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;
96 static inline void __zio_execute(zio_t *zio);
102 vmem_t *data_alloc_arena = NULL;
105 data_alloc_arena = zio_alloc_arena;
107 zio_cache = kmem_cache_create("zio_cache",
108 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
109 zio_link_cache = kmem_cache_create("zio_link_cache",
110 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
113 * For small buffers, we want a cache for each multiple of
114 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
115 * for each quarter-power of 2. For large buffers, we want
116 * a cache for each multiple of PAGESIZE.
118 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
119 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
123 while (p2 & (p2 - 1))
126 if (size <= 4 * SPA_MINBLOCKSIZE) {
127 align = SPA_MINBLOCKSIZE;
128 } else if (P2PHASE(size, PAGESIZE) == 0) {
130 } else if (P2PHASE(size, p2 >> 2) == 0) {
136 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
137 zio_buf_cache[c] = kmem_cache_create(name, size,
138 align, NULL, NULL, NULL, NULL, NULL,
139 size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
141 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
142 zio_data_buf_cache[c] = kmem_cache_create(name, size,
143 align, NULL, NULL, NULL, NULL, data_alloc_arena,
144 size > zio_buf_debug_limit ? KMC_NODEBUG : 0);
149 ASSERT(zio_buf_cache[c] != NULL);
150 if (zio_buf_cache[c - 1] == NULL)
151 zio_buf_cache[c - 1] = zio_buf_cache[c];
153 ASSERT(zio_data_buf_cache[c] != NULL);
154 if (zio_data_buf_cache[c - 1] == NULL)
155 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
165 kmem_cache_t *last_cache = NULL;
166 kmem_cache_t *last_data_cache = NULL;
168 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
169 if (zio_buf_cache[c] != last_cache) {
170 last_cache = zio_buf_cache[c];
171 kmem_cache_destroy(zio_buf_cache[c]);
173 zio_buf_cache[c] = NULL;
175 if (zio_data_buf_cache[c] != last_data_cache) {
176 last_data_cache = zio_data_buf_cache[c];
177 kmem_cache_destroy(zio_data_buf_cache[c]);
179 zio_data_buf_cache[c] = NULL;
182 kmem_cache_destroy(zio_link_cache);
183 kmem_cache_destroy(zio_cache);
189 * ==========================================================================
190 * Allocate and free I/O buffers
191 * ==========================================================================
195 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
196 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
197 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
198 * excess / transient data in-core during a crashdump.
201 zio_buf_alloc(size_t size)
203 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
205 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
207 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
211 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
212 * crashdump if the kernel panics. This exists so that we will limit the amount
213 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
214 * of kernel heap dumped to disk when the kernel panics)
217 zio_data_buf_alloc(size_t size)
219 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
221 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
223 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
227 zio_buf_free(void *buf, size_t size)
229 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
231 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
233 kmem_cache_free(zio_buf_cache[c], buf);
237 zio_data_buf_free(void *buf, size_t size)
239 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
241 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
243 kmem_cache_free(zio_data_buf_cache[c], buf);
247 * ==========================================================================
248 * Push and pop I/O transform buffers
249 * ==========================================================================
252 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
253 zio_transform_func_t *transform)
255 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
257 zt->zt_orig_data = zio->io_data;
258 zt->zt_orig_size = zio->io_size;
259 zt->zt_bufsize = bufsize;
260 zt->zt_transform = transform;
262 zt->zt_next = zio->io_transform_stack;
263 zio->io_transform_stack = zt;
270 zio_pop_transforms(zio_t *zio)
274 while ((zt = zio->io_transform_stack) != NULL) {
275 if (zt->zt_transform != NULL)
276 zt->zt_transform(zio,
277 zt->zt_orig_data, zt->zt_orig_size);
279 if (zt->zt_bufsize != 0)
280 zio_buf_free(zio->io_data, zt->zt_bufsize);
282 zio->io_data = zt->zt_orig_data;
283 zio->io_size = zt->zt_orig_size;
284 zio->io_transform_stack = zt->zt_next;
286 kmem_free(zt, sizeof (zio_transform_t));
291 * ==========================================================================
292 * I/O transform callbacks for subblocks and decompression
293 * ==========================================================================
296 zio_subblock(zio_t *zio, void *data, uint64_t size)
298 ASSERT(zio->io_size > size);
300 if (zio->io_type == ZIO_TYPE_READ)
301 bcopy(zio->io_data, data, size);
305 zio_decompress(zio_t *zio, void *data, uint64_t size)
307 if (zio->io_error == 0 &&
308 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
309 zio->io_data, data, zio->io_size, size) != 0)
314 * ==========================================================================
315 * I/O parent/child relationships and pipeline interlocks
316 * ==========================================================================
319 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
320 * continue calling these functions until they return NULL.
321 * Otherwise, the next caller will pick up the list walk in
322 * some indeterminate state. (Otherwise every caller would
323 * have to pass in a cookie to keep the state represented by
324 * io_walk_link, which gets annoying.)
327 zio_walk_parents(zio_t *cio)
329 zio_link_t *zl = cio->io_walk_link;
330 list_t *pl = &cio->io_parent_list;
332 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
333 cio->io_walk_link = zl;
338 ASSERT(zl->zl_child == cio);
339 return (zl->zl_parent);
343 zio_walk_children(zio_t *pio)
345 zio_link_t *zl = pio->io_walk_link;
346 list_t *cl = &pio->io_child_list;
348 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
349 pio->io_walk_link = zl;
354 ASSERT(zl->zl_parent == pio);
355 return (zl->zl_child);
359 zio_unique_parent(zio_t *cio)
361 zio_t *pio = zio_walk_parents(cio);
363 VERIFY(zio_walk_parents(cio) == NULL);
368 zio_add_child(zio_t *pio, zio_t *cio)
370 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
374 * Logical I/Os can have logical, gang, or vdev children.
375 * Gang I/Os can have gang or vdev children.
376 * Vdev I/Os can only have vdev children.
377 * The following ASSERT captures all of these constraints.
