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.
26 * Copyright (c) 2019, Klara Inc.
27 * Copyright (c) 2019, Allan Jude
30 #include <sys/sysmacros.h>
31 #include <sys/zfs_context.h>
32 #include <sys/fm/fs/zfs.h>
35 #include <sys/spa_impl.h>
36 #include <sys/vdev_impl.h>
37 #include <sys/vdev_trim.h>
38 #include <sys/zio_impl.h>
39 #include <sys/zio_compress.h>
40 #include <sys/zio_checksum.h>
41 #include <sys/dmu_objset.h>
44 #include <sys/blkptr.h>
45 #include <sys/zfeature.h>
46 #include <sys/dsl_scan.h>
47 #include <sys/metaslab_impl.h>
49 #include <sys/trace_zfs.h>
51 #include <sys/dsl_crypt.h>
55 * ==========================================================================
56 * I/O type descriptions
57 * ==========================================================================
59 const char *zio_type_name
[ZIO_TYPES
] = {
61 * Note: Linux kernel thread name length is limited
62 * so these names will differ from upstream open zfs.
64 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
67 int zio_dva_throttle_enabled
= B_TRUE
;
68 int zio_deadman_log_all
= B_FALSE
;
71 * ==========================================================================
73 * ==========================================================================
75 kmem_cache_t
*zio_cache
;
76 kmem_cache_t
*zio_link_cache
;
77 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
79 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
80 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
81 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
84 /* Mark IOs as "slow" if they take longer than 30 seconds */
85 int zio_slow_io_ms
= (30 * MILLISEC
);
87 #define BP_SPANB(indblkshift, level) \
88 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
89 #define COMPARE_META_LEVEL 0x80000000ul
91 * The following actions directly effect the spa's sync-to-convergence logic.
92 * The values below define the sync pass when we start performing the action.
93 * Care should be taken when changing these values as they directly impact
94 * spa_sync() performance. Tuning these values may introduce subtle performance
95 * pathologies and should only be done in the context of performance analysis.
96 * These tunables will eventually be removed and replaced with #defines once
97 * enough analysis has been done to determine optimal values.
99 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
100 * regular blocks are not deferred.
102 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
103 * compression (including of metadata). In practice, we don't have this
104 * many sync passes, so this has no effect.
106 * The original intent was that disabling compression would help the sync
107 * passes to converge. However, in practice disabling compression increases
108 * the average number of sync passes, because when we turn compression off, a
109 * lot of block's size will change and thus we have to re-allocate (not
110 * overwrite) them. It also increases the number of 128KB allocations (e.g.
111 * for indirect blocks and spacemaps) because these will not be compressed.
112 * The 128K allocations are especially detrimental to performance on highly
113 * fragmented systems, which may have very few free segments of this size,
114 * and may need to load new metaslabs to satisfy 128K allocations.
116 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
117 int zfs_sync_pass_dont_compress
= 8; /* don't compress starting in this pass */
118 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
121 * An allocating zio is one that either currently has the DVA allocate
122 * stage set or will have it later in its lifetime.
124 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
127 * Enable smaller cores by excluding metadata
128 * allocations as well.
130 int zio_exclude_metadata
= 0;
131 int zio_requeue_io_start_cut_in_line
= 1;
134 int zio_buf_debug_limit
= 16384;
136 int zio_buf_debug_limit
= 0;
139 static inline void __zio_execute(zio_t
*zio
);
141 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
147 vmem_t
*data_alloc_arena
= NULL
;
149 zio_cache
= kmem_cache_create("zio_cache",
150 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
151 zio_link_cache
= kmem_cache_create("zio_link_cache",
152 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
155 * For small buffers, we want a cache for each multiple of
156 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
157 * for each quarter-power of 2.
159 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
160 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
163 size_t data_cflags
, cflags
;
165 data_cflags
= KMC_NODEBUG
;
166 cflags
= (zio_exclude_metadata
|| size
> zio_buf_debug_limit
) ?
169 #if defined(_ILP32) && defined(_KERNEL)
171 * Cache size limited to 1M on 32-bit platforms until ARC
172 * buffers no longer require virtual address space.
174 if (size
> zfs_max_recordsize
)
183 * If we are using watchpoints, put each buffer on its own page,
184 * to eliminate the performance overhead of trapping to the
185 * kernel when modifying a non-watched buffer that shares the
186 * page with a watched buffer.
188 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
191 * Here's the problem - on 4K native devices in userland on
192 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
193 * will fail with EINVAL, causing zdb (and others) to coredump.
194 * Since userland probably doesn't need optimized buffer caches,
195 * we just force 4K alignment on everything.
197 align
= 8 * SPA_MINBLOCKSIZE
;
199 if (size
< PAGESIZE
) {
200 align
= SPA_MINBLOCKSIZE
;
201 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
208 (void) snprintf(name
, sizeof (name
), "zio_buf_%lu",
210 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
211 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
213 (void) snprintf(name
, sizeof (name
), "zio_data_buf_%lu",
215 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
216 align
, NULL
, NULL
, NULL
, NULL
,
217 data_alloc_arena
, data_cflags
);
222 ASSERT(zio_buf_cache
[c
] != NULL
);
223 if (zio_buf_cache
[c
- 1] == NULL
)
224 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
226 ASSERT(zio_data_buf_cache
[c
] != NULL
);
227 if (zio_data_buf_cache
[c
- 1] == NULL
)
228 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
240 kmem_cache_t
*last_cache
= NULL
;
241 kmem_cache_t
*last_data_cache
= NULL
;
243 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
246 * Cache size limited to 1M on 32-bit platforms until ARC
247 * buffers no longer require virtual address space.
249 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
252 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
253 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
254 (void) printf("zio_fini: [%d] %llu != %llu\n",
255 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
256 (long long unsigned)zio_buf_cache_allocs
[c
],
257 (long long unsigned)zio_buf_cache_frees
[c
]);
259 if (zio_buf_cache
[c
] != last_cache
) {
260 last_cache
= zio_buf_cache
[c
];
261 kmem_cache_destroy(zio_buf_cache
[c
]);
263 zio_buf_cache
[c
] = NULL
;
265 if (zio_data_buf_cache
[c
] != last_data_cache
) {
266 last_data_cache
= zio_data_buf_cache
[c
];
267 kmem_cache_destroy(zio_data_buf_cache
[c
]);
269 zio_data_buf_cache
[c
] = NULL
;
272 kmem_cache_destroy(zio_link_cache
);
273 kmem_cache_destroy(zio_cache
);
281 * ==========================================================================
282 * Allocate and free I/O buffers
283 * ==========================================================================
287 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
288 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
289 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
290 * excess / transient data in-core during a crashdump.
293 zio_buf_alloc(size_t size
)
295 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
297 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
302 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
306 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
307 * crashdump if the kernel panics. This exists so that we will limit the amount
308 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
309 * of kernel heap dumped to disk when the kernel panics)
312 zio_data_buf_alloc(size_t size
)
314 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
316 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
318 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
322 zio_buf_free(void *buf
, size_t size
)
324 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
326 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
327 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
328 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
331 kmem_cache_free(zio_buf_cache
[c
], buf
);
335 zio_data_buf_free(void *buf
, size_t size
)
337 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
339 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
341 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
345 zio_abd_free(void *abd
, size_t size
)
347 abd_free((abd_t
*)abd
);
351 * ==========================================================================
352 * Push and pop I/O transform buffers
353 * ==========================================================================
356 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
357 zio_transform_func_t
*transform
)
359 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
361 zt
->zt_orig_abd
= zio
->io_abd
;
362 zt
->zt_orig_size
= zio
->io_size
;
363 zt
->zt_bufsize
= bufsize
;
364 zt
->zt_transform
= transform
;
366 zt
->zt_next
= zio
->io_transform_stack
;
367 zio
->io_transform_stack
= zt
;
374 zio_pop_transforms(zio_t
*zio
)
378 while ((zt
= zio
->io_transform_stack
) != NULL
) {
379 if (zt
->zt_transform
!= NULL
)
380 zt
->zt_transform(zio
,
381 zt
->zt_orig_abd
, zt
->zt_orig_size
);
383 if (zt
->zt_bufsize
!= 0)
384 abd_free(zio
->io_abd
);
386 zio
->io_abd
= zt
->zt_orig_abd
;
387 zio
->io_size
= zt
->zt_orig_size
;
388 zio
->io_transform_stack
= zt
->zt_next
;
390 kmem_free(zt
, sizeof (zio_transform_t
));
395 * ==========================================================================
396 * I/O transform callbacks for subblocks, decompression, and decryption
397 * ==========================================================================
400 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
402 ASSERT(zio
->io_size
> size
);
404 if (zio
->io_type
== ZIO_TYPE_READ
)
405 abd_copy(data
, zio
->io_abd
, size
);
409 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
411 if (zio
->io_error
== 0) {
412 void *tmp
= abd_borrow_buf(data
, size
);
413 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
414 zio
->io_abd
, tmp
, zio
->io_size
, size
,
415 &zio
->io_prop
.zp_complevel
);
416 abd_return_buf_copy(data
, tmp
, size
);
418 if (zio_injection_enabled
&& ret
== 0)
419 ret
= zio_handle_fault_injection(zio
, EINVAL
);
422 zio
->io_error
= SET_ERROR(EIO
);
427 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
431 blkptr_t
*bp
= zio
->io_bp
;
432 spa_t
*spa
= zio
->io_spa
;
433 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
434 uint64_t lsize
= BP_GET_LSIZE(bp
);
435 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
436 uint8_t salt
[ZIO_DATA_SALT_LEN
];
437 uint8_t iv
[ZIO_DATA_IV_LEN
];
438 uint8_t mac
[ZIO_DATA_MAC_LEN
];
439 boolean_t no_crypt
= B_FALSE
;
441 ASSERT(BP_USES_CRYPT(bp
));
442 ASSERT3U(size
, !=, 0);
444 if (zio
->io_error
!= 0)
448 * Verify the cksum of MACs stored in an indirect bp. It will always
449 * be possible to verify this since it does not require an encryption
452 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
453 zio_crypt_decode_mac_bp(bp
, mac
);
455 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
457 * We haven't decompressed the data yet, but
458 * zio_crypt_do_indirect_mac_checksum() requires
459 * decompressed data to be able to parse out the MACs
460 * from the indirect block. We decompress it now and
461 * throw away the result after we are finished.
463 tmp
= zio_buf_alloc(lsize
);
464 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
465 zio
->io_abd
, tmp
, zio
->io_size
, lsize
,
466 &zio
->io_prop
.zp_complevel
);
468 ret
= SET_ERROR(EIO
);
471 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
472 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
473 zio_buf_free(tmp
, lsize
);
475 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
476 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
478 abd_copy(data
, zio
->io_abd
, size
);
480 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
481 ret
= zio_handle_decrypt_injection(spa
,
482 &zio
->io_bookmark
, ot
, ECKSUM
);
491 * If this is an authenticated block, just check the MAC. It would be
492 * nice to separate this out into its own flag, but for the moment
493 * enum zio_flag is out of bits.
495 if (BP_IS_AUTHENTICATED(bp
)) {
496 if (ot
== DMU_OT_OBJSET
) {
497 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
498 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
500 zio_crypt_decode_mac_bp(bp
, mac
);
501 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
502 zio
->io_abd
, size
, mac
);
503 if (zio_injection_enabled
&& ret
== 0) {
504 ret
= zio_handle_decrypt_injection(spa
,
505 &zio
->io_bookmark
, ot
, ECKSUM
);
508 abd_copy(data
, zio
->io_abd
, size
);
516 zio_crypt_decode_params_bp(bp
, salt
, iv
);
518 if (ot
== DMU_OT_INTENT_LOG
) {
519 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
520 zio_crypt_decode_mac_zil(tmp
, mac
);
521 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
523 zio_crypt_decode_mac_bp(bp
, mac
);
526 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
527 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
528 zio
->io_abd
, &no_crypt
);
530 abd_copy(data
, zio
->io_abd
, size
);
538 /* assert that the key was found unless this was speculative */
539 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
542 * If there was a decryption / authentication error return EIO as
543 * the io_error. If this was not a speculative zio, create an ereport.
546 zio
->io_error
= SET_ERROR(EIO
);
547 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
548 spa_log_error(spa
, &zio
->io_bookmark
);
549 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
550 spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
558 * ==========================================================================
559 * I/O parent/child relationships and pipeline interlocks
560 * ==========================================================================
563 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
565 list_t
*pl
= &cio
->io_parent_list
;
567 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
571 ASSERT((*zl
)->zl_child
== cio
);
572 return ((*zl
)->zl_parent
);
576 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
578 list_t
*cl
= &pio
->io_child_list
;
580 ASSERT(MUTEX_HELD(&pio
->io_lock
));
582 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
586 ASSERT((*zl
)->zl_parent
== pio
);
587 return ((*zl
)->zl_child
);
591 zio_unique_parent(zio_t
*cio
)
593 zio_link_t
*zl
= NULL
;
594 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
596 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
601 zio_add_child(zio_t
*pio
, zio_t
*cio
)
603 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
606 * Logical I/Os can have logical, gang, or vdev children.
607 * Gang I/Os can have gang or vdev children.
608 * Vdev I/Os can only have vdev children.
609 * The following ASSERT captures all of these constraints.
