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) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table
[ZIO_PRIORITY_TABLE_SIZE
] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name
[ZIO_TYPES
] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t
*zio_cache
;
75 kmem_cache_t
*zio_link_cache
;
76 kmem_cache_t
*zio_vdev_cache
;
77 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
79 int zio_bulk_flags
= 0;
80 int zio_delay_max
= ZIO_DELAY_MAX
;
83 extern vmem_t
*zio_alloc_arena
;
85 extern int zfs_mg_alloc_failures
;
88 * An allocating zio is one that either currently has the DVA allocate
89 * stage set or will have it later in its lifetime.
91 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
93 int zio_requeue_io_start_cut_in_line
= 1;
96 int zio_buf_debug_limit
= 16384;
98 int zio_buf_debug_limit
= 0;
101 static inline void __zio_execute(zio_t
*zio
);
104 zio_cons(void *arg
, void *unused
, int kmflag
)
108 bzero(zio
, sizeof (zio_t
));
110 mutex_init(&zio
->io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
111 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
113 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
114 offsetof(zio_link_t
, zl_parent_node
));
115 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
116 offsetof(zio_link_t
, zl_child_node
));
122 zio_dest(void *arg
, void *unused
)
126 mutex_destroy(&zio
->io_lock
);
127 cv_destroy(&zio
->io_cv
);
128 list_destroy(&zio
->io_parent_list
);
129 list_destroy(&zio
->io_child_list
);
136 vmem_t
*data_alloc_arena
= NULL
;
139 data_alloc_arena
= zio_alloc_arena
;
141 zio_cache
= kmem_cache_create("zio_cache", sizeof (zio_t
), 0,
142 zio_cons
, zio_dest
, NULL
, NULL
, NULL
, KMC_KMEM
);
143 zio_link_cache
= kmem_cache_create("zio_link_cache",
144 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, KMC_KMEM
);
145 zio_vdev_cache
= kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t
),
146 PAGESIZE
, NULL
, NULL
, NULL
, NULL
, NULL
, KMC_VMEM
);
149 * For small buffers, we want a cache for each multiple of
150 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
151 * for each quarter-power of 2. For large buffers, we want
152 * a cache for each multiple of PAGESIZE.
154 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
155 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
159 while (p2
& (p2
- 1))
162 if (size
<= 4 * SPA_MINBLOCKSIZE
) {
163 align
= SPA_MINBLOCKSIZE
;
164 } else if (P2PHASE(size
, PAGESIZE
) == 0) {
166 } else if (P2PHASE(size
, p2
>> 2) == 0) {
172 int flags
= zio_bulk_flags
;
175 * The smallest buffers (512b) are heavily used and
176 * experience a lot of churn. The slabs allocated
177 * for them are also relatively small (32K). Thus
178 * in over to avoid expensive calls to vmalloc() we
179 * make an exception to the usual slab allocation
180 * policy and force these buffers to be kmem backed.
182 if (size
== (1 << SPA_MINBLOCKSHIFT
))
185 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
186 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
187 align
, NULL
, NULL
, NULL
, NULL
, NULL
, flags
);
189 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
190 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
191 align
, NULL
, NULL
, NULL
, NULL
,
192 data_alloc_arena
, flags
);
197 ASSERT(zio_buf_cache
[c
] != NULL
);
198 if (zio_buf_cache
[c
- 1] == NULL
)
199 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
201 ASSERT(zio_data_buf_cache
[c
] != NULL
);
202 if (zio_data_buf_cache
[c
- 1] == NULL
)
203 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
207 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
208 * to fail 3 times per txg or 8 failures, whichever is greater.
210 zfs_mg_alloc_failures
= MAX((3 * max_ncpus
/ 2), 8);
219 kmem_cache_t
*last_cache
= NULL
;
220 kmem_cache_t
*last_data_cache
= NULL
;
222 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
223 if (zio_buf_cache
[c
] != last_cache
) {
224 last_cache
= zio_buf_cache
[c
];
225 kmem_cache_destroy(zio_buf_cache
[c
]);
227 zio_buf_cache
[c
] = NULL
;
229 if (zio_data_buf_cache
[c
] != last_data_cache
) {
230 last_data_cache
= zio_data_buf_cache
[c
];
231 kmem_cache_destroy(zio_data_buf_cache
[c
]);
233 zio_data_buf_cache
[c
] = NULL
;
236 kmem_cache_destroy(zio_vdev_cache
);
237 kmem_cache_destroy(zio_link_cache
);
238 kmem_cache_destroy(zio_cache
);
244 * ==========================================================================
245 * Allocate and free I/O buffers
246 * ==========================================================================
250 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
251 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
252 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
253 * excess / transient data in-core during a crashdump.
256 zio_buf_alloc(size_t size
)
258 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
260 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
262 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
| KM_NODEBUG
));
266 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
267 * crashdump if the kernel panics. This exists so that we will limit the amount
268 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
269 * of kernel heap dumped to disk when the kernel panics)
272 zio_data_buf_alloc(size_t size
)
274 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
276 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
278 return (kmem_cache_alloc(zio_data_buf_cache
[c
],
279 KM_PUSHPAGE
| KM_NODEBUG
));
283 zio_buf_free(void *buf
, size_t size
)
285 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
287 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
289 kmem_cache_free(zio_buf_cache
[c
], buf
);
293 zio_data_buf_free(void *buf
, size_t size
)
295 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
297 ASSERT(c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
299 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
303 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
304 * or significantly delays the issuing of a zio. These buffers are used
305 * to aggregate I/O and could be used for raidz stripes.
310 return (kmem_cache_alloc(zio_vdev_cache
, KM_PUSHPAGE
));
314 zio_vdev_free(void *buf
)
316 kmem_cache_free(zio_vdev_cache
, buf
);
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
326 zio_push_transform(zio_t
*zio
, void *data
, uint64_t size
, uint64_t bufsize
,
327 zio_transform_func_t
*transform
)
329 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_PUSHPAGE
);
331 zt
->zt_orig_data
= zio
->io_data
;
332 zt
->zt_orig_size
= zio
->io_size
;
333 zt
->zt_bufsize
= bufsize
;
334 zt
->zt_transform
= transform
;
336 zt
->zt_next
= zio
->io_transform_stack
;
337 zio
->io_transform_stack
= zt
;
344 zio_pop_transforms(zio_t
*zio
)
348 while ((zt
= zio
->io_transform_stack
) != NULL
) {
349 if (zt
->zt_transform
!= NULL
)
350 zt
->zt_transform(zio
,
351 zt
->zt_orig_data
, zt
->zt_orig_size
);
353 if (zt
->zt_bufsize
!= 0)
354 zio_buf_free(zio
->io_data
, zt
->zt_bufsize
);
356 zio
->io_data
= zt
->zt_orig_data
;
357 zio
->io_size
= zt
->zt_orig_size
;
358 zio
->io_transform_stack
= zt
->zt_next
;
360 kmem_free(zt
, sizeof (zio_transform_t
));
365 * ==========================================================================
366 * I/O transform callbacks for subblocks and decompression
367 * ==========================================================================
370 zio_subblock(zio_t
*zio
, void *data
, uint64_t size
)
372 ASSERT(zio
->io_size
> size
);
374 if (zio
->io_type
== ZIO_TYPE_READ
)
375 bcopy(zio
->io_data
, data
, size
);
379 zio_decompress(zio_t
*zio
, void *data
, uint64_t size
)
381 if (zio
->io_error
== 0 &&
382 zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
383 zio
->io_data
, data
, zio
->io_size
, size
) != 0)
388 * ==========================================================================
389 * I/O parent/child relationships and pipeline interlocks
390 * ==========================================================================
393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
394 * continue calling these functions until they return NULL.
395 * Otherwise, the next caller will pick up the list walk in
396 * some indeterminate state. (Otherwise every caller would
397 * have to pass in a cookie to keep the state represented by
398 * io_walk_link, which gets annoying.)
