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, 2014 by Delphix. All rights reserved.
26 /* Portions Copyright 2010 Robert Milkowski */
28 #include <sys/zfs_context.h>
34 #include <sys/resource.h>
36 #include <sys/zil_impl.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/vdev_impl.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/metaslab.h>
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
53 * There is one ZIL per file system. Its on-disk (pool) format consists
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
70 * See zil.h for more information about these fields.
72 zil_stats_t zil_stats
= {
73 { "zil_commit_count", KSTAT_DATA_UINT64
},
74 { "zil_commit_writer_count", KSTAT_DATA_UINT64
},
75 { "zil_itx_count", KSTAT_DATA_UINT64
},
76 { "zil_itx_indirect_count", KSTAT_DATA_UINT64
},
77 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64
},
78 { "zil_itx_copied_count", KSTAT_DATA_UINT64
},
79 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64
},
80 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64
},
81 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64
},
82 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64
},
83 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64
},
84 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64
},
85 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64
},
88 static kstat_t
*zil_ksp
;
91 * Disable intent logging replay. This global ZIL switch affects all pools.
93 int zil_replay_disable
= 0;
96 * Tunable parameter for debugging or performance analysis. Setting
97 * zfs_nocacheflush will cause corruption on power loss if a volatile
98 * out-of-order write cache is enabled.
100 int zfs_nocacheflush
= 0;
102 static kmem_cache_t
*zil_lwb_cache
;
104 static void zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
);
106 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
107 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
111 * ziltest is by and large an ugly hack, but very useful in
112 * checking replay without tedious work.
113 * When running ziltest we want to keep all itx's and so maintain
114 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
115 * We subtract TXG_CONCURRENT_STATES to allow for common code.
117 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
120 zil_bp_compare(const void *x1
, const void *x2
)
122 const dva_t
*dva1
= &((zil_bp_node_t
*)x1
)->zn_dva
;
123 const dva_t
*dva2
= &((zil_bp_node_t
*)x2
)->zn_dva
;
125 if (DVA_GET_VDEV(dva1
) < DVA_GET_VDEV(dva2
))
127 if (DVA_GET_VDEV(dva1
) > DVA_GET_VDEV(dva2
))
130 if (DVA_GET_OFFSET(dva1
) < DVA_GET_OFFSET(dva2
))
132 if (DVA_GET_OFFSET(dva1
) > DVA_GET_OFFSET(dva2
))
139 zil_bp_tree_init(zilog_t
*zilog
)
141 avl_create(&zilog
->zl_bp_tree
, zil_bp_compare
,
142 sizeof (zil_bp_node_t
), offsetof(zil_bp_node_t
, zn_node
));
146 zil_bp_tree_fini(zilog_t
*zilog
)
148 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
152 while ((zn
= avl_destroy_nodes(t
, &cookie
)) != NULL
)
153 kmem_free(zn
, sizeof (zil_bp_node_t
));
159 zil_bp_tree_add(zilog_t
*zilog
, const blkptr_t
*bp
)
161 avl_tree_t
*t
= &zilog
->zl_bp_tree
;
166 if (BP_IS_EMBEDDED(bp
))
169 dva
= BP_IDENTITY(bp
);
171 if (avl_find(t
, dva
, &where
) != NULL
)
172 return (SET_ERROR(EEXIST
));
174 zn
= kmem_alloc(sizeof (zil_bp_node_t
), KM_PUSHPAGE
);
176 avl_insert(t
, zn
, where
);
181 static zil_header_t
*
182 zil_header_in_syncing_context(zilog_t
*zilog
)
184 return ((zil_header_t
*)zilog
->zl_header
);
188 zil_init_log_chain(zilog_t
*zilog
, blkptr_t
*bp
)
190 zio_cksum_t
*zc
= &bp
->blk_cksum
;
192 zc
->zc_word
[ZIL_ZC_GUID_0
] = spa_get_random(-1ULL);
193 zc
->zc_word
[ZIL_ZC_GUID_1
] = spa_get_random(-1ULL);
194 zc
->zc_word
[ZIL_ZC_OBJSET
] = dmu_objset_id(zilog
->zl_os
);
195 zc
->zc_word
[ZIL_ZC_SEQ
] = 1ULL;
199 * Read a log block and make sure it's valid.
202 zil_read_log_block(zilog_t
*zilog
, const blkptr_t
*bp
, blkptr_t
*nbp
, void *dst
,
205 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
206 uint32_t aflags
= ARC_WAIT
;
207 arc_buf_t
*abuf
= NULL
;
211 if (zilog
->zl_header
->zh_claim_txg
== 0)
212 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
214 if (!(zilog
->zl_header
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
215 zio_flags
|= ZIO_FLAG_SPECULATIVE
;
217 SET_BOOKMARK(&zb
, bp
->blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
218 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
, bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
220 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
221 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
224 zio_cksum_t cksum
= bp
->blk_cksum
;
227 * Validate the checksummed log block.
229 * Sequence numbers should be... sequential. The checksum
230 * verifier for the next block should be bp's checksum plus 1.
232 * Also check the log chain linkage and size used.
234 cksum
.zc_word
[ZIL_ZC_SEQ
]++;
236 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
237 zil_chain_t
*zilc
= abuf
->b_data
;
238 char *lr
= (char *)(zilc
+ 1);
239 uint64_t len
= zilc
->zc_nused
- sizeof (zil_chain_t
);
241 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
242 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
)) {
243 error
= SET_ERROR(ECKSUM
);
246 *end
= (char *)dst
+ len
;
247 *nbp
= zilc
->zc_next_blk
;
250 char *lr
= abuf
->b_data
;
251 uint64_t size
= BP_GET_LSIZE(bp
);
252 zil_chain_t
*zilc
= (zil_chain_t
*)(lr
+ size
) - 1;
254 if (bcmp(&cksum
, &zilc
->zc_next_blk
.blk_cksum
,
255 sizeof (cksum
)) || BP_IS_HOLE(&zilc
->zc_next_blk
) ||
256 (zilc
->zc_nused
> (size
- sizeof (*zilc
)))) {
257 error
= SET_ERROR(ECKSUM
);
259 bcopy(lr
, dst
, zilc
->zc_nused
);
260 *end
= (char *)dst
+ zilc
->zc_nused
;
261 *nbp
= zilc
->zc_next_blk
;
265 VERIFY(arc_buf_remove_ref(abuf
, &abuf
));
272 * Read a TX_WRITE log data block.
275 zil_read_log_data(zilog_t
*zilog
, const lr_write_t
*lr
, void *wbuf
)
277 enum zio_flag zio_flags
= ZIO_FLAG_CANFAIL
;
278 const blkptr_t
*bp
= &lr
->lr_blkptr
;
279 uint32_t aflags
= ARC_WAIT
;
280 arc_buf_t
*abuf
= NULL
;
284 if (BP_IS_HOLE(bp
)) {
286 bzero(wbuf
, MAX(BP_GET_LSIZE(bp
), lr
->lr_length
));
290 if (zilog
->zl_header
->zh_claim_txg
== 0)
291 zio_flags
|= ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
;
293 SET_BOOKMARK(&zb
, dmu_objset_id(zilog
->zl_os
), lr
->lr_foid
,
294 ZB_ZIL_LEVEL
, lr
->lr_offset
/ BP_GET_LSIZE(bp
));
296 error
= arc_read(NULL
, zilog
->zl_spa
, bp
, arc_getbuf_func
, &abuf
,
297 ZIO_PRIORITY_SYNC_READ
, zio_flags
, &aflags
, &zb
);
301 bcopy(abuf
->b_data
, wbuf
, arc_buf_size(abuf
));
302 (void) arc_buf_remove_ref(abuf
, &abuf
);
309 * Parse the intent log, and call parse_func for each valid record within.
312 zil_parse(zilog_t
*zilog
, zil_parse_blk_func_t
*parse_blk_func
,
313 zil_parse_lr_func_t
*parse_lr_func
, void *arg
, uint64_t txg
)
315 const zil_header_t
*zh
= zilog
->zl_header
;
316 boolean_t claimed
= !!zh
->zh_claim_txg
;
317 uint64_t claim_blk_seq
= claimed
? zh
->zh_claim_blk_seq
: UINT64_MAX
;
318 uint64_t claim_lr_seq
= claimed
? zh
->zh_claim_lr_seq
: UINT64_MAX
;
319 uint64_t max_blk_seq
= 0;
320 uint64_t max_lr_seq
= 0;
321 uint64_t blk_count
= 0;
322 uint64_t lr_count
= 0;
323 blkptr_t blk
, next_blk
;
327 bzero(&next_blk
, sizeof (blkptr_t
));
330 * Old logs didn't record the maximum zh_claim_lr_seq.
332 if (!(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
))
333 claim_lr_seq
= UINT64_MAX
;
336 * Starting at the block pointed to by zh_log we read the log chain.
337 * For each block in the chain we strongly check that block to
338 * ensure its validity. We stop when an invalid block is found.
339 * For each block pointer in the chain we call parse_blk_func().
340 * For each record in each valid block we call parse_lr_func().
341 * If the log has been claimed, stop if we encounter a sequence
342 * number greater than the highest claimed sequence number.
