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);
440 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
442 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
443 int *errorp = &pio->io_child_error[zio->io_child_type];
445 mutex_enter(&pio->io_lock);
446 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
447 *errorp = zio_worst_error(*errorp, zio->io_error);
448 pio->io_reexecute |= zio->io_reexecute;
449 ASSERT3U(*countp, >, 0);
450 if (--*countp == 0 && pio->io_stall == countp) {
451 pio->io_stall = NULL;
452 mutex_exit(&pio->io_lock);
455 mutex_exit(&pio->io_lock);
460 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
462 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
463 zio->io_error = zio->io_child_error[c];
467 * ==========================================================================
468 * Create the various types of I/O (read, write, free, etc)
469 * ==========================================================================
472 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
473 void *data, uint64_t size, zio_done_func_t *done, void *private,
474 zio_type_t type, int priority, enum zio_flag flags,
475 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
476 enum zio_stage stage, enum zio_stage pipeline)
480 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
481 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
482 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
484 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
485 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
486 ASSERT(vd || stage == ZIO_STAGE_OPEN);
488 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
489 bzero(zio, sizeof (zio_t));
491 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
492 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
494 list_create(&zio->io_parent_list, sizeof (zio_link_t),
495 offsetof(zio_link_t, zl_parent_node));
496 list_create(&zio->io_child_list, sizeof (zio_link_t),
497 offsetof(zio_link_t, zl_child_node));
500 zio->io_child_type = ZIO_CHILD_VDEV;
501 else if (flags & ZIO_FLAG_GANG_CHILD)
502 zio->io_child_type = ZIO_CHILD_GANG;
503 else if (flags & ZIO_FLAG_DDT_CHILD)
504 zio->io_child_type = ZIO_CHILD_DDT;
506 zio->io_child_type = ZIO_CHILD_LOGICAL;
509 zio->io_bp = (blkptr_t *)bp;
510 zio->io_bp_copy = *bp;
511 zio->io_bp_orig = *bp;
512 if (type != ZIO_TYPE_WRITE ||
513 zio->io_child_type == ZIO_CHILD_DDT)
514 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
515 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
516 zio->io_logical = zio;
517 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
518 pipeline |= ZIO_GANG_STAGES;
524 zio->io_private = private;
526 zio->io_priority = priority;
528 zio->io_offset = offset;
529 zio->io_orig_data = zio->io_data = data;
530 zio->io_orig_size = zio->io_size = size;
531 zio->io_orig_flags = zio->io_flags = flags;
532 zio->io_orig_stage = zio->io_stage = stage;
533 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
535 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
536 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
539 zio->io_bookmark = *zb;
542 if (zio->io_logical == NULL)
543 zio->io_logical = pio->io_logical;
544 if (zio->io_child_type == ZIO_CHILD_GANG)
545 zio->io_gang_leader = pio->io_gang_leader;
546 zio_add_child(pio, zio);
553 zio_destroy(zio_t *zio)
555 list_destroy(&zio->io_parent_list);
556 list_destroy(&zio->io_child_list);
557 mutex_destroy(&zio->io_lock);
558 cv_destroy(&zio->io_cv);
559 kmem_cache_free(zio_cache, zio);
563 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
564 void *private, enum zio_flag flags)
568 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
569 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
570 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
576 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
578 return (zio_null(NULL, spa, NULL, done, private, flags));
582 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
583 void *data, uint64_t size, zio_done_func_t *done, void *private,
584 int priority, enum zio_flag flags, const zbookmark_t *zb)
588 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
589 data, size, done, private,
590 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
591 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
592 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
598 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
599 void *data, uint64_t size, const zio_prop_t *zp,
600 zio_done_func_t *ready, zio_done_func_t *done, void *private,
601 int priority, enum zio_flag flags, const zbookmark_t *zb)
605 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
606 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
607 zp->zp_compress >= ZIO_COMPRESS_OFF &&
608 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
609 zp->zp_type < DMU_OT_NUMTYPES &&
612 zp->zp_copies <= spa_max_replication(spa) &&
614 zp->zp_dedup_verify <= 1);
616 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
617 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
618 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
619 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
621 zio->io_ready = ready;
628 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
629 uint64_t size, zio_done_func_t *done, void *private, int priority,
630 enum zio_flag flags, zbookmark_t *zb)
634 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
635 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
636 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
642 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
644 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
645 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
646 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
647 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
649 zio->io_prop.zp_copies = copies;
650 zio->io_bp_override = bp;
654 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
656 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
660 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
665 dprintf_bp(bp, "freeing in txg %llu, pass %u",
666 (longlong_t)txg, spa->spa_sync_pass);
668 ASSERT(!BP_IS_HOLE(bp));
669 ASSERT(spa_syncing_txg(spa) == txg);
670 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
672 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
673 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
674 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
680 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
681 zio_done_func_t *done, void *private, enum zio_flag flags)
686 * A claim is an allocation of a specific block. Claims are needed
687 * to support immediate writes in the intent log. The issue is that
688 * immediate writes contain committed data, but in a txg that was
689 * *not* committed. Upon opening the pool after an unclean shutdown,
690 * the intent log claims all blocks that contain immediate write data
691 * so that the SPA knows they're in use.
693 * All claims *must* be resolved in the first txg -- before the SPA
694 * starts allocating blocks -- so that nothing is allocated twice.
695 * If txg == 0 we just verify that the block is claimable.
697 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
698 ASSERT(txg == spa_first_txg(spa) || txg == 0);
699 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
701 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
702 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
703 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
709 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
710 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
715 if (vd->vdev_children == 0) {
716 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
717 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
718 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
722 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
724 for (c = 0; c < vd->vdev_children; c++)
725 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
726 done, private, priority, flags));
733 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
734 void *data, int checksum, zio_done_func_t *done, void *private,
735 int priority, enum zio_flag flags, boolean_t labels)
739 ASSERT(vd->vdev_children == 0);
740 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
741 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
742 ASSERT3U(offset + size, <=, vd->vdev_psize);
744 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
745 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
746 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
748 zio->io_prop.zp_checksum = checksum;
754 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
755 void *data, int checksum, zio_done_func_t *done, void *private,
756 int priority, enum zio_flag flags, boolean_t labels)
760 ASSERT(vd->vdev_children == 0);
761 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
762 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
763 ASSERT3U(offset + size, <=, vd->vdev_psize);
765 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
766 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
767 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
769 zio->io_prop.zp_checksum = checksum;
771 if (zio_checksum_table[checksum].ci_eck) {
773 * zec checksums are necessarily destructive -- they modify
774 * the end of the write buffer to hold the verifier/checksum.
775 * Therefore, we must make a local copy in case the data is
776 * being written to multiple places in parallel.
778 void *wbuf = zio_buf_alloc(size);
779 bcopy(data, wbuf, size);
780 zio_push_transform(zio, wbuf, size, size, NULL);
787 * Create a child I/O to do some work for us.
