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.
25 /* Portions Copyright 2010 Robert Milkowski */
27 #include <sys/zfs_context.h>
33 #include <sys/resource.h>
35 #include <sys/zil_impl.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/dsl_pool.h>
42 * The zfs intent log (ZIL) saves transaction records of system calls
43 * that change the file system in memory with enough information
44 * to be able to replay them. These are stored in memory until
45 * either the DMU transaction group (txg) commits them to the stable pool
46 * and they can be discarded, or they are flushed to the stable log
47 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
48 * requirement. In the event of a panic or power fail then those log
49 * records (transactions) are replayed.
51 * There is one ZIL per file system. Its on-disk (pool) format consists
58 * A log record holds a system call transaction. Log blocks can
59 * hold many log records and the blocks are chained together.
60 * Each ZIL block contains a block pointer (blkptr_t) to the next
61 * ZIL block in the chain. The ZIL header points to the first
62 * block in the chain. Note there is not a fixed place in the pool
63 * to hold blocks. They are dynamically allocated and freed as
64 * needed from the blocks available. Figure X shows the ZIL structure:
68 * This global ZIL switch affects all pools
70 int zil_replay_disable
= 0; /* disable intent logging replay */
73 * Tunable parameter for debugging or performance analysis. Setting
74 * zfs_nocacheflush will cause corruption on power loss if a volatile
75 * out-of-order write cache is enabled.
77 boolean_t zfs_nocacheflush
= B_FALSE
;
79 static kmem_cache_t
*zil_lwb_cache
;
81 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
83 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
84 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
88 * ziltest is by and large an ugly hack, but very useful in
89 * checking replay without tedious work.
90 * When running ziltest we want to keep all itx's and so maintain
91 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
92 * We subtract TXG_CONCURRENT_STATES to allow for common code.
94 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
97 zil_bp_compare(const void *x1
, const void *x2
)
99 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
100 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
102 if (DVA_GET_VDEV(dva1
) < DVA_GET_VDEV(dva2
))
104 if (DVA_GET_VDEV(dva1
) > DVA_GET_VDEV(dva2
))
107 if (DVA_GET_OFFSET(dva1
) < DVA_GET_OFFSET(dva2
))
109 if (DVA_GET_OFFSET(dva1
) > DVA_GET_OFFSET(dva2
))
116 zil_bp_tree_init(zilog_t
*zilog
)
118 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
119 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
123 zil_bp_tree_fini(zilog_t
*zilog
)
125 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
129 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
130 kmem_free(zn
, sizeof (zil_bp_node_t
));
136 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
138 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
139 const dva_t
*dva
= BP_IDENTITY(bp
);
143 if (avl_find(t
, dva
, &where
) != NULL
)
146 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_SLEEP
);
148 avl_insert(t
, zn
, where
);
153 static zil_header_t
*
154 zil_header_in_syncing_context(zilog_t
*zilog
)
156 return ((zil_header_t
*)zilog
->zl_header
);
160 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
162 zio_cksum_t
*zc
= &bp
->blk_cksum
;
164 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
165 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
166 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
167 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
171 * Read a log block and make sure it's valid.
174 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
177 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
178 uint32_t aflags
= ARC_WAIT
;
179 arc_buf_t
*abuf
= NULL
;
183 if (zilog
->zl_header
->zh_claim_txg
== 0)
184 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
186 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
187 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
189 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
190 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
192 error
= dsl_read_nolock(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
193 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
196 zio_cksum_t cksum
= bp
->blk_cksum
;
199 * Validate the checksummed log block.
201 * Sequence numbers should be... sequential. The checksum
202 * verifier for the next block should be bp's checksum plus 1.
204 * Also check the log chain linkage and size used.
206 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
208 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
209 zil_chain_t
*zilc
= abuf
->b_data
;
210 char *lr
= (char *)(zilc
+ 1);
211 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
213 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
214 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
218 *end
= (char *)dst
+ len
;
219 *nbp
= zilc
->zc_next_blk
;
222 char *lr
= abuf
->b_data
;
223 uint64_t size
= BP_GET_LSIZE(bp
);
224 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
226 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
227 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
228 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
231 bcopy(lr
, dst
, zilc
->zc_nused
);
232 *end
= (char *)dst
+ zilc
->zc_nused
;
233 *nbp
= zilc
->zc_next_blk
;
237 VERIFY(arc_buf_remove_ref(abuf
, &abuf
) == 1);
244 * Read a TX_WRITE log data block.
247 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
249 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
250 const blkptr_t
*bp
= &lr
->lr_blkptr
;
251 uint32_t aflags
= ARC_WAIT
;
252 arc_buf_t
*abuf
= NULL
;
256 if (BP_IS_HOLE(bp
)) {
258 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
262 if (zilog
->zl_header
->zh_claim_txg
== 0)
263 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
265 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
266 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
268 error
= arc_read_nolock(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
269 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
273 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
274 (void) arc_buf_remove_ref(abuf
, &abuf
);
281 * Parse the intent log, and call parse_func for each valid record within.
284 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
285 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
287 const zil_header_t
*zh
= zilog
->zl_header
;
288 boolean_t claimed
= !!zh
->zh_claim_txg
;
289 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
290 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
291 uint64_t max_blk_seq
= 0;
292 uint64_t max_lr_seq
= 0;
293 uint64_t blk_count
= 0;
294 uint64_t lr_count
= 0;
295 blkptr_t blk
, next_blk
;
300 * Old logs didn't record the maximum zh_claim_lr_seq.
