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);
371 * Logical I/Os can have logical, gang, or vdev children.
372 * Gang I/Os can have gang or vdev children.
373 * Vdev I/Os can only have vdev children.
374 * The following ASSERT captures all of these constraints.
376 ASSERT(cio->io_child_type <= pio->io_child_type);
381 mutex_enter(&cio->io_lock);
382 mutex_enter(&pio->io_lock);
384 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
386 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
387 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
389 list_insert_head(&pio->io_child_list, zl);
390 list_insert_head(&cio->io_parent_list, zl);
392 pio->io_child_count++;
393 cio->io_parent_count++;
395 mutex_exit(&pio->io_lock);
396 mutex_exit(&cio->io_lock);
400 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
402 ASSERT(zl->zl_parent == pio);
403 ASSERT(zl->zl_child == cio);
405 mutex_enter(&cio->io_lock);
406 mutex_enter(&pio->io_lock);
408 list_remove(&pio->io_child_list, zl);
409 list_remove(&cio->io_parent_list, zl);
411 pio->io_child_count--;
412 cio->io_parent_count--;
414 mutex_exit(&pio->io_lock);
415 mutex_exit(&cio->io_lock);
417 kmem_cache_free(zio_link_cache, zl);
421 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
423 uint64_t *countp = &zio->io_children[child][wait];
424 boolean_t waiting = B_FALSE;
426 mutex_enter(&zio->io_lock);
427 ASSERT(zio->io_stall == NULL);
430 zio->io_stall = countp;
433 mutex_exit(&zio->io_lock);
439 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
441 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
442 int *errorp = &pio->io_child_error[zio->io_child_type];
444 mutex_enter(&pio->io_lock);
445 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
446 *errorp = zio_worst_error(*errorp, zio->io_error);
447 pio->io_reexecute |= zio->io_reexecute;
448 ASSERT3U(*countp, >, 0);
449 if (--*countp == 0 && pio->io_stall == countp) {
450 pio->io_stall = NULL;
451 mutex_exit(&pio->io_lock);
454 mutex_exit(&pio->io_lock);
459 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
461 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
462 zio->io_error = zio->io_child_error[c];
466 * ==========================================================================
467 * Create the various types of I/O (read, write, free, etc)
468 * ==========================================================================
471 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
472 void *data, uint64_t size, zio_done_func_t *done, void *private,
473 zio_type_t type, int priority, enum zio_flag flags,
474 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
475 enum zio_stage stage, enum zio_stage pipeline)
479 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
480 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
481 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
483 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
484 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
485 ASSERT(vd || stage == ZIO_STAGE_OPEN);
487 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
488 bzero(zio, sizeof (zio_t));
490 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
491 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
493 list_create(&zio->io_parent_list, sizeof (zio_link_t),
494 offsetof(zio_link_t, zl_parent_node));
495 list_create(&zio->io_child_list, sizeof (zio_link_t),
496 offsetof(zio_link_t, zl_child_node));
499 zio->io_child_type = ZIO_CHILD_VDEV;
500 else if (flags & ZIO_FLAG_GANG_CHILD)
501 zio->io_child_type = ZIO_CHILD_GANG;
502 else if (flags & ZIO_FLAG_DDT_CHILD)
503 zio->io_child_type = ZIO_CHILD_DDT;
505 zio->io_child_type = ZIO_CHILD_LOGICAL;
508 zio->io_bp = (blkptr_t *)bp;
509 zio->io_bp_copy = *bp;
510 zio->io_bp_orig = *bp;
511 if (type != ZIO_TYPE_WRITE ||
512 zio->io_child_type == ZIO_CHILD_DDT)
513 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
514 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
515 zio->io_logical = zio;
516 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
517 pipeline |= ZIO_GANG_STAGES;
523 zio->io_private = private;
525 zio->io_priority = priority;
527 zio->io_offset = offset;
528 zio->io_orig_data = zio->io_data = data;
529 zio->io_orig_size = zio->io_size = size;
530 zio->io_orig_flags = zio->io_flags = flags;
531 zio->io_orig_stage = zio->io_stage = stage;
532 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
534 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
535 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
538 zio->io_bookmark = *zb;
541 if (zio->io_logical == NULL)
542 zio->io_logical = pio->io_logical;
543 if (zio->io_child_type == ZIO_CHILD_GANG)
544 zio->io_gang_leader = pio->io_gang_leader;
545 zio_add_child(pio, zio);
552 zio_destroy(zio_t *zio)
554 list_destroy(&zio->io_parent_list);
555 list_destroy(&zio->io_child_list);
556 mutex_destroy(&zio->io_lock);
557 cv_destroy(&zio->io_cv);
558 kmem_cache_free(zio_cache, zio);
562 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
563 void *private, enum zio_flag flags)
567 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
568 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
569 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
575 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
577 return (zio_null(NULL, spa, NULL, done, private, flags));
581 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
582 void *data, uint64_t size, zio_done_func_t *done, void *private,
583 int priority, enum zio_flag flags, const zbookmark_t *zb)
587 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
588 data, size, done, private,
589 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
590 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
591 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
597 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
598 void *data, uint64_t size, const zio_prop_t *zp,
599 zio_done_func_t *ready, zio_done_func_t *done, void *private,
600 int priority, enum zio_flag flags, const zbookmark_t *zb)
604 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
605 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
606 zp->zp_compress >= ZIO_COMPRESS_OFF &&
607 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
608 zp->zp_type < DMU_OT_NUMTYPES &&
611 zp->zp_copies <= spa_max_replication(spa) &&
613 zp->zp_dedup_verify <= 1);
615 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
616 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
617 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
618 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
620 zio->io_ready = ready;
627 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
628 uint64_t size, zio_done_func_t *done, void *private, int priority,
629 enum zio_flag flags, zbookmark_t *zb)
633 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
634 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
635 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
641 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
643 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
644 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
645 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
646 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
648 zio->io_prop.zp_copies = copies;
649 zio->io_bp_override = bp;
653 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
655 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
659 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
664 dprintf_bp(bp, "freeing in txg %llu, pass %u",
665 (longlong_t)txg, spa->spa_sync_pass);
667 ASSERT(!BP_IS_HOLE(bp));
668 ASSERT(spa_syncing_txg(spa) == txg);
669 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
671 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
672 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
673 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
679 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
680 zio_done_func_t *done, void *private, enum zio_flag flags)
685 * A claim is an allocation of a specific block. Claims are needed
686 * to support immediate writes in the intent log. The issue is that
687 * immediate writes contain committed data, but in a txg that was
688 * *not* committed. Upon opening the pool after an unclean shutdown,
689 * the intent log claims all blocks that contain immediate write data
690 * so that the SPA knows they're in use.
692 * All claims *must* be resolved in the first txg -- before the SPA
693 * starts allocating blocks -- so that nothing is allocated twice.
694 * If txg == 0 we just verify that the block is claimable.
