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>
38 SYSCTL_DECL(_vfs_zfs);
39 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
40 static int zio_use_uma = 0;
41 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
42 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
43 "Use uma(9) for ZIO allocations");
46 * ==========================================================================
48 * ==========================================================================
50 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
51 0, /* ZIO_PRIORITY_NOW */
52 0, /* ZIO_PRIORITY_SYNC_READ */
53 0, /* ZIO_PRIORITY_SYNC_WRITE */
54 0, /* ZIO_PRIORITY_LOG_WRITE */
55 1, /* ZIO_PRIORITY_CACHE_FILL */
56 1, /* ZIO_PRIORITY_AGG */
57 4, /* ZIO_PRIORITY_FREE */
58 4, /* ZIO_PRIORITY_ASYNC_WRITE */
59 6, /* ZIO_PRIORITY_ASYNC_READ */
60 10, /* ZIO_PRIORITY_RESILVER */
61 20, /* ZIO_PRIORITY_SCRUB */
62 2, /* ZIO_PRIORITY_DDT_PREFETCH */
66 * ==========================================================================
67 * I/O type descriptions
68 * ==========================================================================
70 char *zio_type_name[ZIO_TYPES] = {
71 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
76 * ==========================================================================
78 * ==========================================================================
80 kmem_cache_t *zio_cache;
81 kmem_cache_t *zio_link_cache;
82 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
86 extern vmem_t *zio_alloc_arena;
90 * An allocating zio is one that either currently has the DVA allocate
91 * stage set or will have it later in its lifetime.
93 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
95 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
98 int zio_buf_debug_limit = 16384;
100 int zio_buf_debug_limit = 0;
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;
122 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
124 while (p2 & (p2 - 1))
127 if (size <= 4 * SPA_MINBLOCKSIZE) {
128 align = SPA_MINBLOCKSIZE;
129 } else if (P2PHASE(size, PAGESIZE) == 0) {
131 } else if (P2PHASE(size, p2 >> 2) == 0) {
137 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
138 zio_buf_cache[c] = kmem_cache_create(name, size,
139 align, NULL, NULL, NULL, NULL, NULL, cflags);
142 * Since zio_data bufs do not appear in crash dumps, we
143 * pass KMC_NOTOUCH so that no allocator metadata is
144 * stored with the buffers.
146 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
147 zio_data_buf_cache[c] = kmem_cache_create(name, size,
148 align, NULL, NULL, NULL, NULL, NULL,
149 cflags | KMC_NOTOUCH);
154 ASSERT(zio_buf_cache[c] != NULL);
155 if (zio_buf_cache[c - 1] == NULL)
156 zio_buf_cache[c - 1] = zio_buf_cache[c];
158 ASSERT(zio_data_buf_cache[c] != NULL);
159 if (zio_data_buf_cache[c - 1] == NULL)
160 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
170 kmem_cache_t *last_cache = NULL;
171 kmem_cache_t *last_data_cache = NULL;
173 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
174 if (zio_buf_cache[c] != last_cache) {
175 last_cache = zio_buf_cache[c];
176 kmem_cache_destroy(zio_buf_cache[c]);
178 zio_buf_cache[c] = NULL;
180 if (zio_data_buf_cache[c] != last_data_cache) {
181 last_data_cache = zio_data_buf_cache[c];
182 kmem_cache_destroy(zio_data_buf_cache[c]);
184 zio_data_buf_cache[c] = NULL;
187 kmem_cache_destroy(zio_link_cache);
188 kmem_cache_destroy(zio_cache);
194 * ==========================================================================
195 * Allocate and free I/O buffers
196 * ==========================================================================
200 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
201 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
202 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
203 * excess / transient data in-core during a crashdump.
206 zio_buf_alloc(size_t size)
208 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
210 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
213 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
215 return (kmem_alloc(size, KM_SLEEP));
219 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
220 * crashdump if the kernel panics. This exists so that we will limit the amount
221 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
222 * of kernel heap dumped to disk when the kernel panics)
225 zio_data_buf_alloc(size_t size)
227 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
229 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
232 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
234 return (kmem_alloc(size, KM_SLEEP));
238 zio_buf_free(void *buf, size_t size)
240 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245 kmem_cache_free(zio_buf_cache[c], buf);
247 kmem_free(buf, size);
251 zio_data_buf_free(void *buf, size_t size)
253 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
255 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
258 kmem_cache_free(zio_data_buf_cache[c], buf);
260 kmem_free(buf, size);
264 * ==========================================================================
265 * Push and pop I/O transform buffers
266 * ==========================================================================
269 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
270 zio_transform_func_t *transform)
272 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
274 zt->zt_orig_data = zio->io_data;
275 zt->zt_orig_size = zio->io_size;
276 zt->zt_bufsize = bufsize;
277 zt->zt_transform = transform;
279 zt->zt_next = zio->io_transform_stack;
280 zio->io_transform_stack = zt;
287 zio_pop_transforms(zio_t *zio)
291 while ((zt = zio->io_transform_stack) != NULL) {
292 if (zt->zt_transform != NULL)
293 zt->zt_transform(zio,
294 zt->zt_orig_data, zt->zt_orig_size);
296 if (zt->zt_bufsize != 0)
297 zio_buf_free(zio->io_data, zt->zt_bufsize);
299 zio->io_data = zt->zt_orig_data;
300 zio->io_size = zt->zt_orig_size;
301 zio->io_transform_stack = zt->zt_next;
303 kmem_free(zt, sizeof (zio_transform_t));
308 * ==========================================================================
309 * I/O transform callbacks for subblocks and decompression
310 * ==========================================================================
313 zio_subblock(zio_t *zio, void *data, uint64_t size)
315 ASSERT(zio->io_size > size);
317 if (zio->io_type == ZIO_TYPE_READ)
318 bcopy(zio->io_data, data, size);
322 zio_decompress(zio_t *zio, void *data, uint64_t size)
324 if (zio->io_error == 0 &&
325 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
326 zio->io_data, data, zio->io_size, size) != 0)
331 * ==========================================================================
332 * I/O parent/child relationships and pipeline interlocks
333 * ==========================================================================
336 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
337 * continue calling these functions until they return NULL.
338 * Otherwise, the next caller will pick up the list walk in
339 * some indeterminate state. (Otherwise every caller would
340 * have to pass in a cookie to keep the state represented by
341 * io_walk_link, which gets annoying.)
344 zio_walk_parents(zio_t *cio)
346 zio_link_t *zl = cio->io_walk_link;
347 list_t *pl = &cio->io_parent_list;
349 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
350 cio->io_walk_link = zl;
355 ASSERT(zl->zl_child == cio);
356 return (zl->zl_parent);
360 zio_walk_children(zio_t *pio)
362 zio_link_t *zl = pio->io_walk_link;
363 list_t *cl = &pio->io_child_list;
365 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
366 pio->io_walk_link = zl;
371 ASSERT(zl->zl_parent == pio);
372 return (zl->zl_child);
376 zio_unique_parent(zio_t *cio)
378 zio_t *pio = zio_walk_parents(cio);
380 VERIFY(zio_walk_parents(cio) == NULL);
385 zio_add_child(zio_t *pio, zio_t *cio)
387 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
390 * Logical I/Os can have logical, gang, or vdev children.
391 * Gang I/Os can have gang or vdev children.
392 * Vdev I/Os can only have vdev children.
393 * The following ASSERT captures all of these constraints.
