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/metaslab_impl.h>
44 #include <sys/trace_zio.h>
46 #include <sys/dsl_crypt.h>
49 * ==========================================================================
50 * I/O type descriptions
51 * ==========================================================================
53 const char *zio_type_name
[ZIO_TYPES
] = {
55 * Note: Linux kernel thread name length is limited
56 * so these names will differ from upstream open zfs.
58 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
61 int zio_dva_throttle_enabled
= B_TRUE
;
64 * ==========================================================================
66 * ==========================================================================
68 kmem_cache_t
*zio_cache
;
69 kmem_cache_t
*zio_link_cache
;
70 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
71 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
72 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
73 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
74 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
77 int zio_delay_max
= ZIO_DELAY_MAX
;
79 #define ZIO_PIPELINE_CONTINUE 0x100
80 #define ZIO_PIPELINE_STOP 0x101
82 #define BP_SPANB(indblkshift, level) \
83 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
84 #define COMPARE_META_LEVEL 0x80000000ul
86 * The following actions directly effect the spa's sync-to-convergence logic.
87 * The values below define the sync pass when we start performing the action.
88 * Care should be taken when changing these values as they directly impact
89 * spa_sync() performance. Tuning these values may introduce subtle performance
90 * pathologies and should only be done in the context of performance analysis.
91 * These tunables will eventually be removed and replaced with #defines once
92 * enough analysis has been done to determine optimal values.
94 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
95 * regular blocks are not deferred.
97 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
98 int zfs_sync_pass_dont_compress
= 5; /* don't compress starting in this pass */
99 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
102 * An allocating zio is one that either currently has the DVA allocate
103 * stage set or will have it later in its lifetime.
105 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
107 int zio_requeue_io_start_cut_in_line
= 1;
110 int zio_buf_debug_limit
= 16384;
112 int zio_buf_debug_limit
= 0;
115 static inline void __zio_execute(zio_t
*zio
);
117 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
123 vmem_t
*data_alloc_arena
= NULL
;
125 zio_cache
= kmem_cache_create("zio_cache",
126 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
127 zio_link_cache
= kmem_cache_create("zio_link_cache",
128 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
131 * For small buffers, we want a cache for each multiple of
132 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
133 * for each quarter-power of 2.
135 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
136 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
139 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
141 #if defined(_ILP32) && defined(_KERNEL)
143 * Cache size limited to 1M on 32-bit platforms until ARC
144 * buffers no longer require virtual address space.
146 if (size
> zfs_max_recordsize
)
155 * If we are using watchpoints, put each buffer on its own page,
156 * to eliminate the performance overhead of trapping to the
157 * kernel when modifying a non-watched buffer that shares the
158 * page with a watched buffer.
160 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
163 * Here's the problem - on 4K native devices in userland on
164 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
165 * will fail with EINVAL, causing zdb (and others) to coredump.
166 * Since userland probably doesn't need optimized buffer caches,
167 * we just force 4K alignment on everything.
169 align
= 8 * SPA_MINBLOCKSIZE
;
171 if (size
< PAGESIZE
) {
172 align
= SPA_MINBLOCKSIZE
;
173 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
180 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
181 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
182 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
184 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
185 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
186 align
, NULL
, NULL
, NULL
, NULL
,
187 data_alloc_arena
, cflags
);
192 ASSERT(zio_buf_cache
[c
] != NULL
);
193 if (zio_buf_cache
[c
- 1] == NULL
)
194 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
196 ASSERT(zio_data_buf_cache
[c
] != NULL
);
197 if (zio_data_buf_cache
[c
- 1] == NULL
)
198 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
210 kmem_cache_t
*last_cache
= NULL
;
211 kmem_cache_t
*last_data_cache
= NULL
;
213 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
216 * Cache size limited to 1M on 32-bit platforms until ARC
217 * buffers no longer require virtual address space.
219 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
222 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
223 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
224 (void) printf("zio_fini: [%d] %llu != %llu\n",
225 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
226 (long long unsigned)zio_buf_cache_allocs
[c
],
227 (long long unsigned)zio_buf_cache_frees
[c
]);
229 if (zio_buf_cache
[c
] != last_cache
) {
230 last_cache
= zio_buf_cache
[c
];
231 kmem_cache_destroy(zio_buf_cache
[c
]);
233 zio_buf_cache
[c
] = NULL
;
235 if (zio_data_buf_cache
[c
] != last_data_cache
) {
236 last_data_cache
= zio_data_buf_cache
[c
];
237 kmem_cache_destroy(zio_data_buf_cache
[c
]);
239 zio_data_buf_cache
[c
] = NULL
;
242 kmem_cache_destroy(zio_link_cache
);
243 kmem_cache_destroy(zio_cache
);
251 * ==========================================================================
252 * Allocate and free I/O buffers
253 * ==========================================================================
257 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
258 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
259 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
260 * excess / transient data in-core during a crashdump.
263 zio_buf_alloc(size_t size
)
265 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
267 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
268 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
269 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
272 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
276 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
277 * crashdump if the kernel panics. This exists so that we will limit the amount
278 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
279 * of kernel heap dumped to disk when the kernel panics)
282 zio_data_buf_alloc(size_t size
)
284 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
286 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
288 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
292 zio_buf_free(void *buf
, size_t size
)
294 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
296 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
297 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
298 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
301 kmem_cache_free(zio_buf_cache
[c
], buf
);
305 zio_data_buf_free(void *buf
, size_t size
)
307 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
309 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
311 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
315 zio_abd_free(void *abd
, size_t size
)
317 abd_free((abd_t
*)abd
);
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
326 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
327 zio_transform_func_t
*transform
)
329 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
332 * Ensure that anyone expecting this zio to contain a linear ABD isn't
333 * going to get a nasty surprise when they try to access the data.
335 IMPLY(abd_is_linear(zio
->io_abd
), abd_is_linear(data
));
337 zt
->zt_orig_abd
= zio
->io_abd
;
338 zt
->zt_orig_size
= zio
->io_size
;
339 zt
->zt_bufsize
= bufsize
;
340 zt
->zt_transform
= transform
;
342 zt
->zt_next
= zio
->io_transform_stack
;
343 zio
->io_transform_stack
= zt
;
350 zio_pop_transforms(zio_t
*zio
)
354 while ((zt
= zio
->io_transform_stack
) != NULL
) {
355 if (zt
->zt_transform
!= NULL
)
356 zt
->zt_transform(zio
,
357 zt
->zt_orig_abd
, zt
->zt_orig_size
);
359 if (zt
->zt_bufsize
!= 0)
360 abd_free(zio
->io_abd
);
362 zio
->io_abd
= zt
->zt_orig_abd
;
363 zio
->io_size
= zt
->zt_orig_size
;
364 zio
->io_transform_stack
= zt
->zt_next
;
366 kmem_free(zt
, sizeof (zio_transform_t
));
371 * ==========================================================================
372 * I/O transform callbacks for subblocks, decompression, and decryption
373 * ==========================================================================
376 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
378 ASSERT(zio
->io_size
> size
);
380 if (zio
->io_type
== ZIO_TYPE_READ
)
381 abd_copy(data
, zio
->io_abd
, size
);
385 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
387 if (zio
->io_error
== 0) {
388 void *tmp
= abd_borrow_buf(data
, size
);
389 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
390 zio
->io_abd
, tmp
, zio
->io_size
, size
);
391 abd_return_buf_copy(data
, tmp
, size
);
394 zio
->io_error
= SET_ERROR(EIO
);
399 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
403 blkptr_t
*bp
= zio
->io_bp
;
404 uint64_t lsize
= BP_GET_LSIZE(bp
);
405 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
406 uint8_t salt
[ZIO_DATA_SALT_LEN
];
407 uint8_t iv
[ZIO_DATA_IV_LEN
];
408 uint8_t mac
[ZIO_DATA_MAC_LEN
];
409 boolean_t no_crypt
= B_FALSE
;
411 ASSERT(BP_USES_CRYPT(bp
));
412 ASSERT3U(size
, !=, 0);
414 if (zio
->io_error
!= 0)
418 * Verify the cksum of MACs stored in an indirect bp. It will always
419 * be possible to verify this since it does not require an encryption
422 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
423 zio_crypt_decode_mac_bp(bp
, mac
);
425 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
427 * We haven't decompressed the data yet, but
428 * zio_crypt_do_indirect_mac_checksum() requires
429 * decompressed data to be able to parse out the MACs
430 * from the indirect block. We decompress it now and
431 * throw away the result after we are finished.
433 tmp
= zio_buf_alloc(lsize
);
434 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
435 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
437 ret
= SET_ERROR(EIO
);
440 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
441 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
442 zio_buf_free(tmp
, lsize
);
444 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
445 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
447 abd_copy(data
, zio
->io_abd
, size
);
456 * If this is an authenticated block, just check the MAC. It would be
457 * nice to separate this out into its own flag, but for the moment
458 * enum zio_flag is out of bits.
460 if (BP_IS_AUTHENTICATED(bp
)) {
461 if (ot
== DMU_OT_OBJSET
) {
462 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, zio
->io_spa
,
463 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, size
,
464 BP_SHOULD_BYTESWAP(bp
));
466 zio_crypt_decode_mac_bp(bp
, mac
);
467 ret
= spa_do_crypt_mac_abd(B_FALSE
, zio
->io_spa
,
468 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, size
, mac
);
470 abd_copy(data
, zio
->io_abd
, size
);
478 zio_crypt_decode_params_bp(bp
, salt
, iv
);
480 if (ot
== DMU_OT_INTENT_LOG
) {
481 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
482 zio_crypt_decode_mac_zil(tmp
, mac
);
483 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
485 zio_crypt_decode_mac_bp(bp
, mac
);
488 ret
= spa_do_crypt_abd(B_FALSE
, zio
->io_spa
, zio
->io_bookmark
.zb_objset
,
489 bp
, bp
->blk_birth
, size
, data
, zio
->io_abd
, iv
, mac
, salt
,
492 abd_copy(data
, zio
->io_abd
, size
);
500 /* assert that the key was found unless this was speculative */
501 ASSERT(ret
!= ENOENT
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
504 * If there was a decryption / authentication error return EIO as
505 * the io_error. If this was not a speculative zio, create an ereport.
508 ret
= SET_ERROR(EIO
);
509 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
510 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
511 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
519 * ==========================================================================
520 * I/O parent/child relationships and pipeline interlocks
521 * ==========================================================================
524 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
526 list_t
*pl
= &cio
->io_parent_list
;
528 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
532 ASSERT((*zl
)->zl_child
== cio
);
533 return ((*zl
)->zl_parent
);
537 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
539 list_t
*cl
= &pio
->io_child_list
;
541 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
545 ASSERT((*zl
)->zl_parent
== pio
);
546 return ((*zl
)->zl_child
);
550 zio_unique_parent(zio_t
*cio
)
552 zio_link_t
*zl
= NULL
;
553 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
555 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
560 zio_add_child(zio_t
*pio
, zio_t
*cio
)
562 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
565 * Logical I/Os can have logical, gang, or vdev children.
566 * Gang I/Os can have gang or vdev children.
567 * Vdev I/Os can only have vdev children.
568 * The following ASSERT captures all of these constraints.
570 ASSERT(cio
->io_child_type
<= pio
->io_child_type
);
575 mutex_enter(&cio
->io_lock
);
576 mutex_enter(&pio
->io_lock
);
578 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
580 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
581 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
583 list_insert_head(&pio
->io_child_list
, zl
);
584 list_insert_head(&cio
->io_parent_list
, zl
);
586 pio
->io_child_count
++;
587 cio
->io_parent_count
++;
589 mutex_exit(&pio
->io_lock
);
590 mutex_exit(&cio
->io_lock
);
594 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
596 ASSERT(zl
->zl_parent
== pio
);
597 ASSERT(zl
->zl_child
== cio
);
599 mutex_enter(&cio
->io_lock
);
600 mutex_enter(&pio
->io_lock
);
602 list_remove(&pio
->io_child_list
, zl
);
603 list_remove(&cio
->io_parent_list
, zl
);
605 pio
->io_child_count
--;
606 cio
->io_parent_count
--;
608 mutex_exit(&pio
->io_lock
);
609 mutex_exit(&cio
->io_lock
);
610 kmem_cache_free(zio_link_cache
, zl
);
614 zio_wait_for_children(zio_t
*zio
, enum zio_child child
, enum zio_wait_type wait
)
616 uint64_t *countp
= &zio
->io_children
[child
][wait
];
617 boolean_t waiting
= B_FALSE
;
619 mutex_enter(&zio
->io_lock
);
620 ASSERT(zio
->io_stall
== NULL
);
623 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
624 zio
->io_stall
= countp
;
627 mutex_exit(&zio
->io_lock
);
632 __attribute__((always_inline
))
634 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
)
636 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
637 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
639 mutex_enter(&pio
->io_lock
);
640 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
641 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
642 pio
->io_reexecute
|= zio
->io_reexecute
;
643 ASSERT3U(*countp
, >, 0);
647 if (*countp
== 0 && pio
->io_stall
== countp
) {
648 zio_taskq_type_t type
=
649 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
651 pio
->io_stall
= NULL
;
652 mutex_exit(&pio
->io_lock
);
654 * Dispatch the parent zio in its own taskq so that
655 * the child can continue to make progress. This also
656 * prevents overflowing the stack when we have deeply nested
657 * parent-child relationships.
