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, 2014 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
86 * The number of "legacy" compression functions which can be set on individual
89 #define ZIO_CHECKSUM_LEGACY_FUNCTIONS ZIO_CHECKSUM_ZILOG2
91 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
92 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
94 #define ZIO_CHECKSUM_MASK 0xffULL
95 #define ZIO_CHECKSUM_VERIFY (1 << 8)
97 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
98 #define ZIO_DEDUPDITTO_MIN 100
101 ZIO_COMPRESS_INHERIT
= 0,
117 ZIO_COMPRESS_FUNCTIONS
121 * The number of "legacy" compression functions which can be set on individual
124 #define ZIO_COMPRESS_LEGACY_FUNCTIONS ZIO_COMPRESS_LZ4
127 * The meaning of "compress = on" selected by the compression features enabled
130 #define ZIO_COMPRESS_LEGACY_ON_VALUE ZIO_COMPRESS_LZJB
131 #define ZIO_COMPRESS_LZ4_ON_VALUE ZIO_COMPRESS_LZ4
133 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
135 #define BOOTFS_COMPRESS_VALID(compress) \
136 ((compress) == ZIO_COMPRESS_LZJB || \
137 (compress) == ZIO_COMPRESS_LZ4 || \
138 (compress) == ZIO_COMPRESS_ON || \
139 (compress) == ZIO_COMPRESS_OFF)
142 * Default Linux timeout for a sd device.
144 #define ZIO_DELAY_MAX (30 * MILLISEC)
146 #define ZIO_FAILURE_MODE_WAIT 0
147 #define ZIO_FAILURE_MODE_CONTINUE 1
148 #define ZIO_FAILURE_MODE_PANIC 2
150 typedef enum zio_priority
{
151 ZIO_PRIORITY_SYNC_READ
,
152 ZIO_PRIORITY_SYNC_WRITE
, /* ZIL */
153 ZIO_PRIORITY_ASYNC_READ
, /* prefetch */
154 ZIO_PRIORITY_ASYNC_WRITE
, /* spa_sync() */
155 ZIO_PRIORITY_SCRUB
, /* asynchronous scrub/resilver reads */
156 ZIO_PRIORITY_NUM_QUEUEABLE
,
158 ZIO_PRIORITY_NOW
/* non-queued i/os (e.g. free) */
163 * Flags inherited by gang, ddt, and vdev children,
164 * and that must be equal for two zios to aggregate
166 ZIO_FLAG_DONT_AGGREGATE
= 1 << 0,
167 ZIO_FLAG_IO_REPAIR
= 1 << 1,
168 ZIO_FLAG_SELF_HEAL
= 1 << 2,
169 ZIO_FLAG_RESILVER
= 1 << 3,
170 ZIO_FLAG_SCRUB
= 1 << 4,
171 ZIO_FLAG_SCAN_THREAD
= 1 << 5,
172 ZIO_FLAG_PHYSICAL
= 1 << 6,
174 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
177 * Flags inherited by ddt, gang, and vdev children.
179 ZIO_FLAG_CANFAIL
= 1 << 7, /* must be first for INHERIT */
180 ZIO_FLAG_SPECULATIVE
= 1 << 8,
181 ZIO_FLAG_CONFIG_WRITER
= 1 << 9,
182 ZIO_FLAG_DONT_RETRY
= 1 << 10,
183 ZIO_FLAG_DONT_CACHE
= 1 << 11,
184 ZIO_FLAG_NODATA
= 1 << 12,
185 ZIO_FLAG_INDUCE_DAMAGE
= 1 << 13,
187 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
188 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
191 * Flags inherited by vdev children.
193 ZIO_FLAG_IO_RETRY
= 1 << 14, /* must be first for INHERIT */
194 ZIO_FLAG_PROBE
= 1 << 15,
195 ZIO_FLAG_TRYHARD
= 1 << 16,
196 ZIO_FLAG_OPTIONAL
= 1 << 17,
198 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
201 * Flags not inherited by any children.
