4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
43 #include <sys/metaslab_impl.h>
47 * ==========================================================================
48 * I/O type descriptions
49 * ==========================================================================
51 const char *zio_type_name
[ZIO_TYPES
] = {
52 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
56 boolean_t zio_dva_throttle_enabled
= B_TRUE
;
59 * ==========================================================================
61 * ==========================================================================
63 kmem_cache_t
*zio_cache
;
64 kmem_cache_t
*zio_link_cache
;
65 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
66 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
69 extern vmem_t
*zio_alloc_arena
;
72 #define ZIO_PIPELINE_CONTINUE 0x100
73 #define ZIO_PIPELINE_STOP 0x101
75 #define BP_SPANB(indblkshift, level) \
76 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
77 #define COMPARE_META_LEVEL 0x80000000ul
79 * The following actions directly effect the spa's sync-to-convergence logic.
80 * The values below define the sync pass when we start performing the action.
81 * Care should be taken when changing these values as they directly impact
82 * spa_sync() performance. Tuning these values may introduce subtle performance
83 * pathologies and should only be done in the context of performance analysis.
84 * These tunables will eventually be removed and replaced with #defines once
85 * enough analysis has been done to determine optimal values.
87 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
88 * regular blocks are not deferred.
90 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
91 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
92 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
95 * An allocating zio is one that either currently has the DVA allocate
96 * stage set or will have it later in its lifetime.
98 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
100 boolean_t zio_requeue_io_start_cut_in_line
= B_TRUE
;
103 int zio_buf_debug_limit
= 16384;
105 int zio_buf_debug_limit
= 0;
108 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
114 vmem_t
*data_alloc_arena
= NULL
;
117 data_alloc_arena
= zio_alloc_arena
;
119 zio_cache
= kmem_cache_create("zio_cache",
120 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
121 zio_link_cache
= kmem_cache_create("zio_link_cache",
122 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
125 * For small buffers, we want a cache for each multiple of
126 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
127 * for each quarter-power of 2.
129 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
130 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
133 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
140 * If we are using watchpoints, put each buffer on its own page,
141 * to eliminate the performance overhead of trapping to the
142 * kernel when modifying a non-watched buffer that shares the
143 * page with a watched buffer.
145 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
148 if (size
<= 4 * SPA_MINBLOCKSIZE
) {
149 align
= SPA_MINBLOCKSIZE
;
150 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
151 align
= MIN(p2
>> 2, PAGESIZE
);
156 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
157 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
158 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
161 * Since zio_data bufs do not appear in crash dumps, we
162 * pass KMC_NOTOUCH so that no allocator metadata is
163 * stored with the buffers.
165 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
166 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
167 align
, NULL
, NULL
, NULL
, NULL
, data_alloc_arena
,
168 cflags
| KMC_NOTOUCH
);
173 ASSERT(zio_buf_cache
[c
] != NULL
);
174 if (zio_buf_cache
[c
- 1] == NULL
)
175 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
177 ASSERT(zio_data_buf_cache
[c
] != NULL
);
178 if (zio_data_buf_cache
[c
- 1] == NULL
)
179 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
189 kmem_cache_t
*last_cache
= NULL
;
190 kmem_cache_t
*last_data_cache
= NULL
;
192 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
193 if (zio_buf_cache
[c
] != last_cache
) {
194 last_cache
= zio_buf_cache
[c
];
195 kmem_cache_destroy(zio_buf_cache
[c
]);
197 zio_buf_cache
[c
] = NULL
;
199 if (zio_data_buf_cache
[c
] != last_data_cache
) {
200 last_data_cache
= zio_data_buf_cache
[c
];
201 kmem_cache_destroy(zio_data_buf_cache
[c
]);
203 zio_data_buf_cache
[c
] = NULL
;
206 kmem_cache_destroy(zio_link_cache
);
207 kmem_cache_destroy(zio_cache
);
213 * ==========================================================================
214 * Allocate and free I/O buffers
215 * ==========================================================================
219 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
220 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
221 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
222 * excess / transient data in-core during a crashdump.
225 zio_buf_alloc(size_t size
)
227 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
229 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
231 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
235 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
236 * crashdump if the kernel panics. This exists so that we will limit the amount
237 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
238 * of kernel heap dumped to disk when the kernel panics)
241 zio_data_buf_alloc(size_t size
)
243 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
245 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
247 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
251 zio_buf_free(void *buf
, size_t size
)
253 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
255 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
257 kmem_cache_free(zio_buf_cache
[c
], buf
);
261 zio_data_buf_free(void *buf
, size_t size
)
263 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
265 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
267 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
271 * ==========================================================================
272 * Push and pop I/O transform buffers
273 * ==========================================================================
276 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
277 zio_transform_func_t
*transform
)
279 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
282 * Ensure that anyone expecting this zio to contain a linear ABD isn't
283 * going to get a nasty surprise when they try to access the data.
285 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
287 zt
->zt_orig_abd
= zio
->io_abd
;
288 zt
->zt_orig_size
= zio
->io_size
;
289 zt
->zt_bufsize
= bufsize
;
290 zt
->zt_transform
= transform
;
292 zt
->zt_next
= zio
->io_transform_stack
;
293 zio
->io_transform_stack
= zt
;
300 zio_pop_transforms(zio_t
*zio
)
304 while ((zt
= zio
->io_transform_stack
) != NULL
) {
305 if (zt
->zt_transform
!= NULL
)
306 zt
->zt_transform(zio
,
307 zt
->zt_orig_abd
, zt
->zt_orig_size
);
309 if (zt
->zt_bufsize
!= 0)
310 abd_free(zio
->io_abd
);
312 zio
->io_abd
= zt
->zt_orig_abd
;
313 zio
->io_size
= zt
->zt_orig_size
;
314 zio
->io_transform_stack
= zt
->zt_next
;
316 kmem_free(zt
, sizeof (zio_transform_t
));
321 * ==========================================================================
322 * I/O transform callbacks for subblocks and decompression
323 * ==========================================================================
326 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
328 ASSERT(zio
->io_size
> size
);
330 if (zio
->io_type
== ZIO_TYPE_READ
)
331 abd_copy(data
, zio
->io_abd
, size
);
335 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
337 if (zio
->io_error
== 0) {
338 void *tmp
= abd_borrow_buf(data
, size
);
339 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
340 zio
->io_abd
, tmp
, zio
->io_size
, size
);
341 abd_return_buf_copy(data
, tmp
, size
);
344 zio
->io_error
= SET_ERROR(EIO
);
349 * ==========================================================================
350 * I/O parent/child relationships and pipeline interlocks
351 * ==========================================================================
354 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
356 list_t
*pl
= &cio
->io_parent_list
;
358 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
362 ASSERT((*zl
)->zl_child
== cio
);
363 return ((*zl
)->zl_parent
);
367 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
369 list_t
*cl
= &pio
->io_child_list
;
371 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
375 ASSERT((*zl
)->zl_parent
== pio
);
376 return ((*zl
)->zl_child
);
380 zio_unique_parent(zio_t
*cio
)
382 zio_link_t
*zl
= NULL
;
383 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
385 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
390 zio_add_child(zio_t
*pio
, zio_t
*cio
)
392 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
395 * Logical I/Os can have logical, gang, or vdev children.
396 * Gang I/Os can have gang or vdev children.
397 * Vdev I/Os can only have vdev children.
398 * The following ASSERT captures all of these constraints.
400 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
405 mutex_enter(&cio
->io_lock
);
406 mutex_enter(&pio
->io_lock
);
408 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
410 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
411 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
413 list_insert_head(&pio
->io_child_list
, zl
);
414 list_insert_head(&cio
->io_parent_list
, zl
);
416 pio
->io_child_count
++;
417 cio
->io_parent_count
++;
419 mutex_exit(&pio
->io_lock
);
420 mutex_exit(&cio
->io_lock
);
424 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
426 ASSERT(zl
->zl_parent
== pio
);
427 ASSERT(zl
->zl_child
== cio
);
429 mutex_enter(&cio
->io_lock
);
430 mutex_enter(&pio
->io_lock
);
432 list_remove(&pio
->io_child_list
, zl
);
433 list_remove(&cio
->io_parent_list
, zl
);
435 pio
->io_child_count
--;
436 cio
->io_parent_count
--;
438 mutex_exit(&pio
->io_lock
);
439 mutex_exit(&cio
->io_lock
);
441 kmem_cache_free(zio_link_cache
, zl
);
445 zio_wait_for_children(zio_t
*zio
, enum zio_child child
, enum zio_wait_type wait
)
447 uint64_t *countp
= &zio
->io_children
[child
][wait
];
448 boolean_t waiting
= B_FALSE
;
450 mutex_enter(&zio
->io_lock
);
451 ASSERT(zio
->io_stall
== NULL
);
454 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
455 zio
->io_stall
= countp
;
458 mutex_exit(&zio
->io_lock
);
464 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
466 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
467 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
469 mutex_enter(&pio
->io_lock
);
470 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
471 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
472 pio
->io_reexecute
|= zio
->io_reexecute
;
473 ASSERT3U(*countp
, >, 0);
477 if (*countp
== 0 && pio
->io_stall
== countp
) {
478 zio_taskq_type_t type
=
479 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
481 pio
->io_stall
= NULL
;
482 mutex_exit(&pio
->io_lock
);
484 * Dispatch the parent zio in its own taskq so that
485 * the child can continue to make progress. This also
486 * prevents overflowing the stack when we have deeply nested
487 * parent-child relationships.
489 zio_taskq_dispatch(pio
, type
, B_FALSE
);
491 mutex_exit(&pio
->io_lock
);
496 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
498 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
499 zio
->io_error
= zio
->io_child_error
[c
];
503 zio_bookmark_compare(const void *x1
, const void *x2
)
505 const zio_t
*z1
= x1
;
506 const zio_t
*z2
= x2
;
508 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
510 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
513 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
515 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
518 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
520 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
523 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
525 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
537 * ==========================================================================
538 * Create the various types of I/O (read, write, free, etc)
539 * ==========================================================================
542 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
543 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
544 void *private, zio_type_t type
, zio_priority_t priority
,
545 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
546 const zbookmark_phys_t
*zb
, enum zio_stage stage
, enum zio_stage pipeline
)
550 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
551 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
552 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
554 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
555 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
556 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
558 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW
) != 0);
560 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
561 bzero(zio
, sizeof (zio_t
));
563 mutex_init(&zio
->io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
564 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
566 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
567 offsetof(zio_link_t
, zl_parent_node
));
568 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
569 offsetof(zio_link_t
, zl_child_node
));
570 metaslab_trace_init(&zio
->io_alloc_list
);
573 zio
->io_child_type
= ZIO_CHILD_VDEV
;
574 else if (flags
& ZIO_FLAG_GANG_CHILD
)
575 zio
->io_child_type
= ZIO_CHILD_GANG
;
576 else if (flags
& ZIO_FLAG_DDT_CHILD
)
577 zio
->io_child_type
= ZIO_CHILD_DDT
;
579 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
582 zio
->io_bp
= (blkptr_t
*)bp
;
583 zio
->io_bp_copy
= *bp
;
584 zio
->io_bp_orig
= *bp
;
585 if (type
!= ZIO_TYPE_WRITE
||
586 zio
->io_child_type
== ZIO_CHILD_DDT
)
587 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
588 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
589 zio
->io_logical
= zio
;
590 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
591 pipeline
|= ZIO_GANG_STAGES
;
597 zio
->io_private
= private;
599 zio
->io_priority
= priority
;
601 zio
->io_offset
= offset
;
602 zio
->io_orig_abd
= zio
->io_abd
= data
;
603 zio
->io_orig_size
= zio
->io_size
= psize
;
604 zio
->io_lsize
= lsize
;
605 zio
->io_orig_flags
= zio
->io_flags
= flags
;
606 zio
->io_orig_stage
= zio
->io_stage
= stage
;
607 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
608 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
610 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
611 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
614 zio
->io_bookmark
= *zb
;
617 if (zio
->io_logical
== NULL
)
618 zio
->io_logical
= pio
->io_logical
;
619 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
620 zio
->io_gang_leader
= pio
->io_gang_leader
;
621 zio_add_child(pio
, zio
);
628 zio_destroy(zio_t
*zio
)
630 metaslab_trace_fini(&zio
->io_alloc_list
);
631 list_destroy(&zio
->io_parent_list
);
632 list_destroy(&zio
->io_child_list
);
633 mutex_destroy(&zio
->io_lock
);
634 cv_destroy(&zio
->io_cv
);
635 kmem_cache_free(zio_cache
, zio
);
639 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
640 void *private, enum zio_flag flags
)
644 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
645 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
646 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
652 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
654 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
658 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
660 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
661 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
662 bp
, (longlong_t
)BP_GET_TYPE(bp
));
664 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
665 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
666 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
667 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
669 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
670 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
671 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
672 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
674 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
675 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
676 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
678 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
679 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
680 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
683 if (BP_IS_EMBEDDED(bp
)) {
684 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
685 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
686 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
691 * Pool-specific checks.
