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 * Do not verify individual DVAs if the config is not trusted. This
692 * will be done once the zio is executed in vdev_mirror_map_alloc.
694 if (!spa
->spa_trust_config
)
698 * Pool-specific checks.
700 * Note: it would be nice to verify that the blk_birth and
701 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
702 * allows the birth time of log blocks (and dmu_sync()-ed blocks
703 * that are in the log) to be arbitrarily large.
705 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
706 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
707 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
708 zfs_panic_recover("blkptr at %p DVA %u has invalid "
710 bp
, i
, (longlong_t
)vdevid
);
713 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
715 zfs_panic_recover("blkptr at %p DVA %u has invalid "
717 bp
, i
, (longlong_t
)vdevid
);
720 if (vd
->vdev_ops
== &vdev_hole_ops
) {
721 zfs_panic_recover("blkptr at %p DVA %u has hole "
723 bp
, i
, (longlong_t
)vdevid
);
726 if (vd
->vdev_ops
== &vdev_missing_ops
) {
728 * "missing" vdevs are valid during import, but we
729 * don't have their detailed info (e.g. asize), so
730 * we can't perform any more checks on them.
734 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
735 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
737 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
738 if (offset
+ asize
> vd
->vdev_asize
) {
739 zfs_panic_recover("blkptr at %p DVA %u has invalid "
741 bp
, i
, (longlong_t
)offset
);
747 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
749 uint64_t vdevid
= DVA_GET_VDEV(dva
);
751 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
754 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
758 if (vd
->vdev_ops
== &vdev_hole_ops
)
761 if (vd
->vdev_ops
== &vdev_missing_ops
) {
765 uint64_t offset
= DVA_GET_OFFSET(dva
);
766 uint64_t asize
= DVA_GET_ASIZE(dva
);
769 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
770 if (offset
+ asize
> vd
->vdev_asize
)
777 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
778 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
779 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
783 zfs_blkptr_verify(spa
, bp
);
785 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
786 data
, size
, size
, done
, private,
787 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
788 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
789 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
795 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
796 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
797 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
798 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
799 void *private, zio_priority_t priority
, enum zio_flag flags
,
800 const zbookmark_phys_t
*zb
)
804 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
805 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
806 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
807 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
808 DMU_OT_IS_VALID(zp
->zp_type
) &&
811 zp
->zp_copies
<= spa_max_replication(spa
));
813 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
814 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
815 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
816 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
818 zio
->io_ready
= ready
;
819 zio
->io_children_ready
= children_ready
;
820 zio
->io_physdone
= physdone
;
824 * Data can be NULL if we are going to call zio_write_override() to
825 * provide the already-allocated BP. But we may need the data to
826 * verify a dedup hit (if requested). In this case, don't try to
827 * dedup (just take the already-allocated BP verbatim).
829 if (data
== NULL
&& zio
->io_prop
.zp_dedup_verify
) {
830 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
837 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
838 uint64_t size
, zio_done_func_t
*done
, void *private,
839 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
843 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
844 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
845 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
851 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
853 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
854 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
855 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
856 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
859 * We must reset the io_prop to match the values that existed
860 * when the bp was first written by dmu_sync() keeping in mind
861 * that nopwrite and dedup are mutually exclusive.
863 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
864 zio
->io_prop
.zp_nopwrite
= nopwrite
;
865 zio
->io_prop
.zp_copies
= copies
;
866 zio
->io_bp_override
= bp
;
870 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
873 zfs_blkptr_verify(spa
, bp
);
876 * The check for EMBEDDED is a performance optimization. We
877 * process the free here (by ignoring it) rather than
878 * putting it on the list and then processing it in zio_free_sync().
880 if (BP_IS_EMBEDDED(bp
))
882 metaslab_check_free(spa
, bp
);
885 * Frees that are for the currently-syncing txg, are not going to be
886 * deferred, and which will not need to do a read (i.e. not GANG or
887 * DEDUP), can be processed immediately. Otherwise, put them on the
888 * in-memory list for later processing.
890 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
891 txg
!= spa
->spa_syncing_txg
||
892 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
893 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
895 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
900 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
904 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
906 ASSERT(!BP_IS_HOLE(bp
));
907 ASSERT(spa_syncing_txg(spa
) == txg
);
908 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
910 if (BP_IS_EMBEDDED(bp
))
911 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
913 metaslab_check_free(spa
, bp
);
917 * GANG and DEDUP blocks can induce a read (for the gang block header,
918 * or the DDT), so issue them asynchronously so that this thread is
921 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
922 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
924 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
925 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
926 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
932 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
933 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
937 zfs_blkptr_verify(spa
, bp
);
939 if (BP_IS_EMBEDDED(bp
))
940 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
943 * A claim is an allocation of a specific block. Claims are needed
944 * to support immediate writes in the intent log. The issue is that
945 * immediate writes contain committed data, but in a txg that was
946 * *not* committed. Upon opening the pool after an unclean shutdown,
947 * the intent log claims all blocks that contain immediate write data
948 * so that the SPA knows they're in use.
950 * All claims *must* be resolved in the first txg -- before the SPA
951 * starts allocating blocks -- so that nothing is allocated twice.
952 * If txg == 0 we just verify that the block is claimable.
954 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
955 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
956 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
958 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
959 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
960 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
961 ASSERT0(zio
->io_queued_timestamp
);
967 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
968 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
973 if (vd
->vdev_children
== 0) {
974 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
975 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
976 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
980 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
982 for (c
= 0; c
< vd
->vdev_children
; c
++)
983 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
984 done
, private, flags
));
991 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
992 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
993 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
997 ASSERT(vd
->vdev_children
== 0);
998 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
999 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1000 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1002 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1003 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1004 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1006 zio
->io_prop
.zp_checksum
= checksum
;
1012 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1013 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1014 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1018 ASSERT(vd
->vdev_children
== 0);
1019 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1020 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1021 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1023 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1024 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1025 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1027 zio
->io_prop
.zp_checksum
= checksum
;
1029 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1031 * zec checksums are necessarily destructive -- they modify
1032 * the end of the write buffer to hold the verifier/checksum.
1033 * Therefore, we must make a local copy in case the data is
1034 * being written to multiple places in parallel.
1036 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1037 abd_copy(wbuf
, data
, size
);
1039 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1046 * Create a child I/O to do some work for us.
1049 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1050 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1051 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1053 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1057 * vdev child I/Os do not propagate their error to the parent.
1058 * Therefore, for correct operation the caller *must* check for
1059 * and handle the error in the child i/o's done callback.
1060 * The only exceptions are i/os that we don't care about
1061 * (OPTIONAL or REPAIR).
1063 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1067 * In the common case, where the parent zio was to a normal vdev,
1068 * the child zio must be to a child vdev of that vdev. Otherwise,
1069 * the child zio must be to a top-level vdev.
1071 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
!= &vdev_indirect_ops
) {
1072 ASSERT3P(vd
->vdev_parent
, ==, pio
->io_vd
);
1074 ASSERT3P(vd
, ==, vd
->vdev_top
);
1077 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1079 * If we have the bp, then the child should perform the
1080 * checksum and the parent need not. This pushes error
1081 * detection as close to the leaves as possible and
1082 * eliminates redundant checksums in the interior nodes.
1084 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1085 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1088 if (vd
->vdev_ops
->vdev_op_leaf
) {
1089 ASSERT0(vd
->vdev_children
);
1090 offset
+= VDEV_LABEL_START_SIZE
;
1093 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1096 * If we've decided to do a repair, the write is not speculative --
1097 * even if the original read was.
1099 if (flags
& ZIO_FLAG_IO_REPAIR
)
1100 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1103 * If we're creating a child I/O that is not associated with a
1104 * top-level vdev, then the child zio is not an allocating I/O.
1105 * If this is a retried I/O then we ignore it since we will
1106 * have already processed the original allocating I/O.
1108 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1109 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1110 metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
);
1112 ASSERT(mc
->mc_alloc_throttle_enabled
);
1113 ASSERT(type
== ZIO_TYPE_WRITE
);
1114 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1115 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1116 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1117 pio
->io_child_type
== ZIO_CHILD_GANG
);
1119 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1122 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1123 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1124 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1125 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1127 zio
->io_physdone
= pio
->io_physdone
;
1128 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1129 zio
->io_logical
->io_phys_children
++;
1135 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1136 int type
, zio_priority_t priority
, enum zio_flag flags
,
1137 zio_done_func_t
*done
, void *private)
1141 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1143 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1144 data
, size
, size
, done
, private, type
, priority
,
1145 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1147 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1153 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1155 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1157 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1161 zio_shrink(zio_t
*zio
, uint64_t size
)
1163 ASSERT3P(zio
->io_executor
, ==, NULL
);
1164 ASSERT3P(zio
->io_orig_size
, ==, zio
->io_size
);
1165 ASSERT3U(size
, <=, zio
->io_size
);
1168 * We don't shrink for raidz because of problems with the
1169 * reconstruction when reading back less than the block size.
1170 * Note, BP_IS_RAIDZ() assumes no compression.
