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.
27 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
33 #include <sys/zio_priority.h>
34 #include <sys/zfs_context.h>
38 #include <sys/fs/zfs.h>
39 #include <sys/zio_impl.h>
48 #define ZEC_MAGIC 0x210da7ab10c7a11ULL
50 typedef struct zio_eck
{
51 uint64_t zec_magic
; /* for validation, endianness */
52 zio_cksum_t zec_cksum
; /* 256-bit checksum */
56 * Gang block headers are self-checksumming and contain an array
59 #define SPA_GANGBLOCKSIZE SPA_MINBLOCKSIZE
60 #define SPA_GBH_NBLKPTRS ((SPA_GANGBLOCKSIZE - \
61 sizeof (zio_eck_t)) / sizeof (blkptr_t))
62 #define SPA_GBH_FILLER ((SPA_GANGBLOCKSIZE - \
63 sizeof (zio_eck_t) - \
64 (SPA_GBH_NBLKPTRS * sizeof (blkptr_t))) /\
67 typedef struct zio_gbh
{
68 blkptr_t zg_blkptr
[SPA_GBH_NBLKPTRS
];
69 uint64_t zg_filler
[SPA_GBH_FILLER
];
74 ZIO_CHECKSUM_INHERIT
= 0,
78 ZIO_CHECKSUM_GANG_HEADER
,
80 ZIO_CHECKSUM_FLETCHER_2
,
81 ZIO_CHECKSUM_FLETCHER_4
,
84 ZIO_CHECKSUM_NOPARITY
,
88 ZIO_CHECKSUM_FUNCTIONS
92 * The number of "legacy" compression functions which can be set on individual
95 #define ZIO_CHECKSUM_LEGACY_FUNCTIONS ZIO_CHECKSUM_ZILOG2
97 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
98 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
100 #define ZIO_CHECKSUM_MASK 0xffULL
101 #define ZIO_CHECKSUM_VERIFY (1 << 8)
103 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
104 #define ZIO_DEDUPDITTO_MIN 100
107 ZIO_COMPRESS_INHERIT
= 0,
123 ZIO_COMPRESS_FUNCTIONS
127 * The number of "legacy" compression functions which can be set on individual
130 #define ZIO_COMPRESS_LEGACY_FUNCTIONS ZIO_COMPRESS_LZ4
133 * The meaning of "compress = on" selected by the compression features enabled
136 #define ZIO_COMPRESS_LEGACY_ON_VALUE ZIO_COMPRESS_LZJB
137 #define ZIO_COMPRESS_LZ4_ON_VALUE ZIO_COMPRESS_LZ4
139 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
141 #define BOOTFS_COMPRESS_VALID(compress) \
142 ((compress) == ZIO_COMPRESS_LZJB || \
143 (compress) == ZIO_COMPRESS_LZ4 || \
144 (compress) == ZIO_COMPRESS_ON || \
145 (compress) == ZIO_COMPRESS_OFF)
147 #define ZIO_FAILURE_MODE_WAIT 0
148 #define ZIO_FAILURE_MODE_CONTINUE 1
149 #define ZIO_FAILURE_MODE_PANIC 2
153 * Flags inherited by gang, ddt, and vdev children,
154 * and that must be equal for two zios to aggregate
156 ZIO_FLAG_DONT_AGGREGATE
= 1 << 0,
157 ZIO_FLAG_IO_REPAIR
= 1 << 1,
158 ZIO_FLAG_SELF_HEAL
= 1 << 2,
159 ZIO_FLAG_RESILVER
= 1 << 3,
160 ZIO_FLAG_SCRUB
= 1 << 4,
161 ZIO_FLAG_SCAN_THREAD
= 1 << 5,
162 ZIO_FLAG_PHYSICAL
= 1 << 6,
164 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
167 * Flags inherited by ddt, gang, and vdev children.
