4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 #include <sys/dsl_scan.h>
43 #include <sys/metaslab_impl.h>
45 #include <sys/trace_zio.h>
47 #include <sys/dsl_crypt.h>
50 * ==========================================================================
51 * I/O type descriptions
52 * ==========================================================================
54 const char *zio_type_name
[ZIO_TYPES
] = {
56 * Note: Linux kernel thread name length is limited
57 * so these names will differ from upstream open zfs.
59 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
62 int zio_dva_throttle_enabled
= B_TRUE
;
65 * ==========================================================================
67 * ==========================================================================
69 kmem_cache_t
*zio_cache
;
70 kmem_cache_t
*zio_link_cache
;
71 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
72 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
73 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
74 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
75 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 int zio_delay_max
= ZIO_DELAY_MAX
;
80 #define ZIO_PIPELINE_CONTINUE 0x100
81 #define ZIO_PIPELINE_STOP 0x101
83 #define BP_SPANB(indblkshift, level) \
84 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
85 #define COMPARE_META_LEVEL 0x80000000ul
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
99 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
100 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
103 * An allocating zio is one that either currently has the DVA allocate
104 * stage set or will have it later in its lifetime.
106 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 int zio_requeue_io_start_cut_in_line
= 1;
111 int zio_buf_debug_limit
= 16384;
113 int zio_buf_debug_limit
= 0;
116 static inline void __zio_execute(zio_t
*zio
);
118 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
124 vmem_t
*data_alloc_arena
= NULL
;
126 zio_cache
= kmem_cache_create("zio_cache",
127 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
128 zio_link_cache
= kmem_cache_create("zio_link_cache",
129 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
132 * For small buffers, we want a cache for each multiple of
133 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
134 * for each quarter-power of 2.
136 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
137 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
140 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
142 #if defined(_ILP32) && defined(_KERNEL)
144 * Cache size limited to 1M on 32-bit platforms until ARC
145 * buffers no longer require virtual address space.
147 if (size
> zfs_max_recordsize
)
156 * If we are using watchpoints, put each buffer on its own page,
157 * to eliminate the performance overhead of trapping to the
158 * kernel when modifying a non-watched buffer that shares the
159 * page with a watched buffer.
161 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
164 * Here's the problem - on 4K native devices in userland on
165 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
166 * will fail with EINVAL, causing zdb (and others) to coredump.
167 * Since userland probably doesn't need optimized buffer caches,
168 * we just force 4K alignment on everything.
170 align
= 8 * SPA_MINBLOCKSIZE
;
172 if (size
< PAGESIZE
) {
173 align
= SPA_MINBLOCKSIZE
;
174 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
181 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
182 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
183 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
185 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
186 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
187 align
, NULL
, NULL
, NULL
, NULL
,
188 data_alloc_arena
, cflags
);
193 ASSERT(zio_buf_cache
[c
] != NULL
);
194 if (zio_buf_cache
[c
- 1] == NULL
)
195 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
197 ASSERT(zio_data_buf_cache
[c
] != NULL
);
198 if (zio_data_buf_cache
[c
- 1] == NULL
)
199 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
211 kmem_cache_t
*last_cache
= NULL
;
212 kmem_cache_t
*last_data_cache
= NULL
;
214 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
217 * Cache size limited to 1M on 32-bit platforms until ARC
218 * buffers no longer require virtual address space.
220 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
223 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
224 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
225 (void) printf("zio_fini: [%d] %llu != %llu\n",
226 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
227 (long long unsigned)zio_buf_cache_allocs
[c
],
228 (long long unsigned)zio_buf_cache_frees
[c
]);
230 if (zio_buf_cache
[c
] != last_cache
) {
231 last_cache
= zio_buf_cache
[c
];
232 kmem_cache_destroy(zio_buf_cache
[c
]);
234 zio_buf_cache
[c
] = NULL
;
236 if (zio_data_buf_cache
[c
] != last_data_cache
) {
237 last_data_cache
= zio_data_buf_cache
[c
];
238 kmem_cache_destroy(zio_data_buf_cache
[c
]);
240 zio_data_buf_cache
[c
] = NULL
;
243 kmem_cache_destroy(zio_link_cache
);
244 kmem_cache_destroy(zio_cache
);
252 * ==========================================================================
253 * Allocate and free I/O buffers
254 * ==========================================================================
258 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
259 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
260 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
261 * excess / transient data in-core during a crashdump.
264 zio_buf_alloc(size_t size
)
266 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
268 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
269 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
270 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
273 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size
)
285 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
287 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
289 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
293 zio_buf_free(void *buf
, size_t size
)
295 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
297 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
302 kmem_cache_free(zio_buf_cache
[c
], buf
);
306 zio_data_buf_free(void *buf
, size_t size
)
308 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
310 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
312 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
316 zio_abd_free(void *abd
, size_t size
)
318 abd_free((abd_t
*)abd
);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
328 zio_transform_func_t
*transform
)
330 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
333 * Ensure that anyone expecting this zio to contain a linear ABD isn't
334 * going to get a nasty surprise when they try to access the data.
336 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
338 zt
->zt_orig_abd
= zio
->io_abd
;
339 zt
->zt_orig_size
= zio
->io_size
;
340 zt
->zt_bufsize
= bufsize
;
341 zt
->zt_transform
= transform
;
343 zt
->zt_next
= zio
->io_transform_stack
;
344 zio
->io_transform_stack
= zt
;
351 zio_pop_transforms(zio_t
*zio
)
355 while ((zt
= zio
->io_transform_stack
) != NULL
) {
356 if (zt
->zt_transform
!= NULL
)
357 zt
->zt_transform(zio
,
358 zt
->zt_orig_abd
, zt
->zt_orig_size
);
360 if (zt
->zt_bufsize
!= 0)
361 abd_free(zio
->io_abd
);
363 zio
->io_abd
= zt
->zt_orig_abd
;
364 zio
->io_size
= zt
->zt_orig_size
;
365 zio
->io_transform_stack
= zt
->zt_next
;
367 kmem_free(zt
, sizeof (zio_transform_t
));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks, decompression, and decryption
374 * ==========================================================================
377 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
379 ASSERT(zio
->io_size
> size
);
381 if (zio
->io_type
== ZIO_TYPE_READ
)
382 abd_copy(data
, zio
->io_abd
, size
);
386 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
388 if (zio
->io_error
== 0) {
389 void *tmp
= abd_borrow_buf(data
, size
);
390 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
391 zio
->io_abd
, tmp
, zio
->io_size
, size
);
392 abd_return_buf_copy(data
, tmp
, size
);
395 zio
->io_error
= SET_ERROR(EIO
);
400 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
404 blkptr_t
*bp
= zio
->io_bp
;
405 uint64_t lsize
= BP_GET_LSIZE(bp
);
406 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
407 uint8_t salt
[ZIO_DATA_SALT_LEN
];
408 uint8_t iv
[ZIO_DATA_IV_LEN
];
409 uint8_t mac
[ZIO_DATA_MAC_LEN
];
410 boolean_t no_crypt
= B_FALSE
;
412 ASSERT(BP_USES_CRYPT(bp
));
413 ASSERT3U(size
, !=, 0);
415 if (zio
->io_error
!= 0)
419 * Verify the cksum of MACs stored in an indirect bp. It will always
420 * be possible to verify this since it does not require an encryption
423 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
424 zio_crypt_decode_mac_bp(bp
, mac
);
426 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
428 * We haven't decompressed the data yet, but
429 * zio_crypt_do_indirect_mac_checksum() requires
430 * decompressed data to be able to parse out the MACs
431 * from the indirect block. We decompress it now and
432 * throw away the result after we are finished.
434 tmp
= zio_buf_alloc(lsize
);
435 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
436 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
438 ret
= SET_ERROR(EIO
);
441 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
442 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
443 zio_buf_free(tmp
, lsize
);
445 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
446 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
448 abd_copy(data
, zio
->io_abd
, size
);
457 * If this is an authenticated block, just check the MAC. It would be
458 * nice to separate this out into its own flag, but for the moment
459 * enum zio_flag is out of bits.
461 if (BP_IS_AUTHENTICATED(bp
)) {
462 if (ot
== DMU_OT_OBJSET
) {
463 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, zio
->io_spa
,
464 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, size
,
465 BP_SHOULD_BYTESWAP(bp
));
467 zio_crypt_decode_mac_bp(bp
, mac
);
468 ret
= spa_do_crypt_mac_abd(B_FALSE
, zio
->io_spa
,
469 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, size
, mac
);
471 abd_copy(data
, zio
->io_abd
, size
);
479 zio_crypt_decode_params_bp(bp
, salt
, iv
);
481 if (ot
== DMU_OT_INTENT_LOG
) {
482 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
483 zio_crypt_decode_mac_zil(tmp
, mac
);
484 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
486 zio_crypt_decode_mac_bp(bp
, mac
);
489 ret
= spa_do_crypt_abd(B_FALSE
, zio
->io_spa
, zio
->io_bookmark
.zb_objset
,
490 bp
, bp
->blk_birth
, size
, data
, zio
->io_abd
, iv
, mac
, salt
,
493 abd_copy(data
, zio
->io_abd
, size
);
501 /* assert that the key was found unless this was speculative */
502 ASSERT(ret
!= ENOENT
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
505 * If there was a decryption / authentication error return EIO as
506 * the io_error. If this was not a speculative zio, create an ereport.
509 ret
= SET_ERROR(EIO
);
510 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
511 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
512 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
520 * ==========================================================================
521 * I/O parent/child relationships and pipeline interlocks
522 * ==========================================================================
525 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
527 list_t
*pl
= &cio
->io_parent_list
;
529 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
533 ASSERT((*zl
)->zl_child
== cio
);
534 return ((*zl
)->zl_parent
);
538 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
540 list_t
*cl
= &pio
->io_child_list
;
542 ASSERT(MUTEX_HELD(&pio
->io_lock
));
544 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
548 ASSERT((*zl
)->zl_parent
== pio
);
549 return ((*zl
)->zl_child
);
553 zio_unique_parent(zio_t
*cio
)
555 zio_link_t
*zl
= NULL
;
556 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
558 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
563 zio_add_child(zio_t
*pio
, zio_t
*cio
)
565 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
568 * Logical I/Os can have logical, gang, or vdev children.
569 * Gang I/Os can have gang or vdev children.
570 * Vdev I/Os can only have vdev children.
571 * The following ASSERT captures all of these constraints.
573 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
578 mutex_enter(&pio
->io_lock
);
579 mutex_enter(&cio
->io_lock
);
581 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
583 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
584 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
586 list_insert_head(&pio
->io_child_list
, zl
);
587 list_insert_head(&cio
->io_parent_list
, zl
);
589 pio
->io_child_count
++;
590 cio
->io_parent_count
++;
592 mutex_exit(&cio
->io_lock
);
593 mutex_exit(&pio
->io_lock
);
597 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
599 ASSERT(zl
->zl_parent
== pio
);
600 ASSERT(zl
->zl_child
== cio
);
602 mutex_enter(&pio
->io_lock
);
603 mutex_enter(&cio
->io_lock
);
605 list_remove(&pio
->io_child_list
, zl
);
606 list_remove(&cio
->io_parent_list
, zl
);
608 pio
->io_child_count
--;
609 cio
->io_parent_count
--;
611 mutex_exit(&cio
->io_lock
);
612 mutex_exit(&pio
->io_lock
);
613 kmem_cache_free(zio_link_cache
, zl
);
617 zio_wait_for_children(zio_t
*zio
, enum zio_child child
, enum zio_wait_type wait
)
619 uint64_t *countp
= &zio
->io_children
[child
][wait
];
620 boolean_t waiting
= B_FALSE
;
622 mutex_enter(&zio
->io_lock
);
623 ASSERT(zio
->io_stall
== NULL
);
626 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
627 zio
->io_stall
= countp
;
630 mutex_exit(&zio
->io_lock
);
635 __attribute__((always_inline
))
637 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
639 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
640 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
642 mutex_enter(&pio
->io_lock
);
643 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
644 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
645 pio
->io_reexecute
|= zio
->io_reexecute
;
646 ASSERT3U(*countp
, >, 0);
650 if (*countp
== 0 && pio
->io_stall
== countp
) {
651 zio_taskq_type_t type
=
652 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
654 pio
->io_stall
= NULL
;
655 mutex_exit(&pio
->io_lock
);
657 * Dispatch the parent zio in its own taskq so that
658 * the child can continue to make progress. This also
659 * prevents overflowing the stack when we have deeply nested
660 * parent-child relationships.