395 ASSERT(cio->io_child_type <= pio->io_child_type);
400 mutex_enter(&cio->io_lock);
401 mutex_enter(&pio->io_lock);
403 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
405 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
406 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
408 list_insert_head(&pio->io_child_list, zl);
409 list_insert_head(&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);
419 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
421 ASSERT(zl->zl_parent == pio);
422 ASSERT(zl->zl_child == cio);
424 mutex_enter(&cio->io_lock);
425 mutex_enter(&pio->io_lock);
427 list_remove(&pio->io_child_list, zl);
428 list_remove(&cio->io_parent_list, zl);
430 pio->io_child_count--;
431 cio->io_parent_count--;
433 mutex_exit(&pio->io_lock);
434 mutex_exit(&cio->io_lock);
436 kmem_cache_free(zio_link_cache, zl);
440 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
442 uint64_t *countp = &zio->io_children[child][wait];
443 boolean_t waiting = B_FALSE;
445 mutex_enter(&zio->io_lock);
446 ASSERT(zio->io_stall == NULL);
449 zio->io_stall = countp;
452 mutex_exit(&zio->io_lock);
458 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
460 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
461 int *errorp = &pio->io_child_error[zio->io_child_type];
463 mutex_enter(&pio->io_lock);
464 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
465 *errorp = zio_worst_error(*errorp, zio->io_error);
466 pio->io_reexecute |= zio->io_reexecute;
467 ASSERT3U(*countp, >, 0);
468 if (--*countp == 0 && pio->io_stall == countp) {
469 pio->io_stall = NULL;
470 mutex_exit(&pio->io_lock);
473 mutex_exit(&pio->io_lock);
478 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
480 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
481 zio->io_error = zio->io_child_error[c];
485 * ==========================================================================
486 * Create the various types of I/O (read, write, free, etc)
487 * ==========================================================================
490 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
491 void *data, uint64_t size, zio_done_func_t *done, void *private,
492 zio_type_t type, int priority, enum zio_flag flags,
493 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
494 enum zio_stage stage, enum zio_stage pipeline)
498 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
499 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
500 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
502 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
503 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
504 ASSERT(vd || stage == ZIO_STAGE_OPEN);
506 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
507 bzero(zio, sizeof (zio_t));
509 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
510 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
512 list_create(&zio->io_parent_list, sizeof (zio_link_t),
513 offsetof(zio_link_t, zl_parent_node));
514 list_create(&zio->io_child_list, sizeof (zio_link_t),
515 offsetof(zio_link_t, zl_child_node));
518 zio->io_child_type = ZIO_CHILD_VDEV;
519 else if (flags & ZIO_FLAG_GANG_CHILD)
520 zio->io_child_type = ZIO_CHILD_GANG;
521 else if (flags & ZIO_FLAG_DDT_CHILD)
522 zio->io_child_type = ZIO_CHILD_DDT;
524 zio->io_child_type = ZIO_CHILD_LOGICAL;
527 zio->io_bp = (blkptr_t *)bp;
528 zio->io_bp_copy = *bp;
529 zio->io_bp_orig = *bp;
530 if (type != ZIO_TYPE_WRITE ||
531 zio->io_child_type == ZIO_CHILD_DDT)
532 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
533 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
534 zio->io_logical = zio;
535 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
536 pipeline |= ZIO_GANG_STAGES;
542 zio->io_private = private;
544 zio->io_priority = priority;
546 zio->io_offset = offset;
547 zio->io_orig_data = zio->io_data = data;
548 zio->io_orig_size = zio->io_size = size;
549 zio->io_orig_flags = zio->io_flags = flags;
550 zio->io_orig_stage = zio->io_stage = stage;
551 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
553 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
554 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
557 zio->io_bookmark = *zb;
560 if (zio->io_logical == NULL)
561 zio->io_logical = pio->io_logical;
562 if (zio->io_child_type == ZIO_CHILD_GANG)
563 zio->io_gang_leader = pio->io_gang_leader;
564 zio_add_child(pio, zio);
571 zio_destroy(zio_t *zio)
573 list_destroy(&zio->io_parent_list);
574 list_destroy(&zio->io_child_list);
575 mutex_destroy(&zio->io_lock);
576 cv_destroy(&zio->io_cv);
577 kmem_cache_free(zio_cache, zio);
581 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
582 void *private, enum zio_flag flags)
586 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
587 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
588 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
594 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
596 return (zio_null(NULL, spa, NULL, done, private, flags));
600 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
601 void *data, uint64_t size, zio_done_func_t *done, void *private,
602 int priority, enum zio_flag flags, const zbookmark_t *zb)
606 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
607 data, size, done, private,
608 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
609 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
610 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
616 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
617 void *data, uint64_t size, const zio_prop_t *zp,
618 zio_done_func_t *ready, zio_done_func_t *done, void *private,
619 int priority, enum zio_flag flags, const zbookmark_t *zb)
623 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
624 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
625 zp->zp_compress >= ZIO_COMPRESS_OFF &&
626 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
627 zp->zp_type < DMU_OT_NUMTYPES &&
630 zp->zp_copies <= spa_max_replication(spa) &&
632 zp->zp_dedup_verify <= 1);
634 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
635 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
636 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
637 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
639 zio->io_ready = ready;
646 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
647 uint64_t size, zio_done_func_t *done, void *private, int priority,
648 enum zio_flag flags, zbookmark_t *zb)
652 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
653 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
654 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
660 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
662 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
663 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
664 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
665 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
667 zio->io_prop.zp_copies = copies;
668 zio->io_bp_override = bp;
672 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
674 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
678 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
683 dprintf_bp(bp, "freeing in txg %llu, pass %u",
684 (longlong_t)txg, spa->spa_sync_pass);
686 ASSERT(!BP_IS_HOLE(bp));
687 ASSERT(spa_syncing_txg(spa) == txg);
688 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
690 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
691 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
692 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
698 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
699 zio_done_func_t *done, void *private, enum zio_flag flags)
704 * A claim is an allocation of a specific block. Claims are needed
705 * to support immediate writes in the intent log. The issue is that
706 * immediate writes contain committed data, but in a txg that was
707 * *not* committed. Upon opening the pool after an unclean shutdown,
708 * the intent log claims all blocks that contain immediate write data
709 * so that the SPA knows they're in use.
711 * All claims *must* be resolved in the first txg -- before the SPA
712 * starts allocating blocks -- so that nothing is allocated twice.
713 * If txg == 0 we just verify that the block is claimable.
715 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
716 ASSERT(txg == spa_first_txg(spa) || txg == 0);
717 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
719 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
720 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
721 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
727 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
728 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
733 if (vd->vdev_children == 0) {
734 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
735 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
736 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
740 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
742 for (c = 0; c < vd->vdev_children; c++)
743 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
744 done, private, priority, flags));
751 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
752 void *data, int checksum, zio_done_func_t *done, void *private,
753 int priority, enum zio_flag flags, boolean_t labels)
757 ASSERT(vd->vdev_children == 0);
758 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
759 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
760 ASSERT3U(offset + size, <=, vd->vdev_psize);
762 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
763 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
764 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
766 zio->io_prop.zp_checksum = checksum;
772 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
773 void *data, int checksum, zio_done_func_t *done, void *private,
774 int priority, enum zio_flag flags, boolean_t labels)
778 ASSERT(vd->vdev_children == 0);
779 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
780 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
781 ASSERT3U(offset + size, <=, vd->vdev_psize);
783 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
784 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
785 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
787 zio->io_prop.zp_checksum = checksum;
789 if (zio_checksum_table[checksum].ci_eck) {
791 * zec checksums are necessarily destructive -- they modify
792 * the end of the write buffer to hold the verifier/checksum.
793 * Therefore, we must make a local copy in case the data is
794 * being written to multiple places in parallel.
796 void *wbuf = zio_buf_alloc(size);
797 bcopy(data, wbuf, size);
798 zio_push_transform(zio, wbuf, size, size, NULL);
805 * Create a child I/O to do some work for us.
808 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
809 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
810 zio_done_func_t *done, void *private)
812 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
815 ASSERT(vd->vdev_parent ==
816 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
818 if (type == ZIO_TYPE_READ && bp != NULL) {
820 * If we have the bp, then the child should perform the
821 * checksum and the parent need not. This pushes error
822 * detection as close to the leaves as possible and
823 * eliminates redundant checksums in the interior nodes.
825 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
826 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
829 if (vd->vdev_children == 0)
830 offset += VDEV_LABEL_START_SIZE;
832 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
835 * If we've decided to do a repair, the write is not speculative --
836 * even if the original read was.
