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
126 #define ZIO_COMPRESS_ON_VALUE ZIO_COMPRESS_LZJB
127 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
129 #define BOOTFS_COMPRESS_VALID(compress) \
130 ((compress) == ZIO_COMPRESS_LZJB || \
131 (compress) == ZIO_COMPRESS_LZ4 || \
132 ((compress) == ZIO_COMPRESS_ON && \
133 ZIO_COMPRESS_ON_VALUE == ZIO_COMPRESS_LZJB) || \
134 (compress) == ZIO_COMPRESS_OFF)
137 * Default Linux timeout for a sd device.
139 #define ZIO_DELAY_MAX (30 * MILLISEC)
141 #define ZIO_FAILURE_MODE_WAIT 0
142 #define ZIO_FAILURE_MODE_CONTINUE 1
143 #define ZIO_FAILURE_MODE_PANIC 2
145 typedef enum zio_priority {
146 ZIO_PRIORITY_SYNC_READ,
147 ZIO_PRIORITY_SYNC_WRITE, /* ZIL */
148 ZIO_PRIORITY_ASYNC_READ, /* prefetch */
149 ZIO_PRIORITY_ASYNC_WRITE, /* spa_sync() */
150 ZIO_PRIORITY_SCRUB, /* asynchronous scrub/resilver reads */
151 ZIO_PRIORITY_NUM_QUEUEABLE,
153 ZIO_PRIORITY_NOW /* non-queued i/os (e.g. free) */
156 #define ZIO_PIPELINE_CONTINUE 0x100
157 #define ZIO_PIPELINE_STOP 0x101
161 * Flags inherited by gang, ddt, and vdev children,
162 * and that must be equal for two zios to aggregate
164 ZIO_FLAG_DONT_AGGREGATE = 1 << 0,
165 ZIO_FLAG_IO_REPAIR = 1 << 1,
166 ZIO_FLAG_SELF_HEAL = 1 << 2,
167 ZIO_FLAG_RESILVER = 1 << 3,
168 ZIO_FLAG_SCRUB = 1 << 4,
169 ZIO_FLAG_SCAN_THREAD = 1 << 5,
170 ZIO_FLAG_PHYSICAL = 1 << 6,
172 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
175 * Flags inherited by ddt, gang, and vdev children.
177 ZIO_FLAG_CANFAIL = 1 << 7, /* must be first for INHERIT */
178 ZIO_FLAG_SPECULATIVE = 1 << 8,
179 ZIO_FLAG_CONFIG_WRITER = 1 << 9,
180 ZIO_FLAG_DONT_RETRY = 1 << 10,
181 ZIO_FLAG_DONT_CACHE = 1 << 11,
182 ZIO_FLAG_NODATA = 1 << 12,
183 ZIO_FLAG_INDUCE_DAMAGE = 1 << 13,
185 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
186 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
189 * Flags inherited by vdev children.
191 ZIO_FLAG_IO_RETRY = 1 << 14, /* must be first for INHERIT */
192 ZIO_FLAG_PROBE = 1 << 15,
193 ZIO_FLAG_TRYHARD = 1 << 16,
194 ZIO_FLAG_OPTIONAL = 1 << 17,
196 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
199 * Flags not inherited by any children.
201 ZIO_FLAG_DONT_QUEUE = 1 << 18, /* must be first for INHERIT */
202 ZIO_FLAG_DONT_PROPAGATE = 1 << 19,
203 ZIO_FLAG_IO_BYPASS = 1 << 20,
204 ZIO_FLAG_IO_REWRITE = 1 << 21,
205 ZIO_FLAG_RAW = 1 << 22,
206 ZIO_FLAG_GANG_CHILD = 1 << 23,
207 ZIO_FLAG_DDT_CHILD = 1 << 24,
208 ZIO_FLAG_GODFATHER = 1 << 25,
209 ZIO_FLAG_NOPWRITE = 1 << 26,
210 ZIO_FLAG_REEXECUTED = 1 << 27,
211 ZIO_FLAG_DELEGATED = 1 << 28,
212 ZIO_FLAG_FASTWRITE = 1 << 29,
215 #define ZIO_FLAG_MUSTSUCCEED 0
217 #define ZIO_DDT_CHILD_FLAGS(zio) \
218 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
219 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
221 #define ZIO_GANG_CHILD_FLAGS(zio) \
222 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
223 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
225 #define ZIO_VDEV_CHILD_FLAGS(zio) \
226 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
244 * We'll take the unused errnos, 'EBADE' and 'EBADR' (from the Convergent
245 * graveyard) to indicate checksum errors and fragmentation.