659 zio_taskq_dispatch(pio
, type
, B_FALSE
);
661 mutex_exit(&pio
->io_lock
);
666 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
668 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
669 zio
->io_error
= zio
->io_child_error
[c
];
673 zio_bookmark_compare(const void *x1
, const void *x2
)
675 const zio_t
*z1
= x1
;
676 const zio_t
*z2
= x2
;
678 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
680 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
683 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
685 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
688 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
690 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
693 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
695 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
707 * ==========================================================================
708 * Create the various types of I/O (read, write, free, etc)
709 * ==========================================================================
712 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
713 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
714 void *private, zio_type_t type
, zio_priority_t priority
,
715 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
716 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
717 enum zio_stage pipeline
)
721 ASSERT3U(psize
, <=, SPA_MAXBLOCKSIZE
);
722 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
723 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
725 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
726 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
727 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
729 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
731 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
732 bzero(zio
, sizeof (zio_t
));
734 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
735 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
737 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
738 offsetof(zio_link_t
, zl_parent_node
));
739 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
740 offsetof(zio_link_t
, zl_child_node
));
741 metaslab_trace_init(&zio
->io_alloc_list
);
744 zio
->io_child_type
= ZIO_CHILD_VDEV
;
745 else if (flags
& ZIO_FLAG_GANG_CHILD
)
746 zio
->io_child_type
= ZIO_CHILD_GANG
;
747 else if (flags
& ZIO_FLAG_DDT_CHILD
)
748 zio
->io_child_type
= ZIO_CHILD_DDT
;
750 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
753 zio
->io_bp
= (blkptr_t
*)bp
;
754 zio
->io_bp_copy
= *bp
;
755 zio
->io_bp_orig
= *bp
;
756 if (type
!= ZIO_TYPE_WRITE
||
757 zio
->io_child_type
== ZIO_CHILD_DDT
)
758 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
759 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
760 zio
->io_logical
= zio
;
761 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
762 pipeline
|= ZIO_GANG_STAGES
;
768 zio
->io_private
= private;
770 zio
->io_priority
= priority
;
772 zio
->io_offset
= offset
;
773 zio
->io_orig_abd
= zio
->io_abd
= data
;
774 zio
->io_orig_size
= zio
->io_size
= psize
;
775 zio
->io_lsize
= lsize
;
776 zio
->io_orig_flags
= zio
->io_flags
= flags
;
777 zio
->io_orig_stage
= zio
->io_stage
= stage
;
778 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
779 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
781 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
782 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
785 zio
->io_bookmark
= *zb
;
788 if (zio
->io_logical
== NULL
)
789 zio
->io_logical
= pio
->io_logical
;
790 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
791 zio
->io_gang_leader
= pio
->io_gang_leader
;
792 zio_add_child(pio
, zio
);
795 taskq_init_ent(&zio
->io_tqent
);
801 zio_destroy(zio_t
*zio
)
803 metaslab_trace_fini(&zio
->io_alloc_list
);
804 list_destroy(&zio
->io_parent_list
);
805 list_destroy(&zio
->io_child_list
);
806 mutex_destroy(&zio
->io_lock
);
807 cv_destroy(&zio
->io_cv
);
808 kmem_cache_free(zio_cache
, zio
);
812 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
813 void *private, enum zio_flag flags
)
817 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
818 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
819 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
825 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
827 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
831 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
)
833 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
834 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
835 bp
, (longlong_t
)BP_GET_TYPE(bp
));
837 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
838 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
839 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
840 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
842 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
843 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
844 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
845 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
847 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
848 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
849 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
851 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
852 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
853 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
856 if (BP_IS_EMBEDDED(bp
)) {
857 if (BPE_GET_ETYPE(bp
) > NUM_BP_EMBEDDED_TYPES
) {
858 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
859 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
864 * Pool-specific checks.
866 * Note: it would be nice to verify that the blk_birth and
867 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
868 * allows the birth time of log blocks (and dmu_sync()-ed blocks
869 * that are in the log) to be arbitrarily large.
871 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
872 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
874 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
875 zfs_panic_recover("blkptr at %p DVA %u has invalid "
877 bp
, i
, (longlong_t
)vdevid
);
880 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
882 zfs_panic_recover("blkptr at %p DVA %u has invalid "
884 bp
, i
, (longlong_t
)vdevid
);
887 if (vd
->vdev_ops
== &vdev_hole_ops
) {
888 zfs_panic_recover("blkptr at %p DVA %u has hole "
890 bp
, i
, (longlong_t
)vdevid
);
893 if (vd
->vdev_ops
== &vdev_missing_ops
) {
895 * "missing" vdevs are valid during import, but we
896 * don't have their detailed info (e.g. asize), so
897 * we can't perform any more checks on them.
901 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
902 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
904 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
905 if (offset
+ asize
> vd
->vdev_asize
) {
906 zfs_panic_recover("blkptr at %p DVA %u has invalid "
908 bp
, i
, (longlong_t
)offset
);
914 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
915 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
916 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
920 zfs_blkptr_verify(spa
, bp
);
922 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
923 data
, size
, size
, done
, private,
924 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
925 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
926 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
932 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
933 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
934 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
935 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
936 void *private, zio_priority_t priority
, enum zio_flag flags
,
937 const zbookmark_phys_t
*zb
)
941 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
942 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
943 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
944 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
945 DMU_OT_IS_VALID(zp
->zp_type
) &&
948 zp
->zp_copies
<= spa_max_replication(spa
));
950 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
951 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
952 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
953 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
955 zio
->io_ready
= ready
;
956 zio
->io_children_ready
= children_ready
;
957 zio
->io_physdone
= physdone
;
961 * Data can be NULL if we are going to call zio_write_override() to
962 * provide the already-allocated BP. But we may need the data to
963 * verify a dedup hit (if requested). In this case, don't try to
964 * dedup (just take the already-allocated BP verbatim). Encrypted
965 * dedup blocks need data as well so we also disable dedup in this
969 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
970 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
977 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
978 uint64_t size
, zio_done_func_t
*done
, void *private,
979 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
983 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
984 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
985 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
991 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
993 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
994 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
995 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
996 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
999 * We must reset the io_prop to match the values that existed
1000 * when the bp was first written by dmu_sync() keeping in mind
1001 * that nopwrite and dedup are mutually exclusive.
1003 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1004 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1005 zio
->io_prop
.zp_copies
= copies
;
1006 zio
->io_bp_override
= bp
;
1010 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1014 * The check for EMBEDDED is a performance optimization. We
1015 * process the free here (by ignoring it) rather than
1016 * putting it on the list and then processing it in zio_free_sync().
1018 if (BP_IS_EMBEDDED(bp
))
1020 metaslab_check_free(spa
, bp
);
1023 * Frees that are for the currently-syncing txg, are not going to be
1024 * deferred, and which will not need to do a read (i.e. not GANG or
1025 * DEDUP), can be processed immediately. Otherwise, put them on the
1026 * in-memory list for later processing.
1028 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
) ||
1029 txg
!= spa
->spa_syncing_txg
||
1030 spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
) {
1031 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1033 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1038 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1039 enum zio_flag flags
)
1042 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1044 ASSERT(!BP_IS_HOLE(bp
));
1045 ASSERT(spa_syncing_txg(spa
) == txg
);
1046 ASSERT(spa_sync_pass(spa
) < zfs_sync_pass_deferred_free
);
1048 if (BP_IS_EMBEDDED(bp
))
1049 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1051 metaslab_check_free(spa
, bp
);
1055 * GANG and DEDUP blocks can induce a read (for the gang block header,
1056 * or the DDT), so issue them asynchronously so that this thread is
1059 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1060 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1062 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1063 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1064 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1070 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1071 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1075 dprintf_bp(bp
, "claiming in txg %llu", txg
);
1077 if (BP_IS_EMBEDDED(bp
))
1078 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1081 * A claim is an allocation of a specific block. Claims are needed
1082 * to support immediate writes in the intent log. The issue is that
1083 * immediate writes contain committed data, but in a txg that was
1084 * *not* committed. Upon opening the pool after an unclean shutdown,
1085 * the intent log claims all blocks that contain immediate write data
1086 * so that the SPA knows they're in use.
1088 * All claims *must* be resolved in the first txg -- before the SPA
1089 * starts allocating blocks -- so that nothing is allocated twice.
1090 * If txg == 0 we just verify that the block is claimable.
1092 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <, spa_first_txg(spa
));
1093 ASSERT(txg
== spa_first_txg(spa
) || txg
== 0);
1094 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1096 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1097 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1098 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1099 ASSERT0(zio
->io_queued_timestamp
);
1105 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1106 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1111 if (vd
->vdev_children
== 0) {
1112 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1113 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1114 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1118 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1120 for (c
= 0; c
< vd
->vdev_children
; c
++)
1121 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1122 done
, private, flags
));
1129 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1130 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1131 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1135 ASSERT(vd
->vdev_children
== 0);
1136 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1137 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1138 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1140 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1141 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1142 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1144 zio
->io_prop
.zp_checksum
= checksum
;
1150 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1151 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1152 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1156 ASSERT(vd
->vdev_children
== 0);
1157 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1158 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1159 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1161 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1162 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1163 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1165 zio
->io_prop
.zp_checksum
= checksum
;
1167 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1169 * zec checksums are necessarily destructive -- they modify
1170 * the end of the write buffer to hold the verifier/checksum.
1171 * Therefore, we must make a local copy in case the data is
1172 * being written to multiple places in parallel.
1174 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1175 abd_copy(wbuf
, data
, size
);
1177 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1184 * Create a child I/O to do some work for us.
1187 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1188 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1189 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1191 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1194 ASSERT(vd
->vdev_parent
==
1195 (pio
->io_vd
? pio
->io_vd
: pio
->io_spa
->spa_root_vdev
));
1197 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1199 * If we have the bp, then the child should perform the
1200 * checksum and the parent need not. This pushes error
1201 * detection as close to the leaves as possible and
1202 * eliminates redundant checksums in the interior nodes.
1204 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1205 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1208 if (vd
->vdev_children
== 0)
1209 offset
+= VDEV_LABEL_START_SIZE
;
1211 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
) | ZIO_FLAG_DONT_PROPAGATE
;
1214 * If we've decided to do a repair, the write is not speculative --
1215 * even if the original read was.
1217 if (flags
& ZIO_FLAG_IO_REPAIR
)
1218 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1221 * If we're creating a child I/O that is not associated with a
1222 * top-level vdev, then the child zio is not an allocating I/O.
1223 * If this is a retried I/O then we ignore it since we will
1224 * have already processed the original allocating I/O.
1226 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1227 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1228 ASSERTV(metaslab_class_t
*mc
= spa_normal_class(pio
->io_spa
));
1230 ASSERT(mc
->mc_alloc_throttle_enabled
);
1231 ASSERT(type
== ZIO_TYPE_WRITE
);
1232 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1233 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1234 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1235 pio
->io_child_type
== ZIO_CHILD_GANG
);
1237 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1241 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1242 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1243 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1244 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1246 zio
->io_physdone
= pio
->io_physdone
;
1247 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1248 zio
->io_logical
->io_phys_children
++;
1254 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1255 int type
, zio_priority_t priority
, enum zio_flag flags
,
1256 zio_done_func_t
*done
, void *private)
1260 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1262 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1263 data
, size
, size
, done
, private, type
, priority
,
1264 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1266 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1272 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1274 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1276 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1280 zio_shrink(zio_t
*zio
, uint64_t size
)
1282 ASSERT(zio
->io_executor
== NULL
);
1283 ASSERT(zio
->io_orig_size
== zio
->io_size
);
1284 ASSERT(size
<= zio
->io_size
);
1287 * We don't shrink for raidz because of problems with the
1288 * reconstruction when reading back less than the block size.
1289 * Note, BP_IS_RAIDZ() assumes no compression.
