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, 2018 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>
45 #include <sys/cityhash.h>
48 * ==========================================================================
49 * I/O type descriptions
50 * ==========================================================================
52 const char *zio_type_name
[ZIO_TYPES
] = {
53 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
57 boolean_t zio_dva_throttle_enabled
= B_TRUE
;
60 * ==========================================================================
62 * ==========================================================================
64 kmem_cache_t
*zio_cache
;
65 kmem_cache_t
*zio_link_cache
;
66 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
67 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
70 extern vmem_t
*zio_alloc_arena
;
73 #define ZIO_PIPELINE_CONTINUE 0x100
74 #define ZIO_PIPELINE_STOP 0x101
76 #define BP_SPANB(indblkshift, level) \
77 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
78 #define COMPARE_META_LEVEL 0x80000000ul
80 * The following actions directly effect the spa's sync-to-convergence logic.
81 * The values below define the sync pass when we start performing the action.
82 * Care should be taken when changing these values as they directly impact
83 * spa_sync() performance. Tuning these values may introduce subtle performance
84 * pathologies and should only be done in the context of performance analysis.
85 * These tunables will eventually be removed and replaced with #defines once
86 * enough analysis has been done to determine optimal values.
88 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
89 * regular blocks are not deferred.
91 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
92 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
93 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
96 * An allocating zio is one that either currently has the DVA allocate
97 * stage set or will have it later in its lifetime.
99 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
101 boolean_t zio_requeue_io_start_cut_in_line
= B_TRUE
;
104 int zio_buf_debug_limit
= 16384;
106 int zio_buf_debug_limit
= 0;
109 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
115 vmem_t
*data_alloc_arena
= NULL
;
118 data_alloc_arena
= zio_alloc_arena
;
120 zio_cache
= kmem_cache_create("zio_cache",
121 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
122 zio_link_cache
= kmem_cache_create("zio_link_cache",
123 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
126 * For small buffers, we want a cache for each multiple of
127 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
128 * for each quarter-power of 2.
130 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
131 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
134 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
141 * If we are using watchpoints, put each buffer on its own page,
142 * to eliminate the performance overhead of trapping to the
143 * kernel when modifying a non-watched buffer that shares the
144 * page with a watched buffer.
146 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
149 if (size
<= 4 * SPA_MINBLOCKSIZE
) {
150 align
= SPA_MINBLOCKSIZE
;
151 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
152 align
= MIN(p2
>> 2, PAGESIZE
);
157 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
158 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
159 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
162 * Since zio_data bufs do not appear in crash dumps, we
163 * pass KMC_NOTOUCH so that no allocator metadata is
164 * stored with the buffers.
166 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
167 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
168 align
, NULL
, NULL
, NULL
, NULL
, data_alloc_arena
,
169 cflags
| KMC_NOTOUCH
);
174 ASSERT(zio_buf_cache
[c
] != NULL
);
175 if (zio_buf_cache
[c
- 1] == NULL
)
176 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
178 ASSERT(zio_data_buf_cache
[c
] != NULL
);
179 if (zio_data_buf_cache
[c
- 1] == NULL
)
180 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
190 kmem_cache_t
*last_cache
= NULL
;
191 kmem_cache_t
*last_data_cache
= NULL
;
193 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
194 if (zio_buf_cache
[c
] != last_cache
) {
195 last_cache
= zio_buf_cache
[c
];
196 kmem_cache_destroy(zio_buf_cache
[c
]);
198 zio_buf_cache
[c
] = NULL
;
200 if (zio_data_buf_cache
[c
] != last_data_cache
) {
201 last_data_cache
= zio_data_buf_cache
[c
];
202 kmem_cache_destroy(zio_data_buf_cache
[c
]);
204 zio_data_buf_cache
[c
] = NULL
;
207 kmem_cache_destroy(zio_link_cache
);
208 kmem_cache_destroy(zio_cache
);
214 * ==========================================================================
215 * Allocate and free I/O buffers
216 * ==========================================================================
220 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
221 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
222 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
223 * excess / transient data in-core during a crashdump.
226 zio_buf_alloc(size_t size
)
228 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
230 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
232 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
236 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
237 * crashdump if the kernel panics. This exists so that we will limit the amount
238 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
239 * of kernel heap dumped to disk when the kernel panics)
242 zio_data_buf_alloc(size_t size
)
244 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
246 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
248 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
252 zio_buf_free(void *buf
, size_t size
)
254 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
256 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
258 kmem_cache_free(zio_buf_cache
[c
], buf
);
262 zio_data_buf_free(void *buf
, size_t size
)
264 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
266 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
268 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
272 * ==========================================================================
273 * Push and pop I/O transform buffers
274 * ==========================================================================
277 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
278 zio_transform_func_t
*transform
)
280 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
283 * Ensure that anyone expecting this zio to contain a linear ABD isn't
284 * going to get a nasty surprise when they try to access the data.
286 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
288 zt
->zt_orig_abd
= zio
->io_abd
;
289 zt
->zt_orig_size
= zio
->io_size
;
290 zt
->zt_bufsize
= bufsize
;
291 zt
->zt_transform
= transform
;
293 zt
->zt_next
= zio
->io_transform_stack
;
294 zio
->io_transform_stack
= zt
;
301 zio_pop_transforms(zio_t
*zio
)
305 while ((zt
= zio
->io_transform_stack
) != NULL
) {
306 if (zt
->zt_transform
!= NULL
)
307 zt
->zt_transform(zio
,
308 zt
->zt_orig_abd
, zt
->zt_orig_size
);
310 if (zt
->zt_bufsize
!= 0)
311 abd_free(zio
->io_abd
);
313 zio
->io_abd
= zt
->zt_orig_abd
;
314 zio
->io_size
= zt
->zt_orig_size
;
315 zio
->io_transform_stack
= zt
->zt_next
;
317 kmem_free(zt
, sizeof (zio_transform_t
));
322 * ==========================================================================
323 * I/O transform callbacks for subblocks and decompression
324 * ==========================================================================
327 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
329 ASSERT(zio
->io_size
> size
);
331 if (zio
->io_type
== ZIO_TYPE_READ
)
332 abd_copy(data
, zio
->io_abd
, size
);
336 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
338 if (zio
->io_error
== 0) {
339 void *tmp
= abd_borrow_buf(data
, size
);
340 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
341 zio
->io_abd
, tmp
, zio
->io_size
, size
);
342 abd_return_buf_copy(data
, tmp
, size
);
345 zio
->io_error
= SET_ERROR(EIO
);
350 * ==========================================================================
351 * I/O parent/child relationships and pipeline interlocks
352 * ==========================================================================
355 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
357 list_t
*pl
= &cio
->io_parent_list
;
359 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
363 ASSERT((*zl
)->zl_child
== cio
);
364 return ((*zl
)->zl_parent
);
368 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
370 list_t
*cl
= &pio
->io_child_list
;
372 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
376 ASSERT((*zl
)->zl_parent
== pio
);
377 return ((*zl
)->zl_child
);
381 zio_unique_parent(zio_t
*cio
)
383 zio_link_t
*zl
= NULL
;
384 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
386 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
391 zio_add_child(zio_t
*pio
, zio_t
*cio
)
393 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
396 * Logical I/Os can have logical, gang, or vdev children.
397 * Gang I/Os can have gang or vdev children.
398 * Vdev I/Os can only have vdev children.
399 * The following ASSERT captures all of these constraints.
401 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
406 mutex_enter(&cio
->io_lock
);
407 mutex_enter(&pio
->io_lock
);
409 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
411 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
412 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
414 list_insert_head(&pio
->io_child_list
, zl
);
415 list_insert_head(&cio
->io_parent_list
, zl
);
417 pio
->io_child_count
++;
418 cio
->io_parent_count
++;
420 mutex_exit(&pio
->io_lock
);
421 mutex_exit(&cio
->io_lock
);
425 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
427 ASSERT(zl
->zl_parent
== pio
);
428 ASSERT(zl
->zl_child
== cio
);
430 mutex_enter(&cio
->io_lock
);
431 mutex_enter(&pio
->io_lock
);
433 list_remove(&pio
->io_child_list
, zl
);
434 list_remove(&cio
->io_parent_list
, zl
);
436 pio
->io_child_count
--;
437 cio
->io_parent_count
--;
439 mutex_exit(&pio
->io_lock
);
440 mutex_exit(&cio
->io_lock
);
442 kmem_cache_free(zio_link_cache
, zl
);
446 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
448 boolean_t waiting
= B_FALSE
;
450 mutex_enter(&zio
->io_lock
);
451 ASSERT(zio
->io_stall
== NULL
);
452 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
453 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
456 uint64_t *countp
= &zio
->io_children
[c
][wait
];
459 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
460 zio
->io_stall
= countp
;
465 mutex_exit(&zio
->io_lock
);
470 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
472 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
473 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
475 mutex_enter(&pio
->io_lock
);
476 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
477 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
478 pio
->io_reexecute
|= zio
->io_reexecute
;
479 ASSERT3U(*countp
, >, 0);
483 if (*countp
== 0 && pio
->io_stall
== countp
) {
484 zio_taskq_type_t type
=
485 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
487 pio
->io_stall
= NULL
;
488 mutex_exit(&pio
->io_lock
);
490 * Dispatch the parent zio in its own taskq so that
491 * the child can continue to make progress. This also
492 * prevents overflowing the stack when we have deeply nested
493 * parent-child relationships.
495 zio_taskq_dispatch(pio
, type
, B_FALSE
);
497 mutex_exit(&pio
->io_lock
);
502 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
504 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
505 zio
->io_error
= zio
->io_child_error
[c
];
509 zio_bookmark_compare(const void *x1
, const void *x2
)
511 const zio_t
*z1
= x1
;
512 const zio_t
*z2
= x2
;
514 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
516 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
519 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
521 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
524 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
526 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
529 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
531 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
543 * ==========================================================================
544 * Create the various types of I/O (read, write, free, etc)
545 * ==========================================================================
548 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
549 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
550 void *private, zio_type_t type
, zio_priority_t priority
,
551 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
552 const zbookmark_phys_t
*zb
, enum zio_stage stage
, enum zio_stage pipeline
)
556 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
557 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
558 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
560 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
561 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
562 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
564 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW
) != 0);
566 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
567 bzero(zio
, sizeof (zio_t
));
569 mutex_init(&zio
->io_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
570 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
572 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
573 offsetof(zio_link_t
, zl_parent_node
));
574 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
575 offsetof(zio_link_t
, zl_child_node
));
576 metaslab_trace_init(&zio
->io_alloc_list
);
579 zio
->io_child_type
= ZIO_CHILD_VDEV
;
580 else if (flags
& ZIO_FLAG_GANG_CHILD
)
581 zio
->io_child_type
= ZIO_CHILD_GANG
;
582 else if (flags
& ZIO_FLAG_DDT_CHILD
)
583 zio
->io_child_type
= ZIO_CHILD_DDT
;
585 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
588 zio
->io_bp
= (blkptr_t
*)bp
;
589 zio
->io_bp_copy
= *bp
;
590 zio
->io_bp_orig
= *bp
;
591 if (type
!= ZIO_TYPE_WRITE
||
592 zio
->io_child_type
== ZIO_CHILD_DDT
)
593 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
594 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
595 zio
->io_logical
= zio
;
596 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
597 pipeline
|= ZIO_GANG_STAGES
;
603 zio
->io_private
= private;
605 zio
->io_priority
= priority
;
607 zio
->io_offset
= offset
;
608 zio
->io_orig_abd
= zio
->io_abd
= data
;
609 zio
->io_orig_size
= zio
->io_size
= psize
;
610 zio
->io_lsize
= lsize
;
611 zio
->io_orig_flags
= zio
->io_flags
= flags
;
612 zio
->io_orig_stage
= zio
->io_stage
= stage
;
613 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
614 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
616 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
617 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
620 zio
->io_bookmark
= *zb
;
623 if (zio
->io_logical
== NULL
)
624 zio
->io_logical
= pio
->io_logical
;
625 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
626 zio
->io_gang_leader
= pio
->io_gang_leader
;
627 zio_add_child(pio
, zio
);
634 zio_destroy(zio_t
*zio
)
636 metaslab_trace_fini(&zio
->io_alloc_list
);
637 list_destroy(&zio
->io_parent_list
);
638 list_destroy(&zio
->io_child_list
);
639 mutex_destroy(&zio
->io_lock
);
640 cv_destroy(&zio
->io_cv
);
641 kmem_cache_free(zio_cache
, zio
);
645 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
646 void *private, enum zio_flag flags
)
650 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
651 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
652 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
658 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
660 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
664 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
666 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
667 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
668 bp
, (longlong_t
)BP_GET_TYPE(bp
));
670 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
671 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
672 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
673 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
675 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
676 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
677 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
678 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
680 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
681 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
682 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
684 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
685 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
686 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
689 if (BP_IS_EMBEDDED(bp
)) {
690 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
691 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
692 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
697 * Do not verify individual DVAs if the config is not trusted. This
698 * will be done once the zio is executed in vdev_mirror_map_alloc.