838 if (flags & ZIO_FLAG_IO_REPAIR)
839 flags &= ~ZIO_FLAG_SPECULATIVE;
841 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
842 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
843 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
849 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
850 int type, int priority, enum zio_flag flags,
851 zio_done_func_t *done, void *private)
855 ASSERT(vd->vdev_ops->vdev_op_leaf);
857 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
858 data, size, done, private, type, priority,
859 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
861 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
867 zio_flush(zio_t *zio, vdev_t *vd)
869 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
870 NULL, NULL, ZIO_PRIORITY_NOW,
871 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
875 zio_shrink(zio_t *zio, uint64_t size)
877 ASSERT(zio->io_executor == NULL);
878 ASSERT(zio->io_orig_size == zio->io_size);
879 ASSERT(size <= zio->io_size);
882 * We don't shrink for raidz because of problems with the
883 * reconstruction when reading back less than the block size.
884 * Note, BP_IS_RAIDZ() assumes no compression.
886 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
887 if (!BP_IS_RAIDZ(zio->io_bp))
888 zio->io_orig_size = zio->io_size = size;
892 * ==========================================================================
893 * Prepare to read and write logical blocks
894 * ==========================================================================
898 zio_read_bp_init(zio_t *zio)
900 blkptr_t *bp = zio->io_bp;
902 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
903 zio->io_child_type == ZIO_CHILD_LOGICAL &&
904 !(zio->io_flags & ZIO_FLAG_RAW)) {
905 uint64_t psize = BP_GET_PSIZE(bp);
906 void *cbuf = zio_buf_alloc(psize);
908 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
911 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
912 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
914 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
915 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
917 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
918 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
920 return (ZIO_PIPELINE_CONTINUE);
924 zio_write_bp_init(zio_t *zio)
926 spa_t *spa = zio->io_spa;
927 zio_prop_t *zp = &zio->io_prop;
928 enum zio_compress compress = zp->zp_compress;
929 blkptr_t *bp = zio->io_bp;
930 uint64_t lsize = zio->io_size;
931 uint64_t psize = lsize;
935 * If our children haven't all reached the ready stage,
936 * wait for them and then repeat this pipeline stage.
938 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
939 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
940 return (ZIO_PIPELINE_STOP);
942 if (!IO_IS_ALLOCATING(zio))
943 return (ZIO_PIPELINE_CONTINUE);
945 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
947 if (zio->io_bp_override) {
948 ASSERT(bp->blk_birth != zio->io_txg);
949 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
951 *bp = *zio->io_bp_override;
952 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
954 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
955 return (ZIO_PIPELINE_CONTINUE);
957 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
958 zp->zp_dedup_verify);
960 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
962 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
963 return (ZIO_PIPELINE_CONTINUE);
965 zio->io_bp_override = NULL;
969 if (bp->blk_birth == zio->io_txg) {
971 * We're rewriting an existing block, which means we're
972 * working on behalf of spa_sync(). For spa_sync() to
973 * converge, it must eventually be the case that we don't
974 * have to allocate new blocks. But compression changes
975 * the blocksize, which forces a reallocate, and makes
976 * convergence take longer. Therefore, after the first
977 * few passes, stop compressing to ensure convergence.
979 pass = spa_sync_pass(spa);
981 ASSERT(zio->io_txg == spa_syncing_txg(spa));
982 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
983 ASSERT(!BP_GET_DEDUP(bp));
985 if (pass > SYNC_PASS_DONT_COMPRESS)
986 compress = ZIO_COMPRESS_OFF;
988 /* Make sure someone doesn't change their mind on overwrites */
989 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
990 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
993 if (compress != ZIO_COMPRESS_OFF) {
994 void *cbuf = zio_buf_alloc(lsize);
995 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
996 if (psize == 0 || psize == lsize) {
997 compress = ZIO_COMPRESS_OFF;
998 zio_buf_free(cbuf, lsize);
1000 ASSERT(psize < lsize);
1001 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1006 * The final pass of spa_sync() must be all rewrites, but the first
1007 * few passes offer a trade-off: allocating blocks defers convergence,
1008 * but newly allocated blocks are sequential, so they can be written
1009 * to disk faster. Therefore, we allow the first few passes of
1010 * spa_sync() to allocate new blocks, but force rewrites after that.
1011 * There should only be a handful of blocks after pass 1 in any case.
1013 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1014 pass > SYNC_PASS_REWRITE) {
1016 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1017 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1018 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1021 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1025 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1027 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1028 BP_SET_LSIZE(bp, lsize);
1029 BP_SET_PSIZE(bp, psize);
1030 BP_SET_COMPRESS(bp, compress);
1031 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1032 BP_SET_TYPE(bp, zp->zp_type);
1033 BP_SET_LEVEL(bp, zp->zp_level);
1034 BP_SET_DEDUP(bp, zp->zp_dedup);
1035 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1037 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1038 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1039 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1043 return (ZIO_PIPELINE_CONTINUE);
1047 zio_free_bp_init(zio_t *zio)
1049 blkptr_t *bp = zio->io_bp;
1051 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1052 if (BP_GET_DEDUP(bp))
1053 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1056 return (ZIO_PIPELINE_CONTINUE);
1060 * ==========================================================================
1061 * Execute the I/O pipeline
1062 * ==========================================================================
1066 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1068 spa_t *spa = zio->io_spa;
1069 zio_type_t t = zio->io_type;
1070 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1072 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1075 * If we're a config writer or a probe, the normal issue and
1076 * interrupt threads may all be blocked waiting for the config lock.
1077 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1079 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1083 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1085 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1089 * If this is a high priority I/O, then use the high priority taskq.
1091 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1092 spa->spa_zio_taskq[t][q + 1] != NULL)
1095 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1097 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1098 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1100 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1101 (task_func_t *)zio_execute, zio, flags);
1106 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1108 kthread_t *executor = zio->io_executor;
1109 spa_t *spa = zio->io_spa;
1111 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1112 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1119 zio_issue_async(zio_t *zio)
1121 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1123 return (ZIO_PIPELINE_STOP);
1127 zio_interrupt(zio_t *zio)
1129 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1133 * Execute the I/O pipeline until one of the following occurs:
1134 * (1) the I/O completes; (2) the pipeline stalls waiting for
1135 * dependent child I/Os; (3) the I/O issues, so we're waiting
1136 * for an I/O completion interrupt; (4) the I/O is delegated by
1137 * vdev-level caching or aggregation; (5) the I/O is deferred
1138 * due to vdev-level queueing; (6) the I/O is handed off to
1139 * another thread. In all cases, the pipeline stops whenever
1140 * there's no CPU work; it never burns a thread in cv_wait().
1142 * There's no locking on io_stage because there's no legitimate way
1143 * for multiple threads to be attempting to process the same I/O.
1145 static zio_pipe_stage_t *zio_pipeline[];
1148 zio_execute(zio_t *zio)
1150 zio->io_executor = curthread;
1152 while (zio->io_stage < ZIO_STAGE_DONE) {
1153 enum zio_stage pipeline = zio->io_pipeline;
1154 enum zio_stage stage = zio->io_stage;
1157 ASSERT(!MUTEX_HELD(&zio->io_lock));
1158 ASSERT(ISP2(stage));
1159 ASSERT(zio->io_stall == NULL);
1163 } while ((stage & pipeline) == 0);
1165 ASSERT(stage <= ZIO_STAGE_DONE);
1168 * If we are in interrupt context and this pipeline stage
1169 * will grab a config lock that is held across I/O,
1170 * or may wait for an I/O that needs an interrupt thread
1171 * to complete, issue async to avoid deadlock.
1173 * For VDEV_IO_START, we cut in line so that the io will
1174 * be sent to disk promptly.
