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, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_trim.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/dsl_scan.h>
45 #include <sys/metaslab_impl.h>
47 #include <sys/trace_zfs.h>
49 #include <sys/dsl_crypt.h>
50 #include <sys/cityhash.h>
53 * ==========================================================================
54 * I/O type descriptions
55 * ==========================================================================
57 const char *zio_type_name
[ZIO_TYPES
] = {
59 * Note: Linux kernel thread name length is limited
60 * so these names will differ from upstream open zfs.
62 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
65 int zio_dva_throttle_enabled
= B_TRUE
;
66 int zio_deadman_log_all
= B_FALSE
;
69 * ==========================================================================
71 * ==========================================================================
73 kmem_cache_t
*zio_cache
;
74 kmem_cache_t
*zio_link_cache
;
75 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
76 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
77 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
78 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
79 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
82 /* Mark IOs as "slow" if they take longer than 30 seconds */
83 int zio_slow_io_ms
= (30 * MILLISEC
);
85 #define BP_SPANB(indblkshift, level) \
86 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
87 #define COMPARE_META_LEVEL 0x80000000ul
89 * The following actions directly effect the spa's sync-to-convergence logic.
90 * The values below define the sync pass when we start performing the action.
91 * Care should be taken when changing these values as they directly impact
92 * spa_sync() performance. Tuning these values may introduce subtle performance
93 * pathologies and should only be done in the context of performance analysis.
94 * These tunables will eventually be removed and replaced with #defines once
95 * enough analysis has been done to determine optimal values.
97 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
98 * regular blocks are not deferred.
100 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
101 * compression (including of metadata). In practice, we don't have this
102 * many sync passes, so this has no effect.
104 * The original intent was that disabling compression would help the sync
105 * passes to converge. However, in practice disabling compression increases
106 * the average number of sync passes, because when we turn compression off, a
107 * lot of block's size will change and thus we have to re-allocate (not
108 * overwrite) them. It also increases the number of 128KB allocations (e.g.
109 * for indirect blocks and spacemaps) because these will not be compressed.
110 * The 128K allocations are especially detrimental to performance on highly
111 * fragmented systems, which may have very few free segments of this size,
112 * and may need to load new metaslabs to satisfy 128K allocations.
114 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
115 int zfs_sync_pass_dont_compress
= 8; /* don't compress starting in this pass */
116 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
119 * An allocating zio is one that either currently has the DVA allocate
120 * stage set or will have it later in its lifetime.
122 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
125 * Enable smaller cores by excluding metadata
126 * allocations as well.
128 int zio_exclude_metadata
= 0;
129 int zio_requeue_io_start_cut_in_line
= 1;
132 int zio_buf_debug_limit
= 16384;
134 int zio_buf_debug_limit
= 0;
137 static inline void __zio_execute(zio_t
*zio
);
139 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
145 vmem_t
*data_alloc_arena
= NULL
;
147 zio_cache
= kmem_cache_create("zio_cache",
148 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
149 zio_link_cache
= kmem_cache_create("zio_link_cache",
150 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
153 * For small buffers, we want a cache for each multiple of
154 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
155 * for each quarter-power of 2.
157 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
158 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
161 size_t data_cflags
, cflags
;
163 data_cflags
= KMC_NODEBUG
;
164 cflags
= (zio_exclude_metadata
|| size
> zio_buf_debug_limit
) ?
167 #if defined(_ILP32) && defined(_KERNEL)
169 * Cache size limited to 1M on 32-bit platforms until ARC
170 * buffers no longer require virtual address space.
172 if (size
> zfs_max_recordsize
)
181 * If we are using watchpoints, put each buffer on its own page,
182 * to eliminate the performance overhead of trapping to the
183 * kernel when modifying a non-watched buffer that shares the
184 * page with a watched buffer.
186 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
189 * Here's the problem - on 4K native devices in userland on
190 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
191 * will fail with EINVAL, causing zdb (and others) to coredump.
192 * Since userland probably doesn't need optimized buffer caches,
193 * we just force 4K alignment on everything.
195 align
= 8 * SPA_MINBLOCKSIZE
;
197 if (size
< PAGESIZE
) {
198 align
= SPA_MINBLOCKSIZE
;
199 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
206 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
207 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
208 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
210 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
211 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
212 align
, NULL
, NULL
, NULL
, NULL
,
213 data_alloc_arena
, data_cflags
);
218 ASSERT(zio_buf_cache
[c
] != NULL
);
219 if (zio_buf_cache
[c
- 1] == NULL
)
220 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
222 ASSERT(zio_data_buf_cache
[c
] != NULL
);
223 if (zio_data_buf_cache
[c
- 1] == NULL
)
224 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
236 kmem_cache_t
*last_cache
= NULL
;
237 kmem_cache_t
*last_data_cache
= NULL
;
239 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
242 * Cache size limited to 1M on 32-bit platforms until ARC
243 * buffers no longer require virtual address space.
245 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
248 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
249 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
250 (void) printf("zio_fini: [%d] %llu != %llu\n",
251 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
252 (long long unsigned)zio_buf_cache_allocs
[c
],
253 (long long unsigned)zio_buf_cache_frees
[c
]);
255 if (zio_buf_cache
[c
] != last_cache
) {
256 last_cache
= zio_buf_cache
[c
];
257 kmem_cache_destroy(zio_buf_cache
[c
]);
259 zio_buf_cache
[c
] = NULL
;
261 if (zio_data_buf_cache
[c
] != last_data_cache
) {
262 last_data_cache
= zio_data_buf_cache
[c
];
263 kmem_cache_destroy(zio_data_buf_cache
[c
]);
265 zio_data_buf_cache
[c
] = NULL
;
268 kmem_cache_destroy(zio_link_cache
);
269 kmem_cache_destroy(zio_cache
);
277 * ==========================================================================
278 * Allocate and free I/O buffers
279 * ==========================================================================
283 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
284 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
285 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
286 * excess / transient data in-core during a crashdump.
289 zio_buf_alloc(size_t size
)
291 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
293 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
294 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
295 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
298 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
302 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
303 * crashdump if the kernel panics. This exists so that we will limit the amount
304 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
305 * of kernel heap dumped to disk when the kernel panics)
308 zio_data_buf_alloc(size_t size
)
310 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
312 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
314 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
318 zio_buf_free(void *buf
, size_t size
)
320 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
322 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
323 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
324 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
327 kmem_cache_free(zio_buf_cache
[c
], buf
);
331 zio_data_buf_free(void *buf
, size_t size
)
333 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
335 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
337 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
341 zio_abd_free(void *abd
, size_t size
)
343 abd_free((abd_t
*)abd
);
347 * ==========================================================================
348 * Push and pop I/O transform buffers
349 * ==========================================================================
352 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
353 zio_transform_func_t
*transform
)
355 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
357 zt
->zt_orig_abd
= zio
->io_abd
;
358 zt
->zt_orig_size
= zio
->io_size
;
359 zt
->zt_bufsize
= bufsize
;
360 zt
->zt_transform
= transform
;
362 zt
->zt_next
= zio
->io_transform_stack
;
363 zio
->io_transform_stack
= zt
;
370 zio_pop_transforms(zio_t
*zio
)
374 while ((zt
= zio
->io_transform_stack
) != NULL
) {
375 if (zt
->zt_transform
!= NULL
)
376 zt
->zt_transform(zio
,
377 zt
->zt_orig_abd
, zt
->zt_orig_size
);
379 if (zt
->zt_bufsize
!= 0)
380 abd_free(zio
->io_abd
);
382 zio
->io_abd
= zt
->zt_orig_abd
;
383 zio
->io_size
= zt
->zt_orig_size
;
384 zio
->io_transform_stack
= zt
->zt_next
;
386 kmem_free(zt
, sizeof (zio_transform_t
));
391 * ==========================================================================
392 * I/O transform callbacks for subblocks, decompression, and decryption
393 * ==========================================================================
396 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
398 ASSERT(zio
->io_size
> size
);
400 if (zio
->io_type
== ZIO_TYPE_READ
)
401 abd_copy(data
, zio
->io_abd
, size
);
405 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
407 if (zio
->io_error
== 0) {
408 void *tmp
= abd_borrow_buf(data
, size
);
409 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
410 zio
->io_abd
, tmp
, zio
->io_size
, size
);
411 abd_return_buf_copy(data
, tmp
, size
);
413 if (zio_injection_enabled
&& ret
== 0)
414 ret
= zio_handle_fault_injection(zio
, EINVAL
);
417 zio
->io_error
= SET_ERROR(EIO
);
422 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
426 blkptr_t
*bp
= zio
->io_bp
;
427 spa_t
*spa
= zio
->io_spa
;
428 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
429 uint64_t lsize
= BP_GET_LSIZE(bp
);
430 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
431 uint8_t salt
[ZIO_DATA_SALT_LEN
];
432 uint8_t iv
[ZIO_DATA_IV_LEN
];
433 uint8_t mac
[ZIO_DATA_MAC_LEN
];
434 boolean_t no_crypt
= B_FALSE
;
436 ASSERT(BP_USES_CRYPT(bp
));
437 ASSERT3U(size
, !=, 0);
439 if (zio
->io_error
!= 0)
443 * Verify the cksum of MACs stored in an indirect bp. It will always
444 * be possible to verify this since it does not require an encryption
447 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
448 zio_crypt_decode_mac_bp(bp
, mac
);
450 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
452 * We haven't decompressed the data yet, but
453 * zio_crypt_do_indirect_mac_checksum() requires
454 * decompressed data to be able to parse out the MACs
455 * from the indirect block. We decompress it now and
456 * throw away the result after we are finished.
458 tmp
= zio_buf_alloc(lsize
);
459 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
460 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
462 ret
= SET_ERROR(EIO
);
465 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
466 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
467 zio_buf_free(tmp
, lsize
);
469 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
470 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
472 abd_copy(data
, zio
->io_abd
, size
);
474 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
475 ret
= zio_handle_decrypt_injection(spa
,
476 &zio
->io_bookmark
, ot
, ECKSUM
);
485 * If this is an authenticated block, just check the MAC. It would be
486 * nice to separate this out into its own flag, but for the moment
487 * enum zio_flag is out of bits.
489 if (BP_IS_AUTHENTICATED(bp
)) {
490 if (ot
== DMU_OT_OBJSET
) {
491 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
492 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
494 zio_crypt_decode_mac_bp(bp
, mac
);
495 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
496 zio
->io_abd
, size
, mac
);
497 if (zio_injection_enabled
&& ret
== 0) {
498 ret
= zio_handle_decrypt_injection(spa
,
499 &zio
->io_bookmark
, ot
, ECKSUM
);
502 abd_copy(data
, zio
->io_abd
, size
);
510 zio_crypt_decode_params_bp(bp
, salt
, iv
);
512 if (ot
== DMU_OT_INTENT_LOG
) {
513 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
514 zio_crypt_decode_mac_zil(tmp
, mac
);
515 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
517 zio_crypt_decode_mac_bp(bp
, mac
);
520 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
521 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
522 zio
->io_abd
, &no_crypt
);
524 abd_copy(data
, zio
->io_abd
, size
);
532 /* assert that the key was found unless this was speculative */
533 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
536 * If there was a decryption / authentication error return EIO as
537 * the io_error. If this was not a speculative zio, create an ereport.
540 zio
->io_error
= SET_ERROR(EIO
);
541 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
542 spa_log_error(spa
, &zio
->io_bookmark
);
543 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
544 spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
552 * ==========================================================================
553 * I/O parent/child relationships and pipeline interlocks
554 * ==========================================================================
557 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
559 list_t
*pl
= &cio
->io_parent_list
;
561 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
565 ASSERT((*zl
)->zl_child
== cio
);
566 return ((*zl
)->zl_parent
);
570 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
572 list_t
*cl
= &pio
->io_child_list
;
574 ASSERT(MUTEX_HELD(&pio
->io_lock
));
576 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
580 ASSERT((*zl
)->zl_parent
== pio
);
581 return ((*zl
)->zl_child
);
585 zio_unique_parent(zio_t
*cio
)
587 zio_link_t
*zl
= NULL
;
588 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
590 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
595 zio_add_child(zio_t
*pio
, zio_t
*cio
)
597 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
600 * Logical I/Os can have logical, gang, or vdev children.
601 * Gang I/Os can have gang or vdev children.
602 * Vdev I/Os can only have vdev children.
603 * The following ASSERT captures all of these constraints.
605 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
610 mutex_enter(&pio
->io_lock
);
611 mutex_enter(&cio
->io_lock
);
613 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
615 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
616 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
618 list_insert_head(&pio
->io_child_list
, zl
);
619 list_insert_head(&cio
->io_parent_list
, zl
);
621 pio
->io_child_count
++;
622 cio
->io_parent_count
++;
624 mutex_exit(&cio
->io_lock
);
625 mutex_exit(&pio
->io_lock
);
629 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
631 ASSERT(zl
->zl_parent
== pio
);
632 ASSERT(zl
->zl_child
== cio
);
634 mutex_enter(&pio
->io_lock
);
635 mutex_enter(&cio
->io_lock
);
637 list_remove(&pio
->io_child_list
, zl
);
638 list_remove(&cio
->io_parent_list
, zl
);
640 pio
->io_child_count
--;
641 cio
->io_parent_count
--;
643 mutex_exit(&cio
->io_lock
);
644 mutex_exit(&pio
->io_lock
);
645 kmem_cache_free(zio_link_cache
, zl
);
649 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
651 boolean_t waiting
= B_FALSE
;
653 mutex_enter(&zio
->io_lock
);
654 ASSERT(zio
->io_stall
== NULL
);
655 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
656 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
659 uint64_t *countp
= &zio
->io_children
[c
][wait
];
662 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
663 zio
->io_stall
= countp
;
668 mutex_exit(&zio
->io_lock
);
672 __attribute__((always_inline
))
674 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
675 zio_t
**next_to_executep
)
677 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
678 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
680 mutex_enter(&pio
->io_lock
);
681 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
682 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
683 pio
->io_reexecute
|= zio
->io_reexecute
;
684 ASSERT3U(*countp
, >, 0);
688 if (*countp
== 0 && pio
->io_stall
== countp
) {
689 zio_taskq_type_t type
=
690 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
692 pio
->io_stall
= NULL
;
693 mutex_exit(&pio
->io_lock
);
696 * If we can tell the caller to execute this parent next, do
697 * so. Otherwise dispatch the parent zio as its own task.
