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) 2013 by Delphix. All rights reserved.
26 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
32 #include <sys/zfs_context.h>
36 #include <sys/fs/zfs.h>
37 #include <sys/zio_impl.h>
46 #define ZEC_MAGIC 0x210da7ab10c7a11ULL
48 typedef struct zio_eck
{
49 uint64_t zec_magic
; /* for validation, endianness */
50 zio_cksum_t zec_cksum
; /* 256-bit checksum */
54 * Gang block headers are self-checksumming and contain an array
57 #define SPA_GANGBLOCKSIZE SPA_MINBLOCKSIZE
58 #define SPA_GBH_NBLKPTRS ((SPA_GANGBLOCKSIZE - \
59 sizeof (zio_eck_t)) / sizeof (blkptr_t))
60 #define SPA_GBH_FILLER ((SPA_GANGBLOCKSIZE - \
61 sizeof (zio_eck_t) - \
62 (SPA_GBH_NBLKPTRS * sizeof (blkptr_t))) /\
65 typedef struct zio_gbh
{
66 blkptr_t zg_blkptr
[SPA_GBH_NBLKPTRS
];
67 uint64_t zg_filler
[SPA_GBH_FILLER
];
72 ZIO_CHECKSUM_INHERIT
= 0,
76 ZIO_CHECKSUM_GANG_HEADER
,
78 ZIO_CHECKSUM_FLETCHER_2
,
79 ZIO_CHECKSUM_FLETCHER_4
,
82 ZIO_CHECKSUM_FUNCTIONS
85 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
86 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
88 #define ZIO_CHECKSUM_MASK 0xffULL
89 #define ZIO_CHECKSUM_VERIFY (1 << 8)
91 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
92 #define ZIO_DEDUPDITTO_MIN 100
95 ZIO_COMPRESS_INHERIT
= 0,
111 ZIO_COMPRESS_FUNCTIONS
114 #define ZIO_COMPRESS_ON_VALUE ZIO_COMPRESS_LZJB
115 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
117 #define BOOTFS_COMPRESS_VALID(compress) \
118 ((compress) == ZIO_COMPRESS_LZJB || \
119 (compress) == ZIO_COMPRESS_LZ4 || \
120 ((compress) == ZIO_COMPRESS_ON && \
121 ZIO_COMPRESS_ON_VALUE == ZIO_COMPRESS_LZJB) || \
122 (compress) == ZIO_COMPRESS_OFF)
125 * Default Linux timeout for a sd device.
127 #define ZIO_DELAY_MAX (30 * MILLISEC)
129 #define ZIO_FAILURE_MODE_WAIT 0
130 #define ZIO_FAILURE_MODE_CONTINUE 1
131 #define ZIO_FAILURE_MODE_PANIC 2
133 typedef enum zio_priority
{
134 ZIO_PRIORITY_SYNC_READ
,
135 ZIO_PRIORITY_SYNC_WRITE
, /* ZIL */
136 ZIO_PRIORITY_ASYNC_READ
, /* prefetch */
137 ZIO_PRIORITY_ASYNC_WRITE
, /* spa_sync() */
138 ZIO_PRIORITY_SCRUB
, /* asynchronous scrub/resilver reads */
139 ZIO_PRIORITY_NUM_QUEUEABLE
,
141 ZIO_PRIORITY_NOW
/* non-queued i/os (e.g. free) */
144 #define ZIO_PIPELINE_CONTINUE 0x100
145 #define ZIO_PIPELINE_STOP 0x101
149 * Flags inherited by gang, ddt, and vdev children,
150 * and that must be equal for two zios to aggregate
152 ZIO_FLAG_DONT_AGGREGATE
= 1 << 0,
153 ZIO_FLAG_IO_REPAIR
= 1 << 1,
154 ZIO_FLAG_SELF_HEAL
= 1 << 2,
155 ZIO_FLAG_RESILVER
= 1 << 3,
156 ZIO_FLAG_SCRUB
= 1 << 4,
157 ZIO_FLAG_SCAN_THREAD
= 1 << 5,
159 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
162 * Flags inherited by ddt, gang, and vdev children.
164 ZIO_FLAG_CANFAIL
= 1 << 6, /* must be first for INHERIT */
165 ZIO_FLAG_SPECULATIVE
= 1 << 7,
166 ZIO_FLAG_CONFIG_WRITER
= 1 << 8,
167 ZIO_FLAG_DONT_RETRY
= 1 << 9,
168 ZIO_FLAG_DONT_CACHE
= 1 << 10,
169 ZIO_FLAG_NODATA
= 1 << 11,
170 ZIO_FLAG_INDUCE_DAMAGE
= 1 << 12,
172 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
173 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
176 * Flags inherited by vdev children.
