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 #include <sys/zfs_context.h>
26 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/dmu_objset.h>
39 * ==========================================================================
41 * ==========================================================================
43 uint8_t zio_priority_table
[ZIO_PRIORITY_TABLE_SIZE
] = {
44 0, /* ZIO_PRIORITY_NOW */
45 0, /* ZIO_PRIORITY_SYNC_READ */
46 0, /* ZIO_PRIORITY_SYNC_WRITE */
47 0, /* ZIO_PRIORITY_LOG_WRITE */
48 1, /* ZIO_PRIORITY_CACHE_FILL */
49 1, /* ZIO_PRIORITY_AGG */
50 4, /* ZIO_PRIORITY_FREE */
51 4, /* ZIO_PRIORITY_ASYNC_WRITE */
52 6, /* ZIO_PRIORITY_ASYNC_READ */
53 10, /* ZIO_PRIORITY_RESILVER */
54 20, /* ZIO_PRIORITY_SCRUB */
55 2, /* ZIO_PRIORITY_DDT_PREFETCH */
59 * ==========================================================================
60 * I/O type descriptions
61 * ==========================================================================
63 char *zio_type_name
[ZIO_TYPES
] = {
64 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
68 * ==========================================================================
70 * ==========================================================================
72 kmem_cache_t
*zio_cache
;
73 kmem_cache_t
*zio_link_cache
;
74 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
75 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
76 int zio_bulk_flags
= 0;
77 int zio_delay_max
= ZIO_DELAY_MAX
;
80 extern vmem_t
*zio_alloc_arena
;
84 * An allocating zio is one that either currently has the DVA allocate
85 * stage set or will have it later in its lifetime.
87 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
89 boolean_t zio_requeue_io_start_cut_in_line
= B_TRUE
;
92 int zio_buf_debug_limit
= 16384;
94 int zio_buf_debug_limit
= 0;
97 static inline void __zio_execute(zio_t
*zio
);
103 vmem_t
*data_alloc_arena
= NULL
;
106 data_alloc_arena
= zio_alloc_arena
;
108 zio_cache
= kmem_cache_create("zio_cache",
109 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
110 zio_link_cache
= kmem_cache_create("zio_link_cache",
111 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
114 * For small buffers, we want a cache for each multiple of
115 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
116 * for each quarter-power of 2. For large buffers, we want
117 * a cache for each multiple of PAGESIZE.
119 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
120 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
124 while (p2
& (p2
- 1))
127 if (size
<= 4 * SPA_MINBLOCKSIZE
) {
128 align
= SPA_MINBLOCKSIZE
;
129 } else if (P2PHASE(size
, PAGESIZE
) == 0) {
131 } else if (P2PHASE(size
, p2
>> 2) == 0) {
137 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
138 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
139 align
, NULL
, NULL
, NULL
, NULL
, NULL
,
140 (size
> zio_buf_debug_limit
? KMC_NODEBUG
: 0) |
143 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
144 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
145 align
, NULL
, NULL
, NULL
, NULL
, data_alloc_arena
,
146 (size
> zio_buf_debug_limit
? KMC_NODEBUG
: 0) |
152 ASSERT(zio_buf_cache
[c
] != NULL
);
153 if (zio_buf_cache
[c
- 1] == NULL
)
154 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
156 ASSERT(zio_data_buf_cache
[c
] != NULL
);
157 if (zio_data_buf_cache
[c
- 1] == NULL
)
158 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
168 kmem_cache_t
*last_cache
= NULL
;
169 kmem_cache_t
*last_data_cache
= NULL
;
171 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
172 if (zio_buf_cache
[c
] != last_cache
) {
173 last_cache
= zio_buf_cache
[c
];
174 kmem_cache_destroy(zio_buf_cache
[c
]);
176 zio_buf_cache
[c
] = NULL
;
178 if (zio_data_buf_cache
[c
] != last_data_cache
) {
179 last_data_cache
= zio_data_buf_cache
[c
];
180 kmem_cache_destroy(zio_data_buf_cache
[c
]);
182 zio_data_buf_cache
[c
] = NULL
;
185 kmem_cache_destroy(zio_link_cache
);
186 kmem_cache_destroy(zio_cache
);
192 * ==========================================================================
193 * Allocate and free I/O buffers
194 * ==========================================================================
198 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
199 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
200 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
201 * excess / transient data in-core during a crashdump.
204 zio_buf_alloc(size_t size
)
206 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
208 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
210 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
214 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
215 * crashdump if the kernel panics. This exists so that we will limit the amount
216 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
217 * of kernel heap dumped to disk when the kernel panics)
220 zio_data_buf_alloc(size_t size
)
222 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
224 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
226 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
230 zio_buf_free(void *buf
, size_t size
)
232 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
234 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
236 kmem_cache_free(zio_buf_cache
[c
], buf
);
240 zio_data_buf_free(void *buf
, size_t size
)
242 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
244 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
246 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
250 * ==========================================================================
251 * Push and pop I/O transform buffers
252 * ==========================================================================
255 zio_push_transform(zio_t
*zio
, void *data
, uint64_t size
, uint64_t bufsize
,
256 zio_transform_func_t
*transform
)
258 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
260 zt
->zt_orig_data
= zio
->io_data
;
261 zt
->zt_orig_size
= zio
->io_size
;
262 zt
->zt_bufsize
= bufsize
;
263 zt
->zt_transform
= transform
;
265 zt
->zt_next
= zio
->io_transform_stack
;
266 zio
->io_transform_stack
= zt
;
273 zio_pop_transforms(zio_t
*zio
)
277 while ((zt
= zio
->io_transform_stack
) != NULL
) {
278 if (zt
->zt_transform
!= NULL
)
279 zt
->zt_transform(zio
,
280 zt
->zt_orig_data
, zt
->zt_orig_size
);
282 if (zt
->zt_bufsize
!= 0)
283 zio_buf_free(zio
->io_data
, zt
->zt_bufsize
);
285 zio
->io_data
= zt
->zt_orig_data
;
286 zio
->io_size
= zt
->zt_orig_size
;
287 zio
->io_transform_stack
= zt
->zt_next
;
289 kmem_free(zt
, sizeof (zio_transform_t
));
294 * ==========================================================================
295 * I/O transform callbacks for subblocks and decompression
296 * ==========================================================================
299 zio_subblock(zio_t
*zio
, void *data
, uint64_t size
)
301 ASSERT(zio
->io_size
> size
);
303 if (zio
->io_type
== ZIO_TYPE_READ
)
304 bcopy(zio
->io_data
, data
, size
);
308 zio_decompress(zio_t
*zio
, void *data
, uint64_t size
)
310 if (zio
->io_error
== 0 &&
311 zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
312 zio
->io_data
, data
, zio
->io_size
, size
) != 0)
317 * ==========================================================================
318 * I/O parent/child relationships and pipeline interlocks
319 * ==========================================================================
322 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
323 * continue calling these functions until they return NULL.
324 * Otherwise, the next caller will pick up the list walk in
325 * some indeterminate state. (Otherwise every caller would
326 * have to pass in a cookie to keep the state represented by
327 * io_walk_link, which gets annoying.)
330 zio_walk_parents(zio_t
*cio
)
332 zio_link_t
*zl
= cio
->io_walk_link
;
333 list_t
*pl
= &cio
->io_parent_list
;
335 zl
= (zl
== NULL
) ? list_head(pl
) : list_next(pl
, zl
);
336 cio
->io_walk_link
= zl
;
341 ASSERT(zl
->zl_child
== cio
);
342 return (zl
->zl_parent
);
346 zio_walk_children(zio_t
*pio
)
348 zio_link_t
*zl
= pio
->io_walk_link
;
349 list_t
*cl
= &pio
->io_child_list
;
351 zl
= (zl
== NULL
) ? list_head(cl
) : list_next(cl
, zl
);
352 pio
->io_walk_link
= zl
;
357 ASSERT(zl
->zl_parent
== pio
);
358 return (zl
->zl_child
);
362 zio_unique_parent(zio_t
*cio
)
364 zio_t
*pio
= zio_walk_parents(cio
);
366 VERIFY(zio_walk_parents(cio
) == NULL
);
371 zio_add_child(zio_t
*pio
, zio_t
*cio
)
373 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
377 * Logical I/Os can have logical, gang, or vdev children.
378 * Gang I/Os can have gang or vdev children.
379 * Vdev I/Os can only have vdev children.
380 * The following ASSERT captures all of these constraints.
