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];
77 int zio_bulk_flags = 0;
78 int zio_delay_max = ZIO_DELAY_MAX;
81 extern vmem_t *zio_alloc_arena;
85 * An allocating zio is one that either currently has the DVA allocate
86 * stage set or will have it later in its lifetime.
88 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
90 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
93 int zio_buf_debug_limit = 16384;
95 int zio_buf_debug_limit = 0;
98 static inline void __zio_execute(zio_t *zio);
104 vmem_t *data_alloc_arena = NULL;
107 data_alloc_arena = zio_alloc_arena;
109 zio_cache = kmem_cache_create("zio_cache",
110 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 zio_link_cache = kmem_cache_create("zio_link_cache",
112 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
115 * For small buffers, we want a cache for each multiple of
116 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
117 * for each quarter-power of 2. For large buffers, we want
118 * a cache for each multiple of PAGESIZE.
120 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
121 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
125 while (p2 & (p2 - 1))
128 if (size <= 4 * SPA_MINBLOCKSIZE) {
129 align = SPA_MINBLOCKSIZE;
130 } else if (P2PHASE(size, PAGESIZE) == 0) {
132 } else if (P2PHASE(size, p2 >> 2) == 0) {
138 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
139 zio_buf_cache[c] = kmem_cache_create(name, size,
140 align, NULL, NULL, NULL, NULL, NULL,
141 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
144 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
145 zio_data_buf_cache[c] = kmem_cache_create(name, size,
146 align, NULL, NULL, NULL, NULL, data_alloc_arena,
147 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
153 ASSERT(zio_buf_cache[c] != NULL);
154 if (zio_buf_cache[c - 1] == NULL)
155 zio_buf_cache[c - 1] = zio_buf_cache[c];
157 ASSERT(zio_data_buf_cache[c] != NULL);
158 if (zio_data_buf_cache[c - 1] == NULL)
159 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
169 kmem_cache_t *last_cache = NULL;
170 kmem_cache_t *last_data_cache = NULL;
172 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
173 if (zio_buf_cache[c] != last_cache) {
174 last_cache = zio_buf_cache[c];
175 kmem_cache_destroy(zio_buf_cache[c]);
177 zio_buf_cache[c] = NULL;
179 if (zio_data_buf_cache[c] != last_data_cache) {
180 last_data_cache = zio_data_buf_cache[c];
181 kmem_cache_destroy(zio_data_buf_cache[c]);
183 zio_data_buf_cache[c] = NULL;
186 kmem_cache_destroy(zio_link_cache);
187 kmem_cache_destroy(zio_cache);
193 * ==========================================================================
194 * Allocate and free I/O buffers
195 * ==========================================================================
199 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
200 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
201 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
202 * excess / transient data in-core during a crashdump.
205 zio_buf_alloc(size_t size)
207 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
209 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
211 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
215 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
216 * crashdump if the kernel panics. This exists so that we will limit the amount
217 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
218 * of kernel heap dumped to disk when the kernel panics)
221 zio_data_buf_alloc(size_t size)
223 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
225 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
227 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
231 zio_buf_free(void *buf, size_t size)
233 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
235 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
237 kmem_cache_free(zio_buf_cache[c], buf);
241 zio_data_buf_free(void *buf, size_t size)
243 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
245 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
247 kmem_cache_free(zio_data_buf_cache[c], buf);
251 * ==========================================================================
252 * Push and pop I/O transform buffers
253 * ==========================================================================
256 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
257 zio_transform_func_t *transform)
259 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
261 zt->zt_orig_data = zio->io_data;
262 zt->zt_orig_size = zio->io_size;
263 zt->zt_bufsize = bufsize;
264 zt->zt_transform = transform;
266 zt->zt_next = zio->io_transform_stack;
267 zio->io_transform_stack = zt;
274 zio_pop_transforms(zio_t *zio)
278 while ((zt = zio->io_transform_stack) != NULL) {
279 if (zt->zt_transform != NULL)
280 zt->zt_transform(zio,
281 zt->zt_orig_data, zt->zt_orig_size);
283 if (zt->zt_bufsize != 0)
284 zio_buf_free(zio->io_data, zt->zt_bufsize);
286 zio->io_data = zt->zt_orig_data;
287 zio->io_size = zt->zt_orig_size;
288 zio->io_transform_stack = zt->zt_next;
290 kmem_free(zt, sizeof (zio_transform_t));
295 * ==========================================================================
296 * I/O transform callbacks for subblocks and decompression
297 * ==========================================================================
300 zio_subblock(zio_t *zio, void *data, uint64_t size)
302 ASSERT(zio->io_size > size);
304 if (zio->io_type == ZIO_TYPE_READ)
305 bcopy(zio->io_data, data, size);
309 zio_decompress(zio_t *zio, void *data, uint64_t size)
311 if (zio->io_error == 0 &&
312 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
313 zio->io_data, data, zio->io_size, size) != 0)
318 * ==========================================================================
319 * I/O parent/child relationships and pipeline interlocks
320 * ==========================================================================
323 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
324 * continue calling these functions until they return NULL.
325 * Otherwise, the next caller will pick up the list walk in
326 * some indeterminate state. (Otherwise every caller would
327 * have to pass in a cookie to keep the state represented by
328 * io_walk_link, which gets annoying.)
331 zio_walk_parents(zio_t *cio)
333 zio_link_t *zl = cio->io_walk_link;
334 list_t *pl = &cio->io_parent_list;
336 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
337 cio->io_walk_link = zl;
342 ASSERT(zl->zl_child == cio);
343 return (zl->zl_parent);
347 zio_walk_children(zio_t *pio)
349 zio_link_t *zl = pio->io_walk_link;
350 list_t *cl = &pio->io_child_list;
352 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
353 pio->io_walk_link = zl;
358 ASSERT(zl->zl_parent == pio);
359 return (zl->zl_child);
363 zio_unique_parent(zio_t *cio)
365 zio_t *pio = zio_walk_parents(cio);
367 VERIFY(zio_walk_parents(cio) == NULL);
372 zio_add_child(zio_t *pio, zio_t *cio)
374 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
378 * Logical I/Os can have logical, gang, or vdev children.
379 * Gang I/Os can have gang or vdev children.
380 * Vdev I/Os can only have vdev children.
381 * The following ASSERT captures all of these constraints.