178 ZIO_FLAG_IO_RETRY
= 1 << 13, /* must be first for INHERIT */
179 ZIO_FLAG_PROBE
= 1 << 14,
180 ZIO_FLAG_TRYHARD
= 1 << 15,
181 ZIO_FLAG_OPTIONAL
= 1 << 16,
183 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
186 * Flags not inherited by any children.
188 ZIO_FLAG_DONT_QUEUE
= 1 << 17, /* must be first for INHERIT */
189 ZIO_FLAG_DONT_PROPAGATE
= 1 << 18,
190 ZIO_FLAG_IO_BYPASS
= 1 << 19,
191 ZIO_FLAG_IO_REWRITE
= 1 << 20,
192 ZIO_FLAG_RAW
= 1 << 21,
193 ZIO_FLAG_GANG_CHILD
= 1 << 22,
194 ZIO_FLAG_DDT_CHILD
= 1 << 23,
195 ZIO_FLAG_GODFATHER
= 1 << 24,
196 ZIO_FLAG_NOPWRITE
= 1 << 25,
197 ZIO_FLAG_REEXECUTED
= 1 << 26,
198 ZIO_FLAG_DELEGATED
= 1 << 27,
199 ZIO_FLAG_FASTWRITE
= 1 << 28
202 #define ZIO_FLAG_MUSTSUCCEED 0
204 #define ZIO_DDT_CHILD_FLAGS(zio) \
205 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
206 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
208 #define ZIO_GANG_CHILD_FLAGS(zio) \
209 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
210 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
212 #define ZIO_VDEV_CHILD_FLAGS(zio) \
213 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
231 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
232 * graveyard) to indicate checksum errors and fragmentation.
237 typedef void zio_done_func_t(zio_t
*zio
);
239 extern const char *zio_type_name
[ZIO_TYPES
];
242 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
243 * identifies any block in the pool. By convention, the meta-objset (MOS)
244 * is objset 0, and the meta-dnode is object 0. This covers all blocks
245 * except root blocks and ZIL blocks, which are defined as follows:
247 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
248 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
249 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
251 * Note: this structure is called a bookmark because its original purpose
252 * was to remember where to resume a pool-wide traverse.
254 * Note: this structure is passed between userland and the kernel.
255 * Therefore it must not change size or alignment between 32/64 bit
256 * compilation options.
266 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
268 (zb)->zb_objset = objset; \
269 (zb)->zb_object = object; \
270 (zb)->zb_level = level; \
271 (zb)->zb_blkid = blkid; \
272 (zb)->zb_func = FTAG; \
275 #define ZB_DESTROYED_OBJSET (-1ULL)
277 #define ZB_ROOT_OBJECT (0ULL)
278 #define ZB_ROOT_LEVEL (-1LL)
279 #define ZB_ROOT_BLKID (0ULL)
281 #define ZB_ZIL_OBJECT (0ULL)
282 #define ZB_ZIL_LEVEL (-2LL)
284 #define ZB_IS_ZERO(zb) \
285 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
286 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
287 #define ZB_IS_ROOT(zb) \
288 ((zb)->zb_object == ZB_ROOT_OBJECT && \
289 (zb)->zb_level == ZB_ROOT_LEVEL && \
290 (zb)->zb_blkid == ZB_ROOT_BLKID)
292 typedef struct zio_prop
{
293 enum zio_checksum zp_checksum
;
294 enum zio_compress zp_compress
;
295 dmu_object_type_t zp_type
;
299 boolean_t zp_dedup_verify
;
300 boolean_t zp_nopwrite
;
303 typedef struct zio_cksum_report zio_cksum_report_t
;
305 typedef void zio_cksum_finish_f(zio_cksum_report_t
*rep
,
306 const void *good_data
);
307 typedef void zio_cksum_free_f(void *cbdata
, size_t size
);
309 struct zio_bad_cksum
; /* defined in zio_checksum.h */
312 struct zio_cksum_report
{
313 struct zio_cksum_report
*zcr_next
;
314 nvlist_t
*zcr_ereport
;
315 nvlist_t
*zcr_detector
;
317 size_t zcr_cbinfo
; /* passed to zcr_free() */
320 zio_cksum_finish_f
*zcr_finish
;
321 zio_cksum_free_f
*zcr_free
;
323 /* internal use only */
324 struct zio_bad_cksum
*zcr_ckinfo
; /* information from failure */
327 typedef void zio_vsd_cksum_report_f(zio_t
*zio
, zio_cksum_report_t
*zcr
,
330 zio_vsd_cksum_report_f zio_vsd_default_cksum_report
;
332 typedef struct zio_vsd_ops
{
333 zio_done_func_t
*vsd_free
;
334 zio_vsd_cksum_report_f
*vsd_cksum_report
;
337 typedef struct zio_gang_node
{
338 zio_gbh_phys_t
*gn_gbh
;
339 struct zio_gang_node
*gn_child
[SPA_GBH_NBLKPTRS
];
342 typedef zio_t
*zio_gang_issue_func_t(zio_t