382 ASSERT(cio
->io_child_type
<= pio
->io_child_type
);
387 mutex_enter(&cio
->io_lock
);
388 mutex_enter(&pio
->io_lock
);
390 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
392 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
393 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
395 list_insert_head(&pio
->io_child_list
, zl
);
396 list_insert_head(&cio
->io_parent_list
, zl
);
398 pio
->io_child_count
++;
399 cio
->io_parent_count
++;
401 mutex_exit(&pio
->io_lock
);
402 mutex_exit(&cio
->io_lock
);
406 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
408 ASSERT(zl
->zl_parent
== pio
);
409 ASSERT(zl
->zl_child
== cio
);
411 mutex_enter(&cio
->io_lock
);
412 mutex_enter(&pio
->io_lock
);
414 list_remove(&pio
->io_child_list
, zl
);
415 list_remove(&cio
->io_parent_list
, zl
);
417 pio
->io_child_count
--;
418 cio
->io_parent_count
--;
420 mutex_exit(&pio
->io_lock
);
421 mutex_exit(&cio
->io_lock
);
423 kmem_cache_free(zio_link_cache
, zl
);
427 zio_wait_for_children(zio_t
*zio
, enum zio_child child
, enum zio_wait_type wait
)
429 uint64_t *countp
= &zio
->io_children
[child
][wait
];
430 boolean_t waiting
= B_FALSE
;
432 mutex_enter(&zio
->io_lock
);
433 ASSERT(zio
->io_stall
== NULL
);
436 zio
->io_stall
= countp
;
439 mutex_exit(&zio
->io_lock
);
444 __attribute__((always_inline
))
446 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
448 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
449 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
451 mutex_enter(&pio
->io_lock
);
452 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
453 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
454 pio
->io_reexecute
|= zio
->io_reexecute
;
455 ASSERT3U(*countp
, >, 0);
456 if (--*countp
== 0 && pio
->io_stall
== countp
) {
457 pio
->io_stall
= NULL
;
458 mutex_exit(&pio
->io_lock
);
461 mutex_exit(&pio
->io_lock
);
466 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
468 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
469 zio
->io_error
= zio
->io_child_error
[c
];
473 * ==========================================================================
474 * Create the various types of I/O (read, write, free, etc)
475 * ==========================================================================
478 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
479 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
480 zio_type_t type
, int priority
, enum zio_flag flags
,
481 vdev_t
*vd
, uint64_t offset
, const zbookmark_t
*zb
,
482 enum zio_stage stage
, enum zio_stage pipeline
)
486 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
487 ASSERT(P2PHASE(size
, SPA_MINBLOCKSIZE
) == 0);
488 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
490 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
491 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
492 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
494 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
495 bzero(zio
, sizeof (zio_t
));
497 mutex_init(&zio
->io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
498 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
500 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
501 offsetof(zio_link_t
, zl_parent_node
));
502 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
503 offsetof(zio_link_t
, zl_child_node
));
506 zio
->io_child_type
= ZIO_CHILD_VDEV
;
507 else if (flags
& ZIO_FLAG_GANG_CHILD
)
508 zio
->io_child_type
= ZIO_CHILD_GANG
;
509 else if (flags
& ZIO_FLAG_DDT_CHILD
)
510 zio
->io_child_type
= ZIO_CHILD_DDT
;
512 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
515 zio
->io_bp
= (blkptr_t
*)bp
;
516 zio
->io_bp_copy
= *bp
;
517 zio
->io_bp_orig
= *bp
;
518 if (type
!= ZIO_TYPE_WRITE
||
519 zio
->io_child_type
== ZIO_CHILD_DDT
)
520 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
521 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
522 zio
->io_logical
= zio
;
523 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
524 pipeline
|= ZIO_GANG_STAGES
;
530 zio
->io_private
= private;
532 zio
->io_priority
= priority
;
534 zio
->io_offset
= offset
;
535 zio
->io_orig_data
= zio
->io_data
= data
;
536 zio
->io_orig_size
= zio
->io_size
= size
;
537 zio
->io_orig_flags
= zio
->io_flags
= flags
;
538 zio
->io_orig_stage
= zio
->io_stage
= stage
;
539 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
541 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
542 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
545 zio
->io_bookmark
= *zb
;
548 if (zio
->io_logical
== NULL
)
549 zio
->io_logical
= pio
->io_logical
;
550 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
551 zio
->io_gang_leader
= pio
->io_gang_leader
;
552 zio_add_child(pio
, zio
);
559 zio_destroy(zio_t
*zio
)
561 list_destroy(&zio
->io_parent_list
);
562 list_destroy(&zio
->io_child_list
);
563 mutex_destroy(&zio
->io_lock
);
564 cv_destroy(&zio
->io_cv
);
565 kmem_cache_free(zio_cache
, zio
);
569 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
570 void *private, enum zio_flag flags
)
574 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, done
, private,
575 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
576 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
582 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
584 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
588 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
589 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
590 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
)
594 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
595 data
, size
, done
, private,
596 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
597 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
598 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
604 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
605 void *data
, uint64_t size
, const zio_prop_t
*zp
,
606 zio_done_func_t
*ready
, zio_done_func_t
*done
, void *private,
607 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
)
611 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
612 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
613 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
614 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
615 zp
->zp_type
< DMU_OT_NUMTYPES
&&
618 zp
->zp_copies
<= spa_max_replication(spa
) &&
620 zp
->zp_dedup_verify
<= 1);
622 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, done
, private,
623 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
624 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
625 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
627 zio
->io_ready
= ready
;
634 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, void *data
,
635 uint64_t size
, zio_done_func_t
*done
, void *private, int priority
,
636 enum zio_flag flags
, zbookmark_t
*zb
)
640 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, done
, private,
641 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
642 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
648 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
)
650 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
651 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
652 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
653 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
655 zio
->io_prop
.zp_copies
= copies
;
656 zio
->io_bp_override
= bp
;
660 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
662 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
666 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
671 dprintf_bp(bp
, "freeing in txg %llu, pass %u",
672 (longlong_t
)txg
, spa
->spa_sync_pass
);
674 ASSERT(!BP_IS_HOLE(bp
));
675 ASSERT(spa_syncing_txg(spa
) == txg
);
676 ASSERT(spa_sync_pass(spa
) <= SYNC_PASS_DEFERRED_FREE
);
678 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
679 NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_FREE
, flags
,
680 NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_FREE_PIPELINE
);
686 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
687 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
692 * A claim is an allocation of a specific block. Claims are needed
693 * to support immediate writes in the intent log. The issue is that
694 * immediate writes contain committed data, but in a txg that was
695 * *not* committed. Upon opening the pool after an unclean shutdown,
696 * the intent log claims all blocks that contain immediate write data
697 * so that the SPA knows they're in use.
699 * All claims *must* be resolved in the first txg -- before the SPA
700 * starts allocating blocks -- so that nothing is allocated twice.
701 * If txg == 0 we just verify that the block is claimable.
703 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
704 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
705 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
707 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
708 done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
, flags
,
709 NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
715 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
716 zio_done_func_t
*done
, void *private, int priority
, enum zio_flag flags
)
721 if (vd
->vdev_children
== 0) {
722 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, done
, private,
723 ZIO_TYPE_IOCTL
, priority
, flags
, vd
, 0, NULL
,
724 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
728 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
730 for (c
= 0; c
< vd
->vdev_children
; c
++)
731 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
732 done
, private, priority
, flags
));
739 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
740 void *data
, int checksum
, zio_done_func_t
*done
, void *private,
741 int priority
, enum zio_flag flags
, boolean_t labels
)
745 ASSERT(vd
->vdev_children
== 0);
746 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
747 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
748 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
750 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, done
, private,
751 ZIO_TYPE_READ
, priority
, flags
, vd
, offset
, NULL
,
752 ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
754 zio
->io_prop
.zp_checksum
= checksum
;
760 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
761 void *data
, int checksum
, zio_done_func_t
*done
, void *private,
762 int priority
, enum zio_flag flags
, boolean_t labels
)
766 ASSERT(vd
->vdev_children
== 0);
767 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
768 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
769 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
771 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, done
, private,
772 ZIO_TYPE_WRITE
, priority
, flags
, vd
, offset
, NULL
,
773 ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
775 zio
->io_prop
.zp_checksum
= checksum
;
777 if (zio_checksum_table
[checksum
].ci_eck
) {
779 * zec checksums are necessarily destructive -- they modify
780 * the end of the write buffer to hold the verifier/checksum.
781 * Therefore, we must make a local copy in case the data is
782 * being written to multiple places in parallel.
784 void *wbuf
= zio_buf_alloc(size
);
785 bcopy(data
, wbuf
, size
);
786 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
793 * Create a child I/O to do some work for us.
796 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
797 void *data
, uint64_t size
, int type
, int priority
, enum zio_flag flags
,
798 zio_done_func_t
*done
, void *private)
800 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
803 ASSERT(vd
->vdev_parent
==
804 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
806 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
808 * If we have the bp, then the child should perform the
809 * checksum and the parent need not. This pushes error
810 * detection as close to the leaves as possible and
811 * eliminates redundant checksums in the interior nodes.
813 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
814 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
817 if (vd
->vdev_children
== 0)
818 offset
+= VDEV_LABEL_START_SIZE
;
820 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
823 * If we've decided to do a repair, the write is not speculative --
824 * even if the original read was.