401 zio_walk_parents(zio_t
*cio
)
403 zio_link_t
*zl
= cio
->io_walk_link
;
404 list_t
*pl
= &cio
->io_parent_list
;
406 zl
= (zl
== NULL
) ? list_head(pl
) : list_next(pl
, zl
);
407 cio
->io_walk_link
= zl
;
412 ASSERT(zl
->zl_child
== cio
);
413 return (zl
->zl_parent
);
417 zio_walk_children(zio_t
*pio
)
419 zio_link_t
*zl
= pio
->io_walk_link
;
420 list_t
*cl
= &pio
->io_child_list
;
422 zl
= (zl
== NULL
) ? list_head(cl
) : list_next(cl
, zl
);
423 pio
->io_walk_link
= zl
;
428 ASSERT(zl
->zl_parent
== pio
);
429 return (zl
->zl_child
);
433 zio_unique_parent(zio_t
*cio
)
435 zio_t
*pio
= zio_walk_parents(cio
);
437 VERIFY(zio_walk_parents(cio
) == NULL
);
442 zio_add_child(zio_t
*pio
, zio_t
*cio
)
444 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_PUSHPAGE
);
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
453 ASSERT(cio
->io_child_type
<= pio
->io_child_type
);
458 mutex_enter(&cio
->io_lock
);
459 mutex_enter(&pio
->io_lock
);
461 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
463 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
464 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
466 list_insert_head(&pio
->io_child_list
, zl
);
467 list_insert_head(&cio
->io_parent_list
, zl
);
469 pio
->io_child_count
++;
470 cio
->io_parent_count
++;
472 mutex_exit(&pio
->io_lock
);
473 mutex_exit(&cio
->io_lock
);
477 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
479 ASSERT(zl
->zl_parent
== pio
);
480 ASSERT(zl
->zl_child
== cio
);
482 mutex_enter(&cio
->io_lock
);
483 mutex_enter(&pio
->io_lock
);
485 list_remove(&pio
->io_child_list
, zl
);
486 list_remove(&cio
->io_parent_list
, zl
);
488 pio
->io_child_count
--;
489 cio
->io_parent_count
--;
491 mutex_exit(&pio
->io_lock
);
492 mutex_exit(&cio
->io_lock
);
494 kmem_cache_free(zio_link_cache
, zl
);
498 zio_wait_for_children(zio_t
*zio
, enum zio_child child
, enum zio_wait_type wait
)
500 uint64_t *countp
= &zio
->io_children
[child
][wait
];
501 boolean_t waiting
= B_FALSE
;
503 mutex_enter(&zio
->io_lock
);
504 ASSERT(zio
->io_stall
== NULL
);
507 zio
->io_stall
= countp
;
510 mutex_exit(&zio
->io_lock
);
515 __attribute__((always_inline
))
517 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
519 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
520 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
522 mutex_enter(&pio
->io_lock
);
523 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
524 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
525 pio
->io_reexecute
|= zio
->io_reexecute
;
526 ASSERT3U(*countp
, >, 0);
527 if (--*countp
== 0 && pio
->io_stall
== countp
) {
528 pio
->io_stall
= NULL
;
529 mutex_exit(&pio
->io_lock
);
532 mutex_exit(&pio
->io_lock
);
537 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
539 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
540 zio
->io_error
= zio
->io_child_error
[c
];
544 * ==========================================================================
545 * Create the various types of I/O (read, write, free, etc)
546 * ==========================================================================
549 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
550 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
551 zio_type_t type
, int priority
, enum zio_flag flags
,
552 vdev_t
*vd
, uint64_t offset
, const zbookmark_t
*zb
,
553 enum zio_stage stage
, enum zio_stage pipeline
)
557 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
558 ASSERT(P2PHASE(size
, SPA_MINBLOCKSIZE
) == 0);
559 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
561 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
562 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
563 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
565 zio
= kmem_cache_alloc(zio_cache
, KM_PUSHPAGE
);
568 zio
->io_child_type
= ZIO_CHILD_VDEV
;
569 else if (flags
& ZIO_FLAG_GANG_CHILD
)
570 zio
->io_child_type
= ZIO_CHILD_GANG
;
571 else if (flags
& ZIO_FLAG_DDT_CHILD
)
572 zio
->io_child_type
= ZIO_CHILD_DDT
;
574 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
577 zio
->io_logical
= NULL
;
578 zio
->io_bp
= (blkptr_t
*)bp
;
579 zio
->io_bp_copy
= *bp
;
580 zio
->io_bp_orig
= *bp
;
581 if (type
!= ZIO_TYPE_WRITE
||
582 zio
->io_child_type
== ZIO_CHILD_DDT
)
583 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
584 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
585 zio
->io_logical
= zio
;
586 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
587 pipeline
|= ZIO_GANG_STAGES
;
589 zio
->io_logical
= NULL
;
591 bzero(&zio
->io_bp_copy
, sizeof (blkptr_t
));
592 bzero(&zio
->io_bp_orig
, sizeof (blkptr_t
));
597 zio
->io_ready
= NULL
;
599 zio
->io_private
= private;
600 zio
->io_prev_space_delta
= 0;
602 zio
->io_priority
= priority
;
605 zio
->io_vsd_ops
= NULL
;
606 zio
->io_offset
= offset
;
607 zio
->io_deadline
= 0;
608 zio
->io_orig_data
= zio
->io_data
= data
;
609 zio
->io_orig_size
= zio
->io_size
= size
;
610 zio
->io_orig_flags
= zio
->io_flags
= flags
;
611 zio
->io_orig_stage
= zio
->io_stage
= stage
;
612 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
613 bzero(&zio
->io_prop
, sizeof (zio_prop_t
));
615 zio
->io_reexecute
= 0;
616 zio
->io_bp_override
= NULL
;
617 zio
->io_walk_link
= NULL
;
618 zio
->io_transform_stack
= NULL
;
621 zio
->io_child_count
= 0;
622 zio
->io_parent_count
= 0;
623 zio
->io_stall
= NULL
;
624 zio
->io_gang_leader
= NULL
;
625 zio
->io_gang_tree
= NULL
;
626 zio
->io_executor
= NULL
;
627 zio
->io_waiter
= NULL
;
628 zio
->io_cksum_report
= NULL
;
630 bzero(zio
->io_child_error
, sizeof (int) * ZIO_CHILD_TYPES
);
631 bzero(zio
->io_children
,
632 sizeof (uint64_t) * ZIO_CHILD_TYPES
* ZIO_WAIT_TYPES
);
633 bzero(&zio
->io_bookmark
, sizeof (zbookmark_t
));
635 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
636 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
639 zio
->io_bookmark
= *zb
;
642 if (zio
->io_logical
== NULL
)
643 zio
->io_logical
= pio
->io_logical
;
644 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
645 zio
->io_gang_leader
= pio
->io_gang_leader
;
646 zio_add_child(pio
, zio
);
649 taskq_init_ent(&zio
->io_tqent
);
655 zio_destroy(zio_t
*zio
)
657 kmem_cache_free(zio_cache
, zio
);
661 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
662 void *private, enum zio_flag flags
)
666 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, done
, private,
667 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
668 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
674 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
676 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
680 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
681 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
682 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
)
686 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
687 data
, size
, done
, private,
688 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
689 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
690 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
696 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
697 void *data
, uint64_t size
, const zio_prop_t
*zp
,
698 zio_done_func_t
*ready
, zio_done_func_t
*done
, void *private,
699 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
)
703 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
704 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
705 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
706 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
707 zp
->zp_type
< DMU_OT_NUMTYPES
&&
710 zp
->zp_copies
<= spa_max_replication(spa
) &&
712 zp
->zp_dedup_verify
<= 1);
714 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, done
, private,
715 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
716 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
717 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
719 zio
->io_ready
= ready
;
726 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, void *data
,
727 uint64_t size
, zio_done_func_t
*done
, void *private, int priority
,
728 enum zio_flag flags
, zbookmark_t
*zb
)
732 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, done
, private,
733 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
734 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
740 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
)
742 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
743 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
744 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
745 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
747 zio
->io_prop
.zp_copies
= copies
;
748 zio
->io_bp_override
= bp
;
752 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
754 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
758 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
763 dprintf_bp(bp
, "freeing in txg %llu, pass %u",
764 (longlong_t
)txg
, spa
->spa_sync_pass
);
766 ASSERT(!BP_IS_HOLE(bp
));
767 ASSERT(spa_syncing_txg(spa
) == txg
);
768 ASSERT(spa_sync_pass(spa
) <= SYNC_PASS_DEFERRED_FREE
);
770 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
771 NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_FREE
, flags
,
772 NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_FREE_PIPELINE
);
778 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
779 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
784 * A claim is an allocation of a specific block. Claims are needed
785 * to support immediate writes in the intent log. The issue is that
786 * immediate writes contain committed data, but in a txg that was
787 * *not* committed. Upon opening the pool after an unclean shutdown,
788 * the intent log claims all blocks that contain immediate write data
789 * so that the SPA knows they're in use.
791 * All claims *must* be resolved in the first txg -- before the SPA
792 * starts allocating blocks -- so that nothing is allocated twice.
793 * If txg == 0 we just verify that the block is claimable.
795 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
796 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
797 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
799 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
800 done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
, flags
,
801 NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
807 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
808 zio_done_func_t
*done
, void *private, int priority
, enum zio_flag flags
)
813 if (vd
->vdev_children
== 0) {
814 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, done
, private,
815 ZIO_TYPE_IOCTL
, priority
, flags
, vd
, 0, NULL
,
816 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
820 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
822 for (c
= 0; c
< vd
->vdev_children
; c
++)
823 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
824 done
, private, priority
, flags
));
831 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
832 void *data
, int checksum
, zio_done_func_t
*done
, void *private,
833 int priority
, enum zio_flag flags
, boolean_t labels
)
837 ASSERT(vd
->vdev_children
== 0);
838 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
839 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
840 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
842 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, done
, private,
843 ZIO_TYPE_READ
, priority
, flags
, vd
, offset
, NULL
,
844 ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
846 zio
->io_prop
.zp_checksum
= checksum
;
852 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
853 void *data
, int checksum
, zio_done_func_t
*done
, void *private,
854 int priority
, enum zio_flag flags
, boolean_t labels
)
858 ASSERT(vd
->vdev_children
== 0);
859 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
860 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
861 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
863 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, done
, private,
864 ZIO_TYPE_WRITE
, priority
, flags
, vd
, offset
, NULL
,
865 ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
867 zio
->io_prop
.zp_checksum
= checksum
;
869 if (zio_checksum_table
[checksum
].ci_eck
) {
871 * zec checksums are necessarily destructive -- they modify
872 * the end of the write buffer to hold the verifier/checksum.
873 * Therefore, we must make a local copy in case the data is
874 * being written to multiple places in parallel.
