4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 #include <sys/dsl_scan.h>
43 #include <sys/metaslab_impl.h>
45 #include <sys/trace_zio.h>
47 #include <sys/dsl_crypt.h>
50 * ==========================================================================
51 * I/O type descriptions
52 * ==========================================================================
54 const char *zio_type_name
[ZIO_TYPES
] = {
56 * Note: Linux kernel thread name length is limited
57 * so these names will differ from upstream open zfs.
59 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
62 int zio_dva_throttle_enabled
= B_TRUE
;
65 * ==========================================================================
67 * ==========================================================================
69 kmem_cache_t
*zio_cache
;
70 kmem_cache_t
*zio_link_cache
;
71 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
72 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
73 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
74 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
75 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
78 int zio_delay_max
= ZIO_DELAY_MAX
;
80 #define ZIO_PIPELINE_CONTINUE 0x100
81 #define ZIO_PIPELINE_STOP 0x101
83 #define BP_SPANB(indblkshift, level) \
84 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
85 #define COMPARE_META_LEVEL 0x80000000ul
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
99 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
100 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
103 * An allocating zio is one that either currently has the DVA allocate
104 * stage set or will have it later in its lifetime.
106 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 int zio_requeue_io_start_cut_in_line
= 1;
111 int zio_buf_debug_limit
= 16384;
113 int zio_buf_debug_limit
= 0;
116 static inline void __zio_execute(zio_t
*zio
);
118 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
124 vmem_t
*data_alloc_arena
= NULL
;
126 zio_cache
= kmem_cache_create("zio_cache",
127 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
128 zio_link_cache
= kmem_cache_create("zio_link_cache",
129 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
132 * For small buffers, we want a cache for each multiple of
133 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
134 * for each quarter-power of 2.
136 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
137 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
140 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
142 #if defined(_ILP32) && defined(_KERNEL)
144 * Cache size limited to 1M on 32-bit platforms until ARC
145 * buffers no longer require virtual address space.
147 if (size
> zfs_max_recordsize
)
156 * If we are using watchpoints, put each buffer on its own page,
157 * to eliminate the performance overhead of trapping to the
158 * kernel when modifying a non-watched buffer that shares the
159 * page with a watched buffer.
161 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
164 * Here's the problem - on 4K native devices in userland on
165 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
166 * will fail with EINVAL, causing zdb (and others) to coredump.
167 * Since userland probably doesn't need optimized buffer caches,
168 * we just force 4K alignment on everything.
170 align
= 8 * SPA_MINBLOCKSIZE
;
172 if (size
< PAGESIZE
) {
173 align
= SPA_MINBLOCKSIZE
;
174 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
181 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
182 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
183 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
185 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
186 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
187 align
, NULL
, NULL
, NULL
, NULL
,
188 data_alloc_arena
, cflags
);
193 ASSERT(zio_buf_cache
[c
] != NULL
);
194 if (zio_buf_cache
[c
- 1] == NULL
)
195 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
197 ASSERT(zio_data_buf_cache
[c
] != NULL
);
198 if (zio_data_buf_cache
[c
- 1] == NULL
)
199 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
211 kmem_cache_t
*last_cache
= NULL
;
212 kmem_cache_t
*last_data_cache
= NULL
;
214 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
217 * Cache size limited to 1M on 32-bit platforms until ARC
218 * buffers no longer require virtual address space.
220 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
223 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
224 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
225 (void) printf("zio_fini: [%d] %llu != %llu\n",
226 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
227 (long long unsigned)zio_buf_cache_allocs
[c
],
228 (long long unsigned)zio_buf_cache_frees
[c
]);
230 if (zio_buf_cache
[c
] != last_cache
) {
231 last_cache
= zio_buf_cache
[c
];
232 kmem_cache_destroy(zio_buf_cache
[c
]);
234 zio_buf_cache
[c
] = NULL
;
236 if (zio_data_buf_cache
[c
] != last_data_cache
) {
237 last_data_cache
= zio_data_buf_cache
[c
];
238 kmem_cache_destroy(zio_data_buf_cache
[c
]);
240 zio_data_buf_cache
[c
] = NULL
;
243 kmem_cache_destroy(zio_link_cache
);
244 kmem_cache_destroy(zio_cache
);
252 * ==========================================================================
253 * Allocate and free I/O buffers
254 * ==========================================================================
258 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
259 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
260 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
261 * excess / transient data in-core during a crashdump.
264 zio_buf_alloc(size_t size
)
266 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
268 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
269 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
270 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
273 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size
)
285 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
287 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
289 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
293 zio_buf_free(void *buf
, size_t size
)
295 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
297 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
302 kmem_cache_free(zio_buf_cache
[c
], buf
);
306 zio_data_buf_free(void *buf
, size_t size
)
308 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
310 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
312 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
316 zio_abd_free(void *abd
, size_t size
)
318 abd_free((abd_t
*)abd
);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
328 zio_transform_func_t
*transform
)
330 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
333 * Ensure that anyone expecting this zio to contain a linear ABD isn't
334 * going to get a nasty surprise when they try to access the data.
336 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
338 zt
->zt_orig_abd
= zio
->io_abd
;
339 zt
->zt_orig_size
= zio
->io_size
;
340 zt
->zt_bufsize
= bufsize
;
341 zt
->zt_transform
= transform
;
343 zt
->zt_next
= zio
->io_transform_stack
;
344 zio
->io_transform_stack
= zt
;
351 zio_pop_transforms(zio_t
*zio
)
355 while ((zt
= zio
->io_transform_stack
) != NULL
) {
356 if (zt
->zt_transform
!= NULL
)
357 zt
->zt_transform(zio
,
358 zt
->zt_orig_abd
, zt
->zt_orig_size
);
360 if (zt
->zt_bufsize
!= 0)
361 abd_free(zio
->io_abd
);
363 zio
->io_abd
= zt
->zt_orig_abd
;
364 zio
->io_size
= zt
->zt_orig_size
;
365 zio
->io_transform_stack
= zt
->zt_next
;
367 kmem_free(zt
, sizeof (zio_transform_t
));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks, decompression, and decryption
374 * ==========================================================================
377 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
379 ASSERT(zio
->io_size
> size
);
381 if (zio
->io_type
== ZIO_TYPE_READ
)
382 abd_copy(data
, zio
->io_abd
, size
);
386 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
388 if (zio
->io_error
== 0) {
389 void *tmp
= abd_borrow_buf(data
, size
);
390 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
391 zio
->io_abd
, tmp
, zio
->io_size
, size
);
392 abd_return_buf_copy(data
, tmp
, size
);
395 zio
->io_error
= SET_ERROR(EIO
);
400 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
404 blkptr_t
*bp
= zio
->io_bp
;
405 spa_t
*spa
= zio
->io_spa
;
406 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
407 uint64_t lsize
= BP_GET_LSIZE(bp
);
408 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
409 uint8_t salt
[ZIO_DATA_SALT_LEN
];
410 uint8_t iv
[ZIO_DATA_IV_LEN
];
411 uint8_t mac
[ZIO_DATA_MAC_LEN
];
412 boolean_t no_crypt
= B_FALSE
;
414 ASSERT(BP_USES_CRYPT(bp
));
415 ASSERT3U(size
, !=, 0);
417 if (zio
->io_error
!= 0)
421 * Verify the cksum of MACs stored in an indirect bp. It will always
422 * be possible to verify this since it does not require an encryption
425 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
426 zio_crypt_decode_mac_bp(bp
, mac
);
428 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
430 * We haven't decompressed the data yet, but
431 * zio_crypt_do_indirect_mac_checksum() requires
432 * decompressed data to be able to parse out the MACs
433 * from the indirect block. We decompress it now and
434 * throw away the result after we are finished.
436 tmp
= zio_buf_alloc(lsize
);
437 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
438 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
440 ret
= SET_ERROR(EIO
);
443 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
444 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
445 zio_buf_free(tmp
, lsize
);
447 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
448 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
450 abd_copy(data
, zio
->io_abd
, size
);
459 * If this is an authenticated block, just check the MAC. It would be
460 * nice to separate this out into its own flag, but for the moment
461 * enum zio_flag is out of bits.
463 if (BP_IS_AUTHENTICATED(bp
)) {
464 if (ot
== DMU_OT_OBJSET
) {
465 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
466 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
468 zio_crypt_decode_mac_bp(bp
, mac
);
469 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
470 zio
->io_abd
, size
, mac
);
472 abd_copy(data
, zio
->io_abd
, size
);
480 zio_crypt_decode_params_bp(bp
, salt
, iv
);
482 if (ot
== DMU_OT_INTENT_LOG
) {
483 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
484 zio_crypt_decode_mac_zil(tmp
, mac
);
485 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
487 zio_crypt_decode_mac_bp(bp
, mac
);
490 ret
= spa_do_crypt_abd(B_FALSE
, spa
, dsobj
, bp
, bp
->blk_birth
,
491 size
, data
, zio
->io_abd
, iv
, mac
, salt
, &no_crypt
);
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 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
, uint8_t childbits
, enum zio_wait_type wait
)
619 boolean_t waiting
= B_FALSE
;
621 mutex_enter(&zio
->io_lock
);
622 ASSERT(zio
->io_stall
== NULL
);
623 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
624 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
627 uint64_t *countp
= &zio
->io_children
[c
][wait
];
630 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
631 zio
->io_stall
= countp
;
636 mutex_exit(&zio
->io_lock
);
640 __attribute__((always_inline
))
642 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
644 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
645 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
647 mutex_enter(&pio
->io_lock
);
648 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
649 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
650 pio
->io_reexecute
|= zio
->io_reexecute
;
651 ASSERT3U(*countp
, >, 0);
655 if (*countp
== 0 && pio
->io_stall
== countp
) {
656 zio_taskq_type_t type
=
657 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
659 pio
->io_stall
= NULL
;
660 mutex_exit(&pio
->io_lock
);
662 * Dispatch the parent zio in its own taskq so that
663 * the child can continue to make progress. This also
664 * prevents overflowing the stack when we have deeply nested
665 * parent-child relationships.
667 zio_taskq_dispatch(pio
, type
, B_FALSE
);
669 mutex_exit(&pio
->io_lock
);
674 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
676 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
677 zio
->io_error
= zio
->io_child_error
[c
];
681 zio_bookmark_compare(const void *x1
, const void *x2
)
683 const zio_t
*z1
= x1
;
684 const zio_t
*z2
= x2
;
686 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
688 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
691 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
693 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
696 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
698 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
701 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
703 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
715 * ==========================================================================
716 * Create the various types of I/O (read, write, free, etc)
717 * ==========================================================================
720 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
721 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
722 void *private, zio_type_t type
, zio_priority_t priority
,
723 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
724 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
725 enum zio_stage pipeline
)
729 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
730 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
731 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
733 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
734 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
735 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
737 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
739 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
740 bzero(zio
, sizeof (zio_t
));
742 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
743 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
745 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
746 offsetof(zio_link_t
, zl_parent_node
));
747 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
748 offsetof(zio_link_t
, zl_child_node
));
749 metaslab_trace_init(&zio
->io_alloc_list
);
752 zio
->io_child_type
= ZIO_CHILD_VDEV
;
753 else if (flags
& ZIO_FLAG_GANG_CHILD
)
754 zio
->io_child_type
= ZIO_CHILD_GANG
;
755 else if (flags
& ZIO_FLAG_DDT_CHILD
)
756 zio
->io_child_type
= ZIO_CHILD_DDT
;
758 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
761 zio
->io_bp
= (blkptr_t
*)bp
;
762 zio
->io_bp_copy
= *bp
;
763 zio
->io_bp_orig
= *bp
;
764 if (type
!= ZIO_TYPE_WRITE
||
765 zio
->io_child_type
== ZIO_CHILD_DDT
)
766 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
767 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
768 zio
->io_logical
= zio
;
769 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
770 pipeline
|= ZIO_GANG_STAGES
;
776 zio
->io_private
= private;
778 zio
->io_priority
= priority
;
780 zio
->io_offset
= offset
;
781 zio
->io_orig_abd
= zio
->io_abd
= data
;
782 zio
->io_orig_size
= zio
->io_size
= psize
;
783 zio
->io_lsize
= lsize
;
784 zio
->io_orig_flags
= zio
->io_flags
= flags
;
785 zio
->io_orig_stage
= zio
->io_stage
= stage
;
786 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
787 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
789 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
790 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
793 zio
->io_bookmark
= *zb
;
796 if (zio
->io_logical
== NULL
)
797 zio
->io_logical
= pio
->io_logical
;
798 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
799 zio
->io_gang_leader
= pio
->io_gang_leader
;
800 zio_add_child(pio
, zio
);
803 taskq_init_ent(&zio
->io_tqent
);
809 zio_destroy(zio_t
*zio
)
811 metaslab_trace_fini(&zio
->io_alloc_list
);
812 list_destroy(&zio
->io_parent_list
);
813 list_destroy(&zio
->io_child_list
);
814 mutex_destroy(&zio
->io_lock
);
815 cv_destroy(&zio
->io_cv
);
816 kmem_cache_free(zio_cache
, zio
);
820 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
821 void *private, enum zio_flag flags
)
825 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
826 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
827 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
833 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
835 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
839 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
841 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
842 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
843 bp
, (longlong_t
)BP_GET_TYPE(bp
));
845 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
846 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
847 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
848 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
850 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
851 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
852 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
853 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
855 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
856 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
857 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
859 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
860 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
861 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
864 if (BP_IS_EMBEDDED(bp
)) {
865 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
866 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
867 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
872 * Pool-specific checks.
