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 fileystem, 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 * Tunable parameter for debugging or performance analysis. Setting
122 * zfs_nocacheflush will cause corruption on power loss if a volatile
123 * out-of-order write cache is enabled.
125 int zfs_nocacheflush
= 0;
128 * Limit SLOG write size per commit executed with synchronous priority.
129 * Any writes above that will be executed with lower (asynchronous) priority
130 * to limit potential SLOG device abuse by single active ZIL writer.
132 unsigned long zil_slog_bulk
= 768 * 1024;
134 static kmem_cache_t
*zil_lwb_cache
;
135 static kmem_cache_t
*zil_zcw_cache
;
137 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
139 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
140 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
143 zil_bp_compare(const void *x1
, const void *x2
)
145 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
146 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
148 int cmp
= AVL_CMP(DVA_GET_VDEV(dva1
), DVA_GET_VDEV(dva2
));
152 return (AVL_CMP(DVA_GET_OFFSET(dva1
), DVA_GET_OFFSET(dva2
)));
156 zil_bp_tree_init(zilog_t
*zilog
)
158 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
159 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
163 zil_bp_tree_fini(zilog_t
*zilog
)
165 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
169 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
170 kmem_free(zn
, sizeof (zil_bp_node_t
));
176 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
178 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
183 if (BP_IS_EMBEDDED(bp
))
186 dva
= BP_IDENTITY(bp
);
188 if (avl_find(t
, dva
, &where
) != NULL
)
189 return (SET_ERROR(EEXIST
));
191 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
193 avl_insert(t
, zn
, where
);
198 static zil_header_t
*
199 zil_header_in_syncing_context(zilog_t
*zilog
)
201 return ((zil_header_t
*)zilog
->zl_header
);
205 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
207 zio_cksum_t
*zc
= &bp
->blk_cksum
;
209 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
210 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
211 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
212 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
216 * Read a log block and make sure it's valid.
219 zil_read_log_block(zilog_t
*zilog
, boolean_t decrypt
, const blkptr_t
*bp
,
220 blkptr_t
*nbp
, void *dst
, char **end
)
222 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
223 arc_flags_t aflags
= ARC_FLAG_WAIT
;
224 arc_buf_t
*abuf
= NULL
;
228 if (zilog
->zl_header
->zh_claim_txg
== 0)
229 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
231 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
232 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
235 zio_flags
|= ZIO_FLAG_RAW
;
237 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
238 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
240 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
,
241 &abuf
, ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
244 zio_cksum_t cksum
= bp
->blk_cksum
;
247 * Validate the checksummed log block.
249 * Sequence numbers should be... sequential. The checksum
250 * verifier for the next block should be bp's checksum plus 1.
252 * Also check the log chain linkage and size used.
254 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
256 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
257 zil_chain_t
*zilc
= abuf
->b_data
;
258 char *lr
= (char *)(zilc
+ 1);
259 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
261 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
262 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
263 error
= SET_ERROR(ECKSUM
);
265 ASSERT3U(len
, <=, SPA_OLD_MAXBLOCKSIZE
);
267 *end
= (char *)dst
+ len
;
268 *nbp
= zilc
->zc_next_blk
;
271 char *lr
= abuf
->b_data
;
272 uint64_t size
= BP_GET_LSIZE(bp
);
273 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
275 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
276 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
277 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
278 error
= SET_ERROR(ECKSUM
);
280 ASSERT3U(zilc
->zc_nused
, <=,
281 SPA_OLD_MAXBLOCKSIZE
);
282 bcopy(lr
, dst
, zilc
->zc_nused
);
283 *end
= (char *)dst
+ zilc
->zc_nused
;
284 *nbp
= zilc
->zc_next_blk
;
288 arc_buf_destroy(abuf
, &abuf
);
295 * Read a TX_WRITE log data block.
298 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
300 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
301 const blkptr_t
*bp
= &lr
->lr_blkptr
;
302 arc_flags_t aflags
= ARC_FLAG_WAIT
;
303 arc_buf_t
*abuf
= NULL
;
307 if (BP_IS_HOLE(bp
)) {
309 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
313 if (zilog
->zl_header
->zh_claim_txg
== 0)
314 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
317 * If we are not using the resulting data, we are just checking that
318 * it hasn't been corrupted so we don't need to waste CPU time
319 * decompressing and decrypting it.
322 zio_flags
|= ZIO_FLAG_RAW
;
324 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
325 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
327 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
328 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
332 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
333 arc_buf_destroy(abuf
, &abuf
);
340 * Parse the intent log, and call parse_func for each valid record within.
343 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
344 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
,
347 const zil_header_t
*zh
= zilog
->zl_header
;
348 boolean_t claimed
= !!zh
->zh_claim_txg
;
349 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
350 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
351 uint64_t max_blk_seq
= 0;
352 uint64_t max_lr_seq
= 0;
353 uint64_t blk_count
= 0;
354 uint64_t lr_count
= 0;
355 blkptr_t blk
, next_blk
;
359 bzero(&next_blk
, sizeof (blkptr_t
));
362 * Old logs didn't record the maximum zh_claim_lr_seq.
364 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
365 claim_lr_seq
= UINT64_MAX
;
368 * Starting at the block pointed to by zh_log we read the log chain.
369 * For each block in the chain we strongly check that block to
370 * ensure its validity. We stop when an invalid block is found.
371 * For each block pointer in the chain we call parse_blk_func().
372 * For each record in each valid block we call parse_lr_func().
373 * If the log has been claimed, stop if we encounter a sequence
374 * number greater than the highest claimed sequence number.
376 lrbuf
= zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE
);
377 zil_bp_tree_init(zilog
);
379 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
380 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
384 if (blk_seq
> claim_blk_seq
)
387 error
= parse_blk_func(zilog
, &blk
, arg
, txg
);
390 ASSERT3U(max_blk_seq
, <, blk_seq
);
391 max_blk_seq
= blk_seq
;
394 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
397 error
= zil_read_log_block(zilog
, decrypt
, &blk
, &next_blk
,
402 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
403 lr_t
*lr
= (lr_t
*)lrp
;
404 reclen
= lr
->lrc_reclen
;
405 ASSERT3U(reclen
, >=, sizeof (lr_t
));
406 if (lr
->lrc_seq
> claim_lr_seq
)
409 error
= parse_lr_func(zilog
, lr
, arg
, txg
);
412 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
413 max_lr_seq
= lr
->lrc_seq
;
418 zilog
->zl_parse_error
= error
;
419 zilog
->zl_parse_blk_seq
= max_blk_seq
;
420 zilog
->zl_parse_lr_seq
= max_lr_seq
;
421 zilog
->zl_parse_blk_count
= blk_count
;
422 zilog
->zl_parse_lr_count
= lr_count
;
424 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
425 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
) ||
426 (decrypt
&& error
== EIO
));
428 zil_bp_tree_fini(zilog
);
429 zio_buf_free(lrbuf
, SPA_OLD_MAXBLOCKSIZE
);
436 zil_clear_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
438 ASSERT(!BP_IS_HOLE(bp
));
441 * As we call this function from the context of a rewind to a
442 * checkpoint, each ZIL block whose txg is later than the txg
443 * that we rewind to is invalid. Thus, we return -1 so
444 * zil_parse() doesn't attempt to read it.
446 if (bp
->blk_birth
>= first_txg
)
449 if (zil_bp_tree_add(zilog
, bp
) != 0)
452 zio_free(zilog
->zl_spa
, first_txg
, bp
);
458 zil_noop_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
464 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
467 * Claim log block if not already committed and not already claimed.
468 * If tx == NULL, just verify that the block is claimable.
470 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
471 zil_bp_tree_add(zilog
, bp
) != 0)
474 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
475 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
476 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
480 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
482 lr_write_t
*lr
= (lr_write_t
*)lrc
;
485 if (lrc
->lrc_txtype
!= TX_WRITE
)
489 * If the block is not readable, don't claim it. This can happen
490 * in normal operation when a log block is written to disk before
491 * some of the dmu_sync() blocks it points to. In this case, the
492 * transaction cannot have been committed to anyone (we would have
493 * waited for all writes to be stable first), so it is semantically
494 * correct to declare this the end of the log.
496 if (lr
->lr_blkptr
.blk_birth
>= first_txg
) {
497 error
= zil_read_log_data(zilog
, lr
, NULL
);
502 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
507 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
509 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
515 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
517 lr_write_t
*lr
= (lr_write_t
*)lrc
;
518 blkptr_t
*bp
= &lr
->lr_blkptr
;
521 * If we previously claimed it, we need to free it.
523 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
524 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
526 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
532 zil_lwb_vdev_compare(const void *x1
, const void *x2
)
534 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
535 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
537 return (AVL_CMP(v1
, v2
));
541 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, boolean_t slog
, uint64_t txg
,
546 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
547 lwb
->lwb_zilog
= zilog
;
549 lwb
->lwb_fastwrite
= fastwrite
;
550 lwb
->lwb_slog
= slog
;
551 lwb
->lwb_state
= LWB_STATE_CLOSED
;
552 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
553 lwb
->lwb_max_txg
= txg
;
554 lwb
->lwb_write_zio
= NULL
;
555 lwb
->lwb_root_zio
= NULL
;
557 lwb
->lwb_issued_timestamp
= 0;
558 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
559 lwb
->lwb_nused
= sizeof (zil_chain_t
);
560 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
563 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
566 mutex_enter(&zilog
->zl_lock
);
567 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
568 mutex_exit(&zilog
->zl_lock
);
570 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
571 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
572 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
573 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
579 zil_free_lwb(zilog_t
*zilog
, lwb_t
*lwb
)
581 ASSERT(MUTEX_HELD(&zilog
->zl_lock
));
582 ASSERT(!MUTEX_HELD(&lwb
->lwb_vdev_lock
));
583 VERIFY(list_is_empty(&lwb
->lwb_waiters
));
584 VERIFY(list_is_empty(&lwb
->lwb_itxs
));
585 ASSERT(avl_is_empty(&lwb
->lwb_vdev_tree
));
586 ASSERT3P(lwb
->lwb_write_zio
, ==, NULL
);
587 ASSERT3P(lwb
->lwb_root_zio
, ==, NULL
);
588 ASSERT3U(lwb
->lwb_max_txg
, <=, spa_syncing_txg(zilog
->zl_spa
));
589 ASSERT(lwb
->lwb_state
== LWB_STATE_CLOSED
||
590 lwb
->lwb_state
== LWB_STATE_DONE
);
593 * Clear the zilog's field to indicate this lwb is no longer
594 * valid, and prevent use-after-free errors.
596 if (zilog
->zl_last_lwb_opened
== lwb
)
597 zilog
->zl_last_lwb_opened
= NULL
;
599 kmem_cache_free(zil_lwb_cache
, lwb
);
603 * Called when we create in-memory log transactions so that we know
604 * to cleanup the itxs at the end of spa_sync().
607 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
609 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
610 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
612 ASSERT(spa_writeable(zilog
->zl_spa
));
614 if (ds
->ds_is_snapshot
)
615 panic("dirtying snapshot!");
617 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
618 /* up the hold count until we can be written out */
619 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
621 zilog
->zl_dirty_max_txg
= MAX(txg
, zilog
->zl_dirty_max_txg
);
626 * Determine if the zil is dirty in the specified txg. Callers wanting to
627 * ensure that the dirty state does not change must hold the itxg_lock for
628 * the specified txg. Holding the lock will ensure that the zil cannot be
629 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
633 zilog_is_dirty_in_txg(zilog_t
*zilog
, uint64_t txg
)
635 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
637 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, txg
& TXG_MASK
))
643 * Determine if the zil is dirty. The zil is considered dirty if it has
644 * any pending itx records that have not been cleaned by zil_clean().
647 zilog_is_dirty(zilog_t
*zilog
)
649 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
651 for (int t
= 0; t
< TXG_SIZE
; t
++) {
652 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
659 * Create an on-disk intent log.
662 zil_create(zilog_t
*zilog
)
664 const zil_header_t
*zh
= zilog
->zl_header
;
670 boolean_t fastwrite
= FALSE
;
671 boolean_t slog
= FALSE
;
674 * Wait for any previous destroy to complete.
676 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
678 ASSERT(zh
->zh_claim_txg
== 0);
679 ASSERT(zh
->zh_replay_seq
== 0);
684 * Allocate an initial log block if:
685 * - there isn't one already
686 * - the existing block is the wrong endianness
688 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
689 tx
= dmu_tx_create(zilog
->zl_os
);
690 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
691 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
692 txg
= dmu_tx_get_txg(tx
);
694 if (!BP_IS_HOLE(&blk
)) {
695 zio_free(zilog
->zl_spa
, txg
, &blk
);
699 error
= zio_alloc_zil(zilog
->zl_spa
, zilog
->zl_os
, txg
, &blk
,
700 ZIL_MIN_BLKSZ
, &slog
);
704 zil_init_log_chain(zilog
, &blk
);
708 * Allocate a log write block (lwb) for the first log block.
