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.
25 #include <sys/zfs_context.h>
26 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/dmu_objset.h>
39 * ==========================================================================
41 * ==========================================================================
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 0, /* ZIO_PRIORITY_NOW */
45 0, /* ZIO_PRIORITY_SYNC_READ */
46 0, /* ZIO_PRIORITY_SYNC_WRITE */
47 0, /* ZIO_PRIORITY_LOG_WRITE */
48 1, /* ZIO_PRIORITY_CACHE_FILL */
49 1, /* ZIO_PRIORITY_AGG */
50 4, /* ZIO_PRIORITY_FREE */
51 4, /* ZIO_PRIORITY_ASYNC_WRITE */
52 6, /* ZIO_PRIORITY_ASYNC_READ */
53 10, /* ZIO_PRIORITY_RESILVER */
54 20, /* ZIO_PRIORITY_SCRUB */
55 2, /* ZIO_PRIORITY_DDT_PREFETCH */
59 * ==========================================================================
60 * I/O type descriptions
61 * ==========================================================================
63 char *zio_type_name[ZIO_TYPES] = {
64 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
69 * ==========================================================================
71 * ==========================================================================
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 extern vmem_t *zio_alloc_arena;
83 * An allocating zio is one that either currently has the DVA allocate
84 * stage set or will have it later in its lifetime.
86 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
88 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
91 int zio_buf_debug_limit = 16384;
93 int zio_buf_debug_limit = 0;
100 vmem_t *data_alloc_arena = NULL;
103 data_alloc_arena = zio_alloc_arena;
105 zio_cache = kmem_cache_create("zio_cache",
106 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
107 zio_link_cache = kmem_cache_create("zio_link_cache",
108 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 * For small buffers, we want a cache for each multiple of
112 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
113 * for each quarter-power of 2. For large buffers, we want
114 * a cache for each multiple of PAGESIZE.
116 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
117 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
121 while (p2 & (p2 - 1))
124 if (size <= 4 * SPA_MINBLOCKSIZE) {
125 align = SPA_MINBLOCKSIZE;
126 } else if (P2PHASE(size, PAGESIZE) == 0) {
128 } else if (P2PHASE(size, p2 >> 2) == 0) {
134 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
135 zio_buf_cache[c] = kmem_cache_create(name, size,
136 align, NULL, NULL, NULL, NULL, NULL,
137 size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
139 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
140 zio_data_buf_cache[c] = kmem_cache_create(name, size,
141 align, NULL, NULL, NULL, NULL, data_alloc_arena,
142 size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
147 ASSERT(zio_buf_cache[c] != NULL);
148 if (zio_buf_cache[c - 1] == NULL)
149 zio_buf_cache[c - 1] = zio_buf_cache[c];
151 ASSERT(zio_data_buf_cache[c] != NULL);
152 if (zio_data_buf_cache[c - 1] == NULL)
153 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
163 kmem_cache_t *last_cache = NULL;
164 kmem_cache_t *last_data_cache = NULL;
166 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
167 if (zio_buf_cache[c] != last_cache) {
168 last_cache = zio_buf_cache[c];
169 kmem_cache_destroy(zio_buf_cache[c]);
171 zio_buf_cache[c] = NULL;
173 if (zio_data_buf_cache[c] != last_data_cache) {
174 last_data_cache = zio_data_buf_cache[c];
175 kmem_cache_destroy(zio_data_buf_cache[c]);
177 zio_data_buf_cache[c] = NULL;
180 kmem_cache_destroy(zio_link_cache);
181 kmem_cache_destroy(zio_cache);
187 * ==========================================================================
188 * Allocate and free I/O buffers
189 * ==========================================================================
193 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
194 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
195 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
196 * excess / transient data in-core during a crashdump.
199 zio_buf_alloc(size_t size)
201 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
203 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
205 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
209 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
210 * crashdump if the kernel panics. This exists so that we will limit the amount
211 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
212 * of kernel heap dumped to disk when the kernel panics)
215 zio_data_buf_alloc(size_t size)
217 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
219 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
221 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
225 zio_buf_free(void *buf, size_t size)
227 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
229 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
231 kmem_cache_free(zio_buf_cache[c], buf);
235 zio_data_buf_free(void *buf, size_t size)
237 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
239 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
241 kmem_cache_free(zio_data_buf_cache[c], buf);
245 * ==========================================================================
246 * Push and pop I/O transform buffers
247 * ==========================================================================
250 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
251 zio_transform_func_t *transform)
253 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
255 zt->zt_orig_data = zio->io_data;
256 zt->zt_orig_size = zio->io_size;
257 zt->zt_bufsize = bufsize;
258 zt->zt_transform = transform;
260 zt->zt_next = zio->io_transform_stack;
261 zio->io_transform_stack = zt;
268 zio_pop_transforms(zio_t *zio)
272 while ((zt = zio->io_transform_stack) != NULL) {
273 if (zt->zt_transform != NULL)
274 zt->zt_transform(zio,
275 zt->zt_orig_data, zt->zt_orig_size);
277 if (zt->zt_bufsize != 0)
278 zio_buf_free(zio->io_data, zt->zt_bufsize);
280 zio->io_data = zt->zt_orig_data;
281 zio->io_size = zt->zt_orig_size;
282 zio->io_transform_stack = zt->zt_next;
284 kmem_free(zt, sizeof (zio_transform_t));
289 * ==========================================================================
290 * I/O transform callbacks for subblocks and decompression
291 * ==========================================================================
294 zio_subblock(zio_t *zio, void *data, uint64_t size)
296 ASSERT(zio->io_size > size);
298 if (zio->io_type == ZIO_TYPE_READ)
299 bcopy(zio->io_data, data, size);
303 zio_decompress(zio_t *zio, void *data, uint64_t size)
305 if (zio->io_error == 0 &&
306 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
307 zio->io_data, data, zio->io_size, size) != 0)
312 * ==========================================================================
313 * I/O parent/child relationships and pipeline interlocks
314 * ==========================================================================
317 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
318 * continue calling these functions until they return NULL.
319 * Otherwise, the next caller will pick up the list walk in
320 * some indeterminate state. (Otherwise every caller would
321 * have to pass in a cookie to keep the state represented by
322 * io_walk_link, which gets annoying.)
325 zio_walk_parents(zio_t *cio)
327 zio_link_t *zl = cio->io_walk_link;
328 list_t *pl = &cio->io_parent_list;
330 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
331 cio->io_walk_link = zl;
336 ASSERT(zl->zl_child == cio);
337 return (zl->zl_parent);
341 zio_walk_children(zio_t *pio)
343 zio_link_t *zl = pio->io_walk_link;
344 list_t *cl = &pio->io_child_list;
346 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
347 pio->io_walk_link = zl;
352 ASSERT(zl->zl_parent == pio);
353 return (zl->zl_child);
357 zio_unique_parent(zio_t *cio)
359 zio_t *pio = zio_walk_parents(cio);
361 VERIFY(zio_walk_parents(cio) == NULL);
366 zio_add_child(zio_t *pio, zio_t *cio)
368 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
372 * Logical I/Os can have logical, gang, or vdev children.
373 * Gang I/Os can have gang or vdev children.
374 * Vdev I/Os can only have vdev children.
375 * The following ASSERT captures all of these constraints.
