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;
83 extern vmem_t *zio_alloc_arena;
85 extern int zfs_mg_alloc_failures;
88 * The following actions directly effect the spa's sync-to-convergence logic.
89 * The values below define the sync pass when we start performing the action.
90 * Care should be taken when changing these values as they directly impact
91 * spa_sync() performance. Tuning these values may introduce subtle performance
92 * pathologies and should only be done in the context of performance analysis.
93 * These tunables will eventually be removed and replaced with #defines once
94 * enough analysis has been done to determine optimal values.
96 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
97 * regular blocks are not deferred.
99 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
100 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
101 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
104 * An allocating zio is one that either currently has the DVA allocate
105 * stage set or will have it later in its lifetime.
107 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
109 int zio_requeue_io_start_cut_in_line = 1;
112 int zio_buf_debug_limit = 16384;
114 int zio_buf_debug_limit = 0;
117 static inline void __zio_execute(zio_t *zio);
120 zio_cons(void *arg, void *unused, int kmflag)
124 bzero(zio, sizeof (zio_t));
126 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
127 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
129 list_create(&zio->io_parent_list, sizeof (zio_link_t),
130 offsetof(zio_link_t, zl_parent_node));
131 list_create(&zio->io_child_list, sizeof (zio_link_t),
132 offsetof(zio_link_t, zl_child_node));
138 zio_dest(void *arg, void *unused)
142 mutex_destroy(&zio->io_lock);
143 cv_destroy(&zio->io_cv);
144 list_destroy(&zio->io_parent_list);
145 list_destroy(&zio->io_child_list);
152 vmem_t *data_alloc_arena = NULL;
155 data_alloc_arena = zio_alloc_arena;
157 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
158 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
159 zio_link_cache = kmem_cache_create("zio_link_cache",
160 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
161 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
162 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
165 * For small buffers, we want a cache for each multiple of
166 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
167 * for each quarter-power of 2. For large buffers, we want
168 * a cache for each multiple of PAGESIZE.
170 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
171 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
175 while (p2 & (p2 - 1))
178 if (size <= 4 * SPA_MINBLOCKSIZE) {
179 align = SPA_MINBLOCKSIZE;
180 } else if (P2PHASE(size, PAGESIZE) == 0) {
182 } else if (P2PHASE(size, p2 >> 2) == 0) {
188 int flags = zio_bulk_flags;
191 * The smallest buffers (512b) are heavily used and
192 * experience a lot of churn. The slabs allocated
193 * for them are also relatively small (32K). Thus
194 * in over to avoid expensive calls to vmalloc() we
195 * make an exception to the usual slab allocation
196 * policy and force these buffers to be kmem backed.
198 if (size == (1 << SPA_MINBLOCKSHIFT))
201 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
202 zio_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL, flags);
205 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
206 zio_data_buf_cache[c] = kmem_cache_create(name, size,
207 align, NULL, NULL, NULL, NULL,
208 data_alloc_arena, flags);
213 ASSERT(zio_buf_cache[c] != NULL);
214 if (zio_buf_cache[c - 1] == NULL)
215 zio_buf_cache[c - 1] = zio_buf_cache[c];
217 ASSERT(zio_data_buf_cache[c] != NULL);
218 if (zio_data_buf_cache[c - 1] == NULL)
219 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
223 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
224 * to fail 3 times per txg or 8 failures, whichever is greater.
226 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
237 kmem_cache_t *last_cache = NULL;
238 kmem_cache_t *last_data_cache = NULL;
240 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
241 if (zio_buf_cache[c] != last_cache) {
242 last_cache = zio_buf_cache[c];
243 kmem_cache_destroy(zio_buf_cache[c]);
245 zio_buf_cache[c] = NULL;
247 if (zio_data_buf_cache[c] != last_data_cache) {
248 last_data_cache = zio_data_buf_cache[c];
249 kmem_cache_destroy(zio_data_buf_cache[c]);
251 zio_data_buf_cache[c] = NULL;
254 kmem_cache_destroy(zio_vdev_cache);
255 kmem_cache_destroy(zio_link_cache);
256 kmem_cache_destroy(zio_cache);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
282 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
286 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
287 * crashdump if the kernel panics. This exists so that we will limit the amount
288 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
289 * of kernel heap dumped to disk when the kernel panics)
292 zio_data_buf_alloc(size_t size)
294 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
296 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 return (kmem_cache_alloc(zio_data_buf_cache[c],
299 KM_PUSHPAGE | KM_NODEBUG));
303 zio_buf_free(void *buf, size_t size)
305 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
307 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
309 kmem_cache_free(zio_buf_cache[c], buf);
313 zio_data_buf_free(void *buf, size_t size)
315 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
317 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
319 kmem_cache_free(zio_data_buf_cache[c], buf);
323 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
324 * or significantly delays the issuing of a zio. These buffers are used
325 * to aggregate I/O and could be used for raidz stripes.
330 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
334 zio_vdev_free(void *buf)
336 kmem_cache_free(zio_vdev_cache, buf);
341 * ==========================================================================
342 * Push and pop I/O transform buffers
343 * ==========================================================================
346 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
347 zio_transform_func_t *transform)
349 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
351 zt->zt_orig_data = zio->io_data;
352 zt->zt_orig_size = zio->io_size;
353 zt->zt_bufsize = bufsize;
354 zt->zt_transform = transform;
356 zt->zt_next = zio->io_transform_stack;
357 zio->io_transform_stack = zt;
364 zio_pop_transforms(zio_t *zio)
368 while ((zt = zio->io_transform_stack) != NULL) {
369 if (zt->zt_transform != NULL)
370 zt->zt_transform(zio,
371 zt->zt_orig_data, zt->zt_orig_size);
373 if (zt->zt_bufsize != 0)
374 zio_buf_free(zio->io_data, zt->zt_bufsize);
376 zio->io_data = zt->zt_orig_data;
377 zio->io_size = zt->zt_orig_size;
378 zio->io_transform_stack = zt->zt_next;
380 kmem_free(zt, sizeof (zio_transform_t));
385 * ==========================================================================
386 * I/O transform callbacks for subblocks and decompression
387 * ==========================================================================
390 zio_subblock(zio_t *zio, void *data, uint64_t size)
392 ASSERT(zio->io_size > size);
394 if (zio->io_type == ZIO_TYPE_READ)
395 bcopy(zio->io_data, data, size);
399 zio_decompress(zio_t *zio, void *data, uint64_t size)
401 if (zio->io_error == 0 &&
402 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
403 zio->io_data, data, zio->io_size, size) != 0)
408 * ==========================================================================
409 * I/O parent/child relationships and pipeline interlocks
410 * ==========================================================================
413 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
414 * continue calling these functions until they return NULL.
415 * Otherwise, the next caller will pick up the list walk in
416 * some indeterminate state. (Otherwise every caller would
417 * have to pass in a cookie to keep the state represented by
418 * io_walk_link, which gets annoying.)