1176 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1177 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1178 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1179 zio_requeue_io_start_cut_in_line : B_FALSE;
1180 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1184 zio->io_stage = stage;
1185 rv = zio_pipeline[highbit(stage) - 1](zio);
1187 if (rv == ZIO_PIPELINE_STOP)
1190 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1195 * ==========================================================================
1196 * Initiate I/O, either sync or async
1197 * ==========================================================================
1200 zio_wait(zio_t *zio)
1204 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1205 ASSERT(zio->io_executor == NULL);
1207 zio->io_waiter = curthread;
1211 mutex_enter(&zio->io_lock);
1212 while (zio->io_executor != NULL)
1213 cv_wait(&zio->io_cv, &zio->io_lock);
1214 mutex_exit(&zio->io_lock);
1216 error = zio->io_error;
1223 zio_nowait(zio_t *zio)
1225 ASSERT(zio->io_executor == NULL);
1227 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1228 zio_unique_parent(zio) == NULL) {
1230 * This is a logical async I/O with no parent to wait for it.
1231 * We add it to the spa_async_root_zio "Godfather" I/O which
1232 * will ensure they complete prior to unloading the pool.
1234 spa_t *spa = zio->io_spa;
1236 zio_add_child(spa->spa_async_zio_root, zio);
1243 * ==========================================================================
1244 * Reexecute or suspend/resume failed I/O
1245 * ==========================================================================
1249 zio_reexecute(zio_t *pio)
1251 zio_t *cio, *cio_next;
1253 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1254 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1255 ASSERT(pio->io_gang_leader == NULL);
1256 ASSERT(pio->io_gang_tree == NULL);
1258 pio->io_flags = pio->io_orig_flags;
1259 pio->io_stage = pio->io_orig_stage;
1260 pio->io_pipeline = pio->io_orig_pipeline;
1261 pio->io_reexecute = 0;
1263 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1264 pio->io_state[w] = 0;
1265 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1266 pio->io_child_error[c] = 0;
1268 if (IO_IS_ALLOCATING(pio))
1269 BP_ZERO(pio->io_bp);
1272 * As we reexecute pio's children, new children could be created.
1273 * New children go to the head of pio's io_child_list, however,
1274 * so we will (correctly) not reexecute them. The key is that
1275 * the remainder of pio's io_child_list, from 'cio_next' onward,
1276 * cannot be affected by any side effects of reexecuting 'cio'.
1278 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1279 cio_next = zio_walk_children(pio);
1280 mutex_enter(&pio->io_lock);
1281 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1282 pio->io_children[cio->io_child_type][w]++;
1283 mutex_exit(&pio->io_lock);
1288 * Now that all children have been reexecuted, execute the parent.
1289 * We don't reexecute "The Godfather" I/O here as it's the
1290 * responsibility of the caller to wait on him.
1292 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1297 zio_suspend(spa_t *spa, zio_t *zio)
1299 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1300 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1301 "failure and the failure mode property for this pool "
1302 "is set to panic.", spa_name(spa));
1304 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1306 mutex_enter(&spa->spa_suspend_lock);
1308 if (spa->spa_suspend_zio_root == NULL)
1309 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1310 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1311 ZIO_FLAG_GODFATHER);
1313 spa->spa_suspended = B_TRUE;
1316 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1317 ASSERT(zio != spa->spa_suspend_zio_root);
1318 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1319 ASSERT(zio_unique_parent(zio) == NULL);
1320 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1321 zio_add_child(spa->spa_suspend_zio_root, zio);
1324 mutex_exit(&spa->spa_suspend_lock);
1328 zio_resume(spa_t *spa)
1333 * Reexecute all previously suspended i/o.
1335 mutex_enter(&spa->spa_suspend_lock);
1336 spa->spa_suspended = B_FALSE;
1337 cv_broadcast(&spa->spa_suspend_cv);
1338 pio = spa->spa_suspend_zio_root;
1339 spa->spa_suspend_zio_root = NULL;
1340 mutex_exit(&spa->spa_suspend_lock);
1346 return (zio_wait(pio));
1350 zio_resume_wait(spa_t *spa)
1352 mutex_enter(&spa->spa_suspend_lock);
1353 while (spa_suspended(spa))
1354 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1355 mutex_exit(&spa->spa_suspend_lock);
1359 * ==========================================================================
1362 * A gang block is a collection of small blocks that looks to the DMU
1363 * like one large block. When zio_dva_allocate() cannot find a block
1364 * of the requested size, due to either severe fragmentation or the pool
1365 * being nearly full, it calls zio_write_gang_block() to construct the
1366 * block from smaller fragments.
1368 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1369 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1370 * an indirect block: it's an array of block pointers. It consumes
1371 * only one sector and hence is allocatable regardless of fragmentation.
1372 * The gang header's bps point to its gang members, which hold the data.
1374 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1375 * as the verifier to ensure uniqueness of the SHA256 checksum.
1376 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1377 * not the gang header. This ensures that data block signatures (needed for
1378 * deduplication) are independent of how the block is physically stored.
1380 * Gang blocks can be nested: a gang member may itself be a gang block.
1381 * Thus every gang block is a tree in which root and all interior nodes are
1382 * gang headers, and the leaves are normal blocks that contain user data.
1383 * The root of the gang tree is called the gang leader.
1385 * To perform any operation (read, rewrite, free, claim) on a gang block,
1386 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1387 * in the io_gang_tree field of the original logical i/o by recursively
1388 * reading the gang leader and all gang headers below it. This yields
1389 * an in-core tree containing the contents of every gang header and the
1390 * bps for every constituent of the gang block.
1392 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1393 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1394 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1395 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1396 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1397 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1398 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1399 * of the gang header plus zio_checksum_compute() of the data to update the
1400 * gang header's blk_cksum as described above.
1402 * The two-phase assemble/issue model solves the problem of partial failure --
1403 * what if you'd freed part of a gang block but then couldn't read the
1404 * gang header for another part? Assembling the entire gang tree first
1405 * ensures that all the necessary gang header I/O has succeeded before
1406 * starting the actual work of free, claim, or write. Once the gang tree
1407 * is assembled, free and claim are in-memory operations that cannot fail.
1409 * In the event that a gang write fails, zio_dva_unallocate() walks the
1410 * gang tree to immediately free (i.e. insert back into the space map)
1411 * everything we've allocated. This ensures that we don't get ENOSPC
1412 * errors during repeated suspend/resume cycles due to a flaky device.
1414 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1415 * the gang tree, we won't modify the block, so we can safely defer the free
1416 * (knowing that the block is still intact). If we *can* assemble the gang
1417 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1418 * each constituent bp and we can allocate a new block on the next sync pass.
1420 * In all cases, the gang tree allows complete recovery from partial failure.
1421 * ==========================================================================
1425 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1430 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1431 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1432 &pio->io_bookmark));
1436 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1441 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1442 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1443 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1445 * As we rewrite each gang header, the pipeline will compute
1446 * a new gang block header checksum for it; but no one will
1447 * compute a new data checksum, so we do that here. The one
1448 * exception is the gang leader: the pipeline already computed
1449 * its data checksum because that stage precedes gang assembly.
1450 * (Presently, nothing actually uses interior data checksums;
1451 * this is just good hygiene.)
1453 if (gn != pio->io_gang_leader->io_gang_tree) {
1454 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1455 data, BP_GET_PSIZE(bp));
1458 * If we are here to damage data for testing purposes,
1459 * leave the GBH alone so that we can detect the damage.
1461 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1462 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1464 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1465 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1466 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1474 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1476 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1477 ZIO_GANG_CHILD_FLAGS(pio)));
1482 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1484 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1485 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1488 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1497 static void zio_gang_tree_assemble_done(zio_t *zio);
1499 static zio_gang_node_t *
1500 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1502 zio_gang_node_t *gn;
1504 ASSERT(*gnpp == NULL);
1506 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1507 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1514 zio_gang_node_free(zio_gang_node_t **gnpp)
1516 zio_gang_node_t *gn = *gnpp;
1518 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1519 ASSERT(gn->gn_child[g] == NULL);
1521 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1522 kmem_free(gn, sizeof (*gn));
1527 zio_gang_tree_free(zio_gang_node_t **gnpp)
1529 zio_gang_node_t *gn = *gnpp;
1534 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1535 zio_gang_tree_free(&gn->gn_child[g]);
1537 zio_gang_node_free(gnpp);
1541 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1543 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1545 ASSERT(gio->io_gang_leader == gio);
1546 ASSERT(BP_IS_GANG(bp));
1548 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1549 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1550 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1554 zio_gang_tree_assemble_done(zio_t *zio)
1556 zio_t *gio = zio->io_gang_leader;
1557 zio_gang_node_t *gn = zio->io_private;
1558 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 (int 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;
1587 ASSERT(BP_IS_GANG(bp) == !!gn);
1588 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1589 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1592 * If you're a gang header, your data is in gn->gn_gbh.
