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
)
837 * The check for EMBEDDED is a performance optimization. We
838 * process the free here (by ignoring it) rather than
839 * putting it on the list and then processing it in zio_free_sync().
841 if (BP_IS_EMBEDDED(bp
))
843 metaslab_check_free(spa
, bp
);
846 * Frees that are for the currently-syncing txg, are not going to be
847 * deferred, and which will not need to do a read (i.e. not GANG or
848 * DEDUP), can be processed immediately. Otherwise, put them on the
849 * in-memory list for later processing.
851 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
852 txg
!= spa
->spa_syncing_txg
||
853 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
854 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
856 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
861 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
865 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
867 ASSERT(!BP_IS_HOLE(bp
));
868 ASSERT(spa_syncing_txg(spa
) == txg
);
869 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
871 if (BP_IS_EMBEDDED(bp
))
872 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
874 metaslab_check_free(spa
, bp
);
878 * GANG and DEDUP blocks can induce a read (for the gang block header,
879 * or the DDT), so issue them asynchronously so that this thread is
882 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
883 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
885 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
886 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
887 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
893 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
894 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
898 dprintf_bp(bp
, "claiming in txg %llu", txg
);
900 if (BP_IS_EMBEDDED(bp
))
901 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
904 * A claim is an allocation of a specific block. Claims are needed
905 * to support immediate writes in the intent log. The issue is that
906 * immediate writes contain committed data, but in a txg that was
907 * *not* committed. Upon opening the pool after an unclean shutdown,
908 * the intent log claims all blocks that contain immediate write data
909 * so that the SPA knows they're in use.
911 * All claims *must* be resolved in the first txg -- before the SPA
912 * starts allocating blocks -- so that nothing is allocated twice.
913 * If txg == 0 we just verify that the block is claimable.
915 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
916 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
917 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(8) */
919 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
920 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
921 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
922 ASSERT0(zio
->io_queued_timestamp
);
928 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
929 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
934 if (vd
->vdev_children
== 0) {
935 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
936 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
937 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
941 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
943 for (c
= 0; c
< vd
->vdev_children
; c
++)
944 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
945 done
, private, flags
));
952 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
953 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
954 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
958 ASSERT(vd
->vdev_children
== 0);
959 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
960 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
961 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
963 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
964 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
965 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
967 zio
->io_prop
.zp_checksum
= checksum
;
973 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
974 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
975 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
979 ASSERT(vd
->vdev_children
== 0);
980 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
981 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
982 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
984 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
985 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
986 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
988 zio
->io_prop
.zp_checksum
= checksum
;
990 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
992 * zec checksums are necessarily destructive -- they modify
993 * the end of the write buffer to hold the verifier/checksum.
994 * Therefore, we must make a local copy in case the data is
995 * being written to multiple places in parallel.
997 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
998 abd_copy(wbuf
, data
, size
);
1000 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1007 * Create a child I/O to do some work for us.
1010 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1011 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1012 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1014 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1017 ASSERT(vd
->vdev_parent
==
1018 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
1020 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1022 * If we have the bp, then the child should perform the
1023 * checksum and the parent need not. This pushes error
1024 * detection as close to the leaves as possible and
1025 * eliminates redundant checksums in the interior nodes.
1027 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1028 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1031 if (vd
->vdev_children
== 0)
1032 offset
+= VDEV_LABEL_START_SIZE
;
1034 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
1037 * If we've decided to do a repair, the write is not speculative --
1038 * even if the original read was.
1040 if (flags
& ZIO_FLAG_IO_REPAIR
)
1041 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1044 * If we're creating a child I/O that is not associated with a
1045 * top-level vdev, then the child zio is not an allocating I/O.
1046 * If this is a retried I/O then we ignore it since we will
1047 * have already processed the original allocating I/O.
1049 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1050 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1051 metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
);
1053 ASSERT(mc
->mc_alloc_throttle_enabled
);
1054 ASSERT(type
== ZIO_TYPE_WRITE
);
1055 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1056 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1057 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1058 pio
->io_child_type
== ZIO_CHILD_GANG
);
1060 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1063 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1064 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1065 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1066 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1068 zio
->io_physdone
= pio
->io_physdone
;
1069 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1070 zio
->io_logical
->io_phys_children
++;
1076 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1077 int type
, zio_priority_t priority
, enum zio_flag flags
,
1078 zio_done_func_t
*done
, void *private)
1082 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1084 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1085 data
, size
, size
, done
, private, type
, priority
,
1086 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1088 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1094 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1096 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1098 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1102 zio_shrink(zio_t
*zio
, uint64_t size
)
1104 ASSERT3P(zio
->io_executor
, ==, NULL
);
1105 ASSERT3P(zio
->io_orig_size
, ==, zio
->io_size
);
1106 ASSERT3U(size
, <=, zio
->io_size
);
1109 * We don't shrink for raidz because of problems with the
1110 * reconstruction when reading back less than the block size.
1111 * Note, BP_IS_RAIDZ() assumes no compression.
1113 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1114 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1115 /* we are not doing a raw write */
1116 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1117 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1122 * ==========================================================================
1123 * Prepare to read and write logical blocks
1124 * ==========================================================================
1128 zio_read_bp_init(zio_t
*zio
)
1130 blkptr_t
*bp
= zio
->io_bp
;
1132 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1133 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1134 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
1136 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1137 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1138 psize
, psize
, zio_decompress
);
1141 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1142 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1144 int psize
= BPE_GET_PSIZE(bp
);
1145 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1146 decode_embedded_bp_compressed(bp
, data
);
1147 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1149 ASSERT(!BP_IS_EMBEDDED(bp
));
1152 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1153 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1155 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1156 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1158 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1159 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1161 return (ZIO_PIPELINE_CONTINUE
);
1165 zio_write_bp_init(zio_t
*zio
)
1167 if (!IO_IS_ALLOCATING(zio
))
1168 return (ZIO_PIPELINE_CONTINUE
);
1170 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1172 if (zio
->io_bp_override
) {
1173 blkptr_t
*bp
= zio
->io_bp
;
1174 zio_prop_t
*zp
= &zio
->io_prop
;
1176 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1177 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1179 *bp
= *zio
->io_bp_override
;
1180 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1182 if (BP_IS_EMBEDDED(bp
))
1183 return (ZIO_PIPELINE_CONTINUE
);
1186 * If we've been overridden and nopwrite is set then
1187 * set the flag accordingly to indicate that a nopwrite
1188 * has already occurred.
1190 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1191 ASSERT(!zp
->zp_dedup
);
1192 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1193 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1194 return (ZIO_PIPELINE_CONTINUE
);
1197 ASSERT(!zp
->zp_nopwrite
);
1199 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1200 return (ZIO_PIPELINE_CONTINUE
);
1202 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1203 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1205 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
1206 BP_SET_DEDUP(bp
, 1);
1207 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1208 return (ZIO_PIPELINE_CONTINUE
);
1212 * We were unable to handle this as an override bp, treat
1213 * it as a regular write I/O.
1215 zio
->io_bp_override
= NULL
;
1216 *bp
= zio
->io_bp_orig
;
1217 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1220 return (ZIO_PIPELINE_CONTINUE
);
1224 zio_write_compress(zio_t
*zio
)
1226 spa_t
*spa
= zio
->io_spa
;
1227 zio_prop_t
*zp
= &zio
->io_prop
;
1228 enum zio_compress compress
= zp
->zp_compress
;
1229 blkptr_t
*bp
= zio
->io_bp
;
1230 uint64_t lsize
= zio
->io_lsize
;
1231 uint64_t psize
= zio
->io_size
;
1234 EQUIV(lsize
!= psize
, (zio
->io_flags
& ZIO_FLAG_RAW
) != 0);
1237 * If our children haven't all reached the ready stage,
1238 * wait for them and then repeat this pipeline stage.
1240 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1241 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1242 return (ZIO_PIPELINE_STOP
);
1244 if (!IO_IS_ALLOCATING(zio
))
1245 return (ZIO_PIPELINE_CONTINUE
);
1247 if (zio
->io_children_ready
!= NULL
) {
1249 * Now that all our children are ready, run the callback
1250 * associated with this zio in case it wants to modify the
1251 * data to be written.
1253 ASSERT3U(zp
->zp_level
, >, 0);
1254 zio
->io_children_ready(zio
);
1257 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1258 ASSERT(zio
->io_bp_override
== NULL
);
1260 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1262 * We're rewriting an existing block, which means we're
1263 * working on behalf of spa_sync(). For spa_sync() to
1264 * converge, it must eventually be the case that we don't
1265 * have to allocate new blocks. But compression changes
1266 * the blocksize, which forces a reallocate, and makes
1267 * convergence take longer. Therefore, after the first
1268 * few passes, stop compressing to ensure convergence.