377 ASSERT(cio->io_child_type <= pio->io_child_type);
382 mutex_enter(&cio->io_lock);
383 mutex_enter(&pio->io_lock);
385 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
387 for (w = 0; w < ZIO_WAIT_TYPES; w++)
388 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
390 list_insert_head(&pio->io_child_list, zl);
391 list_insert_head(&cio->io_parent_list, zl);
393 pio->io_child_count++;
394 cio->io_parent_count++;
396 mutex_exit(&pio->io_lock);
397 mutex_exit(&cio->io_lock);
401 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
403 ASSERT(zl->zl_parent == pio);
404 ASSERT(zl->zl_child == cio);
406 mutex_enter(&cio->io_lock);
407 mutex_enter(&pio->io_lock);
409 list_remove(&pio->io_child_list, zl);
410 list_remove(&cio->io_parent_list, zl);
412 pio->io_child_count--;
413 cio->io_parent_count--;
415 mutex_exit(&pio->io_lock);
416 mutex_exit(&cio->io_lock);
418 kmem_cache_free(zio_link_cache, zl);
422 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
424 uint64_t *countp = &zio->io_children[child][wait];
425 boolean_t waiting = B_FALSE;
427 mutex_enter(&zio->io_lock);
428 ASSERT(zio->io_stall == NULL);
431 zio->io_stall = countp;
434 mutex_exit(&zio->io_lock);
439 __attribute__((always_inline))
441 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
443 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
444 int *errorp = &pio->io_child_error[zio->io_child_type];
446 mutex_enter(&pio->io_lock);
447 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
448 *errorp = zio_worst_error(*errorp, zio->io_error);
449 pio->io_reexecute |= zio->io_reexecute;
450 ASSERT3U(*countp, >, 0);
451 if (--*countp == 0 && pio->io_stall == countp) {
452 pio->io_stall = NULL;
453 mutex_exit(&pio->io_lock);
456 mutex_exit(&pio->io_lock);
461 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
463 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
464 zio->io_error = zio->io_child_error[c];
468 * ==========================================================================
469 * Create the various types of I/O (read, write, free, etc)
470 * ==========================================================================
473 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
474 void *data, uint64_t size, zio_done_func_t *done, void *private,
475 zio_type_t type, int priority, enum zio_flag flags,
476 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
477 enum zio_stage stage, enum zio_stage pipeline)
481 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
482 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
483 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
485 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
486 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
487 ASSERT(vd || stage == ZIO_STAGE_OPEN);
489 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
490 bzero(zio, sizeof (zio_t));
492 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
493 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
495 list_create(&zio->io_parent_list, sizeof (zio_link_t),
496 offsetof(zio_link_t, zl_parent_node));
497 list_create(&zio->io_child_list, sizeof (zio_link_t),
498 offsetof(zio_link_t, zl_child_node));
501 zio->io_child_type = ZIO_CHILD_VDEV;
502 else if (flags & ZIO_FLAG_GANG_CHILD)
503 zio->io_child_type = ZIO_CHILD_GANG;
504 else if (flags & ZIO_FLAG_DDT_CHILD)
505 zio->io_child_type = ZIO_CHILD_DDT;
507 zio->io_child_type = ZIO_CHILD_LOGICAL;
510 zio->io_bp = (blkptr_t *)bp;
511 zio->io_bp_copy = *bp;
512 zio->io_bp_orig = *bp;
513 if (type != ZIO_TYPE_WRITE ||
514 zio->io_child_type == ZIO_CHILD_DDT)
515 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
516 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
517 zio->io_logical = zio;
518 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
519 pipeline |= ZIO_GANG_STAGES;
525 zio->io_private = private;
527 zio->io_priority = priority;
529 zio->io_offset = offset;
530 zio->io_orig_data = zio->io_data = data;
531 zio->io_orig_size = zio->io_size = size;
532 zio->io_orig_flags = zio->io_flags = flags;
533 zio->io_orig_stage = zio->io_stage = stage;
534 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
536 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
537 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
540 zio->io_bookmark = *zb;
543 if (zio->io_logical == NULL)
544 zio->io_logical = pio->io_logical;
545 if (zio->io_child_type == ZIO_CHILD_GANG)
546 zio->io_gang_leader = pio->io_gang_leader;
547 zio_add_child(pio, zio);
554 zio_destroy(zio_t *zio)
556 list_destroy(&zio->io_parent_list);
557 list_destroy(&zio->io_child_list);
558 mutex_destroy(&zio->io_lock);
559 cv_destroy(&zio->io_cv);
560 kmem_cache_free(zio_cache, zio);
564 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
565 void *private, enum zio_flag flags)
569 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
570 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
571 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
577 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
579 return (zio_null(NULL, spa, NULL, done, private, flags));
583 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
584 void *data, uint64_t size, zio_done_func_t *done, void *private,
585 int priority, enum zio_flag flags, const zbookmark_t *zb)
589 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
590 data, size, done, private,
591 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
592 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
593 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
599 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
600 void *data, uint64_t size, const zio_prop_t *zp,
601 zio_done_func_t *ready, zio_done_func_t *done, void *private,
602 int priority, enum zio_flag flags, const zbookmark_t *zb)
606 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
607 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
608 zp->zp_compress >= ZIO_COMPRESS_OFF &&
609 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
610 zp->zp_type < DMU_OT_NUMTYPES &&
613 zp->zp_copies <= spa_max_replication(spa) &&
615 zp->zp_dedup_verify <= 1);
617 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
618 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
619 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
620 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
622 zio->io_ready = ready;
629 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
630 uint64_t size, zio_done_func_t *done, void *private, int priority,
631 enum zio_flag flags, zbookmark_t *zb)
635 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
636 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
637 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
643 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
645 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
646 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
647 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
648 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
650 zio->io_prop.zp_copies = copies;
651 zio->io_bp_override = bp;
655 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
657 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
661 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
666 dprintf_bp(bp, "freeing in txg %llu, pass %u",
667 (longlong_t)txg, spa->spa_sync_pass);
669 ASSERT(!BP_IS_HOLE(bp));
670 ASSERT(spa_syncing_txg(spa) == txg);
671 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
673 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
674 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
675 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
681 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
682 zio_done_func_t *done, void *private, enum zio_flag flags)
687 * A claim is an allocation of a specific block. Claims are needed
688 * to support immediate writes in the intent log. The issue is that
689 * immediate writes contain committed data, but in a txg that was
690 * *not* committed. Upon opening the pool after an unclean shutdown,
691 * the intent log claims all blocks that contain immediate write data
692 * so that the SPA knows they're in use.
694 * All claims *must* be resolved in the first txg -- before the SPA
695 * starts allocating blocks -- so that nothing is allocated twice.
696 * If txg == 0 we just verify that the block is claimable.
698 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
699 ASSERT(txg == spa_first_txg(spa) || txg == 0);
700 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
702 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
703 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
704 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
710 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
711 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
716 if (vd->vdev_children == 0) {
717 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
718 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
719 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
723 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
725 for (c = 0; c < vd->vdev_children; c++)
726 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
727 done, private, priority, flags));
734 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
735 void *data, int checksum, zio_done_func_t *done, void *private,
736 int priority, enum zio_flag flags, boolean_t labels)
740 ASSERT(vd->vdev_children == 0);
741 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
742 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
743 ASSERT3U(offset + size, <=, vd->vdev_psize);
745 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
746 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
747 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
749 zio->io_prop.zp_checksum = checksum;
755 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
756 void *data, int checksum, zio_done_func_t *done, void *private,
757 int priority, enum zio_flag flags, boolean_t labels)
761 ASSERT(vd->vdev_children == 0);
762 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
763 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
764 ASSERT3U(offset + size, <=, vd->vdev_psize);
766 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
767 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
768 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
770 zio->io_prop.zp_checksum = checksum;
772 if (zio_checksum_table[checksum].ci_eck) {
774 * zec checksums are necessarily destructive -- they modify
775 * the end of the write buffer to hold the verifier/checksum.
776 * Therefore, we must make a local copy in case the data is
777 * being written to multiple places in parallel.
779 void *wbuf = zio_buf_alloc(size);
780 bcopy(data, wbuf, size);
781 zio_push_transform(zio, wbuf, size, size, NULL);
788 * Create a child I/O to do some work for us.
791 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
792 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
793 zio_done_func_t *done, void *private)
795 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
798 ASSERT(vd->vdev_parent ==
799 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
801 if (type == ZIO_TYPE_READ && bp != NULL) {
803 * If we have the bp, then the child should perform the
804 * checksum and the parent need not. This pushes error
805 * detection as close to the leaves as possible and
806 * eliminates redundant checksums in the interior nodes.
808 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
809 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
812 if (vd->vdev_children == 0)
813 offset += VDEV_LABEL_START_SIZE;
815 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
818 * If we've decided to do a repair, the write is not speculative --
819 * even if the original read was.
