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.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/trim_map.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
44 SYSCTL_DECL(_vfs_zfs);
45 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
46 #if defined(__amd64__)
47 static int zio_use_uma = 1;
49 static int zio_use_uma = 0;
51 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
52 "Use uma(9) for ZIO allocations");
53 static int zio_exclude_metadata = 0;
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
55 "Exclude metadata buffers from dumps as well");
57 zio_trim_stats_t zio_trim_stats = {
58 { "bytes", KSTAT_DATA_UINT64,
59 "Number of bytes successfully TRIMmed" },
60 { "success", KSTAT_DATA_UINT64,
61 "Number of successful TRIM requests" },
62 { "unsupported", KSTAT_DATA_UINT64,
63 "Number of TRIM requests that failed because TRIM is not supported" },
64 { "failed", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed for reasons other than not supported" },
68 static kstat_t *zio_trim_ksp;
71 * ==========================================================================
72 * I/O type descriptions
73 * ==========================================================================
75 const char *zio_type_name[ZIO_TYPES] = {
76 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
81 * ==========================================================================
83 * ==========================================================================
85 kmem_cache_t *zio_cache;
86 kmem_cache_t *zio_link_cache;
87 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
91 extern vmem_t *zio_alloc_arena;
94 #define ZIO_PIPELINE_CONTINUE 0x100
95 #define ZIO_PIPELINE_STOP 0x101
98 * The following actions directly effect the spa's sync-to-convergence logic.
99 * The values below define the sync pass when we start performing the action.
100 * Care should be taken when changing these values as they directly impact
101 * spa_sync() performance. Tuning these values may introduce subtle performance
102 * pathologies and should only be done in the context of performance analysis.
103 * These tunables will eventually be removed and replaced with #defines once
104 * enough analysis has been done to determine optimal values.
106 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
107 * regular blocks are not deferred.
109 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
110 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
111 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
112 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
113 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
114 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
115 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
116 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
117 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
120 * An allocating zio is one that either currently has the DVA allocate
121 * stage set or will have it later in its lifetime.
123 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
125 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
128 int zio_buf_debug_limit = 16384;
130 int zio_buf_debug_limit = 0;
137 zio_cache = kmem_cache_create("zio_cache",
138 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
139 zio_link_cache = kmem_cache_create("zio_link_cache",
140 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
145 * For small buffers, we want a cache for each multiple of
146 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
147 * for each quarter-power of 2.
149 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
150 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
153 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
161 * If we are using watchpoints, put each buffer on its own page,
162 * to eliminate the performance overhead of trapping to the
163 * kernel when modifying a non-watched buffer that shares the
164 * page with a watched buffer.
166 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
170 if (size <= 4 * SPA_MINBLOCKSIZE) {
171 align = SPA_MINBLOCKSIZE;
172 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
173 align = MIN(p2 >> 2, PAGESIZE);
178 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
179 zio_buf_cache[c] = kmem_cache_create(name, size,
180 align, NULL, NULL, NULL, NULL, NULL, cflags);
183 * Since zio_data bufs do not appear in crash dumps, we
184 * pass KMC_NOTOUCH so that no allocator metadata is
185 * stored with the buffers.