662 zio_taskq_dispatch(pio
, type
, B_FALSE
);
664 mutex_exit(&pio
->io_lock
);
669 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
671 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
672 zio
->io_error
= zio
->io_child_error
[c
];
676 zio_bookmark_compare(const void *x1
, const void *x2
)
678 const zio_t
*z1
= x1
;
679 const zio_t
*z2
= x2
;
681 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
683 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
686 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
688 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
691 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
693 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
696 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
698 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
710 * ==========================================================================
711 * Create the various types of I/O (read, write, free, etc)
712 * ==========================================================================
715 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
716 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
717 void *private, zio_type_t type
, zio_priority_t priority
,
718 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
719 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
720 enum zio_stage pipeline
)
724 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
725 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
726 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
728 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
729 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
730 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
732 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
734 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
735 bzero(zio
, sizeof (zio_t
));
737 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
738 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
740 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
741 offsetof(zio_link_t
, zl_parent_node
));
742 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
743 offsetof(zio_link_t
, zl_child_node
));
744 metaslab_trace_init(&zio
->io_alloc_list
);
747 zio
->io_child_type
= ZIO_CHILD_VDEV
;
748 else if (flags
& ZIO_FLAG_GANG_CHILD
)
749 zio
->io_child_type
= ZIO_CHILD_GANG
;
750 else if (flags
& ZIO_FLAG_DDT_CHILD
)
751 zio
->io_child_type
= ZIO_CHILD_DDT
;
753 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
756 zio
->io_bp
= (blkptr_t
*)bp
;
757 zio
->io_bp_copy
= *bp
;
758 zio
->io_bp_orig
= *bp
;
759 if (type
!= ZIO_TYPE_WRITE
||
760 zio
->io_child_type
== ZIO_CHILD_DDT
)
761 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
762 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
763 zio
->io_logical
= zio
;
764 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
765 pipeline
|= ZIO_GANG_STAGES
;
771 zio
->io_private
= private;
773 zio
->io_priority
= priority
;
775 zio
->io_offset
= offset
;
776 zio
->io_orig_abd
= zio
->io_abd
= data
;
777 zio
->io_orig_size
= zio
->io_size
= psize
;
778 zio
->io_lsize
= lsize
;
779 zio
->io_orig_flags
= zio
->io_flags
= flags
;
780 zio
->io_orig_stage
= zio
->io_stage
= stage
;
781 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
782 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
784 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
785 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
788 zio
->io_bookmark
= *zb
;
791 if (zio
->io_logical
== NULL
)
792 zio
->io_logical
= pio
->io_logical
;
793 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
794 zio
->io_gang_leader
= pio
->io_gang_leader
;
795 zio_add_child(pio
, zio
);
798 taskq_init_ent(&zio
->io_tqent
);
804 zio_destroy(zio_t
*zio
)
806 metaslab_trace_fini(&zio
->io_alloc_list
);
807 list_destroy(&zio
->io_parent_list
);
808 list_destroy(&zio
->io_child_list
);
809 mutex_destroy(&zio
->io_lock
);
810 cv_destroy(&zio
->io_cv
);
811 kmem_cache_free(zio_cache
, zio
);
815 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
816 void *private, enum zio_flag flags
)
820 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
821 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
822 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
828 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
830 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
834 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
836 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
837 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
838 bp
, (longlong_t
)BP_GET_TYPE(bp
));
840 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
841 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
842 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
843 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
845 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
846 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
847 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
848 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
850 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
851 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
852 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
854 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
855 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
856 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
859 if (BP_IS_EMBEDDED(bp
)) {
860 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
861 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
862 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
867 * Pool-specific checks.
869 * Note: it would be nice to verify that the blk_birth and
870 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
871 * allows the birth time of log blocks (and dmu_sync()-ed blocks
872 * that are in the log) to be arbitrarily large.
874 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
875 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
877 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
878 zfs_panic_recover("blkptr at %p DVA %u has invalid "
880 bp
, i
, (longlong_t
)vdevid
);
883 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
885 zfs_panic_recover("blkptr at %p DVA %u has invalid "
887 bp
, i
, (longlong_t
)vdevid
);
890 if (vd
->vdev_ops
== &vdev_hole_ops
) {
891 zfs_panic_recover("blkptr at %p DVA %u has hole "
893 bp
, i
, (longlong_t
)vdevid
);
896 if (vd
->vdev_ops
== &vdev_missing_ops
) {
898 * "missing" vdevs are valid during import, but we
899 * don't have their detailed info (e.g. asize), so
900 * we can't perform any more checks on them.
904 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
905 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
907 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
908 if (offset
+ asize
> vd
->vdev_asize
) {
909 zfs_panic_recover("blkptr at %p DVA %u has invalid "
911 bp
, i
, (longlong_t
)offset
);
917 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
918 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
919 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
923 zfs_blkptr_verify(spa
, bp
);
925 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
926 data
, size
, size
, done
, private,
927 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
928 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
929 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
935 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
936 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
937 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
938 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
939 void *private, zio_priority_t priority
, enum zio_flag flags
,
940 const zbookmark_phys_t
*zb
)
944 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
945 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
946 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
947 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
948 DMU_OT_IS_VALID(zp
->zp_type
) &&
951 zp
->zp_copies
<= spa_max_replication(spa
));
953 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
954 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
955 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
956 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
958 zio
->io_ready
= ready
;
959 zio
->io_children_ready
= children_ready
;
960 zio
->io_physdone
= physdone
;
964 * Data can be NULL if we are going to call zio_write_override() to
965 * provide the already-allocated BP. But we may need the data to
966 * verify a dedup hit (if requested). In this case, don't try to
967 * dedup (just take the already-allocated BP verbatim). Encrypted
968 * dedup blocks need data as well so we also disable dedup in this
972 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
973 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
980 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
981 uint64_t size
, zio_done_func_t
*done
, void *private,
982 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
986 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
987 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
988 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
994 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
996 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
997 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
998 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
999 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1002 * We must reset the io_prop to match the values that existed
1003 * when the bp was first written by dmu_sync() keeping in mind
1004 * that nopwrite and dedup are mutually exclusive.
1006 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1007 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1008 zio
->io_prop
.zp_copies
= copies
;
1009 zio
->io_bp_override
= bp
;
1013 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1017 * The check for EMBEDDED is a performance optimization. We
1018 * process the free here (by ignoring it) rather than
1019 * putting it on the list and then processing it in zio_free_sync().
1021 if (BP_IS_EMBEDDED(bp
))
1023 metaslab_check_free(spa
, bp
);
1026 * Frees that are for the currently-syncing txg, are not going to be
1027 * deferred, and which will not need to do a read (i.e. not GANG or
1028 * DEDUP), can be processed immediately. Otherwise, put them on the
1029 * in-memory list for later processing.
1031 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
1032 txg
!= spa
->spa_syncing_txg
||
1033 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
1034 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1036 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1041 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1042 enum zio_flag flags
)
1045 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1047 ASSERT(!BP_IS_HOLE(bp
));
1048 ASSERT(spa_syncing_txg(spa
) == txg
);
1049 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
1051 if (BP_IS_EMBEDDED(bp
))
1052 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1054 metaslab_check_free(spa
, bp
);
1056 dsl_scan_freed(spa
, bp
);
1059 * GANG and DEDUP blocks can induce a read (for the gang block header,
1060 * or the DDT), so issue them asynchronously so that this thread is
1063 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1064 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1066 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1067 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1068 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1074 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1075 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1079 dprintf_bp(bp
, "claiming in txg %llu", txg
);
1081 if (BP_IS_EMBEDDED(bp
))
1082 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1085 * A claim is an allocation of a specific block. Claims are needed
1086 * to support immediate writes in the intent log. The issue is that
1087 * immediate writes contain committed data, but in a txg that was
1088 * *not* committed. Upon opening the pool after an unclean shutdown,
1089 * the intent log claims all blocks that contain immediate write data
1090 * so that the SPA knows they're in use.
1092 * All claims *must* be resolved in the first txg -- before the SPA
1093 * starts allocating blocks -- so that nothing is allocated twice.
1094 * If txg == 0 we just verify that the block is claimable.
1096 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
1097 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
1098 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1100 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1101 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1102 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1103 ASSERT0(zio
->io_queued_timestamp
);
1109 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1110 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1115 if (vd
->vdev_children
== 0) {
1116 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1117 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1118 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1122 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1124 for (c
= 0; c
< vd
->vdev_children
; c
++)
1125 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1126 done
, private, flags
));
1133 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1134 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1135 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1139 ASSERT(vd
->vdev_children
== 0);
1140 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1141 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1142 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1144 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1145 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1146 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1148 zio
->io_prop
.zp_checksum
= checksum
;
1154 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1155 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1156 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1160 ASSERT(vd
->vdev_children
== 0);
1161 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1162 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1163 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1165 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1166 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1167 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1169 zio
->io_prop
.zp_checksum
= checksum
;
1171 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1173 * zec checksums are necessarily destructive -- they modify
1174 * the end of the write buffer to hold the verifier/checksum.
1175 * Therefore, we must make a local copy in case the data is
1176 * being written to multiple places in parallel.
1178 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1179 abd_copy(wbuf
, data
, size
);
1181 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1188 * Create a child I/O to do some work for us.
1191 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1192 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1193 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1195 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1198 ASSERT(vd
->vdev_parent
==
1199 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
1201 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1203 * If we have the bp, then the child should perform the
1204 * checksum and the parent need not. This pushes error
1205 * detection as close to the leaves as possible and
1206 * eliminates redundant checksums in the interior nodes.
1208 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1209 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1212 if (vd
->vdev_children
== 0)
1213 offset
+= VDEV_LABEL_START_SIZE
;
1215 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
1218 * If we've decided to do a repair, the write is not speculative --
1219 * even if the original read was.
1221 if (flags
& ZIO_FLAG_IO_REPAIR
)
1222 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1225 * If we're creating a child I/O that is not associated with a
1226 * top-level vdev, then the child zio is not an allocating I/O.
1227 * If this is a retried I/O then we ignore it since we will
1228 * have already processed the original allocating I/O.
1230 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1231 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1232 ASSERTV(metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
));
1234 ASSERT(mc
->mc_alloc_throttle_enabled
);
1235 ASSERT(type
== ZIO_TYPE_WRITE
);
1236 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1237 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1238 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1239 pio
->io_child_type
== ZIO_CHILD_GANG
);
1241 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1245 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1246 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1247 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1248 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1250 zio
->io_physdone
= pio
->io_physdone
;
1251 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1252 zio
->io_logical
->io_phys_children
++;
1258 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1259 int type
, zio_priority_t priority
, enum zio_flag flags
,
1260 zio_done_func_t
*done
, void *private)
1264 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1266 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1267 data
, size
, size
, done
, private, type
, priority
,
1268 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1270 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1276 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1278 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1280 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1284 zio_shrink(zio_t
*zio
, uint64_t size
)
1286 ASSERT3P(zio
->io_executor
, ==, NULL
);
1287 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1288 ASSERT3U(size
, <=, zio
->io_size
);
1291 * We don't shrink for raidz because of problems with the
1292 * reconstruction when reading back less than the block size.
1293 * Note, BP_IS_RAIDZ() assumes no compression.
1295 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1296 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1297 /* we are not doing a raw write */
1298 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1299 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1304 * ==========================================================================
1305 * Prepare to read and write logical blocks
1306 * ==========================================================================
1310 zio_read_bp_init(zio_t
*zio
)
1312 blkptr_t
*bp
= zio
->io_bp
;
1314 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1316 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1317 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1318 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1319 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1320 psize
, psize
, zio_decompress
);
1323 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1324 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1325 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1326 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1327 psize
, psize
, zio_decrypt
);
1330 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1331 int psize
= BPE_GET_PSIZE(bp
);
1332 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1334 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1335 decode_embedded_bp_compressed(bp
, data
);
1336 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1338 ASSERT(!BP_IS_EMBEDDED(bp
));
1341 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1342 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1344 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1345 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1347 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1348 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1350 return (ZIO_PIPELINE_CONTINUE
);
1354 zio_write_bp_init(zio_t
*zio
)
1356 if (!IO_IS_ALLOCATING(zio
))
1357 return (ZIO_PIPELINE_CONTINUE
);
1359 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1361 if (zio
->io_bp_override
) {
1362 blkptr_t
*bp
= zio
->io_bp
;
1363 zio_prop_t
*zp
= &zio
->io_prop
;
1365 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1366 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1368 *bp
= *zio
->io_bp_override
;
1369 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1371 if (BP_IS_EMBEDDED(bp
))
1372 return (ZIO_PIPELINE_CONTINUE
);
1375 * If we've been overridden and nopwrite is set then
1376 * set the flag accordingly to indicate that a nopwrite
1377 * has already occurred.
