4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 #include <sys/dsl_scan.h>
43 #include <sys/metaslab_impl.h>
45 #include <sys/trace_zio.h>
47 #include <sys/dsl_crypt.h>
50 * ==========================================================================
51 * I/O type descriptions
52 * ==========================================================================
54 const char *zio_type_name
[ZIO_TYPES
] = {
56 * Note: Linux kernel thread name length is limited
57 * so these names will differ from upstream open zfs.
59 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
62 int zio_dva_throttle_enabled
= B_TRUE
;
65 * ==========================================================================
67 * ==========================================================================
69 kmem_cache_t
*zio_cache
;
70 kmem_cache_t
*zio_link_cache
;
71 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
72 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
73 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
74 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
75 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 int zio_delay_max
= ZIO_DELAY_MAX
;
80 #define ZIO_PIPELINE_CONTINUE 0x100
81 #define ZIO_PIPELINE_STOP 0x101
83 #define BP_SPANB(indblkshift, level) \
84 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
85 #define COMPARE_META_LEVEL 0x80000000ul
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
99 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
100 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
103 * An allocating zio is one that either currently has the DVA allocate
104 * stage set or will have it later in its lifetime.
106 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 int zio_requeue_io_start_cut_in_line
= 1;
111 int zio_buf_debug_limit
= 16384;
113 int zio_buf_debug_limit
= 0;
116 static inline void __zio_execute(zio_t
*zio
);
118 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
124 vmem_t
*data_alloc_arena
= NULL
;
126 zio_cache
= kmem_cache_create("zio_cache",
127 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
128 zio_link_cache
= kmem_cache_create("zio_link_cache",
129 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
132 * For small buffers, we want a cache for each multiple of
133 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
134 * for each quarter-power of 2.
136 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
137 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
140 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
142 #if defined(_ILP32) && defined(_KERNEL)
144 * Cache size limited to 1M on 32-bit platforms until ARC
145 * buffers no longer require virtual address space.
147 if (size
> zfs_max_recordsize
)
156 * If we are using watchpoints, put each buffer on its own page,
157 * to eliminate the performance overhead of trapping to the
158 * kernel when modifying a non-watched buffer that shares the
159 * page with a watched buffer.
161 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
164 * Here's the problem - on 4K native devices in userland on
165 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
166 * will fail with EINVAL, causing zdb (and others) to coredump.
167 * Since userland probably doesn't need optimized buffer caches,
168 * we just force 4K alignment on everything.
170 align
= 8 * SPA_MINBLOCKSIZE
;
172 if (size
< PAGESIZE
) {
173 align
= SPA_MINBLOCKSIZE
;
174 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
181 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
182 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
183 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
185 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
186 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
187 align
, NULL
, NULL
, NULL
, NULL
,
188 data_alloc_arena
, cflags
);
193 ASSERT(zio_buf_cache
[c
] != NULL
);
194 if (zio_buf_cache
[c
- 1] == NULL
)
195 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
197 ASSERT(zio_data_buf_cache
[c
] != NULL
);
198 if (zio_data_buf_cache
[c
- 1] == NULL
)
199 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
211 kmem_cache_t
*last_cache
= NULL
;
212 kmem_cache_t
*last_data_cache
= NULL
;
214 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
217 * Cache size limited to 1M on 32-bit platforms until ARC
218 * buffers no longer require virtual address space.
220 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
223 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
224 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
225 (void) printf("zio_fini: [%d] %llu != %llu\n",
226 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
227 (long long unsigned)zio_buf_cache_allocs
[c
],
228 (long long unsigned)zio_buf_cache_frees
[c
]);
230 if (zio_buf_cache
[c
] != last_cache
) {
231 last_cache
= zio_buf_cache
[c
];
232 kmem_cache_destroy(zio_buf_cache
[c
]);
234 zio_buf_cache
[c
] = NULL
;
236 if (zio_data_buf_cache
[c
] != last_data_cache
) {
237 last_data_cache
= zio_data_buf_cache
[c
];
238 kmem_cache_destroy(zio_data_buf_cache
[c
]);
240 zio_data_buf_cache
[c
] = NULL
;
243 kmem_cache_destroy(zio_link_cache
);
244 kmem_cache_destroy(zio_cache
);
252 * ==========================================================================
253 * Allocate and free I/O buffers
254 * ==========================================================================
258 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
259 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
260 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
261 * excess / transient data in-core during a crashdump.
264 zio_buf_alloc(size_t size
)
266 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
268 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
269 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
270 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
273 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size
)
285 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
287 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
289 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
293 zio_buf_free(void *buf
, size_t size
)
295 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
297 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
302 kmem_cache_free(zio_buf_cache
[c
], buf
);
306 zio_data_buf_free(void *buf
, size_t size
)
308 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
310 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
312 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
316 zio_abd_free(void *abd
, size_t size
)
318 abd_free((abd_t
*)abd
);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
328 zio_transform_func_t
*transform
)
330 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
333 * Ensure that anyone expecting this zio to contain a linear ABD isn't
334 * going to get a nasty surprise when they try to access the data.
336 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
338 zt
->zt_orig_abd
= zio
->io_abd
;
339 zt
->zt_orig_size
= zio
->io_size
;
340 zt
->zt_bufsize
= bufsize
;
341 zt
->zt_transform
= transform
;
343 zt
->zt_next
= zio
->io_transform_stack
;
344 zio
->io_transform_stack
= zt
;
351 zio_pop_transforms(zio_t
*zio
)
355 while ((zt
= zio
->io_transform_stack
) != NULL
) {
356 if (zt
->zt_transform
!= NULL
)
357 zt
->zt_transform(zio
,
358 zt
->zt_orig_abd
, zt
->zt_orig_size
);
360 if (zt
->zt_bufsize
!= 0)
361 abd_free(zio
->io_abd
);
363 zio
->io_abd
= zt
->zt_orig_abd
;
364 zio
->io_size
= zt
->zt_orig_size
;
365 zio
->io_transform_stack
= zt
->zt_next
;
367 kmem_free(zt
, sizeof (zio_transform_t
));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks, decompression, and decryption
374 * ==========================================================================
377 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
379 ASSERT(zio
->io_size
> size
);
381 if (zio
->io_type
== ZIO_TYPE_READ
)
382 abd_copy(data
, zio
->io_abd
, size
);
386 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
388 if (zio
->io_error
== 0) {
389 void *tmp
= abd_borrow_buf(data
, size
);
390 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
391 zio
->io_abd
, tmp
, zio
->io_size
, size
);
392 abd_return_buf_copy(data
, tmp
, size
);
395 zio
->io_error
= SET_ERROR(EIO
);
400 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
404 blkptr_t
*bp
= zio
->io_bp
;
405 spa_t
*spa
= zio
->io_spa
;
406 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
407 uint64_t lsize
= BP_GET_LSIZE(bp
);
408 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
409 uint8_t salt
[ZIO_DATA_SALT_LEN
];
410 uint8_t iv
[ZIO_DATA_IV_LEN
];
411 uint8_t mac
[ZIO_DATA_MAC_LEN
];
412 boolean_t no_crypt
= B_FALSE
;
414 ASSERT(BP_USES_CRYPT(bp
));
415 ASSERT3U(size
, !=, 0);
417 if (zio
->io_error
!= 0)
421 * Verify the cksum of MACs stored in an indirect bp. It will always
422 * be possible to verify this since it does not require an encryption
425 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
426 zio_crypt_decode_mac_bp(bp
, mac
);
428 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
430 * We haven't decompressed the data yet, but
431 * zio_crypt_do_indirect_mac_checksum() requires
432 * decompressed data to be able to parse out the MACs
433 * from the indirect block. We decompress it now and
434 * throw away the result after we are finished.
436 tmp
= zio_buf_alloc(lsize
);
437 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
438 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
440 ret
= SET_ERROR(EIO
);
443 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
444 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
445 zio_buf_free(tmp
, lsize
);
447 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
448 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
450 abd_copy(data
, zio
->io_abd
, size
);
452 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
453 ret
= zio_handle_decrypt_injection(spa
,
454 &zio
->io_bookmark
, ot
, ECKSUM
);
463 * If this is an authenticated block, just check the MAC. It would be
464 * nice to separate this out into its own flag, but for the moment
465 * enum zio_flag is out of bits.
467 if (BP_IS_AUTHENTICATED(bp
)) {
468 if (ot
== DMU_OT_OBJSET
) {
469 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
470 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
472 zio_crypt_decode_mac_bp(bp
, mac
);
473 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
474 zio
->io_abd
, size
, mac
);
475 if (zio_injection_enabled
&& ret
== 0) {
476 ret
= zio_handle_decrypt_injection(spa
,
477 &zio
->io_bookmark
, ot
, ECKSUM
);
480 abd_copy(data
, zio
->io_abd
, size
);
488 zio_crypt_decode_params_bp(bp
, salt
, iv
);
490 if (ot
== DMU_OT_INTENT_LOG
) {
491 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
492 zio_crypt_decode_mac_zil(tmp
, mac
);
493 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
495 zio_crypt_decode_mac_bp(bp
, mac
);
498 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
499 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
500 zio
->io_abd
, &no_crypt
);
502 abd_copy(data
, zio
->io_abd
, size
);
510 /* assert that the key was found unless this was speculative */
511 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
514 * If there was a decryption / authentication error return EIO as
515 * the io_error. If this was not a speculative zio, create an ereport.
518 zio
->io_error
= SET_ERROR(EIO
);
519 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
520 spa_log_error(spa
, &zio
->io_bookmark
);
521 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
522 spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
530 * ==========================================================================
531 * I/O parent/child relationships and pipeline interlocks
532 * ==========================================================================
535 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
537 list_t
*pl
= &cio
->io_parent_list
;
539 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
543 ASSERT((*zl
)->zl_child
== cio
);
544 return ((*zl
)->zl_parent
);
548 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
550 list_t
*cl
= &pio
->io_child_list
;
552 ASSERT(MUTEX_HELD(&pio
->io_lock
));
554 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
558 ASSERT((*zl
)->zl_parent
== pio
);
559 return ((*zl
)->zl_child
);
563 zio_unique_parent(zio_t
*cio
)
565 zio_link_t
*zl
= NULL
;
566 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
568 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
573 zio_add_child(zio_t
*pio
, zio_t
*cio
)
575 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
578 * Logical I/Os can have logical, gang, or vdev children.
579 * Gang I/Os can have gang or vdev children.
580 * Vdev I/Os can only have vdev children.
581 * The following ASSERT captures all of these constraints.
583 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
588 mutex_enter(&pio
->io_lock
);
589 mutex_enter(&cio
->io_lock
);
591 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
593 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
594 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
596 list_insert_head(&pio
->io_child_list
, zl
);
597 list_insert_head(&cio
->io_parent_list
, zl
);
599 pio
->io_child_count
++;
600 cio
->io_parent_count
++;
602 mutex_exit(&cio
->io_lock
);
603 mutex_exit(&pio
->io_lock
);
607 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
609 ASSERT(zl
->zl_parent
== pio
);
610 ASSERT(zl
->zl_child
== cio
);
612 mutex_enter(&pio
->io_lock
);
613 mutex_enter(&cio
->io_lock
);
615 list_remove(&pio
->io_child_list
, zl
);
616 list_remove(&cio
->io_parent_list
, zl
);
618 pio
->io_child_count
--;
619 cio
->io_parent_count
--;
621 mutex_exit(&cio
->io_lock
);
622 mutex_exit(&pio
->io_lock
);
623 kmem_cache_free(zio_link_cache
, zl
);
627 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
629 boolean_t waiting
= B_FALSE
;
631 mutex_enter(&zio
->io_lock
);
632 ASSERT(zio
->io_stall
== NULL
);
633 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
634 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
637 uint64_t *countp
= &zio
->io_children
[c
][wait
];
640 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
641 zio
->io_stall
= countp
;
646 mutex_exit(&zio
->io_lock
);
650 __attribute__((always_inline
))
652 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
654 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
655 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
657 mutex_enter(&pio
->io_lock
);
658 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
659 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
660 pio
->io_reexecute
|= zio
->io_reexecute
;
661 ASSERT3U(*countp
, >, 0);
665 if (*countp
== 0 && pio
->io_stall
== countp
) {
666 zio_taskq_type_t type
=
667 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
669 pio
->io_stall
= NULL
;
670 mutex_exit(&pio
->io_lock
);
672 * Dispatch the parent zio in its own taskq so that
673 * the child can continue to make progress. This also
674 * prevents overflowing the stack when we have deeply nested
675 * parent-child relationships.