302 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
303 claim_lr_seq
= UINT64_MAX
;
306 * Starting at the block pointed to by zh_log we read the log chain.
307 * For each block in the chain we strongly check that block to
308 * ensure its validity. We stop when an invalid block is found.
309 * For each block pointer in the chain we call parse_blk_func().
310 * For each record in each valid block we call parse_lr_func().
311 * If the log has been claimed, stop if we encounter a sequence
312 * number greater than the highest claimed sequence number.
314 lrbuf
= zio_buf_alloc(SPA_MAXBLOCKSIZE
);
315 zil_bp_tree_init(zilog
);
317 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
318 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
322 if (blk_seq
> claim_blk_seq
)
324 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
326 ASSERT3U(max_blk_seq
, <, blk_seq
);
327 max_blk_seq
= blk_seq
;
330 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
333 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
337 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
338 lr_t
*lr
= (lr_t
*)lrp
;
339 reclen
= lr
->lrc_reclen
;
340 ASSERT3U(reclen
, >=, sizeof (lr_t
));
341 if (lr
->lrc_seq
> claim_lr_seq
)
343 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
345 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
346 max_lr_seq
= lr
->lrc_seq
;
351 zilog
->zl_parse_error
= error
;
352 zilog
->zl_parse_blk_seq
= max_blk_seq
;
353 zilog
->zl_parse_lr_seq
= max_lr_seq
;
354 zilog
->zl_parse_blk_count
= blk_count
;
355 zilog
->zl_parse_lr_count
= lr_count
;
357 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
358 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
360 zil_bp_tree_fini(zilog
);
361 zio_buf_free(lrbuf
, SPA_MAXBLOCKSIZE
);
367 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
370 * Claim log block if not already committed and not already claimed.
371 * If tx == NULL, just verify that the block is claimable.
373 if (bp
->blk_birth
< first_txg
|| zil_bp_tree_add(zilog
, bp
) != 0)
376 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
377 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
378 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
382 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
384 lr_write_t
*lr
= (lr_write_t
*)lrc
;
387 if (lrc
->lrc_txtype
!= TX_WRITE
)
391 * If the block is not readable, don't claim it. This can happen
392 * in normal operation when a log block is written to disk before
393 * some of the dmu_sync() blocks it points to. In this case, the
394 * transaction cannot have been committed to anyone (we would have
395 * waited for all writes to be stable first), so it is semantically
396 * correct to declare this the end of the log.
398 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
399 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
401 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
406 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
408 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
414 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
416 lr_write_t
*lr
= (lr_write_t
*)lrc
;
417 blkptr_t
*bp
= &lr
->lr_blkptr
;
420 * If we previously claimed it, we need to free it.
422 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
423 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0)
424 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
430 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, uint64_t txg
)
434 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_SLEEP
);
435 lwb
->lwb_zilog
= zilog
;
437 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
438 lwb
->lwb_max_txg
= txg
;
441 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
442 lwb
->lwb_nused
= sizeof (zil_chain_t
);
443 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
446 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
449 mutex_enter(&zilog
->zl_lock
);
450 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
451 mutex_exit(&zilog
->zl_lock
);
457 * Create an on-disk intent log.
460 zil_create(zilog_t
*zilog
)
462 const zil_header_t
*zh
= zilog
->zl_header
;
470 * Wait for any previous destroy to complete.
472 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
474 ASSERT(zh
->zh_claim_txg
== 0);
475 ASSERT(zh
->zh_replay_seq
== 0);
480 * Allocate an initial log block if:
481 * - there isn't one already
482 * - the existing block is the wrong endianess
484 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
485 tx
= dmu_tx_create(zilog
->zl_os
);
486 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
487 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
488 txg
= dmu_tx_get_txg(tx
);
490 if (!BP_IS_HOLE(&blk
)) {
491 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
495 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
, NULL
,
496 ZIL_MIN_BLKSZ
, zilog
->zl_logbias
== ZFS_LOGBIAS_LATENCY
);
499 zil_init_log_chain(zilog
, &blk
);
503 * Allocate a log write buffer (lwb) for the first log block.
506 lwb
= zil_alloc_lwb(zilog
, &blk
, txg
);
509 * If we just allocated the first log block, commit our transaction
510 * and wait for zil_sync() to stuff the block poiner into zh_log.
511 * (zh is part of the MOS, so we cannot modify it in open context.)
515 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
518 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
524 * In one tx, free all log blocks and clear the log header.
525 * If keep_first is set, then we're replaying a log with no content.
526 * We want to keep the first block, however, so that the first
527 * synchronous transaction doesn't require a txg_wait_synced()
528 * in zil_create(). We don't need to txg_wait_synced() here either
529 * when keep_first is set, because both zil_create() and zil_destroy()
530 * will wait for any in-progress destroys to complete.
533 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
535 const zil_header_t
*zh
= zilog
->zl_header
;
541 * Wait for any previous destroy to complete.