250 typedef void zio_done_func_t(zio_t *zio);
252 extern const char *zio_type_name[ZIO_TYPES];
255 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
256 * identifies any block in the pool. By convention, the meta-objset (MOS)
257 * is objset 0, and the meta-dnode is object 0. This covers all blocks
258 * except root blocks and ZIL blocks, which are defined as follows:
260 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
261 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
262 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
264 * Note: this structure is called a bookmark because its original purpose
265 * was to remember where to resume a pool-wide traverse.
267 * Note: this structure is passed between userland and the kernel, and is
268 * stored on disk (by virtue of being incorporated into other on-disk
269 * structures, e.g. dsl_scan_phys_t).
271 struct zbookmark_phys {
278 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
280 (zb)->zb_objset = objset; \
281 (zb)->zb_object = object; \
282 (zb)->zb_level = level; \
283 (zb)->zb_blkid = blkid; \
286 #define ZB_DESTROYED_OBJSET (-1ULL)
288 #define ZB_ROOT_OBJECT (0ULL)
289 #define ZB_ROOT_LEVEL (-1LL)
290 #define ZB_ROOT_BLKID (0ULL)
292 #define ZB_ZIL_OBJECT (0ULL)
293 #define ZB_ZIL_LEVEL (-2LL)
295 #define ZB_IS_ZERO(zb) \
296 ((zb)->zb_objset == 0 && (zb)->zb_object == 0 && \
297 (zb)->zb_level == 0 && (zb)->zb_blkid == 0)
298 #define ZB_IS_ROOT(zb) \
299 ((zb)->zb_object == ZB_ROOT_OBJECT && \
300 (zb)->zb_level == ZB_ROOT_LEVEL && \
301 (zb)->zb_blkid == ZB_ROOT_BLKID)
303 typedef struct zio_prop {
304 enum zio_checksum zp_checksum;
305 enum zio_compress zp_compress;
306 dmu_object_type_t zp_type;
310 boolean_t zp_dedup_verify;
311 boolean_t zp_nopwrite;
314 typedef struct zio_cksum_report zio_cksum_report_t;
316 typedef void zio_cksum_finish_f(zio_cksum_report_t *rep,
317 const void *good_data);
318 typedef void zio_cksum_free_f(void *cbdata, size_t size);
320 struct zio_bad_cksum; /* defined in zio_checksum.h */
323 struct zio_cksum_report {
324 struct zio_cksum_report *zcr_next;
325 nvlist_t *zcr_ereport;
326 nvlist_t *zcr_detector;
328 size_t zcr_cbinfo; /* passed to zcr_free() */
331 zio_cksum_finish_f *zcr_finish;
332 zio_cksum_free_f *zcr_free;
334 /* internal use only */
335 struct zio_bad_cksum *zcr_ckinfo; /* information from failure */
338 typedef void zio_vsd_cksum_report_f(zio_t *zio, zio_cksum_report_t *zcr,
341 zio_vsd_cksum_report_f zio_vsd_default_cksum_report;
343 typedef struct zio_vsd_ops {
344 zio_done_func_t *vsd_free;
345 zio_vsd_cksum_report_f *vsd_cksum_report;
348 typedef struct zio_gang_node {
349 zio_gbh_phys_t *gn_gbh;
350 struct zio_gang_node *gn_child[SPA_GBH_NBLKPTRS];
353 typedef zio_t *zio_gang_issue_func_t(zio_t *zio, blkptr_t *bp,
354 zio_gang_node_t *gn, void *data);
356 typedef void zio_transform_func_t(zio_t *zio, void *data, uint64_t size);
358 typedef struct zio_transform {
360 uint64_t zt_orig_size;
362 zio_transform_func_t *zt_transform;
363 struct zio_transform *zt_next;
366 typedef int zio_pipe_stage_t(zio_t *zio);
369 * The io_reexecute flags are distinct from io_flags because the child must
370 * be able to propagate them to the parent. The normal io_flags are local
371 * to the zio, not protected by any lock, and not modifiable by children;
372 * the reexecute flags are protected by io_lock, modifiable by children,
373 * and always propagated -- even when ZIO_FLAG_DONT_PROPAGATE is set.