700 if (!spa
->spa_trust_config
)
704 * Pool-specific checks.
706 * Note: it would be nice to verify that the blk_birth and
707 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
708 * allows the birth time of log blocks (and dmu_sync()-ed blocks
709 * that are in the log) to be arbitrarily large.
711 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
712 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
713 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
714 zfs_panic_recover("blkptr at %p DVA %u has invalid "
716 bp
, i
, (longlong_t
)vdevid
);
719 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
721 zfs_panic_recover("blkptr at %p DVA %u has invalid "
723 bp
, i
, (longlong_t
)vdevid
);
726 if (vd
->vdev_ops
== &vdev_hole_ops
) {
727 zfs_panic_recover("blkptr at %p DVA %u has hole "
729 bp
, i
, (longlong_t
)vdevid
);
732 if (vd
->vdev_ops
== &vdev_missing_ops
) {
734 * "missing" vdevs are valid during import, but we
735 * don't have their detailed info (e.g. asize), so
736 * we can't perform any more checks on them.
740 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
741 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
743 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
744 if (offset
+ asize
> vd
->vdev_asize
) {
745 zfs_panic_recover("blkptr at %p DVA %u has invalid "
747 bp
, i
, (longlong_t
)offset
);
753 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
755 uint64_t vdevid
= DVA_GET_VDEV(dva
);
757 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
760 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
764 if (vd
->vdev_ops
== &vdev_hole_ops
)
767 if (vd
->vdev_ops
== &vdev_missing_ops
) {
771 uint64_t offset
= DVA_GET_OFFSET(dva
);
772 uint64_t asize
= DVA_GET_ASIZE(dva
);
775 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
776 if (offset
+ asize
> vd
->vdev_asize
)
783 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
784 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
785 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
789 zfs_blkptr_verify(spa
, bp
);
791 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
792 data
, size
, size
, done
, private,
793 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
794 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
795 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
801 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
802 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
803 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
804 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
805 void *private, zio_priority_t priority
, enum zio_flag flags
,
806 const zbookmark_phys_t
*zb
)
810 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
811 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
812 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
813 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
814 DMU_OT_IS_VALID(zp
->zp_type
) &&
817 zp
->zp_copies
<= spa_max_replication(spa
));
819 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
820 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
821 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
822 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
824 zio
->io_ready
= ready
;
825 zio
->io_children_ready
= children_ready
;
826 zio
->io_physdone
= physdone
;
830 * Data can be NULL if we are going to call zio_write_override() to
831 * provide the already-allocated BP. But we may need the data to
832 * verify a dedup hit (if requested). In this case, don't try to
833 * dedup (just take the already-allocated BP verbatim).
835 if (data
== NULL
&& zio
->io_prop
.zp_dedup_verify
) {
836 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
843 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
844 uint64_t size
, zio_done_func_t
*done
, void *private,
845 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
849 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
850 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
851 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
857 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
859 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
860 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
861 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
862 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
865 * We must reset the io_prop to match the values that existed
866 * when the bp was first written by dmu_sync() keeping in mind
867 * that nopwrite and dedup are mutually exclusive.
869 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
870 zio
->io_prop
.zp_nopwrite
= nopwrite
;
871 zio
->io_prop
.zp_copies
= copies
;
872 zio
->io_bp_override
= bp
;
876 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
879 zfs_blkptr_verify(spa
, bp
);
882 * The check for EMBEDDED is a performance optimization. We
883 * process the free here (by ignoring it) rather than
884 * putting it on the list and then processing it in zio_free_sync().
886 if (BP_IS_EMBEDDED(bp
))
888 metaslab_check_free(spa
, bp
);
891 * Frees that are for the currently-syncing txg, are not going to be
892 * deferred, and which will not need to do a read (i.e. not GANG or
893 * DEDUP), can be processed immediately. Otherwise, put them on the
894 * in-memory list for later processing.
896 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
897 txg
!= spa
->spa_syncing_txg
||
898 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
899 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
901 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
906 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
910 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
912 ASSERT(!BP_IS_HOLE(bp
));
913 ASSERT(spa_syncing_txg(spa
) == txg
);
914 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
916 if (BP_IS_EMBEDDED(bp
))
917 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
919 metaslab_check_free(spa
, bp
);
923 * GANG and DEDUP blocks can induce a read (for the gang block header,
924 * or the DDT), so issue them asynchronously so that this thread is
927 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
928 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
930 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
931 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
932 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
938 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
939 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
943 zfs_blkptr_verify(spa
, bp
);
945 if (BP_IS_EMBEDDED(bp
))
946 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
949 * A claim is an allocation of a specific block. Claims are needed
950 * to support immediate writes in the intent log. The issue is that
951 * immediate writes contain committed data, but in a txg that was
952 * *not* committed. Upon opening the pool after an unclean shutdown,
953 * the intent log claims all blocks that contain immediate write data
954 * so that the SPA knows they're in use.
956 * All claims *must* be resolved in the first txg -- before the SPA
957 * starts allocating blocks -- so that nothing is allocated twice.
958 * If txg == 0 we just verify that the block is claimable.
960 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
961 spa_min_claim_txg(spa
));
962 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
963 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
965 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
966 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
967 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
968 ASSERT0(zio
->io_queued_timestamp
);
974 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
975 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
980 if (vd
->vdev_children
== 0) {
981 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
982 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
983 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
987 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
989 for (c
= 0; c
< vd
->vdev_children
; c
++)
990 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
991 done
, private, flags
));
998 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
999 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1000 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1004 ASSERT(vd
->vdev_children
== 0);
1005 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1006 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1007 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1009 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1010 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1011 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1013 zio
->io_prop
.zp_checksum
= checksum
;
1019 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1020 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1021 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1025 ASSERT(vd
->vdev_children
== 0);
1026 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1027 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1028 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1030 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1031 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1032 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1034 zio
->io_prop
.zp_checksum
= checksum
;
1036 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1038 * zec checksums are necessarily destructive -- they modify
1039 * the end of the write buffer to hold the verifier/checksum.
1040 * Therefore, we must make a local copy in case the data is
1041 * being written to multiple places in parallel.
1043 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1044 abd_copy(wbuf
, data
, size
);
1046 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1053 * Create a child I/O to do some work for us.
1056 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1057 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1058 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1060 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1064 * vdev child I/Os do not propagate their error to the parent.
1065 * Therefore, for correct operation the caller *must* check for
1066 * and handle the error in the child i/o's done callback.
1067 * The only exceptions are i/os that we don't care about
1068 * (OPTIONAL or REPAIR).
1070 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1073 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1075 * If we have the bp, then the child should perform the
1076 * checksum and the parent need not. This pushes error
1077 * detection as close to the leaves as possible and
1078 * eliminates redundant checksums in the interior nodes.
1080 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1081 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1084 if (vd
->vdev_ops
->vdev_op_leaf
) {
1085 ASSERT0(vd
->vdev_children
);
1086 offset
+= VDEV_LABEL_START_SIZE
;
1089 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1092 * If we've decided to do a repair, the write is not speculative --
1093 * even if the original read was.
1095 if (flags
& ZIO_FLAG_IO_REPAIR
)
1096 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1099 * If we're creating a child I/O that is not associated with a
1100 * top-level vdev, then the child zio is not an allocating I/O.
1101 * If this is a retried I/O then we ignore it since we will
1102 * have already processed the original allocating I/O.
1104 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1105 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1106 metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
);
1108 ASSERT(mc
->mc_alloc_throttle_enabled
);
1109 ASSERT(type
== ZIO_TYPE_WRITE
);
1110 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1111 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1112 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1113 pio
->io_child_type
== ZIO_CHILD_GANG
);
1115 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1118 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1119 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1120 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1121 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1123 zio
->io_physdone
= pio
->io_physdone
;
1124 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1125 zio
->io_logical
->io_phys_children
++;
1131 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1132 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1133 zio_done_func_t
*done
, void *private)
1137 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1139 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1140 data
, size
, size
, done
, private, type
, priority
,
1141 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1143 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1149 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1151 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1153 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1157 zio_shrink(zio_t
*zio
, uint64_t size
)
1159 ASSERT3P(zio
->io_executor
, ==, NULL
);
1160 ASSERT3P(zio
->io_orig_size
, ==, zio
->io_size
);
1161 ASSERT3U(size
, <=, zio
->io_size
);
1164 * We don't shrink for raidz because of problems with the
1165 * reconstruction when reading back less than the block size.
1166 * Note, BP_IS_RAIDZ() assumes no compression.
1168 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1169 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1170 /* we are not doing a raw write */
1171 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1172 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1177 * ==========================================================================
1178 * Prepare to read and write logical blocks
1179 * ==========================================================================
1183 zio_read_bp_init(zio_t
*zio
)
1185 blkptr_t
*bp
= zio
->io_bp
;
1187 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1189 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1190 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1191 !(zio
->io_flags
& ZIO_FLAG_RAW
)) {
1193 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1194 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1195 psize
, psize
, zio_decompress
);
1198 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1199 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1201 int psize
= BPE_GET_PSIZE(bp
);
1202 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1203 decode_embedded_bp_compressed(bp
, data
);
1204 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1206 ASSERT(!BP_IS_EMBEDDED(bp
));
1207 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1210 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1211 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1213 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1214 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1216 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1217 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1219 return (ZIO_PIPELINE_CONTINUE
);
1223 zio_write_bp_init(zio_t
*zio
)
1225 if (!IO_IS_ALLOCATING(zio
))
1226 return (ZIO_PIPELINE_CONTINUE
);
1228 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1230 if (zio
->io_bp_override
) {
1231 blkptr_t
*bp
= zio
->io_bp
;
1232 zio_prop_t
*zp
= &zio
->io_prop
;
1234 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1235 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1237 *bp
= *zio
->io_bp_override
;
1238 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1240 if (BP_IS_EMBEDDED(bp
))
1241 return (ZIO_PIPELINE_CONTINUE
);
1244 * If we've been overridden and nopwrite is set then
1245 * set the flag accordingly to indicate that a nopwrite
1246 * has already occurred.