677 zio_taskq_dispatch(pio
, type
, B_FALSE
);
679 mutex_exit(&pio
->io_lock
);
684 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
686 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
687 zio
->io_error
= zio
->io_child_error
[c
];
691 zio_bookmark_compare(const void *x1
, const void *x2
)
693 const zio_t
*z1
= x1
;
694 const zio_t
*z2
= x2
;
696 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
698 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
701 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
703 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
706 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
708 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
711 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
713 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
725 * ==========================================================================
726 * Create the various types of I/O (read, write, free, etc)
727 * ==========================================================================
730 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
731 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
732 void *private, zio_type_t type
, zio_priority_t priority
,
733 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
734 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
735 enum zio_stage pipeline
)
739 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
740 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
741 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
743 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
744 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
745 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
747 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
749 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
750 bzero(zio
, sizeof (zio_t
));
752 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
753 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
755 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
756 offsetof(zio_link_t
, zl_parent_node
));
757 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
758 offsetof(zio_link_t
, zl_child_node
));
759 metaslab_trace_init(&zio
->io_alloc_list
);
762 zio
->io_child_type
= ZIO_CHILD_VDEV
;
763 else if (flags
& ZIO_FLAG_GANG_CHILD
)
764 zio
->io_child_type
= ZIO_CHILD_GANG
;
765 else if (flags
& ZIO_FLAG_DDT_CHILD
)
766 zio
->io_child_type
= ZIO_CHILD_DDT
;
768 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
771 zio
->io_bp
= (blkptr_t
*)bp
;
772 zio
->io_bp_copy
= *bp
;
773 zio
->io_bp_orig
= *bp
;
774 if (type
!= ZIO_TYPE_WRITE
||
775 zio
->io_child_type
== ZIO_CHILD_DDT
)
776 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
777 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
778 zio
->io_logical
= zio
;
779 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
780 pipeline
|= ZIO_GANG_STAGES
;
786 zio
->io_private
= private;
788 zio
->io_priority
= priority
;
790 zio
->io_offset
= offset
;
791 zio
->io_orig_abd
= zio
->io_abd
= data
;
792 zio
->io_orig_size
= zio
->io_size
= psize
;
793 zio
->io_lsize
= lsize
;
794 zio
->io_orig_flags
= zio
->io_flags
= flags
;
795 zio
->io_orig_stage
= zio
->io_stage
= stage
;
796 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
797 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
799 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
800 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
803 zio
->io_bookmark
= *zb
;
806 if (zio
->io_logical
== NULL
)
807 zio
->io_logical
= pio
->io_logical
;
808 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
809 zio
->io_gang_leader
= pio
->io_gang_leader
;
810 zio_add_child(pio
, zio
);
813 taskq_init_ent(&zio
->io_tqent
);
819 zio_destroy(zio_t
*zio
)
821 metaslab_trace_fini(&zio
->io_alloc_list
);
822 list_destroy(&zio
->io_parent_list
);
823 list_destroy(&zio
->io_child_list
);
824 mutex_destroy(&zio
->io_lock
);
825 cv_destroy(&zio
->io_cv
);
826 kmem_cache_free(zio_cache
, zio
);
830 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
831 void *private, enum zio_flag flags
)
835 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
836 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
837 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
843 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
845 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
849 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
851 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
852 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
853 bp
, (longlong_t
)BP_GET_TYPE(bp
));
855 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
856 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
857 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
858 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
860 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
861 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
862 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
863 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
865 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
866 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
867 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
869 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
870 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
871 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
874 if (BP_IS_EMBEDDED(bp
)) {
875 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
876 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
877 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
882 * Do not verify individual DVAs if the config is not trusted. This
883 * will be done once the zio is executed in vdev_mirror_map_alloc.
885 if (!spa
->spa_trust_config
)
889 * Pool-specific checks.
891 * Note: it would be nice to verify that the blk_birth and
892 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
893 * allows the birth time of log blocks (and dmu_sync()-ed blocks
894 * that are in the log) to be arbitrarily large.
896 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
897 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
899 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
900 zfs_panic_recover("blkptr at %p DVA %u has invalid "
902 bp
, i
, (longlong_t
)vdevid
);
905 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
907 zfs_panic_recover("blkptr at %p DVA %u has invalid "
909 bp
, i
, (longlong_t
)vdevid
);
912 if (vd
->vdev_ops
== &vdev_hole_ops
) {
913 zfs_panic_recover("blkptr at %p DVA %u has hole "
915 bp
, i
, (longlong_t
)vdevid
);
918 if (vd
->vdev_ops
== &vdev_missing_ops
) {
920 * "missing" vdevs are valid during import, but we
921 * don't have their detailed info (e.g. asize), so
922 * we can't perform any more checks on them.
926 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
927 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
929 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
930 if (offset
+ asize
> vd
->vdev_asize
) {
931 zfs_panic_recover("blkptr at %p DVA %u has invalid "
933 bp
, i
, (longlong_t
)offset
);
939 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
941 uint64_t vdevid
= DVA_GET_VDEV(dva
);
943 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
946 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
950 if (vd
->vdev_ops
== &vdev_hole_ops
)
953 if (vd
->vdev_ops
== &vdev_missing_ops
) {
957 uint64_t offset
= DVA_GET_OFFSET(dva
);
958 uint64_t asize
= DVA_GET_ASIZE(dva
);
961 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
962 if (offset
+ asize
> vd
->vdev_asize
)
969 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
970 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
971 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
975 zfs_blkptr_verify(spa
, bp
);
977 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
978 data
, size
, size
, done
, private,
979 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
980 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
981 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
987 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
988 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
989 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
990 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
991 void *private, zio_priority_t priority
, enum zio_flag flags
,
992 const zbookmark_phys_t
*zb
)
996 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
997 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
998 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
999 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1000 DMU_OT_IS_VALID(zp
->zp_type
) &&
1001 zp
->zp_level
< 32 &&
1002 zp
->zp_copies
> 0 &&
1003 zp
->zp_copies
<= spa_max_replication(spa
));
1005 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1006 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1007 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1008 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1010 zio
->io_ready
= ready
;
1011 zio
->io_children_ready
= children_ready
;
1012 zio
->io_physdone
= physdone
;
1016 * Data can be NULL if we are going to call zio_write_override() to
1017 * provide the already-allocated BP. But we may need the data to
1018 * verify a dedup hit (if requested). In this case, don't try to
1019 * dedup (just take the already-allocated BP verbatim). Encrypted
1020 * dedup blocks need data as well so we also disable dedup in this
1024 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1025 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1032 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1033 uint64_t size
, zio_done_func_t
*done
, void *private,
1034 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
1038 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1039 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1040 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1046 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1048 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1049 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1050 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1051 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1054 * We must reset the io_prop to match the values that existed
1055 * when the bp was first written by dmu_sync() keeping in mind
1056 * that nopwrite and dedup are mutually exclusive.
1058 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1059 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1060 zio
->io_prop
.zp_copies
= copies
;
1061 zio
->io_bp_override
= bp
;
1065 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1068 zfs_blkptr_verify(spa
, bp
);
1071 * The check for EMBEDDED is a performance optimization. We
1072 * process the free here (by ignoring it) rather than
1073 * putting it on the list and then processing it in zio_free_sync().
1075 if (BP_IS_EMBEDDED(bp
))
1077 metaslab_check_free(spa
, bp
);
1080 * Frees that are for the currently-syncing txg, are not going to be
1081 * deferred, and which will not need to do a read (i.e. not GANG or
1082 * DEDUP), can be processed immediately. Otherwise, put them on the
1083 * in-memory list for later processing.
1085 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
1086 txg
!= spa
->spa_syncing_txg
||
1087 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
1088 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1090 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1095 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1096 enum zio_flag flags
)
1099 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1101 ASSERT(!BP_IS_HOLE(bp
));
1102 ASSERT(spa_syncing_txg(spa
) == txg
);
1103 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
1105 if (BP_IS_EMBEDDED(bp
))
1106 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1108 metaslab_check_free(spa
, bp
);
1110 dsl_scan_freed(spa
, bp
);
1113 * GANG and DEDUP blocks can induce a read (for the gang block header,
1114 * or the DDT), so issue them asynchronously so that this thread is
1117 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1118 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1120 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1121 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1122 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1128 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1129 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1133 zfs_blkptr_verify(spa
, bp
);
1135 if (BP_IS_EMBEDDED(bp
))
1136 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1139 * A claim is an allocation of a specific block. Claims are needed
1140 * to support immediate writes in the intent log. The issue is that
1141 * immediate writes contain committed data, but in a txg that was
1142 * *not* committed. Upon opening the pool after an unclean shutdown,
1143 * the intent log claims all blocks that contain immediate write data
1144 * so that the SPA knows they're in use.
1146 * All claims *must* be resolved in the first txg -- before the SPA
1147 * starts allocating blocks -- so that nothing is allocated twice.
1148 * If txg == 0 we just verify that the block is claimable.
1150 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
1151 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
1152 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1154 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1155 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1156 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1157 ASSERT0(zio
->io_queued_timestamp
);
1163 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1164 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1169 if (vd
->vdev_children
== 0) {
1170 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1171 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1172 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1176 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1178 for (c
= 0; c
< vd
->vdev_children
; c
++)
1179 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1180 done
, private, flags
));
1187 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1188 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1189 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1193 ASSERT(vd
->vdev_children
== 0);
1194 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1195 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1196 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1198 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1199 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1200 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1202 zio
->io_prop
.zp_checksum
= checksum
;
1208 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1209 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1210 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1214 ASSERT(vd
->vdev_children
== 0);
1215 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1216 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1217 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1219 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1220 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1221 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1223 zio
->io_prop
.zp_checksum
= checksum
;
1225 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1227 * zec checksums are necessarily destructive -- they modify
1228 * the end of the write buffer to hold the verifier/checksum.
1229 * Therefore, we must make a local copy in case the data is
1230 * being written to multiple places in parallel.
1232 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1233 abd_copy(wbuf
, data
, size
);
1235 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1242 * Create a child I/O to do some work for us.
1245 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1246 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1247 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1249 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1253 * vdev child I/Os do not propagate their error to the parent.
1254 * Therefore, for correct operation the caller *must* check for
1255 * and handle the error in the child i/o's done callback.
1256 * The only exceptions are i/os that we don't care about
1257 * (OPTIONAL or REPAIR).
1259 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1262 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1264 * If we have the bp, then the child should perform the
1265 * checksum and the parent need not. This pushes error
1266 * detection as close to the leaves as possible and
1267 * eliminates redundant checksums in the interior nodes.
1269 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1270 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1273 if (vd
->vdev_ops
->vdev_op_leaf
) {
1274 ASSERT0(vd
->vdev_children
);
1275 offset
+= VDEV_LABEL_START_SIZE
;
1278 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1281 * If we've decided to do a repair, the write is not speculative --
1282 * even if the original read was.
1284 if (flags
& ZIO_FLAG_IO_REPAIR
)
1285 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1288 * If we're creating a child I/O that is not associated with a
1289 * top-level vdev, then the child zio is not an allocating I/O.
1290 * If this is a retried I/O then we ignore it since we will
1291 * have already processed the original allocating I/O.
1293 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1294 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1295 ASSERTV(metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
));
1297 ASSERT(mc
->mc_alloc_throttle_enabled
);
1298 ASSERT(type
== ZIO_TYPE_WRITE
);
1299 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1300 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1301 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1302 pio
->io_child_type
== ZIO_CHILD_GANG
);
1304 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1308 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1309 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1310 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1311 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1313 zio
->io_physdone
= pio
->io_physdone
;
1314 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1315 zio
->io_logical
->io_phys_children
++;
1321 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1322 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1323 zio_done_func_t
*done
, void *private)
1327 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1329 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1330 data
, size
, size
, done
, private, type
, priority
,
1331 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1333 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1339 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1341 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1343 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1347 zio_shrink(zio_t
*zio
, uint64_t size
)
1349 ASSERT3P(zio
->io_executor
, ==, NULL
);
1350 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1351 ASSERT3U(size
, <=, zio
->io_size
);
1354 * We don't shrink for raidz because of problems with the
1355 * reconstruction when reading back less than the block size.
1356 * Note, BP_IS_RAIDZ() assumes no compression.
1358 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1359 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1360 /* we are not doing a raw write */
1361 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1362 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1367 * ==========================================================================
1368 * Prepare to read and write logical blocks
1369 * ==========================================================================
1373 zio_read_bp_init(zio_t
*zio
)
1375 blkptr_t
*bp
= zio
->io_bp
;
1377 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1379 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1381 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1382 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1383 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1384 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1385 psize
, psize
, zio_decompress
);
1388 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1389 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1390 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1391 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1392 psize
, psize
, zio_decrypt
);
1395 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1396 int psize
= BPE_GET_PSIZE(bp
);
1397 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1399 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1400 decode_embedded_bp_compressed(bp
, data
);
1401 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1403 ASSERT(!BP_IS_EMBEDDED(bp
));
1404 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1407 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1408 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1410 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1411 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1413 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1414 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1416 return (ZIO_PIPELINE_CONTINUE
);
1420 zio_write_bp_init(zio_t
*zio
)
1422 if (!IO_IS_ALLOCATING(zio
))
1423 return (ZIO_PIPELINE_CONTINUE
);
1425 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1427 if (zio
->io_bp_override
) {
1428 blkptr_t
*bp
= zio
->io_bp
;
1429 zio_prop_t
*zp
= &zio
->io_prop
;
1431 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1432 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1434 *bp
= *zio
->io_bp_override
;
1435 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1437 if (BP_IS_EMBEDDED(bp
))
1438 return (ZIO_PIPELINE_CONTINUE
);
1441 * If we've been overridden and nopwrite is set then
1442 * set the flag accordingly to indicate that a nopwrite
1443 * has already occurred.
