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, 2016 by Delphix. All rights reserved.
26 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
32 #include <sys/zio_priority.h>
33 #include <sys/zfs_context.h>
37 #include <sys/kstat.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,
90 ZIO_CHECKSUM_FUNCTIONS
94 * The number of "legacy" compression functions which can be set on individual
97 #define ZIO_CHECKSUM_LEGACY_FUNCTIONS ZIO_CHECKSUM_ZILOG2
99 #define ZIO_CHECKSUM_ON_VALUE ZIO_CHECKSUM_FLETCHER_4
100 #define ZIO_CHECKSUM_DEFAULT ZIO_CHECKSUM_ON
102 #define ZIO_CHECKSUM_MASK 0xffULL
103 #define ZIO_CHECKSUM_VERIFY (1 << 8)
105 #define ZIO_DEDUPCHECKSUM ZIO_CHECKSUM_SHA256
106 #define ZIO_DEDUPDITTO_MIN 100
109 ZIO_COMPRESS_INHERIT = 0,
125 ZIO_COMPRESS_FUNCTIONS
129 * The number of "legacy" compression functions which can be set on individual
132 #define ZIO_COMPRESS_LEGACY_FUNCTIONS ZIO_COMPRESS_LZ4
135 * The meaning of "compress = on" selected by the compression features enabled
138 #define ZIO_COMPRESS_LEGACY_ON_VALUE ZIO_COMPRESS_LZJB
139 #define ZIO_COMPRESS_LZ4_ON_VALUE ZIO_COMPRESS_LZ4
141 #define ZIO_COMPRESS_DEFAULT ZIO_COMPRESS_OFF
143 #define BOOTFS_COMPRESS_VALID(compress) \
144 ((compress) == ZIO_COMPRESS_LZJB || \
145 (compress) == ZIO_COMPRESS_LZ4 || \
146 (compress) == ZIO_COMPRESS_ON || \
147 (compress) == ZIO_COMPRESS_OFF)
149 #define ZIO_FAILURE_MODE_WAIT 0
150 #define ZIO_FAILURE_MODE_CONTINUE 1
151 #define ZIO_FAILURE_MODE_PANIC 2
155 * Flags inherited by gang, ddt, and vdev children,
156 * and that must be equal for two zios to aggregate
158 ZIO_FLAG_DONT_AGGREGATE = 1 << 0,
159 ZIO_FLAG_IO_REPAIR = 1 << 1,
160 ZIO_FLAG_SELF_HEAL = 1 << 2,
161 ZIO_FLAG_RESILVER = 1 << 3,
162 ZIO_FLAG_SCRUB = 1 << 4,
163 ZIO_FLAG_SCAN_THREAD = 1 << 5,
164 ZIO_FLAG_PHYSICAL = 1 << 6,
166 #define ZIO_FLAG_AGG_INHERIT (ZIO_FLAG_CANFAIL - 1)
169 * Flags inherited by ddt, gang, and vdev children.
171 ZIO_FLAG_CANFAIL = 1 << 7, /* must be first for INHERIT */
172 ZIO_FLAG_SPECULATIVE = 1 << 8,
173 ZIO_FLAG_CONFIG_WRITER = 1 << 9,
174 ZIO_FLAG_DONT_RETRY = 1 << 10,
175 ZIO_FLAG_DONT_CACHE = 1 << 11,
176 ZIO_FLAG_NODATA = 1 << 12,
177 ZIO_FLAG_INDUCE_DAMAGE = 1 << 13,
178 ZIO_FLAG_IO_ALLOCATING = 1 << 14,
180 #define ZIO_FLAG_DDT_INHERIT (ZIO_FLAG_IO_RETRY - 1)
181 #define ZIO_FLAG_GANG_INHERIT (ZIO_FLAG_IO_RETRY - 1)
184 * Flags inherited by vdev children.
186 ZIO_FLAG_IO_RETRY = 1 << 15, /* must be first for INHERIT */
187 ZIO_FLAG_PROBE = 1 << 16,
188 ZIO_FLAG_TRYHARD = 1 << 17,
189 ZIO_FLAG_OPTIONAL = 1 << 18,
191 #define ZIO_FLAG_VDEV_INHERIT (ZIO_FLAG_DONT_QUEUE - 1)
194 * Flags not inherited by any children.