543 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
545 zilog
->zl_old_header
= *zh
; /* debugging aid */
547 if (BP_IS_HOLE(&zh
->zh_log
))
550 tx
= dmu_tx_create(zilog
->zl_os
);
551 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
552 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
553 txg
= dmu_tx_get_txg(tx
);
555 mutex_enter(&zilog
->zl_lock
);
557 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
558 zilog
->zl_destroy_txg
= txg
;
559 zilog
->zl_keep_first
= keep_first
;
561 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
562 ASSERT(zh
->zh_claim_txg
== 0);
564 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
565 list_remove(&zilog
->zl_lwb_list
, lwb
);
566 if (lwb
->lwb_buf
!= NULL
)
567 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
568 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
569 kmem_cache_free(zil_lwb_cache
, lwb
);
571 } else if (!keep_first
) {
572 (void) zil_parse(zilog
, zil_free_log_block
,
573 zil_free_log_record
, tx
, zh
->zh_claim_txg
);
575 mutex_exit(&zilog
->zl_lock
);
581 zil_claim(const char *osname
, void *txarg
)
583 dmu_tx_t
*tx
= txarg
;
584 uint64_t first_txg
= dmu_tx_get_txg(tx
);
590 error
= dmu_objset_hold(osname
, FTAG
, &os
);
592 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
596 zilog
= dmu_objset_zil(os
);
597 zh
= zil_header_in_syncing_context(zilog
);
599 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
600 if (!BP_IS_HOLE(&zh
->zh_log
))
601 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
602 BP_ZERO(&zh
->zh_log
);
603 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
604 dmu_objset_rele(os
, FTAG
);
609 * Claim all log blocks if we haven't already done so, and remember
610 * the highest claimed sequence number. This ensures that if we can
611 * read only part of the log now (e.g. due to a missing device),
612 * but we can read the entire log later, we will not try to replay
613 * or destroy beyond the last block we successfully claimed.
615 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
616 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
617 (void) zil_parse(zilog
, zil_claim_log_block
,
618 zil_claim_log_record
, tx
, first_txg
);
619 zh
->zh_claim_txg
= first_txg
;
620 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
621 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
622 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
623 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
624 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
625 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
628 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
629 dmu_objset_rele(os
, FTAG
);
634 * Check the log by walking the log chain.
635 * Checksum errors are ok as they indicate the end of the chain.
636 * Any other error (no device or read failure) returns an error.
639 zil_check_log_chain(const char *osname
, void *tx
)
648 error
= dmu_objset_hold(osname
, FTAG
, &os
);
650 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
654 zilog
= dmu_objset_zil(os
);
655 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
658 * Check the first block and determine if it's on a log device
659 * which may have been removed or faulted prior to loading this
660 * pool. If so, there's no point in checking the rest of the log
661 * as its content should have already been synced to the pool.
663 if (!BP_IS_HOLE(bp
)) {
665 boolean_t valid
= B_TRUE
;
667 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
668 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
669 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
670 valid
= vdev_log_state_valid(vd
);
671 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
674 dmu_objset_rele(os
, FTAG
);
680 * Because tx == NULL, zil_claim_log_block() will not actually claim
681 * any blocks, but just determine whether it is possible to do so.
682 * In addition to checking the log chain, zil_claim_log_block()
683 * will invoke zio_claim() with a done func of spa_claim_notify(),
684 * which will update spa_max_claim_txg. See spa_load() for details.
686 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
687 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
689 dmu_objset_rele(os
, FTAG
);
691 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
695 zil_vdev_compare(const void *x1
, const void *x2
)
697 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
698 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
709 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
711 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
713 zil_vdev_node_t
*zv
, zvsearch
;
714 int ndvas
= BP_GET_NDVAS(bp
);
717 if (zfs_nocacheflush
)
720 ASSERT(zilog
->zl_writer
);
723 * Even though we're zl_writer, we still need a lock because the
724 * zl_get_data() callbacks may have dmu_sync() done callbacks
725 * that will run concurrently.
727 mutex_enter(&zilog
->zl_vdev_lock
);
728 for (i
= 0; i
< ndvas
; i
++) {
729 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
730 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
731 zv
= kmem_alloc(sizeof (*zv
), KM_SLEEP
);
732 zv
->zv_vdev
= zvsearch
.zv_vdev
;
733 avl_insert(t
, zv
, where
);
736 mutex_exit(&zilog
->zl_vdev_lock
);
740 zil_flush_vdevs(zilog_t
*zilog
)
742 spa_t
*spa
= zilog
->zl_spa
;
743 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
748 ASSERT(zilog
->zl_writer
);
751 * We don't need zl_vdev_lock here because we're the zl_writer,
752 * and all zl_get_data() callbacks are done.
754 if (avl_numnodes(t
) == 0)
757 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
759 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
761 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
762 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
765 kmem_free(zv
, sizeof (*zv
));
769 * Wait for all the flushes to complete. Not all devices actually
770 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
772 (void) zio_wait(zio
);
774 spa_config_exit(spa
, SCL_STATE
, FTAG
);
778 * Function called when a log block write completes
781 zil_lwb_write_done(zio_t
*zio
)
783 lwb_t
*lwb
= zio
->io_private
;
784 zilog_t
*zilog
= lwb
->lwb_zilog
;
785 dmu_tx_t
*tx
= lwb
->lwb_tx
;
787 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
788 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
789 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
790 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
791 ASSERT(!BP_IS_GANG(zio
->io_bp
));
792 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
793 ASSERT(zio
->io_bp
->blk_fill
== 0);
796 * Ensure the lwb buffer pointer is cleared before releasing
797 * the txg. If we have had an allocation failure and
798 * the txg is waiting to sync then we want want zil_sync()
799 * to remove the lwb so that it's not picked up as the next new
800 * one in zil_commit_writer(). zil_sync() will only remove
801 * the lwb if lwb_buf is null.
803 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
804 mutex_enter(&zilog
->zl_lock
);
807 mutex_exit(&zilog
->zl_lock
);
810 * Now that we've written this log block, we have a stable pointer
811 * to the next block in the chain, so it's OK to let the txg in
812 * which we allocated the next block sync.
818 * Initialize the io for a log block.