1445 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1446 ASSERT(!zp
->zp_dedup
);
1447 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1448 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1449 return (ZIO_PIPELINE_CONTINUE
);
1452 ASSERT(!zp
->zp_nopwrite
);
1454 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1455 return (ZIO_PIPELINE_CONTINUE
);
1457 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1458 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1460 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1462 BP_SET_DEDUP(bp
, 1);
1463 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1464 return (ZIO_PIPELINE_CONTINUE
);
1468 * We were unable to handle this as an override bp, treat
1469 * it as a regular write I/O.
1471 zio
->io_bp_override
= NULL
;
1472 *bp
= zio
->io_bp_orig
;
1473 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1476 return (ZIO_PIPELINE_CONTINUE
);
1480 zio_write_compress(zio_t
*zio
)
1482 spa_t
*spa
= zio
->io_spa
;
1483 zio_prop_t
*zp
= &zio
->io_prop
;
1484 enum zio_compress compress
= zp
->zp_compress
;
1485 blkptr_t
*bp
= zio
->io_bp
;
1486 uint64_t lsize
= zio
->io_lsize
;
1487 uint64_t psize
= zio
->io_size
;
1491 * If our children haven't all reached the ready stage,
1492 * wait for them and then repeat this pipeline stage.
1494 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1495 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1496 return (ZIO_PIPELINE_STOP
);
1499 if (!IO_IS_ALLOCATING(zio
))
1500 return (ZIO_PIPELINE_CONTINUE
);
1502 if (zio
->io_children_ready
!= NULL
) {
1504 * Now that all our children are ready, run the callback
1505 * associated with this zio in case it wants to modify the
1506 * data to be written.
1508 ASSERT3U(zp
->zp_level
, >, 0);
1509 zio
->io_children_ready(zio
);
1512 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1513 ASSERT(zio
->io_bp_override
== NULL
);
1515 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1517 * We're rewriting an existing block, which means we're
1518 * working on behalf of spa_sync(). For spa_sync() to
1519 * converge, it must eventually be the case that we don't
1520 * have to allocate new blocks. But compression changes
1521 * the blocksize, which forces a reallocate, and makes
1522 * convergence take longer. Therefore, after the first
1523 * few passes, stop compressing to ensure convergence.
1525 pass
= spa_sync_pass(spa
);
1527 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1528 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1529 ASSERT(!BP_GET_DEDUP(bp
));
1531 if (pass
>= zfs_sync_pass_dont_compress
)
1532 compress
= ZIO_COMPRESS_OFF
;
1534 /* Make sure someone doesn't change their mind on overwrites */
1535 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1536 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1539 /* If it's a compressed write that is not raw, compress the buffer. */
1540 if (compress
!= ZIO_COMPRESS_OFF
&&
1541 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1542 void *cbuf
= zio_buf_alloc(lsize
);
1543 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1544 if (psize
== 0 || psize
== lsize
) {
1545 compress
= ZIO_COMPRESS_OFF
;
1546 zio_buf_free(cbuf
, lsize
);
1547 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1548 psize
<= BPE_PAYLOAD_SIZE
&&
1549 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1550 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1551 encode_embedded_bp_compressed(bp
,
1552 cbuf
, compress
, lsize
, psize
);
1553 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1554 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1555 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1556 zio_buf_free(cbuf
, lsize
);
1557 bp
->blk_birth
= zio
->io_txg
;
1558 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1559 ASSERT(spa_feature_is_active(spa
,
1560 SPA_FEATURE_EMBEDDED_DATA
));
1561 return (ZIO_PIPELINE_CONTINUE
);
1564 * Round up compressed size up to the ashift
1565 * of the smallest-ashift device, and zero the tail.
1566 * This ensures that the compressed size of the BP
1567 * (and thus compressratio property) are correct,
1568 * in that we charge for the padding used to fill out
1571 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1572 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1573 1ULL << spa
->spa_min_ashift
);
1574 if (rounded
>= lsize
) {
1575 compress
= ZIO_COMPRESS_OFF
;
1576 zio_buf_free(cbuf
, lsize
);
1579 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1580 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1581 abd_zero_off(cdata
, psize
, rounded
- psize
);
1583 zio_push_transform(zio
, cdata
,
1584 psize
, lsize
, NULL
);
1589 * We were unable to handle this as an override bp, treat
1590 * it as a regular write I/O.
1592 zio
->io_bp_override
= NULL
;
1593 *bp
= zio
->io_bp_orig
;
1594 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1596 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1597 zp
->zp_type
== DMU_OT_DNODE
) {
1599 * The DMU actually relies on the zio layer's compression
1600 * to free metadnode blocks that have had all contained
1601 * dnodes freed. As a result, even when doing a raw
1602 * receive, we must check whether the block can be compressed
1605 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1606 zio
->io_abd
, NULL
, lsize
);
1608 compress
= ZIO_COMPRESS_OFF
;
1610 ASSERT3U(psize
, !=, 0);
1614 * The final pass of spa_sync() must be all rewrites, but the first
1615 * few passes offer a trade-off: allocating blocks defers convergence,
1616 * but newly allocated blocks are sequential, so they can be written
1617 * to disk faster. Therefore, we allow the first few passes of
1618 * spa_sync() to allocate new blocks, but force rewrites after that.
1619 * There should only be a handful of blocks after pass 1 in any case.
1621 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1622 BP_GET_PSIZE(bp
) == psize
&&
1623 pass
>= zfs_sync_pass_rewrite
) {
1625 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1626 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1627 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1630 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1634 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1635 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1636 BP_SET_LSIZE(bp
, lsize
);
1637 BP_SET_TYPE(bp
, zp
->zp_type
);
1638 BP_SET_LEVEL(bp
, zp
->zp_level
);
1639 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1641 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1643 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1644 BP_SET_LSIZE(bp
, lsize
);
1645 BP_SET_TYPE(bp
, zp
->zp_type
);
1646 BP_SET_LEVEL(bp
, zp
->zp_level
);
1647 BP_SET_PSIZE(bp
, psize
);
1648 BP_SET_COMPRESS(bp
, compress
);
1649 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1650 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1651 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1653 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1654 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1655 ASSERT(!zp
->zp_encrypt
||
1656 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1657 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1659 if (zp
->zp_nopwrite
) {
1660 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1661 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1662 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1665 return (ZIO_PIPELINE_CONTINUE
);
1669 zio_free_bp_init(zio_t
*zio
)
1671 blkptr_t
*bp
= zio
->io_bp
;
1673 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1674 if (BP_GET_DEDUP(bp
))
1675 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1678 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1680 return (ZIO_PIPELINE_CONTINUE
);
1684 * ==========================================================================
1685 * Execute the I/O pipeline
1686 * ==========================================================================
1690 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1692 spa_t
*spa
= zio
->io_spa
;
1693 zio_type_t t
= zio
->io_type
;
1694 int flags
= (cutinline
? TQ_FRONT
: 0);
1697 * If we're a config writer or a probe, the normal issue and
1698 * interrupt threads may all be blocked waiting for the config lock.
1699 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1701 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1705 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1707 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1711 * If this is a high priority I/O, then use the high priority taskq if
1714 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1715 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1718 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1721 * NB: We are assuming that the zio can only be dispatched
1722 * to a single taskq at a time. It would be a grievous error
1723 * to dispatch the zio to another taskq at the same time.
1725 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1726 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1727 flags
, &zio
->io_tqent
);
1731 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1733 kthread_t
*executor
= zio
->io_executor
;
1734 spa_t
*spa
= zio
->io_spa
;
1736 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1737 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1739 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1740 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1749 zio_issue_async(zio_t
*zio
)
1751 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1753 return (ZIO_PIPELINE_STOP
);
1757 zio_interrupt(zio_t
*zio
)
1759 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1763 zio_delay_interrupt(zio_t
*zio
)
1766 * The timeout_generic() function isn't defined in userspace, so
1767 * rather than trying to implement the function, the zio delay
1768 * functionality has been disabled for userspace builds.
1773 * If io_target_timestamp is zero, then no delay has been registered
1774 * for this IO, thus jump to the end of this function and "skip" the
1775 * delay; issuing it directly to the zio layer.
1777 if (zio
->io_target_timestamp
!= 0) {
1778 hrtime_t now
= gethrtime();
1780 if (now
>= zio
->io_target_timestamp
) {
1782 * This IO has already taken longer than the target
1783 * delay to complete, so we don't want to delay it
1784 * any longer; we "miss" the delay and issue it
1785 * directly to the zio layer. This is likely due to
1786 * the target latency being set to a value less than
1787 * the underlying hardware can satisfy (e.g. delay
1788 * set to 1ms, but the disks take 10ms to complete an
1792 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1798 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1799 clock_t expire_at_tick
= ddi_get_lbolt() +
1802 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1803 hrtime_t
, now
, hrtime_t
, diff
);
1805 if (NSEC_TO_TICK(diff
) == 0) {
1806 /* Our delay is less than a jiffy - just spin */
1807 zfs_sleep_until(zio
->io_target_timestamp
);
1810 * Use taskq_dispatch_delay() in the place of
1811 * OpenZFS's timeout_generic().
1813 tid
= taskq_dispatch_delay(system_taskq
,
1814 (task_func_t
*)zio_interrupt
,
1815 zio
, TQ_NOSLEEP
, expire_at_tick
);
1816 if (tid
== TASKQID_INVALID
) {
1818 * Couldn't allocate a task. Just
1819 * finish the zio without a delay.
1828 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1833 zio_deadman_impl(zio_t
*pio
)
1835 zio_t
*cio
, *cio_next
;
1836 zio_link_t
*zl
= NULL
;
1837 vdev_t
*vd
= pio
->io_vd
;
1839 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
1840 vdev_queue_t
*vq
= &vd
->vdev_queue
;
1841 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1842 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1843 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1845 zfs_dbgmsg("slow zio: zio=%p timestamp=%llu "
1846 "delta=%llu queued=%llu io=%llu "
1847 "path=%s last=%llu "
1848 "type=%d priority=%d flags=0x%x "
1849 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1850 "objset=%llu object=%llu level=%llu blkid=%llu "
1851 "offset=%llu size=%llu error=%d",
1852 pio
, pio
->io_timestamp
,
1853 delta
, pio
->io_delta
, pio
->io_delay
,
1854 vd
->vdev_path
, vq
->vq_io_complete_ts
,
1855 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
1856 pio
->io_state
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
1857 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
1858 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
1859 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
1860 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
1862 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
1863 taskq_empty_ent(&pio
->io_tqent
)) {
1868 mutex_enter(&pio
->io_lock
);
1869 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1870 cio_next
= zio_walk_children(pio
, &zl
);
1871 zio_deadman_impl(cio
);
1873 mutex_exit(&pio
->io_lock
);
1877 * Log the critical information describing this zio and all of its children
1878 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1881 zio_deadman(zio_t
*pio
, char *tag
)
1883 spa_t
*spa
= pio
->io_spa
;
1884 char *name
= spa_name(spa
);
1886 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
1889 zio_deadman_impl(pio
);
1891 switch (spa_get_deadman_failmode(spa
)) {
1892 case ZIO_FAILURE_MODE_WAIT
:
1893 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
1896 case ZIO_FAILURE_MODE_CONTINUE
:
1897 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
1900 case ZIO_FAILURE_MODE_PANIC
:
1901 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
1907 * Execute the I/O pipeline until one of the following occurs:
1908 * (1) the I/O completes; (2) the pipeline stalls waiting for
1909 * dependent child I/Os; (3) the I/O issues, so we're waiting
1910 * for an I/O completion interrupt; (4) the I/O is delegated by
1911 * vdev-level caching or aggregation; (5) the I/O is deferred
1912 * due to vdev-level queueing; (6) the I/O is handed off to
1913 * another thread. In all cases, the pipeline stops whenever
1914 * there's no CPU work; it never burns a thread in cv_wait_io().
1916 * There's no locking on io_stage because there's no legitimate way
1917 * for multiple threads to be attempting to process the same I/O.
1919 static zio_pipe_stage_t
*zio_pipeline
[];
1922 * zio_execute() is a wrapper around the static function
1923 * __zio_execute() so that we can force __zio_execute() to be
1924 * inlined. This reduces stack overhead which is important
1925 * because __zio_execute() is called recursively in several zio
1926 * code paths. zio_execute() itself cannot be inlined because
1927 * it is externally visible.