611 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
616 mutex_enter(&pio
->io_lock
);
617 mutex_enter(&cio
->io_lock
);
619 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
621 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
622 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
624 list_insert_head(&pio
->io_child_list
, zl
);
625 list_insert_head(&cio
->io_parent_list
, zl
);
627 pio
->io_child_count
++;
628 cio
->io_parent_count
++;
630 mutex_exit(&cio
->io_lock
);
631 mutex_exit(&pio
->io_lock
);
635 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
637 ASSERT(zl
->zl_parent
== pio
);
638 ASSERT(zl
->zl_child
== cio
);
640 mutex_enter(&pio
->io_lock
);
641 mutex_enter(&cio
->io_lock
);
643 list_remove(&pio
->io_child_list
, zl
);
644 list_remove(&cio
->io_parent_list
, zl
);
646 pio
->io_child_count
--;
647 cio
->io_parent_count
--;
649 mutex_exit(&cio
->io_lock
);
650 mutex_exit(&pio
->io_lock
);
651 kmem_cache_free(zio_link_cache
, zl
);
655 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
657 boolean_t waiting
= B_FALSE
;
659 mutex_enter(&zio
->io_lock
);
660 ASSERT(zio
->io_stall
== NULL
);
661 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
662 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
665 uint64_t *countp
= &zio
->io_children
[c
][wait
];
668 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
669 zio
->io_stall
= countp
;
674 mutex_exit(&zio
->io_lock
);
678 __attribute__((always_inline
))
680 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
681 zio_t
**next_to_executep
)
683 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
684 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
686 mutex_enter(&pio
->io_lock
);
687 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
688 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
689 pio
->io_reexecute
|= zio
->io_reexecute
;
690 ASSERT3U(*countp
, >, 0);
694 if (*countp
== 0 && pio
->io_stall
== countp
) {
695 zio_taskq_type_t type
=
696 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
698 pio
->io_stall
= NULL
;
699 mutex_exit(&pio
->io_lock
);
702 * If we can tell the caller to execute this parent next, do
703 * so. Otherwise dispatch the parent zio as its own task.
705 * Having the caller execute the parent when possible reduces
706 * locking on the zio taskq's, reduces context switch
707 * overhead, and has no recursion penalty. Note that one
708 * read from disk typically causes at least 3 zio's: a
709 * zio_null(), the logical zio_read(), and then a physical
710 * zio. When the physical ZIO completes, we are able to call
711 * zio_done() on all 3 of these zio's from one invocation of
712 * zio_execute() by returning the parent back to
713 * zio_execute(). Since the parent isn't executed until this
714 * thread returns back to zio_execute(), the caller should do
717 * In other cases, dispatching the parent prevents
718 * overflowing the stack when we have deeply nested
719 * parent-child relationships, as we do with the "mega zio"
720 * of writes for spa_sync(), and the chain of ZIL blocks.
722 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
) {
723 *next_to_executep
= pio
;
725 zio_taskq_dispatch(pio
, type
, B_FALSE
);
728 mutex_exit(&pio
->io_lock
);
733 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
735 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
736 zio
->io_error
= zio
->io_child_error
[c
];
740 zio_bookmark_compare(const void *x1
, const void *x2
)
742 const zio_t
*z1
= x1
;
743 const zio_t
*z2
= x2
;
745 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
747 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
750 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
752 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
755 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
757 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
760 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
762 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
774 * ==========================================================================
775 * Create the various types of I/O (read, write, free, etc)
776 * ==========================================================================
779 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
780 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
781 void *private, zio_type_t type
, zio_priority_t priority
,
782 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
783 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
784 enum zio_stage pipeline
)
788 IMPLY(type
!= ZIO_TYPE_TRIM
, psize
<= SPA_MAXBLOCKSIZE
);
789 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
790 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
792 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
793 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
794 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
796 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
798 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
799 bzero(zio
, sizeof (zio_t
));
801 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
802 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
804 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
805 offsetof(zio_link_t
, zl_parent_node
));
806 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
807 offsetof(zio_link_t
, zl_child_node
));
808 metaslab_trace_init(&zio
->io_alloc_list
);
811 zio
->io_child_type
= ZIO_CHILD_VDEV
;
812 else if (flags
& ZIO_FLAG_GANG_CHILD
)
813 zio
->io_child_type
= ZIO_CHILD_GANG
;
814 else if (flags
& ZIO_FLAG_DDT_CHILD
)
815 zio
->io_child_type
= ZIO_CHILD_DDT
;
817 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
820 zio
->io_bp
= (blkptr_t
*)bp
;
821 zio
->io_bp_copy
= *bp
;
822 zio
->io_bp_orig
= *bp
;
823 if (type
!= ZIO_TYPE_WRITE
||
824 zio
->io_child_type
== ZIO_CHILD_DDT
)
825 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
826 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
827 zio
->io_logical
= zio
;
828 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
829 pipeline
|= ZIO_GANG_STAGES
;
835 zio
->io_private
= private;
837 zio
->io_priority
= priority
;
839 zio
->io_offset
= offset
;
840 zio
->io_orig_abd
= zio
->io_abd
= data
;
841 zio
->io_orig_size
= zio
->io_size
= psize
;
842 zio
->io_lsize
= lsize
;
843 zio
->io_orig_flags
= zio
->io_flags
= flags
;
844 zio
->io_orig_stage
= zio
->io_stage
= stage
;
845 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
846 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
848 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
849 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
852 zio
->io_bookmark
= *zb
;
855 if (zio
->io_metaslab_class
== NULL
)
856 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
857 if (zio
->io_logical
== NULL
)
858 zio
->io_logical
= pio
->io_logical
;
859 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
860 zio
->io_gang_leader
= pio
->io_gang_leader
;
861 zio_add_child(pio
, zio
);
864 taskq_init_ent(&zio
->io_tqent
);
870 zio_destroy(zio_t
*zio
)
872 metaslab_trace_fini(&zio
->io_alloc_list
);
873 list_destroy(&zio
->io_parent_list
);
874 list_destroy(&zio
->io_child_list
);
875 mutex_destroy(&zio
->io_lock
);
876 cv_destroy(&zio
->io_cv
);
877 kmem_cache_free(zio_cache
, zio
);
881 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
882 void *private, enum zio_flag flags
)
886 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
887 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
888 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
894 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
896 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
900 zfs_blkptr_verify_log(spa_t
*spa
, const blkptr_t
*bp
,
901 enum blk_verify_flag blk_verify
, const char *fmt
, ...)
907 (void) vsnprintf(buf
, sizeof (buf
), fmt
, adx
);
910 switch (blk_verify
) {
911 case BLK_VERIFY_HALT
:
912 dprintf_bp(bp
, "blkptr at %p dprintf_bp():", bp
);
913 zfs_panic_recover("%s: %s", spa_name(spa
), buf
);
916 zfs_dbgmsg("%s: %s", spa_name(spa
), buf
);
918 case BLK_VERIFY_ONLY
:
926 * Verify the block pointer fields contain reasonable values. This means
927 * it only contains known object types, checksum/compression identifiers,
928 * block sizes within the maximum allowed limits, valid DVAs, etc.
930 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
931 * argument controls the behavior when an invalid field is detected.
933 * Modes for zfs_blkptr_verify:
934 * 1) BLK_VERIFY_ONLY (evaluate the block)
935 * 2) BLK_VERIFY_LOG (evaluate the block and log problems)
936 * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
939 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
, boolean_t config_held
,
940 enum blk_verify_flag blk_verify
)
944 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
945 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
946 "blkptr at %p has invalid TYPE %llu",
947 bp
, (longlong_t
)BP_GET_TYPE(bp
));
949 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
950 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
951 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
952 "blkptr at %p has invalid CHECKSUM %llu",
953 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
955 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
956 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
957 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
958 "blkptr at %p has invalid COMPRESS %llu",
959 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
961 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
962 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
963 "blkptr at %p has invalid LSIZE %llu",
964 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
966 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
967 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
968 "blkptr at %p has invalid PSIZE %llu",
969 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
972 if (BP_IS_EMBEDDED(bp
)) {
973 if (BPE_GET_ETYPE(bp
) >= NUM_BP_EMBEDDED_TYPES
) {
974 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
975 "blkptr at %p has invalid ETYPE %llu",
976 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
981 * Do not verify individual DVAs if the config is not trusted. This
982 * will be done once the zio is executed in vdev_mirror_map_alloc.
984 if (!spa
->spa_trust_config
)
988 spa_config_enter(spa
, SCL_VDEV
, bp
, RW_READER
);
990 ASSERT(spa_config_held(spa
, SCL_VDEV
, RW_WRITER
));
992 * Pool-specific checks.
994 * Note: it would be nice to verify that the blk_birth and
995 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
996 * allows the birth time of log blocks (and dmu_sync()-ed blocks
997 * that are in the log) to be arbitrarily large.
999 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
1000 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
1002 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
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 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1010 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1011 "blkptr at %p DVA %u has invalid VDEV %llu",
1012 bp
, i
, (longlong_t
)vdevid
);
1015 if (vd
->vdev_ops
== &vdev_hole_ops
) {
1016 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1017 "blkptr at %p DVA %u has hole VDEV %llu",
1018 bp
, i
, (longlong_t
)vdevid
);
1021 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1023 * "missing" vdevs are valid during import, but we
1024 * don't have their detailed info (e.g. asize), so
1025 * we can't perform any more checks on them.
1029 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
1030 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
1032 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1033 if (offset
+ asize
> vd
->vdev_asize
) {
1034 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1035 "blkptr at %p DVA %u has invalid OFFSET %llu",
1036 bp
, i
, (longlong_t
)offset
);
1040 dprintf_bp(bp
, "blkptr at %p dprintf_bp():", bp
);
1042 spa_config_exit(spa
, SCL_VDEV
, bp
);
1044 return (errors
== 0);
1048 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
1050 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1052 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
1055 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1059 if (vd
->vdev_ops
== &vdev_hole_ops
)
1062 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1066 uint64_t offset
= DVA_GET_OFFSET(dva
);
1067 uint64_t asize
= DVA_GET_ASIZE(dva
);
1070 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1071 if (offset
+ asize
> vd
->vdev_asize
)
1078 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
1079 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
1080 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
1084 (void) zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
,
1087 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
1088 data
, size
, size
, done
, private,
1089 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1090 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1091 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1097 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1098 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1099 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1100 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
1101 void *private, zio_priority_t priority
, enum zio_flag flags
,
1102 const zbookmark_phys_t
*zb
)
1106 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1107 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1108 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1109 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1110 DMU_OT_IS_VALID(zp
->zp_type
) &&
1111 zp
->zp_level
< 32 &&
1112 zp
->zp_copies
> 0 &&
1113 zp
->zp_copies
<= spa_max_replication(spa
));
1115 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1116 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1117 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1118 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1120 zio
->io_ready
= ready
;
1121 zio
->io_children_ready
= children_ready
;
1122 zio
->io_physdone
= physdone
;
1126 * Data can be NULL if we are going to call zio_write_override() to
1127 * provide the already-allocated BP. But we may need the data to
1128 * verify a dedup hit (if requested). In this case, don't try to
1129 * dedup (just take the already-allocated BP verbatim). Encrypted
1130 * dedup blocks need data as well so we also disable dedup in this
1134 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1135 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1142 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1143 uint64_t size
, zio_done_func_t
*done
, void *private,
1144 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
1148 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1149 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1150 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1156 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1158 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1159 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1160 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1161 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1164 * We must reset the io_prop to match the values that existed
1165 * when the bp was first written by dmu_sync() keeping in mind
1166 * that nopwrite and dedup are mutually exclusive.
1168 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1169 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1170 zio
->io_prop
.zp_copies
= copies
;
1171 zio
->io_bp_override
= bp
;
1175 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1178 (void) zfs_blkptr_verify(spa
, bp
, B_FALSE
, BLK_VERIFY_HALT
);
1181 * The check for EMBEDDED is a performance optimization. We
1182 * process the free here (by ignoring it) rather than
1183 * putting it on the list and then processing it in zio_free_sync().
1185 if (BP_IS_EMBEDDED(bp
))
1187 metaslab_check_free(spa
, bp
);
1190 * Frees that are for the currently-syncing txg, are not going to be
1191 * deferred, and which will not need to do a read (i.e. not GANG or
1192 * DEDUP), can be processed immediately. Otherwise, put them on the
1193 * in-memory list for later processing.
1195 * Note that we only defer frees after zfs_sync_pass_deferred_free
1196 * when the log space map feature is disabled. [see relevant comment
1197 * in spa_sync_iterate_to_convergence()]
1199 if (BP_IS_GANG(bp
) ||
1201 txg
!= spa
->spa_syncing_txg
||
1202 (spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
&&
1203 !spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
))) {
1204 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1206 VERIFY3P(zio_free_sync(NULL
, spa
, txg
, bp
, 0), ==, NULL
);
1211 * To improve performance, this function may return NULL if we were able
1212 * to do the free immediately. This avoids the cost of creating a zio
1213 * (and linking it to the parent, etc).
1216 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1217 enum zio_flag flags
)
1219 ASSERT(!BP_IS_HOLE(bp
));
1220 ASSERT(spa_syncing_txg(spa
) == txg
);
1222 if (BP_IS_EMBEDDED(bp
))
1225 metaslab_check_free(spa
, bp
);
1227 dsl_scan_freed(spa
, bp
);
1229 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
)) {
1231 * GANG and DEDUP blocks can induce a read (for the gang block
1232 * header, or the DDT), so issue them asynchronously so that
1233 * this thread is not tied up.
1235 enum zio_stage stage
=
1236 ZIO_FREE_PIPELINE
| ZIO_STAGE_ISSUE_ASYNC
;
1238 return (zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1239 BP_GET_PSIZE(bp
), NULL
, NULL
,
1240 ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1241 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
));
1243 metaslab_free(spa
, bp
, txg
, B_FALSE
);
1249 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1250 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1254 (void) zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
,
1257 if (BP_IS_EMBEDDED(bp
))
1258 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1261 * A claim is an allocation of a specific block. Claims are needed
1262 * to support immediate writes in the intent log. The issue is that
1263 * immediate writes contain committed data, but in a txg that was
1264 * *not* committed. Upon opening the pool after an unclean shutdown,
1265 * the intent log claims all blocks that contain immediate write data
1266 * so that the SPA knows they're in use.