1379 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1380 ASSERT(!zp
->zp_dedup
);
1381 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1382 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1383 return (ZIO_PIPELINE_CONTINUE
);
1386 ASSERT(!zp
->zp_nopwrite
);
1388 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1389 return (ZIO_PIPELINE_CONTINUE
);
1391 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1392 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1394 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1396 BP_SET_DEDUP(bp
, 1);
1397 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1398 return (ZIO_PIPELINE_CONTINUE
);
1402 * We were unable to handle this as an override bp, treat
1403 * it as a regular write I/O.
1405 zio
->io_bp_override
= NULL
;
1406 *bp
= zio
->io_bp_orig
;
1407 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1410 return (ZIO_PIPELINE_CONTINUE
);
1414 zio_write_compress(zio_t
*zio
)
1416 spa_t
*spa
= zio
->io_spa
;
1417 zio_prop_t
*zp
= &zio
->io_prop
;
1418 enum zio_compress compress
= zp
->zp_compress
;
1419 blkptr_t
*bp
= zio
->io_bp
;
1420 uint64_t lsize
= zio
->io_lsize
;
1421 uint64_t psize
= zio
->io_size
;
1425 * If our children haven't all reached the ready stage,
1426 * wait for them and then repeat this pipeline stage.
1428 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1429 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1430 return (ZIO_PIPELINE_STOP
);
1432 if (!IO_IS_ALLOCATING(zio
))
1433 return (ZIO_PIPELINE_CONTINUE
);
1435 if (zio
->io_children_ready
!= NULL
) {
1437 * Now that all our children are ready, run the callback
1438 * associated with this zio in case it wants to modify the
1439 * data to be written.
1441 ASSERT3U(zp
->zp_level
, >, 0);
1442 zio
->io_children_ready(zio
);
1445 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1446 ASSERT(zio
->io_bp_override
== NULL
);
1448 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1450 * We're rewriting an existing block, which means we're
1451 * working on behalf of spa_sync(). For spa_sync() to
1452 * converge, it must eventually be the case that we don't
1453 * have to allocate new blocks. But compression changes
1454 * the blocksize, which forces a reallocate, and makes
1455 * convergence take longer. Therefore, after the first
1456 * few passes, stop compressing to ensure convergence.
1458 pass
= spa_sync_pass(spa
);
1460 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1461 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1462 ASSERT(!BP_GET_DEDUP(bp
));
1464 if (pass
>= zfs_sync_pass_dont_compress
)
1465 compress
= ZIO_COMPRESS_OFF
;
1467 /* Make sure someone doesn't change their mind on overwrites */
1468 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1469 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1472 /* If it's a compressed write that is not raw, compress the buffer. */
1473 if (compress
!= ZIO_COMPRESS_OFF
&&
1474 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1475 void *cbuf
= zio_buf_alloc(lsize
);
1476 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1477 if (psize
== 0 || psize
== lsize
) {
1478 compress
= ZIO_COMPRESS_OFF
;
1479 zio_buf_free(cbuf
, lsize
);
1480 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1481 psize
<= BPE_PAYLOAD_SIZE
&&
1482 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1483 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1484 encode_embedded_bp_compressed(bp
,
1485 cbuf
, compress
, lsize
, psize
);
1486 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1487 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1488 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1489 zio_buf_free(cbuf
, lsize
);
1490 bp
->blk_birth
= zio
->io_txg
;
1491 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1492 ASSERT(spa_feature_is_active(spa
,
1493 SPA_FEATURE_EMBEDDED_DATA
));
1494 return (ZIO_PIPELINE_CONTINUE
);
1497 * Round up compressed size up to the ashift
1498 * of the smallest-ashift device, and zero the tail.
1499 * This ensures that the compressed size of the BP
1500 * (and thus compressratio property) are correct,
1501 * in that we charge for the padding used to fill out
1504 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1505 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1506 1ULL << spa
->spa_min_ashift
);
1507 if (rounded
>= lsize
) {
1508 compress
= ZIO_COMPRESS_OFF
;
1509 zio_buf_free(cbuf
, lsize
);
1512 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1513 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1514 abd_zero_off(cdata
, psize
, rounded
- psize
);
1516 zio_push_transform(zio
, cdata
,
1517 psize
, lsize
, NULL
);
1522 * We were unable to handle this as an override bp, treat
1523 * it as a regular write I/O.
1525 zio
->io_bp_override
= NULL
;
1526 *bp
= zio
->io_bp_orig
;
1527 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1530 ASSERT3U(psize
, !=, 0);
1535 * The final pass of spa_sync() must be all rewrites, but the first
1536 * few passes offer a trade-off: allocating blocks defers convergence,
1537 * but newly allocated blocks are sequential, so they can be written
1538 * to disk faster. Therefore, we allow the first few passes of
1539 * spa_sync() to allocate new blocks, but force rewrites after that.
1540 * There should only be a handful of blocks after pass 1 in any case.
1542 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1543 BP_GET_PSIZE(bp
) == psize
&&
1544 pass
>= zfs_sync_pass_rewrite
) {
1546 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1547 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1548 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1551 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1555 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1556 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1557 BP_SET_LSIZE(bp
, lsize
);
1558 BP_SET_TYPE(bp
, zp
->zp_type
);
1559 BP_SET_LEVEL(bp
, zp
->zp_level
);
1560 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1562 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1564 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1565 BP_SET_LSIZE(bp
, lsize
);
1566 BP_SET_TYPE(bp
, zp
->zp_type
);
1567 BP_SET_LEVEL(bp
, zp
->zp_level
);
1568 BP_SET_PSIZE(bp
, psize
);
1569 BP_SET_COMPRESS(bp
, compress
);
1570 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1571 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1572 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1574 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1575 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1576 ASSERT(!zp
->zp_encrypt
||
1577 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1578 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1580 if (zp
->zp_nopwrite
) {
1581 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1582 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1583 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1586 return (ZIO_PIPELINE_CONTINUE
);
1590 zio_free_bp_init(zio_t
*zio
)
1592 blkptr_t
*bp
= zio
->io_bp
;
1594 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1595 if (BP_GET_DEDUP(bp
))
1596 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1599 return (ZIO_PIPELINE_CONTINUE
);
1603 * ==========================================================================
1604 * Execute the I/O pipeline
1605 * ==========================================================================
1609 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1611 spa_t
*spa
= zio
->io_spa
;
1612 zio_type_t t
= zio
->io_type
;
1613 int flags
= (cutinline
? TQ_FRONT
: 0);
1616 * If we're a config writer or a probe, the normal issue and
1617 * interrupt threads may all be blocked waiting for the config lock.
1618 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1620 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1624 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1626 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1630 * If this is a high priority I/O, then use the high priority taskq if
1633 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1634 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1637 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1640 * NB: We are assuming that the zio can only be dispatched
1641 * to a single taskq at a time. It would be a grievous error
1642 * to dispatch the zio to another taskq at the same time.
1644 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1645 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1646 flags
, &zio
->io_tqent
);
1650 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1652 kthread_t
*executor
= zio
->io_executor
;
1653 spa_t
*spa
= zio
->io_spa
;
1655 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1656 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1658 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1659 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1668 zio_issue_async(zio_t
*zio
)
1670 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1672 return (ZIO_PIPELINE_STOP
);
1676 zio_interrupt(zio_t
*zio
)
1678 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1682 zio_delay_interrupt(zio_t
*zio
)
1685 * The timeout_generic() function isn't defined in userspace, so
1686 * rather than trying to implement the function, the zio delay
1687 * functionality has been disabled for userspace builds.
1692 * If io_target_timestamp is zero, then no delay has been registered
1693 * for this IO, thus jump to the end of this function and "skip" the
1694 * delay; issuing it directly to the zio layer.
1696 if (zio
->io_target_timestamp
!= 0) {
1697 hrtime_t now
= gethrtime();
1699 if (now
>= zio
->io_target_timestamp
) {
1701 * This IO has already taken longer than the target
1702 * delay to complete, so we don't want to delay it
1703 * any longer; we "miss" the delay and issue it
1704 * directly to the zio layer. This is likely due to
1705 * the target latency being set to a value less than
1706 * the underlying hardware can satisfy (e.g. delay
1707 * set to 1ms, but the disks take 10ms to complete an
1711 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1717 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1718 clock_t expire_at_tick
= ddi_get_lbolt() +
1721 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1722 hrtime_t
, now
, hrtime_t
, diff
);
1724 if (NSEC_TO_TICK(diff
) == 0) {
1725 /* Our delay is less than a jiffy - just spin */
1726 zfs_sleep_until(zio
->io_target_timestamp
);
1729 * Use taskq_dispatch_delay() in the place of
1730 * OpenZFS's timeout_generic().
1732 tid
= taskq_dispatch_delay(system_taskq
,
1733 (task_func_t
*)zio_interrupt
,
1734 zio
, TQ_NOSLEEP
, expire_at_tick
);
1735 if (tid
== TASKQID_INVALID
) {
1737 * Couldn't allocate a task. Just
1738 * finish the zio without a delay.
1747 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1752 * Execute the I/O pipeline until one of the following occurs:
1753 * (1) the I/O completes; (2) the pipeline stalls waiting for
1754 * dependent child I/Os; (3) the I/O issues, so we're waiting
1755 * for an I/O completion interrupt; (4) the I/O is delegated by
1756 * vdev-level caching or aggregation; (5) the I/O is deferred
1757 * due to vdev-level queueing; (6) the I/O is handed off to
1758 * another thread. In all cases, the pipeline stops whenever
1759 * there's no CPU work; it never burns a thread in cv_wait_io().
1761 * There's no locking on io_stage because there's no legitimate way
1762 * for multiple threads to be attempting to process the same I/O.
1764 static zio_pipe_stage_t
*zio_pipeline
[];
1767 * zio_execute() is a wrapper around the static function
1768 * __zio_execute() so that we can force __zio_execute() to be
1769 * inlined. This reduces stack overhead which is important
1770 * because __zio_execute() is called recursively in several zio
1771 * code paths. zio_execute() itself cannot be inlined because
1772 * it is externally visible.
1775 zio_execute(zio_t
*zio
)
1777 fstrans_cookie_t cookie
;
1779 cookie
= spl_fstrans_mark();
1781 spl_fstrans_unmark(cookie
);
1785 * Used to determine if in the current context the stack is sized large
1786 * enough to allow zio_execute() to be called recursively. A minimum
1787 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1790 zio_execute_stack_check(zio_t
*zio
)
1792 #if !defined(HAVE_LARGE_STACKS)
1793 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1795 /* Executing in txg_sync_thread() context. */
1796 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1799 /* Pool initialization outside of zio_taskq context. */
1800 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1801 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1802 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1804 #endif /* HAVE_LARGE_STACKS */
1809 __attribute__((always_inline
))
1811 __zio_execute(zio_t
*zio
)
1813 zio
->io_executor
= curthread
;
1815 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1817 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1818 enum zio_stage pipeline
= zio
->io_pipeline
;
1819 enum zio_stage stage
= zio
->io_stage
;
1822 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1823 ASSERT(ISP2(stage
));
1824 ASSERT(zio
->io_stall
== NULL
);
1828 } while ((stage
& pipeline
) == 0);
1830 ASSERT(stage
<= ZIO_STAGE_DONE
);
1833 * If we are in interrupt context and this pipeline stage
1834 * will grab a config lock that is held across I/O,
1835 * or may wait for an I/O that needs an interrupt thread
1836 * to complete, issue async to avoid deadlock.
1838 * For VDEV_IO_START, we cut in line so that the io will
1839 * be sent to disk promptly.
1841 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1842 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1843 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1844 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1845 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1850 * If the current context doesn't have large enough stacks
1851 * the zio must be issued asynchronously to prevent overflow.