711 lwb
= zil_alloc_lwb(zilog
, &blk
, slog
, txg
, fastwrite
);
714 * If we just allocated the first log block, commit our transaction
715 * and wait for zil_sync() to stuff the block pointer into zh_log.
716 * (zh is part of the MOS, so we cannot modify it in open context.)
720 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
723 ASSERT(error
!= 0 || bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
724 IMPLY(error
== 0, lwb
!= NULL
);
730 * In one tx, free all log blocks and clear the log header. If keep_first
731 * is set, then we're replaying a log with no content. We want to keep the
732 * first block, however, so that the first synchronous transaction doesn't
733 * require a txg_wait_synced() in zil_create(). We don't need to
734 * txg_wait_synced() here either when keep_first is set, because both
735 * zil_create() and zil_destroy() will wait for any in-progress destroys
739 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
741 const zil_header_t
*zh
= zilog
->zl_header
;
747 * Wait for any previous destroy to complete.
749 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
751 zilog
->zl_old_header
= *zh
; /* debugging aid */
753 if (BP_IS_HOLE(&zh
->zh_log
))
756 tx
= dmu_tx_create(zilog
->zl_os
);
757 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
758 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
759 txg
= dmu_tx_get_txg(tx
);
761 mutex_enter(&zilog
->zl_lock
);
763 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
764 zilog
->zl_destroy_txg
= txg
;
765 zilog
->zl_keep_first
= keep_first
;
767 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
768 ASSERT(zh
->zh_claim_txg
== 0);
770 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
771 if (lwb
->lwb_fastwrite
)
772 metaslab_fastwrite_unmark(zilog
->zl_spa
,
775 list_remove(&zilog
->zl_lwb_list
, lwb
);
776 if (lwb
->lwb_buf
!= NULL
)
777 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
778 zio_free(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
779 zil_free_lwb(zilog
, lwb
);
781 } else if (!keep_first
) {
782 zil_destroy_sync(zilog
, tx
);
784 mutex_exit(&zilog
->zl_lock
);
790 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
792 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
793 (void) zil_parse(zilog
, zil_free_log_block
,
794 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
, B_FALSE
);
798 zil_claim(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *txarg
)
800 dmu_tx_t
*tx
= txarg
;
807 error
= dmu_objset_own_obj(dp
, ds
->ds_object
,
808 DMU_OST_ANY
, B_FALSE
, B_FALSE
, FTAG
, &os
);
811 * EBUSY indicates that the objset is inconsistent, in which
812 * case it can not have a ZIL.
814 if (error
!= EBUSY
) {
815 cmn_err(CE_WARN
, "can't open objset for %llu, error %u",
816 (unsigned long long)ds
->ds_object
, error
);
822 zilog
= dmu_objset_zil(os
);
823 zh
= zil_header_in_syncing_context(zilog
);
824 ASSERT3U(tx
->tx_txg
, ==, spa_first_txg(zilog
->zl_spa
));
825 first_txg
= spa_min_claim_txg(zilog
->zl_spa
);
828 * If the spa_log_state is not set to be cleared, check whether
829 * the current uberblock is a checkpoint one and if the current
830 * header has been claimed before moving on.
832 * If the current uberblock is a checkpointed uberblock then
833 * one of the following scenarios took place:
835 * 1] We are currently rewinding to the checkpoint of the pool.
836 * 2] We crashed in the middle of a checkpoint rewind but we
837 * did manage to write the checkpointed uberblock to the
838 * vdev labels, so when we tried to import the pool again
839 * the checkpointed uberblock was selected from the import
842 * In both cases we want to zero out all the ZIL blocks, except
843 * the ones that have been claimed at the time of the checkpoint
844 * (their zh_claim_txg != 0). The reason is that these blocks
845 * may be corrupted since we may have reused their locations on
846 * disk after we took the checkpoint.
848 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
849 * when we first figure out whether the current uberblock is
850 * checkpointed or not. Unfortunately, that would discard all
851 * the logs, including the ones that are claimed, and we would
854 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
||
855 (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
856 zh
->zh_claim_txg
== 0)) {
857 if (!BP_IS_HOLE(&zh
->zh_log
)) {
858 (void) zil_parse(zilog
, zil_clear_log_block
,
859 zil_noop_log_record
, tx
, first_txg
, B_FALSE
);
861 BP_ZERO(&zh
->zh_log
);
862 if (os
->os_encrypted
)
863 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
864 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
865 dmu_objset_disown(os
, B_FALSE
, FTAG
);
870 * If we are not rewinding and opening the pool normally, then
871 * the min_claim_txg should be equal to the first txg of the pool.
873 ASSERT3U(first_txg
, ==, spa_first_txg(zilog
->zl_spa
));
876 * Claim all log blocks if we haven't already done so, and remember
877 * the highest claimed sequence number. This ensures that if we can
878 * read only part of the log now (e.g. due to a missing device),
879 * but we can read the entire log later, we will not try to replay
880 * or destroy beyond the last block we successfully claimed.
882 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
883 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
884 (void) zil_parse(zilog
, zil_claim_log_block
,
885 zil_claim_log_record
, tx
, first_txg
, B_FALSE
);
886 zh
->zh_claim_txg
= first_txg
;
887 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
888 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
889 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
890 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
891 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
892 if (os
->os_encrypted
)
893 os
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
894 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
897 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
898 dmu_objset_disown(os
, B_FALSE
, FTAG
);
903 * Check the log by walking the log chain.
904 * Checksum errors are ok as they indicate the end of the chain.
905 * Any other error (no device or read failure) returns an error.
909 zil_check_log_chain(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *tx
)
918 error
= dmu_objset_from_ds(ds
, &os
);
920 cmn_err(CE_WARN
, "can't open objset %llu, error %d",
921 (unsigned long long)ds
->ds_object
, error
);
925 zilog
= dmu_objset_zil(os
);
926 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
928 if (!BP_IS_HOLE(bp
)) {
930 boolean_t valid
= B_TRUE
;
933 * Check the first block and determine if it's on a log device
934 * which may have been removed or faulted prior to loading this
935 * pool. If so, there's no point in checking the rest of the
936 * log as its content should have already been synced to the
939 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
940 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
941 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
942 valid
= vdev_log_state_valid(vd
);
943 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
949 * Check whether the current uberblock is checkpointed (e.g.
950 * we are rewinding) and whether the current header has been
951 * claimed or not. If it hasn't then skip verifying it. We
952 * do this because its ZIL blocks may be part of the pool's
953 * state before the rewind, which is no longer valid.
955 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
956 if (zilog
->zl_spa
->spa_uberblock
.ub_checkpoint_txg
!= 0 &&
957 zh
->zh_claim_txg
== 0)
962 * Because tx == NULL, zil_claim_log_block() will not actually claim
963 * any blocks, but just determine whether it is possible to do so.
964 * In addition to checking the log chain, zil_claim_log_block()
965 * will invoke zio_claim() with a done func of spa_claim_notify(),
966 * which will update spa_max_claim_txg. See spa_load() for details.
968 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
969 zilog
->zl_header
->zh_claim_txg
? -1ULL :
970 spa_min_claim_txg(os
->os_spa
), B_FALSE
);
972 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
976 * When an itx is "skipped", this function is used to properly mark the
977 * waiter as "done, and signal any thread(s) waiting on it. An itx can
978 * be skipped (and not committed to an lwb) for a variety of reasons,
979 * one of them being that the itx was committed via spa_sync(), prior to
980 * it being committed to an lwb; this can happen if a thread calling
981 * zil_commit() is racing with spa_sync().
984 zil_commit_waiter_skip(zil_commit_waiter_t
*zcw
)
986 mutex_enter(&zcw
->zcw_lock
);
987 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
988 zcw
->zcw_done
= B_TRUE
;
989 cv_broadcast(&zcw
->zcw_cv
);
990 mutex_exit(&zcw
->zcw_lock
);
994 * This function is used when the given waiter is to be linked into an
995 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
996 * At this point, the waiter will no longer be referenced by the itx,
997 * and instead, will be referenced by the lwb.
1000 zil_commit_waiter_link_lwb(zil_commit_waiter_t
*zcw
, lwb_t
*lwb
)
1003 * The lwb_waiters field of the lwb is protected by the zilog's
1004 * zl_lock, thus it must be held when calling this function.
1006 ASSERT(MUTEX_HELD(&lwb
->lwb_zilog
->zl_lock
));
1008 mutex_enter(&zcw
->zcw_lock
);
1009 ASSERT(!list_link_active(&zcw
->zcw_node
));
1010 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1011 ASSERT3P(lwb
, !=, NULL
);
1012 ASSERT(lwb
->lwb_state
== LWB_STATE_OPENED
||
1013 lwb
->lwb_state
== LWB_STATE_ISSUED
);
1015 list_insert_tail(&lwb
->lwb_waiters
, zcw
);
1017 mutex_exit(&zcw
->zcw_lock
);
1021 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1022 * block, and the given waiter must be linked to the "nolwb waiters"
1023 * list inside of zil_process_commit_list().
1026 zil_commit_waiter_link_nolwb(zil_commit_waiter_t
*zcw
, list_t
*nolwb
)
1028 mutex_enter(&zcw
->zcw_lock
);
1029 ASSERT(!list_link_active(&zcw
->zcw_node
));
1030 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
1031 list_insert_tail(nolwb
, zcw
);
1032 mutex_exit(&zcw
->zcw_lock
);
1036 zil_lwb_add_block(lwb_t
*lwb
, const blkptr_t
*bp
)
1038 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1040 zil_vdev_node_t
*zv
, zvsearch
;
1041 int ndvas
= BP_GET_NDVAS(bp
);
1044 if (zfs_nocacheflush
)
1047 mutex_enter(&lwb
->lwb_vdev_lock
);
1048 for (i
= 0; i
< ndvas
; i
++) {
1049 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1050 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
1051 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
1052 zv
->zv_vdev
= zvsearch
.zv_vdev
;
1053 avl_insert(t
, zv
, where
);
1056 mutex_exit(&lwb
->lwb_vdev_lock
);
1060 zil_lwb_add_txg(lwb_t
*lwb
, uint64_t txg
)
1062 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1066 * This function is a called after all VDEVs associated with a given lwb
1067 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1068 * as the lwb write completes, if "zfs_nocacheflush" is set.
1070 * The intention is for this function to be called as soon as the
1071 * contents of an lwb are considered "stable" on disk, and will survive
1072 * any sudden loss of power. At this point, any threads waiting for the
1073 * lwb to reach this state are signalled, and the "waiter" structures
1074 * are marked "done".
1077 zil_lwb_flush_vdevs_done(zio_t
*zio
)
1079 lwb_t
*lwb
= zio
->io_private
;
1080 zilog_t
*zilog
= lwb
->lwb_zilog
;
1081 dmu_tx_t
*tx
= lwb
->lwb_tx
;
1082 zil_commit_waiter_t
*zcw
;
1085 spa_config_exit(zilog
->zl_spa
, SCL_STATE
, lwb
);
1087 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1089 mutex_enter(&zilog
->zl_lock
);
1092 * Ensure the lwb buffer pointer is cleared before releasing the
1093 * txg. If we have had an allocation failure and the txg is
1094 * waiting to sync then we want zil_sync() to remove the lwb so
1095 * that it's not picked up as the next new one in
1096 * zil_process_commit_list(). zil_sync() will only remove the
1097 * lwb if lwb_buf is null.
1099 lwb
->lwb_buf
= NULL
;
1102 ASSERT3U(lwb
->lwb_issued_timestamp
, >, 0);
1103 zilog
->zl_last_lwb_latency
= gethrtime() - lwb
->lwb_issued_timestamp
;
1105 lwb
->lwb_root_zio
= NULL
;
1106 lwb
->lwb_state
= LWB_STATE_DONE
;
1108 if (zilog
->zl_last_lwb_opened
== lwb
) {
1110 * Remember the highest committed log sequence number
1111 * for ztest. We only update this value when all the log
1112 * writes succeeded, because ztest wants to ASSERT that
1113 * it got the whole log chain.