1268 * All claims *must* be resolved in the first txg -- before the SPA
1269 * starts allocating blocks -- so that nothing is allocated twice.
1270 * If txg == 0 we just verify that the block is claimable.
1272 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
1273 spa_min_claim_txg(spa
));
1274 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1275 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1277 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1278 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1279 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1280 ASSERT0(zio
->io_queued_timestamp
);
1286 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1287 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1292 if (vd
->vdev_children
== 0) {
1293 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1294 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1295 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1299 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1301 for (c
= 0; c
< vd
->vdev_children
; c
++)
1302 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1303 done
, private, flags
));
1310 zio_trim(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1311 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1312 enum zio_flag flags
, enum trim_flag trim_flags
)
1316 ASSERT0(vd
->vdev_children
);
1317 ASSERT0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
1318 ASSERT0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
1319 ASSERT3U(size
, !=, 0);
1321 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, NULL
, size
, size
, done
,
1322 private, ZIO_TYPE_TRIM
, priority
, flags
| ZIO_FLAG_PHYSICAL
,
1323 vd
, offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_TRIM_PIPELINE
);
1324 zio
->io_trim_flags
= trim_flags
;
1330 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1331 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1332 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1336 ASSERT(vd
->vdev_children
== 0);
1337 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1338 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1339 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1341 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1342 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1343 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1345 zio
->io_prop
.zp_checksum
= checksum
;
1351 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1352 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1353 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1357 ASSERT(vd
->vdev_children
== 0);
1358 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1359 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1360 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1362 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1363 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1364 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1366 zio
->io_prop
.zp_checksum
= checksum
;
1368 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1370 * zec checksums are necessarily destructive -- they modify
1371 * the end of the write buffer to hold the verifier/checksum.
1372 * Therefore, we must make a local copy in case the data is
1373 * being written to multiple places in parallel.
1375 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1376 abd_copy(wbuf
, data
, size
);
1378 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1385 * Create a child I/O to do some work for us.
1388 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1389 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1390 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1392 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1396 * vdev child I/Os do not propagate their error to the parent.
1397 * Therefore, for correct operation the caller *must* check for
1398 * and handle the error in the child i/o's done callback.
1399 * The only exceptions are i/os that we don't care about
1400 * (OPTIONAL or REPAIR).
1402 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1405 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1407 * If we have the bp, then the child should perform the
1408 * checksum and the parent need not. This pushes error
1409 * detection as close to the leaves as possible and
1410 * eliminates redundant checksums in the interior nodes.
1412 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1413 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1416 if (vd
->vdev_ops
->vdev_op_leaf
) {
1417 ASSERT0(vd
->vdev_children
);
1418 offset
+= VDEV_LABEL_START_SIZE
;
1421 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1424 * If we've decided to do a repair, the write is not speculative --
1425 * even if the original read was.
1427 if (flags
& ZIO_FLAG_IO_REPAIR
)
1428 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1431 * If we're creating a child I/O that is not associated with a
1432 * top-level vdev, then the child zio is not an allocating I/O.
1433 * If this is a retried I/O then we ignore it since we will
1434 * have already processed the original allocating I/O.
1436 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1437 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1438 ASSERT(pio
->io_metaslab_class
!= NULL
);
1439 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1440 ASSERT(type
== ZIO_TYPE_WRITE
);
1441 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1442 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1443 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1444 pio
->io_child_type
== ZIO_CHILD_GANG
);
1446 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1450 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1451 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1452 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1453 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1455 zio
->io_physdone
= pio
->io_physdone
;
1456 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1457 zio
->io_logical
->io_phys_children
++;
1463 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1464 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1465 zio_done_func_t
*done
, void *private)
1469 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1471 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1472 data
, size
, size
, done
, private, type
, priority
,
1473 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1475 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1481 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1483 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1485 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1489 zio_shrink(zio_t
*zio
, uint64_t size
)
1491 ASSERT3P(zio
->io_executor
, ==, NULL
);
1492 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1493 ASSERT3U(size
, <=, zio
->io_size
);
1496 * We don't shrink for raidz because of problems with the
1497 * reconstruction when reading back less than the block size.
1498 * Note, BP_IS_RAIDZ() assumes no compression.
1500 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1501 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1502 /* we are not doing a raw write */
1503 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1504 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1509 * ==========================================================================
1510 * Prepare to read and write logical blocks
1511 * ==========================================================================
1515 zio_read_bp_init(zio_t
*zio
)
1517 blkptr_t
*bp
= zio
->io_bp
;
1519 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1521 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1523 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1524 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1525 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1526 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1527 psize
, psize
, zio_decompress
);
1530 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1531 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1532 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1533 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1534 psize
, psize
, zio_decrypt
);
1537 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1538 int psize
= BPE_GET_PSIZE(bp
);
1539 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1541 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1542 decode_embedded_bp_compressed(bp
, data
);
1543 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1545 ASSERT(!BP_IS_EMBEDDED(bp
));
1546 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1549 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1550 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1552 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1553 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1555 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1556 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1562 zio_write_bp_init(zio_t
*zio
)
1564 if (!IO_IS_ALLOCATING(zio
))
1567 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1569 if (zio
->io_bp_override
) {
1570 blkptr_t
*bp
= zio
->io_bp
;
1571 zio_prop_t
*zp
= &zio
->io_prop
;
1573 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1574 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1576 *bp
= *zio
->io_bp_override
;
1577 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1579 if (BP_IS_EMBEDDED(bp
))
1583 * If we've been overridden and nopwrite is set then
1584 * set the flag accordingly to indicate that a nopwrite
1585 * has already occurred.
1587 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1588 ASSERT(!zp
->zp_dedup
);
1589 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1590 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1594 ASSERT(!zp
->zp_nopwrite
);
1596 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1599 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1600 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1602 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1604 BP_SET_DEDUP(bp
, 1);
1605 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1610 * We were unable to handle this as an override bp, treat
1611 * it as a regular write I/O.
1613 zio
->io_bp_override
= NULL
;
1614 *bp
= zio
->io_bp_orig
;
1615 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1622 zio_write_compress(zio_t
*zio
)
1624 spa_t
*spa
= zio
->io_spa
;
1625 zio_prop_t
*zp
= &zio
->io_prop
;
1626 enum zio_compress compress
= zp
->zp_compress
;
1627 blkptr_t
*bp
= zio
->io_bp
;
1628 uint64_t lsize
= zio
->io_lsize
;
1629 uint64_t psize
= zio
->io_size
;
1633 * If our children haven't all reached the ready stage,
1634 * wait for them and then repeat this pipeline stage.
1636 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1637 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1641 if (!IO_IS_ALLOCATING(zio
))
1644 if (zio
->io_children_ready
!= NULL
) {
1646 * Now that all our children are ready, run the callback
1647 * associated with this zio in case it wants to modify the
1648 * data to be written.
1650 ASSERT3U(zp
->zp_level
, >, 0);
1651 zio
->io_children_ready(zio
);
1654 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1655 ASSERT(zio
->io_bp_override
== NULL
);
1657 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1659 * We're rewriting an existing block, which means we're
1660 * working on behalf of spa_sync(). For spa_sync() to
1661 * converge, it must eventually be the case that we don't
1662 * have to allocate new blocks. But compression changes
1663 * the blocksize, which forces a reallocate, and makes
1664 * convergence take longer. Therefore, after the first
1665 * few passes, stop compressing to ensure convergence.
1667 pass
= spa_sync_pass(spa
);
1669 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1670 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1671 ASSERT(!BP_GET_DEDUP(bp
));
1673 if (pass
>= zfs_sync_pass_dont_compress
)
1674 compress
= ZIO_COMPRESS_OFF
;
1676 /* Make sure someone doesn't change their mind on overwrites */
1677 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1678 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1681 /* If it's a compressed write that is not raw, compress the buffer. */
1682 if (compress
!= ZIO_COMPRESS_OFF
&&
1683 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1684 void *cbuf
= zio_buf_alloc(lsize
);
1685 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
,
1687 if (psize
== 0 || psize
>= lsize
) {
1688 compress
= ZIO_COMPRESS_OFF
;
1689 zio_buf_free(cbuf
, lsize
);
1690 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1691 psize
<= BPE_PAYLOAD_SIZE
&&
1692 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1693 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1694 encode_embedded_bp_compressed(bp
,
1695 cbuf
, compress
, lsize
, psize
);
1696 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1697 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1698 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1699 zio_buf_free(cbuf
, lsize
);
1700 bp
->blk_birth
= zio
->io_txg
;
1701 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1702 ASSERT(spa_feature_is_active(spa
,
1703 SPA_FEATURE_EMBEDDED_DATA
));
1707 * Round up compressed size up to the ashift
1708 * of the smallest-ashift device, and zero the tail.
1709 * This ensures that the compressed size of the BP
1710 * (and thus compressratio property) are correct,
1711 * in that we charge for the padding used to fill out
1714 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1715 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1716 1ULL << spa
->spa_min_ashift
);
1717 if (rounded
>= lsize
) {
1718 compress
= ZIO_COMPRESS_OFF
;
1719 zio_buf_free(cbuf
, lsize
);
1722 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1723 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1724 abd_zero_off(cdata
, psize
, rounded
- psize
);
1726 zio_push_transform(zio
, cdata
,
1727 psize
, lsize
, NULL
);
1732 * We were unable to handle this as an override bp, treat
1733 * it as a regular write I/O.
1735 zio
->io_bp_override
= NULL
;
1736 *bp
= zio
->io_bp_orig
;
1737 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1739 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1740 zp
->zp_type
== DMU_OT_DNODE
) {
1742 * The DMU actually relies on the zio layer's compression
1743 * to free metadnode blocks that have had all contained
1744 * dnodes freed. As a result, even when doing a raw
1745 * receive, we must check whether the block can be compressed
1748 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1749 zio
->io_abd
, NULL
, lsize
, zp
->zp_complevel
);
1750 if (psize
== 0 || psize
>= lsize
)
1751 compress
= ZIO_COMPRESS_OFF
;
1753 ASSERT3U(psize
, !=, 0);
1757 * The final pass of spa_sync() must be all rewrites, but the first
1758 * few passes offer a trade-off: allocating blocks defers convergence,
1759 * but newly allocated blocks are sequential, so they can be written
1760 * to disk faster. Therefore, we allow the first few passes of
1761 * spa_sync() to allocate new blocks, but force rewrites after that.
1762 * There should only be a handful of blocks after pass 1 in any case.
1764 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1765 BP_GET_PSIZE(bp
) == psize
&&
1766 pass
>= zfs_sync_pass_rewrite
) {
1767 VERIFY3U(psize
, !=, 0);
1768 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1770 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1771 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1774 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1778 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1779 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1780 BP_SET_LSIZE(bp
, lsize
);
1781 BP_SET_TYPE(bp
, zp
->zp_type
);
1782 BP_SET_LEVEL(bp
, zp
->zp_level
);
1783 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1785 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1787 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1788 BP_SET_LSIZE(bp
, lsize
);
1789 BP_SET_TYPE(bp
, zp
->zp_type
);
1790 BP_SET_LEVEL(bp
, zp
->zp_level
);
1791 BP_SET_PSIZE(bp
, psize
);
1792 BP_SET_COMPRESS(bp
, compress
);
1793 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1794 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1795 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1797 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1798 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1799 ASSERT(!zp
->zp_encrypt
||
1800 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1801 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1803 if (zp
->zp_nopwrite
) {
1804 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1805 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1806 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1813 zio_free_bp_init(zio_t
*zio
)
1815 blkptr_t
*bp
= zio
->io_bp
;
1817 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1818 if (BP_GET_DEDUP(bp
))
1819 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1822 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1828 * ==========================================================================
1829 * Execute the I/O pipeline
1830 * ==========================================================================
1834 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1836 spa_t
*spa
= zio
->io_spa
;
1837 zio_type_t t
= zio
->io_type
;
1838 int flags
= (cutinline
? TQ_FRONT
: 0);
1841 * If we're a config writer or a probe, the normal issue and
1842 * interrupt threads may all be blocked waiting for the config lock.
1843 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1845 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1849 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1851 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1855 * If this is a high priority I/O, then use the high priority taskq if
1858 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1859 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1860 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1863 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1866 * NB: We are assuming that the zio can only be dispatched
1867 * to a single taskq at a time. It would be a grievous error
1868 * to dispatch the zio to another taskq at the same time.
1870 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1871 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1872 flags
, &zio
->io_tqent
);
1876 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1878 spa_t
*spa
= zio
->io_spa
;
1880 taskq_t
*tq
= taskq_of_curthread();
1882 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1883 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1885 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1886 if (tqs
->stqs_taskq
[i
] == tq
)
1895 zio_issue_async(zio_t
*zio
)
1897 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1903 zio_interrupt(zio_t
*zio
)
1905 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1909 zio_delay_interrupt(zio_t
*zio
)
1912 * The timeout_generic() function isn't defined in userspace, so
1913 * rather than trying to implement the function, the zio delay
1914 * functionality has been disabled for userspace builds.
1919 * If io_target_timestamp is zero, then no delay has been registered
1920 * for this IO, thus jump to the end of this function and "skip" the
1921 * delay; issuing it directly to the zio layer.