1593 * If you're a gang member, your data is in 'data' and gn == NULL.
1595 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1598 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1600 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1601 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1602 if (BP_IS_HOLE(gbp))
1604 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1605 data = (char *)data + BP_GET_PSIZE(gbp);
1609 if (gn == gio->io_gang_tree)
1610 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1617 zio_gang_assemble(zio_t *zio)
1619 blkptr_t *bp = zio->io_bp;
1621 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1622 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1624 zio->io_gang_leader = zio;
1626 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1628 return (ZIO_PIPELINE_CONTINUE);
1632 zio_gang_issue(zio_t *zio)
1634 blkptr_t *bp = zio->io_bp;
1636 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1637 return (ZIO_PIPELINE_STOP);
1639 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1640 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1642 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1643 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1645 zio_gang_tree_free(&zio->io_gang_tree);
1647 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1649 return (ZIO_PIPELINE_CONTINUE);
1653 zio_write_gang_member_ready(zio_t *zio)
1655 zio_t *pio = zio_unique_parent(zio);
1656 zio_t *gio = zio->io_gang_leader;
1657 dva_t *cdva = zio->io_bp->blk_dva;
1658 dva_t *pdva = pio->io_bp->blk_dva;
1661 if (BP_IS_HOLE(zio->io_bp))
1664 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1666 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1667 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1668 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1669 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1670 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1672 mutex_enter(&pio->io_lock);
1673 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1674 ASSERT(DVA_GET_GANG(&pdva[d]));
1675 asize = DVA_GET_ASIZE(&pdva[d]);
1676 asize += DVA_GET_ASIZE(&cdva[d]);
1677 DVA_SET_ASIZE(&pdva[d], asize);
1679 mutex_exit(&pio->io_lock);
1683 zio_write_gang_block(zio_t *pio)
1685 spa_t *spa = pio->io_spa;
1686 blkptr_t *bp = pio->io_bp;
1687 zio_t *gio = pio->io_gang_leader;
1689 zio_gang_node_t *gn, **gnpp;
1690 zio_gbh_phys_t *gbh;
1691 uint64_t txg = pio->io_txg;
1692 uint64_t resid = pio->io_size;
1694 int copies = gio->io_prop.zp_copies;
1695 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1699 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1700 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1701 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1703 pio->io_error = error;
1704 return (ZIO_PIPELINE_CONTINUE);
1708 gnpp = &gio->io_gang_tree;
1710 gnpp = pio->io_private;
1711 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1714 gn = zio_gang_node_alloc(gnpp);
1716 bzero(gbh, SPA_GANGBLOCKSIZE);
1719 * Create the gang header.
1721 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1722 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1725 * Create and nowait the gang children.
1727 for (int g = 0; resid != 0; resid -= lsize, g++) {
1728 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1730 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1732 zp.zp_checksum = gio->io_prop.zp_checksum;
1733 zp.zp_compress = ZIO_COMPRESS_OFF;
1734 zp.zp_type = DMU_OT_NONE;
1736 zp.zp_copies = gio->io_prop.zp_copies;
1738 zp.zp_dedup_verify = 0;
1740 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1741 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1742 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1743 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1744 &pio->io_bookmark));
1748 * Set pio's pipeline to just wait for zio to finish.
1750 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1754 return (ZIO_PIPELINE_CONTINUE);
1758 * ==========================================================================
1760 * ==========================================================================
1763 zio_ddt_child_read_done(zio_t *zio)
1765 blkptr_t *bp = zio->io_bp;
1766 ddt_entry_t *dde = zio->io_private;
1768 zio_t *pio = zio_unique_parent(zio);
1770 mutex_enter(&pio->io_lock);
1771 ddp = ddt_phys_select(dde, bp);
1772 if (zio->io_error == 0)
1773 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1774 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1775 dde->dde_repair_data = zio->io_data;
1777 zio_buf_free(zio->io_data, zio->io_size);
1778 mutex_exit(&pio->io_lock);
1782 zio_ddt_read_start(zio_t *zio)
1784 blkptr_t *bp = zio->io_bp;
1786 ASSERT(BP_GET_DEDUP(bp));
1787 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1788 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1790 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1791 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1792 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1793 ddt_phys_t *ddp = dde->dde_phys;
1794 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1797 ASSERT(zio->io_vsd == NULL);
1800 if (ddp_self == NULL)
1801 return (ZIO_PIPELINE_CONTINUE);
1803 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1804 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1806 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1808 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1809 zio_buf_alloc(zio->io_size), zio->io_size,
1810 zio_ddt_child_read_done, dde, zio->io_priority,
1811 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1812 &zio->io_bookmark));
1814 return (ZIO_PIPELINE_CONTINUE);
1817 zio_nowait(zio_read(zio, zio->io_spa, bp,
1818 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1819 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1821 return (ZIO_PIPELINE_CONTINUE);
1825 zio_ddt_read_done(zio_t *zio)
1827 blkptr_t *bp = zio->io_bp;
1829 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1830 return (ZIO_PIPELINE_STOP);
1832 ASSERT(BP_GET_DEDUP(bp));
1833 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1834 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1836 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1837 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1838 ddt_entry_t *dde = zio->io_vsd;
1840 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1841 return (ZIO_PIPELINE_CONTINUE);
1844 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1845 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1846 return (ZIO_PIPELINE_STOP);
1848 if (dde->dde_repair_data != NULL) {
1849 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1850 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1852 ddt_repair_done(ddt, dde);
1856 ASSERT(zio->io_vsd == NULL);
1858 return (ZIO_PIPELINE_CONTINUE);
1862 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1864 spa_t *spa = zio->io_spa;
1867 * Note: we compare the original data, not the transformed data,
1868 * because when zio->io_bp is an override bp, we will not have
1869 * pushed the I/O transforms. That's an important optimization
1870 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1872 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1873 zio_t *lio = dde->dde_lead_zio[p];
1876 return (lio->io_orig_size != zio->io_orig_size ||
1877 bcmp(zio->io_orig_data, lio->io_orig_data,
1878 zio->io_orig_size) != 0);
1882 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1883 ddt_phys_t *ddp = &dde->dde_phys[p];
1885 if (ddp->ddp_phys_birth != 0) {
1886 arc_buf_t *abuf = NULL;
1887 uint32_t aflags = ARC_WAIT;
1888 blkptr_t blk = *zio->io_bp;
1891 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1895 error = arc_read_nolock(NULL, spa, &blk,
1896 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1897 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1898 &aflags, &zio->io_bookmark);
1901 if (arc_buf_size(abuf) != zio->io_orig_size ||
1902 bcmp(abuf->b_data, zio->io_orig_data,
1903 zio->io_orig_size) != 0)
1905 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1909 return (error != 0);
1917 zio_ddt_child_write_ready(zio_t *zio)
1919 int p = zio->io_prop.zp_copies;
1920 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1921 ddt_entry_t *dde = zio->io_private;
1922 ddt_phys_t *ddp = &dde->dde_phys[p];
1930 ASSERT(dde->dde_lead_zio[p] == zio);
1932 ddt_phys_fill(ddp, zio->io_bp);
1934 while ((pio = zio_walk_parents(zio)) != NULL)
1935 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1941 zio_ddt_child_write_done(zio_t *zio)
1943 int p = zio->io_prop.zp_copies;
1944 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1945 ddt_entry_t *dde = zio->io_private;
1946 ddt_phys_t *ddp = &dde->dde_phys[p];
1950 ASSERT(ddp->ddp_refcnt == 0);
1951 ASSERT(dde->dde_lead_zio[p] == zio);
1952 dde->dde_lead_zio[p] = NULL;
1954 if (zio->io_error == 0) {
1955 while (zio_walk_parents(zio) != NULL)
1956 ddt_phys_addref(ddp);
1958 ddt_phys_clear(ddp);
1965 zio_ddt_ditto_write_done(zio_t *zio)
1967 int p = DDT_PHYS_DITTO;
1968 zio_prop_t *zp = &zio->io_prop;
1969 blkptr_t *bp = zio->io_bp;
1970 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1971 ddt_entry_t *dde = zio->io_private;
1972 ddt_phys_t *ddp = &dde->dde_phys[p];
1973 ddt_key_t *ddk = &dde->dde_key;
1977 ASSERT(ddp->ddp_refcnt == 0);
1978 ASSERT(dde->dde_lead_zio[p] == zio);
1979 dde->dde_lead_zio[p] = NULL;
1981 if (zio->io_error == 0) {
1982 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1983 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1984 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1985 if (ddp->ddp_phys_birth != 0)
1986 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1987 ddt_phys_fill(ddp, bp);
1994 zio_ddt_write(zio_t *zio)
1996 spa_t *spa = zio->io_spa;
1997 blkptr_t *bp = zio->io_bp;
1998 uint64_t txg = zio->io_txg;
1999 zio_prop_t *zp = &zio->io_prop;
2000 int p = zp->zp_copies;
2004 ddt_t *ddt = ddt_select(spa, bp);
2008 ASSERT(BP_GET_DEDUP(bp));
2009 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2010 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2013 dde = ddt_lookup(ddt, bp, B_TRUE);
2014 ddp = &dde->dde_phys[p];
2016 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2018 * If we're using a weak checksum, upgrade to a strong checksum
2019 * and try again. If we're already using a strong checksum,
2020 * we can't resolve it, so just convert to an ordinary write.
