3556 * commits, so a subsequent claim can spuriously fail with
3557 * EEXIST. So if we receive any error we try syncing out
3558 * any removes then retry the transaction. Note that we
3559 * specify B_FALSE for byteswap now, so we don't do it twice.
3561 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
3562 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
3564 return (zil_replay_error(zilog
, lr
, error
));
3571 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
3573 zilog
->zl_replay_blks
++;
3579 * If this dataset has a non-empty intent log, replay it and destroy it.
3582 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t
*replay_func
[TX_MAX_TYPE
])
3584 zilog_t
*zilog
= dmu_objset_zil(os
);
3585 const zil_header_t
*zh
= zilog
->zl_header
;
3586 zil_replay_arg_t zr
;
3588 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
3589 zil_destroy(zilog
, B_TRUE
);
3593 zr
.zr_replay
= replay_func
;
3595 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
3596 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
3599 * Wait for in-progress removes to sync before starting replay.
3601 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3603 zilog
->zl_replay
= B_TRUE
;
3604 zilog
->zl_replay_time
= ddi_get_lbolt();
3605 ASSERT(zilog
->zl_replay_blks
== 0);
3606 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
3607 zh
->zh_claim_txg
, B_TRUE
);
3608 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
3610 zil_destroy(zilog
, B_FALSE
);
3611 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
3612 zilog
->zl_replay
= B_FALSE
;
3616 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
3618 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3621 if (zilog
->zl_replay
) {
3622 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
3623 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
3624 zilog
->zl_replaying_seq
;
3633 zil_reset(const char *osname
, void *arg
)
3637 error
= zil_suspend(osname
, NULL
);
3638 /* EACCES means crypto key not loaded */
3639 if ((error
== EACCES
) || (error
== EBUSY
))
3640 return (SET_ERROR(error
));
3642 return (SET_ERROR(EEXIST
));
3646 #if defined(_KERNEL)
3647 EXPORT_SYMBOL(zil_alloc
);
3648 EXPORT_SYMBOL(zil_free
);
3649 EXPORT_SYMBOL(zil_open
);
3650 EXPORT_SYMBOL(zil_close
);
3651 EXPORT_SYMBOL(zil_replay
);
3652 EXPORT_SYMBOL(zil_replaying
);
3653 EXPORT_SYMBOL(zil_destroy
);
3654 EXPORT_SYMBOL(zil_destroy_sync
);
3655 EXPORT_SYMBOL(zil_itx_create
);
3656 EXPORT_SYMBOL(zil_itx_destroy
);
3657 EXPORT_SYMBOL(zil_itx_assign
);
3658 EXPORT_SYMBOL(zil_commit
);
3659 EXPORT_SYMBOL(zil_claim
);
3660 EXPORT_SYMBOL(zil_check_log_chain
);
3661 EXPORT_SYMBOL(zil_sync
);
3662 EXPORT_SYMBOL(zil_clean
);
3663 EXPORT_SYMBOL(zil_suspend
);
3664 EXPORT_SYMBOL(zil_resume
);
3665 EXPORT_SYMBOL(zil_lwb_add_block
);
3666 EXPORT_SYMBOL(zil_bp_tree_add
);
3667 EXPORT_SYMBOL(zil_set_sync
);
3668 EXPORT_SYMBOL(zil_set_logbias
);
3671 module_param(zfs_commit_timeout_pct
, int, 0644);
3672 MODULE_PARM_DESC(zfs_commit_timeout_pct
, "ZIL block open timeout percentage");
3674 module_param(zil_replay_disable
, int, 0644);
3675 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
3677 module_param(zil_nocacheflush
, int, 0644);
3678 MODULE_PARM_DESC(zil_nocacheflush
, "Disable ZIL cache flushes");
3680 module_param(zil_slog_bulk
, ulong
, 0644);
3681 MODULE_PARM_DESC(zil_slog_bulk
, "Limit in bytes slog sync writes per commit");
3683 module_param(zil_maxblocksize
, int, 0644);
3684 MODULE_PARM_DESC(zil_maxblocksize
, "Limit in bytes of ZIL log block size");