4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
43 #include <sys/dsl_scan.h>
44 #include <sys/metaslab_impl.h>
46 #include <sys/trace_zio.h>
48 #include <sys/dsl_crypt.h>
49 #include <sys/cityhash.h>
52 * ==========================================================================
53 * I/O type descriptions
54 * ==========================================================================
56 const char *zio_type_name
[ZIO_TYPES
] = {
58 * Note: Linux kernel thread name length is limited
59 * so these names will differ from upstream open zfs.
61 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
64 int zio_dva_throttle_enabled
= B_TRUE
;
65 int zio_deadman_log_all
= B_FALSE
;
68 * ==========================================================================
70 * ==========================================================================
72 kmem_cache_t
*zio_cache
;
73 kmem_cache_t
*zio_link_cache
;
74 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
75 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
76 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
77 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
81 /* Mark IOs as "slow" if they take longer than 30 seconds */
82 int zio_slow_io_ms
= (30 * MILLISEC
);
84 #define BP_SPANB(indblkshift, level) \
85 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
86 #define COMPARE_META_LEVEL 0x80000000ul
88 * The following actions directly effect the spa's sync-to-convergence logic.
89 * The values below define the sync pass when we start performing the action.
90 * Care should be taken when changing these values as they directly impact
91 * spa_sync() performance. Tuning these values may introduce subtle performance
92 * pathologies and should only be done in the context of performance analysis.
93 * These tunables will eventually be removed and replaced with #defines once
94 * enough analysis has been done to determine optimal values.
96 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
97 * regular blocks are not deferred.
99 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
100 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
101 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
104 * An allocating zio is one that either currently has the DVA allocate
105 * stage set or will have it later in its lifetime.
107 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
109 int zio_requeue_io_start_cut_in_line
= 1;
112 int zio_buf_debug_limit
= 16384;
114 int zio_buf_debug_limit
= 0;
117 static inline void __zio_execute(zio_t
*zio
);
119 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
125 vmem_t
*data_alloc_arena
= NULL
;
127 zio_cache
= kmem_cache_create("zio_cache",
128 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
129 zio_link_cache
= kmem_cache_create("zio_link_cache",
130 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
133 * For small buffers, we want a cache for each multiple of
134 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
135 * for each quarter-power of 2.
137 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
138 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
141 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
143 #if defined(_ILP32) && defined(_KERNEL)
145 * Cache size limited to 1M on 32-bit platforms until ARC
146 * buffers no longer require virtual address space.
148 if (size
> zfs_max_recordsize
)
157 * If we are using watchpoints, put each buffer on its own page,
158 * to eliminate the performance overhead of trapping to the
159 * kernel when modifying a non-watched buffer that shares the
160 * page with a watched buffer.
162 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
165 * Here's the problem - on 4K native devices in userland on
166 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
167 * will fail with EINVAL, causing zdb (and others) to coredump.
168 * Since userland probably doesn't need optimized buffer caches,
169 * we just force 4K alignment on everything.
171 align
= 8 * SPA_MINBLOCKSIZE
;
173 if (size
< PAGESIZE
) {
174 align
= SPA_MINBLOCKSIZE
;
175 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
182 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
183 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
184 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
186 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
187 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
188 align
, NULL
, NULL
, NULL
, NULL
,
189 data_alloc_arena
, cflags
);
194 ASSERT(zio_buf_cache
[c
] != NULL
);
195 if (zio_buf_cache
[c
- 1] == NULL
)
196 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
198 ASSERT(zio_data_buf_cache
[c
] != NULL
);
199 if (zio_data_buf_cache
[c
- 1] == NULL
)
200 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
212 kmem_cache_t
*last_cache
= NULL
;
213 kmem_cache_t
*last_data_cache
= NULL
;
215 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
218 * Cache size limited to 1M on 32-bit platforms until ARC
219 * buffers no longer require virtual address space.
221 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
224 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
225 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
226 (void) printf("zio_fini: [%d] %llu != %llu\n",
227 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
228 (long long unsigned)zio_buf_cache_allocs
[c
],
229 (long long unsigned)zio_buf_cache_frees
[c
]);
231 if (zio_buf_cache
[c
] != last_cache
) {
232 last_cache
= zio_buf_cache
[c
];
233 kmem_cache_destroy(zio_buf_cache
[c
]);
235 zio_buf_cache
[c
] = NULL
;
237 if (zio_data_buf_cache
[c
] != last_data_cache
) {
238 last_data_cache
= zio_data_buf_cache
[c
];
239 kmem_cache_destroy(zio_data_buf_cache
[c
]);
241 zio_data_buf_cache
[c
] = NULL
;
244 kmem_cache_destroy(zio_link_cache
);
245 kmem_cache_destroy(zio_cache
);
253 * ==========================================================================
254 * Allocate and free I/O buffers
255 * ==========================================================================
259 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
260 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
261 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
262 * excess / transient data in-core during a crashdump.
265 zio_buf_alloc(size_t size
)
267 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
269 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
270 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
271 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
274 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
278 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
279 * crashdump if the kernel panics. This exists so that we will limit the amount
280 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
281 * of kernel heap dumped to disk when the kernel panics)
284 zio_data_buf_alloc(size_t size
)
286 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
288 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
290 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
294 zio_buf_free(void *buf
, size_t size
)
296 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
298 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
299 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
300 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
303 kmem_cache_free(zio_buf_cache
[c
], buf
);
307 zio_data_buf_free(void *buf
, size_t size
)
309 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
311 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
313 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
317 zio_abd_free(void *abd
, size_t size
)
319 abd_free((abd_t
*)abd
);
323 * ==========================================================================
324 * Push and pop I/O transform buffers
325 * ==========================================================================
328 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
329 zio_transform_func_t
*transform
)
331 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
334 * Ensure that anyone expecting this zio to contain a linear ABD isn't
335 * going to get a nasty surprise when they try to access the data.
337 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
339 zt
->zt_orig_abd
= zio
->io_abd
;
340 zt
->zt_orig_size
= zio
->io_size
;
341 zt
->zt_bufsize
= bufsize
;
342 zt
->zt_transform
= transform
;
344 zt
->zt_next
= zio
->io_transform_stack
;
345 zio
->io_transform_stack
= zt
;
352 zio_pop_transforms(zio_t
*zio
)
356 while ((zt
= zio
->io_transform_stack
) != NULL
) {
357 if (zt
->zt_transform
!= NULL
)
358 zt
->zt_transform(zio
,
359 zt
->zt_orig_abd
, zt
->zt_orig_size
);
361 if (zt
->zt_bufsize
!= 0)
362 abd_free(zio
->io_abd
);
364 zio
->io_abd
= zt
->zt_orig_abd
;
365 zio
->io_size
= zt
->zt_orig_size
;
366 zio
->io_transform_stack
= zt
->zt_next
;
368 kmem_free(zt
, sizeof (zio_transform_t
));
373 * ==========================================================================
374 * I/O transform callbacks for subblocks, decompression, and decryption
375 * ==========================================================================
378 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
380 ASSERT(zio
->io_size
> size
);
382 if (zio
->io_type
== ZIO_TYPE_READ
)
383 abd_copy(data
, zio
->io_abd
, size
);
387 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
389 if (zio
->io_error
== 0) {
390 void *tmp
= abd_borrow_buf(data
, size
);
391 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
392 zio
->io_abd
, tmp
, zio
->io_size
, size
);
393 abd_return_buf_copy(data
, tmp
, size
);
395 if (zio_injection_enabled
&& ret
== 0)
396 ret
= zio_handle_fault_injection(zio
, EINVAL
);
399 zio
->io_error
= SET_ERROR(EIO
);
404 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
408 blkptr_t
*bp
= zio
->io_bp
;
409 spa_t
*spa
= zio
->io_spa
;
410 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
411 uint64_t lsize
= BP_GET_LSIZE(bp
);
412 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
413 uint8_t salt
[ZIO_DATA_SALT_LEN
];
414 uint8_t iv
[ZIO_DATA_IV_LEN
];
415 uint8_t mac
[ZIO_DATA_MAC_LEN
];
416 boolean_t no_crypt
= B_FALSE
;
418 ASSERT(BP_USES_CRYPT(bp
));
419 ASSERT3U(size
, !=, 0);
421 if (zio
->io_error
!= 0)
425 * Verify the cksum of MACs stored in an indirect bp. It will always
426 * be possible to verify this since it does not require an encryption
429 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
430 zio_crypt_decode_mac_bp(bp
, mac
);
432 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
434 * We haven't decompressed the data yet, but
435 * zio_crypt_do_indirect_mac_checksum() requires
436 * decompressed data to be able to parse out the MACs
437 * from the indirect block. We decompress it now and
438 * throw away the result after we are finished.
440 tmp
= zio_buf_alloc(lsize
);
441 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
442 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
444 ret
= SET_ERROR(EIO
);
447 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
448 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
449 zio_buf_free(tmp
, lsize
);
451 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
452 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
454 abd_copy(data
, zio
->io_abd
, size
);
456 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
457 ret
= zio_handle_decrypt_injection(spa
,
458 &zio
->io_bookmark
, ot
, ECKSUM
);
467 * If this is an authenticated block, just check the MAC. It would be
468 * nice to separate this out into its own flag, but for the moment
469 * enum zio_flag is out of bits.
471 if (BP_IS_AUTHENTICATED(bp
)) {
472 if (ot
== DMU_OT_OBJSET
) {
473 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
474 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
476 zio_crypt_decode_mac_bp(bp
, mac
);
477 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
478 zio
->io_abd
, size
, mac
);
479 if (zio_injection_enabled
&& ret
== 0) {
480 ret
= zio_handle_decrypt_injection(spa
,
481 &zio
->io_bookmark
, ot
, ECKSUM
);
484 abd_copy(data
, zio
->io_abd
, size
);
492 zio_crypt_decode_params_bp(bp
, salt
, iv
);
494 if (ot
== DMU_OT_INTENT_LOG
) {
495 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
496 zio_crypt_decode_mac_zil(tmp
, mac
);
497 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
499 zio_crypt_decode_mac_bp(bp
, mac
);
502 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
503 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
504 zio
->io_abd
, &no_crypt
);
506 abd_copy(data
, zio
->io_abd
, size
);
514 /* assert that the key was found unless this was speculative */
515 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
518 * If there was a decryption / authentication error return EIO as
519 * the io_error. If this was not a speculative zio, create an ereport.
522 zio
->io_error
= SET_ERROR(EIO
);
523 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
524 spa_log_error(spa
, &zio
->io_bookmark
);
525 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
526 spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
534 * ==========================================================================
535 * I/O parent/child relationships and pipeline interlocks
536 * ==========================================================================
539 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
541 list_t
*pl
= &cio
->io_parent_list
;
543 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
547 ASSERT((*zl
)->zl_child
== cio
);
548 return ((*zl
)->zl_parent
);
552 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
554 list_t
*cl
= &pio
->io_child_list
;
556 ASSERT(MUTEX_HELD(&pio
->io_lock
));
558 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
562 ASSERT((*zl
)->zl_parent
== pio
);
563 return ((*zl
)->zl_child
);
567 zio_unique_parent(zio_t
*cio
)
569 zio_link_t
*zl
= NULL
;
570 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
572 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
577 zio_add_child(zio_t
*pio
, zio_t
*cio
)
579 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
582 * Logical I/Os can have logical, gang, or vdev children.
583 * Gang I/Os can have gang or vdev children.
584 * Vdev I/Os can only have vdev children.
585 * The following ASSERT captures all of these constraints.
587 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
592 mutex_enter(&pio
->io_lock
);
593 mutex_enter(&cio
->io_lock
);
595 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
597 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
598 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
600 list_insert_head(&pio
->io_child_list
, zl
);
601 list_insert_head(&cio
->io_parent_list
, zl
);
603 pio
->io_child_count
++;
604 cio
->io_parent_count
++;
606 mutex_exit(&cio
->io_lock
);
607 mutex_exit(&pio
->io_lock
);
611 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
613 ASSERT(zl
->zl_parent
== pio
);
614 ASSERT(zl
->zl_child
== cio
);
616 mutex_enter(&pio
->io_lock
);
617 mutex_enter(&cio
->io_lock
);
619 list_remove(&pio
->io_child_list
, zl
);
620 list_remove(&cio
->io_parent_list
, zl
);
622 pio
->io_child_count
--;
623 cio
->io_parent_count
--;
625 mutex_exit(&cio
->io_lock
);
626 mutex_exit(&pio
->io_lock
);
627 kmem_cache_free(zio_link_cache
, zl
);
631 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
633 boolean_t waiting
= B_FALSE
;
635 mutex_enter(&zio
->io_lock
);
636 ASSERT(zio
->io_stall
== NULL
);
637 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
638 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
641 uint64_t *countp
= &zio
->io_children
[c
][wait
];
644 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
645 zio
->io_stall
= countp
;
650 mutex_exit(&zio
->io_lock
);
654 __attribute__((always_inline
))
656 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
657 zio_t
**next_to_executep
)
659 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
660 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
662 mutex_enter(&pio
->io_lock
);
663 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
664 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
665 pio
->io_reexecute
|= zio
->io_reexecute
;
666 ASSERT3U(*countp
, >, 0);
670 if (*countp
== 0 && pio
->io_stall
== countp
) {
671 zio_taskq_type_t type
=
672 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
674 pio
->io_stall
= NULL
;
675 mutex_exit(&pio
->io_lock
);
678 * If we can tell the caller to execute this parent next, do
679 * so. Otherwise dispatch the parent zio as its own task.