696 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
697 ASSERT(txg == spa_first_txg(spa) || txg == 0);
698 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
700 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
701 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
702 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
708 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
709 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
714 if (vd->vdev_children == 0) {
715 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
716 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
717 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
721 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
723 for (c = 0; c < vd->vdev_children; c++)
724 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
725 done, private, priority, flags));
732 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
733 void *data, int checksum, zio_done_func_t *done, void *private,
734 int priority, enum zio_flag flags, boolean_t labels)
738 ASSERT(vd->vdev_children == 0);
739 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
740 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
741 ASSERT3U(offset + size, <=, vd->vdev_psize);
743 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
744 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
745 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
747 zio->io_prop.zp_checksum = checksum;
753 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
754 void *data, int checksum, zio_done_func_t *done, void *private,
755 int priority, enum zio_flag flags, boolean_t labels)
759 ASSERT(vd->vdev_children == 0);
760 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
761 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
762 ASSERT3U(offset + size, <=, vd->vdev_psize);
764 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
765 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
766 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
768 zio->io_prop.zp_checksum = checksum;
770 if (zio_checksum_table[checksum].ci_eck) {
772 * zec checksums are necessarily destructive -- they modify
773 * the end of the write buffer to hold the verifier/checksum.
774 * Therefore, we must make a local copy in case the data is
775 * being written to multiple places in parallel.
777 void *wbuf = zio_buf_alloc(size);
778 bcopy(data, wbuf, size);
779 zio_push_transform(zio, wbuf, size, size, NULL);
786 * Create a child I/O to do some work for us.
789 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
790 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
791 zio_done_func_t *done, void *private)
793 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
796 ASSERT(vd->vdev_parent ==
797 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
799 if (type == ZIO_TYPE_READ && bp != NULL) {
801 * If we have the bp, then the child should perform the
802 * checksum and the parent need not. This pushes error
803 * detection as close to the leaves as possible and
804 * eliminates redundant checksums in the interior nodes.
806 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
807 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
810 if (vd->vdev_children == 0)
811 offset += VDEV_LABEL_START_SIZE;
813 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
816 * If we've decided to do a repair, the write is not speculative --
817 * even if the original read was.
819 if (flags & ZIO_FLAG_IO_REPAIR)
820 flags &= ~ZIO_FLAG_SPECULATIVE;
822 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
823 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
824 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
830 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
831 int type, int priority, enum zio_flag flags,
832 zio_done_func_t *done, void *private)
836 ASSERT(vd->vdev_ops->vdev_op_leaf);
838 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
839 data, size, done, private, type, priority,
840 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
842 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
848 zio_flush(zio_t *zio, vdev_t *vd)
850 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
851 NULL, NULL, ZIO_PRIORITY_NOW,
852 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
856 zio_shrink(zio_t *zio, uint64_t size)
858 ASSERT(zio->io_executor == NULL);
859 ASSERT(zio->io_orig_size == zio->io_size);
860 ASSERT(size <= zio->io_size);
863 * We don't shrink for raidz because of problems with the
864 * reconstruction when reading back less than the block size.
865 * Note, BP_IS_RAIDZ() assumes no compression.
867 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
868 if (!BP_IS_RAIDZ(zio->io_bp))
869 zio->io_orig_size = zio->io_size = size;
873 * ==========================================================================
874 * Prepare to read and write logical blocks
875 * ==========================================================================
879 zio_read_bp_init(zio_t *zio)
881 blkptr_t *bp = zio->io_bp;
883 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
884 zio->io_child_type == ZIO_CHILD_LOGICAL &&
885 !(zio->io_flags & ZIO_FLAG_RAW)) {
886 uint64_t psize = BP_GET_PSIZE(bp);
887 void *cbuf = zio_buf_alloc(psize);
889 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
892 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
893 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
895 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
896 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
898 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
899 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
901 return (ZIO_PIPELINE_CONTINUE);
905 zio_write_bp_init(zio_t *zio)
907 spa_t *spa = zio->io_spa;
908 zio_prop_t *zp = &zio->io_prop;
909 enum zio_compress compress = zp->zp_compress;
910 blkptr_t *bp = zio->io_bp;
911 uint64_t lsize = zio->io_size;
912 uint64_t psize = lsize;
916 * If our children haven't all reached the ready stage,
917 * wait for them and then repeat this pipeline stage.
919 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
920 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
921 return (ZIO_PIPELINE_STOP);
923 if (!IO_IS_ALLOCATING(zio))
924 return (ZIO_PIPELINE_CONTINUE);
926 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
928 if (zio->io_bp_override) {
929 ASSERT(bp->blk_birth != zio->io_txg);
930 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
932 *bp = *zio->io_bp_override;
933 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
935 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
936 return (ZIO_PIPELINE_CONTINUE);
938 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
939 zp->zp_dedup_verify);
941 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
943 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
944 return (ZIO_PIPELINE_CONTINUE);
946 zio->io_bp_override = NULL;
950 if (bp->blk_birth == zio->io_txg) {
952 * We're rewriting an existing block, which means we're
953 * working on behalf of spa_sync(). For spa_sync() to
954 * converge, it must eventually be the case that we don't
955 * have to allocate new blocks. But compression changes
956 * the blocksize, which forces a reallocate, and makes
957 * convergence take longer. Therefore, after the first
958 * few passes, stop compressing to ensure convergence.
960 pass = spa_sync_pass(spa);
962 ASSERT(zio->io_txg == spa_syncing_txg(spa));
963 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
964 ASSERT(!BP_GET_DEDUP(bp));
966 if (pass > SYNC_PASS_DONT_COMPRESS)
967 compress = ZIO_COMPRESS_OFF;
969 /* Make sure someone doesn't change their mind on overwrites */
970 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
971 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
974 if (compress != ZIO_COMPRESS_OFF) {
975 void *cbuf = zio_buf_alloc(lsize);
976 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
977 if (psize == 0 || psize == lsize) {
978 compress = ZIO_COMPRESS_OFF;
979 zio_buf_free(cbuf, lsize);
981 ASSERT(psize < lsize);
982 zio_push_transform(zio, cbuf, psize, lsize, NULL);
987 * The final pass of spa_sync() must be all rewrites, but the first
988 * few passes offer a trade-off: allocating blocks defers convergence,
989 * but newly allocated blocks are sequential, so they can be written
990 * to disk faster. Therefore, we allow the first few passes of
991 * spa_sync() to allocate new blocks, but force rewrites after that.
992 * There should only be a handful of blocks after pass 1 in any case.
994 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
995 pass > SYNC_PASS_REWRITE) {
997 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
998 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
999 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1002 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1006 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1008 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1009 BP_SET_LSIZE(bp, lsize);
1010 BP_SET_PSIZE(bp, psize);
1011 BP_SET_COMPRESS(bp, compress);
1012 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1013 BP_SET_TYPE(bp, zp->zp_type);
1014 BP_SET_LEVEL(bp, zp->zp_level);
1015 BP_SET_DEDUP(bp, zp->zp_dedup);
1016 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1018 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1019 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1020 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1024 return (ZIO_PIPELINE_CONTINUE);
1028 zio_free_bp_init(zio_t *zio)
1030 blkptr_t *bp = zio->io_bp;
1032 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1033 if (BP_GET_DEDUP(bp))
1034 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1037 return (ZIO_PIPELINE_CONTINUE);
1041 * ==========================================================================
1042 * Execute the I/O pipeline
1043 * ==========================================================================
1047 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1049 spa_t *spa = zio->io_spa;
1050 zio_type_t t = zio->io_type;
1051 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1054 * If we're a config writer or a probe, the normal issue and
1055 * interrupt threads may all be blocked waiting for the config lock.