1270 pass
= spa_sync_pass(spa
);
1272 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1273 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1274 ASSERT(!BP_GET_DEDUP(bp
));
1276 if (pass
>= zfs_sync_pass_dont_compress
)
1277 compress
= ZIO_COMPRESS_OFF
;
1279 /* Make sure someone doesn't change their mind on overwrites */
1280 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1281 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1284 /* If it's a compressed write that is not raw, compress the buffer. */
1285 if (compress
!= ZIO_COMPRESS_OFF
&& psize
== lsize
) {
1286 void *cbuf
= zio_buf_alloc(lsize
);
1287 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1288 if (psize
== 0 || psize
== lsize
) {
1289 compress
= ZIO_COMPRESS_OFF
;
1290 zio_buf_free(cbuf
, lsize
);
1291 } else if (!zp
->zp_dedup
&& psize
<= BPE_PAYLOAD_SIZE
&&
1292 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1293 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1294 encode_embedded_bp_compressed(bp
,
1295 cbuf
, compress
, lsize
, psize
);
1296 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1297 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1298 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1299 zio_buf_free(cbuf
, lsize
);
1300 bp
->blk_birth
= zio
->io_txg
;
1301 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1302 ASSERT(spa_feature_is_active(spa
,
1303 SPA_FEATURE_EMBEDDED_DATA
));
1304 return (ZIO_PIPELINE_CONTINUE
);
1307 * Round up compressed size up to the ashift
1308 * of the smallest-ashift device, and zero the tail.
1309 * This ensures that the compressed size of the BP
1310 * (and thus compressratio property) are correct,
1311 * in that we charge for the padding used to fill out
1314 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1315 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1316 1ULL << spa
->spa_min_ashift
);
1317 if (rounded
>= lsize
) {
1318 compress
= ZIO_COMPRESS_OFF
;
1319 zio_buf_free(cbuf
, lsize
);
1322 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1323 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1324 abd_zero_off(cdata
, psize
, rounded
- psize
);
1326 zio_push_transform(zio
, cdata
,
1327 psize
, lsize
, NULL
);
1332 * We were unable to handle this as an override bp, treat
1333 * it as a regular write I/O.
1335 zio
->io_bp_override
= NULL
;
1336 *bp
= zio
->io_bp_orig
;
1337 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1339 ASSERT3U(psize
, !=, 0);
1343 * The final pass of spa_sync() must be all rewrites, but the first
1344 * few passes offer a trade-off: allocating blocks defers convergence,
1345 * but newly allocated blocks are sequential, so they can be written
1346 * to disk faster. Therefore, we allow the first few passes of
1347 * spa_sync() to allocate new blocks, but force rewrites after that.
1348 * There should only be a handful of blocks after pass 1 in any case.
1350 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1351 BP_GET_PSIZE(bp
) == psize
&&
1352 pass
>= zfs_sync_pass_rewrite
) {
1354 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1355 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1356 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1359 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1363 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1364 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1365 BP_SET_LSIZE(bp
, lsize
);
1366 BP_SET_TYPE(bp
, zp
->zp_type
);
1367 BP_SET_LEVEL(bp
, zp
->zp_level
);
1368 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1370 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1372 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1373 BP_SET_LSIZE(bp
, lsize
);
1374 BP_SET_TYPE(bp
, zp
->zp_type
);
1375 BP_SET_LEVEL(bp
, zp
->zp_level
);
1376 BP_SET_PSIZE(bp
, psize
);
1377 BP_SET_COMPRESS(bp
, compress
);
1378 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1379 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1380 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1382 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1383 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1384 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1386 if (zp
->zp_nopwrite
) {
1387 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1388 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1389 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1392 return (ZIO_PIPELINE_CONTINUE
);
1396 zio_free_bp_init(zio_t
*zio
)
1398 blkptr_t
*bp
= zio
->io_bp
;
1400 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1401 if (BP_GET_DEDUP(bp
))
1402 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1405 return (ZIO_PIPELINE_CONTINUE
);
1409 * ==========================================================================
1410 * Execute the I/O pipeline
1411 * ==========================================================================
1415 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1417 spa_t
*spa
= zio
->io_spa
;
1418 zio_type_t t
= zio
->io_type
;
1419 int flags
= (cutinline
? TQ_FRONT
: 0);
1422 * If we're a config writer or a probe, the normal issue and
1423 * interrupt threads may all be blocked waiting for the config lock.
1424 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1426 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1430 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1432 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1436 * If this is a high priority I/O, then use the high priority taskq if
1439 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1440 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1443 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1446 * NB: We are assuming that the zio can only be dispatched
1447 * to a single taskq at a time. It would be a grievous error
1448 * to dispatch the zio to another taskq at the same time.
1450 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
1451 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1452 flags
, &zio
->io_tqent
);
1456 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1458 kthread_t
*executor
= zio
->io_executor
;
1459 spa_t
*spa
= zio
->io_spa
;
1461 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1462 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1464 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1465 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1474 zio_issue_async(zio_t
*zio
)
1476 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1478 return (ZIO_PIPELINE_STOP
);
1482 zio_interrupt(zio_t
*zio
)
1484 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1488 zio_delay_interrupt(zio_t
*zio
)
1491 * The timeout_generic() function isn't defined in userspace, so
1492 * rather than trying to implement the function, the zio delay
1493 * functionality has been disabled for userspace builds.
1498 * If io_target_timestamp is zero, then no delay has been registered
1499 * for this IO, thus jump to the end of this function and "skip" the
1500 * delay; issuing it directly to the zio layer.
1502 if (zio
->io_target_timestamp
!= 0) {
1503 hrtime_t now
= gethrtime();
1505 if (now
>= zio
->io_target_timestamp
) {
1507 * This IO has already taken longer than the target
1508 * delay to complete, so we don't want to delay it
1509 * any longer; we "miss" the delay and issue it
1510 * directly to the zio layer. This is likely due to
1511 * the target latency being set to a value less than
1512 * the underlying hardware can satisfy (e.g. delay
1513 * set to 1ms, but the disks take 10ms to complete an
1517 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1522 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1524 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1525 hrtime_t
, now
, hrtime_t
, diff
);
1527 (void) timeout_generic(CALLOUT_NORMAL
,
1528 (void (*)(void *))zio_interrupt
, zio
, diff
, 1, 0);
1535 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1540 * Execute the I/O pipeline until one of the following occurs:
1542 * (1) the I/O completes
1543 * (2) the pipeline stalls waiting for dependent child I/Os
1544 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1545 * (4) the I/O is delegated by vdev-level caching or aggregation
1546 * (5) the I/O is deferred due to vdev-level queueing
1547 * (6) the I/O is handed off to another thread.
1549 * In all cases, the pipeline stops whenever there's no CPU work; it never
1550 * burns a thread in cv_wait().
1552 * There's no locking on io_stage because there's no legitimate way
1553 * for multiple threads to be attempting to process the same I/O.
1555 static zio_pipe_stage_t
*zio_pipeline
[];
1558 zio_execute(zio_t
*zio
)
1560 zio
->io_executor
= curthread
;
1562 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1564 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1565 enum zio_stage pipeline
= zio
->io_pipeline
;
1566 enum zio_stage stage
= zio
->io_stage
;
1569 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1570 ASSERT(ISP2(stage
));
1571 ASSERT(zio
->io_stall
== NULL
);
1575 } while ((stage
& pipeline
) == 0);
1577 ASSERT(stage
<= ZIO_STAGE_DONE
);
1580 * If we are in interrupt context and this pipeline stage
1581 * will grab a config lock that is held across I/O,
1582 * or may wait for an I/O that needs an interrupt thread
1583 * to complete, issue async to avoid deadlock.
1585 * For VDEV_IO_START, we cut in line so that the io will
1586 * be sent to disk promptly.
1588 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1589 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1590 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1591 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1592 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1596 zio
->io_stage
= stage
;
1597 zio
->io_pipeline_trace
|= zio
->io_stage
;
1598 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1600 if (rv
== ZIO_PIPELINE_STOP
)
1603 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1608 * ==========================================================================
1609 * Initiate I/O, either sync or async
1610 * ==========================================================================
1613 zio_wait(zio_t
*zio
)
1617 ASSERT3P(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1618 ASSERT3P(zio
->io_executor
, ==, NULL
);
1620 zio
->io_waiter
= curthread
;
1621 ASSERT0(zio
->io_queued_timestamp
);
1622 zio
->io_queued_timestamp
= gethrtime();
1626 mutex_enter(&zio
->io_lock
);
1627 while (zio
->io_executor
!= NULL
)
1628 cv_wait(&zio
->io_cv
, &zio
->io_lock
);
1629 mutex_exit(&zio
->io_lock
);
1631 error
= zio
->io_error
;
1638 zio_nowait(zio_t
*zio
)
1640 ASSERT3P(zio
->io_executor
, ==, NULL
);
1642 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1643 zio_unique_parent(zio
) == NULL
) {
1645 * This is a logical async I/O with no parent to wait for it.
1646 * We add it to the spa_async_root_zio "Godfather" I/O which
1647 * will ensure they complete prior to unloading the pool.
1649 spa_t
*spa
= zio
->io_spa
;
1651 zio_add_child(spa
->spa_async_zio_root
[CPU_SEQID
], zio
);
1654 ASSERT0(zio
->io_queued_timestamp
);
1655 zio
->io_queued_timestamp
= gethrtime();
1660 * ==========================================================================
1661 * Reexecute, cancel, or suspend/resume failed I/O
1662 * ==========================================================================
1666 zio_reexecute(zio_t
*pio
)
1668 zio_t
*cio
, *cio_next
;
1670 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1671 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1672 ASSERT(pio
->io_gang_leader
== NULL
);
1673 ASSERT(pio
->io_gang_tree
== NULL
);
1675 pio
->io_flags
= pio
->io_orig_flags
;
1676 pio
->io_stage
= pio
->io_orig_stage
;
1677 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1678 pio
->io_reexecute
= 0;
1679 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1680 pio
->io_pipeline_trace
= 0;
1682 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1683 pio
->io_state
[w
] = 0;
1684 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1685 pio
->io_child_error
[c
] = 0;
1687 if (IO_IS_ALLOCATING(pio
))
1688 BP_ZERO(pio
->io_bp
);
1691 * As we reexecute pio's children, new children could be created.
