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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_vdev_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 int zio_bulk_flags = 0;
80 int zio_delay_max = ZIO_DELAY_MAX;
82 extern int zfs_mg_alloc_failures;
85 * The following actions directly effect the spa's sync-to-convergence logic.
86 * The values below define the sync pass when we start performing the action.
87 * Care should be taken when changing these values as they directly impact
88 * spa_sync() performance. Tuning these values may introduce subtle performance
89 * pathologies and should only be done in the context of performance analysis.
90 * These tunables will eventually be removed and replaced with #defines once
91 * enough analysis has been done to determine optimal values.
93 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
94 * regular blocks are not deferred.
96 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
97 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
98 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
101 * An allocating zio is one that either currently has the DVA allocate
102 * stage set or will have it later in its lifetime.
104 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
106 int zio_requeue_io_start_cut_in_line = 1;
109 int zio_buf_debug_limit = 16384;
111 int zio_buf_debug_limit = 0;
114 static inline void __zio_execute(zio_t *zio);
117 zio_cons(void *arg, void *unused, int kmflag)
121 bzero(zio, sizeof (zio_t));
123 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
124 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
126 list_create(&zio->io_parent_list, sizeof (zio_link_t),
127 offsetof(zio_link_t, zl_parent_node));
128 list_create(&zio->io_child_list, sizeof (zio_link_t),
129 offsetof(zio_link_t, zl_child_node));
135 zio_dest(void *arg, void *unused)
139 mutex_destroy(&zio->io_lock);
140 cv_destroy(&zio->io_cv);
141 list_destroy(&zio->io_parent_list);
142 list_destroy(&zio->io_child_list);
149 vmem_t *data_alloc_arena = NULL;
151 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
152 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
153 zio_link_cache = kmem_cache_create("zio_link_cache",
154 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
155 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
156 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
159 * For small buffers, we want a cache for each multiple of
160 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
161 * for each quarter-power of 2. For large buffers, we want
162 * a cache for each multiple of PAGESIZE.
164 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
165 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
169 while (p2 & (p2 - 1))
172 if (size <= 4 * SPA_MINBLOCKSIZE) {
173 align = SPA_MINBLOCKSIZE;
174 } else if (P2PHASE(size, PAGESIZE) == 0) {
176 } else if (P2PHASE(size, p2 >> 2) == 0) {
182 int flags = zio_bulk_flags;
185 * The smallest buffers (512b) are heavily used and
186 * experience a lot of churn. The slabs allocated
187 * for them are also relatively small (32K). Thus
188 * in over to avoid expensive calls to vmalloc() we
189 * make an exception to the usual slab allocation
190 * policy and force these buffers to be kmem backed.
192 if (size == (1 << SPA_MINBLOCKSHIFT))
195 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
196 zio_buf_cache[c] = kmem_cache_create(name, size,
197 align, NULL, NULL, NULL, NULL, NULL, flags);
199 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
200 zio_data_buf_cache[c] = kmem_cache_create(name, size,
201 align, NULL, NULL, NULL, NULL,
202 data_alloc_arena, flags);
207 ASSERT(zio_buf_cache[c] != NULL);
208 if (zio_buf_cache[c - 1] == NULL)
209 zio_buf_cache[c - 1] = zio_buf_cache[c];
211 ASSERT(zio_data_buf_cache[c] != NULL);
212 if (zio_data_buf_cache[c - 1] == NULL)
213 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
217 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
218 * to fail 3 times per txg or 8 failures, whichever is greater.
220 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
231 kmem_cache_t *last_cache = NULL;
232 kmem_cache_t *last_data_cache = NULL;
234 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
235 if (zio_buf_cache[c] != last_cache) {
236 last_cache = zio_buf_cache[c];
237 kmem_cache_destroy(zio_buf_cache[c]);
239 zio_buf_cache[c] = NULL;
241 if (zio_data_buf_cache[c] != last_data_cache) {
242 last_data_cache = zio_data_buf_cache[c];
243 kmem_cache_destroy(zio_data_buf_cache[c]);
245 zio_data_buf_cache[c] = NULL;
248 kmem_cache_destroy(zio_vdev_cache);
249 kmem_cache_destroy(zio_link_cache);
250 kmem_cache_destroy(zio_cache);
258 * ==========================================================================
259 * Allocate and free I/O buffers
260 * ==========================================================================
264 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
265 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
266 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
267 * excess / transient data in-core during a crashdump.
270 zio_buf_alloc(size_t size)
272 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
274 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
276 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
280 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
281 * crashdump if the kernel panics. This exists so that we will limit the amount
282 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
283 * of kernel heap dumped to disk when the kernel panics)
286 zio_data_buf_alloc(size_t size)
288 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
290 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
292 return (kmem_cache_alloc(zio_data_buf_cache[c],
293 KM_PUSHPAGE | KM_NODEBUG));
297 zio_buf_free(void *buf, size_t size)
299 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
301 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 kmem_cache_free(zio_buf_cache[c], buf);
307 zio_data_buf_free(void *buf, size_t size)
309 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
311 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
313 kmem_cache_free(zio_data_buf_cache[c], buf);
317 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
318 * or significantly delays the issuing of a zio. These buffers are used
319 * to aggregate I/O and could be used for raidz stripes.
324 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
328 zio_vdev_free(void *buf)
330 kmem_cache_free(zio_vdev_cache, buf);
335 * ==========================================================================
336 * Push and pop I/O transform buffers
337 * ==========================================================================
340 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
341 zio_transform_func_t *transform)
343 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
345 zt->zt_orig_data = zio->io_data;
346 zt->zt_orig_size = zio->io_size;
347 zt->zt_bufsize = bufsize;
348 zt->zt_transform = transform;
350 zt->zt_next = zio->io_transform_stack;
351 zio->io_transform_stack = zt;
358 zio_pop_transforms(zio_t *zio)
362 while ((zt = zio->io_transform_stack) != NULL) {
363 if (zt->zt_transform != NULL)
364 zt->zt_transform(zio,
365 zt->zt_orig_data, zt->zt_orig_size);
367 if (zt->zt_bufsize != 0)
368 zio_buf_free(zio->io_data, zt->zt_bufsize);
370 zio->io_data = zt->zt_orig_data;
371 zio->io_size = zt->zt_orig_size;
372 zio->io_transform_stack = zt->zt_next;
374 kmem_free(zt, sizeof (zio_transform_t));
379 * ==========================================================================
380 * I/O transform callbacks for subblocks and decompression
381 * ==========================================================================
384 zio_subblock(zio_t *zio, void *data, uint64_t size)
386 ASSERT(zio->io_size > size);
388 if (zio->io_type == ZIO_TYPE_READ)
389 bcopy(zio->io_data, data, size);
393 zio_decompress(zio_t *zio, void *data, uint64_t size)
395 if (zio->io_error == 0 &&
396 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
397 zio->io_data, data, zio->io_size, size) != 0)
402 * ==========================================================================
403 * I/O parent/child relationships and pipeline interlocks
404 * ==========================================================================
407 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
408 * continue calling these functions until they return NULL.
409 * Otherwise, the next caller will pick up the list walk in
410 * some indeterminate state. (Otherwise every caller would
411 * have to pass in a cookie to keep the state represented by
412 * io_walk_link, which gets annoying.)
415 zio_walk_parents(zio_t *cio)
417 zio_link_t *zl = cio->io_walk_link;
418 list_t *pl = &cio->io_parent_list;
420 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
421 cio->io_walk_link = zl;
426 ASSERT(zl->zl_child == cio);
427 return (zl->zl_parent);
431 zio_walk_children(zio_t *pio)
433 zio_link_t *zl = pio->io_walk_link;
434 list_t *cl = &pio->io_child_list;
436 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
437 pio->io_walk_link = zl;
442 ASSERT(zl->zl_parent == pio);
443 return (zl->zl_child);
447 zio_unique_parent(zio_t *cio)
449 zio_t *pio = zio_walk_parents(cio);
451 VERIFY(zio_walk_parents(cio) == NULL);
456 zio_add_child(zio_t *pio, zio_t *cio)
458 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
462 * Logical I/Os can have logical, gang, or vdev children.
463 * Gang I/Os can have gang or vdev children.
464 * Vdev I/Os can only have vdev children.
465 * The following ASSERT captures all of these constraints.