821 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
825 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
826 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
827 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
829 if (zilog
->zl_root_zio
== NULL
) {
830 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
833 if (lwb
->lwb_zio
== NULL
) {
834 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
835 0, &lwb
->lwb_blk
, lwb
->lwb_buf
, BP_GET_LSIZE(&lwb
->lwb_blk
),
836 zil_lwb_write_done
, lwb
, ZIO_PRIORITY_LOG_WRITE
,
837 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
, &zb
);
842 * Define a limited set of intent log block sizes.
843 * These must be a multiple of 4KB. Note only the amount used (again
844 * aligned to 4KB) actually gets written. However, we can't always just
845 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
847 uint64_t zil_block_buckets
[] = {
848 4096, /* non TX_WRITE */
849 8192+4096, /* data base */
850 32*1024 + 4096, /* NFS writes */
855 * Use the slog as long as the logbias is 'latency' and the current commit size
856 * is less than the limit or the total list size is less than 2X the limit.
857 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
859 uint64_t zil_slog_limit
= 1024 * 1024;
860 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
861 (((zilog)->zl_cur_used < zil_slog_limit) || \
862 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
865 * Start a log block write and advance to the next log block.
866 * Calls are serialized.
869 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
873 spa_t
*spa
= zilog
->zl_spa
;
877 uint64_t zil_blksz
, wsz
;
880 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
881 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
882 bp
= &zilc
->zc_next_blk
;
884 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
885 bp
= &zilc
->zc_next_blk
;
888 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
891 * Allocate the next block and save its address in this block
892 * before writing it in order to establish the log chain.
893 * Note that if the allocation of nlwb synced before we wrote
894 * the block that points at it (lwb), we'd leak it if we crashed.
895 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
896 * We dirty the dataset to ensure that zil_sync() will be called
897 * to clean up in the event of allocation failure or I/O failure.
899 tx
= dmu_tx_create(zilog
->zl_os
);
900 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
901 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
902 txg
= dmu_tx_get_txg(tx
);
907 * Log blocks are pre-allocated. Here we select the size of the next
908 * block, based on size used in the last block.
909 * - first find the smallest bucket that will fit the block from a
910 * limited set of block sizes. This is because it's faster to write
911 * blocks allocated from the same metaslab as they are adjacent or
913 * - next find the maximum from the new suggested size and an array of
914 * previous sizes. This lessens a picket fence effect of wrongly
915 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
918 * Note we only write what is used, but we can't just allocate
919 * the maximum block size because we can exhaust the available
922 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
923 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
925 zil_blksz
= zil_block_buckets
[i
];
926 if (zil_blksz
== UINT64_MAX
)
927 zil_blksz
= SPA_MAXBLOCKSIZE
;
928 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
929 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
930 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
931 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
934 /* pass the old blkptr in order to spread log blocks across devs */
935 error
= zio_alloc_zil(spa
, txg
, bp
, &lwb
->lwb_blk
, zil_blksz
,
938 ASSERT3U(bp
->blk_birth
, ==, txg
);
939 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
940 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
943 * Allocate a new log write buffer (lwb).
945 nlwb
= zil_alloc_lwb(zilog
, bp
, txg
);
947 /* Record the block for later vdev flushing */
948 zil_add_block(zilog
, &lwb
->lwb_blk
);
951 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
952 /* For Slim ZIL only write what is used. */
953 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
954 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
955 zio_shrink(lwb
->lwb_zio
, wsz
);
962 zilc
->zc_nused
= lwb
->lwb_nused
;
963 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
966 * clear unused data for security
968 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
970 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
973 * If there was an allocation failure then nlwb will be null which
974 * forces a txg_wait_synced().
980 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
982 lr_t
*lrc
= &itx
->itx_lr
; /* common log record */
983 lr_write_t
*lrw
= (lr_write_t
*)lrc
;
985 uint64_t txg
= lrc
->lrc_txg
;
986 uint64_t reclen
= lrc
->lrc_reclen
;
992 ASSERT(lwb
->lwb_buf
!= NULL
);
994 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
)
995 dlen
= P2ROUNDUP_TYPED(
996 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
998 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1000 zil_lwb_write_init(zilog
, lwb
);
1003 * If this record won't fit in the current log block, start a new one.
1005 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1006 lwb
= zil_lwb_write_start(zilog
, lwb
);
1009 zil_lwb_write_init(zilog
, lwb
);
1010 ASSERT(LWB_EMPTY(lwb
));
1011 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1012 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1017 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1018 bcopy(lrc
, lr_buf
, reclen
);
1019 lrc
= (lr_t
*)lr_buf
;
1020 lrw
= (lr_write_t
*)lrc
;
1023 * If it's a write, fetch the data or get its blkptr as appropriate.
1025 if (lrc
->lrc_txtype
== TX_WRITE
) {
1026 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1027 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1028 if (itx
->itx_wr_state
!= WR_COPIED
) {
1033 ASSERT(itx
->itx_wr_state
== WR_NEED_COPY
);
1034 dbuf
= lr_buf
+ reclen
;
1035 lrw
->lr_common
.lrc_reclen
+= dlen
;
1037 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1040 error
= zilog
->zl_get_data(
1041 itx
->itx_private
, lrw
, dbuf
, lwb
->lwb_zio
);
1043 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1047 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1055 * We're actually making an entry, so update lrc_seq to be the
1056 * log record sequence number. Note that this is generally not
1057 * equal to the itx sequence number because not all transactions
1058 * are synchronous, and sometimes spa_sync() gets there first.