187 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
188 zio_data_buf_cache[c] = kmem_cache_create(name, size,
189 align, NULL, NULL, NULL, NULL, NULL,
190 cflags | KMC_NOTOUCH | KMC_NODEBUG);
195 ASSERT(zio_buf_cache[c] != NULL);
196 if (zio_buf_cache[c - 1] == NULL)
197 zio_buf_cache[c - 1] = zio_buf_cache[c];
199 ASSERT(zio_data_buf_cache[c] != NULL);
200 if (zio_data_buf_cache[c - 1] == NULL)
201 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
207 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
209 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
212 if (zio_trim_ksp != NULL) {
213 zio_trim_ksp->ks_data = &zio_trim_stats;
214 kstat_install(zio_trim_ksp);
222 kmem_cache_t *last_cache = NULL;
223 kmem_cache_t *last_data_cache = NULL;
225 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
226 if (zio_buf_cache[c] != last_cache) {
227 last_cache = zio_buf_cache[c];
228 kmem_cache_destroy(zio_buf_cache[c]);
230 zio_buf_cache[c] = NULL;
232 if (zio_data_buf_cache[c] != last_data_cache) {
233 last_data_cache = zio_data_buf_cache[c];
234 kmem_cache_destroy(zio_data_buf_cache[c]);
236 zio_data_buf_cache[c] = NULL;
239 kmem_cache_destroy(zio_link_cache);
240 kmem_cache_destroy(zio_cache);
244 if (zio_trim_ksp != NULL) {
245 kstat_delete(zio_trim_ksp);
251 * ==========================================================================
252 * Allocate and free I/O buffers
253 * ==========================================================================
257 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
258 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
259 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
260 * excess / transient data in-core during a crashdump.
263 zio_buf_alloc(size_t size)
265 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
266 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
268 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
271 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
273 return (kmem_alloc(size, KM_SLEEP|flags));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size)
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
287 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
290 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
292 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
296 zio_buf_free(void *buf, size_t size)
298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 kmem_cache_free(zio_buf_cache[c], buf);
305 kmem_free(buf, size);
309 zio_data_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316 kmem_cache_free(zio_data_buf_cache[c], buf);
318 kmem_free(buf, size);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
328 zio_transform_func_t *transform)
330 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
332 zt->zt_orig_data = zio->io_data;
333 zt->zt_orig_size = zio->io_size;
334 zt->zt_bufsize = bufsize;
335 zt->zt_transform = transform;
337 zt->zt_next = zio->io_transform_stack;
338 zio->io_transform_stack = zt;
345 zio_pop_transforms(zio_t *zio)
349 while ((zt = zio->io_transform_stack) != NULL) {
350 if (zt->zt_transform != NULL)
351 zt->zt_transform(zio,
352 zt->zt_orig_data, zt->zt_orig_size);
354 if (zt->zt_bufsize != 0)
355 zio_buf_free(zio->io_data, zt->zt_bufsize);
357 zio->io_data = zt->zt_orig_data;
358 zio->io_size = zt->zt_orig_size;
359 zio->io_transform_stack = zt->zt_next;
361 kmem_free(zt, sizeof (zio_transform_t));
366 * ==========================================================================
367 * I/O transform callbacks for subblocks and decompression
368 * ==========================================================================
371 zio_subblock(zio_t *zio, void *data, uint64_t size)
373 ASSERT(zio->io_size > size);
375 if (zio->io_type == ZIO_TYPE_READ)
376 bcopy(zio->io_data, data, size);
380 zio_decompress(zio_t *zio, void *data, uint64_t size)
382 if (zio->io_error == 0 &&
383 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
384 zio->io_data, data, zio->io_size, size) != 0)
385 zio->io_error = SET_ERROR(EIO);
389 * ==========================================================================
390 * I/O parent/child relationships and pipeline interlocks
391 * ==========================================================================
394 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
395 * continue calling these functions until they return NULL.
396 * Otherwise, the next caller will pick up the list walk in
397 * some indeterminate state. (Otherwise every caller would
398 * have to pass in a cookie to keep the state represented by
399 * io_walk_link, which gets annoying.)