1172 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1173 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1174 /* we are not doing a raw write */
1175 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1176 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1181 * ==========================================================================
1182 * Prepare to read and write logical blocks
1183 * ==========================================================================
1187 zio_read_bp_init(zio_t
*zio
)
1189 blkptr_t
*bp
= zio
->io_bp
;
1191 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1193 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1194 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1195 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
1197 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1198 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1199 psize
, psize
, zio_decompress
);
1202 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1203 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1205 int psize
= BPE_GET_PSIZE(bp
);
1206 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1207 decode_embedded_bp_compressed(bp
, data
);
1208 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1210 ASSERT(!BP_IS_EMBEDDED(bp
));
1211 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1214 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1215 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1217 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1218 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1220 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1221 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1223 return (ZIO_PIPELINE_CONTINUE
);
1227 zio_write_bp_init(zio_t
*zio
)
1229 if (!IO_IS_ALLOCATING(zio
))
1230 return (ZIO_PIPELINE_CONTINUE
);
1232 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1234 if (zio
->io_bp_override
) {
1235 blkptr_t
*bp
= zio
->io_bp
;
1236 zio_prop_t
*zp
= &zio
->io_prop
;
1238 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1239 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1241 *bp
= *zio
->io_bp_override
;
1242 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1244 if (BP_IS_EMBEDDED(bp
))
1245 return (ZIO_PIPELINE_CONTINUE
);
1248 * If we've been overridden and nopwrite is set then
1249 * set the flag accordingly to indicate that a nopwrite
1250 * has already occurred.
1252 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1253 ASSERT(!zp
->zp_dedup
);
1254 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1255 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1256 return (ZIO_PIPELINE_CONTINUE
);
1259 ASSERT(!zp
->zp_nopwrite
);
1261 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1262 return (ZIO_PIPELINE_CONTINUE
);
1264 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1265 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1267 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
1268 BP_SET_DEDUP(bp
, 1);
1269 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1270 return (ZIO_PIPELINE_CONTINUE
);
1274 * We were unable to handle this as an override bp, treat
1275 * it as a regular write I/O.
1277 zio
->io_bp_override
= NULL
;
1278 *bp
= zio
->io_bp_orig
;
1279 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1282 return (ZIO_PIPELINE_CONTINUE
);
1286 zio_write_compress(zio_t
*zio
)
1288 spa_t
*spa
= zio
->io_spa
;
1289 zio_prop_t
*zp
= &zio
->io_prop
;
1290 enum zio_compress compress
= zp
->zp_compress
;
1291 blkptr_t
*bp
= zio
->io_bp
;
1292 uint64_t lsize
= zio
->io_lsize
;
1293 uint64_t psize
= zio
->io_size
;
1296 EQUIV(lsize
!= psize
, (zio
->io_flags
& ZIO_FLAG_RAW
) != 0);
1299 * If our children haven't all reached the ready stage,
1300 * wait for them and then repeat this pipeline stage.
1302 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1303 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1304 return (ZIO_PIPELINE_STOP
);
1306 if (!IO_IS_ALLOCATING(zio
))
1307 return (ZIO_PIPELINE_CONTINUE
);
1309 if (zio
->io_children_ready
!= NULL
) {
1311 * Now that all our children are ready, run the callback
1312 * associated with this zio in case it wants to modify the
1313 * data to be written.
1315 ASSERT3U(zp
->zp_level
, >, 0);
1316 zio
->io_children_ready(zio
);
1319 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1320 ASSERT(zio
->io_bp_override
== NULL
);
1322 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1324 * We're rewriting an existing block, which means we're
1325 * working on behalf of spa_sync(). For spa_sync() to
1326 * converge, it must eventually be the case that we don't
1327 * have to allocate new blocks. But compression changes
1328 * the blocksize, which forces a reallocate, and makes
1329 * convergence take longer. Therefore, after the first
1330 * few passes, stop compressing to ensure convergence.
1332 pass
= spa_sync_pass(spa
);
1334 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1335 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1336 ASSERT(!BP_GET_DEDUP(bp
));
1338 if (pass
>= zfs_sync_pass_dont_compress
)
1339 compress
= ZIO_COMPRESS_OFF
;
1341 /* Make sure someone doesn't change their mind on overwrites */
1342 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1343 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1346 /* If it's a compressed write that is not raw, compress the buffer. */
1347 if (compress
!= ZIO_COMPRESS_OFF
&& psize
== lsize
) {
1348 void *cbuf
= zio_buf_alloc(lsize
);
1349 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1350 if (psize
== 0 || psize
== lsize
) {
1351 compress
= ZIO_COMPRESS_OFF
;
1352 zio_buf_free(cbuf
, lsize
);
1353 } else if (!zp
->zp_dedup
&& psize
<= BPE_PAYLOAD_SIZE
&&
1354 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1355 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1356 encode_embedded_bp_compressed(bp
,
1357 cbuf
, compress
, lsize
, psize
);
1358 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1359 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1360 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1361 zio_buf_free(cbuf
, lsize
);
1362 bp
->blk_birth
= zio
->io_txg
;
1363 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1364 ASSERT(spa_feature_is_active(spa
,
1365 SPA_FEATURE_EMBEDDED_DATA
));
1366 return (ZIO_PIPELINE_CONTINUE
);
1369 * Round up compressed size up to the ashift
1370 * of the smallest-ashift device, and zero the tail.
1371 * This ensures that the compressed size of the BP
1372 * (and thus compressratio property) are correct,
1373 * in that we charge for the padding used to fill out
1376 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1377 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1378 1ULL << spa
->spa_min_ashift
);
1379 if (rounded
>= lsize
) {
1380 compress
= ZIO_COMPRESS_OFF
;
1381 zio_buf_free(cbuf
, lsize
);
1384 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1385 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1386 abd_zero_off(cdata
, psize
, rounded
- psize
);
1388 zio_push_transform(zio
, cdata
,
1389 psize
, lsize
, NULL
);
1394 * We were unable to handle this as an override bp, treat
1395 * it as a regular write I/O.
1397 zio
->io_bp_override
= NULL
;
1398 *bp
= zio
->io_bp_orig
;
1399 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1401 ASSERT3U(psize
, !=, 0);
1405 * The final pass of spa_sync() must be all rewrites, but the first
1406 * few passes offer a trade-off: allocating blocks defers convergence,
1407 * but newly allocated blocks are sequential, so they can be written
1408 * to disk faster. Therefore, we allow the first few passes of
1409 * spa_sync() to allocate new blocks, but force rewrites after that.
1410 * There should only be a handful of blocks after pass 1 in any case.
1412 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1413 BP_GET_PSIZE(bp
) == psize
&&
1414 pass
>= zfs_sync_pass_rewrite
) {
1416 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1417 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1418 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1421 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1425 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1426 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1427 BP_SET_LSIZE(bp
, lsize
);
1428 BP_SET_TYPE(bp
, zp
->zp_type
);
1429 BP_SET_LEVEL(bp
, zp
->zp_level
);
1430 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1432 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1434 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1435 BP_SET_LSIZE(bp
, lsize
);
1436 BP_SET_TYPE(bp
, zp
->zp_type
);
1437 BP_SET_LEVEL(bp
, zp
->zp_level
);
1438 BP_SET_PSIZE(bp
, psize
);
1439 BP_SET_COMPRESS(bp
, compress
);
1440 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1441 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1442 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1444 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1445 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1446 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1448 if (zp
->zp_nopwrite
) {
1449 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1450 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1451 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1454 return (ZIO_PIPELINE_CONTINUE
);
1458 zio_free_bp_init(zio_t
*zio
)
1460 blkptr_t
*bp
= zio
->io_bp
;
1462 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1463 if (BP_GET_DEDUP(bp
))
1464 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1467 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1469 return (ZIO_PIPELINE_CONTINUE
);
1473 * ==========================================================================
1474 * Execute the I/O pipeline
1475 * ==========================================================================
1479 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1481 spa_t
*spa
= zio
->io_spa
;
1482 zio_type_t t
= zio
->io_type
;
1483 int flags
= (cutinline
? TQ_FRONT
: 0);
1486 * If we're a config writer or a probe, the normal issue and
1487 * interrupt threads may all be blocked waiting for the config lock.
1488 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1490 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1494 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1496 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1500 * If this is a high priority I/O, then use the high priority taskq if
1503 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1504 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1507 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1510 * NB: We are assuming that the zio can only be dispatched
1511 * to a single taskq at a time. It would be a grievous error
1512 * to dispatch the zio to another taskq at the same time.
1514 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
1515 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1516 flags
, &zio
->io_tqent
);
1520 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1522 kthread_t
*executor
= zio
->io_executor
;
1523 spa_t
*spa
= zio
->io_spa
;
1525 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1526 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1528 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1529 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1538 zio_issue_async(zio_t
*zio
)
1540 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1542 return (ZIO_PIPELINE_STOP
);
1546 zio_interrupt(zio_t
*zio
)
1548 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1552 zio_delay_interrupt(zio_t
*zio
)
1555 * The timeout_generic() function isn't defined in userspace, so
1556 * rather than trying to implement the function, the zio delay
1557 * functionality has been disabled for userspace builds.
1562 * If io_target_timestamp is zero, then no delay has been registered
1563 * for this IO, thus jump to the end of this function and "skip" the
1564 * delay; issuing it directly to the zio layer.
1566 if (zio
->io_target_timestamp
!= 0) {
1567 hrtime_t now
= gethrtime();
1569 if (now
>= zio
->io_target_timestamp
) {
1571 * This IO has already taken longer than the target
1572 * delay to complete, so we don't want to delay it
1573 * any longer; we "miss" the delay and issue it
1574 * directly to the zio layer. This is likely due to
1575 * the target latency being set to a value less than
1576 * the underlying hardware can satisfy (e.g. delay
1577 * set to 1ms, but the disks take 10ms to complete an
1581 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1586 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1588 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1589 hrtime_t
, now
, hrtime_t
, diff
);
1591 (void) timeout_generic(CALLOUT_NORMAL
,
1592 (void (*)(void *))zio_interrupt
, zio
, diff
, 1, 0);
1599 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1604 * Execute the I/O pipeline until one of the following occurs:
1606 * (1) the I/O completes
1607 * (2) the pipeline stalls waiting for dependent child I/Os
1608 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1609 * (4) the I/O is delegated by vdev-level caching or aggregation
1610 * (5) the I/O is deferred due to vdev-level queueing
1611 * (6) the I/O is handed off to another thread.