821 if (flags & ZIO_FLAG_IO_REPAIR)
822 flags &= ~ZIO_FLAG_SPECULATIVE;
824 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
825 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
826 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
832 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
833 int type, int priority, enum zio_flag flags,
834 zio_done_func_t *done, void *private)
838 ASSERT(vd->vdev_ops->vdev_op_leaf);
840 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
841 data, size, done, private, type, priority,
842 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
844 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
850 zio_flush(zio_t *zio, vdev_t *vd)
852 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
853 NULL, NULL, ZIO_PRIORITY_NOW,
854 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
858 zio_shrink(zio_t *zio, uint64_t size)
860 ASSERT(zio->io_executor == NULL);
861 ASSERT(zio->io_orig_size == zio->io_size);
862 ASSERT(size <= zio->io_size);
865 * We don't shrink for raidz because of problems with the
866 * reconstruction when reading back less than the block size.
867 * Note, BP_IS_RAIDZ() assumes no compression.
869 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
870 if (!BP_IS_RAIDZ(zio->io_bp))
871 zio->io_orig_size = zio->io_size = size;
875 * ==========================================================================
876 * Prepare to read and write logical blocks
877 * ==========================================================================
881 zio_read_bp_init(zio_t *zio)
883 blkptr_t *bp = zio->io_bp;
885 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
886 zio->io_child_type == ZIO_CHILD_LOGICAL &&
887 !(zio->io_flags & ZIO_FLAG_RAW)) {
888 uint64_t psize = BP_GET_PSIZE(bp);
889 void *cbuf = zio_buf_alloc(psize);
891 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
894 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
895 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
897 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
898 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
900 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
901 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
903 return (ZIO_PIPELINE_CONTINUE);
907 zio_write_bp_init(zio_t *zio)
909 spa_t *spa = zio->io_spa;
910 zio_prop_t *zp = &zio->io_prop;
911 enum zio_compress compress = zp->zp_compress;
912 blkptr_t *bp = zio->io_bp;
913 uint64_t lsize = zio->io_size;
914 uint64_t psize = lsize;
918 * If our children haven't all reached the ready stage,
919 * wait for them and then repeat this pipeline stage.
921 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
922 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
923 return (ZIO_PIPELINE_STOP);
925 if (!IO_IS_ALLOCATING(zio))
926 return (ZIO_PIPELINE_CONTINUE);
928 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
930 if (zio->io_bp_override) {
931 ASSERT(bp->blk_birth != zio->io_txg);
932 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
934 *bp = *zio->io_bp_override;
935 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
937 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
938 return (ZIO_PIPELINE_CONTINUE);
940 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
941 zp->zp_dedup_verify);
943 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
945 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
946 return (ZIO_PIPELINE_CONTINUE);
948 zio->io_bp_override = NULL;
952 if (bp->blk_birth == zio->io_txg) {
954 * We're rewriting an existing block, which means we're
955 * working on behalf of spa_sync(). For spa_sync() to
956 * converge, it must eventually be the case that we don't
957 * have to allocate new blocks. But compression changes
958 * the blocksize, which forces a reallocate, and makes
959 * convergence take longer. Therefore, after the first
960 * few passes, stop compressing to ensure convergence.
962 pass = spa_sync_pass(spa);
964 ASSERT(zio->io_txg == spa_syncing_txg(spa));
965 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
966 ASSERT(!BP_GET_DEDUP(bp));
968 if (pass > SYNC_PASS_DONT_COMPRESS)
969 compress = ZIO_COMPRESS_OFF;
971 /* Make sure someone doesn't change their mind on overwrites */
972 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
973 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
976 if (compress != ZIO_COMPRESS_OFF) {
977 void *cbuf = zio_buf_alloc(lsize);
978 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
979 if (psize == 0 || psize == lsize) {
980 compress = ZIO_COMPRESS_OFF;
981 zio_buf_free(cbuf, lsize);
983 ASSERT(psize < lsize);
984 zio_push_transform(zio, cbuf, psize, lsize, NULL);
989 * The final pass of spa_sync() must be all rewrites, but the first
990 * few passes offer a trade-off: allocating blocks defers convergence,
991 * but newly allocated blocks are sequential, so they can be written
992 * to disk faster. Therefore, we allow the first few passes of
993 * spa_sync() to allocate new blocks, but force rewrites after that.
994 * There should only be a handful of blocks after pass 1 in any case.
996 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
997 pass > SYNC_PASS_REWRITE) {
998 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1000 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1001 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1004 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1008 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1010 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1011 BP_SET_LSIZE(bp, lsize);
1012 BP_SET_PSIZE(bp, psize);
1013 BP_SET_COMPRESS(bp, compress);
1014 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1015 BP_SET_TYPE(bp, zp->zp_type);
1016 BP_SET_LEVEL(bp, zp->zp_level);
1017 BP_SET_DEDUP(bp, zp->zp_dedup);
1018 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1020 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1021 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1022 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1026 return (ZIO_PIPELINE_CONTINUE);
1030 zio_free_bp_init(zio_t *zio)
1032 blkptr_t *bp = zio->io_bp;
1034 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1035 if (BP_GET_DEDUP(bp))
1036 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1039 return (ZIO_PIPELINE_CONTINUE);
1043 * ==========================================================================
1044 * Execute the I/O pipeline
1045 * ==========================================================================
1049 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1051 spa_t *spa = zio->io_spa;
1052 zio_type_t t = zio->io_type;
1053 int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
1056 * If we're a config writer or a probe, the normal issue and
1057 * interrupt threads may all be blocked waiting for the config lock.
1058 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1060 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1064 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1066 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1070 * If this is a high priority I/O, then use the high priority taskq.
1072 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1073 spa->spa_zio_taskq[t][q + 1] != NULL)
1076 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1078 while (taskq_dispatch(spa->spa_zio_taskq[t][q],
1079 (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
1083 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1085 kthread_t *executor = zio->io_executor;
1086 spa_t *spa = zio->io_spa;
1089 for (t = 0; t < ZIO_TYPES; t++)
1090 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1097 zio_issue_async(zio_t *zio)
1099 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1101 return (ZIO_PIPELINE_STOP);
1105 zio_interrupt(zio_t *zio)
1107 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1111 * Execute the I/O pipeline until one of the following occurs:
1112 * (1) the I/O completes; (2) the pipeline stalls waiting for
1113 * dependent child I/Os; (3) the I/O issues, so we're waiting
1114 * for an I/O completion interrupt; (4) the I/O is delegated by
1115 * vdev-level caching or aggregation; (5) the I/O is deferred
1116 * due to vdev-level queueing; (6) the I/O is handed off to
1117 * another thread. In all cases, the pipeline stops whenever
1118 * there's no CPU work; it never burns a thread in cv_wait().
1120 * There's no locking on io_stage because there's no legitimate way
1121 * for multiple threads to be attempting to process the same I/O.
1123 static zio_pipe_stage_t *zio_pipeline[];
1126 zio_execute(zio_t *zio)
1128 zio->io_executor = curthread;
1130 while (zio->io_stage < ZIO_STAGE_DONE) {
1131 enum zio_stage pipeline = zio->io_pipeline;
1132 enum zio_stage stage = zio->io_stage;
1135 ASSERT(!MUTEX_HELD(&zio->io_lock));
1136 ASSERT(ISP2(stage));
1137 ASSERT(zio->io_stall == NULL);
1141 } while ((stage & pipeline) == 0);
1143 ASSERT(stage <= ZIO_STAGE_DONE);
1146 * If we are in interrupt context and this pipeline stage
1147 * will grab a config lock that is held across I/O,
1148 * or may wait for an I/O that needs an interrupt thread
1149 * to complete, issue async to avoid deadlock.
1151 * For VDEV_IO_START, we cut in line so that the io will
1152 * be sent to disk promptly.