344 lrbuf
= zio_buf_alloc(SPA_MAXBLOCKSIZE
);
345 zil_bp_tree_init(zilog
);
347 for (blk
= zh
->zh_log
; !BP_IS_HOLE(&blk
); blk
= next_blk
) {
348 uint64_t blk_seq
= blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
];
352 if (blk_seq
> claim_blk_seq
)
354 if ((error
= parse_blk_func(zilog
, &blk
, arg
, txg
)) != 0)
356 ASSERT3U(max_blk_seq
, <, blk_seq
);
357 max_blk_seq
= blk_seq
;
360 if (max_lr_seq
== claim_lr_seq
&& max_blk_seq
== claim_blk_seq
)
363 error
= zil_read_log_block(zilog
, &blk
, &next_blk
, lrbuf
, &end
);
367 for (lrp
= lrbuf
; lrp
< end
; lrp
+= reclen
) {
368 lr_t
*lr
= (lr_t
*)lrp
;
369 reclen
= lr
->lrc_reclen
;
370 ASSERT3U(reclen
, >=, sizeof (lr_t
));
371 if (lr
->lrc_seq
> claim_lr_seq
)
373 if ((error
= parse_lr_func(zilog
, lr
, arg
, txg
)) != 0)
375 ASSERT3U(max_lr_seq
, <, lr
->lrc_seq
);
376 max_lr_seq
= lr
->lrc_seq
;
381 zilog
->zl_parse_error
= error
;
382 zilog
->zl_parse_blk_seq
= max_blk_seq
;
383 zilog
->zl_parse_lr_seq
= max_lr_seq
;
384 zilog
->zl_parse_blk_count
= blk_count
;
385 zilog
->zl_parse_lr_count
= lr_count
;
387 ASSERT(!claimed
|| !(zh
->zh_flags
& ZIL_CLAIM_LR_SEQ_VALID
) ||
388 (max_blk_seq
== claim_blk_seq
&& max_lr_seq
== claim_lr_seq
));
390 zil_bp_tree_fini(zilog
);
391 zio_buf_free(lrbuf
, SPA_MAXBLOCKSIZE
);
397 zil_claim_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t first_txg
)
400 * Claim log block if not already committed and not already claimed.
401 * If tx == NULL, just verify that the block is claimable.
403 if (BP_IS_HOLE(bp
) || bp
->blk_birth
< first_txg
||
404 zil_bp_tree_add(zilog
, bp
) != 0)
407 return (zio_wait(zio_claim(NULL
, zilog
->zl_spa
,
408 tx
== NULL
? 0 : first_txg
, bp
, spa_claim_notify
, NULL
,
409 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_SCRUB
)));
413 zil_claim_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t first_txg
)
415 lr_write_t
*lr
= (lr_write_t
*)lrc
;
418 if (lrc
->lrc_txtype
!= TX_WRITE
)
422 * If the block is not readable, don't claim it. This can happen
423 * in normal operation when a log block is written to disk before
424 * some of the dmu_sync() blocks it points to. In this case, the
425 * transaction cannot have been committed to anyone (we would have
426 * waited for all writes to be stable first), so it is semantically
427 * correct to declare this the end of the log.
429 if (lr
->lr_blkptr
.blk_birth
>= first_txg
&&
430 (error
= zil_read_log_data(zilog
, lr
, NULL
)) != 0)
432 return (zil_claim_log_block(zilog
, &lr
->lr_blkptr
, tx
, first_txg
));
437 zil_free_log_block(zilog_t
*zilog
, blkptr_t
*bp
, void *tx
, uint64_t claim_txg
)
439 zio_free_zil(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
445 zil_free_log_record(zilog_t
*zilog
, lr_t
*lrc
, void *tx
, uint64_t claim_txg
)
447 lr_write_t
*lr
= (lr_write_t
*)lrc
;
448 blkptr_t
*bp
= &lr
->lr_blkptr
;
451 * If we previously claimed it, we need to free it.
453 if (claim_txg
!= 0 && lrc
->lrc_txtype
== TX_WRITE
&&
454 bp
->blk_birth
>= claim_txg
&& zil_bp_tree_add(zilog
, bp
) == 0 &&
456 zio_free(zilog
->zl_spa
, dmu_tx_get_txg(tx
), bp
);
462 zil_alloc_lwb(zilog_t
*zilog
, blkptr_t
*bp
, uint64_t txg
, boolean_t fastwrite
)
466 lwb
= kmem_cache_alloc(zil_lwb_cache
, KM_PUSHPAGE
);
467 lwb
->lwb_zilog
= zilog
;
469 lwb
->lwb_fastwrite
= fastwrite
;
470 lwb
->lwb_buf
= zio_buf_alloc(BP_GET_LSIZE(bp
));
471 lwb
->lwb_max_txg
= txg
;
474 if (BP_GET_CHECKSUM(bp
) == ZIO_CHECKSUM_ZILOG2
) {
475 lwb
->lwb_nused
= sizeof (zil_chain_t
);
476 lwb
->lwb_sz
= BP_GET_LSIZE(bp
);
479 lwb
->lwb_sz
= BP_GET_LSIZE(bp
) - sizeof (zil_chain_t
);
482 mutex_enter(&zilog
->zl_lock
);
483 list_insert_tail(&zilog
->zl_lwb_list
, lwb
);
484 mutex_exit(&zilog
->zl_lock
);
490 * Called when we create in-memory log transactions so that we know
491 * to cleanup the itxs at the end of spa_sync().
494 zilog_dirty(zilog_t
*zilog
, uint64_t txg
)
496 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
497 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
499 if (dsl_dataset_is_snapshot(ds
))
500 panic("dirtying snapshot!");
502 if (txg_list_add(&dp
->dp_dirty_zilogs
, zilog
, txg
)) {
503 /* up the hold count until we can be written out */
504 dmu_buf_add_ref(ds
->ds_dbuf
, zilog
);
509 zilog_is_dirty(zilog_t
*zilog
)
511 dsl_pool_t
*dp
= zilog
->zl_dmu_pool
;
514 for (t
= 0; t
< TXG_SIZE
; t
++) {
515 if (txg_list_member(&dp
->dp_dirty_zilogs
, zilog
, t
))
522 * Create an on-disk intent log.
525 zil_create(zilog_t
*zilog
)
527 const zil_header_t
*zh
= zilog
->zl_header
;
533 boolean_t fastwrite
= FALSE
;
536 * Wait for any previous destroy to complete.
538 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
540 ASSERT(zh
->zh_claim_txg
== 0);
541 ASSERT(zh
->zh_replay_seq
== 0);
546 * Allocate an initial log block if:
547 * - there isn't one already
548 * - the existing block is the wrong endianess
550 if (BP_IS_HOLE(&blk
) || BP_SHOULD_BYTESWAP(&blk
)) {
551 tx
= dmu_tx_create(zilog
->zl_os
);
552 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
553 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
554 txg
= dmu_tx_get_txg(tx
);
556 if (!BP_IS_HOLE(&blk
)) {
557 zio_free_zil(zilog
->zl_spa
, txg
, &blk
);
561 error
= zio_alloc_zil(zilog
->zl_spa
, txg
, &blk
,
562 ZIL_MIN_BLKSZ
, B_TRUE
);
566 zil_init_log_chain(zilog
, &blk
);
570 * Allocate a log write buffer (lwb) for the first log block.
573 lwb
= zil_alloc_lwb(zilog
, &blk
, txg
, fastwrite
);
576 * If we just allocated the first log block, commit our transaction
577 * and wait for zil_sync() to stuff the block poiner into zh_log.
578 * (zh is part of the MOS, so we cannot modify it in open context.)
582 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
585 ASSERT(bcmp(&blk
, &zh
->zh_log
, sizeof (blk
)) == 0);
591 * In one tx, free all log blocks and clear the log header.
592 * If keep_first is set, then we're replaying a log with no content.
593 * We want to keep the first block, however, so that the first
594 * synchronous transaction doesn't require a txg_wait_synced()
595 * in zil_create(). We don't need to txg_wait_synced() here either
596 * when keep_first is set, because both zil_create() and zil_destroy()
597 * will wait for any in-progress destroys to complete.
600 zil_destroy(zilog_t
*zilog
, boolean_t keep_first
)
602 const zil_header_t
*zh
= zilog
->zl_header
;
608 * Wait for any previous destroy to complete.
610 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
612 zilog
->zl_old_header
= *zh
; /* debugging aid */
614 if (BP_IS_HOLE(&zh
->zh_log
))
617 tx
= dmu_tx_create(zilog
->zl_os
);
618 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
619 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
620 txg
= dmu_tx_get_txg(tx
);
622 mutex_enter(&zilog
->zl_lock
);
624 ASSERT3U(zilog
->zl_destroy_txg
, <, txg
);
625 zilog
->zl_destroy_txg
= txg
;
626 zilog
->zl_keep_first
= keep_first
;
628 if (!list_is_empty(&zilog
->zl_lwb_list
)) {
629 ASSERT(zh
->zh_claim_txg
== 0);
631 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
632 ASSERT(lwb
->lwb_zio
== NULL
);
633 if (lwb
->lwb_fastwrite
)
634 metaslab_fastwrite_unmark(zilog
->zl_spa
,
636 list_remove(&zilog
->zl_lwb_list
, lwb
);
637 if (lwb
->lwb_buf
!= NULL
)
638 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
639 zio_free_zil(zilog
->zl_spa
, txg
, &lwb
->lwb_blk
);
640 kmem_cache_free(zil_lwb_cache
, lwb
);
642 } else if (!keep_first
) {
643 zil_destroy_sync(zilog
, tx
);
645 mutex_exit(&zilog
->zl_lock
);
651 zil_destroy_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
653 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
654 (void) zil_parse(zilog
, zil_free_log_block
,
655 zil_free_log_record
, tx
, zilog
->zl_header
->zh_claim_txg
);
659 zil_claim(const char *osname
, void *txarg
)
661 dmu_tx_t
*tx
= txarg
;
662 uint64_t first_txg
= dmu_tx_get_txg(tx
);
668 error
= dmu_objset_own(osname
, DMU_OST_ANY
, B_FALSE
, FTAG
, &os
);
671 * EBUSY indicates that the objset is inconsistent, in which
672 * case it can not have a ZIL.