1291 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1292 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1293 /* we are not doing a raw write */
1294 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1295 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1300 * ==========================================================================
1301 * Prepare to read and write logical blocks
1302 * ==========================================================================
1306 zio_read_bp_init(zio_t
*zio
)
1308 blkptr_t
*bp
= zio
->io_bp
;
1310 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1312 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1313 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1314 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1315 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1316 psize
, psize
, zio_decompress
);
1319 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1320 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1321 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1322 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1323 psize
, psize
, zio_decrypt
);
1326 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1327 int psize
= BPE_GET_PSIZE(bp
);
1328 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1330 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1331 decode_embedded_bp_compressed(bp
, data
);
1332 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1334 ASSERT(!BP_IS_EMBEDDED(bp
));
1337 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1338 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1340 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1341 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1343 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1344 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1346 return (ZIO_PIPELINE_CONTINUE
);
1350 zio_write_bp_init(zio_t
*zio
)
1352 if (!IO_IS_ALLOCATING(zio
))
1353 return (ZIO_PIPELINE_CONTINUE
);
1355 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1357 if (zio
->io_bp_override
) {
1358 blkptr_t
*bp
= zio
->io_bp
;
1359 zio_prop_t
*zp
= &zio
->io_prop
;
1361 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1362 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1364 *bp
= *zio
->io_bp_override
;
1365 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1367 if (BP_IS_EMBEDDED(bp
))
1368 return (ZIO_PIPELINE_CONTINUE
);
1371 * If we've been overridden and nopwrite is set then
1372 * set the flag accordingly to indicate that a nopwrite
1373 * has already occurred.
1375 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1376 ASSERT(!zp
->zp_dedup
);
1377 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1378 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1379 return (ZIO_PIPELINE_CONTINUE
);
1382 ASSERT(!zp
->zp_nopwrite
);
1384 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1385 return (ZIO_PIPELINE_CONTINUE
);
1387 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1388 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1390 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1392 BP_SET_DEDUP(bp
, 1);
1393 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1394 return (ZIO_PIPELINE_CONTINUE
);
1398 * We were unable to handle this as an override bp, treat
1399 * it as a regular write I/O.
1401 zio
->io_bp_override
= NULL
;
1402 *bp
= zio
->io_bp_orig
;
1403 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1406 return (ZIO_PIPELINE_CONTINUE
);
1410 zio_write_compress(zio_t
*zio
)
1412 spa_t
*spa
= zio
->io_spa
;
1413 zio_prop_t
*zp
= &zio
->io_prop
;
1414 enum zio_compress compress
= zp
->zp_compress
;
1415 blkptr_t
*bp
= zio
->io_bp
;
1416 uint64_t lsize
= zio
->io_lsize
;
1417 uint64_t psize
= zio
->io_size
;
1421 * If our children haven't all reached the ready stage,
1422 * wait for them and then repeat this pipeline stage.
1424 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
1425 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_READY
))
1426 return (ZIO_PIPELINE_STOP
);
1428 if (!IO_IS_ALLOCATING(zio
))
1429 return (ZIO_PIPELINE_CONTINUE
);
1431 if (zio
->io_children_ready
!= NULL
) {
1433 * Now that all our children are ready, run the callback
1434 * associated with this zio in case it wants to modify the
1435 * data to be written.
1437 ASSERT3U(zp
->zp_level
, >, 0);
1438 zio
->io_children_ready(zio
);
1441 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1442 ASSERT(zio
->io_bp_override
== NULL
);
1444 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1446 * We're rewriting an existing block, which means we're
1447 * working on behalf of spa_sync(). For spa_sync() to
1448 * converge, it must eventually be the case that we don't
1449 * have to allocate new blocks. But compression changes
1450 * the blocksize, which forces a reallocate, and makes
1451 * convergence take longer. Therefore, after the first
1452 * few passes, stop compressing to ensure convergence.
1454 pass
= spa_sync_pass(spa
);
1456 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1457 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1458 ASSERT(!BP_GET_DEDUP(bp
));
1460 if (pass
>= zfs_sync_pass_dont_compress
)
1461 compress
= ZIO_COMPRESS_OFF
;
1463 /* Make sure someone doesn't change their mind on overwrites */
1464 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1465 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1468 /* If it's a compressed write that is not raw, compress the buffer. */
1469 if (compress
!= ZIO_COMPRESS_OFF
&&
1470 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1471 void *cbuf
= zio_buf_alloc(lsize
);
1472 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1473 if (psize
== 0 || psize
== lsize
) {
1474 compress
= ZIO_COMPRESS_OFF
;
1475 zio_buf_free(cbuf
, lsize
);
1476 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1477 psize
<= BPE_PAYLOAD_SIZE
&&
1478 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1479 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1480 encode_embedded_bp_compressed(bp
,
1481 cbuf
, compress
, lsize
, psize
);
1482 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1483 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1484 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1485 zio_buf_free(cbuf
, lsize
);
1486 bp
->blk_birth
= zio
->io_txg
;
1487 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1488 ASSERT(spa_feature_is_active(spa
,
1489 SPA_FEATURE_EMBEDDED_DATA
));
1490 return (ZIO_PIPELINE_CONTINUE
);
1493 * Round up compressed size up to the ashift
1494 * of the smallest-ashift device, and zero the tail.
1495 * This ensures that the compressed size of the BP
1496 * (and thus compressratio property) are correct,
1497 * in that we charge for the padding used to fill out
1500 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1501 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1502 1ULL << spa
->spa_min_ashift
);
1503 if (rounded
>= lsize
) {
1504 compress
= ZIO_COMPRESS_OFF
;
1505 zio_buf_free(cbuf
, lsize
);
1508 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1509 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1510 abd_zero_off(cdata
, psize
, rounded
- psize
);
1512 zio_push_transform(zio
, cdata
,
1513 psize
, lsize
, NULL
);
1518 * We were unable to handle this as an override bp, treat
1519 * it as a regular write I/O.
1521 zio
->io_bp_override
= NULL
;
1522 *bp
= zio
->io_bp_orig
;
1523 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1526 ASSERT3U(psize
, !=, 0);
1531 * The final pass of spa_sync() must be all rewrites, but the first
1532 * few passes offer a trade-off: allocating blocks defers convergence,
1533 * but newly allocated blocks are sequential, so they can be written
1534 * to disk faster. Therefore, we allow the first few passes of
1535 * spa_sync() to allocate new blocks, but force rewrites after that.
1536 * There should only be a handful of blocks after pass 1 in any case.
1538 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1539 BP_GET_PSIZE(bp
) == psize
&&
1540 pass
>= zfs_sync_pass_rewrite
) {
1542 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1543 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1544 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1547 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1551 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1552 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1553 BP_SET_LSIZE(bp
, lsize
);
1554 BP_SET_TYPE(bp
, zp
->zp_type
);
1555 BP_SET_LEVEL(bp
, zp
->zp_level
);
1556 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1558 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1560 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1561 BP_SET_LSIZE(bp
, lsize
);
1562 BP_SET_TYPE(bp
, zp
->zp_type
);
1563 BP_SET_LEVEL(bp
, zp
->zp_level
);
1564 BP_SET_PSIZE(bp
, psize
);
1565 BP_SET_COMPRESS(bp
, compress
);
1566 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1567 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1568 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1570 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1571 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1572 ASSERT(!zp
->zp_encrypt
||
1573 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1574 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1576 if (zp
->zp_nopwrite
) {
1577 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1578 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1579 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1582 return (ZIO_PIPELINE_CONTINUE
);
1586 zio_free_bp_init(zio_t
*zio
)
1588 blkptr_t
*bp
= zio
->io_bp
;
1590 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1591 if (BP_GET_DEDUP(bp
))
1592 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1595 return (ZIO_PIPELINE_CONTINUE
);
1599 * ==========================================================================
1600 * Execute the I/O pipeline
1601 * ==========================================================================
1605 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1607 spa_t
*spa
= zio
->io_spa
;
1608 zio_type_t t
= zio
->io_type
;
1609 int flags
= (cutinline
? TQ_FRONT
: 0);
1612 * If we're a config writer or a probe, the normal issue and
1613 * interrupt threads may all be blocked waiting for the config lock.
1614 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1616 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1620 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1622 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1626 * If this is a high priority I/O, then use the high priority taskq if
1629 if (zio
->io_priority
== ZIO_PRIORITY_NOW
&&
1630 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1633 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1636 * NB: We are assuming that the zio can only be dispatched
1637 * to a single taskq at a time. It would be a grievous error
1638 * to dispatch the zio to another taskq at the same time.
1640 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1641 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1642 flags
, &zio
->io_tqent
);
1646 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1648 kthread_t
*executor
= zio
->io_executor
;
1649 spa_t
*spa
= zio
->io_spa
;
1651 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1652 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1654 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1655 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1664 zio_issue_async(zio_t
*zio
)
1666 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1668 return (ZIO_PIPELINE_STOP
);
1672 zio_interrupt(zio_t
*zio
)
1674 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1678 zio_delay_interrupt(zio_t
*zio
)
1681 * The timeout_generic() function isn't defined in userspace, so
1682 * rather than trying to implement the function, the zio delay
1683 * functionality has been disabled for userspace builds.
1688 * If io_target_timestamp is zero, then no delay has been registered
1689 * for this IO, thus jump to the end of this function and "skip" the
1690 * delay; issuing it directly to the zio layer.
1692 if (zio
->io_target_timestamp
!= 0) {
1693 hrtime_t now
= gethrtime();
1695 if (now
>= zio
->io_target_timestamp
) {
1697 * This IO has already taken longer than the target
1698 * delay to complete, so we don't want to delay it
1699 * any longer; we "miss" the delay and issue it
1700 * directly to the zio layer. This is likely due to
1701 * the target latency being set to a value less than
1702 * the underlying hardware can satisfy (e.g. delay
1703 * set to 1ms, but the disks take 10ms to complete an
1707 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
1713 hrtime_t diff
= zio
->io_target_timestamp
- now
;
1714 clock_t expire_at_tick
= ddi_get_lbolt() +
1717 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
1718 hrtime_t
, now
, hrtime_t
, diff
);
1720 if (NSEC_TO_TICK(diff
) == 0) {
1721 /* Our delay is less than a jiffy - just spin */
1722 zfs_sleep_until(zio
->io_target_timestamp
);
1725 * Use taskq_dispatch_delay() in the place of
1726 * OpenZFS's timeout_generic().
1728 tid
= taskq_dispatch_delay(system_taskq
,
1729 (task_func_t
*)zio_interrupt
,
1730 zio
, TQ_NOSLEEP
, expire_at_tick
);
1731 if (tid
== TASKQID_INVALID
) {
1733 * Couldn't allocate a task. Just
1734 * finish the zio without a delay.
1743 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
1748 * Execute the I/O pipeline until one of the following occurs:
1749 * (1) the I/O completes; (2) the pipeline stalls waiting for
1750 * dependent child I/Os; (3) the I/O issues, so we're waiting
1751 * for an I/O completion interrupt; (4) the I/O is delegated by
1752 * vdev-level caching or aggregation; (5) the I/O is deferred
1753 * due to vdev-level queueing; (6) the I/O is handed off to
1754 * another thread. In all cases, the pipeline stops whenever
1755 * there's no CPU work; it never burns a thread in cv_wait_io().
1757 * There's no locking on io_stage because there's no legitimate way
1758 * for multiple threads to be attempting to process the same I/O.
1760 static zio_pipe_stage_t
*zio_pipeline
[];
1763 * zio_execute() is a wrapper around the static function
1764 * __zio_execute() so that we can force __zio_execute() to be
1765 * inlined. This reduces stack overhead which is important
1766 * because __zio_execute() is called recursively in several zio
1767 * code paths. zio_execute() itself cannot be inlined because
1768 * it is externally visible.
1771 zio_execute(zio_t
*zio
)
1773 fstrans_cookie_t cookie
;
1775 cookie
= spl_fstrans_mark();
1777 spl_fstrans_unmark(cookie
);
1781 * Used to determine if in the current context the stack is sized large
1782 * enough to allow zio_execute() to be called recursively. A minimum
1783 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1786 zio_execute_stack_check(zio_t
*zio
)
1788 #if !defined(HAVE_LARGE_STACKS)
1789 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1791 /* Executing in txg_sync_thread() context. */
1792 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1795 /* Pool initialization outside of zio_taskq context. */
1796 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1797 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1798 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1800 #endif /* HAVE_LARGE_STACKS */
1805 __attribute__((always_inline
))
1807 __zio_execute(zio_t
*zio
)
1809 zio
->io_executor
= curthread
;
1811 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1813 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1814 enum zio_stage pipeline
= zio
->io_pipeline
;
1815 enum zio_stage stage
= zio
->io_stage
;
1818 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1819 ASSERT(ISP2(stage
));
1820 ASSERT(zio
->io_stall
== NULL
);
1824 } while ((stage
& pipeline
) == 0);
1826 ASSERT(stage
<= ZIO_STAGE_DONE
);
1829 * If we are in interrupt context and this pipeline stage
1830 * will grab a config lock that is held across I/O,
1831 * or may wait for an I/O that needs an interrupt thread
1832 * to complete, issue async to avoid deadlock.
1834 * For VDEV_IO_START, we cut in line so that the io will
1835 * be sent to disk promptly.
1837 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1838 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1839 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1840 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1841 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1846 * If the current context doesn't have large enough stacks
1847 * the zio must be issued asynchronously to prevent overflow.