790 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
791 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
792 zio_done_func_t *done, void *private)
794 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
797 ASSERT(vd->vdev_parent ==
798 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
800 if (type == ZIO_TYPE_READ && bp != NULL) {
802 * If we have the bp, then the child should perform the
803 * checksum and the parent need not. This pushes error
804 * detection as close to the leaves as possible and
805 * eliminates redundant checksums in the interior nodes.
807 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
808 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
811 if (vd->vdev_children == 0)
812 offset += VDEV_LABEL_START_SIZE;
814 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
817 * If we've decided to do a repair, the write is not speculative --
818 * even if the original read was.
820 if (flags & ZIO_FLAG_IO_REPAIR)
821 flags &= ~ZIO_FLAG_SPECULATIVE;
823 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
824 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
825 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
831 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
832 int type, int priority, enum zio_flag flags,
833 zio_done_func_t *done, void *private)
837 ASSERT(vd->vdev_ops->vdev_op_leaf);
839 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
840 data, size, done, private, type, priority,
841 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
843 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
849 zio_flush(zio_t *zio, vdev_t *vd)
851 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
852 NULL, NULL, ZIO_PRIORITY_NOW,
853 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
857 zio_shrink(zio_t *zio, uint64_t size)
859 ASSERT(zio->io_executor == NULL);
860 ASSERT(zio->io_orig_size == zio->io_size);
861 ASSERT(size <= zio->io_size);
864 * We don't shrink for raidz because of problems with the
865 * reconstruction when reading back less than the block size.
866 * Note, BP_IS_RAIDZ() assumes no compression.
868 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
869 if (!BP_IS_RAIDZ(zio->io_bp))
870 zio->io_orig_size = zio->io_size = size;
874 * ==========================================================================
875 * Prepare to read and write logical blocks
876 * ==========================================================================
880 zio_read_bp_init(zio_t *zio)
882 blkptr_t *bp = zio->io_bp;
884 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
885 zio->io_child_type == ZIO_CHILD_LOGICAL &&
886 !(zio->io_flags & ZIO_FLAG_RAW)) {
887 uint64_t psize = BP_GET_PSIZE(bp);
888 void *cbuf = zio_buf_alloc(psize);
890 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
893 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
894 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
896 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
897 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
899 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
900 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
902 return (ZIO_PIPELINE_CONTINUE);
906 zio_write_bp_init(zio_t *zio)
908 spa_t *spa = zio->io_spa;
909 zio_prop_t *zp = &zio->io_prop;
910 enum zio_compress compress = zp->zp_compress;
911 blkptr_t *bp = zio->io_bp;
912 uint64_t lsize = zio->io_size;
913 uint64_t psize = lsize;
917 * If our children haven't all reached the ready stage,
918 * wait for them and then repeat this pipeline stage.
920 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
921 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
922 return (ZIO_PIPELINE_STOP);
924 if (!IO_IS_ALLOCATING(zio))
925 return (ZIO_PIPELINE_CONTINUE);
927 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
929 if (zio->io_bp_override) {
930 ASSERT(bp->blk_birth != zio->io_txg);
931 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
933 *bp = *zio->io_bp_override;
934 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
936 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
937 return (ZIO_PIPELINE_CONTINUE);
939 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
940 zp->zp_dedup_verify);
942 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
944 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
945 return (ZIO_PIPELINE_CONTINUE);
947 zio->io_bp_override = NULL;
951 if (bp->blk_birth == zio->io_txg) {
953 * We're rewriting an existing block, which means we're
954 * working on behalf of spa_sync(). For spa_sync() to
955 * converge, it must eventually be the case that we don't
956 * have to allocate new blocks. But compression changes
957 * the blocksize, which forces a reallocate, and makes
958 * convergence take longer. Therefore, after the first
959 * few passes, stop compressing to ensure convergence.
961 pass = spa_sync_pass(spa);
963 ASSERT(zio->io_txg == spa_syncing_txg(spa));
964 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
965 ASSERT(!BP_GET_DEDUP(bp));
967 if (pass > SYNC_PASS_DONT_COMPRESS)
968 compress = ZIO_COMPRESS_OFF;
970 /* Make sure someone doesn't change their mind on overwrites */
971 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
972 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
975 if (compress != ZIO_COMPRESS_OFF) {
976 void *cbuf = zio_buf_alloc(lsize);
977 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
978 if (psize == 0 || psize == lsize) {
979 compress = ZIO_COMPRESS_OFF;
980 zio_buf_free(cbuf, lsize);
982 ASSERT(psize < lsize);
983 zio_push_transform(zio, cbuf, psize, lsize, NULL);
988 * The final pass of spa_sync() must be all rewrites, but the first
989 * few passes offer a trade-off: allocating blocks defers convergence,
990 * but newly allocated blocks are sequential, so they can be written
991 * to disk faster. Therefore, we allow the first few passes of
992 * spa_sync() to allocate new blocks, but force rewrites after that.
993 * There should only be a handful of blocks after pass 1 in any case.
995 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
996 pass > SYNC_PASS_REWRITE) {
997 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
999 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1000 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1003 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1007 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1009 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1010 BP_SET_LSIZE(bp, lsize);
1011 BP_SET_PSIZE(bp, psize);
1012 BP_SET_COMPRESS(bp, compress);
1013 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1014 BP_SET_TYPE(bp, zp->zp_type);
1015 BP_SET_LEVEL(bp, zp->zp_level);
1016 BP_SET_DEDUP(bp, zp->zp_dedup);
1017 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1019 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1020 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1021 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1025 return (ZIO_PIPELINE_CONTINUE);
1029 zio_free_bp_init(zio_t *zio)
1031 blkptr_t *bp = zio->io_bp;
1033 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1034 if (BP_GET_DEDUP(bp))
1035 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1038 return (ZIO_PIPELINE_CONTINUE);
1042 * ==========================================================================
1043 * Execute the I/O pipeline
1044 * ==========================================================================
1048 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1050 spa_t *spa = zio->io_spa;
1051 zio_type_t t = zio->io_type;
1052 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1055 * If we're a config writer or a probe, the normal issue and
1056 * interrupt threads may all be blocked waiting for the config lock.
1057 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1059 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1063 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1065 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1069 * If this is a high priority I/O, then use the high priority taskq.
1071 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1072 spa->spa_zio_taskq[t][q + 1] != NULL)
1075 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1076 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1077 (task_func_t *)zio_execute, zio, flags);
1081 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1083 kthread_t *executor = zio->io_executor;
1084 spa_t *spa = zio->io_spa;
1087 for (t = 0; t < ZIO_TYPES; t++)
1088 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1095 zio_issue_async(zio_t *zio)
1097 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1099 return (ZIO_PIPELINE_STOP);
1103 zio_interrupt(zio_t *zio)
1105 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1109 * Execute the I/O pipeline until one of the following occurs:
1110 * (1) the I/O completes; (2) the pipeline stalls waiting for
1111 * dependent child I/Os; (3) the I/O issues, so we're waiting
1112 * for an I/O completion interrupt; (4) the I/O is delegated by
1113 * vdev-level caching or aggregation; (5) the I/O is deferred
1114 * due to vdev-level queueing; (6) the I/O is handed off to
1115 * another thread. In all cases, the pipeline stops whenever
1116 * there's no CPU work; it never burns a thread in cv_wait().