1613 * In all cases, the pipeline stops whenever there's no CPU work; it never
1614 * burns a thread in cv_wait().
1616 * There's no locking on io_stage because there's no legitimate way
1617 * for multiple threads to be attempting to process the same I/O.
1619 static zio_pipe_stage_t
*zio_pipeline
[];
1622 zio_execute(zio_t
*zio
)
1624 zio
->io_executor
= curthread
;
1626 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1628 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1629 enum zio_stage pipeline
= zio
->io_pipeline
;
1630 enum zio_stage stage
= zio
->io_stage
;
1633 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1634 ASSERT(ISP2(stage
));
1635 ASSERT(zio
->io_stall
== NULL
);
1639 } while ((stage
& pipeline
) == 0);
1641 ASSERT(stage
<= ZIO_STAGE_DONE
);
1644 * If we are in interrupt context and this pipeline stage
1645 * will grab a config lock that is held across I/O,
1646 * or may wait for an I/O that needs an interrupt thread
1647 * to complete, issue async to avoid deadlock.
1649 * For VDEV_IO_START, we cut in line so that the io will
1650 * be sent to disk promptly.
1652 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1653 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1654 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1655 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1656 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1660 zio
->io_stage
= stage
;
1661 zio
->io_pipeline_trace
|= zio
->io_stage
;
1662 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1664 if (rv
== ZIO_PIPELINE_STOP
)
1667 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1672 * ==========================================================================
1673 * Initiate I/O, either sync or async
1674 * ==========================================================================
1677 zio_wait(zio_t
*zio
)
1681 ASSERT3P(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1682 ASSERT3P(zio
->io_executor
, ==, NULL
);
1684 zio
->io_waiter
= curthread
;
1685 ASSERT0(zio
->io_queued_timestamp
);
1686 zio
->io_queued_timestamp
= gethrtime();
1690 mutex_enter(&zio
->io_lock
);
1691 while (zio
->io_executor
!= NULL
)
1692 cv_wait(&zio
->io_cv
, &zio
->io_lock
);
1693 mutex_exit(&zio
->io_lock
);
1695 error
= zio
->io_error
;
1702 zio_nowait(zio_t
*zio
)
1704 ASSERT3P(zio
->io_executor
, ==, NULL
);
1706 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1707 zio_unique_parent(zio
) == NULL
) {
1709 * This is a logical async I/O with no parent to wait for it.
1710 * We add it to the spa_async_root_zio "Godfather" I/O which
1711 * will ensure they complete prior to unloading the pool.
1713 spa_t
*spa
= zio
->io_spa
;
1715 zio_add_child(spa
->spa_async_zio_root
[CPU_SEQID
], zio
);
1718 ASSERT0(zio
->io_queued_timestamp
);
1719 zio
->io_queued_timestamp
= gethrtime();
1724 * ==========================================================================
1725 * Reexecute, cancel, or suspend/resume failed I/O
1726 * ==========================================================================
1730 zio_reexecute(zio_t
*pio
)
1732 zio_t
*cio
, *cio_next
;
1734 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1735 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1736 ASSERT(pio
->io_gang_leader
== NULL
);
1737 ASSERT(pio
->io_gang_tree
== NULL
);
1739 pio
->io_flags
= pio
->io_orig_flags
;
1740 pio
->io_stage
= pio
->io_orig_stage
;
1741 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1742 pio
->io_reexecute
= 0;
1743 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1744 pio
->io_pipeline_trace
= 0;
1746 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1747 pio
->io_state
[w
] = 0;
1748 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1749 pio
->io_child_error
[c
] = 0;
1751 if (IO_IS_ALLOCATING(pio
))
1752 BP_ZERO(pio
->io_bp
);
1755 * As we reexecute pio's children, new children could be created.
1756 * New children go to the head of pio's io_child_list, however,
1757 * so we will (correctly) not reexecute them. The key is that
1758 * the remainder of pio's io_child_list, from 'cio_next' onward,
1759 * cannot be affected by any side effects of reexecuting 'cio'.
1761 zio_link_t
*zl
= NULL
;
1762 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1763 cio_next
= zio_walk_children(pio
, &zl
);
1764 mutex_enter(&pio
->io_lock
);
1765 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1766 pio
->io_children
[cio
->io_child_type
][w
]++;
1767 mutex_exit(&pio
->io_lock
);
1772 * Now that all children have been reexecuted, execute the parent.
1773 * We don't reexecute "The Godfather" I/O here as it's the
1774 * responsibility of the caller to wait on it.
1776 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1777 pio
->io_queued_timestamp
= gethrtime();
1783 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1785 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1786 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1787 "failure and the failure mode property for this pool "
1788 "is set to panic.", spa_name(spa
));
1790 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1792 mutex_enter(&spa
->spa_suspend_lock
);
1794 if (spa
->spa_suspend_zio_root
== NULL
)
1795 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1796 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1797 ZIO_FLAG_GODFATHER
);
1799 spa
->spa_suspended
= B_TRUE
;
1802 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1803 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1804 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1805 ASSERT(zio_unique_parent(zio
) == NULL
);
1806 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1807 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1810 mutex_exit(&spa
->spa_suspend_lock
);
1814 zio_resume(spa_t
*spa
)
1819 * Reexecute all previously suspended i/o.
1821 mutex_enter(&spa
->spa_suspend_lock
);
1822 spa
->spa_suspended
= B_FALSE
;
1823 cv_broadcast(&spa
->spa_suspend_cv
);
1824 pio
= spa
->spa_suspend_zio_root
;
1825 spa
->spa_suspend_zio_root
= NULL
;
1826 mutex_exit(&spa
->spa_suspend_lock
);
1832 return (zio_wait(pio
));
1836 zio_resume_wait(spa_t
*spa
)
1838 mutex_enter(&spa
->spa_suspend_lock
);
1839 while (spa_suspended(spa
))
1840 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1841 mutex_exit(&spa
->spa_suspend_lock
);
1845 * ==========================================================================
1848 * A gang block is a collection of small blocks that looks to the DMU
1849 * like one large block. When zio_dva_allocate() cannot find a block
1850 * of the requested size, due to either severe fragmentation or the pool
1851 * being nearly full, it calls zio_write_gang_block() to construct the
1852 * block from smaller fragments.
1854 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1855 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1856 * an indirect block: it's an array of block pointers. It consumes
1857 * only one sector and hence is allocatable regardless of fragmentation.
1858 * The gang header's bps point to its gang members, which hold the data.
1860 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1861 * as the verifier to ensure uniqueness of the SHA256 checksum.
1862 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1863 * not the gang header. This ensures that data block signatures (needed for
1864 * deduplication) are independent of how the block is physically stored.
1866 * Gang blocks can be nested: a gang member may itself be a gang block.
1867 * Thus every gang block is a tree in which root and all interior nodes are
1868 * gang headers, and the leaves are normal blocks that contain user data.
1869 * The root of the gang tree is called the gang leader.
1871 * To perform any operation (read, rewrite, free, claim) on a gang block,
1872 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1873 * in the io_gang_tree field of the original logical i/o by recursively
1874 * reading the gang leader and all gang headers below it. This yields
1875 * an in-core tree containing the contents of every gang header and the
1876 * bps for every constituent of the gang block.
1878 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1879 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1880 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1881 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1882 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1883 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1884 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1885 * of the gang header plus zio_checksum_compute() of the data to update the
1886 * gang header's blk_cksum as described above.
1888 * The two-phase assemble/issue model solves the problem of partial failure --
1889 * what if you'd freed part of a gang block but then couldn't read the
1890 * gang header for another part? Assembling the entire gang tree first
1891 * ensures that all the necessary gang header I/O has succeeded before
1892 * starting the actual work of free, claim, or write. Once the gang tree
1893 * is assembled, free and claim are in-memory operations that cannot fail.
1895 * In the event that a gang write fails, zio_dva_unallocate() walks the
1896 * gang tree to immediately free (i.e. insert back into the space map)
1897 * everything we've allocated. This ensures that we don't get ENOSPC
1898 * errors during repeated suspend/resume cycles due to a flaky device.
1900 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1901 * the gang tree, we won't modify the block, so we can safely defer the free
1902 * (knowing that the block is still intact). If we *can* assemble the gang
1903 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1904 * each constituent bp and we can allocate a new block on the next sync pass.
1906 * In all cases, the gang tree allows complete recovery from partial failure.
1907 * ==========================================================================
1911 zio_gang_issue_func_done(zio_t
*zio
)
1913 abd_put(zio
->io_abd
);
1917 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1923 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
1924 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
1925 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1926 &pio
->io_bookmark
));
1930 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1937 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1938 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1939 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
1940 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1943 * As we rewrite each gang header, the pipeline will compute
1944 * a new gang block header checksum for it; but no one will
1945 * compute a new data checksum, so we do that here. The one
1946 * exception is the gang leader: the pipeline already computed
1947 * its data checksum because that stage precedes gang assembly.
1948 * (Presently, nothing actually uses interior data checksums;
1949 * this is just good hygiene.)
1951 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1952 abd_t
*buf
= abd_get_offset(data
, offset
);
1954 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1955 buf
, BP_GET_PSIZE(bp
));
1960 * If we are here to damage data for testing purposes,
1961 * leave the GBH alone so that we can detect the damage.