1115 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1118 while ((itx
= list_head(&lwb
->lwb_itxs
)) != NULL
) {
1119 list_remove(&lwb
->lwb_itxs
, itx
);
1120 zil_itx_destroy(itx
);
1123 while ((zcw
= list_head(&lwb
->lwb_waiters
)) != NULL
) {
1124 mutex_enter(&zcw
->zcw_lock
);
1126 ASSERT(list_link_active(&zcw
->zcw_node
));
1127 list_remove(&lwb
->lwb_waiters
, zcw
);
1129 ASSERT3P(zcw
->zcw_lwb
, ==, lwb
);
1130 zcw
->zcw_lwb
= NULL
;
1132 zcw
->zcw_zio_error
= zio
->io_error
;
1134 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
1135 zcw
->zcw_done
= B_TRUE
;
1136 cv_broadcast(&zcw
->zcw_cv
);
1138 mutex_exit(&zcw
->zcw_lock
);
1141 mutex_exit(&zilog
->zl_lock
);
1144 * Now that we've written this log block, we have a stable pointer
1145 * to the next block in the chain, so it's OK to let the txg in
1146 * which we allocated the next block sync.
1152 * This is called when an lwb write completes. This means, this specific
1153 * lwb was written to disk, and all dependent lwb have also been
1156 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to
1157 * the VDEVs involved in writing out this specific lwb. The lwb will be
1158 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the
1159 * zio completion callback for the lwb's root zio.
1162 zil_lwb_write_done(zio_t
*zio
)
1164 lwb_t
*lwb
= zio
->io_private
;
1165 spa_t
*spa
= zio
->io_spa
;
1166 zilog_t
*zilog
= lwb
->lwb_zilog
;
1167 avl_tree_t
*t
= &lwb
->lwb_vdev_tree
;
1168 void *cookie
= NULL
;
1169 zil_vdev_node_t
*zv
;
1171 ASSERT3S(spa_config_held(spa
, SCL_STATE
, RW_READER
), !=, 0);
1173 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1174 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
1175 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
1176 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
1177 ASSERT(!BP_IS_GANG(zio
->io_bp
));
1178 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
1179 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
1181 abd_put(zio
->io_abd
);
1183 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_ISSUED
);
1185 mutex_enter(&zilog
->zl_lock
);
1186 lwb
->lwb_write_zio
= NULL
;
1187 lwb
->lwb_fastwrite
= FALSE
;
1188 mutex_exit(&zilog
->zl_lock
);
1190 if (avl_numnodes(t
) == 0)
1194 * If there was an IO error, we're not going to call zio_flush()
1195 * on these vdevs, so we simply empty the tree and free the
1196 * nodes. We avoid calling zio_flush() since there isn't any
1197 * good reason for doing so, after the lwb block failed to be
1200 if (zio
->io_error
!= 0) {
1201 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
1202 kmem_free(zv
, sizeof (*zv
));
1206 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1207 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
1209 zio_flush(lwb
->lwb_root_zio
, vd
);
1210 kmem_free(zv
, sizeof (*zv
));
1215 * This function's purpose is to "open" an lwb such that it is ready to
1216 * accept new itxs being committed to it. To do this, the lwb's zio
1217 * structures are created, and linked to the lwb. This function is
1218 * idempotent; if the passed in lwb has already been opened, this
1219 * function is essentially a no-op.
1222 zil_lwb_write_open(zilog_t
*zilog
, lwb_t
*lwb
)
1224 zbookmark_phys_t zb
;
1225 zio_priority_t prio
;
1227 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1228 ASSERT3P(lwb
, !=, NULL
);
1229 EQUIV(lwb
->lwb_root_zio
== NULL
, lwb
->lwb_state
== LWB_STATE_CLOSED
);
1230 EQUIV(lwb
->lwb_root_zio
!= NULL
, lwb
->lwb_state
== LWB_STATE_OPENED
);
1232 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
1233 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
1234 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
1236 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1237 mutex_enter(&zilog
->zl_lock
);
1238 if (lwb
->lwb_root_zio
== NULL
) {
1239 abd_t
*lwb_abd
= abd_get_from_buf(lwb
->lwb_buf
,
1240 BP_GET_LSIZE(&lwb
->lwb_blk
));
1242 if (!lwb
->lwb_fastwrite
) {
1243 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1244 lwb
->lwb_fastwrite
= 1;
1247 if (!lwb
->lwb_slog
|| zilog
->zl_cur_used
<= zil_slog_bulk
)
1248 prio
= ZIO_PRIORITY_SYNC_WRITE
;
1250 prio
= ZIO_PRIORITY_ASYNC_WRITE
;
1252 lwb
->lwb_root_zio
= zio_root(zilog
->zl_spa
,
1253 zil_lwb_flush_vdevs_done
, lwb
, ZIO_FLAG_CANFAIL
);
1254 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1256 lwb
->lwb_write_zio
= zio_rewrite(lwb
->lwb_root_zio
,
1257 zilog
->zl_spa
, 0, &lwb
->lwb_blk
, lwb_abd
,
1258 BP_GET_LSIZE(&lwb
->lwb_blk
), zil_lwb_write_done
, lwb
,
1259 prio
, ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
1260 ZIO_FLAG_FASTWRITE
, &zb
);
1261 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1263 lwb
->lwb_state
= LWB_STATE_OPENED
;
1266 * The zilog's "zl_last_lwb_opened" field is used to
1267 * build the lwb/zio dependency chain, which is used to
1268 * preserve the ordering of lwb completions that is
1269 * required by the semantics of the ZIL. Each new lwb
1270 * zio becomes a parent of the "previous" lwb zio, such
1271 * that the new lwb's zio cannot complete until the
1272 * "previous" lwb's zio completes.
1274 * This is required by the semantics of zil_commit();
1275 * the commit waiters attached to the lwbs will be woken
1276 * in the lwb zio's completion callback, so this zio
1277 * dependency graph ensures the waiters are woken in the
1278 * correct order (the same order the lwbs were created).
1280 lwb_t
*last_lwb_opened
= zilog
->zl_last_lwb_opened
;
1281 if (last_lwb_opened
!= NULL
&&
1282 last_lwb_opened
->lwb_state
!= LWB_STATE_DONE
) {
1283 ASSERT(last_lwb_opened
->lwb_state
== LWB_STATE_OPENED
||
1284 last_lwb_opened
->lwb_state
== LWB_STATE_ISSUED
);
1285 ASSERT3P(last_lwb_opened
->lwb_root_zio
, !=, NULL
);
1286 zio_add_child(lwb
->lwb_root_zio
,
1287 last_lwb_opened
->lwb_root_zio
);
1289 zilog
->zl_last_lwb_opened
= lwb
;
1291 mutex_exit(&zilog
->zl_lock
);
1293 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1294 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1295 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1299 * Define a limited set of intent log block sizes.
1301 * These must be a multiple of 4KB. Note only the amount used (again
1302 * aligned to 4KB) actually gets written. However, we can't always just
1303 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1305 uint64_t zil_block_buckets
[] = {
1306 4096, /* non TX_WRITE */
1307 8192+4096, /* data base */
1308 32*1024 + 4096, /* NFS writes */
1313 * Start a log block write and advance to the next log block.
1314 * Calls are serialized.
1317 zil_lwb_write_issue(zilog_t
*zilog
, lwb_t
*lwb
)
1321 spa_t
*spa
= zilog
->zl_spa
;
1325 uint64_t zil_blksz
, wsz
;
1329 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1330 ASSERT3P(lwb
->lwb_root_zio
, !=, NULL
);
1331 ASSERT3P(lwb
->lwb_write_zio
, !=, NULL
);
1332 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
1334 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1335 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
1336 bp
= &zilc
->zc_next_blk
;
1338 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
1339 bp
= &zilc
->zc_next_blk
;
1342 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
1345 * Allocate the next block and save its address in this block
1346 * before writing it in order to establish the log chain.
1347 * Note that if the allocation of nlwb synced before we wrote
1348 * the block that points at it (lwb), we'd leak it if we crashed.
1349 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
1350 * We dirty the dataset to ensure that zil_sync() will be called
1351 * to clean up in the event of allocation failure or I/O failure.
1354 tx
= dmu_tx_create(zilog
->zl_os
);
1357 * Since we are not going to create any new dirty data, and we
1358 * can even help with clearing the existing dirty data, we
1359 * should not be subject to the dirty data based delays. We
1360 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1362 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
| TXG_NOTHROTTLE
));
1364 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1365 txg
= dmu_tx_get_txg(tx
);
1370 * Log blocks are pre-allocated. Here we select the size of the next
1371 * block, based on size used in the last block.
1372 * - first find the smallest bucket that will fit the block from a
1373 * limited set of block sizes. This is because it's faster to write
1374 * blocks allocated from the same metaslab as they are adjacent or
1376 * - next find the maximum from the new suggested size and an array of
1377 * previous sizes. This lessens a picket fence effect of wrongly
1378 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1381 * Note we only write what is used, but we can't just allocate
1382 * the maximum block size because we can exhaust the available
1385 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1386 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1388 zil_blksz
= zil_block_buckets
[i
];
1389 if (zil_blksz
== UINT64_MAX
)
1390 zil_blksz
= SPA_OLD_MAXBLOCKSIZE
;
1391 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1392 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1393 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1394 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1397 error
= zio_alloc_zil(spa
, zilog
->zl_os
, txg
, bp
, zil_blksz
, &slog
);
1399 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1400 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1402 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1403 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1406 ASSERT3U(bp
->blk_birth
, ==, txg
);
1407 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1408 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1411 * Allocate a new log write block (lwb).
1413 nlwb
= zil_alloc_lwb(zilog
, bp
, slog
, txg
, TRUE
);
1416 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1417 /* For Slim ZIL only write what is used. */
1418 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1419 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1420 zio_shrink(lwb
->lwb_write_zio
, wsz
);
1427 zilc
->zc_nused
= lwb
->lwb_nused
;
1428 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1431 * clear unused data for security
1433 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1435 spa_config_enter(zilog
->zl_spa
, SCL_STATE
, lwb
, RW_READER
);
1437 zil_lwb_add_block(lwb
, &lwb
->lwb_blk
);
1438 lwb
->lwb_issued_timestamp
= gethrtime();
1439 lwb
->lwb_state
= LWB_STATE_ISSUED
;
1441 zio_nowait(lwb
->lwb_root_zio
);
1442 zio_nowait(lwb
->lwb_write_zio
);
1445 * If there was an allocation failure then nlwb will be null which
1446 * forces a txg_wait_synced().
1452 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1455 lr_write_t
*lrwb
, *lrw
;
1457 uint64_t dlen
, dnow
, lwb_sp
, reclen
, txg
;
1459 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1460 ASSERT3P(lwb
, !=, NULL
);
1461 ASSERT3P(lwb
->lwb_buf
, !=, NULL
);
1463 zil_lwb_write_open(zilog
, lwb
);
1466 lrw
= (lr_write_t
*)lrc
;
1469 * A commit itx doesn't represent any on-disk state; instead
1470 * it's simply used as a place holder on the commit list, and
1471 * provides a mechanism for attaching a "commit waiter" onto the
1472 * correct lwb (such that the waiter can be signalled upon
1473 * completion of that lwb). Thus, we don't process this itx's
1474 * log record if it's a commit itx (these itx's don't have log
1475 * records), and instead link the itx's waiter onto the lwb's
1478 * For more details, see the comment above zil_commit().
1480 if (lrc
->lrc_txtype
== TX_COMMIT
) {
1481 mutex_enter(&zilog
->zl_lock
);
1482 zil_commit_waiter_link_lwb(itx
->itx_private
, lwb
);
1483 itx
->itx_private
= NULL
;
1484 mutex_exit(&zilog
->zl_lock
);
1488 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
) {
1489 dlen
= P2ROUNDUP_TYPED(
1490 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1494 reclen
= lrc
->lrc_reclen
;
1495 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1498 ASSERT3U(zilog
->zl_cur_used
, <, UINT64_MAX
- (reclen
+ dlen
));
1502 * If this record won't fit in the current log block, start a new one.
1503 * For WR_NEED_COPY optimize layout for minimal number of chunks.