1853 if (zio_execute_stack_check(zio
)) {
1854 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1855 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1856 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1860 zio
->io_stage
= stage
;
1861 zio
->io_pipeline_trace
|= zio
->io_stage
;
1862 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1864 if (rv
== ZIO_PIPELINE_STOP
)
1867 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1873 * ==========================================================================
1874 * Initiate I/O, either sync or async
1875 * ==========================================================================
1878 zio_wait(zio_t
*zio
)
1882 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1883 ASSERT3P(zio
->io_executor
, ==, NULL
);
1885 zio
->io_waiter
= curthread
;
1886 ASSERT0(zio
->io_queued_timestamp
);
1887 zio
->io_queued_timestamp
= gethrtime();
1891 mutex_enter(&zio
->io_lock
);
1892 while (zio
->io_executor
!= NULL
)
1893 cv_wait_io(&zio
->io_cv
, &zio
->io_lock
);
1894 mutex_exit(&zio
->io_lock
);
1896 error
= zio
->io_error
;
1903 zio_nowait(zio_t
*zio
)
1905 ASSERT3P(zio
->io_executor
, ==, NULL
);
1907 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1908 zio_unique_parent(zio
) == NULL
) {
1912 * This is a logical async I/O with no parent to wait for it.
1913 * We add it to the spa_async_root_zio "Godfather" I/O which
1914 * will ensure they complete prior to unloading the pool.
1916 spa_t
*spa
= zio
->io_spa
;
1918 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
1921 zio_add_child(pio
, zio
);
1924 ASSERT0(zio
->io_queued_timestamp
);
1925 zio
->io_queued_timestamp
= gethrtime();
1930 * ==========================================================================
1931 * Reexecute, cancel, or suspend/resume failed I/O
1932 * ==========================================================================
1936 zio_reexecute(zio_t
*pio
)
1938 zio_t
*cio
, *cio_next
;
1940 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1941 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1942 ASSERT(pio
->io_gang_leader
== NULL
);
1943 ASSERT(pio
->io_gang_tree
== NULL
);
1945 pio
->io_flags
= pio
->io_orig_flags
;
1946 pio
->io_stage
= pio
->io_orig_stage
;
1947 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1948 pio
->io_reexecute
= 0;
1949 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1950 pio
->io_pipeline_trace
= 0;
1952 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1953 pio
->io_state
[w
] = 0;
1954 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1955 pio
->io_child_error
[c
] = 0;
1957 if (IO_IS_ALLOCATING(pio
))
1958 BP_ZERO(pio
->io_bp
);
1961 * As we reexecute pio's children, new children could be created.
1962 * New children go to the head of pio's io_child_list, however,
1963 * so we will (correctly) not reexecute them. The key is that
1964 * the remainder of pio's io_child_list, from 'cio_next' onward,
1965 * cannot be affected by any side effects of reexecuting 'cio'.
1967 zio_link_t
*zl
= NULL
;
1968 mutex_enter(&pio
->io_lock
);
1969 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1970 cio_next
= zio_walk_children(pio
, &zl
);
1971 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1972 pio
->io_children
[cio
->io_child_type
][w
]++;
1973 mutex_exit(&pio
->io_lock
);
1975 mutex_enter(&pio
->io_lock
);
1977 mutex_exit(&pio
->io_lock
);
1980 * Now that all children have been reexecuted, execute the parent.
1981 * We don't reexecute "The Godfather" I/O here as it's the
1982 * responsibility of the caller to wait on it.
1984 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1985 pio
->io_queued_timestamp
= gethrtime();
1991 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1993 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1994 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1995 "failure and the failure mode property for this pool "
1996 "is set to panic.", spa_name(spa
));
1998 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
1999 "failure and has been suspended.\n", spa_name(spa
));
2001 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2004 mutex_enter(&spa
->spa_suspend_lock
);
2006 if (spa
->spa_suspend_zio_root
== NULL
)
2007 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2008 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2009 ZIO_FLAG_GODFATHER
);
2011 spa
->spa_suspended
= B_TRUE
;
2014 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2015 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2016 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2017 ASSERT(zio_unique_parent(zio
) == NULL
);
2018 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2019 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2022 mutex_exit(&spa
->spa_suspend_lock
);
2026 zio_resume(spa_t
*spa
)
2031 * Reexecute all previously suspended i/o.
2033 mutex_enter(&spa
->spa_suspend_lock
);
2034 spa
->spa_suspended
= B_FALSE
;
2035 cv_broadcast(&spa
->spa_suspend_cv
);
2036 pio
= spa
->spa_suspend_zio_root
;
2037 spa
->spa_suspend_zio_root
= NULL
;
2038 mutex_exit(&spa
->spa_suspend_lock
);
2044 return (zio_wait(pio
));
2048 zio_resume_wait(spa_t
*spa
)
2050 mutex_enter(&spa
->spa_suspend_lock
);
2051 while (spa_suspended(spa
))
2052 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2053 mutex_exit(&spa
->spa_suspend_lock
);
2057 * ==========================================================================
2060 * A gang block is a collection of small blocks that looks to the DMU
2061 * like one large block. When zio_dva_allocate() cannot find a block
2062 * of the requested size, due to either severe fragmentation or the pool
2063 * being nearly full, it calls zio_write_gang_block() to construct the
2064 * block from smaller fragments.
2066 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2067 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2068 * an indirect block: it's an array of block pointers. It consumes
2069 * only one sector and hence is allocatable regardless of fragmentation.
2070 * The gang header's bps point to its gang members, which hold the data.
2072 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2073 * as the verifier to ensure uniqueness of the SHA256 checksum.
2074 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2075 * not the gang header. This ensures that data block signatures (needed for
2076 * deduplication) are independent of how the block is physically stored.
2078 * Gang blocks can be nested: a gang member may itself be a gang block.
2079 * Thus every gang block is a tree in which root and all interior nodes are
2080 * gang headers, and the leaves are normal blocks that contain user data.
2081 * The root of the gang tree is called the gang leader.
2083 * To perform any operation (read, rewrite, free, claim) on a gang block,
2084 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2085 * in the io_gang_tree field of the original logical i/o by recursively
2086 * reading the gang leader and all gang headers below it. This yields
2087 * an in-core tree containing the contents of every gang header and the
2088 * bps for every constituent of the gang block.
2090 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2091 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2092 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2093 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2094 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2095 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2096 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2097 * of the gang header plus zio_checksum_compute() of the data to update the
2098 * gang header's blk_cksum as described above.
2100 * The two-phase assemble/issue model solves the problem of partial failure --
2101 * what if you'd freed part of a gang block but then couldn't read the
2102 * gang header for another part? Assembling the entire gang tree first
2103 * ensures that all the necessary gang header I/O has succeeded before
2104 * starting the actual work of free, claim, or write. Once the gang tree
2105 * is assembled, free and claim are in-memory operations that cannot fail.
2107 * In the event that a gang write fails, zio_dva_unallocate() walks the
2108 * gang tree to immediately free (i.e. insert back into the space map)
2109 * everything we've allocated. This ensures that we don't get ENOSPC
2110 * errors during repeated suspend/resume cycles due to a flaky device.
2112 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2113 * the gang tree, we won't modify the block, so we can safely defer the free
2114 * (knowing that the block is still intact). If we *can* assemble the gang
2115 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2116 * each constituent bp and we can allocate a new block on the next sync pass.
2118 * In all cases, the gang tree allows complete recovery from partial failure.
2119 * ==========================================================================
2123 zio_gang_issue_func_done(zio_t
*zio
)
2125 abd_put(zio
->io_abd
);
2129 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2135 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2136 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2137 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2138 &pio
->io_bookmark
));
2142 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2149 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2150 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2151 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2152 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2155 * As we rewrite each gang header, the pipeline will compute
2156 * a new gang block header checksum for it; but no one will
2157 * compute a new data checksum, so we do that here. The one
2158 * exception is the gang leader: the pipeline already computed
2159 * its data checksum because that stage precedes gang assembly.
2160 * (Presently, nothing actually uses interior data checksums;
2161 * this is just good hygiene.)
2163 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2164 abd_t
*buf
= abd_get_offset(data
, offset
);
2166 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2167 buf
, BP_GET_PSIZE(bp
));
2172 * If we are here to damage data for testing purposes,
2173 * leave the GBH alone so that we can detect the damage.
2175 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2176 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2178 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2179 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2180 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2181 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2189 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2192 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2193 ZIO_GANG_CHILD_FLAGS(pio
)));
2198 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2201 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2202 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2205 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2214 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2216 static zio_gang_node_t
*
2217 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2219 zio_gang_node_t
*gn
;
2221 ASSERT(*gnpp
== NULL
);
2223 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2224 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2231 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2233 zio_gang_node_t
*gn
= *gnpp
;
2235 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2236 ASSERT(gn
->gn_child
[g
] == NULL
);
2238 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2239 kmem_free(gn
, sizeof (*gn
));
2244 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2246 zio_gang_node_t
*gn
= *gnpp
;
2251 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2252 zio_gang_tree_free(&gn
->gn_child
[g
]);
2254 zio_gang_node_free(gnpp
);
2258 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2260 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2261 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2263 ASSERT(gio
->io_gang_leader
== gio
);
2264 ASSERT(BP_IS_GANG(bp
));
2266 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2267 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2268 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2272 zio_gang_tree_assemble_done(zio_t
*zio
)
2274 zio_t
*gio
= zio
->io_gang_leader
;
2275 zio_gang_node_t
*gn
= zio
->io_private
;
2276 blkptr_t
*bp
= zio
->io_bp
;
2278 ASSERT(gio
== zio_unique_parent(zio
));
2279 ASSERT(zio
->io_child_count
== 0);
2284 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2285 if (BP_SHOULD_BYTESWAP(bp
))
2286 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2288 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2289 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2290 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2292 abd_put(zio
->io_abd
);
2294 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2295 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2296 if (!BP_IS_GANG(gbp
))
2298 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2303 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2306 zio_t
*gio
= pio
->io_gang_leader
;
2309 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2310 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2311 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2314 * If you're a gang header, your data is in gn->gn_gbh.
2315 * If you're a gang member, your data is in 'data' and gn == NULL.
2317 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2320 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2322 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2323 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2324 if (BP_IS_HOLE(gbp
))
2326 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2328 offset
+= BP_GET_PSIZE(gbp
);
2332 if (gn
== gio
->io_gang_tree
)
2333 ASSERT3U(gio
->io_size
, ==, offset
);
2340 zio_gang_assemble(zio_t
*zio
)
2342 blkptr_t
*bp
= zio
->io_bp
;
2344 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2345 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2347 zio
->io_gang_leader
= zio
;
2349 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2351 return (ZIO_PIPELINE_CONTINUE
);
2355 zio_gang_issue(zio_t
*zio
)
2357 blkptr_t
*bp
= zio
->io_bp
;
2359 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
2360 return (ZIO_PIPELINE_STOP
);
2362 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2363 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2365 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2366 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2369 zio_gang_tree_free(&zio
->io_gang_tree
);
2371 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2373 return (ZIO_PIPELINE_CONTINUE
);
2377 zio_write_gang_member_ready(zio_t
*zio
)
2379 zio_t
*pio
= zio_unique_parent(zio
);
2380 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2381 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2383 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
);
2385 if (BP_IS_HOLE(zio
->io_bp
))
2388 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2390 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2391 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2392 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2393 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2394 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2396 mutex_enter(&pio
->io_lock
);
2397 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2398 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2399 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2400 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2401 DVA_SET_ASIZE(&pdva
[d
], asize
);
2403 mutex_exit(&pio
->io_lock
);
2407 zio_write_gang_done(zio_t
*zio
)
2409 abd_put(zio
->io_abd
);
2413 zio_write_gang_block(zio_t
*pio
)
2415 spa_t
*spa
= pio
->io_spa
;
2416 metaslab_class_t
*mc
= spa_normal_class(spa
);
2417 blkptr_t
*bp
= pio
->io_bp
;
2418 zio_t
*gio
= pio
->io_gang_leader
;
2420 zio_gang_node_t
*gn
, **gnpp
;
2421 zio_gbh_phys_t
*gbh
;
2423 uint64_t txg
= pio
->io_txg
;
2424 uint64_t resid
= pio
->io_size
;
2426 int copies
= gio
->io_prop
.zp_copies
;
2432 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2433 * have a third copy.
