4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_trim.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/dsl_scan.h>
45 #include <sys/metaslab_impl.h>
47 #include <sys/trace_zfs.h>
49 #include <sys/dsl_crypt.h>
50 #include <sys/cityhash.h>
53 * ==========================================================================
54 * I/O type descriptions
55 * ==========================================================================
57 const char *zio_type_name
[ZIO_TYPES
] = {
59 * Note: Linux kernel thread name length is limited
60 * so these names will differ from upstream open zfs.
62 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
65 int zio_dva_throttle_enabled
= B_TRUE
;
66 int zio_deadman_log_all
= B_FALSE
;
69 * ==========================================================================
71 * ==========================================================================
73 kmem_cache_t
*zio_cache
;
74 kmem_cache_t
*zio_link_cache
;
75 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
76 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
77 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
78 uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
79 uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
82 /* Mark IOs as "slow" if they take longer than 30 seconds */
83 int zio_slow_io_ms
= (30 * MILLISEC
);
85 #define BP_SPANB(indblkshift, level) \
86 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
87 #define COMPARE_META_LEVEL 0x80000000ul
89 * The following actions directly effect the spa's sync-to-convergence logic.
90 * The values below define the sync pass when we start performing the action.
91 * Care should be taken when changing these values as they directly impact
92 * spa_sync() performance. Tuning these values may introduce subtle performance
93 * pathologies and should only be done in the context of performance analysis.
94 * These tunables will eventually be removed and replaced with #defines once
95 * enough analysis has been done to determine optimal values.
97 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
98 * regular blocks are not deferred.
100 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
101 * compression (including of metadata). In practice, we don't have this
102 * many sync passes, so this has no effect.
104 * The original intent was that disabling compression would help the sync
105 * passes to converge. However, in practice disabling compression increases
106 * the average number of sync passes, because when we turn compression off, a
107 * lot of block's size will change and thus we have to re-allocate (not
108 * overwrite) them. It also increases the number of 128KB allocations (e.g.
109 * for indirect blocks and spacemaps) because these will not be compressed.
110 * The 128K allocations are especially detrimental to performance on highly
111 * fragmented systems, which may have very few free segments of this size,
112 * and may need to load new metaslabs to satisfy 128K allocations.
114 int zfs_sync_pass_deferred_free
= 2; /* defer frees starting in this pass */
115 int zfs_sync_pass_dont_compress
= 8; /* don't compress starting in this pass */
116 int zfs_sync_pass_rewrite
= 2; /* rewrite new bps starting in this pass */
119 * An allocating zio is one that either currently has the DVA allocate
120 * stage set or will have it later in its lifetime.
122 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
124 int zio_requeue_io_start_cut_in_line
= 1;
127 int zio_buf_debug_limit
= 16384;
129 int zio_buf_debug_limit
= 0;
132 static inline void __zio_execute(zio_t
*zio
);
134 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
140 vmem_t
*data_alloc_arena
= NULL
;
142 zio_cache
= kmem_cache_create("zio_cache",
143 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
144 zio_link_cache
= kmem_cache_create("zio_link_cache",
145 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
148 * For small buffers, we want a cache for each multiple of
149 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
150 * for each quarter-power of 2.
152 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
153 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
156 size_t cflags
= (size
> zio_buf_debug_limit
) ? KMC_NODEBUG
: 0;
158 #if defined(_ILP32) && defined(_KERNEL)
160 * Cache size limited to 1M on 32-bit platforms until ARC
161 * buffers no longer require virtual address space.
163 if (size
> zfs_max_recordsize
)
172 * If we are using watchpoints, put each buffer on its own page,
173 * to eliminate the performance overhead of trapping to the
174 * kernel when modifying a non-watched buffer that shares the
175 * page with a watched buffer.
177 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
180 * Here's the problem - on 4K native devices in userland on
181 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
182 * will fail with EINVAL, causing zdb (and others) to coredump.
183 * Since userland probably doesn't need optimized buffer caches,
184 * we just force 4K alignment on everything.
186 align
= 8 * SPA_MINBLOCKSIZE
;
188 if (size
< PAGESIZE
) {
189 align
= SPA_MINBLOCKSIZE
;
190 } else if (IS_P2ALIGNED(size
, p2
>> 2)) {
197 (void) sprintf(name
, "zio_buf_%lu", (ulong_t
)size
);
198 zio_buf_cache
[c
] = kmem_cache_create(name
, size
,
199 align
, NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
201 (void) sprintf(name
, "zio_data_buf_%lu", (ulong_t
)size
);
202 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
,
203 align
, NULL
, NULL
, NULL
, NULL
,
204 data_alloc_arena
, cflags
);
209 ASSERT(zio_buf_cache
[c
] != NULL
);
210 if (zio_buf_cache
[c
- 1] == NULL
)
211 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
213 ASSERT(zio_data_buf_cache
[c
] != NULL
);
214 if (zio_data_buf_cache
[c
- 1] == NULL
)
215 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
227 kmem_cache_t
*last_cache
= NULL
;
228 kmem_cache_t
*last_data_cache
= NULL
;
230 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
233 * Cache size limited to 1M on 32-bit platforms until ARC
234 * buffers no longer require virtual address space.
236 if (((c
+ 1) << SPA_MINBLOCKSHIFT
) > zfs_max_recordsize
)
239 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
240 if (zio_buf_cache_allocs
[c
] != zio_buf_cache_frees
[c
])
241 (void) printf("zio_fini: [%d] %llu != %llu\n",
242 (int)((c
+ 1) << SPA_MINBLOCKSHIFT
),
243 (long long unsigned)zio_buf_cache_allocs
[c
],
244 (long long unsigned)zio_buf_cache_frees
[c
]);
246 if (zio_buf_cache
[c
] != last_cache
) {
247 last_cache
= zio_buf_cache
[c
];
248 kmem_cache_destroy(zio_buf_cache
[c
]);
250 zio_buf_cache
[c
] = NULL
;
252 if (zio_data_buf_cache
[c
] != last_data_cache
) {
253 last_data_cache
= zio_data_buf_cache
[c
];
254 kmem_cache_destroy(zio_data_buf_cache
[c
]);
256 zio_data_buf_cache
[c
] = NULL
;
259 kmem_cache_destroy(zio_link_cache
);
260 kmem_cache_destroy(zio_cache
);
268 * ==========================================================================
269 * Allocate and free I/O buffers
270 * ==========================================================================
274 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
275 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
276 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
277 * excess / transient data in-core during a crashdump.
280 zio_buf_alloc(size_t size
)
282 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
284 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
285 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
286 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
289 return (kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
));
293 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
294 * crashdump if the kernel panics. This exists so that we will limit the amount
295 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
296 * of kernel heap dumped to disk when the kernel panics)
299 zio_data_buf_alloc(size_t size
)
301 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
303 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
305 return (kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
));
309 zio_buf_free(void *buf
, size_t size
)
311 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
313 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
314 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
315 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
318 kmem_cache_free(zio_buf_cache
[c
], buf
);
322 zio_data_buf_free(void *buf
, size_t size
)
324 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
326 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
328 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
332 zio_abd_free(void *abd
, size_t size
)
334 abd_free((abd_t
*)abd
);
338 * ==========================================================================
339 * Push and pop I/O transform buffers
340 * ==========================================================================
343 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
344 zio_transform_func_t
*transform
)
346 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
348 zt
->zt_orig_abd
= zio
->io_abd
;
349 zt
->zt_orig_size
= zio
->io_size
;
350 zt
->zt_bufsize
= bufsize
;
351 zt
->zt_transform
= transform
;
353 zt
->zt_next
= zio
->io_transform_stack
;
354 zio
->io_transform_stack
= zt
;
361 zio_pop_transforms(zio_t
*zio
)
365 while ((zt
= zio
->io_transform_stack
) != NULL
) {
366 if (zt
->zt_transform
!= NULL
)
367 zt
->zt_transform(zio
,
368 zt
->zt_orig_abd
, zt
->zt_orig_size
);
370 if (zt
->zt_bufsize
!= 0)
371 abd_free(zio
->io_abd
);
373 zio
->io_abd
= zt
->zt_orig_abd
;
374 zio
->io_size
= zt
->zt_orig_size
;
375 zio
->io_transform_stack
= zt
->zt_next
;
377 kmem_free(zt
, sizeof (zio_transform_t
));
382 * ==========================================================================
383 * I/O transform callbacks for subblocks, decompression, and decryption
384 * ==========================================================================
387 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
389 ASSERT(zio
->io_size
> size
);
391 if (zio
->io_type
== ZIO_TYPE_READ
)
392 abd_copy(data
, zio
->io_abd
, size
);
396 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
398 if (zio
->io_error
== 0) {
399 void *tmp
= abd_borrow_buf(data
, size
);
400 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
401 zio
->io_abd
, tmp
, zio
->io_size
, size
);
402 abd_return_buf_copy(data
, tmp
, size
);
404 if (zio_injection_enabled
&& ret
== 0)
405 ret
= zio_handle_fault_injection(zio
, EINVAL
);
408 zio
->io_error
= SET_ERROR(EIO
);
413 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
417 blkptr_t
*bp
= zio
->io_bp
;
418 spa_t
*spa
= zio
->io_spa
;
419 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
420 uint64_t lsize
= BP_GET_LSIZE(bp
);
421 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
422 uint8_t salt
[ZIO_DATA_SALT_LEN
];
423 uint8_t iv
[ZIO_DATA_IV_LEN
];
424 uint8_t mac
[ZIO_DATA_MAC_LEN
];
425 boolean_t no_crypt
= B_FALSE
;
427 ASSERT(BP_USES_CRYPT(bp
));
428 ASSERT3U(size
, !=, 0);
430 if (zio
->io_error
!= 0)
434 * Verify the cksum of MACs stored in an indirect bp. It will always
435 * be possible to verify this since it does not require an encryption
438 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
439 zio_crypt_decode_mac_bp(bp
, mac
);
441 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
443 * We haven't decompressed the data yet, but
444 * zio_crypt_do_indirect_mac_checksum() requires
445 * decompressed data to be able to parse out the MACs
446 * from the indirect block. We decompress it now and
447 * throw away the result after we are finished.
449 tmp
= zio_buf_alloc(lsize
);
450 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
451 zio
->io_abd
, tmp
, zio
->io_size
, lsize
);
453 ret
= SET_ERROR(EIO
);
456 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
457 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
458 zio_buf_free(tmp
, lsize
);
460 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
461 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
463 abd_copy(data
, zio
->io_abd
, size
);
465 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
466 ret
= zio_handle_decrypt_injection(spa
,
467 &zio
->io_bookmark
, ot
, ECKSUM
);
476 * If this is an authenticated block, just check the MAC. It would be
477 * nice to separate this out into its own flag, but for the moment
478 * enum zio_flag is out of bits.
480 if (BP_IS_AUTHENTICATED(bp
)) {
481 if (ot
== DMU_OT_OBJSET
) {
482 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
483 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
485 zio_crypt_decode_mac_bp(bp
, mac
);
486 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
487 zio
->io_abd
, size
, mac
);
488 if (zio_injection_enabled
&& ret
== 0) {
489 ret
= zio_handle_decrypt_injection(spa
,
490 &zio
->io_bookmark
, ot
, ECKSUM
);
493 abd_copy(data
, zio
->io_abd
, size
);
501 zio_crypt_decode_params_bp(bp
, salt
, iv
);
503 if (ot
== DMU_OT_INTENT_LOG
) {
504 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
505 zio_crypt_decode_mac_zil(tmp
, mac
);
506 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
508 zio_crypt_decode_mac_bp(bp
, mac
);
511 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
512 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
513 zio
->io_abd
, &no_crypt
);
515 abd_copy(data
, zio
->io_abd
, size
);
523 /* assert that the key was found unless this was speculative */
524 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
527 * If there was a decryption / authentication error return EIO as
528 * the io_error. If this was not a speculative zio, create an ereport.
531 zio
->io_error
= SET_ERROR(EIO
);
532 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
533 spa_log_error(spa
, &zio
->io_bookmark
);
534 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
535 spa
, NULL
, &zio
->io_bookmark
, zio
, 0, 0);
543 * ==========================================================================
544 * I/O parent/child relationships and pipeline interlocks
545 * ==========================================================================
548 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
550 list_t
*pl
= &cio
->io_parent_list
;
552 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
556 ASSERT((*zl
)->zl_child
== cio
);
557 return ((*zl
)->zl_parent
);
561 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
563 list_t
*cl
= &pio
->io_child_list
;
565 ASSERT(MUTEX_HELD(&pio
->io_lock
));
567 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
571 ASSERT((*zl
)->zl_parent
== pio
);
572 return ((*zl
)->zl_child
);
576 zio_unique_parent(zio_t
*cio
)
578 zio_link_t
*zl
= NULL
;
579 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
581 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
586 zio_add_child(zio_t
*pio
, zio_t
*cio
)
588 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
591 * Logical I/Os can have logical, gang, or vdev children.
592 * Gang I/Os can have gang or vdev children.
593 * Vdev I/Os can only have vdev children.
594 * The following ASSERT captures all of these constraints.