1505 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1506 if (reclen
> lwb_sp
|| (reclen
+ dlen
> lwb_sp
&&
1507 lwb_sp
< ZIL_MAX_WASTE_SPACE
&& (dlen
% ZIL_MAX_LOG_DATA
== 0 ||
1508 lwb_sp
< reclen
+ dlen
% ZIL_MAX_LOG_DATA
))) {
1509 lwb
= zil_lwb_write_issue(zilog
, lwb
);
1512 zil_lwb_write_open(zilog
, lwb
);
1513 ASSERT(LWB_EMPTY(lwb
));
1514 lwb_sp
= lwb
->lwb_sz
- lwb
->lwb_nused
;
1515 ASSERT3U(reclen
+ MIN(dlen
, sizeof (uint64_t)), <=, lwb_sp
);
1518 dnow
= MIN(dlen
, lwb_sp
- reclen
);
1519 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1520 bcopy(lrc
, lr_buf
, reclen
);
1521 lrcb
= (lr_t
*)lr_buf
; /* Like lrc, but inside lwb. */
1522 lrwb
= (lr_write_t
*)lrcb
; /* Like lrw, but inside lwb. */
1524 ZIL_STAT_BUMP(zil_itx_count
);
1527 * If it's a write, fetch the data or get its blkptr as appropriate.
1529 if (lrc
->lrc_txtype
== TX_WRITE
) {
1530 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1531 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1532 if (itx
->itx_wr_state
== WR_COPIED
) {
1533 ZIL_STAT_BUMP(zil_itx_copied_count
);
1534 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1539 if (itx
->itx_wr_state
== WR_NEED_COPY
) {
1540 dbuf
= lr_buf
+ reclen
;
1541 lrcb
->lrc_reclen
+= dnow
;
1542 if (lrwb
->lr_length
> dnow
)
1543 lrwb
->lr_length
= dnow
;
1544 lrw
->lr_offset
+= dnow
;
1545 lrw
->lr_length
-= dnow
;
1546 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1547 ZIL_STAT_INCR(zil_itx_needcopy_bytes
, dnow
);
1549 ASSERT3S(itx
->itx_wr_state
, ==, WR_INDIRECT
);
1551 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1552 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1557 * We pass in the "lwb_write_zio" rather than
1558 * "lwb_root_zio" so that the "lwb_write_zio"
1559 * becomes the parent of any zio's created by
1560 * the "zl_get_data" callback. The vdevs are
1561 * flushed after the "lwb_write_zio" completes,
1562 * so we want to make sure that completion
1563 * callback waits for these additional zio's,
1564 * such that the vdevs used by those zio's will
1565 * be included in the lwb's vdev tree, and those
1566 * vdevs will be properly flushed. If we passed
1567 * in "lwb_root_zio" here, then these additional
1568 * vdevs may not be flushed; e.g. if these zio's
1569 * completed after "lwb_write_zio" completed.
1571 error
= zilog
->zl_get_data(itx
->itx_private
,
1572 lrwb
, dbuf
, lwb
, lwb
->lwb_write_zio
);
1575 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1579 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1587 * We're actually making an entry, so update lrc_seq to be the
1588 * log record sequence number. Note that this is generally not
1589 * equal to the itx sequence number because not all transactions
1590 * are synchronous, and sometimes spa_sync() gets there first.
1592 lrcb
->lrc_seq
= ++zilog
->zl_lr_seq
;
1593 lwb
->lwb_nused
+= reclen
+ dnow
;
1595 zil_lwb_add_txg(lwb
, txg
);
1597 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1598 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1602 zilog
->zl_cur_used
+= reclen
;
1610 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1615 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1616 itxsize
= offsetof(itx_t
, itx_lr
) + lrsize
;
1618 itx
= zio_data_buf_alloc(itxsize
);
1619 itx
->itx_lr
.lrc_txtype
= txtype
;
1620 itx
->itx_lr
.lrc_reclen
= lrsize
;
1621 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1622 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1623 itx
->itx_callback
= NULL
;
1624 itx
->itx_callback_data
= NULL
;
1625 itx
->itx_size
= itxsize
;
1631 zil_itx_destroy(itx_t
*itx
)
1633 IMPLY(itx
->itx_lr
.lrc_txtype
== TX_COMMIT
, itx
->itx_callback
== NULL
);
1634 IMPLY(itx
->itx_callback
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
1636 if (itx
->itx_callback
!= NULL
)
1637 itx
->itx_callback(itx
->itx_callback_data
);
1639 zio_data_buf_free(itx
, itx
->itx_size
);
1643 * Free up the sync and async itxs. The itxs_t has already been detached
1644 * so no locks are needed.
1647 zil_itxg_clean(itxs_t
*itxs
)
1653 itx_async_node_t
*ian
;
1655 list
= &itxs
->i_sync_list
;
1656 while ((itx
= list_head(list
)) != NULL
) {
1658 * In the general case, commit itxs will not be found
1659 * here, as they'll be committed to an lwb via
1660 * zil_lwb_commit(), and free'd in that function. Having
1661 * said that, it is still possible for commit itxs to be
1662 * found here, due to the following race:
1664 * - a thread calls zil_commit() which assigns the
1665 * commit itx to a per-txg i_sync_list
1666 * - zil_itxg_clean() is called (e.g. via spa_sync())
1667 * while the waiter is still on the i_sync_list
1669 * There's nothing to prevent syncing the txg while the
1670 * waiter is on the i_sync_list. This normally doesn't
1671 * happen because spa_sync() is slower than zil_commit(),
1672 * but if zil_commit() calls txg_wait_synced() (e.g.
1673 * because zil_create() or zil_commit_writer_stall() is
1674 * called) we will hit this case.
1676 if (itx
->itx_lr
.lrc_txtype
== TX_COMMIT
)
1677 zil_commit_waiter_skip(itx
->itx_private
);
1679 list_remove(list
, itx
);
1680 zil_itx_destroy(itx
);
1684 t
= &itxs
->i_async_tree
;
1685 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1686 list
= &ian
->ia_list
;
1687 while ((itx
= list_head(list
)) != NULL
) {
1688 list_remove(list
, itx
);
1689 /* commit itxs should never be on the async lists. */
1690 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
1691 zil_itx_destroy(itx
);
1694 kmem_free(ian
, sizeof (itx_async_node_t
));
1698 kmem_free(itxs
, sizeof (itxs_t
));
1702 zil_aitx_compare(const void *x1
, const void *x2
)
1704 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1705 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1707 return (AVL_CMP(o1
, o2
));
1711 * Remove all async itx with the given oid.
1714 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1717 itx_async_node_t
*ian
;
1724 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1726 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1729 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1731 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1732 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1734 mutex_enter(&itxg
->itxg_lock
);
1735 if (itxg
->itxg_txg
!= txg
) {
1736 mutex_exit(&itxg
->itxg_lock
);
1741 * Locate the object node and append its list.
1743 t
= &itxg
->itxg_itxs
->i_async_tree
;
1744 ian
= avl_find(t
, &oid
, &where
);
1746 list_move_tail(&clean_list
, &ian
->ia_list
);
1747 mutex_exit(&itxg
->itxg_lock
);
1749 while ((itx
= list_head(&clean_list
)) != NULL
) {
1750 list_remove(&clean_list
, itx
);
1751 /* commit itxs should never be on the async lists. */
1752 ASSERT3U(itx
->itx_lr
.lrc_txtype
, !=, TX_COMMIT
);
1753 zil_itx_destroy(itx
);
1755 list_destroy(&clean_list
);
1759 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1763 itxs_t
*itxs
, *clean
= NULL
;
1766 * Object ids can be re-instantiated in the next txg so
1767 * remove any async transactions to avoid future leaks.
1768 * This can happen if a fsync occurs on the re-instantiated
1769 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1770 * the new file data and flushes a write record for the old object.
1772 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1773 zil_remove_async(zilog
, itx
->itx_oid
);
1776 * Ensure the data of a renamed file is committed before the rename.
1778 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1779 zil_async_to_sync(zilog
, itx
->itx_oid
);
1781 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1784 txg
= dmu_tx_get_txg(tx
);
1786 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1787 mutex_enter(&itxg
->itxg_lock
);
1788 itxs
= itxg
->itxg_itxs
;
1789 if (itxg
->itxg_txg
!= txg
) {
1792 * The zil_clean callback hasn't got around to cleaning
1793 * this itxg. Save the itxs for release below.
1794 * This should be rare.
1796 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
1797 "txg %llu", itxg
->itxg_txg
);
1798 clean
= itxg
->itxg_itxs
;
1800 itxg
->itxg_txg
= txg
;
1801 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1804 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1805 offsetof(itx_t
, itx_node
));
1806 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1807 sizeof (itx_async_node_t
),
1808 offsetof(itx_async_node_t
, ia_node
));
1810 if (itx
->itx_sync
) {
1811 list_insert_tail(&itxs
->i_sync_list
, itx
);
1813 avl_tree_t
*t
= &itxs
->i_async_tree
;
1815 LR_FOID_GET_OBJ(((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
);
1816 itx_async_node_t
*ian
;
1819 ian
= avl_find(t
, &foid
, &where
);
1821 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1823 list_create(&ian
->ia_list
, sizeof (itx_t
),
1824 offsetof(itx_t
, itx_node
));
1825 ian
->ia_foid
= foid
;
1826 avl_insert(t
, ian
, where
);
1828 list_insert_tail(&ian
->ia_list
, itx
);
1831 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1834 * We don't want to dirty the ZIL using ZILTEST_TXG, because
1835 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
1836 * need to be careful to always dirty the ZIL using the "real"
1837 * TXG (not itxg_txg) even when the SPA is frozen.
1839 zilog_dirty(zilog
, dmu_tx_get_txg(tx
));
1840 mutex_exit(&itxg
->itxg_lock
);
1842 /* Release the old itxs now we've dropped the lock */
1844 zil_itxg_clean(clean
);
1848 * If there are any in-memory intent log transactions which have now been
1849 * synced then start up a taskq to free them. We should only do this after we
1850 * have written out the uberblocks (i.e. txg has been comitted) so that
1851 * don't inadvertently clean out in-memory log records that would be required
1855 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1857 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1860 ASSERT3U(synced_txg
, <, ZILTEST_TXG
);
1862 mutex_enter(&itxg
->itxg_lock
);
1863 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1864 mutex_exit(&itxg
->itxg_lock
);
1867 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1868 ASSERT3U(itxg
->itxg_txg
, !=, 0);
1869 clean_me
= itxg
->itxg_itxs
;
1870 itxg
->itxg_itxs
= NULL
;
1872 mutex_exit(&itxg
->itxg_lock
);
1874 * Preferably start a task queue to free up the old itxs but
1875 * if taskq_dispatch can't allocate resources to do that then
1876 * free it in-line. This should be rare. Note, using TQ_SLEEP
1877 * created a bad performance problem.
1879 ASSERT3P(zilog
->zl_dmu_pool
, !=, NULL
);
1880 ASSERT3P(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
, !=, NULL
);
1881 taskqid_t id
= taskq_dispatch(zilog
->zl_dmu_pool
->dp_zil_clean_taskq
,
1882 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
);
1883 if (id
== TASKQID_INVALID
)
1884 zil_itxg_clean(clean_me
);
1888 * This function will traverse the queue of itxs that need to be
1889 * committed, and move them onto the ZIL's zl_itx_commit_list.
1892 zil_get_commit_list(zilog_t
*zilog
)
1895 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1897 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
1899 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1902 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1905 * This is inherently racy, since there is nothing to prevent
1906 * the last synced txg from changing. That's okay since we'll
1907 * only commit things in the future.
1909 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1910 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1912 mutex_enter(&itxg
->itxg_lock
);
1913 if (itxg
->itxg_txg
!= txg
) {
1914 mutex_exit(&itxg
->itxg_lock
);
1919 * If we're adding itx records to the zl_itx_commit_list,
1920 * then the zil better be dirty in this "txg". We can assert
1921 * that here since we're holding the itxg_lock which will
1922 * prevent spa_sync from cleaning it. Once we add the itxs
1923 * to the zl_itx_commit_list we must commit it to disk even
1924 * if it's unnecessary (i.e. the txg was synced).
1926 ASSERT(zilog_is_dirty_in_txg(zilog
, txg
) ||
1927 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1928 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1930 mutex_exit(&itxg
->itxg_lock
);
1935 * Move the async itxs for a specified object to commit into sync lists.
1938 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1941 itx_async_node_t
*ian
;
1945 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1948 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1951 * This is inherently racy, since there is nothing to prevent
1952 * the last synced txg from changing.
1954 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1955 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1957 mutex_enter(&itxg
->itxg_lock
);
1958 if (itxg
->itxg_txg
!= txg
) {
1959 mutex_exit(&itxg
->itxg_lock
);
1964 * If a foid is specified then find that node and append its
1965 * list. Otherwise walk the tree appending all the lists
1966 * to the sync list. We add to the end rather than the
1967 * beginning to ensure the create has happened.