467 ASSERT(cio->io_child_type <= pio->io_child_type);
472 mutex_enter(&cio->io_lock);
473 mutex_enter(&pio->io_lock);
475 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
477 for (w = 0; w < ZIO_WAIT_TYPES; w++)
478 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
480 list_insert_head(&pio->io_child_list, zl);
481 list_insert_head(&cio->io_parent_list, zl);
483 pio->io_child_count++;
484 cio->io_parent_count++;
486 mutex_exit(&pio->io_lock);
487 mutex_exit(&cio->io_lock);
491 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
493 ASSERT(zl->zl_parent == pio);
494 ASSERT(zl->zl_child == cio);
496 mutex_enter(&cio->io_lock);
497 mutex_enter(&pio->io_lock);
499 list_remove(&pio->io_child_list, zl);
500 list_remove(&cio->io_parent_list, zl);
502 pio->io_child_count--;
503 cio->io_parent_count--;
505 mutex_exit(&pio->io_lock);
506 mutex_exit(&cio->io_lock);
508 kmem_cache_free(zio_link_cache, zl);
512 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
514 uint64_t *countp = &zio->io_children[child][wait];
515 boolean_t waiting = B_FALSE;
517 mutex_enter(&zio->io_lock);
518 ASSERT(zio->io_stall == NULL);
521 zio->io_stall = countp;
524 mutex_exit(&zio->io_lock);
529 __attribute__((always_inline))
531 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
533 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
534 int *errorp = &pio->io_child_error[zio->io_child_type];
536 mutex_enter(&pio->io_lock);
537 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
538 *errorp = zio_worst_error(*errorp, zio->io_error);
539 pio->io_reexecute |= zio->io_reexecute;
540 ASSERT3U(*countp, >, 0);
541 if (--*countp == 0 && pio->io_stall == countp) {
542 pio->io_stall = NULL;
543 mutex_exit(&pio->io_lock);
546 mutex_exit(&pio->io_lock);
551 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
553 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
554 zio->io_error = zio->io_child_error[c];
558 * ==========================================================================
559 * Create the various types of I/O (read, write, free, etc)
560 * ==========================================================================
563 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
564 void *data, uint64_t size, zio_done_func_t *done, void *private,
565 zio_type_t type, int priority, enum zio_flag flags,
566 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
567 enum zio_stage stage, enum zio_stage pipeline)
571 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
572 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
573 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
575 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
576 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
577 ASSERT(vd || stage == ZIO_STAGE_OPEN);
579 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
582 zio->io_child_type = ZIO_CHILD_VDEV;
583 else if (flags & ZIO_FLAG_GANG_CHILD)
584 zio->io_child_type = ZIO_CHILD_GANG;
585 else if (flags & ZIO_FLAG_DDT_CHILD)
586 zio->io_child_type = ZIO_CHILD_DDT;
588 zio->io_child_type = ZIO_CHILD_LOGICAL;
591 zio->io_logical = NULL;
592 zio->io_bp = (blkptr_t *)bp;
593 zio->io_bp_copy = *bp;
594 zio->io_bp_orig = *bp;
595 if (type != ZIO_TYPE_WRITE ||
596 zio->io_child_type == ZIO_CHILD_DDT)
597 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
598 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
599 zio->io_logical = zio;
600 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
601 pipeline |= ZIO_GANG_STAGES;
603 zio->io_logical = NULL;
605 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
606 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
611 zio->io_ready = NULL;
613 zio->io_private = private;
614 zio->io_prev_space_delta = 0;
616 zio->io_priority = priority;
619 zio->io_vsd_ops = NULL;
620 zio->io_offset = offset;
621 zio->io_deadline = 0;
622 zio->io_timestamp = 0;
625 zio->io_orig_data = zio->io_data = data;
626 zio->io_orig_size = zio->io_size = size;
627 zio->io_orig_flags = zio->io_flags = flags;
628 zio->io_orig_stage = zio->io_stage = stage;
629 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
630 bzero(&zio->io_prop, sizeof (zio_prop_t));
632 zio->io_reexecute = 0;
633 zio->io_bp_override = NULL;
634 zio->io_walk_link = NULL;
635 zio->io_transform_stack = NULL;
637 zio->io_child_count = 0;
638 zio->io_parent_count = 0;
639 zio->io_stall = NULL;
640 zio->io_gang_leader = NULL;
641 zio->io_gang_tree = NULL;
642 zio->io_executor = NULL;
643 zio->io_waiter = NULL;
644 zio->io_cksum_report = NULL;
646 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
647 bzero(zio->io_children,
648 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
649 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
651 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
652 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
655 zio->io_bookmark = *zb;
658 if (zio->io_logical == NULL)
659 zio->io_logical = pio->io_logical;
660 if (zio->io_child_type == ZIO_CHILD_GANG)
661 zio->io_gang_leader = pio->io_gang_leader;
662 zio_add_child(pio, zio);
665 taskq_init_ent(&zio->io_tqent);
671 zio_destroy(zio_t *zio)
673 kmem_cache_free(zio_cache, zio);
677 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
678 void *private, enum zio_flag flags)
682 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
683 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
684 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
690 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
692 return (zio_null(NULL, spa, NULL, done, private, flags));
696 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
697 void *data, uint64_t size, zio_done_func_t *done, void *private,
698 int priority, enum zio_flag flags, const zbookmark_t *zb)
702 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
703 data, size, done, private,
704 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
705 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
706 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
712 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
713 void *data, uint64_t size, const zio_prop_t *zp,
714 zio_done_func_t *ready, zio_done_func_t *done, void *private,
715 int priority, enum zio_flag flags, const zbookmark_t *zb)
719 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
720 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
721 zp->zp_compress >= ZIO_COMPRESS_OFF &&
722 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
723 DMU_OT_IS_VALID(zp->zp_type) &&
726 zp->zp_copies <= spa_max_replication(spa) &&
728 zp->zp_dedup_verify <= 1);
730 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
731 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
732 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
733 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
735 zio->io_ready = ready;
742 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
743 uint64_t size, zio_done_func_t *done, void *private, int priority,
744 enum zio_flag flags, zbookmark_t *zb)
748 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
749 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
750 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
756 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
758 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
759 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
760 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
761 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
763 zio->io_prop.zp_copies = copies;
764 zio->io_bp_override = bp;
768 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
774 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 dprintf_bp(bp, "freeing in txg %llu, pass %u",
780 (longlong_t)txg, spa->spa_sync_pass);
782 ASSERT(!BP_IS_HOLE(bp));
783 ASSERT(spa_syncing_txg(spa) == txg);
784 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
788 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
789 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
790 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
796 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
797 zio_done_func_t *done, void *private, enum zio_flag flags)
802 * A claim is an allocation of a specific block. Claims are needed
803 * to support immediate writes in the intent log. The issue is that
804 * immediate writes contain committed data, but in a txg that was
805 * *not* committed. Upon opening the pool after an unclean shutdown,
806 * the intent log claims all blocks that contain immediate write data
807 * so that the SPA knows they're in use.
809 * All claims *must* be resolved in the first txg -- before the SPA
810 * starts allocating blocks -- so that nothing is allocated twice.
811 * If txg == 0 we just verify that the block is claimable.
813 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
814 ASSERT(txg == spa_first_txg(spa) || txg == 0);
815 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
817 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
818 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
819 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
825 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
826 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
831 if (vd->vdev_children == 0) {
832 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
833 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
834 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
838 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
840 for (c = 0; c < vd->vdev_children; c++)
841 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
842 done, private, priority, flags));
849 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
850 void *data, int checksum, zio_done_func_t *done, void *private,
851 int priority, enum zio_flag flags, boolean_t labels)
855 ASSERT(vd->vdev_children == 0);
856 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
857 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
858 ASSERT3U(offset + size, <=, vd->vdev_psize);
860 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
861 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
862 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
864 zio->io_prop.zp_checksum = checksum;
870 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
871 void *data, int checksum, zio_done_func_t *done, void *private,
872 int priority, enum zio_flag flags, boolean_t labels)
876 ASSERT(vd->vdev_children == 0);
877 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
878 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
879 ASSERT3U(offset + size, <=, vd->vdev_psize);
881 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
882 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
883 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
885 zio->io_prop.zp_checksum = checksum;
887 if (zio_checksum_table[checksum].ci_eck) {
889 * zec checksums are necessarily destructive -- they modify
890 * the end of the write buffer to hold the verifier/checksum.
891 * Therefore, we must make a local copy in case the data is
892 * being written to multiple places in parallel.
894 void *wbuf = zio_buf_alloc(size);
895 bcopy(data, wbuf, size);
896 zio_push_transform(zio, wbuf, size, size, NULL);
903 * Create a child I/O to do some work for us.
906 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
907 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
908 zio_done_func_t *done, void *private)
910 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
913 ASSERT(vd->vdev_parent ==
914 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
916 if (type == ZIO_TYPE_READ && bp != NULL) {
918 * If we have the bp, then the child should perform the
919 * checksum and the parent need not. This pushes error
920 * detection as close to the leaves as possible and
921 * eliminates redundant checksums in the interior nodes.
923 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
924 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
927 if (vd->vdev_children == 0)
928 offset += VDEV_LABEL_START_SIZE;
930 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
933 * If we've decided to do a repair, the write is not speculative --
934 * even if the original read was.