*zio
, blkptr_t
*bp
,
343 zio_gang_node_t
*gn
, void *data
);
345 typedef void zio_transform_func_t(zio_t
*zio
, void *data
, uint64_t size
);
347 typedef struct zio_transform
{
349 uint64_t zt_orig_size
;
351 zio_transform_func_t
*zt_transform
;
352 struct zio_transform
*zt_next
;
355 typedef int zio_pipe_stage_t(zio_t
*zio
);
358 * The io_reexecute flags are distinct from io_flags because the child must
359 * be able to propagate them to the parent. The normal io_flags are local
360 * to the zio, not protected by any lock, and not modifiable by children;
361 * the reexecute flags are protected by io_lock, modifiable by children,
362 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
364 #define ZIO_REEXECUTE_NOW 0x01
365 #define ZIO_REEXECUTE_SUSPEND 0x02
367 typedef struct zio_link
{
370 list_node_t zl_parent_node
;
371 list_node_t zl_child_node
;
375 /* Core information about this I/O */
376 zbookmark_t io_bookmark
;
379 enum zio_child io_child_type
;
381 zio_priority_t io_priority
;
382 uint8_t io_reexecute
;
383 uint8_t io_state
[ZIO_WAIT_TYPES
];
387 blkptr_t
*io_bp_override
;
389 list_t io_parent_list
;
390 list_t io_child_list
;
391 zio_link_t
*io_walk_link
;
393 zio_transform_t
*io_transform_stack
;
396 zio_done_func_t
*io_ready
;
397 zio_done_func_t
*io_physdone
;
398 zio_done_func_t
*io_done
;
400 int64_t io_prev_space_delta
; /* DMU private */
403 /* Data represented by this I/O */
407 uint64_t io_orig_size
;
409 /* Stuff for the vdev stack */
412 const zio_vsd_ops_t
*io_vsd_ops
;
415 hrtime_t io_timestamp
; /* submitted at */
416 hrtime_t io_delta
; /* vdev queue service delta */
417 uint64_t io_delay
; /* vdev disk service delta (ticks) */
418 avl_node_t io_queue_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
;
428 int io_child_error
[ZIO_CHILD_TYPES
];
429 uint64_t io_children
[ZIO_CHILD_TYPES
][ZIO_WAIT_TYPES
];
430 uint64_t io_child_count
;
431 uint64_t io_phys_children
;
432 uint64_t io_parent_count
;
434 zio_t
*io_gang_leader
;
435 zio_gang_node_t
*io_gang_tree
;
442 zio_cksum_report_t
*io_cksum_report
;
445 /* Taskq dispatching state */
446 taskq_ent_t io_tqent
;
449 extern zio_t
*zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
,
450 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
452 extern zio_t
*zio_root(spa_t
*spa
,
453 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
455 extern zio_t
*zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, void *data
,
456 uint64_t size
, zio_done_func_t
*done
, void *private,
457 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_t
*zb
);
459 extern zio_t
*zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
460 void *data
, uint64_t size
, const zio_prop_t
*zp
,
461 zio_done_func_t
*ready
, zio_done_func_t
*physdone
, zio_done_func_t
*done
,
463 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_t
*zb
);
465 extern zio_t
*zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
466 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
467 zio_priority_t priority
, enum zio_flag flags
, zbookmark_t
*zb
);
469 extern void zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
,
472 extern void zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
);
474 extern zio_t
*zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
476 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
478 extern zio_t
*zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
479 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
481 extern zio_t
*zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
482 uint64_t size
, void *data
, int checksum
,
483 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
484 enum zio_flag flags
, boolean_t labels
);
486 extern zio_t
*zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
487 uint64_t size
, void *data
, int checksum
,
488 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
489 enum zio_flag flags
, boolean_t labels
);
491 extern zio_t
*zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
492 const blkptr_t
*bp
, enum zio_flag flags
);