1969 t
= &itxg
->itxg_itxs
->i_async_tree
;
1971 ian
= avl_find(t
, &foid
, &where
);
1973 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1977 void *cookie
= NULL
;
1979 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1980 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1982 list_destroy(&ian
->ia_list
);
1983 kmem_free(ian
, sizeof (itx_async_node_t
));
1986 mutex_exit(&itxg
->itxg_lock
);
1991 * This function will prune commit itxs that are at the head of the
1992 * commit list (it won't prune past the first non-commit itx), and
1993 * either: a) attach them to the last lwb that's still pending
1994 * completion, or b) skip them altogether.
1996 * This is used as a performance optimization to prevent commit itxs
1997 * from generating new lwbs when it's unnecessary to do so.
2000 zil_prune_commit_list(zilog_t
*zilog
)
2004 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2006 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2007 lr_t
*lrc
= &itx
->itx_lr
;
2008 if (lrc
->lrc_txtype
!= TX_COMMIT
)
2011 mutex_enter(&zilog
->zl_lock
);
2013 lwb_t
*last_lwb
= zilog
->zl_last_lwb_opened
;
2014 if (last_lwb
== NULL
|| last_lwb
->lwb_state
== LWB_STATE_DONE
) {
2016 * All of the itxs this waiter was waiting on
2017 * must have already completed (or there were
2018 * never any itx's for it to wait on), so it's
2019 * safe to skip this waiter and mark it done.
2021 zil_commit_waiter_skip(itx
->itx_private
);
2023 zil_commit_waiter_link_lwb(itx
->itx_private
, last_lwb
);
2024 itx
->itx_private
= NULL
;
2027 mutex_exit(&zilog
->zl_lock
);
2029 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2030 zil_itx_destroy(itx
);
2033 IMPLY(itx
!= NULL
, itx
->itx_lr
.lrc_txtype
!= TX_COMMIT
);
2037 zil_commit_writer_stall(zilog_t
*zilog
)
2040 * When zio_alloc_zil() fails to allocate the next lwb block on
2041 * disk, we must call txg_wait_synced() to ensure all of the
2042 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2043 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2044 * to zil_process_commit_list()) will have to call zil_create(),
2045 * and start a new ZIL chain.
2047 * Since zil_alloc_zil() failed, the lwb that was previously
2048 * issued does not have a pointer to the "next" lwb on disk.
2049 * Thus, if another ZIL writer thread was to allocate the "next"
2050 * on-disk lwb, that block could be leaked in the event of a
2051 * crash (because the previous lwb on-disk would not point to
2054 * We must hold the zilog's zl_issuer_lock while we do this, to
2055 * ensure no new threads enter zil_process_commit_list() until
2056 * all lwb's in the zl_lwb_list have been synced and freed
2057 * (which is achieved via the txg_wait_synced() call).
2059 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2060 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2061 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
2065 * This function will traverse the commit list, creating new lwbs as
2066 * needed, and committing the itxs from the commit list to these newly
2067 * created lwbs. Additionally, as a new lwb is created, the previous
2068 * lwb will be issued to the zio layer to be written to disk.
2071 zil_process_commit_list(zilog_t
*zilog
)
2073 spa_t
*spa
= zilog
->zl_spa
;
2075 list_t nolwb_waiters
;
2079 ASSERT(MUTEX_HELD(&zilog
->zl_issuer_lock
));
2082 * Return if there's nothing to commit before we dirty the fs by
2083 * calling zil_create().
2085 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
)
2088 list_create(&nolwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2089 list_create(&nolwb_waiters
, sizeof (zil_commit_waiter_t
),
2090 offsetof(zil_commit_waiter_t
, zcw_node
));
2092 lwb
= list_tail(&zilog
->zl_lwb_list
);
2094 lwb
= zil_create(zilog
);
2096 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2097 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_DONE
);
2100 while ((itx
= list_head(&zilog
->zl_itx_commit_list
)) != NULL
) {
2101 lr_t
*lrc
= &itx
->itx_lr
;
2102 uint64_t txg
= lrc
->lrc_txg
;
2104 ASSERT3U(txg
, !=, 0);
2106 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2107 DTRACE_PROBE2(zil__process__commit__itx
,
2108 zilog_t
*, zilog
, itx_t
*, itx
);
2110 DTRACE_PROBE2(zil__process__normal__itx
,
2111 zilog_t
*, zilog
, itx_t
*, itx
);
2114 list_remove(&zilog
->zl_itx_commit_list
, itx
);
2116 boolean_t synced
= txg
<= spa_last_synced_txg(spa
);
2117 boolean_t frozen
= txg
> spa_freeze_txg(spa
);
2120 * If the txg of this itx has already been synced out, then
2121 * we don't need to commit this itx to an lwb. This is
2122 * because the data of this itx will have already been
2123 * written to the main pool. This is inherently racy, and
2124 * it's still ok to commit an itx whose txg has already
2125 * been synced; this will result in a write that's
2126 * unnecessary, but will do no harm.
2128 * With that said, we always want to commit TX_COMMIT itxs
2129 * to an lwb, regardless of whether or not that itx's txg
2130 * has been synced out. We do this to ensure any OPENED lwb
2131 * will always have at least one zil_commit_waiter_t linked
2134 * As a counter-example, if we skipped TX_COMMIT itx's
2135 * whose txg had already been synced, the following
2136 * situation could occur if we happened to be racing with
2139 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2140 * itx's txg is 10 and the last synced txg is 9.
2141 * 2. spa_sync finishes syncing out txg 10.
2142 * 3. We move to the next itx in the list, it's a TX_COMMIT
2143 * whose txg is 10, so we skip it rather than committing
2144 * it to the lwb used in (1).
2146 * If the itx that is skipped in (3) is the last TX_COMMIT
2147 * itx in the commit list, than it's possible for the lwb
2148 * used in (1) to remain in the OPENED state indefinitely.
2150 * To prevent the above scenario from occurring, ensuring
2151 * that once an lwb is OPENED it will transition to ISSUED
2152 * and eventually DONE, we always commit TX_COMMIT itx's to
2153 * an lwb here, even if that itx's txg has already been
2156 * Finally, if the pool is frozen, we _always_ commit the
2157 * itx. The point of freezing the pool is to prevent data
2158 * from being written to the main pool via spa_sync, and
2159 * instead rely solely on the ZIL to persistently store the
2160 * data; i.e. when the pool is frozen, the last synced txg
2161 * value can't be trusted.
2163 if (frozen
|| !synced
|| lrc
->lrc_txtype
== TX_COMMIT
) {
2165 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
2168 list_insert_tail(&nolwb_itxs
, itx
);
2170 list_insert_tail(&lwb
->lwb_itxs
, itx
);
2172 if (lrc
->lrc_txtype
== TX_COMMIT
) {
2173 zil_commit_waiter_link_nolwb(
2174 itx
->itx_private
, &nolwb_waiters
);
2177 list_insert_tail(&nolwb_itxs
, itx
);
2180 ASSERT3S(lrc
->lrc_txtype
, !=, TX_COMMIT
);
2181 zil_itx_destroy(itx
);
2187 * This indicates zio_alloc_zil() failed to allocate the
2188 * "next" lwb on-disk. When this happens, we must stall
2189 * the ZIL write pipeline; see the comment within
2190 * zil_commit_writer_stall() for more details.
2192 zil_commit_writer_stall(zilog
);
2195 * Additionally, we have to signal and mark the "nolwb"
2196 * waiters as "done" here, since without an lwb, we
2197 * can't do this via zil_lwb_flush_vdevs_done() like
2200 zil_commit_waiter_t
*zcw
;
2201 while ((zcw
= list_head(&nolwb_waiters
)) != NULL
) {
2202 zil_commit_waiter_skip(zcw
);
2203 list_remove(&nolwb_waiters
, zcw
);
2207 * And finally, we have to destroy the itx's that
2208 * couldn't be committed to an lwb; this will also call
2209 * the itx's callback if one exists for the itx.
2211 while ((itx
= list_head(&nolwb_itxs
)) != NULL
) {
2212 list_remove(&nolwb_itxs
, itx
);
2213 zil_itx_destroy(itx
);
2216 ASSERT(list_is_empty(&nolwb_waiters
));
2217 ASSERT3P(lwb
, !=, NULL
);
2218 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
2219 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_DONE
);
2222 * At this point, the ZIL block pointed at by the "lwb"
2223 * variable is in one of the following states: "closed"
2226 * If it's "closed", then no itxs have been committed to
2227 * it, so there's no point in issuing its zio (i.e. it's
2230 * If it's "open", then it contains one or more itxs that
2231 * eventually need to be committed to stable storage. In
2232 * this case we intentionally do not issue the lwb's zio
2233 * to disk yet, and instead rely on one of the following
2234 * two mechanisms for issuing the zio:
2236 * 1. Ideally, there will be more ZIL activity occurring
2237 * on the system, such that this function will be
2238 * immediately called again (not necessarily by the same
2239 * thread) and this lwb's zio will be issued via
2240 * zil_lwb_commit(). This way, the lwb is guaranteed to
2241 * be "full" when it is issued to disk, and we'll make
2242 * use of the lwb's size the best we can.
2244 * 2. If there isn't sufficient ZIL activity occurring on
2245 * the system, such that this lwb's zio isn't issued via
2246 * zil_lwb_commit(), zil_commit_waiter() will issue the
2247 * lwb's zio. If this occurs, the lwb is not guaranteed
2248 * to be "full" by the time its zio is issued, and means
2249 * the size of the lwb was "too large" given the amount
2250 * of ZIL activity occurring on the system at that time.
2252 * We do this for a couple of reasons:
2254 * 1. To try and reduce the number of IOPs needed to
2255 * write the same number of itxs. If an lwb has space
2256 * available in its buffer for more itxs, and more itxs
2257 * will be committed relatively soon (relative to the
2258 * latency of performing a write), then it's beneficial
2259 * to wait for these "next" itxs. This way, more itxs
2260 * can be committed to stable storage with fewer writes.
2262 * 2. To try and use the largest lwb block size that the
2263 * incoming rate of itxs can support. Again, this is to
2264 * try and pack as many itxs into as few lwbs as
2265 * possible, without significantly impacting the latency
2266 * of each individual itx.
2272 * This function is responsible for ensuring the passed in commit waiter
2273 * (and associated commit itx) is committed to an lwb. If the waiter is
2274 * not already committed to an lwb, all itxs in the zilog's queue of
2275 * itxs will be processed. The assumption is the passed in waiter's
2276 * commit itx will found in the queue just like the other non-commit
2277 * itxs, such that when the entire queue is processed, the waiter will
2278 * have been committed to an lwb.
2280 * The lwb associated with the passed in waiter is not guaranteed to
2281 * have been issued by the time this function completes. If the lwb is
2282 * not issued, we rely on future calls to zil_commit_writer() to issue
2283 * the lwb, or the timeout mechanism found in zil_commit_waiter().
2286 zil_commit_writer(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2288 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2289 ASSERT(spa_writeable(zilog
->zl_spa
));
2291 mutex_enter(&zilog
->zl_issuer_lock
);
2293 if (zcw
->zcw_lwb
!= NULL
|| zcw
->zcw_done
) {
2295 * It's possible that, while we were waiting to acquire
2296 * the "zl_issuer_lock", another thread committed this
2297 * waiter to an lwb. If that occurs, we bail out early,
2298 * without processing any of the zilog's queue of itxs.
2300 * On certain workloads and system configurations, the
2301 * "zl_issuer_lock" can become highly contended. In an
2302 * attempt to reduce this contention, we immediately drop
2303 * the lock if the waiter has already been processed.
2305 * We've measured this optimization to reduce CPU spent
2306 * contending on this lock by up to 5%, using a system
2307 * with 32 CPUs, low latency storage (~50 usec writes),
2308 * and 1024 threads performing sync writes.
2313 ZIL_STAT_BUMP(zil_commit_writer_count
);
2315 zil_get_commit_list(zilog
);
2316 zil_prune_commit_list(zilog
);
2317 zil_process_commit_list(zilog
);
2320 mutex_exit(&zilog
->zl_issuer_lock
);
2324 zil_commit_waiter_timeout(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2326 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2327 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2328 ASSERT3B(zcw
->zcw_done
, ==, B_FALSE
);
2330 lwb_t
*lwb
= zcw
->zcw_lwb
;
2331 ASSERT3P(lwb
, !=, NULL
);
2332 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_CLOSED
);
2335 * If the lwb has already been issued by another thread, we can
2336 * immediately return since there's no work to be done (the
2337 * point of this function is to issue the lwb). Additionally, we
2338 * do this prior to acquiring the zl_issuer_lock, to avoid
2339 * acquiring it when it's not necessary to do so.