874 * Note: it would be nice to verify that the blk_birth and
875 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
876 * allows the birth time of log blocks (and dmu_sync()-ed blocks
877 * that are in the log) to be arbitrarily large.
879 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
880 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
882 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
883 zfs_panic_recover("blkptr at %p DVA %u has invalid "
885 bp
, i
, (longlong_t
)vdevid
);
888 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
890 zfs_panic_recover("blkptr at %p DVA %u has invalid "
892 bp
, i
, (longlong_t
)vdevid
);
895 if (vd
->vdev_ops
== &vdev_hole_ops
) {
896 zfs_panic_recover("blkptr at %p DVA %u has hole "
898 bp
, i
, (longlong_t
)vdevid
);
901 if (vd
->vdev_ops
== &vdev_missing_ops
) {
903 * "missing" vdevs are valid during import, but we
904 * don't have their detailed info (e.g. asize), so
905 * we can't perform any more checks on them.
909 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
910 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
912 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
913 if (offset
+ asize
> vd
->vdev_asize
) {
914 zfs_panic_recover("blkptr at %p DVA %u has invalid "
916 bp
, i
, (longlong_t
)offset
);
922 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
923 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
924 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
928 zfs_blkptr_verify(spa
, bp
);
930 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
931 data
, size
, size
, done
, private,
932 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
933 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
934 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
940 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
941 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
942 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
943 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
944 void *private, zio_priority_t priority
, enum zio_flag flags
,
945 const zbookmark_phys_t
*zb
)
949 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
950 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
951 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
952 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
953 DMU_OT_IS_VALID(zp
->zp_type
) &&
956 zp
->zp_copies
<= spa_max_replication(spa
));
958 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
959 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
960 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
961 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
963 zio
->io_ready
= ready
;
964 zio
->io_children_ready
= children_ready
;
965 zio
->io_physdone
= physdone
;
969 * Data can be NULL if we are going to call zio_write_override() to
970 * provide the already-allocated BP. But we may need the data to
971 * verify a dedup hit (if requested). In this case, don't try to
972 * dedup (just take the already-allocated BP verbatim). Encrypted
973 * dedup blocks need data as well so we also disable dedup in this
977 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
978 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
985 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
986 uint64_t size
, zio_done_func_t
*done
, void *private,
987 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
991 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
992 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
993 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
999 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1001 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1002 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1003 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1004 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1007 * We must reset the io_prop to match the values that existed
1008 * when the bp was first written by dmu_sync() keeping in mind
1009 * that nopwrite and dedup are mutually exclusive.
1011 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1012 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1013 zio
->io_prop
.zp_copies
= copies
;
1014 zio
->io_bp_override
= bp
;
1018 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1022 * The check for EMBEDDED is a performance optimization. We
1023 * process the free here (by ignoring it) rather than
1024 * putting it on the list and then processing it in zio_free_sync().
1026 if (BP_IS_EMBEDDED(bp
))
1028 metaslab_check_free(spa
, bp
);
1031 * Frees that are for the currently-syncing txg, are not going to be
1032 * deferred, and which will not need to do a read (i.e. not GANG or
1033 * DEDUP), can be processed immediately. Otherwise, put them on the
1034 * in-memory list for later processing.
1036 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
1037 txg
!= spa
->spa_syncing_txg
||
1038 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
1039 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1041 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1046 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1047 enum zio_flag flags
)
1050 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1052 ASSERT(!BP_IS_HOLE(bp
));
1053 ASSERT(spa_syncing_txg(spa
) == txg
);
1054 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
1056 if (BP_IS_EMBEDDED(bp
))
1057 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1059 metaslab_check_free(spa
, bp
);
1061 dsl_scan_freed(spa
, bp
);
1064 * GANG and DEDUP blocks can induce a read (for the gang block header,
1065 * or the DDT), so issue them asynchronously so that this thread is
1068 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1069 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1071 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1072 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1073 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1079 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1080 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1084 dprintf_bp(bp
, "claiming in txg %llu", txg
);
1086 if (BP_IS_EMBEDDED(bp
))
1087 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1090 * A claim is an allocation of a specific block. Claims are needed
1091 * to support immediate writes in the intent log. The issue is that
1092 * immediate writes contain committed data, but in a txg that was
1093 * *not* committed. Upon opening the pool after an unclean shutdown,
1094 * the intent log claims all blocks that contain immediate write data
1095 * so that the SPA knows they're in use.
1097 * All claims *must* be resolved in the first txg -- before the SPA
1098 * starts allocating blocks -- so that nothing is allocated twice.
1099 * If txg == 0 we just verify that the block is claimable.
1101 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
1102 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
1103 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1105 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1106 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1107 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1108 ASSERT0(zio
->io_queued_timestamp
);
1114 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1115 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1120 if (vd
->vdev_children
== 0) {
1121 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1122 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1123 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1127 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1129 for (c
= 0; c
< vd
->vdev_children
; c
++)
1130 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1131 done
, private, flags
));
1138 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1139 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1140 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1144 ASSERT(vd
->vdev_children
== 0);
1145 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1146 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1147 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1149 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1150 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1151 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1153 zio
->io_prop
.zp_checksum
= checksum
;
1159 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1160 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1161 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1165 ASSERT(vd
->vdev_children
== 0);
1166 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1167 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1168 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1170 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1171 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1172 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1174 zio
->io_prop
.zp_checksum
= checksum
;
1176 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1178 * zec checksums are necessarily destructive -- they modify
1179 * the end of the write buffer to hold the verifier/checksum.
1180 * Therefore, we must make a local copy in case the data is
1181 * being written to multiple places in parallel.
1183 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1184 abd_copy(wbuf
, data
, size
);
1186 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1193 * Create a child I/O to do some work for us.
1196 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1197 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1198 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1200 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1203 ASSERT(vd
->vdev_parent
==
1204 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
1206 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1208 * If we have the bp, then the child should perform the
1209 * checksum and the parent need not. This pushes error
1210 * detection as close to the leaves as possible and
1211 * eliminates redundant checksums in the interior nodes.
1213 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1214 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1217 if (vd
->vdev_children
== 0)
1218 offset
+= VDEV_LABEL_START_SIZE
;
1220 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
1223 * If we've decided to do a repair, the write is not speculative --
1224 * even if the original read was.
1226 if (flags
& ZIO_FLAG_IO_REPAIR
)
1227 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1230 * If we're creating a child I/O that is not associated with a
1231 * top-level vdev, then the child zio is not an allocating I/O.
1232 * If this is a retried I/O then we ignore it since we will
1233 * have already processed the original allocating I/O.
1235 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1236 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1237 ASSERTV(metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
));
1239 ASSERT(mc
->mc_alloc_throttle_enabled
);
1240 ASSERT(type
== ZIO_TYPE_WRITE
);
1241 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1242 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1243 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1244 pio
->io_child_type
== ZIO_CHILD_GANG
);
1246 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1250 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1251 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1252 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1253 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1255 zio
->io_physdone
= pio
->io_physdone
;
1256 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1257 zio
->io_logical
->io_phys_children
++;
1263 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1264 int type
, zio_priority_t priority
, enum zio_flag flags
,
1265 zio_done_func_t
*done
, void *private)
1269 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1271 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1272 data
, size
, size
, done
, private, type
, priority
,
1273 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1275 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1281 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1283 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1285 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1289 zio_shrink(zio_t
*zio
, uint64_t size
)
1291 ASSERT3P(zio
->io_executor
, ==, NULL
);
1292 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1293 ASSERT3U(size
, <=, zio
->io_size
);
1296 * We don't shrink for raidz because of problems with the
1297 * reconstruction when reading back less than the block size.
1298 * Note, BP_IS_RAIDZ() assumes no compression.
1300 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1301 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1302 /* we are not doing a raw write */
1303 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1304 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1309 * ==========================================================================
1310 * Prepare to read and write logical blocks
1311 * ==========================================================================
1315 zio_read_bp_init(zio_t
*zio
)
1317 blkptr_t
*bp
= zio
->io_bp
;
1319 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1321 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1322 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1323 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1324 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1325 psize
, psize
, zio_decompress
);
1328 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1329 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1330 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1331 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1332 psize
, psize
, zio_decrypt
);
1335 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1336 int psize
= BPE_GET_PSIZE(bp
);
1337 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1339 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1340 decode_embedded_bp_compressed(bp
, data
);
1341 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1343 ASSERT(!BP_IS_EMBEDDED(bp
));
1346 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1347 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1349 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1350 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1352 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1353 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1355 return (ZIO_PIPELINE_CONTINUE
);
1359 zio_write_bp_init(zio_t
*zio
)
1361 if (!IO_IS_ALLOCATING(zio
))
1362 return (ZIO_PIPELINE_CONTINUE
);
1364 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1366 if (zio
->io_bp_override
) {
1367 blkptr_t
*bp
= zio
->io_bp
;
1368 zio_prop_t
*zp
= &zio
->io_prop
;
1370 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1371 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1373 *bp
= *zio
->io_bp_override
;
1374 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1376 if (BP_IS_EMBEDDED(bp
))
1377 return (ZIO_PIPELINE_CONTINUE
);
1380 * If we've been overridden and nopwrite is set then
1381 * set the flag accordingly to indicate that a nopwrite
1382 * has already occurred.
1384 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1385 ASSERT(!zp
->zp_dedup
);
1386 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1387 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1388 return (ZIO_PIPELINE_CONTINUE
);
1391 ASSERT(!zp
->zp_nopwrite
);
1393 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1394 return (ZIO_PIPELINE_CONTINUE
);
1396 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1397 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1399 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1401 BP_SET_DEDUP(bp
, 1);
1402 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1403 return (ZIO_PIPELINE_CONTINUE
);
1407 * We were unable to handle this as an override bp, treat
1408 * it as a regular write I/O.
1410 zio
->io_bp_override
= NULL
;
1411 *bp
= zio
->io_bp_orig
;
1412 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1415 return (ZIO_PIPELINE_CONTINUE
);
1419 zio_write_compress(zio_t
*zio
)
1421 spa_t
*spa
= zio
->io_spa
;
1422 zio_prop_t
*zp
= &zio
->io_prop
;
1423 enum zio_compress compress
= zp
->zp_compress
;
1424 blkptr_t
*bp
= zio
->io_bp
;
1425 uint64_t lsize
= zio
->io_lsize
;
1426 uint64_t psize
= zio
->io_size
;
1430 * If our children haven't all reached the ready stage,
1431 * wait for them and then repeat this pipeline stage.
1433 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1434 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1435 return (ZIO_PIPELINE_STOP
);
1438 if (!IO_IS_ALLOCATING(zio
))
1439 return (ZIO_PIPELINE_CONTINUE
);
1441 if (zio
->io_children_ready
!= NULL
) {
1443 * Now that all our children are ready, run the callback
1444 * associated with this zio in case it wants to modify the
1445 * data to be written.
1447 ASSERT3U(zp
->zp_level
, >, 0);
1448 zio
->io_children_ready(zio
);
1451 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1452 ASSERT(zio
->io_bp_override
== NULL
);
1454 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1456 * We're rewriting an existing block, which means we're
1457 * working on behalf of spa_sync(). For spa_sync() to
1458 * converge, it must eventually be the case that we don't
1459 * have to allocate new blocks. But compression changes
1460 * the blocksize, which forces a reallocate, and makes
1461 * convergence take longer. Therefore, after the first
1462 * few passes, stop compressing to ensure convergence.