1923 if (zio
->io_target_timestamp
!= 0) {
1924 hrtime_t now
= gethrtime();
1926 if (now
>= zio
->io_target_timestamp
) {
1928 * This IO has already taken longer than the target
1929 * delay to complete, so we don't want to delay it
1930 * any longer; we "miss" the delay and issue it
1931 * directly to the zio layer. This is likely due to
1932 * the target latency being set to a value less than
1933 * the underlying hardware can satisfy (e.g. delay
1934 * set to 1ms, but the disks take 10ms to complete an
1938 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1944 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1945 clock_t expire_at_tick
= ddi_get_lbolt() +
1948 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1949 hrtime_t
, now
, hrtime_t
, diff
);
1951 if (NSEC_TO_TICK(diff
) == 0) {
1952 /* Our delay is less than a jiffy - just spin */
1953 zfs_sleep_until(zio
->io_target_timestamp
);
1957 * Use taskq_dispatch_delay() in the place of
1958 * OpenZFS's timeout_generic().
1960 tid
= taskq_dispatch_delay(system_taskq
,
1961 (task_func_t
*)zio_interrupt
,
1962 zio
, TQ_NOSLEEP
, expire_at_tick
);
1963 if (tid
== TASKQID_INVALID
) {
1965 * Couldn't allocate a task. Just
1966 * finish the zio without a delay.
1975 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1980 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
1982 zio_t
*cio
, *cio_next
;
1983 zio_link_t
*zl
= NULL
;
1984 vdev_t
*vd
= pio
->io_vd
;
1986 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
1987 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
1988 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1989 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1990 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1992 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
1993 "delta=%llu queued=%llu io=%llu "
1994 "path=%s last=%llu "
1995 "type=%d priority=%d flags=0x%x "
1996 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1997 "objset=%llu object=%llu level=%llu blkid=%llu "
1998 "offset=%llu size=%llu error=%d",
1999 ziodepth
, pio
, pio
->io_timestamp
,
2000 delta
, pio
->io_delta
, pio
->io_delay
,
2001 vd
? vd
->vdev_path
: "NULL", vq
? vq
->vq_io_complete_ts
: 0,
2002 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
2003 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
2004 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
2005 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
2006 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
2007 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
2009 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
2010 taskq_empty_ent(&pio
->io_tqent
)) {
2015 mutex_enter(&pio
->io_lock
);
2016 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2017 cio_next
= zio_walk_children(pio
, &zl
);
2018 zio_deadman_impl(cio
, ziodepth
+ 1);
2020 mutex_exit(&pio
->io_lock
);
2024 * Log the critical information describing this zio and all of its children
2025 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2028 zio_deadman(zio_t
*pio
, char *tag
)
2030 spa_t
*spa
= pio
->io_spa
;
2031 char *name
= spa_name(spa
);
2033 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
2036 zio_deadman_impl(pio
, 0);
2038 switch (spa_get_deadman_failmode(spa
)) {
2039 case ZIO_FAILURE_MODE_WAIT
:
2040 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
2043 case ZIO_FAILURE_MODE_CONTINUE
:
2044 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
2047 case ZIO_FAILURE_MODE_PANIC
:
2048 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
2054 * Execute the I/O pipeline until one of the following occurs:
2055 * (1) the I/O completes; (2) the pipeline stalls waiting for
2056 * dependent child I/Os; (3) the I/O issues, so we're waiting
2057 * for an I/O completion interrupt; (4) the I/O is delegated by
2058 * vdev-level caching or aggregation; (5) the I/O is deferred
2059 * due to vdev-level queueing; (6) the I/O is handed off to
2060 * another thread. In all cases, the pipeline stops whenever
2061 * there's no CPU work; it never burns a thread in cv_wait_io().
2063 * There's no locking on io_stage because there's no legitimate way
2064 * for multiple threads to be attempting to process the same I/O.
2066 static zio_pipe_stage_t
*zio_pipeline
[];
2069 * zio_execute() is a wrapper around the static function
2070 * __zio_execute() so that we can force __zio_execute() to be
2071 * inlined. This reduces stack overhead which is important
2072 * because __zio_execute() is called recursively in several zio
2073 * code paths. zio_execute() itself cannot be inlined because
2074 * it is externally visible.
2077 zio_execute(zio_t
*zio
)
2079 fstrans_cookie_t cookie
;
2081 cookie
= spl_fstrans_mark();
2083 spl_fstrans_unmark(cookie
);
2087 * Used to determine if in the current context the stack is sized large
2088 * enough to allow zio_execute() to be called recursively. A minimum
2089 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2092 zio_execute_stack_check(zio_t
*zio
)
2094 #if !defined(HAVE_LARGE_STACKS)
2095 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
2097 /* Executing in txg_sync_thread() context. */
2098 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
2101 /* Pool initialization outside of zio_taskq context. */
2102 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
2103 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
2104 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
2106 #endif /* HAVE_LARGE_STACKS */
2111 __attribute__((always_inline
))
2113 __zio_execute(zio_t
*zio
)
2115 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2117 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
2118 enum zio_stage pipeline
= zio
->io_pipeline
;
2119 enum zio_stage stage
= zio
->io_stage
;
2121 zio
->io_executor
= curthread
;
2123 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
2124 ASSERT(ISP2(stage
));
2125 ASSERT(zio
->io_stall
== NULL
);
2129 } while ((stage
& pipeline
) == 0);
2131 ASSERT(stage
<= ZIO_STAGE_DONE
);
2134 * If we are in interrupt context and this pipeline stage
2135 * will grab a config lock that is held across I/O,
2136 * or may wait for an I/O that needs an interrupt thread
2137 * to complete, issue async to avoid deadlock.
2139 * For VDEV_IO_START, we cut in line so that the io will
2140 * be sent to disk promptly.
2142 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
2143 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
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
);
2151 * If the current context doesn't have large enough stacks
2152 * the zio must be issued asynchronously to prevent overflow.
2154 if (zio_execute_stack_check(zio
)) {
2155 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2156 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2157 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2161 zio
->io_stage
= stage
;
2162 zio
->io_pipeline_trace
|= zio
->io_stage
;
2165 * The zio pipeline stage returns the next zio to execute
2166 * (typically the same as this one), or NULL if we should
2169 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2178 * ==========================================================================
2179 * Initiate I/O, either sync or async
2180 * ==========================================================================
2183 zio_wait(zio_t
*zio
)
2186 * Some routines, like zio_free_sync(), may return a NULL zio
2187 * to avoid the performance overhead of creating and then destroying
2188 * an unneeded zio. For the callers' simplicity, we accept a NULL
2189 * zio and ignore it.
2194 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2197 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2198 ASSERT3P(zio
->io_executor
, ==, NULL
);
2200 zio
->io_waiter
= curthread
;
2201 ASSERT0(zio
->io_queued_timestamp
);
2202 zio
->io_queued_timestamp
= gethrtime();
2206 mutex_enter(&zio
->io_lock
);
2207 while (zio
->io_executor
!= NULL
) {
2208 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2209 ddi_get_lbolt() + timeout
);
2211 if (zfs_deadman_enabled
&& error
== -1 &&
2212 gethrtime() - zio
->io_queued_timestamp
>
2213 spa_deadman_ziotime(zio
->io_spa
)) {
2214 mutex_exit(&zio
->io_lock
);
2215 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2216 zio_deadman(zio
, FTAG
);
2217 mutex_enter(&zio
->io_lock
);
2220 mutex_exit(&zio
->io_lock
);
2222 error
= zio
->io_error
;
2229 zio_nowait(zio_t
*zio
)
2232 * See comment in zio_wait().
2237 ASSERT3P(zio
->io_executor
, ==, NULL
);
2239 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2240 zio_unique_parent(zio
) == NULL
) {
2244 * This is a logical async I/O with no parent to wait for it.
2245 * We add it to the spa_async_root_zio "Godfather" I/O which
2246 * will ensure they complete prior to unloading the pool.
2248 spa_t
*spa
= zio
->io_spa
;
2250 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2253 zio_add_child(pio
, zio
);
2256 ASSERT0(zio
->io_queued_timestamp
);
2257 zio
->io_queued_timestamp
= gethrtime();
2262 * ==========================================================================
2263 * Reexecute, cancel, or suspend/resume failed I/O
2264 * ==========================================================================
2268 zio_reexecute(zio_t
*pio
)
2270 zio_t
*cio
, *cio_next
;
2272 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2273 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2274 ASSERT(pio
->io_gang_leader
== NULL
);
2275 ASSERT(pio
->io_gang_tree
== NULL
);
2277 pio
->io_flags
= pio
->io_orig_flags
;
2278 pio
->io_stage
= pio
->io_orig_stage
;
2279 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2280 pio
->io_reexecute
= 0;
2281 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2282 pio
->io_pipeline_trace
= 0;
2284 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2285 pio
->io_state
[w
] = 0;
2286 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2287 pio
->io_child_error
[c
] = 0;
2289 if (IO_IS_ALLOCATING(pio
))
2290 BP_ZERO(pio
->io_bp
);
2293 * As we reexecute pio's children, new children could be created.
2294 * New children go to the head of pio's io_child_list, however,
2295 * so we will (correctly) not reexecute them. The key is that
2296 * the remainder of pio's io_child_list, from 'cio_next' onward,
2297 * cannot be affected by any side effects of reexecuting 'cio'.
2299 zio_link_t
*zl
= NULL
;
2300 mutex_enter(&pio
->io_lock
);
2301 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2302 cio_next
= zio_walk_children(pio
, &zl
);
2303 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2304 pio
->io_children
[cio
->io_child_type
][w
]++;
2305 mutex_exit(&pio
->io_lock
);
2307 mutex_enter(&pio
->io_lock
);
2309 mutex_exit(&pio
->io_lock
);
2312 * Now that all children have been reexecuted, execute the parent.
2313 * We don't reexecute "The Godfather" I/O here as it's the
2314 * responsibility of the caller to wait on it.
2316 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2317 pio
->io_queued_timestamp
= gethrtime();
2323 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2325 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2326 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2327 "failure and the failure mode property for this pool "
2328 "is set to panic.", spa_name(spa
));
2330 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2331 "failure and has been suspended.\n", spa_name(spa
));
2333 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2336 mutex_enter(&spa
->spa_suspend_lock
);
2338 if (spa
->spa_suspend_zio_root
== NULL
)
2339 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2340 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2341 ZIO_FLAG_GODFATHER
);
2343 spa
->spa_suspended
= reason
;
2346 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2347 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2348 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2349 ASSERT(zio_unique_parent(zio
) == NULL
);
2350 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2351 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2354 mutex_exit(&spa
->spa_suspend_lock
);
2358 zio_resume(spa_t
*spa
)
2363 * Reexecute all previously suspended i/o.
2365 mutex_enter(&spa
->spa_suspend_lock
);
2366 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2367 cv_broadcast(&spa
->spa_suspend_cv
);
2368 pio
= spa
->spa_suspend_zio_root
;
2369 spa
->spa_suspend_zio_root
= NULL
;
2370 mutex_exit(&spa
->spa_suspend_lock
);
2376 return (zio_wait(pio
));
2380 zio_resume_wait(spa_t
*spa
)
2382 mutex_enter(&spa
->spa_suspend_lock
);
2383 while (spa_suspended(spa
))
2384 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2385 mutex_exit(&spa
->spa_suspend_lock
);
2389 * ==========================================================================
2392 * A gang block is a collection of small blocks that looks to the DMU
2393 * like one large block. When zio_dva_allocate() cannot find a block
2394 * of the requested size, due to either severe fragmentation or the pool
2395 * being nearly full, it calls zio_write_gang_block() to construct the
2396 * block from smaller fragments.
2398 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2399 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2400 * an indirect block: it's an array of block pointers. It consumes
2401 * only one sector and hence is allocatable regardless of fragmentation.
2402 * The gang header's bps point to its gang members, which hold the data.
2404 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2405 * as the verifier to ensure uniqueness of the SHA256 checksum.
2406 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2407 * not the gang header. This ensures that data block signatures (needed for
2408 * deduplication) are independent of how the block is physically stored.
2410 * Gang blocks can be nested: a gang member may itself be a gang block.
2411 * Thus every gang block is a tree in which root and all interior nodes are
2412 * gang headers, and the leaves are normal blocks that contain user data.
2413 * The root of the gang tree is called the gang leader.
2415 * To perform any operation (read, rewrite, free, claim) on a gang block,
2416 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2417 * in the io_gang_tree field of the original logical i/o by recursively
2418 * reading the gang leader and all gang headers below it. This yields
2419 * an in-core tree containing the contents of every gang header and the
2420 * bps for every constituent of the gang block.
2422 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2423 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2424 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2425 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2426 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2427 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2428 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2429 * of the gang header plus zio_checksum_compute() of the data to update the
2430 * gang header's blk_cksum as described above.
2432 * The two-phase assemble/issue model solves the problem of partial failure --
2433 * what if you'd freed part of a gang block but then couldn't read the
2434 * gang header for another part? Assembling the entire gang tree first
2435 * ensures that all the necessary gang header I/O has succeeded before
2436 * starting the actual work of free, claim, or write. Once the gang tree
2437 * is assembled, free and claim are in-memory operations that cannot fail.
2439 * In the event that a gang write fails, zio_dva_unallocate() walks the
2440 * gang tree to immediately free (i.e. insert back into the space map)
2441 * everything we've allocated. This ensures that we don't get ENOSPC
2442 * errors during repeated suspend/resume cycles due to a flaky device.
2444 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2445 * the gang tree, we won't modify the block, so we can safely defer the free
2446 * (knowing that the block is still intact). If we *can* assemble the gang
2447 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2448 * each constituent bp and we can allocate a new block on the next sync pass.
2450 * In all cases, the gang tree allows complete recovery from partial failure.
2451 * ==========================================================================
2455 zio_gang_issue_func_done(zio_t
*zio
)
2457 abd_put(zio
->io_abd
);
2461 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2467 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2468 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2469 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2470 &pio
->io_bookmark
));
2474 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2481 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2482 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2483 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2484 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2487 * As we rewrite each gang header, the pipeline will compute
2488 * a new gang block header checksum for it; but no one will
2489 * compute a new data checksum, so we do that here. The one
2490 * exception is the gang leader: the pipeline already computed
2491 * its data checksum because that stage precedes gang assembly.
