4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
26 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
32 #include <sys/zio_priority.h>
33 #include <sys/zfs_context.h>
37 #include <sys/fs/zfs.h>
38 #include <sys/zio_impl.h>
47 #define ZEC_MAGIC 0x210da7ab10c7a11ULL
49 typedef struct zio_eck
{
50 uint64_t zec_magic
; /* for validation, endianness */
51 zio_cksum_t zec_cksum
; /* 256-bit checksum */
55 * Gang block headers are self-checksumming and contain an array
58 #define SPA_GANGBLOCKSIZE SPA_MINBLOCKSIZE
59 #define SPA_GBH_NBLKPTRS ((SPA_GANGBLOCKSIZE - \
60 sizeof (zio_eck_t)) / sizeof (blkptr_t))
61 #define SPA_GBH_FILLER ((SPA_GANGBLOCKSIZE - \
62 sizeof (zio_eck_t) - \
63 (SPA_GBH_NBLKPTRS * sizeof (blkptr_t))) /\
66 typedef struct zio_gbh
{
67 blkptr_t zg_blkptr
[SPA_GBH_NBLKPTRS
];
68 uint64_t zg_filler
[SPA_GBH_FILLER
];
73 ZIO_CHECKSUM_INHERIT
= 0,
77 ZIO_CHECKSUM_GANG_HEADER
,
79 ZIO_CHECKSUM_FLETCHER_2
,
80 ZIO_CHECKSUM_FLETCHER_4
,
83 ZIO_CHECKSUM_NOPARITY
,
87 ZIO_CHECKSUM_FUNCTIONS
91 * The number of "legacy" compression functions which can be set on individual
94 #define ZIO_CHECKSUM_LEGACY_FUNCTIONS ZIO_CHECKSUM_ZILOG2
96 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
97 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
99 #define ZIO_CHECKSUM_MASK 0xffULL
100 #define ZIO_CHECKSUM_VERIFY (1 << 8)
102 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
103 #define ZIO_DEDUPDITTO_MIN 100
106 * The number of "legacy" compression functions which can be set on individual
109 #define ZIO_COMPRESS_LEGACY_FUNCTIONS ZIO_COMPRESS_LZ4
112 * The meaning of "compress = on" selected by the compression features enabled
115 #define ZIO_COMPRESS_LEGACY_ON_VALUE ZIO_COMPRESS_LZJB
116 #define ZIO_COMPRESS_LZ4_ON_VALUE ZIO_COMPRESS_LZ4
118 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
120 #define BOOTFS_COMPRESS_VALID(compress) \
121 ((compress) == ZIO_COMPRESS_LZJB || \
122 (compress) == ZIO_COMPRESS_LZ4 || \
123 (compress) == ZIO_COMPRESS_ON || \
124 (compress) == ZIO_COMPRESS_OFF)
127 * Default Linux timeout for a sd device.
129 #define ZIO_DELAY_MAX (30 * MILLISEC)
131 #define ZIO_FAILURE_MODE_WAIT 0
132 #define ZIO_FAILURE_MODE_CONTINUE 1
133 #define ZIO_FAILURE_MODE_PANIC 2
137 * Flags inherited by gang, ddt, and vdev children,
138 * and that must be equal for two zios to aggregate
140 ZIO_FLAG_DONT_AGGREGATE
= 1 << 0,
141 ZIO_FLAG_IO_REPAIR
= 1 << 1,
142 ZIO_FLAG_SELF_HEAL
= 1 << 2,
143 ZIO_FLAG_RESILVER
= 1 << 3,
144 ZIO_FLAG_SCRUB
= 1 << 4,
145 ZIO_FLAG_SCAN_THREAD
= 1 << 5,
146 ZIO_FLAG_PHYSICAL
= 1 << 6,
148 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
151 * Flags inherited by ddt, gang, and vdev children.