1692 * New children go to the head of pio's io_child_list, however,
1693 * so we will (correctly) not reexecute them. The key is that
1694 * the remainder of pio's io_child_list, from 'cio_next' onward,
1695 * cannot be affected by any side effects of reexecuting 'cio'.
1697 zio_link_t
*zl
= NULL
;
1698 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1699 cio_next
= zio_walk_children(pio
, &zl
);
1700 mutex_enter(&pio
->io_lock
);
1701 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1702 pio
->io_children
[cio
->io_child_type
][w
]++;
1703 mutex_exit(&pio
->io_lock
);
1708 * Now that all children have been reexecuted, execute the parent.
1709 * We don't reexecute "The Godfather" I/O here as it's the
1710 * responsibility of the caller to wait on it.
1712 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1713 pio
->io_queued_timestamp
= gethrtime();
1719 zio_cancel(zio_t
*zio
)
1722 * Disallow cancellation of a zio that's already been issued.
1724 VERIFY3P(zio
->io_executor
, ==, NULL
);
1726 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1727 zio
->io_done
= NULL
;
1733 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1735 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1736 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1737 "failure and the failure mode property for this pool "
1738 "is set to panic.", spa_name(spa
));
1740 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1742 mutex_enter(&spa
->spa_suspend_lock
);
1744 if (spa
->spa_suspend_zio_root
== NULL
)
1745 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1746 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1747 ZIO_FLAG_GODFATHER
);
1749 spa
->spa_suspended
= B_TRUE
;
1752 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1753 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1754 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1755 ASSERT(zio_unique_parent(zio
) == NULL
);
1756 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1757 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1760 mutex_exit(&spa
->spa_suspend_lock
);
1764 zio_resume(spa_t
*spa
)
1769 * Reexecute all previously suspended i/o.
1771 mutex_enter(&spa
->spa_suspend_lock
);
1772 spa
->spa_suspended
= B_FALSE
;
1773 cv_broadcast(&spa
->spa_suspend_cv
);
1774 pio
= spa
->spa_suspend_zio_root
;
1775 spa
->spa_suspend_zio_root
= NULL
;
1776 mutex_exit(&spa
->spa_suspend_lock
);
1782 return (zio_wait(pio
));
1786 zio_resume_wait(spa_t
*spa
)
1788 mutex_enter(&spa
->spa_suspend_lock
);
1789 while (spa_suspended(spa
))
1790 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1791 mutex_exit(&spa
->spa_suspend_lock
);
1795 * ==========================================================================
1798 * A gang block is a collection of small blocks that looks to the DMU
1799 * like one large block. When zio_dva_allocate() cannot find a block
1800 * of the requested size, due to either severe fragmentation or the pool
1801 * being nearly full, it calls zio_write_gang_block() to construct the
1802 * block from smaller fragments.
1804 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1805 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1806 * an indirect block: it's an array of block pointers. It consumes
1807 * only one sector and hence is allocatable regardless of fragmentation.
1808 * The gang header's bps point to its gang members, which hold the data.
1810 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1811 * as the verifier to ensure uniqueness of the SHA256 checksum.
1812 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1813 * not the gang header. This ensures that data block signatures (needed for
1814 * deduplication) are independent of how the block is physically stored.
1816 * Gang blocks can be nested: a gang member may itself be a gang block.
1817 * Thus every gang block is a tree in which root and all interior nodes are
1818 * gang headers, and the leaves are normal blocks that contain user data.
1819 * The root of the gang tree is called the gang leader.
1821 * To perform any operation (read, rewrite, free, claim) on a gang block,
1822 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1823 * in the io_gang_tree field of the original logical i/o by recursively
1824 * reading the gang leader and all gang headers below it. This yields
1825 * an in-core tree containing the contents of every gang header and the
1826 * bps for every constituent of the gang block.
1828 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1829 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1830 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1831 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1832 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1833 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1834 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1835 * of the gang header plus zio_checksum_compute() of the data to update the
1836 * gang header's blk_cksum as described above.
1838 * The two-phase assemble/issue model solves the problem of partial failure --
1839 * what if you'd freed part of a gang block but then couldn't read the
1840 * gang header for another part? Assembling the entire gang tree first
1841 * ensures that all the necessary gang header I/O has succeeded before
1842 * starting the actual work of free, claim, or write. Once the gang tree
1843 * is assembled, free and claim are in-memory operations that cannot fail.
1845 * In the event that a gang write fails, zio_dva_unallocate() walks the
1846 * gang tree to immediately free (i.e. insert back into the space map)
1847 * everything we've allocated. This ensures that we don't get ENOSPC
1848 * errors during repeated suspend/resume cycles due to a flaky device.
1850 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1851 * the gang tree, we won't modify the block, so we can safely defer the free
1852 * (knowing that the block is still intact). If we *can* assemble the gang
1853 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1854 * each constituent bp and we can allocate a new block on the next sync pass.
1856 * In all cases, the gang tree allows complete recovery from partial failure.
1857 * ==========================================================================
1861 zio_gang_issue_func_done(zio_t
*zio
)
1863 abd_put(zio
->io_abd
);
1867 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1873 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
1874 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
1875 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1876 &pio
->io_bookmark
));
1880 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1887 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1888 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1889 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
1890 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1893 * As we rewrite each gang header, the pipeline will compute
1894 * a new gang block header checksum for it; but no one will
1895 * compute a new data checksum, so we do that here. The one
1896 * exception is the gang leader: the pipeline already computed
1897 * its data checksum because that stage precedes gang assembly.
1898 * (Presently, nothing actually uses interior data checksums;
1899 * this is just good hygiene.)
1901 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1902 abd_t
*buf
= abd_get_offset(data
, offset
);
1904 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1905 buf
, BP_GET_PSIZE(bp
));
1910 * If we are here to damage data for testing purposes,
1911 * leave the GBH alone so that we can detect the damage.
1913 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1914 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1916 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1917 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
1918 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
1919 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1927 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1930 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1931 ZIO_GANG_CHILD_FLAGS(pio
)));
1936 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1939 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1940 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1943 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
1952 static void zio_gang_tree_assemble_done(zio_t
*zio
);
1954 static zio_gang_node_t
*
1955 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
1957 zio_gang_node_t
*gn
;
1959 ASSERT(*gnpp
== NULL
);
1961 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
1962 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
1969 zio_gang_node_free(zio_gang_node_t
**gnpp
)
1971 zio_gang_node_t
*gn
= *gnpp
;
1973 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1974 ASSERT(gn
->gn_child
[g
] == NULL
);
1976 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1977 kmem_free(gn
, sizeof (*gn
));
1982 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
1984 zio_gang_node_t
*gn
= *gnpp
;
1989 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
1990 zio_gang_tree_free(&gn
->gn_child
[g
]);
1992 zio_gang_node_free(gnpp
);
1996 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
1998 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
1999 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2001 ASSERT(gio
->io_gang_leader
== gio
);
2002 ASSERT(BP_IS_GANG(bp
));
2004 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2005 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2006 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2010 zio_gang_tree_assemble_done(zio_t
*zio
)
2012 zio_t
*gio
= zio
->io_gang_leader
;
2013 zio_gang_node_t
*gn
= zio
->io_private
;
2014 blkptr_t
*bp
= zio
->io_bp
;
2016 ASSERT(gio
== zio_unique_parent(zio
));
2017 ASSERT(zio
->io_child_count
== 0);
2022 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2023 if (BP_SHOULD_BYTESWAP(bp
))
2024 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2026 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2027 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2028 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2030 abd_put(zio
->io_abd
);
2032 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2033 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2034 if (!BP_IS_GANG(gbp
))
2036 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2041 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2044 zio_t
*gio
= pio
->io_gang_leader
;
2047 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2048 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2049 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2052 * If you're a gang header, your data is in gn->gn_gbh.
2053 * If you're a gang member, your data is in 'data' and gn == NULL.