1963 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1964 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1966 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1967 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
1968 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
1969 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1977 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1980 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1981 ZIO_GANG_CHILD_FLAGS(pio
)));
1986 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1989 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1990 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1993 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2002 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2004 static zio_gang_node_t
*
2005 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2007 zio_gang_node_t
*gn
;
2009 ASSERT(*gnpp
== NULL
);
2011 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2012 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2019 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2021 zio_gang_node_t
*gn
= *gnpp
;
2023 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2024 ASSERT(gn
->gn_child
[g
] == NULL
);
2026 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2027 kmem_free(gn
, sizeof (*gn
));
2032 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2034 zio_gang_node_t
*gn
= *gnpp
;
2039 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2040 zio_gang_tree_free(&gn
->gn_child
[g
]);
2042 zio_gang_node_free(gnpp
);
2046 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2048 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2049 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2051 ASSERT(gio
->io_gang_leader
== gio
);
2052 ASSERT(BP_IS_GANG(bp
));
2054 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2055 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2056 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2060 zio_gang_tree_assemble_done(zio_t
*zio
)
2062 zio_t
*gio
= zio
->io_gang_leader
;
2063 zio_gang_node_t
*gn
= zio
->io_private
;
2064 blkptr_t
*bp
= zio
->io_bp
;
2066 ASSERT(gio
== zio_unique_parent(zio
));
2067 ASSERT(zio
->io_child_count
== 0);
2072 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2073 if (BP_SHOULD_BYTESWAP(bp
))
2074 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2076 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2077 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2078 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2080 abd_put(zio
->io_abd
);
2082 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2083 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2084 if (!BP_IS_GANG(gbp
))
2086 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2091 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2094 zio_t
*gio
= pio
->io_gang_leader
;
2097 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2098 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2099 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2102 * If you're a gang header, your data is in gn->gn_gbh.
2103 * If you're a gang member, your data is in 'data' and gn == NULL.
2105 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2108 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2110 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2111 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2112 if (BP_IS_HOLE(gbp
))
2114 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2116 offset
+= BP_GET_PSIZE(gbp
);
2120 if (gn
== gio
->io_gang_tree
)
2121 ASSERT3U(gio
->io_size
, ==, offset
);
2128 zio_gang_assemble(zio_t
*zio
)
2130 blkptr_t
*bp
= zio
->io_bp
;
2132 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2133 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2135 zio
->io_gang_leader
= zio
;
2137 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2139 return (ZIO_PIPELINE_CONTINUE
);
2143 zio_gang_issue(zio_t
*zio
)
2145 blkptr_t
*bp
= zio
->io_bp
;
2147 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
2148 return (ZIO_PIPELINE_STOP
);
2150 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2151 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2153 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2154 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2157 zio_gang_tree_free(&zio
->io_gang_tree
);
2159 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2161 return (ZIO_PIPELINE_CONTINUE
);
2165 zio_write_gang_member_ready(zio_t
*zio
)
2167 zio_t
*pio
= zio_unique_parent(zio
);
2168 zio_t
*gio
= zio
->io_gang_leader
;
2169 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2170 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2173 if (BP_IS_HOLE(zio
->io_bp
))
2176 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2178 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2179 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2180 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2181 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2182 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2184 mutex_enter(&pio
->io_lock
);
2185 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2186 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2187 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2188 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2189 DVA_SET_ASIZE(&pdva
[d
], asize
);
2191 mutex_exit(&pio
->io_lock
);
2195 zio_write_gang_done(zio_t
*zio
)
2197 abd_put(zio
->io_abd
);
2201 zio_write_gang_block(zio_t
*pio
)
2203 spa_t
*spa
= pio
->io_spa
;
2204 metaslab_class_t
*mc
= spa_normal_class(spa
);
2205 blkptr_t
*bp
= pio
->io_bp
;
2206 zio_t
*gio
= pio
->io_gang_leader
;
2208 zio_gang_node_t
*gn
, **gnpp
;
2209 zio_gbh_phys_t
*gbh
;
2211 uint64_t txg
= pio
->io_txg
;
2212 uint64_t resid
= pio
->io_size
;
2214 int copies
= gio
->io_prop
.zp_copies
;
2215 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2219 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2220 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2221 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2222 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2224 flags
|= METASLAB_ASYNC_ALLOC
;
2225 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2228 * The logical zio has already placed a reservation for
2229 * 'copies' allocation slots but gang blocks may require
2230 * additional copies. These additional copies
2231 * (i.e. gbh_copies - copies) are guaranteed to succeed
2232 * since metaslab_class_throttle_reserve() always allows
2233 * additional reservations for gang blocks.
2235 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2239 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2240 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2241 &pio
->io_alloc_list
, pio
);
2243 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2244 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2245 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2248 * If we failed to allocate the gang block header then
2249 * we remove any additional allocation reservations that
2250 * we placed here. The original reservation will
2251 * be removed when the logical I/O goes to the ready
2254 metaslab_class_throttle_unreserve(mc
,
2255 gbh_copies
- copies
, pio
);
2257 pio
->io_error
= error
;
2258 return (ZIO_PIPELINE_CONTINUE
);
2262 gnpp
= &gio
->io_gang_tree
;
2264 gnpp
= pio
->io_private
;
2265 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2268 gn
= zio_gang_node_alloc(gnpp
);
2270 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2271 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2274 * Create the gang header.
2276 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2277 zio_write_gang_done
, NULL
, pio
->io_priority
,
2278 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2281 * Create and nowait the gang children.
2283 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2284 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2286 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2288 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2289 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2290 zp
.zp_type
= DMU_OT_NONE
;
2292 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2293 zp
.zp_dedup
= B_FALSE
;
2294 zp
.zp_dedup_verify
= B_FALSE
;
2295 zp
.zp_nopwrite
= B_FALSE
;
2297 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2298 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2299 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2300 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2301 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2303 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2304 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2305 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2308 * Gang children won't throttle but we should
2309 * account for their work, so reserve an allocation
2310 * slot for them here.
2312 VERIFY(metaslab_class_throttle_reserve(mc
,
2313 zp
.zp_copies
, cio
, flags
));
2319 * Set pio's pipeline to just wait for zio to finish.
2321 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2325 return (ZIO_PIPELINE_CONTINUE
);
2329 * The zio_nop_write stage in the pipeline determines if allocating a
2330 * new bp is necessary. The nopwrite feature can handle writes in
2331 * either syncing or open context (i.e. zil writes) and as a result is
2332 * mutually exclusive with dedup.
2334 * By leveraging a cryptographically secure checksum, such as SHA256, we
2335 * can compare the checksums of the new data and the old to determine if
2336 * allocating a new block is required. Note that our requirements for
2337 * cryptographic strength are fairly weak: there can't be any accidental
2338 * hash collisions, but we don't need to be secure against intentional
2339 * (malicious) collisions. To trigger a nopwrite, you have to be able
2340 * to write the file to begin with, and triggering an incorrect (hash
2341 * collision) nopwrite is no worse than simply writing to the file.
2342 * That said, there are no known attacks against the checksum algorithms
2343 * used for nopwrite, assuming that the salt and the checksums
2344 * themselves remain secret.
2347 zio_nop_write(zio_t
*zio
)
2349 blkptr_t
*bp
= zio
->io_bp
;
2350 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2351 zio_prop_t
*zp
= &zio
->io_prop
;
2353 ASSERT(BP_GET_LEVEL(bp
) == 0);
2354 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2355 ASSERT(zp
->zp_nopwrite
);
2356 ASSERT(!zp
->zp_dedup
);
2357 ASSERT(zio
->io_bp_override
== NULL
);
2358 ASSERT(IO_IS_ALLOCATING(zio
));
2361 * Check to see if the original bp and the new bp have matching
2362 * characteristics (i.e. same checksum, compression algorithms, etc).
2363 * If they don't then just continue with the pipeline which will
2364 * allocate a new bp.
2366 if (BP_IS_HOLE(bp_orig
) ||
2367 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2368 ZCHECKSUM_FLAG_NOPWRITE
) ||
2369 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2370 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2371 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2372 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2373 return (ZIO_PIPELINE_CONTINUE
);
2376 * If the checksums match then reset the pipeline so that we
2377 * avoid allocating a new bp and issuing any I/O.