421 zio_walk_parents(zio_t *cio)
423 zio_link_t *zl = cio->io_walk_link;
424 list_t *pl = &cio->io_parent_list;
426 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
427 cio->io_walk_link = zl;
432 ASSERT(zl->zl_child == cio);
433 return (zl->zl_parent);
437 zio_walk_children(zio_t *pio)
439 zio_link_t *zl = pio->io_walk_link;
440 list_t *cl = &pio->io_child_list;
442 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
443 pio->io_walk_link = zl;
448 ASSERT(zl->zl_parent == pio);
449 return (zl->zl_child);
453 zio_unique_parent(zio_t *cio)
455 zio_t *pio = zio_walk_parents(cio);
457 VERIFY(zio_walk_parents(cio) == NULL);
462 zio_add_child(zio_t *pio, zio_t *cio)
464 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT(cio->io_child_type <= pio->io_child_type);
478 mutex_enter(&cio->io_lock);
479 mutex_enter(&pio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&pio->io_lock);
493 mutex_exit(&cio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&cio->io_lock);
503 mutex_enter(&pio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&pio->io_lock);
512 mutex_exit(&cio->io_lock);
514 kmem_cache_free(zio_link_cache, zl);
518 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
520 uint64_t *countp = &zio->io_children[child][wait];
521 boolean_t waiting = B_FALSE;
523 mutex_enter(&zio->io_lock);
524 ASSERT(zio->io_stall == NULL);
527 zio->io_stall = countp;
530 mutex_exit(&zio->io_lock);
535 __attribute__((always_inline))
537 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
539 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
540 int *errorp = &pio->io_child_error[zio->io_child_type];
542 mutex_enter(&pio->io_lock);
543 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
544 *errorp = zio_worst_error(*errorp, zio->io_error);
545 pio->io_reexecute |= zio->io_reexecute;
546 ASSERT3U(*countp, >, 0);
547 if (--*countp == 0 && pio->io_stall == countp) {
548 pio->io_stall = NULL;
549 mutex_exit(&pio->io_lock);
552 mutex_exit(&pio->io_lock);
557 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
559 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
560 zio->io_error = zio->io_child_error[c];
564 * ==========================================================================
565 * Create the various types of I/O (read, write, free, etc)
566 * ==========================================================================
569 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
570 void *data, uint64_t size, zio_done_func_t *done, void *private,
571 zio_type_t type, int priority, enum zio_flag flags,
572 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
573 enum zio_stage stage, enum zio_stage pipeline)
577 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
578 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
579 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
581 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
582 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
583 ASSERT(vd || stage == ZIO_STAGE_OPEN);
585 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
588 zio->io_child_type = ZIO_CHILD_VDEV;
589 else if (flags & ZIO_FLAG_GANG_CHILD)
590 zio->io_child_type = ZIO_CHILD_GANG;
591 else if (flags & ZIO_FLAG_DDT_CHILD)
592 zio->io_child_type = ZIO_CHILD_DDT;
594 zio->io_child_type = ZIO_CHILD_LOGICAL;
597 zio->io_logical = NULL;
598 zio->io_bp = (blkptr_t *)bp;
599 zio->io_bp_copy = *bp;
600 zio->io_bp_orig = *bp;
601 if (type != ZIO_TYPE_WRITE ||
602 zio->io_child_type == ZIO_CHILD_DDT)
603 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
604 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
605 zio->io_logical = zio;
606 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
607 pipeline |= ZIO_GANG_STAGES;
609 zio->io_logical = NULL;
611 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
612 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
617 zio->io_ready = NULL;
619 zio->io_private = private;
620 zio->io_prev_space_delta = 0;
622 zio->io_priority = priority;
625 zio->io_vsd_ops = NULL;
626 zio->io_offset = offset;
627 zio->io_deadline = 0;
628 zio->io_timestamp = 0;
631 zio->io_orig_data = zio->io_data = data;
632 zio->io_orig_size = zio->io_size = size;
633 zio->io_orig_flags = zio->io_flags = flags;
634 zio->io_orig_stage = zio->io_stage = stage;
635 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
636 bzero(&zio->io_prop, sizeof (zio_prop_t));
638 zio->io_reexecute = 0;
639 zio->io_bp_override = NULL;
640 zio->io_walk_link = NULL;
641 zio->io_transform_stack = NULL;
643 zio->io_child_count = 0;
644 zio->io_parent_count = 0;
645 zio->io_stall = NULL;
646 zio->io_gang_leader = NULL;
647 zio->io_gang_tree = NULL;
648 zio->io_executor = NULL;
649 zio->io_waiter = NULL;
650 zio->io_cksum_report = NULL;
652 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
653 bzero(zio->io_children,
654 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
655 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
657 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
658 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
661 zio->io_bookmark = *zb;
664 if (zio->io_logical == NULL)
665 zio->io_logical = pio->io_logical;
666 if (zio->io_child_type == ZIO_CHILD_GANG)
667 zio->io_gang_leader = pio->io_gang_leader;
668 zio_add_child(pio, zio);
671 taskq_init_ent(&zio->io_tqent);
677 zio_destroy(zio_t *zio)
679 kmem_cache_free(zio_cache, zio);
683 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
684 void *private, enum zio_flag flags)
688 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
689 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
690 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
696 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
698 return (zio_null(NULL, spa, NULL, done, private, flags));
702 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
703 void *data, uint64_t size, zio_done_func_t *done, void *private,
704 int priority, enum zio_flag flags, const zbookmark_t *zb)
708 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
709 data, size, done, private,
710 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
711 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
712 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
718 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
719 void *data, uint64_t size, const zio_prop_t *zp,
720 zio_done_func_t *ready, zio_done_func_t *done, void *private,
721 int priority, enum zio_flag flags, const zbookmark_t *zb)
725 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
726 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
727 zp->zp_compress >= ZIO_COMPRESS_OFF &&
728 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
729 DMU_OT_IS_VALID(zp->zp_type) &&
732 zp->zp_copies <= spa_max_replication(spa) &&
734 zp->zp_dedup_verify <= 1);
736 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
737 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
738 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
739 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
741 zio->io_ready = ready;
748 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
749 uint64_t size, zio_done_func_t *done, void *private, int priority,
750 enum zio_flag flags, zbookmark_t *zb)
754 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
755 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
756 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
762 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
764 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
765 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
766 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
767 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
769 zio->io_prop.zp_copies = copies;
770 zio->io_bp_override = bp;
774 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
776 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
780 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
785 dprintf_bp(bp, "freeing in txg %llu, pass %u",
786 (longlong_t)txg, spa->spa_sync_pass);
788 ASSERT(!BP_IS_HOLE(bp));
789 ASSERT(spa_syncing_txg(spa) == txg);
790 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
792 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
793 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
794 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
800 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
801 zio_done_func_t *done, void *private, enum zio_flag flags)
806 * A claim is an allocation of a specific block. Claims are needed
807 * to support immediate writes in the intent log. The issue is that
808 * immediate writes contain committed data, but in a txg that was
809 * *not* committed. Upon opening the pool after an unclean shutdown,
810 * the intent log claims all blocks that contain immediate write data
811 * so that the SPA knows they're in use.
813 * All claims *must* be resolved in the first txg -- before the SPA
814 * starts allocating blocks -- so that nothing is allocated twice.
815 * If txg == 0 we just verify that the block is claimable.
817 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
818 ASSERT(txg == spa_first_txg(spa) || txg == 0);
819 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
821 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
822 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
823 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
829 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
830 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
835 if (vd->vdev_children == 0) {
836 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
837 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
838 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
842 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
844 for (c = 0; c < vd->vdev_children; c++)
845 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
846 done, private, priority, flags));
853 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
854 void *data, int checksum, zio_done_func_t *done, void *private,
855 int priority, enum zio_flag flags, boolean_t labels)
859 ASSERT(vd->vdev_children == 0);
860 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
861 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
862 ASSERT3U(offset + size, <=, vd->vdev_psize);
864 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
865 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
866 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
868 zio->io_prop.zp_checksum = checksum;
874 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
875 void *data, int checksum, zio_done_func_t *done, void *private,
876 int priority, enum zio_flag flags, boolean_t labels)
880 ASSERT(vd->vdev_children == 0);
881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
883 ASSERT3U(offset + size, <=, vd->vdev_psize);
885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
886 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
887 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
889 zio->io_prop.zp_checksum = checksum;
891 if (zio_checksum_table[checksum].ci_eck) {
893 * zec checksums are necessarily destructive -- they modify
894 * the end of the write buffer to hold the verifier/checksum.
895 * Therefore, we must make a local copy in case the data is
896 * being written to multiple places in parallel.
898 void *wbuf = zio_buf_alloc(size);
899 bcopy(data, wbuf, size);
900 zio_push_transform(zio, wbuf, size, size, NULL);
907 * Create a child I/O to do some work for us.
910 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
911 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
912 zio_done_func_t *done, void *private)
914 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
917 ASSERT(vd->vdev_parent ==
918 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
920 if (type == ZIO_TYPE_READ && bp != NULL) {
922 * If we have the bp, then the child should perform the
923 * checksum and the parent need not. This pushes error
924 * detection as close to the leaves as possible and
925 * eliminates redundant checksums in the interior nodes.
927 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
928 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
931 if (vd->vdev_children == 0)
932 offset += VDEV_LABEL_START_SIZE;
934 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
937 * If we've decided to do a repair, the write is not speculative --
938 * even if the original read was.
940 if (flags & ZIO_FLAG_IO_REPAIR)
941 flags &= ~ZIO_FLAG_SPECULATIVE;
943 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
944 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
945 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
951 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
952 int type, int priority, enum zio_flag flags,
953 zio_done_func_t *done, void *private)
957 ASSERT(vd->vdev_ops->vdev_op_leaf);
959 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
960 data, size, done, private, type, priority,
961 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
963 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
969 zio_flush(zio_t *zio, vdev_t *vd)
971 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
972 NULL, NULL, ZIO_PRIORITY_NOW,
973 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
977 zio_shrink(zio_t *zio, uint64_t size)
979 ASSERT(zio->io_executor == NULL);
980 ASSERT(zio->io_orig_size == zio->io_size);
981 ASSERT(size <= zio->io_size);
984 * We don't shrink for raidz because of problems with the
985 * reconstruction when reading back less than the block size.
