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.
26 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2012 by Delphix. All rights reserved.
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_FUNCTIONS
86 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
87 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
89 #define ZIO_CHECKSUM_MASK 0xffULL
90 #define ZIO_CHECKSUM_VERIFY (1 << 8)
92 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
93 #define ZIO_DEDUPDITTO_MIN 100
96 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_ON && \
120 ZIO_COMPRESS_ON_VALUE == ZIO_COMPRESS_LZJB) || \
121 (compress) == ZIO_COMPRESS_OFF)
124 * Default Linux timeout for a sd device.
126 #define ZIO_DELAY_MAX (30 * MILLISEC)
128 #define ZIO_FAILURE_MODE_WAIT 0
129 #define ZIO_FAILURE_MODE_CONTINUE 1
130 #define ZIO_FAILURE_MODE_PANIC 2
132 #define ZIO_PRIORITY_NOW (zio_priority_table[0])
133 #define ZIO_PRIORITY_SYNC_READ (zio_priority_table[1])
134 #define ZIO_PRIORITY_SYNC_WRITE (zio_priority_table[2])
135 #define ZIO_PRIORITY_LOG_WRITE (zio_priority_table[3])
136 #define ZIO_PRIORITY_CACHE_FILL (zio_priority_table[4])
137 #define ZIO_PRIORITY_AGG (zio_priority_table[5])
138 #define ZIO_PRIORITY_FREE (zio_priority_table[6])
139 #define ZIO_PRIORITY_ASYNC_WRITE (zio_priority_table[7])
140 #define ZIO_PRIORITY_ASYNC_READ (zio_priority_table[8])
141 #define ZIO_PRIORITY_RESILVER (zio_priority_table[9])
142 #define ZIO_PRIORITY_SCRUB (zio_priority_table[10])
143 #define ZIO_PRIORITY_DDT_PREFETCH (zio_priority_table[11])
144 #define ZIO_PRIORITY_TABLE_SIZE 12
146 #define ZIO_PIPELINE_CONTINUE 0x100
147 #define ZIO_PIPELINE_STOP 0x101
151 * Flags inherited by gang, ddt, and vdev children,
152 * and that must be equal for two zios to aggregate
154 ZIO_FLAG_DONT_AGGREGATE
= 1 << 0,
155 ZIO_FLAG_IO_REPAIR
= 1 << 1,
156 ZIO_FLAG_SELF_HEAL
= 1 << 2,
157 ZIO_FLAG_RESILVER
= 1 << 3,
158 ZIO_FLAG_SCRUB
= 1 << 4,
159 ZIO_FLAG_SCAN_THREAD
= 1 << 5,
161 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
164 * Flags inherited by ddt, gang, and vdev children.
166 ZIO_FLAG_CANFAIL
= 1 << 6, /* must be first for INHERIT */
167 ZIO_FLAG_SPECULATIVE
= 1 << 7,
168 ZIO_FLAG_CONFIG_WRITER
= 1 << 8,
169 ZIO_FLAG_DONT_RETRY
= 1 << 9,
170 ZIO_FLAG_DONT_CACHE
= 1 << 10,
171 ZIO_FLAG_NODATA
= 1 << 11,
172 ZIO_FLAG_INDUCE_DAMAGE
= 1 << 12,
174 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
175 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
178 * Flags inherited by vdev children.
180 ZIO_FLAG_IO_RETRY
= 1 << 13, /* must be first for INHERIT */
181 ZIO_FLAG_PROBE
= 1 << 14,
182 ZIO_FLAG_TRYHARD
= 1 << 15,
183 ZIO_FLAG_OPTIONAL
= 1 << 16,
185 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
188 * Flags not inherited by any children.