1849 if (zio_execute_stack_check(zio
)) {
1850 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1851 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1852 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
1856 zio
->io_stage
= stage
;
1857 zio
->io_pipeline_trace
|= zio
->io_stage
;
1858 rv
= zio_pipeline
[highbit64(stage
) - 1](zio
);
1860 if (rv
== ZIO_PIPELINE_STOP
)
1863 ASSERT(rv
== ZIO_PIPELINE_CONTINUE
);
1869 * ==========================================================================
1870 * Initiate I/O, either sync or async
1871 * ==========================================================================
1874 zio_wait(zio_t
*zio
)
1878 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1879 ASSERT(zio
->io_executor
== NULL
);
1881 zio
->io_waiter
= curthread
;
1882 ASSERT0(zio
->io_queued_timestamp
);
1883 zio
->io_queued_timestamp
= gethrtime();
1887 mutex_enter(&zio
->io_lock
);
1888 while (zio
->io_executor
!= NULL
)
1889 cv_wait_io(&zio
->io_cv
, &zio
->io_lock
);
1890 mutex_exit(&zio
->io_lock
);
1892 error
= zio
->io_error
;
1899 zio_nowait(zio_t
*zio
)
1901 ASSERT(zio
->io_executor
== NULL
);
1903 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1904 zio_unique_parent(zio
) == NULL
) {
1908 * This is a logical async I/O with no parent to wait for it.
1909 * We add it to the spa_async_root_zio "Godfather" I/O which
1910 * will ensure they complete prior to unloading the pool.
1912 spa_t
*spa
= zio
->io_spa
;
1914 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
1917 zio_add_child(pio
, zio
);
1920 ASSERT0(zio
->io_queued_timestamp
);
1921 zio
->io_queued_timestamp
= gethrtime();
1926 * ==========================================================================
1927 * Reexecute or suspend/resume failed I/O
1928 * ==========================================================================
1932 zio_reexecute(zio_t
*pio
)
1934 zio_t
*cio
, *cio_next
;
1936 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1937 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
1938 ASSERT(pio
->io_gang_leader
== NULL
);
1939 ASSERT(pio
->io_gang_tree
== NULL
);
1941 pio
->io_flags
= pio
->io_orig_flags
;
1942 pio
->io_stage
= pio
->io_orig_stage
;
1943 pio
->io_pipeline
= pio
->io_orig_pipeline
;
1944 pio
->io_reexecute
= 0;
1945 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
1946 pio
->io_pipeline_trace
= 0;
1948 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1949 pio
->io_state
[w
] = 0;
1950 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
1951 pio
->io_child_error
[c
] = 0;
1953 if (IO_IS_ALLOCATING(pio
))
1954 BP_ZERO(pio
->io_bp
);
1957 * As we reexecute pio's children, new children could be created.
1958 * New children go to the head of pio's io_child_list, however,
1959 * so we will (correctly) not reexecute them. The key is that
1960 * the remainder of pio's io_child_list, from 'cio_next' onward,
1961 * cannot be affected by any side effects of reexecuting 'cio'.
1963 zio_link_t
*zl
= NULL
;
1964 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1965 cio_next
= zio_walk_children(pio
, &zl
);
1966 mutex_enter(&pio
->io_lock
);
1967 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
1968 pio
->io_children
[cio
->io_child_type
][w
]++;
1969 mutex_exit(&pio
->io_lock
);
1974 * Now that all children have been reexecuted, execute the parent.
1975 * We don't reexecute "The Godfather" I/O here as it's the
1976 * responsibility of the caller to wait on it.
1978 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
1979 pio
->io_queued_timestamp
= gethrtime();
1985 zio_suspend(spa_t
*spa
, zio_t
*zio
)
1987 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
1988 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1989 "failure and the failure mode property for this pool "
1990 "is set to panic.", spa_name(spa
));
1992 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
1993 "failure and has been suspended.\n", spa_name(spa
));
1995 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
1998 mutex_enter(&spa
->spa_suspend_lock
);
2000 if (spa
->spa_suspend_zio_root
== NULL
)
2001 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2002 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2003 ZIO_FLAG_GODFATHER
);
2005 spa
->spa_suspended
= B_TRUE
;
2008 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2009 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2010 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2011 ASSERT(zio_unique_parent(zio
) == NULL
);
2012 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2013 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2016 mutex_exit(&spa
->spa_suspend_lock
);
2020 zio_resume(spa_t
*spa
)
2025 * Reexecute all previously suspended i/o.
2027 mutex_enter(&spa
->spa_suspend_lock
);
2028 spa
->spa_suspended
= B_FALSE
;
2029 cv_broadcast(&spa
->spa_suspend_cv
);
2030 pio
= spa
->spa_suspend_zio_root
;
2031 spa
->spa_suspend_zio_root
= NULL
;
2032 mutex_exit(&spa
->spa_suspend_lock
);
2038 return (zio_wait(pio
));
2042 zio_resume_wait(spa_t
*spa
)
2044 mutex_enter(&spa
->spa_suspend_lock
);
2045 while (spa_suspended(spa
))
2046 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2047 mutex_exit(&spa
->spa_suspend_lock
);
2051 * ==========================================================================
2054 * A gang block is a collection of small blocks that looks to the DMU
2055 * like one large block. When zio_dva_allocate() cannot find a block
2056 * of the requested size, due to either severe fragmentation or the pool
2057 * being nearly full, it calls zio_write_gang_block() to construct the
2058 * block from smaller fragments.
2060 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2061 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2062 * an indirect block: it's an array of block pointers. It consumes
2063 * only one sector and hence is allocatable regardless of fragmentation.
2064 * The gang header's bps point to its gang members, which hold the data.
2066 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2067 * as the verifier to ensure uniqueness of the SHA256 checksum.
2068 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2069 * not the gang header. This ensures that data block signatures (needed for
2070 * deduplication) are independent of how the block is physically stored.
2072 * Gang blocks can be nested: a gang member may itself be a gang block.
2073 * Thus every gang block is a tree in which root and all interior nodes are
2074 * gang headers, and the leaves are normal blocks that contain user data.
2075 * The root of the gang tree is called the gang leader.
2077 * To perform any operation (read, rewrite, free, claim) on a gang block,
2078 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2079 * in the io_gang_tree field of the original logical i/o by recursively
2080 * reading the gang leader and all gang headers below it. This yields
2081 * an in-core tree containing the contents of every gang header and the
2082 * bps for every constituent of the gang block.
2084 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2085 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2086 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2087 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2088 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2089 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2090 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2091 * of the gang header plus zio_checksum_compute() of the data to update the
2092 * gang header's blk_cksum as described above.
2094 * The two-phase assemble/issue model solves the problem of partial failure --
2095 * what if you'd freed part of a gang block but then couldn't read the
2096 * gang header for another part? Assembling the entire gang tree first
2097 * ensures that all the necessary gang header I/O has succeeded before
2098 * starting the actual work of free, claim, or write. Once the gang tree
2099 * is assembled, free and claim are in-memory operations that cannot fail.
2101 * In the event that a gang write fails, zio_dva_unallocate() walks the
2102 * gang tree to immediately free (i.e. insert back into the space map)
2103 * everything we've allocated. This ensures that we don't get ENOSPC
2104 * errors during repeated suspend/resume cycles due to a flaky device.
2106 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2107 * the gang tree, we won't modify the block, so we can safely defer the free
2108 * (knowing that the block is still intact). If we *can* assemble the gang
2109 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2110 * each constituent bp and we can allocate a new block on the next sync pass.
2112 * In all cases, the gang tree allows complete recovery from partial failure.
2113 * ==========================================================================
2117 zio_gang_issue_func_done(zio_t
*zio
)
2119 abd_put(zio
->io_abd
);
2123 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2129 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2130 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2131 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2132 &pio
->io_bookmark
));
2136 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2143 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2144 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2145 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2146 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2149 * As we rewrite each gang header, the pipeline will compute
2150 * a new gang block header checksum for it; but no one will
2151 * compute a new data checksum, so we do that here. The one
2152 * exception is the gang leader: the pipeline already computed
2153 * its data checksum because that stage precedes gang assembly.
2154 * (Presently, nothing actually uses interior data checksums;
2155 * this is just good hygiene.)
2157 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2158 abd_t
*buf
= abd_get_offset(data
, offset
);
2160 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2161 buf
, BP_GET_PSIZE(bp
));
2166 * If we are here to damage data for testing purposes,
2167 * leave the GBH alone so that we can detect the damage.
2169 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2170 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2172 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2173 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2174 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2175 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2183 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2186 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2187 ZIO_GANG_CHILD_FLAGS(pio
)));
2192 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2195 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2196 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2199 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2208 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2210 static zio_gang_node_t
*
2211 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2213 zio_gang_node_t
*gn
;
2215 ASSERT(*gnpp
== NULL
);
2217 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2218 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2225 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2227 zio_gang_node_t
*gn
= *gnpp
;
2229 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2230 ASSERT(gn
->gn_child
[g
] == NULL
);
2232 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2233 kmem_free(gn
, sizeof (*gn
));
2238 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2240 zio_gang_node_t
*gn
= *gnpp
;
2245 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2246 zio_gang_tree_free(&gn
->gn_child
[g
]);
2248 zio_gang_node_free(gnpp
);
2252 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2254 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2255 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2257 ASSERT(gio
->io_gang_leader
== gio
);
2258 ASSERT(BP_IS_GANG(bp
));
2260 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2261 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2262 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2266 zio_gang_tree_assemble_done(zio_t
*zio
)
2268 zio_t
*gio
= zio
->io_gang_leader
;
2269 zio_gang_node_t
*gn
= zio
->io_private
;
2270 blkptr_t
*bp
= zio
->io_bp
;
2272 ASSERT(gio
== zio_unique_parent(zio
));
2273 ASSERT(zio
->io_child_count
== 0);
2278 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2279 if (BP_SHOULD_BYTESWAP(bp
))
2280 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2282 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2283 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2284 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2286 abd_put(zio
->io_abd
);
2288 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2289 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2290 if (!BP_IS_GANG(gbp
))
2292 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2297 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2300 zio_t
*gio
= pio
->io_gang_leader
;
2303 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2304 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2305 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2308 * If you're a gang header, your data is in gn->gn_gbh.
2309 * If you're a gang member, your data is in 'data' and gn == NULL.
2311 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2314 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2316 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2317 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2318 if (BP_IS_HOLE(gbp
))
2320 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2322 offset
+= BP_GET_PSIZE(gbp
);
2326 if (gn
== gio
->io_gang_tree
)
2327 ASSERT3U(gio
->io_size
, ==, offset
);
2334 zio_gang_assemble(zio_t
*zio
)
2336 blkptr_t
*bp
= zio
->io_bp
;
2338 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2339 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2341 zio
->io_gang_leader
= zio
;
2343 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2345 return (ZIO_PIPELINE_CONTINUE
);
2349 zio_gang_issue(zio_t
*zio
)
2351 blkptr_t
*bp
= zio
->io_bp
;
2353 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
))
2354 return (ZIO_PIPELINE_STOP
);
2356 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2357 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2359 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2360 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2363 zio_gang_tree_free(&zio
->io_gang_tree
);
2365 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2367 return (ZIO_PIPELINE_CONTINUE
);
2371 zio_write_gang_member_ready(zio_t
*zio
)
2373 zio_t
*pio
= zio_unique_parent(zio
);
2374 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2375 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2377 ASSERTV(zio_t
*gio
= zio
->io_gang_leader
);
2379 if (BP_IS_HOLE(zio
->io_bp
))
2382 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2384 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2385 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2386 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2387 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2388 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2390 mutex_enter(&pio
->io_lock
);
2391 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2392 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2393 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2394 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2395 DVA_SET_ASIZE(&pdva
[d
], asize
);
2397 mutex_exit(&pio
->io_lock
);
2401 zio_write_gang_done(zio_t
*zio
)
2403 abd_put(zio
->io_abd
);
2407 zio_write_gang_block(zio_t
*pio
)
2409 spa_t
*spa
= pio
->io_spa
;
2410 metaslab_class_t
*mc
= spa_normal_class(spa
);
2411 blkptr_t
*bp
= pio
->io_bp
;
2412 zio_t
*gio
= pio
->io_gang_leader
;
2414 zio_gang_node_t
*gn
, **gnpp
;
2415 zio_gbh_phys_t
*gbh
;
2417 uint64_t txg
= pio
->io_txg
;
2418 uint64_t resid
= pio
->io_size
;
2420 int copies
= gio
->io_prop
.zp_copies
;
2426 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2427 * have a third copy.
2429 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2430 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2431 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2433 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2434 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2435 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2436 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2438 flags
|= METASLAB_ASYNC_ALLOC
;
2439 VERIFY(refcount_held(&mc
->mc_alloc_slots
, pio
));
2442 * The logical zio has already placed a reservation for
2443 * 'copies' allocation slots but gang blocks may require
2444 * additional copies. These additional copies
2445 * (i.e. gbh_copies - copies) are guaranteed to succeed
2446 * since metaslab_class_throttle_reserve() always allows
2447 * additional reservations for gang blocks.