986 * Note, BP_IS_RAIDZ() assumes no compression.
988 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
989 if (!BP_IS_RAIDZ(zio->io_bp))
990 zio->io_orig_size = zio->io_size = size;
994 * ==========================================================================
995 * Prepare to read and write logical blocks
996 * ==========================================================================
1000 zio_read_bp_init(zio_t *zio)
1002 blkptr_t *bp = zio->io_bp;
1004 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1005 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1006 !(zio->io_flags & ZIO_FLAG_RAW)) {
1007 uint64_t psize = BP_GET_PSIZE(bp);
1008 void *cbuf = zio_buf_alloc(psize);
1010 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1013 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1014 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1016 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1017 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1019 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1020 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1022 return (ZIO_PIPELINE_CONTINUE);
1026 zio_write_bp_init(zio_t *zio)
1028 spa_t *spa = zio->io_spa;
1029 zio_prop_t *zp = &zio->io_prop;
1030 enum zio_compress compress = zp->zp_compress;
1031 blkptr_t *bp = zio->io_bp;
1032 uint64_t lsize = zio->io_size;
1033 uint64_t psize = lsize;
1037 * If our children haven't all reached the ready stage,
1038 * wait for them and then repeat this pipeline stage.
1040 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1041 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1042 return (ZIO_PIPELINE_STOP);
1044 if (!IO_IS_ALLOCATING(zio))
1045 return (ZIO_PIPELINE_CONTINUE);
1047 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1049 if (zio->io_bp_override) {
1050 ASSERT(bp->blk_birth != zio->io_txg);
1051 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1053 *bp = *zio->io_bp_override;
1054 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1056 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1057 return (ZIO_PIPELINE_CONTINUE);
1059 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1060 zp->zp_dedup_verify);
1062 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1063 BP_SET_DEDUP(bp, 1);
1064 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1065 return (ZIO_PIPELINE_CONTINUE);
1067 zio->io_bp_override = NULL;
1071 if (bp->blk_birth == zio->io_txg) {
1073 * We're rewriting an existing block, which means we're
1074 * working on behalf of spa_sync(). For spa_sync() to
1075 * converge, it must eventually be the case that we don't
1076 * have to allocate new blocks. But compression changes
1077 * the blocksize, which forces a reallocate, and makes
1078 * convergence take longer. Therefore, after the first
1079 * few passes, stop compressing to ensure convergence.
1081 pass = spa_sync_pass(spa);
1083 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1084 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1085 ASSERT(!BP_GET_DEDUP(bp));
1087 if (pass >= zfs_sync_pass_dont_compress)
1088 compress = ZIO_COMPRESS_OFF;
1090 /* Make sure someone doesn't change their mind on overwrites */
1091 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1092 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1095 if (compress != ZIO_COMPRESS_OFF) {
1096 void *cbuf = zio_buf_alloc(lsize);
1097 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1098 if (psize == 0 || psize == lsize) {
1099 compress = ZIO_COMPRESS_OFF;
1100 zio_buf_free(cbuf, lsize);
1102 ASSERT(psize < lsize);
1103 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1108 * The final pass of spa_sync() must be all rewrites, but the first
1109 * few passes offer a trade-off: allocating blocks defers convergence,
1110 * but newly allocated blocks are sequential, so they can be written
1111 * to disk faster. Therefore, we allow the first few passes of
1112 * spa_sync() to allocate new blocks, but force rewrites after that.
1113 * There should only be a handful of blocks after pass 1 in any case.
1115 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1116 pass >= zfs_sync_pass_rewrite) {
1117 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1119 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1120 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1123 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1127 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1129 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1130 BP_SET_LSIZE(bp, lsize);
1131 BP_SET_PSIZE(bp, psize);
1132 BP_SET_COMPRESS(bp, compress);
1133 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1134 BP_SET_TYPE(bp, zp->zp_type);
1135 BP_SET_LEVEL(bp, zp->zp_level);
1136 BP_SET_DEDUP(bp, zp->zp_dedup);
1137 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1139 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1140 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1141 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1145 return (ZIO_PIPELINE_CONTINUE);
1149 zio_free_bp_init(zio_t *zio)
1151 blkptr_t *bp = zio->io_bp;
1153 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1154 if (BP_GET_DEDUP(bp))
1155 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1158 return (ZIO_PIPELINE_CONTINUE);
1162 * ==========================================================================
1163 * Execute the I/O pipeline
1164 * ==========================================================================
1168 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1170 spa_t *spa = zio->io_spa;
1171 zio_type_t t = zio->io_type;
1172 int flags = (cutinline ? TQ_FRONT : 0);
1175 * If we're a config writer or a probe, the normal issue and
1176 * interrupt threads may all be blocked waiting for the config lock.
1177 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1179 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1183 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1185 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1189 * If this is a high priority I/O, then use the high priority taskq if
1192 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1193 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1196 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1199 * NB: We are assuming that the zio can only be dispatched
1200 * to a single taskq at a time. It would be a grievous error
1201 * to dispatch the zio to another taskq at the same time.
1203 ASSERT(taskq_empty_ent(&zio->io_tqent));
1204 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1205 flags, &zio->io_tqent);
1209 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1211 kthread_t *executor = zio->io_executor;
1212 spa_t *spa = zio->io_spa;
1215 for (t = 0; t < ZIO_TYPES; t++) {
1216 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1218 for (i = 0; i < tqs->stqs_count; i++) {
1219 if (taskq_member(tqs->stqs_taskq[i], executor))
1228 zio_issue_async(zio_t *zio)
1230 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1232 return (ZIO_PIPELINE_STOP);
1236 zio_interrupt(zio_t *zio)
1238 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1242 * Execute the I/O pipeline until one of the following occurs:
1243 * (1) the I/O completes; (2) the pipeline stalls waiting for
1244 * dependent child I/Os; (3) the I/O issues, so we're waiting
1245 * for an I/O completion interrupt; (4) the I/O is delegated by
1246 * vdev-level caching or aggregation; (5) the I/O is deferred
1247 * due to vdev-level queueing; (6) the I/O is handed off to
1248 * another thread. In all cases, the pipeline stops whenever
1249 * there's no CPU work; it never burns a thread in cv_wait().
1251 * There's no locking on io_stage because there's no legitimate way
1252 * for multiple threads to be attempting to process the same I/O.
1254 static zio_pipe_stage_t *zio_pipeline[];
1257 * zio_execute() is a wrapper around the static function
1258 * __zio_execute() so that we can force __zio_execute() to be
1259 * inlined. This reduces stack overhead which is important
1260 * because __zio_execute() is called recursively in several zio
1261 * code paths. zio_execute() itself cannot be inlined because
1262 * it is externally visible.
1265 zio_execute(zio_t *zio)
1270 __attribute__((always_inline))
1272 __zio_execute(zio_t *zio)
1274 zio->io_executor = curthread;
1276 while (zio->io_stage < ZIO_STAGE_DONE) {
1277 enum zio_stage pipeline = zio->io_pipeline;
1278 enum zio_stage stage = zio->io_stage;
1283 ASSERT(!MUTEX_HELD(&zio->io_lock));
1284 ASSERT(ISP2(stage));
1285 ASSERT(zio->io_stall == NULL);
1289 } while ((stage & pipeline) == 0);
1291 ASSERT(stage <= ZIO_STAGE_DONE);
1293 dp = spa_get_dsl(zio->io_spa);
1294 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1295 zio_requeue_io_start_cut_in_line : B_FALSE;
1298 * If we are in interrupt context and this pipeline stage
1299 * will grab a config lock that is held across I/O,
1300 * or may wait for an I/O that needs an interrupt thread
1301 * to complete, issue async to avoid deadlock.
1303 * For VDEV_IO_START, we cut in line so that the io will
1304 * be sent to disk promptly.
1306 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1307 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1308 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1314 * If we executing in the context of the tx_sync_thread,
1315 * or we are performing pool initialization outside of a
1316 * zio_taskq[ZIO_TASKQ_ISSUE] context. Then issue the zio
1317 * async to minimize stack usage for these deep call paths.