826 if (flags
& ZIO_FLAG_IO_REPAIR
)
827 flags
&= ~ZIO_FLAG_SPECULATIVE
;
829 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
,
830 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
831 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
837 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, void *data
, uint64_t size
,
838 int type
, int priority
, enum zio_flag flags
,
839 zio_done_func_t
*done
, void *private)
843 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
845 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
846 data
, size
, done
, private, type
, priority
,
847 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
,
849 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
855 zio_flush(zio_t
*zio
, vdev_t
*vd
)
857 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
858 NULL
, NULL
, ZIO_PRIORITY_NOW
,
859 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
863 zio_shrink(zio_t
*zio
, uint64_t size
)
865 ASSERT(zio
->io_executor
== NULL
);
866 ASSERT(zio
->io_orig_size
== zio
->io_size
);
867 ASSERT(size
<= zio
->io_size
);
870 * We don't shrink for raidz because of problems with the
871 * reconstruction when reading back less than the block size.
872 * Note, BP_IS_RAIDZ() assumes no compression.
874 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
875 if (!BP_IS_RAIDZ(zio
->io_bp
))
876 zio
->io_orig_size
= zio
->io_size
= size
;
880 * ==========================================================================
881 * Prepare to read and write logical blocks
882 * ==========================================================================
886 zio_read_bp_init(zio_t
*zio
)
888 blkptr_t
*bp
= zio
->io_bp
;
890 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
891 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
892 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
893 uint64_t psize
= BP_GET_PSIZE(bp
);
894 void *cbuf
= zio_buf_alloc(psize
);
896 zio_push_transform(zio
, cbuf
, psize
, psize
, zio_decompress
);
899 if (!dmu_ot
[BP_GET_TYPE(bp
)].ot_metadata
&& BP_GET_LEVEL(bp
) == 0)
900 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
902 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
903 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
905 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
906 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
908 return (ZIO_PIPELINE_CONTINUE
);
912 zio_write_bp_init(zio_t
*zio
)
914 spa_t
*spa
= zio
->io_spa
;
915 zio_prop_t
*zp
= &zio
->io_prop
;
916 enum zio_compress compress
= zp
->zp_compress
;
917 blkptr_t
*bp
= zio
->io_bp
;
918 uint64_t lsize
= zio
->io_size
;
919 uint64_t psize
= lsize
;
923 * If our children haven't all reached the ready stage,
924 * wait for them and then repeat this pipeline stage.
926 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
927 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
928 return (ZIO_PIPELINE_STOP
);
930 if (!IO_IS_ALLOCATING(zio
))
931 return (ZIO_PIPELINE_CONTINUE
);
933 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
935 if (zio
->io_bp_override
) {
936 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
937 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
939 *bp
= *zio
->io_bp_override
;
940 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
942 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
943 return (ZIO_PIPELINE_CONTINUE
);
945 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_dedup
||
946 zp
->zp_dedup_verify
);
948 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
950 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
951 return (ZIO_PIPELINE_CONTINUE
);
953 zio
->io_bp_override
= NULL
;
957 if (bp
->blk_birth
== zio
->io_txg
) {
959 * We're rewriting an existing block, which means we're
960 * working on behalf of spa_sync(). For spa_sync() to
961 * converge, it must eventually be the case that we don't
962 * have to allocate new blocks. But compression changes
963 * the blocksize, which forces a reallocate, and makes
964 * convergence take longer. Therefore, after the first
965 * few passes, stop compressing to ensure convergence.
967 pass
= spa_sync_pass(spa
);
969 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
970 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
971 ASSERT(!BP_GET_DEDUP(bp
));
973 if (pass
> SYNC_PASS_DONT_COMPRESS
)
974 compress
= ZIO_COMPRESS_OFF
;
976 /* Make sure someone doesn't change their mind on overwrites */
977 ASSERT(MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
978 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
981 if (compress
!= ZIO_COMPRESS_OFF
) {
982 void *cbuf
= zio_buf_alloc(lsize
);
983 psize
= zio_compress_data(compress
, zio
->io_data
, cbuf
, lsize
);
984 if (psize
== 0 || psize
== lsize
) {
985 compress
= ZIO_COMPRESS_OFF
;
986 zio_buf_free(cbuf
, lsize
);
988 ASSERT(psize
< lsize
);
989 zio_push_transform(zio
, cbuf
, psize
, lsize
, NULL
);
994 * The final pass of spa_sync() must be all rewrites, but the first
995 * few passes offer a trade-off: allocating blocks defers convergence,
996 * but newly allocated blocks are sequential, so they can be written
997 * to disk faster. Therefore, we allow the first few passes of
998 * spa_sync() to allocate new blocks, but force rewrites after that.
999 * There should only be a handful of blocks after pass 1 in any case.
1001 if (bp
->blk_birth
== zio
->io_txg
&& BP_GET_PSIZE(bp
) == psize
&&
1002 pass
> SYNC_PASS_REWRITE
) {
1003 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1005 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1006 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1009 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1013 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1015 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1016 BP_SET_LSIZE(bp
, lsize
);
1017 BP_SET_PSIZE(bp
, psize
);
1018 BP_SET_COMPRESS(bp
, compress
);
1019 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1020 BP_SET_TYPE(bp
, zp
->zp_type
);
1021 BP_SET_LEVEL(bp
, zp
->zp_level
);
1022 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1023 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1025 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1026 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1027 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1031 return (ZIO_PIPELINE_CONTINUE
);
1035 zio_free_bp_init(zio_t
*zio
)
1037 blkptr_t
*bp
= zio
->io_bp
;
1039 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1040 if (BP_GET_DEDUP(bp
))
1041 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1044 return (ZIO_PIPELINE_CONTINUE
);
1048 * ==========================================================================
1049 * Execute the I/O pipeline
1050 * ==========================================================================
1054 zio_taskq_dispatch(zio_t
*zio
, enum zio_taskq_type q
, boolean_t cutinline
)
1056 spa_t
*spa
= zio
->io_spa
;
1057 zio_type_t t
= zio
->io_type
;
1058 int flags
= TQ_NOSLEEP
| (cutinline
? TQ_FRONT
: 0);
1061 * If we're a config writer or a probe, the normal issue and
1062 * interrupt threads may all be blocked waiting for the config lock.
1063 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1065 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1069 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1071 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1075 * If this is a high priority I/O, then use the high priority taskq.
1077 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1078 spa
->spa_zio_taskq
[t
][q
+ 1] != NULL
)
1081 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1083 while (taskq_dispatch(spa
->spa_zio_taskq
[t
][q
],
1084 (task_func_t
*)zio_execute
, zio
, flags
) == 0); /* do nothing */
1088 zio_taskq_member(zio_t
*zio
, enum zio_taskq_type q
)
1090 kthread_t
*executor
= zio
->io_executor
;
1091 spa_t
*spa
= zio
->io_spa
;
1094 for (t
= 0; t
< ZIO_TYPES
; t
++)
1095 if (taskq_member(spa
->spa_zio_taskq
[t
][q
], executor
))
1102 zio_issue_async(zio_t
*zio
)
1104 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1106 return (ZIO_PIPELINE_STOP
);
1110 zio_interrupt(zio_t
*zio
)
1112 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1116 * Execute the I/O pipeline until one of the following occurs:
1117 * (1) the I/O completes; (2) the pipeline stalls waiting for
1118 * dependent child I/Os; (3) the I/O issues, so we're waiting
1119 * for an I/O completion interrupt; (4) the I/O is delegated by
1120 * vdev-level caching or aggregation; (5) the I/O is deferred
1121 * due to vdev-level queueing; (6) the I/O is handed off to
1122 * another thread. In all cases, the pipeline stops whenever
1123 * there's no CPU work; it never burns a thread in cv_wait().
1125 * There's no locking on io_stage because there's no legitimate way
1126 * for multiple threads to be attempting to process the same I/O.
1128 static zio_pipe_stage_t
*zio_pipeline
[];
1131 * zio_execute() is a wrapper around the static function
1132 * __zio_execute() so that we can force __zio_execute() to be
1133 * inlined. This reduces stack overhead which is important
1134 * because __zio_execute() is called recursively in several zio
1135 * code paths. zio_execute() itself cannot be inlined because
1136 * it is externally visible.
1139 zio_execute(zio_t
*zio
)
1144 __attribute__((always_inline
))
1146 __zio_execute(zio_t
*zio
)
1148 zio
->io_executor
= curthread
;
1150 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1151 enum zio_stage pipeline
= zio
->io_pipeline
;
1152 enum zio_stage stage
= zio
->io_stage
;
1155 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1156 ASSERT(ISP2(stage
));
1157 ASSERT(zio
->io_stall
== NULL
);
1161 } while ((stage
& pipeline
) == 0);
1163 ASSERT(stage
<= ZIO_STAGE_DONE
);
1166 * If we are in interrupt context and this pipeline stage
1167 * will grab a config lock that is held across I/O,
1168 * or may wait for an I/O that needs an interrupt thread
1169 * to complete, issue async to avoid deadlock.
1171 * For VDEV_IO_START, we cut in line so that the io will
1172 * be sent to disk promptly.