2435 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2436 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2437 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2439 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2440 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2441 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2442 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2444 flags
|= METASLAB_ASYNC_ALLOC
;
2445 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2448 * The logical zio has already placed a reservation for
2449 * 'copies' allocation slots but gang blocks may require
2450 * additional copies. These additional copies
2451 * (i.e. gbh_copies - copies) are guaranteed to succeed
2452 * since metaslab_class_throttle_reserve() always allows
2453 * additional reservations for gang blocks.
2455 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2459 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2460 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2461 &pio
->io_alloc_list
, pio
);
2463 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2464 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2465 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2468 * If we failed to allocate the gang block header then
2469 * we remove any additional allocation reservations that
2470 * we placed here. The original reservation will
2471 * be removed when the logical I/O goes to the ready
2474 metaslab_class_throttle_unreserve(mc
,
2475 gbh_copies
- copies
, pio
);
2478 pio
->io_error
= error
;
2479 return (ZIO_PIPELINE_CONTINUE
);
2483 gnpp
= &gio
->io_gang_tree
;
2485 gnpp
= pio
->io_private
;
2486 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2489 gn
= zio_gang_node_alloc(gnpp
);
2491 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2492 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2495 * Create the gang header.
2497 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2498 zio_write_gang_done
, NULL
, pio
->io_priority
,
2499 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2502 * Create and nowait the gang children.
2504 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2505 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2507 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2509 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2510 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2511 zp
.zp_type
= DMU_OT_NONE
;
2513 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2514 zp
.zp_dedup
= B_FALSE
;
2515 zp
.zp_dedup_verify
= B_FALSE
;
2516 zp
.zp_nopwrite
= B_FALSE
;
2517 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2518 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2519 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2520 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2521 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2523 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2524 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2525 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2526 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2527 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2529 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2530 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2531 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2534 * Gang children won't throttle but we should
2535 * account for their work, so reserve an allocation
2536 * slot for them here.
2538 VERIFY(metaslab_class_throttle_reserve(mc
,
2539 zp
.zp_copies
, cio
, flags
));
2545 * Set pio's pipeline to just wait for zio to finish.
2547 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2550 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2552 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2556 return (ZIO_PIPELINE_CONTINUE
);
2560 * The zio_nop_write stage in the pipeline determines if allocating a
2561 * new bp is necessary. The nopwrite feature can handle writes in
2562 * either syncing or open context (i.e. zil writes) and as a result is
2563 * mutually exclusive with dedup.
2565 * By leveraging a cryptographically secure checksum, such as SHA256, we
2566 * can compare the checksums of the new data and the old to determine if
2567 * allocating a new block is required. Note that our requirements for
2568 * cryptographic strength are fairly weak: there can't be any accidental
2569 * hash collisions, but we don't need to be secure against intentional
2570 * (malicious) collisions. To trigger a nopwrite, you have to be able
2571 * to write the file to begin with, and triggering an incorrect (hash
2572 * collision) nopwrite is no worse than simply writing to the file.
2573 * That said, there are no known attacks against the checksum algorithms
2574 * used for nopwrite, assuming that the salt and the checksums
2575 * themselves remain secret.
2578 zio_nop_write(zio_t
*zio
)
2580 blkptr_t
*bp
= zio
->io_bp
;
2581 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2582 zio_prop_t
*zp
= &zio
->io_prop
;
2584 ASSERT(BP_GET_LEVEL(bp
) == 0);
2585 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2586 ASSERT(zp
->zp_nopwrite
);
2587 ASSERT(!zp
->zp_dedup
);
2588 ASSERT(zio
->io_bp_override
== NULL
);
2589 ASSERT(IO_IS_ALLOCATING(zio
));
2592 * Check to see if the original bp and the new bp have matching
2593 * characteristics (i.e. same checksum, compression algorithms, etc).
2594 * If they don't then just continue with the pipeline which will
2595 * allocate a new bp.
2597 if (BP_IS_HOLE(bp_orig
) ||
2598 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2599 ZCHECKSUM_FLAG_NOPWRITE
) ||
2600 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2601 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2602 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2603 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2604 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2605 return (ZIO_PIPELINE_CONTINUE
);
2608 * If the checksums match then reset the pipeline so that we
2609 * avoid allocating a new bp and issuing any I/O.
2611 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2612 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2613 ZCHECKSUM_FLAG_NOPWRITE
);
2614 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2615 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2616 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2617 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2618 sizeof (uint64_t)) == 0);
2621 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2622 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2625 return (ZIO_PIPELINE_CONTINUE
);
2629 * ==========================================================================
2631 * ==========================================================================
2634 zio_ddt_child_read_done(zio_t
*zio
)
2636 blkptr_t
*bp
= zio
->io_bp
;
2637 ddt_entry_t
*dde
= zio
->io_private
;
2639 zio_t
*pio
= zio_unique_parent(zio
);
2641 mutex_enter(&pio
->io_lock
);
2642 ddp
= ddt_phys_select(dde
, bp
);
2643 if (zio
->io_error
== 0)
2644 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2646 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2647 dde
->dde_repair_abd
= zio
->io_abd
;
2649 abd_free(zio
->io_abd
);
2650 mutex_exit(&pio
->io_lock
);
2654 zio_ddt_read_start(zio_t
*zio
)
2656 blkptr_t
*bp
= zio
->io_bp
;
2658 ASSERT(BP_GET_DEDUP(bp
));
2659 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2660 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2662 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2663 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2664 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2665 ddt_phys_t
*ddp
= dde
->dde_phys
;
2666 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2669 ASSERT(zio
->io_vsd
== NULL
);
2672 if (ddp_self
== NULL
)
2673 return (ZIO_PIPELINE_CONTINUE
);
2675 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2676 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2678 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2680 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2681 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2682 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2683 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2684 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2686 return (ZIO_PIPELINE_CONTINUE
);
2689 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2690 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2691 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2693 return (ZIO_PIPELINE_CONTINUE
);
2697 zio_ddt_read_done(zio_t
*zio
)
2699 blkptr_t
*bp
= zio
->io_bp
;
2701 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
2702 return (ZIO_PIPELINE_STOP
);
2704 ASSERT(BP_GET_DEDUP(bp
));
2705 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2706 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2708 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2709 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2710 ddt_entry_t
*dde
= zio
->io_vsd
;
2712 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2713 return (ZIO_PIPELINE_CONTINUE
);
2716 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2717 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2718 return (ZIO_PIPELINE_STOP
);
2720 if (dde
->dde_repair_abd
!= NULL
) {
2721 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2723 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2725 ddt_repair_done(ddt
, dde
);
2729 ASSERT(zio
->io_vsd
== NULL
);
2731 return (ZIO_PIPELINE_CONTINUE
);
2735 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2737 spa_t
*spa
= zio
->io_spa
;
2738 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2740 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2743 * Note: we compare the original data, not the transformed data,
2744 * because when zio->io_bp is an override bp, we will not have
2745 * pushed the I/O transforms. That's an important optimization
2746 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2747 * However, we should never get a raw, override zio so in these
2748 * cases we can compare the io_abd directly. This is useful because
2749 * it allows us to do dedup verification even if we don't have access
2750 * to the original data (for instance, if the encryption keys aren't
2754 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2755 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2757 if (lio
!= NULL
&& do_raw
) {
2758 return (lio
->io_size
!= zio
->io_size
||
2759 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2760 } else if (lio
!= NULL
) {
2761 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2762 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2766 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2767 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2769 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2770 blkptr_t blk
= *zio
->io_bp
;
2775 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2776 psize
= BP_GET_PSIZE(&blk
);
2778 if (psize
!= zio
->io_size
)
2783 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2785 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2786 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2787 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2788 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
2791 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
2792 error
= SET_ERROR(ENOENT
);
2797 return (error
!= 0);
2798 } else if (ddp
->ddp_phys_birth
!= 0) {
2799 arc_buf_t
*abuf
= NULL
;
2800 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2801 blkptr_t blk
= *zio
->io_bp
;
2804 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2806 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
2811 error
= arc_read(NULL
, spa
, &blk
,
2812 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2813 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2814 &aflags
, &zio
->io_bookmark
);
2817 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2818 zio
->io_orig_size
) != 0)
2819 error
= SET_ERROR(ENOENT
);
2820 arc_buf_destroy(abuf
, &abuf
);
2824 return (error
!= 0);
2832 zio_ddt_child_write_ready(zio_t
*zio
)
2834 int p
= zio
->io_prop
.zp_copies
;
2835 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2836 ddt_entry_t
*dde
= zio
->io_private
;
2837 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2845 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2847 ddt_phys_fill(ddp
, zio
->io_bp
);
2849 zio_link_t
*zl
= NULL
;
2850 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2851 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2857 zio_ddt_child_write_done(zio_t
*zio
)
2859 int p
= zio
->io_prop
.zp_copies
;
2860 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2861 ddt_entry_t
*dde
= zio
->io_private
;
2862 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2866 ASSERT(ddp
->ddp_refcnt
== 0);
2867 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2868 dde
->dde_lead_zio
[p
] = NULL
;
2870 if (zio
->io_error
== 0) {
2871 zio_link_t
*zl
= NULL
;
2872 while (zio_walk_parents(zio
, &zl
) != NULL
)
2873 ddt_phys_addref(ddp
);
2875 ddt_phys_clear(ddp
);
2882 zio_ddt_ditto_write_done(zio_t
*zio
)
2884 int p
= DDT_PHYS_DITTO
;
2885 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
2886 blkptr_t
*bp
= zio
->io_bp
;
2887 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2888 ddt_entry_t
*dde
= zio
->io_private
;
2889 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2890 ddt_key_t
*ddk
= &dde
->dde_key
;
2894 ASSERT(ddp
->ddp_refcnt
== 0);
2895 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2896 dde
->dde_lead_zio
[p
] = NULL
;
2898 if (zio
->io_error
== 0) {
2899 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2900 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2901 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2902 if (ddp
->ddp_phys_birth
!= 0)
2903 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2904 ddt_phys_fill(ddp
, bp
);
2911 zio_ddt_write(zio_t
*zio
)
2913 spa_t
*spa
= zio
->io_spa
;
2914 blkptr_t
*bp
= zio
->io_bp
;
2915 uint64_t txg
= zio
->io_txg
;
2916 zio_prop_t
*zp
= &zio
->io_prop
;
2917 int p
= zp
->zp_copies
;
2921 ddt_t
*ddt
= ddt_select(spa
, bp
);
2925 ASSERT(BP_GET_DEDUP(bp
));
2926 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2927 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2928 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2931 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2932 ddp
= &dde
->dde_phys
[p
];
2934 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2936 * If we're using a weak checksum, upgrade to a strong checksum
2937 * and try again. If we're already using a strong checksum,
2938 * we can't resolve it, so just convert to an ordinary write.
2939 * (And automatically e-mail a paper to Nature?)
2941 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2942 ZCHECKSUM_FLAG_DEDUP
)) {
2943 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2944 zio_pop_transforms(zio
);
2945 zio
->io_stage
= ZIO_STAGE_OPEN
;
2948 zp
->zp_dedup
= B_FALSE
;
2950 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2952 return (ZIO_PIPELINE_CONTINUE
);
2955 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2956 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2958 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2959 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2960 zio_prop_t czp
= *zp
;
2962 czp
.zp_copies
= ditto_copies
;
2965 * If we arrived here with an override bp, we won't have run
2966 * the transform stack, so we won't have the data we need to
2967 * generate a child i/o. So, toss the override bp and restart.
2968 * This is safe, because using the override bp is just an
2969 * optimization; and it's rare, so the cost doesn't matter.