1319 if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
1320 (dp && spa_is_initializing(dp->dp_spa) &&
1321 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE))) {
1322 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1327 zio->io_stage = stage;
1328 rv = zio_pipeline[highbit(stage) - 1](zio);
1330 if (rv == ZIO_PIPELINE_STOP)
1333 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1339 * ==========================================================================
1340 * Initiate I/O, either sync or async
1341 * ==========================================================================
1344 zio_wait(zio_t *zio)
1348 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1349 ASSERT(zio->io_executor == NULL);
1351 zio->io_waiter = curthread;
1355 mutex_enter(&zio->io_lock);
1356 while (zio->io_executor != NULL)
1357 cv_wait_io(&zio->io_cv, &zio->io_lock);
1358 mutex_exit(&zio->io_lock);
1360 error = zio->io_error;
1367 zio_nowait(zio_t *zio)
1369 ASSERT(zio->io_executor == NULL);
1371 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1372 zio_unique_parent(zio) == NULL) {
1374 * This is a logical async I/O with no parent to wait for it.
1375 * We add it to the spa_async_root_zio "Godfather" I/O which
1376 * will ensure they complete prior to unloading the pool.
1378 spa_t *spa = zio->io_spa;
1380 zio_add_child(spa->spa_async_zio_root, zio);
1387 * ==========================================================================
1388 * Reexecute or suspend/resume failed I/O
1389 * ==========================================================================
1393 zio_reexecute(zio_t *pio)
1395 zio_t *cio, *cio_next;
1398 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1399 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1400 ASSERT(pio->io_gang_leader == NULL);
1401 ASSERT(pio->io_gang_tree == NULL);
1403 pio->io_flags = pio->io_orig_flags;
1404 pio->io_stage = pio->io_orig_stage;
1405 pio->io_pipeline = pio->io_orig_pipeline;
1406 pio->io_reexecute = 0;
1408 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1409 pio->io_state[w] = 0;
1410 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1411 pio->io_child_error[c] = 0;
1413 if (IO_IS_ALLOCATING(pio))
1414 BP_ZERO(pio->io_bp);
1417 * As we reexecute pio's children, new children could be created.
1418 * New children go to the head of pio's io_child_list, however,
1419 * so we will (correctly) not reexecute them. The key is that
1420 * the remainder of pio's io_child_list, from 'cio_next' onward,
1421 * cannot be affected by any side effects of reexecuting 'cio'.
1423 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1424 cio_next = zio_walk_children(pio);
1425 mutex_enter(&pio->io_lock);
1426 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1427 pio->io_children[cio->io_child_type][w]++;
1428 mutex_exit(&pio->io_lock);
1433 * Now that all children have been reexecuted, execute the parent.
1434 * We don't reexecute "The Godfather" I/O here as it's the
1435 * responsibility of the caller to wait on him.
1437 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1442 zio_suspend(spa_t *spa, zio_t *zio)
1444 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1445 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1446 "failure and the failure mode property for this pool "
1447 "is set to panic.", spa_name(spa));
1449 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1451 mutex_enter(&spa->spa_suspend_lock);
1453 if (spa->spa_suspend_zio_root == NULL)
1454 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1455 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1456 ZIO_FLAG_GODFATHER);
1458 spa->spa_suspended = B_TRUE;
1461 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1462 ASSERT(zio != spa->spa_suspend_zio_root);
1463 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1464 ASSERT(zio_unique_parent(zio) == NULL);
1465 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1466 zio_add_child(spa->spa_suspend_zio_root, zio);
1469 mutex_exit(&spa->spa_suspend_lock);
1473 zio_resume(spa_t *spa)
1478 * Reexecute all previously suspended i/o.
1480 mutex_enter(&spa->spa_suspend_lock);
1481 spa->spa_suspended = B_FALSE;
1482 cv_broadcast(&spa->spa_suspend_cv);
1483 pio = spa->spa_suspend_zio_root;
1484 spa->spa_suspend_zio_root = NULL;
1485 mutex_exit(&spa->spa_suspend_lock);
1491 return (zio_wait(pio));
1495 zio_resume_wait(spa_t *spa)
1497 mutex_enter(&spa->spa_suspend_lock);
1498 while (spa_suspended(spa))
1499 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1500 mutex_exit(&spa->spa_suspend_lock);
1504 * ==========================================================================
1507 * A gang block is a collection of small blocks that looks to the DMU
1508 * like one large block. When zio_dva_allocate() cannot find a block
1509 * of the requested size, due to either severe fragmentation or the pool
1510 * being nearly full, it calls zio_write_gang_block() to construct the
1511 * block from smaller fragments.
1513 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1514 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1515 * an indirect block: it's an array of block pointers. It consumes
1516 * only one sector and hence is allocatable regardless of fragmentation.
1517 * The gang header's bps point to its gang members, which hold the data.
1519 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1520 * as the verifier to ensure uniqueness of the SHA256 checksum.
1521 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1522 * not the gang header. This ensures that data block signatures (needed for
1523 * deduplication) are independent of how the block is physically stored.
1525 * Gang blocks can be nested: a gang member may itself be a gang block.
1526 * Thus every gang block is a tree in which root and all interior nodes are
1527 * gang headers, and the leaves are normal blocks that contain user data.
1528 * The root of the gang tree is called the gang leader.
1530 * To perform any operation (read, rewrite, free, claim) on a gang block,
1531 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1532 * in the io_gang_tree field of the original logical i/o by recursively
1533 * reading the gang leader and all gang headers below it. This yields
1534 * an in-core tree containing the contents of every gang header and the
1535 * bps for every constituent of the gang block.
1537 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1538 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1539 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1540 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1541 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1542 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1543 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1544 * of the gang header plus zio_checksum_compute() of the data to update the
1545 * gang header's blk_cksum as described above.
1547 * The two-phase assemble/issue model solves the problem of partial failure --
1548 * what if you'd freed part of a gang block but then couldn't read the
1549 * gang header for another part? Assembling the entire gang tree first
1550 * ensures that all the necessary gang header I/O has succeeded before
1551 * starting the actual work of free, claim, or write. Once the gang tree
1552 * is assembled, free and claim are in-memory operations that cannot fail.
1554 * In the event that a gang write fails, zio_dva_unallocate() walks the
1555 * gang tree to immediately free (i.e. insert back into the space map)
1556 * everything we've allocated. This ensures that we don't get ENOSPC
1557 * errors during repeated suspend/resume cycles due to a flaky device.
1559 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1560 * the gang tree, we won't modify the block, so we can safely defer the free
1561 * (knowing that the block is still intact). If we *can* assemble the gang
1562 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1563 * each constituent bp and we can allocate a new block on the next sync pass.
1565 * In all cases, the gang tree allows complete recovery from partial failure.
1566 * ==========================================================================
1570 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1575 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1576 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1577 &pio->io_bookmark));
1581 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1586 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1587 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1588 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1590 * As we rewrite each gang header, the pipeline will compute
1591 * a new gang block header checksum for it; but no one will
1592 * compute a new data checksum, so we do that here. The one
1593 * exception is the gang leader: the pipeline already computed
1594 * its data checksum because that stage precedes gang assembly.
1595 * (Presently, nothing actually uses interior data checksums;
1596 * this is just good hygiene.)
1598 if (gn != pio->io_gang_leader->io_gang_tree) {
1599 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1600 data, BP_GET_PSIZE(bp));
1603 * If we are here to damage data for testing purposes,
1604 * leave the GBH alone so that we can detect the damage.
1606 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1607 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1609 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1610 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1611 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1619 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1621 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1622 ZIO_GANG_CHILD_FLAGS(pio)));
1627 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1629 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1630 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1633 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1642 static void zio_gang_tree_assemble_done(zio_t *zio);
1644 static zio_gang_node_t *
1645 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1647 zio_gang_node_t *gn;
1649 ASSERT(*gnpp == NULL);
1651 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1652 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1659 zio_gang_node_free(zio_gang_node_t **gnpp)
1661 zio_gang_node_t *gn = *gnpp;
1664 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1665 ASSERT(gn->gn_child[g] == NULL);
1667 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1668 kmem_free(gn, sizeof (*gn));
1673 zio_gang_tree_free(zio_gang_node_t **gnpp)
1675 zio_gang_node_t *gn = *gnpp;
1681 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1682 zio_gang_tree_free(&gn->gn_child[g]);
1684 zio_gang_node_free(gnpp);
1688 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1690 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1692 ASSERT(gio->io_gang_leader == gio);
1693 ASSERT(BP_IS_GANG(bp));
1695 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1696 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1697 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1701 zio_gang_tree_assemble_done(zio_t *zio)
1703 zio_t *gio = zio->io_gang_leader;
1704 zio_gang_node_t *gn = zio->io_private;
1705 blkptr_t *bp = zio->io_bp;
1708 ASSERT(gio == zio_unique_parent(zio));
1709 ASSERT(zio->io_child_count == 0);
1714 if (BP_SHOULD_BYTESWAP(bp))
1715 byteswap_uint64_array(zio->io_data, zio->io_size);
1717 ASSERT(zio->io_data == gn->gn_gbh);
1718 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1719 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1721 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1722 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1723 if (!BP_IS_GANG(gbp))
1725 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1730 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1732 zio_t *gio = pio->io_gang_leader;
1736 ASSERT(BP_IS_GANG(bp) == !!gn);
1737 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1738 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1741 * If you're a gang header, your data is in gn->gn_gbh.