1118 * There's no locking on io_stage because there's no legitimate way
1119 * for multiple threads to be attempting to process the same I/O.
1121 static zio_pipe_stage_t *zio_pipeline[];
1124 zio_execute(zio_t *zio)
1126 zio->io_executor = curthread;
1128 while (zio->io_stage < ZIO_STAGE_DONE) {
1129 enum zio_stage pipeline = zio->io_pipeline;
1130 enum zio_stage stage = zio->io_stage;
1133 ASSERT(!MUTEX_HELD(&zio->io_lock));
1134 ASSERT(ISP2(stage));
1135 ASSERT(zio->io_stall == NULL);
1139 } while ((stage & pipeline) == 0);
1141 ASSERT(stage <= ZIO_STAGE_DONE);
1144 * If we are in interrupt context and this pipeline stage
1145 * will grab a config lock that is held across I/O,
1146 * or may wait for an I/O that needs an interrupt thread
1147 * to complete, issue async to avoid deadlock.
1149 * For VDEV_IO_START, we cut in line so that the io will
1150 * be sent to disk promptly.
1152 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1153 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1154 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1155 zio_requeue_io_start_cut_in_line : B_FALSE;
1156 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1160 zio->io_stage = stage;
1161 rv = zio_pipeline[highbit(stage) - 1](zio);
1163 if (rv == ZIO_PIPELINE_STOP)
1166 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1171 * ==========================================================================
1172 * Initiate I/O, either sync or async
1173 * ==========================================================================
1176 zio_wait(zio_t *zio)
1180 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1181 ASSERT(zio->io_executor == NULL);
1183 zio->io_waiter = curthread;
1187 mutex_enter(&zio->io_lock);
1188 while (zio->io_executor != NULL)
1189 cv_wait(&zio->io_cv, &zio->io_lock);
1190 mutex_exit(&zio->io_lock);
1192 error = zio->io_error;
1199 zio_nowait(zio_t *zio)
1201 ASSERT(zio->io_executor == NULL);
1203 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1204 zio_unique_parent(zio) == NULL) {
1206 * This is a logical async I/O with no parent to wait for it.
1207 * We add it to the spa_async_root_zio "Godfather" I/O which
1208 * will ensure they complete prior to unloading the pool.
1210 spa_t *spa = zio->io_spa;
1212 zio_add_child(spa->spa_async_zio_root, zio);
1219 * ==========================================================================
1220 * Reexecute or suspend/resume failed I/O
1221 * ==========================================================================
1225 zio_reexecute(zio_t *pio)
1227 zio_t *cio, *cio_next;
1230 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1231 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1232 ASSERT(pio->io_gang_leader == NULL);
1233 ASSERT(pio->io_gang_tree == NULL);
1235 pio->io_flags = pio->io_orig_flags;
1236 pio->io_stage = pio->io_orig_stage;
1237 pio->io_pipeline = pio->io_orig_pipeline;
1238 pio->io_reexecute = 0;
1240 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1241 pio->io_state[w] = 0;
1242 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1243 pio->io_child_error[c] = 0;
1245 if (IO_IS_ALLOCATING(pio))
1246 BP_ZERO(pio->io_bp);
1249 * As we reexecute pio's children, new children could be created.
1250 * New children go to the head of pio's io_child_list, however,
1251 * so we will (correctly) not reexecute them. The key is that
1252 * the remainder of pio's io_child_list, from 'cio_next' onward,
1253 * cannot be affected by any side effects of reexecuting 'cio'.
1255 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1256 cio_next = zio_walk_children(pio);
1257 mutex_enter(&pio->io_lock);
1258 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1259 pio->io_children[cio->io_child_type][w]++;
1260 mutex_exit(&pio->io_lock);
1265 * Now that all children have been reexecuted, execute the parent.
1266 * We don't reexecute "The Godfather" I/O here as it's the
1267 * responsibility of the caller to wait on him.
1269 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1274 zio_suspend(spa_t *spa, zio_t *zio)
1276 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1277 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1278 "failure and the failure mode property for this pool "
1279 "is set to panic.", spa_name(spa));
1281 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1283 mutex_enter(&spa->spa_suspend_lock);
1285 if (spa->spa_suspend_zio_root == NULL)
1286 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1287 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1288 ZIO_FLAG_GODFATHER);
1290 spa->spa_suspended = B_TRUE;
1293 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1294 ASSERT(zio != spa->spa_suspend_zio_root);
1295 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1296 ASSERT(zio_unique_parent(zio) == NULL);
1297 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1298 zio_add_child(spa->spa_suspend_zio_root, zio);
1301 mutex_exit(&spa->spa_suspend_lock);
1305 zio_resume(spa_t *spa)
1310 * Reexecute all previously suspended i/o.
1312 mutex_enter(&spa->spa_suspend_lock);
1313 spa->spa_suspended = B_FALSE;
1314 cv_broadcast(&spa->spa_suspend_cv);
1315 pio = spa->spa_suspend_zio_root;
1316 spa->spa_suspend_zio_root = NULL;
1317 mutex_exit(&spa->spa_suspend_lock);
1323 return (zio_wait(pio));
1327 zio_resume_wait(spa_t *spa)
1329 mutex_enter(&spa->spa_suspend_lock);
1330 while (spa_suspended(spa))
1331 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1332 mutex_exit(&spa->spa_suspend_lock);
1336 * ==========================================================================
1339 * A gang block is a collection of small blocks that looks to the DMU
1340 * like one large block. When zio_dva_allocate() cannot find a block
1341 * of the requested size, due to either severe fragmentation or the pool
1342 * being nearly full, it calls zio_write_gang_block() to construct the
1343 * block from smaller fragments.
1345 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1346 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1347 * an indirect block: it's an array of block pointers. It consumes
1348 * only one sector and hence is allocatable regardless of fragmentation.
1349 * The gang header's bps point to its gang members, which hold the data.
1351 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1352 * as the verifier to ensure uniqueness of the SHA256 checksum.
1353 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1354 * not the gang header. This ensures that data block signatures (needed for
1355 * deduplication) are independent of how the block is physically stored.
1357 * Gang blocks can be nested: a gang member may itself be a gang block.
1358 * Thus every gang block is a tree in which root and all interior nodes are
1359 * gang headers, and the leaves are normal blocks that contain user data.
1360 * The root of the gang tree is called the gang leader.
1362 * To perform any operation (read, rewrite, free, claim) on a gang block,
1363 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1364 * in the io_gang_tree field of the original logical i/o by recursively
1365 * reading the gang leader and all gang headers below it. This yields
1366 * an in-core tree containing the contents of every gang header and the
1367 * bps for every constituent of the gang block.
1369 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1370 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1371 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1372 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1373 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1374 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1375 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1376 * of the gang header plus zio_checksum_compute() of the data to update the
1377 * gang header's blk_cksum as described above.