190 ZIO_FLAG_DONT_QUEUE
= 1 << 17, /* must be first for INHERIT */
191 ZIO_FLAG_DONT_PROPAGATE
= 1 << 18,
192 ZIO_FLAG_IO_BYPASS
= 1 << 19,
193 ZIO_FLAG_IO_REWRITE
= 1 << 20,
194 ZIO_FLAG_RAW
= 1 << 21,
195 ZIO_FLAG_GANG_CHILD
= 1 << 22,
196 ZIO_FLAG_DDT_CHILD
= 1 << 23,
197 ZIO_FLAG_GODFATHER
= 1 << 24,
198 ZIO_FLAG_FASTWRITE
= 1 << 25
201 #define ZIO_FLAG_MUSTSUCCEED 0
203 #define ZIO_DDT_CHILD_FLAGS(zio) \
204 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
205 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
207 #define ZIO_GANG_CHILD_FLAGS(zio) \
208 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
209 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
211 #define ZIO_VDEV_CHILD_FLAGS(zio) \
212 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
230 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
231 * graveyard) to indicate checksum errors and fragmentation.
236 typedef void zio_done_func_t(zio_t
*zio
);
238 extern uint8_t zio_priority_table
[ZIO_PRIORITY_TABLE_SIZE
];
239 extern 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.
258 typedef struct zbookmark
{
265 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
267 (zb)->zb_objset = objset; \
268 (zb)->zb_object = object; \
269 (zb)->zb_level = level; \
270 (zb)->zb_blkid = blkid; \
273 #define ZB_DESTROYED_OBJSET (-1ULL)
275 #define ZB_ROOT_OBJECT (0ULL)
276 #define ZB_ROOT_LEVEL (-1LL)
277 #define ZB_ROOT_BLKID (0ULL)
279 #define ZB_ZIL_OBJECT (0ULL)
280 #define ZB_ZIL_LEVEL (-2LL)
282 #define ZB_IS_ZERO(zb) \
283 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
284 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
285 #define ZB_IS_ROOT(zb) \
286 ((zb)->zb_object == ZB_ROOT_OBJECT && \
287 (zb)->zb_level == ZB_ROOT_LEVEL && \
288 (zb)->zb_blkid == ZB_ROOT_BLKID)
290 typedef struct zio_prop
{
291 enum zio_checksum zp_checksum
;
292 enum zio_compress zp_compress
;
293 dmu_object_type_t zp_type
;
297 uint8_t zp_dedup_verify
;
300 typedef struct zio_cksum_report zio_cksum_report_t
;
302 typedef void zio_cksum_finish_f(zio_cksum_report_t
*rep
,
303 const void *good_data
);
304 typedef void zio_cksum_free_f(void *cbdata
, size_t size
);
306 struct zio_bad_cksum
; /* defined in zio_checksum.h */
309 struct zio_cksum_report
{
310 struct zio_cksum_report
*zcr_next
;
311 nvlist_t
*zcr_ereport
;
312 nvlist_t
*zcr_detector
;
314 size_t zcr_cbinfo
; /* passed to zcr_free() */
317 zio_cksum_finish_f
*zcr_finish
;
318 zio_cksum_free_f
*zcr_free
;
320 /* internal use only */
321 struct zio_bad_cksum
*zcr_ckinfo
; /* information from failure */
324 typedef void zio_vsd_cksum_report_f(zio_t
*zio
, zio_cksum_report_t
*zcr
,
327 zio_vsd_cksum_report_f zio_vsd_default_cksum_report
;
329 typedef struct zio_vsd_ops
{
330 zio_done_func_t
*vsd_free
;
331 zio_vsd_cksum_report_f
*vsd_cksum_report
;
334 typedef struct zio_gang_node
{
335 zio_gbh_phys_t
*gn_gbh
;
336 struct zio_gang_node
*gn_child
[SPA_GBH_NBLKPTRS
];
339 typedef zio_t
*zio_gang_issue_func_t(zio_t
*zio
, blkptr_t
*bp
,
340 zio_gang_node_t
*gn
, void *data
);
342 typedef void zio_transform_func_t(zio_t
*zio
, void *data
, uint64_t size
);
344 typedef struct zio_transform
{
346 uint64_t zt_orig_size
;
348 zio_transform_func_t
*zt_transform
;
349 struct zio_transform
*zt_next
;
352 typedef int zio_pipe_stage_t(zio_t
*zio
);
355 * The io_reexecute flags are distinct from io_flags because the child must
356 * be able to propagate them to the parent. The normal io_flags are local
357 * to the zio, not protected by any lock, and not modifiable by children;
358 * the reexecute flags are protected by io_lock, modifiable by children,