2021 * (And automatically e-mail a paper to Nature?)
2023 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2024 zp->zp_checksum = spa_dedup_checksum(spa);
2025 zio_pop_transforms(zio);
2026 zio->io_stage = ZIO_STAGE_OPEN;
2031 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2033 return (ZIO_PIPELINE_CONTINUE);
2036 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2037 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2039 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2040 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2041 zio_prop_t czp = *zp;
2043 czp.zp_copies = ditto_copies;
2046 * If we arrived here with an override bp, we won't have run
2047 * the transform stack, so we won't have the data we need to
2048 * generate a child i/o. So, toss the override bp and restart.
2049 * This is safe, because using the override bp is just an
2050 * optimization; and it's rare, so the cost doesn't matter.
2052 if (zio->io_bp_override) {
2053 zio_pop_transforms(zio);
2054 zio->io_stage = ZIO_STAGE_OPEN;
2055 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2056 zio->io_bp_override = NULL;
2059 return (ZIO_PIPELINE_CONTINUE);
2062 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2063 zio->io_orig_size, &czp, NULL,
2064 zio_ddt_ditto_write_done, dde, zio->io_priority,
2065 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2067 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2068 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2071 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2072 if (ddp->ddp_phys_birth != 0)
2073 ddt_bp_fill(ddp, bp, txg);
2074 if (dde->dde_lead_zio[p] != NULL)
2075 zio_add_child(zio, dde->dde_lead_zio[p]);
2077 ddt_phys_addref(ddp);
2078 } else if (zio->io_bp_override) {
2079 ASSERT(bp->blk_birth == txg);
2080 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2081 ddt_phys_fill(ddp, bp);
2082 ddt_phys_addref(ddp);
2084 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2085 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2086 zio_ddt_child_write_done, dde, zio->io_priority,
2087 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2089 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2090 dde->dde_lead_zio[p] = cio;
2100 return (ZIO_PIPELINE_CONTINUE);
2103 ddt_entry_t *freedde; /* for debugging */
2106 zio_ddt_free(zio_t *zio)
2108 spa_t *spa = zio->io_spa;
2109 blkptr_t *bp = zio->io_bp;
2110 ddt_t *ddt = ddt_select(spa, bp);
2114 ASSERT(BP_GET_DEDUP(bp));
2115 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2118 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2119 ddp = ddt_phys_select(dde, bp);
2120 ddt_phys_decref(ddp);
2123 return (ZIO_PIPELINE_CONTINUE);
2127 * ==========================================================================
2128 * Allocate and free blocks
2129 * ==========================================================================
2132 zio_dva_allocate(zio_t *zio)
2134 spa_t *spa = zio->io_spa;
2135 metaslab_class_t *mc = spa_normal_class(spa);
2136 blkptr_t *bp = zio->io_bp;
2139 if (zio->io_gang_leader == NULL) {
2140 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2141 zio->io_gang_leader = zio;
2144 ASSERT(BP_IS_HOLE(bp));
2145 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2146 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2147 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2148 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2150 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2151 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2154 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2155 return (zio_write_gang_block(zio));
2156 zio->io_error = error;
2159 return (ZIO_PIPELINE_CONTINUE);
2163 zio_dva_free(zio_t *zio)
2165 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2167 return (ZIO_PIPELINE_CONTINUE);
2171 zio_dva_claim(zio_t *zio)
2175 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2177 zio->io_error = error;
2179 return (ZIO_PIPELINE_CONTINUE);
2183 * Undo an allocation. This is used by zio_done() when an I/O fails
2184 * and we want to give back the block we just allocated.
2185 * This handles both normal blocks and gang blocks.
2188 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2190 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2191 ASSERT(zio->io_bp_override == NULL);
2193 if (!BP_IS_HOLE(bp))
2194 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2197 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2198 zio_dva_unallocate(zio, gn->gn_child[g],
2199 &gn->gn_gbh->zg_blkptr[g]);
2205 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2208 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2209 uint64_t size, boolean_t use_slog)
2213 ASSERT(txg > spa_syncing_txg(spa));
2216 error = metaslab_alloc(spa, spa_log_class(spa), size,
2217 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2220 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2221 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2224 BP_SET_LSIZE(new_bp, size);
2225 BP_SET_PSIZE(new_bp, size);
2226 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2227 BP_SET_CHECKSUM(new_bp,
2228 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2229 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2230 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2231 BP_SET_LEVEL(new_bp, 0);
2232 BP_SET_DEDUP(new_bp, 0);
2233 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2240 * Free an intent log block.
2243 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2245 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2246 ASSERT(!BP_IS_GANG(bp));
2248 zio_free(spa, txg, bp);
2252 * ==========================================================================
2253 * Read and write to physical devices
2254 * ==========================================================================
2257 zio_vdev_io_start(zio_t *zio)
2259 vdev_t *vd = zio->io_vd;
2261 spa_t *spa = zio->io_spa;
2263 ASSERT(zio->io_error == 0);
2264 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2267 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2268 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2271 * The mirror_ops handle multiple DVAs in a single BP.
2273 return (vdev_mirror_ops.vdev_op_io_start(zio));
2277 * We keep track of time-sensitive I/Os so that the scan thread
2278 * can quickly react to certain workloads. In particular, we care
2279 * about non-scrubbing, top-level reads and writes with the following
2281 * - synchronous writes of user data to non-slog devices
2282 * - any reads of user data
2283 * When these conditions are met, adjust the timestamp of spa_last_io
2284 * which allows the scan thread to adjust its workload accordingly.