203 ZIO_FLAG_DONT_QUEUE
= 1 << 18, /* must be first for INHERIT */
204 ZIO_FLAG_DONT_PROPAGATE
= 1 << 19,
205 ZIO_FLAG_IO_BYPASS
= 1 << 20,
206 ZIO_FLAG_IO_REWRITE
= 1 << 21,
207 ZIO_FLAG_RAW
= 1 << 22,
208 ZIO_FLAG_GANG_CHILD
= 1 << 23,
209 ZIO_FLAG_DDT_CHILD
= 1 << 24,
210 ZIO_FLAG_GODFATHER
= 1 << 25,
211 ZIO_FLAG_NOPWRITE
= 1 << 26,
212 ZIO_FLAG_REEXECUTED
= 1 << 27,
213 ZIO_FLAG_DELEGATED
= 1 << 28,
214 ZIO_FLAG_FASTWRITE
= 1 << 29,
217 #define ZIO_FLAG_MUSTSUCCEED 0
219 #define ZIO_DDT_CHILD_FLAGS(zio) \
220 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
221 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
223 #define ZIO_GANG_CHILD_FLAGS(zio) \
224 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
225 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
227 #define ZIO_VDEV_CHILD_FLAGS(zio) \
228 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
246 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
247 * graveyard) to indicate checksum errors and fragmentation.
252 typedef void zio_done_func_t(zio_t
*zio
);
254 extern const char *zio_type_name
[ZIO_TYPES
];
257 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
258 * identifies any block in the pool. By convention, the meta-objset (MOS)
259 * is objset 0, and the meta-dnode is object 0. This covers all blocks
260 * except root blocks and ZIL blocks, which are defined as follows:
262 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
263 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
264 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
266 * Note: this structure is called a bookmark because its original purpose
267 * was to remember where to resume a pool-wide traverse.
269 * Note: this structure is passed between userland and the kernel, and is
270 * stored on disk (by virtue of being incorporated into other on-disk
271 * structures, e.g. dsl_scan_phys_t).
273 struct zbookmark_phys
{
280 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
282 (zb)->zb_objset = objset; \
283 (zb)->zb_object = object; \
284 (zb)->zb_level = level; \
285 (zb)->zb_blkid = blkid; \
288 #define ZB_DESTROYED_OBJSET (-1ULL)
290 #define ZB_ROOT_OBJECT (0ULL)
291 #define ZB_ROOT_LEVEL (-1LL)
292 #define ZB_ROOT_BLKID (0ULL)
294 #define ZB_ZIL_OBJECT (0ULL)
295 #define ZB_ZIL_LEVEL (-2LL)
297 #define ZB_IS_ZERO(zb) \
298 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
299 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
300 #define ZB_IS_ROOT(zb) \
301 ((zb)->zb_object == ZB_ROOT_OBJECT && \
302 (zb)->zb_level == ZB_ROOT_LEVEL && \
303 (zb)->zb_blkid == ZB_ROOT_BLKID)
305 typedef struct zio_prop
{
306 enum zio_checksum zp_checksum
;
307 enum zio_compress zp_compress
;
308 dmu_object_type_t zp_type
;
312 boolean_t zp_dedup_verify
;
313 boolean_t zp_nopwrite
;
316 typedef struct zio_cksum_report zio_cksum_report_t
;
318 typedef void zio_cksum_finish_f(zio_cksum_report_t
*rep
,
319 const void *good_data
);
320 typedef void zio_cksum_free_f(void *cbdata
, size_t size
);
322 struct zio_bad_cksum
; /* defined in zio_checksum.h */
325 struct zio_cksum_report
{
326 struct zio_cksum_report
*zcr_next
;
327 nvlist_t
*zcr_ereport
;
328 nvlist_t
*zcr_detector
;
330 size_t zcr_cbinfo
; /* passed to zcr_free() */
333 zio_cksum_finish_f
*zcr_finish
;
334 zio_cksum_free_f
*zcr_free
;
336 /* internal use only */
337 struct zio_bad_cksum
*zcr_ckinfo
; /* information from failure */
340 typedef void zio_vsd_cksum_report_f(zio_t
*zio
, zio_cksum_report_t
*zcr
,
343 zio_vsd_cksum_report_f zio_vsd_default_cksum_report
;
345 typedef struct zio_vsd_ops
{
346 zio_done_func_t
*vsd_free
;
347 zio_vsd_cksum_report_f
*vsd_cksum_report
;
350 typedef struct zio_gang_node
{
351 zio_gbh_phys_t
*gn_gbh
;
352 struct zio_gang_node
*gn_child
[SPA_GBH_NBLKPTRS
];
355 typedef zio_t
*zio_gang_issue_func_t(zio_t
*zio
, blkptr_t
*bp
,
356 zio_gang_node_t
*gn
, void *data
);
358 typedef void zio_transform_func_t(zio_t
*zio
, void *data
, uint64_t size
);
360 typedef struct zio_transform
{
362 uint64_t zt_orig_size
;
364 zio_transform_func_t