402 zio_walk_parents(zio_t *cio)
404 zio_link_t *zl = cio->io_walk_link;
405 list_t *pl = &cio->io_parent_list;
407 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
408 cio->io_walk_link = zl;
413 ASSERT(zl->zl_child == cio);
414 return (zl->zl_parent);
418 zio_walk_children(zio_t *pio)
420 zio_link_t *zl = pio->io_walk_link;
421 list_t *cl = &pio->io_child_list;
423 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
424 pio->io_walk_link = zl;
429 ASSERT(zl->zl_parent == pio);
430 return (zl->zl_child);
434 zio_unique_parent(zio_t *cio)
436 zio_t *pio = zio_walk_parents(cio);
438 VERIFY(zio_walk_parents(cio) == NULL);
443 zio_add_child(zio_t *pio, zio_t *cio)
445 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
453 ASSERT(cio->io_child_type <= pio->io_child_type);
458 mutex_enter(&cio->io_lock);
459 mutex_enter(&pio->io_lock);
461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
463 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
466 list_insert_head(&pio->io_child_list, zl);
467 list_insert_head(&cio->io_parent_list, zl);
469 pio->io_child_count++;
470 cio->io_parent_count++;
472 mutex_exit(&pio->io_lock);
473 mutex_exit(&cio->io_lock);
477 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
479 ASSERT(zl->zl_parent == pio);
480 ASSERT(zl->zl_child == cio);
482 mutex_enter(&cio->io_lock);
483 mutex_enter(&pio->io_lock);
485 list_remove(&pio->io_child_list, zl);
486 list_remove(&cio->io_parent_list, zl);
488 pio->io_child_count--;
489 cio->io_parent_count--;
491 mutex_exit(&pio->io_lock);
492 mutex_exit(&cio->io_lock);
494 kmem_cache_free(zio_link_cache, zl);
498 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
500 uint64_t *countp = &zio->io_children[child][wait];
501 boolean_t waiting = B_FALSE;
503 mutex_enter(&zio->io_lock);
504 ASSERT(zio->io_stall == NULL);
507 zio->io_stall = countp;
510 mutex_exit(&zio->io_lock);
516 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
518 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
519 int *errorp = &pio->io_child_error[zio->io_child_type];
521 mutex_enter(&pio->io_lock);
522 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
523 *errorp = zio_worst_error(*errorp, zio->io_error);
524 pio->io_reexecute |= zio->io_reexecute;
525 ASSERT3U(*countp, >, 0);
529 if (*countp == 0 && pio->io_stall == countp) {
530 pio->io_stall = NULL;
531 mutex_exit(&pio->io_lock);
534 mutex_exit(&pio->io_lock);
539 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
541 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
542 zio->io_error = zio->io_child_error[c];
546 * ==========================================================================
547 * Create the various types of I/O (read, write, free, etc)
548 * ==========================================================================
551 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
552 void *data, uint64_t size, zio_done_func_t *done, void *private,
553 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
554 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
555 enum zio_stage stage, enum zio_stage pipeline)
559 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
560 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
561 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
563 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
564 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
565 ASSERT(vd || stage == ZIO_STAGE_OPEN);
567 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
568 bzero(zio, sizeof (zio_t));
570 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
571 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
573 list_create(&zio->io_parent_list, sizeof (zio_link_t),
574 offsetof(zio_link_t, zl_parent_node));
575 list_create(&zio->io_child_list, sizeof (zio_link_t),
576 offsetof(zio_link_t, zl_child_node));
579 zio->io_child_type = ZIO_CHILD_VDEV;
580 else if (flags & ZIO_FLAG_GANG_CHILD)
581 zio->io_child_type = ZIO_CHILD_GANG;
582 else if (flags & ZIO_FLAG_DDT_CHILD)
583 zio->io_child_type = ZIO_CHILD_DDT;
585 zio->io_child_type = ZIO_CHILD_LOGICAL;
588 zio->io_bp = (blkptr_t *)bp;
589 zio->io_bp_copy = *bp;
590 zio->io_bp_orig = *bp;
591 if (type != ZIO_TYPE_WRITE ||
592 zio->io_child_type == ZIO_CHILD_DDT)
593 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
594 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
595 zio->io_logical = zio;
596 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
597 pipeline |= ZIO_GANG_STAGES;
603 zio->io_private = private;
605 zio->io_priority = priority;
607 zio->io_offset = offset;
608 zio->io_orig_data = zio->io_data = data;
609 zio->io_orig_size = zio->io_size = size;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
614 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
615 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
618 zio->io_bookmark = *zb;
621 if (zio->io_logical == NULL)
622 zio->io_logical = pio->io_logical;
623 if (zio->io_child_type == ZIO_CHILD_GANG)
624 zio->io_gang_leader = pio->io_gang_leader;
625 zio_add_child(pio, zio);
632 zio_destroy(zio_t *zio)
634 list_destroy(&zio->io_parent_list);
635 list_destroy(&zio->io_child_list);
636 mutex_destroy(&zio->io_lock);
637 cv_destroy(&zio->io_cv);
638 kmem_cache_free(zio_cache, zio);
642 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
643 void *private, enum zio_flag flags)
647 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
648 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
649 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
655 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
657 return (zio_null(NULL, spa, NULL, done, private, flags));
661 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
663 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
664 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
665 bp, (longlong_t)BP_GET_TYPE(bp));
667 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
668 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
669 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
670 bp, (longlong_t)BP_GET_CHECKSUM(bp));
672 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
673 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
674 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
675 bp, (longlong_t)BP_GET_COMPRESS(bp));
677 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
678 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
679 bp, (longlong_t)BP_GET_LSIZE(bp));
681 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
682 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
683 bp, (longlong_t)BP_GET_PSIZE(bp));
686 if (BP_IS_EMBEDDED(bp)) {
687 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
688 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
689 bp, (longlong_t)BPE_GET_ETYPE(bp));
694 * Pool-specific checks.