699 * Having the caller execute the parent when possible reduces
700 * locking on the zio taskq's, reduces context switch
701 * overhead, and has no recursion penalty. Note that one
702 * read from disk typically causes at least 3 zio's: a
703 * zio_null(), the logical zio_read(), and then a physical
704 * zio. When the physical ZIO completes, we are able to call
705 * zio_done() on all 3 of these zio's from one invocation of
706 * zio_execute() by returning the parent back to
707 * zio_execute(). Since the parent isn't executed until this
708 * thread returns back to zio_execute(), the caller should do
711 * In other cases, dispatching the parent prevents
712 * overflowing the stack when we have deeply nested
713 * parent-child relationships, as we do with the "mega zio"
714 * of writes for spa_sync(), and the chain of ZIL blocks.
716 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
) {
717 *next_to_executep
= pio
;
719 zio_taskq_dispatch(pio
, type
, B_FALSE
);
722 mutex_exit(&pio
->io_lock
);
727 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
729 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
730 zio
->io_error
= zio
->io_child_error
[c
];
734 zio_bookmark_compare(const void *x1
, const void *x2
)
736 const zio_t
*z1
= x1
;
737 const zio_t
*z2
= x2
;
739 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
741 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
744 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
746 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
749 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
751 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
754 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
756 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
768 * ==========================================================================
769 * Create the various types of I/O (read, write, free, etc)
770 * ==========================================================================
773 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
774 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
775 void *private, zio_type_t type
, zio_priority_t priority
,
776 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
777 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
778 enum zio_stage pipeline
)
782 IMPLY(type
!= ZIO_TYPE_TRIM
, psize
<= SPA_MAXBLOCKSIZE
);
783 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
784 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
786 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
787 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
788 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
790 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
792 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
793 bzero(zio
, sizeof (zio_t
));
795 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
796 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
798 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
799 offsetof(zio_link_t
, zl_parent_node
));
800 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
801 offsetof(zio_link_t
, zl_child_node
));
802 metaslab_trace_init(&zio
->io_alloc_list
);
805 zio
->io_child_type
= ZIO_CHILD_VDEV
;
806 else if (flags
& ZIO_FLAG_GANG_CHILD
)
807 zio
->io_child_type
= ZIO_CHILD_GANG
;
808 else if (flags
& ZIO_FLAG_DDT_CHILD
)
809 zio
->io_child_type
= ZIO_CHILD_DDT
;
811 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
814 zio
->io_bp
= (blkptr_t
*)bp
;
815 zio
->io_bp_copy
= *bp
;
816 zio
->io_bp_orig
= *bp
;
817 if (type
!= ZIO_TYPE_WRITE
||
818 zio
->io_child_type
== ZIO_CHILD_DDT
)
819 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
820 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
821 zio
->io_logical
= zio
;
822 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
823 pipeline
|= ZIO_GANG_STAGES
;
829 zio
->io_private
= private;
831 zio
->io_priority
= priority
;
833 zio
->io_offset
= offset
;
834 zio
->io_orig_abd
= zio
->io_abd
= data
;
835 zio
->io_orig_size
= zio
->io_size
= psize
;
836 zio
->io_lsize
= lsize
;
837 zio
->io_orig_flags
= zio
->io_flags
= flags
;
838 zio
->io_orig_stage
= zio
->io_stage
= stage
;
839 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
840 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
842 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
843 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
846 zio
->io_bookmark
= *zb
;
849 if (zio
->io_metaslab_class
== NULL
)
850 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
851 if (zio
->io_logical
== NULL
)
852 zio
->io_logical
= pio
->io_logical
;
853 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
854 zio
->io_gang_leader
= pio
->io_gang_leader
;
855 zio_add_child(pio
, zio
);
858 taskq_init_ent(&zio
->io_tqent
);
864 zio_destroy(zio_t
*zio
)
866 metaslab_trace_fini(&zio
->io_alloc_list
);
867 list_destroy(&zio
->io_parent_list
);
868 list_destroy(&zio
->io_child_list
);
869 mutex_destroy(&zio
->io_lock
);
870 cv_destroy(&zio
->io_cv
);
871 kmem_cache_free(zio_cache
, zio
);
875 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
876 void *private, enum zio_flag flags
)
880 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
881 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
882 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
888 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
890 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
894 zfs_blkptr_verify_log(spa_t
*spa
, const blkptr_t
*bp
,
895 enum blk_verify_flag blk_verify
, const char *fmt
, ...)
901 (void) vsnprintf(buf
, sizeof (buf
), fmt
, adx
);
904 switch (blk_verify
) {
905 case BLK_VERIFY_HALT
:
906 zfs_panic_recover("%s: %s", spa_name(spa
), buf
);
909 zfs_dbgmsg("%s: %s", spa_name(spa
), buf
);
911 case BLK_VERIFY_ONLY
:
919 * Verify the block pointer fields contain reasonable values. This means
920 * it only contains known object types, checksum/compression identifiers,
921 * block sizes within the maximum allowed limits, valid DVAs, etc.
923 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
924 * argument controls the behavior when an invalid field is detected.
926 * Modes for zfs_blkptr_verify:
927 * 1) BLK_VERIFY_ONLY (evaluate the block)
928 * 2) BLK_VERIFY_LOG (evaluate the block and log problems)
929 * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
932 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
, boolean_t config_held
,
933 enum blk_verify_flag blk_verify
)
937 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
938 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
939 "blkptr at %p has invalid TYPE %llu",
940 bp
, (longlong_t
)BP_GET_TYPE(bp
));
942 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
943 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
944 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
945 "blkptr at %p has invalid CHECKSUM %llu",
946 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
948 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
949 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
950 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
951 "blkptr at %p has invalid COMPRESS %llu",
952 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
954 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
955 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
956 "blkptr at %p has invalid LSIZE %llu",
957 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
959 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
960 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
961 "blkptr at %p has invalid PSIZE %llu",
962 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
965 if (BP_IS_EMBEDDED(bp
)) {
966 if (BPE_GET_ETYPE(bp
) >= NUM_BP_EMBEDDED_TYPES
) {
967 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
968 "blkptr at %p has invalid ETYPE %llu",
969 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
974 * Do not verify individual DVAs if the config is not trusted. This
975 * will be done once the zio is executed in vdev_mirror_map_alloc.
977 if (!spa
->spa_trust_config
)
981 spa_config_enter(spa
, SCL_VDEV
, bp
, RW_READER
);
983 ASSERT(spa_config_held(spa
, SCL_VDEV
, RW_WRITER
));
985 * Pool-specific checks.
987 * Note: it would be nice to verify that the blk_birth and
988 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
989 * allows the birth time of log blocks (and dmu_sync()-ed blocks
990 * that are in the log) to be arbitrarily large.
992 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
993 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
995 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
996 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
997 "blkptr at %p DVA %u has invalid VDEV %llu",
998 bp
, i
, (longlong_t
)vdevid
);
1001 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1003 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1004 "blkptr at %p DVA %u has invalid VDEV %llu",
1005 bp
, i
, (longlong_t
)vdevid
);
1008 if (vd
->vdev_ops
== &vdev_hole_ops
) {
1009 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1010 "blkptr at %p DVA %u has hole VDEV %llu",
1011 bp
, i
, (longlong_t
)vdevid
);
1014 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1016 * "missing" vdevs are valid during import, but we
1017 * don't have their detailed info (e.g. asize), so
1018 * we can't perform any more checks on them.
1022 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
1023 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
1025 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1026 if (offset
+ asize
> vd
->vdev_asize
) {
1027 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1028 "blkptr at %p DVA %u has invalid OFFSET %llu",
1029 bp
, i
, (longlong_t
)offset
);
1033 spa_config_exit(spa
, SCL_VDEV
, bp
);
1035 return (errors
== 0);
1039 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
1041 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1043 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
1046 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1050 if (vd
->vdev_ops
== &vdev_hole_ops
)
1053 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1057 uint64_t offset
= DVA_GET_OFFSET(dva
);
1058 uint64_t asize
= DVA_GET_ASIZE(dva
);
1061 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1062 if (offset
+ asize
> vd
->vdev_asize
)
1069 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
1070 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
1071 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
1075 (void) zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
,
1078 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
1079 data
, size
, size
, done
, private,
1080 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1081 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1082 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1088 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1089 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1090 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1091 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
1092 void *private, zio_priority_t priority
, enum zio_flag flags
,
1093 const zbookmark_phys_t
*zb
)
1097 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1098 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1099 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1100 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1101 DMU_OT_IS_VALID(zp
->zp_type
) &&
1102 zp
->zp_level
< 32 &&
1103 zp
->zp_copies
> 0 &&
1104 zp
->zp_copies
<= spa_max_replication(spa
));
1106 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1107 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1108 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1109 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1111 zio
->io_ready
= ready
;
1112 zio
->io_children_ready
= children_ready
;
1113 zio
->io_physdone
= physdone
;
1117 * Data can be NULL if we are going to call zio_write_override() to
1118 * provide the already-allocated BP. But we may need the data to
1119 * verify a dedup hit (if requested). In this case, don't try to
1120 * dedup (just take the already-allocated BP verbatim). Encrypted
1121 * dedup blocks need data as well so we also disable dedup in this
1125 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1126 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1133 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1134 uint64_t size
, zio_done_func_t
*done
, void *private,
1135 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
1139 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1140 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1141 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1147 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1149 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1150 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1151 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1152 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1155 * We must reset the io_prop to match the values that existed
1156 * when the bp was first written by dmu_sync() keeping in mind
1157 * that nopwrite and dedup are mutually exclusive.
1159 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1160 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1161 zio
->io_prop
.zp_copies
= copies
;
1162 zio
->io_bp_override
= bp
;
1166 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1169 (void) zfs_blkptr_verify(spa
, bp
, B_FALSE
, BLK_VERIFY_HALT
);
1172 * The check for EMBEDDED is a performance optimization. We
1173 * process the free here (by ignoring it) rather than
1174 * putting it on the list and then processing it in zio_free_sync().
1176 if (BP_IS_EMBEDDED(bp
))
1178 metaslab_check_free(spa
, bp
);
1181 * Frees that are for the currently-syncing txg, are not going to be
1182 * deferred, and which will not need to do a read (i.e. not GANG or
1183 * DEDUP), can be processed immediately. Otherwise, put them on the
1184 * in-memory list for later processing.
1186 * Note that we only defer frees after zfs_sync_pass_deferred_free
1187 * when the log space map feature is disabled. [see relevant comment
1188 * in spa_sync_iterate_to_convergence()]
1190 if (BP_IS_GANG(bp
) ||
1192 txg
!= spa
->spa_syncing_txg
||
1193 (spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
&&
1194 !spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
))) {
1195 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1197 VERIFY3P(zio_free_sync(NULL
, spa
, txg
, bp
, 0), ==, NULL
);
1202 * To improve performance, this function may return NULL if we were able
1203 * to do the free immediately. This avoids the cost of creating a zio
1204 * (and linking it to the parent, etc).
1207 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1208 enum zio_flag flags
)
1210 ASSERT(!BP_IS_HOLE(bp
));
1211 ASSERT(spa_syncing_txg(spa
) == txg
);
1213 if (BP_IS_EMBEDDED(bp
))
1216 metaslab_check_free(spa
, bp
);
1218 dsl_scan_freed(spa
, bp
);
1220 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
)) {
1222 * GANG and DEDUP blocks can induce a read (for the gang block
1223 * header, or the DDT), so issue them asynchronously so that
1224 * this thread is not tied up.
1226 enum zio_stage stage
=
1227 ZIO_FREE_PIPELINE
| ZIO_STAGE_ISSUE_ASYNC
;
1229 return (zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1230 BP_GET_PSIZE(bp
), NULL
, NULL
,
1231 ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1232 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
));
1234 metaslab_free(spa
, bp
, txg
, B_FALSE
);
1240 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1241 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1245 (void) zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
,
1248 if (BP_IS_EMBEDDED(bp
))
1249 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1252 * A claim is an allocation of a specific block. Claims are needed
1253 * to support immediate writes in the intent log. The issue is that
1254 * immediate writes contain committed data, but in a txg that was
1255 * *not* committed. Upon opening the pool after an unclean shutdown,
1256 * the intent log claims all blocks that contain immediate write data
1257 * so that the SPA knows they're in use.
1259 * All claims *must* be resolved in the first txg -- before the SPA
1260 * starts allocating blocks -- so that nothing is allocated twice.
1261 * If txg == 0 we just verify that the block is claimable.
