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 https://opensource.org/licenses/CDDL-1.0.
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, 2022 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
26 * Copyright (c) 2019, Klara Inc.
27 * Copyright (c) 2019, Allan Jude
28 * Copyright (c) 2021, Datto, Inc.
31 #include <sys/sysmacros.h>
32 #include <sys/zfs_context.h>
33 #include <sys/fm/fs/zfs.h>
36 #include <sys/spa_impl.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/vdev_trim.h>
39 #include <sys/zio_impl.h>
40 #include <sys/zio_compress.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/dmu_objset.h>
46 #include <sys/blkptr.h>
47 #include <sys/zfeature.h>
48 #include <sys/dsl_scan.h>
49 #include <sys/metaslab_impl.h>
51 #include <sys/trace_zfs.h>
53 #include <sys/dsl_crypt.h>
57 * ==========================================================================
58 * I/O type descriptions
59 * ==========================================================================
61 const char *const zio_type_name
[ZIO_TYPES
] = {
63 * Note: Linux kernel thread name length is limited
64 * so these names will differ from upstream open zfs.
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
69 int zio_dva_throttle_enabled
= B_TRUE
;
70 static int zio_deadman_log_all
= B_FALSE
;
73 * ==========================================================================
75 * ==========================================================================
77 static kmem_cache_t
*zio_cache
;
78 static kmem_cache_t
*zio_link_cache
;
79 kmem_cache_t
*zio_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
80 kmem_cache_t
*zio_data_buf_cache
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
81 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
82 static uint64_t zio_buf_cache_allocs
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
83 static uint64_t zio_buf_cache_frees
[SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
];
86 /* Mark IOs as "slow" if they take longer than 30 seconds */
87 static uint_t zio_slow_io_ms
= (30 * MILLISEC
);
89 #define BP_SPANB(indblkshift, level) \
90 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
91 #define COMPARE_META_LEVEL 0x80000000ul
93 * The following actions directly effect the spa's sync-to-convergence logic.
94 * The values below define the sync pass when we start performing the action.
95 * Care should be taken when changing these values as they directly impact
96 * spa_sync() performance. Tuning these values may introduce subtle performance
97 * pathologies and should only be done in the context of performance analysis.
98 * These tunables will eventually be removed and replaced with #defines once
99 * enough analysis has been done to determine optimal values.
101 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
102 * regular blocks are not deferred.
104 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
105 * compression (including of metadata). In practice, we don't have this
106 * many sync passes, so this has no effect.
108 * The original intent was that disabling compression would help the sync
109 * passes to converge. However, in practice disabling compression increases
110 * the average number of sync passes, because when we turn compression off, a
111 * lot of block's size will change and thus we have to re-allocate (not
112 * overwrite) them. It also increases the number of 128KB allocations (e.g.
113 * for indirect blocks and spacemaps) because these will not be compressed.
114 * The 128K allocations are especially detrimental to performance on highly
115 * fragmented systems, which may have very few free segments of this size,
116 * and may need to load new metaslabs to satisfy 128K allocations.
119 /* defer frees starting in this pass */
120 uint_t zfs_sync_pass_deferred_free
= 2;
122 /* don't compress starting in this pass */
123 static uint_t zfs_sync_pass_dont_compress
= 8;
125 /* rewrite new bps starting in this pass */
126 static uint_t zfs_sync_pass_rewrite
= 2;
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 * Enable smaller cores by excluding metadata
136 * allocations as well.
138 int zio_exclude_metadata
= 0;
139 static int zio_requeue_io_start_cut_in_line
= 1;
142 static const int zio_buf_debug_limit
= 16384;
144 static const int zio_buf_debug_limit
= 0;
147 static inline void __zio_execute(zio_t
*zio
);
149 static void zio_taskq_dispatch(zio_t
*, zio_taskq_type_t
, boolean_t
);
156 zio_cache
= kmem_cache_create("zio_cache",
157 sizeof (zio_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
158 zio_link_cache
= kmem_cache_create("zio_link_cache",
159 sizeof (zio_link_t
), 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
161 for (c
= 0; c
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; c
++) {
162 size_t size
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
163 size_t align
, cflags
, data_cflags
;
167 * Create cache for each half-power of 2 size, starting from
168 * SPA_MINBLOCKSIZE. It should give us memory space efficiency
169 * of ~7/8, sufficient for transient allocations mostly using
175 if (!IS_P2ALIGNED(size
, p2
/ 2))
180 * If we are using watchpoints, put each buffer on its own page,
181 * to eliminate the performance overhead of trapping to the
182 * kernel when modifying a non-watched buffer that shares the
183 * page with a watched buffer.
185 if (arc_watch
&& !IS_P2ALIGNED(size
, PAGESIZE
))
189 if (IS_P2ALIGNED(size
, PAGESIZE
))
192 align
= 1 << (highbit64(size
^ (size
- 1)) - 1);
194 cflags
= (zio_exclude_metadata
|| size
> zio_buf_debug_limit
) ?
196 data_cflags
= KMC_NODEBUG
;
197 if (cflags
== data_cflags
) {
199 * Resulting kmem caches would be identical.
200 * Save memory by creating only one.
202 (void) snprintf(name
, sizeof (name
),
203 "zio_buf_comb_%lu", (ulong_t
)size
);
204 zio_buf_cache
[c
] = kmem_cache_create(name
, size
, align
,
205 NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
206 zio_data_buf_cache
[c
] = zio_buf_cache
[c
];
209 (void) snprintf(name
, sizeof (name
), "zio_buf_%lu",
211 zio_buf_cache
[c
] = kmem_cache_create(name
, size
, align
,
212 NULL
, NULL
, NULL
, NULL
, NULL
, cflags
);
214 (void) snprintf(name
, sizeof (name
), "zio_data_buf_%lu",
216 zio_data_buf_cache
[c
] = kmem_cache_create(name
, size
, align
,
217 NULL
, NULL
, NULL
, NULL
, NULL
, data_cflags
);
221 ASSERT(zio_buf_cache
[c
] != NULL
);
222 if (zio_buf_cache
[c
- 1] == NULL
)
223 zio_buf_cache
[c
- 1] = zio_buf_cache
[c
];
225 ASSERT(zio_data_buf_cache
[c
] != NULL
);
226 if (zio_data_buf_cache
[c
- 1] == NULL
)
227 zio_data_buf_cache
[c
- 1] = zio_data_buf_cache
[c
];
238 size_t n
= SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
;
240 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
241 for (size_t i
= 0; i
< n
; i
++) {
242 if (zio_buf_cache_allocs
[i
] != zio_buf_cache_frees
[i
])
243 (void) printf("zio_fini: [%d] %llu != %llu\n",
244 (int)((i
+ 1) << SPA_MINBLOCKSHIFT
),
245 (long long unsigned)zio_buf_cache_allocs
[i
],
246 (long long unsigned)zio_buf_cache_frees
[i
]);
251 * The same kmem cache can show up multiple times in both zio_buf_cache
252 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
255 for (size_t i
= 0; i
< n
; i
++) {
256 kmem_cache_t
*cache
= zio_buf_cache
[i
];
259 for (size_t j
= i
; j
< n
; j
++) {
260 if (cache
== zio_buf_cache
[j
])
261 zio_buf_cache
[j
] = NULL
;
262 if (cache
== zio_data_buf_cache
[j
])
263 zio_data_buf_cache
[j
] = NULL
;
265 kmem_cache_destroy(cache
);
268 for (size_t i
= 0; i
< n
; i
++) {
269 kmem_cache_t
*cache
= zio_data_buf_cache
[i
];
272 for (size_t j
= i
; j
< n
; j
++) {
273 if (cache
== zio_data_buf_cache
[j
])
274 zio_data_buf_cache
[j
] = NULL
;
276 kmem_cache_destroy(cache
);
279 for (size_t i
= 0; i
< n
; i
++) {
280 VERIFY3P(zio_buf_cache
[i
], ==, NULL
);
281 VERIFY3P(zio_data_buf_cache
[i
], ==, NULL
);
284 kmem_cache_destroy(zio_link_cache
);
285 kmem_cache_destroy(zio_cache
);
293 * ==========================================================================
294 * Allocate and free I/O buffers
295 * ==========================================================================
299 static const ulong_t zio_buf_canary
= (ulong_t
)0xdeadc0dedead210b;
303 * Use empty space after the buffer to detect overflows.
305 * Since zio_init() creates kmem caches only for certain set of buffer sizes,
306 * allocations of different sizes may have some unused space after the data.
307 * Filling part of that space with a known pattern on allocation and checking
308 * it on free should allow us to detect some buffer overflows.
311 zio_buf_put_canary(ulong_t
*p
, size_t size
, kmem_cache_t
**cache
, size_t c
)
314 size_t off
= P2ROUNDUP(size
, sizeof (ulong_t
));
315 ulong_t
*canary
= p
+ off
/ sizeof (ulong_t
);
316 size_t asize
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
317 if (c
+ 1 < SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
&&
318 cache
[c
] == cache
[c
+ 1])
319 asize
= (c
+ 2) << SPA_MINBLOCKSHIFT
;
320 for (; off
< asize
; canary
++, off
+= sizeof (ulong_t
))
321 *canary
= zio_buf_canary
;
326 zio_buf_check_canary(ulong_t
*p
, size_t size
, kmem_cache_t
**cache
, size_t c
)
329 size_t off
= P2ROUNDUP(size
, sizeof (ulong_t
));
330 ulong_t
*canary
= p
+ off
/ sizeof (ulong_t
);
331 size_t asize
= (c
+ 1) << SPA_MINBLOCKSHIFT
;
332 if (c
+ 1 < SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
&&
333 cache
[c
] == cache
[c
+ 1])
334 asize
= (c
+ 2) << SPA_MINBLOCKSHIFT
;
335 for (; off
< asize
; canary
++, off
+= sizeof (ulong_t
)) {
336 if (unlikely(*canary
!= zio_buf_canary
)) {
337 PANIC("ZIO buffer overflow %p (%zu) + %zu %#lx != %#lx",
338 p
, size
, (canary
- p
) * sizeof (ulong_t
),
339 *canary
, zio_buf_canary
);
346 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
347 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
348 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
349 * excess / transient data in-core during a crashdump.
352 zio_buf_alloc(size_t size
)
354 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
356 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
357 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
358 atomic_add_64(&zio_buf_cache_allocs
[c
], 1);
361 void *p
= kmem_cache_alloc(zio_buf_cache
[c
], KM_PUSHPAGE
);
362 zio_buf_put_canary(p
, size
, zio_buf_cache
, c
);
367 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
368 * crashdump if the kernel panics. This exists so that we will limit the amount
369 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
370 * of kernel heap dumped to disk when the kernel panics)
373 zio_data_buf_alloc(size_t size
)
375 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
377 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
379 void *p
= kmem_cache_alloc(zio_data_buf_cache
[c
], KM_PUSHPAGE
);
380 zio_buf_put_canary(p
, size
, zio_data_buf_cache
, c
);
385 zio_buf_free(void *buf
, size_t size
)
387 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
389 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
390 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
391 atomic_add_64(&zio_buf_cache_frees
[c
], 1);
394 zio_buf_check_canary(buf
, size
, zio_buf_cache
, c
);
395 kmem_cache_free(zio_buf_cache
[c
], buf
);
399 zio_data_buf_free(void *buf
, size_t size
)
401 size_t c
= (size
- 1) >> SPA_MINBLOCKSHIFT
;
403 VERIFY3U(c
, <, SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
);
405 zio_buf_check_canary(buf
, size
, zio_data_buf_cache
, c
);
406 kmem_cache_free(zio_data_buf_cache
[c
], buf
);
410 zio_abd_free(void *abd
, size_t size
)
413 abd_free((abd_t
*)abd
);
417 * ==========================================================================
418 * Push and pop I/O transform buffers
419 * ==========================================================================
422 zio_push_transform(zio_t
*zio
, abd_t
*data
, uint64_t size
, uint64_t bufsize
,
423 zio_transform_func_t
*transform
)
425 zio_transform_t
*zt
= kmem_alloc(sizeof (zio_transform_t
), KM_SLEEP
);
427 zt
->zt_orig_abd
= zio
->io_abd
;
428 zt
->zt_orig_size
= zio
->io_size
;
429 zt
->zt_bufsize
= bufsize
;
430 zt
->zt_transform
= transform
;
432 zt
->zt_next
= zio
->io_transform_stack
;
433 zio
->io_transform_stack
= zt
;
440 zio_pop_transforms(zio_t
*zio
)
444 while ((zt
= zio
->io_transform_stack
) != NULL
) {
445 if (zt
->zt_transform
!= NULL
)
446 zt
->zt_transform(zio
,
447 zt
->zt_orig_abd
, zt
->zt_orig_size
);
449 if (zt
->zt_bufsize
!= 0)
450 abd_free(zio
->io_abd
);
452 zio
->io_abd
= zt
->zt_orig_abd
;
453 zio
->io_size
= zt
->zt_orig_size
;
454 zio
->io_transform_stack
= zt
->zt_next
;
456 kmem_free(zt
, sizeof (zio_transform_t
));
461 * ==========================================================================
462 * I/O transform callbacks for subblocks, decompression, and decryption
463 * ==========================================================================
466 zio_subblock(zio_t
*zio
, abd_t
*data
, uint64_t size
)
468 ASSERT(zio
->io_size
> size
);
470 if (zio
->io_type
== ZIO_TYPE_READ
)
471 abd_copy(data
, zio
->io_abd
, size
);
475 zio_decompress(zio_t
*zio
, abd_t
*data
, uint64_t size
)
477 if (zio
->io_error
== 0) {
478 void *tmp
= abd_borrow_buf(data
, size
);
479 int ret
= zio_decompress_data(BP_GET_COMPRESS(zio
->io_bp
),
480 zio
->io_abd
, tmp
, zio
->io_size
, size
,
481 &zio
->io_prop
.zp_complevel
);
482 abd_return_buf_copy(data
, tmp
, size
);
484 if (zio_injection_enabled
&& ret
== 0)
485 ret
= zio_handle_fault_injection(zio
, EINVAL
);
488 zio
->io_error
= SET_ERROR(EIO
);
493 zio_decrypt(zio_t
*zio
, abd_t
*data
, uint64_t size
)
497 blkptr_t
*bp
= zio
->io_bp
;
498 spa_t
*spa
= zio
->io_spa
;
499 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
500 uint64_t lsize
= BP_GET_LSIZE(bp
);
501 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
502 uint8_t salt
[ZIO_DATA_SALT_LEN
];
503 uint8_t iv
[ZIO_DATA_IV_LEN
];
504 uint8_t mac
[ZIO_DATA_MAC_LEN
];
505 boolean_t no_crypt
= B_FALSE
;
507 ASSERT(BP_USES_CRYPT(bp
));
508 ASSERT3U(size
, !=, 0);
510 if (zio
->io_error
!= 0)
514 * Verify the cksum of MACs stored in an indirect bp. It will always
515 * be possible to verify this since it does not require an encryption
518 if (BP_HAS_INDIRECT_MAC_CKSUM(bp
)) {
519 zio_crypt_decode_mac_bp(bp
, mac
);
521 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
) {
523 * We haven't decompressed the data yet, but
524 * zio_crypt_do_indirect_mac_checksum() requires
525 * decompressed data to be able to parse out the MACs
526 * from the indirect block. We decompress it now and
527 * throw away the result after we are finished.
529 tmp
= zio_buf_alloc(lsize
);
530 ret
= zio_decompress_data(BP_GET_COMPRESS(bp
),
531 zio
->io_abd
, tmp
, zio
->io_size
, lsize
,
532 &zio
->io_prop
.zp_complevel
);
534 ret
= SET_ERROR(EIO
);
537 ret
= zio_crypt_do_indirect_mac_checksum(B_FALSE
,
538 tmp
, lsize
, BP_SHOULD_BYTESWAP(bp
), mac
);
539 zio_buf_free(tmp
, lsize
);
541 ret
= zio_crypt_do_indirect_mac_checksum_abd(B_FALSE
,
542 zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
), mac
);
544 abd_copy(data
, zio
->io_abd
, size
);
546 if (zio_injection_enabled
&& ot
!= DMU_OT_DNODE
&& ret
== 0) {
547 ret
= zio_handle_decrypt_injection(spa
,
548 &zio
->io_bookmark
, ot
, ECKSUM
);
557 * If this is an authenticated block, just check the MAC. It would be
558 * nice to separate this out into its own flag, but when this was done,
559 * we had run out of bits in what is now zio_flag_t. Future cleanup
560 * could make this a flag bit.
562 if (BP_IS_AUTHENTICATED(bp
)) {
563 if (ot
== DMU_OT_OBJSET
) {
564 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
,
565 dsobj
, zio
->io_abd
, size
, BP_SHOULD_BYTESWAP(bp
));
567 zio_crypt_decode_mac_bp(bp
, mac
);
568 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
,
569 zio
->io_abd
, size
, mac
);
570 if (zio_injection_enabled
&& ret
== 0) {
571 ret
= zio_handle_decrypt_injection(spa
,
572 &zio
->io_bookmark
, ot
, ECKSUM
);
575 abd_copy(data
, zio
->io_abd
, size
);
583 zio_crypt_decode_params_bp(bp
, salt
, iv
);
585 if (ot
== DMU_OT_INTENT_LOG
) {
586 tmp
= abd_borrow_buf_copy(zio
->io_abd
, sizeof (zil_chain_t
));
587 zio_crypt_decode_mac_zil(tmp
, mac
);
588 abd_return_buf(zio
->io_abd
, tmp
, sizeof (zil_chain_t
));
590 zio_crypt_decode_mac_bp(bp
, mac
);
593 ret
= spa_do_crypt_abd(B_FALSE
, spa
, &zio
->io_bookmark
, BP_GET_TYPE(bp
),
594 BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
), salt
, iv
, mac
, size
, data
,
595 zio
->io_abd
, &no_crypt
);
597 abd_copy(data
, zio
->io_abd
, size
);
605 /* assert that the key was found unless this was speculative */
606 ASSERT(ret
!= EACCES
|| (zio
->io_flags
& ZIO_FLAG_SPECULATIVE
));
609 * If there was a decryption / authentication error return EIO as
610 * the io_error. If this was not a speculative zio, create an ereport.
613 zio
->io_error
= SET_ERROR(EIO
);
614 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
615 spa_log_error(spa
, &zio
->io_bookmark
,
616 BP_GET_LOGICAL_BIRTH(zio
->io_bp
));
617 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
618 spa
, NULL
, &zio
->io_bookmark
, zio
, 0);
626 * ==========================================================================
627 * I/O parent/child relationships and pipeline interlocks
628 * ==========================================================================
631 zio_walk_parents(zio_t
*cio
, zio_link_t
**zl
)
633 list_t
*pl
= &cio
->io_parent_list
;
635 *zl
= (*zl
== NULL
) ? list_head(pl
) : list_next(pl
, *zl
);
639 ASSERT((*zl
)->zl_child
== cio
);
640 return ((*zl
)->zl_parent
);
644 zio_walk_children(zio_t
*pio
, zio_link_t
**zl
)
646 list_t
*cl
= &pio
->io_child_list
;
648 ASSERT(MUTEX_HELD(&pio
->io_lock
));
650 *zl
= (*zl
== NULL
) ? list_head(cl
) : list_next(cl
, *zl
);
654 ASSERT((*zl
)->zl_parent
== pio
);
655 return ((*zl
)->zl_child
);
659 zio_unique_parent(zio_t
*cio
)
661 zio_link_t
*zl
= NULL
;
662 zio_t
*pio
= zio_walk_parents(cio
, &zl
);
664 VERIFY3P(zio_walk_parents(cio
, &zl
), ==, NULL
);
669 zio_add_child(zio_t
*pio
, zio_t
*cio
)
672 * Logical I/Os can have logical, gang, or vdev children.
673 * Gang I/Os can have gang or vdev children.
674 * Vdev I/Os can only have vdev children.
675 * The following ASSERT captures all of these constraints.
677 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
679 /* Parent should not have READY stage if child doesn't have it. */
680 IMPLY((cio
->io_pipeline
& ZIO_STAGE_READY
) == 0 &&
681 (cio
->io_child_type
!= ZIO_CHILD_VDEV
),
682 (pio
->io_pipeline
& ZIO_STAGE_READY
) == 0);
684 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
688 mutex_enter(&pio
->io_lock
);
689 mutex_enter(&cio
->io_lock
);
691 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
693 uint64_t *countp
= pio
->io_children
[cio
->io_child_type
];
694 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
695 countp
[w
] += !cio
->io_state
[w
];
697 list_insert_head(&pio
->io_child_list
, zl
);
698 list_insert_head(&cio
->io_parent_list
, zl
);
700 mutex_exit(&cio
->io_lock
);
701 mutex_exit(&pio
->io_lock
);
705 zio_add_child_first(zio_t
*pio
, zio_t
*cio
)
708 * Logical I/Os can have logical, gang, or vdev children.
709 * Gang I/Os can have gang or vdev children.
710 * Vdev I/Os can only have vdev children.
711 * The following ASSERT captures all of these constraints.
713 ASSERT3S(cio
->io_child_type
, <=, pio
->io_child_type
);
715 /* Parent should not have READY stage if child doesn't have it. */
716 IMPLY((cio
->io_pipeline
& ZIO_STAGE_READY
) == 0 &&
717 (cio
->io_child_type
!= ZIO_CHILD_VDEV
),
718 (pio
->io_pipeline
& ZIO_STAGE_READY
) == 0);
720 zio_link_t
*zl
= kmem_cache_alloc(zio_link_cache
, KM_SLEEP
);
724 ASSERT(list_is_empty(&cio
->io_parent_list
));
725 list_insert_head(&cio
->io_parent_list
, zl
);
727 mutex_enter(&pio
->io_lock
);
729 ASSERT(pio
->io_state
[ZIO_WAIT_DONE
] == 0);
731 uint64_t *countp
= pio
->io_children
[cio
->io_child_type
];
732 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
733 countp
[w
] += !cio
->io_state
[w
];
735 list_insert_head(&pio
->io_child_list
, zl
);
737 mutex_exit(&pio
->io_lock
);
741 zio_remove_child(zio_t
*pio
, zio_t
*cio
, zio_link_t
*zl
)
743 ASSERT(zl
->zl_parent
== pio
);
744 ASSERT(zl
->zl_child
== cio
);
746 mutex_enter(&pio
->io_lock
);
747 mutex_enter(&cio
->io_lock
);
749 list_remove(&pio
->io_child_list
, zl
);
750 list_remove(&cio
->io_parent_list
, zl
);
752 mutex_exit(&cio
->io_lock
);
753 mutex_exit(&pio
->io_lock
);
754 kmem_cache_free(zio_link_cache
, zl
);
758 zio_wait_for_children(zio_t
*zio
, uint8_t childbits
, enum zio_wait_type wait
)
760 boolean_t waiting
= B_FALSE
;
762 mutex_enter(&zio
->io_lock
);
763 ASSERT(zio
->io_stall
== NULL
);
764 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++) {
765 if (!(ZIO_CHILD_BIT_IS_SET(childbits
, c
)))
768 uint64_t *countp
= &zio
->io_children
[c
][wait
];
771 ASSERT3U(zio
->io_stage
, !=, ZIO_STAGE_OPEN
);
772 zio
->io_stall
= countp
;
777 mutex_exit(&zio
->io_lock
);
781 __attribute__((always_inline
))
783 zio_notify_parent(zio_t
*pio
, zio_t
*zio
, enum zio_wait_type wait
,
784 zio_t
**next_to_executep
)
786 uint64_t *countp
= &pio
->io_children
[zio
->io_child_type
][wait
];
787 int *errorp
= &pio
->io_child_error
[zio
->io_child_type
];
789 mutex_enter(&pio
->io_lock
);
790 if (zio
->io_error
&& !(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
))
791 *errorp
= zio_worst_error(*errorp
, zio
->io_error
);
792 pio
->io_reexecute
|= zio
->io_reexecute
;
793 ASSERT3U(*countp
, >, 0);
797 if (*countp
== 0 && pio
->io_stall
== countp
) {
798 zio_taskq_type_t type
=
799 pio
->io_stage
< ZIO_STAGE_VDEV_IO_START
? ZIO_TASKQ_ISSUE
:
801 pio
->io_stall
= NULL
;
802 mutex_exit(&pio
->io_lock
);
805 * If we can tell the caller to execute this parent next, do
806 * so. We only do this if the parent's zio type matches the
807 * child's type. Otherwise dispatch the parent zio in its
810 * Having the caller execute the parent when possible reduces
811 * locking on the zio taskq's, reduces context switch
812 * overhead, and has no recursion penalty. Note that one
813 * read from disk typically causes at least 3 zio's: a
814 * zio_null(), the logical zio_read(), and then a physical
815 * zio. When the physical ZIO completes, we are able to call
816 * zio_done() on all 3 of these zio's from one invocation of
817 * zio_execute() by returning the parent back to
818 * zio_execute(). Since the parent isn't executed until this
819 * thread returns back to zio_execute(), the caller should do
822 * In other cases, dispatching the parent prevents
823 * overflowing the stack when we have deeply nested
824 * parent-child relationships, as we do with the "mega zio"
825 * of writes for spa_sync(), and the chain of ZIL blocks.