1379 * The two-phase assemble/issue model solves the problem of partial failure --
1380 * what if you'd freed part of a gang block but then couldn't read the
1381 * gang header for another part? Assembling the entire gang tree first
1382 * ensures that all the necessary gang header I/O has succeeded before
1383 * starting the actual work of free, claim, or write. Once the gang tree
1384 * is assembled, free and claim are in-memory operations that cannot fail.
1386 * In the event that a gang write fails, zio_dva_unallocate() walks the
1387 * gang tree to immediately free (i.e. insert back into the space map)
1388 * everything we've allocated. This ensures that we don't get ENOSPC
1389 * errors during repeated suspend/resume cycles due to a flaky device.
1391 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1392 * the gang tree, we won't modify the block, so we can safely defer the free
1393 * (knowing that the block is still intact). If we *can* assemble the gang
1394 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1395 * each constituent bp and we can allocate a new block on the next sync pass.
1397 * In all cases, the gang tree allows complete recovery from partial failure.
1398 * ==========================================================================
1402 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1407 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1408 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1409 &pio->io_bookmark));
1413 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1418 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1419 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1420 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1422 * As we rewrite each gang header, the pipeline will compute
1423 * a new gang block header checksum for it; but no one will
1424 * compute a new data checksum, so we do that here. The one
1425 * exception is the gang leader: the pipeline already computed
1426 * its data checksum because that stage precedes gang assembly.
1427 * (Presently, nothing actually uses interior data checksums;
1428 * this is just good hygiene.)
1430 if (gn != pio->io_gang_leader->io_gang_tree) {
1431 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1432 data, BP_GET_PSIZE(bp));
1435 * If we are here to damage data for testing purposes,
1436 * leave the GBH alone so that we can detect the damage.
1438 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1439 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1441 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1442 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1443 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1451 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1453 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1454 ZIO_GANG_CHILD_FLAGS(pio)));
1459 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1461 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1462 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1465 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1474 static void zio_gang_tree_assemble_done(zio_t *zio);
1476 static zio_gang_node_t *
1477 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1479 zio_gang_node_t *gn;
1481 ASSERT(*gnpp == NULL);
1483 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1484 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1491 zio_gang_node_free(zio_gang_node_t **gnpp)
1493 zio_gang_node_t *gn = *gnpp;
1496 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1497 ASSERT(gn->gn_child[g] == NULL);
1499 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1500 kmem_free(gn, sizeof (*gn));
1505 zio_gang_tree_free(zio_gang_node_t **gnpp)
1507 zio_gang_node_t *gn = *gnpp;
1513 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1514 zio_gang_tree_free(&gn->gn_child[g]);
1516 zio_gang_node_free(gnpp);
1520 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1522 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1524 ASSERT(gio->io_gang_leader == gio);
1525 ASSERT(BP_IS_GANG(bp));
1527 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1528 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1529 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1533 zio_gang_tree_assemble_done(zio_t *zio)
1535 zio_t *gio = zio->io_gang_leader;
1536 zio_gang_node_t *gn = zio->io_private;
1537 blkptr_t *bp = zio->io_bp;
1540 ASSERT(gio == zio_unique_parent(zio));
1541 ASSERT(zio->io_child_count == 0);
1546 if (BP_SHOULD_BYTESWAP(bp))
1547 byteswap_uint64_array(zio->io_data, zio->io_size);
1549 ASSERT(zio->io_data == gn->gn_gbh);
1550 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1551 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1553 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1554 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1555 if (!BP_IS_GANG(gbp))
1557 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1562 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1564 zio_t *gio = pio->io_gang_leader;
1568 ASSERT(BP_IS_GANG(bp) == !!gn);
1569 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1570 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1573 * If you're a gang header, your data is in gn->gn_gbh.
1574 * If you're a gang member, your data is in 'data' and gn == NULL.
1576 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1579 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1581 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1582 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1583 if (BP_IS_HOLE(gbp))
1585 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1586 data = (char *)data + BP_GET_PSIZE(gbp);
1590 if (gn == gio->io_gang_tree)
1591 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1598 zio_gang_assemble(zio_t *zio)
1600 blkptr_t *bp = zio->io_bp;
1602 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1603 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1605 zio->io_gang_leader = zio;
1607 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1609 return (ZIO_PIPELINE_CONTINUE);
1613 zio_gang_issue(zio_t *zio)
1615 blkptr_t *bp = zio->io_bp;
1617 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1618 return (ZIO_PIPELINE_STOP);
1620 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1621 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1623 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1624 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1626 zio_gang_tree_free(&zio->io_gang_tree);
1628 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1630 return (ZIO_PIPELINE_CONTINUE);
1634 zio_write_gang_member_ready(zio_t *zio)
1636 zio_t *pio = zio_unique_parent(zio);
1637 zio_t *gio = zio->io_gang_leader;
1638 dva_t *cdva = zio->io_bp->blk_dva;
1639 dva_t *pdva = pio->io_bp->blk_dva;
1643 if (BP_IS_HOLE(zio->io_bp))
1646 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1648 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1649 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1650 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1651 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1652 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1654 mutex_enter(&pio->io_lock);
1655 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1656 ASSERT(DVA_GET_GANG(&pdva[d]));
1657 asize = DVA_GET_ASIZE(&pdva[d]);
1658 asize += DVA_GET_ASIZE(&cdva[d]);
1659 DVA_SET_ASIZE(&pdva[d], asize);
1661 mutex_exit(&pio->io_lock);
1665 zio_write_gang_block(zio_t *pio)
1667 spa_t *spa = pio->io_spa;
1668 blkptr_t *bp = pio->io_bp;
1669 zio_t *gio = pio->io_gang_leader;
1671 zio_gang_node_t *gn, **gnpp;
1672 zio_gbh_phys_t *gbh;
1673 uint64_t txg = pio->io_txg;
1674 uint64_t resid = pio->io_size;
1676 int copies = gio->io_prop.zp_copies;
1677 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1681 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1682 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1683 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1685 pio->io_error = error;
1686 return (ZIO_PIPELINE_CONTINUE);
1690 gnpp = &gio->io_gang_tree;
1692 gnpp = pio->io_private;
1693 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1696 gn = zio_gang_node_alloc(gnpp);
1698 bzero(gbh, SPA_GANGBLOCKSIZE);
1701 * Create the gang header.
1703 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1704 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1707 * Create and nowait the gang children.
1709 for (g = 0; resid != 0; resid -= lsize, g++) {
1710 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1712 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1714 zp.zp_checksum = gio->io_prop.zp_checksum;
1715 zp.zp_compress = ZIO_COMPRESS_OFF;
1716 zp.zp_type = DMU_OT_NONE;
1718 zp.zp_copies = gio->io_prop.zp_copies;
1720 zp.zp_dedup_verify = 0;
1722 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1723 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1724 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1725 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1726 &pio->io_bookmark));
1730 * Set pio's pipeline to just wait for zio to finish.