1154 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1155 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1156 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1157 zio_requeue_io_start_cut_in_line : B_FALSE;
1158 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1162 zio->io_stage = stage;
1163 rv = zio_pipeline[highbit(stage) - 1](zio);
1165 if (rv == ZIO_PIPELINE_STOP)
1168 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1173 * ==========================================================================
1174 * Initiate I/O, either sync or async
1175 * ==========================================================================
1178 zio_wait(zio_t *zio)
1182 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1183 ASSERT(zio->io_executor == NULL);
1185 zio->io_waiter = curthread;
1189 mutex_enter(&zio->io_lock);
1190 while (zio->io_executor != NULL)
1191 cv_wait(&zio->io_cv, &zio->io_lock);
1192 mutex_exit(&zio->io_lock);
1194 error = zio->io_error;
1201 zio_nowait(zio_t *zio)
1203 ASSERT(zio->io_executor == NULL);
1205 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1206 zio_unique_parent(zio) == NULL) {
1208 * This is a logical async I/O with no parent to wait for it.
1209 * We add it to the spa_async_root_zio "Godfather" I/O which
1210 * will ensure they complete prior to unloading the pool.
1212 spa_t *spa = zio->io_spa;
1214 zio_add_child(spa->spa_async_zio_root, zio);
1221 * ==========================================================================
1222 * Reexecute or suspend/resume failed I/O
1223 * ==========================================================================
1227 zio_reexecute(zio_t *pio)
1229 zio_t *cio, *cio_next;
1232 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1233 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1234 ASSERT(pio->io_gang_leader == NULL);
1235 ASSERT(pio->io_gang_tree == NULL);
1237 pio->io_flags = pio->io_orig_flags;
1238 pio->io_stage = pio->io_orig_stage;
1239 pio->io_pipeline = pio->io_orig_pipeline;
1240 pio->io_reexecute = 0;
1242 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1243 pio->io_state[w] = 0;
1244 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1245 pio->io_child_error[c] = 0;
1247 if (IO_IS_ALLOCATING(pio))
1248 BP_ZERO(pio->io_bp);
1251 * As we reexecute pio's children, new children could be created.
1252 * New children go to the head of pio's io_child_list, however,
1253 * so we will (correctly) not reexecute them. The key is that
1254 * the remainder of pio's io_child_list, from 'cio_next' onward,
1255 * cannot be affected by any side effects of reexecuting 'cio'.
1257 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1258 cio_next = zio_walk_children(pio);
1259 mutex_enter(&pio->io_lock);
1260 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1261 pio->io_children[cio->io_child_type][w]++;
1262 mutex_exit(&pio->io_lock);
1267 * Now that all children have been reexecuted, execute the parent.
1268 * We don't reexecute "The Godfather" I/O here as it's the
1269 * responsibility of the caller to wait on him.
1271 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1276 zio_suspend(spa_t *spa, zio_t *zio)
1278 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1279 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1280 "failure and the failure mode property for this pool "
1281 "is set to panic.", spa_name(spa));
1283 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1285 mutex_enter(&spa->spa_suspend_lock);
1287 if (spa->spa_suspend_zio_root == NULL)
1288 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1289 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1290 ZIO_FLAG_GODFATHER);
1292 spa->spa_suspended = B_TRUE;
1295 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1296 ASSERT(zio != spa->spa_suspend_zio_root);
1297 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1298 ASSERT(zio_unique_parent(zio) == NULL);
1299 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1300 zio_add_child(spa->spa_suspend_zio_root, zio);
1303 mutex_exit(&spa->spa_suspend_lock);
1307 zio_resume(spa_t *spa)
1312 * Reexecute all previously suspended i/o.
1314 mutex_enter(&spa->spa_suspend_lock);
1315 spa->spa_suspended = B_FALSE;
1316 cv_broadcast(&spa->spa_suspend_cv);
1317 pio = spa->spa_suspend_zio_root;
1318 spa->spa_suspend_zio_root = NULL;
1319 mutex_exit(&spa->spa_suspend_lock);
1325 return (zio_wait(pio));
1329 zio_resume_wait(spa_t *spa)
1331 mutex_enter(&spa->spa_suspend_lock);
1332 while (spa_suspended(spa))
1333 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1334 mutex_exit(&spa->spa_suspend_lock);
1338 * ==========================================================================
1341 * A gang block is a collection of small blocks that looks to the DMU
1342 * like one large block. When zio_dva_allocate() cannot find a block
1343 * of the requested size, due to either severe fragmentation or the pool
1344 * being nearly full, it calls zio_write_gang_block() to construct the
1345 * block from smaller fragments.
1347 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1348 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1349 * an indirect block: it's an array of block pointers. It consumes
1350 * only one sector and hence is allocatable regardless of fragmentation.
1351 * The gang header's bps point to its gang members, which hold the data.
1353 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1354 * as the verifier to ensure uniqueness of the SHA256 checksum.
1355 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1356 * not the gang header. This ensures that data block signatures (needed for
1357 * deduplication) are independent of how the block is physically stored.
1359 * Gang blocks can be nested: a gang member may itself be a gang block.
1360 * Thus every gang block is a tree in which root and all interior nodes are
1361 * gang headers, and the leaves are normal blocks that contain user data.
1362 * The root of the gang tree is called the gang leader.
1364 * To perform any operation (read, rewrite, free, claim) on a gang block,
1365 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1366 * in the io_gang_tree field of the original logical i/o by recursively
1367 * reading the gang leader and all gang headers below it. This yields
1368 * an in-core tree containing the contents of every gang header and the
1369 * bps for every constituent of the gang block.
1371 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1372 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1373 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1374 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1375 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1376 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1377 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1378 * of the gang header plus zio_checksum_compute() of the data to update the
1379 * gang header's blk_cksum as described above.
1381 * The two-phase assemble/issue model solves the problem of partial failure --
1382 * what if you'd freed part of a gang block but then couldn't read the
1383 * gang header for another part? Assembling the entire gang tree first
1384 * ensures that all the necessary gang header I/O has succeeded before
1385 * starting the actual work of free, claim, or write. Once the gang tree
1386 * is assembled, free and claim are in-memory operations that cannot fail.
1388 * In the event that a gang write fails, zio_dva_unallocate() walks the
1389 * gang tree to immediately free (i.e. insert back into the space map)
1390 * everything we've allocated. This ensures that we don't get ENOSPC
1391 * errors during repeated suspend/resume cycles due to a flaky device.
1393 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1394 * the gang tree, we won't modify the block, so we can safely defer the free
1395 * (knowing that the block is still intact). If we *can* assemble the gang
1396 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1397 * each constituent bp and we can allocate a new block on the next sync pass.
1399 * In all cases, the gang tree allows complete recovery from partial failure.
1400 * ==========================================================================
1404 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1409 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1410 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1411 &pio->io_bookmark));
1415 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1420 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1421 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1422 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1424 * As we rewrite each gang header, the pipeline will compute
1425 * a new gang block header checksum for it; but no one will
1426 * compute a new data checksum, so we do that here. The one
1427 * exception is the gang leader: the pipeline already computed
1428 * its data checksum because that stage precedes gang assembly.
1429 * (Presently, nothing actually uses interior data checksums;
1430 * this is just good hygiene.)
1432 if (gn != pio->io_gang_leader->io_gang_tree) {
1433 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1434 data, BP_GET_PSIZE(bp));
1437 * If we are here to damage data for testing purposes,
1438 * leave the GBH alone so that we can detect the damage.