1263 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
1264 spa_min_claim_txg(spa
));
1265 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1266 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1268 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1269 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1270 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1271 ASSERT0(zio
->io_queued_timestamp
);
1277 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1278 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1283 if (vd
->vdev_children
== 0) {
1284 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1285 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1286 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1290 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1292 for (c
= 0; c
< vd
->vdev_children
; c
++)
1293 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1294 done
, private, flags
));
1301 zio_trim(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1302 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1303 enum zio_flag flags
, enum trim_flag trim_flags
)
1307 ASSERT0(vd
->vdev_children
);
1308 ASSERT0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
1309 ASSERT0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
1310 ASSERT3U(size
, !=, 0);
1312 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, NULL
, size
, size
, done
,
1313 private, ZIO_TYPE_TRIM
, priority
, flags
| ZIO_FLAG_PHYSICAL
,
1314 vd
, offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_TRIM_PIPELINE
);
1315 zio
->io_trim_flags
= trim_flags
;
1321 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1322 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1323 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1327 ASSERT(vd
->vdev_children
== 0);
1328 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1329 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1330 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1332 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1333 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1334 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1336 zio
->io_prop
.zp_checksum
= checksum
;
1342 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1343 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1344 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1348 ASSERT(vd
->vdev_children
== 0);
1349 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1350 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1351 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1353 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1354 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1355 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1357 zio
->io_prop
.zp_checksum
= checksum
;
1359 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1361 * zec checksums are necessarily destructive -- they modify
1362 * the end of the write buffer to hold the verifier/checksum.
1363 * Therefore, we must make a local copy in case the data is
1364 * being written to multiple places in parallel.
1366 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1367 abd_copy(wbuf
, data
, size
);
1369 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1376 * Create a child I/O to do some work for us.
1379 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1380 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1381 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1383 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1387 * vdev child I/Os do not propagate their error to the parent.
1388 * Therefore, for correct operation the caller *must* check for
1389 * and handle the error in the child i/o's done callback.
1390 * The only exceptions are i/os that we don't care about
1391 * (OPTIONAL or REPAIR).
1393 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1396 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1398 * If we have the bp, then the child should perform the
1399 * checksum and the parent need not. This pushes error
1400 * detection as close to the leaves as possible and
1401 * eliminates redundant checksums in the interior nodes.
1403 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1404 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1407 if (vd
->vdev_ops
->vdev_op_leaf
) {
1408 ASSERT0(vd
->vdev_children
);
1409 offset
+= VDEV_LABEL_START_SIZE
;
1412 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1415 * If we've decided to do a repair, the write is not speculative --
1416 * even if the original read was.
1418 if (flags
& ZIO_FLAG_IO_REPAIR
)
1419 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1422 * If we're creating a child I/O that is not associated with a
1423 * top-level vdev, then the child zio is not an allocating I/O.
1424 * If this is a retried I/O then we ignore it since we will
1425 * have already processed the original allocating I/O.
1427 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1428 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1429 ASSERT(pio
->io_metaslab_class
!= NULL
);
1430 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1431 ASSERT(type
== ZIO_TYPE_WRITE
);
1432 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1433 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1434 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1435 pio
->io_child_type
== ZIO_CHILD_GANG
);
1437 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1441 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1442 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1443 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1444 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1446 zio
->io_physdone
= pio
->io_physdone
;
1447 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1448 zio
->io_logical
->io_phys_children
++;
1454 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1455 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1456 zio_done_func_t
*done
, void *private)
1460 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1462 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1463 data
, size
, size
, done
, private, type
, priority
,
1464 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1466 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1472 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1474 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1476 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1480 zio_shrink(zio_t
*zio
, uint64_t size
)
1482 ASSERT3P(zio
->io_executor
, ==, NULL
);
1483 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1484 ASSERT3U(size
, <=, zio
->io_size
);
1487 * We don't shrink for raidz because of problems with the
1488 * reconstruction when reading back less than the block size.
1489 * Note, BP_IS_RAIDZ() assumes no compression.
1491 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1492 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1493 /* we are not doing a raw write */
1494 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1495 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1500 * ==========================================================================
1501 * Prepare to read and write logical blocks
1502 * ==========================================================================
1506 zio_read_bp_init(zio_t
*zio
)
1508 blkptr_t
*bp
= zio
->io_bp
;
1510 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1512 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1514 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1515 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1516 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1517 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1518 psize
, psize
, zio_decompress
);
1521 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1522 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1523 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1524 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1525 psize
, psize
, zio_decrypt
);
1528 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1529 int psize
= BPE_GET_PSIZE(bp
);
1530 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1532 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1533 decode_embedded_bp_compressed(bp
, data
);
1534 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1536 ASSERT(!BP_IS_EMBEDDED(bp
));
1537 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1540 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1541 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1543 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1544 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1546 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1547 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1553 zio_write_bp_init(zio_t
*zio
)
1555 if (!IO_IS_ALLOCATING(zio
))
1558 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1560 if (zio
->io_bp_override
) {
1561 blkptr_t
*bp
= zio
->io_bp
;
1562 zio_prop_t
*zp
= &zio
->io_prop
;
1564 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1565 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1567 *bp
= *zio
->io_bp_override
;
1568 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1570 if (BP_IS_EMBEDDED(bp
))
1574 * If we've been overridden and nopwrite is set then
1575 * set the flag accordingly to indicate that a nopwrite
1576 * has already occurred.
1578 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1579 ASSERT(!zp
->zp_dedup
);
1580 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1581 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1585 ASSERT(!zp
->zp_nopwrite
);
1587 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1590 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1591 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1593 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1595 BP_SET_DEDUP(bp
, 1);
1596 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1601 * We were unable to handle this as an override bp, treat
1602 * it as a regular write I/O.
1604 zio
->io_bp_override
= NULL
;
1605 *bp
= zio
->io_bp_orig
;
1606 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1613 zio_write_compress(zio_t
*zio
)
1615 spa_t
*spa
= zio
->io_spa
;
1616 zio_prop_t
*zp
= &zio
->io_prop
;
1617 enum zio_compress compress
= zp
->zp_compress
;
1618 blkptr_t
*bp
= zio
->io_bp
;
1619 uint64_t lsize
= zio
->io_lsize
;
1620 uint64_t psize
= zio
->io_size
;
1624 * If our children haven't all reached the ready stage,
1625 * wait for them and then repeat this pipeline stage.
1627 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1628 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1632 if (!IO_IS_ALLOCATING(zio
))
1635 if (zio
->io_children_ready
!= NULL
) {
1637 * Now that all our children are ready, run the callback
1638 * associated with this zio in case it wants to modify the
1639 * data to be written.
1641 ASSERT3U(zp
->zp_level
, >, 0);
1642 zio
->io_children_ready(zio
);
1645 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1646 ASSERT(zio
->io_bp_override
== NULL
);
1648 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1650 * We're rewriting an existing block, which means we're
1651 * working on behalf of spa_sync(). For spa_sync() to
1652 * converge, it must eventually be the case that we don't
1653 * have to allocate new blocks. But compression changes
1654 * the blocksize, which forces a reallocate, and makes
1655 * convergence take longer. Therefore, after the first
1656 * few passes, stop compressing to ensure convergence.
1658 pass
= spa_sync_pass(spa
);
1660 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1661 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1662 ASSERT(!BP_GET_DEDUP(bp
));
1664 if (pass
>= zfs_sync_pass_dont_compress
)
1665 compress
= ZIO_COMPRESS_OFF
;
1667 /* Make sure someone doesn't change their mind on overwrites */
1668 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1669 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1672 /* If it's a compressed write that is not raw, compress the buffer. */
1673 if (compress
!= ZIO_COMPRESS_OFF
&&
1674 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1675 void *cbuf
= zio_buf_alloc(lsize
);
1676 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1677 if (psize
== 0 || psize
== lsize
) {
1678 compress
= ZIO_COMPRESS_OFF
;
1679 zio_buf_free(cbuf
, lsize
);
1680 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1681 psize
<= BPE_PAYLOAD_SIZE
&&
1682 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1683 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1684 encode_embedded_bp_compressed(bp
,
1685 cbuf
, compress
, lsize
, psize
);
1686 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1687 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1688 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1689 zio_buf_free(cbuf
, lsize
);
1690 bp
->blk_birth
= zio
->io_txg
;
1691 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1692 ASSERT(spa_feature_is_active(spa
,
1693 SPA_FEATURE_EMBEDDED_DATA
));
1697 * Round up compressed size up to the ashift
1698 * of the smallest-ashift device, and zero the tail.
1699 * This ensures that the compressed size of the BP
1700 * (and thus compressratio property) are correct,
1701 * in that we charge for the padding used to fill out
1704 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1705 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1706 1ULL << spa
->spa_min_ashift
);
1707 if (rounded
>= lsize
) {
1708 compress
= ZIO_COMPRESS_OFF
;
1709 zio_buf_free(cbuf
, lsize
);
1712 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1713 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1714 abd_zero_off(cdata
, psize
, rounded
- psize
);
1716 zio_push_transform(zio
, cdata
,
1717 psize
, lsize
, NULL
);
1722 * We were unable to handle this as an override bp, treat
1723 * it as a regular write I/O.
1725 zio
->io_bp_override
= NULL
;
1726 *bp
= zio
->io_bp_orig
;
1727 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1729 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1730 zp
->zp_type
== DMU_OT_DNODE
) {
1732 * The DMU actually relies on the zio layer's compression
1733 * to free metadnode blocks that have had all contained
1734 * dnodes freed. As a result, even when doing a raw
1735 * receive, we must check whether the block can be compressed
1738 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1739 zio
->io_abd
, NULL
, lsize
);
1741 compress
= ZIO_COMPRESS_OFF
;
1743 ASSERT3U(psize
, !=, 0);
1747 * The final pass of spa_sync() must be all rewrites, but the first
1748 * few passes offer a trade-off: allocating blocks defers convergence,
1749 * but newly allocated blocks are sequential, so they can be written
1750 * to disk faster. Therefore, we allow the first few passes of
1751 * spa_sync() to allocate new blocks, but force rewrites after that.
1752 * There should only be a handful of blocks after pass 1 in any case.
1754 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1755 BP_GET_PSIZE(bp
) == psize
&&
1756 pass
>= zfs_sync_pass_rewrite
) {
1757 VERIFY3U(psize
, !=, 0);
1758 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1760 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1761 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1764 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1768 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1769 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1770 BP_SET_LSIZE(bp
, lsize
);
1771 BP_SET_TYPE(bp
, zp
->zp_type
);
1772 BP_SET_LEVEL(bp
, zp
->zp_level
);
1773 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1775 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1777 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1778 BP_SET_LSIZE(bp
, lsize
);
1779 BP_SET_TYPE(bp
, zp
->zp_type
);
1780 BP_SET_LEVEL(bp
, zp
->zp_level
);
1781 BP_SET_PSIZE(bp
, psize
);
1782 BP_SET_COMPRESS(bp
, compress
);
1783 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1784 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1785 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1787 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1788 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1789 ASSERT(!zp
->zp_encrypt
||
1790 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1791 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1793 if (zp
->zp_nopwrite
) {
1794 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1795 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1796 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1803 zio_free_bp_init(zio_t
*zio
)
1805 blkptr_t
*bp
= zio
->io_bp
;
1807 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1808 if (BP_GET_DEDUP(bp
))
1809 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1812 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1818 * ==========================================================================
1819 * Execute the I/O pipeline
1820 * ==========================================================================
1824 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1826 spa_t
*spa
= zio
->io_spa
;
1827 zio_type_t t
= zio
->io_type
;
1828 int flags
= (cutinline
? TQ_FRONT
: 0);
1831 * If we're a config writer or a probe, the normal issue and
1832 * interrupt threads may all be blocked waiting for the config lock.
1833 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1835 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1839 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1841 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1845 * If this is a high priority I/O, then use the high priority taskq if
1848 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1849 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1850 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1853 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1856 * NB: We are assuming that the zio can only be dispatched
1857 * to a single taskq at a time. It would be a grievous error
1858 * to dispatch the zio to another taskq at the same time.
1860 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1861 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1862 flags
, &zio
->io_tqent
);
1866 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1868 kthread_t
*executor
= zio
->io_executor
;
1869 spa_t
*spa
= zio
->io_spa
;
1871 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1872 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1874 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1875 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1884 zio_issue_async(zio_t
*zio
)
1886 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1892 zio_interrupt(zio_t
*zio
)
1894 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1898 zio_delay_interrupt(zio_t
*zio
)
1901 * The timeout_generic() function isn't defined in userspace, so
1902 * rather than trying to implement the function, the zio delay
1903 * functionality has been disabled for userspace builds.
1908 * If io_target_timestamp is zero, then no delay has been registered
1909 * for this IO, thus jump to the end of this function and "skip" the
1910 * delay; issuing it directly to the zio layer.
1912 if (zio
->io_target_timestamp
!= 0) {
1913 hrtime_t now
= gethrtime();
1915 if (now
>= zio
->io_target_timestamp
) {
1917 * This IO has already taken longer than the target
1918 * delay to complete, so we don't want to delay it
1919 * any longer; we "miss" the delay and issue it
1920 * directly to the zio layer. This is likely due to
1921 * the target latency being set to a value less than
1922 * the underlying hardware can satisfy (e.g. delay
1923 * set to 1ms, but the disks take 10ms to complete an
1927 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1933 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1934 clock_t expire_at_tick
= ddi_get_lbolt() +
1937 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1938 hrtime_t
, now
, hrtime_t
, diff
);
1940 if (NSEC_TO_TICK(diff
) == 0) {
1941 /* Our delay is less than a jiffy - just spin */
1942 zfs_sleep_until(zio
->io_target_timestamp
);
1946 * Use taskq_dispatch_delay() in the place of
1947 * OpenZFS's timeout_generic().