1248 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1249 ASSERT(!zp
->zp_dedup
);
1250 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1251 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1252 return (ZIO_PIPELINE_CONTINUE
);
1255 ASSERT(!zp
->zp_nopwrite
);
1257 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1258 return (ZIO_PIPELINE_CONTINUE
);
1260 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1261 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1263 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
) {
1264 BP_SET_DEDUP(bp
, 1);
1265 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1266 return (ZIO_PIPELINE_CONTINUE
);
1270 * We were unable to handle this as an override bp, treat
1271 * it as a regular write I/O.
1273 zio
->io_bp_override
= NULL
;
1274 *bp
= zio
->io_bp_orig
;
1275 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1278 return (ZIO_PIPELINE_CONTINUE
);
1282 zio_write_compress(zio_t
*zio
)
1284 spa_t
*spa
= zio
->io_spa
;
1285 zio_prop_t
*zp
= &zio
->io_prop
;
1286 enum zio_compress compress
= zp
->zp_compress
;
1287 blkptr_t
*bp
= zio
->io_bp
;
1288 uint64_t lsize
= zio
->io_lsize
;
1289 uint64_t psize
= zio
->io_size
;
1292 EQUIV(lsize
!= psize
, (zio
->io_flags
& ZIO_FLAG_RAW
) != 0);
1295 * If our children haven't all reached the ready stage,
1296 * wait for them and then repeat this pipeline stage.
1298 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1299 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1300 return (ZIO_PIPELINE_STOP
);
1303 if (!IO_IS_ALLOCATING(zio
))
1304 return (ZIO_PIPELINE_CONTINUE
);
1306 if (zio
->io_children_ready
!= NULL
) {
1308 * Now that all our children are ready, run the callback
1309 * associated with this zio in case it wants to modify the
1310 * data to be written.
1312 ASSERT3U(zp
->zp_level
, >, 0);
1313 zio
->io_children_ready(zio
);
1316 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1317 ASSERT(zio
->io_bp_override
== NULL
);
1319 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1321 * We're rewriting an existing block, which means we're
1322 * working on behalf of spa_sync(). For spa_sync() to
1323 * converge, it must eventually be the case that we don't
1324 * have to allocate new blocks. But compression changes
1325 * the blocksize, which forces a reallocate, and makes
1326 * convergence take longer. Therefore, after the first
1327 * few passes, stop compressing to ensure convergence.
1329 pass
= spa_sync_pass(spa
);
1331 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1332 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1333 ASSERT(!BP_GET_DEDUP(bp
));
1335 if (pass
>= zfs_sync_pass_dont_compress
)
1336 compress
= ZIO_COMPRESS_OFF
;
1338 /* Make sure someone doesn't change their mind on overwrites */
1339 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1340 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1343 /* If it's a compressed write that is not raw, compress the buffer. */
1344 if (compress
!= ZIO_COMPRESS_OFF
&& psize
== lsize
) {
1345 void *cbuf
= zio_buf_alloc(lsize
);
1346 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1347 if (psize
== 0 || psize
== lsize
) {
1348 compress
= ZIO_COMPRESS_OFF
;
1349 zio_buf_free(cbuf
, lsize
);
1350 } else if (!zp
->zp_dedup
&& psize
<= BPE_PAYLOAD_SIZE
&&
1351 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1352 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1353 encode_embedded_bp_compressed(bp
,
1354 cbuf
, compress
, lsize
, psize
);
1355 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1356 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1357 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1358 zio_buf_free(cbuf
, lsize
);
1359 bp
->blk_birth
= zio
->io_txg
;
1360 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1361 ASSERT(spa_feature_is_active(spa
,
1362 SPA_FEATURE_EMBEDDED_DATA
));
1363 return (ZIO_PIPELINE_CONTINUE
);
1366 * Round up compressed size up to the ashift
1367 * of the smallest-ashift device, and zero the tail.
1368 * This ensures that the compressed size of the BP
1369 * (and thus compressratio property) are correct,
1370 * in that we charge for the padding used to fill out
1373 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1374 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1375 1ULL << spa
->spa_min_ashift
);
1376 if (rounded
>= lsize
) {
1377 compress
= ZIO_COMPRESS_OFF
;
1378 zio_buf_free(cbuf
, lsize
);
1381 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1382 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1383 abd_zero_off(cdata
, psize
, rounded
- psize
);
1385 zio_push_transform(zio
, cdata
,
1386 psize
, lsize
, NULL
);
1391 * We were unable to handle this as an override bp, treat
1392 * it as a regular write I/O.
1394 zio
->io_bp_override
= NULL
;
1395 *bp
= zio
->io_bp_orig
;
1396 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1398 ASSERT3U(psize
, !=, 0);
1402 * The final pass of spa_sync() must be all rewrites, but the first
1403 * few passes offer a trade-off: allocating blocks defers convergence,
1404 * but newly allocated blocks are sequential, so they can be written
1405 * to disk faster. Therefore, we allow the first few passes of
1406 * spa_sync() to allocate new blocks, but force rewrites after that.
1407 * There should only be a handful of blocks after pass 1 in any case.
1409 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1410 BP_GET_PSIZE(bp
) == psize
&&
1411 pass
>= zfs_sync_pass_rewrite
) {
1413 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1414 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1415 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1418 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1422 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1423 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1424 BP_SET_LSIZE(bp
, lsize
);
1425 BP_SET_TYPE(bp
, zp
->zp_type
);
1426 BP_SET_LEVEL(bp
, zp
->zp_level
);
1427 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1429 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1431 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1432 BP_SET_LSIZE(bp
, lsize
);
1433 BP_SET_TYPE(bp
, zp
->zp_type
);
1434 BP_SET_LEVEL(bp
, zp
->zp_level
);
1435 BP_SET_PSIZE(bp
, psize
);
1436 BP_SET_COMPRESS(bp
, compress
);
1437 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1438 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1439 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1441 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1442 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1443 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1445 if (zp
->zp_nopwrite
) {
1446 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1447 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1448 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1451 return (ZIO_PIPELINE_CONTINUE
);
1455 zio_free_bp_init(zio_t
*zio
)
1457 blkptr_t
*bp
= zio
->io_bp
;
1459 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1460 if (BP_GET_DEDUP(bp
))
1461 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1464 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1466 return (ZIO_PIPELINE_CONTINUE
);
1470 * ==========================================================================
1471 * Execute the I/O pipeline
1472 * ==========================================================================
1476 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1478 spa_t
*spa
= zio
->io_spa
;
1479 zio_type_t t
= zio
->io_type
;
1480 int flags
= (cutinline
? TQ_FRONT
: 0);
1483 * If we're a config writer or a probe, the normal issue and
1484 * interrupt threads may all be blocked waiting for the config lock.
1485 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1487 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1491 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1493 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1497 * If this is a high priority I/O, then use the high priority taskq if
1500 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1501 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1502 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1505 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1508 * NB: We are assuming that the zio can only be dispatched
1509 * to a single taskq at a time. It would be a grievous error
1510 * to dispatch the zio to another taskq at the same time.
1512 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
1513 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1514 flags
, &zio
->io_tqent
);
1518 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1520 kthread_t
*executor
= zio
->io_executor
;
1521 spa_t
*spa
= zio
->io_spa
;
1523 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1524 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1526 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1527 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1536 zio_issue_async(zio_t
*zio
)
1538 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1540 return (ZIO_PIPELINE_STOP
);
1544 zio_interrupt(zio_t
*zio
)
1546 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1550 zio_delay_interrupt(zio_t
*zio
)
1553 * The timeout_generic() function isn't defined in userspace, so
1554 * rather than trying to implement the function, the zio delay
1555 * functionality has been disabled for userspace builds.
1560 * If io_target_timestamp is zero, then no delay has been registered
1561 * for this IO, thus jump to the end of this function and "skip" the
1562 * delay; issuing it directly to the zio layer.
1564 if (zio
->io_target_timestamp
!= 0) {
1565 hrtime_t now
= gethrtime();
1567 if (now
>= zio
->io_target_timestamp
) {
1569 * This IO has already taken longer than the target
1570 * delay to complete, so we don't want to delay it
1571 * any longer; we "miss" the delay and issue it
1572 * directly to the zio layer. This is likely due to
1573 * the target latency being set to a value less than
1574 * the underlying hardware can satisfy (e.g. delay
1575 * set to 1ms, but the disks take 10ms to complete an
1579 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1584 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1586 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1587 hrtime_t
, now
, hrtime_t
, diff
);
1589 (void) timeout_generic(CALLOUT_NORMAL
,
1590 (void (*)(void *))zio_interrupt
, zio
, diff
, 1, 0);
1597 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1602 * Execute the I/O pipeline until one of the following occurs:
1604 * (1) the I/O completes
1605 * (2) the pipeline stalls waiting for dependent child I/Os
1606 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1607 * (4) the I/O is delegated by vdev-level caching or aggregation
1608 * (5) the I/O is deferred due to vdev-level queueing
1609 * (6) the I/O is handed off to another thread.
1611 * In all cases, the pipeline stops whenever there's no CPU work; it never
1612 * burns a thread in cv_wait().
1614 * There's no locking on io_stage because there's no legitimate way
1615 * for multiple threads to be attempting to process the same I/O.
1617 static zio_pipe_stage_t
*zio_pipeline
[];
1620 zio_execute(zio_t
*zio
)
1622 zio
->io_executor
= curthread
;
1624 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1626 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1627 enum zio_stage pipeline
= zio
->io_pipeline
;
1628 enum zio_stage stage
= zio
->io_stage
;
1631 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1632 ASSERT(ISP2(stage
));
1633 ASSERT(zio
->io_stall
== NULL
);
1637 } while ((stage
& pipeline
) == 0);
1639 ASSERT(stage
<= ZIO_STAGE_DONE
);
1642 * If we are in interrupt context and this pipeline stage
1643 * will grab a config lock that is held across I/O,
1644 * or may wait for an I/O that needs an interrupt thread
1645 * to complete, issue async to avoid deadlock.
1647 * For VDEV_IO_START, we cut in line so that the io will
1648 * be sent to disk promptly.
1650 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1651 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1652 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1653 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1654 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1658 zio
->io_stage
= stage
;
1659 zio
->io_pipeline_trace
|= zio
->io_stage
;
1660 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1662 if (rv
== ZIO_PIPELINE_STOP
)
1665 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1670 * ==========================================================================
1671 * Initiate I/O, either sync or async
1672 * ==========================================================================
1675 zio_wait(zio_t
*zio
)
1679 ASSERT3P(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1680 ASSERT3P(zio
->io_executor
, ==, NULL
);
1682 zio
->io_waiter
= curthread
;
1683 ASSERT0(zio
->io_queued_timestamp
);
1684 zio
->io_queued_timestamp
= gethrtime();
1688 mutex_enter(&zio
->io_lock
);
1689 while (zio
->io_executor
!= NULL
)
1690 cv_wait(&zio
->io_cv
, &zio
->io_lock
);
1691 mutex_exit(&zio
->io_lock
);
1693 error
= zio
->io_error
;
1700 zio_nowait(zio_t
*zio
)
1702 ASSERT3P(zio
->io_executor
, ==, NULL
);
1704 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1705 zio_unique_parent(zio
) == NULL
) {
1707 * This is a logical async I/O with no parent to wait for it.
1708 * We add it to the spa_async_root_zio "Godfather" I/O which
1709 * will ensure they complete prior to unloading the pool.