379 ASSERT(cio->io_child_type <= pio->io_child_type);
384 mutex_enter(&cio->io_lock);
385 mutex_enter(&pio->io_lock);
387 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
389 for (w = 0; w < ZIO_WAIT_TYPES; w++)
390 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
392 list_insert_head(&pio->io_child_list, zl);
393 list_insert_head(&cio->io_parent_list, zl);
395 pio->io_child_count++;
396 cio->io_parent_count++;
398 mutex_exit(&pio->io_lock);
399 mutex_exit(&cio->io_lock);
403 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
405 ASSERT(zl->zl_parent == pio);
406 ASSERT(zl->zl_child == cio);
408 mutex_enter(&cio->io_lock);
409 mutex_enter(&pio->io_lock);
411 list_remove(&pio->io_child_list, zl);
412 list_remove(&cio->io_parent_list, zl);
414 pio->io_child_count--;
415 cio->io_parent_count--;
417 mutex_exit(&pio->io_lock);
418 mutex_exit(&cio->io_lock);
420 kmem_cache_free(zio_link_cache, zl);
424 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
426 uint64_t *countp = &zio->io_children[child][wait];
427 boolean_t waiting = B_FALSE;
429 mutex_enter(&zio->io_lock);
430 ASSERT(zio->io_stall == NULL);
433 zio->io_stall = countp;
436 mutex_exit(&zio->io_lock);
441 __attribute__((always_inline))
443 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
445 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
446 int *errorp = &pio->io_child_error[zio->io_child_type];
448 mutex_enter(&pio->io_lock);
449 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
450 *errorp = zio_worst_error(*errorp, zio->io_error);
451 pio->io_reexecute |= zio->io_reexecute;
452 ASSERT3U(*countp, >, 0);
453 if (--*countp == 0 && pio->io_stall == countp) {
454 pio->io_stall = NULL;
455 mutex_exit(&pio->io_lock);
458 mutex_exit(&pio->io_lock);
463 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
465 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
466 zio->io_error = zio->io_child_error[c];
470 * ==========================================================================
471 * Create the various types of I/O (read, write, free, etc)
472 * ==========================================================================
475 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
476 void *data, uint64_t size, zio_done_func_t *done, void *private,
477 zio_type_t type, int priority, enum zio_flag flags,
478 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
479 enum zio_stage stage, enum zio_stage pipeline)
483 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
484 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
485 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
487 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
488 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
489 ASSERT(vd || stage == ZIO_STAGE_OPEN);
491 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
492 bzero(zio, sizeof (zio_t));
494 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
495 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
497 list_create(&zio->io_parent_list, sizeof (zio_link_t),
498 offsetof(zio_link_t, zl_parent_node));
499 list_create(&zio->io_child_list, sizeof (zio_link_t),
500 offsetof(zio_link_t, zl_child_node));
503 zio->io_child_type = ZIO_CHILD_VDEV;
504 else if (flags & ZIO_FLAG_GANG_CHILD)
505 zio->io_child_type = ZIO_CHILD_GANG;
506 else if (flags & ZIO_FLAG_DDT_CHILD)
507 zio->io_child_type = ZIO_CHILD_DDT;
509 zio->io_child_type = ZIO_CHILD_LOGICAL;
512 zio->io_bp = (blkptr_t *)bp;
513 zio->io_bp_copy = *bp;
514 zio->io_bp_orig = *bp;
515 if (type != ZIO_TYPE_WRITE ||
516 zio->io_child_type == ZIO_CHILD_DDT)
517 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
518 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
519 zio->io_logical = zio;
520 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
521 pipeline |= ZIO_GANG_STAGES;
527 zio->io_private = private;
529 zio->io_priority = priority;
531 zio->io_offset = offset;
532 zio->io_orig_data = zio->io_data = data;
533 zio->io_orig_size = zio->io_size = size;
534 zio->io_orig_flags = zio->io_flags = flags;
535 zio->io_orig_stage = zio->io_stage = stage;
536 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
538 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
539 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
542 zio->io_bookmark = *zb;
545 if (zio->io_logical == NULL)
546 zio->io_logical = pio->io_logical;
547 if (zio->io_child_type == ZIO_CHILD_GANG)
548 zio->io_gang_leader = pio->io_gang_leader;
549 zio_add_child(pio, zio);
556 zio_destroy(zio_t *zio)
558 list_destroy(&zio->io_parent_list);
559 list_destroy(&zio->io_child_list);
560 mutex_destroy(&zio->io_lock);
561 cv_destroy(&zio->io_cv);
562 kmem_cache_free(zio_cache, zio);
566 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
567 void *private, enum zio_flag flags)
571 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
572 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
573 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
579 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
581 return (zio_null(NULL, spa, NULL, done, private, flags));
585 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
586 void *data, uint64_t size, zio_done_func_t *done, void *private,
587 int priority, enum zio_flag flags, const zbookmark_t *zb)
591 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
592 data, size, done, private,
593 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
594 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
595 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
601 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
602 void *data, uint64_t size, const zio_prop_t *zp,
603 zio_done_func_t *ready, zio_done_func_t *done, void *private,
604 int priority, enum zio_flag flags, const zbookmark_t *zb)
608 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
609 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
610 zp->zp_compress >= ZIO_COMPRESS_OFF &&
611 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
612 zp->zp_type < DMU_OT_NUMTYPES &&
615 zp->zp_copies <= spa_max_replication(spa) &&
617 zp->zp_dedup_verify <= 1);
619 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
620 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
621 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
622 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
624 zio->io_ready = ready;
631 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
632 uint64_t size, zio_done_func_t *done, void *private, int priority,
633 enum zio_flag flags, zbookmark_t *zb)
637 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
638 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
639 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
645 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
647 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
648 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
649 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
650 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
652 zio->io_prop.zp_copies = copies;
653 zio->io_bp_override = bp;
657 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
659 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
663 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
668 dprintf_bp(bp, "freeing in txg %llu, pass %u",
669 (longlong_t)txg, spa->spa_sync_pass);
671 ASSERT(!BP_IS_HOLE(bp));
672 ASSERT(spa_syncing_txg(spa) == txg);
673 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
675 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
676 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
677 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
683 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
684 zio_done_func_t *done, void *private, enum zio_flag flags)
689 * A claim is an allocation of a specific block. Claims are needed
690 * to support immediate writes in the intent log. The issue is that
691 * immediate writes contain committed data, but in a txg that was
692 * *not* committed. Upon opening the pool after an unclean shutdown,
693 * the intent log claims all blocks that contain immediate write data
694 * so that the SPA knows they're in use.
696 * All claims *must* be resolved in the first txg -- before the SPA
697 * starts allocating blocks -- so that nothing is allocated twice.
698 * If txg == 0 we just verify that the block is claimable.
700 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
701 ASSERT(txg == spa_first_txg(spa) || txg == 0);
702 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
704 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
705 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
706 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
712 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
713 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
718 if (vd->vdev_children == 0) {
719 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
720 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
721 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
725 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
727 for (c = 0; c < vd->vdev_children; c++)
728 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
729 done, private, priority, flags));
736 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
737 void *data, int checksum, zio_done_func_t *done, void *private,
738 int priority, enum zio_flag flags, boolean_t labels)
742 ASSERT(vd->vdev_children == 0);
743 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
744 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
745 ASSERT3U(offset + size, <=, vd->vdev_psize);
747 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
748 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
749 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
751 zio->io_prop.zp_checksum = checksum;
757 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
758 void *data, int checksum, zio_done_func_t *done, void *private,
759 int priority, enum zio_flag flags, boolean_t labels)
763 ASSERT(vd->vdev_children == 0);
764 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
765 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
766 ASSERT3U(offset + size, <=, vd->vdev_psize);
768 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
769 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
770 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
772 zio->io_prop.zp_checksum = checksum;
774 if (zio_checksum_table[checksum].ci_eck) {
776 * zec checksums are necessarily destructive -- they modify
777 * the end of the write buffer to hold the verifier/checksum.
778 * Therefore, we must make a local copy in case the data is
779 * being written to multiple places in parallel.
781 void *wbuf = zio_buf_alloc(size);
782 bcopy(data, wbuf, size);
783 zio_push_transform(zio, wbuf, size, size, NULL);
790 * Create a child I/O to do some work for us.
793 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
794 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
795 zio_done_func_t *done, void *private)
797 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
800 ASSERT(vd->vdev_parent ==
801 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
803 if (type == ZIO_TYPE_READ && bp != NULL) {
805 * If we have the bp, then the child should perform the
806 * checksum and the parent need not. This pushes error
807 * detection as close to the leaves as possible and
808 * eliminates redundant checksums in the interior nodes.
810 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
811 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
814 if (vd->vdev_children == 0)
815 offset += VDEV_LABEL_START_SIZE;
817 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
820 * If we've decided to do a repair, the write is not speculative --
821 * even if the original read was.
823 if (flags & ZIO_FLAG_IO_REPAIR)
824 flags &= ~ZIO_FLAG_SPECULATIVE;
826 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
827 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
828 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
834 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
835 int type, int priority, enum zio_flag flags,
836 zio_done_func_t *done, void *private)
840 ASSERT(vd->vdev_ops->vdev_op_leaf);
842 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
843 data, size, done, private, type, priority,
844 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
846 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
852 zio_flush(zio_t *zio, vdev_t *vd)
854 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
855 NULL, NULL, ZIO_PRIORITY_NOW,
856 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
860 zio_shrink(zio_t *zio, uint64_t size)
862 ASSERT(zio->io_executor == NULL);
863 ASSERT(zio->io_orig_size == zio->io_size);
864 ASSERT(size <= zio->io_size);
867 * We don't shrink for raidz because of problems with the
868 * reconstruction when reading back less than the block size.
869 * Note, BP_IS_RAIDZ() assumes no compression.