1464 pass
= spa_sync_pass(spa
);
1466 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1467 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1468 ASSERT(!BP_GET_DEDUP(bp
));
1470 if (pass
>= zfs_sync_pass_dont_compress
)
1471 compress
= ZIO_COMPRESS_OFF
;
1473 /* Make sure someone doesn't change their mind on overwrites */
1474 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1475 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1478 /* If it's a compressed write that is not raw, compress the buffer. */
1479 if (compress
!= ZIO_COMPRESS_OFF
&&
1480 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1481 void *cbuf
= zio_buf_alloc(lsize
);
1482 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1483 if (psize
== 0 || psize
== lsize
) {
1484 compress
= ZIO_COMPRESS_OFF
;
1485 zio_buf_free(cbuf
, lsize
);
1486 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1487 psize
<= BPE_PAYLOAD_SIZE
&&
1488 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1489 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1490 encode_embedded_bp_compressed(bp
,
1491 cbuf
, compress
, lsize
, psize
);
1492 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1493 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1494 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1495 zio_buf_free(cbuf
, lsize
);
1496 bp
->blk_birth
= zio
->io_txg
;
1497 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1498 ASSERT(spa_feature_is_active(spa
,
1499 SPA_FEATURE_EMBEDDED_DATA
));
1500 return (ZIO_PIPELINE_CONTINUE
);
1503 * Round up compressed size up to the ashift
1504 * of the smallest-ashift device, and zero the tail.
1505 * This ensures that the compressed size of the BP
1506 * (and thus compressratio property) are correct,
1507 * in that we charge for the padding used to fill out
1510 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1511 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1512 1ULL << spa
->spa_min_ashift
);
1513 if (rounded
>= lsize
) {
1514 compress
= ZIO_COMPRESS_OFF
;
1515 zio_buf_free(cbuf
, lsize
);
1518 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1519 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1520 abd_zero_off(cdata
, psize
, rounded
- psize
);
1522 zio_push_transform(zio
, cdata
,
1523 psize
, lsize
, NULL
);
1528 * We were unable to handle this as an override bp, treat
1529 * it as a regular write I/O.
1531 zio
->io_bp_override
= NULL
;
1532 *bp
= zio
->io_bp_orig
;
1533 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1535 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1536 zp
->zp_type
== DMU_OT_DNODE
) {
1538 * The DMU actually relies on the zio layer's compression
1539 * to free metadnode blocks that have had all contained
1540 * dnodes freed. As a result, even when doing a raw
1541 * receive, we must check whether the block can be compressed
1544 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1545 zio
->io_abd
, NULL
, lsize
);
1547 compress
= ZIO_COMPRESS_OFF
;
1549 ASSERT3U(psize
, !=, 0);
1553 * The final pass of spa_sync() must be all rewrites, but the first
1554 * few passes offer a trade-off: allocating blocks defers convergence,
1555 * but newly allocated blocks are sequential, so they can be written
1556 * to disk faster. Therefore, we allow the first few passes of
1557 * spa_sync() to allocate new blocks, but force rewrites after that.
1558 * There should only be a handful of blocks after pass 1 in any case.
1560 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1561 BP_GET_PSIZE(bp
) == psize
&&
1562 pass
>= zfs_sync_pass_rewrite
) {
1564 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1565 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1566 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1569 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1573 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1574 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1575 BP_SET_LSIZE(bp
, lsize
);
1576 BP_SET_TYPE(bp
, zp
->zp_type
);
1577 BP_SET_LEVEL(bp
, zp
->zp_level
);
1578 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1580 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1582 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1583 BP_SET_LSIZE(bp
, lsize
);
1584 BP_SET_TYPE(bp
, zp
->zp_type
);
1585 BP_SET_LEVEL(bp
, zp
->zp_level
);
1586 BP_SET_PSIZE(bp
, psize
);
1587 BP_SET_COMPRESS(bp
, compress
);
1588 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1589 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1590 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1592 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1593 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1594 ASSERT(!zp
->zp_encrypt
||
1595 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1596 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1598 if (zp
->zp_nopwrite
) {
1599 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1600 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1601 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1604 return (ZIO_PIPELINE_CONTINUE
);
1608 zio_free_bp_init(zio_t
*zio
)
1610 blkptr_t
*bp
= zio
->io_bp
;
1612 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1613 if (BP_GET_DEDUP(bp
))
1614 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1617 return (ZIO_PIPELINE_CONTINUE
);
1621 * ==========================================================================
1622 * Execute the I/O pipeline
1623 * ==========================================================================
1627 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1629 spa_t
*spa
= zio
->io_spa
;
1630 zio_type_t t
= zio
->io_type
;
1631 int flags
= (cutinline
? TQ_FRONT
: 0);
1634 * If we're a config writer or a probe, the normal issue and
1635 * interrupt threads may all be blocked waiting for the config lock.
1636 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1638 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1642 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1644 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1648 * If this is a high priority I/O, then use the high priority taskq if
1651 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1652 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1655 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1658 * NB: We are assuming that the zio can only be dispatched
1659 * to a single taskq at a time. It would be a grievous error
1660 * to dispatch the zio to another taskq at the same time.
1662 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1663 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1664 flags
, &zio
->io_tqent
);
1668 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1670 kthread_t
*executor
= zio
->io_executor
;
1671 spa_t
*spa
= zio
->io_spa
;
1673 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1674 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1676 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1677 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1686 zio_issue_async(zio_t
*zio
)
1688 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1690 return (ZIO_PIPELINE_STOP
);
1694 zio_interrupt(zio_t
*zio
)
1696 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1700 zio_delay_interrupt(zio_t
*zio
)
1703 * The timeout_generic() function isn't defined in userspace, so
1704 * rather than trying to implement the function, the zio delay
1705 * functionality has been disabled for userspace builds.
1710 * If io_target_timestamp is zero, then no delay has been registered
1711 * for this IO, thus jump to the end of this function and "skip" the
1712 * delay; issuing it directly to the zio layer.
1714 if (zio
->io_target_timestamp
!= 0) {
1715 hrtime_t now
= gethrtime();
1717 if (now
>= zio
->io_target_timestamp
) {
1719 * This IO has already taken longer than the target
1720 * delay to complete, so we don't want to delay it
1721 * any longer; we "miss" the delay and issue it
1722 * directly to the zio layer. This is likely due to
1723 * the target latency being set to a value less than
1724 * the underlying hardware can satisfy (e.g. delay
1725 * set to 1ms, but the disks take 10ms to complete an
1729 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1735 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1736 clock_t expire_at_tick
= ddi_get_lbolt() +
1739 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1740 hrtime_t
, now
, hrtime_t
, diff
);
1742 if (NSEC_TO_TICK(diff
) == 0) {
1743 /* Our delay is less than a jiffy - just spin */
1744 zfs_sleep_until(zio
->io_target_timestamp
);
1747 * Use taskq_dispatch_delay() in the place of
1748 * OpenZFS's timeout_generic().
1750 tid
= taskq_dispatch_delay(system_taskq
,
1751 (task_func_t
*)zio_interrupt
,
1752 zio
, TQ_NOSLEEP
, expire_at_tick
);
1753 if (tid
== TASKQID_INVALID
) {
1755 * Couldn't allocate a task. Just
1756 * finish the zio without a delay.
1765 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1770 zio_deadman_impl(zio_t
*pio
)
1772 zio_t
*cio
, *cio_next
;
1773 zio_link_t
*zl
= NULL
;
1774 vdev_t
*vd
= pio
->io_vd
;
1776 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
1777 vdev_queue_t
*vq
= &vd
->vdev_queue
;
1778 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1779 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1780 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1782 zfs_dbgmsg("slow zio: zio=%p timestamp=%llu "
1783 "delta=%llu queued=%llu io=%llu "
1784 "path=%s last=%llu "
1785 "type=%d priority=%d flags=0x%x "
1786 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1787 "objset=%llu object=%llu level=%llu blkid=%llu "
1788 "offset=%llu size=%llu error=%d",
1789 pio
, pio
->io_timestamp
,
1790 delta
, pio
->io_delta
, pio
->io_delay
,
1791 vd
->vdev_path
, vq
->vq_io_complete_ts
,
1792 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
1793 pio
->io_state
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
1794 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
1795 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
1796 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
1797 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
1799 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
1800 taskq_empty_ent(&pio
->io_tqent
)) {
1805 mutex_enter(&pio
->io_lock
);
1806 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1807 cio_next
= zio_walk_children(pio
, &zl
);
1808 zio_deadman_impl(cio
);
1810 mutex_exit(&pio
->io_lock
);
1814 * Log the critical information describing this zio and all of its children
1815 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1818 zio_deadman(zio_t
*pio
, char *tag
)
1820 spa_t
*spa
= pio
->io_spa
;
1821 char *name
= spa_name(spa
);
1823 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
1826 zio_deadman_impl(pio
);
1828 switch (spa_get_deadman_failmode(spa
)) {
1829 case ZIO_FAILURE_MODE_WAIT
:
1830 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
1833 case ZIO_FAILURE_MODE_CONTINUE
:
1834 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
1837 case ZIO_FAILURE_MODE_PANIC
:
1838 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
1844 * Execute the I/O pipeline until one of the following occurs:
1845 * (1) the I/O completes; (2) the pipeline stalls waiting for
1846 * dependent child I/Os; (3) the I/O issues, so we're waiting
1847 * for an I/O completion interrupt; (4) the I/O is delegated by
1848 * vdev-level caching or aggregation; (5) the I/O is deferred
1849 * due to vdev-level queueing; (6) the I/O is handed off to
1850 * another thread. In all cases, the pipeline stops whenever
1851 * there's no CPU work; it never burns a thread in cv_wait_io().
1853 * There's no locking on io_stage because there's no legitimate way
1854 * for multiple threads to be attempting to process the same I/O.
1856 static zio_pipe_stage_t
*zio_pipeline
[];
1859 * zio_execute() is a wrapper around the static function
1860 * __zio_execute() so that we can force __zio_execute() to be
1861 * inlined. This reduces stack overhead which is important
1862 * because __zio_execute() is called recursively in several zio
1863 * code paths. zio_execute() itself cannot be inlined because
1864 * it is externally visible.
1867 zio_execute(zio_t
*zio
)
1869 fstrans_cookie_t cookie
;
1871 cookie
= spl_fstrans_mark();
1873 spl_fstrans_unmark(cookie
);
1877 * Used to determine if in the current context the stack is sized large
1878 * enough to allow zio_execute() to be called recursively. A minimum
1879 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1882 zio_execute_stack_check(zio_t
*zio
)
1884 #if !defined(HAVE_LARGE_STACKS)
1885 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1887 /* Executing in txg_sync_thread() context. */
1888 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1891 /* Pool initialization outside of zio_taskq context. */
1892 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1893 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1894 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1896 #endif /* HAVE_LARGE_STACKS */
1901 __attribute__((always_inline
))
1903 __zio_execute(zio_t
*zio
)
1905 zio
->io_executor
= curthread
;
1907 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1909 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1910 enum zio_stage pipeline
= zio
->io_pipeline
;
1911 enum zio_stage stage
= zio
->io_stage
;
1914 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1915 ASSERT(ISP2(stage
));
1916 ASSERT(zio
->io_stall
== NULL
);
1920 } while ((stage
& pipeline
) == 0);
1922 ASSERT(stage
<= ZIO_STAGE_DONE
);
1925 * If we are in interrupt context and this pipeline stage
1926 * will grab a config lock that is held across I/O,
1927 * or may wait for an I/O that needs an interrupt thread
1928 * to complete, issue async to avoid deadlock.
1930 * For VDEV_IO_START, we cut in line so that the io will
1931 * be sent to disk promptly.
1933 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1934 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1935 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1936 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1937 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1942 * If the current context doesn't have large enough stacks
1943 * the zio must be issued asynchronously to prevent overflow.