153 ZIO_FLAG_CANFAIL
= 1 << 7, /* must be first for INHERIT */
154 ZIO_FLAG_SPECULATIVE
= 1 << 8,
155 ZIO_FLAG_CONFIG_WRITER
= 1 << 9,
156 ZIO_FLAG_DONT_RETRY
= 1 << 10,
157 ZIO_FLAG_DONT_CACHE
= 1 << 11,
158 ZIO_FLAG_NODATA
= 1 << 12,
159 ZIO_FLAG_INDUCE_DAMAGE
= 1 << 13,
160 ZIO_FLAG_IO_ALLOCATING
= 1 << 14,
162 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
163 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
166 * Flags inherited by vdev children.
168 ZIO_FLAG_IO_RETRY
= 1 << 15, /* must be first for INHERIT */
169 ZIO_FLAG_PROBE
= 1 << 16,
170 ZIO_FLAG_TRYHARD
= 1 << 17,
171 ZIO_FLAG_OPTIONAL
= 1 << 18,
173 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
176 * Flags not inherited by any children.
178 ZIO_FLAG_DONT_QUEUE
= 1 << 19, /* must be first for INHERIT */
179 ZIO_FLAG_DONT_PROPAGATE
= 1 << 20,
180 ZIO_FLAG_IO_BYPASS
= 1 << 21,
181 ZIO_FLAG_IO_REWRITE
= 1 << 22,
182 ZIO_FLAG_RAW
= 1 << 23,
183 ZIO_FLAG_GANG_CHILD
= 1 << 24,
184 ZIO_FLAG_DDT_CHILD
= 1 << 25,
185 ZIO_FLAG_GODFATHER
= 1 << 26,
186 ZIO_FLAG_NOPWRITE
= 1 << 27,
187 ZIO_FLAG_REEXECUTED
= 1 << 28,
188 ZIO_FLAG_DELEGATED
= 1 << 29,
189 ZIO_FLAG_FASTWRITE
= 1 << 30
192 #define ZIO_FLAG_MUSTSUCCEED 0
194 #define ZIO_DDT_CHILD_FLAGS(zio) \
195 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
196 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
198 #define ZIO_GANG_CHILD_FLAGS(zio) \
199 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
200 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
202 #define ZIO_VDEV_CHILD_FLAGS(zio) \
203 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
221 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
222 * graveyard) to indicate checksum errors and fragmentation.
227 typedef void zio_done_func_t(zio_t
*zio
);
229 extern int zio_dva_throttle_enabled
;
230 extern const char *zio_type_name
[ZIO_TYPES
];
233 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
234 * identifies any block in the pool. By convention, the meta-objset (MOS)
235 * is objset 0, and the meta-dnode is object 0. This covers all blocks
236 * except root blocks and ZIL blocks, which are defined as follows:
238 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
239 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
240 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
241 * dnode visit bookmarks are <objset, object id of dnode, -3, 0>.
243 * Note: this structure is called a bookmark because its original purpose
244 * was to remember where to resume a pool-wide traverse.
246 * Note: this structure is passed between userland and the kernel, and is
247 * stored on disk (by virtue of being incorporated into other on-disk
248 * structures, e.g. dsl_scan_phys_t).