1056 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1058 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1062 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1064 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1068 * If this is a high priority I/O, then use the high priority taskq.
1070 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1071 spa->spa_zio_taskq[t][q + 1] != NULL)
1074 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1075 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1076 (task_func_t *)zio_execute, zio, flags);
1080 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1082 kthread_t *executor = zio->io_executor;
1083 spa_t *spa = zio->io_spa;
1085 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1086 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1093 zio_issue_async(zio_t *zio)
1095 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1097 return (ZIO_PIPELINE_STOP);
1101 zio_interrupt(zio_t *zio)
1103 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1107 * Execute the I/O pipeline until one of the following occurs:
1108 * (1) the I/O completes; (2) the pipeline stalls waiting for
1109 * dependent child I/Os; (3) the I/O issues, so we're waiting
1110 * for an I/O completion interrupt; (4) the I/O is delegated by
1111 * vdev-level caching or aggregation; (5) the I/O is deferred
1112 * due to vdev-level queueing; (6) the I/O is handed off to
1113 * another thread. In all cases, the pipeline stops whenever
1114 * there's no CPU work; it never burns a thread in cv_wait().
1116 * There's no locking on io_stage because there's no legitimate way
1117 * for multiple threads to be attempting to process the same I/O.
1119 static zio_pipe_stage_t *zio_pipeline[];
1122 zio_execute(zio_t *zio)
1124 zio->io_executor = curthread;
1126 while (zio->io_stage < ZIO_STAGE_DONE) {
1127 enum zio_stage pipeline = zio->io_pipeline;
1128 enum zio_stage stage = zio->io_stage;
1131 ASSERT(!MUTEX_HELD(&zio->io_lock));
1132 ASSERT(ISP2(stage));
1133 ASSERT(zio->io_stall == NULL);
1137 } while ((stage & pipeline) == 0);
1139 ASSERT(stage <= ZIO_STAGE_DONE);
1142 * If we are in interrupt context and this pipeline stage
1143 * will grab a config lock that is held across I/O,
1144 * or may wait for an I/O that needs an interrupt thread
1145 * to complete, issue async to avoid deadlock.
1147 * For VDEV_IO_START, we cut in line so that the io will
1148 * be sent to disk promptly.
1150 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1151 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1152 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1153 zio_requeue_io_start_cut_in_line : B_FALSE;
1154 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1158 zio->io_stage = stage;
1159 rv = zio_pipeline[highbit(stage) - 1](zio);
1161 if (rv == ZIO_PIPELINE_STOP)
1164 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1169 * ==========================================================================
1170 * Initiate I/O, either sync or async
1171 * ==========================================================================
1174 zio_wait(zio_t *zio)
1178 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1179 ASSERT(zio->io_executor == NULL);
1181 zio->io_waiter = curthread;
1185 mutex_enter(&zio->io_lock);
1186 while (zio->io_executor != NULL)
1187 cv_wait(&zio->io_cv, &zio->io_lock);
1188 mutex_exit(&zio->io_lock);
1190 error = zio->io_error;
1197 zio_nowait(zio_t *zio)
1199 ASSERT(zio->io_executor == NULL);
1201 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1202 zio_unique_parent(zio) == NULL) {
1204 * This is a logical async I/O with no parent to wait for it.
1205 * We add it to the spa_async_root_zio "Godfather" I/O which
1206 * will ensure they complete prior to unloading the pool.
1208 spa_t *spa = zio->io_spa;
1210 zio_add_child(spa->spa_async_zio_root, zio);
1217 * ==========================================================================
1218 * Reexecute or suspend/resume failed I/O
1219 * ==========================================================================
1223 zio_reexecute(zio_t *pio)
1225 zio_t *cio, *cio_next;
1227 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1228 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1229 ASSERT(pio->io_gang_leader == NULL);
1230 ASSERT(pio->io_gang_tree == NULL);
1232 pio->io_flags = pio->io_orig_flags;
1233 pio->io_stage = pio->io_orig_stage;
1234 pio->io_pipeline = pio->io_orig_pipeline;
1235 pio->io_reexecute = 0;
1237 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1238 pio->io_state[w] = 0;
1239 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1240 pio->io_child_error[c] = 0;
1242 if (IO_IS_ALLOCATING(pio))
1243 BP_ZERO(pio->io_bp);
1246 * As we reexecute pio's children, new children could be created.
1247 * New children go to the head of pio's io_child_list, however,
1248 * so we will (correctly) not reexecute them. The key is that
1249 * the remainder of pio's io_child_list, from 'cio_next' onward,
1250 * cannot be affected by any side effects of reexecuting 'cio'.
1252 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1253 cio_next = zio_walk_children(pio);
1254 mutex_enter(&pio->io_lock);
1255 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1256 pio->io_children[cio->io_child_type][w]++;
1257 mutex_exit(&pio->io_lock);
1262 * Now that all children have been reexecuted, execute the parent.
1263 * We don't reexecute "The Godfather" I/O here as it's the
1264 * responsibility of the caller to wait on him.
1266 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1271 zio_suspend(spa_t *spa, zio_t *zio)
1273 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1274 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1275 "failure and the failure mode property for this pool "
1276 "is set to panic.", spa_name(spa));
1278 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1280 mutex_enter(&spa->spa_suspend_lock);
1282 if (spa->spa_suspend_zio_root == NULL)
1283 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1284 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1285 ZIO_FLAG_GODFATHER);
1287 spa->spa_suspended = B_TRUE;
1290 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1291 ASSERT(zio != spa->spa_suspend_zio_root);
1292 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1293 ASSERT(zio_unique_parent(zio) == NULL);
1294 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1295 zio_add_child(spa->spa_suspend_zio_root, zio);
1298 mutex_exit(&spa->spa_suspend_lock);
1302 zio_resume(spa_t *spa)
1307 * Reexecute all previously suspended i/o.
1309 mutex_enter(&spa->spa_suspend_lock);
1310 spa->spa_suspended = B_FALSE;
1311 cv_broadcast(&spa->spa_suspend_cv);
1312 pio = spa->spa_suspend_zio_root;
1313 spa->spa_suspend_zio_root = NULL;
1314 mutex_exit(&spa->spa_suspend_lock);
1320 return (zio_wait(pio));
1324 zio_resume_wait(spa_t *spa)
1326 mutex_enter(&spa->spa_suspend_lock);
1327 while (spa_suspended(spa))
1328 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1329 mutex_exit(&spa->spa_suspend_lock);
1333 * ==========================================================================
1336 * A gang block is a collection of small blocks that looks to the DMU
1337 * like one large block. When zio_dva_allocate() cannot find a block
1338 * of the requested size, due to either severe fragmentation or the pool
1339 * being nearly full, it calls zio_write_gang_block() to construct the
1340 * block from smaller fragments.