1945 if (zio_execute_stack_check(zio
)) {
1946 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1947 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1948 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1952 zio
->io_stage
= stage
;
1953 zio
->io_pipeline_trace
|= zio
->io_stage
;
1954 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1956 if (rv
== ZIO_PIPELINE_STOP
)
1959 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1965 * ==========================================================================
1966 * Initiate I/O, either sync or async
1967 * ==========================================================================
1970 zio_wait(zio_t
*zio
)
1972 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
1975 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
1976 ASSERT3P(zio
->io_executor
, ==, NULL
);
1978 zio
->io_waiter
= curthread
;
1979 ASSERT0(zio
->io_queued_timestamp
);
1980 zio
->io_queued_timestamp
= gethrtime();
1984 mutex_enter(&zio
->io_lock
);
1985 while (zio
->io_executor
!= NULL
) {
1986 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
1987 ddi_get_lbolt() + timeout
);
1989 if (zfs_deadman_enabled
&& error
== -1 &&
1990 gethrtime() - zio
->io_queued_timestamp
>
1991 spa_deadman_ziotime(zio
->io_spa
)) {
1992 mutex_exit(&zio
->io_lock
);
1993 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
1994 zio_deadman(zio
, FTAG
);
1995 mutex_enter(&zio
->io_lock
);
1998 mutex_exit(&zio
->io_lock
);
2000 error
= zio
->io_error
;
2007 zio_nowait(zio_t
*zio
)
2009 ASSERT3P(zio
->io_executor
, ==, NULL
);
2011 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2012 zio_unique_parent(zio
) == NULL
) {
2016 * This is a logical async I/O with no parent to wait for it.
2017 * We add it to the spa_async_root_zio "Godfather" I/O which
2018 * will ensure they complete prior to unloading the pool.
2020 spa_t
*spa
= zio
->io_spa
;
2022 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2025 zio_add_child(pio
, zio
);
2028 ASSERT0(zio
->io_queued_timestamp
);
2029 zio
->io_queued_timestamp
= gethrtime();
2034 * ==========================================================================
2035 * Reexecute, cancel, or suspend/resume failed I/O
2036 * ==========================================================================
2040 zio_reexecute(zio_t
*pio
)
2042 zio_t
*cio
, *cio_next
;
2044 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2045 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2046 ASSERT(pio
->io_gang_leader
== NULL
);
2047 ASSERT(pio
->io_gang_tree
== NULL
);
2049 pio
->io_flags
= pio
->io_orig_flags
;
2050 pio
->io_stage
= pio
->io_orig_stage
;
2051 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2052 pio
->io_reexecute
= 0;
2053 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2054 pio
->io_pipeline_trace
= 0;
2056 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2057 pio
->io_state
[w
] = 0;
2058 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2059 pio
->io_child_error
[c
] = 0;
2061 if (IO_IS_ALLOCATING(pio
))
2062 BP_ZERO(pio
->io_bp
);
2065 * As we reexecute pio's children, new children could be created.
2066 * New children go to the head of pio's io_child_list, however,
2067 * so we will (correctly) not reexecute them. The key is that
2068 * the remainder of pio's io_child_list, from 'cio_next' onward,
2069 * cannot be affected by any side effects of reexecuting 'cio'.
2071 zio_link_t
*zl
= NULL
;
2072 mutex_enter(&pio
->io_lock
);
2073 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2074 cio_next
= zio_walk_children(pio
, &zl
);
2075 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2076 pio
->io_children
[cio
->io_child_type
][w
]++;
2077 mutex_exit(&pio
->io_lock
);
2079 mutex_enter(&pio
->io_lock
);
2081 mutex_exit(&pio
->io_lock
);
2084 * Now that all children have been reexecuted, execute the parent.
2085 * We don't reexecute "The Godfather" I/O here as it's the
2086 * responsibility of the caller to wait on it.
2088 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2089 pio
->io_queued_timestamp
= gethrtime();
2095 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2097 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2098 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2099 "failure and the failure mode property for this pool "
2100 "is set to panic.", spa_name(spa
));
2102 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2103 "failure and has been suspended.\n", spa_name(spa
));
2105 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2108 mutex_enter(&spa
->spa_suspend_lock
);
2110 if (spa
->spa_suspend_zio_root
== NULL
)
2111 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2112 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2113 ZIO_FLAG_GODFATHER
);
2115 spa
->spa_suspended
= reason
;
2118 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2119 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2120 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2121 ASSERT(zio_unique_parent(zio
) == NULL
);
2122 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2123 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2126 mutex_exit(&spa
->spa_suspend_lock
);
2130 zio_resume(spa_t
*spa
)
2135 * Reexecute all previously suspended i/o.
2137 mutex_enter(&spa
->spa_suspend_lock
);
2138 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2139 cv_broadcast(&spa
->spa_suspend_cv
);
2140 pio
= spa
->spa_suspend_zio_root
;
2141 spa
->spa_suspend_zio_root
= NULL
;
2142 mutex_exit(&spa
->spa_suspend_lock
);
2148 return (zio_wait(pio
));
2152 zio_resume_wait(spa_t
*spa
)
2154 mutex_enter(&spa
->spa_suspend_lock
);
2155 while (spa_suspended(spa
))
2156 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2157 mutex_exit(&spa
->spa_suspend_lock
);
2161 * ==========================================================================
2164 * A gang block is a collection of small blocks that looks to the DMU
2165 * like one large block. When zio_dva_allocate() cannot find a block
2166 * of the requested size, due to either severe fragmentation or the pool
2167 * being nearly full, it calls zio_write_gang_block() to construct the
2168 * block from smaller fragments.
2170 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2171 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2172 * an indirect block: it's an array of block pointers. It consumes
2173 * only one sector and hence is allocatable regardless of fragmentation.
2174 * The gang header's bps point to its gang members, which hold the data.
2176 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2177 * as the verifier to ensure uniqueness of the SHA256 checksum.
2178 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2179 * not the gang header. This ensures that data block signatures (needed for
2180 * deduplication) are independent of how the block is physically stored.
2182 * Gang blocks can be nested: a gang member may itself be a gang block.
2183 * Thus every gang block is a tree in which root and all interior nodes are
2184 * gang headers, and the leaves are normal blocks that contain user data.
2185 * The root of the gang tree is called the gang leader.
2187 * To perform any operation (read, rewrite, free, claim) on a gang block,
2188 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2189 * in the io_gang_tree field of the original logical i/o by recursively
2190 * reading the gang leader and all gang headers below it. This yields
2191 * an in-core tree containing the contents of every gang header and the
2192 * bps for every constituent of the gang block.
2194 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2195 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2196 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2197 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2198 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2199 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2200 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2201 * of the gang header plus zio_checksum_compute() of the data to update the
2202 * gang header's blk_cksum as described above.
2204 * The two-phase assemble/issue model solves the problem of partial failure --
2205 * what if you'd freed part of a gang block but then couldn't read the
2206 * gang header for another part? Assembling the entire gang tree first
2207 * ensures that all the necessary gang header I/O has succeeded before
2208 * starting the actual work of free, claim, or write. Once the gang tree
2209 * is assembled, free and claim are in-memory operations that cannot fail.
2211 * In the event that a gang write fails, zio_dva_unallocate() walks the
2212 * gang tree to immediately free (i.e. insert back into the space map)
2213 * everything we've allocated. This ensures that we don't get ENOSPC
2214 * errors during repeated suspend/resume cycles due to a flaky device.
2216 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2217 * the gang tree, we won't modify the block, so we can safely defer the free
2218 * (knowing that the block is still intact). If we *can* assemble the gang
2219 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2220 * each constituent bp and we can allocate a new block on the next sync pass.
2222 * In all cases, the gang tree allows complete recovery from partial failure.
2223 * ==========================================================================
2227 zio_gang_issue_func_done(zio_t
*zio
)
2229 abd_put(zio
->io_abd
);
2233 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2239 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2240 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2241 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2242 &pio
->io_bookmark
));
2246 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2253 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2254 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2255 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2256 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2259 * As we rewrite each gang header, the pipeline will compute
2260 * a new gang block header checksum for it; but no one will
2261 * compute a new data checksum, so we do that here. The one
2262 * exception is the gang leader: the pipeline already computed
2263 * its data checksum because that stage precedes gang assembly.
2264 * (Presently, nothing actually uses interior data checksums;
2265 * this is just good hygiene.)
2267 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2268 abd_t
*buf
= abd_get_offset(data
, offset
);
2270 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2271 buf
, BP_GET_PSIZE(bp
));
2276 * If we are here to damage data for testing purposes,
2277 * leave the GBH alone so that we can detect the damage.
2279 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2280 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2282 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2283 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2284 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2285 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2293 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2296 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2297 ZIO_GANG_CHILD_FLAGS(pio
)));
2302 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2305 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2306 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2309 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2318 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2320 static zio_gang_node_t
*
2321 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2323 zio_gang_node_t
*gn
;
2325 ASSERT(*gnpp
== NULL
);
2327 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2328 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2335 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2337 zio_gang_node_t
*gn
= *gnpp
;
2339 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2340 ASSERT(gn
->gn_child
[g
] == NULL
);
2342 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2343 kmem_free(gn
, sizeof (*gn
));
2348 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2350 zio_gang_node_t
*gn
= *gnpp
;
2355 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2356 zio_gang_tree_free(&gn
->gn_child
[g
]);
2358 zio_gang_node_free(gnpp
);
2362 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2364 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2365 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2367 ASSERT(gio
->io_gang_leader
== gio
);
2368 ASSERT(BP_IS_GANG(bp
));
2370 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2371 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2372 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2376 zio_gang_tree_assemble_done(zio_t
*zio
)
2378 zio_t
*gio
= zio
->io_gang_leader
;
2379 zio_gang_node_t
*gn
= zio
->io_private
;
2380 blkptr_t
*bp
= zio
->io_bp
;
2382 ASSERT(gio
== zio_unique_parent(zio
));
2383 ASSERT(zio
->io_child_count
== 0);
2388 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2389 if (BP_SHOULD_BYTESWAP(bp
))
2390 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2392 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2393 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2394 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2396 abd_put(zio
->io_abd
);
2398 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2399 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2400 if (!BP_IS_GANG(gbp
))
2402 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2407 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2410 zio_t
*gio
= pio
->io_gang_leader
;
2413 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2414 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2415 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2418 * If you're a gang header, your data is in gn->gn_gbh.
2419 * If you're a gang member, your data is in 'data' and gn == NULL.
2421 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2424 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2426 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2427 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2428 if (BP_IS_HOLE(gbp
))
2430 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2432 offset
+= BP_GET_PSIZE(gbp
);
2436 if (gn
== gio
->io_gang_tree
)
2437 ASSERT3U(gio
->io_size
, ==, offset
);
2444 zio_gang_assemble(zio_t
*zio
)
2446 blkptr_t
*bp
= zio
->io_bp
;
2448 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2449 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2451 zio
->io_gang_leader
= zio
;
2453 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2455 return (ZIO_PIPELINE_CONTINUE
);
2459 zio_gang_issue(zio_t
*zio
)
2461 blkptr_t
*bp
= zio
->io_bp
;
2463 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2464 return (ZIO_PIPELINE_STOP
);
2467 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2468 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2470 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2471 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2474 zio_gang_tree_free(&zio
->io_gang_tree
);
2476 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2478 return (ZIO_PIPELINE_CONTINUE
);
2482 zio_write_gang_member_ready(zio_t
*zio
)
2484 zio_t
*pio
= zio_unique_parent(zio
);
2485 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2486 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2488 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
);
2490 if (BP_IS_HOLE(zio
->io_bp
))
2493 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2495 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2496 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2497 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2498 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2499 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2501 mutex_enter(&pio
->io_lock
);
2502 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2503 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2504 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2505 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2506 DVA_SET_ASIZE(&pdva
[d
], asize
);
2508 mutex_exit(&pio
->io_lock
);
2512 zio_write_gang_done(zio_t
*zio
)
2514 abd_put(zio
->io_abd
);
2518 zio_write_gang_block(zio_t
*pio
)
2520 spa_t
*spa
= pio
->io_spa
;
2521 metaslab_class_t
*mc
= spa_normal_class(spa
);
2522 blkptr_t
*bp
= pio
->io_bp
;
2523 zio_t
*gio
= pio
->io_gang_leader
;
2525 zio_gang_node_t
*gn
, **gnpp
;
2526 zio_gbh_phys_t
*gbh
;
2528 uint64_t txg
= pio
->io_txg
;
2529 uint64_t resid
= pio
->io_size
;
2531 int copies
= gio
->io_prop
.zp_copies
;
2537 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2538 * have a third copy.