827 if (next_to_executep
!= NULL
&& *next_to_executep
== NULL
&&
828 pio
->io_type
== zio
->io_type
) {
829 *next_to_executep
= pio
;
831 zio_taskq_dispatch(pio
, type
, B_FALSE
);
834 mutex_exit(&pio
->io_lock
);
839 zio_inherit_child_errors(zio_t
*zio
, enum zio_child c
)
841 if (zio
->io_child_error
[c
] != 0 && zio
->io_error
== 0)
842 zio
->io_error
= zio
->io_child_error
[c
];
846 zio_bookmark_compare(const void *x1
, const void *x2
)
848 const zio_t
*z1
= x1
;
849 const zio_t
*z2
= x2
;
851 if (z1
->io_bookmark
.zb_objset
< z2
->io_bookmark
.zb_objset
)
853 if (z1
->io_bookmark
.zb_objset
> z2
->io_bookmark
.zb_objset
)
856 if (z1
->io_bookmark
.zb_object
< z2
->io_bookmark
.zb_object
)
858 if (z1
->io_bookmark
.zb_object
> z2
->io_bookmark
.zb_object
)
861 if (z1
->io_bookmark
.zb_level
< z2
->io_bookmark
.zb_level
)
863 if (z1
->io_bookmark
.zb_level
> z2
->io_bookmark
.zb_level
)
866 if (z1
->io_bookmark
.zb_blkid
< z2
->io_bookmark
.zb_blkid
)
868 if (z1
->io_bookmark
.zb_blkid
> z2
->io_bookmark
.zb_blkid
)
880 * ==========================================================================
881 * Create the various types of I/O (read, write, free, etc)
882 * ==========================================================================
885 zio_create(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
886 abd_t
*data
, uint64_t lsize
, uint64_t psize
, zio_done_func_t
*done
,
887 void *private, zio_type_t type
, zio_priority_t priority
,
888 zio_flag_t flags
, vdev_t
*vd
, uint64_t offset
,
889 const zbookmark_phys_t
*zb
, enum zio_stage stage
,
890 enum zio_stage pipeline
)
894 IMPLY(type
!= ZIO_TYPE_TRIM
, psize
<= SPA_MAXBLOCKSIZE
);
895 ASSERT(P2PHASE(psize
, SPA_MINBLOCKSIZE
) == 0);
896 ASSERT(P2PHASE(offset
, SPA_MINBLOCKSIZE
) == 0);
898 ASSERT(!vd
|| spa_config_held(spa
, SCL_STATE_ALL
, RW_READER
));
899 ASSERT(!bp
|| !(flags
& ZIO_FLAG_CONFIG_WRITER
));
900 ASSERT(vd
|| stage
== ZIO_STAGE_OPEN
);
902 IMPLY(lsize
!= psize
, (flags
& ZIO_FLAG_RAW_COMPRESS
) != 0);
904 zio
= kmem_cache_alloc(zio_cache
, KM_SLEEP
);
905 memset(zio
, 0, sizeof (zio_t
));
907 mutex_init(&zio
->io_lock
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
908 cv_init(&zio
->io_cv
, NULL
, CV_DEFAULT
, NULL
);
910 list_create(&zio
->io_parent_list
, sizeof (zio_link_t
),
911 offsetof(zio_link_t
, zl_parent_node
));
912 list_create(&zio
->io_child_list
, sizeof (zio_link_t
),
913 offsetof(zio_link_t
, zl_child_node
));
914 metaslab_trace_init(&zio
->io_alloc_list
);
917 zio
->io_child_type
= ZIO_CHILD_VDEV
;
918 else if (flags
& ZIO_FLAG_GANG_CHILD
)
919 zio
->io_child_type
= ZIO_CHILD_GANG
;
920 else if (flags
& ZIO_FLAG_DDT_CHILD
)
921 zio
->io_child_type
= ZIO_CHILD_DDT
;
923 zio
->io_child_type
= ZIO_CHILD_LOGICAL
;
926 if (type
!= ZIO_TYPE_WRITE
||
927 zio
->io_child_type
== ZIO_CHILD_DDT
) {
928 zio
->io_bp_copy
= *bp
;
929 zio
->io_bp
= &zio
->io_bp_copy
; /* so caller can free */
931 zio
->io_bp
= (blkptr_t
*)bp
;
933 zio
->io_bp_orig
= *bp
;
934 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
935 zio
->io_logical
= zio
;
936 if (zio
->io_child_type
> ZIO_CHILD_GANG
&& BP_IS_GANG(bp
))
937 pipeline
|= ZIO_GANG_STAGES
;
943 zio
->io_private
= private;
945 zio
->io_priority
= priority
;
947 zio
->io_offset
= offset
;
948 zio
->io_orig_abd
= zio
->io_abd
= data
;
949 zio
->io_orig_size
= zio
->io_size
= psize
;
950 zio
->io_lsize
= lsize
;
951 zio
->io_orig_flags
= zio
->io_flags
= flags
;
952 zio
->io_orig_stage
= zio
->io_stage
= stage
;
953 zio
->io_orig_pipeline
= zio
->io_pipeline
= pipeline
;
954 zio
->io_pipeline_trace
= ZIO_STAGE_OPEN
;
955 zio
->io_allocator
= ZIO_ALLOCATOR_NONE
;
957 zio
->io_state
[ZIO_WAIT_READY
] = (stage
>= ZIO_STAGE_READY
) ||
958 (pipeline
& ZIO_STAGE_READY
) == 0;
959 zio
->io_state
[ZIO_WAIT_DONE
] = (stage
>= ZIO_STAGE_DONE
);
962 zio
->io_bookmark
= *zb
;
965 zio
->io_metaslab_class
= pio
->io_metaslab_class
;
966 if (zio
->io_logical
== NULL
)
967 zio
->io_logical
= pio
->io_logical
;
968 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
969 zio
->io_gang_leader
= pio
->io_gang_leader
;
970 zio_add_child_first(pio
, zio
);
973 taskq_init_ent(&zio
->io_tqent
);
979 zio_destroy(zio_t
*zio
)
981 metaslab_trace_fini(&zio
->io_alloc_list
);
982 list_destroy(&zio
->io_parent_list
);
983 list_destroy(&zio
->io_child_list
);
984 mutex_destroy(&zio
->io_lock
);
985 cv_destroy(&zio
->io_cv
);
986 kmem_cache_free(zio_cache
, zio
);
990 * ZIO intended to be between others. Provides synchronization at READY
991 * and DONE pipeline stages and calls the respective callbacks.
994 zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, zio_done_func_t
*done
,
995 void *private, zio_flag_t flags
)
999 zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1000 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1001 ZIO_STAGE_OPEN
, ZIO_INTERLOCK_PIPELINE
);
1007 * ZIO intended to be a root of a tree. Unlike null ZIO does not have a
1008 * READY pipeline stage (is ready on creation), so it should not be used
1009 * as child of any ZIO that may need waiting for grandchildren READY stage
1010 * (any other ZIO type).
1013 zio_root(spa_t
*spa
, zio_done_func_t
*done
, void *private, zio_flag_t flags
)
1017 zio
= zio_create(NULL
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1018 ZIO_TYPE_NULL
, ZIO_PRIORITY_NOW
, flags
, NULL
, 0, NULL
,
1019 ZIO_STAGE_OPEN
, ZIO_ROOT_PIPELINE
);
1025 zfs_blkptr_verify_log(spa_t
*spa
, const blkptr_t
*bp
,
1026 enum blk_verify_flag blk_verify
, const char *fmt
, ...)
1032 (void) vsnprintf(buf
, sizeof (buf
), fmt
, adx
);
1035 zfs_dbgmsg("bad blkptr at %px: "
1036 "DVA[0]=%#llx/%#llx "
1037 "DVA[1]=%#llx/%#llx "
1038 "DVA[2]=%#llx/%#llx "
1044 "cksum=%#llx/%#llx/%#llx/%#llx",
1046 (long long)bp
->blk_dva
[0].dva_word
[0],
1047 (long long)bp
->blk_dva
[0].dva_word
[1],
1048 (long long)bp
->blk_dva
[1].dva_word
[0],
1049 (long long)bp
->blk_dva
[1].dva_word
[1],
1050 (long long)bp
->blk_dva
[2].dva_word
[0],
1051 (long long)bp
->blk_dva
[2].dva_word
[1],
1052 (long long)bp
->blk_prop
,
1053 (long long)bp
->blk_pad
[0],
1054 (long long)bp
->blk_pad
[1],
1055 (long long)BP_GET_PHYSICAL_BIRTH(bp
),
1056 (long long)BP_GET_LOGICAL_BIRTH(bp
),
1057 (long long)bp
->blk_fill
,
1058 (long long)bp
->blk_cksum
.zc_word
[0],
1059 (long long)bp
->blk_cksum
.zc_word
[1],
1060 (long long)bp
->blk_cksum
.zc_word
[2],
1061 (long long)bp
->blk_cksum
.zc_word
[3]);
1062 switch (blk_verify
) {
1063 case BLK_VERIFY_HALT
:
1064 zfs_panic_recover("%s: %s", spa_name(spa
), buf
);
1066 case BLK_VERIFY_LOG
:
1067 zfs_dbgmsg("%s: %s", spa_name(spa
), buf
);
1069 case BLK_VERIFY_ONLY
:
1077 * Verify the block pointer fields contain reasonable values. This means
1078 * it only contains known object types, checksum/compression identifiers,
1079 * block sizes within the maximum allowed limits, valid DVAs, etc.
1081 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
1082 * argument controls the behavior when an invalid field is detected.
1084 * Values for blk_verify_flag:
1085 * BLK_VERIFY_ONLY: evaluate the block
1086 * BLK_VERIFY_LOG: evaluate the block and log problems
1087 * BLK_VERIFY_HALT: call zfs_panic_recover on error
1089 * Values for blk_config_flag:
1090 * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
1091 * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
1092 * obtained for reader
1093 * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
1097 zfs_blkptr_verify(spa_t
*spa
, const blkptr_t
*bp
,
1098 enum blk_config_flag blk_config
, enum blk_verify_flag blk_verify
)
1102 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp
))) {
1103 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1104 "blkptr at %px has invalid TYPE %llu",
1105 bp
, (longlong_t
)BP_GET_TYPE(bp
));
1107 if (BP_GET_CHECKSUM(bp
) >= ZIO_CHECKSUM_FUNCTIONS
) {
1108 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1109 "blkptr at %px has invalid CHECKSUM %llu",
1110 bp
, (longlong_t
)BP_GET_CHECKSUM(bp
));
1112 if (BP_GET_COMPRESS(bp
) >= ZIO_COMPRESS_FUNCTIONS
) {
1113 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1114 "blkptr at %px has invalid COMPRESS %llu",
1115 bp
, (longlong_t
)BP_GET_COMPRESS(bp
));
1117 if (BP_GET_LSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
1118 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1119 "blkptr at %px has invalid LSIZE %llu",
1120 bp
, (longlong_t
)BP_GET_LSIZE(bp
));
1122 if (BP_GET_PSIZE(bp
) > SPA_MAXBLOCKSIZE
) {
1123 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1124 "blkptr at %px has invalid PSIZE %llu",
1125 bp
, (longlong_t
)BP_GET_PSIZE(bp
));
1128 if (BP_IS_EMBEDDED(bp
)) {
1129 if (BPE_GET_ETYPE(bp
) >= NUM_BP_EMBEDDED_TYPES
) {
1130 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1131 "blkptr at %px has invalid ETYPE %llu",
1132 bp
, (longlong_t
)BPE_GET_ETYPE(bp
));
1137 * Do not verify individual DVAs if the config is not trusted. This
1138 * will be done once the zio is executed in vdev_mirror_map_alloc.
1140 if (!spa
->spa_trust_config
)
1141 return (errors
== 0);
1143 switch (blk_config
) {
1144 case BLK_CONFIG_HELD
:
1145 ASSERT(spa_config_held(spa
, SCL_VDEV
, RW_WRITER
));
1147 case BLK_CONFIG_NEEDED
:
1148 spa_config_enter(spa
, SCL_VDEV
, bp
, RW_READER
);
1150 case BLK_CONFIG_SKIP
:
1151 return (errors
== 0);
1153 panic("invalid blk_config %u", blk_config
);
1157 * Pool-specific checks.
1159 * Note: it would be nice to verify that the logical birth
1160 * and physical birth are not too large. However,
1161 * spa_freeze() allows the birth time of log blocks (and
1162 * dmu_sync()-ed blocks that are in the log) to be arbitrarily
1165 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
1166 const dva_t
*dva
= &bp
->blk_dva
[i
];
1167 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1169 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
) {
1170 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1171 "blkptr at %px DVA %u has invalid VDEV %llu",
1172 bp
, i
, (longlong_t
)vdevid
);
1175 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1177 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1178 "blkptr at %px DVA %u has invalid VDEV %llu",
1179 bp
, i
, (longlong_t
)vdevid
);
1182 if (vd
->vdev_ops
== &vdev_hole_ops
) {
1183 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1184 "blkptr at %px DVA %u has hole VDEV %llu",
1185 bp
, i
, (longlong_t
)vdevid
);
1188 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1190 * "missing" vdevs are valid during import, but we
1191 * don't have their detailed info (e.g. asize), so
1192 * we can't perform any more checks on them.
1196 uint64_t offset
= DVA_GET_OFFSET(dva
);
1197 uint64_t asize
= DVA_GET_ASIZE(dva
);
1198 if (DVA_GET_GANG(dva
))
1199 asize
= vdev_gang_header_asize(vd
);
1200 if (offset
+ asize
> vd
->vdev_asize
) {
1201 errors
+= zfs_blkptr_verify_log(spa
, bp
, blk_verify
,
1202 "blkptr at %px DVA %u has invalid OFFSET %llu",
1203 bp
, i
, (longlong_t
)offset
);
1206 if (blk_config
== BLK_CONFIG_NEEDED
)
1207 spa_config_exit(spa
, SCL_VDEV
, bp
);
1209 return (errors
== 0);
1213 zfs_dva_valid(spa_t
*spa
, const dva_t
*dva
, const blkptr_t
*bp
)
1216 uint64_t vdevid
= DVA_GET_VDEV(dva
);
1218 if (vdevid
>= spa
->spa_root_vdev
->vdev_children
)
1221 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[vdevid
];
1225 if (vd
->vdev_ops
== &vdev_hole_ops
)
1228 if (vd
->vdev_ops
== &vdev_missing_ops
) {
1232 uint64_t offset
= DVA_GET_OFFSET(dva
);
1233 uint64_t asize
= DVA_GET_ASIZE(dva
);
1235 if (DVA_GET_GANG(dva
))
1236 asize
= vdev_gang_header_asize(vd
);
1237 if (offset
+ asize
> vd
->vdev_asize
)
1244 zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
1245 abd_t
*data
, uint64_t size
, zio_done_func_t
*done
, void *private,
1246 zio_priority_t priority
, zio_flag_t flags
, const zbookmark_phys_t
*zb
)
1250 zio
= zio_create(pio
, spa
, BP_GET_BIRTH(bp
), bp
,
1251 data
, size
, size
, done
, private,
1252 ZIO_TYPE_READ
, priority
, flags
, NULL
, 0, zb
,
1253 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1254 ZIO_DDT_CHILD_READ_PIPELINE
: ZIO_READ_PIPELINE
);
1260 zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
1261 abd_t
*data
, uint64_t lsize
, uint64_t psize
, const zio_prop_t
*zp
,
1262 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
1263 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1264 zio_flag_t flags
, const zbookmark_phys_t
*zb
)
1268 ASSERT(zp
->zp_checksum
>= ZIO_CHECKSUM_OFF
&&
1269 zp
->zp_checksum
< ZIO_CHECKSUM_FUNCTIONS
&&
1270 zp
->zp_compress
>= ZIO_COMPRESS_OFF
&&
1271 zp
->zp_compress
< ZIO_COMPRESS_FUNCTIONS
&&
1272 DMU_OT_IS_VALID(zp
->zp_type
) &&
1273 zp
->zp_level
< 32 &&
1274 zp
->zp_copies
> 0 &&
1275 zp
->zp_copies
<= spa_max_replication(spa
));
1277 zio
= zio_create(pio
, spa
, txg
, bp
, data
, lsize
, psize
, done
, private,
1278 ZIO_TYPE_WRITE
, priority
, flags
, NULL
, 0, zb
,
1279 ZIO_STAGE_OPEN
, (flags
& ZIO_FLAG_DDT_CHILD
) ?
1280 ZIO_DDT_CHILD_WRITE_PIPELINE
: ZIO_WRITE_PIPELINE
);
1282 zio
->io_ready
= ready
;
1283 zio
->io_children_ready
= children_ready
;
1287 * Data can be NULL if we are going to call zio_write_override() to
1288 * provide the already-allocated BP. But we may need the data to
1289 * verify a dedup hit (if requested). In this case, don't try to
1290 * dedup (just take the already-allocated BP verbatim). Encrypted
1291 * dedup blocks need data as well so we also disable dedup in this
1295 (zio
->io_prop
.zp_dedup_verify
|| zio
->io_prop
.zp_encrypt
)) {
1296 zio
->io_prop
.zp_dedup
= zio
->io_prop
.zp_dedup_verify
= B_FALSE
;
1303 zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, abd_t
*data
,
1304 uint64_t size
, zio_done_func_t
*done
, void *private,
1305 zio_priority_t priority
, zio_flag_t flags
, zbookmark_phys_t
*zb
)
1309 zio
= zio_create(pio
, spa
, txg
, bp
, data
, size
, size
, done
, private,
1310 ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_IO_REWRITE
, NULL
, 0, zb
,
1311 ZIO_STAGE_OPEN
, ZIO_REWRITE_PIPELINE
);
1317 zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
, boolean_t nopwrite
,
1320 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
1321 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1322 ASSERT(zio
->io_stage
== ZIO_STAGE_OPEN
);
1323 ASSERT(zio
->io_txg
== spa_syncing_txg(zio
->io_spa
));
1324 ASSERT(!brtwrite
|| !nopwrite
);
1327 * We must reset the io_prop to match the values that existed
1328 * when the bp was first written by dmu_sync() keeping in mind
1329 * that nopwrite and dedup are mutually exclusive.
1331 zio
->io_prop
.zp_dedup
= nopwrite
? B_FALSE
: zio
->io_prop
.zp_dedup
;
1332 zio
->io_prop
.zp_nopwrite
= nopwrite
;
1333 zio
->io_prop
.zp_brtwrite
= brtwrite
;
1334 zio
->io_prop
.zp_copies
= copies
;
1335 zio
->io_bp_override
= bp
;
1339 zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
)
1342 (void) zfs_blkptr_verify(spa
, bp
, BLK_CONFIG_NEEDED
, BLK_VERIFY_HALT
);
1345 * The check for EMBEDDED is a performance optimization. We
1346 * process the free here (by ignoring it) rather than
1347 * putting it on the list and then processing it in zio_free_sync().
1349 if (BP_IS_EMBEDDED(bp
))
1353 * Frees that are for the currently-syncing txg, are not going to be
1354 * deferred, and which will not need to do a read (i.e. not GANG or
1355 * DEDUP), can be processed immediately. Otherwise, put them on the
1356 * in-memory list for later processing.
1358 * Note that we only defer frees after zfs_sync_pass_deferred_free
1359 * when the log space map feature is disabled. [see relevant comment
1360 * in spa_sync_iterate_to_convergence()]
1362 if (BP_IS_GANG(bp
) ||
1364 txg
!= spa
->spa_syncing_txg
||
1365 (spa_sync_pass(spa
) >= zfs_sync_pass_deferred_free
&&
1366 !spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
)) ||
1367 brt_maybe_exists(spa
, bp
)) {
1368 metaslab_check_free(spa
, bp
);
1369 bplist_append(&spa
->spa_free_bplist
[txg
& TXG_MASK
], bp
);
1371 VERIFY3P(zio_free_sync(NULL
, spa
, txg
, bp
, 0), ==, NULL
);
1376 * To improve performance, this function may return NULL if we were able
1377 * to do the free immediately. This avoids the cost of creating a zio
1378 * (and linking it to the parent, etc).
1381 zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1384 ASSERT(!BP_IS_HOLE(bp
));
1385 ASSERT(spa_syncing_txg(spa
) == txg
);
1387 if (BP_IS_EMBEDDED(bp
))
1390 metaslab_check_free(spa
, bp
);
1392 dsl_scan_freed(spa
, bp
);
1394 if (BP_IS_GANG(bp
) ||
1396 brt_maybe_exists(spa
, bp
)) {
1398 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1399 * block header, the DDT or the BRT), so issue them
1400 * asynchronously so that this thread is not tied up.
1402 enum zio_stage stage
=
1403 ZIO_FREE_PIPELINE
| ZIO_STAGE_ISSUE_ASYNC
;
1405 return (zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1406 BP_GET_PSIZE(bp
), NULL
, NULL
,
1407 ZIO_TYPE_FREE
, ZIO_PRIORITY_NOW
,
1408 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, stage
));
1410 metaslab_free(spa
, bp
, txg
, B_FALSE
);
1416 zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
,
1417 zio_done_func_t
*done
, void *private, zio_flag_t flags
)
1421 (void) zfs_blkptr_verify(spa
, bp
, (flags
& ZIO_FLAG_CONFIG_WRITER
) ?