2055 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2058 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2060 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2061 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2062 if (BP_IS_HOLE(gbp
))
2064 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2066 offset
+= BP_GET_PSIZE(gbp
);
2070 if (gn
== gio
->io_gang_tree
)
2071 ASSERT3U(gio
->io_size
, ==, offset
);
2078 zio_gang_assemble(zio_t
*zio
)
2080 blkptr_t
*bp
= zio
->io_bp
;
2082 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2083 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2085 zio
->io_gang_leader
= zio
;
2087 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2089 return (ZIO_PIPELINE_CONTINUE
);
2093 zio_gang_issue(zio_t
*zio
)
2095 blkptr_t
*bp
= zio
->io_bp
;
2097 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
2098 return (ZIO_PIPELINE_STOP
);
2100 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2101 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2103 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2104 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2107 zio_gang_tree_free(&zio
->io_gang_tree
);
2109 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2111 return (ZIO_PIPELINE_CONTINUE
);
2115 zio_write_gang_member_ready(zio_t
*zio
)
2117 zio_t
*pio
= zio_unique_parent(zio
);
2118 zio_t
*gio
= zio
->io_gang_leader
;
2119 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2120 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2123 if (BP_IS_HOLE(zio
->io_bp
))
2126 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2128 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2129 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2130 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2131 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2132 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2134 mutex_enter(&pio
->io_lock
);
2135 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2136 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2137 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2138 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2139 DVA_SET_ASIZE(&pdva
[d
], asize
);
2141 mutex_exit(&pio
->io_lock
);
2145 zio_write_gang_done(zio_t
*zio
)
2147 abd_put(zio
->io_abd
);
2151 zio_write_gang_block(zio_t
*pio
)
2153 spa_t
*spa
= pio
->io_spa
;
2154 metaslab_class_t
*mc
= spa_normal_class(spa
);
2155 blkptr_t
*bp
= pio
->io_bp
;
2156 zio_t
*gio
= pio
->io_gang_leader
;
2158 zio_gang_node_t
*gn
, **gnpp
;
2159 zio_gbh_phys_t
*gbh
;
2161 uint64_t txg
= pio
->io_txg
;
2162 uint64_t resid
= pio
->io_size
;
2164 int copies
= gio
->io_prop
.zp_copies
;
2165 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2169 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2170 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2171 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2172 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2174 flags
|= METASLAB_ASYNC_ALLOC
;
2175 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2178 * The logical zio has already placed a reservation for
2179 * 'copies' allocation slots but gang blocks may require
2180 * additional copies. These additional copies
2181 * (i.e. gbh_copies - copies) are guaranteed to succeed
2182 * since metaslab_class_throttle_reserve() always allows
2183 * additional reservations for gang blocks.
2185 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2189 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2190 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2191 &pio
->io_alloc_list
, pio
);
2193 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2194 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2195 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2198 * If we failed to allocate the gang block header then
2199 * we remove any additional allocation reservations that
2200 * we placed here. The original reservation will
2201 * be removed when the logical I/O goes to the ready
2204 metaslab_class_throttle_unreserve(mc
,
2205 gbh_copies
- copies
, pio
);
2207 pio
->io_error
= error
;
2208 return (ZIO_PIPELINE_CONTINUE
);
2212 gnpp
= &gio
->io_gang_tree
;
2214 gnpp
= pio
->io_private
;
2215 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2218 gn
= zio_gang_node_alloc(gnpp
);
2220 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2221 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2224 * Create the gang header.
2226 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2227 zio_write_gang_done
, NULL
, pio
->io_priority
,
2228 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2231 * Create and nowait the gang children.
2233 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2234 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2236 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2238 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2239 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2240 zp
.zp_type
= DMU_OT_NONE
;
2242 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2243 zp
.zp_dedup
= B_FALSE
;
2244 zp
.zp_dedup_verify
= B_FALSE
;
2245 zp
.zp_nopwrite
= B_FALSE
;
2247 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2248 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2249 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2250 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2251 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2253 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2254 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2255 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2258 * Gang children won't throttle but we should
2259 * account for their work, so reserve an allocation
2260 * slot for them here.
2262 VERIFY(metaslab_class_throttle_reserve(mc
,
2263 zp
.zp_copies
, cio
, flags
));
2269 * Set pio's pipeline to just wait for zio to finish.
2271 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2275 return (ZIO_PIPELINE_CONTINUE
);
2279 * The zio_nop_write stage in the pipeline determines if allocating a
2280 * new bp is necessary. The nopwrite feature can handle writes in
2281 * either syncing or open context (i.e. zil writes) and as a result is
2282 * mutually exclusive with dedup.
2284 * By leveraging a cryptographically secure checksum, such as SHA256, we
2285 * can compare the checksums of the new data and the old to determine if
2286 * allocating a new block is required. Note that our requirements for
2287 * cryptographic strength are fairly weak: there can't be any accidental
2288 * hash collisions, but we don't need to be secure against intentional
2289 * (malicious) collisions. To trigger a nopwrite, you have to be able
2290 * to write the file to begin with, and triggering an incorrect (hash
2291 * collision) nopwrite is no worse than simply writing to the file.
2292 * That said, there are no known attacks against the checksum algorithms
2293 * used for nopwrite, assuming that the salt and the checksums
2294 * themselves remain secret.
2297 zio_nop_write(zio_t
*zio
)
2299 blkptr_t
*bp
= zio
->io_bp
;
2300 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2301 zio_prop_t
*zp
= &zio
->io_prop
;
2303 ASSERT(BP_GET_LEVEL(bp
) == 0);
2304 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2305 ASSERT(zp
->zp_nopwrite
);
2306 ASSERT(!zp
->zp_dedup
);
2307 ASSERT(zio
->io_bp_override
== NULL
);
2308 ASSERT(IO_IS_ALLOCATING(zio
));
2311 * Check to see if the original bp and the new bp have matching
2312 * characteristics (i.e. same checksum, compression algorithms, etc).
2313 * If they don't then just continue with the pipeline which will
2314 * allocate a new bp.
2316 if (BP_IS_HOLE(bp_orig
) ||
2317 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2318 ZCHECKSUM_FLAG_NOPWRITE
) ||
2319 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2320 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2321 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2322 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2323 return (ZIO_PIPELINE_CONTINUE
);
2326 * If the checksums match then reset the pipeline so that we
2327 * avoid allocating a new bp and issuing any I/O.
2329 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2330 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2331 ZCHECKSUM_FLAG_NOPWRITE
);
2332 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2333 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2334 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2335 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2336 sizeof (uint64_t)) == 0);
2339 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2340 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2343 return (ZIO_PIPELINE_CONTINUE
);
2347 * ==========================================================================
2349 * ==========================================================================
2352 zio_ddt_child_read_done(zio_t
*zio
)
2354 blkptr_t
*bp
= zio
->io_bp
;
2355 ddt_entry_t
*dde
= zio
->io_private
;
2357 zio_t
*pio
= zio_unique_parent(zio
);
2359 mutex_enter(&pio
->io_lock
);
2360 ddp
= ddt_phys_select(dde
, bp
);
2361 if (zio
->io_error
== 0)
2362 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2364 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2365 dde
->dde_repair_abd
= zio
->io_abd
;
2367 abd_free(zio
->io_abd
);
2368 mutex_exit(&pio
->io_lock
);
2372 zio_ddt_read_start(zio_t
*zio
)
2374 blkptr_t
*bp
= zio
->io_bp
;
2376 ASSERT(BP_GET_DEDUP(bp
));
2377 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2378 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2380 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2381 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2382 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2383 ddt_phys_t
*ddp
= dde
->dde_phys
;
2384 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2387 ASSERT(zio
->io_vsd
== NULL
);
2390 if (ddp_self
== NULL
)
2391 return (ZIO_PIPELINE_CONTINUE
);
2393 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2394 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2396 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2398 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2399 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2400 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2401 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2402 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2404 return (ZIO_PIPELINE_CONTINUE
);
2407 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2408 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2409 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2411 return (ZIO_PIPELINE_CONTINUE
);
2415 zio_ddt_read_done(zio_t
*zio
)
2417 blkptr_t
*bp
= zio
->io_bp
;
2419 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
2420 return (ZIO_PIPELINE_STOP
);
2422 ASSERT(BP_GET_DEDUP(bp
));
2423 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2424 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2426 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2427 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2428 ddt_entry_t
*dde
= zio
->io_vsd
;
2430 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2431 return (ZIO_PIPELINE_CONTINUE
);
2434 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2435 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2436 return (ZIO_PIPELINE_STOP
);
2438 if (dde
->dde_repair_abd
!= NULL
) {
2439 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2441 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2443 ddt_repair_done(ddt
, dde
);
2447 ASSERT(zio
->io_vsd
== NULL
);