871 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
872 if (!BP_IS_RAIDZ(zio->io_bp))
873 zio->io_orig_size = zio->io_size = size;
877 * ==========================================================================
878 * Prepare to read and write logical blocks
879 * ==========================================================================
883 zio_read_bp_init(zio_t *zio)
885 blkptr_t *bp = zio->io_bp;
887 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
888 zio->io_child_type == ZIO_CHILD_LOGICAL &&
889 !(zio->io_flags & ZIO_FLAG_RAW)) {
890 uint64_t psize = BP_GET_PSIZE(bp);
891 void *cbuf = zio_buf_alloc(psize);
893 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
896 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
897 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
899 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
900 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
902 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
903 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
905 return (ZIO_PIPELINE_CONTINUE);
909 zio_write_bp_init(zio_t *zio)
911 spa_t *spa = zio->io_spa;
912 zio_prop_t *zp = &zio->io_prop;
913 enum zio_compress compress = zp->zp_compress;
914 blkptr_t *bp = zio->io_bp;
915 uint64_t lsize = zio->io_size;
916 uint64_t psize = lsize;
920 * If our children haven't all reached the ready stage,
921 * wait for them and then repeat this pipeline stage.
923 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
924 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
925 return (ZIO_PIPELINE_STOP);
927 if (!IO_IS_ALLOCATING(zio))
928 return (ZIO_PIPELINE_CONTINUE);
930 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
932 if (zio->io_bp_override) {
933 ASSERT(bp->blk_birth != zio->io_txg);
934 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
936 *bp = *zio->io_bp_override;
937 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
939 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
940 return (ZIO_PIPELINE_CONTINUE);
942 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
943 zp->zp_dedup_verify);
945 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
947 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
948 return (ZIO_PIPELINE_CONTINUE);
950 zio->io_bp_override = NULL;
954 if (bp->blk_birth == zio->io_txg) {
956 * We're rewriting an existing block, which means we're
957 * working on behalf of spa_sync(). For spa_sync() to
958 * converge, it must eventually be the case that we don't
959 * have to allocate new blocks. But compression changes
960 * the blocksize, which forces a reallocate, and makes
961 * convergence take longer. Therefore, after the first
962 * few passes, stop compressing to ensure convergence.
964 pass = spa_sync_pass(spa);
966 ASSERT(zio->io_txg == spa_syncing_txg(spa));
967 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
968 ASSERT(!BP_GET_DEDUP(bp));
970 if (pass > SYNC_PASS_DONT_COMPRESS)
971 compress = ZIO_COMPRESS_OFF;
973 /* Make sure someone doesn't change their mind on overwrites */
974 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
975 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
978 if (compress != ZIO_COMPRESS_OFF) {
979 void *cbuf = zio_buf_alloc(lsize);
980 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
981 if (psize == 0 || psize == lsize) {
982 compress = ZIO_COMPRESS_OFF;
983 zio_buf_free(cbuf, lsize);
985 ASSERT(psize < lsize);
986 zio_push_transform(zio, cbuf, psize, lsize, NULL);
991 * The final pass of spa_sync() must be all rewrites, but the first
992 * few passes offer a trade-off: allocating blocks defers convergence,
993 * but newly allocated blocks are sequential, so they can be written
994 * to disk faster. Therefore, we allow the first few passes of
995 * spa_sync() to allocate new blocks, but force rewrites after that.
996 * There should only be a handful of blocks after pass 1 in any case.
998 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
999 pass > SYNC_PASS_REWRITE) {
1000 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1002 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1003 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1006 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1010 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1012 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1013 BP_SET_LSIZE(bp, lsize);
1014 BP_SET_PSIZE(bp, psize);
1015 BP_SET_COMPRESS(bp, compress);
1016 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1017 BP_SET_TYPE(bp, zp->zp_type);
1018 BP_SET_LEVEL(bp, zp->zp_level);
1019 BP_SET_DEDUP(bp, zp->zp_dedup);
1020 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1022 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1023 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1024 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1028 return (ZIO_PIPELINE_CONTINUE);
1032 zio_free_bp_init(zio_t *zio)
1034 blkptr_t *bp = zio->io_bp;
1036 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1037 if (BP_GET_DEDUP(bp))
1038 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1041 return (ZIO_PIPELINE_CONTINUE);
1045 * ==========================================================================
1046 * Execute the I/O pipeline
1047 * ==========================================================================
1051 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1053 spa_t *spa = zio->io_spa;
1054 zio_type_t t = zio->io_type;
1055 int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
1058 * If we're a config writer or a probe, the normal issue and
1059 * interrupt threads may all be blocked waiting for the config lock.
1060 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1062 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1066 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1068 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1072 * If this is a high priority I/O, then use the high priority taskq.
1074 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1075 spa->spa_zio_taskq[t][q + 1] != NULL)
1078 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1080 while (taskq_dispatch(spa->spa_zio_taskq[t][q],
1081 (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
1085 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1087 kthread_t *executor = zio->io_executor;
1088 spa_t *spa = zio->io_spa;
1091 for (t = 0; t < ZIO_TYPES; t++)
1092 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1099 zio_issue_async(zio_t *zio)
1101 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1103 return (ZIO_PIPELINE_STOP);
1107 zio_interrupt(zio_t *zio)
1109 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1113 * Execute the I/O pipeline until one of the following occurs:
1114 * (1) the I/O completes; (2) the pipeline stalls waiting for
1115 * dependent child I/Os; (3) the I/O issues, so we're waiting
1116 * for an I/O completion interrupt; (4) the I/O is delegated by
1117 * vdev-level caching or aggregation; (5) the I/O is deferred
1118 * due to vdev-level queueing; (6) the I/O is handed off to
1119 * another thread. In all cases, the pipeline stops whenever
1120 * there's no CPU work; it never burns a thread in cv_wait().
1122 * There's no locking on io_stage because there's no legitimate way
1123 * for multiple threads to be attempting to process the same I/O.
1125 static zio_pipe_stage_t *zio_pipeline[];
1128 * zio_execute() is a wrapper around the static function
1129 * __zio_execute() so that we can force __zio_execute() to be
1130 * inlined. This reduces stack overhead which is important
1131 * because __zio_execute() is called recursively in several zio
1132 * code paths. zio_execute() itself cannot be inlined because
1133 * it is externally visible.
1136 zio_execute(zio_t *zio)
1141 __attribute__((always_inline))
1143 __zio_execute(zio_t *zio)
1145 zio->io_executor = curthread;
1147 while (zio->io_stage < ZIO_STAGE_DONE) {
1148 enum zio_stage pipeline = zio->io_pipeline;
1149 enum zio_stage stage = zio->io_stage;
1152 ASSERT(!MUTEX_HELD(&zio->io_lock));
1153 ASSERT(ISP2(stage));
1154 ASSERT(zio->io_stall == NULL);
1158 } while ((stage & pipeline) == 0);
1160 ASSERT(stage <= ZIO_STAGE_DONE);
1163 * If we are in interrupt context and this pipeline stage
1164 * will grab a config lock that is held across I/O,
1165 * or may wait for an I/O that needs an interrupt thread
1166 * to complete, issue async to avoid deadlock.
1168 * For VDEV_IO_START, we cut in line so that the io will
1169 * be sent to disk promptly.
1171 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1172 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1173 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1174 zio_requeue_io_start_cut_in_line : B_FALSE;
1175 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1179 zio->io_stage = stage;
1180 rv = zio_pipeline[highbit(stage) - 1](zio);
1182 if (rv == ZIO_PIPELINE_STOP)
1185 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1191 * ==========================================================================
1192 * Initiate I/O, either sync or async
1193 * ==========================================================================
1196 zio_wait(zio_t *zio)
1200 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1201 ASSERT(zio->io_executor == NULL);
1203 zio->io_waiter = curthread;
1207 mutex_enter(&zio->io_lock);
1208 while (zio->io_executor != NULL)
1209 cv_wait(&zio->io_cv, &zio->io_lock);
1210 mutex_exit(&zio->io_lock);
1212 error = zio->io_error;
1219 zio_nowait(zio_t *zio)
1221 ASSERT(zio->io_executor == NULL);
1223 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1224 zio_unique_parent(zio) == NULL) {
1226 * This is a logical async I/O with no parent to wait for it.