1060 lrc
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1061 lwb
->lwb_nused
+= reclen
+ dlen
;
1062 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1063 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1064 ASSERT3U(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)), ==, 0);
1070 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1074 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1076 itx
= kmem_alloc(offsetof(itx_t
, itx_lr
) + lrsize
, KM_SLEEP
);
1077 itx
->itx_lr
.lrc_txtype
= txtype
;
1078 itx
->itx_lr
.lrc_reclen
= lrsize
;
1079 itx
->itx_sod
= lrsize
; /* if write & WR_NEED_COPY will be increased */
1080 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1081 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1087 zil_itx_destroy(itx_t
*itx
)
1089 kmem_free(itx
, offsetof(itx_t
, itx_lr
) + itx
->itx_lr
.lrc_reclen
);
1093 * Free up the sync and async itxs. The itxs_t has already been detached
1094 * so no locks are needed.
1097 zil_itxg_clean(itxs_t
*itxs
)
1103 itx_async_node_t
*ian
;
1105 list
= &itxs
->i_sync_list
;
1106 while ((itx
= list_head(list
)) != NULL
) {
1107 list_remove(list
, itx
);
1108 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1109 itx
->itx_lr
.lrc_reclen
);
1113 t
= &itxs
->i_async_tree
;
1114 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1115 list
= &ian
->ia_list
;
1116 while ((itx
= list_head(list
)) != NULL
) {
1117 list_remove(list
, itx
);
1118 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1119 itx
->itx_lr
.lrc_reclen
);
1122 kmem_free(ian
, sizeof (itx_async_node_t
));
1126 kmem_free(itxs
, sizeof (itxs_t
));
1130 zil_aitx_compare(const void *x1
, const void *x2
)
1132 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1133 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1144 * Remove all async itx with the given oid.
1147 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1150 itx_async_node_t
*ian
;
1157 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1159 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1162 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1164 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1165 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1167 mutex_enter(&itxg
->itxg_lock
);
1168 if (itxg
->itxg_txg
!= txg
) {
1169 mutex_exit(&itxg
->itxg_lock
);
1174 * Locate the object node and append its list.
1176 t
= &itxg
->itxg_itxs
->i_async_tree
;
1177 ian
= avl_find(t
, &oid
, &where
);
1179 list_move_tail(&clean_list
, &ian
->ia_list
);
1180 mutex_exit(&itxg
->itxg_lock
);
1182 while ((itx
= list_head(&clean_list
)) != NULL
) {
1183 list_remove(&clean_list
, itx
);
1184 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1185 itx
->itx_lr
.lrc_reclen
);
1187 list_destroy(&clean_list
);
1191 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1195 itxs_t
*itxs
, *clean
= NULL
;
1198 * Object ids can be re-instantiated in the next txg so
1199 * remove any async transactions to avoid future leaks.
1200 * This can happen if a fsync occurs on the re-instantiated
1201 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1202 * the new file data and flushes a write record for the old object.
1204 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1205 zil_remove_async(zilog
, itx
->itx_oid
);
1208 * Ensure the data of a renamed file is committed before the rename.
1210 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1211 zil_async_to_sync(zilog
, itx
->itx_oid
);
1213 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1216 txg
= dmu_tx_get_txg(tx
);
1218 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1219 mutex_enter(&itxg
->itxg_lock
);
1220 itxs
= itxg
->itxg_itxs
;
1221 if (itxg
->itxg_txg
!= txg
) {
1224 * The zil_clean callback hasn't got around to cleaning
1225 * this itxg. Save the itxs for release below.
1226 * This should be rare.
1228 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1230 clean
= itxg
->itxg_itxs
;
1232 ASSERT(itxg
->itxg_sod
== 0);
1233 itxg
->itxg_txg
= txg
;
1234 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
), KM_SLEEP
);
1236 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1237 offsetof(itx_t
, itx_node
));
1238 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1239 sizeof (itx_async_node_t
),
1240 offsetof(itx_async_node_t
, ia_node
));
1242 if (itx
->itx_sync
) {
1243 list_insert_tail(&itxs
->i_sync_list
, itx
);
1244 atomic_add_64(&zilog
->zl_itx_list_sz
, itx
->itx_sod
);
1245 itxg
->itxg_sod
+= itx
->itx_sod
;
1247 avl_tree_t
*t
= &itxs
->i_async_tree
;
1248 uint64_t foid
= ((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
;
1249 itx_async_node_t
*ian
;
1252 ian
= avl_find(t
, &foid
, &where
);
1254 ian
= kmem_alloc(sizeof (itx_async_node_t
), KM_SLEEP
);
1255 list_create(&ian
->ia_list
, sizeof (itx_t
),
1256 offsetof(itx_t
, itx_node
));
1257 ian
->ia_foid
= foid
;
1258 avl_insert(t
, ian
, where
);
1260 list_insert_tail(&ian
->ia_list
, itx
);
1263 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1264 mutex_exit(&itxg
->itxg_lock
);
1266 /* Release the old itxs now we've dropped the lock */
1268 zil_itxg_clean(clean
);
1272 * If there are any in-memory intent log transactions which have now been
1273 * synced then start up a taskq to free them.
1276 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1278 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1281 mutex_enter(&itxg
->itxg_lock
);
1282 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1283 mutex_exit(&itxg
->itxg_lock
);
1286 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1287 ASSERT(itxg
->itxg_txg
!= 0);
1288 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1289 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1291 clean_me
= itxg
->itxg_itxs
;
1292 itxg
->itxg_itxs
= NULL
;
1294 mutex_exit(&itxg
->itxg_lock
);
1296 * Preferably start a task queue to free up the old itxs but
1297 * if taskq_dispatch can't allocate resources to do that then
1298 * free it in-line. This should be rare. Note, using TQ_SLEEP
1299 * created a bad performance problem.