1342 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1343 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1344 * an indirect block: it's an array of block pointers. It consumes
1345 * only one sector and hence is allocatable regardless of fragmentation.
1346 * The gang header's bps point to its gang members, which hold the data.
1348 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1349 * as the verifier to ensure uniqueness of the SHA256 checksum.
1350 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1351 * not the gang header. This ensures that data block signatures (needed for
1352 * deduplication) are independent of how the block is physically stored.
1354 * Gang blocks can be nested: a gang member may itself be a gang block.
1355 * Thus every gang block is a tree in which root and all interior nodes are
1356 * gang headers, and the leaves are normal blocks that contain user data.
1357 * The root of the gang tree is called the gang leader.
1359 * To perform any operation (read, rewrite, free, claim) on a gang block,
1360 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1361 * in the io_gang_tree field of the original logical i/o by recursively
1362 * reading the gang leader and all gang headers below it. This yields
1363 * an in-core tree containing the contents of every gang header and the
1364 * bps for every constituent of the gang block.
1366 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1367 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1368 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1369 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1370 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1371 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1372 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1373 * of the gang header plus zio_checksum_compute() of the data to update the
1374 * gang header's blk_cksum as described above.
1376 * The two-phase assemble/issue model solves the problem of partial failure --
1377 * what if you'd freed part of a gang block but then couldn't read the
1378 * gang header for another part? Assembling the entire gang tree first
1379 * ensures that all the necessary gang header I/O has succeeded before
1380 * starting the actual work of free, claim, or write. Once the gang tree
1381 * is assembled, free and claim are in-memory operations that cannot fail.
1383 * In the event that a gang write fails, zio_dva_unallocate() walks the
1384 * gang tree to immediately free (i.e. insert back into the space map)
1385 * everything we've allocated. This ensures that we don't get ENOSPC
1386 * errors during repeated suspend/resume cycles due to a flaky device.
1388 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1389 * the gang tree, we won't modify the block, so we can safely defer the free
1390 * (knowing that the block is still intact). If we *can* assemble the gang
1391 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1392 * each constituent bp and we can allocate a new block on the next sync pass.
1394 * In all cases, the gang tree allows complete recovery from partial failure.
1395 * ==========================================================================
1399 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1404 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1405 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1406 &pio->io_bookmark));
1410 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1415 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1416 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1417 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1419 * As we rewrite each gang header, the pipeline will compute
1420 * a new gang block header checksum for it; but no one will
1421 * compute a new data checksum, so we do that here. The one
1422 * exception is the gang leader: the pipeline already computed
1423 * its data checksum because that stage precedes gang assembly.
1424 * (Presently, nothing actually uses interior data checksums;
1425 * this is just good hygiene.)
1427 if (gn != pio->io_gang_leader->io_gang_tree) {
1428 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1429 data, BP_GET_PSIZE(bp));
1432 * If we are here to damage data for testing purposes,
1433 * leave the GBH alone so that we can detect the damage.
1435 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1436 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1438 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1439 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1440 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1448 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1450 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1451 ZIO_GANG_CHILD_FLAGS(pio)));
1456 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1458 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1459 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1462 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1471 static void zio_gang_tree_assemble_done(zio_t *zio);
1473 static zio_gang_node_t *
1474 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1476 zio_gang_node_t *gn;
1478 ASSERT(*gnpp == NULL);
1480 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1481 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1488 zio_gang_node_free(zio_gang_node_t **gnpp)
1490 zio_gang_node_t *gn = *gnpp;
1492 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1493 ASSERT(gn->gn_child[g] == NULL);
1495 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1496 kmem_free(gn, sizeof (*gn));
1501 zio_gang_tree_free(zio_gang_node_t **gnpp)
1503 zio_gang_node_t *gn = *gnpp;
1508 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1509 zio_gang_tree_free(&gn->gn_child[g]);
1511 zio_gang_node_free(gnpp);
1515 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1517 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1519 ASSERT(gio->io_gang_leader == gio);
1520 ASSERT(BP_IS_GANG(bp));
1522 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1523 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1524 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1528 zio_gang_tree_assemble_done(zio_t *zio)
1530 zio_t *gio = zio->io_gang_leader;
1531 zio_gang_node_t *gn = zio->io_private;
1532 blkptr_t *bp = zio->io_bp;
1534 ASSERT(gio == zio_unique_parent(zio));
1535 ASSERT(zio->io_child_count == 0);
1540 if (BP_SHOULD_BYTESWAP(bp))
1541 byteswap_uint64_array(zio->io_data, zio->io_size);
1543 ASSERT(zio->io_data == gn->gn_gbh);
1544 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1545 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1547 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1548 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1549 if (!BP_IS_GANG(gbp))
1551 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1556 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1558 zio_t *gio = pio->io_gang_leader;
1561 ASSERT(BP_IS_GANG(bp) == !!gn);
1562 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1563 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1566 * If you're a gang header, your data is in gn->gn_gbh.
1567 * If you're a gang member, your data is in 'data' and gn == NULL.
1569 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1572 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1574 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1575 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1576 if (BP_IS_HOLE(gbp))
1578 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1579 data = (char *)data + BP_GET_PSIZE(gbp);
1583 if (gn == gio->io_gang_tree)
1584 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1591 zio_gang_assemble(zio_t *zio)
1593 blkptr_t *bp = zio->io_bp;
1595 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1596 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1598 zio->io_gang_leader = zio;
1600 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1602 return (ZIO_PIPELINE_CONTINUE);
1606 zio_gang_issue(zio_t *zio)
1608 blkptr_t *bp = zio->io_bp;
1610 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1611 return (ZIO_PIPELINE_STOP);
1613 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1614 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1616 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1617 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1619 zio_gang_tree_free(&zio->io_gang_tree);
1621 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1623 return (ZIO_PIPELINE_CONTINUE);
1627 zio_write_gang_member_ready(zio_t *zio)
1629 zio_t *pio = zio_unique_parent(zio);
1630 zio_t *gio = zio->io_gang_leader;
1631 dva_t *cdva = zio->io_bp->blk_dva;
1632 dva_t *pdva = pio->io_bp->blk_dva;
1635 if (BP_IS_HOLE(zio->io_bp))
1638 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1640 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1641 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1642 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1643 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1644 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1646 mutex_enter(&pio->io_lock);
1647 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1648 ASSERT(DVA_GET_GANG(&pdva[d]));
1649 asize = DVA_GET_ASIZE(&pdva[d]);
1650 asize += DVA_GET_ASIZE(&cdva[d]);
1651 DVA_SET_ASIZE(&pdva[d], asize);
1653 mutex_exit(&pio->io_lock);
1657 zio_write_gang_block(zio_t *pio)
1659 spa_t *spa = pio->io_spa;
1660 blkptr_t *bp = pio->io_bp;
1661 zio_t *gio = pio->io_gang_leader;
1663 zio_gang_node_t *gn, **gnpp;
1664 zio_gbh_phys_t *gbh;
1665 uint64_t txg = pio->io_txg;
1666 uint64_t resid = pio->io_size;
1668 int copies = gio->io_prop.zp_copies;
1669 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1673 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1674 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1675 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1677 pio->io_error = error;
1678 return (ZIO_PIPELINE_CONTINUE);
1682 gnpp = &gio->io_gang_tree;
1684 gnpp = pio->io_private;
1685 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1688 gn = zio_gang_node_alloc(gnpp);
1690 bzero(gbh, SPA_GANGBLOCKSIZE);
1693 * Create the gang header.