2449 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2453 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2454 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2455 &pio
->io_alloc_list
, pio
);
2457 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2458 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2459 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2462 * If we failed to allocate the gang block header then
2463 * we remove any additional allocation reservations that
2464 * we placed here. The original reservation will
2465 * be removed when the logical I/O goes to the ready
2468 metaslab_class_throttle_unreserve(mc
,
2469 gbh_copies
- copies
, pio
);
2472 pio
->io_error
= error
;
2473 return (ZIO_PIPELINE_CONTINUE
);
2477 gnpp
= &gio
->io_gang_tree
;
2479 gnpp
= pio
->io_private
;
2480 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2483 gn
= zio_gang_node_alloc(gnpp
);
2485 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2486 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2489 * Create the gang header.
2491 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2492 zio_write_gang_done
, NULL
, pio
->io_priority
,
2493 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2496 * Create and nowait the gang children.
2498 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2499 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2501 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2503 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2504 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2505 zp
.zp_type
= DMU_OT_NONE
;
2507 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2508 zp
.zp_dedup
= B_FALSE
;
2509 zp
.zp_dedup_verify
= B_FALSE
;
2510 zp
.zp_nopwrite
= B_FALSE
;
2511 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2512 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2513 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2514 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2515 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2517 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2518 abd_get_offset(pio
->io_abd
, pio
->io_size
- resid
), lsize
,
2519 lsize
, &zp
, zio_write_gang_member_ready
, NULL
, NULL
,
2520 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2521 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2523 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2524 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2525 ASSERT(!(pio
->io_flags
& ZIO_FLAG_NODATA
));
2528 * Gang children won't throttle but we should
2529 * account for their work, so reserve an allocation
2530 * slot for them here.
2532 VERIFY(metaslab_class_throttle_reserve(mc
,
2533 zp
.zp_copies
, cio
, flags
));
2539 * Set pio's pipeline to just wait for zio to finish.
2541 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2544 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2546 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2550 return (ZIO_PIPELINE_CONTINUE
);
2554 * The zio_nop_write stage in the pipeline determines if allocating a
2555 * new bp is necessary. The nopwrite feature can handle writes in
2556 * either syncing or open context (i.e. zil writes) and as a result is
2557 * mutually exclusive with dedup.
2559 * By leveraging a cryptographically secure checksum, such as SHA256, we
2560 * can compare the checksums of the new data and the old to determine if
2561 * allocating a new block is required. Note that our requirements for
2562 * cryptographic strength are fairly weak: there can't be any accidental
2563 * hash collisions, but we don't need to be secure against intentional
2564 * (malicious) collisions. To trigger a nopwrite, you have to be able
2565 * to write the file to begin with, and triggering an incorrect (hash
2566 * collision) nopwrite is no worse than simply writing to the file.
2567 * That said, there are no known attacks against the checksum algorithms
2568 * used for nopwrite, assuming that the salt and the checksums
2569 * themselves remain secret.
2572 zio_nop_write(zio_t
*zio
)
2574 blkptr_t
*bp
= zio
->io_bp
;
2575 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2576 zio_prop_t
*zp
= &zio
->io_prop
;
2578 ASSERT(BP_GET_LEVEL(bp
) == 0);
2579 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2580 ASSERT(zp
->zp_nopwrite
);
2581 ASSERT(!zp
->zp_dedup
);
2582 ASSERT(zio
->io_bp_override
== NULL
);
2583 ASSERT(IO_IS_ALLOCATING(zio
));
2586 * Check to see if the original bp and the new bp have matching
2587 * characteristics (i.e. same checksum, compression algorithms, etc).
2588 * If they don't then just continue with the pipeline which will
2589 * allocate a new bp.
2591 if (BP_IS_HOLE(bp_orig
) ||
2592 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2593 ZCHECKSUM_FLAG_NOPWRITE
) ||
2594 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2595 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2596 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2597 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2598 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2599 return (ZIO_PIPELINE_CONTINUE
);
2602 * If the checksums match then reset the pipeline so that we
2603 * avoid allocating a new bp and issuing any I/O.
2605 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2606 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2607 ZCHECKSUM_FLAG_NOPWRITE
);
2608 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2609 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2610 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2611 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2612 sizeof (uint64_t)) == 0);
2615 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2616 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2619 return (ZIO_PIPELINE_CONTINUE
);
2623 * ==========================================================================
2625 * ==========================================================================
2628 zio_ddt_child_read_done(zio_t
*zio
)
2630 blkptr_t
*bp
= zio
->io_bp
;
2631 ddt_entry_t
*dde
= zio
->io_private
;
2633 zio_t
*pio
= zio_unique_parent(zio
);
2635 mutex_enter(&pio
->io_lock
);
2636 ddp
= ddt_phys_select(dde
, bp
);
2637 if (zio
->io_error
== 0)
2638 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2640 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2641 dde
->dde_repair_abd
= zio
->io_abd
;
2643 abd_free(zio
->io_abd
);
2644 mutex_exit(&pio
->io_lock
);
2648 zio_ddt_read_start(zio_t
*zio
)
2650 blkptr_t
*bp
= zio
->io_bp
;
2652 ASSERT(BP_GET_DEDUP(bp
));
2653 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2654 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2656 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2657 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2658 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2659 ddt_phys_t
*ddp
= dde
->dde_phys
;
2660 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2663 ASSERT(zio
->io_vsd
== NULL
);
2666 if (ddp_self
== NULL
)
2667 return (ZIO_PIPELINE_CONTINUE
);
2669 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2670 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2672 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2674 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2675 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2676 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2677 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2678 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2680 return (ZIO_PIPELINE_CONTINUE
);
2683 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2684 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2685 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2687 return (ZIO_PIPELINE_CONTINUE
);
2691 zio_ddt_read_done(zio_t
*zio
)
2693 blkptr_t
*bp
= zio
->io_bp
;
2695 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
))
2696 return (ZIO_PIPELINE_STOP
);
2698 ASSERT(BP_GET_DEDUP(bp
));
2699 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2700 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2702 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2703 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2704 ddt_entry_t
*dde
= zio
->io_vsd
;
2706 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2707 return (ZIO_PIPELINE_CONTINUE
);
2710 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2711 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2712 return (ZIO_PIPELINE_STOP
);
2714 if (dde
->dde_repair_abd
!= NULL
) {
2715 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2717 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2719 ddt_repair_done(ddt
, dde
);
2723 ASSERT(zio
->io_vsd
== NULL
);
2725 return (ZIO_PIPELINE_CONTINUE
);
2729 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2731 spa_t
*spa
= zio
->io_spa
;
2732 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2734 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2737 * Note: we compare the original data, not the transformed data,
2738 * because when zio->io_bp is an override bp, we will not have
2739 * pushed the I/O transforms. That's an important optimization
2740 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2741 * However, we should never get a raw, override zio so in these
2742 * cases we can compare the io_abd directly. This is useful because
2743 * it allows us to do dedup verification even if we don't have access
2744 * to the original data (for instance, if the encryption keys aren't
2748 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2749 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2751 if (lio
!= NULL
&& do_raw
) {
2752 return (lio
->io_size
!= zio
->io_size
||
2753 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2754 } else if (lio
!= NULL
) {
2755 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2756 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2760 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2761 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2763 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2764 blkptr_t blk
= *zio
->io_bp
;
2769 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2770 psize
= BP_GET_PSIZE(&blk
);
2772 if (psize
!= zio
->io_size
)
2777 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2779 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2780 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2781 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2782 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
2785 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
2786 error
= SET_ERROR(ENOENT
);
2791 return (error
!= 0);
2792 } else if (ddp
->ddp_phys_birth
!= 0) {
2793 arc_buf_t
*abuf
= NULL
;
2794 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2795 blkptr_t blk
= *zio
->io_bp
;
2798 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2800 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
2805 error
= arc_read(NULL
, spa
, &blk
,
2806 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
2807 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2808 &aflags
, &zio
->io_bookmark
);
2811 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
2812 zio
->io_orig_size
) != 0)
2813 error
= SET_ERROR(ENOENT
);
2814 arc_buf_destroy(abuf
, &abuf
);
2818 return (error
!= 0);
2826 zio_ddt_child_write_ready(zio_t
*zio
)
2828 int p
= zio
->io_prop
.zp_copies
;
2829 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2830 ddt_entry_t
*dde
= zio
->io_private
;
2831 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2839 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2841 ddt_phys_fill(ddp
, zio
->io_bp
);
2843 zio_link_t
*zl
= NULL
;
2844 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
2845 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
2851 zio_ddt_child_write_done(zio_t
*zio
)
2853 int p
= zio
->io_prop
.zp_copies
;
2854 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
2855 ddt_entry_t
*dde
= zio
->io_private
;
2856 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2860 ASSERT(ddp
->ddp_refcnt
== 0);
2861 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2862 dde
->dde_lead_zio
[p
] = NULL
;
2864 if (zio
->io_error
== 0) {
2865 zio_link_t
*zl
= NULL
;
2866 while (zio_walk_parents(zio
, &zl
) != NULL
)
2867 ddt_phys_addref(ddp
);
2869 ddt_phys_clear(ddp
);
2876 zio_ddt_ditto_write_done(zio_t
*zio
)
2878 int p
= DDT_PHYS_DITTO
;
2879 ASSERTV(zio_prop_t
*zp
= &zio
->io_prop
);
2880 blkptr_t
*bp
= zio
->io_bp
;
2881 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2882 ddt_entry_t
*dde
= zio
->io_private
;
2883 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2884 ddt_key_t
*ddk
= &dde
->dde_key
;
2888 ASSERT(ddp
->ddp_refcnt
== 0);
2889 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
2890 dde
->dde_lead_zio
[p
] = NULL
;
2892 if (zio
->io_error
== 0) {
2893 ASSERT(ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, ddk
->ddk_cksum
));
2894 ASSERT(zp
->zp_copies
< SPA_DVAS_PER_BP
);
2895 ASSERT(zp
->zp_copies
== BP_GET_NDVAS(bp
) - BP_IS_GANG(bp
));
2896 if (ddp
->ddp_phys_birth
!= 0)
2897 ddt_phys_free(ddt
, ddk
, ddp
, zio
->io_txg
);
2898 ddt_phys_fill(ddp
, bp
);
2905 zio_ddt_write(zio_t
*zio
)
2907 spa_t
*spa
= zio
->io_spa
;
2908 blkptr_t
*bp
= zio
->io_bp
;
2909 uint64_t txg
= zio
->io_txg
;
2910 zio_prop_t
*zp
= &zio
->io_prop
;
2911 int p
= zp
->zp_copies
;
2915 ddt_t
*ddt
= ddt_select(spa
, bp
);
2919 ASSERT(BP_GET_DEDUP(bp
));
2920 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
2921 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
2922 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
2925 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
2926 ddp
= &dde
->dde_phys
[p
];
2928 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
2930 * If we're using a weak checksum, upgrade to a strong checksum
2931 * and try again. If we're already using a strong checksum,
2932 * we can't resolve it, so just convert to an ordinary write.
2933 * (And automatically e-mail a paper to Nature?)
2935 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2936 ZCHECKSUM_FLAG_DEDUP
)) {
2937 zp
->zp_checksum
= spa_dedup_checksum(spa
);
2938 zio_pop_transforms(zio
);
2939 zio
->io_stage
= ZIO_STAGE_OPEN
;
2942 zp
->zp_dedup
= B_FALSE
;
2944 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2946 return (ZIO_PIPELINE_CONTINUE
);
2949 ditto_copies
= ddt_ditto_copies_needed(ddt
, dde
, ddp
);
2950 ASSERT(ditto_copies
< SPA_DVAS_PER_BP
);
2952 if (ditto_copies
> ddt_ditto_copies_present(dde
) &&
2953 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] == NULL
) {
2954 zio_prop_t czp
= *zp
;
2956 czp
.zp_copies
= ditto_copies
;
2959 * If we arrived here with an override bp, we won't have run
2960 * the transform stack, so we won't have the data we need to
2961 * generate a child i/o. So, toss the override bp and restart.
2962 * This is safe, because using the override bp is just an
2963 * optimization; and it's rare, so the cost doesn't matter.