1742 * If you're a gang member, your data is in 'data' and gn == NULL.
1744 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1747 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1749 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1750 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1751 if (BP_IS_HOLE(gbp))
1753 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1754 data = (char *)data + BP_GET_PSIZE(gbp);
1758 if (gn == gio->io_gang_tree)
1759 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1766 zio_gang_assemble(zio_t *zio)
1768 blkptr_t *bp = zio->io_bp;
1770 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1771 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1773 zio->io_gang_leader = zio;
1775 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1777 return (ZIO_PIPELINE_CONTINUE);
1781 zio_gang_issue(zio_t *zio)
1783 blkptr_t *bp = zio->io_bp;
1785 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1786 return (ZIO_PIPELINE_STOP);
1788 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1789 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1791 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1792 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1794 zio_gang_tree_free(&zio->io_gang_tree);
1796 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1798 return (ZIO_PIPELINE_CONTINUE);
1802 zio_write_gang_member_ready(zio_t *zio)
1804 zio_t *pio = zio_unique_parent(zio);
1805 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1806 dva_t *cdva = zio->io_bp->blk_dva;
1807 dva_t *pdva = pio->io_bp->blk_dva;
1811 if (BP_IS_HOLE(zio->io_bp))
1814 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1816 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1817 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1818 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1819 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1820 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1822 mutex_enter(&pio->io_lock);
1823 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1824 ASSERT(DVA_GET_GANG(&pdva[d]));
1825 asize = DVA_GET_ASIZE(&pdva[d]);
1826 asize += DVA_GET_ASIZE(&cdva[d]);
1827 DVA_SET_ASIZE(&pdva[d], asize);
1829 mutex_exit(&pio->io_lock);
1833 zio_write_gang_block(zio_t *pio)
1835 spa_t *spa = pio->io_spa;
1836 blkptr_t *bp = pio->io_bp;
1837 zio_t *gio = pio->io_gang_leader;
1839 zio_gang_node_t *gn, **gnpp;
1840 zio_gbh_phys_t *gbh;
1841 uint64_t txg = pio->io_txg;
1842 uint64_t resid = pio->io_size;
1844 int copies = gio->io_prop.zp_copies;
1845 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1849 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1850 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1851 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1853 pio->io_error = error;
1854 return (ZIO_PIPELINE_CONTINUE);
1858 gnpp = &gio->io_gang_tree;
1860 gnpp = pio->io_private;
1861 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1864 gn = zio_gang_node_alloc(gnpp);
1866 bzero(gbh, SPA_GANGBLOCKSIZE);
1869 * Create the gang header.
1871 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1872 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1875 * Create and nowait the gang children.
1877 for (g = 0; resid != 0; resid -= lsize, g++) {
1878 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1880 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1882 zp.zp_checksum = gio->io_prop.zp_checksum;
1883 zp.zp_compress = ZIO_COMPRESS_OFF;
1884 zp.zp_type = DMU_OT_NONE;
1886 zp.zp_copies = gio->io_prop.zp_copies;
1888 zp.zp_dedup_verify = 0;
1890 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1891 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1892 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1893 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1894 &pio->io_bookmark));
1898 * Set pio's pipeline to just wait for zio to finish.
1900 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1903 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1905 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1909 return (ZIO_PIPELINE_CONTINUE);
1913 * ==========================================================================
1915 * ==========================================================================
1918 zio_ddt_child_read_done(zio_t *zio)
1920 blkptr_t *bp = zio->io_bp;
1921 ddt_entry_t *dde = zio->io_private;
1923 zio_t *pio = zio_unique_parent(zio);
1925 mutex_enter(&pio->io_lock);
1926 ddp = ddt_phys_select(dde, bp);
1927 if (zio->io_error == 0)
1928 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1929 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1930 dde->dde_repair_data = zio->io_data;
1932 zio_buf_free(zio->io_data, zio->io_size);
1933 mutex_exit(&pio->io_lock);
1937 zio_ddt_read_start(zio_t *zio)
1939 blkptr_t *bp = zio->io_bp;
1942 ASSERT(BP_GET_DEDUP(bp));
1943 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1944 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1946 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1947 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1948 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1949 ddt_phys_t *ddp = dde->dde_phys;
1950 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1953 ASSERT(zio->io_vsd == NULL);
1956 if (ddp_self == NULL)
1957 return (ZIO_PIPELINE_CONTINUE);
1959 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1960 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1962 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1964 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1965 zio_buf_alloc(zio->io_size), zio->io_size,
1966 zio_ddt_child_read_done, dde, zio->io_priority,
1967 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1968 &zio->io_bookmark));
1970 return (ZIO_PIPELINE_CONTINUE);
1973 zio_nowait(zio_read(zio, zio->io_spa, bp,
1974 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1975 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1977 return (ZIO_PIPELINE_CONTINUE);
1981 zio_ddt_read_done(zio_t *zio)
1983 blkptr_t *bp = zio->io_bp;
1985 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1986 return (ZIO_PIPELINE_STOP);
1988 ASSERT(BP_GET_DEDUP(bp));
1989 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1990 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1992 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1993 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1994 ddt_entry_t *dde = zio->io_vsd;
1996 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1997 return (ZIO_PIPELINE_CONTINUE);
2000 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2001 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2002 return (ZIO_PIPELINE_STOP);
2004 if (dde->dde_repair_data != NULL) {
2005 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2006 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2008 ddt_repair_done(ddt, dde);
2012 ASSERT(zio->io_vsd == NULL);
2014 return (ZIO_PIPELINE_CONTINUE);
2018 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2020 spa_t *spa = zio->io_spa;
2024 * Note: we compare the original data, not the transformed data,
2025 * because when zio->io_bp is an override bp, we will not have
2026 * pushed the I/O transforms. That's an important optimization
2027 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2029 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2030 zio_t *lio = dde->dde_lead_zio[p];
2033 return (lio->io_orig_size != zio->io_orig_size ||
2034 bcmp(zio->io_orig_data, lio->io_orig_data,
2035 zio->io_orig_size) != 0);
2039 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2040 ddt_phys_t *ddp = &dde->dde_phys[p];
2042 if (ddp->ddp_phys_birth != 0) {
2043 arc_buf_t *abuf = NULL;
2044 uint32_t aflags = ARC_WAIT;
2045 blkptr_t blk = *zio->io_bp;
2048 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2052 error = arc_read_nolock(NULL, spa, &blk,
2053 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2054 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2055 &aflags, &zio->io_bookmark);
2058 if (arc_buf_size(abuf) != zio->io_orig_size ||
2059 bcmp(abuf->b_data, zio->io_orig_data,
2060 zio->io_orig_size) != 0)
2062 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2066 return (error != 0);
2074 zio_ddt_child_write_ready(zio_t *zio)
2076 int p = zio->io_prop.zp_copies;
2077 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2078 ddt_entry_t *dde = zio->io_private;
2079 ddt_phys_t *ddp = &dde->dde_phys[p];
2087 ASSERT(dde->dde_lead_zio[p] == zio);
2089 ddt_phys_fill(ddp, zio->io_bp);
2091 while ((pio = zio_walk_parents(zio)) != NULL)
2092 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2098 zio_ddt_child_write_done(zio_t *zio)
2100 int p = zio->io_prop.zp_copies;
2101 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2102 ddt_entry_t *dde = zio->io_private;
2103 ddt_phys_t *ddp = &dde->dde_phys[p];
2107 ASSERT(ddp->ddp_refcnt == 0);
2108 ASSERT(dde->dde_lead_zio[p] == zio);
2109 dde->dde_lead_zio[p] = NULL;
2111 if (zio->io_error == 0) {
2112 while (zio_walk_parents(zio) != NULL)
2113 ddt_phys_addref(ddp);
2115 ddt_phys_clear(ddp);
2122 zio_ddt_ditto_write_done(zio_t *zio)
2124 int p = DDT_PHYS_DITTO;
2125 blkptr_t *bp = zio->io_bp;
2126 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2127 ddt_entry_t *dde = zio->io_private;
2128 ddt_phys_t *ddp = &dde->dde_phys[p];
2129 ddt_key_t *ddk = &dde->dde_key;
2130 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2134 ASSERT(ddp->ddp_refcnt == 0);
2135 ASSERT(dde->dde_lead_zio[p] == zio);
2136 dde->dde_lead_zio[p] = NULL;
2138 if (zio->io_error == 0) {
2139 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2140 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2141 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2142 if (ddp->ddp_phys_birth != 0)
2143 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2144 ddt_phys_fill(ddp, bp);
2151 zio_ddt_write(zio_t *zio)
2153 spa_t *spa = zio->io_spa;
2154 blkptr_t *bp = zio->io_bp;
2155 uint64_t txg = zio->io_txg;
2156 zio_prop_t *zp = &zio->io_prop;
2157 int p = zp->zp_copies;
2161 ddt_t *ddt = ddt_select(spa, bp);
2165 ASSERT(BP_GET_DEDUP(bp));
2166 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2167 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2170 dde = ddt_lookup(ddt, bp, B_TRUE);
2171 ddp = &dde->dde_phys[p];
2173 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2175 * If we're using a weak checksum, upgrade to a strong checksum
2176 * and try again. If we're already using a strong checksum,
2177 * we can't resolve it, so just convert to an ordinary write.