250 struct zbookmark_phys
{
257 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
259 (zb)->zb_objset = objset; \
260 (zb)->zb_object = object; \
261 (zb)->zb_level = level; \
262 (zb)->zb_blkid = blkid; \
265 #define ZB_DESTROYED_OBJSET (-1ULL)
267 #define ZB_ROOT_OBJECT (0ULL)
268 #define ZB_ROOT_LEVEL (-1LL)
269 #define ZB_ROOT_BLKID (0ULL)
271 #define ZB_ZIL_OBJECT (0ULL)
272 #define ZB_ZIL_LEVEL (-2LL)
274 #define ZB_DNODE_LEVEL (-3LL)
275 #define ZB_DNODE_BLKID (0ULL)
277 #define ZB_IS_ZERO(zb) \
278 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
279 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
280 #define ZB_IS_ROOT(zb) \
281 ((zb)->zb_object == ZB_ROOT_OBJECT && \
282 (zb)->zb_level == ZB_ROOT_LEVEL && \
283 (zb)->zb_blkid == ZB_ROOT_BLKID)
285 typedef struct zio_prop
{
286 enum zio_checksum zp_checksum
;
287 enum zio_compress zp_compress
;
288 dmu_object_type_t zp_type
;
292 boolean_t zp_dedup_verify
;
293 boolean_t zp_nopwrite
;
296 typedef struct zio_cksum_report zio_cksum_report_t
;
298 typedef void zio_cksum_finish_f(zio_cksum_report_t
*rep
,
299 const void *good_data
);
300 typedef void zio_cksum_free_f(void *cbdata
, size_t size
);
302 struct zio_bad_cksum
; /* defined in zio_checksum.h */
305 struct zio_cksum_report
{
306 struct zio_cksum_report
*zcr_next
;
307 nvlist_t
*zcr_ereport
;
308 nvlist_t
*zcr_detector
;
310 size_t zcr_cbinfo
; /* passed to zcr_free() */
313 zio_cksum_finish_f
*zcr_finish
;
314 zio_cksum_free_f
*zcr_free
;
316 /* internal use only */
317 struct zio_bad_cksum
*zcr_ckinfo
; /* information from failure */
320 typedef void zio_vsd_cksum_report_f(zio_t
*zio
, zio_cksum_report_t
*zcr
,
323 zio_vsd_cksum_report_f zio_vsd_default_cksum_report
;
325 typedef struct zio_vsd_ops
{
326 zio_done_func_t
*vsd_free
;
327 zio_vsd_cksum_report_f
*vsd_cksum_report
;
330 typedef struct zio_gang_node
{
331 zio_gbh_phys_t
*gn_gbh
;
332 struct zio_gang_node
*gn_child
[SPA_GBH_NBLKPTRS
];
335 typedef zio_t
*zio_gang_issue_func_t(zio_t
*zio
, blkptr_t
*bp
,
336 zio_gang_node_t
*gn
, void *data
);
338 typedef void zio_transform_func_t(zio_t
*zio
, void *data
, uint64_t size
);
340 typedef struct zio_transform
{
342 uint64_t zt_orig_size
;
344 zio_transform_func_t
*zt_transform
;
345 struct zio_transform
*zt_next
;
348 typedef int zio_pipe_stage_t(zio_t
*zio
);
351 * The io_reexecute flags are distinct from io_flags because the child must
352 * be able to propagate them to the parent. The normal io_flags are local
353 * to the zio, not protected by any lock, and not modifiable by children;
354 * the reexecute flags are protected by io_lock, modifiable by children,
355 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
357 #define ZIO_REEXECUTE_NOW 0x01
358 #define ZIO_REEXECUTE_SUSPEND 0x02
360 typedef struct zio_link
{
363 list_node_t zl_parent_node
;
364 list_node_t zl_child_node
;
368 /* Core information about this I/O */
369 zbookmark_phys_t io_bookmark
;
372 enum zio_child io_child_type
;
374 zio_priority_t io_priority
;
375 uint8_t io_reexecute
;
376 uint8_t io_state
[ZIO_WAIT_TYPES
];
380 blkptr_t
*io_bp_override
;
382 list_t io_parent_list
;
383 list_t io_child_list
;
385 zio_transform_t
*io_transform_stack
;
388 zio_done_func_t
*io_ready
;
389 zio_done_func_t
*io_children_ready
;
390 zio_done_func_t
*io_physdone
;
391 zio_done_func_t
*io_done
;
393 int64_t io_prev_space_delta
; /* DMU private */
395 /* io_lsize != io_orig_size iff this is a raw write */
398 /* Data represented by this I/O */
402 uint64_t io_orig_size
;
404 /* Stuff for the vdev stack */
407 const zio_vsd_ops_t
*io_vsd_ops
;
410 hrtime_t io_timestamp
; /* submitted at */
411 hrtime_t io_queued_timestamp
;