169 ZIO_FLAG_CANFAIL
= 1 << 7, /* must be first for INHERIT */
170 ZIO_FLAG_SPECULATIVE
= 1 << 8,
171 ZIO_FLAG_CONFIG_WRITER
= 1 << 9,
172 ZIO_FLAG_DONT_RETRY
= 1 << 10,
173 ZIO_FLAG_DONT_CACHE
= 1 << 11,
174 ZIO_FLAG_NODATA
= 1 << 12,
175 ZIO_FLAG_INDUCE_DAMAGE
= 1 << 13,
177 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
178 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
181 * Flags inherited by vdev children.
183 ZIO_FLAG_IO_RETRY
= 1 << 14, /* must be first for INHERIT */
184 ZIO_FLAG_PROBE
= 1 << 15,
185 ZIO_FLAG_TRYHARD
= 1 << 16,
186 ZIO_FLAG_OPTIONAL
= 1 << 17,
188 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
191 * Flags not inherited by any children.
193 ZIO_FLAG_DONT_QUEUE
= 1 << 18, /* must be first for INHERIT */
194 ZIO_FLAG_DONT_PROPAGATE
= 1 << 19,
195 ZIO_FLAG_IO_BYPASS
= 1 << 20,
196 ZIO_FLAG_IO_REWRITE
= 1 << 21,
197 ZIO_FLAG_RAW
= 1 << 22,
198 ZIO_FLAG_GANG_CHILD
= 1 << 23,
199 ZIO_FLAG_DDT_CHILD
= 1 << 24,
200 ZIO_FLAG_GODFATHER
= 1 << 25,
201 ZIO_FLAG_NOPWRITE
= 1 << 26,
202 ZIO_FLAG_REEXECUTED
= 1 << 27,
203 ZIO_FLAG_DELEGATED
= 1 << 28,
206 #define ZIO_FLAG_MUSTSUCCEED 0
208 #define ZIO_DDT_CHILD_FLAGS(zio) \
209 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
210 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
212 #define ZIO_GANG_CHILD_FLAGS(zio) \
213 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
214 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
216 #define ZIO_VDEV_CHILD_FLAGS(zio) \
217 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
235 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
236 * graveyard) to indicate checksum errors and fragmentation.
241 typedef void zio_done_func_t(zio_t
*zio
);
243 extern const char *zio_type_name
[ZIO_TYPES
];
246 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
247 * identifies any block in the pool. By convention, the meta-objset (MOS)
248 * is objset 0, and the meta-dnode is object 0. This covers all blocks
249 * except root blocks and ZIL blocks, which are defined as follows:
251 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
252 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
253 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
254 * dnode visit bookmarks are <objset, object id of dnode, -3, 0>.
256 * Note: this structure is called a bookmark because its original purpose
257 * was to remember where to resume a pool-wide traverse.
259 * Note: this structure is passed between userland and the kernel, and is
260 * stored on disk (by virtue of being incorporated into other on-disk
261 * structures, e.g. dsl_scan_phys_t).