1174 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1175 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1176 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1177 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1178 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1182 zio
->io_stage
= stage
;
1183 rv
= zio_pipeline
[highbit(stage
) - 1](zio
);
1185 if (rv
== ZIO_PIPELINE_STOP
)
1188 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1194 * ==========================================================================
1195 * Initiate I/O, either sync or async
1196 * ==========================================================================
1199 zio_wait(zio_t
*zio
)
1203 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1204 ASSERT(zio
->io_executor
== NULL
);
1206 zio
->io_waiter
= curthread
;
1210 mutex_enter(&zio
->io_lock
);
1211 while (zio
->io_executor
!= NULL
)
1212 cv_wait(&zio
->io_cv
, &zio
->io_lock
);
1213 mutex_exit(&zio
->io_lock
);
1215 error
= zio
->io_error
;
1222 zio_nowait(zio_t
*zio
)
1224 ASSERT(zio
->io_executor
== NULL
);
1226 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1227 zio_unique_parent(zio
) == NULL
) {
1229 * This is a logical async I/O with no parent to wait for it.
1230 * We add it to the spa_async_root_zio "Godfather" I/O which
1231 * will ensure they complete prior to unloading the pool.
1233 spa_t
*spa
= zio
->io_spa
;
1235 zio_add_child(spa
->spa_async_zio_root
, zio
);
1242 * ==========================================================================
1243 * Reexecute or suspend/resume failed I/O
1244 * ==========================================================================
1248 zio_reexecute(zio_t
*pio
)
1250 zio_t
*cio
, *cio_next
;
1253 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1254 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1255 ASSERT(pio
->io_gang_leader
== NULL
);
1256 ASSERT(pio
->io_gang_tree
== NULL
);
1258 pio
->io_flags
= pio
->io_orig_flags
;
1259 pio
->io_stage
= pio
->io_orig_stage
;
1260 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1261 pio
->io_reexecute
= 0;
1263 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1264 pio
->io_state
[w
] = 0;
1265 for (c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1266 pio
->io_child_error
[c
] = 0;
1268 if (IO_IS_ALLOCATING(pio
))
1269 BP_ZERO(pio
->io_bp
);
1272 * As we reexecute pio's children, new children could be created.
1273 * New children go to the head of pio's io_child_list, however,
1274 * so we will (correctly) not reexecute them. The key is that
1275 * the remainder of pio's io_child_list, from 'cio_next' onward,
1276 * cannot be affected by any side effects of reexecuting 'cio'.
1278 for (cio
= zio_walk_children(pio
); cio
!= NULL
; cio
= cio_next
) {
1279 cio_next
= zio_walk_children(pio
);
1280 mutex_enter(&pio
->io_lock
);
1281 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1282 pio
->io_children
[cio
->io_child_type
][w
]++;
1283 mutex_exit(&pio
->io_lock
);
1288 * Now that all children have been reexecuted, execute the parent.
1289 * We don't reexecute "The Godfather" I/O here as it's the
1290 * responsibility of the caller to wait on him.
1292 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
))
1297 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1299 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1300 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1301 "failure and the failure mode property for this pool "
1302 "is set to panic.", spa_name(spa
));
1304 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1306 mutex_enter(&spa
->spa_suspend_lock
);
1308 if (spa
->spa_suspend_zio_root
== NULL
)
1309 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1310 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1311 ZIO_FLAG_GODFATHER
);
1313 spa
->spa_suspended
= B_TRUE
;
1316 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1317 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1318 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1319 ASSERT(zio_unique_parent(zio
) == NULL
);
1320 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1321 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1324 mutex_exit(&spa
->spa_suspend_lock
);
1328 zio_resume(spa_t
*spa
)
1333 * Reexecute all previously suspended i/o.
1335 mutex_enter(&spa
->spa_suspend_lock
);
1336 spa
->spa_suspended
= B_FALSE
;
1337 cv_broadcast(&spa
->spa_suspend_cv
);
1338 pio
= spa
->spa_suspend_zio_root
;
1339 spa
->spa_suspend_zio_root
= NULL
;
1340 mutex_exit(&spa
->spa_suspend_lock
);
1346 return (zio_wait(pio
));
1350 zio_resume_wait(spa_t
*spa
)
1352 mutex_enter(&spa
->spa_suspend_lock
);
1353 while (spa_suspended(spa
))
1354 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1355 mutex_exit(&spa
->spa_suspend_lock
);
1359 * ==========================================================================
1362 * A gang block is a collection of small blocks that looks to the DMU
1363 * like one large block. When zio_dva_allocate() cannot find a block
1364 * of the requested size, due to either severe fragmentation or the pool
1365 * being nearly full, it calls zio_write_gang_block() to construct the
1366 * block from smaller fragments.
1368 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1369 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1370 * an indirect block: it's an array of block pointers. It consumes
1371 * only one sector and hence is allocatable regardless of fragmentation.
1372 * The gang header's bps point to its gang members, which hold the data.
1374 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1375 * as the verifier to ensure uniqueness of the SHA256 checksum.
1376 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1377 * not the gang header. This ensures that data block signatures (needed for
1378 * deduplication) are independent of how the block is physically stored.
1380 * Gang blocks can be nested: a gang member may itself be a gang block.
1381 * Thus every gang block is a tree in which root and all interior nodes are
1382 * gang headers, and the leaves are normal blocks that contain user data.
1383 * The root of the gang tree is called the gang leader.
1385 * To perform any operation (read, rewrite, free, claim) on a gang block,
1386 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1387 * in the io_gang_tree field of the original logical i/o by recursively
1388 * reading the gang leader and all gang headers below it. This yields
1389 * an in-core tree containing the contents of every gang header and the
1390 * bps for every constituent of the gang block.
1392 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1393 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1394 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1395 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1396 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1397 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1398 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1399 * of the gang header plus zio_checksum_compute() of the data to update the
1400 * gang header's blk_cksum as described above.
1402 * The two-phase assemble/issue model solves the problem of partial failure --
1403 * what if you'd freed part of a gang block but then couldn't read the
1404 * gang header for another part? Assembling the entire gang tree first
1405 * ensures that all the necessary gang header I/O has succeeded before
1406 * starting the actual work of free, claim, or write. Once the gang tree
1407 * is assembled, free and claim are in-memory operations that cannot fail.
1409 * In the event that a gang write fails, zio_dva_unallocate() walks the
1410 * gang tree to immediately free (i.e. insert back into the space map)
1411 * everything we've allocated. This ensures that we don't get ENOSPC
1412 * errors during repeated suspend/resume cycles due to a flaky device.
1414 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1415 * the gang tree, we won't modify the block, so we can safely defer the free
1416 * (knowing that the block is still intact). If we *can* assemble the gang
1417 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1418 * each constituent bp and we can allocate a new block on the next sync pass.
1420 * In all cases, the gang tree allows complete recovery from partial failure.
1421 * ==========================================================================
1425 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1430 return (zio_read(pio
, pio
->io_spa
, bp
, data
, BP_GET_PSIZE(bp
),
1431 NULL
, NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1432 &pio
->io_bookmark
));
1436 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1441 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1442 gn
->gn_gbh
, SPA_GANGBLOCKSIZE
, NULL
, NULL
, pio
->io_priority
,
1443 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1445 * As we rewrite each gang header, the pipeline will compute
1446 * a new gang block header checksum for it; but no one will
1447 * compute a new data checksum, so we do that here. The one
1448 * exception is the gang leader: the pipeline already computed
1449 * its data checksum because that stage precedes gang assembly.
1450 * (Presently, nothing actually uses interior data checksums;
1451 * this is just good hygiene.)
1453 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1454 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1455 data
, BP_GET_PSIZE(bp
));
1458 * If we are here to damage data for testing purposes,
1459 * leave the GBH alone so that we can detect the damage.
1461 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1462 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1464 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1465 data
, BP_GET_PSIZE(bp
), NULL
, NULL
, pio
->io_priority
,
1466 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1474 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1476 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1477 ZIO_GANG_CHILD_FLAGS(pio
)));
1482 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1484 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1485 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1488 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
1497 static void zio_gang_tree_assemble_done(zio_t
*zio
);
1499 static zio_gang_node_t
*
1500 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
1502 zio_gang_node_t
*gn
;
1504 ASSERT(*gnpp
== NULL
);
1506 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
1507 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
1514 zio_gang_node_free(zio_gang_node_t
**gnpp
)
1516 zio_gang_node_t
*gn
= *gnpp
;
1519 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1520 ASSERT(gn
->gn_child
[g
] == NULL
);
1522 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1523 kmem_free(gn
, sizeof (*gn
));
1528 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
1530 zio_gang_node_t
*gn
= *gnpp
;
1536 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1537 zio_gang_tree_free(&gn
->gn_child
[g
]);
1539 zio_gang_node_free(gnpp
);
1543 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
1545 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
1547 ASSERT(gio
->io_gang_leader
== gio
);
1548 ASSERT(BP_IS_GANG(bp
));
1550 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gn
->gn_gbh
,
1551 SPA_GANGBLOCKSIZE
, zio_gang_tree_assemble_done
, gn
,
1552 gio
->io_priority
, ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
1556 zio_gang_tree_assemble_done(zio_t
*zio
)
1558 zio_t
*gio
= zio
->io_gang_leader
;
1559 zio_gang_node_t
*gn
= zio
->io_private
;
1560 blkptr_t
*bp
= zio
->io_bp
;
1563 ASSERT(gio
== zio_unique_parent(zio
));
1564 ASSERT(zio
->io_child_count
== 0);
1569 if (BP_SHOULD_BYTESWAP(bp
))
1570 byteswap_uint64_array(zio
->io_data
, zio
->io_size
);
1572 ASSERT(zio
->io_data
== gn
->gn_gbh
);
1573 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
1574 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
1576 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
1577 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
1578 if (!BP_IS_GANG(gbp
))
1580 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
1585 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, void *data
)
1587 zio_t
*gio
= pio
->io_gang_leader
;
1591 ASSERT(BP_IS_GANG(bp
) == !!gn
);
1592 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
1593 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
1596 * If you're a gang header, your data is in gn->gn_gbh.