383 ASSERT(cio->io_child_type <= pio->io_child_type);
388 mutex_enter(&cio->io_lock);
389 mutex_enter(&pio->io_lock);
391 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
393 for (w = 0; w < ZIO_WAIT_TYPES; w++)
394 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
396 list_insert_head(&pio->io_child_list, zl);
397 list_insert_head(&cio->io_parent_list, zl);
399 pio->io_child_count++;
400 cio->io_parent_count++;
402 mutex_exit(&pio->io_lock);
403 mutex_exit(&cio->io_lock);
407 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
409 ASSERT(zl->zl_parent == pio);
410 ASSERT(zl->zl_child == cio);
412 mutex_enter(&cio->io_lock);
413 mutex_enter(&pio->io_lock);
415 list_remove(&pio->io_child_list, zl);
416 list_remove(&cio->io_parent_list, zl);
418 pio->io_child_count--;
419 cio->io_parent_count--;
421 mutex_exit(&pio->io_lock);
422 mutex_exit(&cio->io_lock);
424 kmem_cache_free(zio_link_cache, zl);
428 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
430 uint64_t *countp = &zio->io_children[child][wait];
431 boolean_t waiting = B_FALSE;
433 mutex_enter(&zio->io_lock);
434 ASSERT(zio->io_stall == NULL);
437 zio->io_stall = countp;
440 mutex_exit(&zio->io_lock);
445 __attribute__((always_inline))
447 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
449 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
450 int *errorp = &pio->io_child_error[zio->io_child_type];
452 mutex_enter(&pio->io_lock);
453 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
454 *errorp = zio_worst_error(*errorp, zio->io_error);
455 pio->io_reexecute |= zio->io_reexecute;
456 ASSERT3U(*countp, >, 0);
457 if (--*countp == 0 && pio->io_stall == countp) {
458 pio->io_stall = NULL;
459 mutex_exit(&pio->io_lock);
462 mutex_exit(&pio->io_lock);
467 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
469 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
470 zio->io_error = zio->io_child_error[c];
474 * ==========================================================================
475 * Create the various types of I/O (read, write, free, etc)
476 * ==========================================================================
479 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
480 void *data, uint64_t size, zio_done_func_t *done, void *private,
481 zio_type_t type, int priority, enum zio_flag flags,
482 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
483 enum zio_stage stage, enum zio_stage pipeline)
487 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
488 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
489 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
491 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
492 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
493 ASSERT(vd || stage == ZIO_STAGE_OPEN);
495 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
496 bzero(zio, sizeof (zio_t));
498 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
499 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
501 list_create(&zio->io_parent_list, sizeof (zio_link_t),
502 offsetof(zio_link_t, zl_parent_node));
503 list_create(&zio->io_child_list, sizeof (zio_link_t),
504 offsetof(zio_link_t, zl_child_node));
507 zio->io_child_type = ZIO_CHILD_VDEV;
508 else if (flags & ZIO_FLAG_GANG_CHILD)
509 zio->io_child_type = ZIO_CHILD_GANG;
510 else if (flags & ZIO_FLAG_DDT_CHILD)
511 zio->io_child_type = ZIO_CHILD_DDT;
513 zio->io_child_type = ZIO_CHILD_LOGICAL;
516 zio->io_bp = (blkptr_t *)bp;
517 zio->io_bp_copy = *bp;
518 zio->io_bp_orig = *bp;
519 if (type != ZIO_TYPE_WRITE ||
520 zio->io_child_type == ZIO_CHILD_DDT)
521 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
522 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
523 zio->io_logical = zio;
524 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
525 pipeline |= ZIO_GANG_STAGES;
531 zio->io_private = private;
533 zio->io_priority = priority;
535 zio->io_offset = offset;
536 zio->io_orig_data = zio->io_data = data;
537 zio->io_orig_size = zio->io_size = size;
538 zio->io_orig_flags = zio->io_flags = flags;
539 zio->io_orig_stage = zio->io_stage = stage;
540 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
542 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
543 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
546 zio->io_bookmark = *zb;
549 if (zio->io_logical == NULL)
550 zio->io_logical = pio->io_logical;
551 if (zio->io_child_type == ZIO_CHILD_GANG)
552 zio->io_gang_leader = pio->io_gang_leader;
553 zio_add_child(pio, zio);
560 zio_destroy(zio_t *zio)
562 list_destroy(&zio->io_parent_list);
563 list_destroy(&zio->io_child_list);
564 mutex_destroy(&zio->io_lock);
565 cv_destroy(&zio->io_cv);
566 kmem_cache_free(zio_cache, zio);
570 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
571 void *private, enum zio_flag flags)
575 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
576 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
577 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
583 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
585 return (zio_null(NULL, spa, NULL, done, private, flags));
589 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
590 void *data, uint64_t size, zio_done_func_t *done, void *private,
591 int priority, enum zio_flag flags, const zbookmark_t *zb)
595 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
596 data, size, done, private,
597 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
598 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
599 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
605 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
606 void *data, uint64_t size, const zio_prop_t *zp,
607 zio_done_func_t *ready, zio_done_func_t *done, void *private,
608 int priority, enum zio_flag flags, const zbookmark_t *zb)
612 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
613 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
614 zp->zp_compress >= ZIO_COMPRESS_OFF &&
615 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
616 zp->zp_type < DMU_OT_NUMTYPES &&
619 zp->zp_copies <= spa_max_replication(spa) &&
621 zp->zp_dedup_verify <= 1);
623 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
624 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
625 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
626 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
628 zio->io_ready = ready;
635 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
636 uint64_t size, zio_done_func_t *done, void *private, int priority,
637 enum zio_flag flags, zbookmark_t *zb)
641 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
642 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
643 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
649 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
651 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
652 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
653 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
654 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
656 zio->io_prop.zp_copies = copies;
657 zio->io_bp_override = bp;
661 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
663 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
667 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
672 dprintf_bp(bp, "freeing in txg %llu, pass %u",
673 (longlong_t)txg, spa->spa_sync_pass);
675 ASSERT(!BP_IS_HOLE(bp));
676 ASSERT(spa_syncing_txg(spa) == txg);
677 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
679 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
680 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
681 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
687 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
688 zio_done_func_t *done, void *private, enum zio_flag flags)
693 * A claim is an allocation of a specific block. Claims are needed
694 * to support immediate writes in the intent log. The issue is that
695 * immediate writes contain committed data, but in a txg that was
696 * *not* committed. Upon opening the pool after an unclean shutdown,
697 * the intent log claims all blocks that contain immediate write data
698 * so that the SPA knows they're in use.
700 * All claims *must* be resolved in the first txg -- before the SPA
701 * starts allocating blocks -- so that nothing is allocated twice.
702 * If txg == 0 we just verify that the block is claimable.
704 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
705 ASSERT(txg == spa_first_txg(spa) || txg == 0);
706 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
708 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
709 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
710 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
716 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
717 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
722 if (vd->vdev_children == 0) {
723 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
724 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
725 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
729 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
731 for (c = 0; c < vd->vdev_children; c++)
732 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
733 done, private, priority, flags));
740 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
741 void *data, int checksum, zio_done_func_t *done, void *private,
742 int priority, enum zio_flag flags, boolean_t labels)
746 ASSERT(vd->vdev_children == 0);
747 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
748 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
749 ASSERT3U(offset + size, <=, vd->vdev_psize);
751 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
752 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
753 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
755 zio->io_prop.zp_checksum = checksum;
761 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
762 void *data, int checksum, zio_done_func_t *done, void *private,
763 int priority, enum zio_flag flags, boolean_t labels)
767 ASSERT(vd->vdev_children == 0);
768 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
769 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
770 ASSERT3U(offset + size, <=, vd->vdev_psize);
772 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
773 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
774 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
776 zio->io_prop.zp_checksum = checksum;
778 if (zio_checksum_table[checksum].ci_eck) {
780 * zec checksums are necessarily destructive -- they modify
781 * the end of the write buffer to hold the verifier/checksum.
782 * Therefore, we must make a local copy in case the data is
783 * being written to multiple places in parallel.
785 void *wbuf = zio_buf_alloc(size);
786 bcopy(data, wbuf, size);
787 zio_push_transform(zio, wbuf, size, size, NULL);
794 * Create a child I/O to do some work for us.
797 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
798 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
799 zio_done_func_t *done, void *private)
801 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
804 ASSERT(vd->vdev_parent ==
805 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
807 if (type == ZIO_TYPE_READ && bp != NULL) {
809 * If we have the bp, then the child should perform the
810 * checksum and the parent need not. This pushes error
811 * detection as close to the leaves as possible and
812 * eliminates redundant checksums in the interior nodes.
814 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
815 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
818 if (vd->vdev_children == 0)
819 offset += VDEV_LABEL_START_SIZE;
821 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
824 * If we've decided to do a repair, the write is not speculative --
825 * even if the original read was.
827 if (flags & ZIO_FLAG_IO_REPAIR)
828 flags &= ~ZIO_FLAG_SPECULATIVE;
830 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
831 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
832 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
838 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
839 int type, int priority, enum zio_flag flags,
840 zio_done_func_t *done, void *private)
844 ASSERT(vd->vdev_ops->vdev_op_leaf);
846 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
847 data, size, done, private, type, priority,
848 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
850 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
856 zio_flush(zio_t *zio, vdev_t *vd)
858 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
859 NULL, NULL, ZIO_PRIORITY_NOW,
860 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
864 zio_shrink(zio_t *zio, uint64_t size)
866 ASSERT(zio->io_executor == NULL);
867 ASSERT(zio->io_orig_size == zio->io_size);
868 ASSERT(size <= zio->io_size);
871 * We don't shrink for raidz because of problems with the
872 * reconstruction when reading back less than the block size.