1227 * We add it to the spa_async_root_zio "Godfather" I/O which
1228 * will ensure they complete prior to unloading the pool.
1230 spa_t *spa = zio->io_spa;
1232 zio_add_child(spa->spa_async_zio_root, zio);
1239 * ==========================================================================
1240 * Reexecute or suspend/resume failed I/O
1241 * ==========================================================================
1245 zio_reexecute(zio_t *pio)
1247 zio_t *cio, *cio_next;
1250 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1251 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1252 ASSERT(pio->io_gang_leader == NULL);
1253 ASSERT(pio->io_gang_tree == NULL);
1255 pio->io_flags = pio->io_orig_flags;
1256 pio->io_stage = pio->io_orig_stage;
1257 pio->io_pipeline = pio->io_orig_pipeline;
1258 pio->io_reexecute = 0;
1260 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1261 pio->io_state[w] = 0;
1262 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1263 pio->io_child_error[c] = 0;
1265 if (IO_IS_ALLOCATING(pio))
1266 BP_ZERO(pio->io_bp);
1269 * As we reexecute pio's children, new children could be created.
1270 * New children go to the head of pio's io_child_list, however,
1271 * so we will (correctly) not reexecute them. The key is that
1272 * the remainder of pio's io_child_list, from 'cio_next' onward,
1273 * cannot be affected by any side effects of reexecuting 'cio'.
1275 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1276 cio_next = zio_walk_children(pio);
1277 mutex_enter(&pio->io_lock);
1278 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1279 pio->io_children[cio->io_child_type][w]++;
1280 mutex_exit(&pio->io_lock);
1285 * Now that all children have been reexecuted, execute the parent.
1286 * We don't reexecute "The Godfather" I/O here as it's the
1287 * responsibility of the caller to wait on him.
1289 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1294 zio_suspend(spa_t *spa, zio_t *zio)
1296 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1297 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1298 "failure and the failure mode property for this pool "
1299 "is set to panic.", spa_name(spa));
1301 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1303 mutex_enter(&spa->spa_suspend_lock);
1305 if (spa->spa_suspend_zio_root == NULL)
1306 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1307 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1308 ZIO_FLAG_GODFATHER);
1310 spa->spa_suspended = B_TRUE;
1313 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1314 ASSERT(zio != spa->spa_suspend_zio_root);
1315 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1316 ASSERT(zio_unique_parent(zio) == NULL);
1317 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1318 zio_add_child(spa->spa_suspend_zio_root, zio);
1321 mutex_exit(&spa->spa_suspend_lock);
1325 zio_resume(spa_t *spa)
1330 * Reexecute all previously suspended i/o.
1332 mutex_enter(&spa->spa_suspend_lock);
1333 spa->spa_suspended = B_FALSE;
1334 cv_broadcast(&spa->spa_suspend_cv);
1335 pio = spa->spa_suspend_zio_root;
1336 spa->spa_suspend_zio_root = NULL;
1337 mutex_exit(&spa->spa_suspend_lock);
1343 return (zio_wait(pio));
1347 zio_resume_wait(spa_t *spa)
1349 mutex_enter(&spa->spa_suspend_lock);
1350 while (spa_suspended(spa))
1351 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1352 mutex_exit(&spa->spa_suspend_lock);
1356 * ==========================================================================
1359 * A gang block is a collection of small blocks that looks to the DMU
1360 * like one large block. When zio_dva_allocate() cannot find a block
1361 * of the requested size, due to either severe fragmentation or the pool
1362 * being nearly full, it calls zio_write_gang_block() to construct the
1363 * block from smaller fragments.
1365 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1366 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1367 * an indirect block: it's an array of block pointers. It consumes
1368 * only one sector and hence is allocatable regardless of fragmentation.
1369 * The gang header's bps point to its gang members, which hold the data.
1371 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1372 * as the verifier to ensure uniqueness of the SHA256 checksum.
1373 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1374 * not the gang header. This ensures that data block signatures (needed for
1375 * deduplication) are independent of how the block is physically stored.
1377 * Gang blocks can be nested: a gang member may itself be a gang block.
1378 * Thus every gang block is a tree in which root and all interior nodes are
1379 * gang headers, and the leaves are normal blocks that contain user data.
1380 * The root of the gang tree is called the gang leader.
1382 * To perform any operation (read, rewrite, free, claim) on a gang block,
1383 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1384 * in the io_gang_tree field of the original logical i/o by recursively
1385 * reading the gang leader and all gang headers below it. This yields
1386 * an in-core tree containing the contents of every gang header and the
1387 * bps for every constituent of the gang block.
1389 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1390 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1391 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1392 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1393 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1394 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1395 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1396 * of the gang header plus zio_checksum_compute() of the data to update the
1397 * gang header's blk_cksum as described above.
1399 * The two-phase assemble/issue model solves the problem of partial failure --
1400 * what if you'd freed part of a gang block but then couldn't read the
1401 * gang header for another part? Assembling the entire gang tree first
1402 * ensures that all the necessary gang header I/O has succeeded before
1403 * starting the actual work of free, claim, or write. Once the gang tree
1404 * is assembled, free and claim are in-memory operations that cannot fail.
1406 * In the event that a gang write fails, zio_dva_unallocate() walks the
1407 * gang tree to immediately free (i.e. insert back into the space map)
1408 * everything we've allocated. This ensures that we don't get ENOSPC
1409 * errors during repeated suspend/resume cycles due to a flaky device.
1411 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1412 * the gang tree, we won't modify the block, so we can safely defer the free
1413 * (knowing that the block is still intact). If we *can* assemble the gang
1414 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1415 * each constituent bp and we can allocate a new block on the next sync pass.
1417 * In all cases, the gang tree allows complete recovery from partial failure.
1418 * ==========================================================================
1422 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1427 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1428 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1429 &pio->io_bookmark));
1433 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1438 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1439 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1440 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1442 * As we rewrite each gang header, the pipeline will compute
1443 * a new gang block header checksum for it; but no one will
1444 * compute a new data checksum, so we do that here. The one
1445 * exception is the gang leader: the pipeline already computed
1446 * its data checksum because that stage precedes gang assembly.
1447 * (Presently, nothing actually uses interior data checksums;
1448 * this is just good hygiene.)
1450 if (gn != pio->io_gang_leader->io_gang_tree) {
1451 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1452 data, BP_GET_PSIZE(bp));
1455 * If we are here to damage data for testing purposes,
1456 * leave the GBH alone so that we can detect the damage.
1458 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1459 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1461 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1462 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1463 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1471 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1473 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1474 ZIO_GANG_CHILD_FLAGS(pio)));
1479 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1481 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1482 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1485 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1494 static void zio_gang_tree_assemble_done(zio_t *zio);
1496 static zio_gang_node_t *
1497 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1499 zio_gang_node_t *gn;
1501 ASSERT(*gnpp == NULL);
1503 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1504 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1511 zio_gang_node_free(zio_gang_node_t **gnpp)
1513 zio_gang_node_t *gn = *gnpp;
1516 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1517 ASSERT(gn->gn_child[g] == NULL);
1519 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1520 kmem_free(gn, sizeof (*gn));
1525 zio_gang_tree_free(zio_gang_node_t **gnpp)
1527 zio_gang_node_t *gn = *gnpp;
1533 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1534 zio_gang_tree_free(&gn->gn_child[g]);
1536 zio_gang_node_free(gnpp);
1540 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1542 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1544 ASSERT(gio->io_gang_leader == gio);
1545 ASSERT(BP_IS_GANG(bp));
1547 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1548 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1549 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1553 zio_gang_tree_assemble_done(zio_t *zio)
1555 zio_t *gio = zio->io_gang_leader;
1556 zio_gang_node_t *gn = zio->io_private;
1557 blkptr_t *bp = zio->io_bp;
1560 ASSERT(gio == zio_unique_parent(zio));
1561 ASSERT(zio->io_child_count == 0);
1566 if (BP_SHOULD_BYTESWAP(bp))
1567 byteswap_uint64_array(zio->io_data, zio->io_size);
1569 ASSERT(zio->io_data == gn->gn_gbh);
1570 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1571 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1573 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1574 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1575 if (!BP_IS_GANG(gbp))
1577 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1582 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1584 zio_t *gio = pio->io_gang_leader;
1588 ASSERT(BP_IS_GANG(bp) == !!gn);
1589 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1590 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1593 * If you're a gang header, your data is in gn->gn_gbh.