1597 * If you're a gang member, your data is in 'data' and gn == NULL.
1599 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
);
1602 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
1604 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
1605 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
1606 if (BP_IS_HOLE(gbp
))
1608 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
);
1609 data
= (char *)data
+ BP_GET_PSIZE(gbp
);
1613 if (gn
== gio
->io_gang_tree
)
1614 ASSERT3P((char *)gio
->io_data
+ gio
->io_size
, ==, data
);
1621 zio_gang_assemble(zio_t
*zio
)
1623 blkptr_t
*bp
= zio
->io_bp
;
1625 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
1626 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
1628 zio
->io_gang_leader
= zio
;
1630 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
1632 return (ZIO_PIPELINE_CONTINUE
);
1636 zio_gang_issue(zio_t
*zio
)
1638 blkptr_t
*bp
= zio
->io_bp
;
1640 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
1641 return (ZIO_PIPELINE_STOP
);
1643 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
1644 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
1646 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
1647 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_data
);
1649 zio_gang_tree_free(&zio
->io_gang_tree
);
1651 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1653 return (ZIO_PIPELINE_CONTINUE
);
1657 zio_write_gang_member_ready(zio_t
*zio
)
1659 zio_t
*pio
= zio_unique_parent(zio
);
1660 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
;)
1661 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
1662 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
1666 if (BP_IS_HOLE(zio
->io_bp
))
1669 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
1671 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
1672 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
1673 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
1674 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
1675 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
1677 mutex_enter(&pio
->io_lock
);
1678 for (d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
1679 ASSERT(DVA_GET_GANG(&pdva
[d
]));
1680 asize
= DVA_GET_ASIZE(&pdva
[d
]);
1681 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
1682 DVA_SET_ASIZE(&pdva
[d
], asize
);
1684 mutex_exit(&pio
->io_lock
);
1688 zio_write_gang_block(zio_t
*pio
)
1690 spa_t
*spa
= pio
->io_spa
;
1691 blkptr_t
*bp
= pio
->io_bp
;
1692 zio_t
*gio
= pio
->io_gang_leader
;
1694 zio_gang_node_t
*gn
, **gnpp
;
1695 zio_gbh_phys_t
*gbh
;
1696 uint64_t txg
= pio
->io_txg
;
1697 uint64_t resid
= pio
->io_size
;
1699 int copies
= gio
->io_prop
.zp_copies
;
1700 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
1704 error
= metaslab_alloc(spa
, spa_normal_class(spa
), SPA_GANGBLOCKSIZE
,
1705 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
,
1706 METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
);
1708 pio
->io_error
= error
;
1709 return (ZIO_PIPELINE_CONTINUE
);
1713 gnpp
= &gio
->io_gang_tree
;
1715 gnpp
= pio
->io_private
;
1716 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
1719 gn
= zio_gang_node_alloc(gnpp
);
1721 bzero(gbh
, SPA_GANGBLOCKSIZE
);
1724 * Create the gang header.
1726 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh
, SPA_GANGBLOCKSIZE
, NULL
, NULL
,
1727 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1730 * Create and nowait the gang children.
1732 for (g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
1733 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
1735 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
1737 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
1738 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
1739 zp
.zp_type
= DMU_OT_NONE
;
1741 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
1743 zp
.zp_dedup_verify
= 0;
1745 zio_nowait(zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
1746 (char *)pio
->io_data
+ (pio
->io_size
- resid
), lsize
, &zp
,
1747 zio_write_gang_member_ready
, NULL
, &gn
->gn_child
[g
],
1748 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1749 &pio
->io_bookmark
));
1753 * Set pio's pipeline to just wait for zio to finish.
1755 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1759 return (ZIO_PIPELINE_CONTINUE
);
1763 * ==========================================================================
1765 * ==========================================================================
1768 zio_ddt_child_read_done(zio_t
*zio
)
1770 blkptr_t
*bp
= zio
->io_bp
;
1771 ddt_entry_t
*dde
= zio
->io_private
;
1773 zio_t
*pio
= zio_unique_parent(zio
);
1775 mutex_enter(&pio
->io_lock
);
1776 ddp
= ddt_phys_select(dde
, bp
);
1777 if (zio
->io_error
== 0)
1778 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
1779 if (zio
->io_error
== 0 && dde
->dde_repair_data
== NULL
)
1780 dde
->dde_repair_data
= zio
->io_data
;
1782 zio_buf_free(zio
->io_data
, zio
->io_size
);
1783 mutex_exit(&pio
->io_lock
);
1787 zio_ddt_read_start(zio_t
*zio
)
1789 blkptr_t
*bp
= zio
->io_bp
;
1792 ASSERT(BP_GET_DEDUP(bp
));
1793 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
1794 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1796 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
1797 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
1798 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
1799 ddt_phys_t
*ddp
= dde
->dde_phys
;
1800 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
1803 ASSERT(zio
->io_vsd
== NULL
);
1806 if (ddp_self
== NULL
)
1807 return (ZIO_PIPELINE_CONTINUE
);
1809 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
1810 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
1812 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
1814 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
1815 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
1816 zio_ddt_child_read_done
, dde
, zio
->io_priority
,
1817 ZIO_DDT_CHILD_FLAGS(zio
) | ZIO_FLAG_DONT_PROPAGATE
,
1818 &zio
->io_bookmark
));
1820 return (ZIO_PIPELINE_CONTINUE
);
1823 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
1824 zio
->io_data
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
1825 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
1827 return (ZIO_PIPELINE_CONTINUE
);
1831 zio_ddt_read_done(zio_t
*zio
)
1833 blkptr_t
*bp
= zio
->io_bp
;
1835 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
1836 return (ZIO_PIPELINE_STOP
);
1838 ASSERT(BP_GET_DEDUP(bp
));
1839 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
1840 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1842 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
1843 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
1844 ddt_entry_t
*dde
= zio
->io_vsd
;
1846 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
1847 return (ZIO_PIPELINE_CONTINUE
);
1850 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
1851 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1852 return (ZIO_PIPELINE_STOP
);
1854 if (dde
->dde_repair_data
!= NULL
) {
1855 bcopy(dde
->dde_repair_data
, zio
->io_data
, zio
->io_size
);
1856 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
1858 ddt_repair_done(ddt
, dde
);
1862 ASSERT(zio
->io_vsd
== NULL
);
1864 return (ZIO_PIPELINE_CONTINUE
);
1868 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
1870 spa_t
*spa
= zio
->io_spa
;
1874 * Note: we compare the original data, not the transformed data,
1875 * because when zio->io_bp is an override bp, we will not have
1876 * pushed the I/O transforms. That's an important optimization
1877 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1879 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
1880 zio_t
*lio
= dde
->dde_lead_zio
[p
];
1883 return (lio
->io_orig_size
!= zio
->io_orig_size
||
1884 bcmp(zio
->io_orig_data
, lio
->io_orig_data
,
1885 zio
->io_orig_size
) != 0);
1889 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
1890 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
1892 if (ddp
->ddp_phys_birth
!= 0) {
1893 arc_buf_t
*abuf
= NULL
;
1894 uint32_t aflags
= ARC_WAIT
;
1895 blkptr_t blk
= *zio
->io_bp
;
1898 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
1902 error
= arc_read_nolock(NULL
, spa
, &blk
,
1903 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
1904 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
1905 &aflags
, &zio
->io_bookmark
);
1908 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
1909 bcmp(abuf
->b_data
, zio
->io_orig_data
,
1910 zio
->io_orig_size
) != 0)
1912 VERIFY(arc_buf_remove_ref(abuf
, &abuf
) == 1);
1916 return (error
!= 0);
1924 zio_ddt_child_write_ready(zio_t
*zio
)
1926 int p
= zio
->io_prop
.zp_copies
;
1927 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
1928 ddt_entry_t
*dde
= zio
->io_private
;
1929 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
1937 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
1939 ddt_phys_fill(ddp
, zio
->io_bp
);
1941 while ((pio
= zio_walk_parents(zio
)) != NULL
)
1942 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
1948 zio_ddt_child_write_done(zio_t
*zio
)
1950 int p
= zio
->io_prop
.zp_copies
;
1951 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
1952 ddt_entry_t
*dde
= zio
->io_private
;
1953 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
1957 ASSERT(ddp
->ddp_refcnt
== 0);
1958 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
1959 dde
->dde_lead_zio
[p
] = NULL
;
1961 if (zio
->io_error
== 0) {
1962 while (zio_walk_parents(zio
) != NULL
)
1963 ddt_phys_addref(ddp
);
1965 ddt_phys_clear(ddp
);
1972 zio_ddt_ditto_write_done(zio_t
*zio
)
1974 int p
= DDT_PHYS_DITTO
;
1975 blkptr_t
*bp
= zio
->io_bp
;
1976 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
1977 ddt_entry_t
*dde
= zio
->io_private
;
1978 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
1979 ddt_key_t
*ddk
= &dde
->dde_key
;
1980 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
1984 ASSERT(ddp
->ddp_refcnt
== 0);
1985 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
1986 dde
->dde_lead_zio
[p
] = NULL
;
1988 if (zio
->io_error
== 0) {
1989 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
1990 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
1991 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
1992 if (ddp
->ddp_phys_birth
!= 0)
1993 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
1994 ddt_phys_fill(ddp
, bp
);
2001 zio_ddt_write(zio_t
*zio
)
2003 spa_t
*spa
= zio
->io_spa
;
2004 blkptr_t
*bp
= zio
->io_bp
;
2005 uint64_t txg
= zio
->io_txg
;
2006 zio_prop_t
*zp
= &zio
->io_prop
;
2007 int p
= zp
->zp_copies
;
2011 ddt_t
*ddt
= ddt_select(spa
, bp
);
2015 ASSERT(BP_GET_DEDUP(bp
));
2016 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2017 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2020 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2021 ddp
= &dde
->dde_phys
[p
];
2023 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2025 * If we're using a weak checksum, upgrade to a strong checksum
2026 * and try again. If we're already using a strong checksum,
2027 * we can't resolve it, so just convert to an ordinary write.