936 if (flags & ZIO_FLAG_IO_REPAIR)
937 flags &= ~ZIO_FLAG_SPECULATIVE;
939 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
940 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
941 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
947 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
948 int type, int priority, enum zio_flag flags,
949 zio_done_func_t *done, void *private)
953 ASSERT(vd->vdev_ops->vdev_op_leaf);
955 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
956 data, size, done, private, type, priority,
957 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
959 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
965 zio_flush(zio_t *zio, vdev_t *vd)
967 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
968 NULL, NULL, ZIO_PRIORITY_NOW,
969 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
973 zio_shrink(zio_t *zio, uint64_t size)
975 ASSERT(zio->io_executor == NULL);
976 ASSERT(zio->io_orig_size == zio->io_size);
977 ASSERT(size <= zio->io_size);
980 * We don't shrink for raidz because of problems with the
981 * reconstruction when reading back less than the block size.
982 * Note, BP_IS_RAIDZ() assumes no compression.
984 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
985 if (!BP_IS_RAIDZ(zio->io_bp))
986 zio->io_orig_size = zio->io_size = size;
990 * ==========================================================================
991 * Prepare to read and write logical blocks
992 * ==========================================================================
996 zio_read_bp_init(zio_t *zio)
998 blkptr_t *bp = zio->io_bp;
1000 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1001 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1002 !(zio->io_flags & ZIO_FLAG_RAW)) {
1003 uint64_t psize = BP_GET_PSIZE(bp);
1004 void *cbuf = zio_buf_alloc(psize);
1006 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1009 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1010 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1012 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1013 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1015 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1016 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1018 return (ZIO_PIPELINE_CONTINUE);
1022 zio_write_bp_init(zio_t *zio)
1024 spa_t *spa = zio->io_spa;
1025 zio_prop_t *zp = &zio->io_prop;
1026 enum zio_compress compress = zp->zp_compress;
1027 blkptr_t *bp = zio->io_bp;
1028 uint64_t lsize = zio->io_size;
1029 uint64_t psize = lsize;
1033 * If our children haven't all reached the ready stage,
1034 * wait for them and then repeat this pipeline stage.
1036 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1037 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1038 return (ZIO_PIPELINE_STOP);
1040 if (!IO_IS_ALLOCATING(zio))
1041 return (ZIO_PIPELINE_CONTINUE);
1043 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1045 if (zio->io_bp_override) {
1046 ASSERT(bp->blk_birth != zio->io_txg);
1047 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1049 *bp = *zio->io_bp_override;
1050 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1052 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1053 return (ZIO_PIPELINE_CONTINUE);
1055 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1056 zp->zp_dedup_verify);
1058 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1059 BP_SET_DEDUP(bp, 1);
1060 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1061 return (ZIO_PIPELINE_CONTINUE);
1063 zio->io_bp_override = NULL;
1067 if (bp->blk_birth == zio->io_txg) {
1069 * We're rewriting an existing block, which means we're
1070 * working on behalf of spa_sync(). For spa_sync() to
1071 * converge, it must eventually be the case that we don't
1072 * have to allocate new blocks. But compression changes
1073 * the blocksize, which forces a reallocate, and makes
1074 * convergence take longer. Therefore, after the first
1075 * few passes, stop compressing to ensure convergence.
1077 pass = spa_sync_pass(spa);
1079 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1080 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1081 ASSERT(!BP_GET_DEDUP(bp));
1083 if (pass >= zfs_sync_pass_dont_compress)
1084 compress = ZIO_COMPRESS_OFF;
1086 /* Make sure someone doesn't change their mind on overwrites */
1087 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1088 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1091 if (compress != ZIO_COMPRESS_OFF) {
1092 void *cbuf = zio_buf_alloc(lsize);
1093 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1094 if (psize == 0 || psize == lsize) {
1095 compress = ZIO_COMPRESS_OFF;
1096 zio_buf_free(cbuf, lsize);
1098 ASSERT(psize < lsize);
1099 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1104 * The final pass of spa_sync() must be all rewrites, but the first
1105 * few passes offer a trade-off: allocating blocks defers convergence,
1106 * but newly allocated blocks are sequential, so they can be written
1107 * to disk faster. Therefore, we allow the first few passes of
1108 * spa_sync() to allocate new blocks, but force rewrites after that.
1109 * There should only be a handful of blocks after pass 1 in any case.
1111 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1112 pass >= zfs_sync_pass_rewrite) {
1113 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1115 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1116 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1119 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1123 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1125 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1126 BP_SET_LSIZE(bp, lsize);
1127 BP_SET_PSIZE(bp, psize);
1128 BP_SET_COMPRESS(bp, compress);
1129 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1130 BP_SET_TYPE(bp, zp->zp_type);
1131 BP_SET_LEVEL(bp, zp->zp_level);
1132 BP_SET_DEDUP(bp, zp->zp_dedup);
1133 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1135 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1136 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1137 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1141 return (ZIO_PIPELINE_CONTINUE);
1145 zio_free_bp_init(zio_t *zio)
1147 blkptr_t *bp = zio->io_bp;
1149 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1150 if (BP_GET_DEDUP(bp))
1151 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1154 return (ZIO_PIPELINE_CONTINUE);
1158 * ==========================================================================
1159 * Execute the I/O pipeline
1160 * ==========================================================================
1164 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1166 spa_t *spa = zio->io_spa;
1167 zio_type_t t = zio->io_type;
1168 int flags = (cutinline ? TQ_FRONT : 0);
1171 * If we're a config writer or a probe, the normal issue and
1172 * interrupt threads may all be blocked waiting for the config lock.
1173 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1175 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1179 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1181 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1185 * If this is a high priority I/O, then use the high priority taskq if
1188 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1189 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1192 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1195 * NB: We are assuming that the zio can only be dispatched
1196 * to a single taskq at a time. It would be a grievous error
1197 * to dispatch the zio to another taskq at the same time.
1199 ASSERT(taskq_empty_ent(&zio->io_tqent));
1200 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1201 flags, &zio->io_tqent);
1205 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1207 kthread_t *executor = zio->io_executor;
1208 spa_t *spa = zio->io_spa;
1211 for (t = 0; t < ZIO_TYPES; t++) {
1212 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1214 for (i = 0; i < tqs->stqs_count; i++) {
1215 if (taskq_member(tqs->stqs_taskq[i], executor))
1224 zio_issue_async(zio_t *zio)
1226 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1228 return (ZIO_PIPELINE_STOP);
1232 zio_interrupt(zio_t *zio)
1234 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1238 * Execute the I/O pipeline until one of the following occurs:
1239 * (1) the I/O completes; (2) the pipeline stalls waiting for
1240 * dependent child I/Os; (3) the I/O issues, so we're waiting
1241 * for an I/O completion interrupt; (4) the I/O is delegated by
1242 * vdev-level caching or aggregation; (5) the I/O is deferred
1243 * due to vdev-level queueing; (6) the I/O is handed off to
1244 * another thread. In all cases, the pipeline stops whenever
1245 * there's no CPU work; it never burns a thread in cv_wait_io().
1247 * There's no locking on io_stage because there's no legitimate way
1248 * for multiple threads to be attempting to process the same I/O.
1250 static zio_pipe_stage_t *zio_pipeline[];
1253 * zio_execute() is a wrapper around the static function
1254 * __zio_execute() so that we can force __zio_execute() to be
1255 * inlined. This reduces stack overhead which is important
1256 * because __zio_execute() is called recursively in several zio
1257 * code paths. zio_execute() itself cannot be inlined because
1258 * it is externally visible.
1261 zio_execute(zio_t *zio)
1266 __attribute__((always_inline))
1268 __zio_execute(zio_t *zio)
1270 zio->io_executor = curthread;
1272 while (zio->io_stage < ZIO_STAGE_DONE) {
1273 enum zio_stage pipeline = zio->io_pipeline;
1274 enum zio_stage stage = zio->io_stage;
1279 ASSERT(!MUTEX_HELD(&zio->io_lock));
1280 ASSERT(ISP2(stage));
1281 ASSERT(zio->io_stall == NULL);
1285 } while ((stage & pipeline) == 0);
1287 ASSERT(stage <= ZIO_STAGE_DONE);
1289 dp = spa_get_dsl(zio->io_spa);
1290 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1291 zio_requeue_io_start_cut_in_line : B_FALSE;
1294 * If we are in interrupt context and this pipeline stage
1295 * will grab a config lock that is held across I/O,
1296 * or may wait for an I/O that needs an interrupt thread
1297 * to complete, issue async to avoid deadlock.
1299 * For VDEV_IO_START, we cut in line so that the io will
1300 * be sent to disk promptly.