693 * Note: it would be nice to verify that the blk_birth and
694 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
695 * allows the birth time of log blocks (and dmu_sync()-ed blocks
696 * that are in the log) to be arbitrarily large.
698 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
699 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
700 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
701 zfs_panic_recover("blkptr at %p DVA %u has invalid "
703 bp
, i
, (longlong_t
)vdevid
);
706 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
708 zfs_panic_recover("blkptr at %p DVA %u has invalid "
710 bp
, i
, (longlong_t
)vdevid
);
713 if (vd
->vdev_ops
== &vdev_hole_ops
) {
714 zfs_panic_recover("blkptr at %p DVA %u has hole "
716 bp
, i
, (longlong_t
)vdevid
);
719 if (vd
->vdev_ops
== &vdev_missing_ops
) {
721 * "missing" vdevs are valid during import, but we
722 * don't have their detailed info (e.g. asize), so
723 * we can't perform any more checks on them.
727 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
728 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
730 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
731 if (offset
+ asize
> vd
->vdev_asize
) {
732 zfs_panic_recover("blkptr at %p DVA %u has invalid "
734 bp
, i
, (longlong_t
)offset
);
740 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
741 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
742 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
746 zfs_blkptr_verify(spa
, bp
);
748 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
749 data
, size
, size
, done
, private,
750 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
751 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
752 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
758 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
759 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
760 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
761 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
762 void *private, zio_priority_t priority
, enum zio_flag flags
,
763 const zbookmark_phys_t
*zb
)
767 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
768 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
769 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
770 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
771 DMU_OT_IS_VALID(zp
->zp_type
) &&
774 zp
->zp_copies
<= spa_max_replication(spa
));
776 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
777 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
778 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
779 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
781 zio
->io_ready
= ready
;
782 zio
->io_children_ready
= children_ready
;
783 zio
->io_physdone
= physdone
;
787 * Data can be NULL if we are going to call zio_write_override() to
788 * provide the already-allocated BP. But we may need the data to
789 * verify a dedup hit (if requested). In this case, don't try to
790 * dedup (just take the already-allocated BP verbatim).
792 if (data
== NULL
&& zio
->io_prop
.zp_dedup_verify
) {
793 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
800 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
801 uint64_t size
, zio_done_func_t
*done
, void *private,
802 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
806 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
807 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
808 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
814 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
816 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
817 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
818 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
819 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
822 * We must reset the io_prop to match the values that existed
823 * when the bp was first written by dmu_sync() keeping in mind
824 * that nopwrite and dedup are mutually exclusive.
826 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
827 zio
->io_prop
.zp_nopwrite
= nopwrite
;
828 zio
->io_prop
.zp_copies
= copies
;
829 zio
->io_bp_override
= bp
;
833 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
836 zfs_blkptr_verify(spa
, bp
);
839 * The check for EMBEDDED is a performance optimization. We
840 * process the free here (by ignoring it) rather than
841 * putting it on the list and then processing it in zio_free_sync().
843 if (BP_IS_EMBEDDED(bp
))
845 metaslab_check_free(spa
, bp
);
848 * Frees that are for the currently-syncing txg, are not going to be
849 * deferred, and which will not need to do a read (i.e. not GANG or
850 * DEDUP), can be processed immediately. Otherwise, put them on the
851 * in-memory list for later processing.
853 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
854 txg
!= spa
->spa_syncing_txg
||
855 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
856 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
858 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
863 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
867 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
869 ASSERT(!BP_IS_HOLE(bp
));
870 ASSERT(spa_syncing_txg(spa
) == txg
);
871 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
873 if (BP_IS_EMBEDDED(bp
))
874 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
876 metaslab_check_free(spa
, bp
);
880 * GANG and DEDUP blocks can induce a read (for the gang block header,
881 * or the DDT), so issue them asynchronously so that this thread is
884 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
885 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
887 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
888 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
889 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
895 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
896 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
900 zfs_blkptr_verify(spa
, bp
);
902 if (BP_IS_EMBEDDED(bp
))
903 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
906 * A claim is an allocation of a specific block. Claims are needed
907 * to support immediate writes in the intent log. The issue is that
908 * immediate writes contain committed data, but in a txg that was
909 * *not* committed. Upon opening the pool after an unclean shutdown,
910 * the intent log claims all blocks that contain immediate write data
911 * so that the SPA knows they're in use.
913 * All claims *must* be resolved in the first txg -- before the SPA
914 * starts allocating blocks -- so that nothing is allocated twice.
915 * If txg == 0 we just verify that the block is claimable.
917 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
918 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
919 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
921 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
922 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
923 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
924 ASSERT0(zio
->io_queued_timestamp
);
930 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
931 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
936 if (vd
->vdev_children
== 0) {
937 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
938 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
939 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
943 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
945 for (c
= 0; c
< vd
->vdev_children
; c
++)
946 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
947 done
, private, flags
));
954 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
955 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
956 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
960 ASSERT(vd
->vdev_children
== 0);
961 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
962 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
963 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
965 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
966 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
967 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
969 zio
->io_prop
.zp_checksum
= checksum
;
975 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
976 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
977 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
981 ASSERT(vd
->vdev_children
== 0);
982 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
983 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
984 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
986 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
987 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
988 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
990 zio
->io_prop
.zp_checksum
= checksum
;
992 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
994 * zec checksums are necessarily destructive -- they modify
995 * the end of the write buffer to hold the verifier/checksum.
996 * Therefore, we must make a local copy in case the data is
997 * being written to multiple places in parallel.
999 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1000 abd_copy(wbuf
, data
, size
);
1002 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1009 * Create a child I/O to do some work for us.
1012 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1013 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1014 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1016 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1020 * vdev child I/Os do not propagate their error to the parent.
1021 * Therefore, for correct operation the caller *must* check for
1022 * and handle the error in the child i/o's done callback.
1023 * The only exceptions are i/os that we don't care about
1024 * (OPTIONAL or REPAIR).
1026 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1030 * In the common case, where the parent zio was to a normal vdev,
1031 * the child zio must be to a child vdev of that vdev. Otherwise,
1032 * the child zio must be to a top-level vdev.
1034 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
!= &vdev_indirect_ops
) {
1035 ASSERT3P(vd
->vdev_parent
, ==, pio
->io_vd
);
1037 ASSERT3P(vd
, ==, vd
->vdev_top
);
1040 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1042 * If we have the bp, then the child should perform the
1043 * checksum and the parent need not. This pushes error
1044 * detection as close to the leaves as possible and
1045 * eliminates redundant checksums in the interior nodes.
1047 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1048 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1051 if (vd
->vdev_ops
->vdev_op_leaf
) {
1052 ASSERT0(vd
->vdev_children
);
1053 offset
+= VDEV_LABEL_START_SIZE
;
1056 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1059 * If we've decided to do a repair, the write is not speculative --
1060 * even if the original read was.
1062 if (flags
& ZIO_FLAG_IO_REPAIR
)
1063 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1066 * If we're creating a child I/O that is not associated with a
1067 * top-level vdev, then the child zio is not an allocating I/O.
1068 * If this is a retried I/O then we ignore it since we will
1069 * have already processed the original allocating I/O.
1071 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1072 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1073 metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
);
1075 ASSERT(mc
->mc_alloc_throttle_enabled
);
1076 ASSERT(type
== ZIO_TYPE_WRITE
);
1077 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1078 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1079 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1080 pio
->io_child_type
== ZIO_CHILD_GANG
);
1082 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1085 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1086 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1087 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1088 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1090 zio
->io_physdone
= pio
->io_physdone
;
1091 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1092 zio
->io_logical
->io_phys_children
++;
1098 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1099 int type
, zio_priority_t priority
, enum zio_flag flags
,
1100 zio_done_func_t
*done
, void *private)
1104 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1106 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1107 data
, size
, size
, done
, private, type
, priority
,
1108 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1110 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1116 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1118 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1120 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1124 zio_shrink(zio_t
*zio
, uint64_t size
)
1126 ASSERT3P(zio
->io_executor
, ==, NULL
);
1127 ASSERT3P(zio
->io_orig_size
, ==, zio
->io_size
);
1128 ASSERT3U(size
, <=, zio
->io_size
);
1131 * We don't shrink for raidz because of problems with the
1132 * reconstruction when reading back less than the block size.
1133 * Note, BP_IS_RAIDZ() assumes no compression.
1135 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1136 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1137 /* we are not doing a raw write */
1138 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1139 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1144 * ==========================================================================
1145 * Prepare to read and write logical blocks
1146 * ==========================================================================
1150 zio_read_bp_init(zio_t
*zio
)
1152 blkptr_t
*bp
= zio
->io_bp
;
1154 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1156 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1157 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1158 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
1160 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1161 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1162 psize
, psize
, zio_decompress
);
1165 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1166 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1168 int psize
= BPE_GET_PSIZE(bp
);
1169 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1170 decode_embedded_bp_compressed(bp
, data
);
1171 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1173 ASSERT(!BP_IS_EMBEDDED(bp
));
1174 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1177 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1178 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1180 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1181 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1183 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1184 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1186 return (ZIO_PIPELINE_CONTINUE
);
1190 zio_write_bp_init(zio_t
*zio
)
1192 if (!IO_IS_ALLOCATING(zio
))
1193 return (ZIO_PIPELINE_CONTINUE
);
1195 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1197 if (zio
->io_bp_override
) {
1198 blkptr_t
*bp
= zio
->io_bp
;
1199 zio_prop_t
*zp
= &zio
->io_prop
;
1201 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1202 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1204 *bp
= *zio
->io_bp_override
;
1205 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1207 if (BP_IS_EMBEDDED(bp
))
1208 return (ZIO_PIPELINE_CONTINUE
);
1211 * If we've been overridden and nopwrite is set then
1212 * set the flag accordingly to indicate that a nopwrite
1213 * has already occurred.