1930 zio_execute(zio_t
*zio
)
1932 fstrans_cookie_t cookie
;
1934 cookie
= spl_fstrans_mark();
1936 spl_fstrans_unmark(cookie
);
1940 * Used to determine if in the current context the stack is sized large
1941 * enough to allow zio_execute() to be called recursively. A minimum
1942 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1945 zio_execute_stack_check(zio_t
*zio
)
1947 #if !defined(HAVE_LARGE_STACKS)
1948 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1950 /* Executing in txg_sync_thread() context. */
1951 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1954 /* Pool initialization outside of zio_taskq context. */
1955 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1956 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1957 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1959 #endif /* HAVE_LARGE_STACKS */
1964 __attribute__((always_inline
))
1966 __zio_execute(zio_t
*zio
)
1968 zio
->io_executor
= curthread
;
1970 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1972 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1973 enum zio_stage pipeline
= zio
->io_pipeline
;
1974 enum zio_stage stage
= zio
->io_stage
;
1977 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1978 ASSERT(ISP2(stage
));
1979 ASSERT(zio
->io_stall
== NULL
);
1983 } while ((stage
& pipeline
) == 0);
1985 ASSERT(stage
<= ZIO_STAGE_DONE
);
1988 * If we are in interrupt context and this pipeline stage
1989 * will grab a config lock that is held across I/O,
1990 * or may wait for an I/O that needs an interrupt thread
1991 * to complete, issue async to avoid deadlock.
1993 * For VDEV_IO_START, we cut in line so that the io will
1994 * be sent to disk promptly.
1996 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1997 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1998 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1999 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2000 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2005 * If the current context doesn't have large enough stacks
2006 * the zio must be issued asynchronously to prevent overflow.
2008 if (zio_execute_stack_check(zio
)) {
2009 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2010 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2011 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2015 zio
->io_stage
= stage
;
2016 zio
->io_pipeline_trace
|= zio
->io_stage
;
2017 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2019 if (rv
== ZIO_PIPELINE_STOP
)
2022 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
2028 * ==========================================================================
2029 * Initiate I/O, either sync or async
2030 * ==========================================================================
2033 zio_wait(zio_t
*zio
)
2035 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2038 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2039 ASSERT3P(zio
->io_executor
, ==, NULL
);
2041 zio
->io_waiter
= curthread
;
2042 ASSERT0(zio
->io_queued_timestamp
);
2043 zio
->io_queued_timestamp
= gethrtime();
2047 mutex_enter(&zio
->io_lock
);
2048 while (zio
->io_executor
!= NULL
) {
2049 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2050 ddi_get_lbolt() + timeout
);
2052 if (zfs_deadman_enabled
&& error
== -1 &&
2053 gethrtime() - zio
->io_queued_timestamp
>
2054 spa_deadman_ziotime(zio
->io_spa
)) {
2055 mutex_exit(&zio
->io_lock
);
2056 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2057 zio_deadman(zio
, FTAG
);
2058 mutex_enter(&zio
->io_lock
);
2061 mutex_exit(&zio
->io_lock
);
2063 error
= zio
->io_error
;
2070 zio_nowait(zio_t
*zio
)
2072 ASSERT3P(zio
->io_executor
, ==, NULL
);
2074 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2075 zio_unique_parent(zio
) == NULL
) {
2079 * This is a logical async I/O with no parent to wait for it.
2080 * We add it to the spa_async_root_zio "Godfather" I/O which
2081 * will ensure they complete prior to unloading the pool.
2083 spa_t
*spa
= zio
->io_spa
;
2085 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2088 zio_add_child(pio
, zio
);
2091 ASSERT0(zio
->io_queued_timestamp
);
2092 zio
->io_queued_timestamp
= gethrtime();
2097 * ==========================================================================
2098 * Reexecute, cancel, or suspend/resume failed I/O
2099 * ==========================================================================
2103 zio_reexecute(zio_t
*pio
)
2105 zio_t
*cio
, *cio_next
;
2107 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2108 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2109 ASSERT(pio
->io_gang_leader
== NULL
);
2110 ASSERT(pio
->io_gang_tree
== NULL
);
2112 pio
->io_flags
= pio
->io_orig_flags
;
2113 pio
->io_stage
= pio
->io_orig_stage
;
2114 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2115 pio
->io_reexecute
= 0;
2116 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2117 pio
->io_pipeline_trace
= 0;
2119 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2120 pio
->io_state
[w
] = 0;
2121 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2122 pio
->io_child_error
[c
] = 0;
2124 if (IO_IS_ALLOCATING(pio
))
2125 BP_ZERO(pio
->io_bp
);
2128 * As we reexecute pio's children, new children could be created.
2129 * New children go to the head of pio's io_child_list, however,
2130 * so we will (correctly) not reexecute them. The key is that
2131 * the remainder of pio's io_child_list, from 'cio_next' onward,
2132 * cannot be affected by any side effects of reexecuting 'cio'.
2134 zio_link_t
*zl
= NULL
;
2135 mutex_enter(&pio
->io_lock
);
2136 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2137 cio_next
= zio_walk_children(pio
, &zl
);
2138 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2139 pio
->io_children
[cio
->io_child_type
][w
]++;
2140 mutex_exit(&pio
->io_lock
);
2142 mutex_enter(&pio
->io_lock
);
2144 mutex_exit(&pio
->io_lock
);
2147 * Now that all children have been reexecuted, execute the parent.
2148 * We don't reexecute "The Godfather" I/O here as it's the
2149 * responsibility of the caller to wait on it.
2151 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2152 pio
->io_queued_timestamp
= gethrtime();
2158 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2160 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2161 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2162 "failure and the failure mode property for this pool "
2163 "is set to panic.", spa_name(spa
));
2165 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2166 "failure and has been suspended.\n", spa_name(spa
));
2168 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2171 mutex_enter(&spa
->spa_suspend_lock
);
2173 if (spa
->spa_suspend_zio_root
== NULL
)
2174 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2175 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2176 ZIO_FLAG_GODFATHER
);
2178 spa
->spa_suspended
= reason
;
2181 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2182 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2183 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2184 ASSERT(zio_unique_parent(zio
) == NULL
);
2185 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2186 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2189 mutex_exit(&spa
->spa_suspend_lock
);
2193 zio_resume(spa_t
*spa
)
2198 * Reexecute all previously suspended i/o.
2200 mutex_enter(&spa
->spa_suspend_lock
);
2201 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2202 cv_broadcast(&spa
->spa_suspend_cv
);
2203 pio
= spa
->spa_suspend_zio_root
;
2204 spa
->spa_suspend_zio_root
= NULL
;
2205 mutex_exit(&spa
->spa_suspend_lock
);
2211 return (zio_wait(pio
));
2215 zio_resume_wait(spa_t
*spa
)
2217 mutex_enter(&spa
->spa_suspend_lock
);
2218 while (spa_suspended(spa
))
2219 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2220 mutex_exit(&spa
->spa_suspend_lock
);
2224 * ==========================================================================
2227 * A gang block is a collection of small blocks that looks to the DMU
2228 * like one large block. When zio_dva_allocate() cannot find a block
2229 * of the requested size, due to either severe fragmentation or the pool
2230 * being nearly full, it calls zio_write_gang_block() to construct the
2231 * block from smaller fragments.
2233 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2234 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2235 * an indirect block: it's an array of block pointers. It consumes
2236 * only one sector and hence is allocatable regardless of fragmentation.
2237 * The gang header's bps point to its gang members, which hold the data.
2239 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2240 * as the verifier to ensure uniqueness of the SHA256 checksum.
2241 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2242 * not the gang header. This ensures that data block signatures (needed for
2243 * deduplication) are independent of how the block is physically stored.
2245 * Gang blocks can be nested: a gang member may itself be a gang block.
2246 * Thus every gang block is a tree in which root and all interior nodes are
2247 * gang headers, and the leaves are normal blocks that contain user data.
2248 * The root of the gang tree is called the gang leader.
2250 * To perform any operation (read, rewrite, free, claim) on a gang block,
2251 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2252 * in the io_gang_tree field of the original logical i/o by recursively
2253 * reading the gang leader and all gang headers below it. This yields
2254 * an in-core tree containing the contents of every gang header and the
2255 * bps for every constituent of the gang block.
2257 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2258 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2259 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2260 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2261 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2262 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2263 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2264 * of the gang header plus zio_checksum_compute() of the data to update the
2265 * gang header's blk_cksum as described above.
2267 * The two-phase assemble/issue model solves the problem of partial failure --
2268 * what if you'd freed part of a gang block but then couldn't read the
2269 * gang header for another part? Assembling the entire gang tree first
2270 * ensures that all the necessary gang header I/O has succeeded before
2271 * starting the actual work of free, claim, or write. Once the gang tree
2272 * is assembled, free and claim are in-memory operations that cannot fail.
2274 * In the event that a gang write fails, zio_dva_unallocate() walks the
2275 * gang tree to immediately free (i.e. insert back into the space map)
2276 * everything we've allocated. This ensures that we don't get ENOSPC
2277 * errors during repeated suspend/resume cycles due to a flaky device.
2279 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2280 * the gang tree, we won't modify the block, so we can safely defer the free
2281 * (knowing that the block is still intact). If we *can* assemble the gang
2282 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2283 * each constituent bp and we can allocate a new block on the next sync pass.
2285 * In all cases, the gang tree allows complete recovery from partial failure.
2286 * ==========================================================================
2290 zio_gang_issue_func_done(zio_t
*zio
)
2292 abd_put(zio
->io_abd
);
2296 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2302 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2303 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2304 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2305 &pio
->io_bookmark
));
2309 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2316 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2317 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2318 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2319 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2322 * As we rewrite each gang header, the pipeline will compute
2323 * a new gang block header checksum for it; but no one will
2324 * compute a new data checksum, so we do that here. The one
2325 * exception is the gang leader: the pipeline already computed
2326 * its data checksum because that stage precedes gang assembly.
2327 * (Presently, nothing actually uses interior data checksums;
2328 * this is just good hygiene.)
2330 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2331 abd_t
*buf
= abd_get_offset(data
, offset
);
2333 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2334 buf
, BP_GET_PSIZE(bp
));
2339 * If we are here to damage data for testing purposes,
2340 * leave the GBH alone so that we can detect the damage.
2342 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2343 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2345 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2346 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2347 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2348 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2356 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2359 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2360 ZIO_GANG_CHILD_FLAGS(pio
)));
2365 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2368 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2369 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2372 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2381 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2383 static zio_gang_node_t
*
2384 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2386 zio_gang_node_t
*gn
;
2388 ASSERT(*gnpp
== NULL
);
2390 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2391 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2398 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2400 zio_gang_node_t
*gn
= *gnpp
;
2402 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2403 ASSERT(gn
->gn_child
[g
] == NULL
);
2405 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2406 kmem_free(gn
, sizeof (*gn
));
2411 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2413 zio_gang_node_t
*gn
= *gnpp
;
2418 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2419 zio_gang_tree_free(&gn
->gn_child
[g
]);
2421 zio_gang_node_free(gnpp
);
2425 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2427 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2428 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2430 ASSERT(gio
->io_gang_leader
== gio
);
2431 ASSERT(BP_IS_GANG(bp
));
2433 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2434 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2435 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2439 zio_gang_tree_assemble_done(zio_t
*zio
)
2441 zio_t
*gio
= zio
->io_gang_leader
;
2442 zio_gang_node_t
*gn
= zio
->io_private
;
2443 blkptr_t
*bp
= zio
->io_bp
;
2445 ASSERT(gio
== zio_unique_parent(zio
));
2446 ASSERT(zio
->io_child_count
== 0);
2451 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2452 if (BP_SHOULD_BYTESWAP(bp
))
2453 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2455 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2456 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2457 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2459 abd_put(zio
->io_abd
);
2461 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2462 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2463 if (!BP_IS_GANG(gbp
))
2465 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2470 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2473 zio_t
*gio
= pio
->io_gang_leader
;
2476 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2477 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2478 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2481 * If you're a gang header, your data is in gn->gn_gbh.
2482 * If you're a gang member, your data is in 'data' and gn == NULL.