1422 BLK_CONFIG_HELD
: BLK_CONFIG_NEEDED
, BLK_VERIFY_HALT
);
1424 if (BP_IS_EMBEDDED(bp
))
1425 return (zio_null(pio
, spa
, NULL
, NULL
, NULL
, 0));
1428 * A claim is an allocation of a specific block. Claims are needed
1429 * to support immediate writes in the intent log. The issue is that
1430 * immediate writes contain committed data, but in a txg that was
1431 * *not* committed. Upon opening the pool after an unclean shutdown,
1432 * the intent log claims all blocks that contain immediate write data
1433 * so that the SPA knows they're in use.
1435 * All claims *must* be resolved in the first txg -- before the SPA
1436 * starts allocating blocks -- so that nothing is allocated twice.
1437 * If txg == 0 we just verify that the block is claimable.
1439 ASSERT3U(BP_GET_LOGICAL_BIRTH(&spa
->spa_uberblock
.ub_rootbp
), <,
1440 spa_min_claim_txg(spa
));
1441 ASSERT(txg
== spa_min_claim_txg(spa
) || txg
== 0);
1442 ASSERT(!BP_GET_DEDUP(bp
) || !spa_writeable(spa
)); /* zdb(8) */
1444 zio
= zio_create(pio
, spa
, txg
, bp
, NULL
, BP_GET_PSIZE(bp
),
1445 BP_GET_PSIZE(bp
), done
, private, ZIO_TYPE_CLAIM
, ZIO_PRIORITY_NOW
,
1446 flags
, NULL
, 0, NULL
, ZIO_STAGE_OPEN
, ZIO_CLAIM_PIPELINE
);
1447 ASSERT0(zio
->io_queued_timestamp
);
1453 zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
1454 zio_done_func_t
*done
, void *private, zio_flag_t flags
)
1456 zio_t
*zio
= zio_create(pio
, spa
, 0, NULL
, NULL
, 0, 0, done
, private,
1457 ZIO_TYPE_IOCTL
, ZIO_PRIORITY_NOW
, flags
, vd
, 0, NULL
,
1458 ZIO_STAGE_OPEN
, ZIO_IOCTL_PIPELINE
);
1464 zio_trim(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1465 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
1466 zio_flag_t flags
, enum trim_flag trim_flags
)
1470 ASSERT0(vd
->vdev_children
);
1471 ASSERT0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
1472 ASSERT0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
1473 ASSERT3U(size
, !=, 0);
1475 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, NULL
, size
, size
, done
,
1476 private, ZIO_TYPE_TRIM
, priority
, flags
| ZIO_FLAG_PHYSICAL
,
1477 vd
, offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_TRIM_PIPELINE
);
1478 zio
->io_trim_flags
= trim_flags
;
1484 zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1485 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1486 zio_priority_t priority
, zio_flag_t flags
, boolean_t labels
)
1490 ASSERT(vd
->vdev_children
== 0);
1491 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1492 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1493 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1495 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1496 private, ZIO_TYPE_READ
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1497 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_READ_PHYS_PIPELINE
);
1499 zio
->io_prop
.zp_checksum
= checksum
;
1505 zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
, uint64_t size
,
1506 abd_t
*data
, int checksum
, zio_done_func_t
*done
, void *private,
1507 zio_priority_t priority
, zio_flag_t flags
, boolean_t labels
)
1511 ASSERT(vd
->vdev_children
== 0);
1512 ASSERT(!labels
|| offset
+ size
<= VDEV_LABEL_START_SIZE
||
1513 offset
>= vd
->vdev_psize
- VDEV_LABEL_END_SIZE
);
1514 ASSERT3U(offset
+ size
, <=, vd
->vdev_psize
);
1516 zio
= zio_create(pio
, vd
->vdev_spa
, 0, NULL
, data
, size
, size
, done
,
1517 private, ZIO_TYPE_WRITE
, priority
, flags
| ZIO_FLAG_PHYSICAL
, vd
,
1518 offset
, NULL
, ZIO_STAGE_OPEN
, ZIO_WRITE_PHYS_PIPELINE
);
1520 zio
->io_prop
.zp_checksum
= checksum
;
1522 if (zio_checksum_table
[checksum
].ci_flags
& ZCHECKSUM_FLAG_EMBEDDED
) {
1524 * zec checksums are necessarily destructive -- they modify
1525 * the end of the write buffer to hold the verifier/checksum.
1526 * Therefore, we must make a local copy in case the data is
1527 * being written to multiple places in parallel.
1529 abd_t
*wbuf
= abd_alloc_sametype(data
, size
);
1530 abd_copy(wbuf
, data
, size
);
1532 zio_push_transform(zio
, wbuf
, size
, size
, NULL
);
1539 * Create a child I/O to do some work for us.
1542 zio_vdev_child_io(zio_t
*pio
, blkptr_t
*bp
, vdev_t
*vd
, uint64_t offset
,
1543 abd_t
*data
, uint64_t size
, int type
, zio_priority_t priority
,
1544 zio_flag_t flags
, zio_done_func_t
*done
, void *private)
1546 enum zio_stage pipeline
= ZIO_VDEV_CHILD_PIPELINE
;
1550 * vdev child I/Os do not propagate their error to the parent.
1551 * Therefore, for correct operation the caller *must* check for
1552 * and handle the error in the child i/o's done callback.
1553 * The only exceptions are i/os that we don't care about
1554 * (OPTIONAL or REPAIR).
1556 ASSERT((flags
& ZIO_FLAG_OPTIONAL
) || (flags
& ZIO_FLAG_IO_REPAIR
) ||
1559 if (type
== ZIO_TYPE_READ
&& bp
!= NULL
) {
1561 * If we have the bp, then the child should perform the
1562 * checksum and the parent need not. This pushes error
1563 * detection as close to the leaves as possible and
1564 * eliminates redundant checksums in the interior nodes.
1566 pipeline
|= ZIO_STAGE_CHECKSUM_VERIFY
;
1567 pio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
1570 if (vd
->vdev_ops
->vdev_op_leaf
) {
1571 ASSERT0(vd
->vdev_children
);
1572 offset
+= VDEV_LABEL_START_SIZE
;
1575 flags
|= ZIO_VDEV_CHILD_FLAGS(pio
);
1578 * If we've decided to do a repair, the write is not speculative --
1579 * even if the original read was.
1581 if (flags
& ZIO_FLAG_IO_REPAIR
)
1582 flags
&= ~ZIO_FLAG_SPECULATIVE
;
1585 * If we're creating a child I/O that is not associated with a
1586 * top-level vdev, then the child zio is not an allocating I/O.
1587 * If this is a retried I/O then we ignore it since we will
1588 * have already processed the original allocating I/O.
1590 if (flags
& ZIO_FLAG_IO_ALLOCATING
&&
1591 (vd
!= vd
->vdev_top
|| (flags
& ZIO_FLAG_IO_RETRY
))) {
1592 ASSERT(pio
->io_metaslab_class
!= NULL
);
1593 ASSERT(pio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
1594 ASSERT(type
== ZIO_TYPE_WRITE
);
1595 ASSERT(priority
== ZIO_PRIORITY_ASYNC_WRITE
);
1596 ASSERT(!(flags
& ZIO_FLAG_IO_REPAIR
));
1597 ASSERT(!(pio
->io_flags
& ZIO_FLAG_IO_REWRITE
) ||
1598 pio
->io_child_type
== ZIO_CHILD_GANG
);
1600 flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
1603 zio
= zio_create(pio
, pio
->io_spa
, pio
->io_txg
, bp
, data
, size
, size
,
1604 done
, private, type
, priority
, flags
, vd
, offset
, &pio
->io_bookmark
,
1605 ZIO_STAGE_VDEV_IO_START
>> 1, pipeline
);
1606 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
1612 zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
, abd_t
*data
, uint64_t size
,
1613 zio_type_t type
, zio_priority_t priority
, zio_flag_t flags
,
1614 zio_done_func_t
*done
, void *private)
1618 ASSERT(vd
->vdev_ops
->vdev_op_leaf
);
1620 zio
= zio_create(NULL
, vd
->vdev_spa
, 0, NULL
,
1621 data
, size
, size
, done
, private, type
, priority
,
1622 flags
| ZIO_FLAG_CANFAIL
| ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_DELEGATED
,
1624 ZIO_STAGE_VDEV_IO_START
>> 1, ZIO_VDEV_CHILD_PIPELINE
);
1630 zio_flush(zio_t
*pio
, vdev_t
*vd
)
1632 if (vd
->vdev_nowritecache
)
1634 if (vd
->vdev_children
== 0) {
1635 zio_nowait(zio_ioctl(pio
, vd
->vdev_spa
, vd
,
1636 DKIOCFLUSHWRITECACHE
, NULL
, NULL
, ZIO_FLAG_CANFAIL
|
1637 ZIO_FLAG_DONT_PROPAGATE
| ZIO_FLAG_DONT_RETRY
));
1639 for (uint64_t c
= 0; c
< vd
->vdev_children
; c
++)
1640 zio_flush(pio
, vd
->vdev_child
[c
]);
1645 zio_shrink(zio_t
*zio
, uint64_t size
)
1647 ASSERT3P(zio
->io_executor
, ==, NULL
);
1648 ASSERT3U(zio
->io_orig_size
, ==, zio
->io_size
);
1649 ASSERT3U(size
, <=, zio
->io_size
);
1652 * We don't shrink for raidz because of problems with the
1653 * reconstruction when reading back less than the block size.
1654 * Note, BP_IS_RAIDZ() assumes no compression.
1656 ASSERT(BP_GET_COMPRESS(zio
->io_bp
) == ZIO_COMPRESS_OFF
);
1657 if (!BP_IS_RAIDZ(zio
->io_bp
)) {
1658 /* we are not doing a raw write */
1659 ASSERT3U(zio
->io_size
, ==, zio
->io_lsize
);
1660 zio
->io_orig_size
= zio
->io_size
= zio
->io_lsize
= size
;
1665 * Round provided allocation size up to a value that can be allocated
1666 * by at least some vdev(s) in the pool with minimum or no additional
1667 * padding and without extra space usage on others
1670 zio_roundup_alloc_size(spa_t
*spa
, uint64_t size
)
1672 if (size
> spa
->spa_min_alloc
)
1673 return (roundup(size
, spa
->spa_gcd_alloc
));
1674 return (spa
->spa_min_alloc
);
1678 * ==========================================================================
1679 * Prepare to read and write logical blocks
1680 * ==========================================================================
1684 zio_read_bp_init(zio_t
*zio
)
1686 blkptr_t
*bp
= zio
->io_bp
;
1688 BP_IS_EMBEDDED(bp
) ? BPE_GET_PSIZE(bp
) : BP_GET_PSIZE(bp
);
1690 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
1692 if (BP_GET_COMPRESS(bp
) != ZIO_COMPRESS_OFF
&&
1693 zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
1694 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1695 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1696 psize
, psize
, zio_decompress
);
1699 if (((BP_IS_PROTECTED(bp
) && !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) ||
1700 BP_HAS_INDIRECT_MAC_CKSUM(bp
)) &&
1701 zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
1702 zio_push_transform(zio
, abd_alloc_sametype(zio
->io_abd
, psize
),
1703 psize
, psize
, zio_decrypt
);
1706 if (BP_IS_EMBEDDED(bp
) && BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
) {
1707 int psize
= BPE_GET_PSIZE(bp
);
1708 void *data
= abd_borrow_buf(zio
->io_abd
, psize
);
1710 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1711 decode_embedded_bp_compressed(bp
, data
);
1712 abd_return_buf_copy(zio
->io_abd
, data
, psize
);
1714 ASSERT(!BP_IS_EMBEDDED(bp
));
1717 if (BP_GET_DEDUP(bp
) && zio
->io_child_type
== ZIO_CHILD_LOGICAL
)
1718 zio
->io_pipeline
= ZIO_DDT_READ_PIPELINE
;
1724 zio_write_bp_init(zio_t
*zio
)
1726 if (!IO_IS_ALLOCATING(zio
))
1729 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1731 if (zio
->io_bp_override
) {
1732 blkptr_t
*bp
= zio
->io_bp
;
1733 zio_prop_t
*zp
= &zio
->io_prop
;
1735 ASSERT(BP_GET_LOGICAL_BIRTH(bp
) != zio
->io_txg
);
1737 *bp
= *zio
->io_bp_override
;
1738 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1740 if (zp
->zp_brtwrite
)
1743 ASSERT(!BP_GET_DEDUP(zio
->io_bp_override
));
1745 if (BP_IS_EMBEDDED(bp
))
1749 * If we've been overridden and nopwrite is set then
1750 * set the flag accordingly to indicate that a nopwrite
1751 * has already occurred.
1753 if (!BP_IS_HOLE(bp
) && zp
->zp_nopwrite
) {
1754 ASSERT(!zp
->zp_dedup
);
1755 ASSERT3U(BP_GET_CHECKSUM(bp
), ==, zp
->zp_checksum
);
1756 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
1760 ASSERT(!zp
->zp_nopwrite
);
1762 if (BP_IS_HOLE(bp
) || !zp
->zp_dedup
)
1765 ASSERT((zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
1766 ZCHECKSUM_FLAG_DEDUP
) || zp
->zp_dedup_verify
);
1768 if (BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
&&
1770 BP_SET_DEDUP(bp
, 1);
1771 zio
->io_pipeline
|= ZIO_STAGE_DDT_WRITE
;
1776 * We were unable to handle this as an override bp, treat
1777 * it as a regular write I/O.
1779 zio
->io_bp_override
= NULL
;
1780 *bp
= zio
->io_bp_orig
;
1781 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1788 zio_write_compress(zio_t
*zio
)
1790 spa_t
*spa
= zio
->io_spa
;
1791 zio_prop_t
*zp
= &zio
->io_prop
;
1792 enum zio_compress compress
= zp
->zp_compress
;
1793 blkptr_t
*bp
= zio
->io_bp
;
1794 uint64_t lsize
= zio
->io_lsize
;
1795 uint64_t psize
= zio
->io_size
;
1799 * If our children haven't all reached the ready stage,
1800 * wait for them and then repeat this pipeline stage.
1802 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
1803 ZIO_CHILD_GANG_BIT
, ZIO_WAIT_READY
)) {
1807 if (!IO_IS_ALLOCATING(zio
))
1810 if (zio
->io_children_ready
!= NULL
) {
1812 * Now that all our children are ready, run the callback
1813 * associated with this zio in case it wants to modify the
1814 * data to be written.
1816 ASSERT3U(zp
->zp_level
, >, 0);
1817 zio
->io_children_ready(zio
);
1820 ASSERT(zio
->io_child_type
!= ZIO_CHILD_DDT
);
1821 ASSERT(zio
->io_bp_override
== NULL
);
1823 if (!BP_IS_HOLE(bp
) && BP_GET_LOGICAL_BIRTH(bp
) == zio
->io_txg
) {
1825 * We're rewriting an existing block, which means we're
1826 * working on behalf of spa_sync(). For spa_sync() to
1827 * converge, it must eventually be the case that we don't
1828 * have to allocate new blocks. But compression changes
1829 * the blocksize, which forces a reallocate, and makes
1830 * convergence take longer. Therefore, after the first
1831 * few passes, stop compressing to ensure convergence.
1833 pass
= spa_sync_pass(spa
);
1835 ASSERT(zio
->io_txg
== spa_syncing_txg(spa
));
1836 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1837 ASSERT(!BP_GET_DEDUP(bp
));
1839 if (pass
>= zfs_sync_pass_dont_compress
)
1840 compress
= ZIO_COMPRESS_OFF
;
1842 /* Make sure someone doesn't change their mind on overwrites */
1843 ASSERT(BP_IS_EMBEDDED(bp
) || BP_IS_GANG(bp
) ||
1844 MIN(zp
->zp_copies
, spa_max_replication(spa
))
1845 == BP_GET_NDVAS(bp
));
1848 /* If it's a compressed write that is not raw, compress the buffer. */
1849 if (compress
!= ZIO_COMPRESS_OFF
&&
1850 !(zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
)) {
1852 psize
= zio_compress_data(compress
, zio
->io_abd
, &cbuf
, lsize
,
1855 compress
= ZIO_COMPRESS_OFF
;
1856 } else if (psize
>= lsize
) {
1857 compress
= ZIO_COMPRESS_OFF
;
1859 zio_buf_free(cbuf
, lsize
);
1860 } else if (!zp
->zp_dedup
&& !zp
->zp_encrypt
&&
1861 psize
<= BPE_PAYLOAD_SIZE
&&
1862 zp
->zp_level
== 0 && !DMU_OT_HAS_FILL(zp
->zp_type
) &&
1863 spa_feature_is_enabled(spa
, SPA_FEATURE_EMBEDDED_DATA
)) {
1864 encode_embedded_bp_compressed(bp
,
1865 cbuf
, compress
, lsize
, psize
);
1866 BPE_SET_ETYPE(bp
, BP_EMBEDDED_TYPE_DATA
);
1867 BP_SET_TYPE(bp
, zio
->io_prop
.zp_type
);
1868 BP_SET_LEVEL(bp
, zio
->io_prop
.zp_level
);
1869 zio_buf_free(cbuf
, lsize
);
1870 BP_SET_LOGICAL_BIRTH(bp
, zio
->io_txg
);
1871 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1872 ASSERT(spa_feature_is_active(spa
,
1873 SPA_FEATURE_EMBEDDED_DATA
));
1877 * Round compressed size up to the minimum allocation
1878 * size of the smallest-ashift device, and zero the
1879 * tail. This ensures that the compressed size of the
1880 * BP (and thus compressratio property) are correct,
1881 * in that we charge for the padding used to fill out
1884 size_t rounded
= (size_t)zio_roundup_alloc_size(spa
,
1886 if (rounded
>= lsize
) {
1887 compress
= ZIO_COMPRESS_OFF
;
1888 zio_buf_free(cbuf
, lsize
);
1891 abd_t
*cdata
= abd_get_from_buf(cbuf
, lsize
);
1892 abd_take_ownership_of_buf(cdata
, B_TRUE
);
1893 abd_zero_off(cdata
, psize
, rounded
- psize
);
1895 zio_push_transform(zio
, cdata
,
1896 psize
, lsize
, NULL
);
1901 * We were unable to handle this as an override bp, treat
1902 * it as a regular write I/O.
1904 zio
->io_bp_override
= NULL
;
1905 *bp
= zio
->io_bp_orig
;
1906 zio
->io_pipeline
= zio
->io_orig_pipeline
;
1908 } else if ((zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0 &&
1909 zp
->zp_type
== DMU_OT_DNODE
) {
1911 * The DMU actually relies on the zio layer's compression
1912 * to free metadnode blocks that have had all contained
1913 * dnodes freed. As a result, even when doing a raw
1914 * receive, we must check whether the block can be compressed
1917 psize
= zio_compress_data(ZIO_COMPRESS_EMPTY
,
1918 zio
->io_abd
, NULL
, lsize
, zp
->zp_complevel
);
1919 if (psize
== 0 || psize
>= lsize
)
1920 compress
= ZIO_COMPRESS_OFF
;
1921 } else if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
&&
1922 !(zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
)) {
1924 * If we are raw receiving an encrypted dataset we should not
1925 * take this codepath because it will change the on-disk block
1926 * and decryption will fail.
1928 size_t rounded
= MIN((size_t)zio_roundup_alloc_size(spa
, psize
),
1931 if (rounded
!= psize
) {
1932 abd_t
*cdata
= abd_alloc_linear(rounded
, B_TRUE
);
1933 abd_zero_off(cdata
, psize
, rounded
- psize
);
1934 abd_copy_off(cdata
, zio
->io_abd
, 0, 0, psize
);
1936 zio_push_transform(zio
, cdata
,
1937 psize
, rounded
, NULL
);
1940 ASSERT3U(psize
, !=, 0);
1944 * The final pass of spa_sync() must be all rewrites, but the first
1945 * few passes offer a trade-off: allocating blocks defers convergence,
1946 * but newly allocated blocks are sequential, so they can be written
1947 * to disk faster. Therefore, we allow the first few passes of
1948 * spa_sync() to allocate new blocks, but force rewrites after that.
1949 * There should only be a handful of blocks after pass 1 in any case.
1951 if (!BP_IS_HOLE(bp
) && BP_GET_LOGICAL_BIRTH(bp
) == zio
->io_txg
&&
1952 BP_GET_PSIZE(bp
) == psize
&&
1953 pass
>= zfs_sync_pass_rewrite
) {
1954 VERIFY3U(psize
, !=, 0);
1955 enum zio_stage gang_stages
= zio
->io_pipeline
& ZIO_GANG_STAGES
;
1957 zio
->io_pipeline
= ZIO_REWRITE_PIPELINE
| gang_stages
;
1958 zio
->io_flags
|= ZIO_FLAG_IO_REWRITE
;
1961 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
1965 if (BP_GET_LOGICAL_BIRTH(&zio
->io_bp_orig
) != 0 &&
1966 spa_feature_is_active(spa
, SPA_FEATURE_HOLE_BIRTH
)) {
1967 BP_SET_LSIZE(bp
, lsize
);
1968 BP_SET_TYPE(bp
, zp
->zp_type
);
1969 BP_SET_LEVEL(bp
, zp
->zp_level
);
1970 BP_SET_BIRTH(bp
, zio
->io_txg
, 0);
1972 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
1974 ASSERT(zp
->zp_checksum
!= ZIO_CHECKSUM_GANG_HEADER
);
1975 BP_SET_LSIZE(bp
, lsize
);
1976 BP_SET_TYPE(bp
, zp
->zp_type
);
1977 BP_SET_LEVEL(bp
, zp
->zp_level
);
1978 BP_SET_PSIZE(bp
, psize
);
1979 BP_SET_COMPRESS(bp
, compress
);
1980 BP_SET_CHECKSUM(bp
, zp
->zp_checksum
);
1981 BP_SET_DEDUP(bp
, zp
->zp_dedup
);
1982 BP_SET_BYTEORDER(bp
, ZFS_HOST_BYTEORDER
);
1984 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1985 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1986 ASSERT(!zp
->zp_encrypt
||
1987 DMU_OT_IS_ENCRYPTED(zp
->zp_type
));
1988 zio
->io_pipeline
= ZIO_DDT_WRITE_PIPELINE
;
1990 if (zp
->zp_nopwrite
) {
1991 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
1992 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
1993 zio
->io_pipeline
|= ZIO_STAGE_NOP_WRITE
;
2000 zio_free_bp_init(zio_t
*zio
)
2002 blkptr_t
*bp
= zio
->io_bp
;
2004 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
) {
2005 if (BP_GET_DEDUP(bp
))
2006 zio
->io_pipeline
= ZIO_DDT_FREE_PIPELINE
;
2009 ASSERT3P(zio
->io_bp
, ==, &zio
->io_bp_copy
);
2015 * ==========================================================================
2016 * Execute the I/O pipeline
2017 * ==========================================================================
2021 zio_taskq_dispatch(zio_t
*zio
, zio_taskq_type_t q
, boolean_t cutinline
)
2023 spa_t
*spa
= zio
->io_spa
;
2024 zio_type_t t
= zio
->io_type
;
2025 int flags
= (cutinline
? TQ_FRONT
: 0);
2028 * If we're a config writer or a probe, the normal issue and
2029 * interrupt threads may all be blocked waiting for the config lock.