2540 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2541 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2542 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2544 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2545 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2546 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2547 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2549 flags
|= METASLAB_ASYNC_ALLOC
;
2550 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2553 * The logical zio has already placed a reservation for
2554 * 'copies' allocation slots but gang blocks may require
2555 * additional copies. These additional copies
2556 * (i.e. gbh_copies - copies) are guaranteed to succeed
2557 * since metaslab_class_throttle_reserve() always allows
2558 * additional reservations for gang blocks.
2560 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2564 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2565 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2566 &pio
->io_alloc_list
, pio
);
2568 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2569 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2570 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2573 * If we failed to allocate the gang block header then
2574 * we remove any additional allocation reservations that
2575 * we placed here. The original reservation will
2576 * be removed when the logical I/O goes to the ready
2579 metaslab_class_throttle_unreserve(mc
,
2580 gbh_copies
- copies
, pio
);
2583 pio
->io_error
= error
;
2584 return (ZIO_PIPELINE_CONTINUE
);
2588 gnpp
= &gio
->io_gang_tree
;
2590 gnpp
= pio
->io_private
;
2591 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2594 gn
= zio_gang_node_alloc(gnpp
);
2596 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2597 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2600 * Create the gang header.
2602 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2603 zio_write_gang_done
, NULL
, pio
->io_priority
,
2604 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2607 * Create and nowait the gang children.
2609 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2610 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2612 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2614 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2615 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2616 zp
.zp_type
= DMU_OT_NONE
;
2618 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2619 zp
.zp_dedup
= B_FALSE
;
2620 zp
.zp_dedup_verify
= B_FALSE
;
2621 zp
.zp_nopwrite
= B_FALSE
;
2622 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2623 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2624 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2625 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2626 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2628 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2629 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2630 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2631 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2632 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2634 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2635 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2636 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2639 * Gang children won't throttle but we should
2640 * account for their work, so reserve an allocation
2641 * slot for them here.
2643 VERIFY(metaslab_class_throttle_reserve(mc
,
2644 zp
.zp_copies
, cio
, flags
));
2650 * Set pio's pipeline to just wait for zio to finish.
2652 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2655 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2657 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2661 return (ZIO_PIPELINE_CONTINUE
);
2665 * The zio_nop_write stage in the pipeline determines if allocating a
2666 * new bp is necessary. The nopwrite feature can handle writes in
2667 * either syncing or open context (i.e. zil writes) and as a result is
2668 * mutually exclusive with dedup.
2670 * By leveraging a cryptographically secure checksum, such as SHA256, we
2671 * can compare the checksums of the new data and the old to determine if
2672 * allocating a new block is required. Note that our requirements for
2673 * cryptographic strength are fairly weak: there can't be any accidental
2674 * hash collisions, but we don't need to be secure against intentional
2675 * (malicious) collisions. To trigger a nopwrite, you have to be able
2676 * to write the file to begin with, and triggering an incorrect (hash
2677 * collision) nopwrite is no worse than simply writing to the file.
2678 * That said, there are no known attacks against the checksum algorithms
2679 * used for nopwrite, assuming that the salt and the checksums
2680 * themselves remain secret.
2683 zio_nop_write(zio_t
*zio
)
2685 blkptr_t
*bp
= zio
->io_bp
;
2686 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2687 zio_prop_t
*zp
= &zio
->io_prop
;
2689 ASSERT(BP_GET_LEVEL(bp
) == 0);
2690 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2691 ASSERT(zp
->zp_nopwrite
);
2692 ASSERT(!zp
->zp_dedup
);
2693 ASSERT(zio
->io_bp_override
== NULL
);
2694 ASSERT(IO_IS_ALLOCATING(zio
));
2697 * Check to see if the original bp and the new bp have matching
2698 * characteristics (i.e. same checksum, compression algorithms, etc).
2699 * If they don't then just continue with the pipeline which will
2700 * allocate a new bp.
2702 if (BP_IS_HOLE(bp_orig
) ||
2703 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2704 ZCHECKSUM_FLAG_NOPWRITE
) ||
2705 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2706 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2707 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2708 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2709 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2710 return (ZIO_PIPELINE_CONTINUE
);
2713 * If the checksums match then reset the pipeline so that we
2714 * avoid allocating a new bp and issuing any I/O.
2716 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2717 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2718 ZCHECKSUM_FLAG_NOPWRITE
);
2719 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2720 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2721 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2722 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2723 sizeof (uint64_t)) == 0);
2726 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2727 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2730 return (ZIO_PIPELINE_CONTINUE
);
2734 * ==========================================================================
2736 * ==========================================================================
2739 zio_ddt_child_read_done(zio_t
*zio
)
2741 blkptr_t
*bp
= zio
->io_bp
;
2742 ddt_entry_t
*dde
= zio
->io_private
;
2744 zio_t
*pio
= zio_unique_parent(zio
);
2746 mutex_enter(&pio
->io_lock
);
2747 ddp
= ddt_phys_select(dde
, bp
);
2748 if (zio
->io_error
== 0)
2749 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2751 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2752 dde
->dde_repair_abd
= zio
->io_abd
;
2754 abd_free(zio
->io_abd
);
2755 mutex_exit(&pio
->io_lock
);
2759 zio_ddt_read_start(zio_t
*zio
)
2761 blkptr_t
*bp
= zio
->io_bp
;
2763 ASSERT(BP_GET_DEDUP(bp
));
2764 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2765 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2767 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2768 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2769 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2770 ddt_phys_t
*ddp
= dde
->dde_phys
;
2771 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2774 ASSERT(zio
->io_vsd
== NULL
);
2777 if (ddp_self
== NULL
)
2778 return (ZIO_PIPELINE_CONTINUE
);
2780 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2781 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2783 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2785 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2786 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2787 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2788 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2789 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2791 return (ZIO_PIPELINE_CONTINUE
);
2794 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2795 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2796 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2798 return (ZIO_PIPELINE_CONTINUE
);
2802 zio_ddt_read_done(zio_t
*zio
)
2804 blkptr_t
*bp
= zio
->io_bp
;
2806 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
2807 return (ZIO_PIPELINE_STOP
);
2810 ASSERT(BP_GET_DEDUP(bp
));
2811 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2812 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2814 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2815 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2816 ddt_entry_t
*dde
= zio
->io_vsd
;
2818 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2819 return (ZIO_PIPELINE_CONTINUE
);
2822 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2823 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2824 return (ZIO_PIPELINE_STOP
);
2826 if (dde
->dde_repair_abd
!= NULL
) {
2827 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2829 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2831 ddt_repair_done(ddt
, dde
);
2835 ASSERT(zio
->io_vsd
== NULL
);
2837 return (ZIO_PIPELINE_CONTINUE
);
2841 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2843 spa_t
*spa
= zio
->io_spa
;
2844 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2846 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2849 * Note: we compare the original data, not the transformed data,
2850 * because when zio->io_bp is an override bp, we will not have
2851 * pushed the I/O transforms. That's an important optimization
2852 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2853 * However, we should never get a raw, override zio so in these
2854 * cases we can compare the io_abd directly. This is useful because
2855 * it allows us to do dedup verification even if we don't have access
2856 * to the original data (for instance, if the encryption keys aren't
2860 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2861 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2863 if (lio
!= NULL
&& do_raw
) {
2864 return (lio
->io_size
!= zio
->io_size
||
2865 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2866 } else if (lio
!= NULL
) {
2867 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2868 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2872 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2873 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2875 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2876 blkptr_t blk
= *zio
->io_bp
;
2881 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2882 psize
= BP_GET_PSIZE(&blk
);
2884 if (psize
!= zio
->io_size
)
2889 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2891 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2892 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2893 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2894 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
2897 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
2898 error
= SET_ERROR(ENOENT
);
2903 return (error
!= 0);
2904 } else if (ddp
->ddp_phys_birth
!= 0) {
2905 arc_buf_t
*abuf
= NULL
;
2906 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2907 blkptr_t blk
= *zio
->io_bp
;
2910 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2912 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
2917 error
= arc_read(NULL
, spa
, &blk
,
2918 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2919 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2920 &aflags
, &zio
->io_bookmark
);
2923 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2924 zio
->io_orig_size
) != 0)
2925 error
= SET_ERROR(ENOENT
);
2926 arc_buf_destroy(abuf
, &abuf
);
2930 return (error
!= 0);
2938 zio_ddt_child_write_ready(zio_t
*zio
)
2940 int p
= zio
->io_prop
.zp_copies
;
2941 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2942 ddt_entry_t
*dde
= zio
->io_private
;
2943 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2951 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2953 ddt_phys_fill(ddp
, zio
->io_bp
);
2955 zio_link_t
*zl
= NULL
;
2956 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2957 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2963 zio_ddt_child_write_done(zio_t
*zio
)
2965 int p
= zio
->io_prop
.zp_copies
;
2966 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2967 ddt_entry_t
*dde
= zio
->io_private
;
2968 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2972 ASSERT(ddp
->ddp_refcnt
== 0);
2973 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2974 dde
->dde_lead_zio
[p
] = NULL
;
2976 if (zio
->io_error
== 0) {
2977 zio_link_t
*zl
= NULL
;
2978 while (zio_walk_parents(zio
, &zl
) != NULL
)
2979 ddt_phys_addref(ddp
);
2981 ddt_phys_clear(ddp
);
2988 zio_ddt_ditto_write_done(zio_t
*zio
)
2990 int p
= DDT_PHYS_DITTO
;
2991 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
2992 blkptr_t
*bp
= zio
->io_bp
;
2993 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2994 ddt_entry_t
*dde
= zio
->io_private
;
2995 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2996 ddt_key_t
*ddk
= &dde
->dde_key
;
3000 ASSERT(ddp
->ddp_refcnt
== 0);
3001 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3002 dde
->dde_lead_zio
[p
] = NULL
;
3004 if (zio
->io_error
== 0) {
3005 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
3006 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
3007 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
3008 if (ddp
->ddp_phys_birth
!= 0)
3009 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
3010 ddt_phys_fill(ddp
, bp
);
3017 zio_ddt_write(zio_t
*zio
)
3019 spa_t
*spa
= zio
->io_spa
;
3020 blkptr_t
*bp
= zio
->io_bp
;
3021 uint64_t txg
= zio
->io_txg
;
3022 zio_prop_t
*zp
= &zio
->io_prop
;
3023 int p
= zp
->zp_copies
;
3027 ddt_t
*ddt
= ddt_select(spa
, bp
);
3031 ASSERT(BP_GET_DEDUP(bp
));
3032 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3033 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3034 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3037 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3038 ddp
= &dde
->dde_phys
[p
];
3040 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3042 * If we're using a weak checksum, upgrade to a strong checksum
3043 * and try again. If we're already using a strong checksum,
3044 * we can't resolve it, so just convert to an ordinary write.
3045 * (And automatically e-mail a paper to Nature?)
3047 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3048 ZCHECKSUM_FLAG_DEDUP
)) {
3049 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3050 zio_pop_transforms(zio
);
3051 zio
->io_stage
= ZIO_STAGE_OPEN
;
3054 zp
->zp_dedup
= B_FALSE
;
3056 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3058 return (ZIO_PIPELINE_CONTINUE
);
3061 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
3062 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
3064 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
3065 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
3066 zio_prop_t czp
= *zp
;
3068 czp
.zp_copies
= ditto_copies
;
3071 * If we arrived here with an override bp, we won't have run
3072 * the transform stack, so we won't have the data we need to
3073 * generate a child i/o. So, toss the override bp and restart.
3074 * This is safe, because using the override bp is just an
3075 * optimization; and it's rare, so the cost doesn't matter.