2492 * (Presently, nothing actually uses interior data checksums;
2493 * this is just good hygiene.)
2495 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2496 abd_t
*buf
= abd_get_offset(data
, offset
);
2498 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2499 buf
, BP_GET_PSIZE(bp
));
2504 * If we are here to damage data for testing purposes,
2505 * leave the GBH alone so that we can detect the damage.
2507 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2508 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2510 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2511 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2512 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2513 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2521 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2524 zio_t
*zio
= zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2525 ZIO_GANG_CHILD_FLAGS(pio
));
2527 zio
= zio_null(pio
, pio
->io_spa
,
2528 NULL
, NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
));
2535 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2538 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2539 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2542 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2551 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2553 static zio_gang_node_t
*
2554 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2556 zio_gang_node_t
*gn
;
2558 ASSERT(*gnpp
== NULL
);
2560 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2561 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2568 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2570 zio_gang_node_t
*gn
= *gnpp
;
2572 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2573 ASSERT(gn
->gn_child
[g
] == NULL
);
2575 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2576 kmem_free(gn
, sizeof (*gn
));
2581 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2583 zio_gang_node_t
*gn
= *gnpp
;
2588 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2589 zio_gang_tree_free(&gn
->gn_child
[g
]);
2591 zio_gang_node_free(gnpp
);
2595 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2597 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2598 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2600 ASSERT(gio
->io_gang_leader
== gio
);
2601 ASSERT(BP_IS_GANG(bp
));
2603 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2604 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2605 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2609 zio_gang_tree_assemble_done(zio_t
*zio
)
2611 zio_t
*gio
= zio
->io_gang_leader
;
2612 zio_gang_node_t
*gn
= zio
->io_private
;
2613 blkptr_t
*bp
= zio
->io_bp
;
2615 ASSERT(gio
== zio_unique_parent(zio
));
2616 ASSERT(zio
->io_child_count
== 0);
2621 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2622 if (BP_SHOULD_BYTESWAP(bp
))
2623 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2625 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2626 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2627 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2629 abd_put(zio
->io_abd
);
2631 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2632 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2633 if (!BP_IS_GANG(gbp
))
2635 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2640 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2643 zio_t
*gio
= pio
->io_gang_leader
;
2646 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2647 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2648 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2651 * If you're a gang header, your data is in gn->gn_gbh.
2652 * If you're a gang member, your data is in 'data' and gn == NULL.
2654 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2657 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2659 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2660 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2661 if (BP_IS_HOLE(gbp
))
2663 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2665 offset
+= BP_GET_PSIZE(gbp
);
2669 if (gn
== gio
->io_gang_tree
)
2670 ASSERT3U(gio
->io_size
, ==, offset
);
2677 zio_gang_assemble(zio_t
*zio
)
2679 blkptr_t
*bp
= zio
->io_bp
;
2681 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2682 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2684 zio
->io_gang_leader
= zio
;
2686 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2692 zio_gang_issue(zio_t
*zio
)
2694 blkptr_t
*bp
= zio
->io_bp
;
2696 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2700 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2701 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2703 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2704 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2707 zio_gang_tree_free(&zio
->io_gang_tree
);
2709 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2715 zio_write_gang_member_ready(zio_t
*zio
)
2717 zio_t
*pio
= zio_unique_parent(zio
);
2718 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2719 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2721 zio_t
*gio __maybe_unused
= zio
->io_gang_leader
;
2723 if (BP_IS_HOLE(zio
->io_bp
))
2726 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2728 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2729 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2730 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2731 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2732 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2734 mutex_enter(&pio
->io_lock
);
2735 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2736 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2737 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2738 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2739 DVA_SET_ASIZE(&pdva
[d
], asize
);
2741 mutex_exit(&pio
->io_lock
);
2745 zio_write_gang_done(zio_t
*zio
)
2748 * The io_abd field will be NULL for a zio with no data. The io_flags
2749 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2750 * check for it here as it is cleared in zio_ready.
2752 if (zio
->io_abd
!= NULL
)
2753 abd_put(zio
->io_abd
);
2757 zio_write_gang_block(zio_t
*pio
)
2759 spa_t
*spa
= pio
->io_spa
;
2760 metaslab_class_t
*mc
= spa_normal_class(spa
);
2761 blkptr_t
*bp
= pio
->io_bp
;
2762 zio_t
*gio
= pio
->io_gang_leader
;
2764 zio_gang_node_t
*gn
, **gnpp
;
2765 zio_gbh_phys_t
*gbh
;
2767 uint64_t txg
= pio
->io_txg
;
2768 uint64_t resid
= pio
->io_size
;
2770 int copies
= gio
->io_prop
.zp_copies
;
2774 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2777 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2778 * have a third copy.
2780 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2781 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2782 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2784 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2785 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2786 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2789 flags
|= METASLAB_ASYNC_ALLOC
;
2790 VERIFY(zfs_refcount_held(&mc
->mc_alloc_slots
[pio
->io_allocator
],
2794 * The logical zio has already placed a reservation for
2795 * 'copies' allocation slots but gang blocks may require
2796 * additional copies. These additional copies
2797 * (i.e. gbh_copies - copies) are guaranteed to succeed
2798 * since metaslab_class_throttle_reserve() always allows
2799 * additional reservations for gang blocks.
2801 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2802 pio
->io_allocator
, pio
, flags
));
2805 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2806 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2807 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2809 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2810 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2814 * If we failed to allocate the gang block header then
2815 * we remove any additional allocation reservations that
2816 * we placed here. The original reservation will
2817 * be removed when the logical I/O goes to the ready
2820 metaslab_class_throttle_unreserve(mc
,
2821 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2824 pio
->io_error
= error
;
2829 gnpp
= &gio
->io_gang_tree
;
2831 gnpp
= pio
->io_private
;
2832 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2835 gn
= zio_gang_node_alloc(gnpp
);
2837 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2838 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2841 * Create the gang header.
2843 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2844 zio_write_gang_done
, NULL
, pio
->io_priority
,
2845 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2848 * Create and nowait the gang children.
2850 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2851 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2853 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2855 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2856 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2857 zp
.zp_complevel
= gio
->io_prop
.zp_complevel
;
2858 zp
.zp_type
= DMU_OT_NONE
;
2860 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2861 zp
.zp_dedup
= B_FALSE
;
2862 zp
.zp_dedup_verify
= B_FALSE
;
2863 zp
.zp_nopwrite
= B_FALSE
;
2864 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2865 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2866 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2867 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2868 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2870 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2871 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2872 resid
) : NULL
, lsize
, lsize
, &zp
,
2873 zio_write_gang_member_ready
, NULL
, NULL
,
2874 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2875 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2877 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2878 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2882 * Gang children won't throttle but we should
2883 * account for their work, so reserve an allocation
2884 * slot for them here.
2886 VERIFY(metaslab_class_throttle_reserve(mc
,
2887 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2893 * Set pio's pipeline to just wait for zio to finish.
2895 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2898 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2900 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2908 * The zio_nop_write stage in the pipeline determines if allocating a
2909 * new bp is necessary. The nopwrite feature can handle writes in
2910 * either syncing or open context (i.e. zil writes) and as a result is
2911 * mutually exclusive with dedup.
2913 * By leveraging a cryptographically secure checksum, such as SHA256, we
2914 * can compare the checksums of the new data and the old to determine if
2915 * allocating a new block is required. Note that our requirements for
2916 * cryptographic strength are fairly weak: there can't be any accidental
2917 * hash collisions, but we don't need to be secure against intentional
2918 * (malicious) collisions. To trigger a nopwrite, you have to be able
2919 * to write the file to begin with, and triggering an incorrect (hash
2920 * collision) nopwrite is no worse than simply writing to the file.
2921 * That said, there are no known attacks against the checksum algorithms
2922 * used for nopwrite, assuming that the salt and the checksums
2923 * themselves remain secret.
2926 zio_nop_write(zio_t
*zio
)
2928 blkptr_t
*bp
= zio
->io_bp
;
2929 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2930 zio_prop_t
*zp
= &zio
->io_prop
;
2932 ASSERT(BP_GET_LEVEL(bp
) == 0);
2933 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2934 ASSERT(zp
->zp_nopwrite
);
2935 ASSERT(!zp
->zp_dedup
);
2936 ASSERT(zio
->io_bp_override
== NULL
);
2937 ASSERT(IO_IS_ALLOCATING(zio
));
2940 * Check to see if the original bp and the new bp have matching
2941 * characteristics (i.e. same checksum, compression algorithms, etc).
2942 * If they don't then just continue with the pipeline which will
2943 * allocate a new bp.
2945 if (BP_IS_HOLE(bp_orig
) ||
2946 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2947 ZCHECKSUM_FLAG_NOPWRITE
) ||
2948 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2949 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2950 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2951 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2952 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2956 * If the checksums match then reset the pipeline so that we
2957 * avoid allocating a new bp and issuing any I/O.
2959 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2960 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2961 ZCHECKSUM_FLAG_NOPWRITE
);
2962 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2963 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2964 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2965 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2966 sizeof (uint64_t)) == 0);
2969 * If we're overwriting a block that is currently on an
2970 * indirect vdev, then ignore the nopwrite request and
2971 * allow a new block to be allocated on a concrete vdev.
2973 spa_config_enter(zio
->io_spa
, SCL_VDEV
, FTAG
, RW_READER
);
2974 vdev_t
*tvd
= vdev_lookup_top(zio
->io_spa
,
2975 DVA_GET_VDEV(&bp
->blk_dva
[0]));
2976 if (tvd
->vdev_ops
== &vdev_indirect_ops
) {
2977 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2980 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2983 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2984 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2991 * ==========================================================================
2993 * ==========================================================================
2996 zio_ddt_child_read_done(zio_t
*zio
)
2998 blkptr_t
*bp
= zio
->io_bp
;
2999 ddt_entry_t
*dde
= zio
->io_private
;
3001 zio_t
*pio
= zio_unique_parent(zio
);
3003 mutex_enter(&pio
->io_lock
);
3004 ddp
= ddt_phys_select(dde
, bp
);
3005 if (zio
->io_error
== 0)
3006 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
3008 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
3009 dde
->dde_repair_abd
= zio
->io_abd
;
3011 abd_free(zio
->io_abd
);
3012 mutex_exit(&pio
->io_lock
);
3016 zio_ddt_read_start(zio_t
*zio
)
3018 blkptr_t
*bp
= zio
->io_bp
;
3020 ASSERT(BP_GET_DEDUP(bp
));
3021 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3022 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3024 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3025 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3026 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
3027 ddt_phys_t
*ddp
= dde
->dde_phys
;
3028 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
3031 ASSERT(zio
->io_vsd
== NULL
);
3034 if (ddp_self
== NULL
)
3037 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
3038 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
3040 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
3042 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
3043 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
3044 zio
->io_size
, zio_ddt_child_read_done
, dde
,
3045 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
3046 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
3051 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
3052 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
3053 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
3059 zio_ddt_read_done(zio_t
*zio
)
3061 blkptr_t
*bp
= zio
->io_bp
;
3063 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
3067 ASSERT(BP_GET_DEDUP(bp
));
3068 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3069 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3071 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3072 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3073 ddt_entry_t
*dde
= zio
->io_vsd
;
3075 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
3079 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
3080 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
3083 if (dde
->dde_repair_abd
!= NULL
) {
3084 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
3086 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
3088 ddt_repair_done(ddt
, dde
);
3092 ASSERT(zio
->io_vsd
== NULL
);
3098 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
3100 spa_t
*spa
= zio
->io_spa
;
3101 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
3103 ASSERT(!(zio
->io_bp_override
&& do_raw
));
3106 * Note: we compare the original data, not the transformed data,
3107 * because when zio->io_bp is an override bp, we will not have
3108 * pushed the I/O transforms. That's an important optimization
3109 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3110 * However, we should never get a raw, override zio so in these
3111 * cases we can compare the io_abd directly. This is useful because
3112 * it allows us to do dedup verification even if we don't have access
3113 * to the original data (for instance, if the encryption keys aren't
3117 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3118 zio_t
*lio
= dde
->dde_lead_zio
[p
];
3120 if (lio
!= NULL
&& do_raw
) {
3121 return (lio
->io_size
!= zio
->io_size
||
3122 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
3123 } else if (lio
!= NULL
) {
3124 return (lio
->io_orig_size
!= zio
->io_orig_size
||
3125 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
3129 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3130 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3132 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
3133 blkptr_t blk
= *zio
->io_bp
;
3138 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3139 psize
= BP_GET_PSIZE(&blk
);
3141 if (psize
!= zio
->io_size
)
3146 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
3148 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
3149 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
3150 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
3151 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
3154 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
3155 error
= SET_ERROR(ENOENT
);
3160 return (error
!= 0);
3161 } else if (ddp
->ddp_phys_birth
!= 0) {
3162 arc_buf_t
*abuf
= NULL
;
3163 arc_flags_t aflags
= ARC_FLAG_WAIT
;
3164 blkptr_t blk
= *zio
->io_bp
;
3167 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3169 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3174 error
= arc_read(NULL
, spa
, &blk
,
3175 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3176 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3177 &aflags
, &zio
->io_bookmark
);
3180 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3181 zio
->io_orig_size
) != 0)
3182 error
= SET_ERROR(ENOENT
);
3183 arc_buf_destroy(abuf
, &abuf
);
3187 return (error
!= 0);
3195 zio_ddt_child_write_ready(zio_t
*zio
)
3197 int p
= zio
->io_prop
.zp_copies
;
3198 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3199 ddt_entry_t
*dde
= zio
->io_private
;
3200 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3208 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3210 ddt_phys_fill(ddp
, zio
->io_bp
);
3212 zio_link_t
*zl
= NULL
;
3213 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3214 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3220 zio_ddt_child_write_done(zio_t
*zio
)
3222 int p
= zio
->io_prop
.zp_copies
;
3223 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3224 ddt_entry_t
*dde
= zio
->io_private
;
3225 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3229 ASSERT(ddp
->ddp_refcnt
== 0);
3230 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3231 dde
->dde_lead_zio
[p
] = NULL
;
3233 if (zio
->io_error
== 0) {
3234 zio_link_t
*zl
= NULL
;
3235 while (zio_walk_parents(zio
, &zl
) != NULL
)
3236 ddt_phys_addref(ddp
);
3238 ddt_phys_clear(ddp
);
3245 zio_ddt_write(zio_t
*zio
)
3247 spa_t
*spa
= zio
->io_spa
;
3248 blkptr_t
*bp
= zio
->io_bp
;
3249 uint64_t txg
= zio
->io_txg
;
3250 zio_prop_t
*zp
= &zio
->io_prop
;
3251 int p
= zp
->zp_copies
;
3253 ddt_t
*ddt
= ddt_select(spa
, bp
);
3257 ASSERT(BP_GET_DEDUP(bp
));
3258 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3259 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3260 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3263 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3264 ddp
= &dde
->dde_phys
[p
];
3266 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3268 * If we're using a weak checksum, upgrade to a strong checksum
3269 * and try again. If we're already using a strong checksum,
3270 * we can't resolve it, so just convert to an ordinary write.