2030 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
2032 if (zio
->io_flags
& (ZIO_FLAG_CONFIG_WRITER
| ZIO_FLAG_PROBE
))
2036 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
2038 if (t
== ZIO_TYPE_WRITE
&& zio
->io_vd
&& zio
->io_vd
->vdev_aux
)
2042 * If this is a high priority I/O, then use the high priority taskq if
2045 if ((zio
->io_priority
== ZIO_PRIORITY_NOW
||
2046 zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
) &&
2047 spa
->spa_zio_taskq
[t
][q
+ 1].stqs_count
!= 0)
2050 ASSERT3U(q
, <, ZIO_TASKQ_TYPES
);
2053 * NB: We are assuming that the zio can only be dispatched
2054 * to a single taskq at a time. It would be a grievous error
2055 * to dispatch the zio to another taskq at the same time.
2057 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
2058 spa_taskq_dispatch_ent(spa
, t
, q
, zio_execute
, zio
, flags
,
2059 &zio
->io_tqent
, zio
);
2063 zio_taskq_member(zio_t
*zio
, zio_taskq_type_t q
)
2065 spa_t
*spa
= zio
->io_spa
;
2067 taskq_t
*tq
= taskq_of_curthread();
2069 for (zio_type_t t
= 0; t
< ZIO_TYPES
; t
++) {
2070 spa_taskqs_t
*tqs
= &spa
->spa_zio_taskq
[t
][q
];
2072 for (i
= 0; i
< tqs
->stqs_count
; i
++) {
2073 if (tqs
->stqs_taskq
[i
] == tq
)
2082 zio_issue_async(zio_t
*zio
)
2084 ASSERT((zio
->io_type
!= ZIO_TYPE_WRITE
) || ZIO_HAS_ALLOCATOR(zio
));
2085 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
2090 zio_interrupt(void *zio
)
2092 zio_taskq_dispatch(zio
, ZIO_TASKQ_INTERRUPT
, B_FALSE
);
2096 zio_delay_interrupt(zio_t
*zio
)
2099 * The timeout_generic() function isn't defined in userspace, so
2100 * rather than trying to implement the function, the zio delay
2101 * functionality has been disabled for userspace builds.
2106 * If io_target_timestamp is zero, then no delay has been registered
2107 * for this IO, thus jump to the end of this function and "skip" the
2108 * delay; issuing it directly to the zio layer.
2110 if (zio
->io_target_timestamp
!= 0) {
2111 hrtime_t now
= gethrtime();
2113 if (now
>= zio
->io_target_timestamp
) {
2115 * This IO has already taken longer than the target
2116 * delay to complete, so we don't want to delay it
2117 * any longer; we "miss" the delay and issue it
2118 * directly to the zio layer. This is likely due to
2119 * the target latency being set to a value less than
2120 * the underlying hardware can satisfy (e.g. delay
2121 * set to 1ms, but the disks take 10ms to complete an
2125 DTRACE_PROBE2(zio__delay__miss
, zio_t
*, zio
,
2131 hrtime_t diff
= zio
->io_target_timestamp
- now
;
2132 clock_t expire_at_tick
= ddi_get_lbolt() +
2135 DTRACE_PROBE3(zio__delay__hit
, zio_t
*, zio
,
2136 hrtime_t
, now
, hrtime_t
, diff
);
2138 if (NSEC_TO_TICK(diff
) == 0) {
2139 /* Our delay is less than a jiffy - just spin */
2140 zfs_sleep_until(zio
->io_target_timestamp
);
2144 * Use taskq_dispatch_delay() in the place of
2145 * OpenZFS's timeout_generic().
2147 tid
= taskq_dispatch_delay(system_taskq
,
2148 zio_interrupt
, zio
, TQ_NOSLEEP
,
2150 if (tid
== TASKQID_INVALID
) {
2152 * Couldn't allocate a task. Just
2153 * finish the zio without a delay.
2162 DTRACE_PROBE1(zio__delay__skip
, zio_t
*, zio
);
2167 zio_deadman_impl(zio_t
*pio
, int ziodepth
)
2169 zio_t
*cio
, *cio_next
;
2170 zio_link_t
*zl
= NULL
;
2171 vdev_t
*vd
= pio
->io_vd
;
2173 if (zio_deadman_log_all
|| (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
)) {
2174 vdev_queue_t
*vq
= vd
? &vd
->vdev_queue
: NULL
;
2175 zbookmark_phys_t
*zb
= &pio
->io_bookmark
;
2176 uint64_t delta
= gethrtime() - pio
->io_timestamp
;
2177 uint64_t failmode
= spa_get_deadman_failmode(pio
->io_spa
);
2179 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2180 "delta=%llu queued=%llu io=%llu "
2182 "last=%llu type=%d "
2183 "priority=%d flags=0x%llx stage=0x%x "
2184 "pipeline=0x%x pipeline-trace=0x%x "
2185 "objset=%llu object=%llu "
2186 "level=%llu blkid=%llu "
2187 "offset=%llu size=%llu "
2189 ziodepth
, pio
, pio
->io_timestamp
,
2190 (u_longlong_t
)delta
, pio
->io_delta
, pio
->io_delay
,
2191 vd
? vd
->vdev_path
: "NULL",
2192 vq
? vq
->vq_io_complete_ts
: 0, pio
->io_type
,
2193 pio
->io_priority
, (u_longlong_t
)pio
->io_flags
,
2194 pio
->io_stage
, pio
->io_pipeline
, pio
->io_pipeline_trace
,
2195 (u_longlong_t
)zb
->zb_objset
, (u_longlong_t
)zb
->zb_object
,
2196 (u_longlong_t
)zb
->zb_level
, (u_longlong_t
)zb
->zb_blkid
,
2197 (u_longlong_t
)pio
->io_offset
, (u_longlong_t
)pio
->io_size
,
2199 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN
,
2200 pio
->io_spa
, vd
, zb
, pio
, 0);
2202 if (failmode
== ZIO_FAILURE_MODE_CONTINUE
&&
2203 taskq_empty_ent(&pio
->io_tqent
)) {
2208 mutex_enter(&pio
->io_lock
);
2209 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2210 cio_next
= zio_walk_children(pio
, &zl
);
2211 zio_deadman_impl(cio
, ziodepth
+ 1);
2213 mutex_exit(&pio
->io_lock
);
2217 * Log the critical information describing this zio and all of its children
2218 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2221 zio_deadman(zio_t
*pio
, const char *tag
)
2223 spa_t
*spa
= pio
->io_spa
;
2224 char *name
= spa_name(spa
);
2226 if (!zfs_deadman_enabled
|| spa_suspended(spa
))
2229 zio_deadman_impl(pio
, 0);
2231 switch (spa_get_deadman_failmode(spa
)) {
2232 case ZIO_FAILURE_MODE_WAIT
:
2233 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag
, name
);
2236 case ZIO_FAILURE_MODE_CONTINUE
:
2237 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag
, name
);
2240 case ZIO_FAILURE_MODE_PANIC
:
2241 fm_panic("%s determined I/O to pool '%s' is hung.", tag
, name
);
2247 * Execute the I/O pipeline until one of the following occurs:
2248 * (1) the I/O completes; (2) the pipeline stalls waiting for
2249 * dependent child I/Os; (3) the I/O issues, so we're waiting
2250 * for an I/O completion interrupt; (4) the I/O is delegated by
2251 * vdev-level caching or aggregation; (5) the I/O is deferred
2252 * due to vdev-level queueing; (6) the I/O is handed off to
2253 * another thread. In all cases, the pipeline stops whenever
2254 * there's no CPU work; it never burns a thread in cv_wait_io().
2256 * There's no locking on io_stage because there's no legitimate way
2257 * for multiple threads to be attempting to process the same I/O.
2259 static zio_pipe_stage_t
*zio_pipeline
[];
2262 * zio_execute() is a wrapper around the static function
2263 * __zio_execute() so that we can force __zio_execute() to be
2264 * inlined. This reduces stack overhead which is important
2265 * because __zio_execute() is called recursively in several zio
2266 * code paths. zio_execute() itself cannot be inlined because
2267 * it is externally visible.
2270 zio_execute(void *zio
)
2272 fstrans_cookie_t cookie
;
2274 cookie
= spl_fstrans_mark();
2276 spl_fstrans_unmark(cookie
);
2280 * Used to determine if in the current context the stack is sized large
2281 * enough to allow zio_execute() to be called recursively. A minimum
2282 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2285 zio_execute_stack_check(zio_t
*zio
)
2287 #if !defined(HAVE_LARGE_STACKS)
2288 dsl_pool_t
*dp
= spa_get_dsl(zio
->io_spa
);
2290 /* Executing in txg_sync_thread() context. */
2291 if (dp
&& curthread
== dp
->dp_tx
.tx_sync_thread
)
2294 /* Pool initialization outside of zio_taskq context. */
2295 if (dp
&& spa_is_initializing(dp
->dp_spa
) &&
2296 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE
) &&
2297 !zio_taskq_member(zio
, ZIO_TASKQ_ISSUE_HIGH
))
2301 #endif /* HAVE_LARGE_STACKS */
2306 __attribute__((always_inline
))
2308 __zio_execute(zio_t
*zio
)
2310 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
2312 while (zio
->io_stage
< ZIO_STAGE_DONE
) {
2313 enum zio_stage pipeline
= zio
->io_pipeline
;
2314 enum zio_stage stage
= zio
->io_stage
;
2316 zio
->io_executor
= curthread
;
2318 ASSERT(!MUTEX_HELD(&zio
->io_lock
));
2319 ASSERT(ISP2(stage
));
2320 ASSERT(zio
->io_stall
== NULL
);
2324 } while ((stage
& pipeline
) == 0);
2326 ASSERT(stage
<= ZIO_STAGE_DONE
);
2329 * If we are in interrupt context and this pipeline stage
2330 * will grab a config lock that is held across I/O,
2331 * or may wait for an I/O that needs an interrupt thread
2332 * to complete, issue async to avoid deadlock.
2334 * For VDEV_IO_START, we cut in line so that the io will
2335 * be sent to disk promptly.
2337 if ((stage
& ZIO_BLOCKING_STAGES
) && zio
->io_vd
== NULL
&&
2338 zio_taskq_member(zio
, ZIO_TASKQ_INTERRUPT
)) {
2339 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2340 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2341 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2346 * If the current context doesn't have large enough stacks
2347 * the zio must be issued asynchronously to prevent overflow.
2349 if (zio_execute_stack_check(zio
)) {
2350 boolean_t cut
= (stage
== ZIO_STAGE_VDEV_IO_START
) ?
2351 zio_requeue_io_start_cut_in_line
: B_FALSE
;
2352 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, cut
);
2356 zio
->io_stage
= stage
;
2357 zio
->io_pipeline_trace
|= zio
->io_stage
;
2360 * The zio pipeline stage returns the next zio to execute
2361 * (typically the same as this one), or NULL if we should
2364 zio
= zio_pipeline
[highbit64(stage
) - 1](zio
);
2373 * ==========================================================================
2374 * Initiate I/O, either sync or async
2375 * ==========================================================================
2378 zio_wait(zio_t
*zio
)
2381 * Some routines, like zio_free_sync(), may return a NULL zio
2382 * to avoid the performance overhead of creating and then destroying
2383 * an unneeded zio. For the callers' simplicity, we accept a NULL
2384 * zio and ignore it.
2389 long timeout
= MSEC_TO_TICK(zfs_deadman_ziotime_ms
);
2392 ASSERT3S(zio
->io_stage
, ==, ZIO_STAGE_OPEN
);
2393 ASSERT3P(zio
->io_executor
, ==, NULL
);
2395 zio
->io_waiter
= curthread
;
2396 ASSERT0(zio
->io_queued_timestamp
);
2397 zio
->io_queued_timestamp
= gethrtime();
2399 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
2400 spa_select_allocator(zio
);
2404 mutex_enter(&zio
->io_lock
);
2405 while (zio
->io_executor
!= NULL
) {
2406 error
= cv_timedwait_io(&zio
->io_cv
, &zio
->io_lock
,
2407 ddi_get_lbolt() + timeout
);
2409 if (zfs_deadman_enabled
&& error
== -1 &&
2410 gethrtime() - zio
->io_queued_timestamp
>
2411 spa_deadman_ziotime(zio
->io_spa
)) {
2412 mutex_exit(&zio
->io_lock
);
2413 timeout
= MSEC_TO_TICK(zfs_deadman_checktime_ms
);
2414 zio_deadman(zio
, FTAG
);
2415 mutex_enter(&zio
->io_lock
);
2418 mutex_exit(&zio
->io_lock
);
2420 error
= zio
->io_error
;
2427 zio_nowait(zio_t
*zio
)
2430 * See comment in zio_wait().
2435 ASSERT3P(zio
->io_executor
, ==, NULL
);
2437 if (zio
->io_child_type
== ZIO_CHILD_LOGICAL
&&
2438 list_is_empty(&zio
->io_parent_list
)) {
2442 * This is a logical async I/O with no parent to wait for it.
2443 * We add it to the spa_async_root_zio "Godfather" I/O which
2444 * will ensure they complete prior to unloading the pool.
2446 spa_t
*spa
= zio
->io_spa
;
2447 pio
= spa
->spa_async_zio_root
[CPU_SEQID_UNSTABLE
];
2449 zio_add_child(pio
, zio
);
2452 ASSERT0(zio
->io_queued_timestamp
);
2453 zio
->io_queued_timestamp
= gethrtime();
2454 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
2455 spa_select_allocator(zio
);
2461 * ==========================================================================
2462 * Reexecute, cancel, or suspend/resume failed I/O
2463 * ==========================================================================
2467 zio_reexecute(void *arg
)
2470 zio_t
*cio
, *cio_next
, *gio
;
2472 ASSERT(pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2473 ASSERT(pio
->io_orig_stage
== ZIO_STAGE_OPEN
);
2474 ASSERT(pio
->io_gang_leader
== NULL
);
2475 ASSERT(pio
->io_gang_tree
== NULL
);
2477 mutex_enter(&pio
->io_lock
);
2478 pio
->io_flags
= pio
->io_orig_flags
;
2479 pio
->io_stage
= pio
->io_orig_stage
;
2480 pio
->io_pipeline
= pio
->io_orig_pipeline
;
2481 pio
->io_reexecute
= 0;
2482 pio
->io_flags
|= ZIO_FLAG_REEXECUTED
;
2483 pio
->io_pipeline_trace
= 0;
2485 pio
->io_state
[ZIO_WAIT_READY
] = (pio
->io_stage
>= ZIO_STAGE_READY
) ||
2486 (pio
->io_pipeline
& ZIO_STAGE_READY
) == 0;
2487 pio
->io_state
[ZIO_WAIT_DONE
] = (pio
->io_stage
>= ZIO_STAGE_DONE
);
2488 zio_link_t
*zl
= NULL
;
2489 while ((gio
= zio_walk_parents(pio
, &zl
)) != NULL
) {
2490 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++) {
2491 gio
->io_children
[pio
->io_child_type
][w
] +=
2495 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
2496 pio
->io_child_error
[c
] = 0;
2498 if (IO_IS_ALLOCATING(pio
))
2499 BP_ZERO(pio
->io_bp
);
2502 * As we reexecute pio's children, new children could be created.
2503 * New children go to the head of pio's io_child_list, however,
2504 * so we will (correctly) not reexecute them. The key is that
2505 * the remainder of pio's io_child_list, from 'cio_next' onward,
2506 * cannot be affected by any side effects of reexecuting 'cio'.
2509 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
2510 cio_next
= zio_walk_children(pio
, &zl
);
2511 mutex_exit(&pio
->io_lock
);
2513 mutex_enter(&pio
->io_lock
);
2515 mutex_exit(&pio
->io_lock
);
2518 * Now that all children have been reexecuted, execute the parent.
2519 * We don't reexecute "The Godfather" I/O here as it's the
2520 * responsibility of the caller to wait on it.
2522 if (!(pio
->io_flags
& ZIO_FLAG_GODFATHER
)) {
2523 pio
->io_queued_timestamp
= gethrtime();
2529 zio_suspend(spa_t
*spa
, zio_t
*zio
, zio_suspend_reason_t reason
)
2531 if (spa_get_failmode(spa
) == ZIO_FAILURE_MODE_PANIC
)
2532 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2533 "failure and the failure mode property for this pool "
2534 "is set to panic.", spa_name(spa
));
2536 cmn_err(CE_WARN
, "Pool '%s' has encountered an uncorrectable I/O "
2537 "failure and has been suspended.\n", spa_name(spa
));
2539 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE
, spa
, NULL
,
2542 mutex_enter(&spa
->spa_suspend_lock
);
2544 if (spa
->spa_suspend_zio_root
== NULL
)
2545 spa
->spa_suspend_zio_root
= zio_root(spa
, NULL
, NULL
,
2546 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
2547 ZIO_FLAG_GODFATHER
);
2549 spa
->spa_suspended
= reason
;
2552 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
2553 ASSERT(zio
!= spa
->spa_suspend_zio_root
);
2554 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
2555 ASSERT(zio_unique_parent(zio
) == NULL
);
2556 ASSERT(zio
->io_stage
== ZIO_STAGE_DONE
);
2557 zio_add_child(spa
->spa_suspend_zio_root
, zio
);
2560 mutex_exit(&spa
->spa_suspend_lock
);
2564 zio_resume(spa_t
*spa
)
2569 * Reexecute all previously suspended i/o.
2571 mutex_enter(&spa
->spa_suspend_lock
);
2572 spa
->spa_suspended
= ZIO_SUSPEND_NONE
;
2573 cv_broadcast(&spa
->spa_suspend_cv
);
2574 pio
= spa
->spa_suspend_zio_root
;
2575 spa
->spa_suspend_zio_root
= NULL
;
2576 mutex_exit(&spa
->spa_suspend_lock
);
2582 return (zio_wait(pio
));
2586 zio_resume_wait(spa_t
*spa
)
2588 mutex_enter(&spa
->spa_suspend_lock
);
2589 while (spa_suspended(spa
))
2590 cv_wait(&spa
->spa_suspend_cv
, &spa
->spa_suspend_lock
);
2591 mutex_exit(&spa
->spa_suspend_lock
);
2595 * ==========================================================================
2598 * A gang block is a collection of small blocks that looks to the DMU
2599 * like one large block. When zio_dva_allocate() cannot find a block
2600 * of the requested size, due to either severe fragmentation or the pool
2601 * being nearly full, it calls zio_write_gang_block() to construct the
2602 * block from smaller fragments.
2604 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2605 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2606 * an indirect block: it's an array of block pointers. It consumes
2607 * only one sector and hence is allocatable regardless of fragmentation.
2608 * The gang header's bps point to its gang members, which hold the data.
2610 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2611 * as the verifier to ensure uniqueness of the SHA256 checksum.
2612 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2613 * not the gang header. This ensures that data block signatures (needed for
2614 * deduplication) are independent of how the block is physically stored.
2616 * Gang blocks can be nested: a gang member may itself be a gang block.
2617 * Thus every gang block is a tree in which root and all interior nodes are
2618 * gang headers, and the leaves are normal blocks that contain user data.
2619 * The root of the gang tree is called the gang leader.
2621 * To perform any operation (read, rewrite, free, claim) on a gang block,
2622 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2623 * in the io_gang_tree field of the original logical i/o by recursively
2624 * reading the gang leader and all gang headers below it. This yields
2625 * an in-core tree containing the contents of every gang header and the
2626 * bps for every constituent of the gang block.
2628 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2629 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2630 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2631 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2632 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2633 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2634 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2635 * of the gang header plus zio_checksum_compute() of the data to update the
2636 * gang header's blk_cksum as described above.
2638 * The two-phase assemble/issue model solves the problem of partial failure --
2639 * what if you'd freed part of a gang block but then couldn't read the
2640 * gang header for another part? Assembling the entire gang tree first
2641 * ensures that all the necessary gang header I/O has succeeded before
2642 * starting the actual work of free, claim, or write. Once the gang tree
2643 * is assembled, free and claim are in-memory operations that cannot fail.
2645 * In the event that a gang write fails, zio_dva_unallocate() walks the
2646 * gang tree to immediately free (i.e. insert back into the space map)
2647 * everything we've allocated. This ensures that we don't get ENOSPC
2648 * errors during repeated suspend/resume cycles due to a flaky device.
2650 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2651 * the gang tree, we won't modify the block, so we can safely defer the free
2652 * (knowing that the block is still intact). If we *can* assemble the gang
2653 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2654 * each constituent bp and we can allocate a new block on the next sync pass.
2656 * In all cases, the gang tree allows complete recovery from partial failure.
2657 * ==========================================================================
2661 zio_gang_issue_func_done(zio_t
*zio
)
2663 abd_free(zio
->io_abd
);
2667 zio_read_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2673 return (zio_read(pio
, pio
->io_spa
, bp
, abd_get_offset(data
, offset
),
2674 BP_GET_PSIZE(bp
), zio_gang_issue_func_done
,
2675 NULL
, pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2676 &pio
->io_bookmark
));
2680 zio_rewrite_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2687 abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2688 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2689 gbh_abd
, SPA_GANGBLOCKSIZE
, zio_gang_issue_func_done
, NULL
,
2690 pio
->io_priority
, ZIO_GANG_CHILD_FLAGS(pio
),
2693 * As we rewrite each gang header, the pipeline will compute
2694 * a new gang block header checksum for it; but no one will
2695 * compute a new data checksum, so we do that here. The one
2696 * exception is the gang leader: the pipeline already computed
2697 * its data checksum because that stage precedes gang assembly.
2698 * (Presently, nothing actually uses interior data checksums;
2699 * this is just good hygiene.)
2701 if (gn
!= pio
->io_gang_leader
->io_gang_tree
) {
2702 abd_t
*buf
= abd_get_offset(data
, offset
);
2704 zio_checksum_compute(zio
, BP_GET_CHECKSUM(bp
),
2705 buf
, BP_GET_PSIZE(bp
));
2710 * If we are here to damage data for testing purposes,
2711 * leave the GBH alone so that we can detect the damage.