412 hrtime_t io_target_timestamp
;
413 hrtime_t io_delta
; /* vdev queue service delta */
414 hrtime_t io_delay
; /* Device access time (disk or */
416 avl_node_t io_queue_node
;
417 avl_node_t io_offset_node
;
418 avl_node_t io_alloc_node
;
420 /* Internal pipeline state */
421 enum zio_flag io_flags
;
422 enum zio_stage io_stage
;
423 enum zio_stage io_pipeline
;
424 enum zio_flag io_orig_flags
;
425 enum zio_stage io_orig_stage
;
426 enum zio_stage io_orig_pipeline
;
427 enum zio_stage io_pipeline_trace
;
429 int io_child_error
[ZIO_CHILD_TYPES
];
430 uint64_t io_children
[ZIO_CHILD_TYPES
][ZIO_WAIT_TYPES
];
431 uint64_t io_child_count
;
432 uint64_t io_phys_children
;
433 uint64_t io_parent_count
;
435 zio_t
*io_gang_leader
;
436 zio_gang_node_t
*io_gang_tree
;
443 zio_cksum_report_t
*io_cksum_report
;
446 /* Taskq dispatching state */
447 taskq_ent_t io_tqent
;
450 extern int zio_timestamp_compare(const void *, const void *);
452 extern zio_t
*zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
,
453 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
455 extern zio_t
*zio_root(spa_t
*spa
,
456 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
458 extern zio_t
*zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, void *data
,
459 uint64_t lsize
, zio_done_func_t
*done
, void *private,
460 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
);
462 extern zio_t
*zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
463 void *data
, uint64_t size
, uint64_t psize
, const zio_prop_t
*zp
,
464 zio_done_func_t
*ready
, zio_done_func_t
*children_ready
,
465 zio_done_func_t
*physdone
, zio_done_func_t
*done
,
466 void *private, zio_priority_t priority
, enum zio_flag flags
,
467 const zbookmark_phys_t
*zb
);
469 extern zio_t
*zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
470 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
471 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
);
473 extern void zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
,
476 extern void zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
);
478 extern zio_t
*zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
480 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
482 extern zio_t
*zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
483 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
485 extern zio_t
*zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
486 uint64_t size
, void *data
, int checksum
,
487 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
488 enum zio_flag flags
, boolean_t labels
);
490 extern zio_t
*zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
491 uint64_t size
, void *data
, int checksum
,
492 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
493 enum zio_flag flags
, boolean_t labels
);
495 extern zio_t
*zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
496 const blkptr_t
*bp
, enum zio_flag flags
);
498 extern int zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
,
499 uint64_t size
, boolean_t use_slog
);
500 extern void zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
);
501 extern void zio_flush(zio_t
*zio
, vdev_t
*vd
);
502 extern void zio_shrink(zio_t
*zio
, uint64_t size
);
504 extern int zio_wait(zio_t
*zio
);
505 extern void zio_nowait(zio_t
*zio
);
506 extern void zio_execute(zio_t
*zio
);
507 extern void zio_interrupt(zio_t
*zio
);
508 extern void zio_delay_init(zio_t
*zio
);
509 extern void zio_delay_interrupt(zio_t
*zio
);
511 extern zio_t
*zio_walk_parents(zio_t
*cio
, zio_link_t
**);
512 extern zio_t
*zio_walk_children(zio_t
*pio
, zio_link_t
**);
513 extern zio_t
*zio_unique_parent(zio_t
*cio
);
514 extern void zio_add_child(zio_t
*pio