2484 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2487 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2489 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2490 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2491 if (BP_IS_HOLE(gbp
))
2493 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2495 offset
+= BP_GET_PSIZE(gbp
);
2499 if (gn
== gio
->io_gang_tree
)
2500 ASSERT3U(gio
->io_size
, ==, offset
);
2507 zio_gang_assemble(zio_t
*zio
)
2509 blkptr_t
*bp
= zio
->io_bp
;
2511 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2512 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2514 zio
->io_gang_leader
= zio
;
2516 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2518 return (ZIO_PIPELINE_CONTINUE
);
2522 zio_gang_issue(zio_t
*zio
)
2524 blkptr_t
*bp
= zio
->io_bp
;
2526 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2527 return (ZIO_PIPELINE_STOP
);
2530 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2531 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2533 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2534 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2537 zio_gang_tree_free(&zio
->io_gang_tree
);
2539 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2541 return (ZIO_PIPELINE_CONTINUE
);
2545 zio_write_gang_member_ready(zio_t
*zio
)
2547 zio_t
*pio
= zio_unique_parent(zio
);
2548 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2549 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2551 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
);
2553 if (BP_IS_HOLE(zio
->io_bp
))
2556 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2558 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2559 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2560 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2561 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2562 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2564 mutex_enter(&pio
->io_lock
);
2565 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2566 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2567 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2568 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2569 DVA_SET_ASIZE(&pdva
[d
], asize
);
2571 mutex_exit(&pio
->io_lock
);
2575 zio_write_gang_done(zio_t
*zio
)
2577 abd_put(zio
->io_abd
);
2581 zio_write_gang_block(zio_t
*pio
)
2583 spa_t
*spa
= pio
->io_spa
;
2584 metaslab_class_t
*mc
= spa_normal_class(spa
);
2585 blkptr_t
*bp
= pio
->io_bp
;
2586 zio_t
*gio
= pio
->io_gang_leader
;
2588 zio_gang_node_t
*gn
, **gnpp
;
2589 zio_gbh_phys_t
*gbh
;
2591 uint64_t txg
= pio
->io_txg
;
2592 uint64_t resid
= pio
->io_size
;
2594 int copies
= gio
->io_prop
.zp_copies
;
2600 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2601 * have a third copy.
2603 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2604 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2605 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2607 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2608 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2609 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2610 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2612 flags
|= METASLAB_ASYNC_ALLOC
;
2613 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2616 * The logical zio has already placed a reservation for
2617 * 'copies' allocation slots but gang blocks may require
2618 * additional copies. These additional copies
2619 * (i.e. gbh_copies - copies) are guaranteed to succeed
2620 * since metaslab_class_throttle_reserve() always allows
2621 * additional reservations for gang blocks.
2623 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2627 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2628 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2629 &pio
->io_alloc_list
, pio
);
2631 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2632 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2633 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2636 * If we failed to allocate the gang block header then
2637 * we remove any additional allocation reservations that
2638 * we placed here. The original reservation will
2639 * be removed when the logical I/O goes to the ready
2642 metaslab_class_throttle_unreserve(mc
,
2643 gbh_copies
- copies
, pio
);
2646 pio
->io_error
= error
;
2647 return (ZIO_PIPELINE_CONTINUE
);
2651 gnpp
= &gio
->io_gang_tree
;
2653 gnpp
= pio
->io_private
;
2654 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2657 gn
= zio_gang_node_alloc(gnpp
);
2659 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2660 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2663 * Create the gang header.
2665 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2666 zio_write_gang_done
, NULL
, pio
->io_priority
,
2667 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2670 * Create and nowait the gang children.
2672 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2673 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2675 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2677 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2678 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2679 zp
.zp_type
= DMU_OT_NONE
;
2681 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2682 zp
.zp_dedup
= B_FALSE
;
2683 zp
.zp_dedup_verify
= B_FALSE
;
2684 zp
.zp_nopwrite
= B_FALSE
;
2685 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2686 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2687 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2688 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2689 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2691 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2692 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2693 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2694 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2695 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2697 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2698 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2699 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2702 * Gang children won't throttle but we should
2703 * account for their work, so reserve an allocation
2704 * slot for them here.
2706 VERIFY(metaslab_class_throttle_reserve(mc
,
2707 zp
.zp_copies
, cio
, flags
));
2713 * Set pio's pipeline to just wait for zio to finish.
2715 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2718 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2720 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2724 return (ZIO_PIPELINE_CONTINUE
);
2728 * The zio_nop_write stage in the pipeline determines if allocating a
2729 * new bp is necessary. The nopwrite feature can handle writes in
2730 * either syncing or open context (i.e. zil writes) and as a result is
2731 * mutually exclusive with dedup.
2733 * By leveraging a cryptographically secure checksum, such as SHA256, we
2734 * can compare the checksums of the new data and the old to determine if
2735 * allocating a new block is required. Note that our requirements for
2736 * cryptographic strength are fairly weak: there can't be any accidental
2737 * hash collisions, but we don't need to be secure against intentional
2738 * (malicious) collisions. To trigger a nopwrite, you have to be able
2739 * to write the file to begin with, and triggering an incorrect (hash
2740 * collision) nopwrite is no worse than simply writing to the file.
2741 * That said, there are no known attacks against the checksum algorithms
2742 * used for nopwrite, assuming that the salt and the checksums
2743 * themselves remain secret.
2746 zio_nop_write(zio_t
*zio
)
2748 blkptr_t
*bp
= zio
->io_bp
;
2749 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2750 zio_prop_t
*zp
= &zio
->io_prop
;
2752 ASSERT(BP_GET_LEVEL(bp
) == 0);
2753 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2754 ASSERT(zp
->zp_nopwrite
);
2755 ASSERT(!zp
->zp_dedup
);
2756 ASSERT(zio
->io_bp_override
== NULL
);
2757 ASSERT(IO_IS_ALLOCATING(zio
));
2760 * Check to see if the original bp and the new bp have matching
2761 * characteristics (i.e. same checksum, compression algorithms, etc).
2762 * If they don't then just continue with the pipeline which will
2763 * allocate a new bp.
2765 if (BP_IS_HOLE(bp_orig
) ||
2766 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2767 ZCHECKSUM_FLAG_NOPWRITE
) ||
2768 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2769 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2770 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2771 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2772 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2773 return (ZIO_PIPELINE_CONTINUE
);
2776 * If the checksums match then reset the pipeline so that we
2777 * avoid allocating a new bp and issuing any I/O.
2779 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2780 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2781 ZCHECKSUM_FLAG_NOPWRITE
);
2782 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2783 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2784 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2785 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2786 sizeof (uint64_t)) == 0);
2789 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2790 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2793 return (ZIO_PIPELINE_CONTINUE
);
2797 * ==========================================================================
2799 * ==========================================================================
2802 zio_ddt_child_read_done(zio_t
*zio
)
2804 blkptr_t
*bp
= zio
->io_bp
;
2805 ddt_entry_t
*dde
= zio
->io_private
;
2807 zio_t
*pio
= zio_unique_parent(zio
);
2809 mutex_enter(&pio
->io_lock
);
2810 ddp
= ddt_phys_select(dde
, bp
);
2811 if (zio
->io_error
== 0)
2812 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2814 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2815 dde
->dde_repair_abd
= zio
->io_abd
;
2817 abd_free(zio
->io_abd
);
2818 mutex_exit(&pio
->io_lock
);
2822 zio_ddt_read_start(zio_t
*zio
)
2824 blkptr_t
*bp
= zio
->io_bp
;
2826 ASSERT(BP_GET_DEDUP(bp
));
2827 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2828 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2830 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2831 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2832 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2833 ddt_phys_t
*ddp
= dde
->dde_phys
;
2834 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2837 ASSERT(zio
->io_vsd
== NULL
);
2840 if (ddp_self
== NULL
)
2841 return (ZIO_PIPELINE_CONTINUE
);
2843 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2844 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2846 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2848 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2849 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2850 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2851 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2852 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2854 return (ZIO_PIPELINE_CONTINUE
);
2857 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2858 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2859 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2861 return (ZIO_PIPELINE_CONTINUE
);
2865 zio_ddt_read_done(zio_t
*zio
)
2867 blkptr_t
*bp
= zio
->io_bp
;
2869 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
2870 return (ZIO_PIPELINE_STOP
);
2873 ASSERT(BP_GET_DEDUP(bp
));
2874 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2875 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2877 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2878 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2879 ddt_entry_t
*dde
= zio
->io_vsd
;
2881 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2882 return (ZIO_PIPELINE_CONTINUE
);
2885 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2886 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2887 return (ZIO_PIPELINE_STOP
);
2889 if (dde
->dde_repair_abd
!= NULL
) {
2890 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2892 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2894 ddt_repair_done(ddt
, dde
);
2898 ASSERT(zio
->io_vsd
== NULL
);
2900 return (ZIO_PIPELINE_CONTINUE
);
2904 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2906 spa_t
*spa
= zio
->io_spa
;
2907 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2909 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2912 * Note: we compare the original data, not the transformed data,
2913 * because when zio->io_bp is an override bp, we will not have
2914 * pushed the I/O transforms. That's an important optimization
2915 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2916 * However, we should never get a raw, override zio so in these
2917 * cases we can compare the io_abd directly. This is useful because
2918 * it allows us to do dedup verification even if we don't have access
2919 * to the original data (for instance, if the encryption keys aren't
2923 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2924 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2926 if (lio
!= NULL
&& do_raw
) {
2927 return (lio
->io_size
!= zio
->io_size
||
2928 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2929 } else if (lio
!= NULL
) {
2930 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2931 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2935 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2936 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2938 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2939 blkptr_t blk
= *zio
->io_bp
;
2944 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2945 psize
= BP_GET_PSIZE(&blk
);
2947 if (psize
!= zio
->io_size
)
2952 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2954 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2955 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2956 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2957 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
2960 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
2961 error
= SET_ERROR(ENOENT
);
2966 return (error
!= 0);
2967 } else if (ddp
->ddp_phys_birth
!= 0) {
2968 arc_buf_t
*abuf
= NULL
;
2969 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2970 blkptr_t blk
= *zio
->io_bp
;
2973 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2975 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
2980 error
= arc_read(NULL
, spa
, &blk
,
2981 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2982 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2983 &aflags
, &zio
->io_bookmark
);
2986 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2987 zio
->io_orig_size
) != 0)
2988 error
= SET_ERROR(ENOENT
);
2989 arc_buf_destroy(abuf
, &abuf
);
2993 return (error
!= 0);
3001 zio_ddt_child_write_ready(zio_t
*zio
)
3003 int p
= zio
->io_prop
.zp_copies
;
3004 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3005 ddt_entry_t
*dde
= zio
->io_private
;
3006 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3014 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3016 ddt_phys_fill(ddp
, zio
->io_bp
);
3018 zio_link_t
*zl
= NULL
;
3019 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3020 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3026 zio_ddt_child_write_done(zio_t
*zio
)
3028 int p
= zio
->io_prop
.zp_copies
;
3029 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3030 ddt_entry_t
*dde
= zio
->io_private
;
3031 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3035 ASSERT(ddp
->ddp_refcnt
== 0);
3036 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3037 dde
->dde_lead_zio
[p
] = NULL
;
3039 if (zio
->io_error
== 0) {
3040 zio_link_t
*zl
= NULL
;
3041 while (zio_walk_parents(zio
, &zl
) != NULL
)
3042 ddt_phys_addref(ddp
);
3044 ddt_phys_clear(ddp
);
3051 zio_ddt_ditto_write_done(zio_t
*zio
)
3053 int p
= DDT_PHYS_DITTO
;
3054 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
3055 blkptr_t
*bp
= zio
->io_bp
;
3056 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3057 ddt_entry_t
*dde
= zio
->io_private
;
3058 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3059 ddt_key_t
*ddk
= &dde
->dde_key
;
3063 ASSERT(ddp
->ddp_refcnt
== 0);
3064 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3065 dde
->dde_lead_zio
[p
] = NULL
;
3067 if (zio
->io_error
== 0) {
3068 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
3069 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
3070 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
3071 if (ddp
->ddp_phys_birth
!= 0)
3072 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
3073 ddt_phys_fill(ddp
, bp
);
3080 zio_ddt_write(zio_t
*zio
)
3082 spa_t
*spa
= zio
->io_spa
;
3083 blkptr_t
*bp
= zio
->io_bp
;
3084 uint64_t txg
= zio
->io_txg
;
3085 zio_prop_t
*zp
= &zio
->io_prop
;
3086 int p
= zp
->zp_copies
;
3090 ddt_t
*ddt
= ddt_select(spa
, bp
);
3094 ASSERT(BP_GET_DEDUP(bp
));
3095 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3096 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3097 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3100 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3101 ddp
= &dde
->dde_phys
[p
];
3103 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3105 * If we're using a weak checksum, upgrade to a strong checksum
3106 * and try again. If we're already using a strong checksum,
3107 * we can't resolve it, so just convert to an ordinary write.
3108 * (And automatically e-mail a paper to Nature?)
3110 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3111 ZCHECKSUM_FLAG_DEDUP
)) {
3112 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3113 zio_pop_transforms(zio
);
3114 zio
->io_stage
= ZIO_STAGE_OPEN
;
3117 zp
->zp_dedup
= B_FALSE
;
3119 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3121 return (ZIO_PIPELINE_CONTINUE
);
3124 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
3125 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
3127 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
3128 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
3129 zio_prop_t czp
= *zp
;
3131 czp
.zp_copies
= ditto_copies
;
3134 * If we arrived here with an override bp, we won't have run
3135 * the transform stack, so we won't have the data we need to
3136 * generate a child i/o. So, toss the override bp and restart.