2965 if (zio
->io_bp_override
) {
2966 zio_pop_transforms(zio
);
2967 zio
->io_stage
= ZIO_STAGE_OPEN
;
2968 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
2969 zio
->io_bp_override
= NULL
;
2972 return (ZIO_PIPELINE_CONTINUE
);
2975 dio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2976 zio
->io_orig_size
, zio
->io_orig_size
, &czp
, NULL
, NULL
,
2977 NULL
, zio_ddt_ditto_write_done
, dde
, zio
->io_priority
,
2978 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
2980 zio_push_transform(dio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
2981 dde
->dde_lead_zio
[DDT_PHYS_DITTO
] = dio
;
2984 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
2985 if (ddp
->ddp_phys_birth
!= 0)
2986 ddt_bp_fill(ddp
, bp
, txg
);
2987 if (dde
->dde_lead_zio
[p
] != NULL
)
2988 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
2990 ddt_phys_addref(ddp
);
2991 } else if (zio
->io_bp_override
) {
2992 ASSERT(bp
->blk_birth
== txg
);
2993 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
2994 ddt_phys_fill(ddp
, bp
);
2995 ddt_phys_addref(ddp
);
2997 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
2998 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
2999 zio_ddt_child_write_ready
, NULL
, NULL
,
3000 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3001 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3003 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3004 dde
->dde_lead_zio
[p
] = cio
;
3014 return (ZIO_PIPELINE_CONTINUE
);
3017 ddt_entry_t
*freedde
; /* for debugging */
3020 zio_ddt_free(zio_t
*zio
)
3022 spa_t
*spa
= zio
->io_spa
;
3023 blkptr_t
*bp
= zio
->io_bp
;
3024 ddt_t
*ddt
= ddt_select(spa
, bp
);
3028 ASSERT(BP_GET_DEDUP(bp
));
3029 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3032 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3034 ddp
= ddt_phys_select(dde
, bp
);
3036 ddt_phys_decref(ddp
);
3040 return (ZIO_PIPELINE_CONTINUE
);
3044 * ==========================================================================
3045 * Allocate and free blocks
3046 * ==========================================================================
3050 zio_io_to_allocate(spa_t
*spa
)
3054 ASSERT(MUTEX_HELD(&spa
->spa_alloc_lock
));
3056 zio
= avl_first(&spa
->spa_alloc_tree
);
3060 ASSERT(IO_IS_ALLOCATING(zio
));
3063 * Try to place a reservation for this zio. If we're unable to
3064 * reserve then we throttle.
3066 if (!metaslab_class_throttle_reserve(spa_normal_class(spa
),
3067 zio
->io_prop
.zp_copies
, zio
, 0)) {
3071 avl_remove(&spa
->spa_alloc_tree
, zio
);
3072 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3078 zio_dva_throttle(zio_t
*zio
)
3080 spa_t
*spa
= zio
->io_spa
;
3083 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3084 !spa_normal_class(zio
->io_spa
)->mc_alloc_throttle_enabled
||
3085 zio
->io_child_type
== ZIO_CHILD_GANG
||
3086 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3087 return (ZIO_PIPELINE_CONTINUE
);
3090 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3092 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3093 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3095 mutex_enter(&spa
->spa_alloc_lock
);
3097 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3098 avl_add(&spa
->spa_alloc_tree
, zio
);
3100 nio
= zio_io_to_allocate(zio
->io_spa
);
3101 mutex_exit(&spa
->spa_alloc_lock
);
3104 return (ZIO_PIPELINE_CONTINUE
);
3107 ASSERT(nio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3109 * We are passing control to a new zio so make sure that
3110 * it is processed by a different thread. We do this to
3111 * avoid stack overflows that can occur when parents are
3112 * throttled and children are making progress. We allow
3113 * it to go to the head of the taskq since it's already
3116 zio_taskq_dispatch(nio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3118 return (ZIO_PIPELINE_STOP
);
3122 zio_allocate_dispatch(spa_t
*spa
)
3126 mutex_enter(&spa
->spa_alloc_lock
);
3127 zio
= zio_io_to_allocate(spa
);
3128 mutex_exit(&spa
->spa_alloc_lock
);
3132 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3133 ASSERT0(zio
->io_error
);
3134 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3138 zio_dva_allocate(zio_t
*zio
)
3140 spa_t
*spa
= zio
->io_spa
;
3141 metaslab_class_t
*mc
= spa_normal_class(spa
);
3142 blkptr_t
*bp
= zio
->io_bp
;
3146 if (zio
->io_gang_leader
== NULL
) {
3147 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3148 zio
->io_gang_leader
= zio
;
3151 ASSERT(BP_IS_HOLE(bp
));
3152 ASSERT0(BP_GET_NDVAS(bp
));
3153 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3154 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3155 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3157 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3158 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3159 flags
|= METASLAB_DONT_THROTTLE
;
3160 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3161 flags
|= METASLAB_GANG_CHILD
;
3162 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3163 flags
|= METASLAB_ASYNC_ALLOC
;
3165 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3166 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3167 &zio
->io_alloc_list
, zio
);
3170 spa_dbgmsg(spa
, "%s: metaslab allocation failure: zio %p, "
3171 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3173 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3174 return (zio_write_gang_block(zio
));
3175 zio
->io_error
= error
;
3178 return (ZIO_PIPELINE_CONTINUE
);
3182 zio_dva_free(zio_t
*zio
)
3184 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3186 return (ZIO_PIPELINE_CONTINUE
);
3190 zio_dva_claim(zio_t
*zio
)
3194 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3196 zio
->io_error
= error
;
3198 return (ZIO_PIPELINE_CONTINUE
);
3202 * Undo an allocation. This is used by zio_done() when an I/O fails
3203 * and we want to give back the block we just allocated.
3204 * This handles both normal blocks and gang blocks.
3207 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3209 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3210 ASSERT(zio
->io_bp_override
== NULL
);
3212 if (!BP_IS_HOLE(bp
))
3213 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3216 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3217 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3218 &gn
->gn_gbh
->zg_blkptr
[g
]);
3224 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3227 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3228 uint64_t size
, boolean_t
*slog
)
3231 zio_alloc_list_t io_alloc_list
;
3233 ASSERT(txg
> spa_syncing_txg(spa
));
3235 metaslab_trace_init(&io_alloc_list
);
3236 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3237 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
);
3241 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3242 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3243 &io_alloc_list
, NULL
);
3247 metaslab_trace_fini(&io_alloc_list
);
3250 BP_SET_LSIZE(new_bp
, size
);
3251 BP_SET_PSIZE(new_bp
, size
);
3252 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3253 BP_SET_CHECKSUM(new_bp
,
3254 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3255 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3256 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3257 BP_SET_LEVEL(new_bp
, 0);
3258 BP_SET_DEDUP(new_bp
, 0);
3259 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3262 * encrypted blocks will require an IV and salt. We generate
3263 * these now since we will not be rewriting the bp at
3266 if (os
->os_encrypted
) {
3267 uint8_t iv
[ZIO_DATA_IV_LEN
];
3268 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3270 BP_SET_CRYPT(new_bp
, B_TRUE
);
3271 VERIFY0(spa_crypt_get_salt(spa
,
3272 dmu_objset_id(os
), salt
));
3273 VERIFY0(zio_crypt_generate_iv(iv
));
3275 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3283 * Free an intent log block.
3286 zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
)
3288 ASSERT(BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
);
3289 ASSERT(!BP_IS_GANG(bp
));
3291 zio_free(spa
, txg
, bp
);
3295 * ==========================================================================
3296 * Read and write to physical devices
3297 * ==========================================================================
3302 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3303 * stops after this stage and will resume upon I/O completion.
3304 * However, there are instances where the vdev layer may need to
3305 * continue the pipeline when an I/O was not issued. Since the I/O
3306 * that was sent to the vdev layer might be different than the one
3307 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3308 * force the underlying vdev layers to call either zio_execute() or
3309 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3312 zio_vdev_io_start(zio_t
*zio
)
3314 vdev_t
*vd
= zio
->io_vd
;
3316 spa_t
*spa
= zio
->io_spa
;
3320 ASSERT(zio
->io_error
== 0);
3321 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3324 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3325 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3328 * The mirror_ops handle multiple DVAs in a single BP.
3330 vdev_mirror_ops
.vdev_op_io_start(zio
);
3331 return (ZIO_PIPELINE_STOP
);
3334 ASSERT3P(zio
->io_logical
, !=, zio
);
3337 * We keep track of time-sensitive I/Os so that the scan thread
3338 * can quickly react to certain workloads. In particular, we care
3339 * about non-scrubbing, top-level reads and writes with the following
3341 * - synchronous writes of user data to non-slog devices
3342 * - any reads of user data
3343 * When these conditions are met, adjust the timestamp of spa_last_io
3344 * which allows the scan thread to adjust its workload accordingly.
3346 if (!(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) && zio
->io_bp
!= NULL
&&
3347 vd
== vd
->vdev_top
&& !vd
->vdev_islog
&&
3348 zio
->io_bookmark
.zb_objset
!= DMU_META_OBJSET
&&
3349 zio
->io_txg
!= spa_syncing_txg(spa
)) {
3350 uint64_t old
= spa
->spa_last_io
;
3351 uint64_t new = ddi_get_lbolt64();
3353 (void) atomic_cas_64(&spa
->spa_last_io
, old
, new);
3356 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3358 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3359 P2PHASE(zio
->io_size
, align
) != 0) {
3360 /* Transform logical writes to be a full physical block size. */
3361 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3362 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3363 ASSERT(vd
== vd
->vdev_top
);
3364 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3365 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3366 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3368 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3372 * If this is not a physical io, make sure that it is properly aligned
3373 * before proceeding.
3375 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3376 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3377 ASSERT0(P2PHASE(zio
->io_size
, align
));
3380 * For physical writes, we allow 512b aligned writes and assume
3381 * the device will perform a read-modify-write as necessary.
3383 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3384 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3387 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3390 * If this is a repair I/O, and there's no self-healing involved --
3391 * that is, we're just resilvering what we expect to resilver --
3392 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3393 * This prevents spurious resilvering with nested replication.
3394 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3395 * A is out of date, we'll read from C+D, then use the data to
3396 * resilver A+B -- but we don't actually want to resilver B, just A.
3397 * The top-level mirror has no way to know this, so instead we just
3398 * discard unnecessary repairs as we work our way down the vdev tree.
3399 * The same logic applies to any form of nested replication:
3400 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3402 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3403 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3404 zio
->io_txg
!= 0 && /* not a delegated i/o */
3405 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3406 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3407 zio_vdev_io_bypass(zio
);
3408 return (ZIO_PIPELINE_CONTINUE
);
3411 if (vd
->vdev_ops
->vdev_op_leaf
&&
3412 (zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
)) {
3414 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3415 return (ZIO_PIPELINE_CONTINUE
);
3417 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3418 return (ZIO_PIPELINE_STOP
);
3420 if (!vdev_accessible(vd
, zio
)) {
3421 zio
->io_error
= SET_ERROR(ENXIO
);
3423 return (ZIO_PIPELINE_STOP
);
3425 zio
->io_delay
= gethrtime();
3428 vd
->vdev_ops
->vdev_op_io_start(zio
);
3429 return (ZIO_PIPELINE_STOP
);
3433 zio_vdev_io_done(zio_t
*zio
)
3435 vdev_t
*vd
= zio
->io_vd
;
3436 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3437 boolean_t unexpected_error
= B_FALSE
;
3439 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3440 return (ZIO_PIPELINE_STOP
);
3442 ASSERT(zio
->io_type
== ZIO_TYPE_READ
|| zio
->io_type
== ZIO_TYPE_WRITE
);
3445 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3447 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3449 vdev_queue_io_done(zio
);
3451 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3452 vdev_cache_write(zio
);
3454 if (zio_injection_enabled
&& zio
->io_error
== 0)
3455 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3458 if (zio_injection_enabled
&& zio
->io_error
== 0)
3459 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3461 if (zio
->io_error
) {
3462 if (!vdev_accessible(vd
, zio
)) {
3463 zio
->io_error
= SET_ERROR(ENXIO
);
3465 unexpected_error
= B_TRUE
;
3470 ops
->vdev_op_io_done(zio
);
3472 if (unexpected_error
)
3473 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3475 return (ZIO_PIPELINE_CONTINUE
);
3479 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3480 * disk, and use that to finish the checksum ereport later.
3483 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3484 const abd_t
*good_buf
)
3486 /* no processing needed */
3487 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3492 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3494 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3496 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3498 zcr
->zcr_cbinfo
= zio
->io_size
;
3499 zcr
->zcr_cbdata
= abd
;
3500 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3501 zcr
->zcr_free
= zio_abd_free
;
3505 zio_vdev_io_assess(zio_t
*zio
)
3507 vdev_t
*vd
= zio
->io_vd
;
3509 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
))
3510 return (ZIO_PIPELINE_STOP
);
3512 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3513 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3515 if (zio
->io_vsd
!= NULL
) {
3516 zio
->io_vsd_ops
->vsd_free(zio
);
3520 if (zio_injection_enabled
&& zio
->io_error
== 0)
3521 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3524 * If the I/O failed, determine whether we should attempt to retry it.
3526 * On retry, we cut in line in the issue queue, since we don't want
3527 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3529 if (zio
->io_error
&& vd
== NULL
&&
3530 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3531 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3532 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3534 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3535 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3536 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3537 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3538 zio_requeue_io_start_cut_in_line
);
3539 return (ZIO_PIPELINE_STOP
);
3543 * If we got an error on a leaf device, convert it to ENXIO
3544 * if the device is not accessible at all.