359 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
361 #define ZIO_REEXECUTE_NOW 0x01
362 #define ZIO_REEXECUTE_SUSPEND 0x02
364 typedef struct zio_link
{
367 list_node_t zl_parent_node
;
368 list_node_t zl_child_node
;
372 /* Core information about this I/O */
373 zbookmark_t io_bookmark
;
376 enum zio_child io_child_type
;
379 uint8_t io_reexecute
;
380 uint8_t io_state
[ZIO_WAIT_TYPES
];
384 blkptr_t
*io_bp_override
;
386 list_t io_parent_list
;
387 list_t io_child_list
;
388 zio_link_t
*io_walk_link
;
390 zio_transform_t
*io_transform_stack
;
393 zio_done_func_t
*io_ready
;
394 zio_done_func_t
*io_done
;
396 int64_t io_prev_space_delta
; /* DMU private */
399 /* Data represented by this I/O */
403 uint64_t io_orig_size
;
405 /* Stuff for the vdev stack */
408 const zio_vsd_ops_t
*io_vsd_ops
;
411 uint64_t io_deadline
;
412 avl_node_t io_offset_node
;
413 avl_node_t io_deadline_node
;
414 avl_tree_t
*io_vdev_tree
;
416 /* Internal pipeline state */
417 enum zio_flag io_flags
;
418 enum zio_stage io_stage
;
419 enum zio_stage io_pipeline
;
420 enum zio_flag io_orig_flags
;
421 enum zio_stage io_orig_stage
;
422 enum zio_stage io_orig_pipeline
;
425 int io_child_error
[ZIO_CHILD_TYPES
];
426 uint64_t io_children
[ZIO_CHILD_TYPES
][ZIO_WAIT_TYPES
];
427 uint64_t io_child_count
;
428 uint64_t io_parent_count
;
430 zio_t
*io_gang_leader
;
431 zio_gang_node_t
*io_gang_tree
;
438 zio_cksum_report_t
*io_cksum_report
;
441 /* Taskq dispatching state */
442 taskq_ent_t io_tqent
;
445 extern zio_t
*zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
,
446 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
448 extern zio_t
*zio_root(spa_t
*spa
,
449 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
451 extern zio_t
*zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, void *data
,
452 uint64_t size
, zio_done_func_t
*done
, void *private,
453 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
);
455 extern zio_t
*zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
456 void *data
, uint64_t size
, const zio_prop_t
*zp
,
457 zio_done_func_t
*ready
, zio_done_func_t
*done
, void *private,
458 int priority
, enum zio_flag flags
, const zbookmark_t
*zb
);
460 extern zio_t
*zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
461 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
462 int priority
, enum zio_flag flags
, zbookmark_t
*zb
);
464 extern void zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
);
466 extern void zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
);
468 extern zio_t
*zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
470 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
472 extern zio_t
*zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
473 zio_done_func_t
*done
, void *private, int priority
, enum zio_flag flags
);
475 extern zio_t
*zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
476 uint64_t size
, void *data
, int checksum
,
477 zio_done_func_t
*done
, void *private, int priority
, enum zio_flag flags
,
480 extern zio_t
*zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
481 uint64_t size
, void *data
, int checksum
,
482 zio_done_func_t
*done
, void *private, int priority
, enum zio_flag flags
,
485 extern zio_t
*zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
486 const blkptr_t
*bp
, enum zio_flag flags
);
488 extern int zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
,
489 uint64_t size
, boolean_t use_slog
);
490 extern void zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
);
491 extern void zio_flush(zio_t
*zio
, vdev_t
*vd
);
492 extern void zio_shrink(zio_t
*zio
, uint64_t size
);
494 extern int zio_wait(zio_t
*zio
);