876 void *wbuf
= zio_buf_alloc(size
);
877 bcopy(data
, wbuf
, size
);
878 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
885 * Create a child I/O to do some work for us.
888 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
889 void *data
, uint64_t size
, int type
, int priority
, enum zio_flag flags
,
890 zio_done_func_t
*done
, void *private)
892 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
895 ASSERT(vd
->vdev_parent
==
896 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
898 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
900 * If we have the bp, then the child should perform the
901 * checksum and the parent need not. This pushes error
902 * detection as close to the leaves as possible and
903 * eliminates redundant checksums in the interior nodes.
905 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
906 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
909 if (vd
->vdev_children
== 0)
910 offset
+= VDEV_LABEL_START_SIZE
;
912 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
915 * If we've decided to do a repair, the write is not speculative --
916 * even if the original read was.
918 if (flags
& ZIO_FLAG_IO_REPAIR
)
919 flags
&= ~ZIO_FLAG_SPECULATIVE
;
921 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
,
922 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
923 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
929 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, void *data
, uint64_t size
,
930 int type
, int priority
, enum zio_flag flags
,
931 zio_done_func_t
*done
, void *private)
935 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
937 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
938 data
, size
, done
, private, type
, priority
,
939 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
,
941 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
947 zio_flush(zio_t
*zio
, vdev_t
*vd
)
949 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
950 NULL
, NULL
, ZIO_PRIORITY_NOW
,
951 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
955 zio_shrink(zio_t
*zio
, uint64_t size
)
957 ASSERT(zio
->io_executor
== NULL
);
958 ASSERT(zio
->io_orig_size
== zio
->io_size
);
959 ASSERT(size
<= zio
->io_size
);
962 * We don't shrink for raidz because of problems with the
963 * reconstruction when reading back less than the block size.
964 * Note, BP_IS_RAIDZ() assumes no compression.
966 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
967 if (!BP_IS_RAIDZ(zio
->io_bp
))
968 zio
->io_orig_size
= zio
->io_size
= size
;
972 * ==========================================================================
973 * Prepare to read and write logical blocks
974 * ==========================================================================
978 zio_read_bp_init(zio_t
*zio
)
980 blkptr_t
*bp
= zio
->io_bp
;
982 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
983 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
984 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
985 uint64_t psize
= BP_GET_PSIZE(bp
);
986 void *cbuf
= zio_buf_alloc(psize
);
988 zio_push_transform(zio
, cbuf
, psize
, psize
, zio_decompress
);
991 if (!dmu_ot
[BP_GET_TYPE(bp
)].ot_metadata
&& BP_GET_LEVEL(bp
) == 0)
992 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
994 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
995 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
997 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
998 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1000 return (ZIO_PIPELINE_CONTINUE
);
1004 zio_write_bp_init(zio_t
*zio
)
1006 spa_t
*spa
= zio
->io_spa
;
1007 zio_prop_t
*zp
= &zio
->io_prop
;
1008 enum zio_compress compress
= zp
->zp_compress
;
1009 blkptr_t
*bp
= zio
->io_bp
;
1010 uint64_t lsize
= zio
->io_size
;
1011 uint64_t psize
= lsize
;
1015 * If our children haven't all reached the ready stage,
1016 * wait for them and then repeat this pipeline stage.
1018 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1019 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1020 return (ZIO_PIPELINE_STOP
);
1022 if (!IO_IS_ALLOCATING(zio
))
1023 return (ZIO_PIPELINE_CONTINUE
);
1025 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1027 if (zio
->io_bp_override
) {
1028 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1029 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1031 *bp
= *zio
->io_bp_override
;
1032 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1034 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1035 return (ZIO_PIPELINE_CONTINUE
);
1037 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_dedup
||
1038 zp
->zp_dedup_verify
);
1040 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
1041 BP_SET_DEDUP(bp
, 1);
1042 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1043 return (ZIO_PIPELINE_CONTINUE
);
1045 zio
->io_bp_override
= NULL
;
1049 if (bp
->blk_birth
== zio
->io_txg
) {
1051 * We're rewriting an existing block, which means we're
1052 * working on behalf of spa_sync(). For spa_sync() to
1053 * converge, it must eventually be the case that we don't
1054 * have to allocate new blocks. But compression changes
1055 * the blocksize, which forces a reallocate, and makes
1056 * convergence take longer. Therefore, after the first
1057 * few passes, stop compressing to ensure convergence.
1059 pass
= spa_sync_pass(spa
);
1061 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1062 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1063 ASSERT(!BP_GET_DEDUP(bp
));
1065 if (pass
> SYNC_PASS_DONT_COMPRESS
)
1066 compress
= ZIO_COMPRESS_OFF
;
1068 /* Make sure someone doesn't change their mind on overwrites */
1069 ASSERT(MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1070 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1073 if (compress
!= ZIO_COMPRESS_OFF
) {
1074 void *cbuf
= zio_buf_alloc(lsize
);
1075 psize
= zio_compress_data(compress
, zio
->io_data
, cbuf
, lsize
);
1076 if (psize
== 0 || psize
== lsize
) {
1077 compress
= ZIO_COMPRESS_OFF
;
1078 zio_buf_free(cbuf
, lsize
);
1080 ASSERT(psize
< lsize
);
1081 zio_push_transform(zio
, cbuf
, psize
, lsize
, NULL
);
1086 * The final pass of spa_sync() must be all rewrites, but the first
1087 * few passes offer a trade-off: allocating blocks defers convergence,
1088 * but newly allocated blocks are sequential, so they can be written
1089 * to disk faster. Therefore, we allow the first few passes of
1090 * spa_sync() to allocate new blocks, but force rewrites after that.
1091 * There should only be a handful of blocks after pass 1 in any case.
1093 if (bp
->blk_birth
== zio
->io_txg
&& BP_GET_PSIZE(bp
) == psize
&&
1094 pass
> SYNC_PASS_REWRITE
) {
1095 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1097 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1098 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1101 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1105 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1107 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1108 BP_SET_LSIZE(bp
, lsize
);
1109 BP_SET_PSIZE(bp
, psize
);
1110 BP_SET_COMPRESS(bp
, compress
);
1111 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1112 BP_SET_TYPE(bp
, zp
->zp_type
);
1113 BP_SET_LEVEL(bp
, zp
->zp_level
);
1114 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1115 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1117 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1118 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1119 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1123 return (ZIO_PIPELINE_CONTINUE
);
1127 zio_free_bp_init(zio_t
*zio
)
1129 blkptr_t
*bp
= zio
->io_bp
;
1131 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1132 if (BP_GET_DEDUP(bp
))
1133 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1136 return (ZIO_PIPELINE_CONTINUE
);
1140 * ==========================================================================
1141 * Execute the I/O pipeline
1142 * ==========================================================================
1146 zio_taskq_dispatch(zio_t
*zio
, enum zio_taskq_type q
, boolean_t cutinline
)
1148 spa_t
*spa
= zio
->io_spa
;
1149 zio_type_t t
= zio
->io_type
;
1150 int flags
= (cutinline
? TQ_FRONT
: 0);
1153 * If we're a config writer or a probe, the normal issue and
1154 * interrupt threads may all be blocked waiting for the config lock.
1155 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1157 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1161 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1163 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1167 * If this is a high priority I/O, then use the high priority taskq.
1169 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1170 spa
->spa_zio_taskq
[t
][q
+ 1] != NULL
)
1173 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1176 * NB: We are assuming that the zio can only be dispatched
1177 * to a single taskq at a time. It would be a grievous error
1178 * to dispatch the zio to another taskq at the same time.
1180 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1181 taskq_dispatch_ent(spa
->spa_zio_taskq
[t
][q
],
1182 (task_func_t
*)zio_execute
, zio
, flags
, &zio
->io_tqent
);
1186 zio_taskq_member(zio_t
*zio
, enum zio_taskq_type q
)
1188 kthread_t
*executor
= zio
->io_executor
;
1189 spa_t
*spa
= zio
->io_spa
;
1192 for (t
= 0; t
< ZIO_TYPES
; t
++)
1193 if (taskq_member(spa
->spa_zio_taskq
[t
][q
], executor
))
1200 zio_issue_async(zio_t
*zio
)
1202 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1204 return (ZIO_PIPELINE_STOP
);
1208 zio_interrupt(zio_t
*zio
)
1210 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1214 * Execute the I/O pipeline until one of the following occurs:
1215 * (1) the I/O completes; (2) the pipeline stalls waiting for
1216 * dependent child I/Os; (3) the I/O issues, so we're waiting
1217 * for an I/O completion interrupt; (4) the I/O is delegated by
1218 * vdev-level caching or aggregation; (5) the I/O is deferred
1219 * due to vdev-level queueing; (6) the I/O is handed off to
1220 * another thread. In all cases, the pipeline stops whenever
1221 * there's no CPU work; it never burns a thread in cv_wait().
1223 * There's no locking on io_stage because there's no legitimate way
1224 * for multiple threads to be attempting to process the same I/O.
1226 static zio_pipe_stage_t
*zio_pipeline
[];
1229 * zio_execute() is a wrapper around the static function
1230 * __zio_execute() so that we can force __zio_execute() to be
1231 * inlined. This reduces stack overhead which is important
1232 * because __zio_execute() is called recursively in several zio
1233 * code paths. zio_execute() itself cannot be inlined because
1234 * it is externally visible.