2449 return (ZIO_PIPELINE_CONTINUE
);
2453 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2455 spa_t
*spa
= zio
->io_spa
;
2456 boolean_t do_raw
= (zio
->io_flags
& ZIO_FLAG_RAW
);
2458 /* We should never get a raw, override zio */
2459 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2462 * Note: we compare the original data, not the transformed data,
2463 * because when zio->io_bp is an override bp, we will not have
2464 * pushed the I/O transforms. That's an important optimization
2465 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2467 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2468 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2471 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2472 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
,
2473 zio
->io_orig_size
) != 0);
2477 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2478 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2480 if (ddp
->ddp_phys_birth
!= 0) {
2481 arc_buf_t
*abuf
= NULL
;
2482 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2483 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2484 blkptr_t blk
= *zio
->io_bp
;
2487 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2492 * Intuitively, it would make more sense to compare
2493 * io_abd than io_orig_abd in the raw case since you
2494 * don't want to look at any transformations that have
2495 * happened to the data. However, for raw I/Os the
2496 * data will actually be the same in io_abd and
2497 * io_orig_abd, so all we have to do is issue this as
2501 zio_flags
|= ZIO_FLAG_RAW
;
2502 ASSERT3U(zio
->io_size
, ==, zio
->io_orig_size
);
2503 ASSERT0(abd_cmp(zio
->io_abd
, zio
->io_orig_abd
,
2505 ASSERT3P(zio
->io_transform_stack
, ==, NULL
);
2508 error
= arc_read(NULL
, spa
, &blk
,
2509 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2510 zio_flags
, &aflags
, &zio
->io_bookmark
);
2513 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
2514 abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2515 zio
->io_orig_size
) != 0)
2516 error
= SET_ERROR(EEXIST
);
2517 arc_buf_destroy(abuf
, &abuf
);
2521 return (error
!= 0);
2529 zio_ddt_child_write_ready(zio_t
*zio
)
2531 int p
= zio
->io_prop
.zp_copies
;
2532 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2533 ddt_entry_t
*dde
= zio
->io_private
;
2534 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2542 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2544 ddt_phys_fill(ddp
, zio
->io_bp
);
2546 zio_link_t
*zl
= NULL
;
2547 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2548 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2554 zio_ddt_child_write_done(zio_t
*zio
)
2556 int p
= zio
->io_prop
.zp_copies
;
2557 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2558 ddt_entry_t
*dde
= zio
->io_private
;
2559 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2563 ASSERT(ddp
->ddp_refcnt
== 0);
2564 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2565 dde
->dde_lead_zio
[p
] = NULL
;
2567 if (zio
->io_error
== 0) {
2568 zio_link_t
*zl
= NULL
;
2569 while (zio_walk_parents(zio
, &zl
) != NULL
)
2570 ddt_phys_addref(ddp
);
2572 ddt_phys_clear(ddp
);
2579 zio_ddt_ditto_write_done(zio_t
*zio
)
2581 int p
= DDT_PHYS_DITTO
;
2582 zio_prop_t
*zp
= &zio
->io_prop
;
2583 blkptr_t
*bp
= zio
->io_bp
;
2584 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2585 ddt_entry_t
*dde
= zio
->io_private
;
2586 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2587 ddt_key_t
*ddk
= &dde
->dde_key
;
2591 ASSERT(ddp
->ddp_refcnt
== 0);
2592 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2593 dde
->dde_lead_zio
[p
] = NULL
;
2595 if (zio
->io_error
== 0) {
2596 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2597 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2598 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2599 if (ddp
->ddp_phys_birth
!= 0)
2600 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2601 ddt_phys_fill(ddp
, bp
);
2608 zio_ddt_write(zio_t
*zio
)
2610 spa_t
*spa
= zio
->io_spa
;
2611 blkptr_t
*bp
= zio
->io_bp
;
2612 uint64_t txg
= zio
->io_txg
;
2613 zio_prop_t
*zp
= &zio
->io_prop
;
2614 int p
= zp
->zp_copies
;
2618 ddt_t
*ddt
= ddt_select(spa
, bp
);
2622 ASSERT(BP_GET_DEDUP(bp
));
2623 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2624 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2625 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2628 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2629 ddp
= &dde
->dde_phys
[p
];
2631 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2633 * If we're using a weak checksum, upgrade to a strong checksum
2634 * and try again. If we're already using a strong checksum,
2635 * we can't resolve it, so just convert to an ordinary write.
2636 * (And automatically e-mail a paper to Nature?)
2638 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2639 ZCHECKSUM_FLAG_DEDUP
)) {
2640 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2641 zio_pop_transforms(zio
);
2642 zio
->io_stage
= ZIO_STAGE_OPEN
;
2645 zp
->zp_dedup
= B_FALSE
;
2646 BP_SET_DEDUP(bp
, B_FALSE
);
2648 ASSERT(!BP_GET_DEDUP(bp
));
2649 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2651 return (ZIO_PIPELINE_CONTINUE
);
2654 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2655 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2657 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2658 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2659 zio_prop_t czp
= *zp
;
2661 czp
.zp_copies
= ditto_copies
;
2664 * If we arrived here with an override bp, we won't have run
2665 * the transform stack, so we won't have the data we need to
2666 * generate a child i/o. So, toss the override bp and restart.
2667 * This is safe, because using the override bp is just an
2668 * optimization; and it's rare, so the cost doesn't matter.
2670 if (zio
->io_bp_override
) {
2671 zio_pop_transforms(zio
);
2672 zio
->io_stage
= ZIO_STAGE_OPEN
;
2673 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2674 zio
->io_bp_override
= NULL
;
2677 return (ZIO_PIPELINE_CONTINUE
);
2680 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2681 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2682 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2683 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2685 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2686 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2689 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2690 if (ddp
->ddp_phys_birth
!= 0)
2691 ddt_bp_fill(ddp
, bp
, txg
);
2692 if (dde
->dde_lead_zio
[p
] != NULL
)
2693 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2695 ddt_phys_addref(ddp
);
2696 } else if (zio
->io_bp_override
) {
2697 ASSERT(bp
->blk_birth
== txg
);
2698 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2699 ddt_phys_fill(ddp
, bp
);
2700 ddt_phys_addref(ddp
);
2702 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2703 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
2704 zio_ddt_child_write_ready
, NULL
, NULL
,
2705 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2706 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2708 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2709 dde
->dde_lead_zio
[p
] = cio
;
2719 return (ZIO_PIPELINE_CONTINUE
);
2722 ddt_entry_t
*freedde
; /* for debugging */
2725 zio_ddt_free(zio_t
*zio
)
2727 spa_t
*spa
= zio
->io_spa
;
2728 blkptr_t
*bp
= zio
->io_bp
;
2729 ddt_t
*ddt
= ddt_select(spa
, bp
);
2733 ASSERT(BP_GET_DEDUP(bp
));
2734 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2737 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2738 ddp
= ddt_phys_select(dde
, bp
);
2739 ddt_phys_decref(ddp
);
2742 return (ZIO_PIPELINE_CONTINUE
);
2746 * ==========================================================================
2747 * Allocate and free blocks
2748 * ==========================================================================
2752 zio_io_to_allocate(spa_t
*spa
)
2756 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
2758 zio
= avl_first(&spa
->spa_alloc_tree
);
2762 ASSERT(IO_IS_ALLOCATING(zio
));
2765 * Try to place a reservation for this zio. If we're unable to
2766 * reserve then we throttle.
2768 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
2769 zio
->io_prop
.zp_copies
, zio
, 0)) {
2773 avl_remove(&spa
->spa_alloc_tree
, zio
);
2774 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
2780 zio_dva_throttle(zio_t
*zio
)
2782 spa_t
*spa
= zio
->io_spa
;
2785 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
2786 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
2787 zio
->io_child_type
== ZIO_CHILD_GANG
||
2788 zio
->io_flags
& ZIO_FLAG_NODATA
) {
2789 return (ZIO_PIPELINE_CONTINUE
);
2792 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2794 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2795 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2797 mutex_enter(&spa
->spa_alloc_lock
);
2799 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2800 avl_add(&spa
->spa_alloc_tree
, zio
);
2802 nio
= zio_io_to_allocate(zio
->io_spa
);
2803 mutex_exit(&spa
->spa_alloc_lock
);
2806 return (ZIO_PIPELINE_CONTINUE
);
2809 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2811 * We are passing control to a new zio so make sure that
2812 * it is processed by a different thread. We do this to
2813 * avoid stack overflows that can occur when parents are
2814 * throttled and children are making progress. We allow
2815 * it to go to the head of the taskq since it's already
2818 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2820 return (ZIO_PIPELINE_STOP
);
2824 zio_allocate_dispatch(spa_t
*spa
)
2828 mutex_enter(&spa
->spa_alloc_lock
);
2829 zio
= zio_io_to_allocate(spa
);
2830 mutex_exit(&spa
->spa_alloc_lock
);
2834 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
2835 ASSERT0(zio
->io_error
);
2836 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2840 zio_dva_allocate(zio_t
*zio
)
2842 spa_t
*spa
= zio
->io_spa
;
2843 metaslab_class_t
*mc
= spa_normal_class(spa
);
2844 blkptr_t
*bp
= zio
->io_bp
;
2848 if (zio
->io_gang_leader
== NULL
) {
2849 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2850 zio
->io_gang_leader
= zio
;
2853 ASSERT(BP_IS_HOLE(bp
));
2854 ASSERT0(BP_GET_NDVAS(bp
));
2855 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2856 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2857 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2859 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2860 flags
|= METASLAB_DONT_THROTTLE
;
2862 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
) {
2863 flags
|= METASLAB_GANG_CHILD
;
2865 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
) {
2866 flags
|= METASLAB_ASYNC_ALLOC
;
2869 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2870 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
2871 &zio
->io_alloc_list
, zio
);
2874 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
2875 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
2877 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2878 return (zio_write_gang_block(zio
));
2879 zio
->io_error
= error
;
2882 return (ZIO_PIPELINE_CONTINUE
);
2886 zio_dva_free(zio_t
*zio
)
2888 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2890 return (ZIO_PIPELINE_CONTINUE
);
2894 zio_dva_claim(zio_t
*zio
)
2898 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2900 zio
->io_error
= error
;
2902 return (ZIO_PIPELINE_CONTINUE
);
2906 * Undo an allocation. This is used by zio_done() when an I/O fails
2907 * and we want to give back the block we just allocated.
2908 * This handles both normal blocks and gang blocks.