596 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
601 mutex_enter(&pio
->io_lock
);
602 mutex_enter(&cio
->io_lock
);
604 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
606 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
607 pio
->io_children
[cio
->io_child_type
][w
] += !cio
->io_state
[w
];
609 list_insert_head(&pio
->io_child_list
, zl
);
610 list_insert_head(&cio
->io_parent_list
, zl
);
612 pio
->io_child_count
++;
613 cio
->io_parent_count
++;
615 mutex_exit(&cio
->io_lock
);
616 mutex_exit(&pio
->io_lock
);
620 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
622 ASSERT(zl
->zl_parent
== pio
);
623 ASSERT(zl
->zl_child
== cio
);
625 mutex_enter(&pio
->io_lock
);
626 mutex_enter(&cio
->io_lock
);
628 list_remove(&pio
->io_child_list
, zl
);
629 list_remove(&cio
->io_parent_list
, zl
);
631 pio
->io_child_count
--;
632 cio
->io_parent_count
--;
634 mutex_exit(&cio
->io_lock
);
635 mutex_exit(&pio
->io_lock
);
636 kmem_cache_free(zio_link_cache
, zl
);
640 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
642 boolean_t waiting
= B_FALSE
;
644 mutex_enter(&zio
->io_lock
);
645 ASSERT(zio
->io_stall
== NULL
);
646 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
647 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
650 uint64_t *countp
= &zio
->io_children
[c
][wait
];
653 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
654 zio
->io_stall
= countp
;
659 mutex_exit(&zio
->io_lock
);
663 __attribute__((always_inline
))
665 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
666 zio_t
**next_to_executep
)
668 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
669 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
671 mutex_enter(&pio
->io_lock
);
672 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
673 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
674 pio
->io_reexecute
|= zio
->io_reexecute
;
675 ASSERT3U(*countp
, >, 0);
679 if (*countp
== 0 && pio
->io_stall
== countp
) {
680 zio_taskq_type_t type
=
681 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
683 pio
->io_stall
= NULL
;
684 mutex_exit(&pio
->io_lock
);
687 * If we can tell the caller to execute this parent next, do
688 * so. Otherwise dispatch the parent zio as its own task.
690 * Having the caller execute the parent when possible reduces
691 * locking on the zio taskq's, reduces context switch
692 * overhead, and has no recursion penalty. Note that one
693 * read from disk typically causes at least 3 zio's: a
694 * zio_null(), the logical zio_read(), and then a physical
695 * zio. When the physical ZIO completes, we are able to call
696 * zio_done() on all 3 of these zio's from one invocation of
697 * zio_execute() by returning the parent back to
698 * zio_execute(). Since the parent isn't executed until this
699 * thread returns back to zio_execute(), the caller should do
702 * In other cases, dispatching the parent prevents
703 * overflowing the stack when we have deeply nested
704 * parent-child relationships, as we do with the "mega zio"
705 * of writes for spa_sync(), and the chain of ZIL blocks.
707 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
) {
708 *next_to_executep
= pio
;
710 zio_taskq_dispatch(pio
, type
, B_FALSE
);
713 mutex_exit(&pio
->io_lock
);
718 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
720 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
721 zio
->io_error
= zio
->io_child_error
[c
];
725 zio_bookmark_compare(const void *x1
, const void *x2
)
727 const zio_t
*z1
= x1
;
728 const zio_t
*z2
= x2
;
730 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
732 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
735 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
737 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
740 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
742 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
745 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
747 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
759 * ==========================================================================
760 * Create the various types of I/O (read, write, free, etc)
761 * ==========================================================================
764 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
765 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
766 void *private, zio_type_t type
, zio_priority_t priority
,
767 enum zio_flag flags
, vdev_t
*vd
, uint64_t offset
,
768 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
769 enum zio_stage pipeline
)
773 IMPLY(type
!= ZIO_TYPE_TRIM
, psize
<= SPA_MAXBLOCKSIZE
);
774 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
775 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
777 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
778 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
779 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
781 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
783 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
784 bzero(zio
, sizeof (zio_t
));
786 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
787 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
789 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
790 offsetof(zio_link_t
, zl_parent_node
));
791 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
792 offsetof(zio_link_t
, zl_child_node
));
793 metaslab_trace_init(&zio
->io_alloc_list
);
796 zio
->io_child_type
= ZIO_CHILD_VDEV
;
797 else if (flags
& ZIO_FLAG_GANG_CHILD
)
798 zio
->io_child_type
= ZIO_CHILD_GANG
;
799 else if (flags
& ZIO_FLAG_DDT_CHILD
)
800 zio
->io_child_type
= ZIO_CHILD_DDT
;
802 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
805 zio
->io_bp
= (blkptr_t
*)bp
;
806 zio
->io_bp_copy
= *bp
;
807 zio
->io_bp_orig
= *bp
;
808 if (type
!= ZIO_TYPE_WRITE
||
809 zio
->io_child_type
== ZIO_CHILD_DDT
)
810 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
811 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
812 zio
->io_logical
= zio
;
813 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
814 pipeline
|= ZIO_GANG_STAGES
;
820 zio
->io_private
= private;
822 zio
->io_priority
= priority
;
824 zio
->io_offset
= offset
;
825 zio
->io_orig_abd
= zio
->io_abd
= data
;
826 zio
->io_orig_size
= zio
->io_size
= psize
;
827 zio
->io_lsize
= lsize
;
828 zio
->io_orig_flags
= zio
->io_flags
= flags
;
829 zio
->io_orig_stage
= zio
->io_stage
= stage
;
830 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
831 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
833 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
);
834 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
837 zio
->io_bookmark
= *zb
;
840 if (zio
->io_metaslab_class
== NULL
)
841 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
842 if (zio
->io_logical
== NULL
)
843 zio
->io_logical
= pio
->io_logical
;
844 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
845 zio
->io_gang_leader
= pio
->io_gang_leader
;
846 zio_add_child(pio
, zio
);
849 taskq_init_ent(&zio
->io_tqent
);
855 zio_destroy(zio_t
*zio
)
857 metaslab_trace_fini(&zio
->io_alloc_list
);
858 list_destroy(&zio
->io_parent_list
);
859 list_destroy(&zio
->io_child_list
);
860 mutex_destroy(&zio
->io_lock
);
861 cv_destroy(&zio
->io_cv
);
862 kmem_cache_free(zio_cache
, zio
);
866 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
867 void *private, enum zio_flag flags
)
871 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
872 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
873 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
879 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, enum zio_flag flags
)
881 return (zio_null(NULL
, spa
, NULL
, done
, private, flags
));
885 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
, boolean_t config_held
)
887 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
888 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
889 bp
, (longlong_t
)BP_GET_TYPE(bp
));
891 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
||
892 BP_GET_CHECKSUM(bp
) <= ZIO_CHECKSUM_ON
) {
893 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
894 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
896 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
||
897 BP_GET_COMPRESS(bp
) <= ZIO_COMPRESS_ON
) {
898 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
899 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
901 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
902 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
903 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
905 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
906 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
907 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
910 if (BP_IS_EMBEDDED(bp
)) {
911 if (BPE_GET_ETYPE(bp
) >= NUM_BP_EMBEDDED_TYPES
) {
912 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
913 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
918 * Do not verify individual DVAs if the config is not trusted. This
919 * will be done once the zio is executed in vdev_mirror_map_alloc.
921 if (!spa
->spa_trust_config
)
925 spa_config_enter(spa
, SCL_VDEV
, bp
, RW_READER
);
927 ASSERT(spa_config_held(spa
, SCL_VDEV
, RW_WRITER
));
929 * Pool-specific checks.
931 * Note: it would be nice to verify that the blk_birth and
932 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
933 * allows the birth time of log blocks (and dmu_sync()-ed blocks
934 * that are in the log) to be arbitrarily large.
936 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
937 uint64_t vdevid
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
939 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
940 zfs_panic_recover("blkptr at %p DVA %u has invalid "
942 bp
, i
, (longlong_t
)vdevid
);
945 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
947 zfs_panic_recover("blkptr at %p DVA %u has invalid "
949 bp
, i
, (longlong_t
)vdevid
);
952 if (vd
->vdev_ops
== &vdev_hole_ops
) {
953 zfs_panic_recover("blkptr at %p DVA %u has hole "
955 bp
, i
, (longlong_t
)vdevid
);
958 if (vd
->vdev_ops
== &vdev_missing_ops
) {
960 * "missing" vdevs are valid during import, but we
961 * don't have their detailed info (e.g. asize), so
962 * we can't perform any more checks on them.
966 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
967 uint64_t asize
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
969 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
970 if (offset
+ asize
> vd
->vdev_asize
) {
971 zfs_panic_recover("blkptr at %p DVA %u has invalid "
973 bp
, i
, (longlong_t
)offset
);
977 spa_config_exit(spa
, SCL_VDEV
, bp
);
981 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
983 uint64_t vdevid
= DVA_GET_VDEV(dva
);
985 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
988 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
992 if (vd
->vdev_ops
== &vdev_hole_ops
)
995 if (vd
->vdev_ops
== &vdev_missing_ops
) {
999 uint64_t offset
= DVA_GET_OFFSET(dva
);
1000 uint64_t asize
= DVA_GET_ASIZE(dva
);
1003 asize
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1004 if (offset
+ asize
> vd
->vdev_asize
)
1011 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
1012 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
1013 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
)
1017 zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
);
1019 zio
= zio_create(pio
, spa
, BP_PHYSICAL_BIRTH(bp
), bp
,
1020 data
, size
, size
, done
, private,
1021 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1022 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1023 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1029 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1030 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1031 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1032 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
1033 void *private, zio_priority_t priority
, enum zio_flag flags
,
1034 const zbookmark_phys_t
*zb
)
1038 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1039 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1040 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1041 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1042 DMU_OT_IS_VALID(zp
->zp_type
) &&
1043 zp
->zp_level
< 32 &&
1044 zp
->zp_copies
> 0 &&
1045 zp
->zp_copies
<= spa_max_replication(spa
));
1047 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1048 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1049 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1050 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1052 zio
->io_ready
= ready
;
1053 zio
->io_children_ready
= children_ready
;
1054 zio
->io_physdone
= physdone
;
1058 * Data can be NULL if we are going to call zio_write_override() to
1059 * provide the already-allocated BP. But we may need the data to
1060 * verify a dedup hit (if requested). In this case, don't try to
1061 * dedup (just take the already-allocated BP verbatim). Encrypted
1062 * dedup blocks need data as well so we also disable dedup in this
1066 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1067 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1074 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1075 uint64_t size
, zio_done_func_t
*done
, void *private,
1076 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
)
1080 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1081 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1082 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1088 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
)
1090 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1091 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1092 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1093 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1096 * We must reset the io_prop to match the values that existed
1097 * when the bp was first written by dmu_sync() keeping in mind
1098 * that nopwrite and dedup are mutually exclusive.
1100 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1101 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1102 zio
->io_prop
.zp_copies
= copies
;
1103 zio
->io_bp_override
= bp
;
1107 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1110 zfs_blkptr_verify(spa
, bp
, B_FALSE
);
1113 * The check for EMBEDDED is a performance optimization. We
1114 * process the free here (by ignoring it) rather than
1115 * putting it on the list and then processing it in zio_free_sync().
1117 if (BP_IS_EMBEDDED(bp
))
1119 metaslab_check_free(spa
, bp
);
1122 * Frees that are for the currently-syncing txg, are not going to be
1123 * deferred, and which will not need to do a read (i.e. not GANG or
1124 * DEDUP), can be processed immediately. Otherwise, put them on the
1125 * in-memory list for later processing.
1127 * Note that we only defer frees after zfs_sync_pass_deferred_free
1128 * when the log space map feature is disabled. [see relevant comment
1129 * in spa_sync_iterate_to_convergence()]
1131 if (BP_IS_GANG(bp
) ||
1133 txg
!= spa
->spa_syncing_txg
||
1134 (spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
&&
1135 !spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
))) {
1136 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1138 VERIFY0(zio_wait(zio_free_sync(NULL
, spa
, txg
, bp
, 0)));
1143 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1144 enum zio_flag flags
)
1147 enum zio_stage stage
= ZIO_FREE_PIPELINE
;
1149 ASSERT(!BP_IS_HOLE(bp
));
1150 ASSERT(spa_syncing_txg(spa
) == txg
);
1152 if (BP_IS_EMBEDDED(bp
))
1153 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1155 metaslab_check_free(spa
, bp
);
1157 dsl_scan_freed(spa
, bp
);
1160 * GANG and DEDUP blocks can induce a read (for the gang block header,
1161 * or the DDT), so issue them asynchronously so that this thread is
1164 if (BP_IS_GANG(bp
) || BP_GET_DEDUP(bp
))
1165 stage
|= ZIO_STAGE_ISSUE_ASYNC
;
1167 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1168 BP_GET_PSIZE(bp
), NULL
, NULL
, ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1169 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
);
1175 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1176 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1180 zfs_blkptr_verify(spa
, bp
, flags
& ZIO_FLAG_CONFIG_WRITER
);
1182 if (BP_IS_EMBEDDED(bp
))
1183 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1186 * A claim is an allocation of a specific block. Claims are needed
1187 * to support immediate writes in the intent log. The issue is that
1188 * immediate writes contain committed data, but in a txg that was
1189 * *not* committed. Upon opening the pool after an unclean shutdown,
1190 * the intent log claims all blocks that contain immediate write data
1191 * so that the SPA knows they're in use.
1193 * All claims *must* be resolved in the first txg -- before the SPA
1194 * starts allocating blocks -- so that nothing is allocated twice.
1195 * If txg == 0 we just verify that the block is claimable.
1197 ASSERT3U(spa
->spa_uberblock
.ub_rootbp
.blk_birth
, <,
1198 spa_min_claim_txg(spa
));
1199 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1200 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(1M) */
1202 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1203 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1204 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1205 ASSERT0(zio
->io_queued_timestamp
);
1211 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1212 zio_done_func_t
*done
, void *private, enum zio_flag flags
)
1217 if (vd
->vdev_children
== 0) {
1218 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1219 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1220 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1224 zio
= zio_null(pio
, spa
, NULL
, NULL
, NULL
, flags
);
1226 for (c
= 0; c
< vd
->vdev_children
; c
++)
1227 zio_nowait(zio_ioctl(zio
, spa
, vd
->vdev_child
[c
], cmd
,
1228 done
, private, flags
));
1235 zio_trim(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1236 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1237 enum zio_flag flags
, enum trim_flag trim_flags
)
1241 ASSERT0(vd
->vdev_children
);
1242 ASSERT0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
1243 ASSERT0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
1244 ASSERT3U(size
, !=, 0);
1246 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, NULL
, size
, size
, done
,
1247 private, ZIO_TYPE_TRIM
, priority
, flags
| ZIO_FLAG_PHYSICAL
,
1248 vd
, offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_TRIM_PIPELINE
);
1249 zio
->io_trim_flags
= trim_flags
;
1255 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1256 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1257 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1261 ASSERT(vd
->vdev_children
== 0);
1262 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1263 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1264 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1266 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1267 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1268 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1270 zio
->io_prop
.zp_checksum
= checksum
;
1276 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1277 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1278 zio_priority_t priority
, enum zio_flag flags
, boolean_t labels
)
1282 ASSERT(vd
->vdev_children
== 0);
1283 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1284 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1285 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1287 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1288 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1289 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1291 zio
->io_prop
.zp_checksum
= checksum
;
1293 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1295 * zec checksums are necessarily destructive -- they modify
1296 * the end of the write buffer to hold the verifier/checksum.