1949 tid
= taskq_dispatch_delay(system_taskq
,
1950 (task_func_t
*)zio_interrupt
,
1951 zio
, TQ_NOSLEEP
, expire_at_tick
);
1952 if (tid
== TASKQID_INVALID
) {
1954 * Couldn't allocate a task. Just
1955 * finish the zio without a delay.
1964 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1969 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
1971 zio_t
*cio
, *cio_next
;
1972 zio_link_t
*zl
= NULL
;
1973 vdev_t
*vd
= pio
->io_vd
;
1975 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
1976 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
1977 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1978 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1979 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1981 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
1982 "delta=%llu queued=%llu io=%llu "
1983 "path=%s last=%llu "
1984 "type=%d priority=%d flags=0x%x "
1985 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1986 "objset=%llu object=%llu level=%llu blkid=%llu "
1987 "offset=%llu size=%llu error=%d",
1988 ziodepth
, pio
, pio
->io_timestamp
,
1989 delta
, pio
->io_delta
, pio
->io_delay
,
1990 vd
? vd
->vdev_path
: "NULL", vq
? vq
->vq_io_complete_ts
: 0,
1991 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
1992 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
1993 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
1994 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
1995 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
1996 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
1998 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
1999 taskq_empty_ent(&pio
->io_tqent
)) {
2004 mutex_enter(&pio
->io_lock
);
2005 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2006 cio_next
= zio_walk_children(pio
, &zl
);
2007 zio_deadman_impl(cio
, ziodepth
+ 1);
2009 mutex_exit(&pio
->io_lock
);
2013 * Log the critical information describing this zio and all of its children
2014 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2017 zio_deadman(zio_t
*pio
, char *tag
)
2019 spa_t
*spa
= pio
->io_spa
;
2020 char *name
= spa_name(spa
);
2022 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
2025 zio_deadman_impl(pio
, 0);
2027 switch (spa_get_deadman_failmode(spa
)) {
2028 case ZIO_FAILURE_MODE_WAIT
:
2029 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
2032 case ZIO_FAILURE_MODE_CONTINUE
:
2033 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
2036 case ZIO_FAILURE_MODE_PANIC
:
2037 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
2043 * Execute the I/O pipeline until one of the following occurs:
2044 * (1) the I/O completes; (2) the pipeline stalls waiting for
2045 * dependent child I/Os; (3) the I/O issues, so we're waiting
2046 * for an I/O completion interrupt; (4) the I/O is delegated by
2047 * vdev-level caching or aggregation; (5) the I/O is deferred
2048 * due to vdev-level queueing; (6) the I/O is handed off to
2049 * another thread. In all cases, the pipeline stops whenever
2050 * there's no CPU work; it never burns a thread in cv_wait_io().
2052 * There's no locking on io_stage because there's no legitimate way
2053 * for multiple threads to be attempting to process the same I/O.
2055 static zio_pipe_stage_t
*zio_pipeline
[];
2058 * zio_execute() is a wrapper around the static function
2059 * __zio_execute() so that we can force __zio_execute() to be
2060 * inlined. This reduces stack overhead which is important
2061 * because __zio_execute() is called recursively in several zio
2062 * code paths. zio_execute() itself cannot be inlined because
2063 * it is externally visible.
2066 zio_execute(zio_t
*zio
)
2068 fstrans_cookie_t cookie
;
2070 cookie
= spl_fstrans_mark();
2072 spl_fstrans_unmark(cookie
);
2076 * Used to determine if in the current context the stack is sized large
2077 * enough to allow zio_execute() to be called recursively. A minimum
2078 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2081 zio_execute_stack_check(zio_t
*zio
)
2083 #if !defined(HAVE_LARGE_STACKS)
2084 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
2086 /* Executing in txg_sync_thread() context. */
2087 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
2090 /* Pool initialization outside of zio_taskq context. */
2091 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
2092 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
2093 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
2095 #endif /* HAVE_LARGE_STACKS */
2100 __attribute__((always_inline
))
2102 __zio_execute(zio_t
*zio
)
2104 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2106 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
2107 enum zio_stage pipeline
= zio
->io_pipeline
;
2108 enum zio_stage stage
= zio
->io_stage
;
2110 zio
->io_executor
= curthread
;
2112 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
2113 ASSERT(ISP2(stage
));
2114 ASSERT(zio
->io_stall
== NULL
);
2118 } while ((stage
& pipeline
) == 0);
2120 ASSERT(stage
<= ZIO_STAGE_DONE
);
2123 * If we are in interrupt context and this pipeline stage
2124 * will grab a config lock that is held across I/O,
2125 * or may wait for an I/O that needs an interrupt thread
2126 * to complete, issue async to avoid deadlock.
2128 * For VDEV_IO_START, we cut in line so that the io will
2129 * be sent to disk promptly.
2131 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
2132 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
2133 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2134 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2135 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2140 * If the current context doesn't have large enough stacks
2141 * the zio must be issued asynchronously to prevent overflow.
2143 if (zio_execute_stack_check(zio
)) {
2144 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2145 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2146 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2150 zio
->io_stage
= stage
;
2151 zio
->io_pipeline_trace
|= zio
->io_stage
;
2154 * The zio pipeline stage returns the next zio to execute
2155 * (typically the same as this one), or NULL if we should
2158 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2167 * ==========================================================================
2168 * Initiate I/O, either sync or async
2169 * ==========================================================================
2172 zio_wait(zio_t
*zio
)
2175 * Some routines, like zio_free_sync(), may return a NULL zio
2176 * to avoid the performance overhead of creating and then destroying
2177 * an unneeded zio. For the callers' simplicity, we accept a NULL
2178 * zio and ignore it.
2183 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2186 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2187 ASSERT3P(zio
->io_executor
, ==, NULL
);
2189 zio
->io_waiter
= curthread
;
2190 ASSERT0(zio
->io_queued_timestamp
);
2191 zio
->io_queued_timestamp
= gethrtime();
2195 mutex_enter(&zio
->io_lock
);
2196 while (zio
->io_executor
!= NULL
) {
2197 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2198 ddi_get_lbolt() + timeout
);
2200 if (zfs_deadman_enabled
&& error
== -1 &&
2201 gethrtime() - zio
->io_queued_timestamp
>
2202 spa_deadman_ziotime(zio
->io_spa
)) {
2203 mutex_exit(&zio
->io_lock
);
2204 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2205 zio_deadman(zio
, FTAG
);
2206 mutex_enter(&zio
->io_lock
);
2209 mutex_exit(&zio
->io_lock
);
2211 error
= zio
->io_error
;
2218 zio_nowait(zio_t
*zio
)
2221 * See comment in zio_wait().
2226 ASSERT3P(zio
->io_executor
, ==, NULL
);
2228 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2229 zio_unique_parent(zio
) == NULL
) {
2233 * This is a logical async I/O with no parent to wait for it.
2234 * We add it to the spa_async_root_zio "Godfather" I/O which
2235 * will ensure they complete prior to unloading the pool.
2237 spa_t
*spa
= zio
->io_spa
;
2239 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2242 zio_add_child(pio
, zio
);
2245 ASSERT0(zio
->io_queued_timestamp
);
2246 zio
->io_queued_timestamp
= gethrtime();
2251 * ==========================================================================
2252 * Reexecute, cancel, or suspend/resume failed I/O
2253 * ==========================================================================
2257 zio_reexecute(zio_t
*pio
)
2259 zio_t
*cio
, *cio_next
;
2261 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2262 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2263 ASSERT(pio
->io_gang_leader
== NULL
);
2264 ASSERT(pio
->io_gang_tree
== NULL
);
2266 pio
->io_flags
= pio
->io_orig_flags
;
2267 pio
->io_stage
= pio
->io_orig_stage
;
2268 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2269 pio
->io_reexecute
= 0;
2270 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2271 pio
->io_pipeline_trace
= 0;
2273 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2274 pio
->io_state
[w
] = 0;
2275 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2276 pio
->io_child_error
[c
] = 0;
2278 if (IO_IS_ALLOCATING(pio
))
2279 BP_ZERO(pio
->io_bp
);
2282 * As we reexecute pio's children, new children could be created.
2283 * New children go to the head of pio's io_child_list, however,
2284 * so we will (correctly) not reexecute them. The key is that
2285 * the remainder of pio's io_child_list, from 'cio_next' onward,
2286 * cannot be affected by any side effects of reexecuting 'cio'.
2288 zio_link_t
*zl
= NULL
;
2289 mutex_enter(&pio
->io_lock
);
2290 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2291 cio_next
= zio_walk_children(pio
, &zl
);
2292 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2293 pio
->io_children
[cio
->io_child_type
][w
]++;
2294 mutex_exit(&pio
->io_lock
);
2296 mutex_enter(&pio
->io_lock
);
2298 mutex_exit(&pio
->io_lock
);
2301 * Now that all children have been reexecuted, execute the parent.
2302 * We don't reexecute "The Godfather" I/O here as it's the
2303 * responsibility of the caller to wait on it.
2305 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2306 pio
->io_queued_timestamp
= gethrtime();
2312 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2314 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2315 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2316 "failure and the failure mode property for this pool "
2317 "is set to panic.", spa_name(spa
));
2319 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2320 "failure and has been suspended.\n", spa_name(spa
));
2322 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2325 mutex_enter(&spa
->spa_suspend_lock
);
2327 if (spa
->spa_suspend_zio_root
== NULL
)
2328 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2329 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2330 ZIO_FLAG_GODFATHER
);
2332 spa
->spa_suspended
= reason
;
2335 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2336 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2337 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2338 ASSERT(zio_unique_parent(zio
) == NULL
);
2339 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2340 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2343 mutex_exit(&spa
->spa_suspend_lock
);
2347 zio_resume(spa_t
*spa
)
2352 * Reexecute all previously suspended i/o.
2354 mutex_enter(&spa
->spa_suspend_lock
);
2355 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2356 cv_broadcast(&spa
->spa_suspend_cv
);
2357 pio
= spa
->spa_suspend_zio_root
;
2358 spa
->spa_suspend_zio_root
= NULL
;
2359 mutex_exit(&spa
->spa_suspend_lock
);
2365 return (zio_wait(pio
));
2369 zio_resume_wait(spa_t
*spa
)
2371 mutex_enter(&spa
->spa_suspend_lock
);
2372 while (spa_suspended(spa
))
2373 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2374 mutex_exit(&spa
->spa_suspend_lock
);
2378 * ==========================================================================
2381 * A gang block is a collection of small blocks that looks to the DMU
2382 * like one large block. When zio_dva_allocate() cannot find a block
2383 * of the requested size, due to either severe fragmentation or the pool
2384 * being nearly full, it calls zio_write_gang_block() to construct the
2385 * block from smaller fragments.
2387 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2388 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2389 * an indirect block: it's an array of block pointers. It consumes
2390 * only one sector and hence is allocatable regardless of fragmentation.
2391 * The gang header's bps point to its gang members, which hold the data.
2393 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2394 * as the verifier to ensure uniqueness of the SHA256 checksum.
2395 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2396 * not the gang header. This ensures that data block signatures (needed for
2397 * deduplication) are independent of how the block is physically stored.
2399 * Gang blocks can be nested: a gang member may itself be a gang block.
2400 * Thus every gang block is a tree in which root and all interior nodes are
2401 * gang headers, and the leaves are normal blocks that contain user data.
2402 * The root of the gang tree is called the gang leader.
2404 * To perform any operation (read, rewrite, free, claim) on a gang block,
2405 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2406 * in the io_gang_tree field of the original logical i/o by recursively
2407 * reading the gang leader and all gang headers below it. This yields
2408 * an in-core tree containing the contents of every gang header and the
2409 * bps for every constituent of the gang block.
2411 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2412 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2413 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2414 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2415 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2416 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2417 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2418 * of the gang header plus zio_checksum_compute() of the data to update the
2419 * gang header's blk_cksum as described above.
2421 * The two-phase assemble/issue model solves the problem of partial failure --
2422 * what if you'd freed part of a gang block but then couldn't read the
2423 * gang header for another part? Assembling the entire gang tree first
2424 * ensures that all the necessary gang header I/O has succeeded before
2425 * starting the actual work of free, claim, or write. Once the gang tree
2426 * is assembled, free and claim are in-memory operations that cannot fail.
2428 * In the event that a gang write fails, zio_dva_unallocate() walks the
2429 * gang tree to immediately free (i.e. insert back into the space map)
2430 * everything we've allocated. This ensures that we don't get ENOSPC
2431 * errors during repeated suspend/resume cycles due to a flaky device.
2433 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2434 * the gang tree, we won't modify the block, so we can safely defer the free
2435 * (knowing that the block is still intact). If we *can* assemble the gang
2436 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2437 * each constituent bp and we can allocate a new block on the next sync pass.
2439 * In all cases, the gang tree allows complete recovery from partial failure.
2440 * ==========================================================================
2444 zio_gang_issue_func_done(zio_t
*zio
)
2446 abd_put(zio
->io_abd
);
2450 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2456 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2457 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2458 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2459 &pio
->io_bookmark
));
2463 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2470 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2471 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2472 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2473 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2476 * As we rewrite each gang header, the pipeline will compute
2477 * a new gang block header checksum for it; but no one will
2478 * compute a new data checksum, so we do that here. The one
2479 * exception is the gang leader: the pipeline already computed
2480 * its data checksum because that stage precedes gang assembly.
2481 * (Presently, nothing actually uses interior data checksums;
2482 * this is just good hygiene.)
2484 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2485 abd_t
*buf
= abd_get_offset(data
, offset
);
2487 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2488 buf
, BP_GET_PSIZE(bp
));
2493 * If we are here to damage data for testing purposes,
2494 * leave the GBH alone so that we can detect the damage.