873 * Note, BP_IS_RAIDZ() assumes no compression.
875 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
876 if (!BP_IS_RAIDZ(zio->io_bp))
877 zio->io_orig_size = zio->io_size = size;
881 * ==========================================================================
882 * Prepare to read and write logical blocks
883 * ==========================================================================
887 zio_read_bp_init(zio_t *zio)
889 blkptr_t *bp = zio->io_bp;
891 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
892 zio->io_child_type == ZIO_CHILD_LOGICAL &&
893 !(zio->io_flags & ZIO_FLAG_RAW)) {
894 uint64_t psize = BP_GET_PSIZE(bp);
895 void *cbuf = zio_buf_alloc(psize);
897 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
900 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
901 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
903 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
904 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
906 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
907 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
909 return (ZIO_PIPELINE_CONTINUE);
913 zio_write_bp_init(zio_t *zio)
915 spa_t *spa = zio->io_spa;
916 zio_prop_t *zp = &zio->io_prop;
917 enum zio_compress compress = zp->zp_compress;
918 blkptr_t *bp = zio->io_bp;
919 uint64_t lsize = zio->io_size;
920 uint64_t psize = lsize;
924 * If our children haven't all reached the ready stage,
925 * wait for them and then repeat this pipeline stage.
927 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
928 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
929 return (ZIO_PIPELINE_STOP);
931 if (!IO_IS_ALLOCATING(zio))
932 return (ZIO_PIPELINE_CONTINUE);
934 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
936 if (zio->io_bp_override) {
937 ASSERT(bp->blk_birth != zio->io_txg);
938 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
940 *bp = *zio->io_bp_override;
941 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
943 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
944 return (ZIO_PIPELINE_CONTINUE);
946 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
947 zp->zp_dedup_verify);
949 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
951 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
952 return (ZIO_PIPELINE_CONTINUE);
954 zio->io_bp_override = NULL;
958 if (bp->blk_birth == zio->io_txg) {
960 * We're rewriting an existing block, which means we're
961 * working on behalf of spa_sync(). For spa_sync() to
962 * converge, it must eventually be the case that we don't
963 * have to allocate new blocks. But compression changes
964 * the blocksize, which forces a reallocate, and makes
965 * convergence take longer. Therefore, after the first
966 * few passes, stop compressing to ensure convergence.
968 pass = spa_sync_pass(spa);
970 ASSERT(zio->io_txg == spa_syncing_txg(spa));
971 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
972 ASSERT(!BP_GET_DEDUP(bp));
974 if (pass > SYNC_PASS_DONT_COMPRESS)
975 compress = ZIO_COMPRESS_OFF;
977 /* Make sure someone doesn't change their mind on overwrites */
978 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
979 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
982 if (compress != ZIO_COMPRESS_OFF) {
983 void *cbuf = zio_buf_alloc(lsize);
984 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
985 if (psize == 0 || psize == lsize) {
986 compress = ZIO_COMPRESS_OFF;
987 zio_buf_free(cbuf, lsize);
989 ASSERT(psize < lsize);
990 zio_push_transform(zio, cbuf, psize, lsize, NULL);
995 * The final pass of spa_sync() must be all rewrites, but the first
996 * few passes offer a trade-off: allocating blocks defers convergence,
997 * but newly allocated blocks are sequential, so they can be written
998 * to disk faster. Therefore, we allow the first few passes of
999 * spa_sync() to allocate new blocks, but force rewrites after that.
1000 * There should only be a handful of blocks after pass 1 in any case.
1002 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1003 pass > SYNC_PASS_REWRITE) {
1004 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1006 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1007 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1010 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1014 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1016 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1017 BP_SET_LSIZE(bp, lsize);
1018 BP_SET_PSIZE(bp, psize);
1019 BP_SET_COMPRESS(bp, compress);
1020 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1021 BP_SET_TYPE(bp, zp->zp_type);
1022 BP_SET_LEVEL(bp, zp->zp_level);
1023 BP_SET_DEDUP(bp, zp->zp_dedup);
1024 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1026 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1027 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1028 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1032 return (ZIO_PIPELINE_CONTINUE);
1036 zio_free_bp_init(zio_t *zio)
1038 blkptr_t *bp = zio->io_bp;
1040 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1041 if (BP_GET_DEDUP(bp))
1042 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1045 return (ZIO_PIPELINE_CONTINUE);
1049 * ==========================================================================
1050 * Execute the I/O pipeline
1051 * ==========================================================================
1055 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1057 spa_t *spa = zio->io_spa;
1058 zio_type_t t = zio->io_type;
1059 int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
1062 * If we're a config writer or a probe, the normal issue and
1063 * interrupt threads may all be blocked waiting for the config lock.
1064 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1066 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1070 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1072 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1076 * If this is a high priority I/O, then use the high priority taskq.
1078 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1079 spa->spa_zio_taskq[t][q + 1] != NULL)
1082 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1084 while (taskq_dispatch(spa->spa_zio_taskq[t][q],
1085 (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
1089 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1091 kthread_t *executor = zio->io_executor;
1092 spa_t *spa = zio->io_spa;
1095 for (t = 0; t < ZIO_TYPES; t++)
1096 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1103 zio_issue_async(zio_t *zio)
1105 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1107 return (ZIO_PIPELINE_STOP);
1111 zio_interrupt(zio_t *zio)
1113 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1117 * Execute the I/O pipeline until one of the following occurs:
1118 * (1) the I/O completes; (2) the pipeline stalls waiting for
1119 * dependent child I/Os; (3) the I/O issues, so we're waiting
1120 * for an I/O completion interrupt; (4) the I/O is delegated by
1121 * vdev-level caching or aggregation; (5) the I/O is deferred
1122 * due to vdev-level queueing; (6) the I/O is handed off to
1123 * another thread. In all cases, the pipeline stops whenever
1124 * there's no CPU work; it never burns a thread in cv_wait().
1126 * There's no locking on io_stage because there's no legitimate way
1127 * for multiple threads to be attempting to process the same I/O.
1129 static zio_pipe_stage_t *zio_pipeline[];
1132 * zio_execute() is a wrapper around the static function
1133 * __zio_execute() so that we can force __zio_execute() to be
1134 * inlined. This reduces stack overhead which is important
1135 * because __zio_execute() is called recursively in several zio
1136 * code paths. zio_execute() itself cannot be inlined because
1137 * it is externally visible.
1140 zio_execute(zio_t *zio)
1145 __attribute__((always_inline))
1147 __zio_execute(zio_t *zio)
1149 zio->io_executor = curthread;
1151 while (zio->io_stage < ZIO_STAGE_DONE) {
1152 enum zio_stage pipeline = zio->io_pipeline;
1153 enum zio_stage stage = zio->io_stage;
1156 ASSERT(!MUTEX_HELD(&zio->io_lock));
1157 ASSERT(ISP2(stage));
1158 ASSERT(zio->io_stall == NULL);
1162 } while ((stage & pipeline) == 0);
1164 ASSERT(stage <= ZIO_STAGE_DONE);
1167 * If we are in interrupt context and this pipeline stage
1168 * will grab a config lock that is held across I/O,
1169 * or may wait for an I/O that needs an interrupt thread
1170 * to complete, issue async to avoid deadlock.
1172 * For VDEV_IO_START, we cut in line so that the io will
1173 * be sent to disk promptly.
1175 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1176 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1177 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1178 zio_requeue_io_start_cut_in_line : B_FALSE;
1179 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1183 zio->io_stage = stage;
1184 rv = zio_pipeline[highbit(stage) - 1](zio);
1186 if (rv == ZIO_PIPELINE_STOP)
1189 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;
1254 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1255 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1256 ASSERT(pio->io_gang_leader == NULL);
1257 ASSERT(pio->io_gang_tree == NULL);
1259 pio->io_flags = pio->io_orig_flags;
1260 pio->io_stage = pio->io_orig_stage;
1261 pio->io_pipeline = pio->io_orig_pipeline;
1262 pio->io_reexecute = 0;
1264 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1265 pio->io_state[w] = 0;
1266 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1267 pio->io_child_error[c] = 0;
1269 if (IO_IS_ALLOCATING(pio))
1270 BP_ZERO(pio->io_bp);
1273 * As we reexecute pio's children, new children could be created.