1297 * Therefore, we must make a local copy in case the data is
1298 * being written to multiple places in parallel.
1300 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1301 abd_copy(wbuf
, data
, size
);
1303 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1310 * Create a child I/O to do some work for us.
1313 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1314 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1315 enum zio_flag flags
, zio_done_func_t
*done
, void *private)
1317 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1321 * vdev child I/Os do not propagate their error to the parent.
1322 * Therefore, for correct operation the caller *must* check for
1323 * and handle the error in the child i/o's done callback.
1324 * The only exceptions are i/os that we don't care about
1325 * (OPTIONAL or REPAIR).
1327 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1330 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1332 * If we have the bp, then the child should perform the
1333 * checksum and the parent need not. This pushes error
1334 * detection as close to the leaves as possible and
1335 * eliminates redundant checksums in the interior nodes.
1337 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1338 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1341 if (vd
->vdev_ops
->vdev_op_leaf
) {
1342 ASSERT0(vd
->vdev_children
);
1343 offset
+= VDEV_LABEL_START_SIZE
;
1346 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1349 * If we've decided to do a repair, the write is not speculative --
1350 * even if the original read was.
1352 if (flags
& ZIO_FLAG_IO_REPAIR
)
1353 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1356 * If we're creating a child I/O that is not associated with a
1357 * top-level vdev, then the child zio is not an allocating I/O.
1358 * If this is a retried I/O then we ignore it since we will
1359 * have already processed the original allocating I/O.
1361 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1362 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1363 ASSERT(pio
->io_metaslab_class
!= NULL
);
1364 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1365 ASSERT(type
== ZIO_TYPE_WRITE
);
1366 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1367 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1368 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1369 pio
->io_child_type
== ZIO_CHILD_GANG
);
1371 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1375 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1376 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1377 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1378 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1380 zio
->io_physdone
= pio
->io_physdone
;
1381 if (vd
->vdev_ops
->vdev_op_leaf
&& zio
->io_logical
!= NULL
)
1382 zio
->io_logical
->io_phys_children
++;
1388 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1389 zio_type_t type
, zio_priority_t priority
, enum zio_flag flags
,
1390 zio_done_func_t
*done
, void *private)
1394 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1396 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1397 data
, size
, size
, done
, private, type
, priority
,
1398 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1400 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1406 zio_flush(zio_t
*zio
, vdev_t
*vd
)
1408 zio_nowait(zio_ioctl(zio
, zio
->io_spa
, vd
, DKIOCFLUSHWRITECACHE
,
1410 ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1414 zio_shrink(zio_t
*zio
, uint64_t size
)
1416 ASSERT3P(zio
->io_executor
, ==, NULL
);
1417 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1418 ASSERT3U(size
, <=, zio
->io_size
);
1421 * We don't shrink for raidz because of problems with the
1422 * reconstruction when reading back less than the block size.
1423 * Note, BP_IS_RAIDZ() assumes no compression.
1425 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1426 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1427 /* we are not doing a raw write */
1428 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1429 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1434 * ==========================================================================
1435 * Prepare to read and write logical blocks
1436 * ==========================================================================
1440 zio_read_bp_init(zio_t
*zio
)
1442 blkptr_t
*bp
= zio
->io_bp
;
1444 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1446 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1448 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1449 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1450 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1451 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1452 psize
, psize
, zio_decompress
);
1455 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1456 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1457 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1458 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1459 psize
, psize
, zio_decrypt
);
1462 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1463 int psize
= BPE_GET_PSIZE(bp
);
1464 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1466 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1467 decode_embedded_bp_compressed(bp
, data
);
1468 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1470 ASSERT(!BP_IS_EMBEDDED(bp
));
1471 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1474 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp
)) && BP_GET_LEVEL(bp
) == 0)
1475 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1477 if (BP_GET_TYPE(bp
) == DMU_OT_DDT_ZAP
)
1478 zio
->io_flags
|= ZIO_FLAG_DONT_CACHE
;
1480 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1481 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1487 zio_write_bp_init(zio_t
*zio
)
1489 if (!IO_IS_ALLOCATING(zio
))
1492 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1494 if (zio
->io_bp_override
) {
1495 blkptr_t
*bp
= zio
->io_bp
;
1496 zio_prop_t
*zp
= &zio
->io_prop
;
1498 ASSERT(bp
->blk_birth
!= zio
->io_txg
);
1499 ASSERT(BP_GET_DEDUP(zio
->io_bp_override
) == 0);
1501 *bp
= *zio
->io_bp_override
;
1502 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1504 if (BP_IS_EMBEDDED(bp
))
1508 * If we've been overridden and nopwrite is set then
1509 * set the flag accordingly to indicate that a nopwrite
1510 * has already occurred.
1512 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1513 ASSERT(!zp
->zp_dedup
);
1514 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1515 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1519 ASSERT(!zp
->zp_nopwrite
);
1521 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1524 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1525 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1527 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1529 BP_SET_DEDUP(bp
, 1);
1530 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1535 * We were unable to handle this as an override bp, treat
1536 * it as a regular write I/O.
1538 zio
->io_bp_override
= NULL
;
1539 *bp
= zio
->io_bp_orig
;
1540 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1547 zio_write_compress(zio_t
*zio
)
1549 spa_t
*spa
= zio
->io_spa
;
1550 zio_prop_t
*zp
= &zio
->io_prop
;
1551 enum zio_compress compress
= zp
->zp_compress
;
1552 blkptr_t
*bp
= zio
->io_bp
;
1553 uint64_t lsize
= zio
->io_lsize
;
1554 uint64_t psize
= zio
->io_size
;
1558 * If our children haven't all reached the ready stage,
1559 * wait for them and then repeat this pipeline stage.
1561 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1562 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1566 if (!IO_IS_ALLOCATING(zio
))
1569 if (zio
->io_children_ready
!= NULL
) {
1571 * Now that all our children are ready, run the callback
1572 * associated with this zio in case it wants to modify the
1573 * data to be written.
1575 ASSERT3U(zp
->zp_level
, >, 0);
1576 zio
->io_children_ready(zio
);
1579 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1580 ASSERT(zio
->io_bp_override
== NULL
);
1582 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
) {
1584 * We're rewriting an existing block, which means we're
1585 * working on behalf of spa_sync(). For spa_sync() to
1586 * converge, it must eventually be the case that we don't
1587 * have to allocate new blocks. But compression changes
1588 * the blocksize, which forces a reallocate, and makes
1589 * convergence take longer. Therefore, after the first
1590 * few passes, stop compressing to ensure convergence.
1592 pass
= spa_sync_pass(spa
);
1594 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1595 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1596 ASSERT(!BP_GET_DEDUP(bp
));
1598 if (pass
>= zfs_sync_pass_dont_compress
)
1599 compress
= ZIO_COMPRESS_OFF
;
1601 /* Make sure someone doesn't change their mind on overwrites */
1602 ASSERT(BP_IS_EMBEDDED(bp
) || MIN(zp
->zp_copies
+ BP_IS_GANG(bp
),
1603 spa_max_replication(spa
)) == BP_GET_NDVAS(bp
));
1606 /* If it's a compressed write that is not raw, compress the buffer. */
1607 if (compress
!= ZIO_COMPRESS_OFF
&&
1608 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1609 void *cbuf
= zio_buf_alloc(lsize
);
1610 psize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1611 if (psize
== 0 || psize
== lsize
) {
1612 compress
= ZIO_COMPRESS_OFF
;
1613 zio_buf_free(cbuf
, lsize
);
1614 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1615 psize
<= BPE_PAYLOAD_SIZE
&&
1616 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1617 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1618 encode_embedded_bp_compressed(bp
,
1619 cbuf
, compress
, lsize
, psize
);
1620 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1621 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1622 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1623 zio_buf_free(cbuf
, lsize
);
1624 bp
->blk_birth
= zio
->io_txg
;
1625 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1626 ASSERT(spa_feature_is_active(spa
,
1627 SPA_FEATURE_EMBEDDED_DATA
));
1631 * Round up compressed size up to the ashift
1632 * of the smallest-ashift device, and zero the tail.
1633 * This ensures that the compressed size of the BP
1634 * (and thus compressratio property) are correct,
1635 * in that we charge for the padding used to fill out
1638 ASSERT3U(spa
->spa_min_ashift
, >=, SPA_MINBLOCKSHIFT
);
1639 size_t rounded
= (size_t)P2ROUNDUP(psize
,
1640 1ULL << spa
->spa_min_ashift
);
1641 if (rounded
>= lsize
) {
1642 compress
= ZIO_COMPRESS_OFF
;
1643 zio_buf_free(cbuf
, lsize
);
1646 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1647 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1648 abd_zero_off(cdata
, psize
, rounded
- psize
);
1650 zio_push_transform(zio
, cdata
,
1651 psize
, lsize
, NULL
);
1656 * We were unable to handle this as an override bp, treat
1657 * it as a regular write I/O.
1659 zio
->io_bp_override
= NULL
;
1660 *bp
= zio
->io_bp_orig
;
1661 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1663 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1664 zp
->zp_type
== DMU_OT_DNODE
) {
1666 * The DMU actually relies on the zio layer's compression
1667 * to free metadnode blocks that have had all contained
1668 * dnodes freed. As a result, even when doing a raw
1669 * receive, we must check whether the block can be compressed
1672 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1673 zio
->io_abd
, NULL
, lsize
);
1675 compress
= ZIO_COMPRESS_OFF
;
1677 ASSERT3U(psize
, !=, 0);
1681 * The final pass of spa_sync() must be all rewrites, but the first
1682 * few passes offer a trade-off: allocating blocks defers convergence,
1683 * but newly allocated blocks are sequential, so they can be written
1684 * to disk faster. Therefore, we allow the first few passes of
1685 * spa_sync() to allocate new blocks, but force rewrites after that.
1686 * There should only be a handful of blocks after pass 1 in any case.
1688 if (!BP_IS_HOLE(bp
) && bp
->blk_birth
== zio
->io_txg
&&
1689 BP_GET_PSIZE(bp
) == psize
&&
1690 pass
>= zfs_sync_pass_rewrite
) {
1691 VERIFY3U(psize
, !=, 0);
1692 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1694 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1695 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1698 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1702 if (zio
->io_bp_orig
.blk_birth
!= 0 &&
1703 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1704 BP_SET_LSIZE(bp
, lsize
);
1705 BP_SET_TYPE(bp
, zp
->zp_type
);
1706 BP_SET_LEVEL(bp
, zp
->zp_level
);
1707 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1709 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1711 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1712 BP_SET_LSIZE(bp
, lsize
);
1713 BP_SET_TYPE(bp
, zp
->zp_type
);
1714 BP_SET_LEVEL(bp
, zp
->zp_level
);
1715 BP_SET_PSIZE(bp
, psize
);
1716 BP_SET_COMPRESS(bp
, compress
);
1717 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1718 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1719 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1721 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1722 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1723 ASSERT(!zp
->zp_encrypt
||
1724 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1725 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1727 if (zp
->zp_nopwrite
) {
1728 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1729 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1730 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
1737 zio_free_bp_init(zio_t
*zio
)
1739 blkptr_t
*bp
= zio
->io_bp
;
1741 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1742 if (BP_GET_DEDUP(bp
))
1743 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
1746 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1752 * ==========================================================================
1753 * Execute the I/O pipeline
1754 * ==========================================================================
1758 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
1760 spa_t
*spa
= zio
->io_spa
;
1761 zio_type_t t
= zio
->io_type
;
1762 int flags
= (cutinline
? TQ_FRONT
: 0);
1765 * If we're a config writer or a probe, the normal issue and
1766 * interrupt threads may all be blocked waiting for the config lock.
1767 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1769 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
1773 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1775 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
1779 * If this is a high priority I/O, then use the high priority taskq if
1782 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
1783 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
1784 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
1787 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
1790 * NB: We are assuming that the zio can only be dispatched
1791 * to a single taskq at a time. It would be a grievous error
1792 * to dispatch the zio to another taskq at the same time.
1794 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
1795 spa_taskq_dispatch_ent(spa
, t
, q
, (task_func_t
*)zio_execute
, zio
,
1796 flags
, &zio
->io_tqent
);
1800 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
1802 kthread_t
*executor
= zio
->io_executor
;
1803 spa_t
*spa
= zio
->io_spa
;
1805 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
1806 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
1808 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
1809 if (taskq_member(tqs
->stqs_taskq
[i
], executor
))
1818 zio_issue_async(zio_t
*zio
)
1820 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
1826 zio_interrupt(zio_t
*zio
)
1828 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
1832 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
1834 zio_t
*cio
, *cio_next
;
1835 zio_link_t
*zl
= NULL
;
1836 vdev_t
*vd
= pio
->io_vd
;
1838 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
1839 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
1840 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
1841 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
1842 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
1844 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
1845 "delta=%llu queued=%llu io=%llu "
1846 "path=%s last=%llu "
1847 "type=%d priority=%d flags=0x%x "
1848 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1849 "objset=%llu object=%llu level=%llu blkid=%llu "
1850 "offset=%llu size=%llu error=%d",
1851 ziodepth
, pio
, pio
->io_timestamp
,
1852 delta
, pio
->io_delta
, pio
->io_delay
,
1853 vd
? vd
->vdev_path
: "NULL", vq
? vq
->vq_io_complete_ts
: 0,
1854 pio
->io_type
, pio
->io_priority
, pio
->io_flags
,
1855 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
1856 zb
->zb_objset
, zb
->zb_object
, zb
->zb_level
, zb
->zb_blkid
,
1857 pio
->io_offset
, pio
->io_size
, pio
->io_error
);
1858 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
1859 pio
->io_spa
, vd
, zb
, pio
, 0, 0);
1861 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
1862 taskq_empty_ent(&pio
->io_tqent
)) {
1867 mutex_enter(&pio
->io_lock
);
1868 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
1869 cio_next
= zio_walk_children(pio
, &zl
);
1870 zio_deadman_impl(cio
, ziodepth
+ 1);
1872 mutex_exit(&pio
->io_lock
);
1876 * Log the critical information describing this zio and all of its children
1877 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1880 zio_deadman(zio_t
*pio
, char *tag
)
1882 spa_t
*spa
= pio
->io_spa
;
1883 char *name
= spa_name(spa
);
1885 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
1888 zio_deadman_impl(pio
, 0);
1890 switch (spa_get_deadman_failmode(spa
)) {
1891 case ZIO_FAILURE_MODE_WAIT
:
1892 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
1895 case ZIO_FAILURE_MODE_CONTINUE
:
1896 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
1899 case ZIO_FAILURE_MODE_PANIC
:
1900 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
1906 * Execute the I/O pipeline until one of the following occurs:
1907 * (1) the I/O completes; (2) the pipeline stalls waiting for
1908 * dependent child I/Os; (3) the I/O issues, so we're waiting
1909 * for an I/O completion interrupt; (4) the I/O is delegated by
1910 * vdev-level caching or aggregation; (5) the I/O is deferred
1911 * due to vdev-level queueing; (6) the I/O is handed off to
1912 * another thread. In all cases, the pipeline stops whenever
1913 * there's no CPU work; it never burns a thread in cv_wait_io().