1711 spa_t
*spa
= zio
->io_spa
;
1713 zio_add_child(spa
->spa_async_zio_root
[CPU_SEQID
], zio
);
1716 ASSERT0(zio
->io_queued_timestamp
);
1717 zio
->io_queued_timestamp
= gethrtime();
1722 * ==========================================================================
1723 * Reexecute, cancel, or suspend/resume failed I/O
1724 * ==========================================================================
1728 zio_reexecute(zio_t
*pio
)
1730 zio_t
*cio
, *cio_next
;
1732 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1733 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1734 ASSERT(pio
->io_gang_leader
== NULL
);
1735 ASSERT(pio
->io_gang_tree
== NULL
);
1737 pio
->io_flags
= pio
->io_orig_flags
;
1738 pio
->io_stage
= pio
->io_orig_stage
;
1739 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1740 pio
->io_reexecute
= 0;
1741 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1742 pio
->io_pipeline_trace
= 0;
1744 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1745 pio
->io_state
[w
] = 0;
1746 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1747 pio
->io_child_error
[c
] = 0;
1749 if (IO_IS_ALLOCATING(pio
))
1750 BP_ZERO(pio
->io_bp
);
1753 * As we reexecute pio's children, new children could be created.
1754 * New children go to the head of pio's io_child_list, however,
1755 * so we will (correctly) not reexecute them. The key is that
1756 * the remainder of pio's io_child_list, from 'cio_next' onward,
1757 * cannot be affected by any side effects of reexecuting 'cio'.
1759 zio_link_t
*zl
= NULL
;
1760 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1761 cio_next
= zio_walk_children(pio
, &zl
);
1762 mutex_enter(&pio
->io_lock
);
1763 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1764 pio
->io_children
[cio
->io_child_type
][w
]++;
1765 mutex_exit(&pio
->io_lock
);
1770 * Now that all children have been reexecuted, execute the parent.
1771 * We don't reexecute "The Godfather" I/O here as it's the
1772 * responsibility of the caller to wait on it.
1774 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1775 pio
->io_queued_timestamp
= gethrtime();
1781 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1783 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1784 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1785 "failure and the failure mode property for this pool "
1786 "is set to panic.", spa_name(spa
));
1788 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
, NULL
, 0, 0);
1790 mutex_enter(&spa
->spa_suspend_lock
);
1792 if (spa
->spa_suspend_zio_root
== NULL
)
1793 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
1794 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
1795 ZIO_FLAG_GODFATHER
);
1797 spa
->spa_suspended
= B_TRUE
;
1800 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
1801 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
1802 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1803 ASSERT(zio_unique_parent(zio
) == NULL
);
1804 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
1805 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
1808 mutex_exit(&spa
->spa_suspend_lock
);
1812 zio_resume(spa_t
*spa
)
1817 * Reexecute all previously suspended i/o.
1819 mutex_enter(&spa
->spa_suspend_lock
);
1820 spa
->spa_suspended
= B_FALSE
;
1821 cv_broadcast(&spa
->spa_suspend_cv
);
1822 pio
= spa
->spa_suspend_zio_root
;
1823 spa
->spa_suspend_zio_root
= NULL
;
1824 mutex_exit(&spa
->spa_suspend_lock
);
1830 return (zio_wait(pio
));
1834 zio_resume_wait(spa_t
*spa
)
1836 mutex_enter(&spa
->spa_suspend_lock
);
1837 while (spa_suspended(spa
))
1838 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
1839 mutex_exit(&spa
->spa_suspend_lock
);
1843 * ==========================================================================
1846 * A gang block is a collection of small blocks that looks to the DMU
1847 * like one large block. When zio_dva_allocate() cannot find a block
1848 * of the requested size, due to either severe fragmentation or the pool
1849 * being nearly full, it calls zio_write_gang_block() to construct the
1850 * block from smaller fragments.
1852 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1853 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1854 * an indirect block: it's an array of block pointers. It consumes
1855 * only one sector and hence is allocatable regardless of fragmentation.
1856 * The gang header's bps point to its gang members, which hold the data.
1858 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1859 * as the verifier to ensure uniqueness of the SHA256 checksum.
1860 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1861 * not the gang header. This ensures that data block signatures (needed for
1862 * deduplication) are independent of how the block is physically stored.
1864 * Gang blocks can be nested: a gang member may itself be a gang block.
1865 * Thus every gang block is a tree in which root and all interior nodes are
1866 * gang headers, and the leaves are normal blocks that contain user data.
1867 * The root of the gang tree is called the gang leader.
1869 * To perform any operation (read, rewrite, free, claim) on a gang block,
1870 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1871 * in the io_gang_tree field of the original logical i/o by recursively
1872 * reading the gang leader and all gang headers below it. This yields
1873 * an in-core tree containing the contents of every gang header and the
1874 * bps for every constituent of the gang block.
1876 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1877 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1878 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1879 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1880 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1881 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1882 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1883 * of the gang header plus zio_checksum_compute() of the data to update the
1884 * gang header's blk_cksum as described above.
1886 * The two-phase assemble/issue model solves the problem of partial failure --
1887 * what if you'd freed part of a gang block but then couldn't read the
1888 * gang header for another part? Assembling the entire gang tree first
1889 * ensures that all the necessary gang header I/O has succeeded before
1890 * starting the actual work of free, claim, or write. Once the gang tree
1891 * is assembled, free and claim are in-memory operations that cannot fail.
1893 * In the event that a gang write fails, zio_dva_unallocate() walks the
1894 * gang tree to immediately free (i.e. insert back into the space map)
1895 * everything we've allocated. This ensures that we don't get ENOSPC
1896 * errors during repeated suspend/resume cycles due to a flaky device.
1898 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1899 * the gang tree, we won't modify the block, so we can safely defer the free
1900 * (knowing that the block is still intact). If we *can* assemble the gang
1901 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1902 * each constituent bp and we can allocate a new block on the next sync pass.
1904 * In all cases, the gang tree allows complete recovery from partial failure.
1905 * ==========================================================================
1909 zio_gang_issue_func_done(zio_t
*zio
)
1911 abd_put(zio
->io_abd
);
1915 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1921 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
1922 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
1923 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1924 &pio
->io_bookmark
));
1928 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1935 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
1936 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1937 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
1938 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
1941 * As we rewrite each gang header, the pipeline will compute
1942 * a new gang block header checksum for it; but no one will
1943 * compute a new data checksum, so we do that here. The one
1944 * exception is the gang leader: the pipeline already computed
1945 * its data checksum because that stage precedes gang assembly.
1946 * (Presently, nothing actually uses interior data checksums;
1947 * this is just good hygiene.)
1949 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
1950 abd_t
*buf
= abd_get_offset(data
, offset
);
1952 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
1953 buf
, BP_GET_PSIZE(bp
));
1958 * If we are here to damage data for testing purposes,
1959 * leave the GBH alone so that we can detect the damage.
1961 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
1962 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
1964 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1965 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
1966 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
1967 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
1975 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1978 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1979 ZIO_GANG_CHILD_FLAGS(pio
)));
1984 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
1987 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
1988 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
1991 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2000 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2002 static zio_gang_node_t
*
2003 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2005 zio_gang_node_t
*gn
;
2007 ASSERT(*gnpp
== NULL
);
2009 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2010 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2017 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2019 zio_gang_node_t
*gn
= *gnpp
;
2021 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2022 ASSERT(gn
->gn_child
[g
] == NULL
);
2024 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2025 kmem_free(gn
, sizeof (*gn
));
2030 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2032 zio_gang_node_t
*gn
= *gnpp
;
2037 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2038 zio_gang_tree_free(&gn
->gn_child
[g
]);
2040 zio_gang_node_free(gnpp
);
2044 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2046 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2047 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2049 ASSERT(gio
->io_gang_leader
== gio
);
2050 ASSERT(BP_IS_GANG(bp
));
2052 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2053 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2054 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2058 zio_gang_tree_assemble_done(zio_t
*zio
)
2060 zio_t
*gio
= zio
->io_gang_leader
;
2061 zio_gang_node_t
*gn
= zio
->io_private
;
2062 blkptr_t
*bp
= zio
->io_bp
;
2064 ASSERT(gio
== zio_unique_parent(zio
));
2065 ASSERT(zio
->io_child_count
== 0);
2070 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2071 if (BP_SHOULD_BYTESWAP(bp
))
2072 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2074 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2075 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2076 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2078 abd_put(zio
->io_abd
);
2080 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2081 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2082 if (!BP_IS_GANG(gbp
))
2084 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2089 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2092 zio_t
*gio
= pio
->io_gang_leader
;
2095 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2096 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2097 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2100 * If you're a gang header, your data is in gn->gn_gbh.
2101 * If you're a gang member, your data is in 'data' and gn == NULL.
2103 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2106 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2108 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2109 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2110 if (BP_IS_HOLE(gbp
))
2112 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2114 offset
+= BP_GET_PSIZE(gbp
);
2118 if (gn
== gio
->io_gang_tree
)
2119 ASSERT3U(gio
->io_size
, ==, offset
);
2126 zio_gang_assemble(zio_t
*zio
)
2128 blkptr_t
*bp
= zio
->io_bp
;
2130 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2131 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2133 zio
->io_gang_leader
= zio
;
2135 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2137 return (ZIO_PIPELINE_CONTINUE
);
2141 zio_gang_issue(zio_t
*zio
)
2143 blkptr_t
*bp
= zio
->io_bp
;
2145 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2146 return (ZIO_PIPELINE_STOP
);
2149 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2150 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2152 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2153 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2156 zio_gang_tree_free(&zio
->io_gang_tree
);
2158 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2160 return (ZIO_PIPELINE_CONTINUE
);
2164 zio_write_gang_member_ready(zio_t
*zio
)
2166 zio_t
*pio
= zio_unique_parent(zio
);
2167 zio_t
*gio
= zio
->io_gang_leader
;
2168 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2169 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2172 if (BP_IS_HOLE(zio
->io_bp
))
2175 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2177 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2178 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2179 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2180 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2181 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2183 mutex_enter(&pio
->io_lock
);
2184 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2185 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2186 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2187 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2188 DVA_SET_ASIZE(&pdva
[d
], asize
);
2190 mutex_exit(&pio
->io_lock
);
2194 zio_write_gang_done(zio_t
*zio
)
2197 * The io_abd field will be NULL for a zio with no data. The io_flags
2198 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2199 * check for it here as it is cleared in zio_ready.
2201 if (zio
->io_abd
!= NULL
)
2202 abd_put(zio
->io_abd
);
2206 zio_write_gang_block(zio_t
*pio
)
2208 spa_t
*spa
= pio
->io_spa
;
2209 metaslab_class_t
*mc
= spa_normal_class(spa
);
2210 blkptr_t
*bp
= pio
->io_bp
;
2211 zio_t
*gio
= pio
->io_gang_leader
;
2213 zio_gang_node_t
*gn
, **gnpp
;
2214 zio_gbh_phys_t
*gbh
;
2216 uint64_t txg
= pio
->io_txg
;
2217 uint64_t resid
= pio
->io_size
;
2219 int copies
= gio
->io_prop
.zp_copies
;
2220 int gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2223 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2225 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2226 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2227 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2230 flags
|= METASLAB_ASYNC_ALLOC
;
2231 VERIFY(refcount_held(&mc
->mc_alloc_slots
[pio
->io_allocator
],
2235 * The logical zio has already placed a reservation for
2236 * 'copies' allocation slots but gang blocks may require
2237 * additional copies. These additional copies
2238 * (i.e. gbh_copies - copies) are guaranteed to succeed
2239 * since metaslab_class_throttle_reserve() always allows
2240 * additional reservations for gang blocks.