3546 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3547 !vdev_accessible(vd
, zio
))
3548 zio
->io_error
= SET_ERROR(ENXIO
);
3551 * If we can't write to an interior vdev (mirror or RAID-Z),
3552 * set vdev_cant_write so that we stop trying to allocate from it.
3554 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3555 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3556 vd
->vdev_cant_write
= B_TRUE
;
3560 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3561 * attempts will ever succeed. In this case we set a persistent bit so
3562 * that we don't bother with it in the future.
3564 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3565 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3566 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3567 vd
->vdev_nowritecache
= B_TRUE
;
3570 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3572 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3573 zio
->io_physdone
!= NULL
) {
3574 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3575 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3576 zio
->io_physdone(zio
->io_logical
);
3579 return (ZIO_PIPELINE_CONTINUE
);
3583 zio_vdev_io_reissue(zio_t
*zio
)
3585 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3586 ASSERT(zio
->io_error
== 0);
3588 zio
->io_stage
>>= 1;
3592 zio_vdev_io_redone(zio_t
*zio
)
3594 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3596 zio
->io_stage
>>= 1;
3600 zio_vdev_io_bypass(zio_t
*zio
)
3602 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3603 ASSERT(zio
->io_error
== 0);
3605 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3606 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3610 * ==========================================================================
3611 * Encrypt and store encryption parameters
3612 * ==========================================================================
3617 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3618 * managing the storage of encryption parameters and passing them to the
3619 * lower-level encryption functions.
3622 zio_encrypt(zio_t
*zio
)
3624 zio_prop_t
*zp
= &zio
->io_prop
;
3625 spa_t
*spa
= zio
->io_spa
;
3626 blkptr_t
*bp
= zio
->io_bp
;
3627 uint64_t psize
= BP_GET_PSIZE(bp
);
3628 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3629 void *enc_buf
= NULL
;
3631 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3632 uint8_t iv
[ZIO_DATA_IV_LEN
];
3633 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3634 boolean_t no_crypt
= B_FALSE
;
3636 /* the root zio already encrypted the data */
3637 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3638 return (ZIO_PIPELINE_CONTINUE
);
3640 /* only ZIL blocks are re-encrypted on rewrite */
3641 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3642 return (ZIO_PIPELINE_CONTINUE
);
3644 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3645 BP_SET_CRYPT(bp
, B_FALSE
);
3646 return (ZIO_PIPELINE_CONTINUE
);
3649 /* if we are doing raw encryption set the provided encryption params */
3650 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3651 BP_SET_CRYPT(bp
, B_TRUE
);
3652 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3653 if (ot
!= DMU_OT_OBJSET
)
3654 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3655 if (DMU_OT_IS_ENCRYPTED(ot
))
3656 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3657 return (ZIO_PIPELINE_CONTINUE
);
3660 /* indirect blocks only maintain a cksum of the lower level MACs */
3661 if (BP_GET_LEVEL(bp
) > 0) {
3662 BP_SET_CRYPT(bp
, B_TRUE
);
3663 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3664 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3666 zio_crypt_encode_mac_bp(bp
, mac
);
3667 return (ZIO_PIPELINE_CONTINUE
);
3671 * Objset blocks are a special case since they have 2 256-bit MACs
3672 * embedded within them.
3674 if (ot
== DMU_OT_OBJSET
) {
3675 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
3676 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3677 BP_SET_CRYPT(bp
, B_TRUE
);
3678 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
,
3679 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, psize
,
3680 BP_SHOULD_BYTESWAP(bp
)));
3681 return (ZIO_PIPELINE_CONTINUE
);
3684 /* unencrypted object types are only authenticated with a MAC */
3685 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
3686 BP_SET_CRYPT(bp
, B_TRUE
);
3687 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
,
3688 zio
->io_bookmark
.zb_objset
, zio
->io_abd
, psize
, mac
));
3689 zio_crypt_encode_mac_bp(bp
, mac
);
3690 return (ZIO_PIPELINE_CONTINUE
);
3694 * Later passes of sync-to-convergence may decide to rewrite data
3695 * in place to avoid more disk reallocations. This presents a problem
3696 * for encryption because this consitutes rewriting the new data with
3697 * the same encryption key and IV. However, this only applies to blocks
3698 * in the MOS (particularly the spacemaps) and we do not encrypt the
3699 * MOS. We assert that the zio is allocating or an intent log write
3702 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
3703 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
3704 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
3705 ASSERT3U(psize
, !=, 0);
3707 enc_buf
= zio_buf_alloc(psize
);
3708 eabd
= abd_get_from_buf(enc_buf
, psize
);
3709 abd_take_ownership_of_buf(eabd
, B_TRUE
);
3712 * For an explanation of what encryption parameters are stored
3713 * where, see the block comment in zio_crypt.c.
3715 if (ot
== DMU_OT_INTENT_LOG
) {
3716 zio_crypt_decode_params_bp(bp
, salt
, iv
);
3718 BP_SET_CRYPT(bp
, B_TRUE
);
3721 /* Perform the encryption. This should not fail */
3722 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, zio
->io_bookmark
.zb_objset
, bp
,
3723 zio
->io_txg
, psize
, zio
->io_abd
, eabd
, iv
, mac
, salt
, &no_crypt
));
3725 /* encode encryption metadata into the bp */
3726 if (ot
== DMU_OT_INTENT_LOG
) {
3728 * ZIL blocks store the MAC in the embedded checksum, so the
3729 * transform must always be applied.
3731 zio_crypt_encode_mac_zil(enc_buf
, mac
);
3732 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3734 BP_SET_CRYPT(bp
, B_TRUE
);
3735 zio_crypt_encode_params_bp(bp
, salt
, iv
);
3736 zio_crypt_encode_mac_bp(bp
, mac
);
3739 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
3742 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3746 return (ZIO_PIPELINE_CONTINUE
);
3750 * ==========================================================================
3751 * Generate and verify checksums
3752 * ==========================================================================
3755 zio_checksum_generate(zio_t
*zio
)
3757 blkptr_t
*bp
= zio
->io_bp
;
3758 enum zio_checksum checksum
;
3762 * This is zio_write_phys().
3763 * We're either generating a label checksum, or none at all.
3765 checksum
= zio
->io_prop
.zp_checksum
;
3767 if (checksum
== ZIO_CHECKSUM_OFF
)
3768 return (ZIO_PIPELINE_CONTINUE
);
3770 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
3772 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
3773 ASSERT(!IO_IS_ALLOCATING(zio
));
3774 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
3776 checksum
= BP_GET_CHECKSUM(bp
);
3780 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
3782 return (ZIO_PIPELINE_CONTINUE
);
3786 zio_checksum_verify(zio_t
*zio
)
3788 zio_bad_cksum_t info
;
3789 blkptr_t
*bp
= zio
->io_bp
;
3792 ASSERT(zio
->io_vd
!= NULL
);
3796 * This is zio_read_phys().
3797 * We're either verifying a label checksum, or nothing at all.
3799 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
3800 return (ZIO_PIPELINE_CONTINUE
);
3802 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
3805 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
3806 zio
->io_error
= error
;
3807 if (error
== ECKSUM
&&
3808 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
3809 zfs_ereport_start_checksum(zio
->io_spa
,
3810 zio
->io_vd
, &zio
->io_bookmark
, zio
,
3811 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
3815 return (ZIO_PIPELINE_CONTINUE
);
3819 * Called by RAID-Z to ensure we don't compute the checksum twice.
3822 zio_checksum_verified(zio_t
*zio
)
3824 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
3828 * ==========================================================================
3829 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3830 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3831 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3832 * indicate errors that are specific to one I/O, and most likely permanent.
3833 * Any other error is presumed to be worse because we weren't expecting it.
3834 * ==========================================================================
3837 zio_worst_error(int e1
, int e2
)
3839 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
3842 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
3843 if (e1
== zio_error_rank
[r1
])
3846 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
3847 if (e2
== zio_error_rank
[r2
])
3850 return (r1
> r2
? e1
: e2
);
3854 * ==========================================================================
3856 * ==========================================================================
3859 zio_ready(zio_t
*zio
)
3861 blkptr_t
*bp
= zio
->io_bp
;
3862 zio_t
*pio
, *pio_next
;
3863 zio_link_t
*zl
= NULL
;
3865 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_READY
) ||
3866 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_READY
))
3867 return (ZIO_PIPELINE_STOP
);
3869 if (zio
->io_ready
) {
3870 ASSERT(IO_IS_ALLOCATING(zio
));
3871 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
3872 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
3873 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
3878 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
3879 zio
->io_bp_copy
= *bp
;
3881 if (zio
->io_error
!= 0) {
3882 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3884 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3885 ASSERT(IO_IS_ALLOCATING(zio
));
3886 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3888 * We were unable to allocate anything, unreserve and
3889 * issue the next I/O to allocate.
3891 metaslab_class_throttle_unreserve(
3892 spa_normal_class(zio
->io_spa
),
3893 zio
->io_prop
.zp_copies
, zio
);
3894 zio_allocate_dispatch(zio
->io_spa
);
3898 mutex_enter(&zio
->io_lock
);
3899 zio
->io_state
[ZIO_WAIT_READY
] = 1;
3900 pio
= zio_walk_parents(zio
, &zl
);
3901 mutex_exit(&zio
->io_lock
);
3904 * As we notify zio's parents, new parents could be added.
3905 * New parents go to the head of zio's io_parent_list, however,
3906 * so we will (correctly) not notify them. The remainder of zio's
3907 * io_parent_list, from 'pio_next' onward, cannot change because
3908 * all parents must wait for us to be done before they can be done.
3910 for (; pio
!= NULL
; pio
= pio_next
) {
3911 pio_next
= zio_walk_parents(zio
, &zl
);
3912 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
);
3915 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
3916 if (BP_IS_GANG(bp
)) {
3917 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
3919 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
3920 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
3924 if (zio_injection_enabled
&&
3925 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
3926 zio_handle_ignored_writes(zio
);
3928 return (ZIO_PIPELINE_CONTINUE
);
3932 * Update the allocation throttle accounting.
3935 zio_dva_throttle_done(zio_t
*zio
)
3937 ASSERTV(zio_t
*lio
= zio
->io_logical
);
3938 zio_t
*pio
= zio_unique_parent(zio
);
3939 vdev_t
*vd
= zio
->io_vd
;
3940 int flags
= METASLAB_ASYNC_ALLOC
;
3942 ASSERT3P(zio
->io_bp
, !=, NULL
);
3943 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
3944 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
3945 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
3947 ASSERT3P(vd
, ==, vd
->vdev_top
);
3948 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
3949 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3950 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
3951 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3952 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
3955 * Parents of gang children can have two flavors -- ones that
3956 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3957 * and ones that allocated the constituent blocks. The allocation
3958 * throttle needs to know the allocating parent zio so we must find
3961 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
3963 * If our parent is a rewrite gang child then our grandparent
3964 * would have been the one that performed the allocation.
3966 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
3967 pio
= zio_unique_parent(pio
);
3968 flags
|= METASLAB_GANG_CHILD
;
3971 ASSERT(IO_IS_ALLOCATING(pio
));
3972 ASSERT3P(zio
, !=, zio
->io_logical
);
3973 ASSERT(zio
->io_logical
!= NULL
);
3974 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
3975 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
3977 mutex_enter(&pio
->io_lock
);
3978 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
);
3979 mutex_exit(&pio
->io_lock
);
3981 metaslab_class_throttle_unreserve(spa_normal_class(zio
->io_spa
),
3985 * Call into the pipeline to see if there is more work that
3986 * needs to be done. If there is work to be done it will be
3987 * dispatched to another taskq thread.
3989 zio_allocate_dispatch(zio
->io_spa
);
3993 zio_done(zio_t
*zio
)
3996 * Always attempt to keep stack usage minimal here since
3997 * we can be called recurisvely up to 19 levels deep.
3999 const uint64_t psize
= zio
->io_size
;
4000 zio_t
*pio
, *pio_next
;
4001 zio_link_t
*zl
= NULL
;
4004 * If our children haven't all completed,
4005 * wait for them and then repeat this pipeline stage.
4007 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV
, ZIO_WAIT_DONE
) ||
4008 zio_wait_for_children(zio
, ZIO_CHILD_GANG
, ZIO_WAIT_DONE
) ||
4009 zio_wait_for_children(zio
, ZIO_CHILD_DDT
, ZIO_WAIT_DONE
) ||
4010 zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL
, ZIO_WAIT_DONE
))
4011 return (ZIO_PIPELINE_STOP
);
4014 * If the allocation throttle is enabled, then update the accounting.
4015 * We only track child I/Os that are part of an allocating async
4016 * write. We must do this since the allocation is performed
4017 * by the logical I/O but the actual write is done by child I/Os.
4019 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4020 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4021 ASSERT(spa_normal_class(
4022 zio
->io_spa
)->mc_alloc_throttle_enabled
);
4023 zio_dva_throttle_done(zio
);
4027 * If the allocation throttle is enabled, verify that
4028 * we have decremented the refcounts for every I/O that was throttled.