2286 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2287 vd == vd->vdev_top && !vd->vdev_islog &&
2288 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2289 zio->io_txg != spa_syncing_txg(spa)) {
2290 uint64_t old = spa->spa_last_io;
2291 uint64_t new = ddi_get_lbolt64();
2293 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2296 align = 1ULL << vd->vdev_top->vdev_ashift;
2298 if (P2PHASE(zio->io_size, align) != 0) {
2299 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2300 char *abuf = zio_buf_alloc(asize);
2301 ASSERT(vd == vd->vdev_top);
2302 if (zio->io_type == ZIO_TYPE_WRITE) {
2303 bcopy(zio->io_data, abuf, zio->io_size);
2304 bzero(abuf + zio->io_size, asize - zio->io_size);
2306 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2309 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2310 ASSERT(P2PHASE(zio->io_size, align) == 0);
2311 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2314 * If this is a repair I/O, and there's no self-healing involved --
2315 * that is, we're just resilvering what we expect to resilver --
2316 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2317 * This prevents spurious resilvering with nested replication.
2318 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2319 * A is out of date, we'll read from C+D, then use the data to
2320 * resilver A+B -- but we don't actually want to resilver B, just A.
2321 * The top-level mirror has no way to know this, so instead we just
2322 * discard unnecessary repairs as we work our way down the vdev tree.
2323 * The same logic applies to any form of nested replication:
2324 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2326 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2327 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2328 zio->io_txg != 0 && /* not a delegated i/o */
2329 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2330 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2331 zio_vdev_io_bypass(zio);
2332 return (ZIO_PIPELINE_CONTINUE);
2335 if (vd->vdev_ops->vdev_op_leaf &&
2336 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2338 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2339 return (ZIO_PIPELINE_CONTINUE);
2341 if ((zio = vdev_queue_io(zio)) == NULL)
2342 return (ZIO_PIPELINE_STOP);
2344 if (!vdev_accessible(vd, zio)) {
2345 zio->io_error = ENXIO;
2347 return (ZIO_PIPELINE_STOP);
2351 return (vd->vdev_ops->vdev_op_io_start(zio));
2355 zio_vdev_io_done(zio_t *zio)
2357 vdev_t *vd = zio->io_vd;
2358 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2359 boolean_t unexpected_error = B_FALSE;
2361 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2362 return (ZIO_PIPELINE_STOP);
2364 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2366 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2368 vdev_queue_io_done(zio);
2370 if (zio->io_type == ZIO_TYPE_WRITE)
2371 vdev_cache_write(zio);
2373 if (zio_injection_enabled && zio->io_error == 0)
2374 zio->io_error = zio_handle_device_injection(vd,
2377 if (zio_injection_enabled && zio->io_error == 0)
2378 zio->io_error = zio_handle_label_injection(zio, EIO);
2380 if (zio->io_error) {
2381 if (!vdev_accessible(vd, zio)) {
2382 zio->io_error = ENXIO;
2384 unexpected_error = B_TRUE;
2389 ops->vdev_op_io_done(zio);
2391 if (unexpected_error)
2392 VERIFY(vdev_probe(vd, zio) == NULL);
2394 return (ZIO_PIPELINE_CONTINUE);
2398 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2399 * disk, and use that to finish the checksum ereport later.
2402 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2403 const void *good_buf)
2405 /* no processing needed */
2406 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2411 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2413 void *buf = zio_buf_alloc(zio->io_size);
2415 bcopy(zio->io_data, buf, zio->io_size);
2417 zcr->zcr_cbinfo = zio->io_size;
2418 zcr->zcr_cbdata = buf;
2419 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2420 zcr->zcr_free = zio_buf_free;
2424 zio_vdev_io_assess(zio_t *zio)
2426 vdev_t *vd = zio->io_vd;
2428 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2429 return (ZIO_PIPELINE_STOP);
2431 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2432 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2434 if (zio->io_vsd != NULL) {
2435 zio->io_vsd_ops->vsd_free(zio);
2439 if (zio_injection_enabled && zio->io_error == 0)
2440 zio->io_error = zio_handle_fault_injection(zio, EIO);
2443 * If the I/O failed, determine whether we should attempt to retry it.
2445 * On retry, we cut in line in the issue queue, since we don't want
2446 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2448 if (zio->io_error && vd == NULL &&
2449 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2450 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2451 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2453 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2454 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2455 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2456 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2457 zio_requeue_io_start_cut_in_line);
2458 return (ZIO_PIPELINE_STOP);
2462 * If we got an error on a leaf device, convert it to ENXIO
2463 * if the device is not accessible at all.
2465 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2466 !vdev_accessible(vd, zio))
2467 zio->io_error = ENXIO;
2470 * If we can't write to an interior vdev (mirror or RAID-Z),
2471 * set vdev_cant_write so that we stop trying to allocate from it.
2473 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2474 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2475 vd->vdev_cant_write = B_TRUE;
2478 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2480 return (ZIO_PIPELINE_CONTINUE);
2484 zio_vdev_io_reissue(zio_t *zio)
2486 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2487 ASSERT(zio->io_error == 0);
2489 zio->io_stage >>= 1;
2493 zio_vdev_io_redone(zio_t *zio)
2495 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2497 zio->io_stage >>= 1;
2501 zio_vdev_io_bypass(zio_t *zio)
2503 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2504 ASSERT(zio->io_error == 0);
2506 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2507 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2511 * ==========================================================================
2512 * Generate and verify checksums
2513 * ==========================================================================
2516 zio_checksum_generate(zio_t *zio)
2518 blkptr_t *bp = zio->io_bp;
2519 enum zio_checksum checksum;
2523 * This is zio_write_phys().
2524 * We're either generating a label checksum, or none at all.
2526 checksum = zio->io_prop.zp_checksum;
2528 if (checksum == ZIO_CHECKSUM_OFF)
2529 return (ZIO_PIPELINE_CONTINUE);
2531 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2533 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2534 ASSERT(!IO_IS_ALLOCATING(zio));
2535 checksum = ZIO_CHECKSUM_GANG_HEADER;
2537 checksum = BP_GET_CHECKSUM(bp);
2541 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2543 return (ZIO_PIPELINE_CONTINUE);
2547 zio_checksum_verify(zio_t *zio)
2549 zio_bad_cksum_t info;
2550 blkptr_t *bp = zio->io_bp;
2553 ASSERT(zio->io_vd != NULL);
2557 * This is zio_read_phys().
2558 * We're either verifying a label checksum, or nothing at all.
2560 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2561 return (ZIO_PIPELINE_CONTINUE);
2563 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2566 if ((error = zio_checksum_error(zio, &info)) != 0) {
2567 zio->io_error = error;
2568 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2569 zfs_ereport_start_checksum(zio->io_spa,
2570 zio->io_vd, zio, zio->io_offset,
2571 zio->io_size, NULL, &info);
2575 return (ZIO_PIPELINE_CONTINUE);
2579 * Called by RAID-Z to ensure we don't compute the checksum twice.
2582 zio_checksum_verified(zio_t *zio)
2584 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2588 * ==========================================================================
2589 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2590 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2591 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2592 * indicate errors that are specific to one I/O, and most likely permanent.
2593 * Any other error is presumed to be worse because we weren't expecting it.
2594 * ==========================================================================
2597 zio_worst_error(int e1, int e2)
2599 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2602 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2603 if (e1 == zio_error_rank[r1])
2606 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2607 if (e2 == zio_error_rank[r2])
2610 return (r1 > r2 ? e1 : e2);
2614 * ==========================================================================
2616 * ==========================================================================
2619 zio_ready(zio_t *zio)
2621 blkptr_t *bp = zio->io_bp;
2622 zio_t *pio, *pio_next;
2624 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2625 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2626 return (ZIO_PIPELINE_STOP);
2628 if (zio->io_ready) {
2629 ASSERT(IO_IS_ALLOCATING(zio));
2630 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2631 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2636 if (bp != NULL && bp != &zio->io_bp_copy)
2637 zio->io_bp_copy = *bp;
2640 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2642 mutex_enter(&zio->io_lock);
2643 zio->io_state[ZIO_WAIT_READY] = 1;
2644 pio = zio_walk_parents(zio);
2645 mutex_exit(&zio->io_lock);
2648 * As we notify zio's parents, new parents could be added.
2649 * New parents go to the head of zio's io_parent_list, however,
2650 * so we will (correctly) not notify them. The remainder of zio's
2651 * io_parent_list, from 'pio_next' onward, cannot change because
2652 * all parents must wait for us to be done before they can be done.