2379 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2380 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2381 ZCHECKSUM_FLAG_NOPWRITE
);
2382 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2383 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2384 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2385 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2386 sizeof (uint64_t)) == 0);
2389 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2390 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2393 return (ZIO_PIPELINE_CONTINUE
);
2397 * ==========================================================================
2399 * ==========================================================================
2402 zio_ddt_child_read_done(zio_t
*zio
)
2404 blkptr_t
*bp
= zio
->io_bp
;
2405 ddt_entry_t
*dde
= zio
->io_private
;
2407 zio_t
*pio
= zio_unique_parent(zio
);
2409 mutex_enter(&pio
->io_lock
);
2410 ddp
= ddt_phys_select(dde
, bp
);
2411 if (zio
->io_error
== 0)
2412 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2414 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2415 dde
->dde_repair_abd
= zio
->io_abd
;
2417 abd_free(zio
->io_abd
);
2418 mutex_exit(&pio
->io_lock
);
2422 zio_ddt_read_start(zio_t
*zio
)
2424 blkptr_t
*bp
= zio
->io_bp
;
2426 ASSERT(BP_GET_DEDUP(bp
));
2427 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2428 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2430 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2431 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2432 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2433 ddt_phys_t
*ddp
= dde
->dde_phys
;
2434 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2437 ASSERT(zio
->io_vsd
== NULL
);
2440 if (ddp_self
== NULL
)
2441 return (ZIO_PIPELINE_CONTINUE
);
2443 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2444 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2446 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2448 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2449 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2450 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2451 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2452 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2454 return (ZIO_PIPELINE_CONTINUE
);
2457 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2458 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2459 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2461 return (ZIO_PIPELINE_CONTINUE
);
2465 zio_ddt_read_done(zio_t
*zio
)
2467 blkptr_t
*bp
= zio
->io_bp
;
2469 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
2470 return (ZIO_PIPELINE_STOP
);
2472 ASSERT(BP_GET_DEDUP(bp
));
2473 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2474 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2476 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2477 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2478 ddt_entry_t
*dde
= zio
->io_vsd
;
2480 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2481 return (ZIO_PIPELINE_CONTINUE
);
2484 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2485 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2486 return (ZIO_PIPELINE_STOP
);
2488 if (dde
->dde_repair_abd
!= NULL
) {
2489 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2491 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2493 ddt_repair_done(ddt
, dde
);
2497 ASSERT(zio
->io_vsd
== NULL
);
2499 return (ZIO_PIPELINE_CONTINUE
);
2503 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2505 spa_t
*spa
= zio
->io_spa
;
2506 boolean_t do_raw
= (zio
->io_flags
& ZIO_FLAG_RAW
);
2508 /* We should never get a raw, override zio */
2509 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2512 * Note: we compare the original data, not the transformed data,
2513 * because when zio->io_bp is an override bp, we will not have
2514 * pushed the I/O transforms. That's an important optimization
2515 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2517 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2518 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2521 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2522 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
,
2523 zio
->io_orig_size
) != 0);
2527 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2528 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2530 if (ddp
->ddp_phys_birth
!= 0) {
2531 arc_buf_t
*abuf
= NULL
;
2532 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2533 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2534 blkptr_t blk
= *zio
->io_bp
;
2537 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2542 * Intuitively, it would make more sense to compare
2543 * io_abd than io_orig_abd in the raw case since you
2544 * don't want to look at any transformations that have
2545 * happened to the data. However, for raw I/Os the
2546 * data will actually be the same in io_abd and
2547 * io_orig_abd, so all we have to do is issue this as
2551 zio_flags
|= ZIO_FLAG_RAW
;
2552 ASSERT3U(zio
->io_size
, ==, zio
->io_orig_size
);
2553 ASSERT0(abd_cmp(zio
->io_abd
, zio
->io_orig_abd
,
2555 ASSERT3P(zio
->io_transform_stack
, ==, NULL
);
2558 error
= arc_read(NULL
, spa
, &blk
,
2559 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2560 zio_flags
, &aflags
, &zio
->io_bookmark
);
2563 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
2564 abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2565 zio
->io_orig_size
) != 0)
2566 error
= SET_ERROR(EEXIST
);
2567 arc_buf_destroy(abuf
, &abuf
);
2571 return (error
!= 0);
2579 zio_ddt_child_write_ready(zio_t
*zio
)
2581 int p
= zio
->io_prop
.zp_copies
;
2582 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2583 ddt_entry_t
*dde
= zio
->io_private
;
2584 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2592 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2594 ddt_phys_fill(ddp
, zio
->io_bp
);
2596 zio_link_t
*zl
= NULL
;
2597 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2598 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2604 zio_ddt_child_write_done(zio_t
*zio
)
2606 int p
= zio
->io_prop
.zp_copies
;
2607 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2608 ddt_entry_t
*dde
= zio
->io_private
;
2609 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2613 ASSERT(ddp
->ddp_refcnt
== 0);
2614 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2615 dde
->dde_lead_zio
[p
] = NULL
;
2617 if (zio
->io_error
== 0) {
2618 zio_link_t
*zl
= NULL
;
2619 while (zio_walk_parents(zio
, &zl
) != NULL
)
2620 ddt_phys_addref(ddp
);
2622 ddt_phys_clear(ddp
);
2629 zio_ddt_ditto_write_done(zio_t
*zio
)
2631 int p
= DDT_PHYS_DITTO
;
2632 zio_prop_t
*zp
= &zio
->io_prop
;
2633 blkptr_t
*bp
= zio
->io_bp
;
2634 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2635 ddt_entry_t
*dde
= zio
->io_private
;
2636 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2637 ddt_key_t
*ddk
= &dde
->dde_key
;
2641 ASSERT(ddp
->ddp_refcnt
== 0);
2642 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2643 dde
->dde_lead_zio
[p
] = NULL
;
2645 if (zio
->io_error
== 0) {
2646 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2647 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2648 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2649 if (ddp
->ddp_phys_birth
!= 0)
2650 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2651 ddt_phys_fill(ddp
, bp
);
2658 zio_ddt_write(zio_t
*zio
)
2660 spa_t
*spa
= zio
->io_spa
;
2661 blkptr_t
*bp
= zio
->io_bp
;
2662 uint64_t txg
= zio
->io_txg
;
2663 zio_prop_t
*zp
= &zio
->io_prop
;
2664 int p
= zp
->zp_copies
;
2668 ddt_t
*ddt
= ddt_select(spa
, bp
);
2672 ASSERT(BP_GET_DEDUP(bp
));
2673 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2674 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2675 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2678 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2679 ddp
= &dde
->dde_phys
[p
];
2681 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2683 * If we're using a weak checksum, upgrade to a strong checksum
2684 * and try again. If we're already using a strong checksum,
2685 * we can't resolve it, so just convert to an ordinary write.
2686 * (And automatically e-mail a paper to Nature?)
2688 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2689 ZCHECKSUM_FLAG_DEDUP
)) {
2690 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2691 zio_pop_transforms(zio
);
2692 zio
->io_stage
= ZIO_STAGE_OPEN
;
2695 zp
->zp_dedup
= B_FALSE
;
2696 BP_SET_DEDUP(bp
, B_FALSE
);
2698 ASSERT(!BP_GET_DEDUP(bp
));
2699 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2701 return (ZIO_PIPELINE_CONTINUE
);
2704 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2705 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2707 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2708 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2709 zio_prop_t czp
= *zp
;
2711 czp
.zp_copies
= ditto_copies
;
2714 * If we arrived here with an override bp, we won't have run
2715 * the transform stack, so we won't have the data we need to
2716 * generate a child i/o. So, toss the override bp and restart.
2717 * This is safe, because using the override bp is just an
2718 * optimization; and it's rare, so the cost doesn't matter.
2720 if (zio
->io_bp_override
) {
2721 zio_pop_transforms(zio
);
2722 zio
->io_stage
= ZIO_STAGE_OPEN
;
2723 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2724 zio
->io_bp_override
= NULL
;
2727 return (ZIO_PIPELINE_CONTINUE
);
2730 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2731 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2732 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2733 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2735 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2736 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2739 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2740 if (ddp
->ddp_phys_birth
!= 0)
2741 ddt_bp_fill(ddp
, bp
, txg
);
2742 if (dde
->dde_lead_zio
[p
] != NULL
)
2743 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2745 ddt_phys_addref(ddp
);
2746 } else if (zio
->io_bp_override
) {
2747 ASSERT(bp
->blk_birth
== txg
);
2748 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2749 ddt_phys_fill(ddp
, bp
);
2750 ddt_phys_addref(ddp
);
2752 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2753 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
2754 zio_ddt_child_write_ready
, NULL
, NULL
,
2755 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2756 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2758 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2759 dde
->dde_lead_zio
[p
] = cio
;
2769 return (ZIO_PIPELINE_CONTINUE
);
2772 ddt_entry_t
*freedde
; /* for debugging */
2775 zio_ddt_free(zio_t
*zio
)
2777 spa_t
*spa
= zio
->io_spa
;
2778 blkptr_t
*bp
= zio
->io_bp
;
2779 ddt_t
*ddt
= ddt_select(spa
, bp
);
2783 ASSERT(BP_GET_DEDUP(bp
));
2784 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2787 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2788 ddp
= ddt_phys_select(dde
, bp
);
2789 ddt_phys_decref(ddp
);
2792 return (ZIO_PIPELINE_CONTINUE
);
2796 * ==========================================================================
2797 * Allocate and free blocks
2798 * ==========================================================================
2802 zio_io_to_allocate(spa_t
*spa
)
2806 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
2808 zio
= avl_first(&spa
->spa_alloc_tree
);
2812 ASSERT(IO_IS_ALLOCATING(zio
));
2815 * Try to place a reservation for this zio. If we're unable to
2816 * reserve then we throttle.
2818 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
2819 zio
->io_prop
.zp_copies
, zio
, 0)) {
2823 avl_remove(&spa
->spa_alloc_tree
, zio
);
2824 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
2830 zio_dva_throttle(zio_t
*zio
)
2832 spa_t
*spa
= zio
->io_spa
;
2835 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
2836 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
2837 zio
->io_child_type
== ZIO_CHILD_GANG
||
2838 zio
->io_flags
& ZIO_FLAG_NODATA
) {
2839 return (ZIO_PIPELINE_CONTINUE
);
2842 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2844 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2845 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2847 mutex_enter(&spa
->spa_alloc_lock
);
2849 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2850 avl_add(&spa
->spa_alloc_tree
, zio
);
2852 nio
= zio_io_to_allocate(zio
->io_spa
);
2853 mutex_exit(&spa
->spa_alloc_lock
);
2856 return (ZIO_PIPELINE_CONTINUE
);
2859 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2861 * We are passing control to a new zio so make sure that
2862 * it is processed by a different thread. We do this to
2863 * avoid stack overflows that can occur when parents are
2864 * throttled and children are making progress. We allow
2865 * it to go to the head of the taskq since it's already
2868 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2870 return (ZIO_PIPELINE_STOP
);
2874 zio_allocate_dispatch(spa_t
*spa
)
2878 mutex_enter(&spa
->spa_alloc_lock
);
2879 zio
= zio_io_to_allocate(spa
);
2880 mutex_exit(&spa
->spa_alloc_lock
);
2884 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
2885 ASSERT0(zio
->io_error
);
2886 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2890 zio_dva_allocate(zio_t
*zio
)
2892 spa_t
*spa
= zio
->io_spa
;
2893 metaslab_class_t
*mc
= spa_normal_class(spa
);
2894 blkptr_t
*bp
= zio
->io_bp
;
2898 if (zio
->io_gang_leader
== NULL
) {
2899 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2900 zio
->io_gang_leader
= zio
;
2903 ASSERT(BP_IS_HOLE(bp
));
2904 ASSERT0(BP_GET_NDVAS(bp
));
2905 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2906 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2907 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2909 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2910 flags
|= METASLAB_DONT_THROTTLE
;
2912 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
) {
2913 flags
|= METASLAB_GANG_CHILD
;
2915 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
) {
2916 flags
|= METASLAB_ASYNC_ALLOC
;
2919 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2920 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
2921 &zio
->io_alloc_list
, zio
);
2924 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
2925 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
2927 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2928 return (zio_write_gang_block(zio
));
2929 zio
->io_error
= error
;
2932 return (ZIO_PIPELINE_CONTINUE
);
2936 zio_dva_free(zio_t
*zio
)
2938 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2940 return (ZIO_PIPELINE_CONTINUE
);
2944 zio_dva_claim(zio_t
*zio
)
2948 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2950 zio
->io_error
= error
;
2952 return (ZIO_PIPELINE_CONTINUE
);
2956 * Undo an allocation. This is used by zio_done() when an I/O fails
2957 * and we want to give back the block we just allocated.