2178 * (And automatically e-mail a paper to Nature?)
2180 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2181 zp->zp_checksum = spa_dedup_checksum(spa);
2182 zio_pop_transforms(zio);
2183 zio->io_stage = ZIO_STAGE_OPEN;
2188 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2190 return (ZIO_PIPELINE_CONTINUE);
2193 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2194 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2196 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2197 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2198 zio_prop_t czp = *zp;
2200 czp.zp_copies = ditto_copies;
2203 * If we arrived here with an override bp, we won't have run
2204 * the transform stack, so we won't have the data we need to
2205 * generate a child i/o. So, toss the override bp and restart.
2206 * This is safe, because using the override bp is just an
2207 * optimization; and it's rare, so the cost doesn't matter.
2209 if (zio->io_bp_override) {
2210 zio_pop_transforms(zio);
2211 zio->io_stage = ZIO_STAGE_OPEN;
2212 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2213 zio->io_bp_override = NULL;
2216 return (ZIO_PIPELINE_CONTINUE);
2219 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2220 zio->io_orig_size, &czp, NULL,
2221 zio_ddt_ditto_write_done, dde, zio->io_priority,
2222 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2224 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2225 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2228 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2229 if (ddp->ddp_phys_birth != 0)
2230 ddt_bp_fill(ddp, bp, txg);
2231 if (dde->dde_lead_zio[p] != NULL)
2232 zio_add_child(zio, dde->dde_lead_zio[p]);
2234 ddt_phys_addref(ddp);
2235 } else if (zio->io_bp_override) {
2236 ASSERT(bp->blk_birth == txg);
2237 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2238 ddt_phys_fill(ddp, bp);
2239 ddt_phys_addref(ddp);
2241 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2242 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2243 zio_ddt_child_write_done, dde, zio->io_priority,
2244 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2246 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2247 dde->dde_lead_zio[p] = cio;
2257 return (ZIO_PIPELINE_CONTINUE);
2260 ddt_entry_t *freedde; /* for debugging */
2263 zio_ddt_free(zio_t *zio)
2265 spa_t *spa = zio->io_spa;
2266 blkptr_t *bp = zio->io_bp;
2267 ddt_t *ddt = ddt_select(spa, bp);
2271 ASSERT(BP_GET_DEDUP(bp));
2272 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2275 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2277 ddp = ddt_phys_select(dde, bp);
2279 ddt_phys_decref(ddp);
2283 return (ZIO_PIPELINE_CONTINUE);
2287 * ==========================================================================
2288 * Allocate and free blocks
2289 * ==========================================================================
2292 zio_dva_allocate(zio_t *zio)
2294 spa_t *spa = zio->io_spa;
2295 metaslab_class_t *mc = spa_normal_class(spa);
2296 blkptr_t *bp = zio->io_bp;
2300 if (zio->io_gang_leader == NULL) {
2301 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2302 zio->io_gang_leader = zio;
2305 ASSERT(BP_IS_HOLE(bp));
2306 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2307 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2308 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2309 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2312 * The dump device does not support gang blocks so allocation on
2313 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2314 * the "fast" gang feature.
2316 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2317 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2318 METASLAB_GANG_CHILD : 0;
2319 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2320 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2321 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2324 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2325 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2327 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2328 return (zio_write_gang_block(zio));
2329 zio->io_error = error;
2332 return (ZIO_PIPELINE_CONTINUE);
2336 zio_dva_free(zio_t *zio)
2338 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2340 return (ZIO_PIPELINE_CONTINUE);
2344 zio_dva_claim(zio_t *zio)
2348 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2350 zio->io_error = error;
2352 return (ZIO_PIPELINE_CONTINUE);
2356 * Undo an allocation. This is used by zio_done() when an I/O fails
2357 * and we want to give back the block we just allocated.
2358 * This handles both normal blocks and gang blocks.
2361 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2365 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2366 ASSERT(zio->io_bp_override == NULL);
2368 if (!BP_IS_HOLE(bp))
2369 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2372 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2373 zio_dva_unallocate(zio, gn->gn_child[g],
2374 &gn->gn_gbh->zg_blkptr[g]);
2380 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2383 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2388 ASSERT(txg > spa_syncing_txg(spa));
2391 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2392 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2393 * when allocating them.
2396 error = metaslab_alloc(spa, spa_log_class(spa), size,
2397 new_bp, 1, txg, NULL,
2398 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2402 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2403 new_bp, 1, txg, NULL,
2404 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2408 BP_SET_LSIZE(new_bp, size);
2409 BP_SET_PSIZE(new_bp, size);
2410 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2411 BP_SET_CHECKSUM(new_bp,
2412 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2413 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2414 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2415 BP_SET_LEVEL(new_bp, 0);
2416 BP_SET_DEDUP(new_bp, 0);
2417 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2424 * Free an intent log block.
2427 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2429 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2430 ASSERT(!BP_IS_GANG(bp));
2432 zio_free(spa, txg, bp);
2436 * ==========================================================================
2437 * Read and write to physical devices
2438 * ==========================================================================
2441 zio_vdev_io_start(zio_t *zio)
2443 vdev_t *vd = zio->io_vd;
2445 spa_t *spa = zio->io_spa;
2447 ASSERT(zio->io_error == 0);
2448 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2451 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2452 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2455 * The mirror_ops handle multiple DVAs in a single BP.
2457 return (vdev_mirror_ops.vdev_op_io_start(zio));
2461 * We keep track of time-sensitive I/Os so that the scan thread
2462 * can quickly react to certain workloads. In particular, we care
2463 * about non-scrubbing, top-level reads and writes with the following
2465 * - synchronous writes of user data to non-slog devices
2466 * - any reads of user data
2467 * When these conditions are met, adjust the timestamp of spa_last_io
2468 * which allows the scan thread to adjust its workload accordingly.
2470 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2471 vd == vd->vdev_top && !vd->vdev_islog &&
2472 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2473 zio->io_txg != spa_syncing_txg(spa)) {
2474 uint64_t old = spa->spa_last_io;
2475 uint64_t new = ddi_get_lbolt64();
2477 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2480 align = 1ULL << vd->vdev_top->vdev_ashift;
2482 if (P2PHASE(zio->io_size, align) != 0) {
2483 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2484 char *abuf = zio_buf_alloc(asize);
2485 ASSERT(vd == vd->vdev_top);
2486 if (zio->io_type == ZIO_TYPE_WRITE) {
2487 bcopy(zio->io_data, abuf, zio->io_size);
2488 bzero(abuf + zio->io_size, asize - zio->io_size);
2490 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2493 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2494 ASSERT(P2PHASE(zio->io_size, align) == 0);
2495 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2498 * If this is a repair I/O, and there's no self-healing involved --
2499 * that is, we're just resilvering what we expect to resilver --
2500 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2501 * This prevents spurious resilvering with nested replication.
2502 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2503 * A is out of date, we'll read from C+D, then use the data to
2504 * resilver A+B -- but we don't actually want to resilver B, just A.
2505 * The top-level mirror has no way to know this, so instead we just
2506 * discard unnecessary repairs as we work our way down the vdev tree.