2911 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2913 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2914 ASSERT(zio
->io_bp_override
== NULL
);
2916 if (!BP_IS_HOLE(bp
))
2917 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2920 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2921 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2922 &gn
->gn_gbh
->zg_blkptr
[g
]);
2928 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2931 zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
, blkptr_t
*old_bp
,
2932 uint64_t size
, boolean_t
*slog
)
2935 zio_alloc_list_t io_alloc_list
;
2937 ASSERT(txg
> spa_syncing_txg(spa
));
2939 metaslab_trace_init(&io_alloc_list
);
2940 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
2941 txg
, old_bp
, METASLAB_HINTBP_AVOID
, &io_alloc_list
, NULL
);
2945 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
2946 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
,
2947 &io_alloc_list
, NULL
);
2951 metaslab_trace_fini(&io_alloc_list
);
2954 BP_SET_LSIZE(new_bp
, size
);
2955 BP_SET_PSIZE(new_bp
, size
);
2956 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
2957 BP_SET_CHECKSUM(new_bp
,
2958 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
2959 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
2960 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
2961 BP_SET_LEVEL(new_bp
, 0);
2962 BP_SET_DEDUP(new_bp
, 0);
2963 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
2965 zfs_dbgmsg("%s: zil block allocation failure: "
2966 "size %llu, error %d", spa_name(spa
), size
, error
);
2973 * Free an intent log block.
2976 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
2978 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
2979 ASSERT(!BP_IS_GANG(bp
));
2981 zio_free(spa
, txg
, bp
);
2985 * ==========================================================================
2986 * Read and write to physical devices
2987 * ==========================================================================
2992 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2993 * stops after this stage and will resume upon I/O completion.
2994 * However, there are instances where the vdev layer may need to
2995 * continue the pipeline when an I/O was not issued. Since the I/O
2996 * that was sent to the vdev layer might be different than the one
2997 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2998 * force the underlying vdev layers to call either zio_execute() or
2999 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3002 zio_vdev_io_start(zio_t
*zio
)
3004 vdev_t
*vd
= zio
->io_vd
;
3006 spa_t
*spa
= zio
->io_spa
;
3008 ASSERT(zio
->io_error
== 0);
3009 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3012 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3013 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3016 * The mirror_ops handle multiple DVAs in a single BP.
3018 vdev_mirror_ops
.vdev_op_io_start(zio
);
3019 return (ZIO_PIPELINE_STOP
);
3022 ASSERT3P(zio
->io_logical
, !=, zio
);
3025 * We keep track of time-sensitive I/Os so that the scan thread
3026 * can quickly react to certain workloads. In particular, we care
3027 * about non-scrubbing, top-level reads and writes with the following
3029 * - synchronous writes of user data to non-slog devices
3030 * - any reads of user data
3031 * When these conditions are met, adjust the timestamp of spa_last_io
3032 * which allows the scan thread to adjust its workload accordingly.
3034 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
3035 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
3036 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
3037 zio
->io_txg
!= spa_syncing_txg(spa
)) {
3038 uint64_t old
= spa
->spa_last_io
;
3039 uint64_t new = ddi_get_lbolt64();
3041 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
3044 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3046 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3047 P2PHASE(zio
->io_size
, align
) != 0) {
3048 /* Transform logical writes to be a full physical block size. */
3049 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3050 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3051 ASSERT(vd
== vd
->vdev_top
);
3052 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3053 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3054 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3056 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3060 * If this is not a physical io, make sure that it is properly aligned
3061 * before proceeding.
3063 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3064 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3065 ASSERT0(P2PHASE(zio
->io_size
, align
));
3068 * For physical writes, we allow 512b aligned writes and assume
3069 * the device will perform a read-modify-write as necessary.
3071 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3072 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3075 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3078 * If this is a repair I/O, and there's no self-healing involved --
3079 * that is, we're just resilvering what we expect to resilver --
3080 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3081 * This prevents spurious resilvering with nested replication.
3082 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3083 * A is out of date, we'll read from C+D, then use the data to
3084 * resilver A+B -- but we don't actually want to resilver B, just A.
3085 * The top-level mirror has no way to know this, so instead we just
3086 * discard unnecessary repairs as we work our way down the vdev tree.
3087 * The same logic applies to any form of nested replication:
3088 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3090 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3091 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3092 zio
->io_txg
!= 0 && /* not a delegated i/o */
3093 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3094 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3095 zio_vdev_io_bypass(zio
);
3096 return (ZIO_PIPELINE_CONTINUE
);
3099 if (vd
->vdev_ops
->vdev_op_leaf
&&
3100 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3102 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3103 return (ZIO_PIPELINE_CONTINUE
);
3105 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3106 return (ZIO_PIPELINE_STOP
);
3108 if (!vdev_accessible(vd
, zio
)) {
3109 zio
->io_error
= SET_ERROR(ENXIO
);
3111 return (ZIO_PIPELINE_STOP
);
3115 vd
->vdev_ops
->vdev_op_io_start(zio
);
3116 return (ZIO_PIPELINE_STOP
);
3120 zio_vdev_io_done(zio_t
*zio
)
3122 vdev_t
*vd
= zio
->io_vd
;
3123 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3124 boolean_t unexpected_error
= B_FALSE
;
3126 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3127 return (ZIO_PIPELINE_STOP
);
3129 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3131 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3133 vdev_queue_io_done(zio
);
3135 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3136 vdev_cache_write(zio
);
3138 if (zio_injection_enabled
&& zio
->io_error
== 0)
3139 zio
->io_error
= zio_handle_device_injection(vd
,
3142 if (zio_injection_enabled
&& zio
->io_error
== 0)
3143 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3145 if (zio
->io_error
) {
3146 if (!vdev_accessible(vd
, zio
)) {
3147 zio
->io_error
= SET_ERROR(ENXIO
);
3149 unexpected_error
= B_TRUE
;
3154 ops
->vdev_op_io_done(zio
);
3156 if (unexpected_error
)
3157 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3159 return (ZIO_PIPELINE_CONTINUE
);
3163 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3164 * disk, and use that to finish the checksum ereport later.
3167 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3168 const void *good_buf
)
3170 /* no processing needed */
3171 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3176 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3178 void *buf
= zio_buf_alloc(zio
->io_size
);
3180 abd_copy_to_buf(buf
, zio
->io_abd
, zio
->io_size
);
3182 zcr
->zcr_cbinfo
= zio
->io_size
;
3183 zcr
->zcr_cbdata
= buf
;
3184 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3185 zcr
->zcr_free
= zio_buf_free
;
3189 zio_vdev_io_assess(zio_t
*zio
)
3191 vdev_t
*vd
= zio
->io_vd
;
3193 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3194 return (ZIO_PIPELINE_STOP
);
3196 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3197 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3199 if (zio
->io_vsd
!= NULL
) {
3200 zio
->io_vsd_ops
->vsd_free(zio
);
3204 if (zio_injection_enabled
&& zio
->io_error
== 0)
3205 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3208 * If the I/O failed, determine whether we should attempt to retry it.
3210 * On retry, we cut in line in the issue queue, since we don't want
3211 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3213 if (zio
->io_error
&& vd
== NULL
&&
3214 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3215 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3216 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3218 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3219 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3220 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3221 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3222 zio_requeue_io_start_cut_in_line
);
3223 return (ZIO_PIPELINE_STOP
);
3227 * If we got an error on a leaf device, convert it to ENXIO
3228 * if the device is not accessible at all.
3230 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3231 !vdev_accessible(vd
, zio
))
3232 zio
->io_error
= SET_ERROR(ENXIO
);
3235 * If we can't write to an interior vdev (mirror or RAID-Z),
3236 * set vdev_cant_write so that we stop trying to allocate from it.
3238 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3239 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3240 vd
->vdev_cant_write
= B_TRUE
;
3244 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3245 * attempts will ever succeed. In this case we set a persistent bit so
3246 * that we don't bother with it in the future.
3248 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3249 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3250 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3251 vd
->vdev_nowritecache
= B_TRUE
;
3254 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3256 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3257 zio
->io_physdone
!= NULL
) {
3258 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3259 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3260 zio
->io_physdone(zio
->io_logical
);
3263 return (ZIO_PIPELINE_CONTINUE
);
3267 zio_vdev_io_reissue(zio_t
*zio
)
3269 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3270 ASSERT(zio
->io_error
== 0);
3272 zio
->io_stage
>>= 1;
3276 zio_vdev_io_redone(zio_t
*zio
)
3278 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3280 zio
->io_stage
>>= 1;
3284 zio_vdev_io_bypass(zio_t
*zio
)
3286 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3287 ASSERT(zio
->io_error
== 0);
3289 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3290 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3294 * ==========================================================================
3295 * Generate and verify checksums
3296 * ==========================================================================
3299 zio_checksum_generate(zio_t
*zio
)
3301 blkptr_t
*bp
= zio
->io_bp
;
3302 enum zio_checksum checksum
;
3306 * This is zio_write_phys().
3307 * We're either generating a label checksum, or none at all.
3309 checksum
= zio
->io_prop
.zp_checksum
;
3311 if (checksum
== ZIO_CHECKSUM_OFF
)
3312 return (ZIO_PIPELINE_CONTINUE
);
3314 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3316 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3317 ASSERT(!IO_IS_ALLOCATING(zio
));
3318 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3320 checksum
= BP_GET_CHECKSUM(bp
);
3324 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3326 return (ZIO_PIPELINE_CONTINUE
);
3330 zio_checksum_verify(zio_t
*zio
)
3332 zio_bad_cksum_t info
;
3333 blkptr_t
*bp
= zio
->io_bp
;
3336 ASSERT(zio
->io_vd
!= NULL
);
3340 * This is zio_read_phys().