1274 * New children go to the head of pio's io_child_list, however,
1275 * so we will (correctly) not reexecute them. The key is that
1276 * the remainder of pio's io_child_list, from 'cio_next' onward,
1277 * cannot be affected by any side effects of reexecuting 'cio'.
1279 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1280 cio_next = zio_walk_children(pio);
1281 mutex_enter(&pio->io_lock);
1282 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1283 pio->io_children[cio->io_child_type][w]++;
1284 mutex_exit(&pio->io_lock);
1289 * Now that all children have been reexecuted, execute the parent.
1290 * We don't reexecute "The Godfather" I/O here as it's the
1291 * responsibility of the caller to wait on him.
1293 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1298 zio_suspend(spa_t *spa, zio_t *zio)
1300 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1301 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1302 "failure and the failure mode property for this pool "
1303 "is set to panic.", spa_name(spa));
1305 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1307 mutex_enter(&spa->spa_suspend_lock);
1309 if (spa->spa_suspend_zio_root == NULL)
1310 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1311 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1312 ZIO_FLAG_GODFATHER);
1314 spa->spa_suspended = B_TRUE;
1317 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1318 ASSERT(zio != spa->spa_suspend_zio_root);
1319 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1320 ASSERT(zio_unique_parent(zio) == NULL);
1321 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1322 zio_add_child(spa->spa_suspend_zio_root, zio);
1325 mutex_exit(&spa->spa_suspend_lock);
1329 zio_resume(spa_t *spa)
1334 * Reexecute all previously suspended i/o.
1336 mutex_enter(&spa->spa_suspend_lock);
1337 spa->spa_suspended = B_FALSE;
1338 cv_broadcast(&spa->spa_suspend_cv);
1339 pio = spa->spa_suspend_zio_root;
1340 spa->spa_suspend_zio_root = NULL;
1341 mutex_exit(&spa->spa_suspend_lock);
1347 return (zio_wait(pio));
1351 zio_resume_wait(spa_t *spa)
1353 mutex_enter(&spa->spa_suspend_lock);
1354 while (spa_suspended(spa))
1355 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1356 mutex_exit(&spa->spa_suspend_lock);
1360 * ==========================================================================
1363 * A gang block is a collection of small blocks that looks to the DMU
1364 * like one large block. When zio_dva_allocate() cannot find a block
1365 * of the requested size, due to either severe fragmentation or the pool
1366 * being nearly full, it calls zio_write_gang_block() to construct the
1367 * block from smaller fragments.
1369 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1370 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1371 * an indirect block: it's an array of block pointers. It consumes
1372 * only one sector and hence is allocatable regardless of fragmentation.
1373 * The gang header's bps point to its gang members, which hold the data.
1375 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1376 * as the verifier to ensure uniqueness of the SHA256 checksum.
1377 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1378 * not the gang header. This ensures that data block signatures (needed for
1379 * deduplication) are independent of how the block is physically stored.
1381 * Gang blocks can be nested: a gang member may itself be a gang block.
1382 * Thus every gang block is a tree in which root and all interior nodes are
1383 * gang headers, and the leaves are normal blocks that contain user data.
1384 * The root of the gang tree is called the gang leader.
1386 * To perform any operation (read, rewrite, free, claim) on a gang block,
1387 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1388 * in the io_gang_tree field of the original logical i/o by recursively
1389 * reading the gang leader and all gang headers below it. This yields
1390 * an in-core tree containing the contents of every gang header and the
1391 * bps for every constituent of the gang block.
1393 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1394 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1395 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1396 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1397 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1398 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1399 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1400 * of the gang header plus zio_checksum_compute() of the data to update the
1401 * gang header's blk_cksum as described above.
1403 * The two-phase assemble/issue model solves the problem of partial failure --
1404 * what if you'd freed part of a gang block but then couldn't read the
1405 * gang header for another part? Assembling the entire gang tree first
1406 * ensures that all the necessary gang header I/O has succeeded before
1407 * starting the actual work of free, claim, or write. Once the gang tree
1408 * is assembled, free and claim are in-memory operations that cannot fail.
1410 * In the event that a gang write fails, zio_dva_unallocate() walks the
1411 * gang tree to immediately free (i.e. insert back into the space map)
1412 * everything we've allocated. This ensures that we don't get ENOSPC
1413 * errors during repeated suspend/resume cycles due to a flaky device.
1415 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1416 * the gang tree, we won't modify the block, so we can safely defer the free
1417 * (knowing that the block is still intact). If we *can* assemble the gang
1418 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1419 * each constituent bp and we can allocate a new block on the next sync pass.
1421 * In all cases, the gang tree allows complete recovery from partial failure.
1422 * ==========================================================================
1426 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1431 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1432 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1433 &pio->io_bookmark));
1437 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1442 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1443 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1444 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1446 * As we rewrite each gang header, the pipeline will compute
1447 * a new gang block header checksum for it; but no one will
1448 * compute a new data checksum, so we do that here. The one
1449 * exception is the gang leader: the pipeline already computed
1450 * its data checksum because that stage precedes gang assembly.
1451 * (Presently, nothing actually uses interior data checksums;
1452 * this is just good hygiene.)
1454 if (gn != pio->io_gang_leader->io_gang_tree) {
1455 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1456 data, BP_GET_PSIZE(bp));
1459 * If we are here to damage data for testing purposes,
1460 * leave the GBH alone so that we can detect the damage.
1462 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1463 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1465 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1466 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1467 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1475 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1477 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1478 ZIO_GANG_CHILD_FLAGS(pio)));
1483 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1485 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1486 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1489 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1498 static void zio_gang_tree_assemble_done(zio_t *zio);
1500 static zio_gang_node_t *
1501 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1503 zio_gang_node_t *gn;
1505 ASSERT(*gnpp == NULL);
1507 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1508 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1515 zio_gang_node_free(zio_gang_node_t **gnpp)
1517 zio_gang_node_t *gn = *gnpp;
1520 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1521 ASSERT(gn->gn_child[g] == NULL);
1523 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1524 kmem_free(gn, sizeof (*gn));
1529 zio_gang_tree_free(zio_gang_node_t **gnpp)
1531 zio_gang_node_t *gn = *gnpp;
1537 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1538 zio_gang_tree_free(&gn->gn_child[g]);
1540 zio_gang_node_free(gnpp);
1544 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1546 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1548 ASSERT(gio->io_gang_leader == gio);
1549 ASSERT(BP_IS_GANG(bp));
1551 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1552 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1553 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1557 zio_gang_tree_assemble_done(zio_t *zio)
1559 zio_t *gio = zio->io_gang_leader;
1560 zio_gang_node_t *gn = zio->io_private;
1561 blkptr_t *bp = zio->io_bp;
1564 ASSERT(gio == zio_unique_parent(zio));
1565 ASSERT(zio->io_child_count == 0);
1570 if (BP_SHOULD_BYTESWAP(bp))
1571 byteswap_uint64_array(zio->io_data, zio->io_size);
1573 ASSERT(zio->io_data == gn->gn_gbh);
1574 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1575 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1577 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1578 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1579 if (!BP_IS_GANG(gbp))
1581 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1586 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1588 zio_t *gio = pio->io_gang_leader;
1592 ASSERT(BP_IS_GANG(bp) == !!gn);
1593 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1594 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1597 * If you're a gang header, your data is in gn->gn_gbh.
1598 * If you're a gang member, your data is in 'data' and gn == NULL.