1215 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1216 ASSERT(!zp
->zp_dedup
);
1217 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1218 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1219 return (ZIO_PIPELINE_CONTINUE
);
1222 ASSERT(!zp
->zp_nopwrite
);
1224 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1225 return (ZIO_PIPELINE_CONTINUE
);
1227 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1228 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1230 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
1231 BP_SET_DEDUP(bp
, 1);
1232 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1233 return (ZIO_PIPELINE_CONTINUE
);
1237 * We were unable to handle this as an override bp, treat
1238 * it as a regular write I/O.
1240 zio
->io_bp_override
= NULL
;
1241 *bp
= zio
->io_bp_orig
;
1242 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1245 return (ZIO_PIPELINE_CONTINUE
);
1249 zio_write_compress(zio_t
*zio
)
1251 spa_t
*spa
= zio
->io_spa
;
1252 zio_prop_t
*zp
= &zio
->io_prop
;
1253 enum zio_compress compress
= zp
->zp_compress
;
1254 blkptr_t
*bp
= zio
->io_bp
;
1255 uint64_t lsize
= zio
->io_lsize
;
1256 uint64_t psize
= zio
->io_size
;
1259 EQUIV(lsize
!= psize
, (zio
->io_flags
& ZIO_FLAG_RAW
) != 0);
1262 * If our children haven't all reached the ready stage,
1263 * wait for them and then repeat this pipeline stage.
1265 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1266 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1267 return (ZIO_PIPELINE_STOP
);
1269 if (!IO_IS_ALLOCATING(zio
))
1270 return (ZIO_PIPELINE_CONTINUE
);
1272 if (zio
->io_children_ready
!= NULL
) {
1274 * Now that all our children are ready, run the callback
1275 * associated with this zio in case it wants to modify the
1276 * data to be written.
1278 ASSERT3U(zp
->zp_level
, >, 0);
1279 zio
->io_children_ready(zio
);
1282 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1283 ASSERT(zio
->io_bp_override
== NULL
);
1285 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1287 * We're rewriting an existing block, which means we're
1288 * working on behalf of spa_sync(). For spa_sync() to
1289 * converge, it must eventually be the case that we don't
1290 * have to allocate new blocks. But compression changes
1291 * the blocksize, which forces a reallocate, and makes
1292 * convergence take longer. Therefore, after the first
1293 * few passes, stop compressing to ensure convergence.
1295 pass
= spa_sync_pass(spa
);
1297 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1298 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1299 ASSERT(!BP_GET_DEDUP(bp
));
1301 if (pass
>= zfs_sync_pass_dont_compress
)
1302 compress
= ZIO_COMPRESS_OFF
;
1304 /* Make sure someone doesn't change their mind on overwrites */
1305 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1306 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1309 /* If it's a compressed write that is not raw, compress the buffer. */
1310 if (compress
!= ZIO_COMPRESS_OFF
&& psize
== lsize
) {
1311 void *cbuf
= zio_buf_alloc(lsize
);
1312 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1313 if (psize
== 0 || psize
== lsize
) {
1314 compress
= ZIO_COMPRESS_OFF
;
1315 zio_buf_free(cbuf
, lsize
);
1316 } else if (!zp
->zp_dedup
&& psize
<= BPE_PAYLOAD_SIZE
&&
1317 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1318 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1319 encode_embedded_bp_compressed(bp
,
1320 cbuf
, compress
, lsize
, psize
);
1321 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1322 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1323 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1324 zio_buf_free(cbuf
, lsize
);
1325 bp
->blk_birth
= zio
->io_txg
;
1326 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1327 ASSERT(spa_feature_is_active(spa
,
1328 SPA_FEATURE_EMBEDDED_DATA
));
1329 return (ZIO_PIPELINE_CONTINUE
);
1332 * Round up compressed size up to the ashift
1333 * of the smallest-ashift device, and zero the tail.
1334 * This ensures that the compressed size of the BP
1335 * (and thus compressratio property) are correct,
1336 * in that we charge for the padding used to fill out
1339 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1340 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1341 1ULL << spa
->spa_min_ashift
);
1342 if (rounded
>= lsize
) {
1343 compress
= ZIO_COMPRESS_OFF
;
1344 zio_buf_free(cbuf
, lsize
);
1347 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1348 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1349 abd_zero_off(cdata
, psize
, rounded
- psize
);
1351 zio_push_transform(zio
, cdata
,
1352 psize
, lsize
, NULL
);
1357 * We were unable to handle this as an override bp, treat
1358 * it as a regular write I/O.
1360 zio
->io_bp_override
= NULL
;
1361 *bp
= zio
->io_bp_orig
;
1362 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1364 ASSERT3U(psize
, !=, 0);
1368 * The final pass of spa_sync() must be all rewrites, but the first
1369 * few passes offer a trade-off: allocating blocks defers convergence,
1370 * but newly allocated blocks are sequential, so they can be written
1371 * to disk faster. Therefore, we allow the first few passes of
1372 * spa_sync() to allocate new blocks, but force rewrites after that.
1373 * There should only be a handful of blocks after pass 1 in any case.
1375 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1376 BP_GET_PSIZE(bp
) == psize
&&
1377 pass
>= zfs_sync_pass_rewrite
) {
1379 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1380 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1381 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1384 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1388 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1389 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1390 BP_SET_LSIZE(bp
, lsize
);
1391 BP_SET_TYPE(bp
, zp
->zp_type
);
1392 BP_SET_LEVEL(bp
, zp
->zp_level
);
1393 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1395 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1397 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1398 BP_SET_LSIZE(bp
, lsize
);
1399 BP_SET_TYPE(bp
, zp
->zp_type
);
1400 BP_SET_LEVEL(bp
, zp
->zp_level
);
1401 BP_SET_PSIZE(bp
, psize
);
1402 BP_SET_COMPRESS(bp
, compress
);
1403 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1404 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1405 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1407 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1408 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1409 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1411 if (zp
->zp_nopwrite
) {
1412 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1413 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1414 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1417 return (ZIO_PIPELINE_CONTINUE
);
1421 zio_free_bp_init(zio_t
*zio
)
1423 blkptr_t
*bp
= zio
->io_bp
;
1425 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1426 if (BP_GET_DEDUP(bp
))
1427 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1430 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1432 return (ZIO_PIPELINE_CONTINUE
);
1436 * ==========================================================================
1437 * Execute the I/O pipeline
1438 * ==========================================================================
1442 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1444 spa_t
*spa
= zio
->io_spa
;
1445 zio_type_t t
= zio
->io_type
;
1446 int flags
= (cutinline
? TQ_FRONT
: 0);
1449 * If we're a config writer or a probe, the normal issue and
1450 * interrupt threads may all be blocked waiting for the config lock.
1451 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1453 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1457 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1459 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1463 * If this is a high priority I/O, then use the high priority taskq if
1466 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1467 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1470 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1473 * NB: We are assuming that the zio can only be dispatched
1474 * to a single taskq at a time. It would be a grievous error
1475 * to dispatch the zio to another taskq at the same time.
1477 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
1478 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1479 flags
, &zio
->io_tqent
);
1483 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1485 kthread_t
*executor
= zio
->io_executor
;
1486 spa_t
*spa
= zio
->io_spa
;
1488 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1489 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1491 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1492 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1501 zio_issue_async(zio_t
*zio
)
1503 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1505 return (ZIO_PIPELINE_STOP
);
1509 zio_interrupt(zio_t
*zio
)
1511 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1515 zio_delay_interrupt(zio_t
*zio
)
1518 * The timeout_generic() function isn't defined in userspace, so
1519 * rather than trying to implement the function, the zio delay
1520 * functionality has been disabled for userspace builds.
1525 * If io_target_timestamp is zero, then no delay has been registered
1526 * for this IO, thus jump to the end of this function and "skip" the
1527 * delay; issuing it directly to the zio layer.
1529 if (zio
->io_target_timestamp
!= 0) {
1530 hrtime_t now
= gethrtime();
1532 if (now
>= zio
->io_target_timestamp
) {
1534 * This IO has already taken longer than the target
1535 * delay to complete, so we don't want to delay it
1536 * any longer; we "miss" the delay and issue it
1537 * directly to the zio layer. This is likely due to
1538 * the target latency being set to a value less than
1539 * the underlying hardware can satisfy (e.g. delay
1540 * set to 1ms, but the disks take 10ms to complete an
1544 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1549 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1551 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1552 hrtime_t
, now
, hrtime_t
, diff
);
1554 (void) timeout_generic(CALLOUT_NORMAL
,
1555 (void (*)(void *))zio_interrupt
, zio
, diff
, 1, 0);
1562 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1567 * Execute the I/O pipeline until one of the following occurs:
1569 * (1) the I/O completes
1570 * (2) the pipeline stalls waiting for dependent child I/Os
1571 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1572 * (4) the I/O is delegated by vdev-level caching or aggregation
1573 * (5) the I/O is deferred due to vdev-level queueing
1574 * (6) the I/O is handed off to another thread.
1576 * In all cases, the pipeline stops whenever there's no CPU work; it never
1577 * burns a thread in cv_wait().
1579 * There's no locking on io_stage because there's no legitimate way
1580 * for multiple threads to be attempting to process the same I/O.
1582 static zio_pipe_stage_t
*zio_pipeline
[];
1585 zio_execute(zio_t
*zio
)
1587 zio
->io_executor
= curthread
;
1589 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1591 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1592 enum zio_stage pipeline
= zio
->io_pipeline
;
1593 enum zio_stage stage
= zio
->io_stage
;
1596 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1597 ASSERT(ISP2(stage
));
1598 ASSERT(zio
->io_stall
== NULL
);
1602 } while ((stage
& pipeline
) == 0);
1604 ASSERT(stage
<= ZIO_STAGE_DONE
);
1607 * If we are in interrupt context and this pipeline stage
1608 * will grab a config lock that is held across I/O,
1609 * or may wait for an I/O that needs an interrupt thread
1610 * to complete, issue async to avoid deadlock.
1612 * For VDEV_IO_START, we cut in line so that the io will
1613 * be sent to disk promptly.
1615 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1616 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1617 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1618 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1619 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1623 zio
->io_stage
= stage
;
1624 zio
->io_pipeline_trace
|= zio
->io_stage
;
1625 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1627 if (rv
== ZIO_PIPELINE_STOP
)
1630 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1635 * ==========================================================================
1636 * Initiate I/O, either sync or async
1637 * ==========================================================================
1640 zio_wait(zio_t
*zio
)
1644 ASSERT3P(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1645 ASSERT3P(zio
->io_executor
, ==, NULL
);
1647 zio
->io_waiter
= curthread
;
1648 ASSERT0(zio
->io_queued_timestamp
);
1649 zio
->io_queued_timestamp
= gethrtime();
1653 mutex_enter(&zio
->io_lock
);
1654 while (zio
->io_executor
!= NULL
)
1655 cv_wait(&zio
->io_cv
, &zio
->io_lock
);
1656 mutex_exit(&zio
->io_lock
);
1658 error
= zio
->io_error
;
1665 zio_nowait(zio_t
*zio
)
1667 ASSERT3P(zio
->io_executor
, ==, NULL
);
1669 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1670 zio_unique_parent(zio
) == NULL
) {
1672 * This is a logical async I/O with no parent to wait for it.