2242 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2243 pio
->io_allocator
, pio
, flags
));
2246 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2247 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2248 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2250 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2251 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2255 * If we failed to allocate the gang block header then
2256 * we remove any additional allocation reservations that
2257 * we placed here. The original reservation will
2258 * be removed when the logical I/O goes to the ready
2261 metaslab_class_throttle_unreserve(mc
,
2262 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2264 pio
->io_error
= error
;
2265 return (ZIO_PIPELINE_CONTINUE
);
2269 gnpp
= &gio
->io_gang_tree
;
2271 gnpp
= pio
->io_private
;
2272 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2275 gn
= zio_gang_node_alloc(gnpp
);
2277 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2278 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2281 * Create the gang header.
2283 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2284 zio_write_gang_done
, NULL
, pio
->io_priority
,
2285 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2288 * Create and nowait the gang children.
2290 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2291 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2293 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2295 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2296 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2297 zp
.zp_type
= DMU_OT_NONE
;
2299 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2300 zp
.zp_dedup
= B_FALSE
;
2301 zp
.zp_dedup_verify
= B_FALSE
;
2302 zp
.zp_nopwrite
= B_FALSE
;
2304 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2305 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2306 resid
) : NULL
, lsize
, lsize
, &zp
,
2307 zio_write_gang_member_ready
, NULL
, NULL
,
2308 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2309 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2311 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2312 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2316 * Gang children won't throttle but we should
2317 * account for their work, so reserve an allocation
2318 * slot for them here.
2320 VERIFY(metaslab_class_throttle_reserve(mc
,
2321 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2327 * Set pio's pipeline to just wait for zio to finish.
2329 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2333 return (ZIO_PIPELINE_CONTINUE
);
2337 * The zio_nop_write stage in the pipeline determines if allocating a
2338 * new bp is necessary. The nopwrite feature can handle writes in
2339 * either syncing or open context (i.e. zil writes) and as a result is
2340 * mutually exclusive with dedup.
2342 * By leveraging a cryptographically secure checksum, such as SHA256, we
2343 * can compare the checksums of the new data and the old to determine if
2344 * allocating a new block is required. Note that our requirements for
2345 * cryptographic strength are fairly weak: there can't be any accidental
2346 * hash collisions, but we don't need to be secure against intentional
2347 * (malicious) collisions. To trigger a nopwrite, you have to be able
2348 * to write the file to begin with, and triggering an incorrect (hash
2349 * collision) nopwrite is no worse than simply writing to the file.
2350 * That said, there are no known attacks against the checksum algorithms
2351 * used for nopwrite, assuming that the salt and the checksums
2352 * themselves remain secret.
2355 zio_nop_write(zio_t
*zio
)
2357 blkptr_t
*bp
= zio
->io_bp
;
2358 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2359 zio_prop_t
*zp
= &zio
->io_prop
;
2361 ASSERT(BP_GET_LEVEL(bp
) == 0);
2362 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2363 ASSERT(zp
->zp_nopwrite
);
2364 ASSERT(!zp
->zp_dedup
);
2365 ASSERT(zio
->io_bp_override
== NULL
);
2366 ASSERT(IO_IS_ALLOCATING(zio
));
2369 * Check to see if the original bp and the new bp have matching
2370 * characteristics (i.e. same checksum, compression algorithms, etc).
2371 * If they don't then just continue with the pipeline which will
2372 * allocate a new bp.
2374 if (BP_IS_HOLE(bp_orig
) ||
2375 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2376 ZCHECKSUM_FLAG_NOPWRITE
) ||
2377 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2378 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2379 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2380 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2381 return (ZIO_PIPELINE_CONTINUE
);
2384 * If the checksums match then reset the pipeline so that we
2385 * avoid allocating a new bp and issuing any I/O.
2387 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2388 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2389 ZCHECKSUM_FLAG_NOPWRITE
);
2390 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2391 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2392 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2393 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2394 sizeof (uint64_t)) == 0);
2397 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2398 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2401 return (ZIO_PIPELINE_CONTINUE
);
2405 * ==========================================================================
2407 * ==========================================================================
2410 zio_ddt_child_read_done(zio_t
*zio
)
2412 blkptr_t
*bp
= zio
->io_bp
;
2413 ddt_entry_t
*dde
= zio
->io_private
;
2415 zio_t
*pio
= zio_unique_parent(zio
);
2417 mutex_enter(&pio
->io_lock
);
2418 ddp
= ddt_phys_select(dde
, bp
);
2419 if (zio
->io_error
== 0)
2420 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2422 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2423 dde
->dde_repair_abd
= zio
->io_abd
;
2425 abd_free(zio
->io_abd
);
2426 mutex_exit(&pio
->io_lock
);
2430 zio_ddt_read_start(zio_t
*zio
)
2432 blkptr_t
*bp
= zio
->io_bp
;
2434 ASSERT(BP_GET_DEDUP(bp
));
2435 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2436 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2438 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2439 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2440 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2441 ddt_phys_t
*ddp
= dde
->dde_phys
;
2442 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2445 ASSERT(zio
->io_vsd
== NULL
);
2448 if (ddp_self
== NULL
)
2449 return (ZIO_PIPELINE_CONTINUE
);
2451 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2452 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2454 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2456 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2457 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2458 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2459 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2460 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2462 return (ZIO_PIPELINE_CONTINUE
);
2465 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2466 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2467 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2469 return (ZIO_PIPELINE_CONTINUE
);
2473 zio_ddt_read_done(zio_t
*zio
)
2475 blkptr_t
*bp
= zio
->io_bp
;
2477 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
2478 return (ZIO_PIPELINE_STOP
);
2481 ASSERT(BP_GET_DEDUP(bp
));
2482 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2483 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2485 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2486 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2487 ddt_entry_t
*dde
= zio
->io_vsd
;
2489 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2490 return (ZIO_PIPELINE_CONTINUE
);
2493 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2494 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2495 return (ZIO_PIPELINE_STOP
);
2497 if (dde
->dde_repair_abd
!= NULL
) {
2498 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2500 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2502 ddt_repair_done(ddt
, dde
);
2506 ASSERT(zio
->io_vsd
== NULL
);
2508 return (ZIO_PIPELINE_CONTINUE
);
2512 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2514 spa_t
*spa
= zio
->io_spa
;
2515 boolean_t do_raw
= (zio
->io_flags
& ZIO_FLAG_RAW
);
2517 /* We should never get a raw, override zio */
2518 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2521 * Note: we compare the original data, not the transformed data,
2522 * because when zio->io_bp is an override bp, we will not have
2523 * pushed the I/O transforms. That's an important optimization
2524 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2526 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2527 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2530 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2531 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
,
2532 zio
->io_orig_size
) != 0);
2536 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2537 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2539 if (ddp
->ddp_phys_birth
!= 0) {
2540 arc_buf_t
*abuf
= NULL
;
2541 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2542 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2543 blkptr_t blk
= *zio
->io_bp
;
2546 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2551 * Intuitively, it would make more sense to compare
2552 * io_abd than io_orig_abd in the raw case since you
2553 * don't want to look at any transformations that have
2554 * happened to the data. However, for raw I/Os the
2555 * data will actually be the same in io_abd and
2556 * io_orig_abd, so all we have to do is issue this as
2560 zio_flags
|= ZIO_FLAG_RAW
;
2561 ASSERT3U(zio
->io_size
, ==, zio
->io_orig_size
);
2562 ASSERT0(abd_cmp(zio
->io_abd
, zio
->io_orig_abd
,
2564 ASSERT3P(zio
->io_transform_stack
, ==, NULL
);
2567 error
= arc_read(NULL
, spa
, &blk
,
2568 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2569 zio_flags
, &aflags
, &zio
->io_bookmark
);
2572 if (arc_buf_size(abuf
) != zio
->io_orig_size
||
2573 abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2574 zio
->io_orig_size
) != 0)
2575 error
= SET_ERROR(EEXIST
);
2576 arc_buf_destroy(abuf
, &abuf
);
2580 return (error
!= 0);
2588 zio_ddt_child_write_ready(zio_t
*zio
)
2590 int p
= zio
->io_prop
.zp_copies
;
2591 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2592 ddt_entry_t
*dde
= zio
->io_private
;
2593 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2601 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2603 ddt_phys_fill(ddp
, zio
->io_bp
);
2605 zio_link_t
*zl
= NULL
;
2606 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2607 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2613 zio_ddt_child_write_done(zio_t
*zio
)
2615 int p
= zio
->io_prop
.zp_copies
;
2616 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2617 ddt_entry_t
*dde
= zio
->io_private
;
2618 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2622 ASSERT(ddp
->ddp_refcnt
== 0);
2623 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2624 dde
->dde_lead_zio
[p
] = NULL
;
2626 if (zio
->io_error
== 0) {
2627 zio_link_t
*zl
= NULL
;
2628 while (zio_walk_parents(zio
, &zl
) != NULL
)
2629 ddt_phys_addref(ddp
);
2631 ddt_phys_clear(ddp
);
2638 zio_ddt_ditto_write_done(zio_t
*zio
)
2640 int p
= DDT_PHYS_DITTO
;
2641 zio_prop_t
*zp
= &zio
->io_prop
;
2642 blkptr_t
*bp
= zio
->io_bp
;
2643 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2644 ddt_entry_t
*dde
= zio
->io_private
;
2645 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2646 ddt_key_t
*ddk
= &dde
->dde_key
;
2650 ASSERT(ddp
->ddp_refcnt
== 0);
2651 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2652 dde
->dde_lead_zio
[p
] = NULL
;
2654 if (zio
->io_error
== 0) {
2655 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2656 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2657 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2658 if (ddp
->ddp_phys_birth
!= 0)
2659 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2660 ddt_phys_fill(ddp
, bp
);
2667 zio_ddt_write(zio_t
*zio
)
2669 spa_t
*spa
= zio
->io_spa
;
2670 blkptr_t
*bp
= zio
->io_bp
;
2671 uint64_t txg
= zio
->io_txg
;
2672 zio_prop_t
*zp
= &zio
->io_prop
;
2673 int p
= zp
->zp_copies
;
2677 ddt_t
*ddt
= ddt_select(spa
, bp
);
2681 ASSERT(BP_GET_DEDUP(bp
));
2682 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2683 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2684 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2687 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2688 ddp
= &dde
->dde_phys
[p
];
2690 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2692 * If we're using a weak checksum, upgrade to a strong checksum
2693 * and try again. If we're already using a strong checksum,
2694 * we can't resolve it, so just convert to an ordinary write.
2695 * (And automatically e-mail a paper to Nature?)
2697 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2698 ZCHECKSUM_FLAG_DEDUP
)) {
2699 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2700 zio_pop_transforms(zio
);
2701 zio
->io_stage
= ZIO_STAGE_OPEN
;
2704 zp
->zp_dedup
= B_FALSE
;
2705 BP_SET_DEDUP(bp
, B_FALSE
);
2707 ASSERT(!BP_GET_DEDUP(bp
));
2708 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2710 return (ZIO_PIPELINE_CONTINUE
);
2713 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2714 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2716 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2717 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2718 zio_prop_t czp
= *zp
;
2720 czp
.zp_copies
= ditto_copies
;
2723 * If we arrived here with an override bp, we won't have run
2724 * the transform stack, so we won't have the data we need to
2725 * generate a child i/o. So, toss the override bp and restart.
2726 * This is safe, because using the override bp is just an
2727 * optimization; and it's rare, so the cost doesn't matter.