1302 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1303 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1304 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1310 * If we executing in the context of the tx_sync_thread,
1311 * or we are performing pool initialization outside of a
1312 * zio_taskq[ZIO_TASKQ_ISSUE|ZIO_TASKQ_ISSUE_HIGH] context.
1313 * Then issue the zio asynchronously to minimize stack usage
1314 * for these deep call paths.
1316 if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
1317 (dp && spa_is_initializing(dp->dp_spa) &&
1318 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
1319 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))) {
1320 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1325 zio->io_stage = stage;
1326 rv = zio_pipeline[highbit(stage) - 1](zio);
1328 if (rv == ZIO_PIPELINE_STOP)
1331 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1337 * ==========================================================================
1338 * Initiate I/O, either sync or async
1339 * ==========================================================================
1342 zio_wait(zio_t *zio)
1346 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1347 ASSERT(zio->io_executor == NULL);
1349 zio->io_waiter = curthread;
1353 mutex_enter(&zio->io_lock);
1354 while (zio->io_executor != NULL)
1355 cv_wait_io(&zio->io_cv, &zio->io_lock);
1356 mutex_exit(&zio->io_lock);
1358 error = zio->io_error;
1365 zio_nowait(zio_t *zio)
1367 ASSERT(zio->io_executor == NULL);
1369 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1370 zio_unique_parent(zio) == NULL) {
1372 * This is a logical async I/O with no parent to wait for it.
1373 * We add it to the spa_async_root_zio "Godfather" I/O which
1374 * will ensure they complete prior to unloading the pool.
1376 spa_t *spa = zio->io_spa;
1378 zio_add_child(spa->spa_async_zio_root, zio);
1385 * ==========================================================================
1386 * Reexecute or suspend/resume failed I/O
1387 * ==========================================================================
1391 zio_reexecute(zio_t *pio)
1393 zio_t *cio, *cio_next;
1396 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1397 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1398 ASSERT(pio->io_gang_leader == NULL);
1399 ASSERT(pio->io_gang_tree == NULL);
1401 pio->io_flags = pio->io_orig_flags;
1402 pio->io_stage = pio->io_orig_stage;
1403 pio->io_pipeline = pio->io_orig_pipeline;
1404 pio->io_reexecute = 0;
1406 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1407 pio->io_state[w] = 0;
1408 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1409 pio->io_child_error[c] = 0;
1411 if (IO_IS_ALLOCATING(pio))
1412 BP_ZERO(pio->io_bp);
1415 * As we reexecute pio's children, new children could be created.
1416 * New children go to the head of pio's io_child_list, however,
1417 * so we will (correctly) not reexecute them. The key is that
1418 * the remainder of pio's io_child_list, from 'cio_next' onward,
1419 * cannot be affected by any side effects of reexecuting 'cio'.
1421 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1422 cio_next = zio_walk_children(pio);
1423 mutex_enter(&pio->io_lock);
1424 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1425 pio->io_children[cio->io_child_type][w]++;
1426 mutex_exit(&pio->io_lock);
1431 * Now that all children have been reexecuted, execute the parent.
1432 * We don't reexecute "The Godfather" I/O here as it's the
1433 * responsibility of the caller to wait on him.
1435 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1440 zio_suspend(spa_t *spa, zio_t *zio)
1442 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1443 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1444 "failure and the failure mode property for this pool "
1445 "is set to panic.", spa_name(spa));
1447 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
1448 "failure and has been suspended.\n", spa_name(spa));
1450 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1452 mutex_enter(&spa->spa_suspend_lock);
1454 if (spa->spa_suspend_zio_root == NULL)
1455 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1456 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1457 ZIO_FLAG_GODFATHER);
1459 spa->spa_suspended = B_TRUE;
1462 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1463 ASSERT(zio != spa->spa_suspend_zio_root);
1464 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1465 ASSERT(zio_unique_parent(zio) == NULL);
1466 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1467 zio_add_child(spa->spa_suspend_zio_root, zio);
1470 mutex_exit(&spa->spa_suspend_lock);
1474 zio_resume(spa_t *spa)
1479 * Reexecute all previously suspended i/o.
1481 mutex_enter(&spa->spa_suspend_lock);
1482 spa->spa_suspended = B_FALSE;
1483 cv_broadcast(&spa->spa_suspend_cv);
1484 pio = spa->spa_suspend_zio_root;
1485 spa->spa_suspend_zio_root = NULL;
1486 mutex_exit(&spa->spa_suspend_lock);
1492 return (zio_wait(pio));
1496 zio_resume_wait(spa_t *spa)
1498 mutex_enter(&spa->spa_suspend_lock);
1499 while (spa_suspended(spa))
1500 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1501 mutex_exit(&spa->spa_suspend_lock);
1505 * ==========================================================================
1508 * A gang block is a collection of small blocks that looks to the DMU
1509 * like one large block. When zio_dva_allocate() cannot find a block
1510 * of the requested size, due to either severe fragmentation or the pool
1511 * being nearly full, it calls zio_write_gang_block() to construct the
1512 * block from smaller fragments.
1514 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1515 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1516 * an indirect block: it's an array of block pointers. It consumes
1517 * only one sector and hence is allocatable regardless of fragmentation.
1518 * The gang header's bps point to its gang members, which hold the data.
1520 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1521 * as the verifier to ensure uniqueness of the SHA256 checksum.
1522 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1523 * not the gang header. This ensures that data block signatures (needed for
1524 * deduplication) are independent of how the block is physically stored.
1526 * Gang blocks can be nested: a gang member may itself be a gang block.
1527 * Thus every gang block is a tree in which root and all interior nodes are
1528 * gang headers, and the leaves are normal blocks that contain user data.
1529 * The root of the gang tree is called the gang leader.
1531 * To perform any operation (read, rewrite, free, claim) on a gang block,
1532 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1533 * in the io_gang_tree field of the original logical i/o by recursively
1534 * reading the gang leader and all gang headers below it. This yields
1535 * an in-core tree containing the contents of every gang header and the
1536 * bps for every constituent of the gang block.
1538 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1539 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1540 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1541 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1542 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1543 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1544 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1545 * of the gang header plus zio_checksum_compute() of the data to update the
1546 * gang header's blk_cksum as described above.
1548 * The two-phase assemble/issue model solves the problem of partial failure --
1549 * what if you'd freed part of a gang block but then couldn't read the
1550 * gang header for another part? Assembling the entire gang tree first
1551 * ensures that all the necessary gang header I/O has succeeded before
1552 * starting the actual work of free, claim, or write. Once the gang tree
1553 * is assembled, free and claim are in-memory operations that cannot fail.
1555 * In the event that a gang write fails, zio_dva_unallocate() walks the
1556 * gang tree to immediately free (i.e. insert back into the space map)
1557 * everything we've allocated. This ensures that we don't get ENOSPC
1558 * errors during repeated suspend/resume cycles due to a flaky device.
1560 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1561 * the gang tree, we won't modify the block, so we can safely defer the free
1562 * (knowing that the block is still intact). If we *can* assemble the gang
1563 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1564 * each constituent bp and we can allocate a new block on the next sync pass.
1566 * In all cases, the gang tree allows complete recovery from partial failure.
1567 * ==========================================================================
1571 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1576 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1577 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1578 &pio->io_bookmark));
1582 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1587 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1588 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1589 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1591 * As we rewrite each gang header, the pipeline will compute
1592 * a new gang block header checksum for it; but no one will
1593 * compute a new data checksum, so we do that here. The one
1594 * exception is the gang leader: the pipeline already computed
1595 * its data checksum because that stage precedes gang assembly.
1596 * (Presently, nothing actually uses interior data checksums;
1597 * this is just good hygiene.)
1599 if (gn != pio->io_gang_leader->io_gang_tree) {
1600 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1601 data, BP_GET_PSIZE(bp));
1604 * If we are here to damage data for testing purposes,
1605 * leave the GBH alone so that we can detect the damage.