681 * Having the caller execute the parent when possible reduces
682 * locking on the zio taskq's, reduces context switch
683 * overhead, and has no recursion penalty. Note that one
684 * read from disk typically causes at least 3 zio's: a
685 * zio_null(), the logical zio_read(), and then a physical
686 * zio. When the physical ZIO completes, we are able to call
687 * zio_done() on all 3 of these zio's from one invocation of
688 * zio_execute() by returning the parent back to
689 * zio_execute(). Since the parent isn't executed until this
690 * thread returns back to zio_execute(), the caller should do
693 * In other cases, dispatching the parent prevents
694 * overflowing the stack when we have deeply nested
695 * parent-child relationships, as we do with the "mega zio"
696 * of writes for spa_sync(), and the chain of ZIL blocks.
698 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
) {
699 *next_to_executep
= pio
;
701 zio_taskq_dispatch(pio
, type
, B_FALSE
);
704 mutex_exit(&pio
->io_lock
);
709 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
711 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
712 zio
->io_error
= zio
->io_child_error
[c
];
716 zio_bookmark_compare(const void *x1
, const void *x2
)
718 const zio_t
*z1
= x1
;
719 const zio_t
*z2
= x2
;
721 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
723 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
726 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
728 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
731 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
733 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
736 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
738 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
750 * ==========================================================================
751 * Create the various types of I/O (read, write, free, etc)
752 * ==========================================================================
755 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
756 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
757 void *private, zio_type_t type
, zio_priority_t priority
,
758 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
759 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
760 enum zio_stage pipeline
)
764 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
765 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
766 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
768 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
769 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
770 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
772 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
774 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
775 bzero(zio
, sizeof (zio_t
));
777 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
778 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
780 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
781 offsetof(zio_link_t
, zl_parent_node
));
782 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
783 offsetof(zio_link_t
, zl_child_node
));
784 metaslab_trace_init(&zio
->io_alloc_list
);
787 zio
->io_child_type
= ZIO_CHILD_VDEV
;
788 else if (flags
& ZIO_FLAG_GANG_CHILD
)
789 zio
->io_child_type
= ZIO_CHILD_GANG
;
790 else if (flags
& ZIO_FLAG_DDT_CHILD
)
791 zio
->io_child_type
= ZIO_CHILD_DDT
;
793 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
796 zio
->io_bp
= (blkptr_t
*)bp
;
797 zio
->io_bp_copy
= *bp
;
798 zio
->io_bp_orig
= *bp
;
799 if (type
!= ZIO_TYPE_WRITE
||
800 zio
->io_child_type
== ZIO_CHILD_DDT
)
801 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
802 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
803 zio
->io_logical
= zio
;
804 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
805 pipeline
|= ZIO_GANG_STAGES
;
811 zio
->io_private
= private;
813 zio
->io_priority
= priority
;
815 zio
->io_offset
= offset
;
816 zio
->io_orig_abd
= zio
->io_abd
= data
;
817 zio
->io_orig_size
= zio
->io_size
= psize
;
818 zio
->io_lsize
= lsize
;
819 zio
->io_orig_flags
= zio
->io_flags
= flags
;
820 zio
->io_orig_stage
= zio
->io_stage
= stage
;
821 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
822 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
824 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
825 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
828 zio
->io_bookmark
= *zb
;
831 if (zio
->io_metaslab_class
== NULL
)
832 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
833 if (zio
->io_logical
== NULL
)
834 zio
->io_logical
= pio
->io_logical
;
835 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
836 zio
->io_gang_leader
= pio
->io_gang_leader
;
837 zio_add_child(pio
, zio
);
840 taskq_init_ent(&zio
->io_tqent
);
846 zio_destroy(zio_t
*zio
)
848 metaslab_trace_fini(&zio
->io_alloc_list
);
849 list_destroy(&zio
->io_parent_list
);
850 list_destroy(&zio
->io_child_list
);
851 mutex_destroy(&zio
->io_lock
);
852 cv_destroy(&zio
->io_cv
);
853 kmem_cache_free(zio_cache
, zio
);
857 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
858 void *private, enum zio_flag flags
)
862 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
863 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
864 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
870 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
872 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
876 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
878 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
879 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
880 bp
, (longlong_t
)BP_GET_TYPE(bp
));
882 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
883 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
884 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
885 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
887 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
888 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
889 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
890 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
892 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
893 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
894 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
896 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
897 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
898 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
901 if (BP_IS_EMBEDDED(bp
)) {
902 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
903 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
904 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
909 * Do not verify individual DVAs if the config is not trusted. This
910 * will be done once the zio is executed in vdev_mirror_map_alloc.
912 if (!spa
->spa_trust_config
)
916 * Pool-specific checks.
918 * Note: it would be nice to verify that the blk_birth and
919 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
920 * allows the birth time of log blocks (and dmu_sync()-ed blocks
921 * that are in the log) to be arbitrarily large.
923 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
924 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
926 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
927 zfs_panic_recover("blkptr at %p DVA %u has invalid "
929 bp
, i
, (longlong_t
)vdevid
);
932 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
934 zfs_panic_recover("blkptr at %p DVA %u has invalid "
936 bp
, i
, (longlong_t
)vdevid
);
939 if (vd
->vdev_ops
== &vdev_hole_ops
) {
940 zfs_panic_recover("blkptr at %p DVA %u has hole "
942 bp
, i
, (longlong_t
)vdevid
);
945 if (vd
->vdev_ops
== &vdev_missing_ops
) {
947 * "missing" vdevs are valid during import, but we
948 * don't have their detailed info (e.g. asize), so
949 * we can't perform any more checks on them.
953 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
954 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
956 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
957 if (offset
+ asize
> vd
->vdev_asize
) {
958 zfs_panic_recover("blkptr at %p DVA %u has invalid "
960 bp
, i
, (longlong_t
)offset
);
966 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
968 uint64_t vdevid
= DVA_GET_VDEV(dva
);
970 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
973 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
977 if (vd
->vdev_ops
== &vdev_hole_ops
)
980 if (vd
->vdev_ops
== &vdev_missing_ops
) {
984 uint64_t offset
= DVA_GET_OFFSET(dva
);
985 uint64_t asize
= DVA_GET_ASIZE(dva
);
988 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
989 if (offset
+ asize
> vd
->vdev_asize
)
996 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
997 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
998 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
1002 zfs_blkptr_verify(spa
, bp
);
1004 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
1005 data
, size
, size
, done
, private,
1006 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1007 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1008 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1014 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1015 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1016 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1017 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
1018 void *private, zio_priority_t priority
, enum zio_flag flags
,
1019 const zbookmark_phys_t
*zb
)
1023 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1024 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1025 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1026 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1027 DMU_OT_IS_VALID(zp
->zp_type
) &&
1028 zp
->zp_level
< 32 &&
1029 zp
->zp_copies
> 0 &&
1030 zp
->zp_copies
<= spa_max_replication(spa
));
1032 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1033 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1034 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1035 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1037 zio
->io_ready
= ready
;
1038 zio
->io_children_ready
= children_ready
;
1039 zio
->io_physdone
= physdone
;
1043 * Data can be NULL if we are going to call zio_write_override() to
1044 * provide the already-allocated BP. But we may need the data to
1045 * verify a dedup hit (if requested). In this case, don't try to
1046 * dedup (just take the already-allocated BP verbatim). Encrypted
1047 * dedup blocks need data as well so we also disable dedup in this
1051 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1052 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1059 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1060 uint64_t size
, zio_done_func_t
*done
, void *private,
1061 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
1065 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1066 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1067 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1073 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1075 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1076 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1077 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1078 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1081 * We must reset the io_prop to match the values that existed
1082 * when the bp was first written by dmu_sync() keeping in mind
1083 * that nopwrite and dedup are mutually exclusive.
1085 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1086 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1087 zio
->io_prop
.zp_copies
= copies
;
1088 zio
->io_bp_override
= bp
;
1092 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1095 zfs_blkptr_verify(spa
, bp
);
1098 * The check for EMBEDDED is a performance optimization. We
1099 * process the free here (by ignoring it) rather than
1100 * putting it on the list and then processing it in zio_free_sync().
1102 if (BP_IS_EMBEDDED(bp
))
1104 metaslab_check_free(spa
, bp
);
1107 * Frees that are for the currently-syncing txg, are not going to be
1108 * deferred, and which will not need to do a read (i.e. not GANG or
1109 * DEDUP), can be processed immediately. Otherwise, put them on the
1110 * in-memory list for later processing.
1112 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
1113 txg
!= spa
->spa_syncing_txg
||
1114 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
1115 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1117 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1122 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1123 enum zio_flag flags
)
1126 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1128 ASSERT(!BP_IS_HOLE(bp
));
1129 ASSERT(spa_syncing_txg(spa
) == txg
);
1130 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
1132 if (BP_IS_EMBEDDED(bp
))
1133 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1135 metaslab_check_free(spa
, bp
);
1137 dsl_scan_freed(spa
, bp
);
1140 * GANG and DEDUP blocks can induce a read (for the gang block header,
1141 * or the DDT), so issue them asynchronously so that this thread is
1144 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1145 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1147 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1148 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1149 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1155 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1156 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1160 zfs_blkptr_verify(spa
, bp
);
1162 if (BP_IS_EMBEDDED(bp
))
1163 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1166 * A claim is an allocation of a specific block. Claims are needed
1167 * to support immediate writes in the intent log. The issue is that
1168 * immediate writes contain committed data, but in a txg that was
1169 * *not* committed. Upon opening the pool after an unclean shutdown,
1170 * the intent log claims all blocks that contain immediate write data
1171 * so that the SPA knows they're in use.
1173 * All claims *must* be resolved in the first txg -- before the SPA
1174 * starts allocating blocks -- so that nothing is allocated twice.
1175 * If txg == 0 we just verify that the block is claimable.
1177 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
1178 spa_min_claim_txg(spa
));
1179 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1180 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1182 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1183 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1184 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1185 ASSERT0(zio
->io_queued_timestamp
);
1191 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1192 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1197 if (vd
->vdev_children
== 0) {
1198 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1199 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1200 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1204 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1206 for (c
= 0; c
< vd
->vdev_children
; c
++)
1207 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1208 done
, private, flags
));
1215 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1216 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1217 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1221 ASSERT(vd
->vdev_children
== 0);
1222 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1223 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1224 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1226 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1227 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1228 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1230 zio
->io_prop
.zp_checksum
= checksum
;
1236 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1237 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1238 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1242 ASSERT(vd
->vdev_children
== 0);
1243 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1244 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1245 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1247 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1248 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1249 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1251 zio
->io_prop
.zp_checksum
= checksum
;
1253 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1255 * zec checksums are necessarily destructive -- they modify
1256 * the end of the write buffer to hold the verifier/checksum.
1257 * Therefore, we must make a local copy in case the data is
1258 * being written to multiple places in parallel.
1260 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1261 abd_copy(wbuf
, data
, size
);
1263 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1270 * Create a child I/O to do some work for us.
1273 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1274 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1275 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1277 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1281 * vdev child I/Os do not propagate their error to the parent.
1282 * Therefore, for correct operation the caller *must* check for
1283 * and handle the error in the child i/o's done callback.
1284 * The only exceptions are i/os that we don't care about
1285 * (OPTIONAL or REPAIR).
1287 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1290 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1292 * If we have the bp, then the child should perform the
1293 * checksum and the parent need not. This pushes error
1294 * detection as close to the leaves as possible and
1295 * eliminates redundant checksums in the interior nodes.
1297 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1298 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1301 if (vd
->vdev_ops
->vdev_op_leaf
) {
1302 ASSERT0(vd
->vdev_children
);
1303 offset
+= VDEV_LABEL_START_SIZE
;
1306 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1309 * If we've decided to do a repair, the write is not speculative --
1310 * even if the original read was.
1312 if (flags
& ZIO_FLAG_IO_REPAIR
)
1313 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1316 * If we're creating a child I/O that is not associated with a
1317 * top-level vdev, then the child zio is not an allocating I/O.
1318 * If this is a retried I/O then we ignore it since we will
1319 * have already processed the original allocating I/O.