196 ZIO_FLAG_DONT_QUEUE = 1 << 19, /* must be first for INHERIT */
197 ZIO_FLAG_DONT_PROPAGATE = 1 << 20,
198 ZIO_FLAG_IO_BYPASS = 1 << 21,
199 ZIO_FLAG_IO_REWRITE = 1 << 22,
200 ZIO_FLAG_RAW = 1 << 23,
201 ZIO_FLAG_GANG_CHILD = 1 << 24,
202 ZIO_FLAG_DDT_CHILD = 1 << 25,
203 ZIO_FLAG_GODFATHER = 1 << 26,
204 ZIO_FLAG_NOPWRITE = 1 << 27,
205 ZIO_FLAG_REEXECUTED = 1 << 28,
206 ZIO_FLAG_DELEGATED = 1 << 29,
209 #define ZIO_FLAG_MUSTSUCCEED 0
211 #define ZIO_DDT_CHILD_FLAGS(zio) \
212 (((zio)->io_flags & ZIO_FLAG_DDT_INHERIT) | \
213 ZIO_FLAG_DDT_CHILD | ZIO_FLAG_CANFAIL)
215 #define ZIO_GANG_CHILD_FLAGS(zio) \
216 (((zio)->io_flags & ZIO_FLAG_GANG_INHERIT) | \
217 ZIO_FLAG_GANG_CHILD | ZIO_FLAG_CANFAIL)
219 #define ZIO_VDEV_CHILD_FLAGS(zio) \
220 (((zio)->io_flags & ZIO_FLAG_VDEV_INHERIT) | \
238 * We'll take the number 122 and 123 to indicate checksum errors and
239 * fragmentation. Those doesn't collide with any errno values as they
240 * are greater than ELAST.
245 typedef void zio_done_func_t(zio_t *zio);
247 extern boolean_t zio_dva_throttle_enabled;
248 extern const char *zio_type_name[ZIO_TYPES];
251 * A bookmark is a four-tuple <objset, object, level, blkid> that uniquely
252 * identifies any block in the pool. By convention, the meta-objset (MOS)
253 * is objset 0, and the meta-dnode is object 0. This covers all blocks
254 * except root blocks and ZIL blocks, which are defined as follows:
256 * Root blocks (objset_phys_t) are object 0, level -1: <objset, 0, -1, 0>.
257 * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
258 * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
259 * dnode visit bookmarks are <objset, object id of dnode, -3, 0>.
261 * Note: this structure is called a bookmark because its original purpose
262 * was to remember where to resume a pool-wide traverse.
264 * Note: this structure is passed between userland and the kernel, and is
265 * stored on disk (by virtue of being incorporated into other on-disk
266 * structures, e.g. dsl_scan_phys_t).
268 typedef struct zbookmark_phys {
275 #define SET_BOOKMARK(zb, objset, object, level, blkid) \
277 (zb)->zb_objset = objset; \
278 (zb)->zb_object = object; \
279 (zb)->zb_level = level; \
280 (zb)->zb_blkid = blkid; \
283 #define ZB_DESTROYED_OBJSET (-1ULL)
285 #define ZB_ROOT_OBJECT (0ULL)
286 #define ZB_ROOT_LEVEL (-1LL)
287 #define ZB_ROOT_BLKID (0ULL)
289 #define ZB_ZIL_OBJECT (0ULL)
290 #define ZB_ZIL_LEVEL (-2LL)
292 #define ZB_DNODE_LEVEL (-3LL)
293 #define ZB_DNODE_BLKID (0ULL)
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 * Used for TRIM kstat.
388 typedef struct zio_trim_stats {
390 * Number of bytes successfully TRIMmed.
395 * Number of successful TRIM requests.
397 kstat_named_t success;
400 * Number of TRIM requests that failed because TRIM is not
403 kstat_named_t unsupported;
406 * Number of TRIM requests that failed for other reasons.