1594 * If you're a gang member, your data is in 'data' and gn == NULL.
1596 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1599 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1601 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1602 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1603 if (BP_IS_HOLE(gbp))
1605 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1606 data = (char *)data + BP_GET_PSIZE(gbp);
1610 if (gn == gio->io_gang_tree)
1611 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1618 zio_gang_assemble(zio_t *zio)
1620 blkptr_t *bp = zio->io_bp;
1622 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1623 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1625 zio->io_gang_leader = zio;
1627 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1629 return (ZIO_PIPELINE_CONTINUE);
1633 zio_gang_issue(zio_t *zio)
1635 blkptr_t *bp = zio->io_bp;
1637 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1638 return (ZIO_PIPELINE_STOP);
1640 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1641 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1643 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1644 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1646 zio_gang_tree_free(&zio->io_gang_tree);
1648 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1650 return (ZIO_PIPELINE_CONTINUE);
1654 zio_write_gang_member_ready(zio_t *zio)
1656 zio_t *pio = zio_unique_parent(zio);
1657 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1658 dva_t *cdva = zio->io_bp->blk_dva;
1659 dva_t *pdva = pio->io_bp->blk_dva;
1663 if (BP_IS_HOLE(zio->io_bp))
1666 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1668 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1669 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1670 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1671 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1672 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1674 mutex_enter(&pio->io_lock);
1675 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1676 ASSERT(DVA_GET_GANG(&pdva[d]));
1677 asize = DVA_GET_ASIZE(&pdva[d]);
1678 asize += DVA_GET_ASIZE(&cdva[d]);
1679 DVA_SET_ASIZE(&pdva[d], asize);
1681 mutex_exit(&pio->io_lock);
1685 zio_write_gang_block(zio_t *pio)
1687 spa_t *spa = pio->io_spa;
1688 blkptr_t *bp = pio->io_bp;
1689 zio_t *gio = pio->io_gang_leader;
1691 zio_gang_node_t *gn, **gnpp;
1692 zio_gbh_phys_t *gbh;
1693 uint64_t txg = pio->io_txg;
1694 uint64_t resid = pio->io_size;
1696 int copies = gio->io_prop.zp_copies;
1697 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1701 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1702 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1703 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1705 pio->io_error = error;
1706 return (ZIO_PIPELINE_CONTINUE);
1710 gnpp = &gio->io_gang_tree;
1712 gnpp = pio->io_private;
1713 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1716 gn = zio_gang_node_alloc(gnpp);
1718 bzero(gbh, SPA_GANGBLOCKSIZE);
1721 * Create the gang header.
1723 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1724 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1727 * Create and nowait the gang children.
1729 for (g = 0; resid != 0; resid -= lsize, g++) {
1730 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1732 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1734 zp.zp_checksum = gio->io_prop.zp_checksum;
1735 zp.zp_compress = ZIO_COMPRESS_OFF;
1736 zp.zp_type = DMU_OT_NONE;
1738 zp.zp_copies = gio->io_prop.zp_copies;
1740 zp.zp_dedup_verify = 0;
1742 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1743 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1744 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1745 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1746 &pio->io_bookmark));
1750 * Set pio's pipeline to just wait for zio to finish.
1752 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1756 return (ZIO_PIPELINE_CONTINUE);
1760 * ==========================================================================
1762 * ==========================================================================
1765 zio_ddt_child_read_done(zio_t *zio)
1767 blkptr_t *bp = zio->io_bp;
1768 ddt_entry_t *dde = zio->io_private;
1770 zio_t *pio = zio_unique_parent(zio);
1772 mutex_enter(&pio->io_lock);
1773 ddp = ddt_phys_select(dde, bp);
1774 if (zio->io_error == 0)
1775 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1776 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1777 dde->dde_repair_data = zio->io_data;
1779 zio_buf_free(zio->io_data, zio->io_size);
1780 mutex_exit(&pio->io_lock);
1784 zio_ddt_read_start(zio_t *zio)
1786 blkptr_t *bp = zio->io_bp;
1789 ASSERT(BP_GET_DEDUP(bp));
1790 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1791 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1793 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1794 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1795 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1796 ddt_phys_t *ddp = dde->dde_phys;
1797 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1800 ASSERT(zio->io_vsd == NULL);
1803 if (ddp_self == NULL)
1804 return (ZIO_PIPELINE_CONTINUE);
1806 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1807 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1809 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1811 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1812 zio_buf_alloc(zio->io_size), zio->io_size,
1813 zio_ddt_child_read_done, dde, zio->io_priority,
1814 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1815 &zio->io_bookmark));
1817 return (ZIO_PIPELINE_CONTINUE);
1820 zio_nowait(zio_read(zio, zio->io_spa, bp,
1821 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1822 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1824 return (ZIO_PIPELINE_CONTINUE);
1828 zio_ddt_read_done(zio_t *zio)
1830 blkptr_t *bp = zio->io_bp;
1832 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1833 return (ZIO_PIPELINE_STOP);
1835 ASSERT(BP_GET_DEDUP(bp));
1836 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1837 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1839 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1840 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1841 ddt_entry_t *dde = zio->io_vsd;
1843 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1844 return (ZIO_PIPELINE_CONTINUE);
1847 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1848 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1849 return (ZIO_PIPELINE_STOP);
1851 if (dde->dde_repair_data != NULL) {
1852 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1853 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1855 ddt_repair_done(ddt, dde);
1859 ASSERT(zio->io_vsd == NULL);
1861 return (ZIO_PIPELINE_CONTINUE);
1865 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1867 spa_t *spa = zio->io_spa;
1871 * Note: we compare the original data, not the transformed data,
1872 * because when zio->io_bp is an override bp, we will not have
1873 * pushed the I/O transforms. That's an important optimization
1874 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1876 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1877 zio_t *lio = dde->dde_lead_zio[p];
1880 return (lio->io_orig_size != zio->io_orig_size ||
1881 bcmp(zio->io_orig_data, lio->io_orig_data,
1882 zio->io_orig_size) != 0);
1886 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1887 ddt_phys_t *ddp = &dde->dde_phys[p];
1889 if (ddp->ddp_phys_birth != 0) {
1890 arc_buf_t *abuf = NULL;
1891 uint32_t aflags = ARC_WAIT;
1892 blkptr_t blk = *zio->io_bp;
1895 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1899 error = arc_read_nolock(NULL, spa, &blk,
1900 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1901 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1902 &aflags, &zio->io_bookmark);
1905 if (arc_buf_size(abuf) != zio->io_orig_size ||
1906 bcmp(abuf->b_data, zio->io_orig_data,
1907 zio->io_orig_size) != 0)
1909 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1913 return (error != 0);
1921 zio_ddt_child_write_ready(zio_t *zio)
1923 int p = zio->io_prop.zp_copies;
1924 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1925 ddt_entry_t *dde = zio->io_private;
1926 ddt_phys_t *ddp = &dde->dde_phys[p];
1934 ASSERT(dde->dde_lead_zio[p] == zio);
1936 ddt_phys_fill(ddp, zio->io_bp);
1938 while ((pio = zio_walk_parents(zio)) != NULL)
1939 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1945 zio_ddt_child_write_done(zio_t *zio)
1947 int p = zio->io_prop.zp_copies;
1948 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1949 ddt_entry_t *dde = zio->io_private;
1950 ddt_phys_t *ddp = &dde->dde_phys[p];
1954 ASSERT(ddp->ddp_refcnt == 0);
1955 ASSERT(dde->dde_lead_zio[p] == zio);
1956 dde->dde_lead_zio[p] = NULL;
1958 if (zio->io_error == 0) {
1959 while (zio_walk_parents(zio) != NULL)
1960 ddt_phys_addref(ddp);
1962 ddt_phys_clear(ddp);
1969 zio_ddt_ditto_write_done(zio_t *zio)
1971 int p = DDT_PHYS_DITTO;
1972 blkptr_t *bp = zio->io_bp;
1973 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1974 ddt_entry_t *dde = zio->io_private;
1975 ddt_phys_t *ddp = &dde->dde_phys[p];
1976 ddt_key_t *ddk = &dde->dde_key;
1977 ASSERTV(zio_prop_t *zp = &zio->io_prop);
1981 ASSERT(ddp->ddp_refcnt == 0);
1982 ASSERT(dde->dde_lead_zio[p] == zio);
1983 dde->dde_lead_zio[p] = NULL;
1985 if (zio->io_error == 0) {
1986 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1987 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1988 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1989 if (ddp->ddp_phys_birth != 0)
1990 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1991 ddt_phys_fill(ddp, bp);
1998 zio_ddt_write(zio_t *zio)
2000 spa_t *spa = zio->io_spa;
2001 blkptr_t *bp = zio->io_bp;
2002 uint64_t txg = zio->io_txg;
2003 zio_prop_t *zp = &zio->io_prop;
2004 int p = zp->zp_copies;
2008 ddt_t *ddt = ddt_select(spa, bp);
2012 ASSERT(BP_GET_DEDUP(bp));
2013 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2014 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2017 dde = ddt_lookup(ddt, bp, B_TRUE);
2018 ddp = &dde->dde_phys[p];
2020 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2022 * If we're using a weak checksum, upgrade to a strong checksum
2023 * and try again. If we're already using a strong checksum,
2024 * we can't resolve it, so just convert to an ordinary write.