1321 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1322 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1323 ASSERT(pio
->io_metaslab_class
!= NULL
);
1324 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1325 ASSERT(type
== ZIO_TYPE_WRITE
);
1326 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1327 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1328 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1329 pio
->io_child_type
== ZIO_CHILD_GANG
);
1331 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1335 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1336 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1337 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1338 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1340 zio
->io_physdone
= pio
->io_physdone
;
1341 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1342 zio
->io_logical
->io_phys_children
++;
1348 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1349 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1350 zio_done_func_t
*done
, void *private)
1354 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1356 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1357 data
, size
, size
, done
, private, type
, priority
,
1358 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1360 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1366 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1368 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1370 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1374 zio_shrink(zio_t
*zio
, uint64_t size
)
1376 ASSERT3P(zio
->io_executor
, ==, NULL
);
1377 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1378 ASSERT3U(size
, <=, zio
->io_size
);
1381 * We don't shrink for raidz because of problems with the
1382 * reconstruction when reading back less than the block size.
1383 * Note, BP_IS_RAIDZ() assumes no compression.
1385 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1386 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1387 /* we are not doing a raw write */
1388 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1389 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1394 * ==========================================================================
1395 * Prepare to read and write logical blocks
1396 * ==========================================================================
1400 zio_read_bp_init(zio_t
*zio
)
1402 blkptr_t
*bp
= zio
->io_bp
;
1404 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1406 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1408 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1409 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1410 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1411 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1412 psize
, psize
, zio_decompress
);
1415 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1416 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1417 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1418 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1419 psize
, psize
, zio_decrypt
);
1422 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1423 int psize
= BPE_GET_PSIZE(bp
);
1424 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1426 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1427 decode_embedded_bp_compressed(bp
, data
);
1428 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1430 ASSERT(!BP_IS_EMBEDDED(bp
));
1431 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1434 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1435 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1437 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1438 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1440 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1441 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1447 zio_write_bp_init(zio_t
*zio
)
1449 if (!IO_IS_ALLOCATING(zio
))
1452 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1454 if (zio
->io_bp_override
) {
1455 blkptr_t
*bp
= zio
->io_bp
;
1456 zio_prop_t
*zp
= &zio
->io_prop
;
1458 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1459 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1461 *bp
= *zio
->io_bp_override
;
1462 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1464 if (BP_IS_EMBEDDED(bp
))
1468 * If we've been overridden and nopwrite is set then
1469 * set the flag accordingly to indicate that a nopwrite
1470 * has already occurred.
1472 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1473 ASSERT(!zp
->zp_dedup
);
1474 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1475 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1479 ASSERT(!zp
->zp_nopwrite
);
1481 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1484 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1485 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1487 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1489 BP_SET_DEDUP(bp
, 1);
1490 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1495 * We were unable to handle this as an override bp, treat
1496 * it as a regular write I/O.
1498 zio
->io_bp_override
= NULL
;
1499 *bp
= zio
->io_bp_orig
;
1500 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1507 zio_write_compress(zio_t
*zio
)
1509 spa_t
*spa
= zio
->io_spa
;
1510 zio_prop_t
*zp
= &zio
->io_prop
;
1511 enum zio_compress compress
= zp
->zp_compress
;
1512 blkptr_t
*bp
= zio
->io_bp
;
1513 uint64_t lsize
= zio
->io_lsize
;
1514 uint64_t psize
= zio
->io_size
;
1518 * If our children haven't all reached the ready stage,
1519 * wait for them and then repeat this pipeline stage.
1521 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1522 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1526 if (!IO_IS_ALLOCATING(zio
))
1529 if (zio
->io_children_ready
!= NULL
) {
1531 * Now that all our children are ready, run the callback
1532 * associated with this zio in case it wants to modify the
1533 * data to be written.
1535 ASSERT3U(zp
->zp_level
, >, 0);
1536 zio
->io_children_ready(zio
);
1539 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1540 ASSERT(zio
->io_bp_override
== NULL
);
1542 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1544 * We're rewriting an existing block, which means we're
1545 * working on behalf of spa_sync(). For spa_sync() to
1546 * converge, it must eventually be the case that we don't
1547 * have to allocate new blocks. But compression changes
1548 * the blocksize, which forces a reallocate, and makes
1549 * convergence take longer. Therefore, after the first
1550 * few passes, stop compressing to ensure convergence.
1552 pass
= spa_sync_pass(spa
);
1554 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1555 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1556 ASSERT(!BP_GET_DEDUP(bp
));
1558 if (pass
>= zfs_sync_pass_dont_compress
)
1559 compress
= ZIO_COMPRESS_OFF
;
1561 /* Make sure someone doesn't change their mind on overwrites */
1562 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1563 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1566 /* If it's a compressed write that is not raw, compress the buffer. */
1567 if (compress
!= ZIO_COMPRESS_OFF
&&
1568 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1569 void *cbuf
= zio_buf_alloc(lsize
);
1570 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1571 if (psize
== 0 || psize
== lsize
) {
1572 compress
= ZIO_COMPRESS_OFF
;
1573 zio_buf_free(cbuf
, lsize
);
1574 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1575 psize
<= BPE_PAYLOAD_SIZE
&&
1576 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1577 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1578 encode_embedded_bp_compressed(bp
,
1579 cbuf
, compress
, lsize
, psize
);
1580 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1581 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1582 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1583 zio_buf_free(cbuf
, lsize
);
1584 bp
->blk_birth
= zio
->io_txg
;
1585 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1586 ASSERT(spa_feature_is_active(spa
,
1587 SPA_FEATURE_EMBEDDED_DATA
));
1591 * Round up compressed size up to the ashift
1592 * of the smallest-ashift device, and zero the tail.
1593 * This ensures that the compressed size of the BP
1594 * (and thus compressratio property) are correct,
1595 * in that we charge for the padding used to fill out
1598 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1599 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1600 1ULL << spa
->spa_min_ashift
);
1601 if (rounded
>= lsize
) {
1602 compress
= ZIO_COMPRESS_OFF
;
1603 zio_buf_free(cbuf
, lsize
);
1606 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1607 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1608 abd_zero_off(cdata
, psize
, rounded
- psize
);
1610 zio_push_transform(zio
, cdata
,
1611 psize
, lsize
, NULL
);
1616 * We were unable to handle this as an override bp, treat
1617 * it as a regular write I/O.
1619 zio
->io_bp_override
= NULL
;
1620 *bp
= zio
->io_bp_orig
;
1621 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1623 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1624 zp
->zp_type
== DMU_OT_DNODE
) {
1626 * The DMU actually relies on the zio layer's compression
1627 * to free metadnode blocks that have had all contained
1628 * dnodes freed. As a result, even when doing a raw
1629 * receive, we must check whether the block can be compressed
1632 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1633 zio
->io_abd
, NULL
, lsize
);
1635 compress
= ZIO_COMPRESS_OFF
;
1637 ASSERT3U(psize
, !=, 0);
1641 * The final pass of spa_sync() must be all rewrites, but the first
1642 * few passes offer a trade-off: allocating blocks defers convergence,
1643 * but newly allocated blocks are sequential, so they can be written
1644 * to disk faster. Therefore, we allow the first few passes of
1645 * spa_sync() to allocate new blocks, but force rewrites after that.
1646 * There should only be a handful of blocks after pass 1 in any case.
1648 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1649 BP_GET_PSIZE(bp
) == psize
&&
1650 pass
>= zfs_sync_pass_rewrite
) {
1651 VERIFY3U(psize
, !=, 0);
1652 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1654 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1655 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1658 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1662 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1663 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1664 BP_SET_LSIZE(bp
, lsize
);
1665 BP_SET_TYPE(bp
, zp
->zp_type
);
1666 BP_SET_LEVEL(bp
, zp
->zp_level
);
1667 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1669 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1671 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1672 BP_SET_LSIZE(bp
, lsize
);
1673 BP_SET_TYPE(bp
, zp
->zp_type
);
1674 BP_SET_LEVEL(bp
, zp
->zp_level
);
1675 BP_SET_PSIZE(bp
, psize
);
1676 BP_SET_COMPRESS(bp
, compress
);
1677 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1678 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1679 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1681 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1682 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1683 ASSERT(!zp
->zp_encrypt
||
1684 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1685 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1687 if (zp
->zp_nopwrite
) {
1688 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1689 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1690 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1697 zio_free_bp_init(zio_t
*zio
)
1699 blkptr_t
*bp
= zio
->io_bp
;
1701 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1702 if (BP_GET_DEDUP(bp
))
1703 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1706 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1712 * ==========================================================================
1713 * Execute the I/O pipeline
1714 * ==========================================================================
1718 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1720 spa_t
*spa
= zio
->io_spa
;
1721 zio_type_t t
= zio
->io_type
;
1722 int flags
= (cutinline
? TQ_FRONT
: 0);
1725 * If we're a config writer or a probe, the normal issue and
1726 * interrupt threads may all be blocked waiting for the config lock.
1727 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1729 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1733 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1735 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1739 * If this is a high priority I/O, then use the high priority taskq if
1742 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1743 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1744 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1747 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1750 * NB: We are assuming that the zio can only be dispatched
1751 * to a single taskq at a time. It would be a grievous error
1752 * to dispatch the zio to another taskq at the same time.
1754 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1755 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1756 flags
, &zio
->io_tqent
);
1760 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1762 kthread_t
*executor
= zio
->io_executor
;
1763 spa_t
*spa
= zio
->io_spa
;
1765 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1766 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1768 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1769 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1778 zio_issue_async(zio_t
*zio
)
1780 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1786 zio_interrupt(zio_t
*zio
)
1788 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1792 zio_delay_interrupt(zio_t
*zio
)
1795 * The timeout_generic() function isn't defined in userspace, so
1796 * rather than trying to implement the function, the zio delay
1797 * functionality has been disabled for userspace builds.
1802 * If io_target_timestamp is zero, then no delay has been registered
1803 * for this IO, thus jump to the end of this function and "skip" the
1804 * delay; issuing it directly to the zio layer.
1806 if (zio
->io_target_timestamp
!= 0) {
1807 hrtime_t now
= gethrtime();
1809 if (now
>= zio
->io_target_timestamp
) {
1811 * This IO has already taken longer than the target
1812 * delay to complete, so we don't want to delay it
1813 * any longer; we "miss" the delay and issue it
1814 * directly to the zio layer. This is likely due to
1815 * the target latency being set to a value less than
1816 * the underlying hardware can satisfy (e.g. delay
1817 * set to 1ms, but the disks take 10ms to complete an
1821 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1827 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1828 clock_t expire_at_tick
= ddi_get_lbolt() +
1831 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1832 hrtime_t
, now
, hrtime_t
, diff
);
1834 if (NSEC_TO_TICK(diff
) == 0) {
1835 /* Our delay is less than a jiffy - just spin */
1836 zfs_sleep_until(zio
->io_target_timestamp
);
1839 * Use taskq_dispatch_delay() in the place of
1840 * OpenZFS's timeout_generic().
1842 tid
= taskq_dispatch_delay(system_taskq
,
1843 (task_func_t
*)zio_interrupt
,
1844 zio
, TQ_NOSLEEP
, expire_at_tick
);
1845 if (tid
== TASKQID_INVALID
) {
1847 * Couldn't allocate a task. Just
1848 * finish the zio without a delay.
1857 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1862 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
1864 zio_t
*cio
, *cio_next
;
1865 zio_link_t
*zl
= NULL
;
1866 vdev_t
*vd
= pio
->io_vd
;
1868 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
1869 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
1870 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1871 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1872 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1874 zfs_dbgmsg("slow zio[%d]: zio=%p timestamp=%llu "
1875 "delta=%llu queued=%llu io=%llu "
1876 "path=%s last=%llu "
1877 "type=%d priority=%d flags=0x%x "
1878 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1879 "objset=%llu object=%llu level=%llu blkid=%llu "
1880 "offset=%llu size=%llu error=%d",
1881 ziodepth
, pio
, pio
->io_timestamp
,
1882 delta
, pio
->io_delta
, pio
->io_delay
,
1883 vd
? vd
->vdev_path
: "NULL", vq
? vq
->vq_io_complete_ts
: 0,
1884 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
1885 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
1886 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
1887 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
1888 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
1889 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
1891 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
1892 taskq_empty_ent(&pio
->io_tqent
)) {
1897 mutex_enter(&pio
->io_lock
);
1898 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1899 cio_next
= zio_walk_children(pio
, &zl
);
1900 zio_deadman_impl(cio
, ziodepth
+ 1);
1902 mutex_exit(&pio
->io_lock
);
1906 * Log the critical information describing this zio and all of its children
1907 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1910 zio_deadman(zio_t
*pio
, char *tag
)
1912 spa_t
*spa
= pio
->io_spa
;
1913 char *name
= spa_name(spa
);
1915 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
1918 zio_deadman_impl(pio
, 0);
1920 switch (spa_get_deadman_failmode(spa
)) {
1921 case ZIO_FAILURE_MODE_WAIT
:
1922 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
1925 case ZIO_FAILURE_MODE_CONTINUE
:
1926 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
1929 case ZIO_FAILURE_MODE_PANIC
:
1930 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
1936 * Execute the I/O pipeline until one of the following occurs:
1937 * (1) the I/O completes; (2) the pipeline stalls waiting for
1938 * dependent child I/Os; (3) the I/O issues, so we're waiting
1939 * for an I/O completion interrupt; (4) the I/O is delegated by
1940 * vdev-level caching or aggregation; (5) the I/O is deferred
1941 * due to vdev-level queueing; (6) the I/O is handed off to
1942 * another thread. In all cases, the pipeline stops whenever
1943 * there's no CPU work; it never burns a thread in cv_wait_io().