674 if (error
!= EBUSY
) {
675 cmn_err(CE_WARN
, "can't open objset for %s, error %u",
682 zilog
= dmu_objset_zil(os
);
683 zh
= zil_header_in_syncing_context(zilog
);
685 if (spa_get_log_state(zilog
->zl_spa
) == SPA_LOG_CLEAR
) {
686 if (!BP_IS_HOLE(&zh
->zh_log
))
687 zio_free_zil(zilog
->zl_spa
, first_txg
, &zh
->zh_log
);
688 BP_ZERO(&zh
->zh_log
);
689 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
690 dmu_objset_disown(os
, FTAG
);
695 * Claim all log blocks if we haven't already done so, and remember
696 * the highest claimed sequence number. This ensures that if we can
697 * read only part of the log now (e.g. due to a missing device),
698 * but we can read the entire log later, we will not try to replay
699 * or destroy beyond the last block we successfully claimed.
701 ASSERT3U(zh
->zh_claim_txg
, <=, first_txg
);
702 if (zh
->zh_claim_txg
== 0 && !BP_IS_HOLE(&zh
->zh_log
)) {
703 (void) zil_parse(zilog
, zil_claim_log_block
,
704 zil_claim_log_record
, tx
, first_txg
);
705 zh
->zh_claim_txg
= first_txg
;
706 zh
->zh_claim_blk_seq
= zilog
->zl_parse_blk_seq
;
707 zh
->zh_claim_lr_seq
= zilog
->zl_parse_lr_seq
;
708 if (zilog
->zl_parse_lr_count
|| zilog
->zl_parse_blk_count
> 1)
709 zh
->zh_flags
|= ZIL_REPLAY_NEEDED
;
710 zh
->zh_flags
|= ZIL_CLAIM_LR_SEQ_VALID
;
711 dsl_dataset_dirty(dmu_objset_ds(os
), tx
);
714 ASSERT3U(first_txg
, ==, (spa_last_synced_txg(zilog
->zl_spa
) + 1));
715 dmu_objset_disown(os
, FTAG
);
720 * Check the log by walking the log chain.
721 * Checksum errors are ok as they indicate the end of the chain.
722 * Any other error (no device or read failure) returns an error.
725 zil_check_log_chain(const char *osname
, void *tx
)
734 error
= dmu_objset_hold(osname
, FTAG
, &os
);
736 cmn_err(CE_WARN
, "can't open objset for %s", osname
);
740 zilog
= dmu_objset_zil(os
);
741 bp
= (blkptr_t
*)&zilog
->zl_header
->zh_log
;
744 * Check the first block and determine if it's on a log device
745 * which may have been removed or faulted prior to loading this
746 * pool. If so, there's no point in checking the rest of the log
747 * as its content should have already been synced to the pool.
749 if (!BP_IS_HOLE(bp
)) {
751 boolean_t valid
= B_TRUE
;
753 spa_config_enter(os
->os_spa
, SCL_STATE
, FTAG
, RW_READER
);
754 vd
= vdev_lookup_top(os
->os_spa
, DVA_GET_VDEV(&bp
->blk_dva
[0]));
755 if (vd
->vdev_islog
&& vdev_is_dead(vd
))
756 valid
= vdev_log_state_valid(vd
);
757 spa_config_exit(os
->os_spa
, SCL_STATE
, FTAG
);
760 dmu_objset_rele(os
, FTAG
);
766 * Because tx == NULL, zil_claim_log_block() will not actually claim
767 * any blocks, but just determine whether it is possible to do so.
768 * In addition to checking the log chain, zil_claim_log_block()
769 * will invoke zio_claim() with a done func of spa_claim_notify(),
770 * which will update spa_max_claim_txg. See spa_load() for details.
772 error
= zil_parse(zilog
, zil_claim_log_block
, zil_claim_log_record
, tx
,
773 zilog
->zl_header
->zh_claim_txg
? -1ULL : spa_first_txg(os
->os_spa
));
775 dmu_objset_rele(os
, FTAG
);
777 return ((error
== ECKSUM
|| error
== ENOENT
) ? 0 : error
);
781 zil_vdev_compare(const void *x1
, const void *x2
)
783 const uint64_t v1
= ((zil_vdev_node_t
*)x1
)->zv_vdev
;
784 const uint64_t v2
= ((zil_vdev_node_t
*)x2
)->zv_vdev
;
795 zil_add_block(zilog_t
*zilog
, const blkptr_t
*bp
)
797 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
799 zil_vdev_node_t
*zv
, zvsearch
;
800 int ndvas
= BP_GET_NDVAS(bp
);
803 if (zfs_nocacheflush
)
806 ASSERT(zilog
->zl_writer
);
809 * Even though we're zl_writer, we still need a lock because the
810 * zl_get_data() callbacks may have dmu_sync() done callbacks
811 * that will run concurrently.
813 mutex_enter(&zilog
->zl_vdev_lock
);
814 for (i
= 0; i
< ndvas
; i
++) {
815 zvsearch
.zv_vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
816 if (avl_find(t
, &zvsearch
, &where
) == NULL
) {
817 zv
= kmem_alloc(sizeof (*zv
), KM_PUSHPAGE
);
818 zv
->zv_vdev
= zvsearch
.zv_vdev
;
819 avl_insert(t
, zv
, where
);
822 mutex_exit(&zilog
->zl_vdev_lock
);
826 zil_flush_vdevs(zilog_t
*zilog
)
828 spa_t
*spa
= zilog
->zl_spa
;
829 avl_tree_t
*t
= &zilog
->zl_vdev_tree
;
834 ASSERT(zilog
->zl_writer
);
837 * We don't need zl_vdev_lock here because we're the zl_writer,
838 * and all zl_get_data() callbacks are done.
840 if (avl_numnodes(t
) == 0)
843 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
845 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
847 while ((zv
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
848 vdev_t
*vd
= vdev_lookup_top(spa
, zv
->zv_vdev
);
851 kmem_free(zv
, sizeof (*zv
));
855 * Wait for all the flushes to complete. Not all devices actually
856 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
858 (void) zio_wait(zio
);
860 spa_config_exit(spa
, SCL_STATE
, FTAG
);
864 * Function called when a log block write completes
867 zil_lwb_write_done(zio_t
*zio
)
869 lwb_t
*lwb
= zio
->io_private
;
870 zilog_t
*zilog
= lwb
->lwb_zilog
;
871 dmu_tx_t
*tx
= lwb
->lwb_tx
;
873 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
874 ASSERT(BP_GET_TYPE(zio
->io_bp
) == DMU_OT_INTENT_LOG
);
875 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
876 ASSERT(BP_GET_BYTEORDER(zio
->io_bp
) == ZFS_HOST_BYTEORDER
);
877 ASSERT(!BP_IS_GANG(zio
->io_bp
));
878 ASSERT(!BP_IS_HOLE(zio
->io_bp
));
879 ASSERT(BP_GET_FILL(zio
->io_bp
) == 0);
882 * Ensure the lwb buffer pointer is cleared before releasing
883 * the txg. If we have had an allocation failure and
884 * the txg is waiting to sync then we want want zil_sync()
885 * to remove the lwb so that it's not picked up as the next new
886 * one in zil_commit_writer(). zil_sync() will only remove
887 * the lwb if lwb_buf is null.
889 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
890 mutex_enter(&zilog
->zl_lock
);
892 lwb
->lwb_fastwrite
= FALSE
;
895 mutex_exit(&zilog
->zl_lock
);
898 * Now that we've written this log block, we have a stable pointer
899 * to the next block in the chain, so it's OK to let the txg in
900 * which we allocated the next block sync.
906 * Initialize the io for a log block.
909 zil_lwb_write_init(zilog_t
*zilog
, lwb_t
*lwb
)
913 SET_BOOKMARK(&zb
, lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_OBJSET
],
914 ZB_ZIL_OBJECT
, ZB_ZIL_LEVEL
,
915 lwb
->lwb_blk
.blk_cksum
.zc_word
[ZIL_ZC_SEQ
]);
917 if (zilog
->zl_root_zio
== NULL
) {
918 zilog
->zl_root_zio
= zio_root(zilog
->zl_spa
, NULL
, NULL
,
922 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
923 mutex_enter(&zilog
->zl_lock
);
924 if (lwb
->lwb_zio
== NULL
) {
925 if (!lwb
->lwb_fastwrite
) {
926 metaslab_fastwrite_mark(zilog
->zl_spa
, &lwb
->lwb_blk
);
927 lwb
->lwb_fastwrite
= 1;
929 lwb
->lwb_zio
= zio_rewrite(zilog
->zl_root_zio
, zilog
->zl_spa
,
930 0, &lwb
->lwb_blk
, lwb
->lwb_buf
, BP_GET_LSIZE(&lwb
->lwb_blk
),
931 zil_lwb_write_done
, lwb
, ZIO_PRIORITY_SYNC_WRITE
,
932 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
|
933 ZIO_FLAG_FASTWRITE
, &zb
);
935 mutex_exit(&zilog
->zl_lock
);
939 * Define a limited set of intent log block sizes.