2025 * (And automatically e-mail a paper to Nature?)
2027 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2028 zp->zp_checksum = spa_dedup_checksum(spa);
2029 zio_pop_transforms(zio);
2030 zio->io_stage = ZIO_STAGE_OPEN;
2035 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2037 return (ZIO_PIPELINE_CONTINUE);
2040 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2041 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2043 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2044 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2045 zio_prop_t czp = *zp;
2047 czp.zp_copies = ditto_copies;
2050 * If we arrived here with an override bp, we won't have run
2051 * the transform stack, so we won't have the data we need to
2052 * generate a child i/o. So, toss the override bp and restart.
2053 * This is safe, because using the override bp is just an
2054 * optimization; and it's rare, so the cost doesn't matter.
2056 if (zio->io_bp_override) {
2057 zio_pop_transforms(zio);
2058 zio->io_stage = ZIO_STAGE_OPEN;
2059 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2060 zio->io_bp_override = NULL;
2063 return (ZIO_PIPELINE_CONTINUE);
2066 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2067 zio->io_orig_size, &czp, NULL,
2068 zio_ddt_ditto_write_done, dde, zio->io_priority,
2069 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2071 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2072 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2075 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2076 if (ddp->ddp_phys_birth != 0)
2077 ddt_bp_fill(ddp, bp, txg);
2078 if (dde->dde_lead_zio[p] != NULL)
2079 zio_add_child(zio, dde->dde_lead_zio[p]);
2081 ddt_phys_addref(ddp);
2082 } else if (zio->io_bp_override) {
2083 ASSERT(bp->blk_birth == txg);
2084 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2085 ddt_phys_fill(ddp, bp);
2086 ddt_phys_addref(ddp);
2088 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2089 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2090 zio_ddt_child_write_done, dde, zio->io_priority,
2091 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2093 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2094 dde->dde_lead_zio[p] = cio;
2104 return (ZIO_PIPELINE_CONTINUE);
2107 ddt_entry_t *freedde; /* for debugging */
2110 zio_ddt_free(zio_t *zio)
2112 spa_t *spa = zio->io_spa;
2113 blkptr_t *bp = zio->io_bp;
2114 ddt_t *ddt = ddt_select(spa, bp);
2118 ASSERT(BP_GET_DEDUP(bp));
2119 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2122 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2123 ddp = ddt_phys_select(dde, bp);
2124 ddt_phys_decref(ddp);
2127 return (ZIO_PIPELINE_CONTINUE);
2131 * ==========================================================================
2132 * Allocate and free blocks
2133 * ==========================================================================
2136 zio_dva_allocate(zio_t *zio)
2138 spa_t *spa = zio->io_spa;
2139 metaslab_class_t *mc = spa_normal_class(spa);
2140 blkptr_t *bp = zio->io_bp;
2143 if (zio->io_gang_leader == NULL) {
2144 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2145 zio->io_gang_leader = zio;
2148 ASSERT(BP_IS_HOLE(bp));
2149 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2150 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2151 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2152 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2154 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2155 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2158 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2159 return (zio_write_gang_block(zio));
2160 zio->io_error = error;
2163 return (ZIO_PIPELINE_CONTINUE);
2167 zio_dva_free(zio_t *zio)
2169 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2171 return (ZIO_PIPELINE_CONTINUE);
2175 zio_dva_claim(zio_t *zio)
2179 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2181 zio->io_error = error;
2183 return (ZIO_PIPELINE_CONTINUE);
2187 * Undo an allocation. This is used by zio_done() when an I/O fails
2188 * and we want to give back the block we just allocated.
2189 * This handles both normal blocks and gang blocks.
2192 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2196 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2197 ASSERT(zio->io_bp_override == NULL);
2199 if (!BP_IS_HOLE(bp))
2200 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2203 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2204 zio_dva_unallocate(zio, gn->gn_child[g],
2205 &gn->gn_gbh->zg_blkptr[g]);
2211 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2214 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2215 uint64_t size, boolean_t use_slog)
2219 ASSERT(txg > spa_syncing_txg(spa));
2222 error = metaslab_alloc(spa, spa_log_class(spa), size,
2223 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2226 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2227 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2230 BP_SET_LSIZE(new_bp, size);
2231 BP_SET_PSIZE(new_bp, size);
2232 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2233 BP_SET_CHECKSUM(new_bp,
2234 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2235 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2236 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2237 BP_SET_LEVEL(new_bp, 0);
2238 BP_SET_DEDUP(new_bp, 0);
2239 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2246 * Free an intent log block.
2249 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2251 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2252 ASSERT(!BP_IS_GANG(bp));
2254 zio_free(spa, txg, bp);
2258 * ==========================================================================
2259 * Read and write to physical devices
2260 * ==========================================================================
2263 zio_vdev_io_start(zio_t *zio)
2265 vdev_t *vd = zio->io_vd;
2267 spa_t *spa = zio->io_spa;
2269 ASSERT(zio->io_error == 0);
2270 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2273 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2274 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2277 * The mirror_ops handle multiple DVAs in a single BP.
2279 return (vdev_mirror_ops.vdev_op_io_start(zio));
2283 * We keep track of time-sensitive I/Os so that the scan thread
2284 * can quickly react to certain workloads. In particular, we care
2285 * about non-scrubbing, top-level reads and writes with the following
2287 * - synchronous writes of user data to non-slog devices
2288 * - any reads of user data
2289 * When these conditions are met, adjust the timestamp of spa_last_io
2290 * which allows the scan thread to adjust its workload accordingly.