1237 zio_execute(zio_t
*zio
)
1242 __attribute__((always_inline
))
1244 __zio_execute(zio_t
*zio
)
1246 zio
->io_executor
= curthread
;
1248 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1249 enum zio_stage pipeline
= zio
->io_pipeline
;
1250 enum zio_stage stage
= zio
->io_stage
;
1255 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1256 ASSERT(ISP2(stage
));
1257 ASSERT(zio
->io_stall
== NULL
);
1261 } while ((stage
& pipeline
) == 0);
1263 ASSERT(stage
<= ZIO_STAGE_DONE
);
1265 dsl
= spa_get_dsl(zio
->io_spa
);
1266 cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1267 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1270 * If we are in interrupt context and this pipeline stage
1271 * will grab a config lock that is held across I/O,
1272 * or may wait for an I/O that needs an interrupt thread
1273 * to complete, issue async to avoid deadlock.
1275 * If we are in the txg_sync_thread or being called
1276 * during pool init issue async to minimize stack depth.
1277 * Both of these call paths may be recursively called.
1279 * For VDEV_IO_START, we cut in line so that the io will
1280 * be sent to disk promptly.
1282 if (((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1283 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) ||
1284 (dsl
!= NULL
&& dsl_pool_sync_context(dsl
))) {
1285 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1289 zio
->io_stage
= stage
;
1290 rv
= zio_pipeline
[highbit(stage
) - 1](zio
);
1292 if (rv
== ZIO_PIPELINE_STOP
)
1295 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1301 * ==========================================================================
1302 * Initiate I/O, either sync or async
1303 * ==========================================================================
1306 zio_wait(zio_t
*zio
)
1310 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1311 ASSERT(zio
->io_executor
== NULL
);
1313 zio
->io_waiter
= curthread
;
1317 mutex_enter(&zio
->io_lock
);
1318 while (zio
->io_executor
!= NULL
)
1319 cv_wait_io(&zio
->io_cv
, &zio
->io_lock
);
1320 mutex_exit(&zio
->io_lock
);
1322 error
= zio
->io_error
;
1329 zio_nowait(zio_t
*zio
)
1331 ASSERT(zio
->io_executor
== NULL
);
1333 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1334 zio_unique_parent(zio
) == NULL
) {
1336 * This is a logical async I/O with no parent to wait for it.
1337 * We add it to the spa_async_root_zio "Godfather" I/O which
1338 * will ensure they complete prior to unloading the pool.
1340 spa_t
*spa
= zio
->io_spa
;
1342 zio_add_child(spa
->spa_async_zio_root
, zio
);
1349 * ==========================================================================
1350 * Reexecute or suspend/resume failed I/O
1351 * ==========================================================================
1355 zio_reexecute(zio_t
*pio
)
1357 zio_t
*cio
, *cio_next
;
1360 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1361 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1362 ASSERT(pio
->io_gang_leader
== NULL
);
1363 ASSERT(pio
->io_gang_tree
== NULL
);
1365 pio
->io_flags
= pio
->io_orig_flags
;
1366 pio
->io_stage
= pio
->io_orig_stage
;
1367 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1368 pio
->io_reexecute
= 0;
1370 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1371 pio
->io_state
[w
] = 0;
1372 for (c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1373 pio
->io_child_error
[c
] = 0;
1375 if (IO_IS_ALLOCATING(pio
))
1376 BP_ZERO(pio
->io_bp
);
1379 * As we reexecute pio's children, new children could be created.
1380 * New children go to the head of pio's io_child_list, however,
1381 * so we will (correctly) not reexecute them. The key is that
1382 * the remainder of pio's io_child_list, from 'cio_next' onward,
1383 * cannot be affected by any side effects of reexecuting 'cio'.
1385 for (cio
= zio_walk_children(pio
); cio
!= NULL
; cio
= cio_next
) {
1386 cio_next
= zio_walk_children(pio
);
1387 mutex_enter(&pio
->io_lock
);
1388 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1389 pio
->io_children
[cio
->io_child_type
][w
]++;
1390 mutex_exit(&pio
->io_lock
);
1395 * Now that all children have been reexecuted, execute the parent.
1396 * We don't reexecute "The Godfather" I/O here as it's the
1397 * responsibility of the caller to wait on him.
1399 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
))
1404 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1406 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1407 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1408 "failure and the failure mode property for this pool "
1409 "is set to panic.", spa_name(spa
));
1411 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1413 mutex_enter(&spa
->spa_suspend_lock
);
1415 if (spa
->spa_suspend_zio_root
== NULL
)
1416 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1417 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1418 ZIO_FLAG_GODFATHER
);
1420 spa
->spa_suspended
= B_TRUE
;
1423 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1424 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1425 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1426 ASSERT(zio_unique_parent(zio
) == NULL
);
1427 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1428 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1431 mutex_exit(&spa
->spa_suspend_lock
);
1435 zio_resume(spa_t
*spa
)
1440 * Reexecute all previously suspended i/o.
1442 mutex_enter(&spa
->spa_suspend_lock
);
1443 spa
->spa_suspended
= B_FALSE
;
1444 cv_broadcast(&spa
->spa_suspend_cv
);
1445 pio
= spa
->spa_suspend_zio_root
;
1446 spa
->spa_suspend_zio_root
= NULL
;
1447 mutex_exit(&spa
->spa_suspend_lock
);
1453 return (zio_wait(pio
));
1457 zio_resume_wait(spa_t
*spa
)
1459 mutex_enter(&spa
->spa_suspend_lock
);
1460 while (spa_suspended(spa
))
1461 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1462 mutex_exit(&spa
->spa_suspend_lock
);
1466 * ==========================================================================
1469 * A gang block is a collection of small blocks that looks to the DMU
1470 * like one large block. When zio_dva_allocate() cannot find a block
1471 * of the requested size, due to either severe fragmentation or the pool
1472 * being nearly full, it calls zio_write_gang_block() to construct the
1473 * block from smaller fragments.
1475 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1476 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1477 * an indirect block: it's an array of block pointers. It consumes
1478 * only one sector and hence is allocatable regardless of fragmentation.
1479 * The gang header's bps point to its gang members, which hold the data.
1481 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1482 * as the verifier to ensure uniqueness of the SHA256 checksum.
1483 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1484 * not the gang header. This ensures that data block signatures (needed for
1485 * deduplication) are independent of how the block is physically stored.
1487 * Gang blocks can be nested: a gang member may itself be a gang block.
1488 * Thus every gang block is a tree in which root and all interior nodes are
1489 * gang headers, and the leaves are normal blocks that contain user data.
1490 * The root of the gang tree is called the gang leader.
1492 * To perform any operation (read, rewrite, free, claim) on a gang block,
1493 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1494 * in the io_gang_tree field of the original logical i/o by recursively
1495 * reading the gang leader and all gang headers below it. This yields
1496 * an in-core tree containing the contents of every gang header and the
1497 * bps for every constituent of the gang block.
1499 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1500 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1501 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1502 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1503 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1504 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1505 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1506 * of the gang header plus zio_checksum_compute() of the data to update the
1507 * gang header's blk_cksum as described above.
1509 * The two-phase assemble/issue model solves the problem of partial failure --
1510 * what if you'd freed part of a gang block but then couldn't read the
1511 * gang header for another part? Assembling the entire gang tree first
1512 * ensures that all the necessary gang header I/O has succeeded before
1513 * starting the actual work of free, claim, or write. Once the gang tree
1514 * is assembled, free and claim are in-memory operations that cannot fail.
1516 * In the event that a gang write fails, zio_dva_unallocate() walks the
1517 * gang tree to immediately free (i.e. insert back into the space map)
1518 * everything we've allocated. This ensures that we don't get ENOSPC
1519 * errors during repeated suspend/resume cycles due to a flaky device.
1521 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1522 * the gang tree, we won't modify the block, so we can safely defer the free
1523 * (knowing that the block is still intact). If we *can* assemble the gang
1524 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1525 * each constituent bp and we can allocate a new block on the next sync pass.
1527 * In all cases, the gang tree allows complete recovery from partial failure.
1528 * ==========================================================================
1532 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1537 return (zio_read(pio
, pio
->io_spa
, bp
, data
, BP_GET_PSIZE(bp
),
1538 NULL
, NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1539 &pio
->io_bookmark
));
1543 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1548 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1549 gn
->gn_gbh
, SPA_GANGBLOCKSIZE
, NULL
, NULL
, pio
->io_priority
,
1550 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1552 * As we rewrite each gang header, the pipeline will compute
1553 * a new gang block header checksum for it; but no one will
1554 * compute a new data checksum, so we do that here. The one
1555 * exception is the gang leader: the pipeline already computed
1556 * its data checksum because that stage precedes gang assembly.
1557 * (Presently, nothing actually uses interior data checksums;
1558 * this is just good hygiene.)
1560 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1561 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1562 data
, BP_GET_PSIZE(bp
));
1565 * If we are here to damage data for testing purposes,
1566 * leave the GBH alone so that we can detect the damage.