494 extern int zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
,
495 uint64_t size
, boolean_t use_slog
);
496 extern void zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
);
497 extern void zio_flush(zio_t
*zio
, vdev_t
*vd
);
498 extern void zio_shrink(zio_t
*zio
, uint64_t size
);
500 extern int zio_wait(zio_t
*zio
);
501 extern void zio_nowait(zio_t
*zio
);
502 extern void zio_execute(zio_t
*zio
);
503 extern void zio_interrupt(zio_t
*zio
);
505 extern zio_t
*zio_walk_parents(zio_t
*cio
);
506 extern zio_t
*zio_walk_children(zio_t
*pio
);
507 extern zio_t
*zio_unique_parent(zio_t
*cio
);
508 extern void zio_add_child(zio_t
*pio
, zio_t
*cio
);
510 extern void *zio_buf_alloc(size_t size
);
511 extern void zio_buf_free(void *buf
, size_t size
);
512 extern void *zio_data_buf_alloc(size_t size
);
513 extern void zio_data_buf_free(void *buf
, size_t size
);
514 extern void *zio_vdev_alloc(void);
515 extern void zio_vdev_free(void *buf
);
517 extern void zio_resubmit_stage_async(void *);
519 extern zio_t
*zio_vdev_child_io(zio_t
*zio
, blkptr_t
*bp
, vdev_t
*vd
,
520 uint64_t offset
, void *data
, uint64_t size
, int type
,
521 zio_priority_t priority
, enum zio_flag flags
,
522 zio_done_func_t
*done
, void *private);
524 extern zio_t
*zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
,
525 void *data
, uint64_t size
, int type
, zio_priority_t priority
,
526 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
528 extern void zio_vdev_io_bypass(zio_t
*zio
);
529 extern void zio_vdev_io_reissue(zio_t
*zio
);
530 extern void zio_vdev_io_redone(zio_t
*zio
);
532 extern void zio_checksum_verified(zio_t
*zio
);
533 extern int zio_worst_error(int e1
, int e2
);
535 extern enum zio_checksum
zio_checksum_select(enum zio_checksum child
,
536 enum zio_checksum parent
);
537 extern enum zio_checksum
zio_checksum_dedup_select(spa_t
*spa
,
538 enum zio_checksum child
, enum zio_checksum parent
);
539 extern enum zio_compress
zio_compress_select(enum zio_compress child
,
540 enum zio_compress parent
);
542 extern void zio_suspend(spa_t
*spa
, zio_t
*zio
);
543 extern int zio_resume(spa_t
*spa
);
544 extern void zio_resume_wait(spa_t
*spa
);
547 * Initial setup and teardown.
549 extern void zio_init(void);
550 extern void zio_fini(void);
555 struct zinject_record
;
556 extern uint32_t zio_injection_enabled
;
557 extern int zio_inject_fault(char *name
, int flags
, int *id
,
558 struct zinject_record
*record
);
559 extern int zio_inject_list_next(int *id
, char *name
, size_t buflen
,
560 struct zinject_record
*record
);
561 extern int zio_clear_fault(int id
);
562 extern void zio_handle_panic_injection(spa_t
*spa
, char *tag
, uint64_t type
);
563 extern int zio_handle_fault_injection(zio_t
*zio
, int error
);
564 extern int zio_handle_device_injection(vdev_t
*vd
, zio_t
*zio
, int error
);
565 extern int zio_handle_label_injection(zio_t
*zio
, int error
);
566 extern void zio_handle_ignored_writes(zio_t
*zio
);
567 extern uint64_t zio_handle_io_delay(zio_t
*zio
);
570 * Checksum ereport functions
572 extern void zfs_ereport_start_checksum(spa_t
*spa
, vdev_t
*vd
, struct zio
*zio
,
573 uint64_t offset
, uint64_t length
, void *arg
, struct zio_bad_cksum
*info
);
574 extern void zfs_ereport_finish_checksum(zio_cksum_report_t
*report
,
575 const void *good_data
, const void *bad_data
, boolean_t drop_if_identical
);
577 extern void zfs_ereport_send_interim_checksum(zio_cksum_report_t
*report
);
578 extern void zfs_ereport_free_checksum(zio_cksum_report_t
*report
);
580 /* If we have the good data in hand, this function can be used */
581 extern void zfs_ereport_post_checksum(spa_t
*spa
, vdev_t
*vd
,
582 struct zio
*zio
, uint64_t offset
, uint64_t length
,
583 const void *good_data
, const void *bad_data
, struct zio_bad_cksum
*info
);
585 /* Called from spa_sync(), but primarily an injection handler */
586 extern void spa_handle_ignored_writes(spa_t
*spa
);
588 /* zbookmark functions */
589 boolean_t
zbookmark_is_before(const struct dnode_phys
*dnp
,
590 const zbookmark_t
*zb1
, const zbookmark_t
*zb2
);