1732 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1736 return (ZIO_PIPELINE_CONTINUE);
1740 * ==========================================================================
1742 * ==========================================================================
1745 zio_ddt_child_read_done(zio_t *zio)
1747 blkptr_t *bp = zio->io_bp;
1748 ddt_entry_t *dde = zio->io_private;
1750 zio_t *pio = zio_unique_parent(zio);
1752 mutex_enter(&pio->io_lock);
1753 ddp = ddt_phys_select(dde, bp);
1754 if (zio->io_error == 0)
1755 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1756 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1757 dde->dde_repair_data = zio->io_data;
1759 zio_buf_free(zio->io_data, zio->io_size);
1760 mutex_exit(&pio->io_lock);
1764 zio_ddt_read_start(zio_t *zio)
1766 blkptr_t *bp = zio->io_bp;
1769 ASSERT(BP_GET_DEDUP(bp));
1770 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1771 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1773 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1774 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1775 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1776 ddt_phys_t *ddp = dde->dde_phys;
1777 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1780 ASSERT(zio->io_vsd == NULL);
1783 if (ddp_self == NULL)
1784 return (ZIO_PIPELINE_CONTINUE);
1786 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1787 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1789 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1791 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1792 zio_buf_alloc(zio->io_size), zio->io_size,
1793 zio_ddt_child_read_done, dde, zio->io_priority,
1794 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1795 &zio->io_bookmark));
1797 return (ZIO_PIPELINE_CONTINUE);
1800 zio_nowait(zio_read(zio, zio->io_spa, bp,
1801 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1802 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1804 return (ZIO_PIPELINE_CONTINUE);
1808 zio_ddt_read_done(zio_t *zio)
1810 blkptr_t *bp = zio->io_bp;
1812 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1813 return (ZIO_PIPELINE_STOP);
1815 ASSERT(BP_GET_DEDUP(bp));
1816 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1817 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1819 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1820 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1821 ddt_entry_t *dde = zio->io_vsd;
1823 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1824 return (ZIO_PIPELINE_CONTINUE);
1827 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1828 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1829 return (ZIO_PIPELINE_STOP);
1831 if (dde->dde_repair_data != NULL) {
1832 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1833 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1835 ddt_repair_done(ddt, dde);
1839 ASSERT(zio->io_vsd == NULL);
1841 return (ZIO_PIPELINE_CONTINUE);
1845 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1847 spa_t *spa = zio->io_spa;
1851 * Note: we compare the original data, not the transformed data,
1852 * because when zio->io_bp is an override bp, we will not have
1853 * pushed the I/O transforms. That's an important optimization
1854 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1856 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1857 zio_t *lio = dde->dde_lead_zio[p];
1860 return (lio->io_orig_size != zio->io_orig_size ||
1861 bcmp(zio->io_orig_data, lio->io_orig_data,
1862 zio->io_orig_size) != 0);
1866 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1867 ddt_phys_t *ddp = &dde->dde_phys[p];
1869 if (ddp->ddp_phys_birth != 0) {
1870 arc_buf_t *abuf = NULL;
1871 uint32_t aflags = ARC_WAIT;
1872 blkptr_t blk = *zio->io_bp;
1875 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1879 error = arc_read_nolock(NULL, spa, &blk,
1880 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1881 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1882 &aflags, &zio->io_bookmark);
1885 if (arc_buf_size(abuf) != zio->io_orig_size ||
1886 bcmp(abuf->b_data, zio->io_orig_data,
1887 zio->io_orig_size) != 0)
1889 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1893 return (error != 0);
1901 zio_ddt_child_write_ready(zio_t *zio)
1903 int p = zio->io_prop.zp_copies;
1904 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1905 ddt_entry_t *dde = zio->io_private;
1906 ddt_phys_t *ddp = &dde->dde_phys[p];
1914 ASSERT(dde->dde_lead_zio[p] == zio);
1916 ddt_phys_fill(ddp, zio->io_bp);
1918 while ((pio = zio_walk_parents(zio)) != NULL)
1919 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1925 zio_ddt_child_write_done(zio_t *zio)
1927 int p = zio->io_prop.zp_copies;
1928 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1929 ddt_entry_t *dde = zio->io_private;
1930 ddt_phys_t *ddp = &dde->dde_phys[p];
1934 ASSERT(ddp->ddp_refcnt == 0);
1935 ASSERT(dde->dde_lead_zio[p] == zio);
1936 dde->dde_lead_zio[p] = NULL;
1938 if (zio->io_error == 0) {
1939 while (zio_walk_parents(zio) != NULL)
1940 ddt_phys_addref(ddp);
1942 ddt_phys_clear(ddp);
1949 zio_ddt_ditto_write_done(zio_t *zio)
1951 int p = DDT_PHYS_DITTO;
1952 zio_prop_t *zp = &zio->io_prop;
1953 blkptr_t *bp = zio->io_bp;
1954 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1955 ddt_entry_t *dde = zio->io_private;
1956 ddt_phys_t *ddp = &dde->dde_phys[p];
1957 ddt_key_t *ddk = &dde->dde_key;
1961 ASSERT(ddp->ddp_refcnt == 0);
1962 ASSERT(dde->dde_lead_zio[p] == zio);
1963 dde->dde_lead_zio[p] = NULL;
1965 if (zio->io_error == 0) {
1966 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1967 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1968 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1969 if (ddp->ddp_phys_birth != 0)
1970 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1971 ddt_phys_fill(ddp, bp);
1978 zio_ddt_write(zio_t *zio)
1980 spa_t *spa = zio->io_spa;
1981 blkptr_t *bp = zio->io_bp;
1982 uint64_t txg = zio->io_txg;
1983 zio_prop_t *zp = &zio->io_prop;
1984 int p = zp->zp_copies;
1988 ddt_t *ddt = ddt_select(spa, bp);
1992 ASSERT(BP_GET_DEDUP(bp));
1993 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1994 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1997 dde = ddt_lookup(ddt, bp, B_TRUE);
1998 ddp = &dde->dde_phys[p];
2000 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2002 * If we're using a weak checksum, upgrade to a strong checksum
2003 * and try again. If we're already using a strong checksum,
2004 * we can't resolve it, so just convert to an ordinary write.
2005 * (And automatically e-mail a paper to Nature?)
2007 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2008 zp->zp_checksum = spa_dedup_checksum(spa);
2009 zio_pop_transforms(zio);
2010 zio->io_stage = ZIO_STAGE_OPEN;
2015 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2017 return (ZIO_PIPELINE_CONTINUE);
2020 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2021 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2023 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2024 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2025 zio_prop_t czp = *zp;
2027 czp.zp_copies = ditto_copies;
2030 * If we arrived here with an override bp, we won't have run
2031 * the transform stack, so we won't have the data we need to
2032 * generate a child i/o. So, toss the override bp and restart.
2033 * This is safe, because using the override bp is just an
2034 * optimization; and it's rare, so the cost doesn't matter.