1673 * We add it to the spa_async_root_zio "Godfather" I/O which
1674 * will ensure they complete prior to unloading the pool.
1676 spa_t
*spa
= zio
->io_spa
;
1678 zio_add_child(spa
->spa_async_zio_root
[CPU_SEQID
], zio
);
1681 ASSERT0(zio
->io_queued_timestamp
);
1682 zio
->io_queued_timestamp
= gethrtime();
1687 * ==========================================================================
1688 * Reexecute, cancel, or suspend/resume failed I/O
1689 * ==========================================================================
1693 zio_reexecute(zio_t
*pio
)
1695 zio_t
*cio
, *cio_next
;
1697 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1698 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1699 ASSERT(pio
->io_gang_leader
== NULL
);
1700 ASSERT(pio
->io_gang_tree
== NULL
);
1702 pio
->io_flags
= pio
->io_orig_flags
;
1703 pio
->io_stage
= pio
->io_orig_stage
;
1704 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1705 pio
->io_reexecute
= 0;
1706 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1707 pio
->io_pipeline_trace
= 0;
1709 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1710 pio
->io_state
[w
] = 0;
1711 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1712 pio
->io_child_error
[c
] = 0;
1714 if (IO_IS_ALLOCATING(pio
))
1715 BP_ZERO(pio
->io_bp
);
1718 * As we reexecute pio's children, new children could be created.
1719 * New children go to the head of pio's io_child_list, however,
1720 * so we will (correctly) not reexecute them. The key is that
1721 * the remainder of pio's io_child_list, from 'cio_next' onward,
1722 * cannot be affected by any side effects of reexecuting 'cio'.
1724 zio_link_t
*zl
= NULL
;
1725 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1726 cio_next
= zio_walk_children(pio
, &zl
);
1727 mutex_enter(&pio
->io_lock
);
1728 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1729 pio
->io_children
[cio
->io_child_type
][w
]++;
1730 mutex_exit(&pio
->io_lock
);
1735 * Now that all children have been reexecuted, execute the parent.
1736 * We don't reexecute "The Godfather" I/O here as it's the
1737 * responsibility of the caller to wait on it.
1739 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1740 pio
->io_queued_timestamp
= gethrtime();
1746 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1748 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1749 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1750 "failure and the failure mode property for this pool "
1751 "is set to panic.", spa_name(spa
));
1753 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1755 mutex_enter(&spa
->spa_suspend_lock
);
1757 if (spa
->spa_suspend_zio_root
== NULL
)
1758 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1759 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1760 ZIO_FLAG_GODFATHER
);
1762 spa
->spa_suspended
= B_TRUE
;
1765 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1766 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1767 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1768 ASSERT(zio_unique_parent(zio
) == NULL
);
1769 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1770 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1773 mutex_exit(&spa
->spa_suspend_lock
);
1777 zio_resume(spa_t
*spa
)
1782 * Reexecute all previously suspended i/o.
1784 mutex_enter(&spa
->spa_suspend_lock
);
1785 spa
->spa_suspended
= B_FALSE
;
1786 cv_broadcast(&spa
->spa_suspend_cv
);
1787 pio
= spa
->spa_suspend_zio_root
;
1788 spa
->spa_suspend_zio_root
= NULL
;
1789 mutex_exit(&spa
->spa_suspend_lock
);
1795 return (zio_wait(pio
));
1799 zio_resume_wait(spa_t
*spa
)
1801 mutex_enter(&spa
->spa_suspend_lock
);
1802 while (spa_suspended(spa
))
1803 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1804 mutex_exit(&spa
->spa_suspend_lock
);
1808 * ==========================================================================
1811 * A gang block is a collection of small blocks that looks to the DMU
1812 * like one large block. When zio_dva_allocate() cannot find a block
1813 * of the requested size, due to either severe fragmentation or the pool
1814 * being nearly full, it calls zio_write_gang_block() to construct the
1815 * block from smaller fragments.
1817 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1818 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1819 * an indirect block: it's an array of block pointers. It consumes
1820 * only one sector and hence is allocatable regardless of fragmentation.
1821 * The gang header's bps point to its gang members, which hold the data.
1823 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1824 * as the verifier to ensure uniqueness of the SHA256 checksum.
1825 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1826 * not the gang header. This ensures that data block signatures (needed for
1827 * deduplication) are independent of how the block is physically stored.
1829 * Gang blocks can be nested: a gang member may itself be a gang block.
1830 * Thus every gang block is a tree in which root and all interior nodes are
1831 * gang headers, and the leaves are normal blocks that contain user data.
1832 * The root of the gang tree is called the gang leader.
1834 * To perform any operation (read, rewrite, free, claim) on a gang block,
1835 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1836 * in the io_gang_tree field of the original logical i/o by recursively
1837 * reading the gang leader and all gang headers below it. This yields
1838 * an in-core tree containing the contents of every gang header and the
1839 * bps for every constituent of the gang block.
1841 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1842 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1843 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1844 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1845 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1846 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1847 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1848 * of the gang header plus zio_checksum_compute() of the data to update the
1849 * gang header's blk_cksum as described above.
1851 * The two-phase assemble/issue model solves the problem of partial failure --
1852 * what if you'd freed part of a gang block but then couldn't read the
1853 * gang header for another part? Assembling the entire gang tree first
1854 * ensures that all the necessary gang header I/O has succeeded before
1855 * starting the actual work of free, claim, or write. Once the gang tree
1856 * is assembled, free and claim are in-memory operations that cannot fail.
1858 * In the event that a gang write fails, zio_dva_unallocate() walks the
1859 * gang tree to immediately free (i.e. insert back into the space map)
1860 * everything we've allocated. This ensures that we don't get ENOSPC
1861 * errors during repeated suspend/resume cycles due to a flaky device.
1863 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1864 * the gang tree, we won't modify the block, so we can safely defer the free
1865 * (knowing that the block is still intact). If we *can* assemble the gang
1866 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1867 * each constituent bp and we can allocate a new block on the next sync pass.
1869 * In all cases, the gang tree allows complete recovery from partial failure.
1870 * ==========================================================================
1874 zio_gang_issue_func_done(zio_t
*zio
)
1876 abd_put(zio
->io_abd
);
1880 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1886 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
1887 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
1888 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1889 &pio
->io_bookmark
));
1893 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1900 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1901 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1902 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
1903 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1906 * As we rewrite each gang header, the pipeline will compute
1907 * a new gang block header checksum for it; but no one will
1908 * compute a new data checksum, so we do that here. The one
1909 * exception is the gang leader: the pipeline already computed
1910 * its data checksum because that stage precedes gang assembly.
1911 * (Presently, nothing actually uses interior data checksums;
1912 * this is just good hygiene.)
1914 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1915 abd_t
*buf
= abd_get_offset(data
, offset
);
1917 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1918 buf
, BP_GET_PSIZE(bp
));
1923 * If we are here to damage data for testing purposes,
1924 * leave the GBH alone so that we can detect the damage.
1926 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1927 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1929 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1930 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
1931 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
1932 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1940 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1943 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1944 ZIO_GANG_CHILD_FLAGS(pio
)));
1949 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1952 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1953 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1956 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
1965 static void zio_gang_tree_assemble_done(zio_t
*zio
);
1967 static zio_gang_node_t
*
1968 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
1970 zio_gang_node_t
*gn
;
1972 ASSERT(*gnpp
== NULL
);
1974 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
1975 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
1982 zio_gang_node_free(zio_gang_node_t
**gnpp
)
1984 zio_gang_node_t
*gn
= *gnpp
;
1986 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1987 ASSERT(gn
->gn_child
[g
] == NULL
);
1989 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1990 kmem_free(gn
, sizeof (*gn
));
1995 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
1997 zio_gang_node_t
*gn
= *gnpp
;
2002 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2003 zio_gang_tree_free(&gn
->gn_child
[g
]);
2005 zio_gang_node_free(gnpp
);
2009 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2011 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2012 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2014 ASSERT(gio
->io_gang_leader
== gio
);
2015 ASSERT(BP_IS_GANG(bp
));
2017 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2018 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2019 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2023 zio_gang_tree_assemble_done(zio_t
*zio
)
2025 zio_t
*gio
= zio
->io_gang_leader
;
2026 zio_gang_node_t
*gn
= zio
->io_private
;
2027 blkptr_t
*bp
= zio
->io_bp
;
2029 ASSERT(gio
== zio_unique_parent(zio
));
2030 ASSERT(zio
->io_child_count
== 0);
2035 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2036 if (BP_SHOULD_BYTESWAP(bp
))
2037 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2039 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2040 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2041 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2043 abd_put(zio
->io_abd
);
2045 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2046 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2047 if (!BP_IS_GANG(gbp
))
2049 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2054 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2057 zio_t
*gio
= pio
->io_gang_leader
;
2060 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2061 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2062 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2065 * If you're a gang header, your data is in gn->gn_gbh.
2066 * If you're a gang member, your data is in 'data' and gn == NULL.
2068 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2071 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2073 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2074 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2075 if (BP_IS_HOLE(gbp
))
2077 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2079 offset
+= BP_GET_PSIZE(gbp
);
2083 if (gn
== gio
->io_gang_tree
)
2084 ASSERT3U(gio
->io_size
, ==, offset
);
2091 zio_gang_assemble(zio_t
*zio
)
2093 blkptr_t
*bp
= zio
->io_bp
;
2095 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2096 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2098 zio
->io_gang_leader
= zio
;
2100 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2102 return (ZIO_PIPELINE_CONTINUE
);
2106 zio_gang_issue(zio_t
*zio
)
2108 blkptr_t
*bp
= zio
->io_bp
;
2110 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
2111 return (ZIO_PIPELINE_STOP
);
2113 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2114 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2116 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2117 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2120 zio_gang_tree_free(&zio
->io_gang_tree
);
2122 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2124 return (ZIO_PIPELINE_CONTINUE
);
2128 zio_write_gang_member_ready(zio_t
*zio
)
2130 zio_t
*pio
= zio_unique_parent(zio
);
2131 zio_t
*gio
= zio
->io_gang_leader
;
2132 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2133 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2136 if (BP_IS_HOLE(zio
->io_bp
))
2139 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2141 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2142 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2143 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2144 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2145 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2147 mutex_enter(&pio
->io_lock
);
2148 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2149 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2150 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2151 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2152 DVA_SET_ASIZE(&pdva
[d
], asize
);
2154 mutex_exit(&pio
->io_lock
);
2158 zio_write_gang_done(zio_t
*zio
)
2160 abd_put(zio
->io_abd
);
2164 zio_write_gang_block(zio_t
*pio
)
2166 spa_t
*spa
= pio
->io_spa
;
2167 metaslab_class_t
*mc
= spa_normal_class(spa
);
2168 blkptr_t
*bp
= pio
->io_bp
;
2169 zio_t
*gio
= pio
->io_gang_leader
;
2171 zio_gang_node_t
*gn
, **gnpp
;
2172 zio_gbh_phys_t
*gbh
;
2174 uint64_t txg
= pio
->io_txg
;
2175 uint64_t resid
= pio
->io_size
;
2177 int copies
= gio
->io_prop
.zp_copies
;
2178 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2182 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2183 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2184 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2185 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2187 flags
|= METASLAB_ASYNC_ALLOC
;
2188 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2191 * The logical zio has already placed a reservation for
2192 * 'copies' allocation slots but gang blocks may require
2193 * additional copies. These additional copies
2194 * (i.e. gbh_copies - copies) are guaranteed to succeed
2195 * since metaslab_class_throttle_reserve() always allows
2196 * additional reservations for gang blocks.