941 * These must be a multiple of 4KB. Note only the amount used (again
942 * aligned to 4KB) actually gets written. However, we can't always just
943 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
945 uint64_t zil_block_buckets
[] = {
946 4096, /* non TX_WRITE */
947 8192+4096, /* data base */
948 32*1024 + 4096, /* NFS writes */
953 * Use the slog as long as the current commit size is less than the
954 * limit or the total list size is less than 2X the limit. Limit
955 * checking is disabled by setting zil_slog_limit to UINT64_MAX.
957 unsigned long zil_slog_limit
= 1024 * 1024;
958 #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \
959 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))
962 * Start a log block write and advance to the next log block.
963 * Calls are serialized.
966 zil_lwb_write_start(zilog_t
*zilog
, lwb_t
*lwb
)
970 spa_t
*spa
= zilog
->zl_spa
;
974 uint64_t zil_blksz
, wsz
;
978 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
979 zilc
= (zil_chain_t
*)lwb
->lwb_buf
;
980 bp
= &zilc
->zc_next_blk
;
982 zilc
= (zil_chain_t
*)(lwb
->lwb_buf
+ lwb
->lwb_sz
);
983 bp
= &zilc
->zc_next_blk
;
986 ASSERT(lwb
->lwb_nused
<= lwb
->lwb_sz
);
989 * Allocate the next block and save its address in this block
990 * before writing it in order to establish the log chain.
991 * Note that if the allocation of nlwb synced before we wrote
992 * the block that points at it (lwb), we'd leak it if we crashed.
993 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
994 * We dirty the dataset to ensure that zil_sync() will be called
995 * to clean up in the event of allocation failure or I/O failure.
997 tx
= dmu_tx_create(zilog
->zl_os
);
998 VERIFY(dmu_tx_assign(tx
, TXG_WAIT
) == 0);
999 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
1000 txg
= dmu_tx_get_txg(tx
);
1005 * Log blocks are pre-allocated. Here we select the size of the next
1006 * block, based on size used in the last block.
1007 * - first find the smallest bucket that will fit the block from a
1008 * limited set of block sizes. This is because it's faster to write
1009 * blocks allocated from the same metaslab as they are adjacent or
1011 * - next find the maximum from the new suggested size and an array of
1012 * previous sizes. This lessens a picket fence effect of wrongly
1013 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1016 * Note we only write what is used, but we can't just allocate
1017 * the maximum block size because we can exhaust the available
1020 zil_blksz
= zilog
->zl_cur_used
+ sizeof (zil_chain_t
);
1021 for (i
= 0; zil_blksz
> zil_block_buckets
[i
]; i
++)
1023 zil_blksz
= zil_block_buckets
[i
];
1024 if (zil_blksz
== UINT64_MAX
)
1025 zil_blksz
= SPA_MAXBLOCKSIZE
;
1026 zilog
->zl_prev_blks
[zilog
->zl_prev_rotor
] = zil_blksz
;
1027 for (i
= 0; i
< ZIL_PREV_BLKS
; i
++)
1028 zil_blksz
= MAX(zil_blksz
, zilog
->zl_prev_blks
[i
]);
1029 zilog
->zl_prev_rotor
= (zilog
->zl_prev_rotor
+ 1) & (ZIL_PREV_BLKS
- 1);
1032 use_slog
= USE_SLOG(zilog
);
1033 error
= zio_alloc_zil(spa
, txg
, bp
, zil_blksz
,
1036 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count
);
1037 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes
, lwb
->lwb_nused
);
1039 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count
);
1040 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes
, lwb
->lwb_nused
);
1043 ASSERT3U(bp
->blk_birth
, ==, txg
);
1044 bp
->blk_cksum
= lwb
->lwb_blk
.blk_cksum
;
1045 bp
->blk_cksum
.zc_word
[ZIL_ZC_SEQ
]++;
1048 * Allocate a new log write buffer (lwb).
1050 nlwb
= zil_alloc_lwb(zilog
, bp
, txg
, TRUE
);
1052 /* Record the block for later vdev flushing */
1053 zil_add_block(zilog
, &lwb
->lwb_blk
);
1056 if (BP_GET_CHECKSUM(&lwb
->lwb_blk
) == ZIO_CHECKSUM_ZILOG2
) {
1057 /* For Slim ZIL only write what is used. */
1058 wsz
= P2ROUNDUP_TYPED(lwb
->lwb_nused
, ZIL_MIN_BLKSZ
, uint64_t);
1059 ASSERT3U(wsz
, <=, lwb
->lwb_sz
);
1060 zio_shrink(lwb
->lwb_zio
, wsz
);
1067 zilc
->zc_nused
= lwb
->lwb_nused
;
1068 zilc
->zc_eck
.zec_cksum
= lwb
->lwb_blk
.blk_cksum
;
1071 * clear unused data for security
1073 bzero(lwb
->lwb_buf
+ lwb
->lwb_nused
, wsz
- lwb
->lwb_nused
);
1075 zio_nowait(lwb
->lwb_zio
); /* Kick off the write for the old log block */
1078 * If there was an allocation failure then nlwb will be null which
1079 * forces a txg_wait_synced().
1085 zil_lwb_commit(zilog_t
*zilog
, itx_t
*itx
, lwb_t
*lwb
)
1087 lr_t
*lrc
= &itx
->itx_lr
; /* common log record */
1088 lr_write_t
*lrw
= (lr_write_t
*)lrc
;
1090 uint64_t txg
= lrc
->lrc_txg
;
1091 uint64_t reclen
= lrc
->lrc_reclen
;
1097 ASSERT(lwb
->lwb_buf
!= NULL
);
1098 ASSERT(zilog_is_dirty(zilog
) ||
1099 spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
);
1101 if (lrc
->lrc_txtype
== TX_WRITE
&& itx
->itx_wr_state
== WR_NEED_COPY
)
1102 dlen
= P2ROUNDUP_TYPED(
1103 lrw
->lr_length
, sizeof (uint64_t), uint64_t);
1105 zilog
->zl_cur_used
+= (reclen
+ dlen
);
1107 zil_lwb_write_init(zilog
, lwb
);
1110 * If this record won't fit in the current log block, start a new one.
1112 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1113 lwb
= zil_lwb_write_start(zilog
, lwb
);
1116 zil_lwb_write_init(zilog
, lwb
);
1117 ASSERT(LWB_EMPTY(lwb
));
1118 if (lwb
->lwb_nused
+ reclen
+ dlen
> lwb
->lwb_sz
) {
1119 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1124 lr_buf
= lwb
->lwb_buf
+ lwb
->lwb_nused
;
1125 bcopy(lrc
, lr_buf
, reclen
);
1126 lrc
= (lr_t
*)lr_buf
;
1127 lrw
= (lr_write_t
*)lrc
;
1129 ZIL_STAT_BUMP(zil_itx_count
);
1132 * If it's a write, fetch the data or get its blkptr as appropriate.
1134 if (lrc
->lrc_txtype
== TX_WRITE
) {
1135 if (txg
> spa_freeze_txg(zilog
->zl_spa
))
1136 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1137 if (itx
->itx_wr_state
== WR_COPIED
) {
1138 ZIL_STAT_BUMP(zil_itx_copied_count
);
1139 ZIL_STAT_INCR(zil_itx_copied_bytes
, lrw
->lr_length
);
1145 ASSERT(itx
->itx_wr_state
== WR_NEED_COPY
);
1146 dbuf
= lr_buf
+ reclen
;
1147 lrw
->lr_common
.lrc_reclen
+= dlen
;
1148 ZIL_STAT_BUMP(zil_itx_needcopy_count
);
1149 ZIL_STAT_INCR(zil_itx_needcopy_bytes
,
1152 ASSERT(itx
->itx_wr_state
== WR_INDIRECT
);
1154 ZIL_STAT_BUMP(zil_itx_indirect_count
);
1155 ZIL_STAT_INCR(zil_itx_indirect_bytes
,
1158 error
= zilog
->zl_get_data(
1159 itx
->itx_private
, lrw
, dbuf
, lwb
->lwb_zio
);
1161 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1165 ASSERT(error
== ENOENT
|| error
== EEXIST
||
1173 * We're actually making an entry, so update lrc_seq to be the
1174 * log record sequence number. Note that this is generally not
1175 * equal to the itx sequence number because not all transactions
1176 * are synchronous, and sometimes spa_sync() gets there first.
1178 lrc
->lrc_seq
= ++zilog
->zl_lr_seq
; /* we are single threaded */
1179 lwb
->lwb_nused
+= reclen
+ dlen
;
1180 lwb
->lwb_max_txg
= MAX(lwb
->lwb_max_txg
, txg
);
1181 ASSERT3U(lwb
->lwb_nused
, <=, lwb
->lwb_sz
);
1182 ASSERT0(P2PHASE(lwb
->lwb_nused
, sizeof (uint64_t)));
1188 zil_itx_create(uint64_t txtype
, size_t lrsize
)
1192 lrsize
= P2ROUNDUP_TYPED(lrsize
, sizeof (uint64_t), size_t);
1194 itx
= kmem_alloc(offsetof(itx_t
, itx_lr
) + lrsize
,
1195 KM_PUSHPAGE
| KM_NODEBUG
);
1196 itx
->itx_lr
.lrc_txtype
= txtype
;
1197 itx
->itx_lr
.lrc_reclen
= lrsize
;
1198 itx
->itx_sod
= lrsize
; /* if write & WR_NEED_COPY will be increased */
1199 itx
->itx_lr
.lrc_seq
= 0; /* defensive */
1200 itx
->itx_sync
= B_TRUE
; /* default is synchronous */
1201 itx
->itx_callback
= NULL
;
1202 itx
->itx_callback_data
= NULL
;
1208 zil_itx_destroy(itx_t
*itx
)
1210 kmem_free(itx
, offsetof(itx_t
, itx_lr
) + itx
->itx_lr
.lrc_reclen
);
1214 * Free up the sync and async itxs. The itxs_t has already been detached
1215 * so no locks are needed.