2971 if (zio
->io_bp_override
) {
2972 zio_pop_transforms(zio
);
2973 zio
->io_stage
= ZIO_STAGE_OPEN
;
2974 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2975 zio
->io_bp_override
= NULL
;
2978 return (ZIO_PIPELINE_CONTINUE
);
2981 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2982 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2983 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2984 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2986 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2987 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2990 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2991 if (ddp
->ddp_phys_birth
!= 0)
2992 ddt_bp_fill(ddp
, bp
, txg
);
2993 if (dde
->dde_lead_zio
[p
] != NULL
)
2994 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2996 ddt_phys_addref(ddp
);
2997 } else if (zio
->io_bp_override
) {
2998 ASSERT(bp
->blk_birth
== txg
);
2999 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3000 ddt_phys_fill(ddp
, bp
);
3001 ddt_phys_addref(ddp
);
3003 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3004 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3005 zio_ddt_child_write_ready
, NULL
, NULL
,
3006 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3007 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3009 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3010 dde
->dde_lead_zio
[p
] = cio
;
3020 return (ZIO_PIPELINE_CONTINUE
);
3023 ddt_entry_t
*freedde
; /* for debugging */
3026 zio_ddt_free(zio_t
*zio
)
3028 spa_t
*spa
= zio
->io_spa
;
3029 blkptr_t
*bp
= zio
->io_bp
;
3030 ddt_t
*ddt
= ddt_select(spa
, bp
);
3034 ASSERT(BP_GET_DEDUP(bp
));
3035 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3038 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3040 ddp
= ddt_phys_select(dde
, bp
);
3042 ddt_phys_decref(ddp
);
3046 return (ZIO_PIPELINE_CONTINUE
);
3050 * ==========================================================================
3051 * Allocate and free blocks
3052 * ==========================================================================
3056 zio_io_to_allocate(spa_t
*spa
)
3060 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
3062 zio
= avl_first(&spa
->spa_alloc_tree
);
3066 ASSERT(IO_IS_ALLOCATING(zio
));
3069 * Try to place a reservation for this zio. If we're unable to
3070 * reserve then we throttle.
3072 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
3073 zio
->io_prop
.zp_copies
, zio
, 0)) {
3077 avl_remove(&spa
->spa_alloc_tree
, zio
);
3078 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3084 zio_dva_throttle(zio_t
*zio
)
3086 spa_t
*spa
= zio
->io_spa
;
3089 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3090 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
3091 zio
->io_child_type
== ZIO_CHILD_GANG
||
3092 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3093 return (ZIO_PIPELINE_CONTINUE
);
3096 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3098 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3099 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3101 mutex_enter(&spa
->spa_alloc_lock
);
3103 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3104 avl_add(&spa
->spa_alloc_tree
, zio
);
3106 nio
= zio_io_to_allocate(zio
->io_spa
);
3107 mutex_exit(&spa
->spa_alloc_lock
);
3110 return (ZIO_PIPELINE_CONTINUE
);
3113 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3115 * We are passing control to a new zio so make sure that
3116 * it is processed by a different thread. We do this to
3117 * avoid stack overflows that can occur when parents are
3118 * throttled and children are making progress. We allow
3119 * it to go to the head of the taskq since it's already
3122 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3124 return (ZIO_PIPELINE_STOP
);
3128 zio_allocate_dispatch(spa_t
*spa
)
3132 mutex_enter(&spa
->spa_alloc_lock
);
3133 zio
= zio_io_to_allocate(spa
);
3134 mutex_exit(&spa
->spa_alloc_lock
);
3138 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3139 ASSERT0(zio
->io_error
);
3140 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3144 zio_dva_allocate(zio_t
*zio
)
3146 spa_t
*spa
= zio
->io_spa
;
3147 metaslab_class_t
*mc
= spa_normal_class(spa
);
3148 blkptr_t
*bp
= zio
->io_bp
;
3152 if (zio
->io_gang_leader
== NULL
) {
3153 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3154 zio
->io_gang_leader
= zio
;
3157 ASSERT(BP_IS_HOLE(bp
));
3158 ASSERT0(BP_GET_NDVAS(bp
));
3159 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3160 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3161 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3163 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3164 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3165 flags
|= METASLAB_DONT_THROTTLE
;
3166 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3167 flags
|= METASLAB_GANG_CHILD
;
3168 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3169 flags
|= METASLAB_ASYNC_ALLOC
;
3171 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3172 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3173 &zio
->io_alloc_list
, zio
);
3176 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
3177 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3179 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3180 return (zio_write_gang_block(zio
));
3181 zio
->io_error
= error
;
3184 return (ZIO_PIPELINE_CONTINUE
);
3188 zio_dva_free(zio_t
*zio
)
3190 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3192 return (ZIO_PIPELINE_CONTINUE
);
3196 zio_dva_claim(zio_t
*zio
)
3200 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3202 zio
->io_error
= error
;
3204 return (ZIO_PIPELINE_CONTINUE
);
3208 * Undo an allocation. This is used by zio_done() when an I/O fails
3209 * and we want to give back the block we just allocated.
3210 * This handles both normal blocks and gang blocks.
3213 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3215 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3216 ASSERT(zio
->io_bp_override
== NULL
);
3218 if (!BP_IS_HOLE(bp
))
3219 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3222 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3223 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3224 &gn
->gn_gbh
->zg_blkptr
[g
]);
3230 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3233 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3234 uint64_t size
, boolean_t
*slog
)
3237 zio_alloc_list_t io_alloc_list
;
3239 ASSERT(txg
> spa_syncing_txg(spa
));
3241 metaslab_trace_init(&io_alloc_list
);
3242 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3243 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
);
3247 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3248 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3249 &io_alloc_list
, NULL
);
3253 metaslab_trace_fini(&io_alloc_list
);
3256 BP_SET_LSIZE(new_bp
, size
);
3257 BP_SET_PSIZE(new_bp
, size
);
3258 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3259 BP_SET_CHECKSUM(new_bp
,
3260 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3261 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3262 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3263 BP_SET_LEVEL(new_bp
, 0);
3264 BP_SET_DEDUP(new_bp
, 0);
3265 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3268 * encrypted blocks will require an IV and salt. We generate
3269 * these now since we will not be rewriting the bp at
3272 if (os
->os_encrypted
) {
3273 uint8_t iv
[ZIO_DATA_IV_LEN
];
3274 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3276 BP_SET_CRYPT(new_bp
, B_TRUE
);
3277 VERIFY0(spa_crypt_get_salt(spa
,
3278 dmu_objset_id(os
), salt
));
3279 VERIFY0(zio_crypt_generate_iv(iv
));
3281 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3284 zfs_dbgmsg("%s: zil block allocation failure: "
3285 "size %llu, error %d", spa_name(spa
), size
, error
);
3292 * Free an intent log block.
3295 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
3297 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
3298 ASSERT(!BP_IS_GANG(bp
));
3300 zio_free(spa
, txg
, bp
);
3304 * ==========================================================================
3305 * Read and write to physical devices
3306 * ==========================================================================
3311 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3312 * stops after this stage and will resume upon I/O completion.
3313 * However, there are instances where the vdev layer may need to
3314 * continue the pipeline when an I/O was not issued. Since the I/O
3315 * that was sent to the vdev layer might be different than the one
3316 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3317 * force the underlying vdev layers to call either zio_execute() or
3318 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3321 zio_vdev_io_start(zio_t
*zio
)
3323 vdev_t
*vd
= zio
->io_vd
;
3325 spa_t
*spa
= zio
->io_spa
;
3329 ASSERT(zio
->io_error
== 0);
3330 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3333 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3334 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3337 * The mirror_ops handle multiple DVAs in a single BP.
3339 vdev_mirror_ops
.vdev_op_io_start(zio
);
3340 return (ZIO_PIPELINE_STOP
);
3343 ASSERT3P(zio
->io_logical
, !=, zio
);
3345 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3347 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3348 P2PHASE(zio
->io_size
, align
) != 0) {
3349 /* Transform logical writes to be a full physical block size. */
3350 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3351 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3352 ASSERT(vd
== vd
->vdev_top
);
3353 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3354 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3355 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3357 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3361 * If this is not a physical io, make sure that it is properly aligned
3362 * before proceeding.
3364 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3365 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3366 ASSERT0(P2PHASE(zio
->io_size
, align
));
3369 * For physical writes, we allow 512b aligned writes and assume
3370 * the device will perform a read-modify-write as necessary.
3372 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3373 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3376 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3379 * If this is a repair I/O, and there's no self-healing involved --
3380 * that is, we're just resilvering what we expect to resilver --
3381 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3382 * This prevents spurious resilvering with nested replication.
3383 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3384 * A is out of date, we'll read from C+D, then use the data to
3385 * resilver A+B -- but we don't actually want to resilver B, just A.
3386 * The top-level mirror has no way to know this, so instead we just
3387 * discard unnecessary repairs as we work our way down the vdev tree.
3388 * The same logic applies to any form of nested replication:
3389 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3391 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3392 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3393 zio
->io_txg
!= 0 && /* not a delegated i/o */
3394 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3395 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3396 zio_vdev_io_bypass(zio
);
3397 return (ZIO_PIPELINE_CONTINUE
);
3400 if (vd
->vdev_ops
->vdev_op_leaf
&&
3401 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3403 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3404 return (ZIO_PIPELINE_CONTINUE
);
3406 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3407 return (ZIO_PIPELINE_STOP
);
3409 if (!vdev_accessible(vd
, zio
)) {
3410 zio
->io_error
= SET_ERROR(ENXIO
);
3412 return (ZIO_PIPELINE_STOP
);
3414 zio
->io_delay
= gethrtime();
3417 vd
->vdev_ops
->vdev_op_io_start(zio
);
3418 return (ZIO_PIPELINE_STOP
);
3422 zio_vdev_io_done(zio_t
*zio
)
3424 vdev_t
*vd
= zio
->io_vd
;
3425 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3426 boolean_t unexpected_error
= B_FALSE
;
3428 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3429 return (ZIO_PIPELINE_STOP
);
3431 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3434 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3436 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3438 vdev_queue_io_done(zio
);
3440 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3441 vdev_cache_write(zio
);
3443 if (zio_injection_enabled
&& zio
->io_error
== 0)
3444 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3447 if (zio_injection_enabled
&& zio
->io_error
== 0)
3448 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3450 if (zio
->io_error
) {
3451 if (!vdev_accessible(vd
, zio
)) {
3452 zio
->io_error
= SET_ERROR(ENXIO
);
3454 unexpected_error
= B_TRUE
;
3459 ops
->vdev_op_io_done(zio
);
3461 if (unexpected_error
)
3462 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3464 return (ZIO_PIPELINE_CONTINUE
);
3468 * This function is used to change the priority of an existing zio that is
3469 * currently in-flight. This is used by the arc to upgrade priority in the
3470 * event that a demand read is made for a block that is currently queued
3471 * as a scrub or async read IO. Otherwise, the high priority read request
3472 * would end up having to wait for the lower priority IO.
3475 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3477 zio_t
*cio
, *cio_next
;
3478 zio_link_t
*zl
= NULL
;
3480 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3482 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3483 vdev_queue_change_io_priority(pio
, priority
);
3485 pio
->io_priority
= priority
;
3488 mutex_enter(&pio
->io_lock
);
3489 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3490 cio_next
= zio_walk_children(pio
, &zl
);
3491 zio_change_priority(cio
, priority
);
3493 mutex_exit(&pio
->io_lock
);
3497 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3498 * disk, and use that to finish the checksum ereport later.
3501 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3502 const abd_t
*good_buf
)
3504 /* no processing needed */
3505 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3510 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3512 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3514 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3516 zcr
->zcr_cbinfo
= zio
->io_size
;
3517 zcr
->zcr_cbdata
= abd
;
3518 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3519 zcr
->zcr_free
= zio_abd_free
;
3523 zio_vdev_io_assess(zio_t
*zio
)
3525 vdev_t
*vd
= zio
->io_vd
;
3527 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3528 return (ZIO_PIPELINE_STOP
);
3530 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3531 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3533 if (zio
->io_vsd
!= NULL
) {
3534 zio
->io_vsd_ops
->vsd_free(zio
);
3538 if (zio_injection_enabled
&& zio
->io_error
== 0)
3539 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3542 * If the I/O failed, determine whether we should attempt to retry it.
3544 * On retry, we cut in line in the issue queue, since we don't want
3545 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3547 if (zio
->io_error
&& vd
== NULL
&&
3548 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3549 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3550 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3552 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3553 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3554 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3555 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3556 zio_requeue_io_start_cut_in_line
);
3557 return (ZIO_PIPELINE_STOP
);
3561 * If we got an error on a leaf device, convert it to ENXIO
3562 * if the device is not accessible at all.
3564 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3565 !vdev_accessible(vd
, zio
))
3566 zio
->io_error
= SET_ERROR(ENXIO
);
3569 * If we can't write to an interior vdev (mirror or RAID-Z),
3570 * set vdev_cant_write so that we stop trying to allocate from it.
3572 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3573 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3574 vd
->vdev_cant_write
= B_TRUE
;
3578 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3579 * attempts will ever succeed. In this case we set a persistent bit so
3580 * that we don't bother with it in the future.