2028 * (And automatically e-mail a paper to Nature?)
2030 if (!zio_checksum_table
[zp
->zp_checksum
].ci_dedup
) {
2031 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2032 zio_pop_transforms(zio
);
2033 zio
->io_stage
= ZIO_STAGE_OPEN
;
2038 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2040 return (ZIO_PIPELINE_CONTINUE
);
2043 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2044 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2046 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2047 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2048 zio_prop_t czp
= *zp
;
2050 czp
.zp_copies
= ditto_copies
;
2053 * If we arrived here with an override bp, we won't have run
2054 * the transform stack, so we won't have the data we need to
2055 * generate a child i/o. So, toss the override bp and restart.
2056 * This is safe, because using the override bp is just an
2057 * optimization; and it's rare, so the cost doesn't matter.
2059 if (zio
->io_bp_override
) {
2060 zio_pop_transforms(zio
);
2061 zio
->io_stage
= ZIO_STAGE_OPEN
;
2062 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2063 zio
->io_bp_override
= NULL
;
2066 return (ZIO_PIPELINE_CONTINUE
);
2069 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_data
,
2070 zio
->io_orig_size
, &czp
, NULL
,
2071 zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2072 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2074 zio_push_transform(dio
, zio
->io_data
, zio
->io_size
, 0, NULL
);
2075 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2078 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2079 if (ddp
->ddp_phys_birth
!= 0)
2080 ddt_bp_fill(ddp
, bp
, txg
);
2081 if (dde
->dde_lead_zio
[p
] != NULL
)
2082 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2084 ddt_phys_addref(ddp
);
2085 } else if (zio
->io_bp_override
) {
2086 ASSERT(bp
->blk_birth
== txg
);
2087 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2088 ddt_phys_fill(ddp
, bp
);
2089 ddt_phys_addref(ddp
);
2091 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_data
,
2092 zio
->io_orig_size
, zp
, zio_ddt_child_write_ready
,
2093 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2094 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2096 zio_push_transform(cio
, zio
->io_data
, zio
->io_size
, 0, NULL
);
2097 dde
->dde_lead_zio
[p
] = cio
;
2107 return (ZIO_PIPELINE_CONTINUE
);
2110 ddt_entry_t
*freedde
; /* for debugging */
2113 zio_ddt_free(zio_t
*zio
)
2115 spa_t
*spa
= zio
->io_spa
;
2116 blkptr_t
*bp
= zio
->io_bp
;
2117 ddt_t
*ddt
= ddt_select(spa
, bp
);
2121 ASSERT(BP_GET_DEDUP(bp
));
2122 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2125 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2126 ddp
= ddt_phys_select(dde
, bp
);
2127 ddt_phys_decref(ddp
);
2130 return (ZIO_PIPELINE_CONTINUE
);
2134 * ==========================================================================
2135 * Allocate and free blocks
2136 * ==========================================================================
2139 zio_dva_allocate(zio_t
*zio
)
2141 spa_t
*spa
= zio
->io_spa
;
2142 metaslab_class_t
*mc
= spa_normal_class(spa
);
2143 blkptr_t
*bp
= zio
->io_bp
;
2146 if (zio
->io_gang_leader
== NULL
) {
2147 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2148 zio
->io_gang_leader
= zio
;
2151 ASSERT(BP_IS_HOLE(bp
));
2152 ASSERT3U(BP_GET_NDVAS(bp
), ==, 0);
2153 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2154 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2155 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2157 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2158 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, 0);
2161 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2162 return (zio_write_gang_block(zio
));
2163 zio
->io_error
= error
;
2166 return (ZIO_PIPELINE_CONTINUE
);
2170 zio_dva_free(zio_t
*zio
)
2172 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2174 return (ZIO_PIPELINE_CONTINUE
);
2178 zio_dva_claim(zio_t
*zio
)
2182 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2184 zio
->io_error
= error
;
2186 return (ZIO_PIPELINE_CONTINUE
);
2190 * Undo an allocation. This is used by zio_done() when an I/O fails
2191 * and we want to give back the block we just allocated.
2192 * This handles both normal blocks and gang blocks.
2195 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2199 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2200 ASSERT(zio
->io_bp_override
== NULL
);
2202 if (!BP_IS_HOLE(bp
))
2203 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2206 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2207 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2208 &gn
->gn_gbh
->zg_blkptr
[g
]);
2214 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2217 zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
, blkptr_t
*old_bp
,
2218 uint64_t size
, boolean_t use_slog
)
2222 ASSERT(txg
> spa_syncing_txg(spa
));
2225 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
,
2226 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
);
2229 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
2230 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
);
2233 BP_SET_LSIZE(new_bp
, size
);
2234 BP_SET_PSIZE(new_bp
, size
);
2235 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
2236 BP_SET_CHECKSUM(new_bp
,
2237 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
2238 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
2239 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
2240 BP_SET_LEVEL(new_bp
, 0);
2241 BP_SET_DEDUP(new_bp
, 0);
2242 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
2249 * Free an intent log block.
2252 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
2254 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
2255 ASSERT(!BP_IS_GANG(bp
));
2257 zio_free(spa
, txg
, bp
);
2261 * ==========================================================================
2262 * Read and write to physical devices
2263 * ==========================================================================
2266 zio_vdev_io_start(zio_t
*zio
)
2268 vdev_t
*vd
= zio
->io_vd
;
2270 spa_t
*spa
= zio
->io_spa
;
2272 ASSERT(zio
->io_error
== 0);
2273 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
2276 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
2277 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
2280 * The mirror_ops handle multiple DVAs in a single BP.
2282 return (vdev_mirror_ops
.vdev_op_io_start(zio
));
2286 * We keep track of time-sensitive I/Os so that the scan thread
2287 * can quickly react to certain workloads. In particular, we care
2288 * about non-scrubbing, top-level reads and writes with the following
2290 * - synchronous writes of user data to non-slog devices
2291 * - any reads of user data
2292 * When these conditions are met, adjust the timestamp of spa_last_io
2293 * which allows the scan thread to adjust its workload accordingly.
2295 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
2296 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
2297 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
2298 zio
->io_txg
!= spa_syncing_txg(spa
)) {
2299 uint64_t old
= spa
->spa_last_io
;
2300 uint64_t new = ddi_get_lbolt64();
2302 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
2305 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
2307 if (P2PHASE(zio
->io_size
, align
) != 0) {
2308 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
2309 char *abuf
= zio_buf_alloc(asize
);
2310 ASSERT(vd
== vd
->vdev_top
);
2311 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
2312 bcopy(zio
->io_data
, abuf
, zio
->io_size
);
2313 bzero(abuf
+ zio
->io_size
, asize
- zio
->io_size
);
2315 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
2318 ASSERT(P2PHASE(zio
->io_offset
, align
) == 0);
2319 ASSERT(P2PHASE(zio
->io_size
, align
) == 0);
2320 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
2323 * If this is a repair I/O, and there's no self-healing involved --
2324 * that is, we're just resilvering what we expect to resilver --
2325 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2326 * This prevents spurious resilvering with nested replication.