2958 * This handles both normal blocks and gang blocks.
2961 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2963 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2964 ASSERT(zio
->io_bp_override
== NULL
);
2966 if (!BP_IS_HOLE(bp
))
2967 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2970 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2971 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2972 &gn
->gn_gbh
->zg_blkptr
[g
]);
2978 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2981 zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
, blkptr_t
*old_bp
,
2982 uint64_t size
, boolean_t
*slog
)
2985 zio_alloc_list_t io_alloc_list
;
2987 ASSERT(txg
> spa_syncing_txg(spa
));
2989 metaslab_trace_init(&io_alloc_list
);
2990 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
2991 txg
, old_bp
, METASLAB_HINTBP_AVOID
, &io_alloc_list
, NULL
);
2995 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
2996 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
,
2997 &io_alloc_list
, NULL
);
3001 metaslab_trace_fini(&io_alloc_list
);
3004 BP_SET_LSIZE(new_bp
, size
);
3005 BP_SET_PSIZE(new_bp
, size
);
3006 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3007 BP_SET_CHECKSUM(new_bp
,
3008 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3009 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3010 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3011 BP_SET_LEVEL(new_bp
, 0);
3012 BP_SET_DEDUP(new_bp
, 0);
3013 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3015 zfs_dbgmsg("%s: zil block allocation failure: "
3016 "size %llu, error %d", spa_name(spa
), size
, error
);
3023 * Free an intent log block.
3026 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
3028 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
3029 ASSERT(!BP_IS_GANG(bp
));
3031 zio_free(spa
, txg
, bp
);
3035 * ==========================================================================
3036 * Read and write to physical devices
3037 * ==========================================================================
3042 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3043 * stops after this stage and will resume upon I/O completion.
3044 * However, there are instances where the vdev layer may need to
3045 * continue the pipeline when an I/O was not issued. Since the I/O
3046 * that was sent to the vdev layer might be different than the one
3047 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3048 * force the underlying vdev layers to call either zio_execute() or
3049 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3052 zio_vdev_io_start(zio_t
*zio
)
3054 vdev_t
*vd
= zio
->io_vd
;
3056 spa_t
*spa
= zio
->io_spa
;
3058 ASSERT(zio
->io_error
== 0);
3059 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3062 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3063 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3066 * The mirror_ops handle multiple DVAs in a single BP.
3068 vdev_mirror_ops
.vdev_op_io_start(zio
);
3069 return (ZIO_PIPELINE_STOP
);
3072 ASSERT3P(zio
->io_logical
, !=, zio
);
3073 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3074 ASSERT(spa
->spa_trust_config
);
3076 if (zio
->io_vd
->vdev_removing
) {
3077 ASSERT(zio
->io_flags
&
3078 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3079 ZIO_FLAG_INDUCE_DAMAGE
));
3084 * We keep track of time-sensitive I/Os so that the scan thread
3085 * can quickly react to certain workloads. In particular, we care
3086 * about non-scrubbing, top-level reads and writes with the following
3088 * - synchronous writes of user data to non-slog devices
3089 * - any reads of user data
3090 * When these conditions are met, adjust the timestamp of spa_last_io
3091 * which allows the scan thread to adjust its workload accordingly.
3093 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
3094 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
3095 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
3096 zio
->io_txg
!= spa_syncing_txg(spa
)) {
3097 uint64_t old
= spa
->spa_last_io
;
3098 uint64_t new = ddi_get_lbolt64();
3100 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
3103 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3105 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3106 P2PHASE(zio
->io_size
, align
) != 0) {
3107 /* Transform logical writes to be a full physical block size. */
3108 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3109 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3110 ASSERT(vd
== vd
->vdev_top
);
3111 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3112 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3113 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3115 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3119 * If this is not a physical io, make sure that it is properly aligned
3120 * before proceeding.
3122 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3123 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3124 ASSERT0(P2PHASE(zio
->io_size
, align
));
3127 * For physical writes, we allow 512b aligned writes and assume
3128 * the device will perform a read-modify-write as necessary.
3130 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3131 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3134 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3137 * If this is a repair I/O, and there's no self-healing involved --
3138 * that is, we're just resilvering what we expect to resilver --
3139 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3140 * This prevents spurious resilvering with nested replication.
3141 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3142 * A is out of date, we'll read from C+D, then use the data to
3143 * resilver A+B -- but we don't actually want to resilver B, just A.
3144 * The top-level mirror has no way to know this, so instead we just
3145 * discard unnecessary repairs as we work our way down the vdev tree.
3146 * The same logic applies to any form of nested replication:
3147 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3149 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3150 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3151 zio
->io_txg
!= 0 && /* not a delegated i/o */
3152 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3153 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3154 zio_vdev_io_bypass(zio
);
3155 return (ZIO_PIPELINE_CONTINUE
);
3158 if (vd
->vdev_ops
->vdev_op_leaf
&&
3159 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3161 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3162 return (ZIO_PIPELINE_CONTINUE
);
3164 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3165 return (ZIO_PIPELINE_STOP
);
3167 if (!vdev_accessible(vd
, zio
)) {
3168 zio
->io_error
= SET_ERROR(ENXIO
);
3170 return (ZIO_PIPELINE_STOP
);
3174 vd
->vdev_ops
->vdev_op_io_start(zio
);
3175 return (ZIO_PIPELINE_STOP
);
3179 zio_vdev_io_done(zio_t
*zio
)
3181 vdev_t
*vd
= zio
->io_vd
;
3182 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3183 boolean_t unexpected_error
= B_FALSE
;
3185 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3186 return (ZIO_PIPELINE_STOP
);
3188 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3190 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3192 vdev_queue_io_done(zio
);
3194 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3195 vdev_cache_write(zio
);
3197 if (zio_injection_enabled
&& zio
->io_error
== 0)
3198 zio
->io_error
= zio_handle_device_injection(vd
,
3201 if (zio_injection_enabled
&& zio
->io_error
== 0)
3202 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3204 if (zio
->io_error
) {
3205 if (!vdev_accessible(vd
, zio
)) {
3206 zio
->io_error
= SET_ERROR(ENXIO
);
3208 unexpected_error
= B_TRUE
;
3213 ops
->vdev_op_io_done(zio
);
3215 if (unexpected_error
)
3216 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3218 return (ZIO_PIPELINE_CONTINUE
);
3222 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3223 * disk, and use that to finish the checksum ereport later.
3226 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3227 const void *good_buf
)
3229 /* no processing needed */
3230 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3235 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3237 void *buf
= zio_buf_alloc(zio
->io_size
);
3239 abd_copy_to_buf(buf
, zio
->io_abd
, zio
->io_size
);
3241 zcr
->zcr_cbinfo
= zio
->io_size
;
3242 zcr
->zcr_cbdata
= buf
;
3243 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3244 zcr
->zcr_free
= zio_buf_free
;
3248 zio_vdev_io_assess(zio_t
*zio
)
3250 vdev_t
*vd
= zio
->io_vd
;
3252 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3253 return (ZIO_PIPELINE_STOP
);
3255 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3256 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3258 if (zio
->io_vsd
!= NULL
) {
3259 zio
->io_vsd_ops
->vsd_free(zio
);
3263 if (zio_injection_enabled
&& zio
->io_error
== 0)
3264 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3267 * If the I/O failed, determine whether we should attempt to retry it.
3269 * On retry, we cut in line in the issue queue, since we don't want
3270 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3272 if (zio
->io_error
&& vd
== NULL
&&
3273 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3274 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3275 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3277 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3278 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3279 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3280 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3281 zio_requeue_io_start_cut_in_line
);
3282 return (ZIO_PIPELINE_STOP
);
3286 * If we got an error on a leaf device, convert it to ENXIO
3287 * if the device is not accessible at all.
3289 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3290 !vdev_accessible(vd
, zio
))
3291 zio
->io_error
= SET_ERROR(ENXIO
);
3294 * If we can't write to an interior vdev (mirror or RAID-Z),
3295 * set vdev_cant_write so that we stop trying to allocate from it.
3297 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3298 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3299 vd
->vdev_cant_write
= B_TRUE
;
3303 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3304 * attempts will ever succeed. In this case we set a persistent bit so
3305 * that we don't bother with it in the future.