1945 * There's no locking on io_stage because there's no legitimate way
1946 * for multiple threads to be attempting to process the same I/O.
1948 static zio_pipe_stage_t
*zio_pipeline
[];
1951 * zio_execute() is a wrapper around the static function
1952 * __zio_execute() so that we can force __zio_execute() to be
1953 * inlined. This reduces stack overhead which is important
1954 * because __zio_execute() is called recursively in several zio
1955 * code paths. zio_execute() itself cannot be inlined because
1956 * it is externally visible.
1959 zio_execute(zio_t
*zio
)
1961 fstrans_cookie_t cookie
;
1963 cookie
= spl_fstrans_mark();
1965 spl_fstrans_unmark(cookie
);
1969 * Used to determine if in the current context the stack is sized large
1970 * enough to allow zio_execute() to be called recursively. A minimum
1971 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1974 zio_execute_stack_check(zio_t
*zio
)
1976 #if !defined(HAVE_LARGE_STACKS)
1977 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1979 /* Executing in txg_sync_thread() context. */
1980 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1983 /* Pool initialization outside of zio_taskq context. */
1984 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1985 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1986 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1988 #endif /* HAVE_LARGE_STACKS */
1993 __attribute__((always_inline
))
1995 __zio_execute(zio_t
*zio
)
1997 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1999 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
2000 enum zio_stage pipeline
= zio
->io_pipeline
;
2001 enum zio_stage stage
= zio
->io_stage
;
2003 zio
->io_executor
= curthread
;
2005 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
2006 ASSERT(ISP2(stage
));
2007 ASSERT(zio
->io_stall
== NULL
);
2011 } while ((stage
& pipeline
) == 0);
2013 ASSERT(stage
<= ZIO_STAGE_DONE
);
2016 * If we are in interrupt context and this pipeline stage
2017 * will grab a config lock that is held across I/O,
2018 * or may wait for an I/O that needs an interrupt thread
2019 * to complete, issue async to avoid deadlock.
2021 * For VDEV_IO_START, we cut in line so that the io will
2022 * be sent to disk promptly.
2024 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
2025 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
2026 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2027 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2028 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2033 * If the current context doesn't have large enough stacks
2034 * the zio must be issued asynchronously to prevent overflow.
2036 if (zio_execute_stack_check(zio
)) {
2037 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2038 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2039 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2043 zio
->io_stage
= stage
;
2044 zio
->io_pipeline_trace
|= zio
->io_stage
;
2047 * The zio pipeline stage returns the next zio to execute
2048 * (typically the same as this one), or NULL if we should
2051 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2060 * ==========================================================================
2061 * Initiate I/O, either sync or async
2062 * ==========================================================================
2065 zio_wait(zio_t
*zio
)
2067 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2070 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2071 ASSERT3P(zio
->io_executor
, ==, NULL
);
2073 zio
->io_waiter
= curthread
;
2074 ASSERT0(zio
->io_queued_timestamp
);
2075 zio
->io_queued_timestamp
= gethrtime();
2079 mutex_enter(&zio
->io_lock
);
2080 while (zio
->io_executor
!= NULL
) {
2081 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2082 ddi_get_lbolt() + timeout
);
2084 if (zfs_deadman_enabled
&& error
== -1 &&
2085 gethrtime() - zio
->io_queued_timestamp
>
2086 spa_deadman_ziotime(zio
->io_spa
)) {
2087 mutex_exit(&zio
->io_lock
);
2088 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2089 zio_deadman(zio
, FTAG
);
2090 mutex_enter(&zio
->io_lock
);
2093 mutex_exit(&zio
->io_lock
);
2095 error
= zio
->io_error
;
2102 zio_nowait(zio_t
*zio
)
2104 ASSERT3P(zio
->io_executor
, ==, NULL
);
2106 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2107 zio_unique_parent(zio
) == NULL
) {
2111 * This is a logical async I/O with no parent to wait for it.
2112 * We add it to the spa_async_root_zio "Godfather" I/O which
2113 * will ensure they complete prior to unloading the pool.
2115 spa_t
*spa
= zio
->io_spa
;
2117 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2120 zio_add_child(pio
, zio
);
2123 ASSERT0(zio
->io_queued_timestamp
);
2124 zio
->io_queued_timestamp
= gethrtime();
2129 * ==========================================================================
2130 * Reexecute, cancel, or suspend/resume failed I/O
2131 * ==========================================================================
2135 zio_reexecute(zio_t
*pio
)
2137 zio_t
*cio
, *cio_next
;
2139 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2140 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2141 ASSERT(pio
->io_gang_leader
== NULL
);
2142 ASSERT(pio
->io_gang_tree
== NULL
);
2144 pio
->io_flags
= pio
->io_orig_flags
;
2145 pio
->io_stage
= pio
->io_orig_stage
;
2146 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2147 pio
->io_reexecute
= 0;
2148 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2149 pio
->io_pipeline_trace
= 0;
2151 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2152 pio
->io_state
[w
] = 0;
2153 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2154 pio
->io_child_error
[c
] = 0;
2156 if (IO_IS_ALLOCATING(pio
))
2157 BP_ZERO(pio
->io_bp
);
2160 * As we reexecute pio's children, new children could be created.
2161 * New children go to the head of pio's io_child_list, however,
2162 * so we will (correctly) not reexecute them. The key is that
2163 * the remainder of pio's io_child_list, from 'cio_next' onward,
2164 * cannot be affected by any side effects of reexecuting 'cio'.
2166 zio_link_t
*zl
= NULL
;
2167 mutex_enter(&pio
->io_lock
);
2168 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2169 cio_next
= zio_walk_children(pio
, &zl
);
2170 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2171 pio
->io_children
[cio
->io_child_type
][w
]++;
2172 mutex_exit(&pio
->io_lock
);
2174 mutex_enter(&pio
->io_lock
);
2176 mutex_exit(&pio
->io_lock
);
2179 * Now that all children have been reexecuted, execute the parent.
2180 * We don't reexecute "The Godfather" I/O here as it's the
2181 * responsibility of the caller to wait on it.
2183 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2184 pio
->io_queued_timestamp
= gethrtime();
2190 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2192 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2193 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2194 "failure and the failure mode property for this pool "
2195 "is set to panic.", spa_name(spa
));
2197 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2198 "failure and has been suspended.\n", spa_name(spa
));
2200 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2203 mutex_enter(&spa
->spa_suspend_lock
);
2205 if (spa
->spa_suspend_zio_root
== NULL
)
2206 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2207 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2208 ZIO_FLAG_GODFATHER
);
2210 spa
->spa_suspended
= reason
;
2213 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2214 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2215 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2216 ASSERT(zio_unique_parent(zio
) == NULL
);
2217 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2218 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2221 mutex_exit(&spa
->spa_suspend_lock
);
2225 zio_resume(spa_t
*spa
)
2230 * Reexecute all previously suspended i/o.
2232 mutex_enter(&spa
->spa_suspend_lock
);
2233 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2234 cv_broadcast(&spa
->spa_suspend_cv
);
2235 pio
= spa
->spa_suspend_zio_root
;
2236 spa
->spa_suspend_zio_root
= NULL
;
2237 mutex_exit(&spa
->spa_suspend_lock
);
2243 return (zio_wait(pio
));
2247 zio_resume_wait(spa_t
*spa
)
2249 mutex_enter(&spa
->spa_suspend_lock
);
2250 while (spa_suspended(spa
))
2251 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2252 mutex_exit(&spa
->spa_suspend_lock
);
2256 * ==========================================================================
2259 * A gang block is a collection of small blocks that looks to the DMU
2260 * like one large block. When zio_dva_allocate() cannot find a block
2261 * of the requested size, due to either severe fragmentation or the pool
2262 * being nearly full, it calls zio_write_gang_block() to construct the
2263 * block from smaller fragments.
2265 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2266 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2267 * an indirect block: it's an array of block pointers. It consumes
2268 * only one sector and hence is allocatable regardless of fragmentation.
2269 * The gang header's bps point to its gang members, which hold the data.
2271 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2272 * as the verifier to ensure uniqueness of the SHA256 checksum.
2273 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2274 * not the gang header. This ensures that data block signatures (needed for
2275 * deduplication) are independent of how the block is physically stored.
2277 * Gang blocks can be nested: a gang member may itself be a gang block.
2278 * Thus every gang block is a tree in which root and all interior nodes are
2279 * gang headers, and the leaves are normal blocks that contain user data.
2280 * The root of the gang tree is called the gang leader.
2282 * To perform any operation (read, rewrite, free, claim) on a gang block,
2283 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2284 * in the io_gang_tree field of the original logical i/o by recursively
2285 * reading the gang leader and all gang headers below it. This yields
2286 * an in-core tree containing the contents of every gang header and the
2287 * bps for every constituent of the gang block.
2289 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2290 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2291 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2292 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2293 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2294 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2295 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2296 * of the gang header plus zio_checksum_compute() of the data to update the
2297 * gang header's blk_cksum as described above.
2299 * The two-phase assemble/issue model solves the problem of partial failure --
2300 * what if you'd freed part of a gang block but then couldn't read the
2301 * gang header for another part? Assembling the entire gang tree first
2302 * ensures that all the necessary gang header I/O has succeeded before
2303 * starting the actual work of free, claim, or write. Once the gang tree
2304 * is assembled, free and claim are in-memory operations that cannot fail.
2306 * In the event that a gang write fails, zio_dva_unallocate() walks the
2307 * gang tree to immediately free (i.e. insert back into the space map)
2308 * everything we've allocated. This ensures that we don't get ENOSPC
2309 * errors during repeated suspend/resume cycles due to a flaky device.
2311 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2312 * the gang tree, we won't modify the block, so we can safely defer the free
2313 * (knowing that the block is still intact). If we *can* assemble the gang
2314 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2315 * each constituent bp and we can allocate a new block on the next sync pass.
2317 * In all cases, the gang tree allows complete recovery from partial failure.
2318 * ==========================================================================
2322 zio_gang_issue_func_done(zio_t
*zio
)
2324 abd_put(zio
->io_abd
);
2328 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2334 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2335 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2336 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2337 &pio
->io_bookmark
));
2341 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2348 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2349 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2350 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2351 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2354 * As we rewrite each gang header, the pipeline will compute
2355 * a new gang block header checksum for it; but no one will
2356 * compute a new data checksum, so we do that here. The one
2357 * exception is the gang leader: the pipeline already computed
2358 * its data checksum because that stage precedes gang assembly.
2359 * (Presently, nothing actually uses interior data checksums;
2360 * this is just good hygiene.)
2362 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2363 abd_t
*buf
= abd_get_offset(data
, offset
);
2365 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2366 buf
, BP_GET_PSIZE(bp
));
2371 * If we are here to damage data for testing purposes,
2372 * leave the GBH alone so that we can detect the damage.
2374 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2375 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2377 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2378 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2379 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2380 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2388 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2391 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2392 ZIO_GANG_CHILD_FLAGS(pio
)));
2397 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2400 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2401 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2404 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2413 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2415 static zio_gang_node_t
*
2416 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2418 zio_gang_node_t
*gn
;
2420 ASSERT(*gnpp
== NULL
);
2422 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2423 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2430 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2432 zio_gang_node_t
*gn
= *gnpp
;
2434 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2435 ASSERT(gn
->gn_child
[g
] == NULL
);
2437 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2438 kmem_free(gn
, sizeof (*gn
));
2443 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2445 zio_gang_node_t
*gn
= *gnpp
;
2450 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2451 zio_gang_tree_free(&gn
->gn_child
[g
]);
2453 zio_gang_node_free(gnpp
);
2457 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2459 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2460 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2462 ASSERT(gio
->io_gang_leader
== gio
);
2463 ASSERT(BP_IS_GANG(bp
));
2465 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2466 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2467 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2471 zio_gang_tree_assemble_done(zio_t
*zio
)
2473 zio_t
*gio
= zio
->io_gang_leader
;
2474 zio_gang_node_t
*gn
= zio
->io_private
;
2475 blkptr_t
*bp
= zio
->io_bp
;
2477 ASSERT(gio
== zio_unique_parent(zio
));
2478 ASSERT(zio
->io_child_count
== 0);
2483 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2484 if (BP_SHOULD_BYTESWAP(bp
))
2485 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2487 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2488 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2489 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2491 abd_put(zio
->io_abd
);
2493 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2494 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2495 if (!BP_IS_GANG(gbp
))
2497 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2502 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2505 zio_t
*gio
= pio
->io_gang_leader
;
2508 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2509 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2510 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2513 * If you're a gang header, your data is in gn->gn_gbh.
2514 * If you're a gang member, your data is in 'data' and gn == NULL.