2198 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2202 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2203 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2204 &pio
->io_alloc_list
, pio
);
2206 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2207 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2208 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2211 * If we failed to allocate the gang block header then
2212 * we remove any additional allocation reservations that
2213 * we placed here. The original reservation will
2214 * be removed when the logical I/O goes to the ready
2217 metaslab_class_throttle_unreserve(mc
,
2218 gbh_copies
- copies
, pio
);
2220 pio
->io_error
= error
;
2221 return (ZIO_PIPELINE_CONTINUE
);
2225 gnpp
= &gio
->io_gang_tree
;
2227 gnpp
= pio
->io_private
;
2228 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2231 gn
= zio_gang_node_alloc(gnpp
);
2233 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2234 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2237 * Create the gang header.
2239 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2240 zio_write_gang_done
, NULL
, pio
->io_priority
,
2241 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2244 * Create and nowait the gang children.
2246 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2247 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2249 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2251 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2252 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2253 zp
.zp_type
= DMU_OT_NONE
;
2255 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2256 zp
.zp_dedup
= B_FALSE
;
2257 zp
.zp_dedup_verify
= B_FALSE
;
2258 zp
.zp_nopwrite
= B_FALSE
;
2260 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2261 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2262 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2263 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2264 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2266 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2267 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2268 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2271 * Gang children won't throttle but we should
2272 * account for their work, so reserve an allocation
2273 * slot for them here.
2275 VERIFY(metaslab_class_throttle_reserve(mc
,
2276 zp
.zp_copies
, cio
, flags
));
2282 * Set pio's pipeline to just wait for zio to finish.
2284 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2288 return (ZIO_PIPELINE_CONTINUE
);
2292 * The zio_nop_write stage in the pipeline determines if allocating a
2293 * new bp is necessary. The nopwrite feature can handle writes in
2294 * either syncing or open context (i.e. zil writes) and as a result is
2295 * mutually exclusive with dedup.
2297 * By leveraging a cryptographically secure checksum, such as SHA256, we
2298 * can compare the checksums of the new data and the old to determine if
2299 * allocating a new block is required. Note that our requirements for
2300 * cryptographic strength are fairly weak: there can't be any accidental
2301 * hash collisions, but we don't need to be secure against intentional
2302 * (malicious) collisions. To trigger a nopwrite, you have to be able
2303 * to write the file to begin with, and triggering an incorrect (hash
2304 * collision) nopwrite is no worse than simply writing to the file.
2305 * That said, there are no known attacks against the checksum algorithms
2306 * used for nopwrite, assuming that the salt and the checksums
2307 * themselves remain secret.
2310 zio_nop_write(zio_t
*zio
)
2312 blkptr_t
*bp
= zio
->io_bp
;
2313 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2314 zio_prop_t
*zp
= &zio
->io_prop
;
2316 ASSERT(BP_GET_LEVEL(bp
) == 0);
2317 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2318 ASSERT(zp
->zp_nopwrite
);
2319 ASSERT(!zp
->zp_dedup
);
2320 ASSERT(zio
->io_bp_override
== NULL
);
2321 ASSERT(IO_IS_ALLOCATING(zio
));
2324 * Check to see if the original bp and the new bp have matching
2325 * characteristics (i.e. same checksum, compression algorithms, etc).
2326 * If they don't then just continue with the pipeline which will
2327 * allocate a new bp.
2329 if (BP_IS_HOLE(bp_orig
) ||
2330 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2331 ZCHECKSUM_FLAG_NOPWRITE
) ||
2332 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2333 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2334 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2335 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2336 return (ZIO_PIPELINE_CONTINUE
);
2339 * If the checksums match then reset the pipeline so that we
2340 * avoid allocating a new bp and issuing any I/O.
2342 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2343 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2344 ZCHECKSUM_FLAG_NOPWRITE
);
2345 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2346 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2347 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2348 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2349 sizeof (uint64_t)) == 0);
2352 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2353 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2356 return (ZIO_PIPELINE_CONTINUE
);
2360 * ==========================================================================
2362 * ==========================================================================
2365 zio_ddt_child_read_done(zio_t
*zio
)
2367 blkptr_t
*bp
= zio
->io_bp
;
2368 ddt_entry_t
*dde
= zio
->io_private
;
2370 zio_t
*pio
= zio_unique_parent(zio
);
2372 mutex_enter(&pio
->io_lock
);
2373 ddp
= ddt_phys_select(dde
, bp
);
2374 if (zio
->io_error
== 0)
2375 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2377 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2378 dde
->dde_repair_abd
= zio
->io_abd
;
2380 abd_free(zio
->io_abd
);
2381 mutex_exit(&pio
->io_lock
);
2385 zio_ddt_read_start(zio_t
*zio
)
2387 blkptr_t
*bp
= zio
->io_bp
;
2389 ASSERT(BP_GET_DEDUP(bp
));
2390 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2391 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2393 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2394 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2395 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2396 ddt_phys_t
*ddp
= dde
->dde_phys
;
2397 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2400 ASSERT(zio
->io_vsd
== NULL
);
2403 if (ddp_self
== NULL
)
2404 return (ZIO_PIPELINE_CONTINUE
);
2406 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2407 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2409 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2411 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2412 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2413 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2414 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2415 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2417 return (ZIO_PIPELINE_CONTINUE
);
2420 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2421 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2422 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2424 return (ZIO_PIPELINE_CONTINUE
);
2428 zio_ddt_read_done(zio_t
*zio
)
2430 blkptr_t
*bp
= zio
->io_bp
;
2432 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
2433 return (ZIO_PIPELINE_STOP
);
2435 ASSERT(BP_GET_DEDUP(bp
));
2436 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2437 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2439 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2440 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2441 ddt_entry_t
*dde
= zio
->io_vsd
;
2443 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2444 return (ZIO_PIPELINE_CONTINUE
);
2447 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2448 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2449 return (ZIO_PIPELINE_STOP
);
2451 if (dde
->dde_repair_abd
!= NULL
) {
2452 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2454 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2456 ddt_repair_done(ddt
, dde
);
2460 ASSERT(zio
->io_vsd
== NULL
);
2462 return (ZIO_PIPELINE_CONTINUE
);
2466 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2468 spa_t
*spa
= zio
->io_spa
;
2469 boolean_t do_raw
= (zio
->io_flags
& ZIO_FLAG_RAW
);
2471 /* We should never get a raw, override zio */
2472 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2475 * Note: we compare the original data, not the transformed data,
2476 * because when zio->io_bp is an override bp, we will not have
2477 * pushed the I/O transforms. That's an important optimization
2478 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2480 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2481 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2484 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2485 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
,
2486 zio
->io_orig_size
) != 0);
2490 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2491 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2493 if (ddp
->ddp_phys_birth
!= 0) {
2494 arc_buf_t
*abuf
= NULL
;
2495 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2496 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2497 blkptr_t blk
= *zio
->io_bp
;
2500 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2505 * Intuitively, it would make more sense to compare
2506 * io_abd than io_orig_abd in the raw case since you
2507 * don't want to look at any transformations that have
2508 * happened to the data. However, for raw I/Os the
2509 * data will actually be the same in io_abd and
2510 * io_orig_abd, so all we have to do is issue this as
2514 zio_flags
|= ZIO_FLAG_RAW
;
2515 ASSERT3U(zio
->io_size
, ==, zio
->io_orig_size
);
2516 ASSERT0(abd_cmp(zio
->io_abd
, zio
->io_orig_abd
,
2518 ASSERT3P(zio
->io_transform_stack
, ==, NULL
);
2521 error
= arc_read(NULL
, spa
, &blk
,
2522 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2523 zio_flags
, &aflags
, &zio
->io_bookmark
);
2526 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
2527 abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2528 zio
->io_orig_size
) != 0)
2529 error
= SET_ERROR(EEXIST
);
2530 arc_buf_destroy(abuf
, &abuf
);
2534 return (error
!= 0);
2542 zio_ddt_child_write_ready(zio_t
*zio
)
2544 int p
= zio
->io_prop
.zp_copies
;
2545 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2546 ddt_entry_t
*dde
= zio
->io_private
;
2547 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2555 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2557 ddt_phys_fill(ddp
, zio
->io_bp
);
2559 zio_link_t
*zl
= NULL
;
2560 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2561 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2567 zio_ddt_child_write_done(zio_t
*zio
)
2569 int p
= zio
->io_prop
.zp_copies
;
2570 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2571 ddt_entry_t
*dde
= zio
->io_private
;
2572 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2576 ASSERT(ddp
->ddp_refcnt
== 0);
2577 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2578 dde
->dde_lead_zio
[p
] = NULL
;
2580 if (zio
->io_error
== 0) {
2581 zio_link_t
*zl
= NULL
;
2582 while (zio_walk_parents(zio
, &zl
) != NULL
)
2583 ddt_phys_addref(ddp
);
2585 ddt_phys_clear(ddp
);
2592 zio_ddt_ditto_write_done(zio_t
*zio
)
2594 int p
= DDT_PHYS_DITTO
;
2595 zio_prop_t
*zp
= &zio
->io_prop
;
2596 blkptr_t
*bp
= zio
->io_bp
;
2597 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2598 ddt_entry_t
*dde
= zio
->io_private
;
2599 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2600 ddt_key_t
*ddk
= &dde
->dde_key
;
2604 ASSERT(ddp
->ddp_refcnt
== 0);
2605 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2606 dde
->dde_lead_zio
[p
] = NULL
;
2608 if (zio
->io_error
== 0) {
2609 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2610 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2611 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2612 if (ddp
->ddp_phys_birth
!= 0)
2613 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2614 ddt_phys_fill(ddp
, bp
);
2621 zio_ddt_write(zio_t
*zio
)
2623 spa_t
*spa
= zio
->io_spa
;
2624 blkptr_t
*bp
= zio
->io_bp
;
2625 uint64_t txg
= zio
->io_txg
;
2626 zio_prop_t
*zp
= &zio
->io_prop
;
2627 int p
= zp
->zp_copies
;
2631 ddt_t
*ddt
= ddt_select(spa
, bp
);
2635 ASSERT(BP_GET_DEDUP(bp
));
2636 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2637 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2638 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2641 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2642 ddp
= &dde
->dde_phys
[p
];
2644 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2646 * If we're using a weak checksum, upgrade to a strong checksum
2647 * and try again. If we're already using a strong checksum,
2648 * we can't resolve it, so just convert to an ordinary write.
2649 * (And automatically e-mail a paper to Nature?)