4030 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4031 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4032 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4033 ASSERT(zio
->io_bp
!= NULL
);
4034 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
);
4035 VERIFY(refcount_not_held(
4036 &(spa_normal_class(zio
->io_spa
)->mc_alloc_slots
), zio
));
4040 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4041 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4042 ASSERT(zio
->io_children
[c
][w
] == 0);
4044 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4045 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4046 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4047 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4048 sizeof (blkptr_t
)) == 0 ||
4049 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4050 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4051 zio
->io_bp_override
== NULL
&&
4052 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4053 ASSERT(!BP_SHOULD_BYTESWAP(zio
->io_bp
));
4054 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4055 BP_GET_NDVAS(zio
->io_bp
));
4056 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4057 (BP_COUNT_GANG(zio
->io_bp
) ==
4058 BP_GET_NDVAS(zio
->io_bp
)));
4060 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4061 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4065 * If there were child vdev/gang/ddt errors, they apply to us now.
4067 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4068 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4069 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4072 * If the I/O on the transformed data was successful, generate any
4073 * checksum reports now while we still have the transformed data.
4075 if (zio
->io_error
== 0) {
4076 while (zio
->io_cksum_report
!= NULL
) {
4077 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4078 uint64_t align
= zcr
->zcr_align
;
4079 uint64_t asize
= P2ROUNDUP(psize
, align
);
4080 abd_t
*adata
= zio
->io_abd
;
4082 if (asize
!= psize
) {
4083 adata
= abd_alloc(asize
, B_TRUE
);
4084 abd_copy(adata
, zio
->io_abd
, psize
);
4085 abd_zero_off(adata
, psize
, asize
- psize
);
4088 zio
->io_cksum_report
= zcr
->zcr_next
;
4089 zcr
->zcr_next
= NULL
;
4090 zcr
->zcr_finish(zcr
, adata
);
4091 zfs_ereport_free_checksum(zcr
);
4098 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4100 vdev_stat_update(zio
, psize
);
4103 * If this I/O is attached to a particular vdev is slow, exceeding
4104 * 30 seconds to complete, post an error described the I/O delay.
4105 * We ignore these errors if the device is currently unavailable.
4107 if (zio
->io_delay
>= MSEC2NSEC(zio_delay_max
)) {
4108 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
))
4109 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
, zio
->io_spa
,
4110 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4113 if (zio
->io_error
) {
4115 * If this I/O is attached to a particular vdev,
4116 * generate an error message describing the I/O failure
4117 * at the block level. We ignore these errors if the
4118 * device is currently unavailable.
4120 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4121 !vdev_is_dead(zio
->io_vd
))
4122 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4123 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4125 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4126 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4127 zio
== zio
->io_logical
) {
4129 * For logical I/O requests, tell the SPA to log the
4130 * error and generate a logical data ereport.
4132 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4133 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4134 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4138 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4140 * Determine whether zio should be reexecuted. This will
4141 * propagate all the way to the root via zio_notify_parent().
4143 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4144 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4146 if (IO_IS_ALLOCATING(zio
) &&
4147 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4148 if (zio
->io_error
!= ENOSPC
)
4149 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4151 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4154 if ((zio
->io_type
== ZIO_TYPE_READ
||
4155 zio
->io_type
== ZIO_TYPE_FREE
) &&
4156 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4157 zio
->io_error
== ENXIO
&&
4158 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4159 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4160 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4162 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4163 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4166 * Here is a possibly good place to attempt to do
4167 * either combinatorial reconstruction or error correction
4168 * based on checksums. It also might be a good place
4169 * to send out preliminary ereports before we suspend
4175 * If there were logical child errors, they apply to us now.
4176 * We defer this until now to avoid conflating logical child
4177 * errors with errors that happened to the zio itself when
4178 * updating vdev stats and reporting FMA events above.
4180 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4182 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4183 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4184 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4185 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4187 zio_gang_tree_free(&zio
->io_gang_tree
);
4190 * Godfather I/Os should never suspend.
4192 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4193 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4194 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4196 if (zio
->io_reexecute
) {
4198 * This is a logical I/O that wants to reexecute.
4200 * Reexecute is top-down. When an i/o fails, if it's not
4201 * the root, it simply notifies its parent and sticks around.
4202 * The parent, seeing that it still has children in zio_done(),
4203 * does the same. This percolates all the way up to the root.
4204 * The root i/o will reexecute or suspend the entire tree.
4206 * This approach ensures that zio_reexecute() honors
4207 * all the original i/o dependency relationships, e.g.
4208 * parents not executing until children are ready.
4210 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4212 zio
->io_gang_leader
= NULL
;
4214 mutex_enter(&zio
->io_lock
);
4215 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4216 mutex_exit(&zio
->io_lock
);
4219 * "The Godfather" I/O monitors its children but is
4220 * not a true parent to them. It will track them through
4221 * the pipeline but severs its ties whenever they get into
4222 * trouble (e.g. suspended). This allows "The Godfather"
4223 * I/O to return status without blocking.
4226 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4228 zio_link_t
*remove_zl
= zl
;
4229 pio_next
= zio_walk_parents(zio
, &zl
);
4231 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4232 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4233 zio_remove_child(pio
, zio
, remove_zl
);
4234 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4238 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4240 * We're not a root i/o, so there's nothing to do
4241 * but notify our parent. Don't propagate errors
4242 * upward since we haven't permanently failed yet.
4244 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4245 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4246 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4247 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4249 * We'd fail again if we reexecuted now, so suspend
4250 * until conditions improve (e.g. device comes online).
4252 zio_suspend(zio
->io_spa
, zio
);
4255 * Reexecution is potentially a huge amount of work.
4256 * Hand it off to the otherwise-unused claim taskq.
4258 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4259 spa_taskq_dispatch_ent(zio
->io_spa
,
4260 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4261 (task_func_t
*)zio_reexecute
, zio
, 0,
4264 return (ZIO_PIPELINE_STOP
);
4267 ASSERT(zio
->io_child_count
== 0);
4268 ASSERT(zio
->io_reexecute
== 0);
4269 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4272 * Report any checksum errors, since the I/O is complete.
4274 while (zio
->io_cksum_report
!= NULL
) {
4275 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4276 zio
->io_cksum_report
= zcr
->zcr_next
;
4277 zcr
->zcr_next
= NULL
;
4278 zcr
->zcr_finish(zcr
, NULL
);
4279 zfs_ereport_free_checksum(zcr
);
4282 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4283 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4284 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4285 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4289 * It is the responsibility of the done callback to ensure that this
4290 * particular zio is no longer discoverable for adoption, and as
4291 * such, cannot acquire any new parents.
4296 mutex_enter(&zio
->io_lock
);
4297 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4298 mutex_exit(&zio
->io_lock
);
4301 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4302 zio_link_t
*remove_zl
= zl
;
4303 pio_next
= zio_walk_parents(zio
, &zl
);
4304 zio_remove_child(pio
, zio
, remove_zl
);
4305 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
);
4308 if (zio
->io_waiter
!= NULL
) {
4309 mutex_enter(&zio
->io_lock
);
4310 zio
->io_executor
= NULL
;
4311 cv_broadcast(&zio
->io_cv
);
4312 mutex_exit(&zio
->io_lock
);
4317 return (ZIO_PIPELINE_STOP
);
4321 * ==========================================================================
4322 * I/O pipeline definition
4323 * ==========================================================================
4325 static zio_pipe_stage_t
*zio_pipeline
[] = {
4333 zio_checksum_generate
,
4349 zio_checksum_verify
,
4357 * Compare two zbookmark_phys_t's to see which we would reach first in a
4358 * pre-order traversal of the object tree.
4360 * This is simple in every case aside from the meta-dnode object. For all other
4361 * objects, we traverse them in order (object 1 before object 2, and so on).
4362 * However, all of these objects are traversed while traversing object 0, since
4363 * the data it points to is the list of objects. Thus, we need to convert to a
4364 * canonical representation so we can compare meta-dnode bookmarks to
4365 * non-meta-dnode bookmarks.
4367 * We do this by calculating "equivalents" for each field of the zbookmark.
4368 * zbookmarks outside of the meta-dnode use their own object and level, and
4369 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4370 * blocks this bookmark refers to) by multiplying their blkid by their span
4371 * (the number of L0 blocks contained within one block at their level).
4372 * zbookmarks inside the meta-dnode calculate their object equivalent
4373 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4374 * level + 1<<31 (any value larger than a level could ever be) for their level.
4375 * This causes them to always compare before a bookmark in their object
4376 * equivalent, compare appropriately to bookmarks in other objects, and to
4377 * compare appropriately to other bookmarks in the meta-dnode.
4380 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4381 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4384 * These variables represent the "equivalent" values for the zbookmark,
4385 * after converting zbookmarks inside the meta dnode to their
4386 * normal-object equivalents.
4388 uint64_t zb1obj
, zb2obj
;
4389 uint64_t zb1L0
, zb2L0
;
4390 uint64_t zb1level
, zb2level
;
4392 if (zb1
->zb_object
== zb2
->zb_object
&&
4393 zb1
->zb_level
== zb2
->zb_level
&&
4394 zb1
->zb_blkid
== zb2
->zb_blkid
)
4398 * BP_SPANB calculates the span in blocks.
4400 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4401 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4403 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4404 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4406 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4408 zb1obj
= zb1
->zb_object
;
4409 zb1level
= zb1
->zb_level
;
4412 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4413 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4415 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4417 zb2obj
= zb2
->zb_object
;
4418 zb2level
= zb2
->zb_level
;
4421 /* Now that we have a canonical representation, do the comparison. */
4422 if (zb1obj
!= zb2obj
)
4423 return (zb1obj
< zb2obj
? -1 : 1);
4424 else if (zb1L0
!= zb2L0
)
4425 return (zb1L0
< zb2L0
? -1 : 1);
4426 else if (zb1level
!= zb2level
)
4427 return (zb1level
> zb2level
? -1 : 1);
4429 * This can (theoretically) happen if the bookmarks have the same object
4430 * and level, but different blkids, if the block sizes are not the same.
4431 * There is presently no way to change the indirect block sizes
4437 * This function checks the following: given that last_block is the place that
4438 * our traversal stopped last time, does that guarantee that we've visited
4439 * every node under subtree_root? Therefore, we can't just use the raw output
4440 * of zbookmark_compare. We have to pass in a modified version of
4441 * subtree_root; by incrementing the block id, and then checking whether
4442 * last_block is before or equal to that, we can tell whether or not having
4443 * visited last_block implies that all of subtree_root's children have been
4447 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4448 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4450 zbookmark_phys_t mod_zb
= *subtree_root
;
4452 ASSERT(last_block
->zb_level
== 0);
4454 /* The objset_phys_t isn't before anything. */
4459 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4460 * data block size in sectors, because that variable is only used if
4461 * the bookmark refers to a block in the meta-dnode. Since we don't
4462 * know without examining it what object it refers to, and there's no
4463 * harm in passing in this value in other cases, we always pass it in.
4465 * We pass in 0 for the indirect block size shift because zb2 must be
4466 * level 0. The indirect block size is only used to calculate the span
4467 * of the bookmark, but since the bookmark must be level 0, the span is
4468 * always 1, so the math works out.
4470 * If you make changes to how the zbookmark_compare code works, be sure
4471 * to make sure that this code still works afterwards.
4473 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4474 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4478 #if defined(_KERNEL) && defined(HAVE_SPL)
4479 EXPORT_SYMBOL(zio_type_name
);
4480 EXPORT_SYMBOL(zio_buf_alloc
);
4481 EXPORT_SYMBOL(zio_data_buf_alloc
);
4482 EXPORT_SYMBOL(zio_buf_free
);
4483 EXPORT_SYMBOL(zio_data_buf_free
);
4485 module_param(zio_delay_max
, int, 0644);
4486 MODULE_PARM_DESC(zio_delay_max
, "Max zio millisec delay before posting event");
4488 module_param(zio_requeue_io_start_cut_in_line
, int, 0644);
4489 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line
, "Prioritize requeued I/O");
4491 module_param(zfs_sync_pass_deferred_free
, int, 0644);
4492 MODULE_PARM_DESC(zfs_sync_pass_deferred_free
,
4493 "Defer frees starting in this pass");
4495 module_param(zfs_sync_pass_dont_compress
, int, 0644);
4496 MODULE_PARM_DESC(zfs_sync_pass_dont_compress
,
4497 "Don't compress starting in this pass");
4499 module_param(zfs_sync_pass_rewrite
, int, 0644);
4500 MODULE_PARM_DESC(zfs_sync_pass_rewrite
,
4501 "Rewrite new bps starting in this pass");
4503 module_param(zio_dva_throttle_enabled
, int, 0644);
4504 MODULE_PARM_DESC(zio_dva_throttle_enabled
,
4505 "Throttle block allocations in the ZIO pipeline");