1915 * There's no locking on io_stage because there's no legitimate way
1916 * for multiple threads to be attempting to process the same I/O.
1918 static zio_pipe_stage_t
*zio_pipeline
[];
1921 * zio_execute() is a wrapper around the static function
1922 * __zio_execute() so that we can force __zio_execute() to be
1923 * inlined. This reduces stack overhead which is important
1924 * because __zio_execute() is called recursively in several zio
1925 * code paths. zio_execute() itself cannot be inlined because
1926 * it is externally visible.
1929 zio_execute(zio_t
*zio
)
1931 fstrans_cookie_t cookie
;
1933 cookie
= spl_fstrans_mark();
1935 spl_fstrans_unmark(cookie
);
1939 * Used to determine if in the current context the stack is sized large
1940 * enough to allow zio_execute() to be called recursively. A minimum
1941 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1944 zio_execute_stack_check(zio_t
*zio
)
1946 #if !defined(HAVE_LARGE_STACKS)
1947 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
1949 /* Executing in txg_sync_thread() context. */
1950 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
1953 /* Pool initialization outside of zio_taskq context. */
1954 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
1955 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
1956 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
1958 #endif /* HAVE_LARGE_STACKS */
1963 __attribute__((always_inline
))
1965 __zio_execute(zio_t
*zio
)
1967 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
1969 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
1970 enum zio_stage pipeline
= zio
->io_pipeline
;
1971 enum zio_stage stage
= zio
->io_stage
;
1973 zio
->io_executor
= curthread
;
1975 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
1976 ASSERT(ISP2(stage
));
1977 ASSERT(zio
->io_stall
== NULL
);
1981 } while ((stage
& pipeline
) == 0);
1983 ASSERT(stage
<= ZIO_STAGE_DONE
);
1986 * If we are in interrupt context and this pipeline stage
1987 * will grab a config lock that is held across I/O,
1988 * or may wait for an I/O that needs an interrupt thread
1989 * to complete, issue async to avoid deadlock.
1991 * For VDEV_IO_START, we cut in line so that the io will
1992 * be sent to disk promptly.
1994 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
1995 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
1996 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
1997 zio_requeue_io_start_cut_in_line
: B_FALSE
;
1998 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2003 * If the current context doesn't have large enough stacks
2004 * the zio must be issued asynchronously to prevent overflow.
2006 if (zio_execute_stack_check(zio
)) {
2007 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2008 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2009 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2013 zio
->io_stage
= stage
;
2014 zio
->io_pipeline_trace
|= zio
->io_stage
;
2017 * The zio pipeline stage returns the next zio to execute
2018 * (typically the same as this one), or NULL if we should
2021 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2030 * ==========================================================================
2031 * Initiate I/O, either sync or async
2032 * ==========================================================================
2035 zio_wait(zio_t
*zio
)
2037 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2040 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2041 ASSERT3P(zio
->io_executor
, ==, NULL
);
2043 zio
->io_waiter
= curthread
;
2044 ASSERT0(zio
->io_queued_timestamp
);
2045 zio
->io_queued_timestamp
= gethrtime();
2049 mutex_enter(&zio
->io_lock
);
2050 while (zio
->io_executor
!= NULL
) {
2051 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2052 ddi_get_lbolt() + timeout
);
2054 if (zfs_deadman_enabled
&& error
== -1 &&
2055 gethrtime() - zio
->io_queued_timestamp
>
2056 spa_deadman_ziotime(zio
->io_spa
)) {
2057 mutex_exit(&zio
->io_lock
);
2058 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2059 zio_deadman(zio
, FTAG
);
2060 mutex_enter(&zio
->io_lock
);
2063 mutex_exit(&zio
->io_lock
);
2065 error
= zio
->io_error
;
2072 zio_nowait(zio_t
*zio
)
2074 ASSERT3P(zio
->io_executor
, ==, NULL
);
2076 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2077 zio_unique_parent(zio
) == NULL
) {
2081 * This is a logical async I/O with no parent to wait for it.
2082 * We add it to the spa_async_root_zio "Godfather" I/O which
2083 * will ensure they complete prior to unloading the pool.
2085 spa_t
*spa
= zio
->io_spa
;
2087 pio
= spa
->spa_async_zio_root
[CPU_SEQID
];
2090 zio_add_child(pio
, zio
);
2093 ASSERT0(zio
->io_queued_timestamp
);
2094 zio
->io_queued_timestamp
= gethrtime();
2099 * ==========================================================================
2100 * Reexecute, cancel, or suspend/resume failed I/O
2101 * ==========================================================================
2105 zio_reexecute(zio_t
*pio
)
2107 zio_t
*cio
, *cio_next
;
2109 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2110 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2111 ASSERT(pio
->io_gang_leader
== NULL
);
2112 ASSERT(pio
->io_gang_tree
== NULL
);
2114 pio
->io_flags
= pio
->io_orig_flags
;
2115 pio
->io_stage
= pio
->io_orig_stage
;
2116 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2117 pio
->io_reexecute
= 0;
2118 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2119 pio
->io_pipeline_trace
= 0;
2121 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2122 pio
->io_state
[w
] = 0;
2123 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2124 pio
->io_child_error
[c
] = 0;
2126 if (IO_IS_ALLOCATING(pio
))
2127 BP_ZERO(pio
->io_bp
);
2130 * As we reexecute pio's children, new children could be created.
2131 * New children go to the head of pio's io_child_list, however,
2132 * so we will (correctly) not reexecute them. The key is that
2133 * the remainder of pio's io_child_list, from 'cio_next' onward,
2134 * cannot be affected by any side effects of reexecuting 'cio'.
2136 zio_link_t
*zl
= NULL
;
2137 mutex_enter(&pio
->io_lock
);
2138 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2139 cio_next
= zio_walk_children(pio
, &zl
);
2140 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
2141 pio
->io_children
[cio
->io_child_type
][w
]++;
2142 mutex_exit(&pio
->io_lock
);
2144 mutex_enter(&pio
->io_lock
);
2146 mutex_exit(&pio
->io_lock
);
2149 * Now that all children have been reexecuted, execute the parent.
2150 * We don't reexecute "The Godfather" I/O here as it's the
2151 * responsibility of the caller to wait on it.
2153 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2154 pio
->io_queued_timestamp
= gethrtime();
2160 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2162 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2163 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2164 "failure and the failure mode property for this pool "
2165 "is set to panic.", spa_name(spa
));
2167 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2168 "failure and has been suspended.\n", spa_name(spa
));
2170 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2173 mutex_enter(&spa
->spa_suspend_lock
);
2175 if (spa
->spa_suspend_zio_root
== NULL
)
2176 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2177 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2178 ZIO_FLAG_GODFATHER
);
2180 spa
->spa_suspended
= reason
;
2183 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2184 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2185 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2186 ASSERT(zio_unique_parent(zio
) == NULL
);
2187 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2188 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2191 mutex_exit(&spa
->spa_suspend_lock
);
2195 zio_resume(spa_t
*spa
)
2200 * Reexecute all previously suspended i/o.
2202 mutex_enter(&spa
->spa_suspend_lock
);
2203 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2204 cv_broadcast(&spa
->spa_suspend_cv
);
2205 pio
= spa
->spa_suspend_zio_root
;
2206 spa
->spa_suspend_zio_root
= NULL
;
2207 mutex_exit(&spa
->spa_suspend_lock
);
2213 return (zio_wait(pio
));
2217 zio_resume_wait(spa_t
*spa
)
2219 mutex_enter(&spa
->spa_suspend_lock
);
2220 while (spa_suspended(spa
))
2221 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2222 mutex_exit(&spa
->spa_suspend_lock
);
2226 * ==========================================================================
2229 * A gang block is a collection of small blocks that looks to the DMU
2230 * like one large block. When zio_dva_allocate() cannot find a block
2231 * of the requested size, due to either severe fragmentation or the pool
2232 * being nearly full, it calls zio_write_gang_block() to construct the
2233 * block from smaller fragments.
2235 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2236 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2237 * an indirect block: it's an array of block pointers. It consumes
2238 * only one sector and hence is allocatable regardless of fragmentation.
2239 * The gang header's bps point to its gang members, which hold the data.
2241 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2242 * as the verifier to ensure uniqueness of the SHA256 checksum.
2243 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2244 * not the gang header. This ensures that data block signatures (needed for
2245 * deduplication) are independent of how the block is physically stored.
2247 * Gang blocks can be nested: a gang member may itself be a gang block.
2248 * Thus every gang block is a tree in which root and all interior nodes are
2249 * gang headers, and the leaves are normal blocks that contain user data.
2250 * The root of the gang tree is called the gang leader.
2252 * To perform any operation (read, rewrite, free, claim) on a gang block,
2253 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2254 * in the io_gang_tree field of the original logical i/o by recursively
2255 * reading the gang leader and all gang headers below it. This yields
2256 * an in-core tree containing the contents of every gang header and the
2257 * bps for every constituent of the gang block.
2259 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2260 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2261 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2262 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2263 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2264 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2265 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2266 * of the gang header plus zio_checksum_compute() of the data to update the
2267 * gang header's blk_cksum as described above.
2269 * The two-phase assemble/issue model solves the problem of partial failure --
2270 * what if you'd freed part of a gang block but then couldn't read the
2271 * gang header for another part? Assembling the entire gang tree first
2272 * ensures that all the necessary gang header I/O has succeeded before
2273 * starting the actual work of free, claim, or write. Once the gang tree
2274 * is assembled, free and claim are in-memory operations that cannot fail.
2276 * In the event that a gang write fails, zio_dva_unallocate() walks the
2277 * gang tree to immediately free (i.e. insert back into the space map)
2278 * everything we've allocated. This ensures that we don't get ENOSPC
2279 * errors during repeated suspend/resume cycles due to a flaky device.
2281 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2282 * the gang tree, we won't modify the block, so we can safely defer the free
2283 * (knowing that the block is still intact). If we *can* assemble the gang
2284 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2285 * each constituent bp and we can allocate a new block on the next sync pass.
2287 * In all cases, the gang tree allows complete recovery from partial failure.
2288 * ==========================================================================
2292 zio_gang_issue_func_done(zio_t
*zio
)
2294 abd_put(zio
->io_abd
);
2298 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2304 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2305 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2306 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2307 &pio
->io_bookmark
));
2311 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2318 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2319 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2320 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2321 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2324 * As we rewrite each gang header, the pipeline will compute
2325 * a new gang block header checksum for it; but no one will
2326 * compute a new data checksum, so we do that here. The one
2327 * exception is the gang leader: the pipeline already computed
2328 * its data checksum because that stage precedes gang assembly.
2329 * (Presently, nothing actually uses interior data checksums;
2330 * this is just good hygiene.)
2332 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2333 abd_t
*buf
= abd_get_offset(data
, offset
);
2335 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2336 buf
, BP_GET_PSIZE(bp
));
2341 * If we are here to damage data for testing purposes,
2342 * leave the GBH alone so that we can detect the damage.
2344 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2345 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2347 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2348 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2349 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2350 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2358 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2361 return (zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2362 ZIO_GANG_CHILD_FLAGS(pio
)));
2367 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2370 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2371 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2374 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2383 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2385 static zio_gang_node_t
*
2386 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2388 zio_gang_node_t
*gn
;
2390 ASSERT(*gnpp
== NULL
);
2392 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2393 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2400 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2402 zio_gang_node_t
*gn
= *gnpp
;
2404 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2405 ASSERT(gn
->gn_child
[g
] == NULL
);
2407 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2408 kmem_free(gn
, sizeof (*gn
));
2413 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2415 zio_gang_node_t
*gn
= *gnpp
;
2420 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2421 zio_gang_tree_free(&gn
->gn_child
[g
]);
2423 zio_gang_node_free(gnpp
);
2427 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2429 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2430 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2432 ASSERT(gio
->io_gang_leader
== gio
);
2433 ASSERT(BP_IS_GANG(bp
));
2435 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2436 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2437 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2441 zio_gang_tree_assemble_done(zio_t
*zio
)
2443 zio_t
*gio
= zio
->io_gang_leader
;
2444 zio_gang_node_t
*gn
= zio
->io_private
;
2445 blkptr_t
*bp
= zio
->io_bp
;
2447 ASSERT(gio
== zio_unique_parent(zio
));
2448 ASSERT(zio
->io_child_count
== 0);
2453 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2454 if (BP_SHOULD_BYTESWAP(bp
))
2455 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2457 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2458 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2459 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2461 abd_put(zio
->io_abd
);
2463 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2464 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2465 if (!BP_IS_GANG(gbp
))
2467 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2472 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2475 zio_t
*gio
= pio
->io_gang_leader
;
2478 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2479 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2480 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2483 * If you're a gang header, your data is in gn->gn_gbh.