2713 if (pio
->io_gang_leader
->io_flags
& ZIO_FLAG_INDUCE_DAMAGE
)
2714 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
2716 zio
= zio_rewrite(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2717 abd_get_offset(data
, offset
), BP_GET_PSIZE(bp
),
2718 zio_gang_issue_func_done
, NULL
, pio
->io_priority
,
2719 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
2726 zio_free_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2729 (void) gn
, (void) data
, (void) offset
;
2731 zio_t
*zio
= zio_free_sync(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2732 ZIO_GANG_CHILD_FLAGS(pio
));
2734 zio
= zio_null(pio
, pio
->io_spa
,
2735 NULL
, NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
));
2741 zio_claim_gang(zio_t
*pio
, blkptr_t
*bp
, zio_gang_node_t
*gn
, abd_t
*data
,
2744 (void) gn
, (void) data
, (void) offset
;
2745 return (zio_claim(pio
, pio
->io_spa
, pio
->io_txg
, bp
,
2746 NULL
, NULL
, ZIO_GANG_CHILD_FLAGS(pio
)));
2749 static zio_gang_issue_func_t
*zio_gang_issue_func
[ZIO_TYPES
] = {
2758 static void zio_gang_tree_assemble_done(zio_t
*zio
);
2760 static zio_gang_node_t
*
2761 zio_gang_node_alloc(zio_gang_node_t
**gnpp
)
2763 zio_gang_node_t
*gn
;
2765 ASSERT(*gnpp
== NULL
);
2767 gn
= kmem_zalloc(sizeof (*gn
), KM_SLEEP
);
2768 gn
->gn_gbh
= zio_buf_alloc(SPA_GANGBLOCKSIZE
);
2775 zio_gang_node_free(zio_gang_node_t
**gnpp
)
2777 zio_gang_node_t
*gn
= *gnpp
;
2779 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2780 ASSERT(gn
->gn_child
[g
] == NULL
);
2782 zio_buf_free(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2783 kmem_free(gn
, sizeof (*gn
));
2788 zio_gang_tree_free(zio_gang_node_t
**gnpp
)
2790 zio_gang_node_t
*gn
= *gnpp
;
2795 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++)
2796 zio_gang_tree_free(&gn
->gn_child
[g
]);
2798 zio_gang_node_free(gnpp
);
2802 zio_gang_tree_assemble(zio_t
*gio
, blkptr_t
*bp
, zio_gang_node_t
**gnpp
)
2804 zio_gang_node_t
*gn
= zio_gang_node_alloc(gnpp
);
2805 abd_t
*gbh_abd
= abd_get_from_buf(gn
->gn_gbh
, SPA_GANGBLOCKSIZE
);
2807 ASSERT(gio
->io_gang_leader
== gio
);
2808 ASSERT(BP_IS_GANG(bp
));
2810 zio_nowait(zio_read(gio
, gio
->io_spa
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
2811 zio_gang_tree_assemble_done
, gn
, gio
->io_priority
,
2812 ZIO_GANG_CHILD_FLAGS(gio
), &gio
->io_bookmark
));
2816 zio_gang_tree_assemble_done(zio_t
*zio
)
2818 zio_t
*gio
= zio
->io_gang_leader
;
2819 zio_gang_node_t
*gn
= zio
->io_private
;
2820 blkptr_t
*bp
= zio
->io_bp
;
2822 ASSERT(gio
== zio_unique_parent(zio
));
2823 ASSERT(list_is_empty(&zio
->io_child_list
));
2828 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2829 if (BP_SHOULD_BYTESWAP(bp
))
2830 byteswap_uint64_array(abd_to_buf(zio
->io_abd
), zio
->io_size
);
2832 ASSERT3P(abd_to_buf(zio
->io_abd
), ==, gn
->gn_gbh
);
2833 ASSERT(zio
->io_size
== SPA_GANGBLOCKSIZE
);
2834 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2836 abd_free(zio
->io_abd
);
2838 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2839 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2840 if (!BP_IS_GANG(gbp
))
2842 zio_gang_tree_assemble(gio
, gbp
, &gn
->gn_child
[g
]);
2847 zio_gang_tree_issue(zio_t
*pio
, zio_gang_node_t
*gn
, blkptr_t
*bp
, abd_t
*data
,
2850 zio_t
*gio
= pio
->io_gang_leader
;
2853 ASSERT(BP_IS_GANG(bp
) == !!gn
);
2854 ASSERT(BP_GET_CHECKSUM(bp
) == BP_GET_CHECKSUM(gio
->io_bp
));
2855 ASSERT(BP_GET_LSIZE(bp
) == BP_GET_PSIZE(bp
) || gn
== gio
->io_gang_tree
);
2858 * If you're a gang header, your data is in gn->gn_gbh.
2859 * If you're a gang member, your data is in 'data' and gn == NULL.
2861 zio
= zio_gang_issue_func
[gio
->io_type
](pio
, bp
, gn
, data
, offset
);
2864 ASSERT(gn
->gn_gbh
->zg_tail
.zec_magic
== ZEC_MAGIC
);
2866 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
2867 blkptr_t
*gbp
= &gn
->gn_gbh
->zg_blkptr
[g
];
2868 if (BP_IS_HOLE(gbp
))
2870 zio_gang_tree_issue(zio
, gn
->gn_child
[g
], gbp
, data
,
2872 offset
+= BP_GET_PSIZE(gbp
);
2876 if (gn
== gio
->io_gang_tree
)
2877 ASSERT3U(gio
->io_size
, ==, offset
);
2884 zio_gang_assemble(zio_t
*zio
)
2886 blkptr_t
*bp
= zio
->io_bp
;
2888 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== NULL
);
2889 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2891 zio
->io_gang_leader
= zio
;
2893 zio_gang_tree_assemble(zio
, bp
, &zio
->io_gang_tree
);
2899 zio_gang_issue(zio_t
*zio
)
2901 blkptr_t
*bp
= zio
->io_bp
;
2903 if (zio_wait_for_children(zio
, ZIO_CHILD_GANG_BIT
, ZIO_WAIT_DONE
)) {
2907 ASSERT(BP_IS_GANG(bp
) && zio
->io_gang_leader
== zio
);
2908 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
2910 if (zio
->io_child_error
[ZIO_CHILD_GANG
] == 0)
2911 zio_gang_tree_issue(zio
, zio
->io_gang_tree
, bp
, zio
->io_abd
,
2914 zio_gang_tree_free(&zio
->io_gang_tree
);
2916 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
2922 zio_gang_inherit_allocator(zio_t
*pio
, zio_t
*cio
)
2924 cio
->io_allocator
= pio
->io_allocator
;
2925 cio
->io_wr_iss_tq
= pio
->io_wr_iss_tq
;
2929 zio_write_gang_member_ready(zio_t
*zio
)
2931 zio_t
*pio
= zio_unique_parent(zio
);
2932 dva_t
*cdva
= zio
->io_bp
->blk_dva
;
2933 dva_t
*pdva
= pio
->io_bp
->blk_dva
;
2935 zio_t
*gio __maybe_unused
= zio
->io_gang_leader
;
2937 if (BP_IS_HOLE(zio
->io_bp
))
2940 ASSERT(BP_IS_HOLE(&zio
->io_bp_orig
));
2942 ASSERT(zio
->io_child_type
== ZIO_CHILD_GANG
);
2943 ASSERT3U(zio
->io_prop
.zp_copies
, ==, gio
->io_prop
.zp_copies
);
2944 ASSERT3U(zio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(zio
->io_bp
));
2945 ASSERT3U(pio
->io_prop
.zp_copies
, <=, BP_GET_NDVAS(pio
->io_bp
));
2946 VERIFY3U(BP_GET_NDVAS(zio
->io_bp
), <=, BP_GET_NDVAS(pio
->io_bp
));
2948 mutex_enter(&pio
->io_lock
);
2949 for (int d
= 0; d
< BP_GET_NDVAS(zio
->io_bp
); d
++) {
2950 ASSERT(DVA_GET_GANG(&pdva
[d
]));
2951 asize
= DVA_GET_ASIZE(&pdva
[d
]);
2952 asize
+= DVA_GET_ASIZE(&cdva
[d
]);
2953 DVA_SET_ASIZE(&pdva
[d
], asize
);
2955 mutex_exit(&pio
->io_lock
);
2959 zio_write_gang_done(zio_t
*zio
)
2962 * The io_abd field will be NULL for a zio with no data. The io_flags
2963 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2964 * check for it here as it is cleared in zio_ready.
2966 if (zio
->io_abd
!= NULL
)
2967 abd_free(zio
->io_abd
);
2971 zio_write_gang_block(zio_t
*pio
, metaslab_class_t
*mc
)
2973 spa_t
*spa
= pio
->io_spa
;
2974 blkptr_t
*bp
= pio
->io_bp
;
2975 zio_t
*gio
= pio
->io_gang_leader
;
2977 zio_gang_node_t
*gn
, **gnpp
;
2978 zio_gbh_phys_t
*gbh
;
2980 uint64_t txg
= pio
->io_txg
;
2981 uint64_t resid
= pio
->io_size
;
2983 int copies
= gio
->io_prop
.zp_copies
;
2986 boolean_t has_data
= !(pio
->io_flags
& ZIO_FLAG_NODATA
);
2989 * If one copy was requested, store 2 copies of the GBH, so that we
2990 * can still traverse all the data (e.g. to free or scrub) even if a
2991 * block is damaged. Note that we can't store 3 copies of the GBH in
2992 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
2994 int gbh_copies
= copies
;
2995 if (gbh_copies
== 1) {
2996 gbh_copies
= MIN(2, spa_max_replication(spa
));
2999 ASSERT(ZIO_HAS_ALLOCATOR(pio
));
3000 int flags
= METASLAB_HINTBP_FAVOR
| METASLAB_GANG_HEADER
;
3001 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3002 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3005 flags
|= METASLAB_ASYNC_ALLOC
;
3006 VERIFY(zfs_refcount_held(&mc
->mc_allocator
[pio
->io_allocator
].
3007 mca_alloc_slots
, pio
));
3010 * The logical zio has already placed a reservation for
3011 * 'copies' allocation slots but gang blocks may require
3012 * additional copies. These additional copies
3013 * (i.e. gbh_copies - copies) are guaranteed to succeed
3014 * since metaslab_class_throttle_reserve() always allows
3015 * additional reservations for gang blocks.
3017 VERIFY(metaslab_class_throttle_reserve(mc
, gbh_copies
- copies
,
3018 pio
->io_allocator
, pio
, flags
));
3021 error
= metaslab_alloc(spa
, mc
, SPA_GANGBLOCKSIZE
,
3022 bp
, gbh_copies
, txg
, pio
== gio
? NULL
: gio
->io_bp
, flags
,
3023 &pio
->io_alloc_list
, pio
, pio
->io_allocator
);
3025 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3026 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3030 * If we failed to allocate the gang block header then
3031 * we remove any additional allocation reservations that
3032 * we placed here. The original reservation will
3033 * be removed when the logical I/O goes to the ready
3036 metaslab_class_throttle_unreserve(mc
,
3037 gbh_copies
- copies
, pio
->io_allocator
, pio
);
3040 pio
->io_error
= error
;
3045 gnpp
= &gio
->io_gang_tree
;
3047 gnpp
= pio
->io_private
;
3048 ASSERT(pio
->io_ready
== zio_write_gang_member_ready
);
3051 gn
= zio_gang_node_alloc(gnpp
);
3053 memset(gbh
, 0, SPA_GANGBLOCKSIZE
);
3054 gbh_abd
= abd_get_from_buf(gbh
, SPA_GANGBLOCKSIZE
);
3057 * Create the gang header.
3059 zio
= zio_rewrite(pio
, spa
, txg
, bp
, gbh_abd
, SPA_GANGBLOCKSIZE
,
3060 zio_write_gang_done
, NULL
, pio
->io_priority
,
3061 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
3063 zio_gang_inherit_allocator(pio
, zio
);
3066 * Create and nowait the gang children.
3068 for (int g
= 0; resid
!= 0; resid
-= lsize
, g
++) {
3069 lsize
= P2ROUNDUP(resid
/ (SPA_GBH_NBLKPTRS
- g
),
3071 ASSERT(lsize
>= SPA_MINBLOCKSIZE
&& lsize
<= resid
);
3073 zp
.zp_checksum
= gio
->io_prop
.zp_checksum
;
3074 zp
.zp_compress
= ZIO_COMPRESS_OFF
;
3075 zp
.zp_complevel
= gio
->io_prop
.zp_complevel
;
3076 zp
.zp_type
= DMU_OT_NONE
;
3078 zp
.zp_copies
= gio
->io_prop
.zp_copies
;
3079 zp
.zp_dedup
= B_FALSE
;
3080 zp
.zp_dedup_verify
= B_FALSE
;
3081 zp
.zp_nopwrite
= B_FALSE
;
3082 zp
.zp_encrypt
= gio
->io_prop
.zp_encrypt
;
3083 zp
.zp_byteorder
= gio
->io_prop
.zp_byteorder
;
3084 memset(zp
.zp_salt
, 0, ZIO_DATA_SALT_LEN
);
3085 memset(zp
.zp_iv
, 0, ZIO_DATA_IV_LEN
);
3086 memset(zp
.zp_mac
, 0, ZIO_DATA_MAC_LEN
);
3088 zio_t
*cio
= zio_write(zio
, spa
, txg
, &gbh
->zg_blkptr
[g
],
3089 has_data
? abd_get_offset(pio
->io_abd
, pio
->io_size
-
3090 resid
) : NULL
, lsize
, lsize
, &zp
,
3091 zio_write_gang_member_ready
, NULL
,
3092 zio_write_gang_done
, &gn
->gn_child
[g
], pio
->io_priority
,
3093 ZIO_GANG_CHILD_FLAGS(pio
), &pio
->io_bookmark
);
3095 zio_gang_inherit_allocator(zio
, cio
);
3097 if (pio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
3098 ASSERT(pio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
3102 * Gang children won't throttle but we should
3103 * account for their work, so reserve an allocation
3104 * slot for them here.
3106 VERIFY(metaslab_class_throttle_reserve(mc
,
3107 zp
.zp_copies
, cio
->io_allocator
, cio
, flags
));
3113 * Set pio's pipeline to just wait for zio to finish.
3115 pio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3123 * The zio_nop_write stage in the pipeline determines if allocating a
3124 * new bp is necessary. The nopwrite feature can handle writes in
3125 * either syncing or open context (i.e. zil writes) and as a result is
3126 * mutually exclusive with dedup.
3128 * By leveraging a cryptographically secure checksum, such as SHA256, we
3129 * can compare the checksums of the new data and the old to determine if
3130 * allocating a new block is required. Note that our requirements for
3131 * cryptographic strength are fairly weak: there can't be any accidental
3132 * hash collisions, but we don't need to be secure against intentional
3133 * (malicious) collisions. To trigger a nopwrite, you have to be able
3134 * to write the file to begin with, and triggering an incorrect (hash
3135 * collision) nopwrite is no worse than simply writing to the file.
3136 * That said, there are no known attacks against the checksum algorithms
3137 * used for nopwrite, assuming that the salt and the checksums
3138 * themselves remain secret.
3141 zio_nop_write(zio_t
*zio
)
3143 blkptr_t
*bp
= zio
->io_bp
;
3144 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3145 zio_prop_t
*zp
= &zio
->io_prop
;
3147 ASSERT(BP_IS_HOLE(bp
));
3148 ASSERT(BP_GET_LEVEL(bp
) == 0);
3149 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
3150 ASSERT(zp
->zp_nopwrite
);
3151 ASSERT(!zp
->zp_dedup
);
3152 ASSERT(zio
->io_bp_override
== NULL
);
3153 ASSERT(IO_IS_ALLOCATING(zio
));
3156 * Check to see if the original bp and the new bp have matching
3157 * characteristics (i.e. same checksum, compression algorithms, etc).
3158 * If they don't then just continue with the pipeline which will
3159 * allocate a new bp.
3161 if (BP_IS_HOLE(bp_orig
) ||
3162 !(zio_checksum_table
[BP_GET_CHECKSUM(bp
)].ci_flags
&
3163 ZCHECKSUM_FLAG_NOPWRITE
) ||
3164 BP_IS_ENCRYPTED(bp
) || BP_IS_ENCRYPTED(bp_orig
) ||
3165 BP_GET_CHECKSUM(bp
) != BP_GET_CHECKSUM(bp_orig
) ||
3166 BP_GET_COMPRESS(bp
) != BP_GET_COMPRESS(bp_orig
) ||
3167 BP_GET_DEDUP(bp
) != BP_GET_DEDUP(bp_orig
) ||
3168 zp
->zp_copies
!= BP_GET_NDVAS(bp_orig
))
3172 * If the checksums match then reset the pipeline so that we
3173 * avoid allocating a new bp and issuing any I/O.
3175 if (ZIO_CHECKSUM_EQUAL(bp
->blk_cksum
, bp_orig
->blk_cksum
)) {
3176 ASSERT(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3177 ZCHECKSUM_FLAG_NOPWRITE
);
3178 ASSERT3U(BP_GET_PSIZE(bp
), ==, BP_GET_PSIZE(bp_orig
));
3179 ASSERT3U(BP_GET_LSIZE(bp
), ==, BP_GET_LSIZE(bp_orig
));
3180 ASSERT(zp
->zp_compress
!= ZIO_COMPRESS_OFF
);
3181 ASSERT3U(bp
->blk_prop
, ==, bp_orig
->blk_prop
);
3184 * If we're overwriting a block that is currently on an
3185 * indirect vdev, then ignore the nopwrite request and
3186 * allow a new block to be allocated on a concrete vdev.
3188 spa_config_enter(zio
->io_spa
, SCL_VDEV
, FTAG
, RW_READER
);
3189 for (int d
= 0; d
< BP_GET_NDVAS(bp_orig
); d
++) {
3190 vdev_t
*tvd
= vdev_lookup_top(zio
->io_spa
,
3191 DVA_GET_VDEV(&bp_orig
->blk_dva
[d
]));
3192 if (tvd
->vdev_ops
== &vdev_indirect_ops
) {
3193 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
3197 spa_config_exit(zio
->io_spa
, SCL_VDEV
, FTAG
);
3200 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3201 zio
->io_flags
|= ZIO_FLAG_NOPWRITE
;
3208 * ==========================================================================
3209 * Block Reference Table
3210 * ==========================================================================
3213 zio_brt_free(zio_t
*zio
)
3219 if (BP_GET_LEVEL(bp
) > 0 ||
3220 BP_IS_METADATA(bp
) ||
3221 !brt_maybe_exists(zio
->io_spa
, bp
)) {
3225 if (!brt_entry_decref(zio
->io_spa
, bp
)) {
3227 * This isn't the last reference, so we cannot free
3230 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
3237 * ==========================================================================
3239 * ==========================================================================
3242 zio_ddt_child_read_done(zio_t
*zio
)
3244 blkptr_t
*bp
= zio
->io_bp
;
3245 ddt_entry_t
*dde
= zio
->io_private
;
3247 zio_t
*pio
= zio_unique_parent(zio
);
3249 mutex_enter(&pio
->io_lock
);
3250 ddp
= ddt_phys_select(dde
, bp
);
3251 if (zio
->io_error
== 0)
3252 ddt_phys_clear(ddp
); /* this ddp doesn't need repair */
3254 if (zio
->io_error
== 0 && dde
->dde_repair_abd
== NULL
)
3255 dde
->dde_repair_abd
= zio
->io_abd
;
3257 abd_free(zio
->io_abd
);
3258 mutex_exit(&pio
->io_lock
);
3262 zio_ddt_read_start(zio_t
*zio
)
3264 blkptr_t
*bp
= zio
->io_bp
;
3266 ASSERT(BP_GET_DEDUP(bp
));
3267 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3268 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3270 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3271 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3272 ddt_entry_t
*dde
= ddt_repair_start(ddt
, bp
);
3273 ddt_phys_t
*ddp
= dde
->dde_phys
;
3274 ddt_phys_t
*ddp_self
= ddt_phys_select(dde
, bp
);
3277 ASSERT(zio
->io_vsd
== NULL
);
3280 if (ddp_self
== NULL
)
3283 for (int p
= 0; p
< DDT_PHYS_TYPES
; p
++, ddp
++) {
3284 if (ddp
->ddp_phys_birth
== 0 || ddp
== ddp_self
)
3286 ddt_bp_create(ddt
->ddt_checksum
, &dde
->dde_key
, ddp
,
3288 zio_nowait(zio_read(zio
, zio
->io_spa
, &blk
,
3289 abd_alloc_for_io(zio
->io_size
, B_TRUE
),
3290 zio
->io_size
, zio_ddt_child_read_done
, dde
,
3291 zio
->io_priority
, ZIO_DDT_CHILD_FLAGS(zio
) |
3292 ZIO_FLAG_DONT_PROPAGATE
, &zio
->io_bookmark
));
3297 zio_nowait(zio_read(zio
, zio
->io_spa
, bp
,
3298 zio
->io_abd
, zio
->io_size
, NULL
, NULL
, zio
->io_priority
,
3299 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
));
3305 zio_ddt_read_done(zio_t
*zio
)
3307 blkptr_t
*bp
= zio
->io_bp
;
3309 if (zio_wait_for_children(zio
, ZIO_CHILD_DDT_BIT
, ZIO_WAIT_DONE
)) {
3313 ASSERT(BP_GET_DEDUP(bp
));
3314 ASSERT(BP_GET_PSIZE(bp
) == zio
->io_size
);
3315 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3317 if (zio
->io_child_error
[ZIO_CHILD_DDT
]) {
3318 ddt_t
*ddt
= ddt_select(zio
->io_spa
, bp
);
3319 ddt_entry_t
*dde
= zio
->io_vsd
;
3321 ASSERT(spa_load_state(zio
->io_spa
) != SPA_LOAD_NONE
);
3325 zio
->io_stage
= ZIO_STAGE_DDT_READ_START
>> 1;
3326 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_FALSE
);
3329 if (dde
->dde_repair_abd
!= NULL
) {
3330 abd_copy(zio
->io_abd
, dde
->dde_repair_abd
,
3332 zio
->io_child_error
[ZIO_CHILD_DDT
] = 0;
3334 ddt_repair_done(ddt
, dde
);
3338 ASSERT(zio
->io_vsd
== NULL
);
3344 zio_ddt_collision(zio_t
*zio
, ddt_t
*ddt
, ddt_entry_t
*dde
)
3346 spa_t
*spa
= zio
->io_spa
;
3347 boolean_t do_raw
= !!(zio
->io_flags
& ZIO_FLAG_RAW
);
3349 ASSERT(!(zio
->io_bp_override
&& do_raw
));
3352 * Note: we compare the original data, not the transformed data,
3353 * because when zio->io_bp is an override bp, we will not have
3354 * pushed the I/O transforms. That's an important optimization
3355 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3356 * However, we should never get a raw, override zio so in these
3357 * cases we can compare the io_abd directly. This is useful because
3358 * it allows us to do dedup verification even if we don't have access
3359 * to the original data (for instance, if the encryption keys aren't
3363 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3364 zio_t
*lio
= dde
->dde_lead_zio
[p
];
3366 if (lio
!= NULL
&& do_raw
) {
3367 return (lio
->io_size
!= zio
->io_size
||
3368 abd_cmp(zio
->io_abd
, lio
->io_abd
) != 0);
3369 } else if (lio
!= NULL
) {
3370 return (lio
->io_orig_size
!= zio
->io_orig_size
||
3371 abd_cmp(zio
->io_orig_abd
, lio
->io_orig_abd
) != 0);
3375 for (int p
= DDT_PHYS_SINGLE
; p
<= DDT_PHYS_TRIPLE
; p
++) {
3376 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3378 if (ddp
->ddp_phys_birth
!= 0 && do_raw
) {
3379 blkptr_t blk
= *zio
->io_bp
;
3384 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3385 psize
= BP_GET_PSIZE(&blk
);
3387 if (psize
!= zio
->io_size
)
3392 tmpabd
= abd_alloc_for_io(psize
, B_TRUE
);
3394 error
= zio_wait(zio_read(NULL
, spa
, &blk
, tmpabd
,
3395 psize
, NULL
, NULL
, ZIO_PRIORITY_SYNC_READ
,
3396 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
|
3397 ZIO_FLAG_RAW
, &zio
->io_bookmark
));
3400 if (abd_cmp(tmpabd
, zio
->io_abd
) != 0)
3401 error
= SET_ERROR(ENOENT
);
3406 return (error
!= 0);
3407 } else if (ddp
->ddp_phys_birth
!= 0) {
3408 arc_buf_t
*abuf
= NULL
;
3409 arc_flags_t aflags
= ARC_FLAG_WAIT
;
3410 blkptr_t blk
= *zio
->io_bp
;
3413 ddt_bp_fill(ddp
, &blk
, ddp
->ddp_phys_birth
);
3415 if (BP_GET_LSIZE(&blk
) != zio
->io_orig_size
)
3420 error
= arc_read(NULL
, spa
, &blk
,
3421 arc_getbuf_func
, &abuf
, ZIO_PRIORITY_SYNC_READ
,
3422 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
3423 &aflags
, &zio
->io_bookmark
);
3426 if (abd_cmp_buf(zio
->io_orig_abd
, abuf
->b_data
,
3427 zio
->io_orig_size
) != 0)
3428 error
= SET_ERROR(ENOENT
);
3429 arc_buf_destroy(abuf
, &abuf
);
3433 return (error
!= 0);
3441 zio_ddt_child_write_ready(zio_t
*zio
)
3443 int p
= zio
->io_prop
.zp_copies
;
3444 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3445 ddt_entry_t
*dde
= zio
->io_private
;
3446 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3454 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3456 ddt_phys_fill(ddp
, zio
->io_bp
);
3458 zio_link_t
*zl
= NULL
;
3459 while ((pio
= zio_walk_parents(zio
, &zl
)) != NULL
)
3460 ddt_bp_fill(ddp
, pio
->io_bp
, zio
->io_txg
);
3466 zio_ddt_child_write_done(zio_t
*zio
)
3468 int p
= zio
->io_prop
.zp_copies
;
3469 ddt_t
*ddt
= ddt_select(zio
->io_spa
, zio
->io_bp
);
3470 ddt_entry_t
*dde
= zio
->io_private
;
3471 ddt_phys_t
*ddp
= &dde
->dde_phys
[p
];
3475 ASSERT(ddp
->ddp_refcnt
== 0);
3476 ASSERT(dde
->dde_lead_zio
[p
] == zio
);
3477 dde
->dde_lead_zio
[p
] = NULL
;
3479 if (zio
->io_error
== 0) {
3480 zio_link_t
*zl
= NULL
;
3481 while (zio_walk_parents(zio
, &zl
) != NULL
)
3482 ddt_phys_addref(ddp
);
3484 ddt_phys_clear(ddp
);
3491 zio_ddt_write(zio_t
*zio
)
3493 spa_t
*spa
= zio
->io_spa
;
3494 blkptr_t
*bp
= zio
->io_bp
;
3495 uint64_t txg
= zio
->io_txg
;
3496 zio_prop_t
*zp
= &zio
->io_prop
;
3497 int p
= zp
->zp_copies
;
3499 ddt_t
*ddt
= ddt_select(spa
, bp
);
3503 ASSERT(BP_GET_DEDUP(bp
));
3504 ASSERT(BP_GET_CHECKSUM(bp
) == zp
->zp_checksum
);
3505 ASSERT(BP_IS_HOLE(bp
) || zio
->io_bp_override
);
3506 ASSERT(!(zio
->io_bp_override
&& (zio
->io_flags
& ZIO_FLAG_RAW
)));
3509 dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3510 ddp
= &dde
->dde_phys
[p
];
3512 if (zp
->zp_dedup_verify
&& zio_ddt_collision(zio
, ddt
, dde
)) {
3514 * If we're using a weak checksum, upgrade to a strong checksum
3515 * and try again. If we're already using a strong checksum,
3516 * we can't resolve it, so just convert to an ordinary write.