2496 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2497 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2499 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2500 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2501 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2502 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2510 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2513 zio_t
*zio
= zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2514 ZIO_GANG_CHILD_FLAGS(pio
));
2516 zio
= zio_null(pio
, pio
->io_spa
,
2517 NULL
, NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
));
2524 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2527 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2528 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2531 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2540 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2542 static zio_gang_node_t
*
2543 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2545 zio_gang_node_t
*gn
;
2547 ASSERT(*gnpp
== NULL
);
2549 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2550 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2557 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2559 zio_gang_node_t
*gn
= *gnpp
;
2561 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2562 ASSERT(gn
->gn_child
[g
] == NULL
);
2564 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2565 kmem_free(gn
, sizeof (*gn
));
2570 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2572 zio_gang_node_t
*gn
= *gnpp
;
2577 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2578 zio_gang_tree_free(&gn
->gn_child
[g
]);
2580 zio_gang_node_free(gnpp
);
2584 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2586 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2587 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2589 ASSERT(gio
->io_gang_leader
== gio
);
2590 ASSERT(BP_IS_GANG(bp
));
2592 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2593 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2594 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2598 zio_gang_tree_assemble_done(zio_t
*zio
)
2600 zio_t
*gio
= zio
->io_gang_leader
;
2601 zio_gang_node_t
*gn
= zio
->io_private
;
2602 blkptr_t
*bp
= zio
->io_bp
;
2604 ASSERT(gio
== zio_unique_parent(zio
));
2605 ASSERT(zio
->io_child_count
== 0);
2610 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2611 if (BP_SHOULD_BYTESWAP(bp
))
2612 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2614 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2615 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2616 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2618 abd_put(zio
->io_abd
);
2620 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2621 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2622 if (!BP_IS_GANG(gbp
))
2624 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2629 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2632 zio_t
*gio
= pio
->io_gang_leader
;
2635 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2636 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2637 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2640 * If you're a gang header, your data is in gn->gn_gbh.
2641 * If you're a gang member, your data is in 'data' and gn == NULL.
2643 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2646 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2648 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2649 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2650 if (BP_IS_HOLE(gbp
))
2652 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2654 offset
+= BP_GET_PSIZE(gbp
);
2658 if (gn
== gio
->io_gang_tree
)
2659 ASSERT3U(gio
->io_size
, ==, offset
);
2666 zio_gang_assemble(zio_t
*zio
)
2668 blkptr_t
*bp
= zio
->io_bp
;
2670 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2671 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2673 zio
->io_gang_leader
= zio
;
2675 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2681 zio_gang_issue(zio_t
*zio
)
2683 blkptr_t
*bp
= zio
->io_bp
;
2685 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2689 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2690 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2692 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2693 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2696 zio_gang_tree_free(&zio
->io_gang_tree
);
2698 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2704 zio_write_gang_member_ready(zio_t
*zio
)
2706 zio_t
*pio
= zio_unique_parent(zio
);
2707 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2708 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2710 zio_t
*gio __maybe_unused
= zio
->io_gang_leader
;
2712 if (BP_IS_HOLE(zio
->io_bp
))
2715 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2717 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2718 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2719 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2720 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2721 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2723 mutex_enter(&pio
->io_lock
);
2724 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2725 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2726 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2727 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2728 DVA_SET_ASIZE(&pdva
[d
], asize
);
2730 mutex_exit(&pio
->io_lock
);
2734 zio_write_gang_done(zio_t
*zio
)
2737 * The io_abd field will be NULL for a zio with no data. The io_flags
2738 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2739 * check for it here as it is cleared in zio_ready.
2741 if (zio
->io_abd
!= NULL
)
2742 abd_put(zio
->io_abd
);
2746 zio_write_gang_block(zio_t
*pio
)
2748 spa_t
*spa
= pio
->io_spa
;
2749 metaslab_class_t
*mc
= spa_normal_class(spa
);
2750 blkptr_t
*bp
= pio
->io_bp
;
2751 zio_t
*gio
= pio
->io_gang_leader
;
2753 zio_gang_node_t
*gn
, **gnpp
;
2754 zio_gbh_phys_t
*gbh
;
2756 uint64_t txg
= pio
->io_txg
;
2757 uint64_t resid
= pio
->io_size
;
2759 int copies
= gio
->io_prop
.zp_copies
;
2763 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2766 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2767 * have a third copy.
2769 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2770 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2771 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2773 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2774 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2775 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2778 flags
|= METASLAB_ASYNC_ALLOC
;
2779 VERIFY(zfs_refcount_held(&mc
->mc_alloc_slots
[pio
->io_allocator
],
2783 * The logical zio has already placed a reservation for
2784 * 'copies' allocation slots but gang blocks may require
2785 * additional copies. These additional copies
2786 * (i.e. gbh_copies - copies) are guaranteed to succeed
2787 * since metaslab_class_throttle_reserve() always allows
2788 * additional reservations for gang blocks.
2790 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2791 pio
->io_allocator
, pio
, flags
));
2794 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2795 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2796 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2798 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2799 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2803 * If we failed to allocate the gang block header then
2804 * we remove any additional allocation reservations that
2805 * we placed here. The original reservation will
2806 * be removed when the logical I/O goes to the ready
2809 metaslab_class_throttle_unreserve(mc
,
2810 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2813 pio
->io_error
= error
;
2818 gnpp
= &gio
->io_gang_tree
;
2820 gnpp
= pio
->io_private
;
2821 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2824 gn
= zio_gang_node_alloc(gnpp
);
2826 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2827 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2830 * Create the gang header.
2832 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2833 zio_write_gang_done
, NULL
, pio
->io_priority
,
2834 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2837 * Create and nowait the gang children.
2839 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2840 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2842 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2844 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2845 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2846 zp
.zp_type
= DMU_OT_NONE
;
2848 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2849 zp
.zp_dedup
= B_FALSE
;
2850 zp
.zp_dedup_verify
= B_FALSE
;
2851 zp
.zp_nopwrite
= B_FALSE
;
2852 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2853 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2854 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2855 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2856 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2858 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2859 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2860 resid
) : NULL
, lsize
, lsize
, &zp
,
2861 zio_write_gang_member_ready
, NULL
, NULL
,
2862 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2863 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2865 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2866 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2870 * Gang children won't throttle but we should
2871 * account for their work, so reserve an allocation
2872 * slot for them here.
2874 VERIFY(metaslab_class_throttle_reserve(mc
,
2875 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2881 * Set pio's pipeline to just wait for zio to finish.
2883 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2886 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2888 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2896 * The zio_nop_write stage in the pipeline determines if allocating a
2897 * new bp is necessary. The nopwrite feature can handle writes in
2898 * either syncing or open context (i.e. zil writes) and as a result is
2899 * mutually exclusive with dedup.
2901 * By leveraging a cryptographically secure checksum, such as SHA256, we
2902 * can compare the checksums of the new data and the old to determine if
2903 * allocating a new block is required. Note that our requirements for
2904 * cryptographic strength are fairly weak: there can't be any accidental
2905 * hash collisions, but we don't need to be secure against intentional
2906 * (malicious) collisions. To trigger a nopwrite, you have to be able
2907 * to write the file to begin with, and triggering an incorrect (hash
2908 * collision) nopwrite is no worse than simply writing to the file.
2909 * That said, there are no known attacks against the checksum algorithms
2910 * used for nopwrite, assuming that the salt and the checksums
2911 * themselves remain secret.
2914 zio_nop_write(zio_t
*zio
)
2916 blkptr_t
*bp
= zio
->io_bp
;
2917 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2918 zio_prop_t
*zp
= &zio
->io_prop
;
2920 ASSERT(BP_GET_LEVEL(bp
) == 0);
2921 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2922 ASSERT(zp
->zp_nopwrite
);
2923 ASSERT(!zp
->zp_dedup
);
2924 ASSERT(zio
->io_bp_override
== NULL
);
2925 ASSERT(IO_IS_ALLOCATING(zio
));
2928 * Check to see if the original bp and the new bp have matching
2929 * characteristics (i.e. same checksum, compression algorithms, etc).
2930 * If they don't then just continue with the pipeline which will
2931 * allocate a new bp.
2933 if (BP_IS_HOLE(bp_orig
) ||
2934 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2935 ZCHECKSUM_FLAG_NOPWRITE
) ||
2936 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2937 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2938 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2939 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2940 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2944 * If the checksums match then reset the pipeline so that we
2945 * avoid allocating a new bp and issuing any I/O.
2947 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2948 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2949 ZCHECKSUM_FLAG_NOPWRITE
);
2950 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2951 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2952 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2953 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2954 sizeof (uint64_t)) == 0);
2957 * If we're overwriting a block that is currently on an
2958 * indirect vdev, then ignore the nopwrite request and
2959 * allow a new block to be allocated on a concrete vdev.
2961 spa_config_enter(zio
->io_spa
, SCL_VDEV
, FTAG
, RW_READER
);
2962 vdev_t
*tvd
= vdev_lookup_top(zio
->io_spa
,
2963 DVA_GET_VDEV(&bp
->blk_dva
[0]));
2964 if (tvd
->vdev_ops
== &vdev_indirect_ops
) {
2965 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2968 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2971 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2972 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2979 * ==========================================================================
2981 * ==========================================================================
2984 zio_ddt_child_read_done(zio_t
*zio
)
2986 blkptr_t
*bp
= zio
->io_bp
;
2987 ddt_entry_t
*dde
= zio
->io_private
;
2989 zio_t
*pio
= zio_unique_parent(zio
);
2991 mutex_enter(&pio
->io_lock
);
2992 ddp
= ddt_phys_select(dde
, bp
);
2993 if (zio
->io_error
== 0)
2994 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2996 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2997 dde
->dde_repair_abd
= zio
->io_abd
;
2999 abd_free(zio
->io_abd
);
3000 mutex_exit(&pio
->io_lock
);
3004 zio_ddt_read_start(zio_t
*zio
)
3006 blkptr_t
*bp
= zio
->io_bp
;
3008 ASSERT(BP_GET_DEDUP(bp
));
3009 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3010 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3012 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3013 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3014 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
3015 ddt_phys_t
*ddp
= dde
->dde_phys
;
3016 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
3019 ASSERT(zio
->io_vsd
== NULL
);
3022 if (ddp_self
== NULL
)
3025 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
3026 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
3028 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
3030 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
3031 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
3032 zio
->io_size
, zio_ddt_child_read_done
, dde
,
3033 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
3034 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
3039 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
3040 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
3041 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
3047 zio_ddt_read_done(zio_t
*zio
)
3049 blkptr_t
*bp
= zio
->io_bp
;
3051 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
3055 ASSERT(BP_GET_DEDUP(bp
));
3056 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3057 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3059 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3060 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3061 ddt_entry_t
*dde
= zio
->io_vsd
;
3063 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
3067 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
3068 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
3071 if (dde
->dde_repair_abd
!= NULL
) {
3072 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
3074 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
3076 ddt_repair_done(ddt
, dde
);
3080 ASSERT(zio
->io_vsd
== NULL
);
3086 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
3088 spa_t
*spa
= zio
->io_spa
;
3089 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
3091 ASSERT(!(zio
->io_bp_override
&& do_raw
));
3094 * Note: we compare the original data, not the transformed data,
3095 * because when zio->io_bp is an override bp, we will not have
3096 * pushed the I/O transforms. That's an important optimization
3097 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3098 * However, we should never get a raw, override zio so in these
3099 * cases we can compare the io_abd directly. This is useful because
3100 * it allows us to do dedup verification even if we don't have access
3101 * to the original data (for instance, if the encryption keys aren't
3105 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3106 zio_t
*lio
= dde
->dde_lead_zio
[p
];
3108 if (lio
!= NULL
&& do_raw
) {
3109 return (lio
->io_size
!= zio
->io_size
||
3110 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
3111 } else if (lio
!= NULL
) {
3112 return (lio
->io_orig_size
!= zio
->io_orig_size
||
3113 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
3117 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3118 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3120 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
3121 blkptr_t blk
= *zio
->io_bp
;
3126 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3127 psize
= BP_GET_PSIZE(&blk
);
3129 if (psize
!= zio
->io_size
)
3134 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
3136 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
3137 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
3138 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
3139 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
3142 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
3143 error
= SET_ERROR(ENOENT
);
3148 return (error
!= 0);
3149 } else if (ddp
->ddp_phys_birth
!= 0) {
3150 arc_buf_t
*abuf
= NULL
;
3151 arc_flags_t aflags
= ARC_FLAG_WAIT
;
3152 blkptr_t blk
= *zio
->io_bp
;
3155 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3157 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3162 error
= arc_read(NULL
, spa
, &blk
,
3163 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3164 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3165 &aflags
, &zio
->io_bookmark
);
3168 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3169 zio
->io_orig_size
) != 0)
3170 error
= SET_ERROR(ENOENT
);
3171 arc_buf_destroy(abuf
, &abuf
);
3175 return (error
!= 0);
3183 zio_ddt_child_write_ready(zio_t
*zio
)
3185 int p
= zio
->io_prop
.zp_copies
;
3186 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3187 ddt_entry_t
*dde
= zio
->io_private
;
3188 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3196 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3198 ddt_phys_fill(ddp
, zio
->io_bp
);
3200 zio_link_t
*zl
= NULL
;
3201 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3202 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3208 zio_ddt_child_write_done(zio_t
*zio
)
3210 int p
= zio
->io_prop
.zp_copies
;
3211 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3212 ddt_entry_t
*dde
= zio
->io_private
;
3213 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3217 ASSERT(ddp
->ddp_refcnt
== 0);
3218 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3219 dde
->dde_lead_zio
[p
] = NULL
;
3221 if (zio
->io_error
== 0) {
3222 zio_link_t
*zl
= NULL
;
3223 while (zio_walk_parents(zio
, &zl
) != NULL
)
3224 ddt_phys_addref(ddp
);
3226 ddt_phys_clear(ddp
);
3233 zio_ddt_write(zio_t
*zio
)
3235 spa_t
*spa
= zio
->io_spa
;
3236 blkptr_t
*bp
= zio
->io_bp
;
3237 uint64_t txg
= zio
->io_txg
;
3238 zio_prop_t
*zp
= &zio
->io_prop
;
3239 int p
= zp
->zp_copies
;
3241 ddt_t
*ddt
= ddt_select(spa
, bp
);
3245 ASSERT(BP_GET_DEDUP(bp
));
3246 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3247 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3248 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3251 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3252 ddp
= &dde
->dde_phys
[p
];
3254 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3256 * If we're using a weak checksum, upgrade to a strong checksum
3257 * and try again. If we're already using a strong checksum,
3258 * we can't resolve it, so just convert to an ordinary write.