1568 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1569 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1571 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1572 data
, BP_GET_PSIZE(bp
), NULL
, NULL
, pio
->io_priority
,
1573 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1581 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1583 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1584 ZIO_GANG_CHILD_FLAGS(pio
)));
1589 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, void *data
)
1591 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1592 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1595 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
1604 static void zio_gang_tree_assemble_done(zio_t
*zio
);
1606 static zio_gang_node_t
*
1607 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
1609 zio_gang_node_t
*gn
;
1611 ASSERT(*gnpp
== NULL
);
1613 gn
= kmem_zalloc(sizeof (*gn
), KM_PUSHPAGE
);
1614 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
1621 zio_gang_node_free(zio_gang_node_t
**gnpp
)
1623 zio_gang_node_t
*gn
= *gnpp
;
1626 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1627 ASSERT(gn
->gn_child
[g
] == NULL
);
1629 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1630 kmem_free(gn
, sizeof (*gn
));
1635 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
1637 zio_gang_node_t
*gn
= *gnpp
;
1643 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1644 zio_gang_tree_free(&gn
->gn_child
[g
]);
1646 zio_gang_node_free(gnpp
);
1650 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
1652 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
1654 ASSERT(gio
->io_gang_leader
== gio
);
1655 ASSERT(BP_IS_GANG(bp
));
1657 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gn
->gn_gbh
,
1658 SPA_GANGBLOCKSIZE
, zio_gang_tree_assemble_done
, gn
,
1659 gio
->io_priority
, ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
1663 zio_gang_tree_assemble_done(zio_t
*zio
)
1665 zio_t
*gio
= zio
->io_gang_leader
;
1666 zio_gang_node_t
*gn
= zio
->io_private
;
1667 blkptr_t
*bp
= zio
->io_bp
;
1670 ASSERT(gio
== zio_unique_parent(zio
));
1671 ASSERT(zio
->io_child_count
== 0);
1676 if (BP_SHOULD_BYTESWAP(bp
))
1677 byteswap_uint64_array(zio
->io_data
, zio
->io_size
);
1679 ASSERT(zio
->io_data
== gn
->gn_gbh
);
1680 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
1681 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
1683 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
1684 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
1685 if (!BP_IS_GANG(gbp
))
1687 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
1692 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, void *data
)
1694 zio_t
*gio
= pio
->io_gang_leader
;
1698 ASSERT(BP_IS_GANG(bp
) == !!gn
);
1699 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
1700 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
1703 * If you're a gang header, your data is in gn->gn_gbh.
1704 * If you're a gang member, your data is in 'data' and gn == NULL.
1706 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
);
1709 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
1711 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
1712 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
1713 if (BP_IS_HOLE(gbp
))
1715 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
);
1716 data
= (char *)data
+ BP_GET_PSIZE(gbp
);
1720 if (gn
== gio
->io_gang_tree
)
1721 ASSERT3P((char *)gio
->io_data
+ gio
->io_size
, ==, data
);
1728 zio_gang_assemble(zio_t
*zio
)
1730 blkptr_t
*bp
= zio
->io_bp
;
1732 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
1733 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
1735 zio
->io_gang_leader
= zio
;
1737 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
1739 return (ZIO_PIPELINE_CONTINUE
);
1743 zio_gang_issue(zio_t
*zio
)
1745 blkptr_t
*bp
= zio
->io_bp
;
1747 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
1748 return (ZIO_PIPELINE_STOP
);
1750 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
1751 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
1753 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
1754 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_data
);
1756 zio_gang_tree_free(&zio
->io_gang_tree
);
1758 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1760 return (ZIO_PIPELINE_CONTINUE
);
1764 zio_write_gang_member_ready(zio_t
*zio
)
1766 zio_t
*pio
= zio_unique_parent(zio
);
1767 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
;)
1768 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
1769 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
1773 if (BP_IS_HOLE(zio
->io_bp
))
1776 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
1778 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
1779 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
1780 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
1781 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
1782 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
1784 mutex_enter(&pio
->io_lock
);
1785 for (d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
1786 ASSERT(DVA_GET_GANG(&pdva
[d
]));
1787 asize
= DVA_GET_ASIZE(&pdva
[d
]);
1788 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
1789 DVA_SET_ASIZE(&pdva
[d
], asize
);
1791 mutex_exit(&pio
->io_lock
);
1795 zio_write_gang_block(zio_t
*pio
)
1797 spa_t
*spa
= pio
->io_spa
;
1798 blkptr_t
*bp
= pio
->io_bp
;
1799 zio_t
*gio
= pio
->io_gang_leader
;
1801 zio_gang_node_t
*gn
, **gnpp
;
1802 zio_gbh_phys_t
*gbh
;
1803 uint64_t txg
= pio
->io_txg
;
1804 uint64_t resid
= pio
->io_size
;
1806 int copies
= gio
->io_prop
.zp_copies
;
1807 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
1811 error
= metaslab_alloc(spa
, spa_normal_class(spa
), SPA_GANGBLOCKSIZE
,
1812 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
,
1813 METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
);
1815 pio
->io_error
= error
;
1816 return (ZIO_PIPELINE_CONTINUE
);
1820 gnpp
= &gio
->io_gang_tree
;
1822 gnpp
= pio
->io_private
;
1823 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
1826 gn
= zio_gang_node_alloc(gnpp
);
1828 bzero(gbh
, SPA_GANGBLOCKSIZE
);
1831 * Create the gang header.
1833 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh
, SPA_GANGBLOCKSIZE
, NULL
, NULL
,
1834 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1837 * Create and nowait the gang children.
1839 for (g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
1840 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
1842 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
1844 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
1845 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
1846 zp
.zp_type
= DMU_OT_NONE
;
1848 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
1850 zp
.zp_dedup_verify
= 0;
1852 zio_nowait(zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
1853 (char *)pio
->io_data
+ (pio
->io_size
- resid
), lsize
, &zp
,
1854 zio_write_gang_member_ready
, NULL
, &gn
->gn_child
[g
],
1855 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1856 &pio
->io_bookmark
));
1860 * Set pio's pipeline to just wait for zio to finish.
1862 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1865 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1867 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
1871 return (ZIO_PIPELINE_CONTINUE
);
1875 * ==========================================================================
1877 * ==========================================================================
1880 zio_ddt_child_read_done(zio_t
*zio
)
1882 blkptr_t
*bp
= zio
->io_bp
;
1883 ddt_entry_t
*dde
= zio
->io_private
;
1885 zio_t
*pio
= zio_unique_parent(zio
);
1887 mutex_enter(&pio
->io_lock
);
1888 ddp
= ddt_phys_select(dde
, bp
);
1889 if (zio
->io_error
== 0)
1890 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
1891 if (zio
->io_error
== 0 && dde
->dde_repair_data
== NULL
)
1892 dde
->dde_repair_data
= zio
->io_data
;
1894 zio_buf_free(zio
->io_data
, zio
->io_size
);
1895 mutex_exit(&pio
->io_lock
);
1899 zio_ddt_read_start(zio_t
*zio
)
1901 blkptr_t
*bp
= zio
->io_bp
;
1904 ASSERT(BP_GET_DEDUP(bp
));
1905 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
1906 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1908 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
1909 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
1910 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
1911 ddt_phys_t
*ddp
= dde
->dde_phys
;
1912 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
1915 ASSERT(zio
->io_vsd
== NULL
);
1918 if (ddp_self
== NULL
)
1919 return (ZIO_PIPELINE_CONTINUE
);
1921 for (p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
1922 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
1924 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
1926 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
1927 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
1928 zio_ddt_child_read_done
, dde
, zio
->io_priority
,
1929 ZIO_DDT_CHILD_FLAGS(zio
) | ZIO_FLAG_DONT_PROPAGATE
,
1930 &zio
->io_bookmark
));
1932 return (ZIO_PIPELINE_CONTINUE
);
1935 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
1936 zio
->io_data
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
1937 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
1939 return (ZIO_PIPELINE_CONTINUE
);
1943 zio_ddt_read_done(zio_t
*zio
)
1945 blkptr_t
*bp
= zio
->io_bp
;
1947 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
1948 return (ZIO_PIPELINE_STOP
);
1950 ASSERT(BP_GET_DEDUP(bp
));
1951 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
1952 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1954 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
1955 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
1956 ddt_entry_t
*dde
= zio
->io_vsd
;
1958 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
1959 return (ZIO_PIPELINE_CONTINUE
);
1962 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
1963 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1964 return (ZIO_PIPELINE_STOP
);
1966 if (dde
->dde_repair_data
!= NULL
) {
1967 bcopy(dde
->dde_repair_data
, zio
->io_data
, zio
->io_size
);
1968 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
1970 ddt_repair_done(ddt
, dde
);
1974 ASSERT(zio
->io_vsd
== NULL
);
1976 return (ZIO_PIPELINE_CONTINUE
);
1980 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
1982 spa_t
*spa
= zio
->io_spa
;
1986 * Note: we compare the original data, not the transformed data,
1987 * because when zio->io_bp is an override bp, we will not have
1988 * pushed the I/O transforms. That's an important optimization
1989 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1991 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
1992 zio_t
*lio
= dde
->dde_lead_zio
[p
];
1995 return (lio
->io_orig_size
!= zio
->io_orig_size
||
1996 bcmp(zio
->io_orig_data
, lio
->io_orig_data
,
1997 zio
->io_orig_size
) != 0);
2001 for (p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2002 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2004 if (ddp
->ddp_phys_birth
!= 0) {
2005 arc_buf_t
*abuf
= NULL
;
2006 uint32_t aflags
= ARC_WAIT
;
2007 blkptr_t blk
= *zio
->io_bp
;
2010 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2014 error
= arc_read_nolock(NULL
, spa
, &blk
,
2015 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2016 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2017 &aflags
, &zio
->io_bookmark
);
2020 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
2021 bcmp(abuf
->b_data
, zio
->io_orig_data
,
2022 zio
->io_orig_size
) != 0)
2024 VERIFY(arc_buf_remove_ref(abuf
, &abuf
) == 1);
2028 return (error
!= 0);
2036 zio_ddt_child_write_ready(zio_t
*zio
)
2038 int p
= zio
->io_prop
.zp_copies
;
2039 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2040 ddt_entry_t
*dde
= zio
->io_private
;
2041 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2049 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2051 ddt_phys_fill(ddp
, zio
->io_bp
);
2053 while ((pio
= zio_walk_parents(zio
)) != NULL
)
2054 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2060 zio_ddt_child_write_done(zio_t
*zio
)
2062 int p
= zio
->io_prop
.zp_copies
;
2063 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2064 ddt_entry_t
*dde
= zio
->io_private
;
2065 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2069 ASSERT(ddp
->ddp_refcnt
== 0);
2070 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2071 dde
->dde_lead_zio
[p
] = NULL
;
2073 if (zio
->io_error
== 0) {
2074 while (zio_walk_parents(zio
) != NULL
)
2075 ddt_phys_addref(ddp
);
2077 ddt_phys_clear(ddp
);
2084 zio_ddt_ditto_write_done(zio_t
*zio
)
2086 int p
= DDT_PHYS_DITTO
;
2087 blkptr_t
*bp
= zio
->io_bp
;
2088 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2089 ddt_entry_t
*dde
= zio
->io_private
;
2090 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2091 ddt_key_t
*ddk
= &dde
->dde_key
;
2092 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
2096 ASSERT(ddp
->ddp_refcnt
== 0);
2097 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2098 dde
->dde_lead_zio
[p
] = NULL
;
2100 if (zio
->io_error
== 0) {
2101 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2102 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2103 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2104 if (ddp
->ddp_phys_birth
!= 0)
2105 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2106 ddt_phys_fill(ddp
, bp
);
2113 zio_ddt_write(zio_t
*zio
)
2115 spa_t
*spa
= zio
->io_spa
;
2116 blkptr_t
*bp
= zio
->io_bp
;
2117 uint64_t txg
= zio
->io_txg
;
2118 zio_prop_t
*zp
= &zio
->io_prop
;
2119 int p
= zp
->zp_copies
;
2123 ddt_t
*ddt
= ddt_select(spa
, bp
);
2127 ASSERT(BP_GET_DEDUP(bp
));
2128 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2129 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2132 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2133 ddp
= &dde
->dde_phys
[p
];
2135 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2137 * If we're using a weak checksum, upgrade to a strong checksum
2138 * and try again. If we're already using a strong checksum,
2139 * we can't resolve it, so just convert to an ordinary write.