1218 zil_itxg_clean(itxs_t
*itxs
)
1224 itx_async_node_t
*ian
;
1226 list
= &itxs
->i_sync_list
;
1227 while ((itx
= list_head(list
)) != NULL
) {
1228 if (itx
->itx_callback
!= NULL
)
1229 itx
->itx_callback(itx
->itx_callback_data
);
1230 list_remove(list
, itx
);
1231 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1232 itx
->itx_lr
.lrc_reclen
);
1236 t
= &itxs
->i_async_tree
;
1237 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1238 list
= &ian
->ia_list
;
1239 while ((itx
= list_head(list
)) != NULL
) {
1240 if (itx
->itx_callback
!= NULL
)
1241 itx
->itx_callback(itx
->itx_callback_data
);
1242 list_remove(list
, itx
);
1243 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1244 itx
->itx_lr
.lrc_reclen
);
1247 kmem_free(ian
, sizeof (itx_async_node_t
));
1251 kmem_free(itxs
, sizeof (itxs_t
));
1255 zil_aitx_compare(const void *x1
, const void *x2
)
1257 const uint64_t o1
= ((itx_async_node_t
*)x1
)->ia_foid
;
1258 const uint64_t o2
= ((itx_async_node_t
*)x2
)->ia_foid
;
1269 * Remove all async itx with the given oid.
1272 zil_remove_async(zilog_t
*zilog
, uint64_t oid
)
1275 itx_async_node_t
*ian
;
1282 list_create(&clean_list
, sizeof (itx_t
), offsetof(itx_t
, itx_node
));
1284 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1287 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1289 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1290 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1292 mutex_enter(&itxg
->itxg_lock
);
1293 if (itxg
->itxg_txg
!= txg
) {
1294 mutex_exit(&itxg
->itxg_lock
);
1299 * Locate the object node and append its list.
1301 t
= &itxg
->itxg_itxs
->i_async_tree
;
1302 ian
= avl_find(t
, &oid
, &where
);
1304 list_move_tail(&clean_list
, &ian
->ia_list
);
1305 mutex_exit(&itxg
->itxg_lock
);
1307 while ((itx
= list_head(&clean_list
)) != NULL
) {
1308 if (itx
->itx_callback
!= NULL
)
1309 itx
->itx_callback(itx
->itx_callback_data
);
1310 list_remove(&clean_list
, itx
);
1311 kmem_free(itx
, offsetof(itx_t
, itx_lr
) +
1312 itx
->itx_lr
.lrc_reclen
);
1314 list_destroy(&clean_list
);
1318 zil_itx_assign(zilog_t
*zilog
, itx_t
*itx
, dmu_tx_t
*tx
)
1322 itxs_t
*itxs
, *clean
= NULL
;
1325 * Object ids can be re-instantiated in the next txg so
1326 * remove any async transactions to avoid future leaks.
1327 * This can happen if a fsync occurs on the re-instantiated
1328 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1329 * the new file data and flushes a write record for the old object.
1331 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_REMOVE
)
1332 zil_remove_async(zilog
, itx
->itx_oid
);
1335 * Ensure the data of a renamed file is committed before the rename.
1337 if ((itx
->itx_lr
.lrc_txtype
& ~TX_CI
) == TX_RENAME
)
1338 zil_async_to_sync(zilog
, itx
->itx_oid
);
1340 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
)
1343 txg
= dmu_tx_get_txg(tx
);
1345 itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1346 mutex_enter(&itxg
->itxg_lock
);
1347 itxs
= itxg
->itxg_itxs
;
1348 if (itxg
->itxg_txg
!= txg
) {
1351 * The zil_clean callback hasn't got around to cleaning
1352 * this itxg. Save the itxs for release below.
1353 * This should be rare.
1355 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1357 clean
= itxg
->itxg_itxs
;
1359 ASSERT(itxg
->itxg_sod
== 0);
1360 itxg
->itxg_txg
= txg
;
1361 itxs
= itxg
->itxg_itxs
= kmem_zalloc(sizeof (itxs_t
),
1364 list_create(&itxs
->i_sync_list
, sizeof (itx_t
),
1365 offsetof(itx_t
, itx_node
));
1366 avl_create(&itxs
->i_async_tree
, zil_aitx_compare
,
1367 sizeof (itx_async_node_t
),
1368 offsetof(itx_async_node_t
, ia_node
));
1370 if (itx
->itx_sync
) {
1371 list_insert_tail(&itxs
->i_sync_list
, itx
);
1372 atomic_add_64(&zilog
->zl_itx_list_sz
, itx
->itx_sod
);
1373 itxg
->itxg_sod
+= itx
->itx_sod
;
1375 avl_tree_t
*t
= &itxs
->i_async_tree
;
1376 uint64_t foid
= ((lr_ooo_t
*)&itx
->itx_lr
)->lr_foid
;
1377 itx_async_node_t
*ian
;
1380 ian
= avl_find(t
, &foid
, &where
);
1382 ian
= kmem_alloc(sizeof (itx_async_node_t
),
1384 list_create(&ian
->ia_list
, sizeof (itx_t
),
1385 offsetof(itx_t
, itx_node
));
1386 ian
->ia_foid
= foid
;
1387 avl_insert(t
, ian
, where
);
1389 list_insert_tail(&ian
->ia_list
, itx
);
1392 itx
->itx_lr
.lrc_txg
= dmu_tx_get_txg(tx
);
1393 zilog_dirty(zilog
, txg
);
1394 mutex_exit(&itxg
->itxg_lock
);
1396 /* Release the old itxs now we've dropped the lock */
1398 zil_itxg_clean(clean
);
1402 * If there are any in-memory intent log transactions which have now been
1403 * synced then start up a taskq to free them. We should only do this after we
1404 * have written out the uberblocks (i.e. txg has been comitted) so that
1405 * don't inadvertently clean out in-memory log records that would be required
1409 zil_clean(zilog_t
*zilog
, uint64_t synced_txg
)
1411 itxg_t
*itxg
= &zilog
->zl_itxg
[synced_txg
& TXG_MASK
];
1414 mutex_enter(&itxg
->itxg_lock
);
1415 if (itxg
->itxg_itxs
== NULL
|| itxg
->itxg_txg
== ZILTEST_TXG
) {
1416 mutex_exit(&itxg
->itxg_lock
);
1419 ASSERT3U(itxg
->itxg_txg
, <=, synced_txg
);
1420 ASSERT(itxg
->itxg_txg
!= 0);
1421 ASSERT(zilog
->zl_clean_taskq
!= NULL
);
1422 atomic_add_64(&zilog
->zl_itx_list_sz
, -itxg
->itxg_sod
);
1424 clean_me
= itxg
->itxg_itxs
;
1425 itxg
->itxg_itxs
= NULL
;
1427 mutex_exit(&itxg
->itxg_lock
);
1429 * Preferably start a task queue to free up the old itxs but
1430 * if taskq_dispatch can't allocate resources to do that then
1431 * free it in-line. This should be rare. Note, using TQ_SLEEP
1432 * created a bad performance problem.
1434 if (taskq_dispatch(zilog
->zl_clean_taskq
,
1435 (void (*)(void *))zil_itxg_clean
, clean_me
, TQ_NOSLEEP
) == 0)
1436 zil_itxg_clean(clean_me
);
1440 * Get the list of itxs to commit into zl_itx_commit_list.
1443 zil_get_commit_list(zilog_t
*zilog
)
1446 list_t
*commit_list
= &zilog
->zl_itx_commit_list
;
1447 uint64_t push_sod
= 0;
1449 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1452 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1454 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1455 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1457 mutex_enter(&itxg
->itxg_lock
);
1458 if (itxg
->itxg_txg
!= txg
) {
1459 mutex_exit(&itxg
->itxg_lock
);
1463 list_move_tail(commit_list
, &itxg
->itxg_itxs
->i_sync_list
);
1464 push_sod
+= itxg
->itxg_sod
;
1467 mutex_exit(&itxg
->itxg_lock
);
1469 atomic_add_64(&zilog
->zl_itx_list_sz
, -push_sod
);
1473 * Move the async itxs for a specified object to commit into sync lists.
1476 zil_async_to_sync(zilog_t
*zilog
, uint64_t foid
)
1479 itx_async_node_t
*ian
;
1483 if (spa_freeze_txg(zilog
->zl_spa
) != UINT64_MAX
) /* ziltest support */
1486 otxg
= spa_last_synced_txg(zilog
->zl_spa
) + 1;
1488 for (txg
= otxg
; txg
< (otxg
+ TXG_CONCURRENT_STATES
); txg
++) {
1489 itxg_t
*itxg
= &zilog
->zl_itxg
[txg
& TXG_MASK
];
1491 mutex_enter(&itxg
->itxg_lock
);
1492 if (itxg
->itxg_txg
!= txg
) {
1493 mutex_exit(&itxg
->itxg_lock
);
1498 * If a foid is specified then find that node and append its
1499 * list. Otherwise walk the tree appending all the lists
1500 * to the sync list. We add to the end rather than the
1501 * beginning to ensure the create has happened.