2729 if (zio
->io_bp_override
) {
2730 zio_pop_transforms(zio
);
2731 zio
->io_stage
= ZIO_STAGE_OPEN
;
2732 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2733 zio
->io_bp_override
= NULL
;
2736 return (ZIO_PIPELINE_CONTINUE
);
2739 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2740 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2741 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2742 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2744 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2745 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2748 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2749 if (ddp
->ddp_phys_birth
!= 0)
2750 ddt_bp_fill(ddp
, bp
, txg
);
2751 if (dde
->dde_lead_zio
[p
] != NULL
)
2752 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2754 ddt_phys_addref(ddp
);
2755 } else if (zio
->io_bp_override
) {
2756 ASSERT(bp
->blk_birth
== txg
);
2757 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2758 ddt_phys_fill(ddp
, bp
);
2759 ddt_phys_addref(ddp
);
2761 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2762 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
2763 zio_ddt_child_write_ready
, NULL
, NULL
,
2764 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
2765 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2767 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2768 dde
->dde_lead_zio
[p
] = cio
;
2778 return (ZIO_PIPELINE_CONTINUE
);
2781 ddt_entry_t
*freedde
; /* for debugging */
2784 zio_ddt_free(zio_t
*zio
)
2786 spa_t
*spa
= zio
->io_spa
;
2787 blkptr_t
*bp
= zio
->io_bp
;
2788 ddt_t
*ddt
= ddt_select(spa
, bp
);
2792 ASSERT(BP_GET_DEDUP(bp
));
2793 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2796 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2797 ddp
= ddt_phys_select(dde
, bp
);
2798 ddt_phys_decref(ddp
);
2801 return (ZIO_PIPELINE_CONTINUE
);
2805 * ==========================================================================
2806 * Allocate and free blocks
2807 * ==========================================================================
2811 zio_io_to_allocate(spa_t
*spa
, int allocator
)
2815 ASSERT(MUTEX_HELD(&spa
->spa_alloc_locks
[allocator
]));
2817 zio
= avl_first(&spa
->spa_alloc_trees
[allocator
]);
2821 ASSERT(IO_IS_ALLOCATING(zio
));
2824 * Try to place a reservation for this zio. If we're unable to
2825 * reserve then we throttle.
2827 ASSERT3U(zio
->io_allocator
, ==, allocator
);
2828 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
2829 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
, 0)) {
2833 avl_remove(&spa
->spa_alloc_trees
[allocator
], zio
);
2834 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
2840 zio_dva_throttle(zio_t
*zio
)
2842 spa_t
*spa
= zio
->io_spa
;
2845 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
2846 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
2847 zio
->io_child_type
== ZIO_CHILD_GANG
||
2848 zio
->io_flags
& ZIO_FLAG_NODATA
) {
2849 return (ZIO_PIPELINE_CONTINUE
);
2852 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2854 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2855 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2857 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
2859 * We want to try to use as many allocators as possible to help improve
2860 * performance, but we also want logically adjacent IOs to be physically
2861 * adjacent to improve sequential read performance. We chunk each object
2862 * into 2^20 block regions, and then hash based on the objset, object,
2863 * level, and region to accomplish both of these goals.
2865 zio
->io_allocator
= cityhash4(bm
->zb_objset
, bm
->zb_object
,
2866 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
2867 mutex_enter(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
2869 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
2870 avl_add(&spa
->spa_alloc_trees
[zio
->io_allocator
], zio
);
2872 nio
= zio_io_to_allocate(zio
->io_spa
, zio
->io_allocator
);
2873 mutex_exit(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
2876 return (ZIO_PIPELINE_CONTINUE
);
2879 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
2881 * We are passing control to a new zio so make sure that
2882 * it is processed by a different thread. We do this to
2883 * avoid stack overflows that can occur when parents are
2884 * throttled and children are making progress. We allow
2885 * it to go to the head of the taskq since it's already
2888 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2890 return (ZIO_PIPELINE_STOP
);
2894 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
2898 mutex_enter(&spa
->spa_alloc_locks
[allocator
]);
2899 zio
= zio_io_to_allocate(spa
, allocator
);
2900 mutex_exit(&spa
->spa_alloc_locks
[allocator
]);
2904 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
2905 ASSERT0(zio
->io_error
);
2906 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
2910 zio_dva_allocate(zio_t
*zio
)
2912 spa_t
*spa
= zio
->io_spa
;
2913 metaslab_class_t
*mc
= spa_normal_class(spa
);
2914 blkptr_t
*bp
= zio
->io_bp
;
2918 if (zio
->io_gang_leader
== NULL
) {
2919 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2920 zio
->io_gang_leader
= zio
;
2923 ASSERT(BP_IS_HOLE(bp
));
2924 ASSERT0(BP_GET_NDVAS(bp
));
2925 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
2926 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
2927 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
2929 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
2930 flags
|= METASLAB_DONT_THROTTLE
;
2932 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
) {
2933 flags
|= METASLAB_GANG_CHILD
;
2935 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
) {
2936 flags
|= METASLAB_ASYNC_ALLOC
;
2939 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
2940 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
2941 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
2944 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
2945 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
2947 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
2948 return (zio_write_gang_block(zio
));
2949 zio
->io_error
= error
;
2952 return (ZIO_PIPELINE_CONTINUE
);
2956 zio_dva_free(zio_t
*zio
)
2958 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
2960 return (ZIO_PIPELINE_CONTINUE
);
2964 zio_dva_claim(zio_t
*zio
)
2968 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
2970 zio
->io_error
= error
;
2972 return (ZIO_PIPELINE_CONTINUE
);
2976 * Undo an allocation. This is used by zio_done() when an I/O fails
2977 * and we want to give back the block we just allocated.
2978 * This handles both normal blocks and gang blocks.
2981 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
2983 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
2984 ASSERT(zio
->io_bp_override
== NULL
);
2986 if (!BP_IS_HOLE(bp
))
2987 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
2990 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2991 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
2992 &gn
->gn_gbh
->zg_blkptr
[g
]);
2998 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3001 zio_alloc_zil(spa_t
*spa
, uint64_t objset
, uint64_t txg
, blkptr_t
*new_bp
,
3002 blkptr_t
*old_bp
, uint64_t size
, boolean_t
*slog
)
3005 zio_alloc_list_t io_alloc_list
;
3007 ASSERT(txg
> spa_syncing_txg(spa
));
3009 metaslab_trace_init(&io_alloc_list
);
3011 * When allocating a zil block, we don't have information about
3012 * the final destination of the block except the objset it's part
3013 * of, so we just hash the objset ID to pick the allocator to get
3016 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3017 txg
, old_bp
, METASLAB_HINTBP_AVOID
, &io_alloc_list
, NULL
,
3018 cityhash4(0, 0, 0, objset
) % spa
->spa_alloc_count
);
3022 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3023 new_bp
, 1, txg
, old_bp
, METASLAB_HINTBP_AVOID
,
3024 &io_alloc_list
, NULL
, cityhash4(0, 0, 0, objset
) %
3025 spa
->spa_alloc_count
);
3029 metaslab_trace_fini(&io_alloc_list
);
3032 BP_SET_LSIZE(new_bp
, size
);
3033 BP_SET_PSIZE(new_bp
, size
);
3034 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3035 BP_SET_CHECKSUM(new_bp
,
3036 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3037 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3038 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3039 BP_SET_LEVEL(new_bp
, 0);
3040 BP_SET_DEDUP(new_bp
, 0);
3041 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3043 zfs_dbgmsg("%s: zil block allocation failure: "
3044 "size %llu, error %d", spa_name(spa
), size
, error
);
3051 * ==========================================================================
3052 * Read and write to physical devices
3053 * ==========================================================================
3058 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3059 * stops after this stage and will resume upon I/O completion.
3060 * However, there are instances where the vdev layer may need to
3061 * continue the pipeline when an I/O was not issued. Since the I/O
3062 * that was sent to the vdev layer might be different than the one
3063 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3064 * force the underlying vdev layers to call either zio_execute() or
3065 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3068 zio_vdev_io_start(zio_t
*zio
)
3070 vdev_t
*vd
= zio
->io_vd
;
3072 spa_t
*spa
= zio
->io_spa
;
3074 ASSERT(zio
->io_error
== 0);
3075 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3078 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3079 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3082 * The mirror_ops handle multiple DVAs in a single BP.
3084 vdev_mirror_ops
.vdev_op_io_start(zio
);
3085 return (ZIO_PIPELINE_STOP
);
3088 ASSERT3P(zio
->io_logical
, !=, zio
);
3089 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3090 ASSERT(spa
->spa_trust_config
);
3092 if (zio
->io_vd
->vdev_removing
) {
3094 * Note: the code can handle other kinds of writes,
3095 * but we don't expect them.
3097 ASSERT(zio
->io_flags
&
3098 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3099 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3104 * We keep track of time-sensitive I/Os so that the scan thread
3105 * can quickly react to certain workloads. In particular, we care
3106 * about non-scrubbing, top-level reads and writes with the following
3108 * - synchronous writes of user data to non-slog devices
3109 * - any reads of user data
3110 * When these conditions are met, adjust the timestamp of spa_last_io
3111 * which allows the scan thread to adjust its workload accordingly.
3113 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
3114 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
3115 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
3116 zio
->io_txg
!= spa_syncing_txg(spa
)) {
3117 uint64_t old
= spa
->spa_last_io
;
3118 uint64_t new = ddi_get_lbolt64();
3120 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
3123 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3125 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3126 P2PHASE(zio
->io_size
, align
) != 0) {
3127 /* Transform logical writes to be a full physical block size. */
3128 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3129 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3130 ASSERT(vd
== vd
->vdev_top
);
3131 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3132 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3133 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3135 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3139 * If this is not a physical io, make sure that it is properly aligned
3140 * before proceeding.
3142 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3143 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3144 ASSERT0(P2PHASE(zio
->io_size
, align
));
3147 * For physical writes, we allow 512b aligned writes and assume
3148 * the device will perform a read-modify-write as necessary.
3150 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3151 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3154 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3157 * If this is a repair I/O, and there's no self-healing involved --
3158 * that is, we're just resilvering what we expect to resilver --
3159 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3160 * This prevents spurious resilvering.
3162 * There are a few ways that we can end up creating these spurious
3165 * 1. A resilver i/o will be issued if any DVA in the BP has a
3166 * dirty DTL. The mirror code will issue resilver writes to
3167 * each DVA, including the one(s) that are not on vdevs with dirty
3170 * 2. With nested replication, which happens when we have a
3171 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3172 * For example, given mirror(replacing(A+B), C), it's likely that
3173 * only A is out of date (it's the new device). In this case, we'll
3174 * read from C, then use the data to resilver A+B -- but we don't
3175 * actually want to resilver B, just A. The top-level mirror has no
3176 * way to know this, so instead we just discard unnecessary repairs
3177 * as we work our way down the vdev tree.
3179 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3180 * The same logic applies to any form of nested replication: ditto
3181 * + mirror, RAID-Z + replacing, etc.
3183 * However, indirect vdevs point off to other vdevs which may have
3184 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3185 * will be properly bypassed instead.