1600 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1603 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1605 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1606 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1607 if (BP_IS_HOLE(gbp))
1609 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1610 data = (char *)data + BP_GET_PSIZE(gbp);
1614 if (gn == gio->io_gang_tree)
1615 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1622 zio_gang_assemble(zio_t *zio)
1624 blkptr_t *bp = zio->io_bp;
1626 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1627 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1629 zio->io_gang_leader = zio;
1631 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1633 return (ZIO_PIPELINE_CONTINUE);
1637 zio_gang_issue(zio_t *zio)
1639 blkptr_t *bp = zio->io_bp;
1641 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1642 return (ZIO_PIPELINE_STOP);
1644 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1645 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1647 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1648 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1650 zio_gang_tree_free(&zio->io_gang_tree);
1652 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1654 return (ZIO_PIPELINE_CONTINUE);
1658 zio_write_gang_member_ready(zio_t *zio)
1660 zio_t *pio = zio_unique_parent(zio);
1661 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1662 dva_t *cdva = zio->io_bp->blk_dva;
1663 dva_t *pdva = pio->io_bp->blk_dva;
1667 if (BP_IS_HOLE(zio->io_bp))
1670 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1672 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1673 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1674 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1675 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1676 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1678 mutex_enter(&pio->io_lock);
1679 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1680 ASSERT(DVA_GET_GANG(&pdva[d]));
1681 asize = DVA_GET_ASIZE(&pdva[d]);
1682 asize += DVA_GET_ASIZE(&cdva[d]);
1683 DVA_SET_ASIZE(&pdva[d], asize);
1685 mutex_exit(&pio->io_lock);
1689 zio_write_gang_block(zio_t *pio)
1691 spa_t *spa = pio->io_spa;
1692 blkptr_t *bp = pio->io_bp;
1693 zio_t *gio = pio->io_gang_leader;
1695 zio_gang_node_t *gn, **gnpp;
1696 zio_gbh_phys_t *gbh;
1697 uint64_t txg = pio->io_txg;
1698 uint64_t resid = pio->io_size;
1700 int copies = gio->io_prop.zp_copies;
1701 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1705 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1706 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1707 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1709 pio->io_error = error;
1710 return (ZIO_PIPELINE_CONTINUE);
1714 gnpp = &gio->io_gang_tree;
1716 gnpp = pio->io_private;
1717 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1720 gn = zio_gang_node_alloc(gnpp);
1722 bzero(gbh, SPA_GANGBLOCKSIZE);
1725 * Create the gang header.
1727 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1728 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1731 * Create and nowait the gang children.
1733 for (g = 0; resid != 0; resid -= lsize, g++) {
1734 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1736 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1738 zp.zp_checksum = gio->io_prop.zp_checksum;
1739 zp.zp_compress = ZIO_COMPRESS_OFF;
1740 zp.zp_type = DMU_OT_NONE;
1742 zp.zp_copies = gio->io_prop.zp_copies;
1744 zp.zp_dedup_verify = 0;
1746 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1747 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1748 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1749 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1750 &pio->io_bookmark));
1754 * Set pio's pipeline to just wait for zio to finish.
1756 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1760 return (ZIO_PIPELINE_CONTINUE);
1764 * ==========================================================================
1766 * ==========================================================================
1769 zio_ddt_child_read_done(zio_t *zio)
1771 blkptr_t *bp = zio->io_bp;
1772 ddt_entry_t *dde = zio->io_private;
1774 zio_t *pio = zio_unique_parent(zio);
1776 mutex_enter(&pio->io_lock);
1777 ddp = ddt_phys_select(dde, bp);
1778 if (zio->io_error == 0)
1779 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1780 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1781 dde->dde_repair_data = zio->io_data;
1783 zio_buf_free(zio->io_data, zio->io_size);
1784 mutex_exit(&pio->io_lock);
1788 zio_ddt_read_start(zio_t *zio)
1790 blkptr_t *bp = zio->io_bp;
1793 ASSERT(BP_GET_DEDUP(bp));
1794 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1795 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1797 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1798 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1799 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1800 ddt_phys_t *ddp = dde->dde_phys;
1801 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1804 ASSERT(zio->io_vsd == NULL);
1807 if (ddp_self == NULL)
1808 return (ZIO_PIPELINE_CONTINUE);
1810 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1811 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1813 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1815 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1816 zio_buf_alloc(zio->io_size), zio->io_size,
1817 zio_ddt_child_read_done, dde, zio->io_priority,
1818 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1819 &zio->io_bookmark));
1821 return (ZIO_PIPELINE_CONTINUE);
1824 zio_nowait(zio_read(zio, zio->io_spa, bp,
1825 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1826 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1828 return (ZIO_PIPELINE_CONTINUE);
1832 zio_ddt_read_done(zio_t *zio)
1834 blkptr_t *bp = zio->io_bp;
1836 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1837 return (ZIO_PIPELINE_STOP);
1839 ASSERT(BP_GET_DEDUP(bp));
1840 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1841 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1843 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1844 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1845 ddt_entry_t *dde = zio->io_vsd;
1847 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1848 return (ZIO_PIPELINE_CONTINUE);
1851 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1852 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1853 return (ZIO_PIPELINE_STOP);
1855 if (dde->dde_repair_data != NULL) {
1856 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1857 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1859 ddt_repair_done(ddt, dde);
1863 ASSERT(zio->io_vsd == NULL);
1865 return (ZIO_PIPELINE_CONTINUE);
1869 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1871 spa_t *spa = zio->io_spa;
1875 * Note: we compare the original data, not the transformed data,
1876 * because when zio->io_bp is an override bp, we will not have
1877 * pushed the I/O transforms. That's an important optimization
1878 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1880 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1881 zio_t *lio = dde->dde_lead_zio[p];
1884 return (lio->io_orig_size != zio->io_orig_size ||
1885 bcmp(zio->io_orig_data, lio->io_orig_data,
1886 zio->io_orig_size) != 0);
1890 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1891 ddt_phys_t *ddp = &dde->dde_phys[p];
1893 if (ddp->ddp_phys_birth != 0) {
1894 arc_buf_t *abuf = NULL;
1895 uint32_t aflags = ARC_WAIT;
1896 blkptr_t blk = *zio->io_bp;
1899 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1903 error = arc_read_nolock(NULL, spa, &blk,
1904 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1905 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1906 &aflags, &zio->io_bookmark);
1909 if (arc_buf_size(abuf) != zio->io_orig_size ||
1910 bcmp(abuf->b_data, zio->io_orig_data,
1911 zio->io_orig_size) != 0)
1913 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1917 return (error != 0);
1925 zio_ddt_child_write_ready(zio_t *zio)
1927 int p = zio->io_prop.zp_copies;
1928 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1929 ddt_entry_t *dde = zio->io_private;
1930 ddt_phys_t *ddp = &dde->dde_phys[p];
1938 ASSERT(dde->dde_lead_zio[p] == zio);
1940 ddt_phys_fill(ddp, zio->io_bp);
1942 while ((pio = zio_walk_parents(zio)) != NULL)
1943 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1949 zio_ddt_child_write_done(zio_t *zio)
1951 int p = zio->io_prop.zp_copies;
1952 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1953 ddt_entry_t *dde = zio->io_private;
1954 ddt_phys_t *ddp = &dde->dde_phys[p];
1958 ASSERT(ddp->ddp_refcnt == 0);
1959 ASSERT(dde->dde_lead_zio[p] == zio);
1960 dde->dde_lead_zio[p] = NULL;
1962 if (zio->io_error == 0) {
1963 while (zio_walk_parents(zio) != NULL)
1964 ddt_phys_addref(ddp);
1966 ddt_phys_clear(ddp);
1973 zio_ddt_ditto_write_done(zio_t *zio)
1975 int p = DDT_PHYS_DITTO;
1976 blkptr_t *bp = zio->io_bp;
1977 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1978 ddt_entry_t *dde = zio->io_private;
1979 ddt_phys_t *ddp = &dde->dde_phys[p];
1980 ddt_key_t *ddk = &dde->dde_key;
1981 ASSERTV(zio_prop_t *zp = &zio->io_prop);
1985 ASSERT(ddp->ddp_refcnt == 0);
1986 ASSERT(dde->dde_lead_zio[p] == zio);
1987 dde->dde_lead_zio[p] = NULL;
1989 if (zio->io_error == 0) {
1990 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1991 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1992 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1993 if (ddp->ddp_phys_birth != 0)
1994 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1995 ddt_phys_fill(ddp, bp);
2002 zio_ddt_write(zio_t *zio)
2004 spa_t *spa = zio->io_spa;
2005 blkptr_t *bp = zio->io_bp;
2006 uint64_t txg = zio->io_txg;
2007 zio_prop_t *zp = &zio->io_prop;
2008 int p = zp->zp_copies;
2012 ddt_t *ddt = ddt_select(spa, bp);
2016 ASSERT(BP_GET_DEDUP(bp));
2017 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2018 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2021 dde = ddt_lookup(ddt, bp, B_TRUE);
2022 ddp = &dde->dde_phys[p];
2024 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2026 * If we're using a weak checksum, upgrade to a strong checksum
2027 * and try again. If we're already using a strong checksum,
2028 * we can't resolve it, so just convert to an ordinary write.