1440 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1441 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1443 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1444 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1445 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1453 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1455 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1456 ZIO_GANG_CHILD_FLAGS(pio)));
1461 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1463 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1464 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1467 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1476 static void zio_gang_tree_assemble_done(zio_t *zio);
1478 static zio_gang_node_t *
1479 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1481 zio_gang_node_t *gn;
1483 ASSERT(*gnpp == NULL);
1485 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1486 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1493 zio_gang_node_free(zio_gang_node_t **gnpp)
1495 zio_gang_node_t *gn = *gnpp;
1498 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1499 ASSERT(gn->gn_child[g] == NULL);
1501 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1502 kmem_free(gn, sizeof (*gn));
1507 zio_gang_tree_free(zio_gang_node_t **gnpp)
1509 zio_gang_node_t *gn = *gnpp;
1515 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1516 zio_gang_tree_free(&gn->gn_child[g]);
1518 zio_gang_node_free(gnpp);
1522 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1524 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1526 ASSERT(gio->io_gang_leader == gio);
1527 ASSERT(BP_IS_GANG(bp));
1529 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1530 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1531 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1535 zio_gang_tree_assemble_done(zio_t *zio)
1537 zio_t *gio = zio->io_gang_leader;
1538 zio_gang_node_t *gn = zio->io_private;
1539 blkptr_t *bp = zio->io_bp;
1542 ASSERT(gio == zio_unique_parent(zio));
1543 ASSERT(zio->io_child_count == 0);
1548 if (BP_SHOULD_BYTESWAP(bp))
1549 byteswap_uint64_array(zio->io_data, zio->io_size);
1551 ASSERT(zio->io_data == gn->gn_gbh);
1552 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1553 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1555 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1556 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1557 if (!BP_IS_GANG(gbp))
1559 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1564 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1566 zio_t *gio = pio->io_gang_leader;
1570 ASSERT(BP_IS_GANG(bp) == !!gn);
1571 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1572 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1575 * If you're a gang header, your data is in gn->gn_gbh.
1576 * If you're a gang member, your data is in 'data' and gn == NULL.
1578 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1581 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1583 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1584 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1585 if (BP_IS_HOLE(gbp))
1587 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1588 data = (char *)data + BP_GET_PSIZE(gbp);
1592 if (gn == gio->io_gang_tree)
1593 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1600 zio_gang_assemble(zio_t *zio)
1602 blkptr_t *bp = zio->io_bp;
1604 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1605 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1607 zio->io_gang_leader = zio;
1609 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1611 return (ZIO_PIPELINE_CONTINUE);
1615 zio_gang_issue(zio_t *zio)
1617 blkptr_t *bp = zio->io_bp;
1619 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1620 return (ZIO_PIPELINE_STOP);
1622 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1623 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1625 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1626 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1628 zio_gang_tree_free(&zio->io_gang_tree);
1630 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1632 return (ZIO_PIPELINE_CONTINUE);
1636 zio_write_gang_member_ready(zio_t *zio)
1638 zio_t *pio = zio_unique_parent(zio);
1639 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1640 dva_t *cdva = zio->io_bp->blk_dva;
1641 dva_t *pdva = pio->io_bp->blk_dva;
1645 if (BP_IS_HOLE(zio->io_bp))
1648 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1650 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1651 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1652 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1653 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1654 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1656 mutex_enter(&pio->io_lock);
1657 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1658 ASSERT(DVA_GET_GANG(&pdva[d]));
1659 asize = DVA_GET_ASIZE(&pdva[d]);
1660 asize += DVA_GET_ASIZE(&cdva[d]);
1661 DVA_SET_ASIZE(&pdva[d], asize);
1663 mutex_exit(&pio->io_lock);
1667 zio_write_gang_block(zio_t *pio)
1669 spa_t *spa = pio->io_spa;
1670 blkptr_t *bp = pio->io_bp;
1671 zio_t *gio = pio->io_gang_leader;
1673 zio_gang_node_t *gn, **gnpp;
1674 zio_gbh_phys_t *gbh;
1675 uint64_t txg = pio->io_txg;
1676 uint64_t resid = pio->io_size;
1678 int copies = gio->io_prop.zp_copies;
1679 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1683 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1684 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1685 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1687 pio->io_error = error;
1688 return (ZIO_PIPELINE_CONTINUE);
1692 gnpp = &gio->io_gang_tree;
1694 gnpp = pio->io_private;
1695 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1698 gn = zio_gang_node_alloc(gnpp);
1700 bzero(gbh, SPA_GANGBLOCKSIZE);
1703 * Create the gang header.
1705 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1706 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1709 * Create and nowait the gang children.
1711 for (g = 0; resid != 0; resid -= lsize, g++) {
1712 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1714 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1716 zp.zp_checksum = gio->io_prop.zp_checksum;
1717 zp.zp_compress = ZIO_COMPRESS_OFF;
1718 zp.zp_type = DMU_OT_NONE;
1720 zp.zp_copies = gio->io_prop.zp_copies;
1722 zp.zp_dedup_verify = 0;
1724 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1725 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1726 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1727 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1728 &pio->io_bookmark));
1732 * Set pio's pipeline to just wait for zio to finish.
1734 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1738 return (ZIO_PIPELINE_CONTINUE);
1742 * ==========================================================================
1744 * ==========================================================================
1747 zio_ddt_child_read_done(zio_t *zio)
1749 blkptr_t *bp = zio->io_bp;
1750 ddt_entry_t *dde = zio->io_private;
1752 zio_t *pio = zio_unique_parent(zio);
1754 mutex_enter(&pio->io_lock);
1755 ddp = ddt_phys_select(dde, bp);
1756 if (zio->io_error == 0)
1757 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1758 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1759 dde->dde_repair_data = zio->io_data;
1761 zio_buf_free(zio->io_data, zio->io_size);
1762 mutex_exit(&pio->io_lock);
1766 zio_ddt_read_start(zio_t *zio)
1768 blkptr_t *bp = zio->io_bp;
1771 ASSERT(BP_GET_DEDUP(bp));
1772 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1773 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1775 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1776 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1777 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1778 ddt_phys_t *ddp = dde->dde_phys;
1779 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1782 ASSERT(zio->io_vsd == NULL);
1785 if (ddp_self == NULL)
1786 return (ZIO_PIPELINE_CONTINUE);
1788 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1789 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1791 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1793 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1794 zio_buf_alloc(zio->io_size), zio->io_size,
1795 zio_ddt_child_read_done, dde, zio->io_priority,
1796 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1797 &zio->io_bookmark));
1799 return (ZIO_PIPELINE_CONTINUE);
1802 zio_nowait(zio_read(zio, zio->io_spa, bp,
1803 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1804 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1806 return (ZIO_PIPELINE_CONTINUE);
1810 zio_ddt_read_done(zio_t *zio)
1812 blkptr_t *bp = zio->io_bp;
1814 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1815 return (ZIO_PIPELINE_STOP);
1817 ASSERT(BP_GET_DEDUP(bp));
1818 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1819 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1821 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1822 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1823 ddt_entry_t *dde = zio->io_vsd;
1825 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1826 return (ZIO_PIPELINE_CONTINUE);
1829 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1830 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1831 return (ZIO_PIPELINE_STOP);
1833 if (dde->dde_repair_data != NULL) {
1834 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1835 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1837 ddt_repair_done(ddt, dde);
1841 ASSERT(zio->io_vsd == NULL);
1843 return (ZIO_PIPELINE_CONTINUE);
1847 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1849 spa_t *spa = zio->io_spa;
1853 * Note: we compare the original data, not the transformed data,
1854 * because when zio->io_bp is an override bp, we will not have
1855 * pushed the I/O transforms. That's an important optimization
1856 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1858 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1859 zio_t *lio = dde->dde_lead_zio[p];
1862 return (lio->io_orig_size != zio->io_orig_size ||
1863 bcmp(zio->io_orig_data, lio->io_orig_data,
1864 zio->io_orig_size) != 0);
1868 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1869 ddt_phys_t *ddp = &dde->dde_phys[p];
1871 if (ddp->ddp_phys_birth != 0) {
1872 arc_buf_t *abuf = NULL;
1873 uint32_t aflags = ARC_WAIT;
1874 blkptr_t blk = *zio->io_bp;
1877 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1881 error = arc_read_nolock(NULL, spa, &blk,
1882 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1883 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1884 &aflags, &zio->io_bookmark);
1887 if (arc_buf_size(abuf) != zio->io_orig_size ||
1888 bcmp(abuf->b_data, zio->io_orig_data,
1889 zio->io_orig_size) != 0)
1891 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1895 return (error != 0);
1903 zio_ddt_child_write_ready(zio_t *zio)
1905 int p = zio->io_prop.zp_copies;
1906 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1907 ddt_entry_t *dde = zio->io_private;
1908 ddt_phys_t *ddp = &dde->dde_phys[p];
1916 ASSERT(dde->dde_lead_zio[p] == zio);
1918 ddt_phys_fill(ddp, zio->io_bp);
1920 while ((pio = zio_walk_parents(zio)) != NULL)
1921 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1927 zio_ddt_child_write_done(zio_t *zio)
1929 int p = zio->io_prop.zp_copies;
1930 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1931 ddt_entry_t *dde = zio->io_private;
1932 ddt_phys_t *ddp = &dde->dde_phys[p];
1936 ASSERT(ddp->ddp_refcnt == 0);
1937 ASSERT(dde->dde_lead_zio[p] == zio);
1938 dde->dde_lead_zio[p] = NULL;
1940 if (zio->io_error == 0) {
1941 while (zio_walk_parents(zio) != NULL)
1942 ddt_phys_addref(ddp);
1944 ddt_phys_clear(ddp);
1951 zio_ddt_ditto_write_done(zio_t *zio)
1953 int p = DDT_PHYS_DITTO;
1954 blkptr_t *bp = zio->io_bp;
1955 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1956 ddt_entry_t *dde = zio->io_private;
1957 ddt_phys_t *ddp = &dde->dde_phys[p];
1958 ddt_key_t *ddk = &dde->dde_key;
1959 ASSERTV(zio_prop_t *zp = &zio->io_prop);
1963 ASSERT(ddp->ddp_refcnt == 0);
1964 ASSERT(dde->dde_lead_zio[p] == zio);
1965 dde->dde_lead_zio[p] = NULL;
1967 if (zio->io_error == 0) {
1968 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1969 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1970 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1971 if (ddp->ddp_phys_birth != 0)
1972 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1973 ddt_phys_fill(ddp, bp);
1980 zio_ddt_write(zio_t *zio)
1982 spa_t *spa = zio->io_spa;
1983 blkptr_t *bp = zio->io_bp;
1984 uint64_t txg = zio->io_txg;
1985 zio_prop_t *zp = &zio->io_prop;
1986 int p = zp->zp_copies;
1990 ddt_t *ddt = ddt_select(spa, bp);
1994 ASSERT(BP_GET_DEDUP(bp));
1995 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1996 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1999 dde = ddt_lookup(ddt, bp, B_TRUE);
2000 ddp = &dde->dde_phys[p];
2002 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2004 * If we're using a weak checksum, upgrade to a strong checksum
2005 * and try again. If we're already using a strong checksum,
2006 * we can't resolve it, so just convert to an ordinary write.