2036 if (zio->io_bp_override) {
2037 zio_pop_transforms(zio);
2038 zio->io_stage = ZIO_STAGE_OPEN;
2039 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2040 zio->io_bp_override = NULL;
2043 return (ZIO_PIPELINE_CONTINUE);
2046 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2047 zio->io_orig_size, &czp, NULL,
2048 zio_ddt_ditto_write_done, dde, zio->io_priority,
2049 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2051 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2052 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2055 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2056 if (ddp->ddp_phys_birth != 0)
2057 ddt_bp_fill(ddp, bp, txg);
2058 if (dde->dde_lead_zio[p] != NULL)
2059 zio_add_child(zio, dde->dde_lead_zio[p]);
2061 ddt_phys_addref(ddp);
2062 } else if (zio->io_bp_override) {
2063 ASSERT(bp->blk_birth == txg);
2064 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2065 ddt_phys_fill(ddp, bp);
2066 ddt_phys_addref(ddp);
2068 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2069 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2070 zio_ddt_child_write_done, dde, zio->io_priority,
2071 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2073 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2074 dde->dde_lead_zio[p] = cio;
2084 return (ZIO_PIPELINE_CONTINUE);
2087 ddt_entry_t *freedde; /* for debugging */
2090 zio_ddt_free(zio_t *zio)
2092 spa_t *spa = zio->io_spa;
2093 blkptr_t *bp = zio->io_bp;
2094 ddt_t *ddt = ddt_select(spa, bp);
2098 ASSERT(BP_GET_DEDUP(bp));
2099 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2102 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2103 ddp = ddt_phys_select(dde, bp);
2104 ddt_phys_decref(ddp);
2107 return (ZIO_PIPELINE_CONTINUE);
2111 * ==========================================================================
2112 * Allocate and free blocks
2113 * ==========================================================================
2116 zio_dva_allocate(zio_t *zio)
2118 spa_t *spa = zio->io_spa;
2119 metaslab_class_t *mc = spa_normal_class(spa);
2120 blkptr_t *bp = zio->io_bp;
2123 if (zio->io_gang_leader == NULL) {
2124 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2125 zio->io_gang_leader = zio;
2128 ASSERT(BP_IS_HOLE(bp));
2129 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2130 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2131 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2132 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2134 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2135 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2138 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2139 return (zio_write_gang_block(zio));
2140 zio->io_error = error;
2143 return (ZIO_PIPELINE_CONTINUE);
2147 zio_dva_free(zio_t *zio)
2149 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2151 return (ZIO_PIPELINE_CONTINUE);
2155 zio_dva_claim(zio_t *zio)
2159 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2161 zio->io_error = error;
2163 return (ZIO_PIPELINE_CONTINUE);
2167 * Undo an allocation. This is used by zio_done() when an I/O fails
2168 * and we want to give back the block we just allocated.
2169 * This handles both normal blocks and gang blocks.
2172 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2176 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2177 ASSERT(zio->io_bp_override == NULL);
2179 if (!BP_IS_HOLE(bp))
2180 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2183 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2184 zio_dva_unallocate(zio, gn->gn_child[g],
2185 &gn->gn_gbh->zg_blkptr[g]);
2191 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2194 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2195 uint64_t size, boolean_t use_slog)
2199 ASSERT(txg > spa_syncing_txg(spa));
2202 error = metaslab_alloc(spa, spa_log_class(spa), size,
2203 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2206 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2207 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2210 BP_SET_LSIZE(new_bp, size);
2211 BP_SET_PSIZE(new_bp, size);
2212 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2213 BP_SET_CHECKSUM(new_bp,
2214 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2215 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2216 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2217 BP_SET_LEVEL(new_bp, 0);
2218 BP_SET_DEDUP(new_bp, 0);
2219 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2226 * Free an intent log block.
2229 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2231 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2232 ASSERT(!BP_IS_GANG(bp));
2234 zio_free(spa, txg, bp);
2238 * ==========================================================================
2239 * Read and write to physical devices
2240 * ==========================================================================
2243 zio_vdev_io_start(zio_t *zio)
2245 vdev_t *vd = zio->io_vd;
2247 spa_t *spa = zio->io_spa;
2249 ASSERT(zio->io_error == 0);
2250 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2253 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2254 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2257 * The mirror_ops handle multiple DVAs in a single BP.
2259 return (vdev_mirror_ops.vdev_op_io_start(zio));
2263 * We keep track of time-sensitive I/Os so that the scan thread
2264 * can quickly react to certain workloads. In particular, we care
2265 * about non-scrubbing, top-level reads and writes with the following
2267 * - synchronous writes of user data to non-slog devices
2268 * - any reads of user data
2269 * When these conditions are met, adjust the timestamp of spa_last_io
2270 * which allows the scan thread to adjust its workload accordingly.
2272 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2273 vd == vd->vdev_top && !vd->vdev_islog &&
2274 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2275 zio->io_txg != spa_syncing_txg(spa)) {
2276 uint64_t old = spa->spa_last_io;
2277 uint64_t new = ddi_get_lbolt64();
2279 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2282 align = 1ULL << vd->vdev_top->vdev_ashift;
2284 if (P2PHASE(zio->io_size, align) != 0) {
2285 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2286 char *abuf = zio_buf_alloc(asize);
2287 ASSERT(vd == vd->vdev_top);
2288 if (zio->io_type == ZIO_TYPE_WRITE) {
2289 bcopy(zio->io_data, abuf, zio->io_size);
2290 bzero(abuf + zio->io_size, asize - zio->io_size);
2292 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2295 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2296 ASSERT(P2PHASE(zio->io_size, align) == 0);
2297 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2300 * If this is a repair I/O, and there's no self-healing involved --
2301 * that is, we're just resilvering what we expect to resilver --
2302 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2303 * This prevents spurious resilvering with nested replication.
2304 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2305 * A is out of date, we'll read from C+D, then use the data to
2306 * resilver A+B -- but we don't actually want to resilver B, just A.
2307 * The top-level mirror has no way to know this, so instead we just
2308 * discard unnecessary repairs as we work our way down the vdev tree.
2309 * The same logic applies to any form of nested replication:
2310 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2312 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2313 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2314 zio->io_txg != 0 && /* not a delegated i/o */
2315 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2316 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2317 zio_vdev_io_bypass(zio);
2318 return (ZIO_PIPELINE_CONTINUE);
2321 if (vd->vdev_ops->vdev_op_leaf &&
2322 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2324 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2325 return (ZIO_PIPELINE_CONTINUE);
2327 if ((zio = vdev_queue_io(zio)) == NULL)
2328 return (ZIO_PIPELINE_STOP);
2330 if (!vdev_accessible(vd, zio)) {
2331 zio->io_error = ENXIO;
2333 return (ZIO_PIPELINE_STOP);
2337 return (vd->vdev_ops->vdev_op_io_start(zio));
2341 zio_vdev_io_done(zio_t *zio)
2343 vdev_t *vd = zio->io_vd;
2344 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2345 boolean_t unexpected_error = B_FALSE;
2347 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2348 return (ZIO_PIPELINE_STOP);
2350 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2352 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2354 vdev_queue_io_done(zio);
2356 if (zio->io_type == ZIO_TYPE_WRITE)
2357 vdev_cache_write(zio);
2359 if (zio_injection_enabled && zio->io_error == 0)
2360 zio->io_error = zio_handle_device_injection(vd,
2363 if (zio_injection_enabled && zio->io_error == 0)
2364 zio->io_error = zio_handle_label_injection(zio, EIO);
2366 if (zio->io_error) {
2367 if (!vdev_accessible(vd, zio)) {
2368 zio->io_error = ENXIO;
2370 unexpected_error = B_TRUE;
2375 ops->vdev_op_io_done(zio);
2377 if (unexpected_error)
2378 VERIFY(vdev_probe(vd, zio) == NULL);
2380 return (ZIO_PIPELINE_CONTINUE);
2384 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2385 * disk, and use that to finish the checksum ereport later.