3582 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3583 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3584 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3585 vd
->vdev_nowritecache
= B_TRUE
;
3588 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3590 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3591 zio
->io_physdone
!= NULL
) {
3592 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3593 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3594 zio
->io_physdone(zio
->io_logical
);
3597 return (ZIO_PIPELINE_CONTINUE
);
3601 zio_vdev_io_reissue(zio_t
*zio
)
3603 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3604 ASSERT(zio
->io_error
== 0);
3606 zio
->io_stage
>>= 1;
3610 zio_vdev_io_redone(zio_t
*zio
)
3612 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3614 zio
->io_stage
>>= 1;
3618 zio_vdev_io_bypass(zio_t
*zio
)
3620 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3621 ASSERT(zio
->io_error
== 0);
3623 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3624 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3628 * ==========================================================================
3629 * Encrypt and store encryption parameters
3630 * ==========================================================================
3635 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3636 * managing the storage of encryption parameters and passing them to the
3637 * lower-level encryption functions.
3640 zio_encrypt(zio_t
*zio
)
3642 zio_prop_t
*zp
= &zio
->io_prop
;
3643 spa_t
*spa
= zio
->io_spa
;
3644 blkptr_t
*bp
= zio
->io_bp
;
3645 uint64_t psize
= BP_GET_PSIZE(bp
);
3646 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3647 void *enc_buf
= NULL
;
3649 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3650 uint8_t iv
[ZIO_DATA_IV_LEN
];
3651 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3652 boolean_t no_crypt
= B_FALSE
;
3654 /* the root zio already encrypted the data */
3655 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3656 return (ZIO_PIPELINE_CONTINUE
);
3658 /* only ZIL blocks are re-encrypted on rewrite */
3659 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3660 return (ZIO_PIPELINE_CONTINUE
);
3662 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3663 BP_SET_CRYPT(bp
, B_FALSE
);
3664 return (ZIO_PIPELINE_CONTINUE
);
3667 /* if we are doing raw encryption set the provided encryption params */
3668 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3669 BP_SET_CRYPT(bp
, B_TRUE
);
3670 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3671 if (ot
!= DMU_OT_OBJSET
)
3672 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3673 if (DMU_OT_IS_ENCRYPTED(ot
))
3674 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3675 return (ZIO_PIPELINE_CONTINUE
);
3678 /* indirect blocks only maintain a cksum of the lower level MACs */
3679 if (BP_GET_LEVEL(bp
) > 0) {
3680 BP_SET_CRYPT(bp
, B_TRUE
);
3681 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3682 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3684 zio_crypt_encode_mac_bp(bp
, mac
);
3685 return (ZIO_PIPELINE_CONTINUE
);
3689 * Objset blocks are a special case since they have 2 256-bit MACs
3690 * embedded within them.
3692 if (ot
== DMU_OT_OBJSET
) {
3693 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
3694 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3695 BP_SET_CRYPT(bp
, B_TRUE
);
3696 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
,
3697 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, psize
,
3698 BP_SHOULD_BYTESWAP(bp
)));
3699 return (ZIO_PIPELINE_CONTINUE
);
3702 /* unencrypted object types are only authenticated with a MAC */
3703 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
3704 BP_SET_CRYPT(bp
, B_TRUE
);
3705 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
,
3706 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, psize
, mac
));
3707 zio_crypt_encode_mac_bp(bp
, mac
);
3708 return (ZIO_PIPELINE_CONTINUE
);
3712 * Later passes of sync-to-convergence may decide to rewrite data
3713 * in place to avoid more disk reallocations. This presents a problem
3714 * for encryption because this consitutes rewriting the new data with
3715 * the same encryption key and IV. However, this only applies to blocks
3716 * in the MOS (particularly the spacemaps) and we do not encrypt the
3717 * MOS. We assert that the zio is allocating or an intent log write
3720 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
3721 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
3722 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
3723 ASSERT3U(psize
, !=, 0);
3725 enc_buf
= zio_buf_alloc(psize
);
3726 eabd
= abd_get_from_buf(enc_buf
, psize
);
3727 abd_take_ownership_of_buf(eabd
, B_TRUE
);
3730 * For an explanation of what encryption parameters are stored
3731 * where, see the block comment in zio_crypt.c.
3733 if (ot
== DMU_OT_INTENT_LOG
) {
3734 zio_crypt_decode_params_bp(bp
, salt
, iv
);
3736 BP_SET_CRYPT(bp
, B_TRUE
);
3739 /* Perform the encryption. This should not fail */
3740 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, zio
->io_bookmark
.zb_objset
, bp
,
3741 zio
->io_txg
, psize
, zio
->io_abd
, eabd
, iv
, mac
, salt
, &no_crypt
));
3743 /* encode encryption metadata into the bp */
3744 if (ot
== DMU_OT_INTENT_LOG
) {
3746 * ZIL blocks store the MAC in the embedded checksum, so the
3747 * transform must always be applied.
3749 zio_crypt_encode_mac_zil(enc_buf
, mac
);
3750 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3752 BP_SET_CRYPT(bp
, B_TRUE
);
3753 zio_crypt_encode_params_bp(bp
, salt
, iv
);
3754 zio_crypt_encode_mac_bp(bp
, mac
);
3757 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
3760 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3764 return (ZIO_PIPELINE_CONTINUE
);
3768 * ==========================================================================
3769 * Generate and verify checksums
3770 * ==========================================================================
3773 zio_checksum_generate(zio_t
*zio
)
3775 blkptr_t
*bp
= zio
->io_bp
;
3776 enum zio_checksum checksum
;
3780 * This is zio_write_phys().
3781 * We're either generating a label checksum, or none at all.
3783 checksum
= zio
->io_prop
.zp_checksum
;
3785 if (checksum
== ZIO_CHECKSUM_OFF
)
3786 return (ZIO_PIPELINE_CONTINUE
);
3788 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3790 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3791 ASSERT(!IO_IS_ALLOCATING(zio
));
3792 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3794 checksum
= BP_GET_CHECKSUM(bp
);
3798 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3800 return (ZIO_PIPELINE_CONTINUE
);
3804 zio_checksum_verify(zio_t
*zio
)
3806 zio_bad_cksum_t info
;
3807 blkptr_t
*bp
= zio
->io_bp
;
3810 ASSERT(zio
->io_vd
!= NULL
);
3814 * This is zio_read_phys().
3815 * We're either verifying a label checksum, or nothing at all.
3817 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3818 return (ZIO_PIPELINE_CONTINUE
);
3820 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3823 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3824 zio
->io_error
= error
;
3825 if (error
== ECKSUM
&&
3826 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3827 zfs_ereport_start_checksum(zio
->io_spa
,
3828 zio
->io_vd
, &zio
->io_bookmark
, zio
,
3829 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
3833 return (ZIO_PIPELINE_CONTINUE
);
3837 * Called by RAID-Z to ensure we don't compute the checksum twice.
3840 zio_checksum_verified(zio_t
*zio
)
3842 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3846 * ==========================================================================
3847 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3848 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3849 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3850 * indicate errors that are specific to one I/O, and most likely permanent.
3851 * Any other error is presumed to be worse because we weren't expecting it.
3852 * ==========================================================================
3855 zio_worst_error(int e1
, int e2
)
3857 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3860 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3861 if (e1
== zio_error_rank
[r1
])
3864 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3865 if (e2
== zio_error_rank
[r2
])
3868 return (r1
> r2
? e1
: e2
);
3872 * ==========================================================================
3874 * ==========================================================================
3877 zio_ready(zio_t
*zio
)
3879 blkptr_t
*bp
= zio
->io_bp
;
3880 zio_t
*pio
, *pio_next
;
3881 zio_link_t
*zl
= NULL
;
3883 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
3884 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
3885 return (ZIO_PIPELINE_STOP
);
3887 if (zio
->io_ready
) {
3888 ASSERT(IO_IS_ALLOCATING(zio
));
3889 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3890 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3891 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3896 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3897 zio
->io_bp_copy
= *bp
;
3899 if (zio
->io_error
!= 0) {
3900 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3902 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3903 ASSERT(IO_IS_ALLOCATING(zio
));
3904 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3906 * We were unable to allocate anything, unreserve and
3907 * issue the next I/O to allocate.
3909 metaslab_class_throttle_unreserve(
3910 spa_normal_class(zio
->io_spa
),
3911 zio
->io_prop
.zp_copies
, zio
);
3912 zio_allocate_dispatch(zio
->io_spa
);
3916 mutex_enter(&zio
->io_lock
);
3917 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3918 pio
= zio_walk_parents(zio
, &zl
);
3919 mutex_exit(&zio
->io_lock
);
3922 * As we notify zio's parents, new parents could be added.
3923 * New parents go to the head of zio's io_parent_list, however,
3924 * so we will (correctly) not notify them. The remainder of zio's
3925 * io_parent_list, from 'pio_next' onward, cannot change because
3926 * all parents must wait for us to be done before they can be done.
3928 for (; pio
!= NULL
; pio
= pio_next
) {
3929 pio_next
= zio_walk_parents(zio
, &zl
);
3930 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3933 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3934 if (BP_IS_GANG(bp
)) {
3935 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3937 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3938 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3942 if (zio_injection_enabled
&&
3943 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3944 zio_handle_ignored_writes(zio
);
3946 return (ZIO_PIPELINE_CONTINUE
);
3950 * Update the allocation throttle accounting.
3953 zio_dva_throttle_done(zio_t
*zio
)
3955 ASSERTV(zio_t
*lio
= zio
->io_logical
);
3956 zio_t
*pio
= zio_unique_parent(zio
);
3957 vdev_t
*vd
= zio
->io_vd
;
3958 int flags
= METASLAB_ASYNC_ALLOC
;
3960 ASSERT3P(zio
->io_bp
, !=, NULL
);
3961 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3962 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3963 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3965 ASSERT3P(vd
, ==, vd
->vdev_top
);
3966 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
3967 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3968 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3969 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3970 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3973 * Parents of gang children can have two flavors -- ones that
3974 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3975 * and ones that allocated the constituent blocks. The allocation
3976 * throttle needs to know the allocating parent zio so we must find
3979 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3981 * If our parent is a rewrite gang child then our grandparent
3982 * would have been the one that performed the allocation.
3984 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3985 pio
= zio_unique_parent(pio
);
3986 flags
|= METASLAB_GANG_CHILD
;
3989 ASSERT(IO_IS_ALLOCATING(pio
));
3990 ASSERT3P(zio
, !=, zio
->io_logical
);
3991 ASSERT(zio
->io_logical
!= NULL
);
3992 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3993 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3995 mutex_enter(&pio
->io_lock
);
3996 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
3997 mutex_exit(&pio
->io_lock
);
3999 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
4003 * Call into the pipeline to see if there is more work that
4004 * needs to be done. If there is work to be done it will be
4005 * dispatched to another taskq thread.
4007 zio_allocate_dispatch(zio
->io_spa
);
4011 zio_done(zio_t
*zio
)
4014 * Always attempt to keep stack usage minimal here since
4015 * we can be called recurisvely up to 19 levels deep.
4017 const uint64_t psize
= zio
->io_size
;
4018 zio_t
*pio
, *pio_next
;
4019 zio_link_t
*zl
= NULL
;
4022 * If our children haven't all completed,
4023 * wait for them and then repeat this pipeline stage.
4025 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
4026 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
4027 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
4028 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
4029 return (ZIO_PIPELINE_STOP
);
4032 * If the allocation throttle is enabled, then update the accounting.
4033 * We only track child I/Os that are part of an allocating async
4034 * write. We must do this since the allocation is performed
4035 * by the logical I/O but the actual write is done by child I/Os.
4037 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4038 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4039 ASSERT(spa_normal_class(
4040 zio
->io_spa
)->mc_alloc_throttle_enabled
);
4041 zio_dva_throttle_done(zio
);
4045 * If the allocation throttle is enabled, verify that
4046 * we have decremented the refcounts for every I/O that was throttled.