1301 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1302 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1303 zil_itxg_clean(clean_me
);
1307 * Get the list of itxs to commit into zl_itx_commit_list.
1310 zil_get_commit_list(zilog_t
*zilog
)
1313 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1314 uint64_t push_sod
= 0;
1316 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1319 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1321 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1322 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1324 mutex_enter(&itxg
->itxg_lock
);
1325 if (itxg
->itxg_txg
!= txg
) {
1326 mutex_exit(&itxg
->itxg_lock
);
1330 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1331 push_sod
+= itxg
->itxg_sod
;
1334 mutex_exit(&itxg
->itxg_lock
);
1336 atomic_add_64(&zilog
->zl_itx_list_sz
, -push_sod
);
1340 * Move the async itxs for a specified object to commit into sync lists.
1343 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1346 itx_async_node_t
*ian
;
1350 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1353 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1355 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1356 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1358 mutex_enter(&itxg
->itxg_lock
);
1359 if (itxg
->itxg_txg
!= txg
) {
1360 mutex_exit(&itxg
->itxg_lock
);
1365 * If a foid is specified then find that node and append its
1366 * list. Otherwise walk the tree appending all the lists
1367 * to the sync list. We add to the end rather than the
1368 * beginning to ensure the create has happened.
1370 t
= &itxg
->itxg_itxs
->i_async_tree
;
1372 ian
= avl_find(t
, &foid
, &where
);
1374 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1378 void *cookie
= NULL
;
1380 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1381 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1383 list_destroy(&ian
->ia_list
);
1384 kmem_free(ian
, sizeof (itx_async_node_t
));
1387 mutex_exit(&itxg
->itxg_lock
);
1392 zil_commit_writer(zilog_t
*zilog
)
1397 spa_t
*spa
= zilog
->zl_spa
;
1400 ASSERT(zilog
->zl_root_zio
== NULL
);
1402 mutex_exit(&zilog
->zl_lock
);
1404 zil_get_commit_list(zilog
);
1407 * Return if there's nothing to commit before we dirty the fs by
1408 * calling zil_create().
1410 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1411 mutex_enter(&zilog
->zl_lock
);
1415 if (zilog
->zl_suspend
) {
1418 lwb
= list_tail(&zilog
->zl_lwb_list
);
1420 lwb
= zil_create(zilog
);
1423 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1424 while (itx
= list_head(&zilog
->zl_itx_commit_list
)) {
1425 txg
= itx
->itx_lr
.lrc_txg
;
1428 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1429 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1430 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1431 kmem_free(itx
, offsetof(itx_t
, itx_lr
)
1432 + itx
->itx_lr
.lrc_reclen
);
1434 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1436 /* write the last block out */
1437 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1438 lwb
= zil_lwb_write_start(zilog
, lwb
);
1440 zilog
->zl_cur_used
= 0;
1443 * Wait if necessary for the log blocks to be on stable storage.
1445 if (zilog
->zl_root_zio
) {
1446 error
= zio_wait(zilog
->zl_root_zio
);
1447 zilog
->zl_root_zio
= NULL
;
1448 zil_flush_vdevs(zilog
);
1451 if (error
|| lwb
== NULL
)
1452 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1454 mutex_enter(&zilog
->zl_lock
);
1457 * Remember the highest committed log sequence number for ztest.
1458 * We only update this value when all the log writes succeeded,
1459 * because ztest wants to ASSERT that it got the whole log chain.
1461 if (error
== 0 && lwb
!= NULL
)
1462 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1466 * Commit zfs transactions to stable storage.
1467 * If foid is 0 push out all transactions, otherwise push only those
1468 * for that object or might reference that object.
1470 * itxs are committed in batches. In a heavily stressed zil there will be
1471 * a commit writer thread who is writing out a bunch of itxs to the log
1472 * for a set of committing threads (cthreads) in the same batch as the writer.
1473 * Those cthreads are all waiting on the same cv for that batch.
1475 * There will also be a different and growing batch of threads that are
1476 * waiting to commit (qthreads). When the committing batch completes
1477 * a transition occurs such that the cthreads exit and the qthreads become
1478 * cthreads. One of the new cthreads becomes the writer thread for the
1479 * batch. Any new threads arriving become new qthreads.
1481 * Only 2 condition variables are needed and there's no transition
1482 * between the two cvs needed. They just flip-flop between qthreads
1485 * Using this scheme we can efficiently wakeup up only those threads
1486 * that have been committed.
1489 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1493 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1496 /* move the async itxs for the foid to the sync queues */
1497 zil_async_to_sync(zilog
, foid
);
1499 mutex_enter(&zilog
->zl_lock
);
1500 mybatch
= zilog
->zl_next_batch
;
1501 while (zilog
->zl_writer
) {
1502 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1503 if (mybatch
<= zilog
->zl_com_batch
) {
1504 mutex_exit(&zilog
->zl_lock
);
1509 zilog
->zl_next_batch
++;
1510 zilog
->zl_writer
= B_TRUE
;
1511 zil_commit_writer(zilog
);
1512 zilog
->zl_com_batch
= mybatch
;
1513 zilog
->zl_writer
= B_FALSE
;
1514 mutex_exit(&zilog
->zl_lock
);
1516 /* wake up one thread to become the next writer */
1517 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1519 /* wake up all threads waiting for this batch to be committed */
1520 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1524 * Called in syncing context to free committed log blocks and update log header.