1503 t
= &itxg
->itxg_itxs
->i_async_tree
;
1505 ian
= avl_find(t
, &foid
, &where
);
1507 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1511 void *cookie
= NULL
;
1513 while ((ian
= avl_destroy_nodes(t
, &cookie
)) != NULL
) {
1514 list_move_tail(&itxg
->itxg_itxs
->i_sync_list
,
1516 list_destroy(&ian
->ia_list
);
1517 kmem_free(ian
, sizeof (itx_async_node_t
));
1520 mutex_exit(&itxg
->itxg_lock
);
1525 zil_commit_writer(zilog_t
*zilog
)
1530 spa_t
*spa
= zilog
->zl_spa
;
1533 ASSERT(zilog
->zl_root_zio
== NULL
);
1535 mutex_exit(&zilog
->zl_lock
);
1537 zil_get_commit_list(zilog
);
1540 * Return if there's nothing to commit before we dirty the fs by
1541 * calling zil_create().
1543 if (list_head(&zilog
->zl_itx_commit_list
) == NULL
) {
1544 mutex_enter(&zilog
->zl_lock
);
1548 if (zilog
->zl_suspend
) {
1551 lwb
= list_tail(&zilog
->zl_lwb_list
);
1553 lwb
= zil_create(zilog
);
1556 DTRACE_PROBE1(zil__cw1
, zilog_t
*, zilog
);
1557 for (itx
= list_head(&zilog
->zl_itx_commit_list
); itx
!= NULL
;
1558 itx
= list_next(&zilog
->zl_itx_commit_list
, itx
)) {
1559 txg
= itx
->itx_lr
.lrc_txg
;
1562 if (txg
> spa_last_synced_txg(spa
) || txg
> spa_freeze_txg(spa
))
1563 lwb
= zil_lwb_commit(zilog
, itx
, lwb
);
1565 DTRACE_PROBE1(zil__cw2
, zilog_t
*, zilog
);
1567 /* write the last block out */
1568 if (lwb
!= NULL
&& lwb
->lwb_zio
!= NULL
)
1569 lwb
= zil_lwb_write_start(zilog
, lwb
);
1571 zilog
->zl_cur_used
= 0;
1574 * Wait if necessary for the log blocks to be on stable storage.
1576 if (zilog
->zl_root_zio
) {
1577 error
= zio_wait(zilog
->zl_root_zio
);
1578 zilog
->zl_root_zio
= NULL
;
1579 zil_flush_vdevs(zilog
);
1582 if (error
|| lwb
== NULL
)
1583 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
1585 while ((itx
= list_head(&zilog
->zl_itx_commit_list
))) {
1586 txg
= itx
->itx_lr
.lrc_txg
;
1589 if (itx
->itx_callback
!= NULL
)
1590 itx
->itx_callback(itx
->itx_callback_data
);
1591 list_remove(&zilog
->zl_itx_commit_list
, itx
);
1592 kmem_free(itx
, offsetof(itx_t
, itx_lr
)
1593 + itx
->itx_lr
.lrc_reclen
);
1596 mutex_enter(&zilog
->zl_lock
);
1599 * Remember the highest committed log sequence number for ztest.
1600 * We only update this value when all the log writes succeeded,
1601 * because ztest wants to ASSERT that it got the whole log chain.
1603 if (error
== 0 && lwb
!= NULL
)
1604 zilog
->zl_commit_lr_seq
= zilog
->zl_lr_seq
;
1608 * Commit zfs transactions to stable storage.
1609 * If foid is 0 push out all transactions, otherwise push only those
1610 * for that object or might reference that object.
1612 * itxs are committed in batches. In a heavily stressed zil there will be
1613 * a commit writer thread who is writing out a bunch of itxs to the log
1614 * for a set of committing threads (cthreads) in the same batch as the writer.
1615 * Those cthreads are all waiting on the same cv for that batch.
1617 * There will also be a different and growing batch of threads that are
1618 * waiting to commit (qthreads). When the committing batch completes
1619 * a transition occurs such that the cthreads exit and the qthreads become
1620 * cthreads. One of the new cthreads becomes the writer thread for the
1621 * batch. Any new threads arriving become new qthreads.
1623 * Only 2 condition variables are needed and there's no transition
1624 * between the two cvs needed. They just flip-flop between qthreads
1627 * Using this scheme we can efficiently wakeup up only those threads
1628 * that have been committed.
1631 zil_commit(zilog_t
*zilog
, uint64_t foid
)
1635 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
1638 ZIL_STAT_BUMP(zil_commit_count
);
1640 /* move the async itxs for the foid to the sync queues */
1641 zil_async_to_sync(zilog
, foid
);
1643 mutex_enter(&zilog
->zl_lock
);
1644 mybatch
= zilog
->zl_next_batch
;
1645 while (zilog
->zl_writer
) {
1646 cv_wait(&zilog
->zl_cv_batch
[mybatch
& 1], &zilog
->zl_lock
);
1647 if (mybatch
<= zilog
->zl_com_batch
) {
1648 mutex_exit(&zilog
->zl_lock
);
1653 zilog
->zl_next_batch
++;
1654 zilog
->zl_writer
= B_TRUE
;
1655 ZIL_STAT_BUMP(zil_commit_writer_count
);
1656 zil_commit_writer(zilog
);
1657 zilog
->zl_com_batch
= mybatch
;
1658 zilog
->zl_writer
= B_FALSE
;
1660 /* wake up one thread to become the next writer */
1661 cv_signal(&zilog
->zl_cv_batch
[(mybatch
+1) & 1]);
1663 /* wake up all threads waiting for this batch to be committed */
1664 cv_broadcast(&zilog
->zl_cv_batch
[mybatch
& 1]);
1666 mutex_exit(&zilog
->zl_lock
);
1670 * Called in syncing context to free committed log blocks and update log header.
1673 zil_sync(zilog_t
*zilog
, dmu_tx_t
*tx
)
1675 zil_header_t
*zh
= zil_header_in_syncing_context(zilog
);
1676 uint64_t txg
= dmu_tx_get_txg(tx
);
1677 spa_t
*spa
= zilog
->zl_spa
;
1678 uint64_t *replayed_seq
= &zilog
->zl_replayed_seq
[txg
& TXG_MASK
];
1682 * We don't zero out zl_destroy_txg, so make sure we don't try
1683 * to destroy it twice.
1685 if (spa_sync_pass(spa
) != 1)
1688 mutex_enter(&zilog
->zl_lock
);
1690 ASSERT(zilog
->zl_stop_sync
== 0);
1692 if (*replayed_seq
!= 0) {
1693 ASSERT(zh
->zh_replay_seq
< *replayed_seq
);
1694 zh
->zh_replay_seq
= *replayed_seq
;
1698 if (zilog
->zl_destroy_txg
== txg
) {
1699 blkptr_t blk
= zh
->zh_log
;
1701 ASSERT(list_head(&zilog
->zl_lwb_list
) == NULL
);
1703 bzero(zh
, sizeof (zil_header_t
));
1704 bzero(zilog
->zl_replayed_seq
, sizeof (zilog
->zl_replayed_seq
));
1706 if (zilog
->zl_keep_first
) {
1708 * If this block was part of log chain that couldn't
1709 * be claimed because a device was missing during
1710 * zil_claim(), but that device later returns,
1711 * then this block could erroneously appear valid.
1712 * To guard against this, assign a new GUID to the new
1713 * log chain so it doesn't matter what blk points to.
1715 zil_init_log_chain(zilog
, &blk
);
1720 while ((lwb
= list_head(&zilog
->zl_lwb_list
)) != NULL
) {
1721 zh
->zh_log
= lwb
->lwb_blk
;
1722 if (lwb
->lwb_buf
!= NULL
|| lwb
->lwb_max_txg
> txg
)
1725 ASSERT(lwb
->lwb_zio
== NULL
);
1727 list_remove(&zilog
->zl_lwb_list
, lwb
);
1728 zio_free_zil(spa
, txg
, &lwb
->lwb_blk
);
1729 kmem_cache_free(zil_lwb_cache
, lwb
);
1732 * If we don't have anything left in the lwb list then
1733 * we've had an allocation failure and we need to zero
1734 * out the zil_header blkptr so that we don't end
1735 * up freeing the same block twice.
1737 if (list_head(&zilog
->zl_lwb_list
) == NULL
)
1738 BP_ZERO(&zh
->zh_log
);
1742 * Remove fastwrite on any blocks that have been pre-allocated for
1743 * the next commit. This prevents fastwrite counter pollution by
1744 * unused, long-lived LWBs.