2507 * The same logic applies to any form of nested replication:
2508 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2510 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2511 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2512 zio->io_txg != 0 && /* not a delegated i/o */
2513 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2514 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2515 zio_vdev_io_bypass(zio);
2516 return (ZIO_PIPELINE_CONTINUE);
2519 if (vd->vdev_ops->vdev_op_leaf &&
2520 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2522 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2523 return (ZIO_PIPELINE_CONTINUE);
2525 if ((zio = vdev_queue_io(zio)) == NULL)
2526 return (ZIO_PIPELINE_STOP);
2528 if (!vdev_accessible(vd, zio)) {
2529 zio->io_error = ENXIO;
2531 return (ZIO_PIPELINE_STOP);
2535 return (vd->vdev_ops->vdev_op_io_start(zio));
2539 zio_vdev_io_done(zio_t *zio)
2541 vdev_t *vd = zio->io_vd;
2542 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2543 boolean_t unexpected_error = B_FALSE;
2545 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2546 return (ZIO_PIPELINE_STOP);
2548 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2550 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2552 vdev_queue_io_done(zio);
2554 if (zio->io_type == ZIO_TYPE_WRITE)
2555 vdev_cache_write(zio);
2557 if (zio_injection_enabled && zio->io_error == 0)
2558 zio->io_error = zio_handle_device_injection(vd,
2561 if (zio_injection_enabled && zio->io_error == 0)
2562 zio->io_error = zio_handle_label_injection(zio, EIO);
2564 if (zio->io_error) {
2565 if (!vdev_accessible(vd, zio)) {
2566 zio->io_error = ENXIO;
2568 unexpected_error = B_TRUE;
2573 ops->vdev_op_io_done(zio);
2575 if (unexpected_error)
2576 VERIFY(vdev_probe(vd, zio) == NULL);
2578 return (ZIO_PIPELINE_CONTINUE);
2582 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2583 * disk, and use that to finish the checksum ereport later.
2586 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2587 const void *good_buf)
2589 /* no processing needed */
2590 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2595 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2597 void *buf = zio_buf_alloc(zio->io_size);
2599 bcopy(zio->io_data, buf, zio->io_size);
2601 zcr->zcr_cbinfo = zio->io_size;
2602 zcr->zcr_cbdata = buf;
2603 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2604 zcr->zcr_free = zio_buf_free;
2608 zio_vdev_io_assess(zio_t *zio)
2610 vdev_t *vd = zio->io_vd;
2612 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2613 return (ZIO_PIPELINE_STOP);
2615 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2616 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2618 if (zio->io_vsd != NULL) {
2619 zio->io_vsd_ops->vsd_free(zio);
2623 if (zio_injection_enabled && zio->io_error == 0)
2624 zio->io_error = zio_handle_fault_injection(zio, EIO);
2627 * If the I/O failed, determine whether we should attempt to retry it.
2629 * On retry, we cut in line in the issue queue, since we don't want
2630 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2632 if (zio->io_error && vd == NULL &&
2633 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2634 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2635 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2637 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2638 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2639 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2640 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2641 zio_requeue_io_start_cut_in_line);
2642 return (ZIO_PIPELINE_STOP);
2646 * If we got an error on a leaf device, convert it to ENXIO
2647 * if the device is not accessible at all.
2649 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2650 !vdev_accessible(vd, zio))
2651 zio->io_error = ENXIO;
2654 * If we can't write to an interior vdev (mirror or RAID-Z),
2655 * set vdev_cant_write so that we stop trying to allocate from it.
2657 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2658 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2659 vd->vdev_cant_write = B_TRUE;
2662 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2664 return (ZIO_PIPELINE_CONTINUE);
2668 zio_vdev_io_reissue(zio_t *zio)
2670 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2671 ASSERT(zio->io_error == 0);
2673 zio->io_stage >>= 1;
2677 zio_vdev_io_redone(zio_t *zio)
2679 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2681 zio->io_stage >>= 1;
2685 zio_vdev_io_bypass(zio_t *zio)
2687 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2688 ASSERT(zio->io_error == 0);
2690 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2691 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2695 * ==========================================================================
2696 * Generate and verify checksums
2697 * ==========================================================================
2700 zio_checksum_generate(zio_t *zio)
2702 blkptr_t *bp = zio->io_bp;
2703 enum zio_checksum checksum;
2707 * This is zio_write_phys().
2708 * We're either generating a label checksum, or none at all.
2710 checksum = zio->io_prop.zp_checksum;
2712 if (checksum == ZIO_CHECKSUM_OFF)
2713 return (ZIO_PIPELINE_CONTINUE);
2715 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2717 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2718 ASSERT(!IO_IS_ALLOCATING(zio));
2719 checksum = ZIO_CHECKSUM_GANG_HEADER;
2721 checksum = BP_GET_CHECKSUM(bp);
2725 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2727 return (ZIO_PIPELINE_CONTINUE);
2731 zio_checksum_verify(zio_t *zio)
2733 zio_bad_cksum_t info;
2734 blkptr_t *bp = zio->io_bp;
2737 ASSERT(zio->io_vd != NULL);
2741 * This is zio_read_phys().
2742 * We're either verifying a label checksum, or nothing at all.
2744 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2745 return (ZIO_PIPELINE_CONTINUE);
2747 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2750 if ((error = zio_checksum_error(zio, &info)) != 0) {
2751 zio->io_error = error;
2752 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2753 zfs_ereport_start_checksum(zio->io_spa,
2754 zio->io_vd, zio, zio->io_offset,
2755 zio->io_size, NULL, &info);
2759 return (ZIO_PIPELINE_CONTINUE);
2763 * Called by RAID-Z to ensure we don't compute the checksum twice.
2766 zio_checksum_verified(zio_t *zio)
2768 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2772 * ==========================================================================
2773 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2774 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2775 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2776 * indicate errors that are specific to one I/O, and most likely permanent.
2777 * Any other error is presumed to be worse because we weren't expecting it.
2778 * ==========================================================================
2781 zio_worst_error(int e1, int e2)
2783 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2786 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2787 if (e1 == zio_error_rank[r1])
2790 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2791 if (e2 == zio_error_rank[r2])
2794 return (r1 > r2 ? e1 : e2);
2798 * ==========================================================================
2800 * ==========================================================================
2803 zio_ready(zio_t *zio)
2805 blkptr_t *bp = zio->io_bp;
2806 zio_t *pio, *pio_next;
2808 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2809 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2810 return (ZIO_PIPELINE_STOP);
2812 if (zio->io_ready) {
2813 ASSERT(IO_IS_ALLOCATING(zio));
2814 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2815 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2820 if (bp != NULL && bp != &zio->io_bp_copy)
2821 zio->io_bp_copy = *bp;
2824 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2826 mutex_enter(&zio->io_lock);
2827 zio->io_state[ZIO_WAIT_READY] = 1;
2828 pio = zio_walk_parents(zio);
2829 mutex_exit(&zio->io_lock);
2832 * As we notify zio's parents, new parents could be added.
2833 * New parents go to the head of zio's io_parent_list, however,
2834 * so we will (correctly) not notify them. The remainder of zio's
2835 * io_parent_list, from 'pio_next' onward, cannot change because
2836 * all parents must wait for us to be done before they can be done.
2838 for (; pio != NULL; pio = pio_next) {
2839 pio_next = zio_walk_parents(zio);
2840 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2843 if (zio->io_flags & ZIO_FLAG_NODATA) {
2844 if (BP_IS_GANG(bp)) {
2845 zio->io_flags &= ~ZIO_FLAG_NODATA;
2847 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2848 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2852 if (zio_injection_enabled &&
2853 zio->io_spa->spa_syncing_txg == zio->io_txg)
2854 zio_handle_ignored_writes(zio);
2856 return (ZIO_PIPELINE_CONTINUE);
2860 zio_done(zio_t *zio)
2862 zio_t *pio, *pio_next;
2866 * If our children haven't all completed,
2867 * wait for them and then repeat this pipeline stage.
2869 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2870 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2871 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2872 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2873 return (ZIO_PIPELINE_STOP);
2875 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2876 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2877 ASSERT(zio->io_children[c][w] == 0);
2879 if (zio->io_bp != NULL) {
2880 ASSERT(zio->io_bp->blk_pad[0] == 0);
2881 ASSERT(zio->io_bp->blk_pad[1] == 0);
2882 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2883 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2884 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2885 zio->io_bp_override == NULL &&
2886 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2887 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2888 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2889 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2890 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2895 * If there were child vdev/gang/ddt errors, they apply to us now.
2897 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2898 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2899 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2902 * If the I/O on the transformed data was successful, generate any
2903 * checksum reports now while we still have the transformed data.