1527 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1529 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1530 uint64_t txg
= dmu_tx_get_txg(tx
);
1531 spa_t
*spa
= zilog
->zl_spa
;
1532 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1536 * We don't zero out zl_destroy_txg, so make sure we don't try
1537 * to destroy it twice.
1539 if (spa_sync_pass(spa
) != 1)
1542 mutex_enter(&zilog
->zl_lock
);
1544 ASSERT(zilog
->zl_stop_sync
== 0);
1546 if (*replayed_seq
!= 0) {
1547 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1548 zh
->zh_replay_seq
= *replayed_seq
;
1552 if (zilog
->zl_destroy_txg
== txg
) {
1553 blkptr_t blk
= zh
->zh_log
;
1555 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1557 bzero(zh
, sizeof (zil_header_t
));
1558 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1560 if (zilog
->zl_keep_first
) {
1562 * If this block was part of log chain that couldn't
1563 * be claimed because a device was missing during
1564 * zil_claim(), but that device later returns,
1565 * then this block could erroneously appear valid.
1566 * To guard against this, assign a new GUID to the new
1567 * log chain so it doesn't matter what blk points to.
1569 zil_init_log_chain(zilog
, &blk
);
1574 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1575 zh
->zh_log
= lwb
->lwb_blk
;
1576 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1578 list_remove(&zilog
->zl_lwb_list
, lwb
);
1579 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1580 kmem_cache_free(zil_lwb_cache
, lwb
);
1583 * If we don't have anything left in the lwb list then
1584 * we've had an allocation failure and we need to zero
1585 * out the zil_header blkptr so that we don't end
1586 * up freeing the same block twice.
1588 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1589 BP_ZERO(&zh
->zh_log
);
1591 mutex_exit(&zilog
->zl_lock
);
1597 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1598 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1604 kmem_cache_destroy(zil_lwb_cache
);
1608 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1610 zilog
->zl_sync
= sync
;
1614 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1616 zilog
->zl_logbias
= logbias
;
1620 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1625 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_SLEEP
);
1627 zilog
->zl_header
= zh_phys
;
1629 zilog
->zl_spa
= dmu_objset_spa(os
);
1630 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1631 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1632 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1633 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1634 zilog
->zl_next_batch
= 1;
1636 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1638 for (i
= 0; i
< TXG_SIZE
; i
++) {
1639 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1640 MUTEX_DEFAULT
, NULL
);
1643 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1644 offsetof(lwb_t
, lwb_node
));
1646 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1647 offsetof(itx_t
, itx_node
));
1649 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1651 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1652 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1654 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1655 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1656 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1657 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1663 zil_free(zilog_t
*zilog
)
1668 zilog
->zl_stop_sync
= 1;
1671 * After zil_close() there should only be one lwb with a buffer.
1673 head_lwb
= list_head(&zilog
->zl_lwb_list
);
1675 ASSERT(head_lwb
== list_tail(&zilog
->zl_lwb_list
));
1676 list_remove(&zilog
->zl_lwb_list
, head_lwb
);
1677 zio_buf_free(head_lwb
->lwb_buf
, head_lwb
->lwb_sz
);
1678 kmem_cache_free(zil_lwb_cache
, head_lwb
);
1680 list_destroy(&zilog
->zl_lwb_list
);
1682 avl_destroy(&zilog
->zl_vdev_tree
);
1683 mutex_destroy(&zilog
->zl_vdev_lock
);
1685 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1686 list_destroy(&zilog
->zl_itx_commit_list
);
1688 for (i
= 0; i
< TXG_SIZE
; i
++) {
1690 * It's possible for an itx to be generated that doesn't dirty
1691 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1692 * callback to remove the entry. We remove those here.
1694 * Also free up the ziltest itxs.
1696 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1697 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1698 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1701 mutex_destroy(&zilog
->zl_lock
);
1703 cv_destroy(&zilog
->zl_cv_writer
);
1704 cv_destroy(&zilog
->zl_cv_suspend
);
1705 cv_destroy(&zilog
->zl_cv_batch
[0]);
1706 cv_destroy(&zilog
->zl_cv_batch
[1]);
1708 kmem_free(zilog
, sizeof (zilog_t
));
1712 * Open an intent log.
1715 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1717 zilog_t
*zilog
= dmu_objset_zil(os
);
1719 zilog
->zl_get_data
= get_data
;
1720 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, minclsyspri
,
1721 2, 2, TASKQ_PREPOPULATE
);
1727 * Close an intent log.
1730 zil_close(zilog_t
*zilog
)
1735 zil_commit(zilog
, 0); /* commit all itx */
1738 * The lwb_max_txg for the stubby lwb will reflect the last activity
1739 * for the zil. After a txg_wait_synced() on the txg we know all the
1740 * callbacks have occurred that may clean the zil. Only then can we
1741 * destroy the zl_clean_taskq.
1743 mutex_enter(&zilog
->zl_lock
);
1744 tail_lwb
= list_tail(&zilog
->zl_lwb_list
);
1745 if (tail_lwb
!= NULL
)
1746 txg
= tail_lwb
->lwb_max_txg
;
1747 mutex_exit(&zilog
->zl_lock
);
1749 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1751 taskq_destroy(zilog
->zl_clean_taskq
);
1752 zilog
->zl_clean_taskq
= NULL
;
1753 zilog
->zl_get_data
= NULL
;
1757 * Suspend an intent log. While in suspended mode, we still honor
1758 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1759 * We suspend the log briefly when taking a snapshot so that the snapshot
1760 * contains all the data it's supposed to, and has an empty intent log.