263 typedef struct zbookmark_phys
{
270 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
272 (zb)->zb_objset = objset; \
273 (zb)->zb_object = object; \
274 (zb)->zb_level = level; \
275 (zb)->zb_blkid = blkid; \
278 #define ZB_DESTROYED_OBJSET (-1ULL)
280 #define ZB_ROOT_OBJECT (0ULL)
281 #define ZB_ROOT_LEVEL (-1LL)
282 #define ZB_ROOT_BLKID (0ULL)
284 #define ZB_ZIL_OBJECT (0ULL)
285 #define ZB_ZIL_LEVEL (-2LL)
287 #define ZB_DNODE_LEVEL (-3LL)
288 #define ZB_DNODE_BLKID (0ULL)
290 #define ZB_IS_ZERO(zb) \
291 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
292 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
293 #define ZB_IS_ROOT(zb) \
294 ((zb)->zb_object == ZB_ROOT_OBJECT && \
295 (zb)->zb_level == ZB_ROOT_LEVEL && \
296 (zb)->zb_blkid == ZB_ROOT_BLKID)
298 typedef struct zio_prop
{
299 enum zio_checksum zp_checksum
;
300 enum zio_compress zp_compress
;
301 dmu_object_type_t zp_type
;
305 boolean_t zp_dedup_verify
;
306 boolean_t zp_nopwrite
;
309 typedef struct zio_cksum_report zio_cksum_report_t
;
311 typedef void zio_cksum_finish_f(zio_cksum_report_t
*rep
,
312 const void *good_data
);
313 typedef void zio_cksum_free_f(void *cbdata
, size_t size
);
315 struct zio_bad_cksum
; /* defined in zio_checksum.h */
318 struct zio_cksum_report
{
319 struct zio_cksum_report
*zcr_next
;
320 nvlist_t
*zcr_ereport
;
321 nvlist_t
*zcr_detector
;
323 size_t zcr_cbinfo
; /* passed to zcr_free() */
326 zio_cksum_finish_f
*zcr_finish
;
327 zio_cksum_free_f
*zcr_free
;
329 /* internal use only */
330 struct zio_bad_cksum
*zcr_ckinfo
; /* information from failure */
333 typedef void zio_vsd_cksum_report_f(zio_t
*zio
, zio_cksum_report_t
*zcr
,
336 zio_vsd_cksum_report_f zio_vsd_default_cksum_report
;
338 typedef struct zio_vsd_ops
{
339 zio_done_func_t
*vsd_free
;
340 zio_vsd_cksum_report_f
*vsd_cksum_report
;
343 typedef struct zio_gang_node
{
344 zio_gbh_phys_t
*gn_gbh
;
345 struct zio_gang_node
*gn_child
[SPA_GBH_NBLKPTRS
];
348 typedef zio_t
*zio_gang_issue_func_t(zio_t
*zio
, blkptr_t
*bp
,
349 zio_gang_node_t
*gn
, void *data
);
351 typedef void zio_transform_func_t(zio_t
*zio
, void *data
, uint64_t size
);
353 typedef struct zio_transform
{
355 uint64_t zt_orig_size
;
357 zio_transform_func_t
*zt_transform
;
358 struct zio_transform
*zt_next
;
361 typedef int zio_pipe_stage_t(zio_t
*zio
);
364 * The io_reexecute flags are distinct from io_flags because the child must
365 * be able to propagate them to the parent. The normal io_flags are local
366 * to the zio, not protected by any lock, and not modifiable by children;
367 * the reexecute flags are protected by io_lock, modifiable by children,
368 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
370 #define ZIO_REEXECUTE_NOW 0x01
371 #define ZIO_REEXECUTE_SUSPEND 0x02
373 typedef struct zio_link
{
376 list_node_t zl_parent_node
;
377 list_node_t zl_child_node
;
381 /* Core information about this I/O */
382 zbookmark_phys_t io_bookmark
;
385 enum zio_child io_child_type
;
387 zio_priority_t io_priority
;
388 uint8_t io_reexecute
;
389 uint8_t io_state
[ZIO_WAIT_TYPES
];
393 blkptr_t
*io_bp_override
;
395 list_t io_parent_list
;
396 list_t io_child_list
;
397 zio_link_t
*io_walk_link
;
399 zio_transform_t
*io_transform_stack
;
402 zio_done_func_t
*io_ready
;
403 zio_done_func_t
*io_physdone
;
404 zio_done_func_t
*io_done
;
406 int64_t io_prev_space_delta
; /* DMU private */
409 /* Data represented by this I/O */
413 uint64_t io_orig_size
;
415 /* Stuff for the vdev stack */
418 const zio_vsd_ops_t
*io_vsd_ops
;