1746 for (; lwb
!= NULL
; lwb
= list_next(&zilog
->zl_lwb_list
, lwb
)) {
1747 if (lwb
->lwb_fastwrite
&& !lwb
->lwb_zio
) {
1748 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1749 lwb
->lwb_fastwrite
= 0;
1753 mutex_exit(&zilog
->zl_lock
);
1759 zil_lwb_cache
= kmem_cache_create("zil_lwb_cache",
1760 sizeof (struct lwb
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1762 zil_ksp
= kstat_create("zfs", 0, "zil", "misc",
1763 KSTAT_TYPE_NAMED
, sizeof (zil_stats
) / sizeof (kstat_named_t
),
1764 KSTAT_FLAG_VIRTUAL
);
1766 if (zil_ksp
!= NULL
) {
1767 zil_ksp
->ks_data
= &zil_stats
;
1768 kstat_install(zil_ksp
);
1775 kmem_cache_destroy(zil_lwb_cache
);
1777 if (zil_ksp
!= NULL
) {
1778 kstat_delete(zil_ksp
);
1784 zil_set_sync(zilog_t
*zilog
, uint64_t sync
)
1786 zilog
->zl_sync
= sync
;
1790 zil_set_logbias(zilog_t
*zilog
, uint64_t logbias
)
1792 zilog
->zl_logbias
= logbias
;
1796 zil_alloc(objset_t
*os
, zil_header_t
*zh_phys
)
1801 zilog
= kmem_zalloc(sizeof (zilog_t
), KM_PUSHPAGE
);
1803 zilog
->zl_header
= zh_phys
;
1805 zilog
->zl_spa
= dmu_objset_spa(os
);
1806 zilog
->zl_dmu_pool
= dmu_objset_pool(os
);
1807 zilog
->zl_destroy_txg
= TXG_INITIAL
- 1;
1808 zilog
->zl_logbias
= dmu_objset_logbias(os
);
1809 zilog
->zl_sync
= dmu_objset_syncprop(os
);
1810 zilog
->zl_next_batch
= 1;
1812 mutex_init(&zilog
->zl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1814 for (i
= 0; i
< TXG_SIZE
; i
++) {
1815 mutex_init(&zilog
->zl_itxg
[i
].itxg_lock
, NULL
,
1816 MUTEX_DEFAULT
, NULL
);
1819 list_create(&zilog
->zl_lwb_list
, sizeof (lwb_t
),
1820 offsetof(lwb_t
, lwb_node
));
1822 list_create(&zilog
->zl_itx_commit_list
, sizeof (itx_t
),
1823 offsetof(itx_t
, itx_node
));
1825 mutex_init(&zilog
->zl_vdev_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1827 avl_create(&zilog
->zl_vdev_tree
, zil_vdev_compare
,
1828 sizeof (zil_vdev_node_t
), offsetof(zil_vdev_node_t
, zv_node
));
1830 cv_init(&zilog
->zl_cv_writer
, NULL
, CV_DEFAULT
, NULL
);
1831 cv_init(&zilog
->zl_cv_suspend
, NULL
, CV_DEFAULT
, NULL
);
1832 cv_init(&zilog
->zl_cv_batch
[0], NULL
, CV_DEFAULT
, NULL
);
1833 cv_init(&zilog
->zl_cv_batch
[1], NULL
, CV_DEFAULT
, NULL
);
1839 zil_free(zilog_t
*zilog
)
1843 zilog
->zl_stop_sync
= 1;
1845 ASSERT0(zilog
->zl_suspend
);
1846 ASSERT0(zilog
->zl_suspending
);
1848 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1849 list_destroy(&zilog
->zl_lwb_list
);
1851 avl_destroy(&zilog
->zl_vdev_tree
);
1852 mutex_destroy(&zilog
->zl_vdev_lock
);
1854 ASSERT(list_is_empty(&zilog
->zl_itx_commit_list
));
1855 list_destroy(&zilog
->zl_itx_commit_list
);
1857 for (i
= 0; i
< TXG_SIZE
; i
++) {
1859 * It's possible for an itx to be generated that doesn't dirty
1860 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1861 * callback to remove the entry. We remove those here.
1863 * Also free up the ziltest itxs.
1865 if (zilog
->zl_itxg
[i
].itxg_itxs
)
1866 zil_itxg_clean(zilog
->zl_itxg
[i
].itxg_itxs
);
1867 mutex_destroy(&zilog
->zl_itxg
[i
].itxg_lock
);
1870 mutex_destroy(&zilog
->zl_lock
);
1872 cv_destroy(&zilog
->zl_cv_writer
);
1873 cv_destroy(&zilog
->zl_cv_suspend
);
1874 cv_destroy(&zilog
->zl_cv_batch
[0]);
1875 cv_destroy(&zilog
->zl_cv_batch
[1]);
1877 kmem_free(zilog
, sizeof (zilog_t
));
1881 * Open an intent log.
1884 zil_open(objset_t
*os
, zil_get_data_t
*get_data
)
1886 zilog_t
*zilog
= dmu_objset_zil(os
);
1888 ASSERT(zilog
->zl_clean_taskq
== NULL
);
1889 ASSERT(zilog
->zl_get_data
== NULL
);
1890 ASSERT(list_is_empty(&zilog
->zl_lwb_list
));
1892 zilog
->zl_get_data
= get_data
;
1893 zilog
->zl_clean_taskq
= taskq_create("zil_clean", 1, minclsyspri
,
1894 2, 2, TASKQ_PREPOPULATE
);
1900 * Close an intent log.
1903 zil_close(zilog_t
*zilog
)
1908 zil_commit(zilog
, 0); /* commit all itx */
1911 * The lwb_max_txg for the stubby lwb will reflect the last activity
1912 * for the zil. After a txg_wait_synced() on the txg we know all the
1913 * callbacks have occurred that may clean the zil. Only then can we
1914 * destroy the zl_clean_taskq.
1916 mutex_enter(&zilog
->zl_lock
);
1917 lwb
= list_tail(&zilog
->zl_lwb_list
);
1919 txg
= lwb
->lwb_max_txg
;
1920 mutex_exit(&zilog
->zl_lock
);
1922 txg_wait_synced(zilog
->zl_dmu_pool
, txg
);
1923 ASSERT(!zilog_is_dirty(zilog
));
1925 taskq_destroy(zilog
->zl_clean_taskq
);
1926 zilog
->zl_clean_taskq
= NULL
;
1927 zilog
->zl_get_data
= NULL
;
1930 * We should have only one LWB left on the list; remove it now.
1932 mutex_enter(&zilog
->zl_lock
);
1933 lwb
= list_head(&zilog
->zl_lwb_list
);
1935 ASSERT(lwb
== list_tail(&zilog
->zl_lwb_list
));
1936 ASSERT(lwb
->lwb_zio
== NULL
);
1937 if (lwb
->lwb_fastwrite
)
1938 metaslab_fastwrite_unmark(zilog
->zl_spa
, &lwb
->lwb_blk
);
1939 list_remove(&zilog
->zl_lwb_list
, lwb
);
1940 zio_buf_free(lwb
->lwb_buf
, lwb
->lwb_sz
);
1941 kmem_cache_free(zil_lwb_cache
, lwb
);
1943 mutex_exit(&zilog
->zl_lock
);
1946 static char *suspend_tag
= "zil suspending";
1949 * Suspend an intent log. While in suspended mode, we still honor
1950 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1951 * On old version pools, we suspend the log briefly when taking a
1952 * snapshot so that it will have an empty intent log.
1954 * Long holds are not really intended to be used the way we do here --
1955 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1956 * could fail. Therefore we take pains to only put a long hold if it is
1957 * actually necessary. Fortunately, it will only be necessary if the
1958 * objset is currently mounted (or the ZVOL equivalent). In that case it
1959 * will already have a long hold, so we are not really making things any worse.
1961 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1962 * zvol_state_t), and use their mechanism to prevent their hold from being
1963 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1966 * if cookiep == NULL, this does both the suspend & resume.
1967 * Otherwise, it returns with the dataset "long held", and the cookie
1968 * should be passed into zil_resume().
1971 zil_suspend(const char *osname
, void **cookiep
)
1975 const zil_header_t
*zh
;
1978 error
= dmu_objset_hold(osname
, suspend_tag
, &os
);
1981 zilog
= dmu_objset_zil(os
);
1983 mutex_enter(&zilog
->zl_lock
);
1984 zh
= zilog
->zl_header
;
1986 if (zh
->zh_flags
& ZIL_REPLAY_NEEDED
) { /* unplayed log */
1987 mutex_exit(&zilog
->zl_lock
);
1988 dmu_objset_rele(os
, suspend_tag
);
1989 return (SET_ERROR(EBUSY
));
1993 * Don't put a long hold in the cases where we can avoid it. This
1994 * is when there is no cookie so we are doing a suspend & resume
1995 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1996 * for the suspend because it's already suspended, or there's no ZIL.
1998 if (cookiep
== NULL
&& !zilog
->zl_suspending
&&
1999 (zilog
->zl_suspend
> 0 || BP_IS_HOLE(&zh
->zh_log
))) {
2000 mutex_exit(&zilog
->zl_lock
);
2001 dmu_objset_rele(os
, suspend_tag
);
2005 dsl_dataset_long_hold(dmu_objset_ds(os
), suspend_tag
);
2006 dsl_pool_rele(dmu_objset_pool(os
), suspend_tag
);
2008 zilog
->zl_suspend
++;
2010 if (zilog
->zl_suspend
> 1) {
2012 * Someone else is already suspending it.
2013 * Just wait for them to finish.