3187 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3188 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3189 zio
->io_txg
!= 0 && /* not a delegated i/o */
3190 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3191 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3192 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3193 zio_vdev_io_bypass(zio
);
3194 return (ZIO_PIPELINE_CONTINUE
);
3197 if (vd
->vdev_ops
->vdev_op_leaf
&&
3198 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3200 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3201 return (ZIO_PIPELINE_CONTINUE
);
3203 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3204 return (ZIO_PIPELINE_STOP
);
3206 if (!vdev_accessible(vd
, zio
)) {
3207 zio
->io_error
= SET_ERROR(ENXIO
);
3209 return (ZIO_PIPELINE_STOP
);
3213 vd
->vdev_ops
->vdev_op_io_start(zio
);
3214 return (ZIO_PIPELINE_STOP
);
3218 zio_vdev_io_done(zio_t
*zio
)
3220 vdev_t
*vd
= zio
->io_vd
;
3221 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3222 boolean_t unexpected_error
= B_FALSE
;
3224 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3225 return (ZIO_PIPELINE_STOP
);
3228 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3230 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3232 vdev_queue_io_done(zio
);
3234 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3235 vdev_cache_write(zio
);
3237 if (zio_injection_enabled
&& zio
->io_error
== 0)
3238 zio
->io_error
= zio_handle_device_injection(vd
,
3241 if (zio_injection_enabled
&& zio
->io_error
== 0)
3242 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3244 if (zio
->io_error
) {
3245 if (!vdev_accessible(vd
, zio
)) {
3246 zio
->io_error
= SET_ERROR(ENXIO
);
3248 unexpected_error
= B_TRUE
;
3253 ops
->vdev_op_io_done(zio
);
3255 if (unexpected_error
)
3256 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3258 return (ZIO_PIPELINE_CONTINUE
);
3262 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3263 * disk, and use that to finish the checksum ereport later.
3266 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3267 const void *good_buf
)
3269 /* no processing needed */
3270 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3275 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3277 void *buf
= zio_buf_alloc(zio
->io_size
);
3279 abd_copy_to_buf(buf
, zio
->io_abd
, zio
->io_size
);
3281 zcr
->zcr_cbinfo
= zio
->io_size
;
3282 zcr
->zcr_cbdata
= buf
;
3283 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3284 zcr
->zcr_free
= zio_buf_free
;
3288 zio_vdev_io_assess(zio_t
*zio
)
3290 vdev_t
*vd
= zio
->io_vd
;
3292 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3293 return (ZIO_PIPELINE_STOP
);
3296 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3297 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3299 if (zio
->io_vsd
!= NULL
) {
3300 zio
->io_vsd_ops
->vsd_free(zio
);
3304 if (zio_injection_enabled
&& zio
->io_error
== 0)
3305 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3308 * If the I/O failed, determine whether we should attempt to retry it.
3310 * On retry, we cut in line in the issue queue, since we don't want
3311 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3313 if (zio
->io_error
&& vd
== NULL
&&
3314 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3315 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3316 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3318 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3319 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3320 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3321 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3322 zio_requeue_io_start_cut_in_line
);
3323 return (ZIO_PIPELINE_STOP
);
3327 * If we got an error on a leaf device, convert it to ENXIO
3328 * if the device is not accessible at all.
3330 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3331 !vdev_accessible(vd
, zio
))
3332 zio
->io_error
= SET_ERROR(ENXIO
);
3335 * If we can't write to an interior vdev (mirror or RAID-Z),
3336 * set vdev_cant_write so that we stop trying to allocate from it.
3338 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3339 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3340 vd
->vdev_cant_write
= B_TRUE
;
3344 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3345 * attempts will ever succeed. In this case we set a persistent bit so
3346 * that we don't bother with it in the future.
3348 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3349 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3350 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3351 vd
->vdev_nowritecache
= B_TRUE
;
3354 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3356 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3357 zio
->io_physdone
!= NULL
) {
3358 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3359 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3360 zio
->io_physdone(zio
->io_logical
);
3363 return (ZIO_PIPELINE_CONTINUE
);
3367 zio_vdev_io_reissue(zio_t
*zio
)
3369 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3370 ASSERT(zio
->io_error
== 0);
3372 zio
->io_stage
>>= 1;
3376 zio_vdev_io_redone(zio_t
*zio
)
3378 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3380 zio
->io_stage
>>= 1;
3384 zio_vdev_io_bypass(zio_t
*zio
)
3386 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3387 ASSERT(zio
->io_error
== 0);
3389 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3390 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3394 * ==========================================================================
3395 * Generate and verify checksums
3396 * ==========================================================================
3399 zio_checksum_generate(zio_t
*zio
)
3401 blkptr_t
*bp
= zio
->io_bp
;
3402 enum zio_checksum checksum
;
3406 * This is zio_write_phys().
3407 * We're either generating a label checksum, or none at all.
3409 checksum
= zio
->io_prop
.zp_checksum
;
3411 if (checksum
== ZIO_CHECKSUM_OFF
)
3412 return (ZIO_PIPELINE_CONTINUE
);
3414 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3416 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3417 ASSERT(!IO_IS_ALLOCATING(zio
));
3418 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3420 checksum
= BP_GET_CHECKSUM(bp
);
3424 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3426 return (ZIO_PIPELINE_CONTINUE
);
3430 zio_checksum_verify(zio_t
*zio
)
3432 zio_bad_cksum_t info
;
3433 blkptr_t
*bp
= zio
->io_bp
;
3436 ASSERT(zio
->io_vd
!= NULL
);
3440 * This is zio_read_phys().
3441 * We're either verifying a label checksum, or nothing at all.
3443 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3444 return (ZIO_PIPELINE_CONTINUE
);
3446 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3449 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3450 zio
->io_error
= error
;
3451 if (error
== ECKSUM
&&
3452 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3453 zfs_ereport_start_checksum(zio
->io_spa
,
3454 zio
->io_vd
, zio
, zio
->io_offset
,
3455 zio
->io_size
, NULL
, &info
);
3459 return (ZIO_PIPELINE_CONTINUE
);
3463 * Called by RAID-Z to ensure we don't compute the checksum twice.
3466 zio_checksum_verified(zio_t
*zio
)
3468 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3472 * ==========================================================================
3473 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3474 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3475 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3476 * indicate errors that are specific to one I/O, and most likely permanent.
3477 * Any other error is presumed to be worse because we weren't expecting it.
3478 * ==========================================================================
3481 zio_worst_error(int e1
, int e2
)
3483 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3486 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3487 if (e1
== zio_error_rank
[r1
])
3490 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3491 if (e2
== zio_error_rank
[r2
])
3494 return (r1
> r2
? e1
: e2
);
3498 * ==========================================================================
3500 * ==========================================================================
3503 zio_ready(zio_t
*zio
)
3505 blkptr_t
*bp
= zio
->io_bp
;
3506 zio_t
*pio
, *pio_next
;
3507 zio_link_t
*zl
= NULL
;
3509 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
3511 return (ZIO_PIPELINE_STOP
);
3514 if (zio
->io_ready
) {
3515 ASSERT(IO_IS_ALLOCATING(zio
));
3516 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3517 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3518 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3523 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3524 zio
->io_bp_copy
= *bp
;
3526 if (zio
->io_error
!= 0) {
3527 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3529 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3530 ASSERT(IO_IS_ALLOCATING(zio
));
3531 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3533 * We were unable to allocate anything, unreserve and
3534 * issue the next I/O to allocate.
3536 metaslab_class_throttle_unreserve(
3537 spa_normal_class(zio
->io_spa
),
3538 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3539 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
3543 mutex_enter(&zio
->io_lock
);
3544 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3545 pio
= zio_walk_parents(zio
, &zl
);
3546 mutex_exit(&zio
->io_lock
);
3549 * As we notify zio's parents, new parents could be added.
3550 * New parents go to the head of zio's io_parent_list, however,
3551 * so we will (correctly) not notify them. The remainder of zio's
3552 * io_parent_list, from 'pio_next' onward, cannot change because
3553 * all parents must wait for us to be done before they can be done.
3555 for (; pio
!= NULL
; pio
= pio_next
) {
3556 pio_next
= zio_walk_parents(zio
, &zl
);
3557 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3560 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3561 if (BP_IS_GANG(bp
)) {
3562 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3564 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3565 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3569 if (zio_injection_enabled
&&
3570 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3571 zio_handle_ignored_writes(zio
);
3573 return (ZIO_PIPELINE_CONTINUE
);
3577 * Update the allocation throttle accounting.
3580 zio_dva_throttle_done(zio_t
*zio
)
3582 zio_t
*lio
= zio
->io_logical
;
3583 zio_t
*pio
= zio_unique_parent(zio
);
3584 vdev_t
*vd
= zio
->io_vd
;
3585 int flags
= METASLAB_ASYNC_ALLOC
;
3587 ASSERT3P(zio
->io_bp
, !=, NULL
);
3588 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3589 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3590 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3592 ASSERT3P(vd
, ==, vd
->vdev_top
);
3593 ASSERT(!(zio
->io_flags
& (ZIO_FLAG_IO_REPAIR
| ZIO_FLAG_IO_RETRY
)));
3594 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3595 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3596 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3599 * Parents of gang children can have two flavors -- ones that
3600 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3601 * and ones that allocated the constituent blocks. The allocation
3602 * throttle needs to know the allocating parent zio so we must find
3605 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3607 * If our parent is a rewrite gang child then our grandparent
3608 * would have been the one that performed the allocation.
3610 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3611 pio
= zio_unique_parent(pio
);
3612 flags
|= METASLAB_GANG_CHILD
;
3615 ASSERT(IO_IS_ALLOCATING(pio
));
3616 ASSERT3P(zio
, !=, zio
->io_logical
);
3617 ASSERT(zio
->io_logical
!= NULL
);
3618 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3619 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3621 mutex_enter(&pio
->io_lock
);
3622 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
3623 pio
->io_allocator
, B_TRUE
);
3624 mutex_exit(&pio
->io_lock
);
3626 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
3627 1, pio
->io_allocator
, pio
);
3630 * Call into the pipeline to see if there is more work that
3631 * needs to be done. If there is work to be done it will be
3632 * dispatched to another taskq thread.
3634 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
3638 zio_done(zio_t
*zio
)
3640 spa_t
*spa
= zio
->io_spa
;
3641 zio_t
*lio
= zio
->io_logical
;
3642 blkptr_t
*bp
= zio
->io_bp
;
3643 vdev_t
*vd
= zio
->io_vd
;
3644 uint64_t psize
= zio
->io_size
;
3645 zio_t
*pio
, *pio_next
;
3646 metaslab_class_t
*mc
= spa_normal_class(spa
);
3647 zio_link_t
*zl
= NULL
;
3650 * If our children haven't all completed,
3651 * wait for them and then repeat this pipeline stage.
3653 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
3654 return (ZIO_PIPELINE_STOP
);
3658 * If the allocation throttle is enabled, then update the accounting.
3659 * We only track child I/Os that are part of an allocating async
3660 * write. We must do this since the allocation is performed
3661 * by the logical I/O but the actual write is done by child I/Os.
3663 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
3664 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
3665 ASSERT(mc
->mc_alloc_throttle_enabled
);
3666 zio_dva_throttle_done(zio
);
3670 * If the allocation throttle is enabled, verify that
3671 * we have decremented the refcounts for every I/O that was throttled.
3673 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3674 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3675 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3677 metaslab_group_alloc_verify(spa
, zio
->io_bp
, zio
,
3679 VERIFY(refcount_not_held(&mc
->mc_alloc_slots
[zio
->io_allocator
],
3683 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
3684 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
3685 ASSERT(zio
->io_children
[c
][w
] == 0);
3687 if (bp
!= NULL
&& !BP_IS_EMBEDDED(bp
)) {
3688 ASSERT(bp
->blk_pad
[0] == 0);
3689 ASSERT(bp
->blk_pad
[1] == 0);
3690 ASSERT(bcmp(bp
, &zio
->io_bp_copy
, sizeof (blkptr_t
)) == 0 ||
3691 (bp
== zio_unique_parent(zio
)->io_bp
));
3692 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(bp
) &&
3693 zio
->io_bp_override
== NULL
&&
3694 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
3695 ASSERT(!BP_SHOULD_BYTESWAP(bp
));
3696 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(bp
));
3697 ASSERT(BP_COUNT_GANG(bp
) == 0 ||
3698 (BP_COUNT_GANG(bp
) == BP_GET_NDVAS(bp
)));
3700 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
3701 VERIFY(BP_EQUAL(bp
, &zio
->io_bp_orig
));
3705 * If there were child vdev/gang/ddt errors, they apply to us now.