*zt_transform
;
365 struct zio_transform
*zt_next
;
368 typedef int zio_pipe_stage_t(zio_t
*zio
);
371 * The io_reexecute flags are distinct from io_flags because the child must
372 * be able to propagate them to the parent. The normal io_flags are local
373 * to the zio, not protected by any lock, and not modifiable by children;
374 * the reexecute flags are protected by io_lock, modifiable by children,
375 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
377 #define ZIO_REEXECUTE_NOW 0x01
378 #define ZIO_REEXECUTE_SUSPEND 0x02
380 typedef struct zio_link
{
383 list_node_t zl_parent_node
;
384 list_node_t zl_child_node
;
388 /* Core information about this I/O */
389 zbookmark_phys_t io_bookmark
;
392 enum zio_child io_child_type
;
394 zio_priority_t io_priority
;
395 uint8_t io_reexecute
;
396 uint8_t io_state
[ZIO_WAIT_TYPES
];
400 blkptr_t
*io_bp_override
;
402 list_t io_parent_list
;
403 list_t io_child_list
;
404 zio_link_t
*io_walk_link
;
406 zio_transform_t
*io_transform_stack
;
409 zio_done_func_t
*io_ready
;
410 zio_done_func_t
*io_physdone
;
411 zio_done_func_t
*io_done
;
413 int64_t io_prev_space_delta
; /* DMU private */
416 /* Data represented by this I/O */
420 uint64_t io_orig_size
;
422 /* Stuff for the vdev stack */
425 const zio_vsd_ops_t
*io_vsd_ops
;
428 hrtime_t io_timestamp
; /* submitted at */
429 hrtime_t io_delta
; /* vdev queue service delta */
430 uint64_t io_delay
; /* vdev disk service delta (ticks) */
431 avl_node_t io_queue_node
;
432 avl_node_t io_offset_node
;
434 /* Internal pipeline state */
435 enum zio_flag io_flags
;
436 enum zio_stage io_stage
;
437 enum zio_stage io_pipeline
;
438 enum zio_flag io_orig_flags
;
439 enum zio_stage io_orig_stage
;
440 enum zio_stage io_orig_pipeline
;
442 int io_child_error
[ZIO_CHILD_TYPES
];
443 uint64_t io_children
[ZIO_CHILD_TYPES
][ZIO_WAIT_TYPES
];
444 uint64_t io_child_count
;
445 uint64_t io_phys_children
;
446 uint64_t io_parent_count
;
448 zio_t
*io_gang_leader
;
449 zio_gang_node_t
*io_gang_tree
;
456 zio_cksum_report_t
*io_cksum_report
;
459 /* Taskq dispatching state */
460 taskq_ent_t io_tqent
;
463 extern zio_t
*zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
,
464 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
466 extern zio_t
*zio_root(spa_t
*spa
,
467 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
469 extern zio_t
*zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, void *data
,
470 uint64_t size
, zio_done_func_t
*done
, void *private,
471 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
);
473 extern zio_t
*zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
474 void *data
, uint64_t size
, const zio_prop_t
*zp
,
475 zio_done_func_t
*ready
, zio_done_func_t
*physdone
, zio_done_func_t
*done
,
477 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
);
479 extern zio_t
*zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
480 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
481 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
);
483 extern void zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
,
486 extern void zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
);
488 extern zio_t
*zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
490 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
492 extern zio_t
*zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
493 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
495 extern zio_t
*zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
496 uint64_t size
, void *data
, int checksum
,
497 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
498 enum zio_flag flags
, boolean_t labels
);
500 extern zio_t
*zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
501 uint64_t size
, void *data
, int checksum
,
502 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
503 enum zio_flag flags
, boolean_t labels
);
505 extern zio_t
*zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
506 const blkptr_t
*bp
, enum zio_flag flags
);
508 extern int zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