3077 if (zio
->io_bp_override
) {
3078 zio_pop_transforms(zio
);
3079 zio
->io_stage
= ZIO_STAGE_OPEN
;
3080 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3081 zio
->io_bp_override
= NULL
;
3084 return (ZIO_PIPELINE_CONTINUE
);
3087 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3088 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
3089 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
3090 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3092 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3093 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
3096 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3097 if (ddp
->ddp_phys_birth
!= 0)
3098 ddt_bp_fill(ddp
, bp
, txg
);
3099 if (dde
->dde_lead_zio
[p
] != NULL
)
3100 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3102 ddt_phys_addref(ddp
);
3103 } else if (zio
->io_bp_override
) {
3104 ASSERT(bp
->blk_birth
== txg
);
3105 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3106 ddt_phys_fill(ddp
, bp
);
3107 ddt_phys_addref(ddp
);
3109 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3110 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3111 zio_ddt_child_write_ready
, NULL
, NULL
,
3112 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3113 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3115 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3116 dde
->dde_lead_zio
[p
] = cio
;
3126 return (ZIO_PIPELINE_CONTINUE
);
3129 ddt_entry_t
*freedde
; /* for debugging */
3132 zio_ddt_free(zio_t
*zio
)
3134 spa_t
*spa
= zio
->io_spa
;
3135 blkptr_t
*bp
= zio
->io_bp
;
3136 ddt_t
*ddt
= ddt_select(spa
, bp
);
3140 ASSERT(BP_GET_DEDUP(bp
));
3141 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3144 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3146 ddp
= ddt_phys_select(dde
, bp
);
3148 ddt_phys_decref(ddp
);
3152 return (ZIO_PIPELINE_CONTINUE
);
3156 * ==========================================================================
3157 * Allocate and free blocks
3158 * ==========================================================================
3162 zio_io_to_allocate(spa_t
*spa
)
3166 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
3168 zio
= avl_first(&spa
->spa_alloc_tree
);
3172 ASSERT(IO_IS_ALLOCATING(zio
));
3175 * Try to place a reservation for this zio. If we're unable to
3176 * reserve then we throttle.
3178 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
3179 zio
->io_prop
.zp_copies
, zio
, 0)) {
3183 avl_remove(&spa
->spa_alloc_tree
, zio
);
3184 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3190 zio_dva_throttle(zio_t
*zio
)
3192 spa_t
*spa
= zio
->io_spa
;
3195 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3196 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
3197 zio
->io_child_type
== ZIO_CHILD_GANG
||
3198 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3199 return (ZIO_PIPELINE_CONTINUE
);
3202 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3204 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3205 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3207 mutex_enter(&spa
->spa_alloc_lock
);
3209 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3210 avl_add(&spa
->spa_alloc_tree
, zio
);
3212 nio
= zio_io_to_allocate(zio
->io_spa
);
3213 mutex_exit(&spa
->spa_alloc_lock
);
3216 return (ZIO_PIPELINE_CONTINUE
);
3219 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3221 * We are passing control to a new zio so make sure that
3222 * it is processed by a different thread. We do this to
3223 * avoid stack overflows that can occur when parents are
3224 * throttled and children are making progress. We allow
3225 * it to go to the head of the taskq since it's already
3228 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3230 return (ZIO_PIPELINE_STOP
);
3234 zio_allocate_dispatch(spa_t
*spa
)
3238 mutex_enter(&spa
->spa_alloc_lock
);
3239 zio
= zio_io_to_allocate(spa
);
3240 mutex_exit(&spa
->spa_alloc_lock
);
3244 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3245 ASSERT0(zio
->io_error
);
3246 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3250 zio_dva_allocate(zio_t
*zio
)
3252 spa_t
*spa
= zio
->io_spa
;
3253 metaslab_class_t
*mc
= spa_normal_class(spa
);
3254 blkptr_t
*bp
= zio
->io_bp
;
3258 if (zio
->io_gang_leader
== NULL
) {
3259 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3260 zio
->io_gang_leader
= zio
;
3263 ASSERT(BP_IS_HOLE(bp
));
3264 ASSERT0(BP_GET_NDVAS(bp
));
3265 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3266 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3267 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3269 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3270 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3271 flags
|= METASLAB_DONT_THROTTLE
;
3272 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3273 flags
|= METASLAB_GANG_CHILD
;
3274 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3275 flags
|= METASLAB_ASYNC_ALLOC
;
3277 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3278 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3279 &zio
->io_alloc_list
, zio
);
3282 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
3283 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3285 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3286 return (zio_write_gang_block(zio
));
3287 zio
->io_error
= error
;
3290 return (ZIO_PIPELINE_CONTINUE
);
3294 zio_dva_free(zio_t
*zio
)
3296 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3298 return (ZIO_PIPELINE_CONTINUE
);
3302 zio_dva_claim(zio_t
*zio
)
3306 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3308 zio
->io_error
= error
;
3310 return (ZIO_PIPELINE_CONTINUE
);
3314 * Undo an allocation. This is used by zio_done() when an I/O fails
3315 * and we want to give back the block we just allocated.
3316 * This handles both normal blocks and gang blocks.
3319 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3321 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3322 ASSERT(zio
->io_bp_override
== NULL
);
3324 if (!BP_IS_HOLE(bp
))
3325 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3328 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3329 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3330 &gn
->gn_gbh
->zg_blkptr
[g
]);
3336 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3339 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3340 uint64_t size
, boolean_t
*slog
)
3343 zio_alloc_list_t io_alloc_list
;
3345 ASSERT(txg
> spa_syncing_txg(spa
));
3347 metaslab_trace_init(&io_alloc_list
);
3348 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3349 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
);
3353 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3354 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3355 &io_alloc_list
, NULL
);
3359 metaslab_trace_fini(&io_alloc_list
);
3362 BP_SET_LSIZE(new_bp
, size
);
3363 BP_SET_PSIZE(new_bp
, size
);
3364 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3365 BP_SET_CHECKSUM(new_bp
,
3366 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3367 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3368 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3369 BP_SET_LEVEL(new_bp
, 0);
3370 BP_SET_DEDUP(new_bp
, 0);
3371 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3374 * encrypted blocks will require an IV and salt. We generate
3375 * these now since we will not be rewriting the bp at
3378 if (os
->os_encrypted
) {
3379 uint8_t iv
[ZIO_DATA_IV_LEN
];
3380 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3382 BP_SET_CRYPT(new_bp
, B_TRUE
);
3383 VERIFY0(spa_crypt_get_salt(spa
,
3384 dmu_objset_id(os
), salt
));
3385 VERIFY0(zio_crypt_generate_iv(iv
));
3387 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3390 zfs_dbgmsg("%s: zil block allocation failure: "
3391 "size %llu, error %d", spa_name(spa
), size
, error
);
3398 * Free an intent log block.
3401 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
3403 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
3404 ASSERT(!BP_IS_GANG(bp
));
3406 zio_free(spa
, txg
, bp
);
3410 * ==========================================================================
3411 * Read and write to physical devices
3412 * ==========================================================================
3417 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3418 * stops after this stage and will resume upon I/O completion.
3419 * However, there are instances where the vdev layer may need to
3420 * continue the pipeline when an I/O was not issued. Since the I/O
3421 * that was sent to the vdev layer might be different than the one
3422 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3423 * force the underlying vdev layers to call either zio_execute() or
3424 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3427 zio_vdev_io_start(zio_t
*zio
)
3429 vdev_t
*vd
= zio
->io_vd
;
3431 spa_t
*spa
= zio
->io_spa
;
3435 ASSERT(zio
->io_error
== 0);
3436 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3439 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3440 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3443 * The mirror_ops handle multiple DVAs in a single BP.
3445 vdev_mirror_ops
.vdev_op_io_start(zio
);
3446 return (ZIO_PIPELINE_STOP
);
3449 ASSERT3P(zio
->io_logical
, !=, zio
);
3451 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3453 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3454 P2PHASE(zio
->io_size
, align
) != 0) {
3455 /* Transform logical writes to be a full physical block size. */
3456 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3457 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3458 ASSERT(vd
== vd
->vdev_top
);
3459 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3460 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3461 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3463 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3467 * If this is not a physical io, make sure that it is properly aligned
3468 * before proceeding.
3470 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3471 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3472 ASSERT0(P2PHASE(zio
->io_size
, align
));
3475 * For physical writes, we allow 512b aligned writes and assume
3476 * the device will perform a read-modify-write as necessary.
3478 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3479 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3482 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3485 * If this is a repair I/O, and there's no self-healing involved --
3486 * that is, we're just resilvering what we expect to resilver --
3487 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3488 * This prevents spurious resilvering with nested replication.
3489 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3490 * A is out of date, we'll read from C+D, then use the data to
3491 * resilver A+B -- but we don't actually want to resilver B, just A.
3492 * The top-level mirror has no way to know this, so instead we just
3493 * discard unnecessary repairs as we work our way down the vdev tree.
3494 * The same logic applies to any form of nested replication:
3495 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3497 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3498 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3499 zio
->io_txg
!= 0 && /* not a delegated i/o */
3500 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3501 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3502 zio_vdev_io_bypass(zio
);
3503 return (ZIO_PIPELINE_CONTINUE
);
3506 if (vd
->vdev_ops
->vdev_op_leaf
&&
3507 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3509 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3510 return (ZIO_PIPELINE_CONTINUE
);
3512 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3513 return (ZIO_PIPELINE_STOP
);
3515 if (!vdev_accessible(vd
, zio
)) {
3516 zio
->io_error
= SET_ERROR(ENXIO
);
3518 return (ZIO_PIPELINE_STOP
);
3520 zio
->io_delay
= gethrtime();
3523 vd
->vdev_ops
->vdev_op_io_start(zio
);
3524 return (ZIO_PIPELINE_STOP
);
3528 zio_vdev_io_done(zio_t
*zio
)
3530 vdev_t
*vd
= zio
->io_vd
;
3531 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3532 boolean_t unexpected_error
= B_FALSE
;
3534 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3535 return (ZIO_PIPELINE_STOP
);
3538 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3541 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3543 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3545 vdev_queue_io_done(zio
);
3547 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3548 vdev_cache_write(zio
);
3550 if (zio_injection_enabled
&& zio
->io_error
== 0)
3551 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3554 if (zio_injection_enabled
&& zio
->io_error
== 0)
3555 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3557 if (zio
->io_error
) {
3558 if (!vdev_accessible(vd
, zio
)) {
3559 zio
->io_error
= SET_ERROR(ENXIO
);
3561 unexpected_error
= B_TRUE
;
3566 ops
->vdev_op_io_done(zio
);
3568 if (unexpected_error
)
3569 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3571 return (ZIO_PIPELINE_CONTINUE
);
3575 * This function is used to change the priority of an existing zio that is
3576 * currently in-flight. This is used by the arc to upgrade priority in the
3577 * event that a demand read is made for a block that is currently queued
3578 * as a scrub or async read IO. Otherwise, the high priority read request
3579 * would end up having to wait for the lower priority IO.
3582 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3584 zio_t
*cio
, *cio_next
;
3585 zio_link_t
*zl
= NULL
;
3587 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3589 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3590 vdev_queue_change_io_priority(pio
, priority
);
3592 pio
->io_priority
= priority
;
3595 mutex_enter(&pio
->io_lock
);
3596 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3597 cio_next
= zio_walk_children(pio
, &zl
);
3598 zio_change_priority(cio
, priority
);
3600 mutex_exit(&pio
->io_lock
);
3604 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3605 * disk, and use that to finish the checksum ereport later.
3608 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3609 const abd_t
*good_buf
)
3611 /* no processing needed */
3612 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3617 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3619 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3621 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3623 zcr
->zcr_cbinfo
= zio
->io_size
;
3624 zcr
->zcr_cbdata
= abd
;
3625 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3626 zcr
->zcr_free
= zio_abd_free
;
3630 zio_vdev_io_assess(zio_t
*zio
)
3632 vdev_t
*vd
= zio
->io_vd
;
3634 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3635 return (ZIO_PIPELINE_STOP
);
3638 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3639 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3641 if (zio
->io_vsd
!= NULL
) {
3642 zio
->io_vsd_ops
->vsd_free(zio
);
3646 if (zio_injection_enabled
&& zio
->io_error
== 0)
3647 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3650 * If the I/O failed, determine whether we should attempt to retry it.
3652 * On retry, we cut in line in the issue queue, since we don't want
3653 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3655 if (zio
->io_error
&& vd
== NULL
&&
3656 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3657 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3658 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3660 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3661 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3662 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3663 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3664 zio_requeue_io_start_cut_in_line
);
3665 return (ZIO_PIPELINE_STOP
);
3669 * If we got an error on a leaf device, convert it to ENXIO
3670 * if the device is not accessible at all.
3672 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3673 !vdev_accessible(vd
, zio
))
3674 zio
->io_error
= SET_ERROR(ENXIO
);
3677 * If we can't write to an interior vdev (mirror or RAID-Z),
3678 * set vdev_cant_write so that we stop trying to allocate from it.
3680 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3681 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3682 vd
->vdev_cant_write
= B_TRUE
;
3686 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3687 * attempts will ever succeed. In this case we set a persistent bit so
3688 * that we don't bother with it in the future.