1607 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1608 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1610 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1611 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1612 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1620 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1622 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1623 ZIO_GANG_CHILD_FLAGS(pio)));
1628 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1630 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1631 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1634 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1643 static void zio_gang_tree_assemble_done(zio_t *zio);
1645 static zio_gang_node_t *
1646 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1648 zio_gang_node_t *gn;
1650 ASSERT(*gnpp == NULL);
1652 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1653 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1660 zio_gang_node_free(zio_gang_node_t **gnpp)
1662 zio_gang_node_t *gn = *gnpp;
1665 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1666 ASSERT(gn->gn_child[g] == NULL);
1668 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1669 kmem_free(gn, sizeof (*gn));
1674 zio_gang_tree_free(zio_gang_node_t **gnpp)
1676 zio_gang_node_t *gn = *gnpp;
1682 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1683 zio_gang_tree_free(&gn->gn_child[g]);
1685 zio_gang_node_free(gnpp);
1689 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1691 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1693 ASSERT(gio->io_gang_leader == gio);
1694 ASSERT(BP_IS_GANG(bp));
1696 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1697 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1698 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1702 zio_gang_tree_assemble_done(zio_t *zio)
1704 zio_t *gio = zio->io_gang_leader;
1705 zio_gang_node_t *gn = zio->io_private;
1706 blkptr_t *bp = zio->io_bp;
1709 ASSERT(gio == zio_unique_parent(zio));
1710 ASSERT(zio->io_child_count == 0);
1715 if (BP_SHOULD_BYTESWAP(bp))
1716 byteswap_uint64_array(zio->io_data, zio->io_size);
1718 ASSERT(zio->io_data == gn->gn_gbh);
1719 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1720 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1722 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1723 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1724 if (!BP_IS_GANG(gbp))
1726 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1731 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1733 zio_t *gio = pio->io_gang_leader;
1737 ASSERT(BP_IS_GANG(bp) == !!gn);
1738 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1739 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1742 * If you're a gang header, your data is in gn->gn_gbh.
1743 * If you're a gang member, your data is in 'data' and gn == NULL.
1745 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1748 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1750 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1751 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1752 if (BP_IS_HOLE(gbp))
1754 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1755 data = (char *)data + BP_GET_PSIZE(gbp);
1759 if (gn == gio->io_gang_tree)
1760 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1767 zio_gang_assemble(zio_t *zio)
1769 blkptr_t *bp = zio->io_bp;
1771 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1772 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1774 zio->io_gang_leader = zio;
1776 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1778 return (ZIO_PIPELINE_CONTINUE);
1782 zio_gang_issue(zio_t *zio)
1784 blkptr_t *bp = zio->io_bp;
1786 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1787 return (ZIO_PIPELINE_STOP);
1789 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1790 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1792 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1793 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1795 zio_gang_tree_free(&zio->io_gang_tree);
1797 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1799 return (ZIO_PIPELINE_CONTINUE);
1803 zio_write_gang_member_ready(zio_t *zio)
1805 zio_t *pio = zio_unique_parent(zio);
1806 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1807 dva_t *cdva = zio->io_bp->blk_dva;
1808 dva_t *pdva = pio->io_bp->blk_dva;
1812 if (BP_IS_HOLE(zio->io_bp))
1815 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1817 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1818 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1819 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1820 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1821 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1823 mutex_enter(&pio->io_lock);
1824 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1825 ASSERT(DVA_GET_GANG(&pdva[d]));
1826 asize = DVA_GET_ASIZE(&pdva[d]);
1827 asize += DVA_GET_ASIZE(&cdva[d]);
1828 DVA_SET_ASIZE(&pdva[d], asize);
1830 mutex_exit(&pio->io_lock);
1834 zio_write_gang_block(zio_t *pio)
1836 spa_t *spa = pio->io_spa;
1837 blkptr_t *bp = pio->io_bp;
1838 zio_t *gio = pio->io_gang_leader;
1840 zio_gang_node_t *gn, **gnpp;
1841 zio_gbh_phys_t *gbh;
1842 uint64_t txg = pio->io_txg;
1843 uint64_t resid = pio->io_size;
1845 int copies = gio->io_prop.zp_copies;
1846 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1850 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1851 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1852 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1854 pio->io_error = error;
1855 return (ZIO_PIPELINE_CONTINUE);
1859 gnpp = &gio->io_gang_tree;
1861 gnpp = pio->io_private;
1862 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1865 gn = zio_gang_node_alloc(gnpp);
1867 bzero(gbh, SPA_GANGBLOCKSIZE);
1870 * Create the gang header.
1872 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1873 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1876 * Create and nowait the gang children.
1878 for (g = 0; resid != 0; resid -= lsize, g++) {
1879 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1881 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1883 zp.zp_checksum = gio->io_prop.zp_checksum;
1884 zp.zp_compress = ZIO_COMPRESS_OFF;
1885 zp.zp_type = DMU_OT_NONE;
1887 zp.zp_copies = gio->io_prop.zp_copies;
1889 zp.zp_dedup_verify = 0;
1891 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1892 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1893 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1894 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1895 &pio->io_bookmark));
1899 * Set pio's pipeline to just wait for zio to finish.
1901 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1904 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1906 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1910 return (ZIO_PIPELINE_CONTINUE);
1914 * ==========================================================================
1916 * ==========================================================================
1919 zio_ddt_child_read_done(zio_t *zio)
1921 blkptr_t *bp = zio->io_bp;
1922 ddt_entry_t *dde = zio->io_private;
1924 zio_t *pio = zio_unique_parent(zio);
1926 mutex_enter(&pio->io_lock);
1927 ddp = ddt_phys_select(dde, bp);
1928 if (zio->io_error == 0)
1929 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1930 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1931 dde->dde_repair_data = zio->io_data;
1933 zio_buf_free(zio->io_data, zio->io_size);
1934 mutex_exit(&pio->io_lock);
1938 zio_ddt_read_start(zio_t *zio)
1940 blkptr_t *bp = zio->io_bp;
1943 ASSERT(BP_GET_DEDUP(bp));
1944 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1945 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1947 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1948 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1949 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1950 ddt_phys_t *ddp = dde->dde_phys;
1951 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1954 ASSERT(zio->io_vsd == NULL);
1957 if (ddp_self == NULL)
1958 return (ZIO_PIPELINE_CONTINUE);
1960 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1961 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1963 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1965 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1966 zio_buf_alloc(zio->io_size), zio->io_size,
1967 zio_ddt_child_read_done, dde, zio->io_priority,
1968 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1969 &zio->io_bookmark));
1971 return (ZIO_PIPELINE_CONTINUE);
1974 zio_nowait(zio_read(zio, zio->io_spa, bp,
1975 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1976 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1978 return (ZIO_PIPELINE_CONTINUE);
1982 zio_ddt_read_done(zio_t *zio)
1984 blkptr_t *bp = zio->io_bp;
1986 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1987 return (ZIO_PIPELINE_STOP);
1989 ASSERT(BP_GET_DEDUP(bp));
1990 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1991 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1993 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1994 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1995 ddt_entry_t *dde = zio->io_vsd;
1997 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1998 return (ZIO_PIPELINE_CONTINUE);
2001 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2002 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2003 return (ZIO_PIPELINE_STOP);
2005 if (dde->dde_repair_data != NULL) {
2006 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2007 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2009 ddt_repair_done(ddt, dde);
2013 ASSERT(zio->io_vsd == NULL);
2015 return (ZIO_PIPELINE_CONTINUE);
2019 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2021 spa_t *spa = zio->io_spa;
2025 * Note: we compare the original data, not the transformed data,
2026 * because when zio->io_bp is an override bp, we will not have
2027 * pushed the I/O transforms. That's an important optimization
2028 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2030 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2031 zio_t *lio = dde->dde_lead_zio[p];
2034 return (lio->io_orig_size != zio->io_orig_size ||
2035 bcmp(zio->io_orig_data, lio->io_orig_data,
2036 zio->io_orig_size) != 0);
2040 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2041 ddt_phys_t *ddp = &dde->dde_phys[p];
2043 if (ddp->ddp_phys_birth != 0) {
2044 arc_buf_t *abuf = NULL;
2045 uint32_t aflags = ARC_WAIT;
2046 blkptr_t blk = *zio->io_bp;
2049 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2053 error = arc_read(NULL, spa, &blk,
2054 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2055 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2056 &aflags, &zio->io_bookmark);
2059 if (arc_buf_size(abuf) != zio->io_orig_size ||
2060 bcmp(abuf->b_data, zio->io_orig_data,
2061 zio->io_orig_size) != 0)
2063 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2067 return (error != 0);
2075 zio_ddt_child_write_ready(zio_t *zio)
2077 int p = zio->io_prop.zp_copies;
2078 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2079 ddt_entry_t *dde = zio->io_private;
2080 ddt_phys_t *ddp = &dde->dde_phys[p];
2088 ASSERT(dde->dde_lead_zio[p] == zio);
2090 ddt_phys_fill(ddp, zio->io_bp);
2092 while ((pio = zio_walk_parents(zio)) != NULL)
2093 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2099 zio_ddt_child_write_done(zio_t *zio)
2101 int p = zio->io_prop.zp_copies;
2102 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2103 ddt_entry_t *dde = zio->io_private;
2104 ddt_phys_t *ddp = &dde->dde_phys[p];
2108 ASSERT(ddp->ddp_refcnt == 0);
2109 ASSERT(dde->dde_lead_zio[p] == zio);
2110 dde->dde_lead_zio[p] = NULL;
2112 if (zio->io_error == 0) {
2113 while (zio_walk_parents(zio) != NULL)
2114 ddt_phys_addref(ddp);
2116 ddt_phys_clear(ddp);
2123 zio_ddt_ditto_write_done(zio_t *zio)
2125 int p = DDT_PHYS_DITTO;
2126 blkptr_t *bp = zio->io_bp;
2127 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2128 ddt_entry_t *dde = zio->io_private;
2129 ddt_phys_t *ddp = &dde->dde_phys[p];
2130 ddt_key_t *ddk = &dde->dde_key;
2131 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2135 ASSERT(ddp->ddp_refcnt == 0);
2136 ASSERT(dde->dde_lead_zio[p] == zio);
2137 dde->dde_lead_zio[p] = NULL;
2139 if (zio->io_error == 0) {
2140 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2141 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2142 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2143 if (ddp->ddp_phys_birth != 0)
2144 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2145 ddt_phys_fill(ddp, bp);
2152 zio_ddt_write(zio_t *zio)
2154 spa_t *spa = zio->io_spa;
2155 blkptr_t *bp = zio->io_bp;
2156 uint64_t txg = zio->io_txg;
2157 zio_prop_t *zp = &zio->io_prop;
2158 int p = zp->zp_copies;
2162 ddt_t *ddt = ddt_select(spa, bp);
2166 ASSERT(BP_GET_DEDUP(bp));
2167 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2168 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2171 dde = ddt_lookup(ddt, bp, B_TRUE);
2172 ddp = &dde->dde_phys[p];
2174 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2176 * If we're using a weak checksum, upgrade to a strong checksum
2177 * and try again. If we're already using a strong checksum,
2178 * we can't resolve it, so just convert to an ordinary write.