3259 * (And automatically e-mail a paper to Nature?)
3261 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3262 ZCHECKSUM_FLAG_DEDUP
)) {
3263 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3264 zio_pop_transforms(zio
);
3265 zio
->io_stage
= ZIO_STAGE_OPEN
;
3268 zp
->zp_dedup
= B_FALSE
;
3269 BP_SET_DEDUP(bp
, B_FALSE
);
3271 ASSERT(!BP_GET_DEDUP(bp
));
3272 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3277 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3278 if (ddp
->ddp_phys_birth
!= 0)
3279 ddt_bp_fill(ddp
, bp
, txg
);
3280 if (dde
->dde_lead_zio
[p
] != NULL
)
3281 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3283 ddt_phys_addref(ddp
);
3284 } else if (zio
->io_bp_override
) {
3285 ASSERT(bp
->blk_birth
== txg
);
3286 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3287 ddt_phys_fill(ddp
, bp
);
3288 ddt_phys_addref(ddp
);
3290 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3291 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3292 zio_ddt_child_write_ready
, NULL
, NULL
,
3293 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3294 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3296 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3297 dde
->dde_lead_zio
[p
] = cio
;
3307 ddt_entry_t
*freedde
; /* for debugging */
3310 zio_ddt_free(zio_t
*zio
)
3312 spa_t
*spa
= zio
->io_spa
;
3313 blkptr_t
*bp
= zio
->io_bp
;
3314 ddt_t
*ddt
= ddt_select(spa
, bp
);
3318 ASSERT(BP_GET_DEDUP(bp
));
3319 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3322 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3324 ddp
= ddt_phys_select(dde
, bp
);
3326 ddt_phys_decref(ddp
);
3334 * ==========================================================================
3335 * Allocate and free blocks
3336 * ==========================================================================
3340 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3344 ASSERT(MUTEX_HELD(&spa
->spa_alloc_locks
[allocator
]));
3346 zio
= avl_first(&spa
->spa_alloc_trees
[allocator
]);
3350 ASSERT(IO_IS_ALLOCATING(zio
));
3353 * Try to place a reservation for this zio. If we're unable to
3354 * reserve then we throttle.
3356 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3357 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3358 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
, 0)) {
3362 avl_remove(&spa
->spa_alloc_trees
[allocator
], zio
);
3363 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3369 zio_dva_throttle(zio_t
*zio
)
3371 spa_t
*spa
= zio
->io_spa
;
3373 metaslab_class_t
*mc
;
3375 /* locate an appropriate allocation class */
3376 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3377 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3379 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3380 !mc
->mc_alloc_throttle_enabled
||
3381 zio
->io_child_type
== ZIO_CHILD_GANG
||
3382 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3386 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3388 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3389 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3391 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
3393 * We want to try to use as many allocators as possible to help improve
3394 * performance, but we also want logically adjacent IOs to be physically
3395 * adjacent to improve sequential read performance. We chunk each object
3396 * into 2^20 block regions, and then hash based on the objset, object,
3397 * level, and region to accomplish both of these goals.
3399 zio
->io_allocator
= cityhash4(bm
->zb_objset
, bm
->zb_object
,
3400 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
3401 mutex_enter(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3402 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3403 zio
->io_metaslab_class
= mc
;
3404 avl_add(&spa
->spa_alloc_trees
[zio
->io_allocator
], zio
);
3405 nio
= zio_io_to_allocate(spa
, zio
->io_allocator
);
3406 mutex_exit(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3411 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3415 mutex_enter(&spa
->spa_alloc_locks
[allocator
]);
3416 zio
= zio_io_to_allocate(spa
, allocator
);
3417 mutex_exit(&spa
->spa_alloc_locks
[allocator
]);
3421 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3422 ASSERT0(zio
->io_error
);
3423 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3427 zio_dva_allocate(zio_t
*zio
)
3429 spa_t
*spa
= zio
->io_spa
;
3430 metaslab_class_t
*mc
;
3431 blkptr_t
*bp
= zio
->io_bp
;
3435 if (zio
->io_gang_leader
== NULL
) {
3436 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3437 zio
->io_gang_leader
= zio
;
3440 ASSERT(BP_IS_HOLE(bp
));
3441 ASSERT0(BP_GET_NDVAS(bp
));
3442 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3443 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3444 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3446 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3447 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3448 flags
|= METASLAB_DONT_THROTTLE
;
3449 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3450 flags
|= METASLAB_GANG_CHILD
;
3451 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3452 flags
|= METASLAB_ASYNC_ALLOC
;
3455 * if not already chosen, locate an appropriate allocation class
3457 mc
= zio
->io_metaslab_class
;
3459 mc
= spa_preferred_class(spa
, zio
->io_size
,
3460 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3461 zio
->io_prop
.zp_zpl_smallblk
);
3462 zio
->io_metaslab_class
= mc
;
3465 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3466 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3467 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3470 * Fallback to normal class when an alloc class is full
3472 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3474 * If throttling, transfer reservation over to normal class.
3475 * The io_allocator slot can remain the same even though we
3476 * are switching classes.
3478 if (mc
->mc_alloc_throttle_enabled
&&
3479 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3480 metaslab_class_throttle_unreserve(mc
,
3481 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3482 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3484 mc
= spa_normal_class(spa
);
3485 VERIFY(metaslab_class_throttle_reserve(mc
,
3486 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3487 flags
| METASLAB_MUST_RESERVE
));
3489 mc
= spa_normal_class(spa
);
3491 zio
->io_metaslab_class
= mc
;
3493 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3494 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3495 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3499 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3500 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3502 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3503 return (zio_write_gang_block(zio
));
3504 zio
->io_error
= error
;
3511 zio_dva_free(zio_t
*zio
)
3513 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3519 zio_dva_claim(zio_t
*zio
)
3523 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3525 zio
->io_error
= error
;
3531 * Undo an allocation. This is used by zio_done() when an I/O fails
3532 * and we want to give back the block we just allocated.
3533 * This handles both normal blocks and gang blocks.
3536 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3538 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3539 ASSERT(zio
->io_bp_override
== NULL
);
3541 if (!BP_IS_HOLE(bp
))
3542 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3545 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3546 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3547 &gn
->gn_gbh
->zg_blkptr
[g
]);
3553 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3556 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3557 uint64_t size
, boolean_t
*slog
)
3560 zio_alloc_list_t io_alloc_list
;
3562 ASSERT(txg
> spa_syncing_txg(spa
));
3564 metaslab_trace_init(&io_alloc_list
);
3567 * Block pointer fields are useful to metaslabs for stats and debugging.
3568 * Fill in the obvious ones before calling into metaslab_alloc().
3570 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3571 BP_SET_PSIZE(new_bp
, size
);
3572 BP_SET_LEVEL(new_bp
, 0);
3575 * When allocating a zil block, we don't have information about
3576 * the final destination of the block except the objset it's part
3577 * of, so we just hash the objset ID to pick the allocator to get
3580 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3581 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
,
3582 cityhash4(0, 0, 0, os
->os_dsl_dataset
->ds_object
) %
3583 spa
->spa_alloc_count
);
3587 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3588 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3589 &io_alloc_list
, NULL
, cityhash4(0, 0, 0,
3590 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
);
3594 metaslab_trace_fini(&io_alloc_list
);
3597 BP_SET_LSIZE(new_bp
, size
);
3598 BP_SET_PSIZE(new_bp
, size
);
3599 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3600 BP_SET_CHECKSUM(new_bp
,
3601 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3602 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3603 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3604 BP_SET_LEVEL(new_bp
, 0);
3605 BP_SET_DEDUP(new_bp
, 0);
3606 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3609 * encrypted blocks will require an IV and salt. We generate
3610 * these now since we will not be rewriting the bp at
3613 if (os
->os_encrypted
) {
3614 uint8_t iv
[ZIO_DATA_IV_LEN
];
3615 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3617 BP_SET_CRYPT(new_bp
, B_TRUE
);
3618 VERIFY0(spa_crypt_get_salt(spa
,
3619 dmu_objset_id(os
), salt
));
3620 VERIFY0(zio_crypt_generate_iv(iv
));
3622 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3625 zfs_dbgmsg("%s: zil block allocation failure: "
3626 "size %llu, error %d", spa_name(spa
), size
, error
);
3633 * ==========================================================================
3634 * Read and write to physical devices
3635 * ==========================================================================
3639 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3640 * stops after this stage and will resume upon I/O completion.
3641 * However, there are instances where the vdev layer may need to
3642 * continue the pipeline when an I/O was not issued. Since the I/O
3643 * that was sent to the vdev layer might be different than the one
3644 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3645 * force the underlying vdev layers to call either zio_execute() or
3646 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3649 zio_vdev_io_start(zio_t
*zio
)
3651 vdev_t
*vd
= zio
->io_vd
;
3653 spa_t
*spa
= zio
->io_spa
;
3657 ASSERT(zio
->io_error
== 0);
3658 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3661 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3662 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3665 * The mirror_ops handle multiple DVAs in a single BP.
3667 vdev_mirror_ops
.vdev_op_io_start(zio
);
3671 ASSERT3P(zio
->io_logical
, !=, zio
);
3672 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3673 ASSERT(spa
->spa_trust_config
);
3676 * Note: the code can handle other kinds of writes,
3677 * but we don't expect them.
3679 if (zio
->io_vd
->vdev_removing
) {
3680 ASSERT(zio
->io_flags
&
3681 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3682 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3686 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3688 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3689 P2PHASE(zio
->io_size
, align
) != 0) {
3690 /* Transform logical writes to be a full physical block size. */
3691 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3692 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3693 ASSERT(vd
== vd
->vdev_top
);
3694 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3695 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3696 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3698 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3702 * If this is not a physical io, make sure that it is properly aligned
3703 * before proceeding.
3705 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3706 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3707 ASSERT0(P2PHASE(zio
->io_size
, align
));
3710 * For physical writes, we allow 512b aligned writes and assume
3711 * the device will perform a read-modify-write as necessary.
3713 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3714 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3717 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3720 * If this is a repair I/O, and there's no self-healing involved --
3721 * that is, we're just resilvering what we expect to resilver --
3722 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3723 * This prevents spurious resilvering.
3725 * There are a few ways that we can end up creating these spurious
3728 * 1. A resilver i/o will be issued if any DVA in the BP has a
3729 * dirty DTL. The mirror code will issue resilver writes to
3730 * each DVA, including the one(s) that are not on vdevs with dirty
3733 * 2. With nested replication, which happens when we have a
3734 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3735 * For example, given mirror(replacing(A+B), C), it's likely that
3736 * only A is out of date (it's the new device). In this case, we'll
3737 * read from C, then use the data to resilver A+B -- but we don't
3738 * actually want to resilver B, just A. The top-level mirror has no
3739 * way to know this, so instead we just discard unnecessary repairs
3740 * as we work our way down the vdev tree.
3742 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3743 * The same logic applies to any form of nested replication: ditto
3744 * + mirror, RAID-Z + replacing, etc.
3746 * However, indirect vdevs point off to other vdevs which may have
3747 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3748 * will be properly bypassed instead.
3750 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3751 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3752 zio
->io_txg
!= 0 && /* not a delegated i/o */
3753 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3754 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3755 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3756 zio_vdev_io_bypass(zio
);
3760 if (vd
->vdev_ops
->vdev_op_leaf
&& (zio
->io_type
== ZIO_TYPE_READ
||
3761 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
)) {
3763 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3766 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3769 if (!vdev_accessible(vd
, zio
)) {
3770 zio
->io_error
= SET_ERROR(ENXIO
);
3774 zio
->io_delay
= gethrtime();
3777 vd
->vdev_ops
->vdev_op_io_start(zio
);
3782 zio_vdev_io_done(zio_t
*zio
)
3784 vdev_t
*vd
= zio
->io_vd
;
3785 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3786 boolean_t unexpected_error
= B_FALSE
;
3788 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3792 ASSERT(zio
->io_type
== ZIO_TYPE_READ
||
3793 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
);
3796 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3798 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3800 vdev_queue_io_done(zio
);
3802 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3803 vdev_cache_write(zio
);
3805 if (zio_injection_enabled
&& zio
->io_error
== 0)
3806 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3809 if (zio_injection_enabled
&& zio
->io_error
== 0)
3810 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3812 if (zio
->io_error
&& zio
->io_type
!= ZIO_TYPE_TRIM
) {
3813 if (!vdev_accessible(vd
, zio
)) {
3814 zio
->io_error
= SET_ERROR(ENXIO
);
3816 unexpected_error
= B_TRUE
;
3821 ops
->vdev_op_io_done(zio
);
3823 if (unexpected_error
)
3824 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3830 * This function is used to change the priority of an existing zio that is
3831 * currently in-flight. This is used by the arc to upgrade priority in the
3832 * event that a demand read is made for a block that is currently queued
3833 * as a scrub or async read IO. Otherwise, the high priority read request
3834 * would end up having to wait for the lower priority IO.