1763 zil_suspend(zilog_t
*zilog
)
1765 const zil_header_t
*zh
= zilog
->zl_header
;
1767 mutex_enter(&zilog
->zl_lock
);
1768 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
1769 mutex_exit(&zilog
->zl_lock
);
1772 if (zilog
->zl_suspend
++ != 0) {
1774 * Someone else already began a suspend.
1775 * Just wait for them to finish.
1777 while (zilog
->zl_suspending
)
1778 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
1779 mutex_exit(&zilog
->zl_lock
);
1782 zilog
->zl_suspending
= B_TRUE
;
1783 mutex_exit(&zilog
->zl_lock
);
1785 zil_commit(zilog
, 0);
1787 zil_destroy(zilog
, B_FALSE
);
1789 mutex_enter(&zilog
->zl_lock
);
1790 zilog
->zl_suspending
= B_FALSE
;
1791 cv_broadcast(&zilog
->zl_cv_suspend
);
1792 mutex_exit(&zilog
->zl_lock
);
1798 zil_resume(zilog_t
*zilog
)
1800 mutex_enter(&zilog
->zl_lock
);
1801 ASSERT(zilog
->zl_suspend
!= 0);
1802 zilog
->zl_suspend
--;
1803 mutex_exit(&zilog
->zl_lock
);
1806 typedef struct zil_replay_arg
{
1807 zil_replay_func_t
**zr_replay
;
1809 boolean_t zr_byteswap
;
1814 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
1816 char name
[MAXNAMELEN
];
1818 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
1820 dmu_objset_name(zilog
->zl_os
, name
);
1822 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
1823 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
1824 (u_longlong_t
)lr
->lrc_seq
,
1825 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
1826 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
1832 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
1834 zil_replay_arg_t
*zr
= zra
;
1835 const zil_header_t
*zh
= zilog
->zl_header
;
1836 uint64_t reclen
= lr
->lrc_reclen
;
1837 uint64_t txtype
= lr
->lrc_txtype
;
1840 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
1842 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
1845 if (lr
->lrc_txg
< claim_txg
) /* already committed */
1848 /* Strip case-insensitive bit, still present in log record */
1851 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
1852 return (zil_replay_error(zilog
, lr
, EINVAL
));
1855 * If this record type can be logged out of order, the object
1856 * (lr_foid) may no longer exist. That's legitimate, not an error.
1858 if (TX_OOO(txtype
)) {
1859 error
= dmu_object_info(zilog
->zl_os
,
1860 ((lr_ooo_t
*)lr
)->lr_foid
, NULL
);
1861 if (error
== ENOENT
|| error
== EEXIST
)
1866 * Make a copy of the data so we can revise and extend it.
1868 bcopy(lr
, zr
->zr_lr
, reclen
);
1871 * If this is a TX_WRITE with a blkptr, suck in the data.
1873 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
1874 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
1875 zr
->zr_lr
+ reclen
);
1877 return (zil_replay_error(zilog
, lr
, error
));
1881 * The log block containing this lr may have been byteswapped
1882 * so that we can easily examine common fields like lrc_txtype.
1883 * However, the log is a mix of different record types, and only the
1884 * replay vectors know how to byteswap their records. Therefore, if
1885 * the lr was byteswapped, undo it before invoking the replay vector.
1887 if (zr
->zr_byteswap
)
1888 byteswap_uint64_array(zr
->zr_lr
, reclen
);
1891 * We must now do two things atomically: replay this log record,
1892 * and update the log header sequence number to reflect the fact that
1893 * we did so. At the end of each replay function the sequence number
1894 * is updated if we are in replay mode.
1896 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
1899 * The DMU's dnode layer doesn't see removes until the txg
1900 * commits, so a subsequent claim can spuriously fail with
1901 * EEXIST. So if we receive any error we try syncing out
1902 * any removes then retry the transaction. Note that we
1903 * specify B_FALSE for byteswap now, so we don't do it twice.
1905 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
1906 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
1908 return (zil_replay_error(zilog
, lr
, error
));
1915 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
1917 zilog
->zl_replay_blks
++;
1923 * If this dataset has a non-empty intent log, replay it and destroy it.
1926 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t
*replay_func
[TX_MAX_TYPE
])
1928 zilog_t
*zilog
= dmu_objset_zil(os
);
1929 const zil_header_t
*zh
= zilog
->zl_header
;
1930 zil_replay_arg_t zr
;
1932 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
1933 zil_destroy(zilog
, B_TRUE
);
1937 zr
.zr_replay
= replay_func
;
1939 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
1940 zr
.zr_lr
= kmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_SLEEP
);
1943 * Wait for in-progress removes to sync before starting replay.
1945 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1947 zilog
->zl_replay
= B_TRUE
;
1948 zilog
->zl_replay_time
= ddi_get_lbolt();
1949 ASSERT(zilog
->zl_replay_blks
== 0);
1950 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
1952 kmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
1954 zil_destroy(zilog
, B_FALSE
);
1955 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
1956 zilog
->zl_replay
= B_FALSE
;
1960 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
1962 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1965 if (zilog
->zl_replay
) {
1966 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1967 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
1968 zilog
->zl_replaying_seq
;
1977 zil_vdev_offline(const char *osname
, void *arg
)
1983 error
= dmu_objset_hold(osname
, FTAG
, &os
);
1987 zilog
= dmu_objset_zil(os
);
1988 if (zil_suspend(zilog
) != 0)
1992 dmu_objset_rele(os
, FTAG
);