2179 * (And automatically e-mail a paper to Nature?)
2181 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2182 zp->zp_checksum = spa_dedup_checksum(spa);
2183 zio_pop_transforms(zio);
2184 zio->io_stage = ZIO_STAGE_OPEN;
2189 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2191 return (ZIO_PIPELINE_CONTINUE);
2194 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2195 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2197 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2198 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2199 zio_prop_t czp = *zp;
2201 czp.zp_copies = ditto_copies;
2204 * If we arrived here with an override bp, we won't have run
2205 * the transform stack, so we won't have the data we need to
2206 * generate a child i/o. So, toss the override bp and restart.
2207 * This is safe, because using the override bp is just an
2208 * optimization; and it's rare, so the cost doesn't matter.
2210 if (zio->io_bp_override) {
2211 zio_pop_transforms(zio);
2212 zio->io_stage = ZIO_STAGE_OPEN;
2213 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2214 zio->io_bp_override = NULL;
2217 return (ZIO_PIPELINE_CONTINUE);
2220 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2221 zio->io_orig_size, &czp, NULL,
2222 zio_ddt_ditto_write_done, dde, zio->io_priority,
2223 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2225 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2226 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2229 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2230 if (ddp->ddp_phys_birth != 0)
2231 ddt_bp_fill(ddp, bp, txg);
2232 if (dde->dde_lead_zio[p] != NULL)
2233 zio_add_child(zio, dde->dde_lead_zio[p]);
2235 ddt_phys_addref(ddp);
2236 } else if (zio->io_bp_override) {
2237 ASSERT(bp->blk_birth == txg);
2238 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2239 ddt_phys_fill(ddp, bp);
2240 ddt_phys_addref(ddp);
2242 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2243 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2244 zio_ddt_child_write_done, dde, zio->io_priority,
2245 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2247 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2248 dde->dde_lead_zio[p] = cio;
2258 return (ZIO_PIPELINE_CONTINUE);
2261 ddt_entry_t *freedde; /* for debugging */
2264 zio_ddt_free(zio_t *zio)
2266 spa_t *spa = zio->io_spa;
2267 blkptr_t *bp = zio->io_bp;
2268 ddt_t *ddt = ddt_select(spa, bp);
2272 ASSERT(BP_GET_DEDUP(bp));
2273 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2276 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2278 ddp = ddt_phys_select(dde, bp);
2280 ddt_phys_decref(ddp);
2284 return (ZIO_PIPELINE_CONTINUE);
2288 * ==========================================================================
2289 * Allocate and free blocks
2290 * ==========================================================================
2293 zio_dva_allocate(zio_t *zio)
2295 spa_t *spa = zio->io_spa;
2296 metaslab_class_t *mc = spa_normal_class(spa);
2297 blkptr_t *bp = zio->io_bp;
2301 if (zio->io_gang_leader == NULL) {
2302 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2303 zio->io_gang_leader = zio;
2306 ASSERT(BP_IS_HOLE(bp));
2307 ASSERT0(BP_GET_NDVAS(bp));
2308 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2309 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2310 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2313 * The dump device does not support gang blocks so allocation on
2314 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2315 * the "fast" gang feature.
2317 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2318 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2319 METASLAB_GANG_CHILD : 0;
2320 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2321 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2322 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2325 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2326 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2328 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2329 return (zio_write_gang_block(zio));
2330 zio->io_error = error;
2333 return (ZIO_PIPELINE_CONTINUE);
2337 zio_dva_free(zio_t *zio)
2339 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2341 return (ZIO_PIPELINE_CONTINUE);
2345 zio_dva_claim(zio_t *zio)
2349 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2351 zio->io_error = error;
2353 return (ZIO_PIPELINE_CONTINUE);
2357 * Undo an allocation. This is used by zio_done() when an I/O fails
2358 * and we want to give back the block we just allocated.
2359 * This handles both normal blocks and gang blocks.
2362 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2366 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2367 ASSERT(zio->io_bp_override == NULL);
2369 if (!BP_IS_HOLE(bp))
2370 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2373 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2374 zio_dva_unallocate(zio, gn->gn_child[g],
2375 &gn->gn_gbh->zg_blkptr[g]);
2381 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2384 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2389 ASSERT(txg > spa_syncing_txg(spa));
2392 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2393 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2394 * when allocating them.
2397 error = metaslab_alloc(spa, spa_log_class(spa), size,
2398 new_bp, 1, txg, NULL,
2399 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2403 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2404 new_bp, 1, txg, NULL,
2405 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2409 BP_SET_LSIZE(new_bp, size);
2410 BP_SET_PSIZE(new_bp, size);
2411 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2412 BP_SET_CHECKSUM(new_bp,
2413 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2414 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2415 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2416 BP_SET_LEVEL(new_bp, 0);
2417 BP_SET_DEDUP(new_bp, 0);
2418 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2425 * Free an intent log block.
2428 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2430 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2431 ASSERT(!BP_IS_GANG(bp));
2433 zio_free(spa, txg, bp);
2437 * ==========================================================================
2438 * Read and write to physical devices
2439 * ==========================================================================
2442 zio_vdev_io_start(zio_t *zio)
2444 vdev_t *vd = zio->io_vd;
2446 spa_t *spa = zio->io_spa;
2448 ASSERT(zio->io_error == 0);
2449 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2452 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2453 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2456 * The mirror_ops handle multiple DVAs in a single BP.
2458 return (vdev_mirror_ops.vdev_op_io_start(zio));
2462 * We keep track of time-sensitive I/Os so that the scan thread
2463 * can quickly react to certain workloads. In particular, we care
2464 * about non-scrubbing, top-level reads and writes with the following
2466 * - synchronous writes of user data to non-slog devices
2467 * - any reads of user data
2468 * When these conditions are met, adjust the timestamp of spa_last_io
2469 * which allows the scan thread to adjust its workload accordingly.
2471 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2472 vd == vd->vdev_top && !vd->vdev_islog &&
2473 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2474 zio->io_txg != spa_syncing_txg(spa)) {
2475 uint64_t old = spa->spa_last_io;
2476 uint64_t new = ddi_get_lbolt64();
2478 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2481 align = 1ULL << vd->vdev_top->vdev_ashift;
2483 if (P2PHASE(zio->io_size, align) != 0) {
2484 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2485 char *abuf = zio_buf_alloc(asize);
2486 ASSERT(vd == vd->vdev_top);
2487 if (zio->io_type == ZIO_TYPE_WRITE) {
2488 bcopy(zio->io_data, abuf, zio->io_size);
2489 bzero(abuf + zio->io_size, asize - zio->io_size);
2491 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2494 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2495 ASSERT(P2PHASE(zio->io_size, align) == 0);
2496 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2499 * If this is a repair I/O, and there's no self-healing involved --
2500 * that is, we're just resilvering what we expect to resilver --
2501 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2502 * This prevents spurious resilvering with nested replication.
2503 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2504 * A is out of date, we'll read from C+D, then use the data to
2505 * resilver A+B -- but we don't actually want to resilver B, just A.
2506 * The top-level mirror has no way to know this, so instead we just
2507 * discard unnecessary repairs as we work our way down the vdev tree.