3837 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3839 zio_t
*cio
, *cio_next
;
3840 zio_link_t
*zl
= NULL
;
3842 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3844 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3845 vdev_queue_change_io_priority(pio
, priority
);
3847 pio
->io_priority
= priority
;
3850 mutex_enter(&pio
->io_lock
);
3851 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3852 cio_next
= zio_walk_children(pio
, &zl
);
3853 zio_change_priority(cio
, priority
);
3855 mutex_exit(&pio
->io_lock
);
3859 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3860 * disk, and use that to finish the checksum ereport later.
3863 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3864 const abd_t
*good_buf
)
3866 /* no processing needed */
3867 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3872 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3874 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3876 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3878 zcr
->zcr_cbinfo
= zio
->io_size
;
3879 zcr
->zcr_cbdata
= abd
;
3880 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3881 zcr
->zcr_free
= zio_abd_free
;
3885 zio_vdev_io_assess(zio_t
*zio
)
3887 vdev_t
*vd
= zio
->io_vd
;
3889 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3893 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3894 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3896 if (zio
->io_vsd
!= NULL
) {
3897 zio
->io_vsd_ops
->vsd_free(zio
);
3901 if (zio_injection_enabled
&& zio
->io_error
== 0)
3902 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3905 * If the I/O failed, determine whether we should attempt to retry it.
3907 * On retry, we cut in line in the issue queue, since we don't want
3908 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3910 if (zio
->io_error
&& vd
== NULL
&&
3911 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3912 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3913 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3915 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3916 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3917 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3918 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3919 zio_requeue_io_start_cut_in_line
);
3924 * If we got an error on a leaf device, convert it to ENXIO
3925 * if the device is not accessible at all.
3927 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3928 !vdev_accessible(vd
, zio
))
3929 zio
->io_error
= SET_ERROR(ENXIO
);
3932 * If we can't write to an interior vdev (mirror or RAID-Z),
3933 * set vdev_cant_write so that we stop trying to allocate from it.
3935 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3936 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3937 vd
->vdev_cant_write
= B_TRUE
;
3941 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3942 * attempts will ever succeed. In this case we set a persistent
3943 * boolean flag so that we don't bother with it in the future.
3945 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3946 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3947 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3948 vd
->vdev_nowritecache
= B_TRUE
;
3951 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3953 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3954 zio
->io_physdone
!= NULL
) {
3955 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3956 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3957 zio
->io_physdone(zio
->io_logical
);
3964 zio_vdev_io_reissue(zio_t
*zio
)
3966 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3967 ASSERT(zio
->io_error
== 0);
3969 zio
->io_stage
>>= 1;
3973 zio_vdev_io_redone(zio_t
*zio
)
3975 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3977 zio
->io_stage
>>= 1;
3981 zio_vdev_io_bypass(zio_t
*zio
)
3983 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3984 ASSERT(zio
->io_error
== 0);
3986 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3987 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3991 * ==========================================================================
3992 * Encrypt and store encryption parameters
3993 * ==========================================================================
3998 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3999 * managing the storage of encryption parameters and passing them to the
4000 * lower-level encryption functions.
4003 zio_encrypt(zio_t
*zio
)
4005 zio_prop_t
*zp
= &zio
->io_prop
;
4006 spa_t
*spa
= zio
->io_spa
;
4007 blkptr_t
*bp
= zio
->io_bp
;
4008 uint64_t psize
= BP_GET_PSIZE(bp
);
4009 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
4010 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
4011 void *enc_buf
= NULL
;
4013 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4014 uint8_t iv
[ZIO_DATA_IV_LEN
];
4015 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4016 boolean_t no_crypt
= B_FALSE
;
4018 /* the root zio already encrypted the data */
4019 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
4022 /* only ZIL blocks are re-encrypted on rewrite */
4023 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
4026 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
4027 BP_SET_CRYPT(bp
, B_FALSE
);
4031 /* if we are doing raw encryption set the provided encryption params */
4032 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
4033 ASSERT0(BP_GET_LEVEL(bp
));
4034 BP_SET_CRYPT(bp
, B_TRUE
);
4035 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
4036 if (ot
!= DMU_OT_OBJSET
)
4037 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
4039 /* dnode blocks must be written out in the provided byteorder */
4040 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
4041 ot
== DMU_OT_DNODE
) {
4042 void *bswap_buf
= zio_buf_alloc(psize
);
4043 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
4045 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4046 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
4047 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
4050 abd_take_ownership_of_buf(babd
, B_TRUE
);
4051 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
4054 if (DMU_OT_IS_ENCRYPTED(ot
))
4055 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
4059 /* indirect blocks only maintain a cksum of the lower level MACs */
4060 if (BP_GET_LEVEL(bp
) > 0) {
4061 BP_SET_CRYPT(bp
, B_TRUE
);
4062 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
4063 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
4065 zio_crypt_encode_mac_bp(bp
, mac
);
4070 * Objset blocks are a special case since they have 2 256-bit MACs
4071 * embedded within them.
4073 if (ot
== DMU_OT_OBJSET
) {
4074 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
4075 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4076 BP_SET_CRYPT(bp
, B_TRUE
);
4077 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
4078 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
4082 /* unencrypted object types are only authenticated with a MAC */
4083 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
4084 BP_SET_CRYPT(bp
, B_TRUE
);
4085 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
4086 zio
->io_abd
, psize
, mac
));
4087 zio_crypt_encode_mac_bp(bp
, mac
);
4092 * Later passes of sync-to-convergence may decide to rewrite data
4093 * in place to avoid more disk reallocations. This presents a problem
4094 * for encryption because this constitutes rewriting the new data with
4095 * the same encryption key and IV. However, this only applies to blocks
4096 * in the MOS (particularly the spacemaps) and we do not encrypt the
4097 * MOS. We assert that the zio is allocating or an intent log write
4100 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
4101 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
4102 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
4103 ASSERT3U(psize
, !=, 0);
4105 enc_buf
= zio_buf_alloc(psize
);
4106 eabd
= abd_get_from_buf(enc_buf
, psize
);
4107 abd_take_ownership_of_buf(eabd
, B_TRUE
);
4110 * For an explanation of what encryption parameters are stored
4111 * where, see the block comment in zio_crypt.c.
4113 if (ot
== DMU_OT_INTENT_LOG
) {
4114 zio_crypt_decode_params_bp(bp
, salt
, iv
);
4116 BP_SET_CRYPT(bp
, B_TRUE
);
4119 /* Perform the encryption. This should not fail */
4120 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
4121 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
4122 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
4124 /* encode encryption metadata into the bp */
4125 if (ot
== DMU_OT_INTENT_LOG
) {
4127 * ZIL blocks store the MAC in the embedded checksum, so the
4128 * transform must always be applied.
4130 zio_crypt_encode_mac_zil(enc_buf
, mac
);
4131 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4133 BP_SET_CRYPT(bp
, B_TRUE
);
4134 zio_crypt_encode_params_bp(bp
, salt
, iv
);
4135 zio_crypt_encode_mac_bp(bp
, mac
);
4138 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
4141 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4149 * ==========================================================================
4150 * Generate and verify checksums
4151 * ==========================================================================
4154 zio_checksum_generate(zio_t
*zio
)
4156 blkptr_t
*bp
= zio
->io_bp
;
4157 enum zio_checksum checksum
;
4161 * This is zio_write_phys().
4162 * We're either generating a label checksum, or none at all.
4164 checksum
= zio
->io_prop
.zp_checksum
;
4166 if (checksum
== ZIO_CHECKSUM_OFF
)
4169 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4171 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4172 ASSERT(!IO_IS_ALLOCATING(zio
));
4173 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4175 checksum
= BP_GET_CHECKSUM(bp
);
4179 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4185 zio_checksum_verify(zio_t
*zio
)
4187 zio_bad_cksum_t info
;
4188 blkptr_t
*bp
= zio
->io_bp
;
4191 ASSERT(zio
->io_vd
!= NULL
);
4195 * This is zio_read_phys().
4196 * We're either verifying a label checksum, or nothing at all.
4198 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4201 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
4204 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4205 zio
->io_error
= error
;
4206 if (error
== ECKSUM
&&
4207 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4208 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4209 zio
->io_vd
->vdev_stat
.vs_checksum_errors
++;
4210 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4212 zfs_ereport_start_checksum(zio
->io_spa
,
4213 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4214 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
4222 * Called by RAID-Z to ensure we don't compute the checksum twice.
4225 zio_checksum_verified(zio_t
*zio
)
4227 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4231 * ==========================================================================
4232 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4233 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4234 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4235 * indicate errors that are specific to one I/O, and most likely permanent.
4236 * Any other error is presumed to be worse because we weren't expecting it.
4237 * ==========================================================================
4240 zio_worst_error(int e1
, int e2
)
4242 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4245 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4246 if (e1
== zio_error_rank
[r1
])
4249 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4250 if (e2
== zio_error_rank
[r2
])
4253 return (r1
> r2
? e1
: e2
);
4257 * ==========================================================================
4259 * ==========================================================================
4262 zio_ready(zio_t
*zio
)
4264 blkptr_t
*bp
= zio
->io_bp
;
4265 zio_t
*pio
, *pio_next
;
4266 zio_link_t
*zl
= NULL
;
4268 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4273 if (zio
->io_ready
) {
4274 ASSERT(IO_IS_ALLOCATING(zio
));
4275 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4276 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4277 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4282 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4283 zio
->io_bp_copy
= *bp
;
4285 if (zio
->io_error
!= 0) {
4286 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4288 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4289 ASSERT(IO_IS_ALLOCATING(zio
));
4290 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4291 ASSERT(zio
->io_metaslab_class
!= NULL
);
4294 * We were unable to allocate anything, unreserve and
4295 * issue the next I/O to allocate.
4297 metaslab_class_throttle_unreserve(
4298 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4299 zio
->io_allocator
, zio
);
4300 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4304 mutex_enter(&zio
->io_lock
);
4305 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4306 pio
= zio_walk_parents(zio
, &zl
);
4307 mutex_exit(&zio
->io_lock
);
4310 * As we notify zio's parents, new parents could be added.
4311 * New parents go to the head of zio's io_parent_list, however,
4312 * so we will (correctly) not notify them. The remainder of zio's
4313 * io_parent_list, from 'pio_next' onward, cannot change because
4314 * all parents must wait for us to be done before they can be done.
4316 for (; pio
!= NULL
; pio
= pio_next
) {
4317 pio_next
= zio_walk_parents(zio
, &zl
);
4318 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4321 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4322 if (BP_IS_GANG(bp
)) {
4323 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4325 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4326 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4330 if (zio_injection_enabled
&&
4331 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4332 zio_handle_ignored_writes(zio
);
4338 * Update the allocation throttle accounting.
4341 zio_dva_throttle_done(zio_t
*zio
)
4343 zio_t
*lio __maybe_unused
= zio
->io_logical
;
4344 zio_t
*pio
= zio_unique_parent(zio
);
4345 vdev_t
*vd
= zio
->io_vd
;
4346 int flags
= METASLAB_ASYNC_ALLOC
;
4348 ASSERT3P(zio
->io_bp
, !=, NULL
);
4349 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4350 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4351 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4353 ASSERT3P(vd
, ==, vd
->vdev_top
);
4354 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4355 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4356 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4357 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4358 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4361 * Parents of gang children can have two flavors -- ones that
4362 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4363 * and ones that allocated the constituent blocks. The allocation
4364 * throttle needs to know the allocating parent zio so we must find
4367 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4369 * If our parent is a rewrite gang child then our grandparent
4370 * would have been the one that performed the allocation.
4372 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4373 pio
= zio_unique_parent(pio
);
4374 flags
|= METASLAB_GANG_CHILD
;
4377 ASSERT(IO_IS_ALLOCATING(pio
));
4378 ASSERT3P(zio
, !=, zio
->io_logical
);
4379 ASSERT(zio
->io_logical
!= NULL
);
4380 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4381 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4382 ASSERT(zio
->io_metaslab_class
!= NULL
);
4384 mutex_enter(&pio
->io_lock
);
4385 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4386 pio
->io_allocator
, B_TRUE
);
4387 mutex_exit(&pio
->io_lock
);
4389 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4390 pio
->io_allocator
, pio
);
4393 * Call into the pipeline to see if there is more work that
4394 * needs to be done. If there is work to be done it will be
4395 * dispatched to another taskq thread.
4397 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4401 zio_done(zio_t
*zio
)
4404 * Always attempt to keep stack usage minimal here since
4405 * we can be called recursively up to 19 levels deep.
4407 const uint64_t psize
= zio
->io_size
;
4408 zio_t
*pio
, *pio_next
;
4409 zio_link_t
*zl
= NULL
;
4412 * If our children haven't all completed,
4413 * wait for them and then repeat this pipeline stage.
4415 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4420 * If the allocation throttle is enabled, then update the accounting.
4421 * We only track child I/Os that are part of an allocating async
4422 * write. We must do this since the allocation is performed
4423 * by the logical I/O but the actual write is done by child I/Os.
4425 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4426 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4427 ASSERT(zio
->io_metaslab_class
!= NULL
);
4428 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4429 zio_dva_throttle_done(zio
);
4433 * If the allocation throttle is enabled, verify that
4434 * we have decremented the refcounts for every I/O that was throttled.