3271 * (And automatically e-mail a paper to Nature?)
3273 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3274 ZCHECKSUM_FLAG_DEDUP
)) {
3275 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3276 zio_pop_transforms(zio
);
3277 zio
->io_stage
= ZIO_STAGE_OPEN
;
3280 zp
->zp_dedup
= B_FALSE
;
3281 BP_SET_DEDUP(bp
, B_FALSE
);
3283 ASSERT(!BP_GET_DEDUP(bp
));
3284 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3289 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3290 if (ddp
->ddp_phys_birth
!= 0)
3291 ddt_bp_fill(ddp
, bp
, txg
);
3292 if (dde
->dde_lead_zio
[p
] != NULL
)
3293 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3295 ddt_phys_addref(ddp
);
3296 } else if (zio
->io_bp_override
) {
3297 ASSERT(bp
->blk_birth
== txg
);
3298 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3299 ddt_phys_fill(ddp
, bp
);
3300 ddt_phys_addref(ddp
);
3302 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3303 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3304 zio_ddt_child_write_ready
, NULL
, NULL
,
3305 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3306 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3308 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3309 dde
->dde_lead_zio
[p
] = cio
;
3319 ddt_entry_t
*freedde
; /* for debugging */
3322 zio_ddt_free(zio_t
*zio
)
3324 spa_t
*spa
= zio
->io_spa
;
3325 blkptr_t
*bp
= zio
->io_bp
;
3326 ddt_t
*ddt
= ddt_select(spa
, bp
);
3330 ASSERT(BP_GET_DEDUP(bp
));
3331 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3334 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3336 ddp
= ddt_phys_select(dde
, bp
);
3338 ddt_phys_decref(ddp
);
3346 * ==========================================================================
3347 * Allocate and free blocks
3348 * ==========================================================================
3352 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3356 ASSERT(MUTEX_HELD(&spa
->spa_alloc_locks
[allocator
]));
3358 zio
= avl_first(&spa
->spa_alloc_trees
[allocator
]);
3362 ASSERT(IO_IS_ALLOCATING(zio
));
3365 * Try to place a reservation for this zio. If we're unable to
3366 * reserve then we throttle.
3368 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3369 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3370 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
, 0)) {
3374 avl_remove(&spa
->spa_alloc_trees
[allocator
], zio
);
3375 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3381 zio_dva_throttle(zio_t
*zio
)
3383 spa_t
*spa
= zio
->io_spa
;
3385 metaslab_class_t
*mc
;
3387 /* locate an appropriate allocation class */
3388 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3389 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3391 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3392 !mc
->mc_alloc_throttle_enabled
||
3393 zio
->io_child_type
== ZIO_CHILD_GANG
||
3394 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3398 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3400 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3401 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3403 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
3405 * We want to try to use as many allocators as possible to help improve
3406 * performance, but we also want logically adjacent IOs to be physically
3407 * adjacent to improve sequential read performance. We chunk each object
3408 * into 2^20 block regions, and then hash based on the objset, object,
3409 * level, and region to accomplish both of these goals.
3411 zio
->io_allocator
= cityhash4(bm
->zb_objset
, bm
->zb_object
,
3412 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
3413 mutex_enter(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3414 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3415 zio
->io_metaslab_class
= mc
;
3416 avl_add(&spa
->spa_alloc_trees
[zio
->io_allocator
], zio
);
3417 nio
= zio_io_to_allocate(spa
, zio
->io_allocator
);
3418 mutex_exit(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3423 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3427 mutex_enter(&spa
->spa_alloc_locks
[allocator
]);
3428 zio
= zio_io_to_allocate(spa
, allocator
);
3429 mutex_exit(&spa
->spa_alloc_locks
[allocator
]);
3433 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3434 ASSERT0(zio
->io_error
);
3435 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3439 zio_dva_allocate(zio_t
*zio
)
3441 spa_t
*spa
= zio
->io_spa
;
3442 metaslab_class_t
*mc
;
3443 blkptr_t
*bp
= zio
->io_bp
;
3447 if (zio
->io_gang_leader
== NULL
) {
3448 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3449 zio
->io_gang_leader
= zio
;
3452 ASSERT(BP_IS_HOLE(bp
));
3453 ASSERT0(BP_GET_NDVAS(bp
));
3454 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3455 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3456 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3458 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3459 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3460 flags
|= METASLAB_DONT_THROTTLE
;
3461 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3462 flags
|= METASLAB_GANG_CHILD
;
3463 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3464 flags
|= METASLAB_ASYNC_ALLOC
;
3467 * if not already chosen, locate an appropriate allocation class
3469 mc
= zio
->io_metaslab_class
;
3471 mc
= spa_preferred_class(spa
, zio
->io_size
,
3472 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3473 zio
->io_prop
.zp_zpl_smallblk
);
3474 zio
->io_metaslab_class
= mc
;
3477 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3478 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3479 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3482 * Fallback to normal class when an alloc class is full
3484 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3486 * If throttling, transfer reservation over to normal class.
3487 * The io_allocator slot can remain the same even though we
3488 * are switching classes.
3490 if (mc
->mc_alloc_throttle_enabled
&&
3491 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3492 metaslab_class_throttle_unreserve(mc
,
3493 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3494 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3496 mc
= spa_normal_class(spa
);
3497 VERIFY(metaslab_class_throttle_reserve(mc
,
3498 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3499 flags
| METASLAB_MUST_RESERVE
));
3501 mc
= spa_normal_class(spa
);
3503 zio
->io_metaslab_class
= mc
;
3505 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3506 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3507 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3511 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3512 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3514 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3515 return (zio_write_gang_block(zio
));
3516 zio
->io_error
= error
;
3523 zio_dva_free(zio_t
*zio
)
3525 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3531 zio_dva_claim(zio_t
*zio
)
3535 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3537 zio
->io_error
= error
;
3543 * Undo an allocation. This is used by zio_done() when an I/O fails
3544 * and we want to give back the block we just allocated.
3545 * This handles both normal blocks and gang blocks.
3548 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3550 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3551 ASSERT(zio
->io_bp_override
== NULL
);
3553 if (!BP_IS_HOLE(bp
))
3554 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3557 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3558 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3559 &gn
->gn_gbh
->zg_blkptr
[g
]);
3565 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3568 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3569 uint64_t size
, boolean_t
*slog
)
3572 zio_alloc_list_t io_alloc_list
;
3574 ASSERT(txg
> spa_syncing_txg(spa
));
3576 metaslab_trace_init(&io_alloc_list
);
3579 * Block pointer fields are useful to metaslabs for stats and debugging.
3580 * Fill in the obvious ones before calling into metaslab_alloc().
3582 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3583 BP_SET_PSIZE(new_bp
, size
);
3584 BP_SET_LEVEL(new_bp
, 0);
3587 * When allocating a zil block, we don't have information about
3588 * the final destination of the block except the objset it's part
3589 * of, so we just hash the objset ID to pick the allocator to get
3592 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3593 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
,
3594 cityhash4(0, 0, 0, os
->os_dsl_dataset
->ds_object
) %
3595 spa
->spa_alloc_count
);
3599 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3600 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3601 &io_alloc_list
, NULL
, cityhash4(0, 0, 0,
3602 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
);
3606 metaslab_trace_fini(&io_alloc_list
);
3609 BP_SET_LSIZE(new_bp
, size
);
3610 BP_SET_PSIZE(new_bp
, size
);
3611 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3612 BP_SET_CHECKSUM(new_bp
,
3613 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3614 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3615 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3616 BP_SET_LEVEL(new_bp
, 0);
3617 BP_SET_DEDUP(new_bp
, 0);
3618 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3621 * encrypted blocks will require an IV and salt. We generate
3622 * these now since we will not be rewriting the bp at
3625 if (os
->os_encrypted
) {
3626 uint8_t iv
[ZIO_DATA_IV_LEN
];
3627 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3629 BP_SET_CRYPT(new_bp
, B_TRUE
);
3630 VERIFY0(spa_crypt_get_salt(spa
,
3631 dmu_objset_id(os
), salt
));
3632 VERIFY0(zio_crypt_generate_iv(iv
));
3634 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3637 zfs_dbgmsg("%s: zil block allocation failure: "
3638 "size %llu, error %d", spa_name(spa
), size
, error
);
3645 * ==========================================================================
3646 * Read and write to physical devices
3647 * ==========================================================================
3651 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3652 * stops after this stage and will resume upon I/O completion.
3653 * However, there are instances where the vdev layer may need to
3654 * continue the pipeline when an I/O was not issued. Since the I/O
3655 * that was sent to the vdev layer might be different than the one
3656 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3657 * force the underlying vdev layers to call either zio_execute() or
3658 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3661 zio_vdev_io_start(zio_t
*zio
)
3663 vdev_t
*vd
= zio
->io_vd
;
3665 spa_t
*spa
= zio
->io_spa
;
3669 ASSERT(zio
->io_error
== 0);
3670 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3673 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3674 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3677 * The mirror_ops handle multiple DVAs in a single BP.
3679 vdev_mirror_ops
.vdev_op_io_start(zio
);
3683 ASSERT3P(zio
->io_logical
, !=, zio
);
3684 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3685 ASSERT(spa
->spa_trust_config
);
3688 * Note: the code can handle other kinds of writes,
3689 * but we don't expect them.
3691 if (zio
->io_vd
->vdev_removing
) {
3692 ASSERT(zio
->io_flags
&
3693 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3694 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3698 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3700 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3701 P2PHASE(zio
->io_size
, align
) != 0) {
3702 /* Transform logical writes to be a full physical block size. */
3703 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3704 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3705 ASSERT(vd
== vd
->vdev_top
);
3706 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3707 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3708 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3710 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3714 * If this is not a physical io, make sure that it is properly aligned
3715 * before proceeding.
3717 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3718 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3719 ASSERT0(P2PHASE(zio
->io_size
, align
));
3722 * For physical writes, we allow 512b aligned writes and assume
3723 * the device will perform a read-modify-write as necessary.
3725 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3726 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3729 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3732 * If this is a repair I/O, and there's no self-healing involved --
3733 * that is, we're just resilvering what we expect to resilver --
3734 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3735 * This prevents spurious resilvering.
3737 * There are a few ways that we can end up creating these spurious
3740 * 1. A resilver i/o will be issued if any DVA in the BP has a
3741 * dirty DTL. The mirror code will issue resilver writes to
3742 * each DVA, including the one(s) that are not on vdevs with dirty
3745 * 2. With nested replication, which happens when we have a
3746 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3747 * For example, given mirror(replacing(A+B), C), it's likely that
3748 * only A is out of date (it's the new device). In this case, we'll
3749 * read from C, then use the data to resilver A+B -- but we don't
3750 * actually want to resilver B, just A. The top-level mirror has no
3751 * way to know this, so instead we just discard unnecessary repairs
3752 * as we work our way down the vdev tree.
3754 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3755 * The same logic applies to any form of nested replication: ditto
3756 * + mirror, RAID-Z + replacing, etc.
3758 * However, indirect vdevs point off to other vdevs which may have
3759 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3760 * will be properly bypassed instead.
3762 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3763 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3764 zio
->io_txg
!= 0 && /* not a delegated i/o */
3765 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3766 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3767 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3768 zio_vdev_io_bypass(zio
);
3772 if (vd
->vdev_ops
->vdev_op_leaf
&& (zio
->io_type
== ZIO_TYPE_READ
||
3773 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
)) {
3775 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3778 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3781 if (!vdev_accessible(vd
, zio
)) {
3782 zio
->io_error
= SET_ERROR(ENXIO
);
3786 zio
->io_delay
= gethrtime();
3789 vd
->vdev_ops
->vdev_op_io_start(zio
);
3794 zio_vdev_io_done(zio_t
*zio
)
3796 vdev_t
*vd
= zio
->io_vd
;
3797 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3798 boolean_t unexpected_error
= B_FALSE
;
3800 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3804 ASSERT(zio
->io_type
== ZIO_TYPE_READ
||
3805 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
);
3808 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3810 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3812 vdev_queue_io_done(zio
);
3814 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3815 vdev_cache_write(zio
);
3817 if (zio_injection_enabled
&& zio
->io_error
== 0)
3818 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3821 if (zio_injection_enabled
&& zio
->io_error
== 0)
3822 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3824 if (zio
->io_error
&& zio
->io_type
!= ZIO_TYPE_TRIM
) {
3825 if (!vdev_accessible(vd
, zio
)) {
3826 zio
->io_error
= SET_ERROR(ENXIO
);
3828 unexpected_error
= B_TRUE
;
3833 ops
->vdev_op_io_done(zio
);
3835 if (unexpected_error
)
3836 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3842 * This function is used to change the priority of an existing zio that is
3843 * currently in-flight. This is used by the arc to upgrade priority in the
3844 * event that a demand read is made for a block that is currently queued
3845 * as a scrub or async read IO. Otherwise, the high priority read request
3846 * would end up having to wait for the lower priority IO.