375 #define ZIO_REEXECUTE_NOW 0x01
376 #define ZIO_REEXECUTE_SUSPEND 0x02
378 typedef struct zio_link {
381 list_node_t zl_parent_node;
382 list_node_t zl_child_node;
386 /* Core information about this I/O */
387 zbookmark_phys_t io_bookmark;
390 enum zio_child io_child_type;
392 zio_priority_t io_priority;
393 uint8_t io_reexecute;
394 uint8_t io_state[ZIO_WAIT_TYPES];
398 blkptr_t *io_bp_override;
400 list_t io_parent_list;
401 list_t io_child_list;
402 zio_link_t *io_walk_link;
404 zio_transform_t *io_transform_stack;
407 zio_done_func_t *io_ready;
408 zio_done_func_t *io_physdone;
409 zio_done_func_t *io_done;
411 int64_t io_prev_space_delta; /* DMU private */
414 /* Data represented by this I/O */
418 uint64_t io_orig_size;
420 /* Stuff for the vdev stack */
423 const zio_vsd_ops_t *io_vsd_ops;
426 hrtime_t io_timestamp; /* submitted at */
427 hrtime_t io_delta; /* vdev queue service delta */
428 uint64_t io_delay; /* vdev disk service delta (ticks) */
429 avl_node_t io_queue_node;
431 /* Internal pipeline state */
432 enum zio_flag io_flags;
433 enum zio_stage io_stage;
434 enum zio_stage io_pipeline;
435 enum zio_flag io_orig_flags;
436 enum zio_stage io_orig_stage;
437 enum zio_stage io_orig_pipeline;
439 int io_child_error[ZIO_CHILD_TYPES];
440 uint64_t io_children[ZIO_CHILD_TYPES][ZIO_WAIT_TYPES];
441 uint64_t io_child_count;
442 uint64_t io_phys_children;
443 uint64_t io_parent_count;
445 zio_t *io_gang_leader;
446 zio_gang_node_t *io_gang_tree;
453 zio_cksum_report_t *io_cksum_report;
456 /* Taskq dispatching state */
457 taskq_ent_t io_tqent;
460 extern zio_t *zio_null(zio_t *pio, spa_t *spa, vdev_t *vd,
461 zio_done_func_t *done, void *private, enum zio_flag flags);
463 extern zio_t *zio_root(spa_t *spa,
464 zio_done_func_t *done, void *private, enum zio_flag flags);
466 extern zio_t *zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, void *data,
467 uint64_t size, zio_done_func_t *done, void *private,
468 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb);
470 extern zio_t *zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
471 void *data, uint64_t size, const zio_prop_t *zp,
472 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
474 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb);
476 extern zio_t *zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
477 void *data, uint64_t size, zio_done_func_t *done, void *private,
478 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb);
480 extern void zio_write_override(zio_t *zio, blkptr_t *bp, int copies,
483 extern void zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp);
485 extern zio_t *zio_claim(zio_t *pio, spa_t *spa, uint64_t txg,
487 zio_done_func_t *done, void *private, enum zio_flag flags);
489 extern zio_t *zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
490 zio_done_func_t *done, void *private, enum zio_flag flags);
492 extern zio_t *zio_read_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_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset,
498 uint64_t size, void *data, int checksum,
499 zio_done_func_t *done, void *private, zio_priority_t priority,
500 enum zio_flag flags, boolean_t labels);
502 extern zio_t *zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg,
503 const blkptr_t *bp, enum zio_flag flags);
505 extern int zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp,
506 uint64_t size, boolean_t use_slog);
507 extern void zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp);
508 extern void zio_flush(zio_t *zio, vdev_t *vd);
509 extern void zio_shrink(zio_t *zio, uint64_t size);
511 extern int zio_wait(zio_t *zio);
512 extern void zio_nowait(zio_t *zio);
513 extern void zio_execute(zio_t *zio);
514 extern void zio_interrupt(zio_t *zio);
516 extern zio_t *zio_walk_parents(zio_t *cio);
517 extern zio_t *zio_walk_children(zio_t *pio);
518 extern zio_t *zio_unique_parent(zio_t *cio);
519 extern void zio_add_child(zio_t *pio, zio_t *cio);
521 extern void *zio_buf_alloc(size_t size);
522 extern void zio_buf_free(void *buf, size_t size);
523 extern void *zio_data_buf_alloc(size_t size);
524 extern void zio_data_buf_free(void *buf, size_t size);
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(enum zio_compress child,
549 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 uint64_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_send_interim_checksum(zio_cksum_report_t *report);
587 extern void zfs_ereport_free_checksum(zio_cksum_report_t *report);
589 /* If we have the good data in hand, this function can be used */
590 extern void zfs_ereport_post_checksum(spa_t *spa, vdev_t *vd,
591 struct zio *zio, uint64_t offset, uint64_t length,
592 const void *good_data, const void *bad_data, struct zio_bad_cksum *info);
594 /* Called from spa_sync(), but primarily an injection handler */
595 extern void spa_handle_ignored_writes(spa_t *spa);
597 /* zbookmark_phys functions */
598 boolean_t zbookmark_is_before(const struct dnode_phys *dnp,
599 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2);