4436 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4437 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4438 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4439 ASSERT(zio
->io_bp
!= NULL
);
4441 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4443 VERIFY(zfs_refcount_not_held(
4444 &zio
->io_metaslab_class
->mc_alloc_slots
[zio
->io_allocator
],
4449 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4450 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4451 ASSERT(zio
->io_children
[c
][w
] == 0);
4453 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4454 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4455 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4456 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4457 sizeof (blkptr_t
)) == 0 ||
4458 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4459 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4460 zio
->io_bp_override
== NULL
&&
4461 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4462 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4463 BP_GET_NDVAS(zio
->io_bp
));
4464 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4465 (BP_COUNT_GANG(zio
->io_bp
) ==
4466 BP_GET_NDVAS(zio
->io_bp
)));
4468 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4469 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4473 * If there were child vdev/gang/ddt errors, they apply to us now.
4475 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4476 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4477 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4480 * If the I/O on the transformed data was successful, generate any
4481 * checksum reports now while we still have the transformed data.
4483 if (zio
->io_error
== 0) {
4484 while (zio
->io_cksum_report
!= NULL
) {
4485 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4486 uint64_t align
= zcr
->zcr_align
;
4487 uint64_t asize
= P2ROUNDUP(psize
, align
);
4488 abd_t
*adata
= zio
->io_abd
;
4490 if (asize
!= psize
) {
4491 adata
= abd_alloc(asize
, B_TRUE
);
4492 abd_copy(adata
, zio
->io_abd
, psize
);
4493 abd_zero_off(adata
, psize
, asize
- psize
);
4496 zio
->io_cksum_report
= zcr
->zcr_next
;
4497 zcr
->zcr_next
= NULL
;
4498 zcr
->zcr_finish(zcr
, adata
);
4499 zfs_ereport_free_checksum(zcr
);
4506 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4508 vdev_stat_update(zio
, psize
);
4511 * If this I/O is attached to a particular vdev is slow, exceeding
4512 * 30 seconds to complete, post an error described the I/O delay.
4513 * We ignore these errors if the device is currently unavailable.
4515 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4516 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4518 * We want to only increment our slow IO counters if
4519 * the IO is valid (i.e. not if the drive is removed).
4521 * zfs_ereport_post() will also do these checks, but
4522 * it can also ratelimit and have other failures, so we
4523 * need to increment the slow_io counters independent
4526 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4527 zio
->io_spa
, zio
->io_vd
, zio
)) {
4528 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4529 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4530 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4532 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4533 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4539 if (zio
->io_error
) {
4541 * If this I/O is attached to a particular vdev,
4542 * generate an error message describing the I/O failure
4543 * at the block level. We ignore these errors if the
4544 * device is currently unavailable.
4546 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4547 !vdev_is_dead(zio
->io_vd
)) {
4548 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4549 if (zio
->io_type
== ZIO_TYPE_READ
) {
4550 zio
->io_vd
->vdev_stat
.vs_read_errors
++;
4551 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
4552 zio
->io_vd
->vdev_stat
.vs_write_errors
++;
4554 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4556 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4557 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4560 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4561 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4562 zio
== zio
->io_logical
) {
4564 * For logical I/O requests, tell the SPA to log the
4565 * error and generate a logical data ereport.
4567 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4568 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4569 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4573 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4575 * Determine whether zio should be reexecuted. This will
4576 * propagate all the way to the root via zio_notify_parent().
4578 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4579 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4581 if (IO_IS_ALLOCATING(zio
) &&
4582 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4583 if (zio
->io_error
!= ENOSPC
)
4584 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4586 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4589 if ((zio
->io_type
== ZIO_TYPE_READ
||
4590 zio
->io_type
== ZIO_TYPE_FREE
) &&
4591 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4592 zio
->io_error
== ENXIO
&&
4593 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4594 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4595 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4597 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4598 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4601 * Here is a possibly good place to attempt to do
4602 * either combinatorial reconstruction or error correction
4603 * based on checksums. It also might be a good place
4604 * to send out preliminary ereports before we suspend
4610 * If there were logical child errors, they apply to us now.
4611 * We defer this until now to avoid conflating logical child
4612 * errors with errors that happened to the zio itself when
4613 * updating vdev stats and reporting FMA events above.
4615 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4617 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4618 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4619 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4620 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4622 zio_gang_tree_free(&zio
->io_gang_tree
);
4625 * Godfather I/Os should never suspend.
4627 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4628 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4629 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4631 if (zio
->io_reexecute
) {
4633 * This is a logical I/O that wants to reexecute.
4635 * Reexecute is top-down. When an i/o fails, if it's not
4636 * the root, it simply notifies its parent and sticks around.
4637 * The parent, seeing that it still has children in zio_done(),
4638 * does the same. This percolates all the way up to the root.
4639 * The root i/o will reexecute or suspend the entire tree.
4641 * This approach ensures that zio_reexecute() honors
4642 * all the original i/o dependency relationships, e.g.
4643 * parents not executing until children are ready.
4645 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4647 zio
->io_gang_leader
= NULL
;
4649 mutex_enter(&zio
->io_lock
);
4650 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4651 mutex_exit(&zio
->io_lock
);
4654 * "The Godfather" I/O monitors its children but is
4655 * not a true parent to them. It will track them through
4656 * the pipeline but severs its ties whenever they get into
4657 * trouble (e.g. suspended). This allows "The Godfather"
4658 * I/O to return status without blocking.
4661 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4663 zio_link_t
*remove_zl
= zl
;
4664 pio_next
= zio_walk_parents(zio
, &zl
);
4666 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4667 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4668 zio_remove_child(pio
, zio
, remove_zl
);
4670 * This is a rare code path, so we don't
4671 * bother with "next_to_execute".
4673 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4678 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4680 * We're not a root i/o, so there's nothing to do
4681 * but notify our parent. Don't propagate errors
4682 * upward since we haven't permanently failed yet.
4684 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4685 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4687 * This is a rare code path, so we don't bother with
4688 * "next_to_execute".
4690 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4691 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4693 * We'd fail again if we reexecuted now, so suspend
4694 * until conditions improve (e.g. device comes online).
4696 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4699 * Reexecution is potentially a huge amount of work.
4700 * Hand it off to the otherwise-unused claim taskq.
4702 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4703 spa_taskq_dispatch_ent(zio
->io_spa
,
4704 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4705 (task_func_t
*)zio_reexecute
, zio
, 0,
4711 ASSERT(zio
->io_child_count
== 0);
4712 ASSERT(zio
->io_reexecute
== 0);
4713 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4716 * Report any checksum errors, since the I/O is complete.
4718 while (zio
->io_cksum_report
!= NULL
) {
4719 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4720 zio
->io_cksum_report
= zcr
->zcr_next
;
4721 zcr
->zcr_next
= NULL
;
4722 zcr
->zcr_finish(zcr
, NULL
);
4723 zfs_ereport_free_checksum(zcr
);
4726 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4727 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4728 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4729 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4733 * It is the responsibility of the done callback to ensure that this
4734 * particular zio is no longer discoverable for adoption, and as
4735 * such, cannot acquire any new parents.
4740 mutex_enter(&zio
->io_lock
);
4741 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4742 mutex_exit(&zio
->io_lock
);
4745 * We are done executing this zio. We may want to execute a parent
4746 * next. See the comment in zio_notify_parent().
4748 zio_t
*next_to_execute
= NULL
;
4750 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4751 zio_link_t
*remove_zl
= zl
;
4752 pio_next
= zio_walk_parents(zio
, &zl
);
4753 zio_remove_child(pio
, zio
, remove_zl
);
4754 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
4757 if (zio
->io_waiter
!= NULL
) {
4758 mutex_enter(&zio
->io_lock
);
4759 zio
->io_executor
= NULL
;
4760 cv_broadcast(&zio
->io_cv
);
4761 mutex_exit(&zio
->io_lock
);
4766 return (next_to_execute
);
4770 * ==========================================================================
4771 * I/O pipeline definition
4772 * ==========================================================================
4774 static zio_pipe_stage_t
*zio_pipeline
[] = {
4782 zio_checksum_generate
,
4798 zio_checksum_verify
,
4806 * Compare two zbookmark_phys_t's to see which we would reach first in a
4807 * pre-order traversal of the object tree.
4809 * This is simple in every case aside from the meta-dnode object. For all other
4810 * objects, we traverse them in order (object 1 before object 2, and so on).
4811 * However, all of these objects are traversed while traversing object 0, since
4812 * the data it points to is the list of objects. Thus, we need to convert to a
4813 * canonical representation so we can compare meta-dnode bookmarks to
4814 * non-meta-dnode bookmarks.
4816 * We do this by calculating "equivalents" for each field of the zbookmark.
4817 * zbookmarks outside of the meta-dnode use their own object and level, and
4818 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4819 * blocks this bookmark refers to) by multiplying their blkid by their span
4820 * (the number of L0 blocks contained within one block at their level).
4821 * zbookmarks inside the meta-dnode calculate their object equivalent
4822 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4823 * level + 1<<31 (any value larger than a level could ever be) for their level.
4824 * This causes them to always compare before a bookmark in their object
4825 * equivalent, compare appropriately to bookmarks in other objects, and to
4826 * compare appropriately to other bookmarks in the meta-dnode.
4829 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4830 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4833 * These variables represent the "equivalent" values for the zbookmark,
4834 * after converting zbookmarks inside the meta dnode to their
4835 * normal-object equivalents.
4837 uint64_t zb1obj
, zb2obj
;
4838 uint64_t zb1L0
, zb2L0
;
4839 uint64_t zb1level
, zb2level
;
4841 if (zb1
->zb_object
== zb2
->zb_object
&&
4842 zb1
->zb_level
== zb2
->zb_level
&&
4843 zb1
->zb_blkid
== zb2
->zb_blkid
)
4846 IMPLY(zb1
->zb_level
> 0, ibs1
>= SPA_MINBLOCKSHIFT
);
4847 IMPLY(zb2
->zb_level
> 0, ibs2
>= SPA_MINBLOCKSHIFT
);
4850 * BP_SPANB calculates the span in blocks.
4852 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4853 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4855 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4856 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4858 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4860 zb1obj
= zb1
->zb_object
;
4861 zb1level
= zb1
->zb_level
;
4864 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4865 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4867 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4869 zb2obj
= zb2
->zb_object
;
4870 zb2level
= zb2
->zb_level
;
4873 /* Now that we have a canonical representation, do the comparison. */
4874 if (zb1obj
!= zb2obj
)
4875 return (zb1obj
< zb2obj
? -1 : 1);
4876 else if (zb1L0
!= zb2L0
)
4877 return (zb1L0
< zb2L0
? -1 : 1);
4878 else if (zb1level
!= zb2level
)
4879 return (zb1level
> zb2level
? -1 : 1);
4881 * This can (theoretically) happen if the bookmarks have the same object
4882 * and level, but different blkids, if the block sizes are not the same.
4883 * There is presently no way to change the indirect block sizes
4889 * This function checks the following: given that last_block is the place that
4890 * our traversal stopped last time, does that guarantee that we've visited
4891 * every node under subtree_root? Therefore, we can't just use the raw output
4892 * of zbookmark_compare. We have to pass in a modified version of
4893 * subtree_root; by incrementing the block id, and then checking whether
4894 * last_block is before or equal to that, we can tell whether or not having
4895 * visited last_block implies that all of subtree_root's children have been
4899 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4900 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4902 zbookmark_phys_t mod_zb
= *subtree_root
;
4904 ASSERT(last_block
->zb_level
== 0);
4906 /* The objset_phys_t isn't before anything. */
4911 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4912 * data block size in sectors, because that variable is only used if
4913 * the bookmark refers to a block in the meta-dnode. Since we don't
4914 * know without examining it what object it refers to, and there's no
4915 * harm in passing in this value in other cases, we always pass it in.
4917 * We pass in 0 for the indirect block size shift because zb2 must be
4918 * level 0. The indirect block size is only used to calculate the span
4919 * of the bookmark, but since the bookmark must be level 0, the span is
4920 * always 1, so the math works out.
4922 * If you make changes to how the zbookmark_compare code works, be sure
4923 * to make sure that this code still works afterwards.
4925 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4926 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4930 EXPORT_SYMBOL(zio_type_name
);
4931 EXPORT_SYMBOL(zio_buf_alloc
);
4932 EXPORT_SYMBOL(zio_data_buf_alloc
);
4933 EXPORT_SYMBOL(zio_buf_free
);
4934 EXPORT_SYMBOL(zio_data_buf_free
);
4937 ZFS_MODULE_PARAM(zfs_zio
, zio_
, slow_io_ms
, INT
, ZMOD_RW
,
4938 "Max I/O completion time (milliseconds) before marking it as slow");
4940 ZFS_MODULE_PARAM(zfs_zio
, zio_
, requeue_io_start_cut_in_line
, INT
, ZMOD_RW
,
4941 "Prioritize requeued I/O");
4943 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_deferred_free
, INT
, ZMOD_RW
,
4944 "Defer frees starting in this pass");
4946 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_dont_compress
, INT
, ZMOD_RW
,
4947 "Don't compress starting in this pass");
4949 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_rewrite
, INT
, ZMOD_RW
,
4950 "Rewrite new bps starting in this pass");
4952 ZFS_MODULE_PARAM(zfs_zio
, zio_
, dva_throttle_enabled
, INT
, ZMOD_RW
,
4953 "Throttle block allocations in the ZIO pipeline");
4955 ZFS_MODULE_PARAM(zfs_zio
, zio_
, deadman_log_all
, INT
, ZMOD_RW
,
4956 "Log all slow ZIOs, not just those with vdevs");