2388 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2389 const void *good_buf)
2391 /* no processing needed */
2392 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2397 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2399 void *buf = zio_buf_alloc(zio->io_size);
2401 bcopy(zio->io_data, buf, zio->io_size);
2403 zcr->zcr_cbinfo = zio->io_size;
2404 zcr->zcr_cbdata = buf;
2405 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2406 zcr->zcr_free = zio_buf_free;
2410 zio_vdev_io_assess(zio_t *zio)
2412 vdev_t *vd = zio->io_vd;
2414 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2415 return (ZIO_PIPELINE_STOP);
2417 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2418 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2420 if (zio->io_vsd != NULL) {
2421 zio->io_vsd_ops->vsd_free(zio);
2425 if (zio_injection_enabled && zio->io_error == 0)
2426 zio->io_error = zio_handle_fault_injection(zio, EIO);
2429 * If the I/O failed, determine whether we should attempt to retry it.
2431 * On retry, we cut in line in the issue queue, since we don't want
2432 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2434 if (zio->io_error && vd == NULL &&
2435 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2436 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2437 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2439 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2440 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2441 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2442 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2443 zio_requeue_io_start_cut_in_line);
2444 return (ZIO_PIPELINE_STOP);
2448 * If we got an error on a leaf device, convert it to ENXIO
2449 * if the device is not accessible at all.
2451 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2452 !vdev_accessible(vd, zio))
2453 zio->io_error = ENXIO;
2456 * If we can't write to an interior vdev (mirror or RAID-Z),
2457 * set vdev_cant_write so that we stop trying to allocate from it.
2459 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2460 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2461 vd->vdev_cant_write = B_TRUE;
2464 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2466 return (ZIO_PIPELINE_CONTINUE);
2470 zio_vdev_io_reissue(zio_t *zio)
2472 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2473 ASSERT(zio->io_error == 0);
2475 zio->io_stage >>= 1;
2479 zio_vdev_io_redone(zio_t *zio)
2481 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2483 zio->io_stage >>= 1;
2487 zio_vdev_io_bypass(zio_t *zio)
2489 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2490 ASSERT(zio->io_error == 0);
2492 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2493 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2497 * ==========================================================================
2498 * Generate and verify checksums
2499 * ==========================================================================
2502 zio_checksum_generate(zio_t *zio)
2504 blkptr_t *bp = zio->io_bp;
2505 enum zio_checksum checksum;
2509 * This is zio_write_phys().
2510 * We're either generating a label checksum, or none at all.
2512 checksum = zio->io_prop.zp_checksum;
2514 if (checksum == ZIO_CHECKSUM_OFF)
2515 return (ZIO_PIPELINE_CONTINUE);
2517 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2519 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2520 ASSERT(!IO_IS_ALLOCATING(zio));
2521 checksum = ZIO_CHECKSUM_GANG_HEADER;
2523 checksum = BP_GET_CHECKSUM(bp);
2527 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2529 return (ZIO_PIPELINE_CONTINUE);
2533 zio_checksum_verify(zio_t *zio)
2535 zio_bad_cksum_t info;
2536 blkptr_t *bp = zio->io_bp;
2539 ASSERT(zio->io_vd != NULL);
2543 * This is zio_read_phys().
2544 * We're either verifying a label checksum, or nothing at all.
2546 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2547 return (ZIO_PIPELINE_CONTINUE);
2549 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2552 if ((error = zio_checksum_error(zio, &info)) != 0) {
2553 zio->io_error = error;
2554 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2555 zfs_ereport_start_checksum(zio->io_spa,
2556 zio->io_vd, zio, zio->io_offset,
2557 zio->io_size, NULL, &info);
2561 return (ZIO_PIPELINE_CONTINUE);
2565 * Called by RAID-Z to ensure we don't compute the checksum twice.
2568 zio_checksum_verified(zio_t *zio)
2570 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2574 * ==========================================================================
2575 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2576 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2577 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2578 * indicate errors that are specific to one I/O, and most likely permanent.
2579 * Any other error is presumed to be worse because we weren't expecting it.
2580 * ==========================================================================
2583 zio_worst_error(int e1, int e2)
2585 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2588 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2589 if (e1 == zio_error_rank[r1])
2592 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2593 if (e2 == zio_error_rank[r2])
2596 return (r1 > r2 ? e1 : e2);
2600 * ==========================================================================
2602 * ==========================================================================
2605 zio_ready(zio_t *zio)
2607 blkptr_t *bp = zio->io_bp;
2608 zio_t *pio, *pio_next;
2610 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2611 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2612 return (ZIO_PIPELINE_STOP);
2614 if (zio->io_ready) {
2615 ASSERT(IO_IS_ALLOCATING(zio));
2616 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2617 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2622 if (bp != NULL && bp != &zio->io_bp_copy)
2623 zio->io_bp_copy = *bp;
2626 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2628 mutex_enter(&zio->io_lock);
2629 zio->io_state[ZIO_WAIT_READY] = 1;
2630 pio = zio_walk_parents(zio);
2631 mutex_exit(&zio->io_lock);
2634 * As we notify zio's parents, new parents could be added.
2635 * New parents go to the head of zio's io_parent_list, however,
2636 * so we will (correctly) not notify them. The remainder of zio's
2637 * io_parent_list, from 'pio_next' onward, cannot change because
2638 * all parents must wait for us to be done before they can be done.
2640 for (; pio != NULL; pio = pio_next) {
2641 pio_next = zio_walk_parents(zio);
2642 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2645 if (zio->io_flags & ZIO_FLAG_NODATA) {
2646 if (BP_IS_GANG(bp)) {
2647 zio->io_flags &= ~ZIO_FLAG_NODATA;
2649 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2650 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2654 if (zio_injection_enabled &&
2655 zio->io_spa->spa_syncing_txg == zio->io_txg)
2656 zio_handle_ignored_writes(zio);
2658 return (ZIO_PIPELINE_CONTINUE);
2662 zio_done(zio_t *zio)
2664 spa_t *spa = zio->io_spa;
2665 zio_t *lio = zio->io_logical;
2666 blkptr_t *bp = zio->io_bp;
2667 vdev_t *vd = zio->io_vd;
2668 uint64_t psize = zio->io_size;
2669 zio_t *pio, *pio_next;
2673 * If our children haven't all completed,
2674 * wait for them and then repeat this pipeline stage.