2341 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2342 lwb
->lwb_state
== LWB_STATE_DONE
)
2346 * In order to call zil_lwb_write_issue() we must hold the
2347 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
2348 * since we're already holding the commit waiter's "zcw_lock",
2349 * and those two locks are acquired in the opposite order
2352 mutex_exit(&zcw
->zcw_lock
);
2353 mutex_enter(&zilog
->zl_issuer_lock
);
2354 mutex_enter(&zcw
->zcw_lock
);
2357 * Since we just dropped and re-acquired the commit waiter's
2358 * lock, we have to re-check to see if the waiter was marked
2359 * "done" during that process. If the waiter was marked "done",
2360 * the "lwb" pointer is no longer valid (it can be free'd after
2361 * the waiter is marked "done"), so without this check we could
2362 * wind up with a use-after-free error below.
2367 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2370 * We've already checked this above, but since we hadn't acquired
2371 * the zilog's zl_issuer_lock, we have to perform this check a
2372 * second time while holding the lock.
2374 * We don't need to hold the zl_lock since the lwb cannot transition
2375 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
2376 * _can_ transition from ISSUED to DONE, but it's OK to race with
2377 * that transition since we treat the lwb the same, whether it's in
2378 * the ISSUED or DONE states.
2380 * The important thing, is we treat the lwb differently depending on
2381 * if it's ISSUED or OPENED, and block any other threads that might
2382 * attempt to issue this lwb. For that reason we hold the
2383 * zl_issuer_lock when checking the lwb_state; we must not call
2384 * zil_lwb_write_issue() if the lwb had already been issued.
2386 * See the comment above the lwb_state_t structure definition for
2387 * more details on the lwb states, and locking requirements.
2389 if (lwb
->lwb_state
== LWB_STATE_ISSUED
||
2390 lwb
->lwb_state
== LWB_STATE_DONE
)
2393 ASSERT3S(lwb
->lwb_state
, ==, LWB_STATE_OPENED
);
2396 * As described in the comments above zil_commit_waiter() and
2397 * zil_process_commit_list(), we need to issue this lwb's zio
2398 * since we've reached the commit waiter's timeout and it still
2399 * hasn't been issued.
2401 lwb_t
*nlwb
= zil_lwb_write_issue(zilog
, lwb
);
2403 IMPLY(nlwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_OPENED
);
2406 * Since the lwb's zio hadn't been issued by the time this thread
2407 * reached its timeout, we reset the zilog's "zl_cur_used" field
2408 * to influence the zil block size selection algorithm.
2410 * By having to issue the lwb's zio here, it means the size of the
2411 * lwb was too large, given the incoming throughput of itxs. By
2412 * setting "zl_cur_used" to zero, we communicate this fact to the
2413 * block size selection algorithm, so it can take this information
2414 * into account, and potentially select a smaller size for the
2415 * next lwb block that is allocated.
2417 zilog
->zl_cur_used
= 0;
2421 * When zil_lwb_write_issue() returns NULL, this
2422 * indicates zio_alloc_zil() failed to allocate the
2423 * "next" lwb on-disk. When this occurs, the ZIL write
2424 * pipeline must be stalled; see the comment within the
2425 * zil_commit_writer_stall() function for more details.
2427 * We must drop the commit waiter's lock prior to
2428 * calling zil_commit_writer_stall() or else we can wind
2429 * up with the following deadlock:
2431 * - This thread is waiting for the txg to sync while
2432 * holding the waiter's lock; txg_wait_synced() is
2433 * used within txg_commit_writer_stall().
2435 * - The txg can't sync because it is waiting for this
2436 * lwb's zio callback to call dmu_tx_commit().
2438 * - The lwb's zio callback can't call dmu_tx_commit()
2439 * because it's blocked trying to acquire the waiter's
2440 * lock, which occurs prior to calling dmu_tx_commit()
2442 mutex_exit(&zcw
->zcw_lock
);
2443 zil_commit_writer_stall(zilog
);
2444 mutex_enter(&zcw
->zcw_lock
);
2448 mutex_exit(&zilog
->zl_issuer_lock
);
2449 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2453 * This function is responsible for performing the following two tasks:
2455 * 1. its primary responsibility is to block until the given "commit
2456 * waiter" is considered "done".
2458 * 2. its secondary responsibility is to issue the zio for the lwb that
2459 * the given "commit waiter" is waiting on, if this function has
2460 * waited "long enough" and the lwb is still in the "open" state.
2462 * Given a sufficient amount of itxs being generated and written using
2463 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit()
2464 * function. If this does not occur, this secondary responsibility will
2465 * ensure the lwb is issued even if there is not other synchronous
2466 * activity on the system.
2468 * For more details, see zil_process_commit_list(); more specifically,
2469 * the comment at the bottom of that function.
2472 zil_commit_waiter(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2474 ASSERT(!MUTEX_HELD(&zilog
->zl_lock
));
2475 ASSERT(!MUTEX_HELD(&zilog
->zl_issuer_lock
));
2476 ASSERT(spa_writeable(zilog
->zl_spa
));
2478 mutex_enter(&zcw
->zcw_lock
);
2481 * The timeout is scaled based on the lwb latency to avoid
2482 * significantly impacting the latency of each individual itx.
2483 * For more details, see the comment at the bottom of the
2484 * zil_process_commit_list() function.
2486 int pct
= MAX(zfs_commit_timeout_pct
, 1);
2487 hrtime_t sleep
= (zilog
->zl_last_lwb_latency
* pct
) / 100;
2488 hrtime_t wakeup
= gethrtime() + sleep
;
2489 boolean_t timedout
= B_FALSE
;
2491 while (!zcw
->zcw_done
) {
2492 ASSERT(MUTEX_HELD(&zcw
->zcw_lock
));
2494 lwb_t
*lwb
= zcw
->zcw_lwb
;
2497 * Usually, the waiter will have a non-NULL lwb field here,
2498 * but it's possible for it to be NULL as a result of
2499 * zil_commit() racing with spa_sync().
2501 * When zil_clean() is called, it's possible for the itxg
2502 * list (which may be cleaned via a taskq) to contain
2503 * commit itxs. When this occurs, the commit waiters linked
2504 * off of these commit itxs will not be committed to an
2505 * lwb. Additionally, these commit waiters will not be
2506 * marked done until zil_commit_waiter_skip() is called via
2509 * Thus, it's possible for this commit waiter (i.e. the
2510 * "zcw" variable) to be found in this "in between" state;
2511 * where it's "zcw_lwb" field is NULL, and it hasn't yet
2512 * been skipped, so it's "zcw_done" field is still B_FALSE.
2514 IMPLY(lwb
!= NULL
, lwb
->lwb_state
!= LWB_STATE_CLOSED
);
2516 if (lwb
!= NULL
&& lwb
->lwb_state
== LWB_STATE_OPENED
) {
2517 ASSERT3B(timedout
, ==, B_FALSE
);
2520 * If the lwb hasn't been issued yet, then we
2521 * need to wait with a timeout, in case this
2522 * function needs to issue the lwb after the
2523 * timeout is reached; responsibility (2) from
2524 * the comment above this function.
2526 clock_t timeleft
= cv_timedwait_hires(&zcw
->zcw_cv
,
2527 &zcw
->zcw_lock
, wakeup
, USEC2NSEC(1),
2528 CALLOUT_FLAG_ABSOLUTE
);
2530 if (timeleft
>= 0 || zcw
->zcw_done
)
2534 zil_commit_waiter_timeout(zilog
, zcw
);
2536 if (!zcw
->zcw_done
) {
2538 * If the commit waiter has already been
2539 * marked "done", it's possible for the
2540 * waiter's lwb structure to have already
2541 * been freed. Thus, we can only reliably
2542 * make these assertions if the waiter
2545 ASSERT3P(lwb
, ==, zcw
->zcw_lwb
);
2546 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_OPENED
);
2550 * If the lwb isn't open, then it must have already
2551 * been issued. In that case, there's no need to
2552 * use a timeout when waiting for the lwb to
2555 * Additionally, if the lwb is NULL, the waiter
2556 * will soon be signaled and marked done via
2557 * zil_clean() and zil_itxg_clean(), so no timeout
2562 lwb
->lwb_state
== LWB_STATE_ISSUED
||
2563 lwb
->lwb_state
== LWB_STATE_DONE
);
2564 cv_wait(&zcw
->zcw_cv
, &zcw
->zcw_lock
);
2568 mutex_exit(&zcw
->zcw_lock
);
2571 static zil_commit_waiter_t
*
2572 zil_alloc_commit_waiter(void)
2574 zil_commit_waiter_t
*zcw
= kmem_cache_alloc(zil_zcw_cache
, KM_SLEEP
);
2576 cv_init(&zcw
->zcw_cv
, NULL
, CV_DEFAULT
, NULL
);
2577 mutex_init(&zcw
->zcw_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2578 list_link_init(&zcw
->zcw_node
);
2579 zcw
->zcw_lwb
= NULL
;
2580 zcw
->zcw_done
= B_FALSE
;
2581 zcw
->zcw_zio_error
= 0;
2587 zil_free_commit_waiter(zil_commit_waiter_t
*zcw
)
2589 ASSERT(!list_link_active(&zcw
->zcw_node
));
2590 ASSERT3P(zcw
->zcw_lwb
, ==, NULL
);
2591 ASSERT3B(zcw
->zcw_done
, ==, B_TRUE
);
2592 mutex_destroy(&zcw
->zcw_lock
);
2593 cv_destroy(&zcw
->zcw_cv
);
2594 kmem_cache_free(zil_zcw_cache
, zcw
);
2598 * This function is used to create a TX_COMMIT itx and assign it. This
2599 * way, it will be linked into the ZIL's list of synchronous itxs, and
2600 * then later committed to an lwb (or skipped) when
2601 * zil_process_commit_list() is called.
2604 zil_commit_itx_assign(zilog_t
*zilog
, zil_commit_waiter_t
*zcw
)
2606 dmu_tx_t
*tx
= dmu_tx_create(zilog
->zl_os
);
2607 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
2609 itx_t
*itx
= zil_itx_create(TX_COMMIT
, sizeof (lr_t
));
2610 itx
->itx_sync
= B_TRUE
;
2611 itx
->itx_private
= zcw
;
2613 zil_itx_assign(zilog
, itx
, tx
);
2619 * Commit ZFS Intent Log transactions (itxs) to stable storage.
2621 * When writing ZIL transactions to the on-disk representation of the
2622 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
2623 * itxs can be committed to a single lwb. Once a lwb is written and
2624 * committed to stable storage (i.e. the lwb is written, and vdevs have
2625 * been flushed), each itx that was committed to that lwb is also
2626 * considered to be committed to stable storage.
2628 * When an itx is committed to an lwb, the log record (lr_t) contained
2629 * by the itx is copied into the lwb's zio buffer, and once this buffer
2630 * is written to disk, it becomes an on-disk ZIL block.
2632 * As itxs are generated, they're inserted into the ZIL's queue of
2633 * uncommitted itxs. The semantics of zil_commit() are such that it will
2634 * block until all itxs that were in the queue when it was called, are
2635 * committed to stable storage.
2637 * If "foid" is zero, this means all "synchronous" and "asynchronous"
2638 * itxs, for all objects in the dataset, will be committed to stable
2639 * storage prior to zil_commit() returning. If "foid" is non-zero, all
2640 * "synchronous" itxs for all objects, but only "asynchronous" itxs
2641 * that correspond to the foid passed in, will be committed to stable
2642 * storage prior to zil_commit() returning.
2644 * Generally speaking, when zil_commit() is called, the consumer doesn't
2645 * actually care about _all_ of the uncommitted itxs. Instead, they're
2646 * simply trying to waiting for a specific itx to be committed to disk,
2647 * but the interface(s) for interacting with the ZIL don't allow such
2648 * fine-grained communication. A better interface would allow a consumer
2649 * to create and assign an itx, and then pass a reference to this itx to
2650 * zil_commit(); such that zil_commit() would return as soon as that
2651 * specific itx was committed to disk (instead of waiting for _all_
2652 * itxs to be committed).
2654 * When a thread calls zil_commit() a special "commit itx" will be
2655 * generated, along with a corresponding "waiter" for this commit itx.
2656 * zil_commit() will wait on this waiter's CV, such that when the waiter
2657 * is marked done, and signaled, zil_commit() will return.
2659 * This commit itx is inserted into the queue of uncommitted itxs. This
2660 * provides an easy mechanism for determining which itxs were in the
2661 * queue prior to zil_commit() having been called, and which itxs were
2662 * added after zil_commit() was called.
2664 * The commit it is special; it doesn't have any on-disk representation.