2140 * (And automatically e-mail a paper to Nature?)
2142 if (!zio_checksum_table
[zp
->zp_checksum
].ci_dedup
) {
2143 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2144 zio_pop_transforms(zio
);
2145 zio
->io_stage
= ZIO_STAGE_OPEN
;
2150 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2152 return (ZIO_PIPELINE_CONTINUE
);
2155 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2156 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2158 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2159 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2160 zio_prop_t czp
= *zp
;
2162 czp
.zp_copies
= ditto_copies
;
2165 * If we arrived here with an override bp, we won't have run
2166 * the transform stack, so we won't have the data we need to
2167 * generate a child i/o. So, toss the override bp and restart.
2168 * This is safe, because using the override bp is just an
2169 * optimization; and it's rare, so the cost doesn't matter.
2171 if (zio
->io_bp_override
) {
2172 zio_pop_transforms(zio
);
2173 zio
->io_stage
= ZIO_STAGE_OPEN
;
2174 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2175 zio
->io_bp_override
= NULL
;
2178 return (ZIO_PIPELINE_CONTINUE
);
2181 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_data
,
2182 zio
->io_orig_size
, &czp
, NULL
,
2183 zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2184 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2186 zio_push_transform(dio
, zio
->io_data
, zio
->io_size
, 0, NULL
);
2187 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2190 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2191 if (ddp
->ddp_phys_birth
!= 0)
2192 ddt_bp_fill(ddp
, bp
, txg
);
2193 if (dde
->dde_lead_zio
[p
] != NULL
)
2194 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2196 ddt_phys_addref(ddp
);
2197 } else if (zio
->io_bp_override
) {
2198 ASSERT(bp
->blk_birth
== txg
);
2199 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2200 ddt_phys_fill(ddp
, bp
);
2201 ddt_phys_addref(ddp
);
2203 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_data
,
2204 zio
->io_orig_size
, zp
, zio_ddt_child_write_ready
,
2205 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2206 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2208 zio_push_transform(cio
, zio
->io_data
, zio
->io_size
, 0, NULL
);
2209 dde
->dde_lead_zio
[p
] = cio
;
2219 return (ZIO_PIPELINE_CONTINUE
);
2222 ddt_entry_t
*freedde
; /* for debugging */
2225 zio_ddt_free(zio_t
*zio
)
2227 spa_t
*spa
= zio
->io_spa
;
2228 blkptr_t
*bp
= zio
->io_bp
;
2229 ddt_t
*ddt
= ddt_select(spa
, bp
);
2233 ASSERT(BP_GET_DEDUP(bp
));
2234 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2237 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2238 ddp
= ddt_phys_select(dde
, bp
);
2239 ddt_phys_decref(ddp
);
2242 return (ZIO_PIPELINE_CONTINUE
);
2246 * ==========================================================================
2247 * Allocate and free blocks
2248 * ==========================================================================
2251 zio_dva_allocate(zio_t
*zio
)
2253 spa_t
*spa
= zio
->io_spa
;
2254 metaslab_class_t
*mc
= spa_normal_class(spa
);
2255 blkptr_t
*bp
= zio
->io_bp
;
2259 if (zio
->io_gang_leader
== NULL
) {
2260 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2261 zio
->io_gang_leader
= zio
;
2264 ASSERT(BP_IS_HOLE(bp
));
2265 ASSERT3U(BP_GET_NDVAS(bp
), ==, 0);
2266 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2267 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2268 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2271 * The dump device does not support gang blocks so allocation on
2272 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2273 * the "fast" gang feature.
2275 flags
|= (zio
->io_flags
& ZIO_FLAG_NODATA
) ? METASLAB_GANG_AVOID
: 0;
2276 flags
|= (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
) ?
2277 METASLAB_GANG_CHILD
: 0;
2278 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
2279 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2280 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
);
2283 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
2284 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
2286 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2287 return (zio_write_gang_block(zio
));
2288 zio
->io_error
= error
;
2291 return (ZIO_PIPELINE_CONTINUE
);
2295 zio_dva_free(zio_t
*zio
)
2297 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2299 return (ZIO_PIPELINE_CONTINUE
);
2303 zio_dva_claim(zio_t
*zio
)
2307 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2309 zio
->io_error
= error
;
2311 return (ZIO_PIPELINE_CONTINUE
);
2315 * Undo an allocation. This is used by zio_done() when an I/O fails
2316 * and we want to give back the block we just allocated.
2317 * This handles both normal blocks and gang blocks.
2320 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2324 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2325 ASSERT(zio
->io_bp_override
== NULL
);
2327 if (!BP_IS_HOLE(bp
))
2328 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2331 for (g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2332 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2333 &gn
->gn_gbh
->zg_blkptr
[g
]);
2339 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2342 zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
, uint64_t size
,
2347 ASSERT(txg
> spa_syncing_txg(spa
));
2350 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2351 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2352 * when allocating them.
2355 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
,
2356 new_bp
, 1, txg
, NULL
,
2357 METASLAB_FASTWRITE
| METASLAB_GANG_AVOID
);
2361 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
2362 new_bp
, 1, txg
, NULL
,
2363 METASLAB_FASTWRITE
| METASLAB_GANG_AVOID
);
2367 BP_SET_LSIZE(new_bp
, size
);
2368 BP_SET_PSIZE(new_bp
, size
);
2369 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
2370 BP_SET_CHECKSUM(new_bp
,
2371 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
2372 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
2373 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
2374 BP_SET_LEVEL(new_bp
, 0);
2375 BP_SET_DEDUP(new_bp
, 0);
2376 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
2383 * Free an intent log block.
2386 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
2388 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
2389 ASSERT(!BP_IS_GANG(bp
));
2391 zio_free(spa
, txg
, bp
);
2395 * ==========================================================================
2396 * Read and write to physical devices
2397 * ==========================================================================
2400 zio_vdev_io_start(zio_t
*zio
)
2402 vdev_t
*vd
= zio
->io_vd
;
2404 spa_t
*spa
= zio
->io_spa
;
2406 ASSERT(zio
->io_error
== 0);
2407 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
2410 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
2411 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
2414 * The mirror_ops handle multiple DVAs in a single BP.
2416 return (vdev_mirror_ops
.vdev_op_io_start(zio
));
2420 * We keep track of time-sensitive I/Os so that the scan thread
2421 * can quickly react to certain workloads. In particular, we care
2422 * about non-scrubbing, top-level reads and writes with the following
2424 * - synchronous writes of user data to non-slog devices
2425 * - any reads of user data
2426 * When these conditions are met, adjust the timestamp of spa_last_io
2427 * which allows the scan thread to adjust its workload accordingly.