2007 * (And automatically e-mail a paper to Nature?)
2009 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2010 zp->zp_checksum = spa_dedup_checksum(spa);
2011 zio_pop_transforms(zio);
2012 zio->io_stage = ZIO_STAGE_OPEN;
2017 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2019 return (ZIO_PIPELINE_CONTINUE);
2022 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2023 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2025 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2026 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2027 zio_prop_t czp = *zp;
2029 czp.zp_copies = ditto_copies;
2032 * If we arrived here with an override bp, we won't have run
2033 * the transform stack, so we won't have the data we need to
2034 * generate a child i/o. So, toss the override bp and restart.
2035 * This is safe, because using the override bp is just an
2036 * optimization; and it's rare, so the cost doesn't matter.
2038 if (zio->io_bp_override) {
2039 zio_pop_transforms(zio);
2040 zio->io_stage = ZIO_STAGE_OPEN;
2041 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2042 zio->io_bp_override = NULL;
2045 return (ZIO_PIPELINE_CONTINUE);
2048 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2049 zio->io_orig_size, &czp, NULL,
2050 zio_ddt_ditto_write_done, dde, zio->io_priority,
2051 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2053 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2054 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2057 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2058 if (ddp->ddp_phys_birth != 0)
2059 ddt_bp_fill(ddp, bp, txg);
2060 if (dde->dde_lead_zio[p] != NULL)
2061 zio_add_child(zio, dde->dde_lead_zio[p]);
2063 ddt_phys_addref(ddp);
2064 } else if (zio->io_bp_override) {
2065 ASSERT(bp->blk_birth == txg);
2066 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2067 ddt_phys_fill(ddp, bp);
2068 ddt_phys_addref(ddp);
2070 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2071 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2072 zio_ddt_child_write_done, dde, zio->io_priority,
2073 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2075 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2076 dde->dde_lead_zio[p] = cio;
2086 return (ZIO_PIPELINE_CONTINUE);
2089 ddt_entry_t *freedde; /* for debugging */
2092 zio_ddt_free(zio_t *zio)
2094 spa_t *spa = zio->io_spa;
2095 blkptr_t *bp = zio->io_bp;
2096 ddt_t *ddt = ddt_select(spa, bp);
2100 ASSERT(BP_GET_DEDUP(bp));
2101 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2104 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2105 ddp = ddt_phys_select(dde, bp);
2106 ddt_phys_decref(ddp);
2109 return (ZIO_PIPELINE_CONTINUE);
2113 * ==========================================================================
2114 * Allocate and free blocks
2115 * ==========================================================================
2118 zio_dva_allocate(zio_t *zio)
2120 spa_t *spa = zio->io_spa;
2121 metaslab_class_t *mc = spa_normal_class(spa);
2122 blkptr_t *bp = zio->io_bp;
2125 if (zio->io_gang_leader == NULL) {
2126 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2127 zio->io_gang_leader = zio;
2130 ASSERT(BP_IS_HOLE(bp));
2131 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2132 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2133 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2134 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2136 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2137 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2140 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2141 return (zio_write_gang_block(zio));
2142 zio->io_error = error;
2145 return (ZIO_PIPELINE_CONTINUE);
2149 zio_dva_free(zio_t *zio)
2151 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2153 return (ZIO_PIPELINE_CONTINUE);
2157 zio_dva_claim(zio_t *zio)
2161 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2163 zio->io_error = error;
2165 return (ZIO_PIPELINE_CONTINUE);
2169 * Undo an allocation. This is used by zio_done() when an I/O fails
2170 * and we want to give back the block we just allocated.
2171 * This handles both normal blocks and gang blocks.
2174 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2178 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2179 ASSERT(zio->io_bp_override == NULL);
2181 if (!BP_IS_HOLE(bp))
2182 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2185 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2186 zio_dva_unallocate(zio, gn->gn_child[g],
2187 &gn->gn_gbh->zg_blkptr[g]);
2193 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2196 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2197 uint64_t size, boolean_t use_slog)
2201 ASSERT(txg > spa_syncing_txg(spa));
2204 error = metaslab_alloc(spa, spa_log_class(spa), size,
2205 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2208 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2209 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2212 BP_SET_LSIZE(new_bp, size);
2213 BP_SET_PSIZE(new_bp, size);
2214 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2215 BP_SET_CHECKSUM(new_bp,
2216 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2217 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2218 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2219 BP_SET_LEVEL(new_bp, 0);
2220 BP_SET_DEDUP(new_bp, 0);
2221 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2228 * Free an intent log block.
2231 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2233 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2234 ASSERT(!BP_IS_GANG(bp));
2236 zio_free(spa, txg, bp);
2240 * ==========================================================================
2241 * Read and write to physical devices
2242 * ==========================================================================
2245 zio_vdev_io_start(zio_t *zio)
2247 vdev_t *vd = zio->io_vd;
2249 spa_t *spa = zio->io_spa;
2251 ASSERT(zio->io_error == 0);
2252 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2255 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2256 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2259 * The mirror_ops handle multiple DVAs in a single BP.
2261 return (vdev_mirror_ops.vdev_op_io_start(zio));
2265 * We keep track of time-sensitive I/Os so that the scan thread
2266 * can quickly react to certain workloads. In particular, we care
2267 * about non-scrubbing, top-level reads and writes with the following
2269 * - synchronous writes of user data to non-slog devices
2270 * - any reads of user data
2271 * When these conditions are met, adjust the timestamp of spa_last_io
2272 * which allows the scan thread to adjust its workload accordingly.