2292 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2293 vd == vd->vdev_top && !vd->vdev_islog &&
2294 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2295 zio->io_txg != spa_syncing_txg(spa)) {
2296 uint64_t old = spa->spa_last_io;
2297 uint64_t new = ddi_get_lbolt64();
2299 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2302 align = 1ULL << vd->vdev_top->vdev_ashift;
2304 if (P2PHASE(zio->io_size, align) != 0) {
2305 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2306 char *abuf = zio_buf_alloc(asize);
2307 ASSERT(vd == vd->vdev_top);
2308 if (zio->io_type == ZIO_TYPE_WRITE) {
2309 bcopy(zio->io_data, abuf, zio->io_size);
2310 bzero(abuf + zio->io_size, asize - zio->io_size);
2312 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2315 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2316 ASSERT(P2PHASE(zio->io_size, align) == 0);
2317 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2320 * If this is a repair I/O, and there's no self-healing involved --
2321 * that is, we're just resilvering what we expect to resilver --
2322 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2323 * This prevents spurious resilvering with nested replication.
2324 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2325 * A is out of date, we'll read from C+D, then use the data to
2326 * resilver A+B -- but we don't actually want to resilver B, just A.
2327 * The top-level mirror has no way to know this, so instead we just
2328 * discard unnecessary repairs as we work our way down the vdev tree.
2329 * The same logic applies to any form of nested replication:
2330 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2332 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2333 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2334 zio->io_txg != 0 && /* not a delegated i/o */
2335 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2336 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2337 zio_vdev_io_bypass(zio);
2338 return (ZIO_PIPELINE_CONTINUE);
2341 if (vd->vdev_ops->vdev_op_leaf &&
2342 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2344 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2345 return (ZIO_PIPELINE_CONTINUE);
2347 if ((zio = vdev_queue_io(zio)) == NULL)
2348 return (ZIO_PIPELINE_STOP);
2350 if (!vdev_accessible(vd, zio)) {
2351 zio->io_error = ENXIO;
2353 return (ZIO_PIPELINE_STOP);
2357 return (vd->vdev_ops->vdev_op_io_start(zio));
2361 zio_vdev_io_done(zio_t *zio)
2363 vdev_t *vd = zio->io_vd;
2364 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2365 boolean_t unexpected_error = B_FALSE;
2367 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2368 return (ZIO_PIPELINE_STOP);
2370 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2372 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2374 vdev_queue_io_done(zio);
2376 if (zio->io_type == ZIO_TYPE_WRITE)
2377 vdev_cache_write(zio);
2379 if (zio_injection_enabled && zio->io_error == 0)
2380 zio->io_error = zio_handle_device_injection(vd,
2383 if (zio_injection_enabled && zio->io_error == 0)
2384 zio->io_error = zio_handle_label_injection(zio, EIO);
2386 if (zio->io_error) {
2387 if (!vdev_accessible(vd, zio)) {
2388 zio->io_error = ENXIO;
2390 unexpected_error = B_TRUE;
2395 ops->vdev_op_io_done(zio);
2397 if (unexpected_error)
2398 VERIFY(vdev_probe(vd, zio) == NULL);
2400 return (ZIO_PIPELINE_CONTINUE);
2404 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2405 * disk, and use that to finish the checksum ereport later.
2408 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2409 const void *good_buf)
2411 /* no processing needed */
2412 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2417 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2419 void *buf = zio_buf_alloc(zio->io_size);
2421 bcopy(zio->io_data, buf, zio->io_size);
2423 zcr->zcr_cbinfo = zio->io_size;
2424 zcr->zcr_cbdata = buf;
2425 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2426 zcr->zcr_free = zio_buf_free;
2430 zio_vdev_io_assess(zio_t *zio)
2432 vdev_t *vd = zio->io_vd;
2434 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2435 return (ZIO_PIPELINE_STOP);
2437 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2438 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2440 if (zio->io_vsd != NULL) {
2441 zio->io_vsd_ops->vsd_free(zio);
2445 if (zio_injection_enabled && zio->io_error == 0)
2446 zio->io_error = zio_handle_fault_injection(zio, EIO);
2449 * If the I/O failed, determine whether we should attempt to retry it.
2451 * On retry, we cut in line in the issue queue, since we don't want
2452 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2454 if (zio->io_error && vd == NULL &&
2455 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2456 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2457 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2459 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2460 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2461 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2462 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2463 zio_requeue_io_start_cut_in_line);
2464 return (ZIO_PIPELINE_STOP);
2468 * If we got an error on a leaf device, convert it to ENXIO
2469 * if the device is not accessible at all.
2471 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2472 !vdev_accessible(vd, zio))
2473 zio->io_error = ENXIO;
2476 * If we can't write to an interior vdev (mirror or RAID-Z),
2477 * set vdev_cant_write so that we stop trying to allocate from it.
2479 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2480 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2481 vd->vdev_cant_write = B_TRUE;
2484 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2486 return (ZIO_PIPELINE_CONTINUE);
2490 zio_vdev_io_reissue(zio_t *zio)
2492 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2493 ASSERT(zio->io_error == 0);
2495 zio->io_stage >>= 1;
2499 zio_vdev_io_redone(zio_t *zio)
2501 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2503 zio->io_stage >>= 1;
2507 zio_vdev_io_bypass(zio_t *zio)
2509 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2510 ASSERT(zio->io_error == 0);
2512 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2513 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2517 * ==========================================================================
2518 * Generate and verify checksums
2519 * ==========================================================================
2522 zio_checksum_generate(zio_t *zio)
2524 blkptr_t *bp = zio->io_bp;
2525 enum zio_checksum checksum;
2529 * This is zio_write_phys().
2530 * We're either generating a label checksum, or none at all.
2532 checksum = zio->io_prop.zp_checksum;
2534 if (checksum == ZIO_CHECKSUM_OFF)
2535 return (ZIO_PIPELINE_CONTINUE);
2537 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2539 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2540 ASSERT(!IO_IS_ALLOCATING(zio));
2541 checksum = ZIO_CHECKSUM_GANG_HEADER;
2543 checksum = BP_GET_CHECKSUM(bp);
2547 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2549 return (ZIO_PIPELINE_CONTINUE);
2553 zio_checksum_verify(zio_t *zio)
2555 zio_bad_cksum_t info;
2556 blkptr_t *bp = zio->io_bp;
2559 ASSERT(zio->io_vd != NULL);
2563 * This is zio_read_phys().
2564 * We're either verifying a label checksum, or nothing at all.
2566 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2567 return (ZIO_PIPELINE_CONTINUE);
2569 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2572 if ((error = zio_checksum_error(zio, &info)) != 0) {
2573 zio->io_error = error;
2574 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2575 zfs_ereport_start_checksum(zio->io_spa,
2576 zio->io_vd, zio, zio->io_offset,
2577 zio->io_size, NULL, &info);
2581 return (ZIO_PIPELINE_CONTINUE);
2585 * Called by RAID-Z to ensure we don't compute the checksum twice.
2588 zio_checksum_verified(zio_t *zio)
2590 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2594 * ==========================================================================
2595 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2596 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2597 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2598 * indicate errors that are specific to one I/O, and most likely permanent.
2599 * Any other error is presumed to be worse because we weren't expecting it.
2600 * ==========================================================================
2603 zio_worst_error(int e1, int e2)
2605 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2608 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2609 if (e1 == zio_error_rank[r1])
2612 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2613 if (e2 == zio_error_rank[r2])
2616 return (r1 > r2 ? e1 : e2);
2620 * ==========================================================================
2622 * ==========================================================================
2625 zio_ready(zio_t *zio)
2627 blkptr_t *bp = zio->io_bp;
2628 zio_t *pio, *pio_next;
2630 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2631 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2632 return (ZIO_PIPELINE_STOP);
2634 if (zio->io_ready) {
2635 ASSERT(IO_IS_ALLOCATING(zio));
2636 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2637 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2642 if (bp != NULL && bp != &zio->io_bp_copy)
2643 zio->io_bp_copy = *bp;
2646 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2648 mutex_enter(&zio->io_lock);
2649 zio->io_state[ZIO_WAIT_READY] = 1;
2650 pio = zio_walk_parents(zio);
2651 mutex_exit(&zio->io_lock);
2654 * As we notify zio's parents, new parents could be added.