1695 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1696 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1699 * Create and nowait the gang children.
1701 for (int g = 0; resid != 0; resid -= lsize, g++) {
1702 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1704 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1706 zp.zp_checksum = gio->io_prop.zp_checksum;
1707 zp.zp_compress = ZIO_COMPRESS_OFF;
1708 zp.zp_type = DMU_OT_NONE;
1710 zp.zp_copies = gio->io_prop.zp_copies;
1712 zp.zp_dedup_verify = 0;
1714 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1715 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1716 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1717 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1718 &pio->io_bookmark));
1722 * Set pio's pipeline to just wait for zio to finish.
1724 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1728 return (ZIO_PIPELINE_CONTINUE);
1732 * ==========================================================================
1734 * ==========================================================================
1737 zio_ddt_child_read_done(zio_t *zio)
1739 blkptr_t *bp = zio->io_bp;
1740 ddt_entry_t *dde = zio->io_private;
1742 zio_t *pio = zio_unique_parent(zio);
1744 mutex_enter(&pio->io_lock);
1745 ddp = ddt_phys_select(dde, bp);
1746 if (zio->io_error == 0)
1747 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1748 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1749 dde->dde_repair_data = zio->io_data;
1751 zio_buf_free(zio->io_data, zio->io_size);
1752 mutex_exit(&pio->io_lock);
1756 zio_ddt_read_start(zio_t *zio)
1758 blkptr_t *bp = zio->io_bp;
1760 ASSERT(BP_GET_DEDUP(bp));
1761 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1762 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1764 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1765 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1766 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1767 ddt_phys_t *ddp = dde->dde_phys;
1768 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1771 ASSERT(zio->io_vsd == NULL);
1774 if (ddp_self == NULL)
1775 return (ZIO_PIPELINE_CONTINUE);
1777 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1778 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1780 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1782 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1783 zio_buf_alloc(zio->io_size), zio->io_size,
1784 zio_ddt_child_read_done, dde, zio->io_priority,
1785 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1786 &zio->io_bookmark));
1788 return (ZIO_PIPELINE_CONTINUE);
1791 zio_nowait(zio_read(zio, zio->io_spa, bp,
1792 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1793 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1795 return (ZIO_PIPELINE_CONTINUE);
1799 zio_ddt_read_done(zio_t *zio)
1801 blkptr_t *bp = zio->io_bp;
1803 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1804 return (ZIO_PIPELINE_STOP);
1806 ASSERT(BP_GET_DEDUP(bp));
1807 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1808 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1810 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1811 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1812 ddt_entry_t *dde = zio->io_vsd;
1814 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1815 return (ZIO_PIPELINE_CONTINUE);
1818 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1819 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1820 return (ZIO_PIPELINE_STOP);
1822 if (dde->dde_repair_data != NULL) {
1823 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1824 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1826 ddt_repair_done(ddt, dde);
1830 ASSERT(zio->io_vsd == NULL);
1832 return (ZIO_PIPELINE_CONTINUE);
1836 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1838 spa_t *spa = zio->io_spa;
1841 * Note: we compare the original data, not the transformed data,
1842 * because when zio->io_bp is an override bp, we will not have
1843 * pushed the I/O transforms. That's an important optimization
1844 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1846 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1847 zio_t *lio = dde->dde_lead_zio[p];
1850 return (lio->io_orig_size != zio->io_orig_size ||
1851 bcmp(zio->io_orig_data, lio->io_orig_data,
1852 zio->io_orig_size) != 0);
1856 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1857 ddt_phys_t *ddp = &dde->dde_phys[p];
1859 if (ddp->ddp_phys_birth != 0) {
1860 arc_buf_t *abuf = NULL;
1861 uint32_t aflags = ARC_WAIT;
1862 blkptr_t blk = *zio->io_bp;
1865 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1869 error = arc_read_nolock(NULL, spa, &blk,
1870 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1871 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1872 &aflags, &zio->io_bookmark);
1875 if (arc_buf_size(abuf) != zio->io_orig_size ||
1876 bcmp(abuf->b_data, zio->io_orig_data,
1877 zio->io_orig_size) != 0)
1879 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1883 return (error != 0);
1891 zio_ddt_child_write_ready(zio_t *zio)
1893 int p = zio->io_prop.zp_copies;
1894 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1895 ddt_entry_t *dde = zio->io_private;
1896 ddt_phys_t *ddp = &dde->dde_phys[p];
1904 ASSERT(dde->dde_lead_zio[p] == zio);
1906 ddt_phys_fill(ddp, zio->io_bp);
1908 while ((pio = zio_walk_parents(zio)) != NULL)
1909 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1915 zio_ddt_child_write_done(zio_t *zio)
1917 int p = zio->io_prop.zp_copies;
1918 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1919 ddt_entry_t *dde = zio->io_private;
1920 ddt_phys_t *ddp = &dde->dde_phys[p];
1924 ASSERT(ddp->ddp_refcnt == 0);
1925 ASSERT(dde->dde_lead_zio[p] == zio);
1926 dde->dde_lead_zio[p] = NULL;
1928 if (zio->io_error == 0) {
1929 while (zio_walk_parents(zio) != NULL)
1930 ddt_phys_addref(ddp);
1932 ddt_phys_clear(ddp);
1939 zio_ddt_ditto_write_done(zio_t *zio)
1941 int p = DDT_PHYS_DITTO;
1942 zio_prop_t *zp = &zio->io_prop;
1943 blkptr_t *bp = zio->io_bp;
1944 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1945 ddt_entry_t *dde = zio->io_private;
1946 ddt_phys_t *ddp = &dde->dde_phys[p];
1947 ddt_key_t *ddk = &dde->dde_key;
1951 ASSERT(ddp->ddp_refcnt == 0);
1952 ASSERT(dde->dde_lead_zio[p] == zio);
1953 dde->dde_lead_zio[p] = NULL;
1955 if (zio->io_error == 0) {
1956 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1957 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1958 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1959 if (ddp->ddp_phys_birth != 0)
1960 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1961 ddt_phys_fill(ddp, bp);
1968 zio_ddt_write(zio_t *zio)
1970 spa_t *spa = zio->io_spa;
1971 blkptr_t *bp = zio->io_bp;
1972 uint64_t txg = zio->io_txg;
1973 zio_prop_t *zp = &zio->io_prop;
1974 int p = zp->zp_copies;
1978 ddt_t *ddt = ddt_select(spa, bp);
1982 ASSERT(BP_GET_DEDUP(bp));
1983 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
1984 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
1987 dde = ddt_lookup(ddt, bp, B_TRUE);
1988 ddp = &dde->dde_phys[p];
1990 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
1992 * If we're using a weak checksum, upgrade to a strong checksum
1993 * and try again. If we're already using a strong checksum,
1994 * we can't resolve it, so just convert to an ordinary write.