3707 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
3708 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
3709 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
3712 * If the I/O on the transformed data was successful, generate any
3713 * checksum reports now while we still have the transformed data.
3715 if (zio
->io_error
== 0) {
3716 while (zio
->io_cksum_report
!= NULL
) {
3717 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3718 uint64_t align
= zcr
->zcr_align
;
3719 uint64_t asize
= P2ROUNDUP(psize
, align
);
3721 abd_t
*adata
= zio
->io_abd
;
3723 if (asize
!= psize
) {
3724 adata
= abd_alloc_linear(asize
, B_TRUE
);
3725 abd_copy(adata
, zio
->io_abd
, psize
);
3726 abd_zero_off(adata
, psize
, asize
- psize
);
3730 abuf
= abd_borrow_buf_copy(adata
, asize
);
3732 zio
->io_cksum_report
= zcr
->zcr_next
;
3733 zcr
->zcr_next
= NULL
;
3734 zcr
->zcr_finish(zcr
, abuf
);
3735 zfs_ereport_free_checksum(zcr
);
3738 abd_return_buf(adata
, abuf
, asize
);
3745 zio_pop_transforms(zio
); /* note: may set zio->io_error */
3747 vdev_stat_update(zio
, psize
);
3749 if (zio
->io_error
) {
3751 * If this I/O is attached to a particular vdev,
3752 * generate an error message describing the I/O failure
3753 * at the block level. We ignore these errors if the
3754 * device is currently unavailable.
3756 if (zio
->io_error
!= ECKSUM
&& vd
!= NULL
&& !vdev_is_dead(vd
))
3757 zfs_ereport_post(FM_EREPORT_ZFS_IO
, spa
, vd
, zio
, 0, 0);
3759 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
3760 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
3763 * For logical I/O requests, tell the SPA to log the
3764 * error and generate a logical data ereport.
3766 spa_log_error(spa
, zio
);
3767 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, spa
, NULL
, zio
,
3772 if (zio
->io_error
&& zio
== lio
) {
3774 * Determine whether zio should be reexecuted. This will
3775 * propagate all the way to the root via zio_notify_parent().
3777 ASSERT(vd
== NULL
&& bp
!= NULL
);
3778 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3780 if (IO_IS_ALLOCATING(zio
) &&
3781 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
3782 if (zio
->io_error
!= ENOSPC
)
3783 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
3785 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3788 if ((zio
->io_type
== ZIO_TYPE_READ
||
3789 zio
->io_type
== ZIO_TYPE_FREE
) &&
3790 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
3791 zio
->io_error
== ENXIO
&&
3792 spa_load_state(spa
) == SPA_LOAD_NONE
&&
3793 spa_get_failmode(spa
) != ZIO_FAILURE_MODE_CONTINUE
)
3794 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3796 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
3797 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
3800 * Here is a possibly good place to attempt to do
3801 * either combinatorial reconstruction or error correction
3802 * based on checksums. It also might be a good place
3803 * to send out preliminary ereports before we suspend
3809 * If there were logical child errors, they apply to us now.
3810 * We defer this until now to avoid conflating logical child
3811 * errors with errors that happened to the zio itself when
3812 * updating vdev stats and reporting FMA events above.
3814 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
3816 if ((zio
->io_error
|| zio
->io_reexecute
) &&
3817 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
3818 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
3819 zio_dva_unallocate(zio
, zio
->io_gang_tree
, bp
);
3821 zio_gang_tree_free(&zio
->io_gang_tree
);
3824 * Godfather I/Os should never suspend.
3826 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3827 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
3828 zio
->io_reexecute
= 0;
3830 if (zio
->io_reexecute
) {
3832 * This is a logical I/O that wants to reexecute.
3834 * Reexecute is top-down. When an i/o fails, if it's not
3835 * the root, it simply notifies its parent and sticks around.
3836 * The parent, seeing that it still has children in zio_done(),
3837 * does the same. This percolates all the way up to the root.
3838 * The root i/o will reexecute or suspend the entire tree.
3840 * This approach ensures that zio_reexecute() honors
3841 * all the original i/o dependency relationships, e.g.
3842 * parents not executing until children are ready.
3844 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3846 zio
->io_gang_leader
= NULL
;
3848 mutex_enter(&zio
->io_lock
);
3849 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3850 mutex_exit(&zio
->io_lock
);
3853 * "The Godfather" I/O monitors its children but is
3854 * not a true parent to them. It will track them through
3855 * the pipeline but severs its ties whenever they get into
3856 * trouble (e.g. suspended). This allows "The Godfather"
3857 * I/O to return status without blocking.
3860 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
3862 zio_link_t
*remove_zl
= zl
;
3863 pio_next
= zio_walk_parents(zio
, &zl
);
3865 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
3866 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
3867 zio_remove_child(pio
, zio
, remove_zl
);
3868 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3872 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
3874 * We're not a root i/o, so there's nothing to do
3875 * but notify our parent. Don't propagate errors
3876 * upward since we haven't permanently failed yet.
3878 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
3879 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
3880 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3881 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
3883 * We'd fail again if we reexecuted now, so suspend
3884 * until conditions improve (e.g. device comes online).
3886 zio_suspend(spa
, zio
);
3889 * Reexecution is potentially a huge amount of work.
3890 * Hand it off to the otherwise-unused claim taskq.
3892 ASSERT(zio
->io_tqent
.tqent_next
== NULL
);
3893 spa_taskq_dispatch_ent(spa
, ZIO_TYPE_CLAIM
,
3894 ZIO_TASKQ_ISSUE
, (task_func_t
*)zio_reexecute
, zio
,
3897 return (ZIO_PIPELINE_STOP
);
3900 ASSERT(zio
->io_child_count
== 0);
3901 ASSERT(zio
->io_reexecute
== 0);
3902 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
3905 * Report any checksum errors, since the I/O is complete.
3907 while (zio
->io_cksum_report
!= NULL
) {
3908 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
3909 zio
->io_cksum_report
= zcr
->zcr_next
;
3910 zcr
->zcr_next
= NULL
;
3911 zcr
->zcr_finish(zcr
, NULL
);
3912 zfs_ereport_free_checksum(zcr
);
3916 * It is the responsibility of the done callback to ensure that this
3917 * particular zio is no longer discoverable for adoption, and as
3918 * such, cannot acquire any new parents.
3923 mutex_enter(&zio
->io_lock
);
3924 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
3925 mutex_exit(&zio
->io_lock
);
3928 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
3929 zio_link_t
*remove_zl
= zl
;
3930 pio_next
= zio_walk_parents(zio
, &zl
);
3931 zio_remove_child(pio
, zio
, remove_zl
);
3932 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
3935 if (zio
->io_waiter
!= NULL
) {
3936 mutex_enter(&zio
->io_lock
);
3937 zio
->io_executor
= NULL
;
3938 cv_broadcast(&zio
->io_cv
);
3939 mutex_exit(&zio
->io_lock
);
3944 return (ZIO_PIPELINE_STOP
);
3948 * ==========================================================================
3949 * I/O pipeline definition
3950 * ==========================================================================
3952 static zio_pipe_stage_t
*zio_pipeline
[] = {
3959 zio_checksum_generate
,
3975 zio_checksum_verify
,
3983 * Compare two zbookmark_phys_t's to see which we would reach first in a
3984 * pre-order traversal of the object tree.
3986 * This is simple in every case aside from the meta-dnode object. For all other
3987 * objects, we traverse them in order (object 1 before object 2, and so on).
3988 * However, all of these objects are traversed while traversing object 0, since
3989 * the data it points to is the list of objects. Thus, we need to convert to a
3990 * canonical representation so we can compare meta-dnode bookmarks to
3991 * non-meta-dnode bookmarks.
3993 * We do this by calculating "equivalents" for each field of the zbookmark.
3994 * zbookmarks outside of the meta-dnode use their own object and level, and
3995 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3996 * blocks this bookmark refers to) by multiplying their blkid by their span
3997 * (the number of L0 blocks contained within one block at their level).
3998 * zbookmarks inside the meta-dnode calculate their object equivalent
3999 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4000 * level + 1<<31 (any value larger than a level could ever be) for their level.
4001 * This causes them to always compare before a bookmark in their object
4002 * equivalent, compare appropriately to bookmarks in other objects, and to
4003 * compare appropriately to other bookmarks in the meta-dnode.
4006 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4007 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4010 * These variables represent the "equivalent" values for the zbookmark,
4011 * after converting zbookmarks inside the meta dnode to their
4012 * normal-object equivalents.
4014 uint64_t zb1obj
, zb2obj
;
4015 uint64_t zb1L0
, zb2L0
;
4016 uint64_t zb1level
, zb2level
;
4018 if (zb1
->zb_object
== zb2
->zb_object
&&
4019 zb1
->zb_level
== zb2
->zb_level
&&
4020 zb1
->zb_blkid
== zb2
->zb_blkid
)
4024 * BP_SPANB calculates the span in blocks.
4026 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4027 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4029 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4030 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4032 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4034 zb1obj
= zb1
->zb_object
;
4035 zb1level
= zb1
->zb_level
;
4038 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4039 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4041 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4043 zb2obj
= zb2
->zb_object
;
4044 zb2level
= zb2
->zb_level
;
4047 /* Now that we have a canonical representation, do the comparison. */
4048 if (zb1obj
!= zb2obj
)
4049 return (zb1obj
< zb2obj
? -1 : 1);
4050 else if (zb1L0
!= zb2L0
)
4051 return (zb1L0
< zb2L0
? -1 : 1);
4052 else if (zb1level
!= zb2level
)
4053 return (zb1level
> zb2level
? -1 : 1);
4055 * This can (theoretically) happen if the bookmarks have the same object
4056 * and level, but different blkids, if the block sizes are not the same.
4057 * There is presently no way to change the indirect block sizes
4063 * This function checks the following: given that last_block is the place that
4064 * our traversal stopped last time, does that guarantee that we've visited
4065 * every node under subtree_root? Therefore, we can't just use the raw output
4066 * of zbookmark_compare. We have to pass in a modified version of
4067 * subtree_root; by incrementing the block id, and then checking whether
4068 * last_block is before or equal to that, we can tell whether or not having
4069 * visited last_block implies that all of subtree_root's children have been
4073 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4074 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4076 zbookmark_phys_t mod_zb
= *subtree_root
;
4078 ASSERT(last_block
->zb_level
== 0);
4080 /* The objset_phys_t isn't before anything. */
4085 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4086 * data block size in sectors, because that variable is only used if
4087 * the bookmark refers to a block in the meta-dnode. Since we don't
4088 * know without examining it what object it refers to, and there's no
4089 * harm in passing in this value in other cases, we always pass it in.
4091 * We pass in 0 for the indirect block size shift because zb2 must be
4092 * level 0. The indirect block size is only used to calculate the span
4093 * of the bookmark, but since the bookmark must be level 0, the span is
4094 * always 1, so the math works out.
4096 * If you make changes to how the zbookmark_compare code works, be sure
4097 * to make sure that this code still works afterwards.
4099 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4100 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,