495 extern void zio_nowait(zio_t
*zio
);
496 extern void zio_execute(zio_t
*zio
);
497 extern void zio_interrupt(zio_t
*zio
);
499 extern zio_t
*zio_walk_parents(zio_t
*cio
);
500 extern zio_t
*zio_walk_children(zio_t
*pio
);
501 extern zio_t
*zio_unique_parent(zio_t
*cio
);
502 extern void zio_add_child(zio_t
*pio
, zio_t
*cio
);
504 extern void *zio_buf_alloc(size_t size
);
505 extern void zio_buf_free(void *buf
, size_t size
);
506 extern void *zio_data_buf_alloc(size_t size
);
507 extern void zio_data_buf_free(void *buf
, size_t size
);
508 extern void *zio_vdev_alloc(void);
509 extern void zio_vdev_free(void *buf
);
511 extern void zio_resubmit_stage_async(void *);
513 extern zio_t
*zio_vdev_child_io(zio_t
*zio
, blkptr_t
*bp
, vdev_t
*vd
,
514 uint64_t offset
, void *data
, uint64_t size
, int type
, int priority
,
515 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
517 extern zio_t
*zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
,
518 void *data
, uint64_t size
, int type
, int priority
,
519 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
521 extern void zio_vdev_io_bypass(zio_t
*zio
);
522 extern void zio_vdev_io_reissue(zio_t
*zio
);
523 extern void zio_vdev_io_redone(zio_t
*zio
);
525 extern void zio_checksum_verified(zio_t
*zio
);
526 extern int zio_worst_error(int e1
, int e2
);
528 extern enum zio_checksum
zio_checksum_select(enum zio_checksum child
,
529 enum zio_checksum parent
);
530 extern enum zio_checksum
zio_checksum_dedup_select(spa_t
*spa
,
531 enum zio_checksum child
, enum zio_checksum parent
);
532 extern enum zio_compress
zio_compress_select(enum zio_compress child
,
533 enum zio_compress parent
);
535 extern void zio_suspend(spa_t
*spa
, zio_t
*zio
);
536 extern int zio_resume(spa_t
*spa
);
537 extern void zio_resume_wait(spa_t
*spa
);
540 * Initial setup and teardown.
542 extern void zio_init(void);
543 extern void zio_fini(void);
548 struct zinject_record
;
549 extern uint32_t zio_injection_enabled
;
550 extern int zio_inject_fault(char *name
, int flags
, int *id
,
551 struct zinject_record
*record
);
552 extern int zio_inject_list_next(int *id
, char *name
, size_t buflen
,
553 struct zinject_record
*record
);
554 extern int zio_clear_fault(int id
);
555 extern void zio_handle_panic_injection(spa_t
*spa
, char *tag
, uint64_t type
);
556 extern int zio_handle_fault_injection(zio_t
*zio
, int error
);
557 extern int zio_handle_device_injection(vdev_t
*vd
, zio_t
*zio
, int error
);
558 extern int zio_handle_label_injection(zio_t
*zio
, int error
);
559 extern void zio_handle_ignored_writes(zio_t
*zio
);
562 * Checksum ereport functions
564 extern void zfs_ereport_start_checksum(spa_t
*spa
, vdev_t
*vd
, struct zio
*zio
,
565 uint64_t offset
, uint64_t length
, void *arg
, struct zio_bad_cksum
*info
);
566 extern void zfs_ereport_finish_checksum(zio_cksum_report_t
*report
,
567 const void *good_data
, const void *bad_data
, boolean_t drop_if_identical
);
569 extern void zfs_ereport_send_interim_checksum(zio_cksum_report_t
*report
);
570 extern void zfs_ereport_free_checksum(zio_cksum_report_t
*report
);
572 /* If we have the good data in hand, this function can be used */
573 extern void zfs_ereport_post_checksum(spa_t
*spa
, vdev_t
*vd
,
574 struct zio
*zio
, uint64_t offset
, uint64_t length
,
575 const void *good_data
, const void *bad_data
, struct zio_bad_cksum
*info
);
577 /* Called from spa_sync(), but primarily an injection handler */
578 extern void spa_handle_ignored_writes(spa_t
*spa
);
580 /* zbookmark functions */
581 boolean_t
zbookmark_is_before(const struct dnode_phys
*dnp
,
582 const zbookmark_t
*zb1
, const zbookmark_t
*zb2
);