3307 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3308 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3309 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3310 vd
->vdev_nowritecache
= B_TRUE
;
3313 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3315 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3316 zio
->io_physdone
!= NULL
) {
3317 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3318 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3319 zio
->io_physdone(zio
->io_logical
);
3322 return (ZIO_PIPELINE_CONTINUE
);
3326 zio_vdev_io_reissue(zio_t
*zio
)
3328 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3329 ASSERT(zio
->io_error
== 0);
3331 zio
->io_stage
>>= 1;
3335 zio_vdev_io_redone(zio_t
*zio
)
3337 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3339 zio
->io_stage
>>= 1;
3343 zio_vdev_io_bypass(zio_t
*zio
)
3345 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3346 ASSERT(zio
->io_error
== 0);
3348 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3349 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3353 * ==========================================================================
3354 * Generate and verify checksums
3355 * ==========================================================================
3358 zio_checksum_generate(zio_t
*zio
)
3360 blkptr_t
*bp
= zio
->io_bp
;
3361 enum zio_checksum checksum
;
3365 * This is zio_write_phys().
3366 * We're either generating a label checksum, or none at all.
3368 checksum
= zio
->io_prop
.zp_checksum
;
3370 if (checksum
== ZIO_CHECKSUM_OFF
)
3371 return (ZIO_PIPELINE_CONTINUE
);
3373 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3375 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3376 ASSERT(!IO_IS_ALLOCATING(zio
));
3377 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3379 checksum
= BP_GET_CHECKSUM(bp
);
3383 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3385 return (ZIO_PIPELINE_CONTINUE
);
3389 zio_checksum_verify(zio_t
*zio
)
3391 zio_bad_cksum_t info
;
3392 blkptr_t
*bp
= zio
->io_bp
;
3395 ASSERT(zio
->io_vd
!= NULL
);
3399 * This is zio_read_phys().
3400 * We're either verifying a label checksum, or nothing at all.
3402 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3403 return (ZIO_PIPELINE_CONTINUE
);
3405 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3408 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3409 zio
->io_error
= error
;
3410 if (error
== ECKSUM
&&
3411 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3412 zfs_ereport_start_checksum(zio
->io_spa
,
3413 zio
->io_vd
, zio
, zio
->io_offset
,
3414 zio
->io_size
, NULL
, &info
);
3418 return (ZIO_PIPELINE_CONTINUE
);
3422 * Called by RAID-Z to ensure we don't compute the checksum twice.
3425 zio_checksum_verified(zio_t
*zio
)
3427 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3431 * ==========================================================================
3432 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3433 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3434 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3435 * indicate errors that are specific to one I/O, and most likely permanent.
3436 * Any other error is presumed to be worse because we weren't expecting it.
3437 * ==========================================================================
3440 zio_worst_error(int e1
, int e2
)
3442 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3445 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3446 if (e1
== zio_error_rank
[r1
])
3449 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3450 if (e2
== zio_error_rank
[r2
])
3453 return (r1
> r2
? e1
: e2
);
3457 * ==========================================================================
3459 * ==========================================================================
3462 zio_ready(zio_t
*zio
)
3464 blkptr_t
*bp
= zio
->io_bp
;
3465 zio_t
*pio
, *pio_next
;
3466 zio_link_t
*zl
= NULL
;
3468 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
3469 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
3470 return (ZIO_PIPELINE_STOP
);
3472 if (zio
->io_ready
) {
3473 ASSERT(IO_IS_ALLOCATING(zio
));
3474 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3475 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3476 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3481 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3482 zio
->io_bp_copy
= *bp
;
3484 if (zio
->io_error
!= 0) {
3485 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3487 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3488 ASSERT(IO_IS_ALLOCATING(zio
));
3489 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3491 * We were unable to allocate anything, unreserve and
3492 * issue the next I/O to allocate.
3494 metaslab_class_throttle_unreserve(
3495 spa_normal_class(zio
->io_spa
),
3496 zio
->io_prop
.zp_copies
, zio
);
3497 zio_allocate_dispatch(zio
->io_spa
);
3501 mutex_enter(&zio
->io_lock
);
3502 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3503 pio
= zio_walk_parents(zio
, &zl
);
3504 mutex_exit(&zio
->io_lock
);
3507 * As we notify zio's parents, new parents could be added.
3508 * New parents go to the head of zio's io_parent_list, however,
3509 * so we will (correctly) not notify them. The remainder of zio's
3510 * io_parent_list, from 'pio_next' onward, cannot change because
3511 * all parents must wait for us to be done before they can be done.
3513 for (; pio
!= NULL
; pio
= pio_next
) {
3514 pio_next
= zio_walk_parents(zio
, &zl
);
3515 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3518 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3519 if (BP_IS_GANG(bp
)) {
3520 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3522 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3523 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3527 if (zio_injection_enabled
&&
3528 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3529 zio_handle_ignored_writes(zio
);
3531 return (ZIO_PIPELINE_CONTINUE
);
3535 * Update the allocation throttle accounting.
3538 zio_dva_throttle_done(zio_t
*zio
)
3540 zio_t
*lio
= zio
->io_logical
;
3541 zio_t
*pio
= zio_unique_parent(zio
);
3542 vdev_t
*vd
= zio
->io_vd
;
3543 int flags
= METASLAB_ASYNC_ALLOC
;
3545 ASSERT3P(zio
->io_bp
, !=, NULL
);
3546 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3547 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3548 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3550 ASSERT3P(vd
, ==, vd
->vdev_top
);
3551 ASSERT(!(zio
->io_flags
& (ZIO_FLAG_IO_REPAIR
| ZIO_FLAG_IO_RETRY
)));
3552 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3553 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3554 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3557 * Parents of gang children can have two flavors -- ones that
3558 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3559 * and ones that allocated the constituent blocks. The allocation
3560 * throttle needs to know the allocating parent zio so we must find
3563 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3565 * If our parent is a rewrite gang child then our grandparent
3566 * would have been the one that performed the allocation.
3568 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3569 pio
= zio_unique_parent(pio
);
3570 flags
|= METASLAB_GANG_CHILD
;
3573 ASSERT(IO_IS_ALLOCATING(pio
));
3574 ASSERT3P(zio
, !=, zio
->io_logical
);
3575 ASSERT(zio
->io_logical
!= NULL
);
3576 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3577 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3579 mutex_enter(&pio
->io_lock
);
3580 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
3581 mutex_exit(&pio
->io_lock
);
3583 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
3587 * Call into the pipeline to see if there is more work that
3588 * needs to be done. If there is work to be done it will be
3589 * dispatched to another taskq thread.
3591 zio_allocate_dispatch(zio
->io_spa
);
3595 zio_done(zio_t
*zio
)
3597 spa_t
*spa
= zio
->io_spa
;
3598 zio_t
*lio
= zio
->io_logical
;
3599 blkptr_t
*bp
= zio
->io_bp
;
3600 vdev_t
*vd
= zio
->io_vd
;
3601 uint64_t psize
= zio
->io_size
;
3602 zio_t
*pio
, *pio_next
;
3603 metaslab_class_t
*mc
= spa_normal_class(spa
);
3604 zio_link_t
*zl
= NULL
;
3607 * If our children haven't all completed,
3608 * wait for them and then repeat this pipeline stage.
3610 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
3611 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
3612 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
3613 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
3614 return (ZIO_PIPELINE_STOP
);
3617 * If the allocation throttle is enabled, then update the accounting.
3618 * We only track child I/Os that are part of an allocating async
3619 * write. We must do this since the allocation is performed
3620 * by the logical I/O but the actual write is done by child I/Os.
3622 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
3623 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
3624 ASSERT(mc
->mc_alloc_throttle_enabled
);
3625 zio_dva_throttle_done(zio
);
3629 * If the allocation throttle is enabled, verify that
3630 * we have decremented the refcounts for every I/O that was throttled.
3632 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3633 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3634 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3636 metaslab_group_alloc_verify(spa
, zio
->io_bp
, zio
);
3637 VERIFY(refcount_not_held(&mc
->mc_alloc_slots
, zio
));
3640 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
3641 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
3642 ASSERT(zio
->io_children
[c
][w
] == 0);
3644 if (bp
!= NULL
&& !BP_IS_EMBEDDED(bp
)) {
3645 ASSERT(bp
->blk_pad
[0] == 0);
3646 ASSERT(bp
->blk_pad
[1] == 0);
3647 ASSERT(bcmp(bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
3648 (bp
== zio_unique_parent(zio
)->io_bp
));
3649 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(bp
) &&
3650 zio
->io_bp_override
== NULL
&&
3651 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
3652 ASSERT(!BP_SHOULD_BYTESWAP(bp
));
3653 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(bp
));
3654 ASSERT(BP_COUNT_GANG(bp
) == 0 ||
3655 (BP_COUNT_GANG(bp
) == BP_GET_NDVAS(bp
)));
3657 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
3658 VERIFY(BP_EQUAL(bp
, &zio
->io_bp_orig
));
3662 * If there were child vdev/gang/ddt errors, they apply to us now.
3664 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
3665 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
3666 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
3669 * If the I/O on the transformed data was successful, generate any
3670 * checksum reports now while we still have the transformed data.
3672 if (zio
->io_error
== 0) {
3673 while (zio
->io_cksum_report
!= NULL
) {
3674 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3675 uint64_t align
= zcr
->zcr_align
;
3676 uint64_t asize
= P2ROUNDUP(psize
, align
);
3678 abd_t
*adata
= zio
->io_abd
;
3680 if (asize
!= psize
) {
3681 adata
= abd_alloc_linear(asize
, B_TRUE
);
3682 abd_copy(adata
, zio
->io_abd
, psize
);
3683 abd_zero_off(adata
, psize
, asize
- psize
);
3687 abuf
= abd_borrow_buf_copy(adata
, asize
);
3689 zio
->io_cksum_report
= zcr
->zcr_next
;
3690 zcr
->zcr_next
= NULL
;
3691 zcr
->zcr_finish(zcr
, abuf
);
3692 zfs_ereport_free_checksum(zcr
);
3695 abd_return_buf(adata
, abuf
, asize
);
3702 zio_pop_transforms(zio
); /* note: may set zio->io_error */
3704 vdev_stat_update(zio
, psize
);
3706 if (zio
->io_error
) {
3708 * If this I/O is attached to a particular vdev,
3709 * generate an error message describing the I/O failure
3710 * at the block level. We ignore these errors if the
3711 * device is currently unavailable.