3137 * This is safe, because using the override bp is just an
3138 * optimization; and it's rare, so the cost doesn't matter.
3140 if (zio
->io_bp_override
) {
3141 zio_pop_transforms(zio
);
3142 zio
->io_stage
= ZIO_STAGE_OPEN
;
3143 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3144 zio
->io_bp_override
= NULL
;
3147 return (ZIO_PIPELINE_CONTINUE
);
3150 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3151 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
3152 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
3153 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3155 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3156 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
3159 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3160 if (ddp
->ddp_phys_birth
!= 0)
3161 ddt_bp_fill(ddp
, bp
, txg
);
3162 if (dde
->dde_lead_zio
[p
] != NULL
)
3163 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3165 ddt_phys_addref(ddp
);
3166 } else if (zio
->io_bp_override
) {
3167 ASSERT(bp
->blk_birth
== txg
);
3168 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3169 ddt_phys_fill(ddp
, bp
);
3170 ddt_phys_addref(ddp
);
3172 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3173 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3174 zio_ddt_child_write_ready
, NULL
, NULL
,
3175 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3176 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3178 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3179 dde
->dde_lead_zio
[p
] = cio
;
3189 return (ZIO_PIPELINE_CONTINUE
);
3192 ddt_entry_t
*freedde
; /* for debugging */
3195 zio_ddt_free(zio_t
*zio
)
3197 spa_t
*spa
= zio
->io_spa
;
3198 blkptr_t
*bp
= zio
->io_bp
;
3199 ddt_t
*ddt
= ddt_select(spa
, bp
);
3203 ASSERT(BP_GET_DEDUP(bp
));
3204 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3207 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3209 ddp
= ddt_phys_select(dde
, bp
);
3211 ddt_phys_decref(ddp
);
3215 return (ZIO_PIPELINE_CONTINUE
);
3219 * ==========================================================================
3220 * Allocate and free blocks
3221 * ==========================================================================
3225 zio_io_to_allocate(spa_t
*spa
)
3229 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
3231 zio
= avl_first(&spa
->spa_alloc_tree
);
3235 ASSERT(IO_IS_ALLOCATING(zio
));
3238 * Try to place a reservation for this zio. If we're unable to
3239 * reserve then we throttle.
3241 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
3242 zio
->io_prop
.zp_copies
, zio
, 0)) {
3246 avl_remove(&spa
->spa_alloc_tree
, zio
);
3247 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3253 zio_dva_throttle(zio_t
*zio
)
3255 spa_t
*spa
= zio
->io_spa
;
3258 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3259 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
3260 zio
->io_child_type
== ZIO_CHILD_GANG
||
3261 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3262 return (ZIO_PIPELINE_CONTINUE
);
3265 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3267 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3268 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3270 mutex_enter(&spa
->spa_alloc_lock
);
3272 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3273 avl_add(&spa
->spa_alloc_tree
, zio
);
3275 nio
= zio_io_to_allocate(zio
->io_spa
);
3276 mutex_exit(&spa
->spa_alloc_lock
);
3279 return (ZIO_PIPELINE_CONTINUE
);
3282 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3284 * We are passing control to a new zio so make sure that
3285 * it is processed by a different thread. We do this to
3286 * avoid stack overflows that can occur when parents are
3287 * throttled and children are making progress. We allow
3288 * it to go to the head of the taskq since it's already
3291 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3293 return (ZIO_PIPELINE_STOP
);
3297 zio_allocate_dispatch(spa_t
*spa
)
3301 mutex_enter(&spa
->spa_alloc_lock
);
3302 zio
= zio_io_to_allocate(spa
);
3303 mutex_exit(&spa
->spa_alloc_lock
);
3307 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3308 ASSERT0(zio
->io_error
);
3309 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3313 zio_dva_allocate(zio_t
*zio
)
3315 spa_t
*spa
= zio
->io_spa
;
3316 metaslab_class_t
*mc
= spa_normal_class(spa
);
3317 blkptr_t
*bp
= zio
->io_bp
;
3321 if (zio
->io_gang_leader
== NULL
) {
3322 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3323 zio
->io_gang_leader
= zio
;
3326 ASSERT(BP_IS_HOLE(bp
));
3327 ASSERT0(BP_GET_NDVAS(bp
));
3328 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3329 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3330 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3332 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3333 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3334 flags
|= METASLAB_DONT_THROTTLE
;
3335 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3336 flags
|= METASLAB_GANG_CHILD
;
3337 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3338 flags
|= METASLAB_ASYNC_ALLOC
;
3340 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3341 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3342 &zio
->io_alloc_list
, zio
);
3345 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3346 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3348 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3349 return (zio_write_gang_block(zio
));
3350 zio
->io_error
= error
;
3353 return (ZIO_PIPELINE_CONTINUE
);
3357 zio_dva_free(zio_t
*zio
)
3359 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3361 return (ZIO_PIPELINE_CONTINUE
);
3365 zio_dva_claim(zio_t
*zio
)
3369 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3371 zio
->io_error
= error
;
3373 return (ZIO_PIPELINE_CONTINUE
);
3377 * Undo an allocation. This is used by zio_done() when an I/O fails
3378 * and we want to give back the block we just allocated.
3379 * This handles both normal blocks and gang blocks.
3382 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3384 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3385 ASSERT(zio
->io_bp_override
== NULL
);
3387 if (!BP_IS_HOLE(bp
))
3388 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3391 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3392 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3393 &gn
->gn_gbh
->zg_blkptr
[g
]);
3399 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3402 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3403 uint64_t size
, boolean_t
*slog
)
3406 zio_alloc_list_t io_alloc_list
;
3408 ASSERT(txg
> spa_syncing_txg(spa
));
3410 metaslab_trace_init(&io_alloc_list
);
3411 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3412 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
);
3416 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3417 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3418 &io_alloc_list
, NULL
);
3422 metaslab_trace_fini(&io_alloc_list
);
3425 BP_SET_LSIZE(new_bp
, size
);
3426 BP_SET_PSIZE(new_bp
, size
);
3427 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3428 BP_SET_CHECKSUM(new_bp
,
3429 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3430 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3431 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3432 BP_SET_LEVEL(new_bp
, 0);
3433 BP_SET_DEDUP(new_bp
, 0);
3434 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3437 * encrypted blocks will require an IV and salt. We generate
3438 * these now since we will not be rewriting the bp at
3441 if (os
->os_encrypted
) {
3442 uint8_t iv
[ZIO_DATA_IV_LEN
];
3443 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3445 BP_SET_CRYPT(new_bp
, B_TRUE
);
3446 VERIFY0(spa_crypt_get_salt(spa
,
3447 dmu_objset_id(os
), salt
));
3448 VERIFY0(zio_crypt_generate_iv(iv
));
3450 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3453 zfs_dbgmsg("%s: zil block allocation failure: "
3454 "size %llu, error %d", spa_name(spa
), size
, error
);
3461 * Free an intent log block.
3464 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
3466 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
3467 ASSERT(!BP_IS_GANG(bp
));
3469 zio_free(spa
, txg
, bp
);
3473 * ==========================================================================
3474 * Read and write to physical devices
3475 * ==========================================================================
3480 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3481 * stops after this stage and will resume upon I/O completion.
3482 * However, there are instances where the vdev layer may need to
3483 * continue the pipeline when an I/O was not issued. Since the I/O
3484 * that was sent to the vdev layer might be different than the one
3485 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3486 * force the underlying vdev layers to call either zio_execute() or
3487 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3490 zio_vdev_io_start(zio_t
*zio
)
3492 vdev_t
*vd
= zio
->io_vd
;
3494 spa_t
*spa
= zio
->io_spa
;
3498 ASSERT(zio
->io_error
== 0);
3499 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3502 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3503 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3506 * The mirror_ops handle multiple DVAs in a single BP.
3508 vdev_mirror_ops
.vdev_op_io_start(zio
);
3509 return (ZIO_PIPELINE_STOP
);
3512 ASSERT3P(zio
->io_logical
, !=, zio
);
3513 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3514 ASSERT(spa
->spa_trust_config
);
3517 * Note: the code can handle other kinds of writes,
3518 * but we don't expect them.
3520 if (zio
->io_vd
->vdev_removing
) {
3521 ASSERT(zio
->io_flags
&
3522 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3523 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3527 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3529 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3530 P2PHASE(zio
->io_size
, align
) != 0) {
3531 /* Transform logical writes to be a full physical block size. */
3532 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3533 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3534 ASSERT(vd
== vd
->vdev_top
);
3535 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3536 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3537 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3539 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3543 * If this is not a physical io, make sure that it is properly aligned
3544 * before proceeding.
3546 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3547 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3548 ASSERT0(P2PHASE(zio
->io_size
, align
));
3551 * For physical writes, we allow 512b aligned writes and assume
3552 * the device will perform a read-modify-write as necessary.
3554 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3555 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3558 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3561 * If this is a repair I/O, and there's no self-healing involved --
3562 * that is, we're just resilvering what we expect to resilver --
3563 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3564 * This prevents spurious resilvering.
3566 * There are a few ways that we can end up creating these spurious
3569 * 1. A resilver i/o will be issued if any DVA in the BP has a
3570 * dirty DTL. The mirror code will issue resilver writes to
3571 * each DVA, including the one(s) that are not on vdevs with dirty
3574 * 2. With nested replication, which happens when we have a
3575 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3576 * For example, given mirror(replacing(A+B), C), it's likely that
3577 * only A is out of date (it's the new device). In this case, we'll
3578 * read from C, then use the data to resilver A+B -- but we don't
3579 * actually want to resilver B, just A. The top-level mirror has no
3580 * way to know this, so instead we just discard unnecessary repairs
3581 * as we work our way down the vdev tree.
3583 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3584 * The same logic applies to any form of nested replication: ditto
3585 * + mirror, RAID-Z + replacing, etc.
3587 * However, indirect vdevs point off to other vdevs which may have
3588 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3589 * will be properly bypassed instead.
3591 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3592 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3593 zio
->io_txg
!= 0 && /* not a delegated i/o */
3594 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3595 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3596 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3597 zio_vdev_io_bypass(zio
);
3598 return (ZIO_PIPELINE_CONTINUE
);
3601 if (vd
->vdev_ops
->vdev_op_leaf
&&
3602 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3604 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3605 return (ZIO_PIPELINE_CONTINUE
);
3607 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3608 return (ZIO_PIPELINE_STOP
);
3610 if (!vdev_accessible(vd
, zio
)) {
3611 zio
->io_error
= SET_ERROR(ENXIO
);
3613 return (ZIO_PIPELINE_STOP
);
3615 zio
->io_delay
= gethrtime();
3618 vd
->vdev_ops
->vdev_op_io_start(zio
);
3619 return (ZIO_PIPELINE_STOP
);
3623 zio_vdev_io_done(zio_t
*zio
)
3625 vdev_t
*vd
= zio
->io_vd
;
3626 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3627 boolean_t unexpected_error
= B_FALSE
;
3629 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3630 return (ZIO_PIPELINE_STOP
);
3633 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3636 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3638 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3640 vdev_queue_io_done(zio
);
3642 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3643 vdev_cache_write(zio
);
3645 if (zio_injection_enabled
&& zio
->io_error
== 0)
3646 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3649 if (zio_injection_enabled
&& zio
->io_error
== 0)
3650 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3652 if (zio
->io_error
) {
3653 if (!vdev_accessible(vd
, zio
)) {
3654 zio
->io_error
= SET_ERROR(ENXIO
);
3656 unexpected_error
= B_TRUE
;
3661 ops
->vdev_op_io_done(zio
);
3663 if (unexpected_error
)
3664 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3666 return (ZIO_PIPELINE_CONTINUE
);
3670 * This function is used to change the priority of an existing zio that is
3671 * currently in-flight. This is used by the arc to upgrade priority in the
3672 * event that a demand read is made for a block that is currently queued
3673 * as a scrub or async read IO. Otherwise, the high priority read request
3674 * would end up having to wait for the lower priority IO.
3677 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3679 zio_t
*cio
, *cio_next
;
3680 zio_link_t
*zl
= NULL
;
3682 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3684 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3685 vdev_queue_change_io_priority(pio
, priority
);
3687 pio
->io_priority
= priority
;
3690 mutex_enter(&pio
->io_lock
);
3691 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3692 cio_next
= zio_walk_children(pio
, &zl
);
3693 zio_change_priority(cio
, priority
);
3695 mutex_exit(&pio
->io_lock
);
3699 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3700 * disk, and use that to finish the checksum ereport later.