2516 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2519 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2521 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2522 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2523 if (BP_IS_HOLE(gbp
))
2525 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2527 offset
+= BP_GET_PSIZE(gbp
);
2531 if (gn
== gio
->io_gang_tree
)
2532 ASSERT3U(gio
->io_size
, ==, offset
);
2539 zio_gang_assemble(zio_t
*zio
)
2541 blkptr_t
*bp
= zio
->io_bp
;
2543 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2544 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2546 zio
->io_gang_leader
= zio
;
2548 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2554 zio_gang_issue(zio_t
*zio
)
2556 blkptr_t
*bp
= zio
->io_bp
;
2558 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2562 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2563 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2565 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2566 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2569 zio_gang_tree_free(&zio
->io_gang_tree
);
2571 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2577 zio_write_gang_member_ready(zio_t
*zio
)
2579 zio_t
*pio
= zio_unique_parent(zio
);
2580 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2581 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2583 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
);
2585 if (BP_IS_HOLE(zio
->io_bp
))
2588 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2590 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2591 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2592 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2593 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2594 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2596 mutex_enter(&pio
->io_lock
);
2597 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2598 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2599 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2600 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2601 DVA_SET_ASIZE(&pdva
[d
], asize
);
2603 mutex_exit(&pio
->io_lock
);
2607 zio_write_gang_done(zio_t
*zio
)
2610 * The io_abd field will be NULL for a zio with no data. The io_flags
2611 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2612 * check for it here as it is cleared in zio_ready.
2614 if (zio
->io_abd
!= NULL
)
2615 abd_put(zio
->io_abd
);
2619 zio_write_gang_block(zio_t
*pio
)
2621 spa_t
*spa
= pio
->io_spa
;
2622 metaslab_class_t
*mc
= spa_normal_class(spa
);
2623 blkptr_t
*bp
= pio
->io_bp
;
2624 zio_t
*gio
= pio
->io_gang_leader
;
2626 zio_gang_node_t
*gn
, **gnpp
;
2627 zio_gbh_phys_t
*gbh
;
2629 uint64_t txg
= pio
->io_txg
;
2630 uint64_t resid
= pio
->io_size
;
2632 int copies
= gio
->io_prop
.zp_copies
;
2636 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2639 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2640 * have a third copy.
2642 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2643 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2644 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2646 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2647 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2648 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2651 flags
|= METASLAB_ASYNC_ALLOC
;
2652 VERIFY(zfs_refcount_held(&mc
->mc_alloc_slots
[pio
->io_allocator
],
2656 * The logical zio has already placed a reservation for
2657 * 'copies' allocation slots but gang blocks may require
2658 * additional copies. These additional copies
2659 * (i.e. gbh_copies - copies) are guaranteed to succeed
2660 * since metaslab_class_throttle_reserve() always allows
2661 * additional reservations for gang blocks.
2663 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2664 pio
->io_allocator
, pio
, flags
));
2667 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2668 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2669 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2671 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2672 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2676 * If we failed to allocate the gang block header then
2677 * we remove any additional allocation reservations that
2678 * we placed here. The original reservation will
2679 * be removed when the logical I/O goes to the ready
2682 metaslab_class_throttle_unreserve(mc
,
2683 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2686 pio
->io_error
= error
;
2691 gnpp
= &gio
->io_gang_tree
;
2693 gnpp
= pio
->io_private
;
2694 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2697 gn
= zio_gang_node_alloc(gnpp
);
2699 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2700 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2703 * Create the gang header.
2705 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2706 zio_write_gang_done
, NULL
, pio
->io_priority
,
2707 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2710 * Create and nowait the gang children.
2712 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2713 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2715 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2717 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2718 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2719 zp
.zp_type
= DMU_OT_NONE
;
2721 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2722 zp
.zp_dedup
= B_FALSE
;
2723 zp
.zp_dedup_verify
= B_FALSE
;
2724 zp
.zp_nopwrite
= B_FALSE
;
2725 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2726 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2727 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2728 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2729 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2731 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2732 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2733 resid
) : NULL
, lsize
, lsize
, &zp
,
2734 zio_write_gang_member_ready
, NULL
, NULL
,
2735 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2736 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2738 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2739 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2743 * Gang children won't throttle but we should
2744 * account for their work, so reserve an allocation
2745 * slot for them here.
2747 VERIFY(metaslab_class_throttle_reserve(mc
,
2748 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2754 * Set pio's pipeline to just wait for zio to finish.
2756 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2759 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2761 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2769 * The zio_nop_write stage in the pipeline determines if allocating a
2770 * new bp is necessary. The nopwrite feature can handle writes in
2771 * either syncing or open context (i.e. zil writes) and as a result is
2772 * mutually exclusive with dedup.
2774 * By leveraging a cryptographically secure checksum, such as SHA256, we
2775 * can compare the checksums of the new data and the old to determine if
2776 * allocating a new block is required. Note that our requirements for
2777 * cryptographic strength are fairly weak: there can't be any accidental
2778 * hash collisions, but we don't need to be secure against intentional
2779 * (malicious) collisions. To trigger a nopwrite, you have to be able
2780 * to write the file to begin with, and triggering an incorrect (hash
2781 * collision) nopwrite is no worse than simply writing to the file.
2782 * That said, there are no known attacks against the checksum algorithms
2783 * used for nopwrite, assuming that the salt and the checksums
2784 * themselves remain secret.
2787 zio_nop_write(zio_t
*zio
)
2789 blkptr_t
*bp
= zio
->io_bp
;
2790 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2791 zio_prop_t
*zp
= &zio
->io_prop
;
2793 ASSERT(BP_GET_LEVEL(bp
) == 0);
2794 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2795 ASSERT(zp
->zp_nopwrite
);
2796 ASSERT(!zp
->zp_dedup
);
2797 ASSERT(zio
->io_bp_override
== NULL
);
2798 ASSERT(IO_IS_ALLOCATING(zio
));
2801 * Check to see if the original bp and the new bp have matching
2802 * characteristics (i.e. same checksum, compression algorithms, etc).
2803 * If they don't then just continue with the pipeline which will
2804 * allocate a new bp.
2806 if (BP_IS_HOLE(bp_orig
) ||
2807 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2808 ZCHECKSUM_FLAG_NOPWRITE
) ||
2809 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2810 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2811 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2812 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2813 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2817 * If the checksums match then reset the pipeline so that we
2818 * avoid allocating a new bp and issuing any I/O.
2820 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2821 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2822 ZCHECKSUM_FLAG_NOPWRITE
);
2823 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2824 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2825 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2826 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2827 sizeof (uint64_t)) == 0);
2830 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2831 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2838 * ==========================================================================
2840 * ==========================================================================
2843 zio_ddt_child_read_done(zio_t
*zio
)
2845 blkptr_t
*bp
= zio
->io_bp
;
2846 ddt_entry_t
*dde
= zio
->io_private
;
2848 zio_t
*pio
= zio_unique_parent(zio
);
2850 mutex_enter(&pio
->io_lock
);
2851 ddp
= ddt_phys_select(dde
, bp
);
2852 if (zio
->io_error
== 0)
2853 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2855 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2856 dde
->dde_repair_abd
= zio
->io_abd
;
2858 abd_free(zio
->io_abd
);
2859 mutex_exit(&pio
->io_lock
);
2863 zio_ddt_read_start(zio_t
*zio
)
2865 blkptr_t
*bp
= zio
->io_bp
;
2867 ASSERT(BP_GET_DEDUP(bp
));
2868 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2869 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2871 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2872 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2873 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2874 ddt_phys_t
*ddp
= dde
->dde_phys
;
2875 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2878 ASSERT(zio
->io_vsd
== NULL
);
2881 if (ddp_self
== NULL
)
2884 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2885 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2887 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2889 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2890 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2891 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2892 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2893 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2898 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2899 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2900 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2906 zio_ddt_read_done(zio_t
*zio
)
2908 blkptr_t
*bp
= zio
->io_bp
;
2910 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
2914 ASSERT(BP_GET_DEDUP(bp
));
2915 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2916 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2918 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2919 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2920 ddt_entry_t
*dde
= zio
->io_vsd
;
2922 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2926 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2927 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2930 if (dde
->dde_repair_abd
!= NULL
) {
2931 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2933 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2935 ddt_repair_done(ddt
, dde
);
2939 ASSERT(zio
->io_vsd
== NULL
);
2945 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2947 spa_t
*spa
= zio
->io_spa
;
2948 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2950 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2953 * Note: we compare the original data, not the transformed data,
2954 * because when zio->io_bp is an override bp, we will not have
2955 * pushed the I/O transforms. That's an important optimization
2956 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2957 * However, we should never get a raw, override zio so in these
2958 * cases we can compare the io_abd directly. This is useful because
2959 * it allows us to do dedup verification even if we don't have access
2960 * to the original data (for instance, if the encryption keys aren't
2964 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2965 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2967 if (lio
!= NULL
&& do_raw
) {
2968 return (lio
->io_size
!= zio
->io_size
||
2969 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2970 } else if (lio
!= NULL
) {
2971 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2972 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2976 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2977 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2979 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2980 blkptr_t blk
= *zio
->io_bp
;
2985 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2986 psize
= BP_GET_PSIZE(&blk
);
2988 if (psize
!= zio
->io_size
)
2993 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2995 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2996 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2997 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2998 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
3001 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
3002 error
= SET_ERROR(ENOENT
);
3007 return (error
!= 0);
3008 } else if (ddp
->ddp_phys_birth
!= 0) {
3009 arc_buf_t
*abuf
= NULL
;
3010 arc_flags_t aflags
= ARC_FLAG_WAIT
;
3011 blkptr_t blk
= *zio
->io_bp
;
3014 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3016 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3021 error
= arc_read(NULL
, spa
, &blk
,
3022 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3023 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3024 &aflags
, &zio
->io_bookmark
);
3027 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3028 zio
->io_orig_size
) != 0)
3029 error
= SET_ERROR(ENOENT
);
3030 arc_buf_destroy(abuf
, &abuf
);
3034 return (error
!= 0);
3042 zio_ddt_child_write_ready(zio_t
*zio
)
3044 int p
= zio
->io_prop
.zp_copies
;
3045 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3046 ddt_entry_t
*dde
= zio
->io_private
;
3047 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3055 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3057 ddt_phys_fill(ddp
, zio
->io_bp
);
3059 zio_link_t
*zl
= NULL
;
3060 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3061 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3067 zio_ddt_child_write_done(zio_t
*zio
)
3069 int p
= zio
->io_prop
.zp_copies
;
3070 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3071 ddt_entry_t
*dde
= zio
->io_private
;
3072 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3076 ASSERT(ddp
->ddp_refcnt
== 0);
3077 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3078 dde
->dde_lead_zio
[p
] = NULL
;
3080 if (zio
->io_error
== 0) {
3081 zio_link_t
*zl
= NULL
;
3082 while (zio_walk_parents(zio
, &zl
) != NULL
)
3083 ddt_phys_addref(ddp
);
3085 ddt_phys_clear(ddp
);
3092 zio_ddt_ditto_write_done(zio_t
*zio
)
3094 int p
= DDT_PHYS_DITTO
;
3095 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
3096 blkptr_t
*bp
= zio
->io_bp
;
3097 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3098 ddt_entry_t
*dde
= zio
->io_private
;
3099 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3100 ddt_key_t
*ddk
= &dde
->dde_key
;
3104 ASSERT(ddp
->ddp_refcnt
== 0);
3105 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3106 dde
->dde_lead_zio
[p
] = NULL
;
3108 if (zio
->io_error
== 0) {
3109 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
3110 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
3111 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
3112 if (ddp
->ddp_phys_birth
!= 0)
3113 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
3114 ddt_phys_fill(ddp
, bp
);
3121 zio_ddt_write(zio_t
*zio
)
3123 spa_t
*spa
= zio
->io_spa
;
3124 blkptr_t
*bp
= zio
->io_bp
;
3125 uint64_t txg
= zio
->io_txg
;
3126 zio_prop_t
*zp
= &zio
->io_prop
;
3127 int p
= zp
->zp_copies
;
3131 ddt_t
*ddt
= ddt_select(spa
, bp
);
3135 ASSERT(BP_GET_DEDUP(bp
));
3136 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3137 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3138 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3141 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3142 ddp
= &dde
->dde_phys
[p
];
3144 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3146 * If we're using a weak checksum, upgrade to a strong checksum
3147 * and try again. If we're already using a strong checksum,
3148 * we can't resolve it, so just convert to an ordinary write.
3149 * (And automatically e-mail a paper to Nature?)
3151 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3152 ZCHECKSUM_FLAG_DEDUP
)) {
3153 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3154 zio_pop_transforms(zio
);
3155 zio
->io_stage
= ZIO_STAGE_OPEN
;
3158 zp
->zp_dedup
= B_FALSE
;
3160 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3165 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
3166 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
3168 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
3169 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
3170 zio_prop_t czp
= *zp
;
3172 czp
.zp_copies
= ditto_copies
;
3175 * If we arrived here with an override bp, we won't have run
3176 * the transform stack, so we won't have the data we need to
3177 * generate a child i/o. So, toss the override bp and restart.