2274 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2275 vd == vd->vdev_top && !vd->vdev_islog &&
2276 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2277 zio->io_txg != spa_syncing_txg(spa)) {
2278 uint64_t old = spa->spa_last_io;
2279 uint64_t new = ddi_get_lbolt64();
2281 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2284 align = 1ULL << vd->vdev_top->vdev_ashift;
2286 if (P2PHASE(zio->io_size, align) != 0) {
2287 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2288 char *abuf = zio_buf_alloc(asize);
2289 ASSERT(vd == vd->vdev_top);
2290 if (zio->io_type == ZIO_TYPE_WRITE) {
2291 bcopy(zio->io_data, abuf, zio->io_size);
2292 bzero(abuf + zio->io_size, asize - zio->io_size);
2294 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2297 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2298 ASSERT(P2PHASE(zio->io_size, align) == 0);
2299 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2302 * If this is a repair I/O, and there's no self-healing involved --
2303 * that is, we're just resilvering what we expect to resilver --
2304 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2305 * This prevents spurious resilvering with nested replication.
2306 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2307 * A is out of date, we'll read from C+D, then use the data to
2308 * resilver A+B -- but we don't actually want to resilver B, just A.
2309 * The top-level mirror has no way to know this, so instead we just
2310 * discard unnecessary repairs as we work our way down the vdev tree.
2311 * The same logic applies to any form of nested replication:
2312 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2314 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2315 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2316 zio->io_txg != 0 && /* not a delegated i/o */
2317 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2318 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2319 zio_vdev_io_bypass(zio);
2320 return (ZIO_PIPELINE_CONTINUE);
2323 if (vd->vdev_ops->vdev_op_leaf &&
2324 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2326 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2327 return (ZIO_PIPELINE_CONTINUE);
2329 if ((zio = vdev_queue_io(zio)) == NULL)
2330 return (ZIO_PIPELINE_STOP);
2332 if (!vdev_accessible(vd, zio)) {
2333 zio->io_error = ENXIO;
2335 return (ZIO_PIPELINE_STOP);
2339 return (vd->vdev_ops->vdev_op_io_start(zio));
2343 zio_vdev_io_done(zio_t *zio)
2345 vdev_t *vd = zio->io_vd;
2346 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2347 boolean_t unexpected_error = B_FALSE;
2349 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2350 return (ZIO_PIPELINE_STOP);
2352 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2354 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2356 vdev_queue_io_done(zio);
2358 if (zio->io_type == ZIO_TYPE_WRITE)
2359 vdev_cache_write(zio);
2361 if (zio_injection_enabled && zio->io_error == 0)
2362 zio->io_error = zio_handle_device_injection(vd,
2365 if (zio_injection_enabled && zio->io_error == 0)
2366 zio->io_error = zio_handle_label_injection(zio, EIO);
2368 if (zio->io_error) {
2369 if (!vdev_accessible(vd, zio)) {
2370 zio->io_error = ENXIO;
2372 unexpected_error = B_TRUE;
2377 ops->vdev_op_io_done(zio);
2379 if (unexpected_error)
2380 VERIFY(vdev_probe(vd, zio) == NULL);
2382 return (ZIO_PIPELINE_CONTINUE);
2386 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2387 * disk, and use that to finish the checksum ereport later.
2390 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2391 const void *good_buf)
2393 /* no processing needed */
2394 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2399 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2401 void *buf = zio_buf_alloc(zio->io_size);
2403 bcopy(zio->io_data, buf, zio->io_size);
2405 zcr->zcr_cbinfo = zio->io_size;
2406 zcr->zcr_cbdata = buf;
2407 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2408 zcr->zcr_free = zio_buf_free;
2412 zio_vdev_io_assess(zio_t *zio)
2414 vdev_t *vd = zio->io_vd;
2416 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2417 return (ZIO_PIPELINE_STOP);
2419 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2420 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2422 if (zio->io_vsd != NULL) {
2423 zio->io_vsd_ops->vsd_free(zio);
2427 if (zio_injection_enabled && zio->io_error == 0)
2428 zio->io_error = zio_handle_fault_injection(zio, EIO);
2431 * If the I/O failed, determine whether we should attempt to retry it.
2433 * On retry, we cut in line in the issue queue, since we don't want
2434 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2436 if (zio->io_error && vd == NULL &&
2437 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2438 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2439 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2441 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2442 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2443 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2444 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2445 zio_requeue_io_start_cut_in_line);
2446 return (ZIO_PIPELINE_STOP);
2450 * If we got an error on a leaf device, convert it to ENXIO
2451 * if the device is not accessible at all.
2453 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2454 !vdev_accessible(vd, zio))
2455 zio->io_error = ENXIO;
2458 * If we can't write to an interior vdev (mirror or RAID-Z),
2459 * set vdev_cant_write so that we stop trying to allocate from it.
2461 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2462 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2463 vd->vdev_cant_write = B_TRUE;
2466 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2468 return (ZIO_PIPELINE_CONTINUE);
2472 zio_vdev_io_reissue(zio_t *zio)
2474 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2475 ASSERT(zio->io_error == 0);
2477 zio->io_stage >>= 1;
2481 zio_vdev_io_redone(zio_t *zio)
2483 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2485 zio->io_stage >>= 1;
2489 zio_vdev_io_bypass(zio_t *zio)
2491 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2492 ASSERT(zio->io_error == 0);
2494 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2495 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2499 * ==========================================================================
2500 * Generate and verify checksums
2501 * ==========================================================================
2504 zio_checksum_generate(zio_t *zio)
2506 blkptr_t *bp = zio->io_bp;
2507 enum zio_checksum checksum;
2511 * This is zio_write_phys().
2512 * We're either generating a label checksum, or none at all.
2514 checksum = zio->io_prop.zp_checksum;
2516 if (checksum == ZIO_CHECKSUM_OFF)
2517 return (ZIO_PIPELINE_CONTINUE);
2519 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2521 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2522 ASSERT(!IO_IS_ALLOCATING(zio));
2523 checksum = ZIO_CHECKSUM_GANG_HEADER;
2525 checksum = BP_GET_CHECKSUM(bp);
2529 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2531 return (ZIO_PIPELINE_CONTINUE);
2535 zio_checksum_verify(zio_t *zio)
2537 zio_bad_cksum_t info;
2538 blkptr_t *bp = zio->io_bp;
2541 ASSERT(zio->io_vd != NULL);
2545 * This is zio_read_phys().
2546 * We're either verifying a label checksum, or nothing at all.
2548 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2549 return (ZIO_PIPELINE_CONTINUE);
2551 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2554 if ((error = zio_checksum_error(zio, &info)) != 0) {
2555 zio->io_error = error;
2556 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2557 zfs_ereport_start_checksum(zio->io_spa,
2558 zio->io_vd, zio, zio->io_offset,
2559 zio->io_size, NULL, &info);
2563 return (ZIO_PIPELINE_CONTINUE);
2567 * Called by RAID-Z to ensure we don't compute the checksum twice.
2570 zio_checksum_verified(zio_t *zio)
2572 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2576 * ==========================================================================
2577 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2578 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2579 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2580 * indicate errors that are specific to one I/O, and most likely permanent.
2581 * Any other error is presumed to be worse because we weren't expecting it.
2582 * ==========================================================================
2585 zio_worst_error(int e1, int e2)
2587 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2590 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2591 if (e1 == zio_error_rank[r1])
2594 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2595 if (e2 == zio_error_rank[r2])
2598 return (r1 > r2 ? e1 : e2);
2602 * ==========================================================================
2604 * ==========================================================================
2607 zio_ready(zio_t *zio)
2609 blkptr_t *bp = zio->io_bp;
2610 zio_t *pio, *pio_next;
2612 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2613 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2614 return (ZIO_PIPELINE_STOP);
2616 if (zio->io_ready) {
2617 ASSERT(IO_IS_ALLOCATING(zio));
2618 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2619 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2624 if (bp != NULL && bp != &zio->io_bp_copy)
2625 zio->io_bp_copy = *bp;
2628 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2630 mutex_enter(&zio->io_lock);
2631 zio->io_state[ZIO_WAIT_READY] = 1;
2632 pio = zio_walk_parents(zio);
2633 mutex_exit(&zio->io_lock);
2636 * As we notify zio's parents, new parents could be added.