3849 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3851 zio_t
*cio
, *cio_next
;
3852 zio_link_t
*zl
= NULL
;
3854 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3856 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3857 vdev_queue_change_io_priority(pio
, priority
);
3859 pio
->io_priority
= priority
;
3862 mutex_enter(&pio
->io_lock
);
3863 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3864 cio_next
= zio_walk_children(pio
, &zl
);
3865 zio_change_priority(cio
, priority
);
3867 mutex_exit(&pio
->io_lock
);
3871 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3872 * disk, and use that to finish the checksum ereport later.
3875 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3876 const abd_t
*good_buf
)
3878 /* no processing needed */
3879 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3884 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3886 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3888 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3890 zcr
->zcr_cbinfo
= zio
->io_size
;
3891 zcr
->zcr_cbdata
= abd
;
3892 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3893 zcr
->zcr_free
= zio_abd_free
;
3897 zio_vdev_io_assess(zio_t
*zio
)
3899 vdev_t
*vd
= zio
->io_vd
;
3901 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3905 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3906 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3908 if (zio
->io_vsd
!= NULL
) {
3909 zio
->io_vsd_ops
->vsd_free(zio
);
3913 if (zio_injection_enabled
&& zio
->io_error
== 0)
3914 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3917 * If the I/O failed, determine whether we should attempt to retry it.
3919 * On retry, we cut in line in the issue queue, since we don't want
3920 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3922 if (zio
->io_error
&& vd
== NULL
&&
3923 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3924 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3925 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3927 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3928 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3929 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3930 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3931 zio_requeue_io_start_cut_in_line
);
3936 * If we got an error on a leaf device, convert it to ENXIO
3937 * if the device is not accessible at all.
3939 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3940 !vdev_accessible(vd
, zio
))
3941 zio
->io_error
= SET_ERROR(ENXIO
);
3944 * If we can't write to an interior vdev (mirror or RAID-Z),
3945 * set vdev_cant_write so that we stop trying to allocate from it.
3947 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3948 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3949 vd
->vdev_cant_write
= B_TRUE
;
3953 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3954 * attempts will ever succeed. In this case we set a persistent
3955 * boolean flag so that we don't bother with it in the future.
3957 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3958 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3959 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3960 vd
->vdev_nowritecache
= B_TRUE
;
3963 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3965 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3966 zio
->io_physdone
!= NULL
) {
3967 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3968 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3969 zio
->io_physdone(zio
->io_logical
);
3976 zio_vdev_io_reissue(zio_t
*zio
)
3978 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3979 ASSERT(zio
->io_error
== 0);
3981 zio
->io_stage
>>= 1;
3985 zio_vdev_io_redone(zio_t
*zio
)
3987 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3989 zio
->io_stage
>>= 1;
3993 zio_vdev_io_bypass(zio_t
*zio
)
3995 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3996 ASSERT(zio
->io_error
== 0);
3998 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3999 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
4003 * ==========================================================================
4004 * Encrypt and store encryption parameters
4005 * ==========================================================================
4010 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4011 * managing the storage of encryption parameters and passing them to the
4012 * lower-level encryption functions.
4015 zio_encrypt(zio_t
*zio
)
4017 zio_prop_t
*zp
= &zio
->io_prop
;
4018 spa_t
*spa
= zio
->io_spa
;
4019 blkptr_t
*bp
= zio
->io_bp
;
4020 uint64_t psize
= BP_GET_PSIZE(bp
);
4021 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
4022 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
4023 void *enc_buf
= NULL
;
4025 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4026 uint8_t iv
[ZIO_DATA_IV_LEN
];
4027 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4028 boolean_t no_crypt
= B_FALSE
;
4030 /* the root zio already encrypted the data */
4031 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
4034 /* only ZIL blocks are re-encrypted on rewrite */
4035 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
4038 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
4039 BP_SET_CRYPT(bp
, B_FALSE
);
4043 /* if we are doing raw encryption set the provided encryption params */
4044 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
4045 ASSERT0(BP_GET_LEVEL(bp
));
4046 BP_SET_CRYPT(bp
, B_TRUE
);
4047 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
4048 if (ot
!= DMU_OT_OBJSET
)
4049 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
4051 /* dnode blocks must be written out in the provided byteorder */
4052 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
4053 ot
== DMU_OT_DNODE
) {
4054 void *bswap_buf
= zio_buf_alloc(psize
);
4055 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
4057 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4058 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
4059 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
4062 abd_take_ownership_of_buf(babd
, B_TRUE
);
4063 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
4066 if (DMU_OT_IS_ENCRYPTED(ot
))
4067 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
4071 /* indirect blocks only maintain a cksum of the lower level MACs */
4072 if (BP_GET_LEVEL(bp
) > 0) {
4073 BP_SET_CRYPT(bp
, B_TRUE
);
4074 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
4075 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
4077 zio_crypt_encode_mac_bp(bp
, mac
);
4082 * Objset blocks are a special case since they have 2 256-bit MACs
4083 * embedded within them.
4085 if (ot
== DMU_OT_OBJSET
) {
4086 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
4087 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4088 BP_SET_CRYPT(bp
, B_TRUE
);
4089 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
4090 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
4094 /* unencrypted object types are only authenticated with a MAC */
4095 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
4096 BP_SET_CRYPT(bp
, B_TRUE
);
4097 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
4098 zio
->io_abd
, psize
, mac
));
4099 zio_crypt_encode_mac_bp(bp
, mac
);
4104 * Later passes of sync-to-convergence may decide to rewrite data
4105 * in place to avoid more disk reallocations. This presents a problem
4106 * for encryption because this constitutes rewriting the new data with
4107 * the same encryption key and IV. However, this only applies to blocks
4108 * in the MOS (particularly the spacemaps) and we do not encrypt the
4109 * MOS. We assert that the zio is allocating or an intent log write
4112 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
4113 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
4114 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
4115 ASSERT3U(psize
, !=, 0);
4117 enc_buf
= zio_buf_alloc(psize
);
4118 eabd
= abd_get_from_buf(enc_buf
, psize
);
4119 abd_take_ownership_of_buf(eabd
, B_TRUE
);
4122 * For an explanation of what encryption parameters are stored
4123 * where, see the block comment in zio_crypt.c.
4125 if (ot
== DMU_OT_INTENT_LOG
) {
4126 zio_crypt_decode_params_bp(bp
, salt
, iv
);
4128 BP_SET_CRYPT(bp
, B_TRUE
);
4131 /* Perform the encryption. This should not fail */
4132 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
4133 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
4134 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
4136 /* encode encryption metadata into the bp */
4137 if (ot
== DMU_OT_INTENT_LOG
) {
4139 * ZIL blocks store the MAC in the embedded checksum, so the
4140 * transform must always be applied.
4142 zio_crypt_encode_mac_zil(enc_buf
, mac
);
4143 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4145 BP_SET_CRYPT(bp
, B_TRUE
);
4146 zio_crypt_encode_params_bp(bp
, salt
, iv
);
4147 zio_crypt_encode_mac_bp(bp
, mac
);
4150 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
4153 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4161 * ==========================================================================
4162 * Generate and verify checksums
4163 * ==========================================================================
4166 zio_checksum_generate(zio_t
*zio
)
4168 blkptr_t
*bp
= zio
->io_bp
;
4169 enum zio_checksum checksum
;
4173 * This is zio_write_phys().
4174 * We're either generating a label checksum, or none at all.
4176 checksum
= zio
->io_prop
.zp_checksum
;
4178 if (checksum
== ZIO_CHECKSUM_OFF
)
4181 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4183 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4184 ASSERT(!IO_IS_ALLOCATING(zio
));
4185 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4187 checksum
= BP_GET_CHECKSUM(bp
);
4191 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4197 zio_checksum_verify(zio_t
*zio
)
4199 zio_bad_cksum_t info
;
4200 blkptr_t
*bp
= zio
->io_bp
;
4203 ASSERT(zio
->io_vd
!= NULL
);
4207 * This is zio_read_phys().
4208 * We're either verifying a label checksum, or nothing at all.
4210 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4213 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
4216 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4217 zio
->io_error
= error
;
4218 if (error
== ECKSUM
&&
4219 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4220 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4221 zio
->io_vd
->vdev_stat
.vs_checksum_errors
++;
4222 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4224 zfs_ereport_start_checksum(zio
->io_spa
,
4225 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4226 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
4234 * Called by RAID-Z to ensure we don't compute the checksum twice.
4237 zio_checksum_verified(zio_t
*zio
)
4239 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4243 * ==========================================================================
4244 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4245 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4246 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4247 * indicate errors that are specific to one I/O, and most likely permanent.
4248 * Any other error is presumed to be worse because we weren't expecting it.
4249 * ==========================================================================
4252 zio_worst_error(int e1
, int e2
)
4254 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4257 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4258 if (e1
== zio_error_rank
[r1
])
4261 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4262 if (e2
== zio_error_rank
[r2
])
4265 return (r1
> r2
? e1
: e2
);
4269 * ==========================================================================
4271 * ==========================================================================
4274 zio_ready(zio_t
*zio
)
4276 blkptr_t
*bp
= zio
->io_bp
;
4277 zio_t
*pio
, *pio_next
;
4278 zio_link_t
*zl
= NULL
;
4280 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4285 if (zio
->io_ready
) {
4286 ASSERT(IO_IS_ALLOCATING(zio
));
4287 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4288 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4289 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4294 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4295 zio
->io_bp_copy
= *bp
;
4297 if (zio
->io_error
!= 0) {
4298 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4300 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4301 ASSERT(IO_IS_ALLOCATING(zio
));
4302 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4303 ASSERT(zio
->io_metaslab_class
!= NULL
);
4306 * We were unable to allocate anything, unreserve and
4307 * issue the next I/O to allocate.
4309 metaslab_class_throttle_unreserve(
4310 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4311 zio
->io_allocator
, zio
);
4312 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4316 mutex_enter(&zio
->io_lock
);
4317 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4318 pio
= zio_walk_parents(zio
, &zl
);
4319 mutex_exit(&zio
->io_lock
);
4322 * As we notify zio's parents, new parents could be added.
4323 * New parents go to the head of zio's io_parent_list, however,
4324 * so we will (correctly) not notify them. The remainder of zio's
4325 * io_parent_list, from 'pio_next' onward, cannot change because
4326 * all parents must wait for us to be done before they can be done.
4328 for (; pio
!= NULL
; pio
= pio_next
) {
4329 pio_next
= zio_walk_parents(zio
, &zl
);
4330 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4333 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4334 if (BP_IS_GANG(bp
)) {
4335 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4337 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4338 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4342 if (zio_injection_enabled
&&
4343 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4344 zio_handle_ignored_writes(zio
);
4350 * Update the allocation throttle accounting.
4353 zio_dva_throttle_done(zio_t
*zio
)
4355 zio_t
*lio __maybe_unused
= zio
->io_logical
;
4356 zio_t
*pio
= zio_unique_parent(zio
);
4357 vdev_t
*vd
= zio
->io_vd
;
4358 int flags
= METASLAB_ASYNC_ALLOC
;
4360 ASSERT3P(zio
->io_bp
, !=, NULL
);
4361 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4362 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4363 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4365 ASSERT3P(vd
, ==, vd
->vdev_top
);
4366 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4367 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4368 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4369 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4370 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4373 * Parents of gang children can have two flavors -- ones that
4374 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4375 * and ones that allocated the constituent blocks. The allocation
4376 * throttle needs to know the allocating parent zio so we must find
4379 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4381 * If our parent is a rewrite gang child then our grandparent
4382 * would have been the one that performed the allocation.
4384 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4385 pio
= zio_unique_parent(pio
);
4386 flags
|= METASLAB_GANG_CHILD
;
4389 ASSERT(IO_IS_ALLOCATING(pio
));
4390 ASSERT3P(zio
, !=, zio
->io_logical
);
4391 ASSERT(zio
->io_logical
!= NULL
);
4392 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4393 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4394 ASSERT(zio
->io_metaslab_class
!= NULL
);
4396 mutex_enter(&pio
->io_lock
);
4397 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4398 pio
->io_allocator
, B_TRUE
);
4399 mutex_exit(&pio
->io_lock
);
4401 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4402 pio
->io_allocator
, pio
);
4405 * Call into the pipeline to see if there is more work that
4406 * needs to be done. If there is work to be done it will be
4407 * dispatched to another taskq thread.
4409 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4413 zio_done(zio_t
*zio
)
4416 * Always attempt to keep stack usage minimal here since
4417 * we can be called recursively up to 19 levels deep.
4419 const uint64_t psize
= zio
->io_size
;
4420 zio_t
*pio
, *pio_next
;
4421 zio_link_t
*zl
= NULL
;
4424 * If our children haven't all completed,
4425 * wait for them and then repeat this pipeline stage.
4427 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4432 * If the allocation throttle is enabled, then update the accounting.
4433 * We only track child I/Os that are part of an allocating async
4434 * write. We must do this since the allocation is performed
4435 * by the logical I/O but the actual write is done by child I/Os.
4437 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4438 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4439 ASSERT(zio
->io_metaslab_class
!= NULL
);
4440 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4441 zio_dva_throttle_done(zio
);
4445 * If the allocation throttle is enabled, verify that
4446 * we have decremented the refcounts for every I/O that was throttled.