3713 if (zio
->io_error
!= ECKSUM
&& vd
!= NULL
&& !vdev_is_dead(vd
))
3714 zfs_ereport_post(FM_EREPORT_ZFS_IO
, spa
, vd
, zio
, 0, 0);
3716 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
3717 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
3720 * For logical I/O requests, tell the SPA to log the
3721 * error and generate a logical data ereport.
3723 spa_log_error(spa
, zio
);
3724 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, spa
, NULL
, zio
,
3729 if (zio
->io_error
&& zio
== lio
) {
3731 * Determine whether zio should be reexecuted. This will
3732 * propagate all the way to the root via zio_notify_parent().
3734 ASSERT(vd
== NULL
&& bp
!= NULL
);
3735 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3737 if (IO_IS_ALLOCATING(zio
) &&
3738 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
3739 if (zio
->io_error
!= ENOSPC
)
3740 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
3742 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3745 if ((zio
->io_type
== ZIO_TYPE_READ
||
3746 zio
->io_type
== ZIO_TYPE_FREE
) &&
3747 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
3748 zio
->io_error
== ENXIO
&&
3749 spa_load_state(spa
) == SPA_LOAD_NONE
&&
3750 spa_get_failmode(spa
) != ZIO_FAILURE_MODE_CONTINUE
)
3751 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3753 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
3754 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3757 * Here is a possibly good place to attempt to do
3758 * either combinatorial reconstruction or error correction
3759 * based on checksums. It also might be a good place
3760 * to send out preliminary ereports before we suspend
3766 * If there were logical child errors, they apply to us now.
3767 * We defer this until now to avoid conflating logical child
3768 * errors with errors that happened to the zio itself when
3769 * updating vdev stats and reporting FMA events above.
3771 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
3773 if ((zio
->io_error
|| zio
->io_reexecute
) &&
3774 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
3775 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
3776 zio_dva_unallocate(zio
, zio
->io_gang_tree
, bp
);
3778 zio_gang_tree_free(&zio
->io_gang_tree
);
3781 * Godfather I/Os should never suspend.
3783 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3784 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
3785 zio
->io_reexecute
= 0;
3787 if (zio
->io_reexecute
) {
3789 * This is a logical I/O that wants to reexecute.
3791 * Reexecute is top-down. When an i/o fails, if it's not
3792 * the root, it simply notifies its parent and sticks around.
3793 * The parent, seeing that it still has children in zio_done(),
3794 * does the same. This percolates all the way up to the root.
3795 * The root i/o will reexecute or suspend the entire tree.
3797 * This approach ensures that zio_reexecute() honors
3798 * all the original i/o dependency relationships, e.g.
3799 * parents not executing until children are ready.
3801 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3803 zio
->io_gang_leader
= NULL
;
3805 mutex_enter(&zio
->io_lock
);
3806 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3807 mutex_exit(&zio
->io_lock
);
3810 * "The Godfather" I/O monitors its children but is
3811 * not a true parent to them. It will track them through
3812 * the pipeline but severs its ties whenever they get into
3813 * trouble (e.g. suspended). This allows "The Godfather"
3814 * I/O to return status without blocking.
3817 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
3819 zio_link_t
*remove_zl
= zl
;
3820 pio_next
= zio_walk_parents(zio
, &zl
);
3822 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3823 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
3824 zio_remove_child(pio
, zio
, remove_zl
);
3825 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3829 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
3831 * We're not a root i/o, so there's nothing to do
3832 * but notify our parent. Don't propagate errors
3833 * upward since we haven't permanently failed yet.
3835 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
3836 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
3837 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3838 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
3840 * We'd fail again if we reexecuted now, so suspend
3841 * until conditions improve (e.g. device comes online).
3843 zio_suspend(spa
, zio
);
3846 * Reexecution is potentially a huge amount of work.
3847 * Hand it off to the otherwise-unused claim taskq.
3849 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
3850 spa_taskq_dispatch_ent(spa
, ZIO_TYPE_CLAIM
,
3851 ZIO_TASKQ_ISSUE
, (task_func_t
*)zio_reexecute
, zio
,
3854 return (ZIO_PIPELINE_STOP
);
3857 ASSERT(zio
->io_child_count
== 0);
3858 ASSERT(zio
->io_reexecute
== 0);
3859 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
3862 * Report any checksum errors, since the I/O is complete.
3864 while (zio
->io_cksum_report
!= NULL
) {
3865 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3866 zio
->io_cksum_report
= zcr
->zcr_next
;
3867 zcr
->zcr_next
= NULL
;
3868 zcr
->zcr_finish(zcr
, NULL
);
3869 zfs_ereport_free_checksum(zcr
);
3873 * It is the responsibility of the done callback to ensure that this
3874 * particular zio is no longer discoverable for adoption, and as
3875 * such, cannot acquire any new parents.
3880 mutex_enter(&zio
->io_lock
);
3881 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3882 mutex_exit(&zio
->io_lock
);
3885 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
3886 zio_link_t
*remove_zl
= zl
;
3887 pio_next
= zio_walk_parents(zio
, &zl
);
3888 zio_remove_child(pio
, zio
, remove_zl
);
3889 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3892 if (zio
->io_waiter
!= NULL
) {
3893 mutex_enter(&zio
->io_lock
);
3894 zio
->io_executor
= NULL
;
3895 cv_broadcast(&zio
->io_cv
);
3896 mutex_exit(&zio
->io_lock
);
3901 return (ZIO_PIPELINE_STOP
);
3905 * ==========================================================================
3906 * I/O pipeline definition
3907 * ==========================================================================
3909 static zio_pipe_stage_t
*zio_pipeline
[] = {
3916 zio_checksum_generate
,
3932 zio_checksum_verify
,
3940 * Compare two zbookmark_phys_t's to see which we would reach first in a
3941 * pre-order traversal of the object tree.
3943 * This is simple in every case aside from the meta-dnode object. For all other
3944 * objects, we traverse them in order (object 1 before object 2, and so on).
3945 * However, all of these objects are traversed while traversing object 0, since
3946 * the data it points to is the list of objects. Thus, we need to convert to a
3947 * canonical representation so we can compare meta-dnode bookmarks to
3948 * non-meta-dnode bookmarks.
3950 * We do this by calculating "equivalents" for each field of the zbookmark.
3951 * zbookmarks outside of the meta-dnode use their own object and level, and
3952 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3953 * blocks this bookmark refers to) by multiplying their blkid by their span
3954 * (the number of L0 blocks contained within one block at their level).
3955 * zbookmarks inside the meta-dnode calculate their object equivalent
3956 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3957 * level + 1<<31 (any value larger than a level could ever be) for their level.
3958 * This causes them to always compare before a bookmark in their object
3959 * equivalent, compare appropriately to bookmarks in other objects, and to
3960 * compare appropriately to other bookmarks in the meta-dnode.
3963 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
3964 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
3967 * These variables represent the "equivalent" values for the zbookmark,
3968 * after converting zbookmarks inside the meta dnode to their
3969 * normal-object equivalents.
3971 uint64_t zb1obj
, zb2obj
;
3972 uint64_t zb1L0
, zb2L0
;
3973 uint64_t zb1level
, zb2level
;
3975 if (zb1
->zb_object
== zb2
->zb_object
&&
3976 zb1
->zb_level
== zb2
->zb_level
&&
3977 zb1
->zb_blkid
== zb2
->zb_blkid
)
3981 * BP_SPANB calculates the span in blocks.
3983 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
3984 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
3986 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
3987 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3989 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
3991 zb1obj
= zb1
->zb_object
;
3992 zb1level
= zb1
->zb_level
;
3995 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
3996 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
3998 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4000 zb2obj
= zb2
->zb_object
;
4001 zb2level
= zb2
->zb_level
;
4004 /* Now that we have a canonical representation, do the comparison. */
4005 if (zb1obj
!= zb2obj
)
4006 return (zb1obj
< zb2obj
? -1 : 1);
4007 else if (zb1L0
!= zb2L0
)
4008 return (zb1L0
< zb2L0
? -1 : 1);
4009 else if (zb1level
!= zb2level
)
4010 return (zb1level
> zb2level
? -1 : 1);
4012 * This can (theoretically) happen if the bookmarks have the same object
4013 * and level, but different blkids, if the block sizes are not the same.
4014 * There is presently no way to change the indirect block sizes
4020 * This function checks the following: given that last_block is the place that
4021 * our traversal stopped last time, does that guarantee that we've visited
4022 * every node under subtree_root? Therefore, we can't just use the raw output
4023 * of zbookmark_compare. We have to pass in a modified version of
4024 * subtree_root; by incrementing the block id, and then checking whether
4025 * last_block is before or equal to that, we can tell whether or not having
4026 * visited last_block implies that all of subtree_root's children have been
4030 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4031 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4033 zbookmark_phys_t mod_zb
= *subtree_root
;
4035 ASSERT(last_block
->zb_level
== 0);
4037 /* The objset_phys_t isn't before anything. */
4042 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4043 * data block size in sectors, because that variable is only used if
4044 * the bookmark refers to a block in the meta-dnode. Since we don't
4045 * know without examining it what object it refers to, and there's no
4046 * harm in passing in this value in other cases, we always pass it in.
4048 * We pass in 0 for the indirect block size shift because zb2 must be
4049 * level 0. The indirect block size is only used to calculate the span
4050 * of the bookmark, but since the bookmark must be level 0, the span is
4051 * always 1, so the math works out.
4053 * If you make changes to how the zbookmark_compare code works, be sure
4054 * to make sure that this code still works afterwards.
4056 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4057 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,