2676 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2677 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2678 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2679 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2680 return (ZIO_PIPELINE_STOP);
2682 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2683 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2684 ASSERT(zio->io_children[c][w] == 0);
2687 ASSERT(bp->blk_pad[0] == 0);
2688 ASSERT(bp->blk_pad[1] == 0);
2689 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2690 (bp == zio_unique_parent(zio)->io_bp));
2691 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2692 zio->io_bp_override == NULL &&
2693 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2694 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2695 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2696 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2697 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2702 * If there were child vdev/gang/ddt errors, they apply to us now.
2704 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2705 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2706 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2709 * If the I/O on the transformed data was successful, generate any
2710 * checksum reports now while we still have the transformed data.
2712 if (zio->io_error == 0) {
2713 while (zio->io_cksum_report != NULL) {
2714 zio_cksum_report_t *zcr = zio->io_cksum_report;
2715 uint64_t align = zcr->zcr_align;
2716 uint64_t asize = P2ROUNDUP(psize, align);
2717 char *abuf = zio->io_data;
2719 if (asize != psize) {
2720 abuf = zio_buf_alloc(asize);
2721 bcopy(zio->io_data, abuf, psize);
2722 bzero(abuf + psize, asize - psize);
2725 zio->io_cksum_report = zcr->zcr_next;
2726 zcr->zcr_next = NULL;
2727 zcr->zcr_finish(zcr, abuf);
2728 zfs_ereport_free_checksum(zcr);
2731 zio_buf_free(abuf, asize);
2735 zio_pop_transforms(zio); /* note: may set zio->io_error */
2737 vdev_stat_update(zio, psize);
2739 if (zio->io_error) {
2741 * If this I/O is attached to a particular vdev,
2742 * generate an error message describing the I/O failure
2743 * at the block level. We ignore these errors if the
2744 * device is currently unavailable.
2746 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2747 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2749 if ((zio->io_error == EIO || !(zio->io_flags &
2750 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2753 * For logical I/O requests, tell the SPA to log the
2754 * error and generate a logical data ereport.
2756 spa_log_error(spa, zio);
2757 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2762 if (zio->io_error && zio == lio) {
2764 * Determine whether zio should be reexecuted. This will
2765 * propagate all the way to the root via zio_notify_parent().
2767 ASSERT(vd == NULL && bp != NULL);
2768 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2770 if (IO_IS_ALLOCATING(zio) &&
2771 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2772 if (zio->io_error != ENOSPC)
2773 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2775 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2778 if ((zio->io_type == ZIO_TYPE_READ ||
2779 zio->io_type == ZIO_TYPE_FREE) &&
2780 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2781 zio->io_error == ENXIO &&
2782 spa_load_state(spa) == SPA_LOAD_NONE &&
2783 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2784 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2786 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2787 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2790 * Here is a possibly good place to attempt to do
2791 * either combinatorial reconstruction or error correction
2792 * based on checksums. It also might be a good place
2793 * to send out preliminary ereports before we suspend
2799 * If there were logical child errors, they apply to us now.
2800 * We defer this until now to avoid conflating logical child
2801 * errors with errors that happened to the zio itself when
2802 * updating vdev stats and reporting FMA events above.
2804 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2806 if ((zio->io_error || zio->io_reexecute) &&
2807 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2808 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2809 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2811 zio_gang_tree_free(&zio->io_gang_tree);
2814 * Godfather I/Os should never suspend.
2816 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2817 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2818 zio->io_reexecute = 0;
2820 if (zio->io_reexecute) {
2822 * This is a logical I/O that wants to reexecute.
2824 * Reexecute is top-down. When an i/o fails, if it's not
2825 * the root, it simply notifies its parent and sticks around.
2826 * The parent, seeing that it still has children in zio_done(),
2827 * does the same. This percolates all the way up to the root.
2828 * The root i/o will reexecute or suspend the entire tree.
2830 * This approach ensures that zio_reexecute() honors
2831 * all the original i/o dependency relationships, e.g.
2832 * parents not executing until children are ready.
2834 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2836 zio->io_gang_leader = NULL;
2838 mutex_enter(&zio->io_lock);
2839 zio->io_state[ZIO_WAIT_DONE] = 1;
2840 mutex_exit(&zio->io_lock);
2843 * "The Godfather" I/O monitors its children but is
2844 * not a true parent to them. It will track them through
2845 * the pipeline but severs its ties whenever they get into
2846 * trouble (e.g. suspended). This allows "The Godfather"
2847 * I/O to return status without blocking.
2849 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2850 zio_link_t *zl = zio->io_walk_link;
2851 pio_next = zio_walk_parents(zio);
2853 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2854 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2855 zio_remove_child(pio, zio, zl);
2856 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2860 if ((pio = zio_unique_parent(zio)) != NULL) {
2862 * We're not a root i/o, so there's nothing to do
2863 * but notify our parent. Don't propagate errors
2864 * upward since we haven't permanently failed yet.
2866 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2867 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2868 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2869 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2871 * We'd fail again if we reexecuted now, so suspend
2872 * until conditions improve (e.g. device comes online).
2874 zio_suspend(spa, zio);
2877 * Reexecution is potentially a huge amount of work.
2878 * Hand it off to the otherwise-unused claim taskq.
2880 (void) taskq_dispatch(
2881 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2882 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2884 return (ZIO_PIPELINE_STOP);
2887 ASSERT(zio->io_child_count == 0);
2888 ASSERT(zio->io_reexecute == 0);
2889 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2892 * Report any checksum errors, since the I/O is complete.
2894 while (zio->io_cksum_report != NULL) {
2895 zio_cksum_report_t *zcr = zio->io_cksum_report;
2896 zio->io_cksum_report = zcr->zcr_next;
2897 zcr->zcr_next = NULL;
2898 zcr->zcr_finish(zcr, NULL);
2899 zfs_ereport_free_checksum(zcr);
2903 * It is the responsibility of the done callback to ensure that this
2904 * particular zio is no longer discoverable for adoption, and as
2905 * such, cannot acquire any new parents.
2910 mutex_enter(&zio->io_lock);
2911 zio->io_state[ZIO_WAIT_DONE] = 1;
2912 mutex_exit(&zio->io_lock);
2914 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2915 zio_link_t *zl = zio->io_walk_link;
2916 pio_next = zio_walk_parents(zio);
2917 zio_remove_child(pio, zio, zl);
2918 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2921 if (zio->io_waiter != NULL) {
2922 mutex_enter(&zio->io_lock);
2923 zio->io_executor = NULL;
2924 cv_broadcast(&zio->io_cv);
2925 mutex_exit(&zio->io_lock);
2930 return (ZIO_PIPELINE_STOP);
2934 * ==========================================================================
2935 * I/O pipeline definition
2936 * ==========================================================================
2938 static zio_pipe_stage_t *zio_pipeline[] = {
2944 zio_checksum_generate,
2958 zio_checksum_verify,