4048 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4049 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4050 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4051 ASSERT(zio
->io_bp
!= NULL
);
4052 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
);
4053 VERIFY(refcount_not_held(
4054 &(spa_normal_class(zio
->io_spa
)->mc_alloc_slots
), zio
));
4058 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4059 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4060 ASSERT(zio
->io_children
[c
][w
] == 0);
4062 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4063 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4064 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4065 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4066 sizeof (blkptr_t
)) == 0 ||
4067 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4068 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4069 zio
->io_bp_override
== NULL
&&
4070 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4071 ASSERT(!BP_SHOULD_BYTESWAP(zio
->io_bp
));
4072 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4073 BP_GET_NDVAS(zio
->io_bp
));
4074 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4075 (BP_COUNT_GANG(zio
->io_bp
) ==
4076 BP_GET_NDVAS(zio
->io_bp
)));
4078 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4079 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4083 * If there were child vdev/gang/ddt errors, they apply to us now.
4085 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4086 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4087 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4090 * If the I/O on the transformed data was successful, generate any
4091 * checksum reports now while we still have the transformed data.
4093 if (zio
->io_error
== 0) {
4094 while (zio
->io_cksum_report
!= NULL
) {
4095 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4096 uint64_t align
= zcr
->zcr_align
;
4097 uint64_t asize
= P2ROUNDUP(psize
, align
);
4098 abd_t
*adata
= zio
->io_abd
;
4100 if (asize
!= psize
) {
4101 adata
= abd_alloc(asize
, B_TRUE
);
4102 abd_copy(adata
, zio
->io_abd
, psize
);
4103 abd_zero_off(adata
, psize
, asize
- psize
);
4106 zio
->io_cksum_report
= zcr
->zcr_next
;
4107 zcr
->zcr_next
= NULL
;
4108 zcr
->zcr_finish(zcr
, adata
);
4109 zfs_ereport_free_checksum(zcr
);
4116 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4118 vdev_stat_update(zio
, psize
);
4121 * If this I/O is attached to a particular vdev is slow, exceeding
4122 * 30 seconds to complete, post an error described the I/O delay.
4123 * We ignore these errors if the device is currently unavailable.
4125 if (zio
->io_delay
>= MSEC2NSEC(zio_delay_max
)) {
4126 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
4127 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
4128 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4131 if (zio
->io_error
) {
4133 * If this I/O is attached to a particular vdev,
4134 * generate an error message describing the I/O failure
4135 * at the block level. We ignore these errors if the
4136 * device is currently unavailable.
4138 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4139 !vdev_is_dead(zio
->io_vd
))
4140 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4141 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4143 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4144 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4145 zio
== zio
->io_logical
) {
4147 * For logical I/O requests, tell the SPA to log the
4148 * error and generate a logical data ereport.
4150 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4151 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4152 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4156 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4158 * Determine whether zio should be reexecuted. This will
4159 * propagate all the way to the root via zio_notify_parent().
4161 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4162 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4164 if (IO_IS_ALLOCATING(zio
) &&
4165 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4166 if (zio
->io_error
!= ENOSPC
)
4167 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4169 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4172 if ((zio
->io_type
== ZIO_TYPE_READ
||
4173 zio
->io_type
== ZIO_TYPE_FREE
) &&
4174 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4175 zio
->io_error
== ENXIO
&&
4176 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4177 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4178 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4180 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4181 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4184 * Here is a possibly good place to attempt to do
4185 * either combinatorial reconstruction or error correction
4186 * based on checksums. It also might be a good place
4187 * to send out preliminary ereports before we suspend
4193 * If there were logical child errors, they apply to us now.
4194 * We defer this until now to avoid conflating logical child
4195 * errors with errors that happened to the zio itself when
4196 * updating vdev stats and reporting FMA events above.
4198 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4200 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4201 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4202 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4203 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4205 zio_gang_tree_free(&zio
->io_gang_tree
);
4208 * Godfather I/Os should never suspend.
4210 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4211 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4212 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4214 if (zio
->io_reexecute
) {
4216 * This is a logical I/O that wants to reexecute.
4218 * Reexecute is top-down. When an i/o fails, if it's not
4219 * the root, it simply notifies its parent and sticks around.
4220 * The parent, seeing that it still has children in zio_done(),
4221 * does the same. This percolates all the way up to the root.
4222 * The root i/o will reexecute or suspend the entire tree.
4224 * This approach ensures that zio_reexecute() honors
4225 * all the original i/o dependency relationships, e.g.
4226 * parents not executing until children are ready.
4228 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4230 zio
->io_gang_leader
= NULL
;
4232 mutex_enter(&zio
->io_lock
);
4233 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4234 mutex_exit(&zio
->io_lock
);
4237 * "The Godfather" I/O monitors its children but is
4238 * not a true parent to them. It will track them through
4239 * the pipeline but severs its ties whenever they get into
4240 * trouble (e.g. suspended). This allows "The Godfather"
4241 * I/O to return status without blocking.
4244 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4246 zio_link_t
*remove_zl
= zl
;
4247 pio_next
= zio_walk_parents(zio
, &zl
);
4249 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4250 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4251 zio_remove_child(pio
, zio
, remove_zl
);
4252 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4256 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4258 * We're not a root i/o, so there's nothing to do
4259 * but notify our parent. Don't propagate errors
4260 * upward since we haven't permanently failed yet.
4262 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4263 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4264 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4265 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4267 * We'd fail again if we reexecuted now, so suspend
4268 * until conditions improve (e.g. device comes online).
4270 zio_suspend(zio
->io_spa
, zio
);
4273 * Reexecution is potentially a huge amount of work.
4274 * Hand it off to the otherwise-unused claim taskq.
4276 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4277 spa_taskq_dispatch_ent(zio
->io_spa
,
4278 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4279 (task_func_t
*)zio_reexecute
, zio
, 0,
4282 return (ZIO_PIPELINE_STOP
);
4285 ASSERT(zio
->io_child_count
== 0);
4286 ASSERT(zio
->io_reexecute
== 0);
4287 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4290 * Report any checksum errors, since the I/O is complete.
4292 while (zio
->io_cksum_report
!= NULL
) {
4293 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4294 zio
->io_cksum_report
= zcr
->zcr_next
;
4295 zcr
->zcr_next
= NULL
;
4296 zcr
->zcr_finish(zcr
, NULL
);
4297 zfs_ereport_free_checksum(zcr
);
4300 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4301 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4302 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4303 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4307 * It is the responsibility of the done callback to ensure that this
4308 * particular zio is no longer discoverable for adoption, and as
4309 * such, cannot acquire any new parents.
4314 mutex_enter(&zio
->io_lock
);
4315 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4316 mutex_exit(&zio
->io_lock
);
4319 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4320 zio_link_t
*remove_zl
= zl
;
4321 pio_next
= zio_walk_parents(zio
, &zl
);
4322 zio_remove_child(pio
, zio
, remove_zl
);
4323 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4326 if (zio
->io_waiter
!= NULL
) {
4327 mutex_enter(&zio
->io_lock
);
4328 zio
->io_executor
= NULL
;
4329 cv_broadcast(&zio
->io_cv
);
4330 mutex_exit(&zio
->io_lock
);
4335 return (ZIO_PIPELINE_STOP
);
4339 * ==========================================================================
4340 * I/O pipeline definition
4341 * ==========================================================================
4343 static zio_pipe_stage_t
*zio_pipeline
[] = {
4351 zio_checksum_generate
,
4367 zio_checksum_verify
,
4375 * Compare two zbookmark_phys_t's to see which we would reach first in a
4376 * pre-order traversal of the object tree.
4378 * This is simple in every case aside from the meta-dnode object. For all other
4379 * objects, we traverse them in order (object 1 before object 2, and so on).
4380 * However, all of these objects are traversed while traversing object 0, since
4381 * the data it points to is the list of objects. Thus, we need to convert to a
4382 * canonical representation so we can compare meta-dnode bookmarks to
4383 * non-meta-dnode bookmarks.
4385 * We do this by calculating "equivalents" for each field of the zbookmark.
4386 * zbookmarks outside of the meta-dnode use their own object and level, and
4387 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4388 * blocks this bookmark refers to) by multiplying their blkid by their span
4389 * (the number of L0 blocks contained within one block at their level).
4390 * zbookmarks inside the meta-dnode calculate their object equivalent
4391 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4392 * level + 1<<31 (any value larger than a level could ever be) for their level.
4393 * This causes them to always compare before a bookmark in their object
4394 * equivalent, compare appropriately to bookmarks in other objects, and to
4395 * compare appropriately to other bookmarks in the meta-dnode.
4398 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4399 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4402 * These variables represent the "equivalent" values for the zbookmark,
4403 * after converting zbookmarks inside the meta dnode to their
4404 * normal-object equivalents.
4406 uint64_t zb1obj
, zb2obj
;
4407 uint64_t zb1L0
, zb2L0
;
4408 uint64_t zb1level
, zb2level
;
4410 if (zb1
->zb_object
== zb2
->zb_object
&&
4411 zb1
->zb_level
== zb2
->zb_level
&&
4412 zb1
->zb_blkid
== zb2
->zb_blkid
)
4416 * BP_SPANB calculates the span in blocks.
4418 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4419 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4421 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4422 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4424 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4426 zb1obj
= zb1
->zb_object
;
4427 zb1level
= zb1
->zb_level
;
4430 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4431 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4433 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4435 zb2obj
= zb2
->zb_object
;
4436 zb2level
= zb2
->zb_level
;
4439 /* Now that we have a canonical representation, do the comparison. */
4440 if (zb1obj
!= zb2obj
)
4441 return (zb1obj
< zb2obj
? -1 : 1);
4442 else if (zb1L0
!= zb2L0
)
4443 return (zb1L0
< zb2L0
? -1 : 1);
4444 else if (zb1level
!= zb2level
)
4445 return (zb1level
> zb2level
? -1 : 1);
4447 * This can (theoretically) happen if the bookmarks have the same object
4448 * and level, but different blkids, if the block sizes are not the same.
4449 * There is presently no way to change the indirect block sizes
4455 * This function checks the following: given that last_block is the place that
4456 * our traversal stopped last time, does that guarantee that we've visited
4457 * every node under subtree_root? Therefore, we can't just use the raw output
4458 * of zbookmark_compare. We have to pass in a modified version of
4459 * subtree_root; by incrementing the block id, and then checking whether
4460 * last_block is before or equal to that, we can tell whether or not having
4461 * visited last_block implies that all of subtree_root's children have been
4465 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4466 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4468 zbookmark_phys_t mod_zb
= *subtree_root
;
4470 ASSERT(last_block
->zb_level
== 0);
4472 /* The objset_phys_t isn't before anything. */
4477 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4478 * data block size in sectors, because that variable is only used if
4479 * the bookmark refers to a block in the meta-dnode. Since we don't
4480 * know without examining it what object it refers to, and there's no
4481 * harm in passing in this value in other cases, we always pass it in.
4483 * We pass in 0 for the indirect block size shift because zb2 must be
4484 * level 0. The indirect block size is only used to calculate the span
4485 * of the bookmark, but since the bookmark must be level 0, the span is
4486 * always 1, so the math works out.
4488 * If you make changes to how the zbookmark_compare code works, be sure
4489 * to make sure that this code still works afterwards.
4491 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4492 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4496 #if defined(_KERNEL) && defined(HAVE_SPL)
4497 EXPORT_SYMBOL(zio_type_name
);
4498 EXPORT_SYMBOL(zio_buf_alloc
);
4499 EXPORT_SYMBOL(zio_data_buf_alloc
);
4500 EXPORT_SYMBOL(zio_buf_free
);
4501 EXPORT_SYMBOL(zio_data_buf_free
);
4503 module_param(zio_delay_max
, int, 0644);
4504 MODULE_PARM_DESC(zio_delay_max
, "Max zio millisec delay before posting event");
4506 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
4507 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");
4509 module_param(zfs_sync_pass_deferred_free
, int, 0644);
4510 MODULE_PARM_DESC(zfs_sync_pass_deferred_free
,
4511 "Defer frees starting in this pass");
4513 module_param(zfs_sync_pass_dont_compress
, int, 0644);
4514 MODULE_PARM_DESC(zfs_sync_pass_dont_compress
,
4515 "Don't compress starting in this pass");
4517 module_param(zfs_sync_pass_rewrite
, int, 0644);
4518 MODULE_PARM_DESC(zfs_sync_pass_rewrite
,
4519 "Rewrite new bps starting in this pass");
4521 module_param(zio_dva_throttle_enabled
, int, 0644);
4522 MODULE_PARM_DESC(zio_dva_throttle_enabled
,
4523 "Throttle block allocations in the ZIO pipeline");