2651 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2652 ZCHECKSUM_FLAG_DEDUP
)) {
2653 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2654 zio_pop_transforms(zio
);
2655 zio
->io_stage
= ZIO_STAGE_OPEN
;
2658 zp
->zp_dedup
= B_FALSE
;
2659 BP_SET_DEDUP(bp
, B_FALSE
);
2661 ASSERT(!BP_GET_DEDUP(bp
));
2662 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2664 return (ZIO_PIPELINE_CONTINUE
);
2667 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2668 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2670 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2671 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2672 zio_prop_t czp
= *zp
;
2674 czp
.zp_copies
= ditto_copies
;
2677 * If we arrived here with an override bp, we won't have run
2678 * the transform stack, so we won't have the data we need to
2679 * generate a child i/o. So, toss the override bp and restart.
2680 * This is safe, because using the override bp is just an
2681 * optimization; and it's rare, so the cost doesn't matter.
2683 if (zio
->io_bp_override
) {
2684 zio_pop_transforms(zio
);
2685 zio
->io_stage
= ZIO_STAGE_OPEN
;
2686 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2687 zio
->io_bp_override
= NULL
;
2690 return (ZIO_PIPELINE_CONTINUE
);
2693 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2694 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2695 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2696 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2698 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2699 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2702 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2703 if (ddp
->ddp_phys_birth
!= 0)
2704 ddt_bp_fill(ddp
, bp
, txg
);
2705 if (dde
->dde_lead_zio
[p
] != NULL
)
2706 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2708 ddt_phys_addref(ddp
);
2709 } else if (zio
->io_bp_override
) {
2710 ASSERT(bp
->blk_birth
== txg
);
2711 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2712 ddt_phys_fill(ddp
, bp
);
2713 ddt_phys_addref(ddp
);
2715 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2716 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
2717 zio_ddt_child_write_ready
, NULL
, NULL
,
2718 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2719 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2721 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2722 dde
->dde_lead_zio
[p
] = cio
;
2732 return (ZIO_PIPELINE_CONTINUE
);
2735 ddt_entry_t
*freedde
; /* for debugging */
2738 zio_ddt_free(zio_t
*zio
)
2740 spa_t
*spa
= zio
->io_spa
;
2741 blkptr_t
*bp
= zio
->io_bp
;
2742 ddt_t
*ddt
= ddt_select(spa
, bp
);
2746 ASSERT(BP_GET_DEDUP(bp
));
2747 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2750 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2751 ddp
= ddt_phys_select(dde
, bp
);
2752 ddt_phys_decref(ddp
);
2755 return (ZIO_PIPELINE_CONTINUE
);
2759 * ==========================================================================
2760 * Allocate and free blocks
2761 * ==========================================================================
2765 zio_io_to_allocate(spa_t
*spa
)
2769 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
2771 zio
= avl_first(&spa
->spa_alloc_tree
);
2775 ASSERT(IO_IS_ALLOCATING(zio
));
2778 * Try to place a reservation for this zio. If we're unable to
2779 * reserve then we throttle.
2781 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
2782 zio
->io_prop
.zp_copies
, zio
, 0)) {
2786 avl_remove(&spa
->spa_alloc_tree
, zio
);
2787 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
2793 zio_dva_throttle(zio_t
*zio
)
2795 spa_t
*spa
= zio
->io_spa
;
2798 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
2799 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
2800 zio
->io_child_type
== ZIO_CHILD_GANG
||
2801 zio
->io_flags
& ZIO_FLAG_NODATA
) {
2802 return (ZIO_PIPELINE_CONTINUE
);
2805 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2807 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2808 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2810 mutex_enter(&spa
->spa_alloc_lock
);
2812 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2813 avl_add(&spa
->spa_alloc_tree
, zio
);
2815 nio
= zio_io_to_allocate(zio
->io_spa
);
2816 mutex_exit(&spa
->spa_alloc_lock
);
2819 return (ZIO_PIPELINE_CONTINUE
);
2822 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2824 * We are passing control to a new zio so make sure that
2825 * it is processed by a different thread. We do this to
2826 * avoid stack overflows that can occur when parents are
2827 * throttled and children are making progress. We allow
2828 * it to go to the head of the taskq since it's already
2831 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2833 return (ZIO_PIPELINE_STOP
);
2837 zio_allocate_dispatch(spa_t
*spa
)
2841 mutex_enter(&spa
->spa_alloc_lock
);
2842 zio
= zio_io_to_allocate(spa
);
2843 mutex_exit(&spa
->spa_alloc_lock
);
2847 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
2848 ASSERT0(zio
->io_error
);
2849 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2853 zio_dva_allocate(zio_t
*zio
)
2855 spa_t
*spa
= zio
->io_spa
;
2856 metaslab_class_t
*mc
= spa_normal_class(spa
);
2857 blkptr_t
*bp
= zio
->io_bp
;
2861 if (zio
->io_gang_leader
== NULL
) {
2862 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2863 zio
->io_gang_leader
= zio
;
2866 ASSERT(BP_IS_HOLE(bp
));
2867 ASSERT0(BP_GET_NDVAS(bp
));
2868 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2869 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2870 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2872 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2873 flags
|= METASLAB_DONT_THROTTLE
;
2875 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
) {
2876 flags
|= METASLAB_GANG_CHILD
;
2878 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
) {
2879 flags
|= METASLAB_ASYNC_ALLOC
;
2882 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2883 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
2884 &zio
->io_alloc_list
, zio
);
2887 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
2888 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
2890 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2891 return (zio_write_gang_block(zio
));
2892 zio
->io_error
= error
;
2895 return (ZIO_PIPELINE_CONTINUE
);
2899 zio_dva_free(zio_t
*zio
)
2901 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2903 return (ZIO_PIPELINE_CONTINUE
);
2907 zio_dva_claim(zio_t
*zio
)
2911 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2913 zio
->io_error
= error
;
2915 return (ZIO_PIPELINE_CONTINUE
);
2919 * Undo an allocation. This is used by zio_done() when an I/O fails
2920 * and we want to give back the block we just allocated.
2921 * This handles both normal blocks and gang blocks.
2924 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2926 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2927 ASSERT(zio
->io_bp_override
== NULL
);
2929 if (!BP_IS_HOLE(bp
))
2930 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2933 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2934 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2935 &gn
->gn_gbh
->zg_blkptr
[g
]);
2941 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2944 zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
, blkptr_t
*old_bp
,
2945 uint64_t size
, boolean_t
*slog
)
2948 zio_alloc_list_t io_alloc_list
;
2950 ASSERT(txg
> spa_syncing_txg(spa
));
2952 metaslab_trace_init(&io_alloc_list
);
2953 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
2954 txg
, old_bp
, METASLAB_HINTBP_AVOID
, &io_alloc_list
, NULL
);
2958 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
2959 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
,
2960 &io_alloc_list
, NULL
);
2964 metaslab_trace_fini(&io_alloc_list
);
2967 BP_SET_LSIZE(new_bp
, size
);
2968 BP_SET_PSIZE(new_bp
, size
);
2969 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
2970 BP_SET_CHECKSUM(new_bp
,
2971 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
2972 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
2973 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
2974 BP_SET_LEVEL(new_bp
, 0);
2975 BP_SET_DEDUP(new_bp
, 0);
2976 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
2978 zfs_dbgmsg("%s: zil block allocation failure: "
2979 "size %llu, error %d", spa_name(spa
), size
, error
);
2986 * Free an intent log block.
2989 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
2991 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
2992 ASSERT(!BP_IS_GANG(bp
));
2994 zio_free(spa
, txg
, bp
);
2998 * ==========================================================================
2999 * Read and write to physical devices
3000 * ==========================================================================
3005 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3006 * stops after this stage and will resume upon I/O completion.
3007 * However, there are instances where the vdev layer may need to
3008 * continue the pipeline when an I/O was not issued. Since the I/O
3009 * that was sent to the vdev layer might be different than the one
3010 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3011 * force the underlying vdev layers to call either zio_execute() or
3012 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3015 zio_vdev_io_start(zio_t
*zio
)
3017 vdev_t
*vd
= zio
->io_vd
;
3019 spa_t
*spa
= zio
->io_spa
;
3021 ASSERT(zio
->io_error
== 0);
3022 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3025 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3026 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3029 * The mirror_ops handle multiple DVAs in a single BP.
3031 vdev_mirror_ops
.vdev_op_io_start(zio
);
3032 return (ZIO_PIPELINE_STOP
);
3035 ASSERT3P(zio
->io_logical
, !=, zio
);
3036 if (zio
->io_type
== ZIO_TYPE_WRITE
&& zio
->io_vd
->vdev_removing
) {
3037 ASSERT(zio
->io_flags
&
3038 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3039 ZIO_FLAG_INDUCE_DAMAGE
));
3043 * We keep track of time-sensitive I/Os so that the scan thread
3044 * can quickly react to certain workloads. In particular, we care
3045 * about non-scrubbing, top-level reads and writes with the following
3047 * - synchronous writes of user data to non-slog devices
3048 * - any reads of user data
3049 * When these conditions are met, adjust the timestamp of spa_last_io
3050 * which allows the scan thread to adjust its workload accordingly.
3052 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
3053 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
3054 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
3055 zio
->io_txg
!= spa_syncing_txg(spa
)) {
3056 uint64_t old
= spa
->spa_last_io
;
3057 uint64_t new = ddi_get_lbolt64();
3059 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
3062 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3064 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3065 P2PHASE(zio
->io_size
, align
) != 0) {
3066 /* Transform logical writes to be a full physical block size. */
3067 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3068 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3069 ASSERT(vd
== vd
->vdev_top
);
3070 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3071 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3072 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3074 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3078 * If this is not a physical io, make sure that it is properly aligned
3079 * before proceeding.
3081 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3082 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3083 ASSERT0(P2PHASE(zio
->io_size
, align
));
3086 * For physical writes, we allow 512b aligned writes and assume
3087 * the device will perform a read-modify-write as necessary.
3089 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3090 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3093 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3096 * If this is a repair I/O, and there's no self-healing involved --
3097 * that is, we're just resilvering what we expect to resilver --
3098 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3099 * This prevents spurious resilvering with nested replication.
3100 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3101 * A is out of date, we'll read from C+D, then use the data to
3102 * resilver A+B -- but we don't actually want to resilver B, just A.
3103 * The top-level mirror has no way to know this, so instead we just
3104 * discard unnecessary repairs as we work our way down the vdev tree.
3105 * The same logic applies to any form of nested replication:
3106 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3108 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3109 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3110 zio
->io_txg
!= 0 && /* not a delegated i/o */
3111 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3112 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3113 zio_vdev_io_bypass(zio
);
3114 return (ZIO_PIPELINE_CONTINUE
);
3117 if (vd
->vdev_ops
->vdev_op_leaf
&&
3118 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3120 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3121 return (ZIO_PIPELINE_CONTINUE
);
3123 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3124 return (ZIO_PIPELINE_STOP
);
3126 if (!vdev_accessible(vd
, zio
)) {
3127 zio
->io_error
= SET_ERROR(ENXIO
);
3129 return (ZIO_PIPELINE_STOP
);
3133 vd
->vdev_ops
->vdev_op_io_start(zio
);
3134 return (ZIO_PIPELINE_STOP
);
3138 zio_vdev_io_done(zio_t
*zio
)
3140 vdev_t
*vd
= zio
->io_vd
;
3141 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3142 boolean_t unexpected_error
= B_FALSE
;
3144 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3145 return (ZIO_PIPELINE_STOP
);
3147 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3149 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3151 vdev_queue_io_done(zio
);
3153 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3154 vdev_cache_write(zio
);
3156 if (zio_injection_enabled
&& zio
->io_error
== 0)
3157 zio
->io_error
= zio_handle_device_injection(vd
,
3160 if (zio_injection_enabled
&& zio
->io_error
== 0)
3161 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3163 if (zio
->io_error
) {
3164 if (!vdev_accessible(vd
, zio
)) {
3165 zio
->io_error
= SET_ERROR(ENXIO
);
3167 unexpected_error
= B_TRUE
;
3172 ops
->vdev_op_io_done(zio
);
3174 if (unexpected_error
)
3175 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3177 return (ZIO_PIPELINE_CONTINUE
);
3181 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3182 * disk, and use that to finish the checksum ereport later.