3517 * (And automatically e-mail a paper to Nature?)
3519 if (!(zio_checksum_table
[zp
->zp_checksum
].ci_flags
&
3520 ZCHECKSUM_FLAG_DEDUP
)) {
3521 zp
->zp_checksum
= spa_dedup_checksum(spa
);
3522 zio_pop_transforms(zio
);
3523 zio
->io_stage
= ZIO_STAGE_OPEN
;
3526 zp
->zp_dedup
= B_FALSE
;
3527 BP_SET_DEDUP(bp
, B_FALSE
);
3529 ASSERT(!BP_GET_DEDUP(bp
));
3530 zio
->io_pipeline
= ZIO_WRITE_PIPELINE
;
3535 if (ddp
->ddp_phys_birth
!= 0 || dde
->dde_lead_zio
[p
] != NULL
) {
3536 if (ddp
->ddp_phys_birth
!= 0)
3537 ddt_bp_fill(ddp
, bp
, txg
);
3538 if (dde
->dde_lead_zio
[p
] != NULL
)
3539 zio_add_child(zio
, dde
->dde_lead_zio
[p
]);
3541 ddt_phys_addref(ddp
);
3542 } else if (zio
->io_bp_override
) {
3543 ASSERT(BP_GET_LOGICAL_BIRTH(bp
) == txg
);
3544 ASSERT(BP_EQUAL(bp
, zio
->io_bp_override
));
3545 ddt_phys_fill(ddp
, bp
);
3546 ddt_phys_addref(ddp
);
3548 cio
= zio_write(zio
, spa
, txg
, bp
, zio
->io_orig_abd
,
3549 zio
->io_orig_size
, zio
->io_orig_size
, zp
,
3550 zio_ddt_child_write_ready
, NULL
,
3551 zio_ddt_child_write_done
, dde
, zio
->io_priority
,
3552 ZIO_DDT_CHILD_FLAGS(zio
), &zio
->io_bookmark
);
3554 zio_push_transform(cio
, zio
->io_abd
, zio
->io_size
, 0, NULL
);
3555 dde
->dde_lead_zio
[p
] = cio
;
3565 static ddt_entry_t
*freedde
; /* for debugging */
3568 zio_ddt_free(zio_t
*zio
)
3570 spa_t
*spa
= zio
->io_spa
;
3571 blkptr_t
*bp
= zio
->io_bp
;
3572 ddt_t
*ddt
= ddt_select(spa
, bp
);
3576 ASSERT(BP_GET_DEDUP(bp
));
3577 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
3580 freedde
= dde
= ddt_lookup(ddt
, bp
, B_TRUE
);
3582 ddp
= ddt_phys_select(dde
, bp
);
3584 ddt_phys_decref(ddp
);
3592 * ==========================================================================
3593 * Allocate and free blocks
3594 * ==========================================================================
3598 zio_io_to_allocate(spa_t
*spa
, int allocator
)
3602 ASSERT(MUTEX_HELD(&spa
->spa_allocs
[allocator
].spaa_lock
));
3604 zio
= avl_first(&spa
->spa_allocs
[allocator
].spaa_tree
);
3608 ASSERT(IO_IS_ALLOCATING(zio
));
3609 ASSERT(ZIO_HAS_ALLOCATOR(zio
));
3612 * Try to place a reservation for this zio. If we're unable to
3613 * reserve then we throttle.
3615 ASSERT3U(zio
->io_allocator
, ==, allocator
);
3616 if (!metaslab_class_throttle_reserve(zio
->io_metaslab_class
,
3617 zio
->io_prop
.zp_copies
, allocator
, zio
, 0)) {
3621 avl_remove(&spa
->spa_allocs
[allocator
].spaa_tree
, zio
);
3622 ASSERT3U(zio
->io_stage
, <, ZIO_STAGE_DVA_ALLOCATE
);
3628 zio_dva_throttle(zio_t
*zio
)
3630 spa_t
*spa
= zio
->io_spa
;
3632 metaslab_class_t
*mc
;
3634 /* locate an appropriate allocation class */
3635 mc
= spa_preferred_class(spa
, zio
->io_size
, zio
->io_prop
.zp_type
,
3636 zio
->io_prop
.zp_level
, zio
->io_prop
.zp_zpl_smallblk
);
3638 if (zio
->io_priority
== ZIO_PRIORITY_SYNC_WRITE
||
3639 !mc
->mc_alloc_throttle_enabled
||
3640 zio
->io_child_type
== ZIO_CHILD_GANG
||
3641 zio
->io_flags
& ZIO_FLAG_NODATA
) {
3645 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
3646 ASSERT(ZIO_HAS_ALLOCATOR(zio
));
3647 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3648 ASSERT3U(zio
->io_queued_timestamp
, >, 0);
3649 ASSERT(zio
->io_stage
== ZIO_STAGE_DVA_THROTTLE
);
3651 int allocator
= zio
->io_allocator
;
3652 zio
->io_metaslab_class
= mc
;
3653 mutex_enter(&spa
->spa_allocs
[allocator
].spaa_lock
);
3654 avl_add(&spa
->spa_allocs
[allocator
].spaa_tree
, zio
);
3655 nio
= zio_io_to_allocate(spa
, allocator
);
3656 mutex_exit(&spa
->spa_allocs
[allocator
].spaa_lock
);
3661 zio_allocate_dispatch(spa_t
*spa
, int allocator
)
3665 mutex_enter(&spa
->spa_allocs
[allocator
].spaa_lock
);
3666 zio
= zio_io_to_allocate(spa
, allocator
);
3667 mutex_exit(&spa
->spa_allocs
[allocator
].spaa_lock
);
3671 ASSERT3U(zio
->io_stage
, ==, ZIO_STAGE_DVA_THROTTLE
);
3672 ASSERT0(zio
->io_error
);
3673 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
, B_TRUE
);
3677 zio_dva_allocate(zio_t
*zio
)
3679 spa_t
*spa
= zio
->io_spa
;
3680 metaslab_class_t
*mc
;
3681 blkptr_t
*bp
= zio
->io_bp
;
3685 if (zio
->io_gang_leader
== NULL
) {
3686 ASSERT(zio
->io_child_type
> ZIO_CHILD_GANG
);
3687 zio
->io_gang_leader
= zio
;
3690 ASSERT(BP_IS_HOLE(bp
));
3691 ASSERT0(BP_GET_NDVAS(bp
));
3692 ASSERT3U(zio
->io_prop
.zp_copies
, >, 0);
3693 ASSERT3U(zio
->io_prop
.zp_copies
, <=, spa_max_replication(spa
));
3694 ASSERT3U(zio
->io_size
, ==, BP_GET_PSIZE(bp
));
3696 if (zio
->io_flags
& ZIO_FLAG_NODATA
)
3697 flags
|= METASLAB_DONT_THROTTLE
;
3698 if (zio
->io_flags
& ZIO_FLAG_GANG_CHILD
)
3699 flags
|= METASLAB_GANG_CHILD
;
3700 if (zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
)
3701 flags
|= METASLAB_ASYNC_ALLOC
;
3704 * if not already chosen, locate an appropriate allocation class
3706 mc
= zio
->io_metaslab_class
;
3708 mc
= spa_preferred_class(spa
, zio
->io_size
,
3709 zio
->io_prop
.zp_type
, zio
->io_prop
.zp_level
,
3710 zio
->io_prop
.zp_zpl_smallblk
);
3711 zio
->io_metaslab_class
= mc
;
3715 * Try allocating the block in the usual metaslab class.
3716 * If that's full, allocate it in the normal class.
3717 * If that's full, allocate as a gang block,
3718 * and if all are full, the allocation fails (which shouldn't happen).
3720 * Note that we do not fall back on embedded slog (ZIL) space, to
3721 * preserve unfragmented slog space, which is critical for decent
3722 * sync write performance. If a log allocation fails, we will fall
3723 * back to spa_sync() which is abysmal for performance.
3725 ASSERT(ZIO_HAS_ALLOCATOR(zio
));
3726 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3727 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3728 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3731 * Fallback to normal class when an alloc class is full
3733 if (error
== ENOSPC
&& mc
!= spa_normal_class(spa
)) {
3735 * If throttling, transfer reservation over to normal class.
3736 * The io_allocator slot can remain the same even though we
3737 * are switching classes.
3739 if (mc
->mc_alloc_throttle_enabled
&&
3740 (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
)) {
3741 metaslab_class_throttle_unreserve(mc
,
3742 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
);
3743 zio
->io_flags
&= ~ZIO_FLAG_IO_ALLOCATING
;
3745 VERIFY(metaslab_class_throttle_reserve(
3746 spa_normal_class(spa
),
3747 zio
->io_prop
.zp_copies
, zio
->io_allocator
, zio
,
3748 flags
| METASLAB_MUST_RESERVE
));
3750 zio
->io_metaslab_class
= mc
= spa_normal_class(spa
);
3751 if (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
) {
3752 zfs_dbgmsg("%s: metaslab allocation failure, "
3753 "trying normal class: zio %px, size %llu, error %d",
3754 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3758 error
= metaslab_alloc(spa
, mc
, zio
->io_size
, bp
,
3759 zio
->io_prop
.zp_copies
, zio
->io_txg
, NULL
, flags
,
3760 &zio
->io_alloc_list
, zio
, zio
->io_allocator
);
3763 if (error
== ENOSPC
&& zio
->io_size
> SPA_MINBLOCKSIZE
) {
3764 if (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
) {
3765 zfs_dbgmsg("%s: metaslab allocation failure, "
3766 "trying ganging: zio %px, size %llu, error %d",
3767 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3770 return (zio_write_gang_block(zio
, mc
));
3773 if (error
!= ENOSPC
||
3774 (zfs_flags
& ZFS_DEBUG_METASLAB_ALLOC
)) {
3775 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3776 "size %llu, error %d",
3777 spa_name(spa
), zio
, (u_longlong_t
)zio
->io_size
,
3780 zio
->io_error
= error
;
3787 zio_dva_free(zio_t
*zio
)
3789 metaslab_free(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
, B_FALSE
);
3795 zio_dva_claim(zio_t
*zio
)
3799 error
= metaslab_claim(zio
->io_spa
, zio
->io_bp
, zio
->io_txg
);
3801 zio
->io_error
= error
;
3807 * Undo an allocation. This is used by zio_done() when an I/O fails
3808 * and we want to give back the block we just allocated.
3809 * This handles both normal blocks and gang blocks.
3812 zio_dva_unallocate(zio_t
*zio
, zio_gang_node_t
*gn
, blkptr_t
*bp
)
3814 ASSERT(BP_GET_LOGICAL_BIRTH(bp
) == zio
->io_txg
|| BP_IS_HOLE(bp
));
3815 ASSERT(zio
->io_bp_override
== NULL
);
3817 if (!BP_IS_HOLE(bp
)) {
3818 metaslab_free(zio
->io_spa
, bp
, BP_GET_LOGICAL_BIRTH(bp
),
3823 for (int g
= 0; g
< SPA_GBH_NBLKPTRS
; g
++) {
3824 zio_dva_unallocate(zio
, gn
->gn_child
[g
],
3825 &gn
->gn_gbh
->zg_blkptr
[g
]);
3831 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3834 zio_alloc_zil(spa_t
*spa
, objset_t
*os
, uint64_t txg
, blkptr_t
*new_bp
,
3835 uint64_t size
, boolean_t
*slog
)
3838 zio_alloc_list_t io_alloc_list
;
3840 ASSERT(txg
> spa_syncing_txg(spa
));
3842 metaslab_trace_init(&io_alloc_list
);
3845 * Block pointer fields are useful to metaslabs for stats and debugging.
3846 * Fill in the obvious ones before calling into metaslab_alloc().
3848 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3849 BP_SET_PSIZE(new_bp
, size
);
3850 BP_SET_LEVEL(new_bp
, 0);
3853 * When allocating a zil block, we don't have information about
3854 * the final destination of the block except the objset it's part
3855 * of, so we just hash the objset ID to pick the allocator to get
3858 int flags
= METASLAB_ZIL
;
3859 int allocator
= (uint_t
)cityhash4(0, 0, 0,
3860 os
->os_dsl_dataset
->ds_object
) % spa
->spa_alloc_count
;
3861 error
= metaslab_alloc(spa
, spa_log_class(spa
), size
, new_bp
, 1,
3862 txg
, NULL
, flags
, &io_alloc_list
, NULL
, allocator
);
3863 *slog
= (error
== 0);
3865 error
= metaslab_alloc(spa
, spa_embedded_log_class(spa
), size
,
3866 new_bp
, 1, txg
, NULL
, flags
,
3867 &io_alloc_list
, NULL
, allocator
);
3870 error
= metaslab_alloc(spa
, spa_normal_class(spa
), size
,
3871 new_bp
, 1, txg
, NULL
, flags
,
3872 &io_alloc_list
, NULL
, allocator
);
3874 metaslab_trace_fini(&io_alloc_list
);
3877 BP_SET_LSIZE(new_bp
, size
);
3878 BP_SET_PSIZE(new_bp
, size
);
3879 BP_SET_COMPRESS(new_bp
, ZIO_COMPRESS_OFF
);
3880 BP_SET_CHECKSUM(new_bp
,
3881 spa_version(spa
) >= SPA_VERSION_SLIM_ZIL
3882 ? ZIO_CHECKSUM_ZILOG2
: ZIO_CHECKSUM_ZILOG
);
3883 BP_SET_TYPE(new_bp
, DMU_OT_INTENT_LOG
);
3884 BP_SET_LEVEL(new_bp
, 0);
3885 BP_SET_DEDUP(new_bp
, 0);
3886 BP_SET_BYTEORDER(new_bp
, ZFS_HOST_BYTEORDER
);
3889 * encrypted blocks will require an IV and salt. We generate
3890 * these now since we will not be rewriting the bp at
3893 if (os
->os_encrypted
) {
3894 uint8_t iv
[ZIO_DATA_IV_LEN
];
3895 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3897 BP_SET_CRYPT(new_bp
, B_TRUE
);
3898 VERIFY0(spa_crypt_get_salt(spa
,
3899 dmu_objset_id(os
), salt
));
3900 VERIFY0(zio_crypt_generate_iv(iv
));
3902 zio_crypt_encode_params_bp(new_bp
, salt
, iv
);
3905 zfs_dbgmsg("%s: zil block allocation failure: "
3906 "size %llu, error %d", spa_name(spa
), (u_longlong_t
)size
,
3914 * ==========================================================================
3915 * Read and write to physical devices
3916 * ==========================================================================
3920 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3921 * stops after this stage and will resume upon I/O completion.
3922 * However, there are instances where the vdev layer may need to
3923 * continue the pipeline when an I/O was not issued. Since the I/O
3924 * that was sent to the vdev layer might be different than the one
3925 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3926 * force the underlying vdev layers to call either zio_execute() or
3927 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3930 zio_vdev_io_start(zio_t
*zio
)
3932 vdev_t
*vd
= zio
->io_vd
;
3934 spa_t
*spa
= zio
->io_spa
;
3938 ASSERT(zio
->io_error
== 0);
3939 ASSERT(zio
->io_child_error
[ZIO_CHILD_VDEV
] == 0);
3942 if (!(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
3943 spa_config_enter(spa
, SCL_ZIO
, zio
, RW_READER
);
3946 * The mirror_ops handle multiple DVAs in a single BP.
3948 vdev_mirror_ops
.vdev_op_io_start(zio
);
3952 ASSERT3P(zio
->io_logical
, !=, zio
);
3953 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3954 ASSERT(spa
->spa_trust_config
);
3957 * Note: the code can handle other kinds of writes,
3958 * but we don't expect them.
3960 if (zio
->io_vd
->vdev_noalloc
) {
3961 ASSERT(zio
->io_flags
&
3962 (ZIO_FLAG_PHYSICAL
| ZIO_FLAG_SELF_HEAL
|
3963 ZIO_FLAG_RESILVER
| ZIO_FLAG_INDUCE_DAMAGE
));
3967 align
= 1ULL << vd
->vdev_top
->vdev_ashift
;
3969 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
) &&
3970 P2PHASE(zio
->io_size
, align
) != 0) {
3971 /* Transform logical writes to be a full physical block size. */
3972 uint64_t asize
= P2ROUNDUP(zio
->io_size
, align
);
3973 abd_t
*abuf
= abd_alloc_sametype(zio
->io_abd
, asize
);
3974 ASSERT(vd
== vd
->vdev_top
);
3975 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
3976 abd_copy(abuf
, zio
->io_abd
, zio
->io_size
);
3977 abd_zero_off(abuf
, zio
->io_size
, asize
- zio
->io_size
);
3979 zio_push_transform(zio
, abuf
, asize
, asize
, zio_subblock
);
3983 * If this is not a physical io, make sure that it is properly aligned
3984 * before proceeding.
3986 if (!(zio
->io_flags
& ZIO_FLAG_PHYSICAL
)) {
3987 ASSERT0(P2PHASE(zio
->io_offset
, align
));
3988 ASSERT0(P2PHASE(zio
->io_size
, align
));
3991 * For physical writes, we allow 512b aligned writes and assume
3992 * the device will perform a read-modify-write as necessary.
3994 ASSERT0(P2PHASE(zio
->io_offset
, SPA_MINBLOCKSIZE
));
3995 ASSERT0(P2PHASE(zio
->io_size
, SPA_MINBLOCKSIZE
));
3998 VERIFY(zio
->io_type
!= ZIO_TYPE_WRITE
|| spa_writeable(spa
));
4001 * If this is a repair I/O, and there's no self-healing involved --
4002 * that is, we're just resilvering what we expect to resilver --
4003 * then don't do the I/O unless zio's txg is actually in vd's DTL.
4004 * This prevents spurious resilvering.
4006 * There are a few ways that we can end up creating these spurious
4009 * 1. A resilver i/o will be issued if any DVA in the BP has a
4010 * dirty DTL. The mirror code will issue resilver writes to
4011 * each DVA, including the one(s) that are not on vdevs with dirty
4014 * 2. With nested replication, which happens when we have a
4015 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
4016 * For example, given mirror(replacing(A+B), C), it's likely that
4017 * only A is out of date (it's the new device). In this case, we'll
4018 * read from C, then use the data to resilver A+B -- but we don't
4019 * actually want to resilver B, just A. The top-level mirror has no
4020 * way to know this, so instead we just discard unnecessary repairs
4021 * as we work our way down the vdev tree.
4023 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
4024 * The same logic applies to any form of nested replication: ditto
4025 * + mirror, RAID-Z + replacing, etc.
4027 * However, indirect vdevs point off to other vdevs which may have
4028 * DTL's, so we never bypass them. The child i/os on concrete vdevs
4029 * will be properly bypassed instead.
4031 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
4032 * a dRAID spare vdev. For example, when a dRAID spare is first
4033 * used, its spare blocks need to be written to but the leaf vdev's
4034 * of such blocks can have empty DTL_PARTIAL.