3178 * This is safe, because using the override bp is just an
3179 * optimization; and it's rare, so the cost doesn't matter.
3181 if (zio
->io_bp_override
) {
3182 zio_pop_transforms(zio
);
3183 zio
->io_stage
= ZIO_STAGE_OPEN
;
3184 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3185 zio
->io_bp_override
= NULL
;
3191 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3192 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
3193 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
3194 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3196 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3197 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
3200 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3201 if (ddp
->ddp_phys_birth
!= 0)
3202 ddt_bp_fill(ddp
, bp
, txg
);
3203 if (dde
->dde_lead_zio
[p
] != NULL
)
3204 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3206 ddt_phys_addref(ddp
);
3207 } else if (zio
->io_bp_override
) {
3208 ASSERT(bp
->blk_birth
== txg
);
3209 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3210 ddt_phys_fill(ddp
, bp
);
3211 ddt_phys_addref(ddp
);
3213 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3214 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3215 zio_ddt_child_write_ready
, NULL
, NULL
,
3216 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3217 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3219 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3220 dde
->dde_lead_zio
[p
] = cio
;
3233 ddt_entry_t
*freedde
; /* for debugging */
3236 zio_ddt_free(zio_t
*zio
)
3238 spa_t
*spa
= zio
->io_spa
;
3239 blkptr_t
*bp
= zio
->io_bp
;
3240 ddt_t
*ddt
= ddt_select(spa
, bp
);
3244 ASSERT(BP_GET_DEDUP(bp
));
3245 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3248 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3250 ddp
= ddt_phys_select(dde
, bp
);
3252 ddt_phys_decref(ddp
);
3260 * ==========================================================================
3261 * Allocate and free blocks
3262 * ==========================================================================
3266 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3270 ASSERT(MUTEX_HELD(&spa
->spa_alloc_locks
[allocator
]));
3272 zio
= avl_first(&spa
->spa_alloc_trees
[allocator
]);
3276 ASSERT(IO_IS_ALLOCATING(zio
));
3279 * Try to place a reservation for this zio. If we're unable to
3280 * reserve then we throttle.
3282 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3283 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3284 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
, 0)) {
3288 avl_remove(&spa
->spa_alloc_trees
[allocator
], zio
);
3289 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3295 zio_dva_throttle(zio_t
*zio
)
3297 spa_t
*spa
= zio
->io_spa
;
3299 metaslab_class_t
*mc
;
3301 /* locate an appropriate allocation class */
3302 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3303 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3305 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3306 !mc
->mc_alloc_throttle_enabled
||
3307 zio
->io_child_type
== ZIO_CHILD_GANG
||
3308 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3312 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3314 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3315 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3317 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
3319 * We want to try to use as many allocators as possible to help improve
3320 * performance, but we also want logically adjacent IOs to be physically
3321 * adjacent to improve sequential read performance. We chunk each object
3322 * into 2^20 block regions, and then hash based on the objset, object,
3323 * level, and region to accomplish both of these goals.
3325 zio
->io_allocator
= cityhash4(bm
->zb_objset
, bm
->zb_object
,
3326 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
3327 mutex_enter(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3328 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3329 zio
->io_metaslab_class
= mc
;
3330 avl_add(&spa
->spa_alloc_trees
[zio
->io_allocator
], zio
);
3331 nio
= zio_io_to_allocate(spa
, zio
->io_allocator
);
3332 mutex_exit(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3337 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3341 mutex_enter(&spa
->spa_alloc_locks
[allocator
]);
3342 zio
= zio_io_to_allocate(spa
, allocator
);
3343 mutex_exit(&spa
->spa_alloc_locks
[allocator
]);
3347 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3348 ASSERT0(zio
->io_error
);
3349 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3353 zio_dva_allocate(zio_t
*zio
)
3355 spa_t
*spa
= zio
->io_spa
;
3356 metaslab_class_t
*mc
;
3357 blkptr_t
*bp
= zio
->io_bp
;
3361 if (zio
->io_gang_leader
== NULL
) {
3362 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3363 zio
->io_gang_leader
= zio
;
3366 ASSERT(BP_IS_HOLE(bp
));
3367 ASSERT0(BP_GET_NDVAS(bp
));
3368 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3369 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3370 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3372 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3373 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3374 flags
|= METASLAB_DONT_THROTTLE
;
3375 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3376 flags
|= METASLAB_GANG_CHILD
;
3377 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3378 flags
|= METASLAB_ASYNC_ALLOC
;
3381 * if not already chosen, locate an appropriate allocation class
3383 mc
= zio
->io_metaslab_class
;
3385 mc
= spa_preferred_class(spa
, zio
->io_size
,
3386 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3387 zio
->io_prop
.zp_zpl_smallblk
);
3388 zio
->io_metaslab_class
= mc
;
3391 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3392 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3393 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3396 * Fallback to normal class when an alloc class is full
3398 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3400 * If throttling, transfer reservation over to normal class.
3401 * The io_allocator slot can remain the same even though we
3402 * are switching classes.
3404 if (mc
->mc_alloc_throttle_enabled
&&
3405 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3406 metaslab_class_throttle_unreserve(mc
,
3407 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3408 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3410 mc
= spa_normal_class(spa
);
3411 VERIFY(metaslab_class_throttle_reserve(mc
,
3412 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3413 flags
| METASLAB_MUST_RESERVE
));
3415 mc
= spa_normal_class(spa
);
3417 zio
->io_metaslab_class
= mc
;
3419 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3420 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3421 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3425 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3426 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3428 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3429 return (zio_write_gang_block(zio
));
3430 zio
->io_error
= error
;
3437 zio_dva_free(zio_t
*zio
)
3439 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3445 zio_dva_claim(zio_t
*zio
)
3449 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3451 zio
->io_error
= error
;
3457 * Undo an allocation. This is used by zio_done() when an I/O fails
3458 * and we want to give back the block we just allocated.
3459 * This handles both normal blocks and gang blocks.
3462 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3464 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3465 ASSERT(zio
->io_bp_override
== NULL
);
3467 if (!BP_IS_HOLE(bp
))
3468 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3471 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3472 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3473 &gn
->gn_gbh
->zg_blkptr
[g
]);
3479 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3482 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3483 uint64_t size
, boolean_t
*slog
)
3486 zio_alloc_list_t io_alloc_list
;
3488 ASSERT(txg
> spa_syncing_txg(spa
));
3490 metaslab_trace_init(&io_alloc_list
);
3493 * Block pointer fields are useful to metaslabs for stats and debugging.
3494 * Fill in the obvious ones before calling into metaslab_alloc().
3496 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3497 BP_SET_PSIZE(new_bp
, size
);
3498 BP_SET_LEVEL(new_bp
, 0);
3501 * When allocating a zil block, we don't have information about
3502 * the final destination of the block except the objset it's part
3503 * of, so we just hash the objset ID to pick the allocator to get
3506 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3507 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
,
3508 cityhash4(0, 0, 0, os
->os_dsl_dataset
->ds_object
) %
3509 spa
->spa_alloc_count
);
3513 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3514 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3515 &io_alloc_list
, NULL
, cityhash4(0, 0, 0,
3516 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
);
3520 metaslab_trace_fini(&io_alloc_list
);
3523 BP_SET_LSIZE(new_bp
, size
);
3524 BP_SET_PSIZE(new_bp
, size
);
3525 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3526 BP_SET_CHECKSUM(new_bp
,
3527 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3528 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3529 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3530 BP_SET_LEVEL(new_bp
, 0);
3531 BP_SET_DEDUP(new_bp
, 0);
3532 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3535 * encrypted blocks will require an IV and salt. We generate
3536 * these now since we will not be rewriting the bp at
3539 if (os
->os_encrypted
) {
3540 uint8_t iv
[ZIO_DATA_IV_LEN
];
3541 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3543 BP_SET_CRYPT(new_bp
, B_TRUE
);
3544 VERIFY0(spa_crypt_get_salt(spa
,
3545 dmu_objset_id(os
), salt
));
3546 VERIFY0(zio_crypt_generate_iv(iv
));
3548 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3551 zfs_dbgmsg("%s: zil block allocation failure: "
3552 "size %llu, error %d", spa_name(spa
), size
, error
);
3559 * ==========================================================================
3560 * Read and write to physical devices
3561 * ==========================================================================
3566 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3567 * stops after this stage and will resume upon I/O completion.
3568 * However, there are instances where the vdev layer may need to
3569 * continue the pipeline when an I/O was not issued. Since the I/O
3570 * that was sent to the vdev layer might be different than the one
3571 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3572 * force the underlying vdev layers to call either zio_execute() or
3573 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3576 zio_vdev_io_start(zio_t
*zio
)
3578 vdev_t
*vd
= zio
->io_vd
;
3580 spa_t
*spa
= zio
->io_spa
;
3584 ASSERT(zio
->io_error
== 0);
3585 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3588 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3589 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3592 * The mirror_ops handle multiple DVAs in a single BP.
3594 vdev_mirror_ops
.vdev_op_io_start(zio
);
3598 ASSERT3P(zio
->io_logical
, !=, zio
);
3599 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3600 ASSERT(spa
->spa_trust_config
);
3603 * Note: the code can handle other kinds of writes,
3604 * but we don't expect them.
3606 if (zio
->io_vd
->vdev_removing
) {
3607 ASSERT(zio
->io_flags
&
3608 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3609 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3613 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3615 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3616 P2PHASE(zio
->io_size
, align
) != 0) {
3617 /* Transform logical writes to be a full physical block size. */
3618 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3619 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3620 ASSERT(vd
== vd
->vdev_top
);
3621 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3622 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3623 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3625 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3629 * If this is not a physical io, make sure that it is properly aligned
3630 * before proceeding.
3632 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3633 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3634 ASSERT0(P2PHASE(zio
->io_size
, align
));
3637 * For physical writes, we allow 512b aligned writes and assume
3638 * the device will perform a read-modify-write as necessary.
3640 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3641 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3644 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3647 * If this is a repair I/O, and there's no self-healing involved --
3648 * that is, we're just resilvering what we expect to resilver --
3649 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3650 * This prevents spurious resilvering.
3652 * There are a few ways that we can end up creating these spurious
3655 * 1. A resilver i/o will be issued if any DVA in the BP has a
3656 * dirty DTL. The mirror code will issue resilver writes to
3657 * each DVA, including the one(s) that are not on vdevs with dirty
3660 * 2. With nested replication, which happens when we have a
3661 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3662 * For example, given mirror(replacing(A+B), C), it's likely that
3663 * only A is out of date (it's the new device). In this case, we'll
3664 * read from C, then use the data to resilver A+B -- but we don't
3665 * actually want to resilver B, just A. The top-level mirror has no
3666 * way to know this, so instead we just discard unnecessary repairs
3667 * as we work our way down the vdev tree.
3669 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3670 * The same logic applies to any form of nested replication: ditto
3671 * + mirror, RAID-Z + replacing, etc.
3673 * However, indirect vdevs point off to other vdevs which may have
3674 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3675 * will be properly bypassed instead.
3677 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3678 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3679 zio
->io_txg
!= 0 && /* not a delegated i/o */
3680 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3681 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3682 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3683 zio_vdev_io_bypass(zio
);
3687 if (vd
->vdev_ops
->vdev_op_leaf
&&
3688 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3690 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3693 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3696 if (!vdev_accessible(vd
, zio
)) {
3697 zio
->io_error
= SET_ERROR(ENXIO
);
3701 zio
->io_delay
= gethrtime();
3704 vd
->vdev_ops
->vdev_op_io_start(zio
);
3709 zio_vdev_io_done(zio_t
*zio
)
3711 vdev_t
*vd
= zio
->io_vd
;
3712 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3713 boolean_t unexpected_error
= B_FALSE
;
3715 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3719 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3722 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3724 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3726 vdev_queue_io_done(zio
);
3728 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3729 vdev_cache_write(zio
);
3731 if (zio_injection_enabled
&& zio
->io_error
== 0)
3732 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3735 if (zio_injection_enabled
&& zio
->io_error
== 0)
3736 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3738 if (zio
->io_error
) {
3739 if (!vdev_accessible(vd
, zio
)) {
3740 zio
->io_error
= SET_ERROR(ENXIO
);
3742 unexpected_error
= B_TRUE
;
3747 ops
->vdev_op_io_done(zio
);
3749 if (unexpected_error
)
3750 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3756 * This function is used to change the priority of an existing zio that is
3757 * currently in-flight. This is used by the arc to upgrade priority in the
3758 * event that a demand read is made for a block that is currently queued
3759 * as a scrub or async read IO. Otherwise, the high priority read request
3760 * would end up having to wait for the lower priority IO.
3763 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3765 zio_t
*cio
, *cio_next
;
3766 zio_link_t
*zl
= NULL
;
3768 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3770 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3771 vdev_queue_change_io_priority(pio
, priority
);
3773 pio
->io_priority
= priority
;
3776 mutex_enter(&pio
->io_lock
);
3777 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3778 cio_next
= zio_walk_children(pio
, &zl
);
3779 zio_change_priority(cio
, priority
);
3781 mutex_exit(&pio
->io_lock
);
3785 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3786 * disk, and use that to finish the checksum ereport later.