3703 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3704 const abd_t
*good_buf
)
3706 /* no processing needed */
3707 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3712 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3714 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3716 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3718 zcr
->zcr_cbinfo
= zio
->io_size
;
3719 zcr
->zcr_cbdata
= abd
;
3720 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3721 zcr
->zcr_free
= zio_abd_free
;
3725 zio_vdev_io_assess(zio_t
*zio
)
3727 vdev_t
*vd
= zio
->io_vd
;
3729 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3730 return (ZIO_PIPELINE_STOP
);
3733 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3734 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3736 if (zio
->io_vsd
!= NULL
) {
3737 zio
->io_vsd_ops
->vsd_free(zio
);
3741 if (zio_injection_enabled
&& zio
->io_error
== 0)
3742 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3745 * If the I/O failed, determine whether we should attempt to retry it.
3747 * On retry, we cut in line in the issue queue, since we don't want
3748 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3750 if (zio
->io_error
&& vd
== NULL
&&
3751 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3752 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3753 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3755 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3756 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3757 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3758 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3759 zio_requeue_io_start_cut_in_line
);
3760 return (ZIO_PIPELINE_STOP
);
3764 * If we got an error on a leaf device, convert it to ENXIO
3765 * if the device is not accessible at all.
3767 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3768 !vdev_accessible(vd
, zio
))
3769 zio
->io_error
= SET_ERROR(ENXIO
);
3772 * If we can't write to an interior vdev (mirror or RAID-Z),
3773 * set vdev_cant_write so that we stop trying to allocate from it.
3775 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3776 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3777 vd
->vdev_cant_write
= B_TRUE
;
3781 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3782 * attempts will ever succeed. In this case we set a persistent bit so
3783 * that we don't bother with it in the future.
3785 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3786 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3787 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3788 vd
->vdev_nowritecache
= B_TRUE
;
3791 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3793 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3794 zio
->io_physdone
!= NULL
) {
3795 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3796 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3797 zio
->io_physdone(zio
->io_logical
);
3800 return (ZIO_PIPELINE_CONTINUE
);
3804 zio_vdev_io_reissue(zio_t
*zio
)
3806 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3807 ASSERT(zio
->io_error
== 0);
3809 zio
->io_stage
>>= 1;
3813 zio_vdev_io_redone(zio_t
*zio
)
3815 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3817 zio
->io_stage
>>= 1;
3821 zio_vdev_io_bypass(zio_t
*zio
)
3823 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3824 ASSERT(zio
->io_error
== 0);
3826 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3827 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3831 * ==========================================================================
3832 * Encrypt and store encryption parameters
3833 * ==========================================================================
3838 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3839 * managing the storage of encryption parameters and passing them to the
3840 * lower-level encryption functions.
3843 zio_encrypt(zio_t
*zio
)
3845 zio_prop_t
*zp
= &zio
->io_prop
;
3846 spa_t
*spa
= zio
->io_spa
;
3847 blkptr_t
*bp
= zio
->io_bp
;
3848 uint64_t psize
= BP_GET_PSIZE(bp
);
3849 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
3850 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3851 void *enc_buf
= NULL
;
3853 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3854 uint8_t iv
[ZIO_DATA_IV_LEN
];
3855 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3856 boolean_t no_crypt
= B_FALSE
;
3858 /* the root zio already encrypted the data */
3859 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3860 return (ZIO_PIPELINE_CONTINUE
);
3862 /* only ZIL blocks are re-encrypted on rewrite */
3863 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3864 return (ZIO_PIPELINE_CONTINUE
);
3866 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3867 BP_SET_CRYPT(bp
, B_FALSE
);
3868 return (ZIO_PIPELINE_CONTINUE
);
3871 /* if we are doing raw encryption set the provided encryption params */
3872 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3873 ASSERT0(BP_GET_LEVEL(bp
));
3874 BP_SET_CRYPT(bp
, B_TRUE
);
3875 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3876 if (ot
!= DMU_OT_OBJSET
)
3877 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3879 /* dnode blocks must be written out in the provided byteorder */
3880 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
3881 ot
== DMU_OT_DNODE
) {
3882 void *bswap_buf
= zio_buf_alloc(psize
);
3883 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
3885 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3886 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
3887 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
3890 abd_take_ownership_of_buf(babd
, B_TRUE
);
3891 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
3894 if (DMU_OT_IS_ENCRYPTED(ot
))
3895 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3896 return (ZIO_PIPELINE_CONTINUE
);
3899 /* indirect blocks only maintain a cksum of the lower level MACs */
3900 if (BP_GET_LEVEL(bp
) > 0) {
3901 BP_SET_CRYPT(bp
, B_TRUE
);
3902 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3903 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3905 zio_crypt_encode_mac_bp(bp
, mac
);
3906 return (ZIO_PIPELINE_CONTINUE
);
3910 * Objset blocks are a special case since they have 2 256-bit MACs
3911 * embedded within them.
3913 if (ot
== DMU_OT_OBJSET
) {
3914 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
3915 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3916 BP_SET_CRYPT(bp
, B_TRUE
);
3917 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
3918 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
3919 return (ZIO_PIPELINE_CONTINUE
);
3922 /* unencrypted object types are only authenticated with a MAC */
3923 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
3924 BP_SET_CRYPT(bp
, B_TRUE
);
3925 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
3926 zio
->io_abd
, psize
, mac
));
3927 zio_crypt_encode_mac_bp(bp
, mac
);
3928 return (ZIO_PIPELINE_CONTINUE
);
3932 * Later passes of sync-to-convergence may decide to rewrite data
3933 * in place to avoid more disk reallocations. This presents a problem
3934 * for encryption because this consitutes rewriting the new data with
3935 * the same encryption key and IV. However, this only applies to blocks
3936 * in the MOS (particularly the spacemaps) and we do not encrypt the
3937 * MOS. We assert that the zio is allocating or an intent log write
3940 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
3941 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
3942 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
3943 ASSERT3U(psize
, !=, 0);
3945 enc_buf
= zio_buf_alloc(psize
);
3946 eabd
= abd_get_from_buf(enc_buf
, psize
);
3947 abd_take_ownership_of_buf(eabd
, B_TRUE
);
3950 * For an explanation of what encryption parameters are stored
3951 * where, see the block comment in zio_crypt.c.
3953 if (ot
== DMU_OT_INTENT_LOG
) {
3954 zio_crypt_decode_params_bp(bp
, salt
, iv
);
3956 BP_SET_CRYPT(bp
, B_TRUE
);
3959 /* Perform the encryption. This should not fail */
3960 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
3961 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
3962 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
3964 /* encode encryption metadata into the bp */
3965 if (ot
== DMU_OT_INTENT_LOG
) {
3967 * ZIL blocks store the MAC in the embedded checksum, so the
3968 * transform must always be applied.
3970 zio_crypt_encode_mac_zil(enc_buf
, mac
);
3971 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3973 BP_SET_CRYPT(bp
, B_TRUE
);
3974 zio_crypt_encode_params_bp(bp
, salt
, iv
);
3975 zio_crypt_encode_mac_bp(bp
, mac
);
3978 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
3981 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3985 return (ZIO_PIPELINE_CONTINUE
);
3989 * ==========================================================================
3990 * Generate and verify checksums
3991 * ==========================================================================
3994 zio_checksum_generate(zio_t
*zio
)
3996 blkptr_t
*bp
= zio
->io_bp
;
3997 enum zio_checksum checksum
;
4001 * This is zio_write_phys().
4002 * We're either generating a label checksum, or none at all.
4004 checksum
= zio
->io_prop
.zp_checksum
;
4006 if (checksum
== ZIO_CHECKSUM_OFF
)
4007 return (ZIO_PIPELINE_CONTINUE
);
4009 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4011 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4012 ASSERT(!IO_IS_ALLOCATING(zio
));
4013 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4015 checksum
= BP_GET_CHECKSUM(bp
);
4019 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4021 return (ZIO_PIPELINE_CONTINUE
);
4025 zio_checksum_verify(zio_t
*zio
)
4027 zio_bad_cksum_t info
;
4028 blkptr_t
*bp
= zio
->io_bp
;
4031 ASSERT(zio
->io_vd
!= NULL
);
4035 * This is zio_read_phys().
4036 * We're either verifying a label checksum, or nothing at all.
4038 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4039 return (ZIO_PIPELINE_CONTINUE
);
4041 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
4044 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4045 zio
->io_error
= error
;
4046 if (error
== ECKSUM
&&
4047 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4048 zfs_ereport_start_checksum(zio
->io_spa
,
4049 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4050 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
4054 return (ZIO_PIPELINE_CONTINUE
);
4058 * Called by RAID-Z to ensure we don't compute the checksum twice.
4061 zio_checksum_verified(zio_t
*zio
)
4063 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4067 * ==========================================================================
4068 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4069 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4070 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
4071 * indicate errors that are specific to one I/O, and most likely permanent.
4072 * Any other error is presumed to be worse because we weren't expecting it.
4073 * ==========================================================================
4076 zio_worst_error(int e1
, int e2
)
4078 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4081 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4082 if (e1
== zio_error_rank
[r1
])
4085 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4086 if (e2
== zio_error_rank
[r2
])
4089 return (r1
> r2
? e1
: e2
);
4093 * ==========================================================================
4095 * ==========================================================================
4098 zio_ready(zio_t
*zio
)
4100 blkptr_t
*bp
= zio
->io_bp
;
4101 zio_t
*pio
, *pio_next
;
4102 zio_link_t
*zl
= NULL
;
4104 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4106 return (ZIO_PIPELINE_STOP
);
4109 if (zio
->io_ready
) {
4110 ASSERT(IO_IS_ALLOCATING(zio
));
4111 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4112 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4113 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4118 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4119 zio
->io_bp_copy
= *bp
;
4121 if (zio
->io_error
!= 0) {
4122 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4124 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4125 ASSERT(IO_IS_ALLOCATING(zio
));
4126 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4128 * We were unable to allocate anything, unreserve and
4129 * issue the next I/O to allocate.
4131 metaslab_class_throttle_unreserve(
4132 spa_normal_class(zio
->io_spa
),
4133 zio
->io_prop
.zp_copies
, zio
);
4134 zio_allocate_dispatch(zio
->io_spa
);
4138 mutex_enter(&zio
->io_lock
);
4139 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4140 pio
= zio_walk_parents(zio
, &zl
);
4141 mutex_exit(&zio
->io_lock
);
4144 * As we notify zio's parents, new parents could be added.
4145 * New parents go to the head of zio's io_parent_list, however,
4146 * so we will (correctly) not notify them. The remainder of zio's
4147 * io_parent_list, from 'pio_next' onward, cannot change because
4148 * all parents must wait for us to be done before they can be done.
4150 for (; pio
!= NULL
; pio
= pio_next
) {
4151 pio_next
= zio_walk_parents(zio
, &zl
);
4152 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
4155 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4156 if (BP_IS_GANG(bp
)) {
4157 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4159 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4160 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4164 if (zio_injection_enabled
&&
4165 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4166 zio_handle_ignored_writes(zio
);
4168 return (ZIO_PIPELINE_CONTINUE
);
4172 * Update the allocation throttle accounting.
4175 zio_dva_throttle_done(zio_t
*zio
)
4177 ASSERTV(zio_t
*lio
= zio
->io_logical
);
4178 zio_t
*pio
= zio_unique_parent(zio
);
4179 vdev_t
*vd
= zio
->io_vd
;
4180 int flags
= METASLAB_ASYNC_ALLOC
;
4182 ASSERT3P(zio
->io_bp
, !=, NULL
);
4183 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4184 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4185 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4187 ASSERT3P(vd
, ==, vd
->vdev_top
);
4188 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4189 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4190 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4191 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4192 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4195 * Parents of gang children can have two flavors -- ones that
4196 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4197 * and ones that allocated the constituent blocks. The allocation
4198 * throttle needs to know the allocating parent zio so we must find
4201 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4203 * If our parent is a rewrite gang child then our grandparent
4204 * would have been the one that performed the allocation.
4206 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4207 pio
= zio_unique_parent(pio
);
4208 flags
|= METASLAB_GANG_CHILD
;
4211 ASSERT(IO_IS_ALLOCATING(pio
));
4212 ASSERT3P(zio
, !=, zio
->io_logical
);
4213 ASSERT(zio
->io_logical
!= NULL
);
4214 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4215 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4217 mutex_enter(&pio
->io_lock
);
4218 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
4219 mutex_exit(&pio
->io_lock
);
4221 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
4225 * Call into the pipeline to see if there is more work that
4226 * needs to be done. If there is work to be done it will be
4227 * dispatched to another taskq thread.
4229 zio_allocate_dispatch(zio
->io_spa
);
4233 zio_done(zio_t
*zio
)
4236 * Always attempt to keep stack usage minimal here since
4237 * we can be called recurisvely up to 19 levels deep.
4239 const uint64_t psize
= zio
->io_size
;
4240 zio_t
*pio
, *pio_next
;
4241 zio_link_t
*zl
= NULL
;
4244 * If our children haven't all completed,
4245 * wait for them and then repeat this pipeline stage.
4247 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4248 return (ZIO_PIPELINE_STOP
);
4252 * If the allocation throttle is enabled, then update the accounting.