2637 * New parents go to the head of zio's io_parent_list, however,
2638 * so we will (correctly) not notify them. The remainder of zio's
2639 * io_parent_list, from 'pio_next' onward, cannot change because
2640 * all parents must wait for us to be done before they can be done.
2642 for (; pio != NULL; pio = pio_next) {
2643 pio_next = zio_walk_parents(zio);
2644 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2647 if (zio->io_flags & ZIO_FLAG_NODATA) {
2648 if (BP_IS_GANG(bp)) {
2649 zio->io_flags &= ~ZIO_FLAG_NODATA;
2651 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2652 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2656 if (zio_injection_enabled &&
2657 zio->io_spa->spa_syncing_txg == zio->io_txg)
2658 zio_handle_ignored_writes(zio);
2660 return (ZIO_PIPELINE_CONTINUE);
2664 zio_done(zio_t *zio)
2666 spa_t *spa = zio->io_spa;
2667 zio_t *lio = zio->io_logical;
2668 blkptr_t *bp = zio->io_bp;
2669 vdev_t *vd = zio->io_vd;
2670 uint64_t psize = zio->io_size;
2671 zio_t *pio, *pio_next;
2675 * If our children haven't all completed,
2676 * wait for them and then repeat this pipeline stage.
2678 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2679 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2680 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2681 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2682 return (ZIO_PIPELINE_STOP);
2684 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2685 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2686 ASSERT(zio->io_children[c][w] == 0);
2689 ASSERT(bp->blk_pad[0] == 0);
2690 ASSERT(bp->blk_pad[1] == 0);
2691 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2692 (bp == zio_unique_parent(zio)->io_bp));
2693 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2694 zio->io_bp_override == NULL &&
2695 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2696 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2697 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2698 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2699 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2704 * If there were child vdev/gang/ddt errors, they apply to us now.
2706 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2707 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2708 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2711 * If the I/O on the transformed data was successful, generate any
2712 * checksum reports now while we still have the transformed data.
2714 if (zio->io_error == 0) {
2715 while (zio->io_cksum_report != NULL) {
2716 zio_cksum_report_t *zcr = zio->io_cksum_report;
2717 uint64_t align = zcr->zcr_align;
2718 uint64_t asize = P2ROUNDUP(psize, align);
2719 char *abuf = zio->io_data;
2721 if (asize != psize) {
2722 abuf = zio_buf_alloc(asize);
2723 bcopy(zio->io_data, abuf, psize);
2724 bzero(abuf + psize, asize - psize);
2727 zio->io_cksum_report = zcr->zcr_next;
2728 zcr->zcr_next = NULL;
2729 zcr->zcr_finish(zcr, abuf);
2730 zfs_ereport_free_checksum(zcr);
2733 zio_buf_free(abuf, asize);
2737 zio_pop_transforms(zio); /* note: may set zio->io_error */
2739 vdev_stat_update(zio, psize);
2741 if (zio->io_error) {
2743 * If this I/O is attached to a particular vdev,
2744 * generate an error message describing the I/O failure
2745 * at the block level. We ignore these errors if the
2746 * device is currently unavailable.
2748 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2749 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2751 if ((zio->io_error == EIO || !(zio->io_flags &
2752 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2755 * For logical I/O requests, tell the SPA to log the
2756 * error and generate a logical data ereport.
2758 spa_log_error(spa, zio);
2759 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2764 if (zio->io_error && zio == lio) {
2766 * Determine whether zio should be reexecuted. This will
2767 * propagate all the way to the root via zio_notify_parent().
2769 ASSERT(vd == NULL && bp != NULL);
2770 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2772 if (IO_IS_ALLOCATING(zio) &&
2773 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2774 if (zio->io_error != ENOSPC)
2775 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2777 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2780 if ((zio->io_type == ZIO_TYPE_READ ||
2781 zio->io_type == ZIO_TYPE_FREE) &&
2782 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2783 zio->io_error == ENXIO &&
2784 spa_load_state(spa) == SPA_LOAD_NONE &&
2785 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2786 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2788 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2789 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2792 * Here is a possibly good place to attempt to do
2793 * either combinatorial reconstruction or error correction
2794 * based on checksums. It also might be a good place
2795 * to send out preliminary ereports before we suspend
2801 * If there were logical child errors, they apply to us now.
2802 * We defer this until now to avoid conflating logical child
2803 * errors with errors that happened to the zio itself when
2804 * updating vdev stats and reporting FMA events above.
2806 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2808 if ((zio->io_error || zio->io_reexecute) &&
2809 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2810 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2811 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2813 zio_gang_tree_free(&zio->io_gang_tree);
2816 * Godfather I/Os should never suspend.
2818 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2819 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2820 zio->io_reexecute = 0;
2822 if (zio->io_reexecute) {
2824 * This is a logical I/O that wants to reexecute.
2826 * Reexecute is top-down. When an i/o fails, if it's not
2827 * the root, it simply notifies its parent and sticks around.
2828 * The parent, seeing that it still has children in zio_done(),
2829 * does the same. This percolates all the way up to the root.
2830 * The root i/o will reexecute or suspend the entire tree.
2832 * This approach ensures that zio_reexecute() honors
2833 * all the original i/o dependency relationships, e.g.
2834 * parents not executing until children are ready.
2836 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2838 zio->io_gang_leader = NULL;
2840 mutex_enter(&zio->io_lock);
2841 zio->io_state[ZIO_WAIT_DONE] = 1;
2842 mutex_exit(&zio->io_lock);
2845 * "The Godfather" I/O monitors its children but is
2846 * not a true parent to them. It will track them through
2847 * the pipeline but severs its ties whenever they get into
2848 * trouble (e.g. suspended). This allows "The Godfather"
2849 * I/O to return status without blocking.
2851 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2852 zio_link_t *zl = zio->io_walk_link;
2853 pio_next = zio_walk_parents(zio);
2855 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2856 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2857 zio_remove_child(pio, zio, zl);
2858 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2862 if ((pio = zio_unique_parent(zio)) != NULL) {
2864 * We're not a root i/o, so there's nothing to do
2865 * but notify our parent. Don't propagate errors
2866 * upward since we haven't permanently failed yet.
2868 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2869 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2870 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2871 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2873 * We'd fail again if we reexecuted now, so suspend
2874 * until conditions improve (e.g. device comes online).
2876 zio_suspend(spa, zio);
2879 * Reexecution is potentially a huge amount of work.
2880 * Hand it off to the otherwise-unused claim taskq.
2882 (void) taskq_dispatch(
2883 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2884 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2886 return (ZIO_PIPELINE_STOP);
2889 ASSERT(zio->io_child_count == 0);
2890 ASSERT(zio->io_reexecute == 0);
2891 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2894 * Report any checksum errors, since the I/O is complete.
2896 while (zio->io_cksum_report != NULL) {
2897 zio_cksum_report_t *zcr = zio->io_cksum_report;
2898 zio->io_cksum_report = zcr->zcr_next;
2899 zcr->zcr_next = NULL;
2900 zcr->zcr_finish(zcr, NULL);
2901 zfs_ereport_free_checksum(zcr);
2905 * It is the responsibility of the done callback to ensure that this
2906 * particular zio is no longer discoverable for adoption, and as
2907 * such, cannot acquire any new parents.
2912 mutex_enter(&zio->io_lock);
2913 zio->io_state[ZIO_WAIT_DONE] = 1;
2914 mutex_exit(&zio->io_lock);
2916 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2917 zio_link_t *zl = zio->io_walk_link;
2918 pio_next = zio_walk_parents(zio);
2919 zio_remove_child(pio, zio, zl);
2920 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2923 if (zio->io_waiter != NULL) {
2924 mutex_enter(&zio->io_lock);
2925 zio->io_executor = NULL;
2926 cv_broadcast(&zio->io_cv);
2927 mutex_exit(&zio->io_lock);
2932 return (ZIO_PIPELINE_STOP);
2936 * ==========================================================================
2937 * I/O pipeline definition
2938 * ==========================================================================
2940 static zio_pipe_stage_t *zio_pipeline[] = {
2946 zio_checksum_generate,
2960 zio_checksum_verify,