2484 * If you're a gang member, your data is in 'data' and gn == NULL.
2486 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2489 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2491 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2492 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2493 if (BP_IS_HOLE(gbp
))
2495 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2497 offset
+= BP_GET_PSIZE(gbp
);
2501 if (gn
== gio
->io_gang_tree
)
2502 ASSERT3U(gio
->io_size
, ==, offset
);
2509 zio_gang_assemble(zio_t
*zio
)
2511 blkptr_t
*bp
= zio
->io_bp
;
2513 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2514 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2516 zio
->io_gang_leader
= zio
;
2518 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2524 zio_gang_issue(zio_t
*zio
)
2526 blkptr_t
*bp
= zio
->io_bp
;
2528 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2532 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2533 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2535 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2536 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2539 zio_gang_tree_free(&zio
->io_gang_tree
);
2541 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2547 zio_write_gang_member_ready(zio_t
*zio
)
2549 zio_t
*pio
= zio_unique_parent(zio
);
2550 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2551 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2553 zio_t
*gio __maybe_unused
= zio
->io_gang_leader
;
2555 if (BP_IS_HOLE(zio
->io_bp
))
2558 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2560 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2561 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2562 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2563 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2564 ASSERT3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2566 mutex_enter(&pio
->io_lock
);
2567 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2568 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2569 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2570 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2571 DVA_SET_ASIZE(&pdva
[d
], asize
);
2573 mutex_exit(&pio
->io_lock
);
2577 zio_write_gang_done(zio_t
*zio
)
2580 * The io_abd field will be NULL for a zio with no data. The io_flags
2581 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2582 * check for it here as it is cleared in zio_ready.
2584 if (zio
->io_abd
!= NULL
)
2585 abd_put(zio
->io_abd
);
2589 zio_write_gang_block(zio_t
*pio
)
2591 spa_t
*spa
= pio
->io_spa
;
2592 metaslab_class_t
*mc
= spa_normal_class(spa
);
2593 blkptr_t
*bp
= pio
->io_bp
;
2594 zio_t
*gio
= pio
->io_gang_leader
;
2596 zio_gang_node_t
*gn
, **gnpp
;
2597 zio_gbh_phys_t
*gbh
;
2599 uint64_t txg
= pio
->io_txg
;
2600 uint64_t resid
= pio
->io_size
;
2602 int copies
= gio
->io_prop
.zp_copies
;
2606 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2609 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2610 * have a third copy.
2612 gbh_copies
= MIN(copies
+ 1, spa_max_replication(spa
));
2613 if (gio
->io_prop
.zp_encrypt
&& gbh_copies
>= SPA_DVAS_PER_BP
)
2614 gbh_copies
= SPA_DVAS_PER_BP
- 1;
2616 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
2617 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2618 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2621 flags
|= METASLAB_ASYNC_ALLOC
;
2622 VERIFY(zfs_refcount_held(&mc
->mc_alloc_slots
[pio
->io_allocator
],
2626 * The logical zio has already placed a reservation for
2627 * 'copies' allocation slots but gang blocks may require
2628 * additional copies. These additional copies
2629 * (i.e. gbh_copies - copies) are guaranteed to succeed
2630 * since metaslab_class_throttle_reserve() always allows
2631 * additional reservations for gang blocks.
2633 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
2634 pio
->io_allocator
, pio
, flags
));
2637 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
2638 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
2639 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
2641 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2642 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2646 * If we failed to allocate the gang block header then
2647 * we remove any additional allocation reservations that
2648 * we placed here. The original reservation will
2649 * be removed when the logical I/O goes to the ready
2652 metaslab_class_throttle_unreserve(mc
,
2653 gbh_copies
- copies
, pio
->io_allocator
, pio
);
2656 pio
->io_error
= error
;
2661 gnpp
= &gio
->io_gang_tree
;
2663 gnpp
= pio
->io_private
;
2664 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
2667 gn
= zio_gang_node_alloc(gnpp
);
2669 bzero(gbh
, SPA_GANGBLOCKSIZE
);
2670 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
2673 * Create the gang header.
2675 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2676 zio_write_gang_done
, NULL
, pio
->io_priority
,
2677 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2680 * Create and nowait the gang children.
2682 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
2683 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
2685 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
2687 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
2688 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
2689 zp
.zp_type
= DMU_OT_NONE
;
2691 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
2692 zp
.zp_dedup
= B_FALSE
;
2693 zp
.zp_dedup_verify
= B_FALSE
;
2694 zp
.zp_nopwrite
= B_FALSE
;
2695 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
2696 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
2697 bzero(zp
.zp_salt
, ZIO_DATA_SALT_LEN
);
2698 bzero(zp
.zp_iv
, ZIO_DATA_IV_LEN
);
2699 bzero(zp
.zp_mac
, ZIO_DATA_MAC_LEN
);
2701 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
2702 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
2703 resid
) : NULL
, lsize
, lsize
, &zp
,
2704 zio_write_gang_member_ready
, NULL
, NULL
,
2705 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
2706 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2708 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
2709 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
2713 * Gang children won't throttle but we should
2714 * account for their work, so reserve an allocation
2715 * slot for them here.
2717 VERIFY(metaslab_class_throttle_reserve(mc
,
2718 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
2724 * Set pio's pipeline to just wait for zio to finish.
2726 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2729 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2731 pio
->io_flags
&= ~ZIO_FLAG_FASTWRITE
;
2739 * The zio_nop_write stage in the pipeline determines if allocating a
2740 * new bp is necessary. The nopwrite feature can handle writes in
2741 * either syncing or open context (i.e. zil writes) and as a result is
2742 * mutually exclusive with dedup.
2744 * By leveraging a cryptographically secure checksum, such as SHA256, we
2745 * can compare the checksums of the new data and the old to determine if
2746 * allocating a new block is required. Note that our requirements for
2747 * cryptographic strength are fairly weak: there can't be any accidental
2748 * hash collisions, but we don't need to be secure against intentional
2749 * (malicious) collisions. To trigger a nopwrite, you have to be able
2750 * to write the file to begin with, and triggering an incorrect (hash
2751 * collision) nopwrite is no worse than simply writing to the file.
2752 * That said, there are no known attacks against the checksum algorithms
2753 * used for nopwrite, assuming that the salt and the checksums
2754 * themselves remain secret.
2757 zio_nop_write(zio_t
*zio
)
2759 blkptr_t
*bp
= zio
->io_bp
;
2760 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
2761 zio_prop_t
*zp
= &zio
->io_prop
;
2763 ASSERT(BP_GET_LEVEL(bp
) == 0);
2764 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
2765 ASSERT(zp
->zp_nopwrite
);
2766 ASSERT(!zp
->zp_dedup
);
2767 ASSERT(zio
->io_bp_override
== NULL
);
2768 ASSERT(IO_IS_ALLOCATING(zio
));
2771 * Check to see if the original bp and the new bp have matching
2772 * characteristics (i.e. same checksum, compression algorithms, etc).
2773 * If they don't then just continue with the pipeline which will
2774 * allocate a new bp.
2776 if (BP_IS_HOLE(bp_orig
) ||
2777 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
2778 ZCHECKSUM_FLAG_NOPWRITE
) ||
2779 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
2780 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
2781 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
2782 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
2783 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
2787 * If the checksums match then reset the pipeline so that we
2788 * avoid allocating a new bp and issuing any I/O.
2790 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
2791 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
2792 ZCHECKSUM_FLAG_NOPWRITE
);
2793 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
2794 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
2795 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
2796 ASSERT(bcmp(&bp
->blk_prop
, &bp_orig
->blk_prop
,
2797 sizeof (uint64_t)) == 0);
2800 * If we're overwriting a block that is currently on an
2801 * indirect vdev, then ignore the nopwrite request and
2802 * allow a new block to be allocated on a concrete vdev.
2804 spa_config_enter(zio
->io_spa
, SCL_VDEV
, FTAG
, RW_READER
);
2805 vdev_t
*tvd
= vdev_lookup_top(zio
->io_spa
,
2806 DVA_GET_VDEV(&bp
->blk_dva
[0]));
2807 if (tvd
->vdev_ops
== &vdev_indirect_ops
) {
2808 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2811 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
2814 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2815 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
2822 * ==========================================================================
2824 * ==========================================================================
2827 zio_ddt_child_read_done(zio_t
*zio
)
2829 blkptr_t
*bp
= zio
->io_bp
;
2830 ddt_entry_t
*dde
= zio
->io_private
;
2832 zio_t
*pio
= zio_unique_parent(zio
);
2834 mutex_enter(&pio
->io_lock
);
2835 ddp
= ddt_phys_select(dde
, bp
);
2836 if (zio
->io_error
== 0)
2837 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
2839 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
2840 dde
->dde_repair_abd
= zio
->io_abd
;
2842 abd_free(zio
->io_abd
);
2843 mutex_exit(&pio
->io_lock
);
2847 zio_ddt_read_start(zio_t
*zio
)
2849 blkptr_t
*bp
= zio
->io_bp
;
2851 ASSERT(BP_GET_DEDUP(bp
));
2852 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2853 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2855 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2856 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2857 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
2858 ddt_phys_t
*ddp
= dde
->dde_phys
;
2859 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
2862 ASSERT(zio
->io_vsd
== NULL
);
2865 if (ddp_self
== NULL
)
2868 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
2869 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
2871 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
2873 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
2874 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
2875 zio
->io_size
, zio_ddt_child_read_done
, dde
,
2876 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
2877 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
2882 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
2883 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
2884 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
2890 zio_ddt_read_done(zio_t
*zio
)
2892 blkptr_t
*bp
= zio
->io_bp
;
2894 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
2898 ASSERT(BP_GET_DEDUP(bp
));
2899 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
2900 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2902 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
2903 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
2904 ddt_entry_t
*dde
= zio
->io_vsd
;
2906 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
2910 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
2911 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2914 if (dde
->dde_repair_abd
!= NULL
) {
2915 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
2917 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
2919 ddt_repair_done(ddt
, dde
);
2923 ASSERT(zio
->io_vsd
== NULL
);
2929 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
2931 spa_t
*spa
= zio
->io_spa
;
2932 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
2934 ASSERT(!(zio
->io_bp_override
&& do_raw
));
2937 * Note: we compare the original data, not the transformed data,
2938 * because when zio->io_bp is an override bp, we will not have
2939 * pushed the I/O transforms. That's an important optimization
2940 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2941 * However, we should never get a raw, override zio so in these
2942 * cases we can compare the io_abd directly. This is useful because
2943 * it allows us to do dedup verification even if we don't have access
2944 * to the original data (for instance, if the encryption keys aren't
2948 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2949 zio_t
*lio
= dde
->dde_lead_zio
[p
];
2951 if (lio
!= NULL
&& do_raw
) {
2952 return (lio
->io_size
!= zio
->io_size
||
2953 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
2954 } else if (lio
!= NULL
) {
2955 return (lio
->io_orig_size
!= zio
->io_orig_size
||
2956 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
2960 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
2961 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
2963 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
2964 blkptr_t blk
= *zio
->io_bp
;
2969 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
2970 psize
= BP_GET_PSIZE(&blk
);
2972 if (psize
!= zio
->io_size
)
2977 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
2979 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
2980 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
2981 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2982 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
2985 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
2986 error
= SET_ERROR(ENOENT
);
2991 return (error
!= 0);
2992 } else if (ddp
->ddp_phys_birth
!= 0) {
2993 arc_buf_t
*abuf
= NULL
;
2994 arc_flags_t aflags
= ARC_FLAG_WAIT
;
2995 blkptr_t blk
= *zio
->io_bp
;
2998 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3000 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3005 error
= arc_read(NULL
, spa
, &blk
,
3006 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3007 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3008 &aflags
, &zio
->io_bookmark
);
3011 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3012 zio
->io_orig_size
) != 0)
3013 error
= SET_ERROR(ENOENT
);
3014 arc_buf_destroy(abuf
, &abuf
);
3018 return (error
!= 0);
3026 zio_ddt_child_write_ready(zio_t
*zio
)
3028 int p
= zio
->io_prop
.zp_copies
;
3029 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3030 ddt_entry_t
*dde
= zio
->io_private
;
3031 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3039 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3041 ddt_phys_fill(ddp
, zio
->io_bp
);
3043 zio_link_t
*zl
= NULL
;
3044 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3045 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3051 zio_ddt_child_write_done(zio_t
*zio
)
3053 int p
= zio
->io_prop
.zp_copies
;
3054 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3055 ddt_entry_t
*dde
= zio
->io_private
;
3056 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3060 ASSERT(ddp
->ddp_refcnt
== 0);
3061 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3062 dde
->dde_lead_zio
[p
] = NULL
;
3064 if (zio
->io_error
== 0) {
3065 zio_link_t
*zl
= NULL
;
3066 while (zio_walk_parents(zio
, &zl
) != NULL
)
3067 ddt_phys_addref(ddp
);
3069 ddt_phys_clear(ddp
);
3076 zio_ddt_write(zio_t
*zio
)
3078 spa_t
*spa
= zio
->io_spa
;
3079 blkptr_t
*bp
= zio
->io_bp
;
3080 uint64_t txg
= zio
->io_txg
;
3081 zio_prop_t
*zp
= &zio
->io_prop
;
3082 int p
= zp
->zp_copies
;
3084 ddt_t
*ddt
= ddt_select(spa
, bp
);
3088 ASSERT(BP_GET_DEDUP(bp
));
3089 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3090 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3091 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3094 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3095 ddp
= &dde
->dde_phys
[p
];
3097 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3099 * If we're using a weak checksum, upgrade to a strong checksum
3100 * and try again. If we're already using a strong checksum,
3101 * we can't resolve it, so just convert to an ordinary write.
3102 * (And automatically e-mail a paper to Nature?)