2029 * (And automatically e-mail a paper to Nature?)
2031 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2032 zp->zp_checksum = spa_dedup_checksum(spa);
2033 zio_pop_transforms(zio);
2034 zio->io_stage = ZIO_STAGE_OPEN;
2039 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2041 return (ZIO_PIPELINE_CONTINUE);
2044 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2045 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2047 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2048 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2049 zio_prop_t czp = *zp;
2051 czp.zp_copies = ditto_copies;
2054 * If we arrived here with an override bp, we won't have run
2055 * the transform stack, so we won't have the data we need to
2056 * generate a child i/o. So, toss the override bp and restart.
2057 * This is safe, because using the override bp is just an
2058 * optimization; and it's rare, so the cost doesn't matter.
2060 if (zio->io_bp_override) {
2061 zio_pop_transforms(zio);
2062 zio->io_stage = ZIO_STAGE_OPEN;
2063 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2064 zio->io_bp_override = NULL;
2067 return (ZIO_PIPELINE_CONTINUE);
2070 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2071 zio->io_orig_size, &czp, NULL,
2072 zio_ddt_ditto_write_done, dde, zio->io_priority,
2073 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2075 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2076 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2079 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2080 if (ddp->ddp_phys_birth != 0)
2081 ddt_bp_fill(ddp, bp, txg);
2082 if (dde->dde_lead_zio[p] != NULL)
2083 zio_add_child(zio, dde->dde_lead_zio[p]);
2085 ddt_phys_addref(ddp);
2086 } else if (zio->io_bp_override) {
2087 ASSERT(bp->blk_birth == txg);
2088 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2089 ddt_phys_fill(ddp, bp);
2090 ddt_phys_addref(ddp);
2092 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2093 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2094 zio_ddt_child_write_done, dde, zio->io_priority,
2095 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2097 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2098 dde->dde_lead_zio[p] = cio;
2108 return (ZIO_PIPELINE_CONTINUE);
2111 ddt_entry_t *freedde; /* for debugging */
2114 zio_ddt_free(zio_t *zio)
2116 spa_t *spa = zio->io_spa;
2117 blkptr_t *bp = zio->io_bp;
2118 ddt_t *ddt = ddt_select(spa, bp);
2122 ASSERT(BP_GET_DEDUP(bp));
2123 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2126 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2127 ddp = ddt_phys_select(dde, bp);
2128 ddt_phys_decref(ddp);
2131 return (ZIO_PIPELINE_CONTINUE);
2135 * ==========================================================================
2136 * Allocate and free blocks
2137 * ==========================================================================
2140 zio_dva_allocate(zio_t *zio)
2142 spa_t *spa = zio->io_spa;
2143 metaslab_class_t *mc = spa_normal_class(spa);
2144 blkptr_t *bp = zio->io_bp;
2147 if (zio->io_gang_leader == NULL) {
2148 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2149 zio->io_gang_leader = zio;
2152 ASSERT(BP_IS_HOLE(bp));
2153 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2154 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2155 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2156 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2158 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2159 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2162 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2163 return (zio_write_gang_block(zio));
2164 zio->io_error = error;
2167 return (ZIO_PIPELINE_CONTINUE);
2171 zio_dva_free(zio_t *zio)
2173 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2175 return (ZIO_PIPELINE_CONTINUE);
2179 zio_dva_claim(zio_t *zio)
2183 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2185 zio->io_error = error;
2187 return (ZIO_PIPELINE_CONTINUE);
2191 * Undo an allocation. This is used by zio_done() when an I/O fails
2192 * and we want to give back the block we just allocated.
2193 * This handles both normal blocks and gang blocks.
2196 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2200 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2201 ASSERT(zio->io_bp_override == NULL);
2203 if (!BP_IS_HOLE(bp))
2204 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2207 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2208 zio_dva_unallocate(zio, gn->gn_child[g],
2209 &gn->gn_gbh->zg_blkptr[g]);
2215 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2218 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2219 uint64_t size, boolean_t use_slog)
2223 ASSERT(txg > spa_syncing_txg(spa));
2226 error = metaslab_alloc(spa, spa_log_class(spa), size,
2227 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2230 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2231 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2234 BP_SET_LSIZE(new_bp, size);
2235 BP_SET_PSIZE(new_bp, size);
2236 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2237 BP_SET_CHECKSUM(new_bp,
2238 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2239 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2240 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2241 BP_SET_LEVEL(new_bp, 0);
2242 BP_SET_DEDUP(new_bp, 0);
2243 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2250 * Free an intent log block.
2253 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2255 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2256 ASSERT(!BP_IS_GANG(bp));
2258 zio_free(spa, txg, bp);
2262 * ==========================================================================
2263 * Read and write to physical devices
2264 * ==========================================================================
2267 zio_vdev_io_start(zio_t *zio)
2269 vdev_t *vd = zio->io_vd;
2271 spa_t *spa = zio->io_spa;
2273 ASSERT(zio->io_error == 0);
2274 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2277 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2278 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2281 * The mirror_ops handle multiple DVAs in a single BP.
2283 return (vdev_mirror_ops.vdev_op_io_start(zio));
2287 * We keep track of time-sensitive I/Os so that the scan thread
2288 * can quickly react to certain workloads. In particular, we care
2289 * about non-scrubbing, top-level reads and writes with the following
2291 * - synchronous writes of user data to non-slog devices
2292 * - any reads of user data
2293 * When these conditions are met, adjust the timestamp of spa_last_io
2294 * which allows the scan thread to adjust its workload accordingly.
2296 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2297 vd == vd->vdev_top && !vd->vdev_islog &&
2298 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2299 zio->io_txg != spa_syncing_txg(spa)) {
2300 uint64_t old = spa->spa_last_io;
2301 uint64_t new = ddi_get_lbolt64();
2303 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2306 align = 1ULL << vd->vdev_top->vdev_ashift;
2308 if (P2PHASE(zio->io_size, align) != 0) {
2309 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2310 char *abuf = zio_buf_alloc(asize);
2311 ASSERT(vd == vd->vdev_top);
2312 if (zio->io_type == ZIO_TYPE_WRITE) {
2313 bcopy(zio->io_data, abuf, zio->io_size);
2314 bzero(abuf + zio->io_size, asize - zio->io_size);
2316 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2319 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2320 ASSERT(P2PHASE(zio->io_size, align) == 0);
2321 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2324 * If this is a repair I/O, and there's no self-healing involved --
2325 * that is, we're just resilvering what we expect to resilver --
2326 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2327 * This prevents spurious resilvering with nested replication.
2328 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2329 * A is out of date, we'll read from C+D, then use the data to
2330 * resilver A+B -- but we don't actually want to resilver B, just A.
2331 * The top-level mirror has no way to know this, so instead we just
2332 * discard unnecessary repairs as we work our way down the vdev tree.