2508 * The same logic applies to any form of nested replication:
2509 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2511 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2512 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2513 zio->io_txg != 0 && /* not a delegated i/o */
2514 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2515 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2516 zio_vdev_io_bypass(zio);
2517 return (ZIO_PIPELINE_CONTINUE);
2520 if (vd->vdev_ops->vdev_op_leaf &&
2521 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2523 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2524 return (ZIO_PIPELINE_CONTINUE);
2526 if ((zio = vdev_queue_io(zio)) == NULL)
2527 return (ZIO_PIPELINE_STOP);
2529 if (!vdev_accessible(vd, zio)) {
2530 zio->io_error = ENXIO;
2532 return (ZIO_PIPELINE_STOP);
2536 return (vd->vdev_ops->vdev_op_io_start(zio));
2540 zio_vdev_io_done(zio_t *zio)
2542 vdev_t *vd = zio->io_vd;
2543 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2544 boolean_t unexpected_error = B_FALSE;
2546 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2547 return (ZIO_PIPELINE_STOP);
2549 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2551 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2553 vdev_queue_io_done(zio);
2555 if (zio->io_type == ZIO_TYPE_WRITE)
2556 vdev_cache_write(zio);
2558 if (zio_injection_enabled && zio->io_error == 0)
2559 zio->io_error = zio_handle_device_injection(vd,
2562 if (zio_injection_enabled && zio->io_error == 0)
2563 zio->io_error = zio_handle_label_injection(zio, EIO);
2565 if (zio->io_error) {
2566 if (!vdev_accessible(vd, zio)) {
2567 zio->io_error = ENXIO;
2569 unexpected_error = B_TRUE;
2574 ops->vdev_op_io_done(zio);
2576 if (unexpected_error)
2577 VERIFY(vdev_probe(vd, zio) == NULL);
2579 return (ZIO_PIPELINE_CONTINUE);
2583 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2584 * disk, and use that to finish the checksum ereport later.
2587 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2588 const void *good_buf)
2590 /* no processing needed */
2591 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2596 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2598 void *buf = zio_buf_alloc(zio->io_size);
2600 bcopy(zio->io_data, buf, zio->io_size);
2602 zcr->zcr_cbinfo = zio->io_size;
2603 zcr->zcr_cbdata = buf;
2604 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2605 zcr->zcr_free = zio_buf_free;
2609 zio_vdev_io_assess(zio_t *zio)
2611 vdev_t *vd = zio->io_vd;
2613 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2614 return (ZIO_PIPELINE_STOP);
2616 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2617 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2619 if (zio->io_vsd != NULL) {
2620 zio->io_vsd_ops->vsd_free(zio);
2624 if (zio_injection_enabled && zio->io_error == 0)
2625 zio->io_error = zio_handle_fault_injection(zio, EIO);
2628 * If the I/O failed, determine whether we should attempt to retry it.
2630 * On retry, we cut in line in the issue queue, since we don't want
2631 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2633 if (zio->io_error && vd == NULL &&
2634 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2635 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2636 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2638 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2639 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2640 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2641 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2642 zio_requeue_io_start_cut_in_line);
2643 return (ZIO_PIPELINE_STOP);
2647 * If we got an error on a leaf device, convert it to ENXIO
2648 * if the device is not accessible at all.
2650 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2651 !vdev_accessible(vd, zio))
2652 zio->io_error = ENXIO;
2655 * If we can't write to an interior vdev (mirror or RAID-Z),
2656 * set vdev_cant_write so that we stop trying to allocate from it.
2658 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2659 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2660 vd->vdev_cant_write = B_TRUE;
2663 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2665 return (ZIO_PIPELINE_CONTINUE);
2669 zio_vdev_io_reissue(zio_t *zio)
2671 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2672 ASSERT(zio->io_error == 0);
2674 zio->io_stage >>= 1;
2678 zio_vdev_io_redone(zio_t *zio)
2680 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2682 zio->io_stage >>= 1;
2686 zio_vdev_io_bypass(zio_t *zio)
2688 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2689 ASSERT(zio->io_error == 0);
2691 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2692 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2696 * ==========================================================================
2697 * Generate and verify checksums
2698 * ==========================================================================
2701 zio_checksum_generate(zio_t *zio)
2703 blkptr_t *bp = zio->io_bp;
2704 enum zio_checksum checksum;
2708 * This is zio_write_phys().
2709 * We're either generating a label checksum, or none at all.
2711 checksum = zio->io_prop.zp_checksum;
2713 if (checksum == ZIO_CHECKSUM_OFF)
2714 return (ZIO_PIPELINE_CONTINUE);
2716 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2718 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2719 ASSERT(!IO_IS_ALLOCATING(zio));
2720 checksum = ZIO_CHECKSUM_GANG_HEADER;
2722 checksum = BP_GET_CHECKSUM(bp);
2726 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2728 return (ZIO_PIPELINE_CONTINUE);
2732 zio_checksum_verify(zio_t *zio)
2734 zio_bad_cksum_t info;
2735 blkptr_t *bp = zio->io_bp;
2738 ASSERT(zio->io_vd != NULL);
2742 * This is zio_read_phys().
2743 * We're either verifying a label checksum, or nothing at all.
2745 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2746 return (ZIO_PIPELINE_CONTINUE);
2748 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2751 if ((error = zio_checksum_error(zio, &info)) != 0) {
2752 zio->io_error = error;
2753 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2754 zfs_ereport_start_checksum(zio->io_spa,
2755 zio->io_vd, zio, zio->io_offset,
2756 zio->io_size, NULL, &info);
2760 return (ZIO_PIPELINE_CONTINUE);
2764 * Called by RAID-Z to ensure we don't compute the checksum twice.
2767 zio_checksum_verified(zio_t *zio)
2769 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2773 * ==========================================================================
2774 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2775 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2776 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2777 * indicate errors that are specific to one I/O, and most likely permanent.
2778 * Any other error is presumed to be worse because we weren't expecting it.
2779 * ==========================================================================
2782 zio_worst_error(int e1, int e2)
2784 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2787 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2788 if (e1 == zio_error_rank[r1])
2791 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2792 if (e2 == zio_error_rank[r2])
2795 return (r1 > r2 ? e1 : e2);
2799 * ==========================================================================
2801 * ==========================================================================
2804 zio_ready(zio_t *zio)
2806 blkptr_t *bp = zio->io_bp;
2807 zio_t *pio, *pio_next;
2809 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2810 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2811 return (ZIO_PIPELINE_STOP);
2813 if (zio->io_ready) {
2814 ASSERT(IO_IS_ALLOCATING(zio));
2815 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2816 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2821 if (bp != NULL && bp != &zio->io_bp_copy)
2822 zio->io_bp_copy = *bp;
2825 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2827 mutex_enter(&zio->io_lock);
2828 zio->io_state[ZIO_WAIT_READY] = 1;
2829 pio = zio_walk_parents(zio);
2830 mutex_exit(&zio->io_lock);
2833 * As we notify zio's parents, new parents could be added.
2834 * New parents go to the head of zio's io_parent_list, however,
2835 * so we will (correctly) not notify them. The remainder of zio's
2836 * io_parent_list, from 'pio_next' onward, cannot change because
2837 * all parents must wait for us to be done before they can be done.
2839 for (; pio != NULL; pio = pio_next) {
2840 pio_next = zio_walk_parents(zio);
2841 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2844 if (zio->io_flags & ZIO_FLAG_NODATA) {
2845 if (BP_IS_GANG(bp)) {
2846 zio->io_flags &= ~ZIO_FLAG_NODATA;
2848 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2849 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2853 if (zio_injection_enabled &&
2854 zio->io_spa->spa_syncing_txg == zio->io_txg)
2855 zio_handle_ignored_writes(zio);
2857 return (ZIO_PIPELINE_CONTINUE);
2861 zio_done(zio_t *zio)
2863 zio_t *pio, *pio_next;
2867 * If our children haven't all completed,
2868 * wait for them and then repeat this pipeline stage.
2870 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2871 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2872 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2873 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2874 return (ZIO_PIPELINE_STOP);
2876 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2877 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2878 ASSERT(zio->io_children[c][w] == 0);
2880 if (zio->io_bp != NULL) {
2881 ASSERT(zio->io_bp->blk_pad[0] == 0);
2882 ASSERT(zio->io_bp->blk_pad[1] == 0);
2883 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2884 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2885 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2886 zio->io_bp_override == NULL &&
2887 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2888 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2889 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2890 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2891 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2896 * If there were child vdev/gang/ddt errors, they apply to us now.
2898 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2899 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2900 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2903 * If the I/O on the transformed data was successful, generate any
2904 * checksum reports now while we still have the transformed data.