3104 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3105 ZCHECKSUM_FLAG_DEDUP
)) {
3106 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3107 zio_pop_transforms(zio
);
3108 zio
->io_stage
= ZIO_STAGE_OPEN
;
3111 zp
->zp_dedup
= B_FALSE
;
3112 BP_SET_DEDUP(bp
, B_FALSE
);
3114 ASSERT(!BP_GET_DEDUP(bp
));
3115 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3120 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3121 if (ddp
->ddp_phys_birth
!= 0)
3122 ddt_bp_fill(ddp
, bp
, txg
);
3123 if (dde
->dde_lead_zio
[p
] != NULL
)
3124 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3126 ddt_phys_addref(ddp
);
3127 } else if (zio
->io_bp_override
) {
3128 ASSERT(bp
->blk_birth
== txg
);
3129 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3130 ddt_phys_fill(ddp
, bp
);
3131 ddt_phys_addref(ddp
);
3133 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3134 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3135 zio_ddt_child_write_ready
, NULL
, NULL
,
3136 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3137 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3139 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3140 dde
->dde_lead_zio
[p
] = cio
;
3151 ddt_entry_t
*freedde
; /* for debugging */
3154 zio_ddt_free(zio_t
*zio
)
3156 spa_t
*spa
= zio
->io_spa
;
3157 blkptr_t
*bp
= zio
->io_bp
;
3158 ddt_t
*ddt
= ddt_select(spa
, bp
);
3162 ASSERT(BP_GET_DEDUP(bp
));
3163 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3166 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3168 ddp
= ddt_phys_select(dde
, bp
);
3170 ddt_phys_decref(ddp
);
3178 * ==========================================================================
3179 * Allocate and free blocks
3180 * ==========================================================================
3184 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3188 ASSERT(MUTEX_HELD(&spa
->spa_alloc_locks
[allocator
]));
3190 zio
= avl_first(&spa
->spa_alloc_trees
[allocator
]);
3194 ASSERT(IO_IS_ALLOCATING(zio
));
3197 * Try to place a reservation for this zio. If we're unable to
3198 * reserve then we throttle.
3200 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3201 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3202 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
, 0)) {
3206 avl_remove(&spa
->spa_alloc_trees
[allocator
], zio
);
3207 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3213 zio_dva_throttle(zio_t
*zio
)
3215 spa_t
*spa
= zio
->io_spa
;
3217 metaslab_class_t
*mc
;
3219 /* locate an appropriate allocation class */
3220 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3221 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3223 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3224 !mc
->mc_alloc_throttle_enabled
||
3225 zio
->io_child_type
== ZIO_CHILD_GANG
||
3226 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3230 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3232 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3233 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3235 zbookmark_phys_t
*bm
= &zio
->io_bookmark
;
3237 * We want to try to use as many allocators as possible to help improve
3238 * performance, but we also want logically adjacent IOs to be physically
3239 * adjacent to improve sequential read performance. We chunk each object
3240 * into 2^20 block regions, and then hash based on the objset, object,
3241 * level, and region to accomplish both of these goals.
3243 zio
->io_allocator
= cityhash4(bm
->zb_objset
, bm
->zb_object
,
3244 bm
->zb_level
, bm
->zb_blkid
>> 20) % spa
->spa_alloc_count
;
3245 mutex_enter(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3246 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3247 zio
->io_metaslab_class
= mc
;
3248 avl_add(&spa
->spa_alloc_trees
[zio
->io_allocator
], zio
);
3249 nio
= zio_io_to_allocate(spa
, zio
->io_allocator
);
3250 mutex_exit(&spa
->spa_alloc_locks
[zio
->io_allocator
]);
3255 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3259 mutex_enter(&spa
->spa_alloc_locks
[allocator
]);
3260 zio
= zio_io_to_allocate(spa
, allocator
);
3261 mutex_exit(&spa
->spa_alloc_locks
[allocator
]);
3265 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3266 ASSERT0(zio
->io_error
);
3267 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3271 zio_dva_allocate(zio_t
*zio
)
3273 spa_t
*spa
= zio
->io_spa
;
3274 metaslab_class_t
*mc
;
3275 blkptr_t
*bp
= zio
->io_bp
;
3279 if (zio
->io_gang_leader
== NULL
) {
3280 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3281 zio
->io_gang_leader
= zio
;
3284 ASSERT(BP_IS_HOLE(bp
));
3285 ASSERT0(BP_GET_NDVAS(bp
));
3286 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3287 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3288 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3290 flags
|= (zio
->io_flags
& ZIO_FLAG_FASTWRITE
) ? METASLAB_FASTWRITE
: 0;
3291 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3292 flags
|= METASLAB_DONT_THROTTLE
;
3293 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3294 flags
|= METASLAB_GANG_CHILD
;
3295 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3296 flags
|= METASLAB_ASYNC_ALLOC
;
3299 * if not already chosen, locate an appropriate allocation class
3301 mc
= zio
->io_metaslab_class
;
3303 mc
= spa_preferred_class(spa
, zio
->io_size
,
3304 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3305 zio
->io_prop
.zp_zpl_smallblk
);
3306 zio
->io_metaslab_class
= mc
;
3309 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3310 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3311 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3314 * Fallback to normal class when an alloc class is full
3316 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3318 * If throttling, transfer reservation over to normal class.
3319 * The io_allocator slot can remain the same even though we
3320 * are switching classes.
3322 if (mc
->mc_alloc_throttle_enabled
&&
3323 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3324 metaslab_class_throttle_unreserve(mc
,
3325 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3326 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3328 mc
= spa_normal_class(spa
);
3329 VERIFY(metaslab_class_throttle_reserve(mc
,
3330 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3331 flags
| METASLAB_MUST_RESERVE
));
3333 mc
= spa_normal_class(spa
);
3335 zio
->io_metaslab_class
= mc
;
3337 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3338 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3339 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3343 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3344 "size %llu, error %d", spa_name(spa
), zio
, zio
->io_size
,
3346 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
)
3347 return (zio_write_gang_block(zio
));
3348 zio
->io_error
= error
;
3355 zio_dva_free(zio_t
*zio
)
3357 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3363 zio_dva_claim(zio_t
*zio
)
3367 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3369 zio
->io_error
= error
;
3375 * Undo an allocation. This is used by zio_done() when an I/O fails
3376 * and we want to give back the block we just allocated.
3377 * This handles both normal blocks and gang blocks.
3380 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3382 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
));
3383 ASSERT(zio
->io_bp_override
== NULL
);
3385 if (!BP_IS_HOLE(bp
))
3386 metaslab_free(zio
->io_spa
, bp
, bp
->blk_birth
, B_TRUE
);
3389 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3390 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3391 &gn
->gn_gbh
->zg_blkptr
[g
]);
3397 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3400 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3401 uint64_t size
, boolean_t
*slog
)
3404 zio_alloc_list_t io_alloc_list
;
3406 ASSERT(txg
> spa_syncing_txg(spa
));
3408 metaslab_trace_init(&io_alloc_list
);
3411 * Block pointer fields are useful to metaslabs for stats and debugging.
3412 * Fill in the obvious ones before calling into metaslab_alloc().
3414 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3415 BP_SET_PSIZE(new_bp
, size
);
3416 BP_SET_LEVEL(new_bp
, 0);
3419 * When allocating a zil block, we don't have information about
3420 * the final destination of the block except the objset it's part
3421 * of, so we just hash the objset ID to pick the allocator to get
3424 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3425 txg
, NULL
, METASLAB_FASTWRITE
, &io_alloc_list
, NULL
,
3426 cityhash4(0, 0, 0, os
->os_dsl_dataset
->ds_object
) %
3427 spa
->spa_alloc_count
);
3431 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3432 new_bp
, 1, txg
, NULL
, METASLAB_FASTWRITE
,
3433 &io_alloc_list
, NULL
, cityhash4(0, 0, 0,
3434 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
);
3438 metaslab_trace_fini(&io_alloc_list
);
3441 BP_SET_LSIZE(new_bp
, size
);
3442 BP_SET_PSIZE(new_bp
, size
);
3443 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3444 BP_SET_CHECKSUM(new_bp
,
3445 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3446 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3447 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3448 BP_SET_LEVEL(new_bp
, 0);
3449 BP_SET_DEDUP(new_bp
, 0);
3450 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3453 * encrypted blocks will require an IV and salt. We generate
3454 * these now since we will not be rewriting the bp at
3457 if (os
->os_encrypted
) {
3458 uint8_t iv
[ZIO_DATA_IV_LEN
];
3459 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3461 BP_SET_CRYPT(new_bp
, B_TRUE
);
3462 VERIFY0(spa_crypt_get_salt(spa
,
3463 dmu_objset_id(os
), salt
));
3464 VERIFY0(zio_crypt_generate_iv(iv
));
3466 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3469 zfs_dbgmsg("%s: zil block allocation failure: "
3470 "size %llu, error %d", spa_name(spa
), size
, error
);
3477 * ==========================================================================
3478 * Read and write to physical devices
3479 * ==========================================================================
3483 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3484 * stops after this stage and will resume upon I/O completion.
3485 * However, there are instances where the vdev layer may need to
3486 * continue the pipeline when an I/O was not issued. Since the I/O
3487 * that was sent to the vdev layer might be different than the one
3488 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3489 * force the underlying vdev layers to call either zio_execute() or
3490 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3493 zio_vdev_io_start(zio_t
*zio
)
3495 vdev_t
*vd
= zio
->io_vd
;
3497 spa_t
*spa
= zio
->io_spa
;
3501 ASSERT(zio
->io_error
== 0);
3502 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3505 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3506 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3509 * The mirror_ops handle multiple DVAs in a single BP.
3511 vdev_mirror_ops
.vdev_op_io_start(zio
);
3515 ASSERT3P(zio
->io_logical
, !=, zio
);
3516 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3517 ASSERT(spa
->spa_trust_config
);
3520 * Note: the code can handle other kinds of writes,
3521 * but we don't expect them.
3523 if (zio
->io_vd
->vdev_removing
) {
3524 ASSERT(zio
->io_flags
&
3525 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3526 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3530 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3532 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3533 P2PHASE(zio
->io_size
, align
) != 0) {
3534 /* Transform logical writes to be a full physical block size. */
3535 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3536 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3537 ASSERT(vd
== vd
->vdev_top
);
3538 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3539 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3540 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3542 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3546 * If this is not a physical io, make sure that it is properly aligned
3547 * before proceeding.
3549 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3550 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3551 ASSERT0(P2PHASE(zio
->io_size
, align
));
3554 * For physical writes, we allow 512b aligned writes and assume
3555 * the device will perform a read-modify-write as necessary.
3557 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3558 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3561 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
3564 * If this is a repair I/O, and there's no self-healing involved --
3565 * that is, we're just resilvering what we expect to resilver --
3566 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3567 * This prevents spurious resilvering.
3569 * There are a few ways that we can end up creating these spurious
3572 * 1. A resilver i/o will be issued if any DVA in the BP has a
3573 * dirty DTL. The mirror code will issue resilver writes to
3574 * each DVA, including the one(s) that are not on vdevs with dirty
3577 * 2. With nested replication, which happens when we have a
3578 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3579 * For example, given mirror(replacing(A+B), C), it's likely that
3580 * only A is out of date (it's the new device). In this case, we'll
3581 * read from C, then use the data to resilver A+B -- but we don't
3582 * actually want to resilver B, just A. The top-level mirror has no
3583 * way to know this, so instead we just discard unnecessary repairs
3584 * as we work our way down the vdev tree.
3586 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3587 * The same logic applies to any form of nested replication: ditto
3588 * + mirror, RAID-Z + replacing, etc.
3590 * However, indirect vdevs point off to other vdevs which may have
3591 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3592 * will be properly bypassed instead.
3594 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
3595 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
3596 zio
->io_txg
!= 0 && /* not a delegated i/o */
3597 vd
->vdev_ops
!= &vdev_indirect_ops
&&
3598 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
3599 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3600 zio_vdev_io_bypass(zio
);
3604 if (vd
->vdev_ops
->vdev_op_leaf
&& (zio
->io_type
== ZIO_TYPE_READ
||
3605 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
)) {
3607 if (zio
->io_type
== ZIO_TYPE_READ
&& vdev_cache_read(zio
))
3610 if ((zio
= vdev_queue_io(zio
)) == NULL
)
3613 if (!vdev_accessible(vd
, zio
)) {
3614 zio
->io_error
= SET_ERROR(ENXIO
);
3618 zio
->io_delay
= gethrtime();
3621 vd
->vdev_ops
->vdev_op_io_start(zio
);
3626 zio_vdev_io_done(zio_t
*zio
)
3628 vdev_t
*vd
= zio
->io_vd
;
3629 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
3630 boolean_t unexpected_error
= B_FALSE
;
3632 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3636 ASSERT(zio
->io_type
== ZIO_TYPE_READ
||
3637 zio
->io_type
== ZIO_TYPE_WRITE
|| zio
->io_type
== ZIO_TYPE_TRIM
);
3640 zio
->io_delay
= gethrtime() - zio
->io_delay
;
3642 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
) {
3644 vdev_queue_io_done(zio
);
3646 if (zio
->io_type
== ZIO_TYPE_WRITE
)
3647 vdev_cache_write(zio
);
3649 if (zio_injection_enabled
&& zio
->io_error
== 0)
3650 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
3653 if (zio_injection_enabled
&& zio
->io_error
== 0)
3654 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
3656 if (zio
->io_error
&& zio
->io_type
!= ZIO_TYPE_TRIM
) {
3657 if (!vdev_accessible(vd
, zio
)) {
3658 zio
->io_error
= SET_ERROR(ENXIO
);
3660 unexpected_error
= B_TRUE
;
3665 ops
->vdev_op_io_done(zio
);
3667 if (unexpected_error
)
3668 VERIFY(vdev_probe(vd
, zio
) == NULL
);
3674 * This function is used to change the priority of an existing zio that is
3675 * currently in-flight. This is used by the arc to upgrade priority in the
3676 * event that a demand read is made for a block that is currently queued
3677 * as a scrub or async read IO. Otherwise, the high priority read request
3678 * would end up having to wait for the lower priority IO.