421 hrtime_t io_timestamp
;
422 hrtime_t io_target_timestamp
;
423 avl_node_t io_queue_node
;
424 avl_node_t io_offset_node
;
426 /* Internal pipeline state */
427 enum zio_flag io_flags
;
428 enum zio_stage io_stage
;
429 enum zio_stage io_pipeline
;
430 enum zio_flag io_orig_flags
;
431 enum zio_stage io_orig_stage
;
432 enum zio_stage io_orig_pipeline
;
434 int io_child_error
[ZIO_CHILD_TYPES
];
435 uint64_t io_children
[ZIO_CHILD_TYPES
][ZIO_WAIT_TYPES
];
436 uint64_t io_child_count
;
437 uint64_t io_phys_children
;
438 uint64_t io_parent_count
;
440 zio_t
*io_gang_leader
;
441 zio_gang_node_t
*io_gang_tree
;
448 zio_cksum_report_t
*io_cksum_report
;
451 /* Taskq dispatching state */
452 taskq_ent_t io_tqent
;
455 extern zio_t
*zio_null(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
,
456 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
458 extern zio_t
*zio_root(spa_t
*spa
,
459 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
461 extern zio_t
*zio_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, void *data
,
462 uint64_t size
, zio_done_func_t
*done
, void *private,
463 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
);
465 extern zio_t
*zio_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
466 void *data
, uint64_t size
, const zio_prop_t
*zp
,
467 zio_done_func_t
*ready
, zio_done_func_t
*physdone
, zio_done_func_t
*done
,
469 zio_priority_t priority
, enum zio_flag flags
, const zbookmark_phys_t
*zb
);
471 extern zio_t
*zio_rewrite(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
,
472 void *data
, uint64_t size
, zio_done_func_t
*done
, void *private,
473 zio_priority_t priority
, enum zio_flag flags
, zbookmark_phys_t
*zb
);
475 extern void zio_write_override(zio_t
*zio
, blkptr_t
*bp
, int copies
,
478 extern void zio_free(spa_t
*spa
, uint64_t txg
, const blkptr_t
*bp
);
480 extern zio_t
*zio_claim(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
482 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
484 extern zio_t
*zio_ioctl(zio_t
*pio
, spa_t
*spa
, vdev_t
*vd
, int cmd
,
485 zio_done_func_t
*done
, void *private, enum zio_flag flags
);
487 extern zio_t
*zio_read_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
488 uint64_t size
, void *data
, int checksum
,
489 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
490 enum zio_flag flags
, boolean_t labels
);
492 extern zio_t
*zio_write_phys(zio_t
*pio
, vdev_t
*vd
, uint64_t offset
,
493 uint64_t size
, void *data
, int checksum
,
494 zio_done_func_t
*done
, void *private, zio_priority_t priority
,
495 enum zio_flag flags
, boolean_t labels
);
497 extern zio_t
*zio_free_sync(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
498 const blkptr_t
*bp
, enum zio_flag flags
);
500 extern int zio_alloc_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*new_bp
,
501 blkptr_t
*old_bp
, uint64_t size
, boolean_t use_slog
);
502 extern void zio_free_zil(spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
);
503 extern void zio_flush(zio_t
*zio
, vdev_t
*vd
);
504 extern void zio_shrink(zio_t
*zio
, uint64_t size
);
506 extern int zio_wait(zio_t
*zio
);
507 extern void zio_nowait(zio_t
*zio
);
508 extern void zio_execute(zio_t
*zio
);
509 extern void zio_interrupt(zio_t
*zio
);
510 extern void zio_delay_init(zio_t
*zio
);
511 extern void zio_delay_interrupt(zio_t
*zio
);
513 extern zio_t
*zio_walk_parents(zio_t
*cio
);
514 extern zio_t
*zio_walk_children(zio_t
*pio
);
515 extern zio_t
*zio_unique_parent(zio_t
*cio
);
516 extern void zio_add_child(zio_t
*pio
, zio_t
*cio