2655 * New parents go to the head of zio's io_parent_list, however,
2656 * so we will (correctly) not notify them. The remainder of zio's
2657 * io_parent_list, from 'pio_next' onward, cannot change because
2658 * all parents must wait for us to be done before they can be done.
2660 for (; pio != NULL; pio = pio_next) {
2661 pio_next = zio_walk_parents(zio);
2662 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2665 if (zio->io_flags & ZIO_FLAG_NODATA) {
2666 if (BP_IS_GANG(bp)) {
2667 zio->io_flags &= ~ZIO_FLAG_NODATA;
2669 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2670 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2674 if (zio_injection_enabled &&
2675 zio->io_spa->spa_syncing_txg == zio->io_txg)
2676 zio_handle_ignored_writes(zio);
2678 return (ZIO_PIPELINE_CONTINUE);
2682 zio_done(zio_t *zio)
2684 zio_t *pio, *pio_next;
2688 * If our children haven't all completed,
2689 * wait for them and then repeat this pipeline stage.
2691 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2692 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2693 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2694 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2695 return (ZIO_PIPELINE_STOP);
2697 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2698 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2699 ASSERT(zio->io_children[c][w] == 0);
2701 if (zio->io_bp != NULL) {
2702 ASSERT(zio->io_bp->blk_pad[0] == 0);
2703 ASSERT(zio->io_bp->blk_pad[1] == 0);
2704 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2705 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2706 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2707 zio->io_bp_override == NULL &&
2708 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2709 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2710 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2711 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2712 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2717 * If there were child vdev/gang/ddt errors, they apply to us now.
2719 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2720 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2721 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2724 * If the I/O on the transformed data was successful, generate any
2725 * checksum reports now while we still have the transformed data.
2727 if (zio->io_error == 0) {
2728 while (zio->io_cksum_report != NULL) {
2729 zio_cksum_report_t *zcr = zio->io_cksum_report;
2730 uint64_t align = zcr->zcr_align;
2731 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2732 char *abuf = zio->io_data;
2734 if (asize != zio->io_size) {
2735 abuf = zio_buf_alloc(asize);
2736 bcopy(zio->io_data, abuf, zio->io_size);
2737 bzero(abuf + zio->io_size, asize - zio->io_size);
2740 zio->io_cksum_report = zcr->zcr_next;
2741 zcr->zcr_next = NULL;
2742 zcr->zcr_finish(zcr, abuf);
2743 zfs_ereport_free_checksum(zcr);
2745 if (asize != zio->io_size)
2746 zio_buf_free(abuf, asize);
2750 zio_pop_transforms(zio); /* note: may set zio->io_error */
2752 vdev_stat_update(zio, zio->io_size);
2754 if (zio->io_error) {
2756 * If this I/O is attached to a particular vdev,
2757 * generate an error message describing the I/O failure
2758 * at the block level. We ignore these errors if the
2759 * device is currently unavailable.
2761 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2762 !vdev_is_dead(zio->io_vd))
2763 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2764 zio->io_vd, zio, 0, 0);
2766 if ((zio->io_error == EIO || !(zio->io_flags &
2767 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2768 zio == zio->io_logical) {
2770 * For logical I/O requests, tell the SPA to log the
2771 * error and generate a logical data ereport.
2773 spa_log_error(zio->io_spa, zio);
2774 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2779 if (zio->io_error && zio == zio->io_logical) {
2781 * Determine whether zio should be reexecuted. This will
2782 * propagate all the way to the root via zio_notify_parent().
2784 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2785 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2787 if (IO_IS_ALLOCATING(zio) &&
2788 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2789 if (zio->io_error != ENOSPC)
2790 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2792 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2795 if ((zio->io_type == ZIO_TYPE_READ ||
2796 zio->io_type == ZIO_TYPE_FREE) &&
2797 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2798 zio->io_error == ENXIO &&
2799 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2800 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2801 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2803 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2804 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2807 * Here is a possibly good place to attempt to do
2808 * either combinatorial reconstruction or error correction
2809 * based on checksums. It also might be a good place
2810 * to send out preliminary ereports before we suspend
2816 * If there were logical child errors, they apply to us now.
2817 * We defer this until now to avoid conflating logical child
2818 * errors with errors that happened to the zio itself when
2819 * updating vdev stats and reporting FMA events above.
2821 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2823 if ((zio->io_error || zio->io_reexecute) &&
2824 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2825 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2826 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2828 zio_gang_tree_free(&zio->io_gang_tree);
2831 * Godfather I/Os should never suspend.
2833 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2834 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2835 zio->io_reexecute = 0;
2837 if (zio->io_reexecute) {
2839 * This is a logical I/O that wants to reexecute.
2841 * Reexecute is top-down. When an i/o fails, if it's not
2842 * the root, it simply notifies its parent and sticks around.
2843 * The parent, seeing that it still has children in zio_done(),
2844 * does the same. This percolates all the way up to the root.
2845 * The root i/o will reexecute or suspend the entire tree.
2847 * This approach ensures that zio_reexecute() honors
2848 * all the original i/o dependency relationships, e.g.
2849 * parents not executing until children are ready.
2851 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2853 zio->io_gang_leader = NULL;
2855 mutex_enter(&zio->io_lock);
2856 zio->io_state[ZIO_WAIT_DONE] = 1;
2857 mutex_exit(&zio->io_lock);
2860 * "The Godfather" I/O monitors its children but is
2861 * not a true parent to them. It will track them through
2862 * the pipeline but severs its ties whenever they get into
2863 * trouble (e.g. suspended). This allows "The Godfather"
2864 * I/O to return status without blocking.
2866 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2867 zio_link_t *zl = zio->io_walk_link;
2868 pio_next = zio_walk_parents(zio);
2870 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2871 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2872 zio_remove_child(pio, zio, zl);
2873 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2877 if ((pio = zio_unique_parent(zio)) != NULL) {
2879 * We're not a root i/o, so there's nothing to do
2880 * but notify our parent. Don't propagate errors
2881 * upward since we haven't permanently failed yet.
2883 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2884 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2885 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2886 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2888 * We'd fail again if we reexecuted now, so suspend
2889 * until conditions improve (e.g. device comes online).
2891 zio_suspend(zio->io_spa, zio);
2894 * Reexecution is potentially a huge amount of work.
2895 * Hand it off to the otherwise-unused claim taskq.
2897 (void) taskq_dispatch(
2898 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2899 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2901 return (ZIO_PIPELINE_STOP);
2904 ASSERT(zio->io_child_count == 0);
2905 ASSERT(zio->io_reexecute == 0);
2906 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2909 * Report any checksum errors, since the I/O is complete.
2911 while (zio->io_cksum_report != NULL) {
2912 zio_cksum_report_t *zcr = zio->io_cksum_report;
2913 zio->io_cksum_report = zcr->zcr_next;
2914 zcr->zcr_next = NULL;
2915 zcr->zcr_finish(zcr, NULL);
2916 zfs_ereport_free_checksum(zcr);
2920 * It is the responsibility of the done callback to ensure that this
2921 * particular zio is no longer discoverable for adoption, and as
2922 * such, cannot acquire any new parents.
2927 mutex_enter(&zio->io_lock);
2928 zio->io_state[ZIO_WAIT_DONE] = 1;
2929 mutex_exit(&zio->io_lock);
2931 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2932 zio_link_t *zl = zio->io_walk_link;
2933 pio_next = zio_walk_parents(zio);
2934 zio_remove_child(pio, zio, zl);
2935 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2938 if (zio->io_waiter != NULL) {
2939 mutex_enter(&zio->io_lock);
2940 zio->io_executor = NULL;
2941 cv_broadcast(&zio->io_cv);
2942 mutex_exit(&zio->io_lock);
2947 return (ZIO_PIPELINE_STOP);
2951 * ==========================================================================
2952 * I/O pipeline definition
2953 * ==========================================================================
2955 static zio_pipe_stage_t *zio_pipeline[] = {
2961 zio_checksum_generate,
2975 zio_checksum_verify,