1995 * (And automatically e-mail a paper to Nature?)
1997 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
1998 zp->zp_checksum = spa_dedup_checksum(spa);
1999 zio_pop_transforms(zio);
2000 zio->io_stage = ZIO_STAGE_OPEN;
2005 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2007 return (ZIO_PIPELINE_CONTINUE);
2010 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2011 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2013 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2014 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2015 zio_prop_t czp = *zp;
2017 czp.zp_copies = ditto_copies;
2020 * If we arrived here with an override bp, we won't have run
2021 * the transform stack, so we won't have the data we need to
2022 * generate a child i/o. So, toss the override bp and restart.
2023 * This is safe, because using the override bp is just an
2024 * optimization; and it's rare, so the cost doesn't matter.
2026 if (zio->io_bp_override) {
2027 zio_pop_transforms(zio);
2028 zio->io_stage = ZIO_STAGE_OPEN;
2029 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2030 zio->io_bp_override = NULL;
2033 return (ZIO_PIPELINE_CONTINUE);
2036 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2037 zio->io_orig_size, &czp, NULL,
2038 zio_ddt_ditto_write_done, dde, zio->io_priority,
2039 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2041 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2042 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2045 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2046 if (ddp->ddp_phys_birth != 0)
2047 ddt_bp_fill(ddp, bp, txg);
2048 if (dde->dde_lead_zio[p] != NULL)
2049 zio_add_child(zio, dde->dde_lead_zio[p]);
2051 ddt_phys_addref(ddp);
2052 } else if (zio->io_bp_override) {
2053 ASSERT(bp->blk_birth == txg);
2054 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2055 ddt_phys_fill(ddp, bp);
2056 ddt_phys_addref(ddp);
2058 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2059 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2060 zio_ddt_child_write_done, dde, zio->io_priority,
2061 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2063 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2064 dde->dde_lead_zio[p] = cio;
2074 return (ZIO_PIPELINE_CONTINUE);
2077 ddt_entry_t *freedde; /* for debugging */
2080 zio_ddt_free(zio_t *zio)
2082 spa_t *spa = zio->io_spa;
2083 blkptr_t *bp = zio->io_bp;
2084 ddt_t *ddt = ddt_select(spa, bp);
2088 ASSERT(BP_GET_DEDUP(bp));
2089 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2092 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2093 ddp = ddt_phys_select(dde, bp);
2094 ddt_phys_decref(ddp);
2097 return (ZIO_PIPELINE_CONTINUE);
2101 * ==========================================================================
2102 * Allocate and free blocks
2103 * ==========================================================================
2106 zio_dva_allocate(zio_t *zio)
2108 spa_t *spa = zio->io_spa;
2109 metaslab_class_t *mc = spa_normal_class(spa);
2110 blkptr_t *bp = zio->io_bp;
2113 if (zio->io_gang_leader == NULL) {
2114 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2115 zio->io_gang_leader = zio;
2118 ASSERT(BP_IS_HOLE(bp));
2119 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2120 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2121 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2122 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2124 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2125 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2128 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2129 return (zio_write_gang_block(zio));
2130 zio->io_error = error;
2133 return (ZIO_PIPELINE_CONTINUE);
2137 zio_dva_free(zio_t *zio)
2139 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2141 return (ZIO_PIPELINE_CONTINUE);
2145 zio_dva_claim(zio_t *zio)
2149 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2151 zio->io_error = error;
2153 return (ZIO_PIPELINE_CONTINUE);
2157 * Undo an allocation. This is used by zio_done() when an I/O fails
2158 * and we want to give back the block we just allocated.
2159 * This handles both normal blocks and gang blocks.
2162 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2164 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2165 ASSERT(zio->io_bp_override == NULL);
2167 if (!BP_IS_HOLE(bp))
2168 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2171 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2172 zio_dva_unallocate(zio, gn->gn_child[g],
2173 &gn->gn_gbh->zg_blkptr[g]);
2179 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2182 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2183 uint64_t size, boolean_t use_slog)
2187 ASSERT(txg > spa_syncing_txg(spa));
2190 error = metaslab_alloc(spa, spa_log_class(spa), size,
2191 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2194 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2195 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2198 BP_SET_LSIZE(new_bp, size);
2199 BP_SET_PSIZE(new_bp, size);
2200 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2201 BP_SET_CHECKSUM(new_bp,
2202 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2203 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2204 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2205 BP_SET_LEVEL(new_bp, 0);
2206 BP_SET_DEDUP(new_bp, 0);
2207 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2214 * Free an intent log block.
2217 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2219 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2220 ASSERT(!BP_IS_GANG(bp));
2222 zio_free(spa, txg, bp);
2226 * ==========================================================================
2227 * Read and write to physical devices
2228 * ==========================================================================
2231 zio_vdev_io_start(zio_t *zio)
2233 vdev_t *vd = zio->io_vd;
2235 spa_t *spa = zio->io_spa;
2237 ASSERT(zio->io_error == 0);
2238 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2241 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2242 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2245 * The mirror_ops handle multiple DVAs in a single BP.
2247 return (vdev_mirror_ops.vdev_op_io_start(zio));
2250 align = 1ULL << vd->vdev_top->vdev_ashift;
2252 if (P2PHASE(zio->io_size, align) != 0) {
2253 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2254 char *abuf = zio_buf_alloc(asize);
2255 ASSERT(vd == vd->vdev_top);
2256 if (zio->io_type == ZIO_TYPE_WRITE) {
2257 bcopy(zio->io_data, abuf, zio->io_size);
2258 bzero(abuf + zio->io_size, asize - zio->io_size);
2260 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2263 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2264 ASSERT(P2PHASE(zio->io_size, align) == 0);
2265 ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2268 * If this is a repair I/O, and there's no self-healing involved --
2269 * that is, we're just resilvering what we expect to resilver --
2270 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2271 * This prevents spurious resilvering with nested replication.
2272 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2273 * A is out of date, we'll read from C+D, then use the data to
2274 * resilver A+B -- but we don't actually want to resilver B, just A.
2275 * The top-level mirror has no way to know this, so instead we just
2276 * discard unnecessary repairs as we work our way down the vdev tree.