2333 * The same logic applies to any form of nested replication:
2334 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2336 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2337 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2338 zio->io_txg != 0 && /* not a delegated i/o */
2339 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2340 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2341 zio_vdev_io_bypass(zio);
2342 return (ZIO_PIPELINE_CONTINUE);
2345 if (vd->vdev_ops->vdev_op_leaf &&
2346 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2348 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2349 return (ZIO_PIPELINE_CONTINUE);
2351 if ((zio = vdev_queue_io(zio)) == NULL)
2352 return (ZIO_PIPELINE_STOP);
2354 if (!vdev_accessible(vd, zio)) {
2355 zio->io_error = ENXIO;
2357 return (ZIO_PIPELINE_STOP);
2361 return (vd->vdev_ops->vdev_op_io_start(zio));
2365 zio_vdev_io_done(zio_t *zio)
2367 vdev_t *vd = zio->io_vd;
2368 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2369 boolean_t unexpected_error = B_FALSE;
2371 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2372 return (ZIO_PIPELINE_STOP);
2374 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2376 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2378 vdev_queue_io_done(zio);
2380 if (zio->io_type == ZIO_TYPE_WRITE)
2381 vdev_cache_write(zio);
2383 if (zio_injection_enabled && zio->io_error == 0)
2384 zio->io_error = zio_handle_device_injection(vd,
2387 if (zio_injection_enabled && zio->io_error == 0)
2388 zio->io_error = zio_handle_label_injection(zio, EIO);
2390 if (zio->io_error) {
2391 if (!vdev_accessible(vd, zio)) {
2392 zio->io_error = ENXIO;
2394 unexpected_error = B_TRUE;
2399 ops->vdev_op_io_done(zio);
2401 if (unexpected_error)
2402 VERIFY(vdev_probe(vd, zio) == NULL);
2404 return (ZIO_PIPELINE_CONTINUE);
2408 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2409 * disk, and use that to finish the checksum ereport later.
2412 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2413 const void *good_buf)
2415 /* no processing needed */
2416 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2421 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2423 void *buf = zio_buf_alloc(zio->io_size);
2425 bcopy(zio->io_data, buf, zio->io_size);
2427 zcr->zcr_cbinfo = zio->io_size;
2428 zcr->zcr_cbdata = buf;
2429 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2430 zcr->zcr_free = zio_buf_free;
2434 zio_vdev_io_assess(zio_t *zio)
2436 vdev_t *vd = zio->io_vd;
2438 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2439 return (ZIO_PIPELINE_STOP);
2441 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2442 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2444 if (zio->io_vsd != NULL) {
2445 zio->io_vsd_ops->vsd_free(zio);
2449 if (zio_injection_enabled && zio->io_error == 0)
2450 zio->io_error = zio_handle_fault_injection(zio, EIO);
2453 * If the I/O failed, determine whether we should attempt to retry it.
2455 * On retry, we cut in line in the issue queue, since we don't want
2456 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2458 if (zio->io_error && vd == NULL &&
2459 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2460 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2461 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2463 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2464 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2465 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2466 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2467 zio_requeue_io_start_cut_in_line);
2468 return (ZIO_PIPELINE_STOP);
2472 * If we got an error on a leaf device, convert it to ENXIO
2473 * if the device is not accessible at all.
2475 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2476 !vdev_accessible(vd, zio))
2477 zio->io_error = ENXIO;
2480 * If we can't write to an interior vdev (mirror or RAID-Z),
2481 * set vdev_cant_write so that we stop trying to allocate from it.
2483 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2484 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2485 vd->vdev_cant_write = B_TRUE;
2488 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2490 return (ZIO_PIPELINE_CONTINUE);
2494 zio_vdev_io_reissue(zio_t *zio)
2496 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2497 ASSERT(zio->io_error == 0);
2499 zio->io_stage >>= 1;
2503 zio_vdev_io_redone(zio_t *zio)
2505 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2507 zio->io_stage >>= 1;
2511 zio_vdev_io_bypass(zio_t *zio)
2513 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2514 ASSERT(zio->io_error == 0);
2516 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2517 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2521 * ==========================================================================
2522 * Generate and verify checksums
2523 * ==========================================================================
2526 zio_checksum_generate(zio_t *zio)
2528 blkptr_t *bp = zio->io_bp;
2529 enum zio_checksum checksum;
2533 * This is zio_write_phys().
2534 * We're either generating a label checksum, or none at all.
2536 checksum = zio->io_prop.zp_checksum;
2538 if (checksum == ZIO_CHECKSUM_OFF)
2539 return (ZIO_PIPELINE_CONTINUE);
2541 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2543 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2544 ASSERT(!IO_IS_ALLOCATING(zio));
2545 checksum = ZIO_CHECKSUM_GANG_HEADER;
2547 checksum = BP_GET_CHECKSUM(bp);
2551 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2553 return (ZIO_PIPELINE_CONTINUE);
2557 zio_checksum_verify(zio_t *zio)
2559 zio_bad_cksum_t info;
2560 blkptr_t *bp = zio->io_bp;
2563 ASSERT(zio->io_vd != NULL);
2567 * This is zio_read_phys().
2568 * We're either verifying a label checksum, or nothing at all.
2570 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2571 return (ZIO_PIPELINE_CONTINUE);
2573 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2576 if ((error = zio_checksum_error(zio, &info)) != 0) {
2577 zio->io_error = error;
2578 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2579 zfs_ereport_start_checksum(zio->io_spa,
2580 zio->io_vd, zio, zio->io_offset,
2581 zio->io_size, NULL, &info);
2585 return (ZIO_PIPELINE_CONTINUE);
2589 * Called by RAID-Z to ensure we don't compute the checksum twice.
2592 zio_checksum_verified(zio_t *zio)
2594 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2598 * ==========================================================================
2599 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2600 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2601 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2602 * indicate errors that are specific to one I/O, and most likely permanent.
2603 * Any other error is presumed to be worse because we weren't expecting it.
2604 * ==========================================================================
2607 zio_worst_error(int e1, int e2)
2609 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2612 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2613 if (e1 == zio_error_rank[r1])
2616 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2617 if (e2 == zio_error_rank[r2])
2620 return (r1 > r2 ? e1 : e2);
2624 * ==========================================================================
2626 * ==========================================================================
2629 zio_ready(zio_t *zio)
2631 blkptr_t *bp = zio->io_bp;
2632 zio_t *pio, *pio_next;
2634 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2635 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2636 return (ZIO_PIPELINE_STOP);
2638 if (zio->io_ready) {
2639 ASSERT(IO_IS_ALLOCATING(zio));
2640 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2641 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2646 if (bp != NULL && bp != &zio->io_bp_copy)
2647 zio->io_bp_copy = *bp;
2650 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2652 mutex_enter(&zio->io_lock);
2653 zio->io_state[ZIO_WAIT_READY] = 1;
2654 pio = zio_walk_parents(zio);
2655 mutex_exit(&zio->io_lock);
2658 * As we notify zio's parents, new parents could be added.
2659 * New parents go to the head of zio's io_parent_list, however,
2660 * so we will (correctly) not notify them. The remainder of zio's
2661 * io_parent_list, from 'pio_next' onward, cannot change because
2662 * all parents must wait for us to be done before they can be done.
2664 for (; pio != NULL; pio = pio_next) {
2665 pio_next = zio_walk_parents(zio);
2666 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2669 if (zio->io_flags & ZIO_FLAG_NODATA) {
2670 if (BP_IS_GANG(bp)) {
2671 zio->io_flags &= ~ZIO_FLAG_NODATA;
2673 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2674 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2678 if (zio_injection_enabled &&
2679 zio->io_spa->spa_syncing_txg == zio->io_txg)
2680 zio_handle_ignored_writes(zio);
2682 return (ZIO_PIPELINE_CONTINUE);
2686 zio_done(zio_t *zio)
2688 zio_t *pio, *pio_next;
2692 * If our children haven't all completed,
2693 * wait for them and then repeat this pipeline stage.