, zio_t
*cio
);
516 extern void *zio_buf_alloc(size_t size
);
517 extern void zio_buf_free(void *buf
, size_t size
);
518 extern void *zio_data_buf_alloc(size_t size
);
519 extern void zio_data_buf_free(void *buf
, size_t size
);
520 extern void *zio_buf_alloc_flags(size_t size
, int flags
);
522 extern void zio_push_transform(zio_t
*zio
, void *data
, uint64_t size
,
523 uint64_t bufsize
, zio_transform_func_t
*transform
);
524 extern void zio_pop_transforms(zio_t
*zio
);
526 extern void zio_resubmit_stage_async(void *);
528 extern zio_t
*zio_vdev_child_io(zio_t
*zio
, blkptr_t
*bp
, vdev_t
*vd
,
529 uint64_t offset
, void *data
, uint64_t size
, int type
,
530 zio_priority_t priority
, enum zio_flag flags
,
531 zio_done_func_t
*done
, void *private);
533 extern zio_t
*zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
,
534 void *data
, uint64_t size
, int type
, zio_priority_t priority
,
535 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
537 extern void zio_vdev_io_bypass(zio_t
*zio
);
538 extern void zio_vdev_io_reissue(zio_t
*zio
);
539 extern void zio_vdev_io_redone(zio_t
*zio
);
541 extern void zio_checksum_verified(zio_t
*zio
);
542 extern int zio_worst_error(int e1
, int e2
);
544 extern enum zio_checksum
zio_checksum_select(enum zio_checksum child
,
545 enum zio_checksum parent
);
546 extern enum zio_checksum
zio_checksum_dedup_select(spa_t
*spa
,
547 enum zio_checksum child
, enum zio_checksum parent
);
548 extern enum zio_compress
zio_compress_select(spa_t
*spa
,
549 enum zio_compress child
, enum zio_compress parent
);
551 extern void zio_suspend(spa_t
*spa
, zio_t
*zio
);
552 extern int zio_resume(spa_t
*spa
);
553 extern void zio_resume_wait(spa_t
*spa
);
556 * Initial setup and teardown.
558 extern void zio_init(void);
559 extern void zio_fini(void);
564 struct zinject_record
;
565 extern uint32_t zio_injection_enabled
;
566 extern int zio_inject_fault(char *name
, int flags
, int *id
,
567 struct zinject_record
*record
);
568 extern int zio_inject_list_next(int *id
, char *name
, size_t buflen
,
569 struct zinject_record
*record
);
570 extern int zio_clear_fault(int id
);
571 extern void zio_handle_panic_injection(spa_t
*spa
, char *tag
, uint64_t type
);
572 extern int zio_handle_fault_injection(zio_t
*zio
, int error
);
573 extern int zio_handle_device_injection(vdev_t
*vd
, zio_t
*zio
, int error
);
574 extern int zio_handle_label_injection(zio_t
*zio
, int error
);
575 extern void zio_handle_ignored_writes(zio_t
*zio
);
576 extern hrtime_t
zio_handle_io_delay(zio_t
*zio
);
579 * Checksum ereport functions
581 extern void zfs_ereport_start_checksum(spa_t
*spa
, vdev_t
*vd
, struct zio
*zio
,
582 uint64_t offset
, uint64_t length
, void *arg
, struct zio_bad_cksum
*info
);
583 extern void zfs_ereport_finish_checksum(zio_cksum_report_t
*report
,
584 const void *good_data
, const void *bad_data
, boolean_t drop_if_identical
);
586 extern void zfs_ereport_free_checksum(zio_cksum_report_t
*report
);
588 /* If we have the good data in hand, this function can be used */
589 extern void zfs_ereport_post_checksum(spa_t
*spa
, vdev_t
*vd
,
590 struct zio
*zio
, uint64_t offset
, uint64_t length
,
591 const void *good_data
, const void *bad_data
, struct zio_bad_cksum
*info
);
593 /* Called from spa_sync(), but primarily an injection handler */
594 extern void spa_handle_ignored_writes(spa_t
*spa
);
596 /* zbookmark_phys functions */
597 boolean_t
zbookmark_subtree_completed(const struct dnode_phys
*dnp
,
598 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
);
599 int zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
,
600 uint8_t ibs2
, const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
);