4253 * We only track child I/Os that are part of an allocating async
4254 * write. We must do this since the allocation is performed
4255 * by the logical I/O but the actual write is done by child I/Os.
4257 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4258 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4259 ASSERT(spa_normal_class(
4260 zio
->io_spa
)->mc_alloc_throttle_enabled
);
4261 zio_dva_throttle_done(zio
);
4265 * If the allocation throttle is enabled, verify that
4266 * we have decremented the refcounts for every I/O that was throttled.
4268 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4269 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4270 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4271 ASSERT(zio
->io_bp
!= NULL
);
4272 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
);
4273 VERIFY(refcount_not_held(
4274 &(spa_normal_class(zio
->io_spa
)->mc_alloc_slots
), zio
));
4278 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4279 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4280 ASSERT(zio
->io_children
[c
][w
] == 0);
4282 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4283 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4284 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4285 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4286 sizeof (blkptr_t
)) == 0 ||
4287 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4288 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4289 zio
->io_bp_override
== NULL
&&
4290 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4291 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4292 BP_GET_NDVAS(zio
->io_bp
));
4293 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4294 (BP_COUNT_GANG(zio
->io_bp
) ==
4295 BP_GET_NDVAS(zio
->io_bp
)));
4297 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4298 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4302 * If there were child vdev/gang/ddt errors, they apply to us now.
4304 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4305 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4306 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4309 * If the I/O on the transformed data was successful, generate any
4310 * checksum reports now while we still have the transformed data.
4312 if (zio
->io_error
== 0) {
4313 while (zio
->io_cksum_report
!= NULL
) {
4314 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4315 uint64_t align
= zcr
->zcr_align
;
4316 uint64_t asize
= P2ROUNDUP(psize
, align
);
4317 abd_t
*adata
= zio
->io_abd
;
4319 if (asize
!= psize
) {
4320 adata
= abd_alloc(asize
, B_TRUE
);
4321 abd_copy(adata
, zio
->io_abd
, psize
);
4322 abd_zero_off(adata
, psize
, asize
- psize
);
4325 zio
->io_cksum_report
= zcr
->zcr_next
;
4326 zcr
->zcr_next
= NULL
;
4327 zcr
->zcr_finish(zcr
, adata
);
4328 zfs_ereport_free_checksum(zcr
);
4335 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4337 vdev_stat_update(zio
, psize
);
4340 * If this I/O is attached to a particular vdev is slow, exceeding
4341 * 30 seconds to complete, post an error described the I/O delay.
4342 * We ignore these errors if the device is currently unavailable.
4344 if (zio
->io_delay
>= MSEC2NSEC(zio_delay_max
)) {
4345 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
4346 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
4347 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4350 if (zio
->io_error
) {
4352 * If this I/O is attached to a particular vdev,
4353 * generate an error message describing the I/O failure
4354 * at the block level. We ignore these errors if the
4355 * device is currently unavailable.
4357 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4358 !vdev_is_dead(zio
->io_vd
))
4359 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4360 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4362 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4363 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4364 zio
== zio
->io_logical
) {
4366 * For logical I/O requests, tell the SPA to log the
4367 * error and generate a logical data ereport.
4369 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4370 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4371 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4375 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4377 * Determine whether zio should be reexecuted. This will
4378 * propagate all the way to the root via zio_notify_parent().
4380 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4381 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4383 if (IO_IS_ALLOCATING(zio
) &&
4384 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4385 if (zio
->io_error
!= ENOSPC
)
4386 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4388 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4391 if ((zio
->io_type
== ZIO_TYPE_READ
||
4392 zio
->io_type
== ZIO_TYPE_FREE
) &&
4393 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4394 zio
->io_error
== ENXIO
&&
4395 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4396 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4397 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4399 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4400 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4403 * Here is a possibly good place to attempt to do
4404 * either combinatorial reconstruction or error correction
4405 * based on checksums. It also might be a good place
4406 * to send out preliminary ereports before we suspend
4412 * If there were logical child errors, they apply to us now.
4413 * We defer this until now to avoid conflating logical child
4414 * errors with errors that happened to the zio itself when
4415 * updating vdev stats and reporting FMA events above.
4417 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4419 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4420 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4421 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4422 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4424 zio_gang_tree_free(&zio
->io_gang_tree
);
4427 * Godfather I/Os should never suspend.
4429 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4430 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4431 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4433 if (zio
->io_reexecute
) {
4435 * This is a logical I/O that wants to reexecute.
4437 * Reexecute is top-down. When an i/o fails, if it's not
4438 * the root, it simply notifies its parent and sticks around.
4439 * The parent, seeing that it still has children in zio_done(),
4440 * does the same. This percolates all the way up to the root.
4441 * The root i/o will reexecute or suspend the entire tree.
4443 * This approach ensures that zio_reexecute() honors
4444 * all the original i/o dependency relationships, e.g.
4445 * parents not executing until children are ready.
4447 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4449 zio
->io_gang_leader
= NULL
;
4451 mutex_enter(&zio
->io_lock
);
4452 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4453 mutex_exit(&zio
->io_lock
);
4456 * "The Godfather" I/O monitors its children but is
4457 * not a true parent to them. It will track them through
4458 * the pipeline but severs its ties whenever they get into
4459 * trouble (e.g. suspended). This allows "The Godfather"
4460 * I/O to return status without blocking.
4463 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4465 zio_link_t
*remove_zl
= zl
;
4466 pio_next
= zio_walk_parents(zio
, &zl
);
4468 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4469 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4470 zio_remove_child(pio
, zio
, remove_zl
);
4471 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4475 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4477 * We're not a root i/o, so there's nothing to do
4478 * but notify our parent. Don't propagate errors
4479 * upward since we haven't permanently failed yet.
4481 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4482 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4483 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4484 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4486 * We'd fail again if we reexecuted now, so suspend
4487 * until conditions improve (e.g. device comes online).
4489 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4492 * Reexecution is potentially a huge amount of work.
4493 * Hand it off to the otherwise-unused claim taskq.
4495 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4496 spa_taskq_dispatch_ent(zio
->io_spa
,
4497 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4498 (task_func_t
*)zio_reexecute
, zio
, 0,
4501 return (ZIO_PIPELINE_STOP
);
4504 ASSERT(zio
->io_child_count
== 0);
4505 ASSERT(zio
->io_reexecute
== 0);
4506 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4509 * Report any checksum errors, since the I/O is complete.
4511 while (zio
->io_cksum_report
!= NULL
) {
4512 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4513 zio
->io_cksum_report
= zcr
->zcr_next
;
4514 zcr
->zcr_next
= NULL
;
4515 zcr
->zcr_finish(zcr
, NULL
);
4516 zfs_ereport_free_checksum(zcr
);
4519 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4520 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4521 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4522 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4526 * It is the responsibility of the done callback to ensure that this
4527 * particular zio is no longer discoverable for adoption, and as
4528 * such, cannot acquire any new parents.
4533 mutex_enter(&zio
->io_lock
);
4534 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4535 mutex_exit(&zio
->io_lock
);
4538 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4539 zio_link_t
*remove_zl
= zl
;
4540 pio_next
= zio_walk_parents(zio
, &zl
);
4541 zio_remove_child(pio
, zio
, remove_zl
);
4542 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4545 if (zio
->io_waiter
!= NULL
) {
4546 mutex_enter(&zio
->io_lock
);
4547 zio
->io_executor
= NULL
;
4548 cv_broadcast(&zio
->io_cv
);
4549 mutex_exit(&zio
->io_lock
);
4554 return (ZIO_PIPELINE_STOP
);
4558 * ==========================================================================
4559 * I/O pipeline definition
4560 * ==========================================================================
4562 static zio_pipe_stage_t
*zio_pipeline
[] = {
4570 zio_checksum_generate
,
4586 zio_checksum_verify
,
4594 * Compare two zbookmark_phys_t's to see which we would reach first in a
4595 * pre-order traversal of the object tree.
4597 * This is simple in every case aside from the meta-dnode object. For all other
4598 * objects, we traverse them in order (object 1 before object 2, and so on).
4599 * However, all of these objects are traversed while traversing object 0, since
4600 * the data it points to is the list of objects. Thus, we need to convert to a
4601 * canonical representation so we can compare meta-dnode bookmarks to
4602 * non-meta-dnode bookmarks.
4604 * We do this by calculating "equivalents" for each field of the zbookmark.
4605 * zbookmarks outside of the meta-dnode use their own object and level, and
4606 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4607 * blocks this bookmark refers to) by multiplying their blkid by their span
4608 * (the number of L0 blocks contained within one block at their level).
4609 * zbookmarks inside the meta-dnode calculate their object equivalent
4610 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4611 * level + 1<<31 (any value larger than a level could ever be) for their level.
4612 * This causes them to always compare before a bookmark in their object
4613 * equivalent, compare appropriately to bookmarks in other objects, and to
4614 * compare appropriately to other bookmarks in the meta-dnode.
4617 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4618 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4621 * These variables represent the "equivalent" values for the zbookmark,
4622 * after converting zbookmarks inside the meta dnode to their
4623 * normal-object equivalents.
4625 uint64_t zb1obj
, zb2obj
;
4626 uint64_t zb1L0
, zb2L0
;
4627 uint64_t zb1level
, zb2level
;
4629 if (zb1
->zb_object
== zb2
->zb_object
&&
4630 zb1
->zb_level
== zb2
->zb_level
&&
4631 zb1
->zb_blkid
== zb2
->zb_blkid
)
4635 * BP_SPANB calculates the span in blocks.
4637 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4638 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4640 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4641 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4643 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4645 zb1obj
= zb1
->zb_object
;
4646 zb1level
= zb1
->zb_level
;
4649 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4650 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4652 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4654 zb2obj
= zb2
->zb_object
;
4655 zb2level
= zb2
->zb_level
;
4658 /* Now that we have a canonical representation, do the comparison. */
4659 if (zb1obj
!= zb2obj
)
4660 return (zb1obj
< zb2obj
? -1 : 1);
4661 else if (zb1L0
!= zb2L0
)
4662 return (zb1L0
< zb2L0
? -1 : 1);
4663 else if (zb1level
!= zb2level
)
4664 return (zb1level
> zb2level
? -1 : 1);
4666 * This can (theoretically) happen if the bookmarks have the same object
4667 * and level, but different blkids, if the block sizes are not the same.
4668 * There is presently no way to change the indirect block sizes
4674 * This function checks the following: given that last_block is the place that
4675 * our traversal stopped last time, does that guarantee that we've visited
4676 * every node under subtree_root? Therefore, we can't just use the raw output
4677 * of zbookmark_compare. We have to pass in a modified version of
4678 * subtree_root; by incrementing the block id, and then checking whether
4679 * last_block is before or equal to that, we can tell whether or not having
4680 * visited last_block implies that all of subtree_root's children have been
4684 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4685 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4687 zbookmark_phys_t mod_zb
= *subtree_root
;
4689 ASSERT(last_block
->zb_level
== 0);
4691 /* The objset_phys_t isn't before anything. */
4696 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4697 * data block size in sectors, because that variable is only used if
4698 * the bookmark refers to a block in the meta-dnode. Since we don't
4699 * know without examining it what object it refers to, and there's no
4700 * harm in passing in this value in other cases, we always pass it in.
4702 * We pass in 0 for the indirect block size shift because zb2 must be
4703 * level 0. The indirect block size is only used to calculate the span
4704 * of the bookmark, but since the bookmark must be level 0, the span is
4705 * always 1, so the math works out.
4707 * If you make changes to how the zbookmark_compare code works, be sure
4708 * to make sure that this code still works afterwards.
4710 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4711 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4715 #if defined(_KERNEL)
4716 EXPORT_SYMBOL(zio_type_name
);
4717 EXPORT_SYMBOL(zio_buf_alloc
);
4718 EXPORT_SYMBOL(zio_data_buf_alloc
);
4719 EXPORT_SYMBOL(zio_buf_free
);
4720 EXPORT_SYMBOL(zio_data_buf_free
);
4722 module_param(zio_delay_max
, int, 0644);
4723 MODULE_PARM_DESC(zio_delay_max
, "Max zio millisec delay before posting event");
4725 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
4726 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");
4728 module_param(zfs_sync_pass_deferred_free
, int, 0644);
4729 MODULE_PARM_DESC(zfs_sync_pass_deferred_free
,
4730 "Defer frees starting in this pass");
4732 module_param(zfs_sync_pass_dont_compress
, int, 0644);
4733 MODULE_PARM_DESC(zfs_sync_pass_dont_compress
,
4734 "Don't compress starting in this pass");
4736 module_param(zfs_sync_pass_rewrite
, int, 0644);
4737 MODULE_PARM_DESC(zfs_sync_pass_rewrite
,
4738 "Rewrite new bps starting in this pass");
4740 module_param(zio_dva_throttle_enabled
, int, 0644);
4741 MODULE_PARM_DESC(zio_dva_throttle_enabled
,
4742 "Throttle block allocations in the ZIO pipeline");