3789 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3790 const abd_t
*good_buf
)
3792 /* no processing needed */
3793 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3798 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3800 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3802 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3804 zcr
->zcr_cbinfo
= zio
->io_size
;
3805 zcr
->zcr_cbdata
= abd
;
3806 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3807 zcr
->zcr_free
= zio_abd_free
;
3811 zio_vdev_io_assess(zio_t
*zio
)
3813 vdev_t
*vd
= zio
->io_vd
;
3815 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3819 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3820 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3822 if (zio
->io_vsd
!= NULL
) {
3823 zio
->io_vsd_ops
->vsd_free(zio
);
3827 if (zio_injection_enabled
&& zio
->io_error
== 0)
3828 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3831 * If the I/O failed, determine whether we should attempt to retry it.
3833 * On retry, we cut in line in the issue queue, since we don't want
3834 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3836 if (zio
->io_error
&& vd
== NULL
&&
3837 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3838 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3839 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3841 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3842 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3843 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3844 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3845 zio_requeue_io_start_cut_in_line
);
3850 * If we got an error on a leaf device, convert it to ENXIO
3851 * if the device is not accessible at all.
3853 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3854 !vdev_accessible(vd
, zio
))
3855 zio
->io_error
= SET_ERROR(ENXIO
);
3858 * If we can't write to an interior vdev (mirror or RAID-Z),
3859 * set vdev_cant_write so that we stop trying to allocate from it.
3861 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3862 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3863 vd
->vdev_cant_write
= B_TRUE
;
3867 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3868 * attempts will ever succeed. In this case we set a persistent bit so
3869 * that we don't bother with it in the future.
3871 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3872 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3873 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3874 vd
->vdev_nowritecache
= B_TRUE
;
3877 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3879 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3880 zio
->io_physdone
!= NULL
) {
3881 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3882 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3883 zio
->io_physdone(zio
->io_logical
);
3890 zio_vdev_io_reissue(zio_t
*zio
)
3892 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3893 ASSERT(zio
->io_error
== 0);
3895 zio
->io_stage
>>= 1;
3899 zio_vdev_io_redone(zio_t
*zio
)
3901 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3903 zio
->io_stage
>>= 1;
3907 zio_vdev_io_bypass(zio_t
*zio
)
3909 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3910 ASSERT(zio
->io_error
== 0);
3912 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3913 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3917 * ==========================================================================
3918 * Encrypt and store encryption parameters
3919 * ==========================================================================
3924 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3925 * managing the storage of encryption parameters and passing them to the
3926 * lower-level encryption functions.
3929 zio_encrypt(zio_t
*zio
)
3931 zio_prop_t
*zp
= &zio
->io_prop
;
3932 spa_t
*spa
= zio
->io_spa
;
3933 blkptr_t
*bp
= zio
->io_bp
;
3934 uint64_t psize
= BP_GET_PSIZE(bp
);
3935 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
3936 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3937 void *enc_buf
= NULL
;
3939 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3940 uint8_t iv
[ZIO_DATA_IV_LEN
];
3941 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3942 boolean_t no_crypt
= B_FALSE
;
3944 /* the root zio already encrypted the data */
3945 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3948 /* only ZIL blocks are re-encrypted on rewrite */
3949 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3952 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3953 BP_SET_CRYPT(bp
, B_FALSE
);
3957 /* if we are doing raw encryption set the provided encryption params */
3958 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3959 ASSERT0(BP_GET_LEVEL(bp
));
3960 BP_SET_CRYPT(bp
, B_TRUE
);
3961 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3962 if (ot
!= DMU_OT_OBJSET
)
3963 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3965 /* dnode blocks must be written out in the provided byteorder */
3966 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
3967 ot
== DMU_OT_DNODE
) {
3968 void *bswap_buf
= zio_buf_alloc(psize
);
3969 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
3971 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3972 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
3973 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
3976 abd_take_ownership_of_buf(babd
, B_TRUE
);
3977 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
3980 if (DMU_OT_IS_ENCRYPTED(ot
))
3981 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3985 /* indirect blocks only maintain a cksum of the lower level MACs */
3986 if (BP_GET_LEVEL(bp
) > 0) {
3987 BP_SET_CRYPT(bp
, B_TRUE
);
3988 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3989 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3991 zio_crypt_encode_mac_bp(bp
, mac
);
3996 * Objset blocks are a special case since they have 2 256-bit MACs
3997 * embedded within them.
3999 if (ot
== DMU_OT_OBJSET
) {
4000 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
4001 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4002 BP_SET_CRYPT(bp
, B_TRUE
);
4003 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
4004 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
4008 /* unencrypted object types are only authenticated with a MAC */
4009 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
4010 BP_SET_CRYPT(bp
, B_TRUE
);
4011 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
4012 zio
->io_abd
, psize
, mac
));
4013 zio_crypt_encode_mac_bp(bp
, mac
);
4018 * Later passes of sync-to-convergence may decide to rewrite data
4019 * in place to avoid more disk reallocations. This presents a problem
4020 * for encryption because this constitutes rewriting the new data with
4021 * the same encryption key and IV. However, this only applies to blocks
4022 * in the MOS (particularly the spacemaps) and we do not encrypt the
4023 * MOS. We assert that the zio is allocating or an intent log write
4026 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
4027 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
4028 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
4029 ASSERT3U(psize
, !=, 0);
4031 enc_buf
= zio_buf_alloc(psize
);
4032 eabd
= abd_get_from_buf(enc_buf
, psize
);
4033 abd_take_ownership_of_buf(eabd
, B_TRUE
);
4036 * For an explanation of what encryption parameters are stored
4037 * where, see the block comment in zio_crypt.c.
4039 if (ot
== DMU_OT_INTENT_LOG
) {
4040 zio_crypt_decode_params_bp(bp
, salt
, iv
);
4042 BP_SET_CRYPT(bp
, B_TRUE
);
4045 /* Perform the encryption. This should not fail */
4046 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
4047 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
4048 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
4050 /* encode encryption metadata into the bp */
4051 if (ot
== DMU_OT_INTENT_LOG
) {
4053 * ZIL blocks store the MAC in the embedded checksum, so the
4054 * transform must always be applied.
4056 zio_crypt_encode_mac_zil(enc_buf
, mac
);
4057 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4059 BP_SET_CRYPT(bp
, B_TRUE
);
4060 zio_crypt_encode_params_bp(bp
, salt
, iv
);
4061 zio_crypt_encode_mac_bp(bp
, mac
);
4064 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
4067 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4075 * ==========================================================================
4076 * Generate and verify checksums
4077 * ==========================================================================
4080 zio_checksum_generate(zio_t
*zio
)
4082 blkptr_t
*bp
= zio
->io_bp
;
4083 enum zio_checksum checksum
;
4087 * This is zio_write_phys().
4088 * We're either generating a label checksum, or none at all.
4090 checksum
= zio
->io_prop
.zp_checksum
;
4092 if (checksum
== ZIO_CHECKSUM_OFF
)
4095 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4097 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4098 ASSERT(!IO_IS_ALLOCATING(zio
));
4099 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4101 checksum
= BP_GET_CHECKSUM(bp
);
4105 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4111 zio_checksum_verify(zio_t
*zio
)
4113 zio_bad_cksum_t info
;
4114 blkptr_t
*bp
= zio
->io_bp
;
4117 ASSERT(zio
->io_vd
!= NULL
);
4121 * This is zio_read_phys().
4122 * We're either verifying a label checksum, or nothing at all.
4124 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4127 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
4130 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4131 zio
->io_error
= error
;
4132 if (error
== ECKSUM
&&
4133 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4134 zfs_ereport_start_checksum(zio
->io_spa
,
4135 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4136 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
4144 * Called by RAID-Z to ensure we don't compute the checksum twice.
4147 zio_checksum_verified(zio_t
*zio
)
4149 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4153 * ==========================================================================
4154 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4155 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4156 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4157 * indicate errors that are specific to one I/O, and most likely permanent.
4158 * Any other error is presumed to be worse because we weren't expecting it.
4159 * ==========================================================================
4162 zio_worst_error(int e1
, int e2
)
4164 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4167 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4168 if (e1
== zio_error_rank
[r1
])
4171 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4172 if (e2
== zio_error_rank
[r2
])
4175 return (r1
> r2
? e1
: e2
);
4179 * ==========================================================================
4181 * ==========================================================================
4184 zio_ready(zio_t
*zio
)
4186 blkptr_t
*bp
= zio
->io_bp
;
4187 zio_t
*pio
, *pio_next
;
4188 zio_link_t
*zl
= NULL
;
4190 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4195 if (zio
->io_ready
) {
4196 ASSERT(IO_IS_ALLOCATING(zio
));
4197 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4198 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4199 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4204 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4205 zio
->io_bp_copy
= *bp
;
4207 if (zio
->io_error
!= 0) {
4208 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4210 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4211 ASSERT(IO_IS_ALLOCATING(zio
));
4212 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4213 ASSERT(zio
->io_metaslab_class
!= NULL
);
4216 * We were unable to allocate anything, unreserve and
4217 * issue the next I/O to allocate.
4219 metaslab_class_throttle_unreserve(
4220 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4221 zio
->io_allocator
, zio
);
4222 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4226 mutex_enter(&zio
->io_lock
);
4227 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4228 pio
= zio_walk_parents(zio
, &zl
);
4229 mutex_exit(&zio
->io_lock
);
4232 * As we notify zio's parents, new parents could be added.
4233 * New parents go to the head of zio's io_parent_list, however,
4234 * so we will (correctly) not notify them. The remainder of zio's
4235 * io_parent_list, from 'pio_next' onward, cannot change because
4236 * all parents must wait for us to be done before they can be done.
4238 for (; pio
!= NULL
; pio
= pio_next
) {
4239 pio_next
= zio_walk_parents(zio
, &zl
);
4240 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4243 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4244 if (BP_IS_GANG(bp
)) {
4245 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4247 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4248 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4252 if (zio_injection_enabled
&&
4253 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4254 zio_handle_ignored_writes(zio
);
4260 * Update the allocation throttle accounting.
4263 zio_dva_throttle_done(zio_t
*zio
)
4265 ASSERTV(zio_t
*lio
= zio
->io_logical
);
4266 zio_t
*pio
= zio_unique_parent(zio
);
4267 vdev_t
*vd
= zio
->io_vd
;
4268 int flags
= METASLAB_ASYNC_ALLOC
;
4270 ASSERT3P(zio
->io_bp
, !=, NULL
);
4271 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4272 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4273 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4275 ASSERT3P(vd
, ==, vd
->vdev_top
);
4276 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4277 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4278 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4279 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4280 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4283 * Parents of gang children can have two flavors -- ones that
4284 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4285 * and ones that allocated the constituent blocks. The allocation
4286 * throttle needs to know the allocating parent zio so we must find
4289 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4291 * If our parent is a rewrite gang child then our grandparent
4292 * would have been the one that performed the allocation.
4294 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4295 pio
= zio_unique_parent(pio
);
4296 flags
|= METASLAB_GANG_CHILD
;
4299 ASSERT(IO_IS_ALLOCATING(pio
));
4300 ASSERT3P(zio
, !=, zio
->io_logical
);
4301 ASSERT(zio
->io_logical
!= NULL
);
4302 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4303 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4304 ASSERT(zio
->io_metaslab_class
!= NULL
);
4306 mutex_enter(&pio
->io_lock
);
4307 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4308 pio
->io_allocator
, B_TRUE
);
4309 mutex_exit(&pio
->io_lock
);
4311 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4312 pio
->io_allocator
, pio
);
4315 * Call into the pipeline to see if there is more work that
4316 * needs to be done. If there is work to be done it will be
4317 * dispatched to another taskq thread.
4319 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4323 zio_done(zio_t
*zio
)
4326 * Always attempt to keep stack usage minimal here since
4327 * we can be called recursively up to 19 levels deep.
4329 const uint64_t psize
= zio
->io_size
;
4330 zio_t
*pio
, *pio_next
;
4331 zio_link_t
*zl
= NULL
;
4334 * If our children haven't all completed,
4335 * wait for them and then repeat this pipeline stage.
4337 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4342 * If the allocation throttle is enabled, then update the accounting.
4343 * We only track child I/Os that are part of an allocating async
4344 * write. We must do this since the allocation is performed
4345 * by the logical I/O but the actual write is done by child I/Os.
4347 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4348 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4349 ASSERT(zio
->io_metaslab_class
!= NULL
);
4350 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4351 zio_dva_throttle_done(zio
);
4355 * If the allocation throttle is enabled, verify that
4356 * we have decremented the refcounts for every I/O that was throttled.