2429 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
2430 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
2431 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
2432 zio
->io_txg
!= spa_syncing_txg(spa
)) {
2433 uint64_t old
= spa
->spa_last_io
;
2434 uint64_t new = ddi_get_lbolt64();
2436 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
2439 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
2441 if (P2PHASE(zio
->io_size
, align
) != 0) {
2442 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
2443 char *abuf
= zio_buf_alloc(asize
);
2444 ASSERT(vd
== vd
->vdev_top
);
2445 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
2446 bcopy(zio
->io_data
, abuf
, zio
->io_size
);
2447 bzero(abuf
+ zio
->io_size
, asize
- zio
->io_size
);
2449 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
2452 ASSERT(P2PHASE(zio
->io_offset
, align
) == 0);
2453 ASSERT(P2PHASE(zio
->io_size
, align
) == 0);
2454 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
2457 * If this is a repair I/O, and there's no self-healing involved --
2458 * that is, we're just resilvering what we expect to resilver --
2459 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2460 * This prevents spurious resilvering with nested replication.
2461 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2462 * A is out of date, we'll read from C+D, then use the data to
2463 * resilver A+B -- but we don't actually want to resilver B, just A.
2464 * The top-level mirror has no way to know this, so instead we just
2465 * discard unnecessary repairs as we work our way down the vdev tree.
2466 * The same logic applies to any form of nested replication:
2467 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2469 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
2470 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
2471 zio
->io_txg
!= 0 && /* not a delegated i/o */
2472 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
2473 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2474 zio_vdev_io_bypass(zio
);
2475 return (ZIO_PIPELINE_CONTINUE
);
2478 if (vd
->vdev_ops
->vdev_op_leaf
&&
2479 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
2481 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
) == 0)
2482 return (ZIO_PIPELINE_CONTINUE
);
2484 if ((zio
= vdev_queue_io(zio
)) == NULL
)
2485 return (ZIO_PIPELINE_STOP
);
2487 if (!vdev_accessible(vd
, zio
)) {
2488 zio
->io_error
= ENXIO
;
2490 return (ZIO_PIPELINE_STOP
);
2494 return (vd
->vdev_ops
->vdev_op_io_start(zio
));
2498 zio_vdev_io_done(zio_t
*zio
)
2500 vdev_t
*vd
= zio
->io_vd
;
2501 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
2502 boolean_t unexpected_error
= B_FALSE
;
2504 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
2505 return (ZIO_PIPELINE_STOP
);
2507 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
2509 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
2511 vdev_queue_io_done(zio
);
2513 if (zio
->io_type
== ZIO_TYPE_WRITE
)
2514 vdev_cache_write(zio
);
2516 if (zio_injection_enabled
&& zio
->io_error
== 0)
2517 zio
->io_error
= zio_handle_device_injection(vd
,
2520 if (zio_injection_enabled
&& zio
->io_error
== 0)
2521 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
2523 if (zio
->io_error
) {
2524 if (!vdev_accessible(vd
, zio
)) {
2525 zio
->io_error
= ENXIO
;
2527 unexpected_error
= B_TRUE
;
2532 ops
->vdev_op_io_done(zio
);
2534 if (unexpected_error
)
2535 VERIFY(vdev_probe(vd
, zio
) == NULL
);
2537 return (ZIO_PIPELINE_CONTINUE
);
2541 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2542 * disk, and use that to finish the checksum ereport later.
2545 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
2546 const void *good_buf
)
2548 /* no processing needed */
2549 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
2554 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
2556 void *buf
= zio_buf_alloc(zio
->io_size
);
2558 bcopy(zio
->io_data
, buf
, zio
->io_size
);
2560 zcr
->zcr_cbinfo
= zio
->io_size
;
2561 zcr
->zcr_cbdata
= buf
;
2562 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
2563 zcr
->zcr_free
= zio_buf_free
;
2567 zio_vdev_io_assess(zio_t
*zio
)
2569 vdev_t
*vd
= zio
->io_vd
;
2571 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
2572 return (ZIO_PIPELINE_STOP
);
2574 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
2575 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
2577 if (zio
->io_vsd
!= NULL
) {
2578 zio
->io_vsd_ops
->vsd_free(zio
);
2582 if (zio_injection_enabled
&& zio
->io_error
== 0)
2583 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
2586 * If the I/O failed, determine whether we should attempt to retry it.
2588 * On retry, we cut in line in the issue queue, since we don't want
2589 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2591 if (zio
->io_error
&& vd
== NULL
&&
2592 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
2593 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
2594 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
2596 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
2597 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
2598 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
2599 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
2600 zio_requeue_io_start_cut_in_line
);
2601 return (ZIO_PIPELINE_STOP
);
2605 * If we got an error on a leaf device, convert it to ENXIO
2606 * if the device is not accessible at all.
2608 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
2609 !vdev_accessible(vd
, zio
))
2610 zio
->io_error
= ENXIO
;
2613 * If we can't write to an interior vdev (mirror or RAID-Z),
2614 * set vdev_cant_write so that we stop trying to allocate from it.
2616 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
2617 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
)
2618 vd
->vdev_cant_write
= B_TRUE
;
2621 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2623 return (ZIO_PIPELINE_CONTINUE
);
2627 zio_vdev_io_reissue(zio_t
*zio
)
2629 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
2630 ASSERT(zio
->io_error
== 0);
2632 zio
->io_stage
>>= 1;
2636 zio_vdev_io_redone(zio_t
*zio
)
2638 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
2640 zio
->io_stage
>>= 1;
2644 zio_vdev_io_bypass(zio_t
*zio
)
2646 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
2647 ASSERT(zio
->io_error
== 0);
2649 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
2650 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
2654 * ==========================================================================
2655 * Generate and verify checksums
2656 * ==========================================================================
2659 zio_checksum_generate(zio_t
*zio
)
2661 blkptr_t
*bp
= zio
->io_bp
;
2662 enum zio_checksum checksum
;
2666 * This is zio_write_phys().
2667 * We're either generating a label checksum, or none at all.
2669 checksum
= zio
->io_prop
.zp_checksum
;
2671 if (checksum
== ZIO_CHECKSUM_OFF
)
2672 return (ZIO_PIPELINE_CONTINUE
);
2674 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
2676 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
2677 ASSERT(!IO_IS_ALLOCATING(zio
));
2678 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
2680 checksum
= BP_GET_CHECKSUM(bp
);
2684 zio_checksum_compute(zio
, checksum
, zio
->io_data
, zio
->io_size
);
2686 return (ZIO_PIPELINE_CONTINUE
);
2690 zio_checksum_verify(zio_t
*zio
)
2692 zio_bad_cksum_t info
;
2693 blkptr_t
*bp
= zio
->io_bp
;
2696 ASSERT(zio
->io_vd
!= NULL
);
2700 * This is zio_read_phys().
2701 * We're either verifying a label checksum, or nothing at all.
2703 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
2704 return (ZIO_PIPELINE_CONTINUE
);
2706 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
2709 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
2710 zio
->io_error
= error
;
2711 if (!(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
2712 zfs_ereport_start_checksum(zio
->io_spa
,
2713 zio
->io_vd
, zio
, zio
->io_offset
,
2714 zio
->io_size
, NULL
, &info
);
2718 return (ZIO_PIPELINE_CONTINUE
);
2722 * Called by RAID-Z to ensure we don't compute the checksum twice.
2725 zio_checksum_verified(zio_t
*zio
)
2727 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
2731 * ==========================================================================
2732 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2733 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2734 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2735 * indicate errors that are specific to one I/O, and most likely permanent.
2736 * Any other error is presumed to be worse because we weren't expecting it.
2737 * ==========================================================================
2740 zio_worst_error(int e1
, int e2
)
2742 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
2745 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
2746 if (e1
== zio_error_rank
[r1
])
2749 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
2750 if (e2
== zio_error_rank
[r2
])
2753 return (r1
> r2
? e1
: e2
);
2757 * ==========================================================================
2759 * ==========================================================================
2762 zio_ready(zio_t
*zio
)
2764 blkptr_t
*bp
= zio
->io_bp
;
2765 zio_t
*pio
, *pio_next
;
2767 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
2768 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
2769 return (ZIO_PIPELINE_STOP
);
2771 if (zio
->io_ready
) {
2772 ASSERT(IO_IS_ALLOCATING(zio
));
2773 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2774 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
2779 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
2780 zio
->io_bp_copy
= *bp
;
2783 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2785 mutex_enter(&zio
->io_lock
);
2786 zio
->io_state
[ZIO_WAIT_READY
] = 1;
2787 pio
= zio_walk_parents(zio
);
2788 mutex_exit(&zio
->io_lock
);
2791 * As we notify zio's parents, new parents could be added.
2792 * New parents go to the head of zio's io_parent_list, however,
2793 * so we will (correctly) not notify them. The remainder of zio's
2794 * io_parent_list, from 'pio_next' onward, cannot change because
2795 * all parents must wait for us to be done before they can be done.
2797 for (; pio
!= NULL
; pio
= pio_next
) {
2798 pio_next
= zio_walk_parents(zio
);
2799 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
2802 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2803 if (BP_IS_GANG(bp
)) {
2804 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
2806 ASSERT((uintptr_t)zio
->io_data
< SPA_MAXBLOCKSIZE
);
2807 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2811 if (zio_injection_enabled
&&
2812 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
2813 zio_handle_ignored_writes(zio
);
2815 return (ZIO_PIPELINE_CONTINUE
);
2819 zio_done(zio_t
*zio
)
2821 zio_t
*pio
, *pio_next
;
2825 * If our children haven't all completed,
2826 * wait for them and then repeat this pipeline stage.