3185 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3186 const void *good_buf
)
3188 /* no processing needed */
3189 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3194 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3196 void *buf
= zio_buf_alloc(zio
->io_size
);
3198 abd_copy_to_buf(buf
, zio
->io_abd
, zio
->io_size
);
3200 zcr
->zcr_cbinfo
= zio
->io_size
;
3201 zcr
->zcr_cbdata
= buf
;
3202 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3203 zcr
->zcr_free
= zio_buf_free
;
3207 zio_vdev_io_assess(zio_t
*zio
)
3209 vdev_t
*vd
= zio
->io_vd
;
3211 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3212 return (ZIO_PIPELINE_STOP
);
3214 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3215 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3217 if (zio
->io_vsd
!= NULL
) {
3218 zio
->io_vsd_ops
->vsd_free(zio
);
3222 if (zio_injection_enabled
&& zio
->io_error
== 0)
3223 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3226 * If the I/O failed, determine whether we should attempt to retry it.
3228 * On retry, we cut in line in the issue queue, since we don't want
3229 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3231 if (zio
->io_error
&& vd
== NULL
&&
3232 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3233 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3234 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3236 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3237 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3238 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3239 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3240 zio_requeue_io_start_cut_in_line
);
3241 return (ZIO_PIPELINE_STOP
);
3245 * If we got an error on a leaf device, convert it to ENXIO
3246 * if the device is not accessible at all.
3248 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3249 !vdev_accessible(vd
, zio
))
3250 zio
->io_error
= SET_ERROR(ENXIO
);
3253 * If we can't write to an interior vdev (mirror or RAID-Z),
3254 * set vdev_cant_write so that we stop trying to allocate from it.
3256 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3257 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3258 vd
->vdev_cant_write
= B_TRUE
;
3262 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3263 * attempts will ever succeed. In this case we set a persistent bit so
3264 * that we don't bother with it in the future.
3266 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3267 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3268 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3269 vd
->vdev_nowritecache
= B_TRUE
;
3272 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3274 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3275 zio
->io_physdone
!= NULL
) {
3276 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3277 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3278 zio
->io_physdone(zio
->io_logical
);
3281 return (ZIO_PIPELINE_CONTINUE
);
3285 zio_vdev_io_reissue(zio_t
*zio
)
3287 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3288 ASSERT(zio
->io_error
== 0);
3290 zio
->io_stage
>>= 1;
3294 zio_vdev_io_redone(zio_t
*zio
)
3296 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3298 zio
->io_stage
>>= 1;
3302 zio_vdev_io_bypass(zio_t
*zio
)
3304 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3305 ASSERT(zio
->io_error
== 0);
3307 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3308 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3312 * ==========================================================================
3313 * Generate and verify checksums
3314 * ==========================================================================
3317 zio_checksum_generate(zio_t
*zio
)
3319 blkptr_t
*bp
= zio
->io_bp
;
3320 enum zio_checksum checksum
;
3324 * This is zio_write_phys().
3325 * We're either generating a label checksum, or none at all.
3327 checksum
= zio
->io_prop
.zp_checksum
;
3329 if (checksum
== ZIO_CHECKSUM_OFF
)
3330 return (ZIO_PIPELINE_CONTINUE
);
3332 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3334 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3335 ASSERT(!IO_IS_ALLOCATING(zio
));
3336 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3338 checksum
= BP_GET_CHECKSUM(bp
);
3342 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3344 return (ZIO_PIPELINE_CONTINUE
);
3348 zio_checksum_verify(zio_t
*zio
)
3350 zio_bad_cksum_t info
;
3351 blkptr_t
*bp
= zio
->io_bp
;
3354 ASSERT(zio
->io_vd
!= NULL
);
3358 * This is zio_read_phys().
3359 * We're either verifying a label checksum, or nothing at all.
3361 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3362 return (ZIO_PIPELINE_CONTINUE
);
3364 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3367 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3368 zio
->io_error
= error
;
3369 if (error
== ECKSUM
&&
3370 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3371 zfs_ereport_start_checksum(zio
->io_spa
,
3372 zio
->io_vd
, zio
, zio
->io_offset
,
3373 zio
->io_size
, NULL
, &info
);
3377 return (ZIO_PIPELINE_CONTINUE
);
3381 * Called by RAID-Z to ensure we don't compute the checksum twice.
3384 zio_checksum_verified(zio_t
*zio
)
3386 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3390 * ==========================================================================
3391 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3392 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3393 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3394 * indicate errors that are specific to one I/O, and most likely permanent.
3395 * Any other error is presumed to be worse because we weren't expecting it.
3396 * ==========================================================================
3399 zio_worst_error(int e1
, int e2
)
3401 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3404 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3405 if (e1
== zio_error_rank
[r1
])
3408 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3409 if (e2
== zio_error_rank
[r2
])
3412 return (r1
> r2
? e1
: e2
);
3416 * ==========================================================================
3418 * ==========================================================================
3421 zio_ready(zio_t
*zio
)
3423 blkptr_t
*bp
= zio
->io_bp
;
3424 zio_t
*pio
, *pio_next
;
3425 zio_link_t
*zl
= NULL
;
3427 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
3428 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
3429 return (ZIO_PIPELINE_STOP
);
3431 if (zio
->io_ready
) {
3432 ASSERT(IO_IS_ALLOCATING(zio
));
3433 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3434 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3435 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3440 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3441 zio
->io_bp_copy
= *bp
;
3443 if (zio
->io_error
!= 0) {
3444 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3446 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3447 ASSERT(IO_IS_ALLOCATING(zio
));
3448 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3450 * We were unable to allocate anything, unreserve and
3451 * issue the next I/O to allocate.
3453 metaslab_class_throttle_unreserve(
3454 spa_normal_class(zio
->io_spa
),
3455 zio
->io_prop
.zp_copies
, zio
);
3456 zio_allocate_dispatch(zio
->io_spa
);
3460 mutex_enter(&zio
->io_lock
);
3461 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3462 pio
= zio_walk_parents(zio
, &zl
);
3463 mutex_exit(&zio
->io_lock
);
3466 * As we notify zio's parents, new parents could be added.
3467 * New parents go to the head of zio's io_parent_list, however,
3468 * so we will (correctly) not notify them. The remainder of zio's
3469 * io_parent_list, from 'pio_next' onward, cannot change because
3470 * all parents must wait for us to be done before they can be done.
3472 for (; pio
!= NULL
; pio
= pio_next
) {
3473 pio_next
= zio_walk_parents(zio
, &zl
);
3474 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3477 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3478 if (BP_IS_GANG(bp
)) {
3479 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3481 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3482 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3486 if (zio_injection_enabled
&&
3487 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3488 zio_handle_ignored_writes(zio
);
3490 return (ZIO_PIPELINE_CONTINUE
);
3494 * Update the allocation throttle accounting.
3497 zio_dva_throttle_done(zio_t
*zio
)
3499 zio_t
*lio
= zio
->io_logical
;
3500 zio_t
*pio
= zio_unique_parent(zio
);
3501 vdev_t
*vd
= zio
->io_vd
;
3502 int flags
= METASLAB_ASYNC_ALLOC
;
3504 ASSERT3P(zio
->io_bp
, !=, NULL
);
3505 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3506 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3507 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3509 ASSERT3P(vd
, ==, vd
->vdev_top
);
3510 ASSERT(!(zio
->io_flags
& (ZIO_FLAG_IO_REPAIR
| ZIO_FLAG_IO_RETRY
)));
3511 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3512 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3513 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3516 * Parents of gang children can have two flavors -- ones that
3517 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3518 * and ones that allocated the constituent blocks. The allocation
3519 * throttle needs to know the allocating parent zio so we must find
3522 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3524 * If our parent is a rewrite gang child then our grandparent
3525 * would have been the one that performed the allocation.
3527 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3528 pio
= zio_unique_parent(pio
);
3529 flags
|= METASLAB_GANG_CHILD
;
3532 ASSERT(IO_IS_ALLOCATING(pio
));
3533 ASSERT3P(zio
, !=, zio
->io_logical
);
3534 ASSERT(zio
->io_logical
!= NULL
);
3535 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3536 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3538 mutex_enter(&pio
->io_lock
);
3539 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
3540 mutex_exit(&pio
->io_lock
);
3542 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
3546 * Call into the pipeline to see if there is more work that
3547 * needs to be done. If there is work to be done it will be
3548 * dispatched to another taskq thread.
3550 zio_allocate_dispatch(zio
->io_spa
);
3554 zio_done(zio_t
*zio
)
3556 spa_t
*spa
= zio
->io_spa
;
3557 zio_t
*lio
= zio
->io_logical
;
3558 blkptr_t
*bp
= zio
->io_bp
;
3559 vdev_t
*vd
= zio
->io_vd
;
3560 uint64_t psize
= zio
->io_size
;
3561 zio_t
*pio
, *pio_next
;
3562 metaslab_class_t
*mc
= spa_normal_class(spa
);
3563 zio_link_t
*zl
= NULL
;
3566 * If our children haven't all completed,
3567 * wait for them and then repeat this pipeline stage.
3569 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
3570 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
3571 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
3572 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
3573 return (ZIO_PIPELINE_STOP
);
3576 * If the allocation throttle is enabled, then update the accounting.
3577 * We only track child I/Os that are part of an allocating async
3578 * write. We must do this since the allocation is performed
3579 * by the logical I/O but the actual write is done by child I/Os.
3581 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
3582 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
3583 ASSERT(mc
->mc_alloc_throttle_enabled
);
3584 zio_dva_throttle_done(zio
);
3588 * If the allocation throttle is enabled, verify that
3589 * we have decremented the refcounts for every I/O that was throttled.
3591 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3592 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3593 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3595 metaslab_group_alloc_verify(spa
, zio
->io_bp
, zio
);
3596 VERIFY(refcount_not_held(&mc
->mc_alloc_slots
, zio
));
3599 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
3600 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
3601 ASSERT(zio
->io_children
[c
][w
] == 0);
3603 if (bp
!= NULL
&& !BP_IS_EMBEDDED(bp
)) {
3604 ASSERT(bp
->blk_pad
[0] == 0);
3605 ASSERT(bp
->blk_pad
[1] == 0);
3606 ASSERT(bcmp(bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
3607 (bp
== zio_unique_parent(zio
)->io_bp
));
3608 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(bp
) &&
3609 zio
->io_bp_override
== NULL
&&
3610 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
3611 ASSERT(!BP_SHOULD_BYTESWAP(bp
));
3612 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(bp
));
3613 ASSERT(BP_COUNT_GANG(bp
) == 0 ||
3614 (BP_COUNT_GANG(bp
) == BP_GET_NDVAS(bp
)));
3616 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
3617 VERIFY(BP_EQUAL(bp
, &zio
->io_bp_orig
));
3621 * If there were child vdev/gang/ddt errors, they apply to us now.