4036 * There seemed no clean way to allow such writes while bypassing
4037 * spurious ones. At this point, just avoid all bypassing for dRAID
4040 if ((zio
->io_flags
& ZIO_FLAG_IO_REPAIR
) &&
4041 !(zio
->io_flags
& ZIO_FLAG_SELF_HEAL
) &&
4042 zio
->io_txg
!= 0 && /* not a delegated i/o */
4043 vd
->vdev_ops
!= &vdev_indirect_ops
&&
4044 vd
->vdev_top
->vdev_ops
!= &vdev_draid_ops
&&
4045 !vdev_dtl_contains(vd
, DTL_PARTIAL
, zio
->io_txg
, 1)) {
4046 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4047 zio_vdev_io_bypass(zio
);
4052 * Select the next best leaf I/O to process. Distributed spares are
4053 * excluded since they dispatch the I/O directly to a leaf vdev after
4054 * applying the dRAID mapping.
4056 if (vd
->vdev_ops
->vdev_op_leaf
&&
4057 vd
->vdev_ops
!= &vdev_draid_spare_ops
&&
4058 (zio
->io_type
== ZIO_TYPE_READ
||
4059 zio
->io_type
== ZIO_TYPE_WRITE
||
4060 zio
->io_type
== ZIO_TYPE_TRIM
)) {
4062 if ((zio
= vdev_queue_io(zio
)) == NULL
)
4065 if (!vdev_accessible(vd
, zio
)) {
4066 zio
->io_error
= SET_ERROR(ENXIO
);
4070 zio
->io_delay
= gethrtime();
4073 vd
->vdev_ops
->vdev_op_io_start(zio
);
4078 zio_vdev_io_done(zio_t
*zio
)
4080 vdev_t
*vd
= zio
->io_vd
;
4081 vdev_ops_t
*ops
= vd
? vd
->vdev_ops
: &vdev_mirror_ops
;
4082 boolean_t unexpected_error
= B_FALSE
;
4084 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
4088 ASSERT(zio
->io_type
== ZIO_TYPE_READ
||
4089 zio
->io_type
== ZIO_TYPE_WRITE
||
4090 zio
->io_type
== ZIO_TYPE_IOCTL
||
4091 zio
->io_type
== ZIO_TYPE_TRIM
);
4094 zio
->io_delay
= gethrtime() - zio
->io_delay
;
4096 if (vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
4097 vd
->vdev_ops
!= &vdev_draid_spare_ops
) {
4098 if (zio
->io_type
!= ZIO_TYPE_IOCTL
)
4099 vdev_queue_io_done(zio
);
4101 if (zio_injection_enabled
&& zio
->io_error
== 0)
4102 zio
->io_error
= zio_handle_device_injections(vd
, zio
,
4105 if (zio_injection_enabled
&& zio
->io_error
== 0)
4106 zio
->io_error
= zio_handle_label_injection(zio
, EIO
);
4108 if (zio
->io_error
&& zio
->io_type
!= ZIO_TYPE_TRIM
) {
4109 if (!vdev_accessible(vd
, zio
)) {
4110 zio
->io_error
= SET_ERROR(ENXIO
);
4112 unexpected_error
= B_TRUE
;
4117 ops
->vdev_op_io_done(zio
);
4119 if (unexpected_error
&& vd
->vdev_remove_wanted
== B_FALSE
)
4120 VERIFY(vdev_probe(vd
, zio
) == NULL
);
4126 * This function is used to change the priority of an existing zio that is
4127 * currently in-flight. This is used by the arc to upgrade priority in the
4128 * event that a demand read is made for a block that is currently queued
4129 * as a scrub or async read IO. Otherwise, the high priority read request
4130 * would end up having to wait for the lower priority IO.
4133 zio_change_priority(zio_t
*pio
, zio_priority_t priority
)
4135 zio_t
*cio
, *cio_next
;
4136 zio_link_t
*zl
= NULL
;
4138 ASSERT3U(priority
, <, ZIO_PRIORITY_NUM_QUEUEABLE
);
4140 if (pio
->io_vd
!= NULL
&& pio
->io_vd
->vdev_ops
->vdev_op_leaf
) {
4141 vdev_queue_change_io_priority(pio
, priority
);
4143 pio
->io_priority
= priority
;
4146 mutex_enter(&pio
->io_lock
);
4147 for (cio
= zio_walk_children(pio
, &zl
); cio
!= NULL
; cio
= cio_next
) {
4148 cio_next
= zio_walk_children(pio
, &zl
);
4149 zio_change_priority(cio
, priority
);
4151 mutex_exit(&pio
->io_lock
);
4155 * For non-raidz ZIOs, we can just copy aside the bad data read from the
4156 * disk, and use that to finish the checksum ereport later.
4159 zio_vsd_default_cksum_finish(zio_cksum_report_t
*zcr
,
4160 const abd_t
*good_buf
)
4162 /* no processing needed */
4163 zfs_ereport_finish_checksum(zcr
, good_buf
, zcr
->zcr_cbdata
, B_FALSE
);
4167 zio_vsd_default_cksum_report(zio_t
*zio
, zio_cksum_report_t
*zcr
)
4169 void *abd
= abd_alloc_sametype(zio
->io_abd
, zio
->io_size
);
4171 abd_copy(abd
, zio
->io_abd
, zio
->io_size
);
4173 zcr
->zcr_cbinfo
= zio
->io_size
;
4174 zcr
->zcr_cbdata
= abd
;
4175 zcr
->zcr_finish
= zio_vsd_default_cksum_finish
;
4176 zcr
->zcr_free
= zio_abd_free
;
4180 zio_vdev_io_assess(zio_t
*zio
)
4182 vdev_t
*vd
= zio
->io_vd
;
4184 if (zio_wait_for_children(zio
, ZIO_CHILD_VDEV_BIT
, ZIO_WAIT_DONE
)) {
4188 if (vd
== NULL
&& !(zio
->io_flags
& ZIO_FLAG_CONFIG_WRITER
))
4189 spa_config_exit(zio
->io_spa
, SCL_ZIO
, zio
);
4191 if (zio
->io_vsd
!= NULL
) {
4192 zio
->io_vsd_ops
->vsd_free(zio
);
4196 if (zio_injection_enabled
&& zio
->io_error
== 0)
4197 zio
->io_error
= zio_handle_fault_injection(zio
, EIO
);
4200 * If the I/O failed, determine whether we should attempt to retry it.
4202 * On retry, we cut in line in the issue queue, since we don't want
4203 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4205 if (zio
->io_error
&& vd
== NULL
&&
4206 !(zio
->io_flags
& (ZIO_FLAG_DONT_RETRY
| ZIO_FLAG_IO_RETRY
))) {
4207 ASSERT(!(zio
->io_flags
& ZIO_FLAG_DONT_QUEUE
)); /* not a leaf */
4208 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_BYPASS
)); /* not a leaf */
4210 zio
->io_flags
|= ZIO_FLAG_IO_RETRY
| ZIO_FLAG_DONT_AGGREGATE
;
4211 zio
->io_stage
= ZIO_STAGE_VDEV_IO_START
>> 1;
4212 zio_taskq_dispatch(zio
, ZIO_TASKQ_ISSUE
,
4213 zio_requeue_io_start_cut_in_line
);
4218 * If we got an error on a leaf device, convert it to ENXIO
4219 * if the device is not accessible at all.
4221 if (zio
->io_error
&& vd
!= NULL
&& vd
->vdev_ops
->vdev_op_leaf
&&
4222 !vdev_accessible(vd
, zio
))
4223 zio
->io_error
= SET_ERROR(ENXIO
);
4226 * If we can't write to an interior vdev (mirror or RAID-Z),
4227 * set vdev_cant_write so that we stop trying to allocate from it.
4229 if (zio
->io_error
== ENXIO
&& zio
->io_type
== ZIO_TYPE_WRITE
&&
4230 vd
!= NULL
&& !vd
->vdev_ops
->vdev_op_leaf
) {
4231 vdev_dbgmsg(vd
, "zio_vdev_io_assess(zio=%px) setting "
4232 "cant_write=TRUE due to write failure with ENXIO",
4234 vd
->vdev_cant_write
= B_TRUE
;
4238 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4239 * attempts will ever succeed. In this case we set a persistent
4240 * boolean flag so that we don't bother with it in the future.
4242 if ((zio
->io_error
== ENOTSUP
|| zio
->io_error
== ENOTTY
) &&
4243 zio
->io_type
== ZIO_TYPE_IOCTL
&&
4244 zio
->io_cmd
== DKIOCFLUSHWRITECACHE
&& vd
!= NULL
)
4245 vd
->vdev_nowritecache
= B_TRUE
;
4248 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4254 zio_vdev_io_reissue(zio_t
*zio
)
4256 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
4257 ASSERT(zio
->io_error
== 0);
4259 zio
->io_stage
>>= 1;
4263 zio_vdev_io_redone(zio_t
*zio
)
4265 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_DONE
);
4267 zio
->io_stage
>>= 1;
4271 zio_vdev_io_bypass(zio_t
*zio
)
4273 ASSERT(zio
->io_stage
== ZIO_STAGE_VDEV_IO_START
);
4274 ASSERT(zio
->io_error
== 0);
4276 zio
->io_flags
|= ZIO_FLAG_IO_BYPASS
;
4277 zio
->io_stage
= ZIO_STAGE_VDEV_IO_ASSESS
>> 1;
4281 * ==========================================================================
4282 * Encrypt and store encryption parameters
4283 * ==========================================================================
4288 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4289 * managing the storage of encryption parameters and passing them to the
4290 * lower-level encryption functions.
4293 zio_encrypt(zio_t
*zio
)
4295 zio_prop_t
*zp
= &zio
->io_prop
;
4296 spa_t
*spa
= zio
->io_spa
;
4297 blkptr_t
*bp
= zio
->io_bp
;
4298 uint64_t psize
= BP_GET_PSIZE(bp
);
4299 uint64_t dsobj
= zio
->io_bookmark
.zb_objset
;
4300 dmu_object_type_t ot
= BP_GET_TYPE(bp
);
4301 void *enc_buf
= NULL
;
4303 uint8_t salt
[ZIO_DATA_SALT_LEN
];
4304 uint8_t iv
[ZIO_DATA_IV_LEN
];
4305 uint8_t mac
[ZIO_DATA_MAC_LEN
];
4306 boolean_t no_crypt
= B_FALSE
;
4308 /* the root zio already encrypted the data */
4309 if (zio
->io_child_type
== ZIO_CHILD_GANG
)
4312 /* only ZIL blocks are re-encrypted on rewrite */
4313 if (!IO_IS_ALLOCATING(zio
) && ot
!= DMU_OT_INTENT_LOG
)
4316 if (!(zp
->zp_encrypt
|| BP_IS_ENCRYPTED(bp
))) {
4317 BP_SET_CRYPT(bp
, B_FALSE
);
4321 /* if we are doing raw encryption set the provided encryption params */
4322 if (zio
->io_flags
& ZIO_FLAG_RAW_ENCRYPT
) {
4323 ASSERT0(BP_GET_LEVEL(bp
));
4324 BP_SET_CRYPT(bp
, B_TRUE
);
4325 BP_SET_BYTEORDER(bp
, zp
->zp_byteorder
);
4326 if (ot
!= DMU_OT_OBJSET
)
4327 zio_crypt_encode_mac_bp(bp
, zp
->zp_mac
);
4329 /* dnode blocks must be written out in the provided byteorder */
4330 if (zp
->zp_byteorder
!= ZFS_HOST_BYTEORDER
&&
4331 ot
== DMU_OT_DNODE
) {
4332 void *bswap_buf
= zio_buf_alloc(psize
);
4333 abd_t
*babd
= abd_get_from_buf(bswap_buf
, psize
);
4335 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4336 abd_copy_to_buf(bswap_buf
, zio
->io_abd
, psize
);
4337 dmu_ot_byteswap
[DMU_OT_BYTESWAP(ot
)].ob_func(bswap_buf
,
4340 abd_take_ownership_of_buf(babd
, B_TRUE
);
4341 zio_push_transform(zio
, babd
, psize
, psize
, NULL
);
4344 if (DMU_OT_IS_ENCRYPTED(ot
))
4345 zio_crypt_encode_params_bp(bp
, zp
->zp_salt
, zp
->zp_iv
);
4349 /* indirect blocks only maintain a cksum of the lower level MACs */
4350 if (BP_GET_LEVEL(bp
) > 0) {
4351 BP_SET_CRYPT(bp
, B_TRUE
);
4352 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE
,
4353 zio
->io_orig_abd
, BP_GET_LSIZE(bp
), BP_SHOULD_BYTESWAP(bp
),
4355 zio_crypt_encode_mac_bp(bp
, mac
);
4360 * Objset blocks are a special case since they have 2 256-bit MACs
4361 * embedded within them.
4363 if (ot
== DMU_OT_OBJSET
) {
4364 ASSERT0(DMU_OT_IS_ENCRYPTED(ot
));
4365 ASSERT3U(BP_GET_COMPRESS(bp
), ==, ZIO_COMPRESS_OFF
);
4366 BP_SET_CRYPT(bp
, B_TRUE
);
4367 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE
, spa
, dsobj
,
4368 zio
->io_abd
, psize
, BP_SHOULD_BYTESWAP(bp
)));
4372 /* unencrypted object types are only authenticated with a MAC */
4373 if (!DMU_OT_IS_ENCRYPTED(ot
)) {
4374 BP_SET_CRYPT(bp
, B_TRUE
);
4375 VERIFY0(spa_do_crypt_mac_abd(B_TRUE
, spa
, dsobj
,
4376 zio
->io_abd
, psize
, mac
));
4377 zio_crypt_encode_mac_bp(bp
, mac
);
4382 * Later passes of sync-to-convergence may decide to rewrite data
4383 * in place to avoid more disk reallocations. This presents a problem
4384 * for encryption because this constitutes rewriting the new data with
4385 * the same encryption key and IV. However, this only applies to blocks
4386 * in the MOS (particularly the spacemaps) and we do not encrypt the
4387 * MOS. We assert that the zio is allocating or an intent log write
4390 ASSERT(IO_IS_ALLOCATING(zio
) || ot
== DMU_OT_INTENT_LOG
);
4391 ASSERT(BP_GET_LEVEL(bp
) == 0 || ot
== DMU_OT_INTENT_LOG
);
4392 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_ENCRYPTION
));
4393 ASSERT3U(psize
, !=, 0);
4395 enc_buf
= zio_buf_alloc(psize
);
4396 eabd
= abd_get_from_buf(enc_buf
, psize
);
4397 abd_take_ownership_of_buf(eabd
, B_TRUE
);
4400 * For an explanation of what encryption parameters are stored
4401 * where, see the block comment in zio_crypt.c.
4403 if (ot
== DMU_OT_INTENT_LOG
) {
4404 zio_crypt_decode_params_bp(bp
, salt
, iv
);
4406 BP_SET_CRYPT(bp
, B_TRUE
);
4409 /* Perform the encryption. This should not fail */
4410 VERIFY0(spa_do_crypt_abd(B_TRUE
, spa
, &zio
->io_bookmark
,
4411 BP_GET_TYPE(bp
), BP_GET_DEDUP(bp
), BP_SHOULD_BYTESWAP(bp
),
4412 salt
, iv
, mac
, psize
, zio
->io_abd
, eabd
, &no_crypt
));
4414 /* encode encryption metadata into the bp */
4415 if (ot
== DMU_OT_INTENT_LOG
) {
4417 * ZIL blocks store the MAC in the embedded checksum, so the
4418 * transform must always be applied.
4420 zio_crypt_encode_mac_zil(enc_buf
, mac
);
4421 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4423 BP_SET_CRYPT(bp
, B_TRUE
);
4424 zio_crypt_encode_params_bp(bp
, salt
, iv
);
4425 zio_crypt_encode_mac_bp(bp
, mac
);
4428 ASSERT3U(ot
, ==, DMU_OT_DNODE
);
4431 zio_push_transform(zio
, eabd
, psize
, psize
, NULL
);
4439 * ==========================================================================
4440 * Generate and verify checksums
4441 * ==========================================================================
4444 zio_checksum_generate(zio_t
*zio
)
4446 blkptr_t
*bp
= zio
->io_bp
;
4447 enum zio_checksum checksum
;
4451 * This is zio_write_phys().
4452 * We're either generating a label checksum, or none at all.
4454 checksum
= zio
->io_prop
.zp_checksum
;
4456 if (checksum
== ZIO_CHECKSUM_OFF
)
4459 ASSERT(checksum
== ZIO_CHECKSUM_LABEL
);
4461 if (BP_IS_GANG(bp
) && zio
->io_child_type
== ZIO_CHILD_GANG
) {
4462 ASSERT(!IO_IS_ALLOCATING(zio
));
4463 checksum
= ZIO_CHECKSUM_GANG_HEADER
;
4465 checksum
= BP_GET_CHECKSUM(bp
);
4469 zio_checksum_compute(zio
, checksum
, zio
->io_abd
, zio
->io_size
);
4475 zio_checksum_verify(zio_t
*zio
)
4477 zio_bad_cksum_t info
;
4478 blkptr_t
*bp
= zio
->io_bp
;
4481 ASSERT(zio
->io_vd
!= NULL
);
4485 * This is zio_read_phys().
4486 * We're either verifying a label checksum, or nothing at all.
4488 if (zio
->io_prop
.zp_checksum
== ZIO_CHECKSUM_OFF
)
4491 ASSERT3U(zio
->io_prop
.zp_checksum
, ==, ZIO_CHECKSUM_LABEL
);
4494 if ((error
= zio_checksum_error(zio
, &info
)) != 0) {
4495 zio
->io_error
= error
;
4496 if (error
== ECKSUM
&&
4497 !(zio
->io_flags
& ZIO_FLAG_SPECULATIVE
)) {
4498 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4499 zio
->io_vd
->vdev_stat
.vs_checksum_errors
++;
4500 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4501 (void) zfs_ereport_start_checksum(zio
->io_spa
,
4502 zio
->io_vd
, &zio
->io_bookmark
, zio
,
4503 zio
->io_offset
, zio
->io_size
, &info
);
4511 * Called by RAID-Z to ensure we don't compute the checksum twice.
4514 zio_checksum_verified(zio_t
*zio
)
4516 zio
->io_pipeline
&= ~ZIO_STAGE_CHECKSUM_VERIFY
;
4520 * ==========================================================================
4521 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4522 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4523 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4524 * indicate errors that are specific to one I/O, and most likely permanent.
4525 * Any other error is presumed to be worse because we weren't expecting it.
4526 * ==========================================================================
4529 zio_worst_error(int e1
, int e2
)
4531 static int zio_error_rank
[] = { 0, ENXIO
, ECKSUM
, EIO
};
4534 for (r1
= 0; r1
< sizeof (zio_error_rank
) / sizeof (int); r1
++)
4535 if (e1
== zio_error_rank
[r1
])
4538 for (r2
= 0; r2
< sizeof (zio_error_rank
) / sizeof (int); r2
++)
4539 if (e2
== zio_error_rank
[r2
])
4542 return (r1
> r2
? e1
: e2
);
4546 * ==========================================================================
4548 * ==========================================================================
4551 zio_ready(zio_t
*zio
)
4553 blkptr_t
*bp
= zio
->io_bp
;
4554 zio_t
*pio
, *pio_next
;
4555 zio_link_t
*zl
= NULL
;
4557 if (zio_wait_for_children(zio
, ZIO_CHILD_LOGICAL_BIT
|
4558 ZIO_CHILD_GANG_BIT
| ZIO_CHILD_DDT_BIT
, ZIO_WAIT_READY
)) {
4562 if (zio
->io_ready
) {
4563 ASSERT(IO_IS_ALLOCATING(zio
));
4564 ASSERT(BP_GET_LOGICAL_BIRTH(bp
) == zio
->io_txg
||
4565 BP_IS_HOLE(bp
) || (zio
->io_flags
& ZIO_FLAG_NOPWRITE
));
4566 ASSERT(zio
->io_children
[ZIO_CHILD_GANG
][ZIO_WAIT_READY
] == 0);
4572 if (bp
!= NULL
&& bp
!= &zio
->io_bp_copy
)
4573 zio
->io_bp_copy
= *bp
;
4576 if (zio
->io_error
!= 0) {
4577 zio
->io_pipeline
= ZIO_INTERLOCK_PIPELINE
;
4579 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4580 ASSERT(IO_IS_ALLOCATING(zio
));
4581 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4582 ASSERT(zio
->io_metaslab_class
!= NULL
);
4583 ASSERT(ZIO_HAS_ALLOCATOR(zio
));
4586 * We were unable to allocate anything, unreserve and
4587 * issue the next I/O to allocate.
4589 metaslab_class_throttle_unreserve(
4590 zio
->io_metaslab_class
, zio
->io_prop
.zp_copies
,
4591 zio
->io_allocator
, zio
);
4592 zio_allocate_dispatch(zio
->io_spa
, zio
->io_allocator
);
4596 mutex_enter(&zio
->io_lock
);
4597 zio
->io_state
[ZIO_WAIT_READY
] = 1;
4598 pio
= zio_walk_parents(zio
, &zl
);
4599 mutex_exit(&zio
->io_lock
);
4602 * As we notify zio's parents, new parents could be added.
4603 * New parents go to the head of zio's io_parent_list, however,
4604 * so we will (correctly) not notify them. The remainder of zio's
4605 * io_parent_list, from 'pio_next' onward, cannot change because
4606 * all parents must wait for us to be done before they can be done.
4608 for (; pio
!= NULL
; pio
= pio_next
) {
4609 pio_next
= zio_walk_parents(zio
, &zl
);
4610 zio_notify_parent(pio
, zio
, ZIO_WAIT_READY
, NULL
);
4613 if (zio
->io_flags
& ZIO_FLAG_NODATA
) {
4614 if (bp
!= NULL
&& BP_IS_GANG(bp
)) {
4615 zio
->io_flags
&= ~ZIO_FLAG_NODATA
;
4617 ASSERT((uintptr_t)zio
->io_abd
< SPA_MAXBLOCKSIZE
);
4618 zio
->io_pipeline
&= ~ZIO_VDEV_IO_STAGES
;
4622 if (zio_injection_enabled
&&
4623 zio
->io_spa
->spa_syncing_txg
== zio
->io_txg
)
4624 zio_handle_ignored_writes(zio
);
4630 * Update the allocation throttle accounting.
4633 zio_dva_throttle_done(zio_t
*zio
)
4635 zio_t
*lio __maybe_unused
= zio
->io_logical
;
4636 zio_t
*pio
= zio_unique_parent(zio
);
4637 vdev_t
*vd
= zio
->io_vd
;
4638 int flags
= METASLAB_ASYNC_ALLOC
;
4640 ASSERT3P(zio
->io_bp
, !=, NULL
);
4641 ASSERT3U(zio
->io_type
, ==, ZIO_TYPE_WRITE
);
4642 ASSERT3U(zio
->io_priority
, ==, ZIO_PRIORITY_ASYNC_WRITE
);
4643 ASSERT3U(zio
->io_child_type
, ==, ZIO_CHILD_VDEV
);
4645 ASSERT3P(vd
, ==, vd
->vdev_top
);
4646 ASSERT(zio_injection_enabled
|| !(zio
->io_flags
& ZIO_FLAG_IO_RETRY
));
4647 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4648 ASSERT(zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
);
4649 ASSERT(!(lio
->io_flags
& ZIO_FLAG_IO_REWRITE
));
4650 ASSERT(!(lio
->io_orig_flags
& ZIO_FLAG_NODATA
));
4653 * Parents of gang children can have two flavors -- ones that
4654 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4655 * and ones that allocated the constituent blocks. The allocation
4656 * throttle needs to know the allocating parent zio so we must find
4659 if (pio
->io_child_type
== ZIO_CHILD_GANG
) {
4661 * If our parent is a rewrite gang child then our grandparent
4662 * would have been the one that performed the allocation.