2327 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2328 * A is out of date, we'll read from C+D, then use the data to
2329 * resilver A+B -- but we don't actually want to resilver B, just A.
2330 * The top-level mirror has no way to know this, so instead we just
2331 * discard unnecessary repairs as we work our way down the vdev tree.
2332 * The same logic applies to any form of nested replication:
2333 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2335 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
2336 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
2337 zio
->io_txg
!= 0 && /* not a delegated i/o */
2338 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
2339 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2340 zio_vdev_io_bypass(zio
);
2341 return (ZIO_PIPELINE_CONTINUE
);
2344 if (vd
->vdev_ops
->vdev_op_leaf
&&
2345 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
2347 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
) == 0)
2348 return (ZIO_PIPELINE_CONTINUE
);
2350 if ((zio
= vdev_queue_io(zio
)) == NULL
)
2351 return (ZIO_PIPELINE_STOP
);
2353 if (!vdev_accessible(vd
, zio
)) {
2354 zio
->io_error
= ENXIO
;
2356 return (ZIO_PIPELINE_STOP
);
2360 return (vd
->vdev_ops
->vdev_op_io_start(zio
));
2364 zio_vdev_io_done(zio_t
*zio
)
2366 vdev_t
*vd
= zio
->io_vd
;
2367 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
2368 boolean_t unexpected_error
= B_FALSE
;
2370 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
2371 return (ZIO_PIPELINE_STOP
);
2373 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
2375 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
2377 vdev_queue_io_done(zio
);
2379 if (zio
->io_type
== ZIO_TYPE_WRITE
)
2380 vdev_cache_write(zio
);
2382 if (zio_injection_enabled
&& zio
->io_error
== 0)
2383 zio
->io_error
= zio_handle_device_injection(vd
,
2386 if (zio_injection_enabled
&& zio
->io_error
== 0)
2387 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
2389 if (zio
->io_error
) {
2390 if (!vdev_accessible(vd
, zio
)) {
2391 zio
->io_error
= ENXIO
;
2393 unexpected_error
= B_TRUE
;
2398 ops
->vdev_op_io_done(zio
);
2400 if (unexpected_error
)
2401 VERIFY(vdev_probe(vd
, zio
) == NULL
);
2403 return (ZIO_PIPELINE_CONTINUE
);
2407 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2408 * disk, and use that to finish the checksum ereport later.
2411 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
2412 const void *good_buf
)
2414 /* no processing needed */
2415 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
2420 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
2422 void *buf
= zio_buf_alloc(zio
->io_size
);
2424 bcopy(zio
->io_data
, buf
, zio
->io_size
);
2426 zcr
->zcr_cbinfo
= zio
->io_size
;
2427 zcr
->zcr_cbdata
= buf
;
2428 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
2429 zcr
->zcr_free
= zio_buf_free
;
2433 zio_vdev_io_assess(zio_t
*zio
)
2435 vdev_t
*vd
= zio
->io_vd
;
2437 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
2438 return (ZIO_PIPELINE_STOP
);
2440 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
2441 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
2443 if (zio
->io_vsd
!= NULL
) {
2444 zio
->io_vsd_ops
->vsd_free(zio
);
2448 if (zio_injection_enabled
&& zio
->io_error
== 0)
2449 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
2452 * If the I/O failed, determine whether we should attempt to retry it.
2454 * On retry, we cut in line in the issue queue, since we don't want
2455 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2457 if (zio
->io_error
&& vd
== NULL
&&
2458 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
2459 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
2460 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
2462 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
2463 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
2464 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
2465 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
2466 zio_requeue_io_start_cut_in_line
);
2467 return (ZIO_PIPELINE_STOP
);
2471 * If we got an error on a leaf device, convert it to ENXIO
2472 * if the device is not accessible at all.
2474 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
2475 !vdev_accessible(vd
, zio
))
2476 zio
->io_error
= ENXIO
;
2479 * If we can't write to an interior vdev (mirror or RAID-Z),
2480 * set vdev_cant_write so that we stop trying to allocate from it.
2482 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
2483 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
)
2484 vd
->vdev_cant_write
= B_TRUE
;
2487 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2489 return (ZIO_PIPELINE_CONTINUE
);
2493 zio_vdev_io_reissue(zio_t
*zio
)
2495 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
2496 ASSERT(zio
->io_error
== 0);
2498 zio
->io_stage
>>= 1;
2502 zio_vdev_io_redone(zio_t
*zio
)
2504 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
2506 zio
->io_stage
>>= 1;
2510 zio_vdev_io_bypass(zio_t
*zio
)
2512 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
2513 ASSERT(zio
->io_error
== 0);
2515 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
2516 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
2520 * ==========================================================================
2521 * Generate and verify checksums
2522 * ==========================================================================
2525 zio_checksum_generate(zio_t
*zio
)
2527 blkptr_t
*bp
= zio
->io_bp
;
2528 enum zio_checksum checksum
;
2532 * This is zio_write_phys().
2533 * We're either generating a label checksum, or none at all.
2535 checksum
= zio
->io_prop
.zp_checksum
;
2537 if (checksum
== ZIO_CHECKSUM_OFF
)
2538 return (ZIO_PIPELINE_CONTINUE
);
2540 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
2542 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
2543 ASSERT(!IO_IS_ALLOCATING(zio
));
2544 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
2546 checksum
= BP_GET_CHECKSUM(bp
);
2550 zio_checksum_compute(zio
, checksum
, zio
->io_data
, zio
->io_size
);
2552 return (ZIO_PIPELINE_CONTINUE
);
2556 zio_checksum_verify(zio_t
*zio
)
2558 zio_bad_cksum_t info
;
2559 blkptr_t
*bp
= zio
->io_bp
;
2562 ASSERT(zio
->io_vd
!= NULL
);
2566 * This is zio_read_phys().
2567 * We're either verifying a label checksum, or nothing at all.
2569 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
2570 return (ZIO_PIPELINE_CONTINUE
);
2572 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
2575 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
2576 zio
->io_error
= error
;
2577 if (!(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
2578 zfs_ereport_start_checksum(zio
->io_spa
,
2579 zio
->io_vd
, zio
, zio
->io_offset
,
2580 zio
->io_size
, NULL
, &info
);
2584 return (ZIO_PIPELINE_CONTINUE
);
2588 * Called by RAID-Z to ensure we don't compute the checksum twice.
2591 zio_checksum_verified(zio_t
*zio
)
2593 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
2597 * ==========================================================================
2598 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2599 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2600 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2601 * indicate errors that are specific to one I/O, and most likely permanent.
2602 * Any other error is presumed to be worse because we weren't expecting it.
2603 * ==========================================================================
2606 zio_worst_error(int e1
, int e2
)
2608 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
2611 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
2612 if (e1
== zio_error_rank
[r1
])
2615 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
2616 if (e2
== zio_error_rank
[r2
])
2619 return (r1
> r2
? e1
: e2
);
2623 * ==========================================================================
2625 * ==========================================================================
2628 zio_ready(zio_t
*zio
)
2630 blkptr_t
*bp
= zio
->io_bp
;
2631 zio_t
*pio
, *pio_next
;
2633 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
2634 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
2635 return (ZIO_PIPELINE_STOP
);
2637 if (zio
->io_ready
) {
2638 ASSERT(IO_IS_ALLOCATING(zio
));
2639 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2640 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
2645 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
2646 zio
->io_bp_copy
= *bp
;
2649 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2651 mutex_enter(&zio
->io_lock
);
2652 zio
->io_state
[ZIO_WAIT_READY
] = 1;
2653 pio
= zio_walk_parents(zio
);
2654 mutex_exit(&zio
->io_lock
);
2657 * As we notify zio's parents, new parents could be added.
2658 * New parents go to the head of zio's io_parent_list, however,
2659 * so we will (correctly) not notify them. The remainder of zio's
2660 * io_parent_list, from 'pio_next' onward, cannot change because
2661 * all parents must wait for us to be done before they can be done.
2663 for (; pio
!= NULL
; pio
= pio_next
) {
2664 pio_next
= zio_walk_parents(zio
);
2665 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
2668 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2669 if (BP_IS_GANG(bp
)) {
2670 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
2672 ASSERT((uintptr_t)zio
->io_data
< SPA_MAXBLOCKSIZE
);
2673 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2677 if (zio_injection_enabled
&&
2678 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
2679 zio_handle_ignored_writes(zio
);
2681 return (ZIO_PIPELINE_CONTINUE
);
2685 zio_done(zio_t
*zio
)
2687 zio_t
*pio
, *pio_next
;
2691 * If our children haven't all completed,
2692 * wait for them and then repeat this pipeline stage.