2905 if (zio->io_error == 0) {
2906 while (zio->io_cksum_report != NULL) {
2907 zio_cksum_report_t *zcr = zio->io_cksum_report;
2908 uint64_t align = zcr->zcr_align;
2909 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2910 char *abuf = zio->io_data;
2912 if (asize != zio->io_size) {
2913 abuf = zio_buf_alloc(asize);
2914 bcopy(zio->io_data, abuf, zio->io_size);
2915 bzero(abuf + zio->io_size, asize - zio->io_size);
2918 zio->io_cksum_report = zcr->zcr_next;
2919 zcr->zcr_next = NULL;
2920 zcr->zcr_finish(zcr, abuf);
2921 zfs_ereport_free_checksum(zcr);
2923 if (asize != zio->io_size)
2924 zio_buf_free(abuf, asize);
2928 zio_pop_transforms(zio); /* note: may set zio->io_error */
2930 vdev_stat_update(zio, zio->io_size);
2933 * If this I/O is attached to a particular vdev is slow, exceeding
2934 * 30 seconds to complete, post an error described the I/O delay.
2935 * We ignore these errors if the device is currently unavailable.
2937 if (zio->io_delay >= MSEC_TO_TICK(zio_delay_max)) {
2938 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2939 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2940 zio->io_vd, zio, 0, 0);
2943 if (zio->io_error) {
2945 * If this I/O is attached to a particular vdev,
2946 * generate an error message describing the I/O failure
2947 * at the block level. We ignore these errors if the
2948 * device is currently unavailable.
2950 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2951 !vdev_is_dead(zio->io_vd))
2952 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2953 zio->io_vd, zio, 0, 0);
2955 if ((zio->io_error == EIO || !(zio->io_flags &
2956 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2957 zio == zio->io_logical) {
2959 * For logical I/O requests, tell the SPA to log the
2960 * error and generate a logical data ereport.
2962 spa_log_error(zio->io_spa, zio);
2963 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2968 if (zio->io_error && zio == zio->io_logical) {
2970 * Determine whether zio should be reexecuted. This will
2971 * propagate all the way to the root via zio_notify_parent().
2973 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2974 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2976 if (IO_IS_ALLOCATING(zio) &&
2977 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2978 if (zio->io_error != ENOSPC)
2979 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2981 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2984 if ((zio->io_type == ZIO_TYPE_READ ||
2985 zio->io_type == ZIO_TYPE_FREE) &&
2986 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2987 zio->io_error == ENXIO &&
2988 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2989 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2990 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2992 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2993 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2996 * Here is a possibly good place to attempt to do
2997 * either combinatorial reconstruction or error correction
2998 * based on checksums. It also might be a good place
2999 * to send out preliminary ereports before we suspend
3005 * If there were logical child errors, they apply to us now.
3006 * We defer this until now to avoid conflating logical child
3007 * errors with errors that happened to the zio itself when
3008 * updating vdev stats and reporting FMA events above.
3010 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3012 if ((zio->io_error || zio->io_reexecute) &&
3013 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3014 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
3015 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
3017 zio_gang_tree_free(&zio->io_gang_tree);
3020 * Godfather I/Os should never suspend.
3022 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3023 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3024 zio->io_reexecute = 0;
3026 if (zio->io_reexecute) {
3028 * This is a logical I/O that wants to reexecute.
3030 * Reexecute is top-down. When an i/o fails, if it's not
3031 * the root, it simply notifies its parent and sticks around.
3032 * The parent, seeing that it still has children in zio_done(),
3033 * does the same. This percolates all the way up to the root.
3034 * The root i/o will reexecute or suspend the entire tree.
3036 * This approach ensures that zio_reexecute() honors
3037 * all the original i/o dependency relationships, e.g.
3038 * parents not executing until children are ready.
3040 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3042 zio->io_gang_leader = NULL;
3044 mutex_enter(&zio->io_lock);
3045 zio->io_state[ZIO_WAIT_DONE] = 1;
3046 mutex_exit(&zio->io_lock);
3049 * "The Godfather" I/O monitors its children but is
3050 * not a true parent to them. It will track them through
3051 * the pipeline but severs its ties whenever they get into
3052 * trouble (e.g. suspended). This allows "The Godfather"
3053 * I/O to return status without blocking.
3055 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3056 zio_link_t *zl = zio->io_walk_link;
3057 pio_next = zio_walk_parents(zio);
3059 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3060 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3061 zio_remove_child(pio, zio, zl);
3062 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3066 if ((pio = zio_unique_parent(zio)) != NULL) {
3068 * We're not a root i/o, so there's nothing to do
3069 * but notify our parent. Don't propagate errors
3070 * upward since we haven't permanently failed yet.
3072 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3073 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3074 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3075 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3077 * We'd fail again if we reexecuted now, so suspend
3078 * until conditions improve (e.g. device comes online).
3080 zio_suspend(zio->io_spa, zio);
3083 * Reexecution is potentially a huge amount of work.
3084 * Hand it off to the otherwise-unused claim taskq.
3086 ASSERT(taskq_empty_ent(&zio->io_tqent));
3087 spa_taskq_dispatch_ent(zio->io_spa,
3088 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
3089 (task_func_t *)zio_reexecute, zio, 0,
3092 return (ZIO_PIPELINE_STOP);
3095 ASSERT(zio->io_child_count == 0);
3096 ASSERT(zio->io_reexecute == 0);
3097 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3100 * Report any checksum errors, since the I/O is complete.
3102 while (zio->io_cksum_report != NULL) {
3103 zio_cksum_report_t *zcr = zio->io_cksum_report;
3104 zio->io_cksum_report = zcr->zcr_next;
3105 zcr->zcr_next = NULL;
3106 zcr->zcr_finish(zcr, NULL);
3107 zfs_ereport_free_checksum(zcr);
3110 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3111 !BP_IS_HOLE(zio->io_bp)) {
3112 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3116 * It is the responsibility of the done callback to ensure that this
3117 * particular zio is no longer discoverable for adoption, and as
3118 * such, cannot acquire any new parents.
3123 mutex_enter(&zio->io_lock);
3124 zio->io_state[ZIO_WAIT_DONE] = 1;
3125 mutex_exit(&zio->io_lock);
3127 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3128 zio_link_t *zl = zio->io_walk_link;
3129 pio_next = zio_walk_parents(zio);
3130 zio_remove_child(pio, zio, zl);
3131 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3134 if (zio->io_waiter != NULL) {
3135 mutex_enter(&zio->io_lock);
3136 zio->io_executor = NULL;
3137 cv_broadcast(&zio->io_cv);
3138 mutex_exit(&zio->io_lock);
3143 return (ZIO_PIPELINE_STOP);
3147 * ==========================================================================
3148 * I/O pipeline definition
3149 * ==========================================================================
3151 static zio_pipe_stage_t *zio_pipeline[] = {
3157 zio_checksum_generate,
3171 zio_checksum_verify,
3175 /* dnp is the dnode for zb1->zb_object */
3177 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3178 const zbookmark_t *zb2)
3180 uint64_t zb1nextL0, zb2thisobj;
3182 ASSERT(zb1->zb_objset == zb2->zb_objset);
3183 ASSERT(zb2->zb_level == 0);
3186 * A bookmark in the deadlist is considered to be after
3189 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3192 /* The objset_phys_t isn't before anything. */
3196 zb1nextL0 = (zb1->zb_blkid + 1) <<
3197 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3199 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3200 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3202 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3203 uint64_t nextobj = zb1nextL0 *
3204 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3205 return (nextobj <= zb2thisobj);
3208 if (zb1->zb_object < zb2thisobj)
3210 if (zb1->zb_object > zb2thisobj)
3212 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3214 return (zb1nextL0 <= zb2->zb_blkid);
3217 #if defined(_KERNEL) && defined(HAVE_SPL)
3218 /* Fault injection */
3219 EXPORT_SYMBOL(zio_injection_enabled);
3220 EXPORT_SYMBOL(zio_inject_fault);
3221 EXPORT_SYMBOL(zio_inject_list_next);
3222 EXPORT_SYMBOL(zio_clear_fault);
3223 EXPORT_SYMBOL(zio_handle_fault_injection);
3224 EXPORT_SYMBOL(zio_handle_device_injection);
3225 EXPORT_SYMBOL(zio_handle_label_injection);
3226 EXPORT_SYMBOL(zio_priority_table);
3227 EXPORT_SYMBOL(zio_type_name);
3229 module_param(zio_bulk_flags, int, 0644);
3230 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3232 module_param(zio_delay_max, int, 0644);
3233 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3235 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3236 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");