4664 if (pio
->io_flags
& ZIO_FLAG_IO_REWRITE
)
4665 pio
= zio_unique_parent(pio
);
4666 flags
|= METASLAB_GANG_CHILD
;
4669 ASSERT(IO_IS_ALLOCATING(pio
));
4670 ASSERT(ZIO_HAS_ALLOCATOR(pio
));
4671 ASSERT3P(zio
, !=, zio
->io_logical
);
4672 ASSERT(zio
->io_logical
!= NULL
);
4673 ASSERT(!(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
));
4674 ASSERT0(zio
->io_flags
& ZIO_FLAG_NOPWRITE
);
4675 ASSERT(zio
->io_metaslab_class
!= NULL
);
4677 mutex_enter(&pio
->io_lock
);
4678 metaslab_group_alloc_decrement(zio
->io_spa
, vd
->vdev_id
, pio
, flags
,
4679 pio
->io_allocator
, B_TRUE
);
4680 mutex_exit(&pio
->io_lock
);
4682 metaslab_class_throttle_unreserve(zio
->io_metaslab_class
, 1,
4683 pio
->io_allocator
, pio
);
4686 * Call into the pipeline to see if there is more work that
4687 * needs to be done. If there is work to be done it will be
4688 * dispatched to another taskq thread.
4690 zio_allocate_dispatch(zio
->io_spa
, pio
->io_allocator
);
4694 zio_done(zio_t
*zio
)
4697 * Always attempt to keep stack usage minimal here since
4698 * we can be called recursively up to 19 levels deep.
4700 const uint64_t psize
= zio
->io_size
;
4701 zio_t
*pio
, *pio_next
;
4702 zio_link_t
*zl
= NULL
;
4705 * If our children haven't all completed,
4706 * wait for them and then repeat this pipeline stage.
4708 if (zio_wait_for_children(zio
, ZIO_CHILD_ALL_BITS
, ZIO_WAIT_DONE
)) {
4713 * If the allocation throttle is enabled, then update the accounting.
4714 * We only track child I/Os that are part of an allocating async
4715 * write. We must do this since the allocation is performed
4716 * by the logical I/O but the actual write is done by child I/Os.
4718 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
&&
4719 zio
->io_child_type
== ZIO_CHILD_VDEV
) {
4720 ASSERT(zio
->io_metaslab_class
!= NULL
);
4721 ASSERT(zio
->io_metaslab_class
->mc_alloc_throttle_enabled
);
4722 zio_dva_throttle_done(zio
);
4726 * If the allocation throttle is enabled, verify that
4727 * we have decremented the refcounts for every I/O that was throttled.
4729 if (zio
->io_flags
& ZIO_FLAG_IO_ALLOCATING
) {
4730 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
4731 ASSERT(zio
->io_priority
== ZIO_PRIORITY_ASYNC_WRITE
);
4732 ASSERT(zio
->io_bp
!= NULL
);
4733 ASSERT(ZIO_HAS_ALLOCATOR(zio
));
4735 metaslab_group_alloc_verify(zio
->io_spa
, zio
->io_bp
, zio
,
4737 VERIFY(zfs_refcount_not_held(&zio
->io_metaslab_class
->
4738 mc_allocator
[zio
->io_allocator
].mca_alloc_slots
, zio
));
4742 for (int c
= 0; c
< ZIO_CHILD_TYPES
; c
++)
4743 for (int w
= 0; w
< ZIO_WAIT_TYPES
; w
++)
4744 ASSERT(zio
->io_children
[c
][w
] == 0);
4746 if (zio
->io_bp
!= NULL
&& !BP_IS_EMBEDDED(zio
->io_bp
)) {
4747 ASSERT(zio
->io_bp
->blk_pad
[0] == 0);
4748 ASSERT(zio
->io_bp
->blk_pad
[1] == 0);
4749 ASSERT(memcmp(zio
->io_bp
, &zio
->io_bp_copy
,
4750 sizeof (blkptr_t
)) == 0 ||
4751 (zio
->io_bp
== zio_unique_parent(zio
)->io_bp
));
4752 if (zio
->io_type
== ZIO_TYPE_WRITE
&& !BP_IS_HOLE(zio
->io_bp
) &&
4753 zio
->io_bp_override
== NULL
&&
4754 !(zio
->io_flags
& ZIO_FLAG_IO_REPAIR
)) {
4755 ASSERT3U(zio
->io_prop
.zp_copies
, <=,
4756 BP_GET_NDVAS(zio
->io_bp
));
4757 ASSERT(BP_COUNT_GANG(zio
->io_bp
) == 0 ||
4758 (BP_COUNT_GANG(zio
->io_bp
) ==
4759 BP_GET_NDVAS(zio
->io_bp
)));
4761 if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
)
4762 VERIFY(BP_EQUAL(zio
->io_bp
, &zio
->io_bp_orig
));
4766 * If there were child vdev/gang/ddt errors, they apply to us now.
4768 zio_inherit_child_errors(zio
, ZIO_CHILD_VDEV
);
4769 zio_inherit_child_errors(zio
, ZIO_CHILD_GANG
);
4770 zio_inherit_child_errors(zio
, ZIO_CHILD_DDT
);
4773 * If the I/O on the transformed data was successful, generate any
4774 * checksum reports now while we still have the transformed data.
4776 if (zio
->io_error
== 0) {
4777 while (zio
->io_cksum_report
!= NULL
) {
4778 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
4779 uint64_t align
= zcr
->zcr_align
;
4780 uint64_t asize
= P2ROUNDUP(psize
, align
);
4781 abd_t
*adata
= zio
->io_abd
;
4783 if (adata
!= NULL
&& asize
!= psize
) {
4784 adata
= abd_alloc(asize
, B_TRUE
);
4785 abd_copy(adata
, zio
->io_abd
, psize
);
4786 abd_zero_off(adata
, psize
, asize
- psize
);
4789 zio
->io_cksum_report
= zcr
->zcr_next
;
4790 zcr
->zcr_next
= NULL
;
4791 zcr
->zcr_finish(zcr
, adata
);
4792 zfs_ereport_free_checksum(zcr
);
4794 if (adata
!= NULL
&& asize
!= psize
)
4799 zio_pop_transforms(zio
); /* note: may set zio->io_error */
4801 vdev_stat_update(zio
, psize
);
4804 * If this I/O is attached to a particular vdev is slow, exceeding
4805 * 30 seconds to complete, post an error described the I/O delay.
4806 * We ignore these errors if the device is currently unavailable.
4808 if (zio
->io_delay
>= MSEC2NSEC(zio_slow_io_ms
)) {
4809 if (zio
->io_vd
!= NULL
&& !vdev_is_dead(zio
->io_vd
)) {
4811 * We want to only increment our slow IO counters if
4812 * the IO is valid (i.e. not if the drive is removed).
4814 * zfs_ereport_post() will also do these checks, but
4815 * it can also ratelimit and have other failures, so we
4816 * need to increment the slow_io counters independent
4819 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY
,
4820 zio
->io_spa
, zio
->io_vd
, zio
)) {
4821 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4822 zio
->io_vd
->vdev_stat
.vs_slow_ios
++;
4823 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4825 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY
,
4826 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
,
4832 if (zio
->io_error
) {
4834 * If this I/O is attached to a particular vdev,
4835 * generate an error message describing the I/O failure
4836 * at the block level. We ignore these errors if the
4837 * device is currently unavailable.
4839 if (zio
->io_error
!= ECKSUM
&& zio
->io_vd
!= NULL
&&
4840 !vdev_is_dead(zio
->io_vd
)) {
4841 int ret
= zfs_ereport_post(FM_EREPORT_ZFS_IO
,
4842 zio
->io_spa
, zio
->io_vd
, &zio
->io_bookmark
, zio
, 0);
4843 if (ret
!= EALREADY
) {
4844 mutex_enter(&zio
->io_vd
->vdev_stat_lock
);
4845 if (zio
->io_type
== ZIO_TYPE_READ
)
4846 zio
->io_vd
->vdev_stat
.vs_read_errors
++;
4847 else if (zio
->io_type
== ZIO_TYPE_WRITE
)
4848 zio
->io_vd
->vdev_stat
.vs_write_errors
++;
4849 mutex_exit(&zio
->io_vd
->vdev_stat_lock
);
4853 if ((zio
->io_error
== EIO
|| !(zio
->io_flags
&
4854 (ZIO_FLAG_SPECULATIVE
| ZIO_FLAG_DONT_PROPAGATE
))) &&
4855 zio
== zio
->io_logical
) {
4857 * For logical I/O requests, tell the SPA to log the
4858 * error and generate a logical data ereport.
4860 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
,
4861 BP_GET_LOGICAL_BIRTH(zio
->io_bp
));
4862 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA
,
4863 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
, 0);
4867 if (zio
->io_error
&& zio
== zio
->io_logical
) {
4869 * Determine whether zio should be reexecuted. This will
4870 * propagate all the way to the root via zio_notify_parent().
4872 ASSERT(zio
->io_vd
== NULL
&& zio
->io_bp
!= NULL
);
4873 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4875 if (IO_IS_ALLOCATING(zio
) &&
4876 !(zio
->io_flags
& ZIO_FLAG_CANFAIL
)) {
4877 if (zio
->io_error
!= ENOSPC
)
4878 zio
->io_reexecute
|= ZIO_REEXECUTE_NOW
;
4880 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4883 if ((zio
->io_type
== ZIO_TYPE_READ
||
4884 zio
->io_type
== ZIO_TYPE_FREE
) &&
4885 !(zio
->io_flags
& ZIO_FLAG_SCAN_THREAD
) &&
4886 zio
->io_error
== ENXIO
&&
4887 spa_load_state(zio
->io_spa
) == SPA_LOAD_NONE
&&
4888 spa_get_failmode(zio
->io_spa
) != ZIO_FAILURE_MODE_CONTINUE
)
4889 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4891 if (!(zio
->io_flags
& ZIO_FLAG_CANFAIL
) && !zio
->io_reexecute
)
4892 zio
->io_reexecute
|= ZIO_REEXECUTE_SUSPEND
;
4895 * Here is a possibly good place to attempt to do
4896 * either combinatorial reconstruction or error correction
4897 * based on checksums. It also might be a good place
4898 * to send out preliminary ereports before we suspend
4904 * If there were logical child errors, they apply to us now.
4905 * We defer this until now to avoid conflating logical child
4906 * errors with errors that happened to the zio itself when
4907 * updating vdev stats and reporting FMA events above.
4909 zio_inherit_child_errors(zio
, ZIO_CHILD_LOGICAL
);
4911 if ((zio
->io_error
|| zio
->io_reexecute
) &&
4912 IO_IS_ALLOCATING(zio
) && zio
->io_gang_leader
== zio
&&
4913 !(zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)))
4914 zio_dva_unallocate(zio
, zio
->io_gang_tree
, zio
->io_bp
);
4916 zio_gang_tree_free(&zio
->io_gang_tree
);
4919 * Godfather I/Os should never suspend.
4921 if ((zio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4922 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
))
4923 zio
->io_reexecute
&= ~ZIO_REEXECUTE_SUSPEND
;
4925 if (zio
->io_reexecute
) {
4927 * This is a logical I/O that wants to reexecute.
4929 * Reexecute is top-down. When an i/o fails, if it's not
4930 * the root, it simply notifies its parent and sticks around.
4931 * The parent, seeing that it still has children in zio_done(),
4932 * does the same. This percolates all the way up to the root.
4933 * The root i/o will reexecute or suspend the entire tree.
4935 * This approach ensures that zio_reexecute() honors
4936 * all the original i/o dependency relationships, e.g.
4937 * parents not executing until children are ready.
4939 ASSERT(zio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4941 zio
->io_gang_leader
= NULL
;
4943 mutex_enter(&zio
->io_lock
);
4944 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
4945 mutex_exit(&zio
->io_lock
);
4948 * "The Godfather" I/O monitors its children but is
4949 * not a true parent to them. It will track them through
4950 * the pipeline but severs its ties whenever they get into
4951 * trouble (e.g. suspended). This allows "The Godfather"
4952 * I/O to return status without blocking.
4955 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
;
4957 zio_link_t
*remove_zl
= zl
;
4958 pio_next
= zio_walk_parents(zio
, &zl
);
4960 if ((pio
->io_flags
& ZIO_FLAG_GODFATHER
) &&
4961 (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
)) {
4962 zio_remove_child(pio
, zio
, remove_zl
);
4964 * This is a rare code path, so we don't
4965 * bother with "next_to_execute".
4967 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
,
4972 if ((pio
= zio_unique_parent(zio
)) != NULL
) {
4974 * We're not a root i/o, so there's nothing to do
4975 * but notify our parent. Don't propagate errors
4976 * upward since we haven't permanently failed yet.
4978 ASSERT(!(zio
->io_flags
& ZIO_FLAG_GODFATHER
));
4979 zio
->io_flags
|= ZIO_FLAG_DONT_PROPAGATE
;
4981 * This is a rare code path, so we don't bother with
4982 * "next_to_execute".
4984 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, NULL
);
4985 } else if (zio
->io_reexecute
& ZIO_REEXECUTE_SUSPEND
) {
4987 * We'd fail again if we reexecuted now, so suspend
4988 * until conditions improve (e.g. device comes online).
4990 zio_suspend(zio
->io_spa
, zio
, ZIO_SUSPEND_IOERR
);
4993 * Reexecution is potentially a huge amount of work.
4994 * Hand it off to the otherwise-unused claim taskq.
4996 ASSERT(taskq_empty_ent(&zio
->io_tqent
));
4997 spa_taskq_dispatch_ent(zio
->io_spa
,
4998 ZIO_TYPE_CLAIM
, ZIO_TASKQ_ISSUE
,
4999 zio_reexecute
, zio
, 0, &zio
->io_tqent
, NULL
);
5004 ASSERT(list_is_empty(&zio
->io_child_list
));
5005 ASSERT(zio
->io_reexecute
== 0);
5006 ASSERT(zio
->io_error
== 0 || (zio
->io_flags
& ZIO_FLAG_CANFAIL
));
5009 * Report any checksum errors, since the I/O is complete.
5011 while (zio
->io_cksum_report
!= NULL
) {
5012 zio_cksum_report_t
*zcr
= zio
->io_cksum_report
;
5013 zio
->io_cksum_report
= zcr
->zcr_next
;
5014 zcr
->zcr_next
= NULL
;
5015 zcr
->zcr_finish(zcr
, NULL
);
5016 zfs_ereport_free_checksum(zcr
);
5020 * It is the responsibility of the done callback to ensure that this
5021 * particular zio is no longer discoverable for adoption, and as
5022 * such, cannot acquire any new parents.
5027 mutex_enter(&zio
->io_lock
);
5028 zio
->io_state
[ZIO_WAIT_DONE
] = 1;
5029 mutex_exit(&zio
->io_lock
);
5032 * We are done executing this zio. We may want to execute a parent
5033 * next. See the comment in zio_notify_parent().
5035 zio_t
*next_to_execute
= NULL
;
5037 for (pio
= zio_walk_parents(zio
, &zl
); pio
!= NULL
; pio
= pio_next
) {
5038 zio_link_t
*remove_zl
= zl
;
5039 pio_next
= zio_walk_parents(zio
, &zl
);
5040 zio_remove_child(pio
, zio
, remove_zl
);
5041 zio_notify_parent(pio
, zio
, ZIO_WAIT_DONE
, &next_to_execute
);
5044 if (zio
->io_waiter
!= NULL
) {
5045 mutex_enter(&zio
->io_lock
);
5046 zio
->io_executor
= NULL
;
5047 cv_broadcast(&zio
->io_cv
);
5048 mutex_exit(&zio
->io_lock
);
5053 return (next_to_execute
);
5057 * ==========================================================================
5058 * I/O pipeline definition
5059 * ==========================================================================
5061 static zio_pipe_stage_t
*zio_pipeline
[] = {
5069 zio_checksum_generate
,
5086 zio_checksum_verify
,
5094 * Compare two zbookmark_phys_t's to see which we would reach first in a
5095 * pre-order traversal of the object tree.
5097 * This is simple in every case aside from the meta-dnode object. For all other
5098 * objects, we traverse them in order (object 1 before object 2, and so on).
5099 * However, all of these objects are traversed while traversing object 0, since
5100 * the data it points to is the list of objects. Thus, we need to convert to a
5101 * canonical representation so we can compare meta-dnode bookmarks to
5102 * non-meta-dnode bookmarks.
5104 * We do this by calculating "equivalents" for each field of the zbookmark.
5105 * zbookmarks outside of the meta-dnode use their own object and level, and
5106 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
5107 * blocks this bookmark refers to) by multiplying their blkid by their span
5108 * (the number of L0 blocks contained within one block at their level).
5109 * zbookmarks inside the meta-dnode calculate their object equivalent
5110 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
5111 * level + 1<<31 (any value larger than a level could ever be) for their level.
5112 * This causes them to always compare before a bookmark in their object
5113 * equivalent, compare appropriately to bookmarks in other objects, and to
5114 * compare appropriately to other bookmarks in the meta-dnode.
5117 zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
, uint8_t ibs2
,
5118 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
)
5121 * These variables represent the "equivalent" values for the zbookmark,
5122 * after converting zbookmarks inside the meta dnode to their
5123 * normal-object equivalents.
5125 uint64_t zb1obj
, zb2obj
;
5126 uint64_t zb1L0
, zb2L0
;
5127 uint64_t zb1level
, zb2level
;
5129 if (zb1
->zb_object
== zb2
->zb_object
&&
5130 zb1
->zb_level
== zb2
->zb_level
&&
5131 zb1
->zb_blkid
== zb2
->zb_blkid
)
5134 IMPLY(zb1
->zb_level
> 0, ibs1
>= SPA_MINBLOCKSHIFT
);
5135 IMPLY(zb2
->zb_level
> 0, ibs2
>= SPA_MINBLOCKSHIFT
);
5138 * BP_SPANB calculates the span in blocks.
5140 zb1L0
= (zb1
->zb_blkid
) * BP_SPANB(ibs1
, zb1
->zb_level
);
5141 zb2L0
= (zb2
->zb_blkid
) * BP_SPANB(ibs2
, zb2
->zb_level
);
5143 if (zb1
->zb_object
== DMU_META_DNODE_OBJECT
) {
5144 zb1obj
= zb1L0
* (dbss1
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
5146 zb1level
= zb1
->zb_level
+ COMPARE_META_LEVEL
;
5148 zb1obj
= zb1
->zb_object
;
5149 zb1level
= zb1
->zb_level
;
5152 if (zb2
->zb_object
== DMU_META_DNODE_OBJECT
) {
5153 zb2obj
= zb2L0
* (dbss2
<< (SPA_MINBLOCKSHIFT
- DNODE_SHIFT
));
5155 zb2level
= zb2
->zb_level
+ COMPARE_META_LEVEL
;
5157 zb2obj
= zb2
->zb_object
;
5158 zb2level
= zb2
->zb_level
;
5161 /* Now that we have a canonical representation, do the comparison. */
5162 if (zb1obj
!= zb2obj
)
5163 return (zb1obj
< zb2obj
? -1 : 1);
5164 else if (zb1L0
!= zb2L0
)
5165 return (zb1L0
< zb2L0
? -1 : 1);
5166 else if (zb1level
!= zb2level
)
5167 return (zb1level
> zb2level
? -1 : 1);
5169 * This can (theoretically) happen if the bookmarks have the same object
5170 * and level, but different blkids, if the block sizes are not the same.
5171 * There is presently no way to change the indirect block sizes
5177 * This function checks the following: given that last_block is the place that
5178 * our traversal stopped last time, does that guarantee that we've visited
5179 * every node under subtree_root? Therefore, we can't just use the raw output
5180 * of zbookmark_compare. We have to pass in a modified version of
5181 * subtree_root; by incrementing the block id, and then checking whether
5182 * last_block is before or equal to that, we can tell whether or not having
5183 * visited last_block implies that all of subtree_root's children have been
5187 zbookmark_subtree_completed(const dnode_phys_t
*dnp
,
5188 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
5190 zbookmark_phys_t mod_zb
= *subtree_root
;
5192 ASSERT0(last_block
->zb_level
);
5194 /* The objset_phys_t isn't before anything. */
5199 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5200 * data block size in sectors, because that variable is only used if
5201 * the bookmark refers to a block in the meta-dnode. Since we don't
5202 * know without examining it what object it refers to, and there's no
5203 * harm in passing in this value in other cases, we always pass it in.
5205 * We pass in 0 for the indirect block size shift because zb2 must be
5206 * level 0. The indirect block size is only used to calculate the span
5207 * of the bookmark, but since the bookmark must be level 0, the span is
5208 * always 1, so the math works out.
5210 * If you make changes to how the zbookmark_compare code works, be sure
5211 * to make sure that this code still works afterwards.
5213 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
5214 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, &mod_zb
,
5219 * This function is similar to zbookmark_subtree_completed(), but returns true
5220 * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5223 zbookmark_subtree_tbd(const dnode_phys_t
*dnp
,
5224 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
)
5226 ASSERT0(last_block
->zb_level
);
5229 return (zbookmark_compare(dnp
->dn_datablkszsec
, dnp
->dn_indblkshift
,
5230 1ULL << (DNODE_BLOCK_SHIFT
- SPA_MINBLOCKSHIFT
), 0, subtree_root
,
5234 EXPORT_SYMBOL(zio_type_name
);
5235 EXPORT_SYMBOL(zio_buf_alloc
);
5236 EXPORT_SYMBOL(zio_data_buf_alloc
);
5237 EXPORT_SYMBOL(zio_buf_free
);
5238 EXPORT_SYMBOL(zio_data_buf_free
);
5240 ZFS_MODULE_PARAM(zfs_zio
, zio_
, slow_io_ms
, INT
, ZMOD_RW
,
5241 "Max I/O completion time (milliseconds) before marking it as slow");
5243 ZFS_MODULE_PARAM(zfs_zio
, zio_
, requeue_io_start_cut_in_line
, INT
, ZMOD_RW
,
5244 "Prioritize requeued I/O");
5246 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_deferred_free
, UINT
, ZMOD_RW
,
5247 "Defer frees starting in this pass");
5249 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_dont_compress
, UINT
, ZMOD_RW
,
5250 "Don't compress starting in this pass");
5252 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_pass_rewrite
, UINT
, ZMOD_RW
,
5253 "Rewrite new bps starting in this pass");
5255 ZFS_MODULE_PARAM(zfs_zio
, zio_
, dva_throttle_enabled
, INT
, ZMOD_RW
,
5256 "Throttle block allocations in the ZIO pipeline");
5258 ZFS_MODULE_PARAM(zfs_zio
, zio_
, deadman_log_all
, INT
, ZMOD_RW
,
5259 "Log all slow ZIOs, not just those with vdevs");