3341 * We're either verifying a label checksum, or nothing at all.
3343 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3344 return (ZIO_PIPELINE_CONTINUE
);
3346 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3349 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3350 zio
->io_error
= error
;
3351 if (error
== ECKSUM
&&
3352 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3353 zfs_ereport_start_checksum(zio
->io_spa
,
3354 zio
->io_vd
, zio
, zio
->io_offset
,
3355 zio
->io_size
, NULL
, &info
);
3359 return (ZIO_PIPELINE_CONTINUE
);
3363 * Called by RAID-Z to ensure we don't compute the checksum twice.
3366 zio_checksum_verified(zio_t
*zio
)
3368 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3372 * ==========================================================================
3373 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3374 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3375 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3376 * indicate errors that are specific to one I/O, and most likely permanent.
3377 * Any other error is presumed to be worse because we weren't expecting it.
3378 * ==========================================================================
3381 zio_worst_error(int e1
, int e2
)
3383 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3386 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3387 if (e1
== zio_error_rank
[r1
])
3390 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3391 if (e2
== zio_error_rank
[r2
])
3394 return (r1
> r2
? e1
: e2
);
3398 * ==========================================================================
3400 * ==========================================================================
3403 zio_ready(zio_t
*zio
)
3405 blkptr_t
*bp
= zio
->io_bp
;
3406 zio_t
*pio
, *pio_next
;
3407 zio_link_t
*zl
= NULL
;
3409 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
3410 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
3411 return (ZIO_PIPELINE_STOP
);
3413 if (zio
->io_ready
) {
3414 ASSERT(IO_IS_ALLOCATING(zio
));
3415 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3416 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3417 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3422 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3423 zio
->io_bp_copy
= *bp
;
3425 if (zio
->io_error
!= 0) {
3426 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3428 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3429 ASSERT(IO_IS_ALLOCATING(zio
));
3430 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3432 * We were unable to allocate anything, unreserve and
3433 * issue the next I/O to allocate.
3435 metaslab_class_throttle_unreserve(
3436 spa_normal_class(zio
->io_spa
),
3437 zio
->io_prop
.zp_copies
, zio
);
3438 zio_allocate_dispatch(zio
->io_spa
);
3442 mutex_enter(&zio
->io_lock
);
3443 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3444 pio
= zio_walk_parents(zio
, &zl
);
3445 mutex_exit(&zio
->io_lock
);
3448 * As we notify zio's parents, new parents could be added.
3449 * New parents go to the head of zio's io_parent_list, however,
3450 * so we will (correctly) not notify them. The remainder of zio's
3451 * io_parent_list, from 'pio_next' onward, cannot change because
3452 * all parents must wait for us to be done before they can be done.
3454 for (; pio
!= NULL
; pio
= pio_next
) {
3455 pio_next
= zio_walk_parents(zio
, &zl
);
3456 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3459 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3460 if (BP_IS_GANG(bp
)) {
3461 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3463 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3464 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3468 if (zio_injection_enabled
&&
3469 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3470 zio_handle_ignored_writes(zio
);
3472 return (ZIO_PIPELINE_CONTINUE
);
3476 * Update the allocation throttle accounting.
3479 zio_dva_throttle_done(zio_t
*zio
)
3481 zio_t
*lio
= zio
->io_logical
;
3482 zio_t
*pio
= zio_unique_parent(zio
);
3483 vdev_t
*vd
= zio
->io_vd
;
3484 int flags
= METASLAB_ASYNC_ALLOC
;
3486 ASSERT3P(zio
->io_bp
, !=, NULL
);
3487 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3488 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3489 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3491 ASSERT3P(vd
, ==, vd
->vdev_top
);
3492 ASSERT(!(zio
->io_flags
& (ZIO_FLAG_IO_REPAIR
| ZIO_FLAG_IO_RETRY
)));
3493 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3494 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3495 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3498 * Parents of gang children can have two flavors -- ones that
3499 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3500 * and ones that allocated the constituent blocks. The allocation
3501 * throttle needs to know the allocating parent zio so we must find
3504 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3506 * If our parent is a rewrite gang child then our grandparent
3507 * would have been the one that performed the allocation.
3509 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3510 pio
= zio_unique_parent(pio
);
3511 flags
|= METASLAB_GANG_CHILD
;
3514 ASSERT(IO_IS_ALLOCATING(pio
));
3515 ASSERT3P(zio
, !=, zio
->io_logical
);
3516 ASSERT(zio
->io_logical
!= NULL
);
3517 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3518 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3520 mutex_enter(&pio
->io_lock
);
3521 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
3522 mutex_exit(&pio
->io_lock
);
3524 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
3528 * Call into the pipeline to see if there is more work that
3529 * needs to be done. If there is work to be done it will be
3530 * dispatched to another taskq thread.
3532 zio_allocate_dispatch(zio
->io_spa
);
3536 zio_done(zio_t
*zio
)
3538 spa_t
*spa
= zio
->io_spa
;
3539 zio_t
*lio
= zio
->io_logical
;
3540 blkptr_t
*bp
= zio
->io_bp
;
3541 vdev_t
*vd
= zio
->io_vd
;
3542 uint64_t psize
= zio
->io_size
;
3543 zio_t
*pio
, *pio_next
;
3544 metaslab_class_t
*mc
= spa_normal_class(spa
);
3545 zio_link_t
*zl
= NULL
;
3548 * If our children haven't all completed,
3549 * wait for them and then repeat this pipeline stage.
3551 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
3552 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
3553 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
3554 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
3555 return (ZIO_PIPELINE_STOP
);
3558 * If the allocation throttle is enabled, then update the accounting.
3559 * We only track child I/Os that are part of an allocating async
3560 * write. We must do this since the allocation is performed
3561 * by the logical I/O but the actual write is done by child I/Os.
3563 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
3564 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
3565 ASSERT(mc
->mc_alloc_throttle_enabled
);
3566 zio_dva_throttle_done(zio
);
3570 * If the allocation throttle is enabled, verify that
3571 * we have decremented the refcounts for every I/O that was throttled.
3573 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3574 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3575 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3577 metaslab_group_alloc_verify(spa
, zio
->io_bp
, zio
);
3578 VERIFY(refcount_not_held(&mc
->mc_alloc_slots
, zio
));
3581 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
3582 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
3583 ASSERT(zio
->io_children
[c
][w
] == 0);
3585 if (bp
!= NULL
&& !BP_IS_EMBEDDED(bp
)) {
3586 ASSERT(bp
->blk_pad
[0] == 0);
3587 ASSERT(bp
->blk_pad
[1] == 0);
3588 ASSERT(bcmp(bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
3589 (bp
== zio_unique_parent(zio
)->io_bp
));
3590 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(bp
) &&
3591 zio
->io_bp_override
== NULL
&&
3592 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
3593 ASSERT(!BP_SHOULD_BYTESWAP(bp
));
3594 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(bp
));
3595 ASSERT(BP_COUNT_GANG(bp
) == 0 ||
3596 (BP_COUNT_GANG(bp
) == BP_GET_NDVAS(bp
)));
3598 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
3599 VERIFY(BP_EQUAL(bp
, &zio
->io_bp_orig
));
3603 * If there were child vdev/gang/ddt errors, they apply to us now.
3605 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
3606 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
3607 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
3610 * If the I/O on the transformed data was successful, generate any
3611 * checksum reports now while we still have the transformed data.
3613 if (zio
->io_error
== 0) {
3614 while (zio
->io_cksum_report
!= NULL
) {
3615 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3616 uint64_t align
= zcr
->zcr_align
;
3617 uint64_t asize
= P2ROUNDUP(psize
, align
);
3619 abd_t
*adata
= zio
->io_abd
;
3621 if (asize
!= psize
) {
3622 adata
= abd_alloc_linear(asize
, B_TRUE
);
3623 abd_copy(adata
, zio
->io_abd
, psize
);
3624 abd_zero_off(adata
, psize
, asize
- psize
);
3628 abuf
= abd_borrow_buf_copy(adata
, asize
);
3630 zio
->io_cksum_report
= zcr
->zcr_next
;
3631 zcr
->zcr_next
= NULL
;
3632 zcr
->zcr_finish(zcr
, abuf
);
3633 zfs_ereport_free_checksum(zcr
);
3636 abd_return_buf(adata
, abuf
, asize
);
3643 zio_pop_transforms(zio
); /* note: may set zio->io_error */
3645 vdev_stat_update(zio
, psize
);
3647 if (zio
->io_error
) {
3649 * If this I/O is attached to a particular vdev,
3650 * generate an error message describing the I/O failure
3651 * at the block level. We ignore these errors if the
3652 * device is currently unavailable.
3654 if (zio
->io_error
!= ECKSUM
&& vd
!= NULL
&& !vdev_is_dead(vd
))
3655 zfs_ereport_post(FM_EREPORT_ZFS_IO
, spa
, vd
, zio
, 0, 0);
3657 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
3658 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
3661 * For logical I/O requests, tell the SPA to log the
3662 * error and generate a logical data ereport.