);
518 extern void *zio_buf_alloc(size_t size
);
519 extern void zio_buf_free(void *buf
, size_t size
);
520 extern void *zio_data_buf_alloc(size_t size
);
521 extern void zio_data_buf_free(void *buf
, size_t size
);
523 extern void zio_resubmit_stage_async(void *);
525 extern zio_t
*zio_vdev_child_io(zio_t
*zio
, blkptr_t
*bp
, vdev_t
*vd
,
526 uint64_t offset
, void *data
, uint64_t size
, int type
,
527 zio_priority_t priority
, enum zio_flag flags
,
528 zio_done_func_t
*done
, void *private);
530 extern zio_t
*zio_vdev_delegated_io(vdev_t
*vd
, uint64_t offset
,
531 void *data
, uint64_t size
, int type
, zio_priority_t priority
,
532 enum zio_flag flags
, zio_done_func_t
*done
, void *private);
534 extern void zio_vdev_io_bypass(zio_t
*zio
);
535 extern void zio_vdev_io_reissue(zio_t
*zio
);
536 extern void zio_vdev_io_redone(zio_t
*zio
);
538 extern void zio_checksum_verified(zio_t
*zio
);
539 extern int zio_worst_error(int e1
, int e2
);
541 extern enum zio_checksum
zio_checksum_select(enum zio_checksum child
,
542 enum zio_checksum parent
);
543 extern enum zio_checksum
zio_checksum_dedup_select(spa_t
*spa
,
544 enum zio_checksum child
, enum zio_checksum parent
);
545 extern enum zio_compress
zio_compress_select(spa_t
*spa
,
546 enum zio_compress child
, enum zio_compress parent
);
548 extern void zio_suspend(spa_t
*spa
, zio_t
*zio
);
549 extern int zio_resume(spa_t
*spa
);
550 extern void zio_resume_wait(spa_t
*spa
);
553 * Initial setup and teardown.
555 extern void zio_init(void);
556 extern void zio_fini(void);
561 struct zinject_record
;
562 extern uint32_t zio_injection_enabled
;
563 extern int zio_inject_fault(char *name
, int flags
, int *id
,
564 struct zinject_record
*record
);
565 extern int zio_inject_list_next(int *id
, char *name
, size_t buflen
,
566 struct zinject_record
*record
);
567 extern int zio_clear_fault(int id
);
568 extern void zio_handle_panic_injection(spa_t
*spa
, char *tag
, uint64_t type
);
569 extern int zio_handle_fault_injection(zio_t
*zio
, int error
);
570 extern int zio_handle_device_injection(vdev_t
*vd
, zio_t
*zio
, int error
);
571 extern int zio_handle_label_injection(zio_t
*zio
, int error
);
572 extern void zio_handle_ignored_writes(zio_t
*zio
);
573 extern hrtime_t
zio_handle_io_delay(zio_t
*zio
);
576 * Checksum ereport functions
578 extern void zfs_ereport_start_checksum(spa_t
*spa
, vdev_t
*vd
, struct zio
*zio
,
579 uint64_t offset
, uint64_t length
, void *arg
, struct zio_bad_cksum
*info
);
580 extern void zfs_ereport_finish_checksum(zio_cksum_report_t
*report
,
581 const void *good_data
, const void *bad_data
, boolean_t drop_if_identical
);
583 extern void zfs_ereport_send_interim_checksum(zio_cksum_report_t
*report
);
584 extern void zfs_ereport_free_checksum(zio_cksum_report_t
*report
);
586 /* If we have the good data in hand, this function can be used */
587 extern void zfs_ereport_post_checksum(spa_t
*spa
, vdev_t
*vd
,
588 struct zio
*zio
, uint64_t offset
, uint64_t length
,
589 const void *good_data
, const void *bad_data
, struct zio_bad_cksum
*info
);
591 /* Called from spa_sync(), but primarily an injection handler */
592 extern void spa_handle_ignored_writes(spa_t
*spa
);
594 /* zbookmark_phys functions */
595 boolean_t
zbookmark_subtree_completed(const struct dnode_phys
*dnp
,
596 const zbookmark_phys_t
*subtree_root
, const zbookmark_phys_t
*last_block
);
597 int zbookmark_compare(uint16_t dbss1
, uint8_t ibs1
, uint16_t dbss2
,
598 uint8_t ibs2
, const zbookmark_phys_t
*zb1
, const zbookmark_phys_t
*zb2
);