2016 while (zilog
->zl_suspending
)
2017 cv_wait(&zilog
->zl_cv_suspend
, &zilog
->zl_lock
);
2018 mutex_exit(&zilog
->zl_lock
);
2020 if (cookiep
== NULL
)
2028 * If there is no pointer to an on-disk block, this ZIL must not
2029 * be active (e.g. filesystem not mounted), so there's nothing
2032 if (BP_IS_HOLE(&zh
->zh_log
)) {
2033 ASSERT(cookiep
!= NULL
); /* fast path already handled */
2036 mutex_exit(&zilog
->zl_lock
);
2040 zilog
->zl_suspending
= B_TRUE
;
2041 mutex_exit(&zilog
->zl_lock
);
2043 zil_commit(zilog
, 0);
2045 zil_destroy(zilog
, B_FALSE
);
2047 mutex_enter(&zilog
->zl_lock
);
2048 zilog
->zl_suspending
= B_FALSE
;
2049 cv_broadcast(&zilog
->zl_cv_suspend
);
2050 mutex_exit(&zilog
->zl_lock
);
2052 if (cookiep
== NULL
)
2060 zil_resume(void *cookie
)
2062 objset_t
*os
= cookie
;
2063 zilog_t
*zilog
= dmu_objset_zil(os
);
2065 mutex_enter(&zilog
->zl_lock
);
2066 ASSERT(zilog
->zl_suspend
!= 0);
2067 zilog
->zl_suspend
--;
2068 mutex_exit(&zilog
->zl_lock
);
2069 dsl_dataset_long_rele(dmu_objset_ds(os
), suspend_tag
);
2070 dsl_dataset_rele(dmu_objset_ds(os
), suspend_tag
);
2073 typedef struct zil_replay_arg
{
2074 zil_replay_func_t
*zr_replay
;
2076 boolean_t zr_byteswap
;
2081 zil_replay_error(zilog_t
*zilog
, lr_t
*lr
, int error
)
2083 char name
[MAXNAMELEN
];
2085 zilog
->zl_replaying_seq
--; /* didn't actually replay this one */
2087 dmu_objset_name(zilog
->zl_os
, name
);
2089 cmn_err(CE_WARN
, "ZFS replay transaction error %d, "
2090 "dataset %s, seq 0x%llx, txtype %llu %s\n", error
, name
,
2091 (u_longlong_t
)lr
->lrc_seq
,
2092 (u_longlong_t
)(lr
->lrc_txtype
& ~TX_CI
),
2093 (lr
->lrc_txtype
& TX_CI
) ? "CI" : "");
2099 zil_replay_log_record(zilog_t
*zilog
, lr_t
*lr
, void *zra
, uint64_t claim_txg
)
2101 zil_replay_arg_t
*zr
= zra
;
2102 const zil_header_t
*zh
= zilog
->zl_header
;
2103 uint64_t reclen
= lr
->lrc_reclen
;
2104 uint64_t txtype
= lr
->lrc_txtype
;
2107 zilog
->zl_replaying_seq
= lr
->lrc_seq
;
2109 if (lr
->lrc_seq
<= zh
->zh_replay_seq
) /* already replayed */
2112 if (lr
->lrc_txg
< claim_txg
) /* already committed */
2115 /* Strip case-insensitive bit, still present in log record */
2118 if (txtype
== 0 || txtype
>= TX_MAX_TYPE
)
2119 return (zil_replay_error(zilog
, lr
, EINVAL
));
2122 * If this record type can be logged out of order, the object
2123 * (lr_foid) may no longer exist. That's legitimate, not an error.
2125 if (TX_OOO(txtype
)) {
2126 error
= dmu_object_info(zilog
->zl_os
,
2127 ((lr_ooo_t
*)lr
)->lr_foid
, NULL
);
2128 if (error
== ENOENT
|| error
== EEXIST
)
2133 * Make a copy of the data so we can revise and extend it.
2135 bcopy(lr
, zr
->zr_lr
, reclen
);
2138 * If this is a TX_WRITE with a blkptr, suck in the data.
2140 if (txtype
== TX_WRITE
&& reclen
== sizeof (lr_write_t
)) {
2141 error
= zil_read_log_data(zilog
, (lr_write_t
*)lr
,
2142 zr
->zr_lr
+ reclen
);
2144 return (zil_replay_error(zilog
, lr
, error
));
2148 * The log block containing this lr may have been byteswapped
2149 * so that we can easily examine common fields like lrc_txtype.
2150 * However, the log is a mix of different record types, and only the
2151 * replay vectors know how to byteswap their records. Therefore, if
2152 * the lr was byteswapped, undo it before invoking the replay vector.
2154 if (zr
->zr_byteswap
)
2155 byteswap_uint64_array(zr
->zr_lr
, reclen
);
2158 * We must now do two things atomically: replay this log record,
2159 * and update the log header sequence number to reflect the fact that
2160 * we did so. At the end of each replay function the sequence number
2161 * is updated if we are in replay mode.
2163 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, zr
->zr_byteswap
);
2166 * The DMU's dnode layer doesn't see removes until the txg
2167 * commits, so a subsequent claim can spuriously fail with
2168 * EEXIST. So if we receive any error we try syncing out
2169 * any removes then retry the transaction. Note that we
2170 * specify B_FALSE for byteswap now, so we don't do it twice.
2172 txg_wait_synced(spa_get_dsl(zilog
->zl_spa
), 0);
2173 error
= zr
->zr_replay
[txtype
](zr
->zr_arg
, zr
->zr_lr
, B_FALSE
);
2175 return (zil_replay_error(zilog
, lr
, error
));
2182 zil_incr_blks(zilog_t
*zilog
, blkptr_t
*bp
, void *arg
, uint64_t claim_txg
)
2184 zilog
->zl_replay_blks
++;
2190 * If this dataset has a non-empty intent log, replay it and destroy it.
2193 zil_replay(objset_t
*os
, void *arg
, zil_replay_func_t replay_func
[TX_MAX_TYPE
])
2195 zilog_t
*zilog
= dmu_objset_zil(os
);
2196 const zil_header_t
*zh
= zilog
->zl_header
;
2197 zil_replay_arg_t zr
;
2199 if ((zh
->zh_flags
& ZIL_REPLAY_NEEDED
) == 0) {
2200 zil_destroy(zilog
, B_TRUE
);
2204 zr
.zr_replay
= replay_func
;
2206 zr
.zr_byteswap
= BP_SHOULD_BYTESWAP(&zh
->zh_log
);
2207 zr
.zr_lr
= vmem_alloc(2 * SPA_MAXBLOCKSIZE
, KM_PUSHPAGE
);
2210 * Wait for in-progress removes to sync before starting replay.
2212 txg_wait_synced(zilog
->zl_dmu_pool
, 0);
2214 zilog
->zl_replay
= B_TRUE
;
2215 zilog
->zl_replay_time
= ddi_get_lbolt();
2216 ASSERT(zilog
->zl_replay_blks
== 0);
2217 (void) zil_parse(zilog
, zil_incr_blks
, zil_replay_log_record
, &zr
,
2219 vmem_free(zr
.zr_lr
, 2 * SPA_MAXBLOCKSIZE
);
2221 zil_destroy(zilog
, B_FALSE
);
2222 txg_wait_synced(zilog
->zl_dmu_pool
, zilog
->zl_destroy_txg
);
2223 zilog
->zl_replay
= B_FALSE
;
2227 zil_replaying(zilog_t
*zilog
, dmu_tx_t
*tx
)
2229 if (zilog
->zl_sync
== ZFS_SYNC_DISABLED
)
2232 if (zilog
->zl_replay
) {
2233 dsl_dataset_dirty(dmu_objset_ds(zilog
->zl_os
), tx
);
2234 zilog
->zl_replayed_seq
[dmu_tx_get_txg(tx
) & TXG_MASK
] =
2235 zilog
->zl_replaying_seq
;
2244 zil_vdev_offline(const char *osname
, void *arg
)
2248 error
= zil_suspend(osname
, NULL
);
2250 return (SET_ERROR(EEXIST
));
2254 #if defined(_KERNEL) && defined(HAVE_SPL)
2255 EXPORT_SYMBOL(zil_alloc
);
2256 EXPORT_SYMBOL(zil_free
);
2257 EXPORT_SYMBOL(zil_open
);
2258 EXPORT_SYMBOL(zil_close
);
2259 EXPORT_SYMBOL(zil_replay
);
2260 EXPORT_SYMBOL(zil_replaying
);
2261 EXPORT_SYMBOL(zil_destroy
);
2262 EXPORT_SYMBOL(zil_destroy_sync
);
2263 EXPORT_SYMBOL(zil_itx_create
);
2264 EXPORT_SYMBOL(zil_itx_destroy
);
2265 EXPORT_SYMBOL(zil_itx_assign
);
2266 EXPORT_SYMBOL(zil_commit
);
2267 EXPORT_SYMBOL(zil_vdev_offline
);
2268 EXPORT_SYMBOL(zil_claim
);
2269 EXPORT_SYMBOL(zil_check_log_chain
);
2270 EXPORT_SYMBOL(zil_sync
);
2271 EXPORT_SYMBOL(zil_clean
);
2272 EXPORT_SYMBOL(zil_suspend
);
2273 EXPORT_SYMBOL(zil_resume
);
2274 EXPORT_SYMBOL(zil_add_block
);
2275 EXPORT_SYMBOL(zil_bp_tree_add
);
2276 EXPORT_SYMBOL(zil_set_sync
);
2277 EXPORT_SYMBOL(zil_set_logbias
);
2279 module_param(zil_replay_disable
, int, 0644);
2280 MODULE_PARM_DESC(zil_replay_disable
, "Disable intent logging replay");
2282 module_param(zfs_nocacheflush
, int, 0644);
2283 MODULE_PARM_DESC(zfs_nocacheflush
, "Disable cache flushes");
2285 module_param(zil_slog_limit
, ulong
, 0644);
2286 MODULE_PARM_DESC(zil_slog_limit
, "Max commit bytes to separate log device");