,
509 uint64_t size
, boolean_t use_slog
);
510 extern void zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
);
511 extern void zio_flush(zio_t
*zio
, vdev_t
*vd
);
512 extern void zio_shrink(zio_t
*zio
, uint64_t size
);
514 extern int zio_wait(zio_t
*zio
);
515 extern void zio_nowait(zio_t
*zio
);
516 extern void zio_execute(zio_t
*zio
);
517 extern void zio_interrupt(zio_t
*zio
);
519 extern zio_t
*zio_walk_parents(zio_t
*cio
);
520 extern zio_t
*zio_walk_children(zio_t
*pio
);
521 extern zio_t
*zio_unique_parent(zio_t
*cio
);
522 extern void zio_add_child(zio_t
*pio
, zio_t
*cio
);
524 extern void *zio_buf_alloc(size_t size
);
525 extern void zio_buf_free(void *buf
, size_t size
);
526 extern void *zio_data_buf_alloc(size_t size
);
527 extern void zio_data_buf_free(void *buf
, size_t size
);
528 extern void *zio_buf_alloc_flags(size_t size
, int flags
);
530 extern void zio_resubmit_stage_async(void *);
532 extern zio_t
*zio_vdev_child_io(zio_t
*zio
, blkptr_t
*bp
, vdev_t
*vd
,
533 uint64_t offset
, void *data
, uint64_t size
, int type
,
534 zio_priority_t priority
, enum zio_flag flags
,
535 zio_done_func_t
*done
, void *private);
537 extern zio_t
*zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
,
538 void *data
, uint64_t size
, int type
, zio_priority_t priority
,
539 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
541 extern void zio_vdev_io_bypass(zio_t
*zio
);
542 extern void zio_vdev_io_reissue(zio_t
*zio
);
543 extern void zio_vdev_io_redone(zio_t
*zio
);
545 extern void zio_checksum_verified(zio_t
*zio
);
546 extern int zio_worst_error(int e1
, int e2
);
548 extern enum zio_checksum
zio_checksum_select(enum zio_checksum child
,
549 enum zio_checksum parent
);
550 extern enum zio_checksum
zio_checksum_dedup_select(spa_t
*spa
,
551 enum zio_checksum child
, enum zio_checksum parent
);
552 extern enum zio_compress
zio_compress_select(spa_t
*spa
,
553 enum zio_compress child
, enum zio_compress parent
);
555 extern void zio_suspend(spa_t
*spa
, zio_t
*zio
);
556 extern int zio_resume(spa_t
*spa
);
557 extern void zio_resume_wait(spa_t
*spa
);
560 * Initial setup and teardown.
562 extern void zio_init(void);
563 extern void zio_fini(void);
568 struct zinject_record
;
569 extern uint32_t zio_injection_enabled
;
570 extern int zio_inject_fault(char *name
, int flags
, int *id
,
571 struct zinject_record
*record
);
572 extern int zio_inject_list_next(int *id
, char *name
, size_t buflen
,
573 struct zinject_record
*record
);
574 extern int zio_clear_fault(int id
);
575 extern void zio_handle_panic_injection(spa_t
*spa
, char *tag
, uint64_t type
);
576 extern int zio_handle_fault_injection(zio_t
*zio
, int error
);
577 extern int zio_handle_device_injection(vdev_t
*vd
, zio_t
*zio
, int error
);
578 extern int zio_handle_label_injection(zio_t
*zio
, int error
);
579 extern void zio_handle_ignored_writes(zio_t
*zio
);
580 extern uint64_t zio_handle_io_delay(zio_t
*zio
);
583 * Checksum ereport functions
585 extern void zfs_ereport_start_checksum(spa_t
*spa
, vdev_t
*vd
, struct zio
*zio
,
586 uint64_t offset
, uint64_t length
, void *arg
, struct zio_bad_cksum
*info
);
587 extern void zfs_ereport_finish_checksum(zio_cksum_report_t
*report
,
588 const void *good_data
, const void *bad_data
, boolean_t drop_if_identical
);
590 extern void zfs_ereport_send_interim_checksum(zio_cksum_report_t
*report
);
591 extern void zfs_ereport_free_checksum(zio_cksum_report_t
*report
);
593 /* If we have the good data in hand, this function can be used */
594 extern void zfs_ereport_post_checksum(spa_t
*spa
, vdev_t
*vd
,
595 struct zio
*zio
, uint64_t offset
, uint64_t length
,
596 const void *good_data
, const void *bad_data
, struct zio_bad_cksum
*info
);
598 /* Called from spa_sync(), but primarily an injection handler */
599 extern void spa_handle_ignored_writes(spa_t
*spa
);
601 /* zbookmark_phys functions */
602 boolean_t
zbookmark_is_before(const struct dnode_phys
*dnp
,
603 const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
);