2695 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2696 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2697 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2698 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2699 return (ZIO_PIPELINE_STOP);
2701 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2702 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2703 ASSERT(zio->io_children[c][w] == 0);
2705 if (zio->io_bp != NULL) {
2706 ASSERT(zio->io_bp->blk_pad[0] == 0);
2707 ASSERT(zio->io_bp->blk_pad[1] == 0);
2708 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2709 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2710 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2711 zio->io_bp_override == NULL &&
2712 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2713 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2714 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2715 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2716 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2721 * If there were child vdev/gang/ddt errors, they apply to us now.
2723 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2724 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2725 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2728 * If the I/O on the transformed data was successful, generate any
2729 * checksum reports now while we still have the transformed data.
2731 if (zio->io_error == 0) {
2732 while (zio->io_cksum_report != NULL) {
2733 zio_cksum_report_t *zcr = zio->io_cksum_report;
2734 uint64_t align = zcr->zcr_align;
2735 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2736 char *abuf = zio->io_data;
2738 if (asize != zio->io_size) {
2739 abuf = zio_buf_alloc(asize);
2740 bcopy(zio->io_data, abuf, zio->io_size);
2741 bzero(abuf + zio->io_size, asize - zio->io_size);
2744 zio->io_cksum_report = zcr->zcr_next;
2745 zcr->zcr_next = NULL;
2746 zcr->zcr_finish(zcr, abuf);
2747 zfs_ereport_free_checksum(zcr);
2749 if (asize != zio->io_size)
2750 zio_buf_free(abuf, asize);
2754 zio_pop_transforms(zio); /* note: may set zio->io_error */
2756 vdev_stat_update(zio, zio->io_size);
2759 * If this I/O is attached to a particular vdev is slow, exeeding
2760 * 30 seconds to complete, post an error described the I/O delay.
2761 * We ignore these errors if the device is currently unavailable.
2763 if (zio->io_delay >= zio_delay_max) {
2764 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2765 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2766 zio->io_vd, zio, 0, 0);
2769 if (zio->io_error) {
2771 * If this I/O is attached to a particular vdev,
2772 * generate an error message describing the I/O failure
2773 * at the block level. We ignore these errors if the
2774 * device is currently unavailable.
2776 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2777 !vdev_is_dead(zio->io_vd))
2778 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2779 zio->io_vd, zio, 0, 0);
2781 if ((zio->io_error == EIO || !(zio->io_flags &
2782 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2783 zio == zio->io_logical) {
2785 * For logical I/O requests, tell the SPA to log the
2786 * error and generate a logical data ereport.
2788 spa_log_error(zio->io_spa, zio);
2789 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2794 if (zio->io_error && zio == zio->io_logical) {
2796 * Determine whether zio should be reexecuted. This will
2797 * propagate all the way to the root via zio_notify_parent().
2799 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2800 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2802 if (IO_IS_ALLOCATING(zio) &&
2803 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2804 if (zio->io_error != ENOSPC)
2805 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2807 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2810 if ((zio->io_type == ZIO_TYPE_READ ||
2811 zio->io_type == ZIO_TYPE_FREE) &&
2812 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2813 zio->io_error == ENXIO &&
2814 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2815 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2816 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2818 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2819 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2822 * Here is a possibly good place to attempt to do
2823 * either combinatorial reconstruction or error correction
2824 * based on checksums. It also might be a good place
2825 * to send out preliminary ereports before we suspend
2831 * If there were logical child errors, they apply to us now.
2832 * We defer this until now to avoid conflating logical child
2833 * errors with errors that happened to the zio itself when
2834 * updating vdev stats and reporting FMA events above.
2836 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2838 if ((zio->io_error || zio->io_reexecute) &&
2839 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2840 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2841 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2843 zio_gang_tree_free(&zio->io_gang_tree);
2846 * Godfather I/Os should never suspend.
2848 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2849 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2850 zio->io_reexecute = 0;
2852 if (zio->io_reexecute) {
2854 * This is a logical I/O that wants to reexecute.
2856 * Reexecute is top-down. When an i/o fails, if it's not
2857 * the root, it simply notifies its parent and sticks around.
2858 * The parent, seeing that it still has children in zio_done(),
2859 * does the same. This percolates all the way up to the root.
2860 * The root i/o will reexecute or suspend the entire tree.
2862 * This approach ensures that zio_reexecute() honors
2863 * all the original i/o dependency relationships, e.g.
2864 * parents not executing until children are ready.
2866 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2868 zio->io_gang_leader = NULL;
2870 mutex_enter(&zio->io_lock);
2871 zio->io_state[ZIO_WAIT_DONE] = 1;
2872 mutex_exit(&zio->io_lock);
2875 * "The Godfather" I/O monitors its children but is
2876 * not a true parent to them. It will track them through
2877 * the pipeline but severs its ties whenever they get into
2878 * trouble (e.g. suspended). This allows "The Godfather"
2879 * I/O to return status without blocking.
2881 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2882 zio_link_t *zl = zio->io_walk_link;
2883 pio_next = zio_walk_parents(zio);
2885 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2886 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2887 zio_remove_child(pio, zio, zl);
2888 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2892 if ((pio = zio_unique_parent(zio)) != NULL) {
2894 * We're not a root i/o, so there's nothing to do
2895 * but notify our parent. Don't propagate errors
2896 * upward since we haven't permanently failed yet.
2898 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2899 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2900 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2901 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2903 * We'd fail again if we reexecuted now, so suspend
2904 * until conditions improve (e.g. device comes online).
2906 zio_suspend(zio->io_spa, zio);
2909 * Reexecution is potentially a huge amount of work.
2910 * Hand it off to the otherwise-unused claim taskq.
2912 (void) taskq_dispatch(
2913 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2914 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2916 return (ZIO_PIPELINE_STOP);
2919 ASSERT(zio->io_child_count == 0);
2920 ASSERT(zio->io_reexecute == 0);
2921 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2924 * Report any checksum errors, since the I/O is complete.
2926 while (zio->io_cksum_report != NULL) {
2927 zio_cksum_report_t *zcr = zio->io_cksum_report;
2928 zio->io_cksum_report = zcr->zcr_next;
2929 zcr->zcr_next = NULL;
2930 zcr->zcr_finish(zcr, NULL);
2931 zfs_ereport_free_checksum(zcr);
2935 * It is the responsibility of the done callback to ensure that this
2936 * particular zio is no longer discoverable for adoption, and as
2937 * such, cannot acquire any new parents.
2942 mutex_enter(&zio->io_lock);
2943 zio->io_state[ZIO_WAIT_DONE] = 1;
2944 mutex_exit(&zio->io_lock);
2946 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2947 zio_link_t *zl = zio->io_walk_link;
2948 pio_next = zio_walk_parents(zio);
2949 zio_remove_child(pio, zio, zl);
2950 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2953 if (zio->io_waiter != NULL) {
2954 mutex_enter(&zio->io_lock);
2955 zio->io_executor = NULL;
2956 cv_broadcast(&zio->io_cv);
2957 mutex_exit(&zio->io_lock);
2962 return (ZIO_PIPELINE_STOP);
2966 * ==========================================================================
2967 * I/O pipeline definition
2968 * ==========================================================================
2970 static zio_pipe_stage_t *zio_pipeline[] = {
2976 zio_checksum_generate,
2990 zio_checksum_verify,
2994 #if defined(_KERNEL) && defined(HAVE_SPL)
2995 /* Fault injection */
2996 EXPORT_SYMBOL(zio_injection_enabled);
2997 EXPORT_SYMBOL(zio_inject_fault);
2998 EXPORT_SYMBOL(zio_inject_list_next);
2999 EXPORT_SYMBOL(zio_clear_fault);
3000 EXPORT_SYMBOL(zio_handle_fault_injection);
3001 EXPORT_SYMBOL(zio_handle_device_injection);
3002 EXPORT_SYMBOL(zio_handle_label_injection);
3003 EXPORT_SYMBOL(zio_priority_table);
3004 EXPORT_SYMBOL(zio_type_name);
3006 module_param(zio_bulk_flags, int, 0644);
3007 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3009 module_param(zio_delay_max, int, 0644);
3010 MODULE_PARM_DESC(zio_delay_max, "Max zio delay before posting an event (ms)");