2277 * The same logic applies to any form of nested replication:
2278 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2280 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2281 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2282 zio->io_txg != 0 && /* not a delegated i/o */
2283 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2284 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2285 zio_vdev_io_bypass(zio);
2286 return (ZIO_PIPELINE_CONTINUE);
2289 if (vd->vdev_ops->vdev_op_leaf &&
2290 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2292 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2293 return (ZIO_PIPELINE_CONTINUE);
2295 if ((zio = vdev_queue_io(zio)) == NULL)
2296 return (ZIO_PIPELINE_STOP);
2298 if (!vdev_accessible(vd, zio)) {
2299 zio->io_error = ENXIO;
2301 return (ZIO_PIPELINE_STOP);
2305 return (vd->vdev_ops->vdev_op_io_start(zio));
2309 zio_vdev_io_done(zio_t *zio)
2311 vdev_t *vd = zio->io_vd;
2312 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2313 boolean_t unexpected_error = B_FALSE;
2315 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2316 return (ZIO_PIPELINE_STOP);
2318 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2320 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2322 vdev_queue_io_done(zio);
2324 if (zio->io_type == ZIO_TYPE_WRITE)
2325 vdev_cache_write(zio);
2327 if (zio_injection_enabled && zio->io_error == 0)
2328 zio->io_error = zio_handle_device_injection(vd,
2331 if (zio_injection_enabled && zio->io_error == 0)
2332 zio->io_error = zio_handle_label_injection(zio, EIO);
2334 if (zio->io_error) {
2335 if (!vdev_accessible(vd, zio)) {
2336 zio->io_error = ENXIO;
2338 unexpected_error = B_TRUE;
2343 ops->vdev_op_io_done(zio);
2345 if (unexpected_error)
2346 VERIFY(vdev_probe(vd, zio) == NULL);
2348 return (ZIO_PIPELINE_CONTINUE);
2352 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2353 * disk, and use that to finish the checksum ereport later.
2356 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2357 const void *good_buf)
2359 /* no processing needed */
2360 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2365 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2367 void *buf = zio_buf_alloc(zio->io_size);
2369 bcopy(zio->io_data, buf, zio->io_size);
2371 zcr->zcr_cbinfo = zio->io_size;
2372 zcr->zcr_cbdata = buf;
2373 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2374 zcr->zcr_free = zio_buf_free;
2378 zio_vdev_io_assess(zio_t *zio)
2380 vdev_t *vd = zio->io_vd;
2382 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2383 return (ZIO_PIPELINE_STOP);
2385 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2386 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2388 if (zio->io_vsd != NULL) {
2389 zio->io_vsd_ops->vsd_free(zio);
2393 if (zio_injection_enabled && zio->io_error == 0)
2394 zio->io_error = zio_handle_fault_injection(zio, EIO);
2397 * If the I/O failed, determine whether we should attempt to retry it.
2399 * On retry, we cut in line in the issue queue, since we don't want
2400 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2402 if (zio->io_error && vd == NULL &&
2403 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2404 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2405 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2407 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2408 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2409 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2410 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2411 zio_requeue_io_start_cut_in_line);
2412 return (ZIO_PIPELINE_STOP);
2416 * If we got an error on a leaf device, convert it to ENXIO
2417 * if the device is not accessible at all.
2419 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2420 !vdev_accessible(vd, zio))
2421 zio->io_error = ENXIO;
2424 * If we can't write to an interior vdev (mirror or RAID-Z),
2425 * set vdev_cant_write so that we stop trying to allocate from it.
2427 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2428 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2429 vd->vdev_cant_write = B_TRUE;
2432 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2434 return (ZIO_PIPELINE_CONTINUE);
2438 zio_vdev_io_reissue(zio_t *zio)
2440 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2441 ASSERT(zio->io_error == 0);
2443 zio->io_stage >>= 1;
2447 zio_vdev_io_redone(zio_t *zio)
2449 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2451 zio->io_stage >>= 1;
2455 zio_vdev_io_bypass(zio_t *zio)
2457 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2458 ASSERT(zio->io_error == 0);
2460 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2461 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2465 * ==========================================================================
2466 * Generate and verify checksums
2467 * ==========================================================================
2470 zio_checksum_generate(zio_t *zio)
2472 blkptr_t *bp = zio->io_bp;
2473 enum zio_checksum checksum;
2477 * This is zio_write_phys().
2478 * We're either generating a label checksum, or none at all.
2480 checksum = zio->io_prop.zp_checksum;
2482 if (checksum == ZIO_CHECKSUM_OFF)
2483 return (ZIO_PIPELINE_CONTINUE);
2485 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2487 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2488 ASSERT(!IO_IS_ALLOCATING(zio));
2489 checksum = ZIO_CHECKSUM_GANG_HEADER;
2491 checksum = BP_GET_CHECKSUM(bp);
2495 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2497 return (ZIO_PIPELINE_CONTINUE);
2501 zio_checksum_verify(zio_t *zio)
2503 zio_bad_cksum_t info;
2504 blkptr_t *bp = zio->io_bp;
2507 ASSERT(zio->io_vd != NULL);
2511 * This is zio_read_phys().
2512 * We're either verifying a label checksum, or nothing at all.
2514 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2515 return (ZIO_PIPELINE_CONTINUE);
2517 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2520 if ((error = zio_checksum_error(zio, &info)) != 0) {
2521 zio->io_error = error;
2522 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2523 zfs_ereport_start_checksum(zio->io_spa,
2524 zio->io_vd, zio, zio->io_offset,
2525 zio->io_size, NULL, &info);
2529 return (ZIO_PIPELINE_CONTINUE);
2533 * Called by RAID-Z to ensure we don't compute the checksum twice.
2536 zio_checksum_verified(zio_t *zio)
2538 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2542 * ==========================================================================
2543 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2544 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2545 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2546 * indicate errors that are specific to one I/O, and most likely permanent.
2547 * Any other error is presumed to be worse because we weren't expecting it.
2548 * ==========================================================================
2551 zio_worst_error(int e1, int e2)
2553 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2556 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2557 if (e1 == zio_error_rank[r1])
2560 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2561 if (e2 == zio_error_rank[r2])
2564 return (r1 > r2 ? e1 : e2);
2568 * ==========================================================================
2570 * ==========================================================================
2573 zio_ready(zio_t *zio)
2575 blkptr_t *bp = zio->io_bp;
2576 zio_t *pio, *pio_next;
2578 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2579 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2580 return (ZIO_PIPELINE_STOP);
2582 if (zio->io_ready) {
2583 ASSERT(IO_IS_ALLOCATING(zio));
2584 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2585 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2590 if (bp != NULL && bp != &zio->io_bp_copy)
2591 zio->io_bp_copy = *bp;
2594 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2596 mutex_enter(&zio->io_lock);
2597 zio->io_state[ZIO_WAIT_READY] = 1;
2598 pio = zio_walk_parents(zio);
2599 mutex_exit(&zio->io_lock);
2602 * As we notify zio's parents, new parents could be added.
2603 * New parents go to the head of zio's io_parent_list, however,
2604 * so we will (correctly) not notify them. The remainder of zio's
2605 * io_parent_list, from 'pio_next' onward, cannot change because
2606 * all parents must wait for us to be done before they can be done.
2608 for (; pio != NULL; pio = pio_next) {
2609 pio_next = zio_walk_parents(zio);
2610 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2613 if (zio->io_flags & ZIO_FLAG_NODATA) {
2614 if (BP_IS_GANG(bp)) {
2615 zio->io_flags &= ~ZIO_FLAG_NODATA;
2617 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2618 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2622 if (zio_injection_enabled &&
2623 zio->io_spa->spa_syncing_txg == zio->io_txg)
2624 zio_handle_ignored_writes(zio);
2626 return (ZIO_PIPELINE_CONTINUE);
2630 zio_done(zio_t *zio)
2632 spa_t *spa = zio->io_spa;
2633 zio_t *lio = zio->io_logical;
2634 blkptr_t *bp = zio->io_bp;
2635 vdev_t *vd = zio->io_vd;
2636 uint64_t psize = zio->io_size;
2637 zio_t *pio, *pio_next;
2640 * If our children haven't all completed,
2641 * wait for them and then repeat this pipeline stage.