2828 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
2829 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
2830 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
2831 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
2832 return (ZIO_PIPELINE_STOP
);
2834 for (c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2835 for (w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2836 ASSERT(zio
->io_children
[c
][w
] == 0);
2838 if (zio
->io_bp
!= NULL
) {
2839 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
2840 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
2841 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
2842 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
2843 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
2844 zio
->io_bp_override
== NULL
&&
2845 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
2846 ASSERT(!BP_SHOULD_BYTESWAP(zio
->io_bp
));
2847 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2848 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
2849 (BP_COUNT_GANG(zio
->io_bp
) == BP_GET_NDVAS(zio
->io_bp
)));
2854 * If there were child vdev/gang/ddt errors, they apply to us now.
2856 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
2857 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
2858 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
2861 * If the I/O on the transformed data was successful, generate any
2862 * checksum reports now while we still have the transformed data.
2864 if (zio
->io_error
== 0) {
2865 while (zio
->io_cksum_report
!= NULL
) {
2866 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
2867 uint64_t align
= zcr
->zcr_align
;
2868 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
2869 char *abuf
= zio
->io_data
;
2871 if (asize
!= zio
->io_size
) {
2872 abuf
= zio_buf_alloc(asize
);
2873 bcopy(zio
->io_data
, abuf
, zio
->io_size
);
2874 bzero(abuf
+ zio
->io_size
, asize
- zio
->io_size
);
2877 zio
->io_cksum_report
= zcr
->zcr_next
;
2878 zcr
->zcr_next
= NULL
;
2879 zcr
->zcr_finish(zcr
, abuf
);
2880 zfs_ereport_free_checksum(zcr
);
2882 if (asize
!= zio
->io_size
)
2883 zio_buf_free(abuf
, asize
);
2887 zio_pop_transforms(zio
); /* note: may set zio->io_error */
2889 vdev_stat_update(zio
, zio
->io_size
);
2892 * If this I/O is attached to a particular vdev is slow, exeeding
2893 * 30 seconds to complete, post an error described the I/O delay.
2894 * We ignore these errors if the device is currently unavailable.
2896 if (zio
->io_delay
>= zio_delay_max
) {
2897 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
2898 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
2899 zio
->io_vd
, zio
, 0, 0);
2902 if (zio
->io_error
) {
2904 * If this I/O is attached to a particular vdev,
2905 * generate an error message describing the I/O failure
2906 * at the block level. We ignore these errors if the
2907 * device is currently unavailable.
2909 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
2910 !vdev_is_dead(zio
->io_vd
))
2911 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
2912 zio
->io_vd
, zio
, 0, 0);
2914 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
2915 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
2916 zio
== zio
->io_logical
) {
2918 * For logical I/O requests, tell the SPA to log the
2919 * error and generate a logical data ereport.
2921 spa_log_error(zio
->io_spa
, zio
);
2922 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
, NULL
, zio
,
2927 if (zio
->io_error
&& zio
== zio
->io_logical
) {
2929 * Determine whether zio should be reexecuted. This will
2930 * propagate all the way to the root via zio_notify_parent().
2932 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
2933 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2935 if (IO_IS_ALLOCATING(zio
) &&
2936 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
2937 if (zio
->io_error
!= ENOSPC
)
2938 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
2940 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2943 if ((zio
->io_type
== ZIO_TYPE_READ
||
2944 zio
->io_type
== ZIO_TYPE_FREE
) &&
2945 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
2946 zio
->io_error
== ENXIO
&&
2947 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
2948 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
2949 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2951 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
2952 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
2955 * Here is a possibly good place to attempt to do
2956 * either combinatorial reconstruction or error correction
2957 * based on checksums. It also might be a good place
2958 * to send out preliminary ereports before we suspend
2964 * If there were logical child errors, they apply to us now.
2965 * We defer this until now to avoid conflating logical child
2966 * errors with errors that happened to the zio itself when
2967 * updating vdev stats and reporting FMA events above.
2969 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
2971 if ((zio
->io_error
|| zio
->io_reexecute
) &&
2972 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
2973 !(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
))
2974 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
2976 zio_gang_tree_free(&zio
->io_gang_tree
);
2979 * Godfather I/Os should never suspend.
2981 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
2982 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
2983 zio
->io_reexecute
= 0;
2985 if (zio
->io_reexecute
) {
2987 * This is a logical I/O that wants to reexecute.
2989 * Reexecute is top-down. When an i/o fails, if it's not
2990 * the root, it simply notifies its parent and sticks around.
2991 * The parent, seeing that it still has children in zio_done(),
2992 * does the same. This percolates all the way up to the root.
2993 * The root i/o will reexecute or suspend the entire tree.
2995 * This approach ensures that zio_reexecute() honors
2996 * all the original i/o dependency relationships, e.g.
2997 * parents not executing until children are ready.
2999 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3001 zio
->io_gang_leader
= NULL
;
3003 mutex_enter(&zio
->io_lock
);
3004 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3005 mutex_exit(&zio
->io_lock
);
3008 * "The Godfather" I/O monitors its children but is
3009 * not a true parent to them. It will track them through
3010 * the pipeline but severs its ties whenever they get into
3011 * trouble (e.g. suspended). This allows "The Godfather"
3012 * I/O to return status without blocking.
3014 for (pio
= zio_walk_parents(zio
); pio
!= NULL
; pio
= pio_next
) {
3015 zio_link_t
*zl
= zio
->io_walk_link
;
3016 pio_next
= zio_walk_parents(zio
);
3018 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3019 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
3020 zio_remove_child(pio
, zio
, zl
);
3021 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3025 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
3027 * We're not a root i/o, so there's nothing to do
3028 * but notify our parent. Don't propagate errors
3029 * upward since we haven't permanently failed yet.
3031 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
3032 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
3033 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3034 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
3036 * We'd fail again if we reexecuted now, so suspend
3037 * until conditions improve (e.g. device comes online).
3039 zio_suspend(zio
->io_spa
, zio
);
3042 * Reexecution is potentially a huge amount of work.
3043 * Hand it off to the otherwise-unused claim taskq.
3045 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
3046 (void) taskq_dispatch_ent(
3047 zio
->io_spa
->spa_zio_taskq
[ZIO_TYPE_CLAIM
][ZIO_TASKQ_ISSUE
],
3048 (task_func_t
*)zio_reexecute
, zio
, 0,
3051 return (ZIO_PIPELINE_STOP
);
3054 ASSERT(zio
->io_child_count
== 0);
3055 ASSERT(zio
->io_reexecute
== 0);
3056 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
3059 * Report any checksum errors, since the I/O is complete.
3061 while (zio
->io_cksum_report
!= NULL
) {
3062 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3063 zio
->io_cksum_report
= zcr
->zcr_next
;
3064 zcr
->zcr_next
= NULL
;
3065 zcr
->zcr_finish(zcr
, NULL
);
3066 zfs_ereport_free_checksum(zcr
);
3069 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
3070 !BP_IS_HOLE(zio
->io_bp
)) {
3071 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
3075 * It is the responsibility of the done callback to ensure that this
3076 * particular zio is no longer discoverable for adoption, and as
3077 * such, cannot acquire any new parents.
3082 mutex_enter(&zio
->io_lock
);
3083 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3084 mutex_exit(&zio
->io_lock
);
3086 for (pio
= zio_walk_parents(zio
); pio
!= NULL
; pio
= pio_next
) {
3087 zio_link_t
*zl
= zio
->io_walk_link
;
3088 pio_next
= zio_walk_parents(zio
);
3089 zio_remove_child(pio
, zio
, zl
);
3090 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3093 if (zio
->io_waiter
!= NULL
) {
3094 mutex_enter(&zio
->io_lock
);
3095 zio
->io_executor
= NULL
;
3096 cv_broadcast(&zio
->io_cv
);
3097 mutex_exit(&zio
->io_lock
);
3102 return (ZIO_PIPELINE_STOP
);
3106 * ==========================================================================
3107 * I/O pipeline definition
3108 * ==========================================================================
3110 static zio_pipe_stage_t
*zio_pipeline
[] = {
3116 zio_checksum_generate
,
3130 zio_checksum_verify
,
3134 #if defined(_KERNEL) && defined(HAVE_SPL)
3135 /* Fault injection */
3136 EXPORT_SYMBOL(zio_injection_enabled
);
3137 EXPORT_SYMBOL(zio_inject_fault
);
3138 EXPORT_SYMBOL(zio_inject_list_next
);
3139 EXPORT_SYMBOL(zio_clear_fault
);
3140 EXPORT_SYMBOL(zio_handle_fault_injection
);
3141 EXPORT_SYMBOL(zio_handle_device_injection
);
3142 EXPORT_SYMBOL(zio_handle_label_injection
);
3143 EXPORT_SYMBOL(zio_priority_table
);
3144 EXPORT_SYMBOL(zio_type_name
);
3146 module_param(zio_bulk_flags
, int, 0644);
3147 MODULE_PARM_DESC(zio_bulk_flags
, "Additional flags to pass to bulk buffers");
3149 module_param(zio_delay_max
, int, 0644);
3150 MODULE_PARM_DESC(zio_delay_max
, "Max zio millisec delay before posting event");
3152 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
3153 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");