2694 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
2695 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
2696 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
2697 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
2698 return (ZIO_PIPELINE_STOP
);
2700 for (c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2701 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2702 ASSERT(zio
->io_children
[c
][w
] == 0);
2704 if (zio
->io_bp
!= NULL
) {
2705 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
2706 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
2707 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
2708 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
2709 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
2710 zio
->io_bp_override
== NULL
&&
2711 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
2712 ASSERT(!BP_SHOULD_BYTESWAP(zio
->io_bp
));
2713 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2714 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
2715 (BP_COUNT_GANG(zio
->io_bp
) == BP_GET_NDVAS(zio
->io_bp
)));
2720 * If there were child vdev/gang/ddt errors, they apply to us now.
2722 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
2723 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
2724 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
2727 * If the I/O on the transformed data was successful, generate any
2728 * checksum reports now while we still have the transformed data.
2730 if (zio
->io_error
== 0) {
2731 while (zio
->io_cksum_report
!= NULL
) {
2732 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
2733 uint64_t align
= zcr
->zcr_align
;
2734 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
2735 char *abuf
= zio
->io_data
;
2737 if (asize
!= zio
->io_size
) {
2738 abuf
= zio_buf_alloc(asize
);
2739 bcopy(zio
->io_data
, abuf
, zio
->io_size
);
2740 bzero(abuf
+ zio
->io_size
, asize
- zio
->io_size
);
2743 zio
->io_cksum_report
= zcr
->zcr_next
;
2744 zcr
->zcr_next
= NULL
;
2745 zcr
->zcr_finish(zcr
, abuf
);
2746 zfs_ereport_free_checksum(zcr
);
2748 if (asize
!= zio
->io_size
)
2749 zio_buf_free(abuf
, asize
);
2753 zio_pop_transforms(zio
); /* note: may set zio->io_error */
2755 vdev_stat_update(zio
, zio
->io_size
);
2758 * If this I/O is attached to a particular vdev is slow, exeeding
2759 * 30 seconds to complete, post an error described the I/O delay.
2760 * We ignore these errors if the device is currently unavailable.
2762 if (zio
->io_delay
>= zio_delay_max
) {
2763 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
2764 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
2765 zio
->io_vd
, zio
, 0, 0);
2768 if (zio
->io_error
) {
2770 * If this I/O is attached to a particular vdev,
2771 * generate an error message describing the I/O failure
2772 * at the block level. We ignore these errors if the
2773 * device is currently unavailable.
2775 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
2776 !vdev_is_dead(zio
->io_vd
))
2777 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
2778 zio
->io_vd
, zio
, 0, 0);
2780 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
2781 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
2782 zio
== zio
->io_logical
) {
2784 * For logical I/O requests, tell the SPA to log the
2785 * error and generate a logical data ereport.
2787 spa_log_error(zio
->io_spa
, zio
);
2788 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
, NULL
, zio
,
2793 if (zio
->io_error
&& zio
== zio
->io_logical
) {
2795 * Determine whether zio should be reexecuted. This will
2796 * propagate all the way to the root via zio_notify_parent().
2798 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
2799 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2801 if (IO_IS_ALLOCATING(zio
) &&
2802 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
2803 if (zio
->io_error
!= ENOSPC
)
2804 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
2806 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2809 if ((zio
->io_type
== ZIO_TYPE_READ
||
2810 zio
->io_type
== ZIO_TYPE_FREE
) &&
2811 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
2812 zio
->io_error
== ENXIO
&&
2813 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
2814 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
2815 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2817 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
2818 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2821 * Here is a possibly good place to attempt to do
2822 * either combinatorial reconstruction or error correction
2823 * based on checksums. It also might be a good place
2824 * to send out preliminary ereports before we suspend
2830 * If there were logical child errors, they apply to us now.
2831 * We defer this until now to avoid conflating logical child
2832 * errors with errors that happened to the zio itself when
2833 * updating vdev stats and reporting FMA events above.
2835 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
2837 if ((zio
->io_error
|| zio
->io_reexecute
) &&
2838 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
2839 !(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
))
2840 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
2842 zio_gang_tree_free(&zio
->io_gang_tree
);
2845 * Godfather I/Os should never suspend.
2847 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
2848 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
2849 zio
->io_reexecute
= 0;
2851 if (zio
->io_reexecute
) {
2853 * This is a logical I/O that wants to reexecute.
2855 * Reexecute is top-down. When an i/o fails, if it's not
2856 * the root, it simply notifies its parent and sticks around.
2857 * The parent, seeing that it still has children in zio_done(),
2858 * does the same. This percolates all the way up to the root.
2859 * The root i/o will reexecute or suspend the entire tree.
2861 * This approach ensures that zio_reexecute() honors
2862 * all the original i/o dependency relationships, e.g.
2863 * parents not executing until children are ready.
2865 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2867 zio
->io_gang_leader
= NULL
;
2869 mutex_enter(&zio
->io_lock
);
2870 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
2871 mutex_exit(&zio
->io_lock
);
2874 * "The Godfather" I/O monitors its children but is
2875 * not a true parent to them. It will track them through
2876 * the pipeline but severs its ties whenever they get into
2877 * trouble (e.g. suspended). This allows "The Godfather"
2878 * I/O to return status without blocking.
2880 for (pio
= zio_walk_parents(zio
); pio
!= NULL
; pio
= pio_next
) {
2881 zio_link_t
*zl
= zio
->io_walk_link
;
2882 pio_next
= zio_walk_parents(zio
);
2884 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
2885 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
2886 zio_remove_child(pio
, zio
, zl
);
2887 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
2891 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
2893 * We're not a root i/o, so there's nothing to do
2894 * but notify our parent. Don't propagate errors
2895 * upward since we haven't permanently failed yet.
2897 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2898 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
2899 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
2900 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
2902 * We'd fail again if we reexecuted now, so suspend
2903 * until conditions improve (e.g. device comes online).
2905 zio_suspend(zio
->io_spa
, zio
);
2908 * Reexecution is potentially a huge amount of work.
2909 * Hand it off to the otherwise-unused claim taskq.
2911 (void) taskq_dispatch(
2912 zio
->io_spa
->spa_zio_taskq
[ZIO_TYPE_CLAIM
][ZIO_TASKQ_ISSUE
],
2913 (task_func_t
*)zio_reexecute
, zio
, TQ_SLEEP
);
2915 return (ZIO_PIPELINE_STOP
);
2918 ASSERT(zio
->io_child_count
== 0);
2919 ASSERT(zio
->io_reexecute
== 0);
2920 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
2923 * Report any checksum errors, since the I/O is complete.
2925 while (zio
->io_cksum_report
!= NULL
) {
2926 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
2927 zio
->io_cksum_report
= zcr
->zcr_next
;
2928 zcr
->zcr_next
= NULL
;
2929 zcr
->zcr_finish(zcr
, NULL
);
2930 zfs_ereport_free_checksum(zcr
);
2934 * It is the responsibility of the done callback to ensure that this
2935 * particular zio is no longer discoverable for adoption, and as
2936 * such, cannot acquire any new parents.
2941 mutex_enter(&zio
->io_lock
);
2942 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
2943 mutex_exit(&zio
->io_lock
);
2945 for (pio
= zio_walk_parents(zio
); pio
!= NULL
; pio
= pio_next
) {
2946 zio_link_t
*zl
= zio
->io_walk_link
;
2947 pio_next
= zio_walk_parents(zio
);
2948 zio_remove_child(pio
, zio
, zl
);
2949 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
2952 if (zio
->io_waiter
!= NULL
) {
2953 mutex_enter(&zio
->io_lock
);
2954 zio
->io_executor
= NULL
;
2955 cv_broadcast(&zio
->io_cv
);
2956 mutex_exit(&zio
->io_lock
);
2961 return (ZIO_PIPELINE_STOP
);
2965 * ==========================================================================
2966 * I/O pipeline definition
2967 * ==========================================================================
2969 static zio_pipe_stage_t
*zio_pipeline
[] = {
2975 zio_checksum_generate
,
2989 zio_checksum_verify
,
2993 #if defined(_KERNEL) && defined(HAVE_SPL)
2994 /* Fault injection */
2995 EXPORT_SYMBOL(zio_injection_enabled
);
2996 EXPORT_SYMBOL(zio_inject_fault
);
2997 EXPORT_SYMBOL(zio_inject_list_next
);
2998 EXPORT_SYMBOL(zio_clear_fault
);
2999 EXPORT_SYMBOL(zio_handle_fault_injection
);
3000 EXPORT_SYMBOL(zio_handle_device_injection
);
3001 EXPORT_SYMBOL(zio_handle_label_injection
);
3002 EXPORT_SYMBOL(zio_priority_table
);
3003 EXPORT_SYMBOL(zio_type_name
);
3005 module_param(zio_bulk_flags
, int, 0644);
3006 MODULE_PARM_DESC(zio_bulk_flags
, "Additional flags to pass to bulk buffers");
3008 module_param(zio_delay_max
, int, 0644);
3009 MODULE_PARM_DESC(zio_delay_max
, "Max zio delay before posting an event (ms)");