2643 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2644 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2645 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2646 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2647 return (ZIO_PIPELINE_STOP);
2649 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2650 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2651 ASSERT(zio->io_children[c][w] == 0);
2654 ASSERT(bp->blk_pad[0] == 0);
2655 ASSERT(bp->blk_pad[1] == 0);
2656 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2657 (bp == zio_unique_parent(zio)->io_bp));
2658 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2659 zio->io_bp_override == NULL &&
2660 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2661 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2662 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2663 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2664 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2669 * If there were child vdev/gang/ddt errors, they apply to us now.
2671 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2672 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2673 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2676 * If the I/O on the transformed data was successful, generate any
2677 * checksum reports now while we still have the transformed data.
2679 if (zio->io_error == 0) {
2680 while (zio->io_cksum_report != NULL) {
2681 zio_cksum_report_t *zcr = zio->io_cksum_report;
2682 uint64_t align = zcr->zcr_align;
2683 uint64_t asize = P2ROUNDUP(psize, align);
2684 char *abuf = zio->io_data;
2686 if (asize != psize) {
2687 abuf = zio_buf_alloc(asize);
2688 bcopy(zio->io_data, abuf, psize);
2689 bzero(abuf + psize, asize - psize);
2692 zio->io_cksum_report = zcr->zcr_next;
2693 zcr->zcr_next = NULL;
2694 zcr->zcr_finish(zcr, abuf);
2695 zfs_ereport_free_checksum(zcr);
2698 zio_buf_free(abuf, asize);
2702 zio_pop_transforms(zio); /* note: may set zio->io_error */
2704 vdev_stat_update(zio, psize);
2706 if (zio->io_error) {
2708 * If this I/O is attached to a particular vdev,
2709 * generate an error message describing the I/O failure
2710 * at the block level. We ignore these errors if the
2711 * device is currently unavailable.
2713 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2714 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2716 if ((zio->io_error == EIO || !(zio->io_flags &
2717 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2720 * For logical I/O requests, tell the SPA to log the
2721 * error and generate a logical data ereport.
2723 spa_log_error(spa, zio);
2724 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2729 if (zio->io_error && zio == lio) {
2731 * Determine whether zio should be reexecuted. This will
2732 * propagate all the way to the root via zio_notify_parent().
2734 ASSERT(vd == NULL && bp != NULL);
2735 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2737 if (IO_IS_ALLOCATING(zio) &&
2738 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2739 if (zio->io_error != ENOSPC)
2740 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2742 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2745 if ((zio->io_type == ZIO_TYPE_READ ||
2746 zio->io_type == ZIO_TYPE_FREE) &&
2747 zio->io_error == ENXIO &&
2748 spa_load_state(spa) == SPA_LOAD_NONE &&
2749 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2750 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2752 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2753 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2756 * Here is a possibly good place to attempt to do
2757 * either combinatorial reconstruction or error correction
2758 * based on checksums. It also might be a good place
2759 * to send out preliminary ereports before we suspend
2765 * If there were logical child errors, they apply to us now.
2766 * We defer this until now to avoid conflating logical child
2767 * errors with errors that happened to the zio itself when
2768 * updating vdev stats and reporting FMA events above.
2770 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2772 if ((zio->io_error || zio->io_reexecute) &&
2773 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2774 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2775 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2777 zio_gang_tree_free(&zio->io_gang_tree);
2780 * Godfather I/Os should never suspend.
2782 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2783 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2784 zio->io_reexecute = 0;
2786 if (zio->io_reexecute) {
2788 * This is a logical I/O that wants to reexecute.
2790 * Reexecute is top-down. When an i/o fails, if it's not
2791 * the root, it simply notifies its parent and sticks around.
2792 * The parent, seeing that it still has children in zio_done(),
2793 * does the same. This percolates all the way up to the root.
2794 * The root i/o will reexecute or suspend the entire tree.
2796 * This approach ensures that zio_reexecute() honors
2797 * all the original i/o dependency relationships, e.g.
2798 * parents not executing until children are ready.
2800 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2802 zio->io_gang_leader = NULL;
2804 mutex_enter(&zio->io_lock);
2805 zio->io_state[ZIO_WAIT_DONE] = 1;
2806 mutex_exit(&zio->io_lock);
2809 * "The Godfather" I/O monitors its children but is
2810 * not a true parent to them. It will track them through
2811 * the pipeline but severs its ties whenever they get into
2812 * trouble (e.g. suspended). This allows "The Godfather"
2813 * I/O to return status without blocking.
2815 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2816 zio_link_t *zl = zio->io_walk_link;
2817 pio_next = zio_walk_parents(zio);
2819 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2820 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2821 zio_remove_child(pio, zio, zl);
2822 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2826 if ((pio = zio_unique_parent(zio)) != NULL) {
2828 * We're not a root i/o, so there's nothing to do
2829 * but notify our parent. Don't propagate errors
2830 * upward since we haven't permanently failed yet.
2832 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2833 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2834 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2835 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2837 * We'd fail again if we reexecuted now, so suspend
2838 * until conditions improve (e.g. device comes online).
2840 zio_suspend(spa, zio);
2843 * Reexecution is potentially a huge amount of work.
2844 * Hand it off to the otherwise-unused claim taskq.
2846 (void) taskq_dispatch(
2847 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2848 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2850 return (ZIO_PIPELINE_STOP);
2853 ASSERT(zio->io_child_count == 0);
2854 ASSERT(zio->io_reexecute == 0);
2855 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2858 * Report any checksum errors, since the I/O is complete.
2860 while (zio->io_cksum_report != NULL) {
2861 zio_cksum_report_t *zcr = zio->io_cksum_report;
2862 zio->io_cksum_report = zcr->zcr_next;
2863 zcr->zcr_next = NULL;
2864 zcr->zcr_finish(zcr, NULL);
2865 zfs_ereport_free_checksum(zcr);
2869 * It is the responsibility of the done callback to ensure that this
2870 * particular zio is no longer discoverable for adoption, and as
2871 * such, cannot acquire any new parents.
2876 mutex_enter(&zio->io_lock);
2877 zio->io_state[ZIO_WAIT_DONE] = 1;
2878 mutex_exit(&zio->io_lock);
2880 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2881 zio_link_t *zl = zio->io_walk_link;
2882 pio_next = zio_walk_parents(zio);
2883 zio_remove_child(pio, zio, zl);
2884 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2887 if (zio->io_waiter != NULL) {
2888 mutex_enter(&zio->io_lock);
2889 zio->io_executor = NULL;
2890 cv_broadcast(&zio->io_cv);
2891 mutex_exit(&zio->io_lock);
2896 return (ZIO_PIPELINE_STOP);
2900 * ==========================================================================
2901 * I/O pipeline definition
2902 * ==========================================================================
2904 static zio_pipe_stage_t *zio_pipeline[] = {
2910 zio_checksum_generate,
2924 zio_checksum_verify,