4448 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4449 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4450 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4451 ASSERT(zio
->io_bp
!= NULL
);
4453 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4455 VERIFY(zfs_refcount_not_held(
4456 &zio
->io_metaslab_class
->mc_alloc_slots
[zio
->io_allocator
],
4461 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4462 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4463 ASSERT(zio
->io_children
[c
][w
] == 0);
4465 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4466 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4467 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4468 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4469 sizeof (blkptr_t
)) == 0 ||
4470 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4471 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4472 zio
->io_bp_override
== NULL
&&
4473 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4474 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4475 BP_GET_NDVAS(zio
->io_bp
));
4476 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4477 (BP_COUNT_GANG(zio
->io_bp
) ==
4478 BP_GET_NDVAS(zio
->io_bp
)));
4480 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4481 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4485 * If there were child vdev/gang/ddt errors, they apply to us now.
4487 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4488 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4489 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4492 * If the I/O on the transformed data was successful, generate any
4493 * checksum reports now while we still have the transformed data.
4495 if (zio
->io_error
== 0) {
4496 while (zio
->io_cksum_report
!= NULL
) {
4497 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4498 uint64_t align
= zcr
->zcr_align
;
4499 uint64_t asize
= P2ROUNDUP(psize
, align
);
4500 abd_t
*adata
= zio
->io_abd
;
4502 if (asize
!= psize
) {
4503 adata
= abd_alloc(asize
, B_TRUE
);
4504 abd_copy(adata
, zio
->io_abd
, psize
);
4505 abd_zero_off(adata
, psize
, asize
- psize
);
4508 zio
->io_cksum_report
= zcr
->zcr_next
;
4509 zcr
->zcr_next
= NULL
;
4510 zcr
->zcr_finish(zcr
, adata
);
4511 zfs_ereport_free_checksum(zcr
);
4518 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4520 vdev_stat_update(zio
, psize
);
4523 * If this I/O is attached to a particular vdev is slow, exceeding
4524 * 30 seconds to complete, post an error described the I/O delay.
4525 * We ignore these errors if the device is currently unavailable.
4527 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4528 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4530 * We want to only increment our slow IO counters if
4531 * the IO is valid (i.e. not if the drive is removed).
4533 * zfs_ereport_post() will also do these checks, but
4534 * it can also ratelimit and have other failures, so we
4535 * need to increment the slow_io counters independent
4538 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4539 zio
->io_spa
, zio
->io_vd
, zio
)) {
4540 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4541 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4542 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4544 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4545 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4551 if (zio
->io_error
) {
4553 * If this I/O is attached to a particular vdev,
4554 * generate an error message describing the I/O failure
4555 * at the block level. We ignore these errors if the
4556 * device is currently unavailable.
4558 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4559 !vdev_is_dead(zio
->io_vd
)) {
4560 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4561 if (zio
->io_type
== ZIO_TYPE_READ
) {
4562 zio
->io_vd
->vdev_stat
.vs_read_errors
++;
4563 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
4564 zio
->io_vd
->vdev_stat
.vs_write_errors
++;
4566 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4568 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4569 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4572 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4573 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4574 zio
== zio
->io_logical
) {
4576 * For logical I/O requests, tell the SPA to log the
4577 * error and generate a logical data ereport.
4579 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4580 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4581 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4585 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4587 * Determine whether zio should be reexecuted. This will
4588 * propagate all the way to the root via zio_notify_parent().
4590 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4591 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4593 if (IO_IS_ALLOCATING(zio
) &&
4594 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4595 if (zio
->io_error
!= ENOSPC
)
4596 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4598 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4601 if ((zio
->io_type
== ZIO_TYPE_READ
||
4602 zio
->io_type
== ZIO_TYPE_FREE
) &&
4603 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4604 zio
->io_error
== ENXIO
&&
4605 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4606 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4607 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4609 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4610 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4613 * Here is a possibly good place to attempt to do
4614 * either combinatorial reconstruction or error correction
4615 * based on checksums. It also might be a good place
4616 * to send out preliminary ereports before we suspend
4622 * If there were logical child errors, they apply to us now.
4623 * We defer this until now to avoid conflating logical child
4624 * errors with errors that happened to the zio itself when
4625 * updating vdev stats and reporting FMA events above.
4627 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4629 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4630 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4631 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4632 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4634 zio_gang_tree_free(&zio
->io_gang_tree
);
4637 * Godfather I/Os should never suspend.
4639 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4640 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4641 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4643 if (zio
->io_reexecute
) {
4645 * This is a logical I/O that wants to reexecute.
4647 * Reexecute is top-down. When an i/o fails, if it's not
4648 * the root, it simply notifies its parent and sticks around.
4649 * The parent, seeing that it still has children in zio_done(),
4650 * does the same. This percolates all the way up to the root.
4651 * The root i/o will reexecute or suspend the entire tree.
4653 * This approach ensures that zio_reexecute() honors
4654 * all the original i/o dependency relationships, e.g.
4655 * parents not executing until children are ready.
4657 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4659 zio
->io_gang_leader
= NULL
;
4661 mutex_enter(&zio
->io_lock
);
4662 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4663 mutex_exit(&zio
->io_lock
);
4666 * "The Godfather" I/O monitors its children but is
4667 * not a true parent to them. It will track them through
4668 * the pipeline but severs its ties whenever they get into
4669 * trouble (e.g. suspended). This allows "The Godfather"
4670 * I/O to return status without blocking.
4673 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4675 zio_link_t
*remove_zl
= zl
;
4676 pio_next
= zio_walk_parents(zio
, &zl
);
4678 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4679 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4680 zio_remove_child(pio
, zio
, remove_zl
);
4682 * This is a rare code path, so we don't
4683 * bother with "next_to_execute".
4685 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4690 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4692 * We're not a root i/o, so there's nothing to do
4693 * but notify our parent. Don't propagate errors
4694 * upward since we haven't permanently failed yet.
4696 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4697 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4699 * This is a rare code path, so we don't bother with
4700 * "next_to_execute".
4702 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4703 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4705 * We'd fail again if we reexecuted now, so suspend
4706 * until conditions improve (e.g. device comes online).
4708 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4711 * Reexecution is potentially a huge amount of work.
4712 * Hand it off to the otherwise-unused claim taskq.
4714 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4715 spa_taskq_dispatch_ent(zio
->io_spa
,
4716 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4717 (task_func_t
*)zio_reexecute
, zio
, 0,
4723 ASSERT(zio
->io_child_count
== 0);
4724 ASSERT(zio
->io_reexecute
== 0);
4725 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4728 * Report any checksum errors, since the I/O is complete.
4730 while (zio
->io_cksum_report
!= NULL
) {
4731 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4732 zio
->io_cksum_report
= zcr
->zcr_next
;
4733 zcr
->zcr_next
= NULL
;
4734 zcr
->zcr_finish(zcr
, NULL
);
4735 zfs_ereport_free_checksum(zcr
);
4738 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4739 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4740 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4741 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4745 * It is the responsibility of the done callback to ensure that this
4746 * particular zio is no longer discoverable for adoption, and as
4747 * such, cannot acquire any new parents.
4752 mutex_enter(&zio
->io_lock
);
4753 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4754 mutex_exit(&zio
->io_lock
);
4757 * We are done executing this zio. We may want to execute a parent
4758 * next. See the comment in zio_notify_parent().
4760 zio_t
*next_to_execute
= NULL
;
4762 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4763 zio_link_t
*remove_zl
= zl
;
4764 pio_next
= zio_walk_parents(zio
, &zl
);
4765 zio_remove_child(pio
, zio
, remove_zl
);
4766 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
4769 if (zio
->io_waiter
!= NULL
) {
4770 mutex_enter(&zio
->io_lock
);
4771 zio
->io_executor
= NULL
;
4772 cv_broadcast(&zio
->io_cv
);
4773 mutex_exit(&zio
->io_lock
);
4778 return (next_to_execute
);
4782 * ==========================================================================
4783 * I/O pipeline definition
4784 * ==========================================================================
4786 static zio_pipe_stage_t
*zio_pipeline
[] = {
4794 zio_checksum_generate
,
4810 zio_checksum_verify
,
4818 * Compare two zbookmark_phys_t's to see which we would reach first in a
4819 * pre-order traversal of the object tree.
4821 * This is simple in every case aside from the meta-dnode object. For all other
4822 * objects, we traverse them in order (object 1 before object 2, and so on).
4823 * However, all of these objects are traversed while traversing object 0, since
4824 * the data it points to is the list of objects. Thus, we need to convert to a
4825 * canonical representation so we can compare meta-dnode bookmarks to
4826 * non-meta-dnode bookmarks.
4828 * We do this by calculating "equivalents" for each field of the zbookmark.
4829 * zbookmarks outside of the meta-dnode use their own object and level, and
4830 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4831 * blocks this bookmark refers to) by multiplying their blkid by their span
4832 * (the number of L0 blocks contained within one block at their level).
4833 * zbookmarks inside the meta-dnode calculate their object equivalent
4834 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4835 * level + 1<<31 (any value larger than a level could ever be) for their level.
4836 * This causes them to always compare before a bookmark in their object
4837 * equivalent, compare appropriately to bookmarks in other objects, and to
4838 * compare appropriately to other bookmarks in the meta-dnode.
4841 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4842 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4845 * These variables represent the "equivalent" values for the zbookmark,
4846 * after converting zbookmarks inside the meta dnode to their
4847 * normal-object equivalents.
4849 uint64_t zb1obj
, zb2obj
;
4850 uint64_t zb1L0
, zb2L0
;
4851 uint64_t zb1level
, zb2level
;
4853 if (zb1
->zb_object
== zb2
->zb_object
&&
4854 zb1
->zb_level
== zb2
->zb_level
&&
4855 zb1
->zb_blkid
== zb2
->zb_blkid
)
4858 IMPLY(zb1
->zb_level
> 0, ibs1
>= SPA_MINBLOCKSHIFT
);
4859 IMPLY(zb2
->zb_level
> 0, ibs2
>= SPA_MINBLOCKSHIFT
);
4862 * BP_SPANB calculates the span in blocks.
4864 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4865 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4867 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4868 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4870 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4872 zb1obj
= zb1
->zb_object
;
4873 zb1level
= zb1
->zb_level
;
4876 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4877 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4879 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4881 zb2obj
= zb2
->zb_object
;
4882 zb2level
= zb2
->zb_level
;
4885 /* Now that we have a canonical representation, do the comparison. */
4886 if (zb1obj
!= zb2obj
)
4887 return (zb1obj
< zb2obj
? -1 : 1);
4888 else if (zb1L0
!= zb2L0
)
4889 return (zb1L0
< zb2L0
? -1 : 1);
4890 else if (zb1level
!= zb2level
)
4891 return (zb1level
> zb2level
? -1 : 1);
4893 * This can (theoretically) happen if the bookmarks have the same object
4894 * and level, but different blkids, if the block sizes are not the same.
4895 * There is presently no way to change the indirect block sizes
4901 * This function checks the following: given that last_block is the place that
4902 * our traversal stopped last time, does that guarantee that we've visited
4903 * every node under subtree_root? Therefore, we can't just use the raw output
4904 * of zbookmark_compare. We have to pass in a modified version of
4905 * subtree_root; by incrementing the block id, and then checking whether
4906 * last_block is before or equal to that, we can tell whether or not having
4907 * visited last_block implies that all of subtree_root's children have been
4911 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4912 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4914 zbookmark_phys_t mod_zb
= *subtree_root
;
4916 ASSERT(last_block
->zb_level
== 0);
4918 /* The objset_phys_t isn't before anything. */
4923 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4924 * data block size in sectors, because that variable is only used if
4925 * the bookmark refers to a block in the meta-dnode. Since we don't
4926 * know without examining it what object it refers to, and there's no
4927 * harm in passing in this value in other cases, we always pass it in.
4929 * We pass in 0 for the indirect block size shift because zb2 must be
4930 * level 0. The indirect block size is only used to calculate the span
4931 * of the bookmark, but since the bookmark must be level 0, the span is
4932 * always 1, so the math works out.
4934 * If you make changes to how the zbookmark_compare code works, be sure
4935 * to make sure that this code still works afterwards.
4937 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4938 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4942 EXPORT_SYMBOL(zio_type_name
);
4943 EXPORT_SYMBOL(zio_buf_alloc
);
4944 EXPORT_SYMBOL(zio_data_buf_alloc
);
4945 EXPORT_SYMBOL(zio_buf_free
);
4946 EXPORT_SYMBOL(zio_data_buf_free
);
4949 ZFS_MODULE_PARAM(zfs_zio
, zio_
, slow_io_ms
, INT
, ZMOD_RW
,
4950 "Max I/O completion time (milliseconds) before marking it as slow");
4952 ZFS_MODULE_PARAM(zfs_zio
, zio_
, requeue_io_start_cut_in_line
, INT
, ZMOD_RW
,
4953 "Prioritize requeued I/O");
4955 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_deferred_free
, INT
, ZMOD_RW
,
4956 "Defer frees starting in this pass");
4958 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_dont_compress
, INT
, ZMOD_RW
,
4959 "Don't compress starting in this pass");
4961 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_rewrite
, INT
, ZMOD_RW
,
4962 "Rewrite new bps starting in this pass");
4964 ZFS_MODULE_PARAM(zfs_zio
, zio_
, dva_throttle_enabled
, INT
, ZMOD_RW
,
4965 "Throttle block allocations in the ZIO pipeline");
4967 ZFS_MODULE_PARAM(zfs_zio
, zio_
, deadman_log_all
, INT
, ZMOD_RW
,
4968 "Log all slow ZIOs, not just those with vdevs");