408 kstat_named_t failed;
411 extern zio_trim_stats_t zio_trim_stats;
413 #define ZIO_TRIM_STAT_INCR(stat, val) \
414 atomic_add_64(&zio_trim_stats.stat.value.ui64, (val));
415 #define ZIO_TRIM_STAT_BUMP(stat) \
416 ZIO_TRIM_STAT_INCR(stat, 1);
419 /* Core information about this I/O */
420 zbookmark_phys_t io_bookmark;
423 enum zio_child io_child_type;
425 zio_priority_t io_priority;
426 uint8_t io_reexecute;
427 uint8_t io_state[ZIO_WAIT_TYPES];
431 blkptr_t *io_bp_override;
433 list_t io_parent_list;
434 list_t io_child_list;
436 zio_transform_t *io_transform_stack;
439 zio_done_func_t *io_ready;
440 zio_done_func_t *io_children_ready;
441 zio_done_func_t *io_physdone;
442 zio_done_func_t *io_done;
444 int64_t io_prev_space_delta; /* DMU private */
447 /* Data represented by this I/O */
451 uint64_t io_orig_size;
453 /* Stuff for the vdev stack */
456 const zio_vsd_ops_t *io_vsd_ops;
459 hrtime_t io_timestamp;
460 hrtime_t io_queued_timestamp;
461 hrtime_t io_target_timestamp;
462 avl_node_t io_queue_node;
463 avl_node_t io_offset_node;
464 avl_node_t io_alloc_node;
466 /* Internal pipeline state */
467 enum zio_flag io_flags;
468 enum zio_stage io_stage;
469 enum zio_stage io_pipeline;
470 enum zio_flag io_orig_flags;
471 enum zio_stage io_orig_stage;
472 enum zio_stage io_orig_pipeline;
473 enum zio_stage io_pipeline_trace;
475 int io_child_error[ZIO_CHILD_TYPES];
476 uint64_t io_children[ZIO_CHILD_TYPES][ZIO_WAIT_TYPES];
477 uint64_t io_child_count;
478 uint64_t io_phys_children;
479 uint64_t io_parent_count;
481 zio_t *io_gang_leader;
482 zio_gang_node_t *io_gang_tree;
489 zio_cksum_report_t *io_cksum_report;
492 /* Taskq dispatching state */
493 taskq_ent_t io_tqent;
495 avl_node_t io_trim_node;
496 list_node_t io_trim_link;
499 extern int zio_timestamp_compare(const void *, const void *);
501 extern zio_t *zio_null(zio_t *pio, spa_t *spa, vdev_t *vd,
502 zio_done_func_t *done, void *priv, enum zio_flag flags);
504 extern zio_t *zio_root(spa_t *spa,
505 zio_done_func_t *done, void *priv, enum zio_flag flags);
507 extern zio_t *zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, void *data,
508 uint64_t size, zio_done_func_t *done, void *priv,
509 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb);
511 extern zio_t *zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
512 void *data, uint64_t size, const zio_prop_t *zp,
513 zio_done_func_t *ready, zio_done_func_t *children_ready,
514 zio_done_func_t *physdone, zio_done_func_t *done,
515 void *priv, zio_priority_t priority, enum zio_flag flags,
516 const zbookmark_phys_t *zb);
518 extern zio_t *zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
519 void *data, uint64_t size, zio_done_func_t *done, void *priv,
520 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb);
522 extern void zio_write_override(zio_t *zio, blkptr_t *bp, int copies,
525 extern void zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp);
527 extern zio_t *zio_claim(zio_t *pio, spa_t *spa, uint64_t txg,
529 zio_done_func_t *done, void *priv, enum zio_flag flags);
531 extern zio_t *zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
532 uint64_t offset, uint64_t size, zio_done_func_t *done, void *priv,
533 zio_priority_t priority, enum zio_flag flags);
535 extern zio_t *zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset,
536 uint64_t size, void *data, int checksum,
537 zio_done_func_t *done, void *priv, zio_priority_t priority,
538 enum zio_flag flags, boolean_t labels);
540 extern zio_t *zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset,
541 uint64_t size, void *data, int checksum,
542 zio_done_func_t *done, void *priv, zio_priority_t priority,
543 enum zio_flag flags, boolean_t labels);
545 extern zio_t *zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg,
546 const blkptr_t *bp, uint64_t size, enum zio_flag flags);
548 extern int zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp,
549 blkptr_t *old_bp, uint64_t size, boolean_t *slog);
550 extern void zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp);
551 extern void zio_flush(zio_t *zio, vdev_t *vd);
552 extern zio_t *zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset,