2654 for (; pio != NULL; pio = pio_next) {
2655 pio_next = zio_walk_parents(zio);
2656 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2659 if (zio->io_flags & ZIO_FLAG_NODATA) {
2660 if (BP_IS_GANG(bp)) {
2661 zio->io_flags &= ~ZIO_FLAG_NODATA;
2663 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2664 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2668 if (zio_injection_enabled &&
2669 zio->io_spa->spa_syncing_txg == zio->io_txg)
2670 zio_handle_ignored_writes(zio);
2672 return (ZIO_PIPELINE_CONTINUE);
2676 zio_done(zio_t *zio)
2678 spa_t *spa = zio->io_spa;
2679 zio_t *lio = zio->io_logical;
2680 blkptr_t *bp = zio->io_bp;
2681 vdev_t *vd = zio->io_vd;
2682 uint64_t psize = zio->io_size;
2683 zio_t *pio, *pio_next;
2686 * If our children haven't all completed,
2687 * wait for them and then repeat this pipeline stage.
2689 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2690 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2691 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2692 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2693 return (ZIO_PIPELINE_STOP);
2695 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2696 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2697 ASSERT(zio->io_children[c][w] == 0);
2700 ASSERT(bp->blk_pad[0] == 0);
2701 ASSERT(bp->blk_pad[1] == 0);
2702 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2703 (bp == zio_unique_parent(zio)->io_bp));
2704 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2705 zio->io_bp_override == NULL &&
2706 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2707 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2708 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2709 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2710 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2715 * If there were child vdev/gang/ddt errors, they apply to us now.
2717 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2718 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2719 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2722 * If the I/O on the transformed data was successful, generate any
2723 * checksum reports now while we still have the transformed data.
2725 if (zio->io_error == 0) {
2726 while (zio->io_cksum_report != NULL) {
2727 zio_cksum_report_t *zcr = zio->io_cksum_report;
2728 uint64_t align = zcr->zcr_align;
2729 uint64_t asize = P2ROUNDUP(psize, align);
2730 char *abuf = zio->io_data;
2732 if (asize != psize) {
2733 abuf = zio_buf_alloc(asize);
2734 bcopy(zio->io_data, abuf, psize);
2735 bzero(abuf + psize, asize - psize);
2738 zio->io_cksum_report = zcr->zcr_next;
2739 zcr->zcr_next = NULL;
2740 zcr->zcr_finish(zcr, abuf);
2741 zfs_ereport_free_checksum(zcr);
2744 zio_buf_free(abuf, asize);
2748 zio_pop_transforms(zio); /* note: may set zio->io_error */
2750 vdev_stat_update(zio, psize);
2752 if (zio->io_error) {
2754 * If this I/O is attached to a particular vdev,
2755 * generate an error message describing the I/O failure
2756 * at the block level. We ignore these errors if the
2757 * device is currently unavailable.
2759 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2760 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2762 if ((zio->io_error == EIO || !(zio->io_flags &
2763 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2766 * For logical I/O requests, tell the SPA to log the
2767 * error and generate a logical data ereport.
2769 spa_log_error(spa, zio);
2770 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2775 if (zio->io_error && zio == lio) {
2777 * Determine whether zio should be reexecuted. This will
2778 * propagate all the way to the root via zio_notify_parent().
2780 ASSERT(vd == NULL && bp != NULL);
2781 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2783 if (IO_IS_ALLOCATING(zio) &&
2784 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2785 if (zio->io_error != ENOSPC)
2786 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2788 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2791 if ((zio->io_type == ZIO_TYPE_READ ||
2792 zio->io_type == ZIO_TYPE_FREE) &&
2793 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2794 zio->io_error == ENXIO &&
2795 spa_load_state(spa) == SPA_LOAD_NONE &&
2796 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2797 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2799 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2800 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2803 * Here is a possibly good place to attempt to do
2804 * either combinatorial reconstruction or error correction
2805 * based on checksums. It also might be a good place
2806 * to send out preliminary ereports before we suspend
2812 * If there were logical child errors, they apply to us now.
2813 * We defer this until now to avoid conflating logical child
2814 * errors with errors that happened to the zio itself when
2815 * updating vdev stats and reporting FMA events above.
2817 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2819 if ((zio->io_error || zio->io_reexecute) &&
2820 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2821 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2822 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2824 zio_gang_tree_free(&zio->io_gang_tree);
2827 * Godfather I/Os should never suspend.
2829 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2830 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2831 zio->io_reexecute = 0;
2833 if (zio->io_reexecute) {
2835 * This is a logical I/O that wants to reexecute.
2837 * Reexecute is top-down. When an i/o fails, if it's not
2838 * the root, it simply notifies its parent and sticks around.
2839 * The parent, seeing that it still has children in zio_done(),
2840 * does the same. This percolates all the way up to the root.
2841 * The root i/o will reexecute or suspend the entire tree.
2843 * This approach ensures that zio_reexecute() honors
2844 * all the original i/o dependency relationships, e.g.
2845 * parents not executing until children are ready.
2847 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2849 zio->io_gang_leader = NULL;
2851 mutex_enter(&zio->io_lock);
2852 zio->io_state[ZIO_WAIT_DONE] = 1;
2853 mutex_exit(&zio->io_lock);
2856 * "The Godfather" I/O monitors its children but is
2857 * not a true parent to them. It will track them through
2858 * the pipeline but severs its ties whenever they get into
2859 * trouble (e.g. suspended). This allows "The Godfather"
2860 * I/O to return status without blocking.
2862 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2863 zio_link_t *zl = zio->io_walk_link;
2864 pio_next = zio_walk_parents(zio);
2866 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2867 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2868 zio_remove_child(pio, zio, zl);
2869 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2873 if ((pio = zio_unique_parent(zio)) != NULL) {
2875 * We're not a root i/o, so there's nothing to do
2876 * but notify our parent. Don't propagate errors
2877 * upward since we haven't permanently failed yet.
2879 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2880 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2881 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2882 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2884 * We'd fail again if we reexecuted now, so suspend
2885 * until conditions improve (e.g. device comes online).
2887 zio_suspend(spa, zio);
2890 * Reexecution is potentially a huge amount of work.
2891 * Hand it off to the otherwise-unused claim taskq.
2894 (void) taskq_dispatch_safe(
2895 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2896 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
2899 (void) taskq_dispatch(
2900 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2901 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2904 return (ZIO_PIPELINE_STOP);
2907 ASSERT(zio->io_child_count == 0);
2908 ASSERT(zio->io_reexecute == 0);
2909 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2912 * Report any checksum errors, since the I/O is complete.
2914 while (zio->io_cksum_report != NULL) {
2915 zio_cksum_report_t *zcr = zio->io_cksum_report;
2916 zio->io_cksum_report = zcr->zcr_next;
2917 zcr->zcr_next = NULL;
2918 zcr->zcr_finish(zcr, NULL);
2919 zfs_ereport_free_checksum(zcr);
2923 * It is the responsibility of the done callback to ensure that this
2924 * particular zio is no longer discoverable for adoption, and as
2925 * such, cannot acquire any new parents.
2930 mutex_enter(&zio->io_lock);
2931 zio->io_state[ZIO_WAIT_DONE] = 1;
2932 mutex_exit(&zio->io_lock);
2934 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2935 zio_link_t *zl = zio->io_walk_link;
2936 pio_next = zio_walk_parents(zio);
2937 zio_remove_child(pio, zio, zl);
2938 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2941 if (zio->io_waiter != NULL) {
2942 mutex_enter(&zio->io_lock);
2943 zio->io_executor = NULL;
2944 cv_broadcast(&zio->io_cv);
2945 mutex_exit(&zio->io_lock);
2950 return (ZIO_PIPELINE_STOP);
2954 * ==========================================================================
2955 * I/O pipeline definition
2956 * ==========================================================================
2958 static zio_pipe_stage_t *zio_pipeline[] = {
2964 zio_checksum_generate,
2978 zio_checksum_verify,