2665 * When a commit itx is "committed" to an lwb, the waiter associated
2666 * with it is linked onto the lwb's list of waiters. Then, when that lwb
2667 * completes, each waiter on the lwb's list is marked done and signaled
2668 * -- allowing the thread waiting on the waiter to return from zil_commit().
2670 * It's important to point out a few critical factors that allow us
2671 * to make use of the commit itxs, commit waiters, per-lwb lists of
2672 * commit waiters, and zio completion callbacks like we're doing:
2674 * 1. The list of waiters for each lwb is traversed, and each commit
2675 * waiter is marked "done" and signaled, in the zio completion
2676 * callback of the lwb's zio[*].
2678 * * Actually, the waiters are signaled in the zio completion
2679 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands
2680 * that are sent to the vdevs upon completion of the lwb zio.
2682 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
2683 * itxs, the order in which they are inserted is preserved[*]; as
2684 * itxs are added to the queue, they are added to the tail of
2685 * in-memory linked lists.
2687 * When committing the itxs to lwbs (to be written to disk), they
2688 * are committed in the same order in which the itxs were added to
2689 * the uncommitted queue's linked list(s); i.e. the linked list of
2690 * itxs to commit is traversed from head to tail, and each itx is
2691 * committed to an lwb in that order.
2695 * - the order of "sync" itxs is preserved w.r.t. other
2696 * "sync" itxs, regardless of the corresponding objects.
2697 * - the order of "async" itxs is preserved w.r.t. other
2698 * "async" itxs corresponding to the same object.
2699 * - the order of "async" itxs is *not* preserved w.r.t. other
2700 * "async" itxs corresponding to different objects.
2701 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
2702 * versa) is *not* preserved, even for itxs that correspond
2703 * to the same object.
2705 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
2706 * zil_get_commit_list(), and zil_process_commit_list().
2708 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
2709 * lwb cannot be considered committed to stable storage, until its
2710 * "previous" lwb is also committed to stable storage. This fact,
2711 * coupled with the fact described above, means that itxs are
2712 * committed in (roughly) the order in which they were generated.
2713 * This is essential because itxs are dependent on prior itxs.
2714 * Thus, we *must not* deem an itx as being committed to stable
2715 * storage, until *all* prior itxs have also been committed to
2718 * To enforce this ordering of lwb zio's, while still leveraging as
2719 * much of the underlying storage performance as possible, we rely
2720 * on two fundamental concepts:
2722 * 1. The creation and issuance of lwb zio's is protected by
2723 * the zilog's "zl_issuer_lock", which ensures only a single
2724 * thread is creating and/or issuing lwb's at a time
2725 * 2. The "previous" lwb is a child of the "current" lwb
2726 * (leveraging the zio parent-child dependency graph)
2728 * By relying on this parent-child zio relationship, we can have
2729 * many lwb zio's concurrently issued to the underlying storage,
2730 * but the order in which they complete will be the same order in
2731 * which they were created.
2734 zil_commit(zilog_t
*zilog
, uint64_t foid
)
2737 * We should never attempt to call zil_commit on a snapshot for
2738 * a couple of reasons:
2740 * 1. A snapshot may never be modified, thus it cannot have any
2741 * in-flight itxs that would have modified the dataset.
2743 * 2. By design, when zil_commit() is called, a commit itx will
2744 * be assigned to this zilog; as a result, the zilog will be
2745 * dirtied. We must not dirty the zilog of a snapshot; there's
2746 * checks in the code that enforce this invariant, and will
2747 * cause a panic if it's not upheld.
2749 ASSERT3B(dmu_objset_is_snapshot(zilog
->zl_os
), ==, B_FALSE
);
2751 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2754 if (!spa_writeable(zilog
->zl_spa
)) {
2756 * If the SPA is not writable, there should never be any
2757 * pending itxs waiting to be committed to disk. If that
2758 * weren't true, we'd skip writing those itxs out, and
2759 * would break the semantics of zil_commit(); thus, we're
2760 * verifying that truth before we return to the caller.
2762 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
2763 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
2764 for (int i
= 0; i
< TXG_SIZE
; i
++)
2765 ASSERT3P(zilog
->zl_itxg
[i
].itxg_itxs
, ==, NULL
);
2770 * If the ZIL is suspended, we don't want to dirty it by calling
2771 * zil_commit_itx_assign() below, nor can we write out
2772 * lwbs like would be done in zil_commit_write(). Thus, we
2773 * simply rely on txg_wait_synced() to maintain the necessary
2774 * semantics, and avoid calling those functions altogether.
2776 if (zilog
->zl_suspend
> 0) {
2777 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2781 zil_commit_impl(zilog
, foid
);
2785 zil_commit_impl(zilog_t
*zilog
, uint64_t foid
)
2787 ZIL_STAT_BUMP(zil_commit_count
);
2790 * Move the "async" itxs for the specified foid to the "sync"
2791 * queues, such that they will be later committed (or skipped)
2792 * to an lwb when zil_process_commit_list() is called.
2794 * Since these "async" itxs must be committed prior to this
2795 * call to zil_commit returning, we must perform this operation
2796 * before we call zil_commit_itx_assign().
2798 zil_async_to_sync(zilog
, foid
);
2801 * We allocate a new "waiter" structure which will initially be
2802 * linked to the commit itx using the itx's "itx_private" field.
2803 * Since the commit itx doesn't represent any on-disk state,
2804 * when it's committed to an lwb, rather than copying the its
2805 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
2806 * added to the lwb's list of waiters. Then, when the lwb is
2807 * committed to stable storage, each waiter in the lwb's list of
2808 * waiters will be marked "done", and signalled.
2810 * We must create the waiter and assign the commit itx prior to
2811 * calling zil_commit_writer(), or else our specific commit itx
2812 * is not guaranteed to be committed to an lwb prior to calling
2813 * zil_commit_waiter().
2815 zil_commit_waiter_t
*zcw
= zil_alloc_commit_waiter();
2816 zil_commit_itx_assign(zilog
, zcw
);
2818 zil_commit_writer(zilog
, zcw
);
2819 zil_commit_waiter(zilog
, zcw
);
2821 if (zcw
->zcw_zio_error
!= 0) {
2823 * If there was an error writing out the ZIL blocks that
2824 * this thread is waiting on, then we fallback to
2825 * relying on spa_sync() to write out the data this
2826 * thread is waiting on. Obviously this has performance
2827 * implications, but the expectation is for this to be
2828 * an exceptional case, and shouldn't occur often.
2830 DTRACE_PROBE2(zil__commit__io__error
,
2831 zilog_t
*, zilog
, zil_commit_waiter_t
*, zcw
);
2832 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2835 zil_free_commit_waiter(zcw
);
2839 * Called in syncing context to free committed log blocks and update log header.
2842 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
2844 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
2845 uint64_t txg
= dmu_tx_get_txg(tx
);
2846 spa_t
*spa
= zilog
->zl_spa
;
2847 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
2851 * We don't zero out zl_destroy_txg, so make sure we don't try
2852 * to destroy it twice.
2854 if (spa_sync_pass(spa
) != 1)
2857 mutex_enter(&zilog
->zl_lock
);
2859 ASSERT(zilog
->zl_stop_sync
== 0);
2861 if (*replayed_seq
!= 0) {
2862 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
2863 zh
->zh_replay_seq
= *replayed_seq
;
2867 if (zilog
->zl_destroy_txg
== txg
) {
2868 blkptr_t blk
= zh
->zh_log
;
2870 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
2872 bzero(zh
, sizeof (zil_header_t
));
2873 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
2875 if (zilog
->zl_keep_first
) {
2877 * If this block was part of log chain that couldn't
2878 * be claimed because a device was missing during
2879 * zil_claim(), but that device later returns,
2880 * then this block could erroneously appear valid.
2881 * To guard against this, assign a new GUID to the new
2882 * log chain so it doesn't matter what blk points to.
2884 zil_init_log_chain(zilog
, &blk
);
2889 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
2890 zh
->zh_log
= lwb
->lwb_blk
;
2891 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
2893 list_remove(&zilog
->zl_lwb_list
, lwb
);
2894 zio_free(spa
, txg
, &lwb
->lwb_blk
);
2895 zil_free_lwb(zilog
, lwb
);
2898 * If we don't have anything left in the lwb list then
2899 * we've had an allocation failure and we need to zero
2900 * out the zil_header blkptr so that we don't end
2901 * up freeing the same block twice.
2903 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
2904 BP_ZERO(&zh
->zh_log
);
2908 * Remove fastwrite on any blocks that have been pre-allocated for
2909 * the next commit. This prevents fastwrite counter pollution by
2910 * unused, long-lived LWBs.
2912 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
2913 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_write_zio
) {
2914 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
2915 lwb
->lwb_fastwrite
= 0;
2919 mutex_exit(&zilog
->zl_lock
);
2924 zil_lwb_cons(void *vbuf
, void *unused
, int kmflag
)
2927 list_create(&lwb
->lwb_itxs
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
2928 list_create(&lwb
->lwb_waiters
, sizeof (zil_commit_waiter_t
),
2929 offsetof(zil_commit_waiter_t
, zcw_node
));
2930 avl_create(&lwb
->lwb_vdev_tree
, zil_lwb_vdev_compare
,
2931 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
2932 mutex_init(&lwb
->lwb_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2938 zil_lwb_dest(void *vbuf
, void *unused
)
2941 mutex_destroy(&lwb
->lwb_vdev_lock
);
2942 avl_destroy(&lwb
->lwb_vdev_tree
);
2943 list_destroy(&lwb
->lwb_waiters
);
2944 list_destroy(&lwb
->lwb_itxs
);
2950 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
2951 sizeof (lwb_t
), 0, zil_lwb_cons
, zil_lwb_dest
, NULL
, NULL
, NULL
, 0);
2953 zil_zcw_cache
= kmem_cache_create("zil_zcw_cache",
2954 sizeof (zil_commit_waiter_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
2956 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
2957 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
2958 KSTAT_FLAG_VIRTUAL
);
2960 if (zil_ksp
!= NULL
) {
2961 zil_ksp
->ks_data
= &zil_stats
;
2962 kstat_install(zil_ksp
);
2969 kmem_cache_destroy(zil_zcw_cache
);
2970 kmem_cache_destroy(zil_lwb_cache
);
2972 if (zil_ksp
!= NULL
) {
2973 kstat_delete(zil_ksp
);
2979 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
2981 zilog
->zl_sync
= sync
;
2985 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
2987 zilog
->zl_logbias
= logbias
;
2991 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
2995 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
2997 zilog
->zl_header
= zh_phys
;
2999 zilog
->zl_spa
= dmu_objset_spa(os
);
3000 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
3001 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
3002 zilog
->zl_logbias
= dmu_objset_logbias(os
);
3003 zilog
->zl_sync
= dmu_objset_syncprop(os
);
3004 zilog
->zl_dirty_max_txg
= 0;
3005 zilog
->zl_last_lwb_opened
= NULL
;
3006 zilog
->zl_last_lwb_latency
= 0;
3008 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3009 mutex_init(&zilog
->zl_issuer_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3011 for (int i
= 0; i
< TXG_SIZE
; i
++) {
3012 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
3013 MUTEX_DEFAULT
, NULL
);
3016 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
3017 offsetof(lwb_t
, lwb_node
));
3019 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
3020 offsetof(itx_t
, itx_node
));
3022 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
3028 zil_free(zilog_t
*zilog
)
3032 zilog
->zl_stop_sync
= 1;
3034 ASSERT0(zilog
->zl_suspend
);
3035 ASSERT0(zilog
->zl_suspending
);
3037 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3038 list_destroy(&zilog
->zl_lwb_list
);
3040 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
3041 list_destroy(&zilog
->zl_itx_commit_list
);
3043 for (i
= 0; i
< TXG_SIZE
; i
++) {
3045 * It's possible for an itx to be generated that doesn't dirty
3046 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3047 * callback to remove the entry. We remove those here.
3049 * Also free up the ziltest itxs.
3051 if (zilog
->zl_itxg
[i
].itxg_itxs
)
3052 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
3053 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
3056 mutex_destroy(&zilog
->zl_issuer_lock
);
3057 mutex_destroy(&zilog
->zl_lock
);
3059 cv_destroy(&zilog
->zl_cv_suspend
);
3061 kmem_free(zilog
, sizeof (zilog_t
));
3065 * Open an intent log.
3068 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
3070 zilog_t
*zilog
= dmu_objset_zil(os
);
3072 ASSERT3P(zilog
->zl_get_data
, ==, NULL
);
3073 ASSERT3P(zilog
->zl_last_lwb_opened
, ==, NULL
);
3074 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
3076 zilog
->zl_get_data
= get_data
;
3082 * Close an intent log.