3623 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
3624 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
3625 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
3628 * If the I/O on the transformed data was successful, generate any
3629 * checksum reports now while we still have the transformed data.
3631 if (zio
->io_error
== 0) {
3632 while (zio
->io_cksum_report
!= NULL
) {
3633 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3634 uint64_t align
= zcr
->zcr_align
;
3635 uint64_t asize
= P2ROUNDUP(psize
, align
);
3637 abd_t
*adata
= zio
->io_abd
;
3639 if (asize
!= psize
) {
3640 adata
= abd_alloc_linear(asize
, B_TRUE
);
3641 abd_copy(adata
, zio
->io_abd
, psize
);
3642 abd_zero_off(adata
, psize
, asize
- psize
);
3646 abuf
= abd_borrow_buf_copy(adata
, asize
);
3648 zio
->io_cksum_report
= zcr
->zcr_next
;
3649 zcr
->zcr_next
= NULL
;
3650 zcr
->zcr_finish(zcr
, abuf
);
3651 zfs_ereport_free_checksum(zcr
);
3654 abd_return_buf(adata
, abuf
, asize
);
3661 zio_pop_transforms(zio
); /* note: may set zio->io_error */
3663 vdev_stat_update(zio
, psize
);
3665 if (zio
->io_error
) {
3667 * If this I/O is attached to a particular vdev,
3668 * generate an error message describing the I/O failure
3669 * at the block level. We ignore these errors if the
3670 * device is currently unavailable.
3672 if (zio
->io_error
!= ECKSUM
&& vd
!= NULL
&& !vdev_is_dead(vd
))
3673 zfs_ereport_post(FM_EREPORT_ZFS_IO
, spa
, vd
, zio
, 0, 0);
3675 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
3676 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
3679 * For logical I/O requests, tell the SPA to log the
3680 * error and generate a logical data ereport.
3682 spa_log_error(spa
, zio
);
3683 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, spa
, NULL
, zio
,
3688 if (zio
->io_error
&& zio
== lio
) {
3690 * Determine whether zio should be reexecuted. This will
3691 * propagate all the way to the root via zio_notify_parent().
3693 ASSERT(vd
== NULL
&& bp
!= NULL
);
3694 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3696 if (IO_IS_ALLOCATING(zio
) &&
3697 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
3698 if (zio
->io_error
!= ENOSPC
)
3699 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
3701 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3704 if ((zio
->io_type
== ZIO_TYPE_READ
||
3705 zio
->io_type
== ZIO_TYPE_FREE
) &&
3706 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
3707 zio
->io_error
== ENXIO
&&
3708 spa_load_state(spa
) == SPA_LOAD_NONE
&&
3709 spa_get_failmode(spa
) != ZIO_FAILURE_MODE_CONTINUE
)
3710 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3712 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
3713 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3716 * Here is a possibly good place to attempt to do
3717 * either combinatorial reconstruction or error correction
3718 * based on checksums. It also might be a good place
3719 * to send out preliminary ereports before we suspend
3725 * If there were logical child errors, they apply to us now.
3726 * We defer this until now to avoid conflating logical child
3727 * errors with errors that happened to the zio itself when
3728 * updating vdev stats and reporting FMA events above.
3730 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
3732 if ((zio
->io_error
|| zio
->io_reexecute
) &&
3733 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
3734 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
3735 zio_dva_unallocate(zio
, zio
->io_gang_tree
, bp
);
3737 zio_gang_tree_free(&zio
->io_gang_tree
);
3740 * Godfather I/Os should never suspend.
3742 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3743 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
3744 zio
->io_reexecute
= 0;
3746 if (zio
->io_reexecute
) {
3748 * This is a logical I/O that wants to reexecute.
3750 * Reexecute is top-down. When an i/o fails, if it's not
3751 * the root, it simply notifies its parent and sticks around.
3752 * The parent, seeing that it still has children in zio_done(),
3753 * does the same. This percolates all the way up to the root.
3754 * The root i/o will reexecute or suspend the entire tree.
3756 * This approach ensures that zio_reexecute() honors
3757 * all the original i/o dependency relationships, e.g.
3758 * parents not executing until children are ready.
3760 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3762 zio
->io_gang_leader
= NULL
;
3764 mutex_enter(&zio
->io_lock
);
3765 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3766 mutex_exit(&zio
->io_lock
);
3769 * "The Godfather" I/O monitors its children but is
3770 * not a true parent to them. It will track them through
3771 * the pipeline but severs its ties whenever they get into
3772 * trouble (e.g. suspended). This allows "The Godfather"
3773 * I/O to return status without blocking.
3776 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
3778 zio_link_t
*remove_zl
= zl
;
3779 pio_next
= zio_walk_parents(zio
, &zl
);
3781 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3782 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
3783 zio_remove_child(pio
, zio
, remove_zl
);
3784 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3788 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
3790 * We're not a root i/o, so there's nothing to do
3791 * but notify our parent. Don't propagate errors
3792 * upward since we haven't permanently failed yet.
3794 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
3795 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
3796 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3797 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
3799 * We'd fail again if we reexecuted now, so suspend
3800 * until conditions improve (e.g. device comes online).
3802 zio_suspend(spa
, zio
);
3805 * Reexecution is potentially a huge amount of work.
3806 * Hand it off to the otherwise-unused claim taskq.
3808 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
3809 spa_taskq_dispatch_ent(spa
, ZIO_TYPE_CLAIM
,
3810 ZIO_TASKQ_ISSUE
, (task_func_t
*)zio_reexecute
, zio
,
3813 return (ZIO_PIPELINE_STOP
);
3816 ASSERT(zio
->io_child_count
== 0);
3817 ASSERT(zio
->io_reexecute
== 0);
3818 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
3821 * Report any checksum errors, since the I/O is complete.
3823 while (zio
->io_cksum_report
!= NULL
) {
3824 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3825 zio
->io_cksum_report
= zcr
->zcr_next
;
3826 zcr
->zcr_next
= NULL
;
3827 zcr
->zcr_finish(zcr
, NULL
);
3828 zfs_ereport_free_checksum(zcr
);
3832 * It is the responsibility of the done callback to ensure that this
3833 * particular zio is no longer discoverable for adoption, and as
3834 * such, cannot acquire any new parents.
3839 mutex_enter(&zio
->io_lock
);
3840 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3841 mutex_exit(&zio
->io_lock
);
3844 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
3845 zio_link_t
*remove_zl
= zl
;
3846 pio_next
= zio_walk_parents(zio
, &zl
);
3847 zio_remove_child(pio
, zio
, remove_zl
);
3848 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3851 if (zio
->io_waiter
!= NULL
) {
3852 mutex_enter(&zio
->io_lock
);
3853 zio
->io_executor
= NULL
;
3854 cv_broadcast(&zio
->io_cv
);
3855 mutex_exit(&zio
->io_lock
);
3860 return (ZIO_PIPELINE_STOP
);
3864 * ==========================================================================
3865 * I/O pipeline definition
3866 * ==========================================================================
3868 static zio_pipe_stage_t
*zio_pipeline
[] = {
3875 zio_checksum_generate
,
3891 zio_checksum_verify
,
3899 * Compare two zbookmark_phys_t's to see which we would reach first in a
3900 * pre-order traversal of the object tree.
3902 * This is simple in every case aside from the meta-dnode object. For all other
3903 * objects, we traverse them in order (object 1 before object 2, and so on).
3904 * However, all of these objects are traversed while traversing object 0, since
3905 * the data it points to is the list of objects. Thus, we need to convert to a
3906 * canonical representation so we can compare meta-dnode bookmarks to
3907 * non-meta-dnode bookmarks.
3909 * We do this by calculating "equivalents" for each field of the zbookmark.
3910 * zbookmarks outside of the meta-dnode use their own object and level, and
3911 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3912 * blocks this bookmark refers to) by multiplying their blkid by their span
3913 * (the number of L0 blocks contained within one block at their level).
3914 * zbookmarks inside the meta-dnode calculate their object equivalent
3915 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3916 * level + 1<<31 (any value larger than a level could ever be) for their level.
3917 * This causes them to always compare before a bookmark in their object
3918 * equivalent, compare appropriately to bookmarks in other objects, and to
3919 * compare appropriately to other bookmarks in the meta-dnode.
3922 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
3923 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
3926 * These variables represent the "equivalent" values for the zbookmark,
3927 * after converting zbookmarks inside the meta dnode to their
3928 * normal-object equivalents.
3930 uint64_t zb1obj
, zb2obj
;
3931 uint64_t zb1L0
, zb2L0
;
3932 uint64_t zb1level
, zb2level
;
3934 if (zb1
->zb_object
== zb2
->zb_object
&&
3935 zb1
->zb_level
== zb2
->zb_level
&&
3936 zb1
->zb_blkid
== zb2
->zb_blkid
)
3940 * BP_SPANB calculates the span in blocks.
3942 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
3943 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
3945 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
3946 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3948 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
3950 zb1obj
= zb1
->zb_object
;
3951 zb1level
= zb1
->zb_level
;
3954 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
3955 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3957 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
3959 zb2obj
= zb2
->zb_object
;
3960 zb2level
= zb2
->zb_level
;
3963 /* Now that we have a canonical representation, do the comparison. */
3964 if (zb1obj
!= zb2obj
)
3965 return (zb1obj
< zb2obj
? -1 : 1);
3966 else if (zb1L0
!= zb2L0
)
3967 return (zb1L0
< zb2L0
? -1 : 1);
3968 else if (zb1level
!= zb2level
)
3969 return (zb1level
> zb2level
? -1 : 1);
3971 * This can (theoretically) happen if the bookmarks have the same object
3972 * and level, but different blkids, if the block sizes are not the same.
3973 * There is presently no way to change the indirect block sizes
3979 * This function checks the following: given that last_block is the place that
3980 * our traversal stopped last time, does that guarantee that we've visited
3981 * every node under subtree_root? Therefore, we can't just use the raw output
3982 * of zbookmark_compare. We have to pass in a modified version of
3983 * subtree_root; by incrementing the block id, and then checking whether
3984 * last_block is before or equal to that, we can tell whether or not having
3985 * visited last_block implies that all of subtree_root's children have been
3989 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
3990 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
3992 zbookmark_phys_t mod_zb
= *subtree_root
;
3994 ASSERT(last_block
->zb_level
== 0);
3996 /* The objset_phys_t isn't before anything. */
4001 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4002 * data block size in sectors, because that variable is only used if
4003 * the bookmark refers to a block in the meta-dnode. Since we don't
4004 * know without examining it what object it refers to, and there's no
4005 * harm in passing in this value in other cases, we always pass it in.
4007 * We pass in 0 for the indirect block size shift because zb2 must be
4008 * level 0. The indirect block size is only used to calculate the span
4009 * of the bookmark, but since the bookmark must be level 0, the span is
4010 * always 1, so the math works out.
4012 * If you make changes to how the zbookmark_compare code works, be sure
4013 * to make sure that this code still works afterwards.
4015 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4016 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,