3690 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3691 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3692 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3693 vd
->vdev_nowritecache
= B_TRUE
;
3696 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3698 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3699 zio
->io_physdone
!= NULL
) {
3700 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3701 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3702 zio
->io_physdone(zio
->io_logical
);
3705 return (ZIO_PIPELINE_CONTINUE
);
3709 zio_vdev_io_reissue(zio_t
*zio
)
3711 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3712 ASSERT(zio
->io_error
== 0);
3714 zio
->io_stage
>>= 1;
3718 zio_vdev_io_redone(zio_t
*zio
)
3720 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3722 zio
->io_stage
>>= 1;
3726 zio_vdev_io_bypass(zio_t
*zio
)
3728 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3729 ASSERT(zio
->io_error
== 0);
3731 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3732 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3736 * ==========================================================================
3737 * Encrypt and store encryption parameters
3738 * ==========================================================================
3743 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3744 * managing the storage of encryption parameters and passing them to the
3745 * lower-level encryption functions.
3748 zio_encrypt(zio_t
*zio
)
3750 zio_prop_t
*zp
= &zio
->io_prop
;
3751 spa_t
*spa
= zio
->io_spa
;
3752 blkptr_t
*bp
= zio
->io_bp
;
3753 uint64_t psize
= BP_GET_PSIZE(bp
);
3754 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
3755 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3756 void *enc_buf
= NULL
;
3758 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3759 uint8_t iv
[ZIO_DATA_IV_LEN
];
3760 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3761 boolean_t no_crypt
= B_FALSE
;
3763 /* the root zio already encrypted the data */
3764 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3765 return (ZIO_PIPELINE_CONTINUE
);
3767 /* only ZIL blocks are re-encrypted on rewrite */
3768 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3769 return (ZIO_PIPELINE_CONTINUE
);
3771 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3772 BP_SET_CRYPT(bp
, B_FALSE
);
3773 return (ZIO_PIPELINE_CONTINUE
);
3776 /* if we are doing raw encryption set the provided encryption params */
3777 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3778 ASSERT0(BP_GET_LEVEL(bp
));
3779 BP_SET_CRYPT(bp
, B_TRUE
);
3780 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3781 if (ot
!= DMU_OT_OBJSET
)
3782 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3784 /* dnode blocks must be written out in the provided byteorder */
3785 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
3786 ot
== DMU_OT_DNODE
) {
3787 void *bswap_buf
= zio_buf_alloc(psize
);
3788 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
3790 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3791 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
3792 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
3795 abd_take_ownership_of_buf(babd
, B_TRUE
);
3796 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
3799 if (DMU_OT_IS_ENCRYPTED(ot
))
3800 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3801 return (ZIO_PIPELINE_CONTINUE
);
3804 /* indirect blocks only maintain a cksum of the lower level MACs */
3805 if (BP_GET_LEVEL(bp
) > 0) {
3806 BP_SET_CRYPT(bp
, B_TRUE
);
3807 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3808 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3810 zio_crypt_encode_mac_bp(bp
, mac
);
3811 return (ZIO_PIPELINE_CONTINUE
);
3815 * Objset blocks are a special case since they have 2 256-bit MACs
3816 * embedded within them.
3818 if (ot
== DMU_OT_OBJSET
) {
3819 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
3820 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3821 BP_SET_CRYPT(bp
, B_TRUE
);
3822 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
3823 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
3824 return (ZIO_PIPELINE_CONTINUE
);
3827 /* unencrypted object types are only authenticated with a MAC */
3828 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
3829 BP_SET_CRYPT(bp
, B_TRUE
);
3830 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
3831 zio
->io_abd
, psize
, mac
));
3832 zio_crypt_encode_mac_bp(bp
, mac
);
3833 return (ZIO_PIPELINE_CONTINUE
);
3837 * Later passes of sync-to-convergence may decide to rewrite data
3838 * in place to avoid more disk reallocations. This presents a problem
3839 * for encryption because this consitutes rewriting the new data with
3840 * the same encryption key and IV. However, this only applies to blocks
3841 * in the MOS (particularly the spacemaps) and we do not encrypt the
3842 * MOS. We assert that the zio is allocating or an intent log write
3845 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
3846 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
3847 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
3848 ASSERT3U(psize
, !=, 0);
3850 enc_buf
= zio_buf_alloc(psize
);
3851 eabd
= abd_get_from_buf(enc_buf
, psize
);
3852 abd_take_ownership_of_buf(eabd
, B_TRUE
);
3855 * For an explanation of what encryption parameters are stored
3856 * where, see the block comment in zio_crypt.c.
3858 if (ot
== DMU_OT_INTENT_LOG
) {
3859 zio_crypt_decode_params_bp(bp
, salt
, iv
);
3861 BP_SET_CRYPT(bp
, B_TRUE
);
3864 /* Perform the encryption. This should not fail */
3865 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, dsobj
, bp
, zio
->io_txg
,
3866 psize
, zio
->io_abd
, eabd
, iv
, mac
, salt
, &no_crypt
));
3868 /* encode encryption metadata into the bp */
3869 if (ot
== DMU_OT_INTENT_LOG
) {
3871 * ZIL blocks store the MAC in the embedded checksum, so the
3872 * transform must always be applied.
3874 zio_crypt_encode_mac_zil(enc_buf
, mac
);
3875 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3877 BP_SET_CRYPT(bp
, B_TRUE
);
3878 zio_crypt_encode_params_bp(bp
, salt
, iv
);
3879 zio_crypt_encode_mac_bp(bp
, mac
);
3882 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
3885 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3889 return (ZIO_PIPELINE_CONTINUE
);
3893 * ==========================================================================
3894 * Generate and verify checksums
3895 * ==========================================================================
3898 zio_checksum_generate(zio_t
*zio
)
3900 blkptr_t
*bp
= zio
->io_bp
;
3901 enum zio_checksum checksum
;
3905 * This is zio_write_phys().
3906 * We're either generating a label checksum, or none at all.
3908 checksum
= zio
->io_prop
.zp_checksum
;
3910 if (checksum
== ZIO_CHECKSUM_OFF
)
3911 return (ZIO_PIPELINE_CONTINUE
);
3913 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3915 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3916 ASSERT(!IO_IS_ALLOCATING(zio
));
3917 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3919 checksum
= BP_GET_CHECKSUM(bp
);
3923 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3925 return (ZIO_PIPELINE_CONTINUE
);
3929 zio_checksum_verify(zio_t
*zio
)
3931 zio_bad_cksum_t info
;
3932 blkptr_t
*bp
= zio
->io_bp
;
3935 ASSERT(zio
->io_vd
!= NULL
);
3939 * This is zio_read_phys().
3940 * We're either verifying a label checksum, or nothing at all.
3942 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3943 return (ZIO_PIPELINE_CONTINUE
);
3945 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3948 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3949 zio
->io_error
= error
;
3950 if (error
== ECKSUM
&&
3951 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3952 zfs_ereport_start_checksum(zio
->io_spa
,
3953 zio
->io_vd
, &zio
->io_bookmark
, zio
,
3954 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
3958 return (ZIO_PIPELINE_CONTINUE
);
3962 * Called by RAID-Z to ensure we don't compute the checksum twice.
3965 zio_checksum_verified(zio_t
*zio
)
3967 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3971 * ==========================================================================
3972 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3973 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3974 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3975 * indicate errors that are specific to one I/O, and most likely permanent.
3976 * Any other error is presumed to be worse because we weren't expecting it.
3977 * ==========================================================================
3980 zio_worst_error(int e1
, int e2
)
3982 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3985 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3986 if (e1
== zio_error_rank
[r1
])
3989 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3990 if (e2
== zio_error_rank
[r2
])
3993 return (r1
> r2
? e1
: e2
);
3997 * ==========================================================================
3999 * ==========================================================================
4002 zio_ready(zio_t
*zio
)
4004 blkptr_t
*bp
= zio
->io_bp
;
4005 zio_t
*pio
, *pio_next
;
4006 zio_link_t
*zl
= NULL
;
4008 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4010 return (ZIO_PIPELINE_STOP
);
4013 if (zio
->io_ready
) {
4014 ASSERT(IO_IS_ALLOCATING(zio
));
4015 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4016 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4017 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4022 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4023 zio
->io_bp_copy
= *bp
;
4025 if (zio
->io_error
!= 0) {
4026 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4028 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4029 ASSERT(IO_IS_ALLOCATING(zio
));
4030 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4032 * We were unable to allocate anything, unreserve and
4033 * issue the next I/O to allocate.
4035 metaslab_class_throttle_unreserve(
4036 spa_normal_class(zio
->io_spa
),
4037 zio
->io_prop
.zp_copies
, zio
);
4038 zio_allocate_dispatch(zio
->io_spa
);
4042 mutex_enter(&zio
->io_lock
);
4043 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4044 pio
= zio_walk_parents(zio
, &zl
);
4045 mutex_exit(&zio
->io_lock
);
4048 * As we notify zio's parents, new parents could be added.
4049 * New parents go to the head of zio's io_parent_list, however,
4050 * so we will (correctly) not notify them. The remainder of zio's
4051 * io_parent_list, from 'pio_next' onward, cannot change because
4052 * all parents must wait for us to be done before they can be done.
4054 for (; pio
!= NULL
; pio
= pio_next
) {
4055 pio_next
= zio_walk_parents(zio
, &zl
);
4056 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
4059 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4060 if (BP_IS_GANG(bp
)) {
4061 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4063 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4064 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4068 if (zio_injection_enabled
&&
4069 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4070 zio_handle_ignored_writes(zio
);
4072 return (ZIO_PIPELINE_CONTINUE
);
4076 * Update the allocation throttle accounting.
4079 zio_dva_throttle_done(zio_t
*zio
)
4081 ASSERTV(zio_t
*lio
= zio
->io_logical
);
4082 zio_t
*pio
= zio_unique_parent(zio
);
4083 vdev_t
*vd
= zio
->io_vd
;
4084 int flags
= METASLAB_ASYNC_ALLOC
;
4086 ASSERT3P(zio
->io_bp
, !=, NULL
);
4087 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4088 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4089 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4091 ASSERT3P(vd
, ==, vd
->vdev_top
);
4092 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4093 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4094 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4095 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4096 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4099 * Parents of gang children can have two flavors -- ones that
4100 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4101 * and ones that allocated the constituent blocks. The allocation
4102 * throttle needs to know the allocating parent zio so we must find
4105 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4107 * If our parent is a rewrite gang child then our grandparent
4108 * would have been the one that performed the allocation.
4110 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4111 pio
= zio_unique_parent(pio
);
4112 flags
|= METASLAB_GANG_CHILD
;
4115 ASSERT(IO_IS_ALLOCATING(pio
));
4116 ASSERT3P(zio
, !=, zio
->io_logical
);
4117 ASSERT(zio
->io_logical
!= NULL
);
4118 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4119 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4121 mutex_enter(&pio
->io_lock
);
4122 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
4123 mutex_exit(&pio
->io_lock
);
4125 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
4129 * Call into the pipeline to see if there is more work that
4130 * needs to be done. If there is work to be done it will be
4131 * dispatched to another taskq thread.
4133 zio_allocate_dispatch(zio
->io_spa
);
4137 zio_done(zio_t
*zio
)
4140 * Always attempt to keep stack usage minimal here since
4141 * we can be called recurisvely up to 19 levels deep.
4143 const uint64_t psize
= zio
->io_size
;
4144 zio_t
*pio
, *pio_next
;
4145 zio_link_t
*zl
= NULL
;
4148 * If our children haven't all completed,
4149 * wait for them and then repeat this pipeline stage.
4151 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4152 return (ZIO_PIPELINE_STOP
);
4156 * If the allocation throttle is enabled, then update the accounting.
4157 * We only track child I/Os that are part of an allocating async
4158 * write. We must do this since the allocation is performed
4159 * by the logical I/O but the actual write is done by child I/Os.
4161 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4162 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4163 ASSERT(spa_normal_class(
4164 zio
->io_spa
)->mc_alloc_throttle_enabled
);
4165 zio_dva_throttle_done(zio
);
4169 * If the allocation throttle is enabled, verify that
4170 * we have decremented the refcounts for every I/O that was throttled.