3085 zil_close(zilog_t
*zilog
)
3090 if (!dmu_objset_is_snapshot(zilog
->zl_os
)) {
3091 zil_commit(zilog
, 0);
3093 ASSERT3P(list_tail(&zilog
->zl_lwb_list
), ==, NULL
);
3094 ASSERT0(zilog
->zl_dirty_max_txg
);
3095 ASSERT3B(zilog_is_dirty(zilog
), ==, B_FALSE
);
3098 mutex_enter(&zilog
->zl_lock
);
3099 lwb
= list_tail(&zilog
->zl_lwb_list
);
3101 txg
= zilog
->zl_dirty_max_txg
;
3103 txg
= MAX(zilog
->zl_dirty_max_txg
, lwb
->lwb_max_txg
);
3104 mutex_exit(&zilog
->zl_lock
);
3107 * We need to use txg_wait_synced() to wait long enough for the
3108 * ZIL to be clean, and to wait for all pending lwbs to be
3112 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
3114 if (zilog_is_dirty(zilog
))
3115 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog
, txg
);
3116 if (txg
< spa_freeze_txg(zilog
->zl_spa
))
3117 VERIFY(!zilog_is_dirty(zilog
));
3119 zilog
->zl_get_data
= NULL
;
3122 * We should have only one lwb left on the list; remove it now.
3124 mutex_enter(&zilog
->zl_lock
);
3125 lwb
= list_head(&zilog
->zl_lwb_list
);
3127 ASSERT3P(lwb
, ==, list_tail(&zilog
->zl_lwb_list
));
3128 ASSERT3S(lwb
->lwb_state
, !=, LWB_STATE_ISSUED
);
3130 if (lwb
->lwb_fastwrite
)
3131 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
3133 list_remove(&zilog
->zl_lwb_list
, lwb
);
3134 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
3135 zil_free_lwb(zilog
, lwb
);
3137 mutex_exit(&zilog
->zl_lock
);
3140 static char *suspend_tag
= "zil suspending";
3143 * Suspend an intent log. While in suspended mode, we still honor
3144 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3145 * On old version pools, we suspend the log briefly when taking a
3146 * snapshot so that it will have an empty intent log.
3148 * Long holds are not really intended to be used the way we do here --
3149 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3150 * could fail. Therefore we take pains to only put a long hold if it is
3151 * actually necessary. Fortunately, it will only be necessary if the
3152 * objset is currently mounted (or the ZVOL equivalent). In that case it
3153 * will already have a long hold, so we are not really making things any worse.
3155 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3156 * zvol_state_t), and use their mechanism to prevent their hold from being
3157 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
3160 * if cookiep == NULL, this does both the suspend & resume.
3161 * Otherwise, it returns with the dataset "long held", and the cookie
3162 * should be passed into zil_resume().
3165 zil_suspend(const char *osname
, void **cookiep
)
3169 const zil_header_t
*zh
;
3172 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
3175 zilog
= dmu_objset_zil(os
);
3177 mutex_enter(&zilog
->zl_lock
);
3178 zh
= zilog
->zl_header
;
3180 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
3181 mutex_exit(&zilog
->zl_lock
);
3182 dmu_objset_rele(os
, suspend_tag
);
3183 return (SET_ERROR(EBUSY
));
3187 * Don't put a long hold in the cases where we can avoid it. This
3188 * is when there is no cookie so we are doing a suspend & resume
3189 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3190 * for the suspend because it's already suspended, or there's no ZIL.
3192 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
3193 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
3194 mutex_exit(&zilog
->zl_lock
);
3195 dmu_objset_rele(os
, suspend_tag
);
3199 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
3200 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
3202 zilog
->zl_suspend
++;
3204 if (zilog
->zl_suspend
> 1) {
3206 * Someone else is already suspending it.
3207 * Just wait for them to finish.
3210 while (zilog
->zl_suspending
)
3211 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
3212 mutex_exit(&zilog
->zl_lock
);
3214 if (cookiep
== NULL
)
3222 * If there is no pointer to an on-disk block, this ZIL must not
3223 * be active (e.g. filesystem not mounted), so there's nothing
3226 if (BP_IS_HOLE(&zh
->zh_log
)) {
3227 ASSERT(cookiep
!= NULL
); /* fast path already handled */
3230 mutex_exit(&zilog
->zl_lock
);
3235 * The ZIL has work to do. Ensure that the associated encryption
3236 * key will remain mapped while we are committing the log by
3237 * grabbing a reference to it. If the key isn't loaded we have no
3238 * choice but to return an error until the wrapping key is loaded.
3240 if (os
->os_encrypted
&&
3241 dsl_dataset_create_key_mapping(dmu_objset_ds(os
)) != 0) {
3242 zilog
->zl_suspend
--;
3243 mutex_exit(&zilog
->zl_lock
);
3244 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3245 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3246 return (SET_ERROR(EACCES
));
3249 zilog
->zl_suspending
= B_TRUE
;
3250 mutex_exit(&zilog
->zl_lock
);
3253 * We need to use zil_commit_impl to ensure we wait for all
3254 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
3255 * to disk before proceeding. If we used zil_commit instead, it
3256 * would just call txg_wait_synced(), because zl_suspend is set.
3257 * txg_wait_synced() doesn't wait for these lwb's to be
3258 * LWB_STATE_DONE before returning.
3260 zil_commit_impl(zilog
, 0);
3263 * Now that we've ensured all lwb's are LWB_STATE_DONE,
3264 * txg_wait_synced() will be called from within zil_destroy(),
3265 * which will ensure the data from the zilog has migrated to the
3266 * main pool before it returns.
3268 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3270 zil_destroy(zilog
, B_FALSE
);
3272 mutex_enter(&zilog
->zl_lock
);
3273 zilog
->zl_suspending
= B_FALSE
;
3274 cv_broadcast(&zilog
->zl_cv_suspend
);
3275 mutex_exit(&zilog
->zl_lock
);
3277 if (os
->os_encrypted
)
3278 dsl_dataset_remove_key_mapping(dmu_objset_ds(os
));
3280 if (cookiep
== NULL
)
3288 zil_resume(void *cookie
)
3290 objset_t
*os
= cookie
;
3291 zilog_t
*zilog
= dmu_objset_zil(os
);
3293 mutex_enter(&zilog
->zl_lock
);
3294 ASSERT(zilog
->zl_suspend
!= 0);
3295 zilog
->zl_suspend
--;
3296 mutex_exit(&zilog
->zl_lock
);
3297 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
3298 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
3301 typedef struct zil_replay_arg
{
3302 zil_replay_func_t
**zr_replay
;
3304 boolean_t zr_byteswap
;
3309 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
3311 char name
[ZFS_MAX_DATASET_NAME_LEN
];
3313 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
3315 dmu_objset_name(zilog
->zl_os
, name
);
3317 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
3318 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
3319 (u_longlong_t
)lr
->lrc_seq
,
3320 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
3321 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
3327 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
3329 zil_replay_arg_t
*zr
= zra
;
3330 const zil_header_t
*zh
= zilog
->zl_header
;
3331 uint64_t reclen
= lr
->lrc_reclen
;
3332 uint64_t txtype
= lr
->lrc_txtype
;
3335 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
3337 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
3340 if (lr
->lrc_txg
< claim_txg
) /* already committed */
3343 /* Strip case-insensitive bit, still present in log record */
3346 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
3347 return (zil_replay_error(zilog
, lr
, EINVAL
));
3350 * If this record type can be logged out of order, the object
3351 * (lr_foid) may no longer exist. That's legitimate, not an error.
3353 if (TX_OOO(txtype
)) {
3354 error
= dmu_object_info(zilog
->zl_os
,
3355 LR_FOID_GET_OBJ(((lr_ooo_t
*)lr
)->lr_foid
), NULL
);
3356 if (error
== ENOENT
|| error
== EEXIST
)
3361 * Make a copy of the data so we can revise and extend it.
3363 bcopy(lr
, zr
->zr_lr
, reclen
);
3366 * If this is a TX_WRITE with a blkptr, suck in the data.
3368 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
3369 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
3370 zr
->zr_lr
+ reclen
);
3372 return (zil_replay_error(zilog
, lr
, error
));
3376 * The log block containing this lr may have been byteswapped
3377 * so that we can easily examine common fields like lrc_txtype.
3378 * However, the log is a mix of different record types, and only the
3379 * replay vectors know how to byteswap their records. Therefore, if
3380 * the lr was byteswapped, undo it before invoking the replay vector.
3382 if (zr
->zr_byteswap
)
3383 byteswap_uint64_array(zr
->zr_lr
, reclen
);
3386 * We must now do two things atomically: replay this log record,
3387 * and update the log header sequence number to reflect the fact that
3388 * we did so. At the end of each replay function the sequence number
3389 * is updated if we are in replay mode.
3391 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
3394 * The DMU's dnode layer doesn't see removes until the txg
3395 * commits, so a subsequent claim can spuriously fail with
3396 * EEXIST. So if we receive any error we try syncing out
3397 * any removes then retry the transaction. Note that we
3398 * specify B_FALSE for byteswap now, so we don't do it twice.
3400 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
3401 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
3403 return (zil_replay_error(zilog
, lr
, error
));
3410 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
3412 zilog
->zl_replay_blks
++;
3418 * If this dataset has a non-empty intent log, replay it and destroy it.
3421 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t
*replay_func
[TX_MAX_TYPE
])
3423 zilog_t
*zilog
= dmu_objset_zil(os
);
3424 const zil_header_t
*zh
= zilog
->zl_header
;
3425 zil_replay_arg_t zr
;
3427 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
3428 zil_destroy(zilog
, B_TRUE
);
3432 zr
.zr_replay
= replay_func
;
3434 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
3435 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
3438 * Wait for in-progress removes to sync before starting replay.
3440 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
3442 zilog
->zl_replay
= B_TRUE
;
3443 zilog
->zl_replay_time
= ddi_get_lbolt();
3444 ASSERT(zilog
->zl_replay_blks
== 0);
3445 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
3446 zh
->zh_claim_txg
, B_TRUE
);
3447 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
3449 zil_destroy(zilog
, B_FALSE
);
3450 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
3451 zilog
->zl_replay
= B_FALSE
;
3455 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
3457 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
3460 if (zilog
->zl_replay
) {
3461 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
3462 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
3463 zilog
->zl_replaying_seq
;
3472 zil_reset(const char *osname
, void *arg
)
3476 error
= zil_suspend(osname
, NULL
);
3477 /* EACCES means crypto key not loaded */
3478 if ((error
== EACCES
) || (error
== EBUSY
))
3479 return (SET_ERROR(error
));
3481 return (SET_ERROR(EEXIST
));
3485 #if defined(_KERNEL)
3486 EXPORT_SYMBOL(zil_alloc
);
3487 EXPORT_SYMBOL(zil_free
);
3488 EXPORT_SYMBOL(zil_open
);
3489 EXPORT_SYMBOL(zil_close
);
3490 EXPORT_SYMBOL(zil_replay
);
3491 EXPORT_SYMBOL(zil_replaying
);
3492 EXPORT_SYMBOL(zil_destroy
);
3493 EXPORT_SYMBOL(zil_destroy_sync
);
3494 EXPORT_SYMBOL(zil_itx_create
);
3495 EXPORT_SYMBOL(zil_itx_destroy
);
3496 EXPORT_SYMBOL(zil_itx_assign
);
3497 EXPORT_SYMBOL(zil_commit
);
3498 EXPORT_SYMBOL(zil_claim
);
3499 EXPORT_SYMBOL(zil_check_log_chain
);
3500 EXPORT_SYMBOL(zil_sync
);
3501 EXPORT_SYMBOL(zil_clean
);
3502 EXPORT_SYMBOL(zil_suspend
);
3503 EXPORT_SYMBOL(zil_resume
);
3504 EXPORT_SYMBOL(zil_lwb_add_block
);
3505 EXPORT_SYMBOL(zil_bp_tree_add
);
3506 EXPORT_SYMBOL(zil_set_sync
);
3507 EXPORT_SYMBOL(zil_set_logbias
);
3510 module_param(zfs_commit_timeout_pct
, int, 0644);
3511 MODULE_PARM_DESC(zfs_commit_timeout_pct
, "ZIL block open timeout percentage");
3513 module_param(zil_replay_disable
, int, 0644);
3514 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
3516 module_param(zfs_nocacheflush
, int, 0644);
3517 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
3519 module_param(zil_slog_bulk
, ulong
, 0644);
3520 MODULE_PARM_DESC(zil_slog_bulk
, "Limit in bytes slog sync writes per commit");