3664 spa_log_error(spa
, zio
);
3665 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, spa
, NULL
, zio
,
3670 if (zio
->io_error
&& zio
== lio
) {
3672 * Determine whether zio should be reexecuted. This will
3673 * propagate all the way to the root via zio_notify_parent().
3675 ASSERT(vd
== NULL
&& bp
!= NULL
);
3676 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3678 if (IO_IS_ALLOCATING(zio
) &&
3679 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
3680 if (zio
->io_error
!= ENOSPC
)
3681 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
3683 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3686 if ((zio
->io_type
== ZIO_TYPE_READ
||
3687 zio
->io_type
== ZIO_TYPE_FREE
) &&
3688 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
3689 zio
->io_error
== ENXIO
&&
3690 spa_load_state(spa
) == SPA_LOAD_NONE
&&
3691 spa_get_failmode(spa
) != ZIO_FAILURE_MODE_CONTINUE
)
3692 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3694 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
3695 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3698 * Here is a possibly good place to attempt to do
3699 * either combinatorial reconstruction or error correction
3700 * based on checksums. It also might be a good place
3701 * to send out preliminary ereports before we suspend
3707 * If there were logical child errors, they apply to us now.
3708 * We defer this until now to avoid conflating logical child
3709 * errors with errors that happened to the zio itself when
3710 * updating vdev stats and reporting FMA events above.
3712 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
3714 if ((zio
->io_error
|| zio
->io_reexecute
) &&
3715 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
3716 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
3717 zio_dva_unallocate(zio
, zio
->io_gang_tree
, bp
);
3719 zio_gang_tree_free(&zio
->io_gang_tree
);
3722 * Godfather I/Os should never suspend.
3724 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3725 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
3726 zio
->io_reexecute
= 0;
3728 if (zio
->io_reexecute
) {
3730 * This is a logical I/O that wants to reexecute.
3732 * Reexecute is top-down. When an i/o fails, if it's not
3733 * the root, it simply notifies its parent and sticks around.
3734 * The parent, seeing that it still has children in zio_done(),
3735 * does the same. This percolates all the way up to the root.
3736 * The root i/o will reexecute or suspend the entire tree.
3738 * This approach ensures that zio_reexecute() honors
3739 * all the original i/o dependency relationships, e.g.
3740 * parents not executing until children are ready.
3742 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3744 zio
->io_gang_leader
= NULL
;
3746 mutex_enter(&zio
->io_lock
);
3747 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3748 mutex_exit(&zio
->io_lock
);
3751 * "The Godfather" I/O monitors its children but is
3752 * not a true parent to them. It will track them through
3753 * the pipeline but severs its ties whenever they get into
3754 * trouble (e.g. suspended). This allows "The Godfather"
3755 * I/O to return status without blocking.
3758 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
3760 zio_link_t
*remove_zl
= zl
;
3761 pio_next
= zio_walk_parents(zio
, &zl
);
3763 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3764 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
3765 zio_remove_child(pio
, zio
, remove_zl
);
3766 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3770 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
3772 * We're not a root i/o, so there's nothing to do
3773 * but notify our parent. Don't propagate errors
3774 * upward since we haven't permanently failed yet.
3776 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
3777 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
3778 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3779 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
3781 * We'd fail again if we reexecuted now, so suspend
3782 * until conditions improve (e.g. device comes online).
3784 zio_suspend(spa
, zio
);
3787 * Reexecution is potentially a huge amount of work.
3788 * Hand it off to the otherwise-unused claim taskq.
3790 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
3791 spa_taskq_dispatch_ent(spa
, ZIO_TYPE_CLAIM
,
3792 ZIO_TASKQ_ISSUE
, (task_func_t
*)zio_reexecute
, zio
,
3795 return (ZIO_PIPELINE_STOP
);
3798 ASSERT(zio
->io_child_count
== 0);
3799 ASSERT(zio
->io_reexecute
== 0);
3800 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
3803 * Report any checksum errors, since the I/O is complete.
3805 while (zio
->io_cksum_report
!= NULL
) {
3806 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3807 zio
->io_cksum_report
= zcr
->zcr_next
;
3808 zcr
->zcr_next
= NULL
;
3809 zcr
->zcr_finish(zcr
, NULL
);
3810 zfs_ereport_free_checksum(zcr
);
3814 * It is the responsibility of the done callback to ensure that this
3815 * particular zio is no longer discoverable for adoption, and as
3816 * such, cannot acquire any new parents.
3821 mutex_enter(&zio
->io_lock
);
3822 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3823 mutex_exit(&zio
->io_lock
);
3826 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
3827 zio_link_t
*remove_zl
= zl
;
3828 pio_next
= zio_walk_parents(zio
, &zl
);
3829 zio_remove_child(pio
, zio
, remove_zl
);
3830 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3833 if (zio
->io_waiter
!= NULL
) {
3834 mutex_enter(&zio
->io_lock
);
3835 zio
->io_executor
= NULL
;
3836 cv_broadcast(&zio
->io_cv
);
3837 mutex_exit(&zio
->io_lock
);
3842 return (ZIO_PIPELINE_STOP
);
3846 * ==========================================================================
3847 * I/O pipeline definition
3848 * ==========================================================================
3850 static zio_pipe_stage_t
*zio_pipeline
[] = {
3857 zio_checksum_generate
,
3873 zio_checksum_verify
,
3881 * Compare two zbookmark_phys_t's to see which we would reach first in a
3882 * pre-order traversal of the object tree.
3884 * This is simple in every case aside from the meta-dnode object. For all other
3885 * objects, we traverse them in order (object 1 before object 2, and so on).
3886 * However, all of these objects are traversed while traversing object 0, since
3887 * the data it points to is the list of objects. Thus, we need to convert to a
3888 * canonical representation so we can compare meta-dnode bookmarks to
3889 * non-meta-dnode bookmarks.
3891 * We do this by calculating "equivalents" for each field of the zbookmark.
3892 * zbookmarks outside of the meta-dnode use their own object and level, and
3893 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3894 * blocks this bookmark refers to) by multiplying their blkid by their span
3895 * (the number of L0 blocks contained within one block at their level).
3896 * zbookmarks inside the meta-dnode calculate their object equivalent
3897 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3898 * level + 1<<31 (any value larger than a level could ever be) for their level.
3899 * This causes them to always compare before a bookmark in their object
3900 * equivalent, compare appropriately to bookmarks in other objects, and to
3901 * compare appropriately to other bookmarks in the meta-dnode.
3904 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
3905 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
3908 * These variables represent the "equivalent" values for the zbookmark,
3909 * after converting zbookmarks inside the meta dnode to their
3910 * normal-object equivalents.
3912 uint64_t zb1obj
, zb2obj
;
3913 uint64_t zb1L0
, zb2L0
;
3914 uint64_t zb1level
, zb2level
;
3916 if (zb1
->zb_object
== zb2
->zb_object
&&
3917 zb1
->zb_level
== zb2
->zb_level
&&
3918 zb1
->zb_blkid
== zb2
->zb_blkid
)
3922 * BP_SPANB calculates the span in blocks.
3924 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
3925 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
3927 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
3928 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3930 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
3932 zb1obj
= zb1
->zb_object
;
3933 zb1level
= zb1
->zb_level
;
3936 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
3937 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3939 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
3941 zb2obj
= zb2
->zb_object
;
3942 zb2level
= zb2
->zb_level
;
3945 /* Now that we have a canonical representation, do the comparison. */
3946 if (zb1obj
!= zb2obj
)
3947 return (zb1obj
< zb2obj
? -1 : 1);
3948 else if (zb1L0
!= zb2L0
)
3949 return (zb1L0
< zb2L0
? -1 : 1);
3950 else if (zb1level
!= zb2level
)
3951 return (zb1level
> zb2level
? -1 : 1);
3953 * This can (theoretically) happen if the bookmarks have the same object
3954 * and level, but different blkids, if the block sizes are not the same.
3955 * There is presently no way to change the indirect block sizes
3961 * This function checks the following: given that last_block is the place that
3962 * our traversal stopped last time, does that guarantee that we've visited
3963 * every node under subtree_root? Therefore, we can't just use the raw output
3964 * of zbookmark_compare. We have to pass in a modified version of
3965 * subtree_root; by incrementing the block id, and then checking whether
3966 * last_block is before or equal to that, we can tell whether or not having
3967 * visited last_block implies that all of subtree_root's children have been
3971 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
3972 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
3974 zbookmark_phys_t mod_zb
= *subtree_root
;
3976 ASSERT(last_block
->zb_level
== 0);
3978 /* The objset_phys_t isn't before anything. */
3983 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3984 * data block size in sectors, because that variable is only used if
3985 * the bookmark refers to a block in the meta-dnode. Since we don't
3986 * know without examining it what object it refers to, and there's no
3987 * harm in passing in this value in other cases, we always pass it in.
3989 * We pass in 0 for the indirect block size shift because zb2 must be
3990 * level 0. The indirect block size is only used to calculate the span
3991 * of the bookmark, but since the bookmark must be level 0, the span is
3992 * always 1, so the math works out.
3994 * If you make changes to how the zbookmark_compare code works, be sure
3995 * to make sure that this code still works afterwards.
3997 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
3998 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,