3681 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
3683 zio_t
*cio
, *cio_next
;
3684 zio_link_t
*zl
= NULL
;
3686 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
3688 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
3689 vdev_queue_change_io_priority(pio
, priority
);
3691 pio
->io_priority
= priority
;
3694 mutex_enter(&pio
->io_lock
);
3695 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
3696 cio_next
= zio_walk_children(pio
, &zl
);
3697 zio_change_priority(cio
, priority
);
3699 mutex_exit(&pio
->io_lock
);
3703 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3704 * disk, and use that to finish the checksum ereport later.
3707 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
3708 const abd_t
*good_buf
)
3710 /* no processing needed */
3711 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
3716 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
, void *ignored
)
3718 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
3720 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
3722 zcr
->zcr_cbinfo
= zio
->io_size
;
3723 zcr
->zcr_cbdata
= abd
;
3724 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
3725 zcr
->zcr_free
= zio_abd_free
;
3729 zio_vdev_io_assess(zio_t
*zio
)
3731 vdev_t
*vd
= zio
->io_vd
;
3733 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
3737 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3738 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
3740 if (zio
->io_vsd
!= NULL
) {
3741 zio
->io_vsd_ops
->vsd_free(zio
);
3745 if (zio_injection_enabled
&& zio
->io_error
== 0)
3746 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
3749 * If the I/O failed, determine whether we should attempt to retry it.
3751 * On retry, we cut in line in the issue queue, since we don't want
3752 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3754 if (zio
->io_error
&& vd
== NULL
&&
3755 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
3756 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
3757 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
3759 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
|
3760 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_DONT_AGGREGATE
;
3761 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
3762 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
3763 zio_requeue_io_start_cut_in_line
);
3768 * If we got an error on a leaf device, convert it to ENXIO
3769 * if the device is not accessible at all.
3771 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3772 !vdev_accessible(vd
, zio
))
3773 zio
->io_error
= SET_ERROR(ENXIO
);
3776 * If we can't write to an interior vdev (mirror or RAID-Z),
3777 * set vdev_cant_write so that we stop trying to allocate from it.
3779 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
3780 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
3781 vd
->vdev_cant_write
= B_TRUE
;
3785 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3786 * attempts will ever succeed. In this case we set a persistent
3787 * boolean flag so that we don't bother with it in the future.
3789 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
3790 zio
->io_type
== ZIO_TYPE_IOCTL
&&
3791 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
3792 vd
->vdev_nowritecache
= B_TRUE
;
3795 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3797 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
3798 zio
->io_physdone
!= NULL
) {
3799 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DELEGATED
));
3800 ASSERT(zio
->io_child_type
== ZIO_CHILD_VDEV
);
3801 zio
->io_physdone(zio
->io_logical
);
3808 zio_vdev_io_reissue(zio_t
*zio
)
3810 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3811 ASSERT(zio
->io_error
== 0);
3813 zio
->io_stage
>>= 1;
3817 zio_vdev_io_redone(zio_t
*zio
)
3819 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
3821 zio
->io_stage
>>= 1;
3825 zio_vdev_io_bypass(zio_t
*zio
)
3827 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
3828 ASSERT(zio
->io_error
== 0);
3830 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
3831 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
3835 * ==========================================================================
3836 * Encrypt and store encryption parameters
3837 * ==========================================================================
3842 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3843 * managing the storage of encryption parameters and passing them to the
3844 * lower-level encryption functions.
3847 zio_encrypt(zio_t
*zio
)
3849 zio_prop_t
*zp
= &zio
->io_prop
;
3850 spa_t
*spa
= zio
->io_spa
;
3851 blkptr_t
*bp
= zio
->io_bp
;
3852 uint64_t psize
= BP_GET_PSIZE(bp
);
3853 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
3854 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
3855 void *enc_buf
= NULL
;
3857 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3858 uint8_t iv
[ZIO_DATA_IV_LEN
];
3859 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3860 boolean_t no_crypt
= B_FALSE
;
3862 /* the root zio already encrypted the data */
3863 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
3866 /* only ZIL blocks are re-encrypted on rewrite */
3867 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
3870 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
3871 BP_SET_CRYPT(bp
, B_FALSE
);
3875 /* if we are doing raw encryption set the provided encryption params */
3876 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
3877 ASSERT0(BP_GET_LEVEL(bp
));
3878 BP_SET_CRYPT(bp
, B_TRUE
);
3879 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
3880 if (ot
!= DMU_OT_OBJSET
)
3881 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
3883 /* dnode blocks must be written out in the provided byteorder */
3884 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
3885 ot
== DMU_OT_DNODE
) {
3886 void *bswap_buf
= zio_buf_alloc(psize
);
3887 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
3889 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3890 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
3891 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
3894 abd_take_ownership_of_buf(babd
, B_TRUE
);
3895 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
3898 if (DMU_OT_IS_ENCRYPTED(ot
))
3899 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
3903 /* indirect blocks only maintain a cksum of the lower level MACs */
3904 if (BP_GET_LEVEL(bp
) > 0) {
3905 BP_SET_CRYPT(bp
, B_TRUE
);
3906 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
3907 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
3909 zio_crypt_encode_mac_bp(bp
, mac
);
3914 * Objset blocks are a special case since they have 2 256-bit MACs
3915 * embedded within them.
3917 if (ot
== DMU_OT_OBJSET
) {
3918 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
3919 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
3920 BP_SET_CRYPT(bp
, B_TRUE
);
3921 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
3922 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
3926 /* unencrypted object types are only authenticated with a MAC */
3927 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
3928 BP_SET_CRYPT(bp
, B_TRUE
);
3929 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
3930 zio
->io_abd
, psize
, mac
));
3931 zio_crypt_encode_mac_bp(bp
, mac
);
3936 * Later passes of sync-to-convergence may decide to rewrite data
3937 * in place to avoid more disk reallocations. This presents a problem
3938 * for encryption because this constitutes rewriting the new data with
3939 * the same encryption key and IV. However, this only applies to blocks
3940 * in the MOS (particularly the spacemaps) and we do not encrypt the
3941 * MOS. We assert that the zio is allocating or an intent log write
3944 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
3945 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
3946 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
3947 ASSERT3U(psize
, !=, 0);
3949 enc_buf
= zio_buf_alloc(psize
);
3950 eabd
= abd_get_from_buf(enc_buf
, psize
);
3951 abd_take_ownership_of_buf(eabd
, B_TRUE
);
3954 * For an explanation of what encryption parameters are stored
3955 * where, see the block comment in zio_crypt.c.
3957 if (ot
== DMU_OT_INTENT_LOG
) {
3958 zio_crypt_decode_params_bp(bp
, salt
, iv
);
3960 BP_SET_CRYPT(bp
, B_TRUE
);
3963 /* Perform the encryption. This should not fail */
3964 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
3965 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
3966 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
3968 /* encode encryption metadata into the bp */
3969 if (ot
== DMU_OT_INTENT_LOG
) {
3971 * ZIL blocks store the MAC in the embedded checksum, so the
3972 * transform must always be applied.
3974 zio_crypt_encode_mac_zil(enc_buf
, mac
);
3975 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3977 BP_SET_CRYPT(bp
, B_TRUE
);
3978 zio_crypt_encode_params_bp(bp
, salt
, iv
);
3979 zio_crypt_encode_mac_bp(bp
, mac
);
3982 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
3985 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
3993 * ==========================================================================
3994 * Generate and verify checksums
3995 * ==========================================================================
3998 zio_checksum_generate(zio_t
*zio
)
4000 blkptr_t
*bp
= zio
->io_bp
;
4001 enum zio_checksum checksum
;
4005 * This is zio_write_phys().
4006 * We're either generating a label checksum, or none at all.
4008 checksum
= zio
->io_prop
.zp_checksum
;
4010 if (checksum
== ZIO_CHECKSUM_OFF
)
4013 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4015 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4016 ASSERT(!IO_IS_ALLOCATING(zio
));
4017 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4019 checksum
= BP_GET_CHECKSUM(bp
);
4023 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4029 zio_checksum_verify(zio_t
*zio
)
4031 zio_bad_cksum_t info
;
4032 blkptr_t
*bp
= zio
->io_bp
;
4035 ASSERT(zio
->io_vd
!= NULL
);
4039 * This is zio_read_phys().
4040 * We're either verifying a label checksum, or nothing at all.
4042 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4045 ASSERT(zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_LABEL
);
4048 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4049 zio
->io_error
= error
;
4050 if (error
== ECKSUM
&&
4051 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4052 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4053 zio
->io_vd
->vdev_stat
.vs_checksum_errors
++;
4054 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4056 zfs_ereport_start_checksum(zio
->io_spa
,
4057 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4058 zio
->io_offset
, zio
->io_size
, NULL
, &info
);
4066 * Called by RAID-Z to ensure we don't compute the checksum twice.
4069 zio_checksum_verified(zio_t
*zio
)
4071 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4075 * ==========================================================================
4076 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4077 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4078 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4079 * indicate errors that are specific to one I/O, and most likely permanent.
4080 * Any other error is presumed to be worse because we weren't expecting it.
4081 * ==========================================================================
4084 zio_worst_error(int e1
, int e2
)
4086 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4089 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4090 if (e1
== zio_error_rank
[r1
])
4093 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4094 if (e2
== zio_error_rank
[r2
])
4097 return (r1
> r2
? e1
: e2
);
4101 * ==========================================================================
4103 * ==========================================================================
4106 zio_ready(zio_t
*zio
)
4108 blkptr_t
*bp
= zio
->io_bp
;
4109 zio_t
*pio
, *pio_next
;
4110 zio_link_t
*zl
= NULL
;
4112 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
,
4117 if (zio
->io_ready
) {
4118 ASSERT(IO_IS_ALLOCATING(zio
));
4119 ASSERT(bp
->blk_birth
== zio
->io_txg
|| BP_IS_HOLE(bp
) ||
4120 (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4121 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4126 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4127 zio
->io_bp_copy
= *bp
;
4129 if (zio
->io_error
!= 0) {
4130 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4132 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4133 ASSERT(IO_IS_ALLOCATING(zio
));
4134 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4135 ASSERT(zio
->io_metaslab_class
!= NULL
);
4138 * We were unable to allocate anything, unreserve and
4139 * issue the next I/O to allocate.
4141 metaslab_class_throttle_unreserve(
4142 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4143 zio
->io_allocator
, zio
);
4144 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4148 mutex_enter(&zio
->io_lock
);
4149 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4150 pio
= zio_walk_parents(zio
, &zl
);
4151 mutex_exit(&zio
->io_lock
);
4154 * As we notify zio's parents, new parents could be added.
4155 * New parents go to the head of zio's io_parent_list, however,
4156 * so we will (correctly) not notify them. The remainder of zio's
4157 * io_parent_list, from 'pio_next' onward, cannot change because
4158 * all parents must wait for us to be done before they can be done.
4160 for (; pio
!= NULL
; pio
= pio_next
) {
4161 pio_next
= zio_walk_parents(zio
, &zl
);
4162 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4165 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4166 if (BP_IS_GANG(bp
)) {
4167 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4169 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4170 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4174 if (zio_injection_enabled
&&
4175 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4176 zio_handle_ignored_writes(zio
);
4182 * Update the allocation throttle accounting.
4185 zio_dva_throttle_done(zio_t
*zio
)
4187 zio_t
*lio __maybe_unused
= zio
->io_logical
;
4188 zio_t
*pio
= zio_unique_parent(zio
);
4189 vdev_t
*vd
= zio
->io_vd
;
4190 int flags
= METASLAB_ASYNC_ALLOC
;
4192 ASSERT3P(zio
->io_bp
, !=, NULL
);
4193 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4194 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4195 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4197 ASSERT3P(vd
, ==, vd
->vdev_top
);
4198 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4199 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4200 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4201 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4202 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4205 * Parents of gang children can have two flavors -- ones that
4206 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4207 * and ones that allocated the constituent blocks. The allocation
4208 * throttle needs to know the allocating parent zio so we must find
4211 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4213 * If our parent is a rewrite gang child then our grandparent
4214 * would have been the one that performed the allocation.
4216 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4217 pio
= zio_unique_parent(pio
);
4218 flags
|= METASLAB_GANG_CHILD
;
4221 ASSERT(IO_IS_ALLOCATING(pio
));
4222 ASSERT3P(zio
, !=, zio
->io_logical
);
4223 ASSERT(zio
->io_logical
!= NULL
);
4224 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4225 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4226 ASSERT(zio
->io_metaslab_class
!= NULL
);
4228 mutex_enter(&pio
->io_lock
);
4229 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4230 pio
->io_allocator
, B_TRUE
);
4231 mutex_exit(&pio
->io_lock
);
4233 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4234 pio
->io_allocator
, pio
);
4237 * Call into the pipeline to see if there is more work that
4238 * needs to be done. If there is work to be done it will be
4239 * dispatched to another taskq thread.
4241 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4245 zio_done(zio_t
*zio
)
4248 * Always attempt to keep stack usage minimal here since
4249 * we can be called recursively up to 19 levels deep.
4251 const uint64_t psize
= zio
->io_size
;
4252 zio_t
*pio
, *pio_next
;
4253 zio_link_t
*zl
= NULL
;
4256 * If our children haven't all completed,
4257 * wait for them and then repeat this pipeline stage.
4259 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4264 * If the allocation throttle is enabled, then update the accounting.
4265 * We only track child I/Os that are part of an allocating async
4266 * write. We must do this since the allocation is performed
4267 * by the logical I/O but the actual write is done by child I/Os.
4269 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4270 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4271 ASSERT(zio
->io_metaslab_class
!= NULL
);
4272 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4273 zio_dva_throttle_done(zio
);
4277 * If the allocation throttle is enabled, verify that
4278 * we have decremented the refcounts for every I/O that was throttled.