4358 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4359 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4360 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4361 ASSERT(zio
->io_bp
!= NULL
);
4363 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4365 VERIFY(zfs_refcount_not_held(
4366 &zio
->io_metaslab_class
->mc_alloc_slots
[zio
->io_allocator
],
4371 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4372 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4373 ASSERT(zio
->io_children
[c
][w
] == 0);
4375 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4376 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4377 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4378 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4379 sizeof (blkptr_t
)) == 0 ||
4380 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4381 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4382 zio
->io_bp_override
== NULL
&&
4383 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4384 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4385 BP_GET_NDVAS(zio
->io_bp
));
4386 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4387 (BP_COUNT_GANG(zio
->io_bp
) ==
4388 BP_GET_NDVAS(zio
->io_bp
)));
4390 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4391 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4395 * If there were child vdev/gang/ddt errors, they apply to us now.
4397 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4398 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4399 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4402 * If the I/O on the transformed data was successful, generate any
4403 * checksum reports now while we still have the transformed data.
4405 if (zio
->io_error
== 0) {
4406 while (zio
->io_cksum_report
!= NULL
) {
4407 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4408 uint64_t align
= zcr
->zcr_align
;
4409 uint64_t asize
= P2ROUNDUP(psize
, align
);
4410 abd_t
*adata
= zio
->io_abd
;
4412 if (asize
!= psize
) {
4413 adata
= abd_alloc(asize
, B_TRUE
);
4414 abd_copy(adata
, zio
->io_abd
, psize
);
4415 abd_zero_off(adata
, psize
, asize
- psize
);
4418 zio
->io_cksum_report
= zcr
->zcr_next
;
4419 zcr
->zcr_next
= NULL
;
4420 zcr
->zcr_finish(zcr
, adata
);
4421 zfs_ereport_free_checksum(zcr
);
4428 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4430 vdev_stat_update(zio
, psize
);
4433 * If this I/O is attached to a particular vdev is slow, exceeding
4434 * 30 seconds to complete, post an error described the I/O delay.
4435 * We ignore these errors if the device is currently unavailable.
4437 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4438 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4440 * We want to only increment our slow IO counters if
4441 * the IO is valid (i.e. not if the drive is removed).
4443 * zfs_ereport_post() will also do these checks, but
4444 * it can also ratelimit and have other failures, so we
4445 * need to increment the slow_io counters independent
4448 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4449 zio
->io_spa
, zio
->io_vd
, zio
)) {
4450 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4451 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4452 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4454 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4455 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4461 if (zio
->io_error
) {
4463 * If this I/O is attached to a particular vdev,
4464 * generate an error message describing the I/O failure
4465 * at the block level. We ignore these errors if the
4466 * device is currently unavailable.
4468 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4469 !vdev_is_dead(zio
->io_vd
))
4470 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4471 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4473 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4474 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4475 zio
== zio
->io_logical
) {
4477 * For logical I/O requests, tell the SPA to log the
4478 * error and generate a logical data ereport.
4480 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4481 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4482 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4486 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4488 * Determine whether zio should be reexecuted. This will
4489 * propagate all the way to the root via zio_notify_parent().
4491 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4492 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4494 if (IO_IS_ALLOCATING(zio
) &&
4495 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4496 if (zio
->io_error
!= ENOSPC
)
4497 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4499 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4502 if ((zio
->io_type
== ZIO_TYPE_READ
||
4503 zio
->io_type
== ZIO_TYPE_FREE
) &&
4504 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4505 zio
->io_error
== ENXIO
&&
4506 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4507 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4508 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4510 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4511 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4514 * Here is a possibly good place to attempt to do
4515 * either combinatorial reconstruction or error correction
4516 * based on checksums. It also might be a good place
4517 * to send out preliminary ereports before we suspend
4523 * If there were logical child errors, they apply to us now.
4524 * We defer this until now to avoid conflating logical child
4525 * errors with errors that happened to the zio itself when
4526 * updating vdev stats and reporting FMA events above.
4528 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4530 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4531 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4532 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4533 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4535 zio_gang_tree_free(&zio
->io_gang_tree
);
4538 * Godfather I/Os should never suspend.
4540 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4541 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4542 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4544 if (zio
->io_reexecute
) {
4546 * This is a logical I/O that wants to reexecute.
4548 * Reexecute is top-down. When an i/o fails, if it's not
4549 * the root, it simply notifies its parent and sticks around.
4550 * The parent, seeing that it still has children in zio_done(),
4551 * does the same. This percolates all the way up to the root.
4552 * The root i/o will reexecute or suspend the entire tree.
4554 * This approach ensures that zio_reexecute() honors
4555 * all the original i/o dependency relationships, e.g.
4556 * parents not executing until children are ready.
4558 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4560 zio
->io_gang_leader
= NULL
;
4562 mutex_enter(&zio
->io_lock
);
4563 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4564 mutex_exit(&zio
->io_lock
);
4567 * "The Godfather" I/O monitors its children but is
4568 * not a true parent to them. It will track them through
4569 * the pipeline but severs its ties whenever they get into
4570 * trouble (e.g. suspended). This allows "The Godfather"
4571 * I/O to return status without blocking.
4574 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4576 zio_link_t
*remove_zl
= zl
;
4577 pio_next
= zio_walk_parents(zio
, &zl
);
4579 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4580 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4581 zio_remove_child(pio
, zio
, remove_zl
);
4583 * This is a rare code path, so we don't
4584 * bother with "next_to_execute".
4586 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4591 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4593 * We're not a root i/o, so there's nothing to do
4594 * but notify our parent. Don't propagate errors
4595 * upward since we haven't permanently failed yet.
4597 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4598 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4600 * This is a rare code path, so we don't bother with
4601 * "next_to_execute".
4603 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4604 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4606 * We'd fail again if we reexecuted now, so suspend
4607 * until conditions improve (e.g. device comes online).
4609 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4612 * Reexecution is potentially a huge amount of work.
4613 * Hand it off to the otherwise-unused claim taskq.
4615 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4616 spa_taskq_dispatch_ent(zio
->io_spa
,
4617 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4618 (task_func_t
*)zio_reexecute
, zio
, 0,
4624 ASSERT(zio
->io_child_count
== 0);
4625 ASSERT(zio
->io_reexecute
== 0);
4626 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4629 * Report any checksum errors, since the I/O is complete.
4631 while (zio
->io_cksum_report
!= NULL
) {
4632 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4633 zio
->io_cksum_report
= zcr
->zcr_next
;
4634 zcr
->zcr_next
= NULL
;
4635 zcr
->zcr_finish(zcr
, NULL
);
4636 zfs_ereport_free_checksum(zcr
);
4639 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4640 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4641 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4642 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4646 * It is the responsibility of the done callback to ensure that this
4647 * particular zio is no longer discoverable for adoption, and as
4648 * such, cannot acquire any new parents.
4653 mutex_enter(&zio
->io_lock
);
4654 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4655 mutex_exit(&zio
->io_lock
);
4658 * We are done executing this zio. We may want to execute a parent
4659 * next. See the comment in zio_notify_parent().
4661 zio_t
*next_to_execute
= NULL
;
4663 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4664 zio_link_t
*remove_zl
= zl
;
4665 pio_next
= zio_walk_parents(zio
, &zl
);
4666 zio_remove_child(pio
, zio
, remove_zl
);
4667 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
4670 if (zio
->io_waiter
!= NULL
) {
4671 mutex_enter(&zio
->io_lock
);
4672 zio
->io_executor
= NULL
;
4673 cv_broadcast(&zio
->io_cv
);
4674 mutex_exit(&zio
->io_lock
);
4679 return (next_to_execute
);
4683 * ==========================================================================
4684 * I/O pipeline definition
4685 * ==========================================================================
4687 static zio_pipe_stage_t
*zio_pipeline
[] = {
4695 zio_checksum_generate
,
4711 zio_checksum_verify
,
4719 * Compare two zbookmark_phys_t's to see which we would reach first in a
4720 * pre-order traversal of the object tree.
4722 * This is simple in every case aside from the meta-dnode object. For all other
4723 * objects, we traverse them in order (object 1 before object 2, and so on).
4724 * However, all of these objects are traversed while traversing object 0, since
4725 * the data it points to is the list of objects. Thus, we need to convert to a
4726 * canonical representation so we can compare meta-dnode bookmarks to
4727 * non-meta-dnode bookmarks.
4729 * We do this by calculating "equivalents" for each field of the zbookmark.
4730 * zbookmarks outside of the meta-dnode use their own object and level, and
4731 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4732 * blocks this bookmark refers to) by multiplying their blkid by their span
4733 * (the number of L0 blocks contained within one block at their level).
4734 * zbookmarks inside the meta-dnode calculate their object equivalent
4735 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4736 * level + 1<<31 (any value larger than a level could ever be) for their level.
4737 * This causes them to always compare before a bookmark in their object
4738 * equivalent, compare appropriately to bookmarks in other objects, and to
4739 * compare appropriately to other bookmarks in the meta-dnode.
4742 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4743 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4746 * These variables represent the "equivalent" values for the zbookmark,
4747 * after converting zbookmarks inside the meta dnode to their
4748 * normal-object equivalents.
4750 uint64_t zb1obj
, zb2obj
;
4751 uint64_t zb1L0
, zb2L0
;
4752 uint64_t zb1level
, zb2level
;
4754 if (zb1
->zb_object
== zb2
->zb_object
&&
4755 zb1
->zb_level
== zb2
->zb_level
&&
4756 zb1
->zb_blkid
== zb2
->zb_blkid
)
4760 * BP_SPANB calculates the span in blocks.
4762 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4763 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4765 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4766 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4768 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4770 zb1obj
= zb1
->zb_object
;
4771 zb1level
= zb1
->zb_level
;
4774 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4775 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4777 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4779 zb2obj
= zb2
->zb_object
;
4780 zb2level
= zb2
->zb_level
;
4783 /* Now that we have a canonical representation, do the comparison. */
4784 if (zb1obj
!= zb2obj
)
4785 return (zb1obj
< zb2obj
? -1 : 1);
4786 else if (zb1L0
!= zb2L0
)
4787 return (zb1L0
< zb2L0
? -1 : 1);
4788 else if (zb1level
!= zb2level
)
4789 return (zb1level
> zb2level
? -1 : 1);
4791 * This can (theoretically) happen if the bookmarks have the same object
4792 * and level, but different blkids, if the block sizes are not the same.
4793 * There is presently no way to change the indirect block sizes
4799 * This function checks the following: given that last_block is the place that
4800 * our traversal stopped last time, does that guarantee that we've visited
4801 * every node under subtree_root? Therefore, we can't just use the raw output
4802 * of zbookmark_compare. We have to pass in a modified version of
4803 * subtree_root; by incrementing the block id, and then checking whether
4804 * last_block is before or equal to that, we can tell whether or not having
4805 * visited last_block implies that all of subtree_root's children have been
4809 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4810 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4812 zbookmark_phys_t mod_zb
= *subtree_root
;
4814 ASSERT(last_block
->zb_level
== 0);
4816 /* The objset_phys_t isn't before anything. */
4821 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4822 * data block size in sectors, because that variable is only used if
4823 * the bookmark refers to a block in the meta-dnode. Since we don't
4824 * know without examining it what object it refers to, and there's no
4825 * harm in passing in this value in other cases, we always pass it in.
4827 * We pass in 0 for the indirect block size shift because zb2 must be
4828 * level 0. The indirect block size is only used to calculate the span
4829 * of the bookmark, but since the bookmark must be level 0, the span is
4830 * always 1, so the math works out.
4832 * If you make changes to how the zbookmark_compare code works, be sure
4833 * to make sure that this code still works afterwards.
4835 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4836 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4840 #if defined(_KERNEL)
4841 EXPORT_SYMBOL(zio_type_name
);
4842 EXPORT_SYMBOL(zio_buf_alloc
);
4843 EXPORT_SYMBOL(zio_data_buf_alloc
);
4844 EXPORT_SYMBOL(zio_buf_free
);
4845 EXPORT_SYMBOL(zio_data_buf_free
);
4847 module_param(zio_slow_io_ms
, int, 0644);
4848 MODULE_PARM_DESC(zio_slow_io_ms
,
4849 "Max I/O completion time (milliseconds) before marking it as slow");
4851 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
4852 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");
4854 module_param(zfs_sync_pass_deferred_free
, int, 0644);
4855 MODULE_PARM_DESC(zfs_sync_pass_deferred_free
,
4856 "Defer frees starting in this pass");
4858 module_param(zfs_sync_pass_dont_compress
, int, 0644);
4859 MODULE_PARM_DESC(zfs_sync_pass_dont_compress
,
4860 "Don't compress starting in this pass");
4862 module_param(zfs_sync_pass_rewrite
, int, 0644);
4863 MODULE_PARM_DESC(zfs_sync_pass_rewrite
,
4864 "Rewrite new bps starting in this pass");
4866 module_param(zio_dva_throttle_enabled
, int, 0644);
4867 MODULE_PARM_DESC(zio_dva_throttle_enabled
,
4868 "Throttle block allocations in the ZIO pipeline");
4870 module_param(zio_deadman_log_all
, int, 0644);
4871 MODULE_PARM_DESC(zio_deadman_log_all
,
4872 "Log all slow ZIOs, not just those with vdevs");