2906 if (zio->io_error == 0) {
2907 while (zio->io_cksum_report != NULL) {
2908 zio_cksum_report_t *zcr = zio->io_cksum_report;
2909 uint64_t align = zcr->zcr_align;
2910 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2911 char *abuf = zio->io_data;
2913 if (asize != zio->io_size) {
2914 abuf = zio_buf_alloc(asize);
2915 bcopy(zio->io_data, abuf, zio->io_size);
2916 bzero(abuf + zio->io_size, asize - zio->io_size);
2919 zio->io_cksum_report = zcr->zcr_next;
2920 zcr->zcr_next = NULL;
2921 zcr->zcr_finish(zcr, abuf);
2922 zfs_ereport_free_checksum(zcr);
2924 if (asize != zio->io_size)
2925 zio_buf_free(abuf, asize);
2929 zio_pop_transforms(zio); /* note: may set zio->io_error */
2931 vdev_stat_update(zio, zio->io_size);
2934 * If this I/O is attached to a particular vdev is slow, exceeding
2935 * 30 seconds to complete, post an error described the I/O delay.
2936 * We ignore these errors if the device is currently unavailable.
2938 if (zio->io_delay >= MSEC_TO_TICK(zio_delay_max)) {
2939 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2940 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2941 zio->io_vd, zio, 0, 0);
2944 if (zio->io_error) {
2946 * If this I/O is attached to a particular vdev,
2947 * generate an error message describing the I/O failure
2948 * at the block level. We ignore these errors if the
2949 * device is currently unavailable.
2951 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2952 !vdev_is_dead(zio->io_vd))
2953 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2954 zio->io_vd, zio, 0, 0);
2956 if ((zio->io_error == EIO || !(zio->io_flags &
2957 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2958 zio == zio->io_logical) {
2960 * For logical I/O requests, tell the SPA to log the
2961 * error and generate a logical data ereport.
2963 spa_log_error(zio->io_spa, zio);
2964 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2969 if (zio->io_error && zio == zio->io_logical) {
2971 * Determine whether zio should be reexecuted. This will
2972 * propagate all the way to the root via zio_notify_parent().
2974 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2975 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2977 if (IO_IS_ALLOCATING(zio) &&
2978 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2979 if (zio->io_error != ENOSPC)
2980 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2982 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2985 if ((zio->io_type == ZIO_TYPE_READ ||
2986 zio->io_type == ZIO_TYPE_FREE) &&
2987 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2988 zio->io_error == ENXIO &&
2989 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2990 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2991 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2993 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2994 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2997 * Here is a possibly good place to attempt to do
2998 * either combinatorial reconstruction or error correction
2999 * based on checksums. It also might be a good place
3000 * to send out preliminary ereports before we suspend
3006 * If there were logical child errors, they apply to us now.
3007 * We defer this until now to avoid conflating logical child
3008 * errors with errors that happened to the zio itself when
3009 * updating vdev stats and reporting FMA events above.
3011 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3013 if ((zio->io_error || zio->io_reexecute) &&
3014 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3015 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
3016 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
3018 zio_gang_tree_free(&zio->io_gang_tree);
3021 * Godfather I/Os should never suspend.
3023 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3024 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3025 zio->io_reexecute = 0;
3027 if (zio->io_reexecute) {
3029 * This is a logical I/O that wants to reexecute.
3031 * Reexecute is top-down. When an i/o fails, if it's not
3032 * the root, it simply notifies its parent and sticks around.
3033 * The parent, seeing that it still has children in zio_done(),
3034 * does the same. This percolates all the way up to the root.
3035 * The root i/o will reexecute or suspend the entire tree.
3037 * This approach ensures that zio_reexecute() honors
3038 * all the original i/o dependency relationships, e.g.
3039 * parents not executing until children are ready.
3041 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3043 zio->io_gang_leader = NULL;
3045 mutex_enter(&zio->io_lock);
3046 zio->io_state[ZIO_WAIT_DONE] = 1;
3047 mutex_exit(&zio->io_lock);
3050 * "The Godfather" I/O monitors its children but is
3051 * not a true parent to them. It will track them through
3052 * the pipeline but severs its ties whenever they get into
3053 * trouble (e.g. suspended). This allows "The Godfather"
3054 * I/O to return status without blocking.
3056 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3057 zio_link_t *zl = zio->io_walk_link;
3058 pio_next = zio_walk_parents(zio);
3060 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3061 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3062 zio_remove_child(pio, zio, zl);
3063 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3067 if ((pio = zio_unique_parent(zio)) != NULL) {
3069 * We're not a root i/o, so there's nothing to do
3070 * but notify our parent. Don't propagate errors
3071 * upward since we haven't permanently failed yet.
3073 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3074 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3075 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3076 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3078 * We'd fail again if we reexecuted now, so suspend
3079 * until conditions improve (e.g. device comes online).
3081 zio_suspend(zio->io_spa, zio);
3084 * Reexecution is potentially a huge amount of work.
3085 * Hand it off to the otherwise-unused claim taskq.
3087 ASSERT(taskq_empty_ent(&zio->io_tqent));
3088 spa_taskq_dispatch_ent(zio->io_spa,
3089 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
3090 (task_func_t *)zio_reexecute, zio, 0,
3093 return (ZIO_PIPELINE_STOP);
3096 ASSERT(zio->io_child_count == 0);
3097 ASSERT(zio->io_reexecute == 0);
3098 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3101 * Report any checksum errors, since the I/O is complete.
3103 while (zio->io_cksum_report != NULL) {
3104 zio_cksum_report_t *zcr = zio->io_cksum_report;
3105 zio->io_cksum_report = zcr->zcr_next;
3106 zcr->zcr_next = NULL;
3107 zcr->zcr_finish(zcr, NULL);
3108 zfs_ereport_free_checksum(zcr);
3111 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3112 !BP_IS_HOLE(zio->io_bp)) {
3113 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3117 * It is the responsibility of the done callback to ensure that this
3118 * particular zio is no longer discoverable for adoption, and as
3119 * such, cannot acquire any new parents.
3124 mutex_enter(&zio->io_lock);
3125 zio->io_state[ZIO_WAIT_DONE] = 1;
3126 mutex_exit(&zio->io_lock);
3128 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3129 zio_link_t *zl = zio->io_walk_link;
3130 pio_next = zio_walk_parents(zio);
3131 zio_remove_child(pio, zio, zl);
3132 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3135 if (zio->io_waiter != NULL) {
3136 mutex_enter(&zio->io_lock);
3137 zio->io_executor = NULL;
3138 cv_broadcast(&zio->io_cv);
3139 mutex_exit(&zio->io_lock);
3144 return (ZIO_PIPELINE_STOP);
3148 * ==========================================================================
3149 * I/O pipeline definition
3150 * ==========================================================================
3152 static zio_pipe_stage_t *zio_pipeline[] = {
3158 zio_checksum_generate,
3172 zio_checksum_verify,
3176 /* dnp is the dnode for zb1->zb_object */
3178 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3179 const zbookmark_t *zb2)
3181 uint64_t zb1nextL0, zb2thisobj;
3183 ASSERT(zb1->zb_objset == zb2->zb_objset);
3184 ASSERT(zb2->zb_level == 0);
3187 * A bookmark in the deadlist is considered to be after
3190 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3193 /* The objset_phys_t isn't before anything. */
3197 zb1nextL0 = (zb1->zb_blkid + 1) <<
3198 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3200 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3201 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3203 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3204 uint64_t nextobj = zb1nextL0 *
3205 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3206 return (nextobj <= zb2thisobj);
3209 if (zb1->zb_object < zb2thisobj)
3211 if (zb1->zb_object > zb2thisobj)
3213 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3215 return (zb1nextL0 <= zb2->zb_blkid);
3218 #if defined(_KERNEL) && defined(HAVE_SPL)
3219 /* Fault injection */
3220 EXPORT_SYMBOL(zio_injection_enabled);
3221 EXPORT_SYMBOL(zio_inject_fault);
3222 EXPORT_SYMBOL(zio_inject_list_next);
3223 EXPORT_SYMBOL(zio_clear_fault);
3224 EXPORT_SYMBOL(zio_handle_fault_injection);
3225 EXPORT_SYMBOL(zio_handle_device_injection);
3226 EXPORT_SYMBOL(zio_handle_label_injection);
3227 EXPORT_SYMBOL(zio_priority_table);
3228 EXPORT_SYMBOL(zio_type_name);
3230 module_param(zio_bulk_flags, int, 0644);
3231 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3233 module_param(zio_delay_max, int, 0644);
3234 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3236 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3237 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
3239 module_param(zfs_sync_pass_deferred_free, int, 0644);
3240 MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
3241 "defer frees starting in this pass");
3243 module_param(zfs_sync_pass_dont_compress, int, 0644);
3244 MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
3245 "don't compress starting in this pass");
3247 module_param(zfs_sync_pass_rewrite, int, 0644);
3248 MODULE_PARM_DESC(zfs_sync_pass_rewrite,
3249 "rewrite new bps starting in this pass");