4280 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4281 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4282 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4283 ASSERT(zio
->io_bp
!= NULL
);
4285 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4287 VERIFY(zfs_refcount_not_held(
4288 &zio
->io_metaslab_class
->mc_alloc_slots
[zio
->io_allocator
],
4293 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4294 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4295 ASSERT(zio
->io_children
[c
][w
] == 0);
4297 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4298 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4299 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4300 ASSERT(bcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4301 sizeof (blkptr_t
)) == 0 ||
4302 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4303 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4304 zio
->io_bp_override
== NULL
&&
4305 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4306 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4307 BP_GET_NDVAS(zio
->io_bp
));
4308 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4309 (BP_COUNT_GANG(zio
->io_bp
) ==
4310 BP_GET_NDVAS(zio
->io_bp
)));
4312 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4313 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4317 * If there were child vdev/gang/ddt errors, they apply to us now.
4319 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4320 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4321 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4324 * If the I/O on the transformed data was successful, generate any
4325 * checksum reports now while we still have the transformed data.
4327 if (zio
->io_error
== 0) {
4328 while (zio
->io_cksum_report
!= NULL
) {
4329 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4330 uint64_t align
= zcr
->zcr_align
;
4331 uint64_t asize
= P2ROUNDUP(psize
, align
);
4332 abd_t
*adata
= zio
->io_abd
;
4334 if (asize
!= psize
) {
4335 adata
= abd_alloc(asize
, B_TRUE
);
4336 abd_copy(adata
, zio
->io_abd
, psize
);
4337 abd_zero_off(adata
, psize
, asize
- psize
);
4340 zio
->io_cksum_report
= zcr
->zcr_next
;
4341 zcr
->zcr_next
= NULL
;
4342 zcr
->zcr_finish(zcr
, adata
);
4343 zfs_ereport_free_checksum(zcr
);
4350 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4352 vdev_stat_update(zio
, psize
);
4355 * If this I/O is attached to a particular vdev is slow, exceeding
4356 * 30 seconds to complete, post an error described the I/O delay.
4357 * We ignore these errors if the device is currently unavailable.
4359 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4360 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4362 * We want to only increment our slow IO counters if
4363 * the IO is valid (i.e. not if the drive is removed).
4365 * zfs_ereport_post() will also do these checks, but
4366 * it can also ratelimit and have other failures, so we
4367 * need to increment the slow_io counters independent
4370 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4371 zio
->io_spa
, zio
->io_vd
, zio
)) {
4372 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4373 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4374 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4376 zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4377 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4383 if (zio
->io_error
) {
4385 * If this I/O is attached to a particular vdev,
4386 * generate an error message describing the I/O failure
4387 * at the block level. We ignore these errors if the
4388 * device is currently unavailable.
4390 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4391 !vdev_is_dead(zio
->io_vd
)) {
4392 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4393 if (zio
->io_type
== ZIO_TYPE_READ
) {
4394 zio
->io_vd
->vdev_stat
.vs_read_errors
++;
4395 } else if (zio
->io_type
== ZIO_TYPE_WRITE
) {
4396 zio
->io_vd
->vdev_stat
.vs_write_errors
++;
4398 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4400 zfs_ereport_post(FM_EREPORT_ZFS_IO
, zio
->io_spa
,
4401 zio
->io_vd
, &zio
->io_bookmark
, zio
, 0, 0);
4404 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4405 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4406 zio
== zio
->io_logical
) {
4408 * For logical I/O requests, tell the SPA to log the
4409 * error and generate a logical data ereport.
4411 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
4412 zfs_ereport_post(FM_EREPORT_ZFS_DATA
, zio
->io_spa
,
4413 NULL
, &zio
->io_bookmark
, zio
, 0, 0);
4417 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4419 * Determine whether zio should be reexecuted. This will
4420 * propagate all the way to the root via zio_notify_parent().
4422 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4423 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4425 if (IO_IS_ALLOCATING(zio
) &&
4426 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4427 if (zio
->io_error
!= ENOSPC
)
4428 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4430 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4433 if ((zio
->io_type
== ZIO_TYPE_READ
||
4434 zio
->io_type
== ZIO_TYPE_FREE
) &&
4435 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4436 zio
->io_error
== ENXIO
&&
4437 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4438 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4439 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4441 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4442 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4445 * Here is a possibly good place to attempt to do
4446 * either combinatorial reconstruction or error correction
4447 * based on checksums. It also might be a good place
4448 * to send out preliminary ereports before we suspend
4454 * If there were logical child errors, they apply to us now.
4455 * We defer this until now to avoid conflating logical child
4456 * errors with errors that happened to the zio itself when
4457 * updating vdev stats and reporting FMA events above.
4459 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4461 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4462 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4463 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4464 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4466 zio_gang_tree_free(&zio
->io_gang_tree
);
4469 * Godfather I/Os should never suspend.
4471 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4472 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4473 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4475 if (zio
->io_reexecute
) {
4477 * This is a logical I/O that wants to reexecute.
4479 * Reexecute is top-down. When an i/o fails, if it's not
4480 * the root, it simply notifies its parent and sticks around.
4481 * The parent, seeing that it still has children in zio_done(),
4482 * does the same. This percolates all the way up to the root.
4483 * The root i/o will reexecute or suspend the entire tree.
4485 * This approach ensures that zio_reexecute() honors
4486 * all the original i/o dependency relationships, e.g.
4487 * parents not executing until children are ready.
4489 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4491 zio
->io_gang_leader
= NULL
;
4493 mutex_enter(&zio
->io_lock
);
4494 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4495 mutex_exit(&zio
->io_lock
);
4498 * "The Godfather" I/O monitors its children but is
4499 * not a true parent to them. It will track them through
4500 * the pipeline but severs its ties whenever they get into
4501 * trouble (e.g. suspended). This allows "The Godfather"
4502 * I/O to return status without blocking.
4505 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4507 zio_link_t
*remove_zl
= zl
;
4508 pio_next
= zio_walk_parents(zio
, &zl
);
4510 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4511 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4512 zio_remove_child(pio
, zio
, remove_zl
);
4514 * This is a rare code path, so we don't
4515 * bother with "next_to_execute".
4517 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4522 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4524 * We're not a root i/o, so there's nothing to do
4525 * but notify our parent. Don't propagate errors
4526 * upward since we haven't permanently failed yet.
4528 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4529 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4531 * This is a rare code path, so we don't bother with
4532 * "next_to_execute".
4534 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4535 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4537 * We'd fail again if we reexecuted now, so suspend
4538 * until conditions improve (e.g. device comes online).
4540 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4543 * Reexecution is potentially a huge amount of work.
4544 * Hand it off to the otherwise-unused claim taskq.
4546 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4547 spa_taskq_dispatch_ent(zio
->io_spa
,
4548 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4549 (task_func_t
*)zio_reexecute
, zio
, 0,
4555 ASSERT(zio
->io_child_count
== 0);
4556 ASSERT(zio
->io_reexecute
== 0);
4557 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
4560 * Report any checksum errors, since the I/O is complete.
4562 while (zio
->io_cksum_report
!= NULL
) {
4563 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4564 zio
->io_cksum_report
= zcr
->zcr_next
;
4565 zcr
->zcr_next
= NULL
;
4566 zcr
->zcr_finish(zcr
, NULL
);
4567 zfs_ereport_free_checksum(zcr
);
4570 if (zio
->io_flags
& ZIO_FLAG_FASTWRITE
&& zio
->io_bp
&&
4571 !BP_IS_HOLE(zio
->io_bp
) && !BP_IS_EMBEDDED(zio
->io_bp
) &&
4572 !(zio
->io_flags
& ZIO_FLAG_NOPWRITE
)) {
4573 metaslab_fastwrite_unmark(zio
->io_spa
, zio
->io_bp
);
4577 * It is the responsibility of the done callback to ensure that this
4578 * particular zio is no longer discoverable for adoption, and as
4579 * such, cannot acquire any new parents.
4584 mutex_enter(&zio
->io_lock
);
4585 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4586 mutex_exit(&zio
->io_lock
);
4589 * We are done executing this zio. We may want to execute a parent
4590 * next. See the comment in zio_notify_parent().
4592 zio_t
*next_to_execute
= NULL
;
4594 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
4595 zio_link_t
*remove_zl
= zl
;
4596 pio_next
= zio_walk_parents(zio
, &zl
);
4597 zio_remove_child(pio
, zio
, remove_zl
);
4598 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
4601 if (zio
->io_waiter
!= NULL
) {
4602 mutex_enter(&zio
->io_lock
);
4603 zio
->io_executor
= NULL
;
4604 cv_broadcast(&zio
->io_cv
);
4605 mutex_exit(&zio
->io_lock
);
4610 return (next_to_execute
);
4614 * ==========================================================================
4615 * I/O pipeline definition
4616 * ==========================================================================
4618 static zio_pipe_stage_t
*zio_pipeline
[] = {
4626 zio_checksum_generate
,
4642 zio_checksum_verify
,
4650 * Compare two zbookmark_phys_t's to see which we would reach first in a
4651 * pre-order traversal of the object tree.
4653 * This is simple in every case aside from the meta-dnode object. For all other
4654 * objects, we traverse them in order (object 1 before object 2, and so on).
4655 * However, all of these objects are traversed while traversing object 0, since
4656 * the data it points to is the list of objects. Thus, we need to convert to a
4657 * canonical representation so we can compare meta-dnode bookmarks to
4658 * non-meta-dnode bookmarks.
4660 * We do this by calculating "equivalents" for each field of the zbookmark.
4661 * zbookmarks outside of the meta-dnode use their own object and level, and
4662 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4663 * blocks this bookmark refers to) by multiplying their blkid by their span
4664 * (the number of L0 blocks contained within one block at their level).
4665 * zbookmarks inside the meta-dnode calculate their object equivalent
4666 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4667 * level + 1<<31 (any value larger than a level could ever be) for their level.
4668 * This causes them to always compare before a bookmark in their object
4669 * equivalent, compare appropriately to bookmarks in other objects, and to
4670 * compare appropriately to other bookmarks in the meta-dnode.
4673 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
4674 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
4677 * These variables represent the "equivalent" values for the zbookmark,
4678 * after converting zbookmarks inside the meta dnode to their
4679 * normal-object equivalents.
4681 uint64_t zb1obj
, zb2obj
;
4682 uint64_t zb1L0
, zb2L0
;
4683 uint64_t zb1level
, zb2level
;
4685 if (zb1
->zb_object
== zb2
->zb_object
&&
4686 zb1
->zb_level
== zb2
->zb_level
&&
4687 zb1
->zb_blkid
== zb2
->zb_blkid
)
4690 IMPLY(zb1
->zb_level
> 0, ibs1
>= SPA_MINBLOCKSHIFT
);
4691 IMPLY(zb2
->zb_level
> 0, ibs2
>= SPA_MINBLOCKSHIFT
);
4694 * BP_SPANB calculates the span in blocks.
4696 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
4697 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
4699 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
4700 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4702 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
4704 zb1obj
= zb1
->zb_object
;
4705 zb1level
= zb1
->zb_level
;
4708 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
4709 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
4711 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
4713 zb2obj
= zb2
->zb_object
;
4714 zb2level
= zb2
->zb_level
;
4717 /* Now that we have a canonical representation, do the comparison. */
4718 if (zb1obj
!= zb2obj
)
4719 return (zb1obj
< zb2obj
? -1 : 1);
4720 else if (zb1L0
!= zb2L0
)
4721 return (zb1L0
< zb2L0
? -1 : 1);
4722 else if (zb1level
!= zb2level
)
4723 return (zb1level
> zb2level
? -1 : 1);
4725 * This can (theoretically) happen if the bookmarks have the same object
4726 * and level, but different blkids, if the block sizes are not the same.
4727 * There is presently no way to change the indirect block sizes
4733 * This function checks the following: given that last_block is the place that
4734 * our traversal stopped last time, does that guarantee that we've visited
4735 * every node under subtree_root? Therefore, we can't just use the raw output
4736 * of zbookmark_compare. We have to pass in a modified version of
4737 * subtree_root; by incrementing the block id, and then checking whether
4738 * last_block is before or equal to that, we can tell whether or not having
4739 * visited last_block implies that all of subtree_root's children have been
4743 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
4744 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
4746 zbookmark_phys_t mod_zb
= *subtree_root
;
4748 ASSERT(last_block
->zb_level
== 0);
4750 /* The objset_phys_t isn't before anything. */
4755 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4756 * data block size in sectors, because that variable is only used if
4757 * the bookmark refers to a block in the meta-dnode. Since we don't
4758 * know without examining it what object it refers to, and there's no
4759 * harm in passing in this value in other cases, we always pass it in.
4761 * We pass in 0 for the indirect block size shift because zb2 must be
4762 * level 0. The indirect block size is only used to calculate the span
4763 * of the bookmark, but since the bookmark must be level 0, the span is
4764 * always 1, so the math works out.
4766 * If you make changes to how the zbookmark_compare code works, be sure
4767 * to make sure that this code still works afterwards.
4769 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
4770 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
4774 EXPORT_SYMBOL(zio_type_name
);
4775 EXPORT_SYMBOL(zio_buf_alloc
);
4776 EXPORT_SYMBOL(zio_data_buf_alloc
);
4777 EXPORT_SYMBOL(zio_buf_free
);
4778 EXPORT_SYMBOL(zio_data_buf_free
);
4781 ZFS_MODULE_PARAM(zfs_zio
, zio_
, slow_io_ms
, INT
, ZMOD_RW
,
4782 "Max I/O completion time (milliseconds) before marking it as slow");
4784 ZFS_MODULE_PARAM(zfs_zio
, zio_
, requeue_io_start_cut_in_line
, INT
, ZMOD_RW
,
4785 "Prioritize requeued I/O");
4787 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_deferred_free
, INT
, ZMOD_RW
,
4788 "Defer frees starting in this pass");
4790 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_dont_compress
, INT
, ZMOD_RW
,
4791 "Don't compress starting in this pass");
4793 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_rewrite
, INT
, ZMOD_RW
,
4794 "Rewrite new bps starting in this pass");
4796 ZFS_MODULE_PARAM(zfs_zio
, zio_
, dva_throttle_enabled
, INT
, ZMOD_RW
,
4797 "Throttle block allocations in the ZIO pipeline");
4799 ZFS_MODULE_PARAM(zfs_zio
, zio_
, deadman_log_all
, INT
, ZMOD_RW
,
4800 "Log all slow ZIOs, not just those with vdevs");