554 extern void zio_shrink(zio_t *zio, uint64_t size);
556 extern int zio_wait(zio_t *zio);
557 extern void zio_nowait(zio_t *zio);
558 extern void zio_execute(zio_t *zio);
559 extern void zio_interrupt(zio_t *zio);
560 extern void zio_delay_init(zio_t *zio);
561 extern void zio_delay_interrupt(zio_t *zio);
563 extern zio_t *zio_walk_parents(zio_t *cio, zio_link_t **);
564 extern zio_t *zio_walk_children(zio_t *pio, zio_link_t **);
565 extern zio_t *zio_unique_parent(zio_t *cio);
566 extern void zio_add_child(zio_t *pio, zio_t *cio);
568 extern void *zio_buf_alloc(size_t size);
569 extern void *zio_buf_alloc_nowait(size_t size);
570 extern void zio_buf_free(void *buf, size_t size);
571 extern void *zio_data_buf_alloc(size_t size);
572 extern void zio_data_buf_free(void *buf, size_t size);
574 extern void zio_push_transform(zio_t *zio, void *data, uint64_t size,
575 uint64_t bufsize, zio_transform_func_t *transform);
576 extern void zio_pop_transforms(zio_t *zio);
578 extern void zio_resubmit_stage_async(void *);
580 extern zio_t *zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd,
581 uint64_t offset, void *data, uint64_t size, int type,
582 zio_priority_t priority, enum zio_flag flags,
583 zio_done_func_t *done, void *priv);
585 extern zio_t *zio_vdev_delegated_io(vdev_t *vd, uint64_t offset,
586 void *data, uint64_t size, int type, zio_priority_t priority,
587 enum zio_flag flags, zio_done_func_t *done, void *priv);
589 extern void zio_vdev_io_bypass(zio_t *zio);
590 extern void zio_vdev_io_reissue(zio_t *zio);
591 extern void zio_vdev_io_redone(zio_t *zio);
593 extern void zio_checksum_verified(zio_t *zio);
594 extern int zio_worst_error(int e1, int e2);
596 extern enum zio_checksum zio_checksum_select(enum zio_checksum child,
597 enum zio_checksum parent);
598 extern enum zio_checksum zio_checksum_dedup_select(spa_t *spa,
599 enum zio_checksum child, enum zio_checksum parent);
600 extern enum zio_compress zio_compress_select(spa_t *spa,
601 enum zio_compress child, enum zio_compress parent);
603 extern void zio_suspend(spa_t *spa, zio_t *zio);
604 extern int zio_resume(spa_t *spa);
605 extern void zio_resume_wait(spa_t *spa);
608 * Initial setup and teardown.
610 extern void zio_init(void);
611 extern void zio_fini(void);
616 struct zinject_record;
617 extern uint32_t zio_injection_enabled;
618 extern int zio_inject_fault(char *name, int flags, int *id,
619 struct zinject_record *record);
620 extern int zio_inject_list_next(int *id, char *name, size_t buflen,
621 struct zinject_record *record);
622 extern int zio_clear_fault(int id);
623 extern void zio_handle_panic_injection(spa_t *spa, char *tag, uint64_t type);
624 extern int zio_handle_fault_injection(zio_t *zio, int error);
625 extern int zio_handle_device_injection(vdev_t *vd, zio_t *zio, int error);
626 extern int zio_handle_label_injection(zio_t *zio, int error);
627 extern void zio_handle_ignored_writes(zio_t *zio);
628 extern hrtime_t zio_handle_io_delay(zio_t *zio);
631 * Checksum ereport functions
633 extern void zfs_ereport_start_checksum(spa_t *spa, vdev_t *vd, struct zio *zio,
634 uint64_t offset, uint64_t length, void *arg, struct zio_bad_cksum *info);
635 extern void zfs_ereport_finish_checksum(zio_cksum_report_t *report,
636 const void *good_data, const void *bad_data, boolean_t drop_if_identical);
638 extern void zfs_ereport_send_interim_checksum(zio_cksum_report_t *report);
639 extern void zfs_ereport_free_checksum(zio_cksum_report_t *report);
641 /* If we have the good data in hand, this function can be used */
642 extern void zfs_ereport_post_checksum(spa_t *spa, vdev_t *vd,
643 struct zio *zio, uint64_t offset, uint64_t length,
644 const void *good_data, const void *bad_data, struct zio_bad_cksum *info);
646 /* Called from spa_sync(), but primarily an injection handler */
647 extern void spa_handle_ignored_writes(spa_t *spa);
649 /* zbookmark_phys functions */
650 boolean_t zbookmark_subtree_completed(const struct dnode_phys *dnp,
651 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block);
652 int zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2,
653 uint8_t ibs2, const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2);