4172 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4173 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4174 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4175 ASSERT(zio
->io_bp
!= NULL
);
4176 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
);
4177 VERIFY(refcount_not_held(
4178 &(spa_normal_class(zio
->io_spa
)->mc_alloc_slots
), zio
));
4182 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4183 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4184 ASSERT(zio
->io_children
[c
][w
] == 0);
4186 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4187 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4188 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4189 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4190 sizeof (blkptr_t
)) == 0 ||
4191 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4192 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4193 zio
->io_bp_override
== NULL
&&
4194 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4195 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4196 BP_GET_NDVAS(zio
->io_bp
));
4197 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4198 (BP_COUNT_GANG(zio
->io_bp
) ==
4199 BP_GET_NDVAS(zio
->io_bp
)));
4201 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4202 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4206 * If there were child vdev/gang/ddt errors, they apply to us now.
4208 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4209 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4210 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4213 * If the I/O on the transformed data was successful, generate any
4214 * checksum reports now while we still have the transformed data.
4216 if (zio
->io_error
== 0) {
4217 while (zio
->io_cksum_report
!= NULL
) {
4218 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4219 uint64_t align
= zcr
->zcr_align
;
4220 uint64_t asize
= P2ROUNDUP(psize
, align
);
4221 abd_t
*adata
= zio
->io_abd
;
4223 if (asize
!= psize
) {
4224 adata
= abd_alloc(asize
, B_TRUE
);
4225 abd_copy(adata
, zio
->io_abd
, psize
);
4226 abd_zero_off(adata
, psize
, asize
- psize
);
4229 zio
->io_cksum_report
= zcr
->zcr_next
;
4230 zcr
->zcr_next
= NULL
;
4231 zcr
->zcr_finish(zcr
, adata
);
4232 zfs_ereport_free_checksum(zcr
);
4239 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4241 vdev_stat_update(zio
, psize
);
4244 * If this I/O is attached to a particular vdev is slow, exceeding
4245 * 30 seconds to complete, post an error described the I/O delay.
4246 * We ignore these errors if the device is currently unavailable.
4248 if (zio
->io_delay
>= MSEC2NSEC(zio_delay_max
)) {
4249 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
4250 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
4251 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4254 if (zio
->io_error
) {
4256 * If this I/O is attached to a particular vdev,
4257 * generate an error message describing the I/O failure
4258 * at the block level. We ignore these errors if the
4259 * device is currently unavailable.
4261 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4262 !vdev_is_dead(zio
->io_vd
))
4263 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4264 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4266 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4267 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4268 zio
== zio
->io_logical
) {
4270 * For logical I/O requests, tell the SPA to log the
4271 * error and generate a logical data ereport.
4273 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4274 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4275 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4279 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4281 * Determine whether zio should be reexecuted. This will
4282 * propagate all the way to the root via zio_notify_parent().
4284 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4285 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4287 if (IO_IS_ALLOCATING(zio
) &&
4288 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4289 if (zio
->io_error
!= ENOSPC
)
4290 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4292 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4295 if ((zio
->io_type
== ZIO_TYPE_READ
||
4296 zio
->io_type
== ZIO_TYPE_FREE
) &&
4297 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4298 zio
->io_error
== ENXIO
&&
4299 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4300 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4301 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4303 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4304 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4307 * Here is a possibly good place to attempt to do
4308 * either combinatorial reconstruction or error correction
4309 * based on checksums. It also might be a good place
4310 * to send out preliminary ereports before we suspend
4316 * If there were logical child errors, they apply to us now.
4317 * We defer this until now to avoid conflating logical child
4318 * errors with errors that happened to the zio itself when
4319 * updating vdev stats and reporting FMA events above.
4321 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4323 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4324 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4325 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4326 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4328 zio_gang_tree_free(&zio
->io_gang_tree
);
4331 * Godfather I/Os should never suspend.
4333 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4334 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4335 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4337 if (zio
->io_reexecute
) {
4339 * This is a logical I/O that wants to reexecute.
4341 * Reexecute is top-down. When an i/o fails, if it's not
4342 * the root, it simply notifies its parent and sticks around.
4343 * The parent, seeing that it still has children in zio_done(),
4344 * does the same. This percolates all the way up to the root.
4345 * The root i/o will reexecute or suspend the entire tree.
4347 * This approach ensures that zio_reexecute() honors
4348 * all the original i/o dependency relationships, e.g.
4349 * parents not executing until children are ready.
4351 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4353 zio
->io_gang_leader
= NULL
;
4355 mutex_enter(&zio
->io_lock
);
4356 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4357 mutex_exit(&zio
->io_lock
);
4360 * "The Godfather" I/O monitors its children but is
4361 * not a true parent to them. It will track them through
4362 * the pipeline but severs its ties whenever they get into
4363 * trouble (e.g. suspended). This allows "The Godfather"
4364 * I/O to return status without blocking.
4367 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4369 zio_link_t
*remove_zl
= zl
;
4370 pio_next
= zio_walk_parents(zio
, &zl
);
4372 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4373 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4374 zio_remove_child(pio
, zio
, remove_zl
);
4375 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4379 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4381 * We're not a root i/o, so there's nothing to do
4382 * but notify our parent. Don't propagate errors
4383 * upward since we haven't permanently failed yet.
4385 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4386 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4387 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4388 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4390 * We'd fail again if we reexecuted now, so suspend
4391 * until conditions improve (e.g. device comes online).
4393 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4396 * Reexecution is potentially a huge amount of work.
4397 * Hand it off to the otherwise-unused claim taskq.
4399 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4400 spa_taskq_dispatch_ent(zio
->io_spa
,
4401 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4402 (task_func_t
*)zio_reexecute
, zio
, 0,
4405 return (ZIO_PIPELINE_STOP
);
4408 ASSERT(zio
->io_child_count
== 0);
4409 ASSERT(zio
->io_reexecute
== 0);
4410 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4413 * Report any checksum errors, since the I/O is complete.
4415 while (zio
->io_cksum_report
!= NULL
) {
4416 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4417 zio
->io_cksum_report
= zcr
->zcr_next
;
4418 zcr
->zcr_next
= NULL
;
4419 zcr
->zcr_finish(zcr
, NULL
);
4420 zfs_ereport_free_checksum(zcr
);
4423 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4424 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4425 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4426 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4430 * It is the responsibility of the done callback to ensure that this
4431 * particular zio is no longer discoverable for adoption, and as
4432 * such, cannot acquire any new parents.
4437 mutex_enter(&zio
->io_lock
);
4438 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4439 mutex_exit(&zio
->io_lock
);
4442 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4443 zio_link_t
*remove_zl
= zl
;
4444 pio_next
= zio_walk_parents(zio
, &zl
);
4445 zio_remove_child(pio
, zio
, remove_zl
);
4446 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4449 if (zio
->io_waiter
!= NULL
) {
4450 mutex_enter(&zio
->io_lock
);
4451 zio
->io_executor
= NULL
;
4452 cv_broadcast(&zio
->io_cv
);
4453 mutex_exit(&zio
->io_lock
);
4458 return (ZIO_PIPELINE_STOP
);
4462 * ==========================================================================
4463 * I/O pipeline definition
4464 * ==========================================================================
4466 static zio_pipe_stage_t
*zio_pipeline
[] = {
4474 zio_checksum_generate
,
4490 zio_checksum_verify
,
4498 * Compare two zbookmark_phys_t's to see which we would reach first in a
4499 * pre-order traversal of the object tree.
4501 * This is simple in every case aside from the meta-dnode object. For all other
4502 * objects, we traverse them in order (object 1 before object 2, and so on).
4503 * However, all of these objects are traversed while traversing object 0, since
4504 * the data it points to is the list of objects. Thus, we need to convert to a
4505 * canonical representation so we can compare meta-dnode bookmarks to
4506 * non-meta-dnode bookmarks.
4508 * We do this by calculating "equivalents" for each field of the zbookmark.
4509 * zbookmarks outside of the meta-dnode use their own object and level, and
4510 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4511 * blocks this bookmark refers to) by multiplying their blkid by their span
4512 * (the number of L0 blocks contained within one block at their level).
4513 * zbookmarks inside the meta-dnode calculate their object equivalent
4514 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4515 * level + 1<<31 (any value larger than a level could ever be) for their level.
4516 * This causes them to always compare before a bookmark in their object
4517 * equivalent, compare appropriately to bookmarks in other objects, and to
4518 * compare appropriately to other bookmarks in the meta-dnode.
4521 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4522 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4525 * These variables represent the "equivalent" values for the zbookmark,
4526 * after converting zbookmarks inside the meta dnode to their
4527 * normal-object equivalents.
4529 uint64_t zb1obj
, zb2obj
;
4530 uint64_t zb1L0
, zb2L0
;
4531 uint64_t zb1level
, zb2level
;
4533 if (zb1
->zb_object
== zb2
->zb_object
&&
4534 zb1
->zb_level
== zb2
->zb_level
&&
4535 zb1
->zb_blkid
== zb2
->zb_blkid
)
4539 * BP_SPANB calculates the span in blocks.
4541 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4542 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4544 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4545 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4547 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4549 zb1obj
= zb1
->zb_object
;
4550 zb1level
= zb1
->zb_level
;
4553 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4554 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4556 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4558 zb2obj
= zb2
->zb_object
;
4559 zb2level
= zb2
->zb_level
;
4562 /* Now that we have a canonical representation, do the comparison. */
4563 if (zb1obj
!= zb2obj
)
4564 return (zb1obj
< zb2obj
? -1 : 1);
4565 else if (zb1L0
!= zb2L0
)
4566 return (zb1L0
< zb2L0
? -1 : 1);
4567 else if (zb1level
!= zb2level
)
4568 return (zb1level
> zb2level
? -1 : 1);
4570 * This can (theoretically) happen if the bookmarks have the same object
4571 * and level, but different blkids, if the block sizes are not the same.
4572 * There is presently no way to change the indirect block sizes
4578 * This function checks the following: given that last_block is the place that
4579 * our traversal stopped last time, does that guarantee that we've visited
4580 * every node under subtree_root? Therefore, we can't just use the raw output
4581 * of zbookmark_compare. We have to pass in a modified version of
4582 * subtree_root; by incrementing the block id, and then checking whether
4583 * last_block is before or equal to that, we can tell whether or not having
4584 * visited last_block implies that all of subtree_root's children have been
4588 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4589 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4591 zbookmark_phys_t mod_zb
= *subtree_root
;
4593 ASSERT(last_block
->zb_level
== 0);
4595 /* The objset_phys_t isn't before anything. */
4600 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4601 * data block size in sectors, because that variable is only used if
4602 * the bookmark refers to a block in the meta-dnode. Since we don't
4603 * know without examining it what object it refers to, and there's no
4604 * harm in passing in this value in other cases, we always pass it in.
4606 * We pass in 0 for the indirect block size shift because zb2 must be
4607 * level 0. The indirect block size is only used to calculate the span
4608 * of the bookmark, but since the bookmark must be level 0, the span is
4609 * always 1, so the math works out.
4611 * If you make changes to how the zbookmark_compare code works, be sure
4612 * to make sure that this code still works afterwards.
4614 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4615 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4619 #if defined(_KERNEL) && defined(HAVE_SPL)
4620 EXPORT_SYMBOL(zio_type_name
);
4621 EXPORT_SYMBOL(zio_buf_alloc
);
4622 EXPORT_SYMBOL(zio_data_buf_alloc
);
4623 EXPORT_SYMBOL(zio_buf_free
);
4624 EXPORT_SYMBOL(zio_data_buf_free
);
4626 module_param(zio_delay_max
, int, 0644);
4627 MODULE_PARM_DESC(zio_delay_max
, "Max zio millisec delay before posting event");
4629 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
4630 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");
4632 module_param(zfs_sync_pass_deferred_free
, int, 0644);
4633 MODULE_PARM_DESC(zfs_sync_pass_deferred_free
,
4634 "Defer frees starting in this pass");
4636 module_param(zfs_sync_pass_dont_compress
, int, 0644);
4637 MODULE_PARM_DESC(zfs_sync_pass_dont_compress
,
4638 "Don't compress starting in this pass");
4640 module_param(zfs_sync_pass_rewrite
, int, 0644);
4641 MODULE_PARM_DESC(zfs_sync_pass_rewrite
,
4642 "Rewrite new bps starting in this pass");
4644 module_param(zio_dva_throttle_enabled
, int, 0644);
4645 MODULE_PARM_DESC(zio_dva_throttle_enabled
,
4646 "Throttle block allocations in the ZIO pipeline");