696 * Note: it would be nice to verify that the blk_birth and
697 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
698 * allows the birth time of log blocks (and dmu_sync()-ed blocks
699 * that are in the log) to be arbitrarily large.
701 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
702 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
703 if (vdevid >= spa->spa_root_vdev->vdev_children) {
704 zfs_panic_recover("blkptr at %p DVA %u has invalid "
706 bp, i, (longlong_t)vdevid);
708 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
710 zfs_panic_recover("blkptr at %p DVA %u has invalid "
712 bp, i, (longlong_t)vdevid);
714 if (vd->vdev_ops == &vdev_hole_ops) {
715 zfs_panic_recover("blkptr at %p DVA %u has hole "
717 bp, i, (longlong_t)vdevid);
720 if (vd->vdev_ops == &vdev_missing_ops) {
722 * "missing" vdevs are valid during import, but we
723 * don't have their detailed info (e.g. asize), so
724 * we can't perform any more checks on them.
728 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
729 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
731 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
732 if (offset + asize > vd->vdev_asize) {
733 zfs_panic_recover("blkptr at %p DVA %u has invalid "
735 bp, i, (longlong_t)offset);
741 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
742 void *data, uint64_t size, zio_done_func_t *done, void *private,
743 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
747 zfs_blkptr_verify(spa, bp);
749 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
750 data, size, done, private,
751 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
752 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
753 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
759 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
760 void *data, uint64_t size, const zio_prop_t *zp,
761 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
763 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
767 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
768 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
769 zp->zp_compress >= ZIO_COMPRESS_OFF &&
770 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
771 DMU_OT_IS_VALID(zp->zp_type) &&
774 zp->zp_copies <= spa_max_replication(spa));
776 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
777 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
778 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
779 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
781 zio->io_ready = ready;
782 zio->io_physdone = physdone;
786 * Data can be NULL if we are going to call zio_write_override() to
787 * provide the already-allocated BP. But we may need the data to
788 * verify a dedup hit (if requested). In this case, don't try to
789 * dedup (just take the already-allocated BP verbatim).
791 if (data == NULL && zio->io_prop.zp_dedup_verify) {
792 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
799 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
800 uint64_t size, zio_done_func_t *done, void *private,
801 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
805 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
806 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
807 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
813 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
815 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
816 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
817 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
818 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
821 * We must reset the io_prop to match the values that existed
822 * when the bp was first written by dmu_sync() keeping in mind
823 * that nopwrite and dedup are mutually exclusive.
825 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
826 zio->io_prop.zp_nopwrite = nopwrite;
827 zio->io_prop.zp_copies = copies;
828 zio->io_bp_override = bp;
832 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
836 * The check for EMBEDDED is a performance optimization. We
837 * process the free here (by ignoring it) rather than
838 * putting it on the list and then processing it in zio_free_sync().
840 if (BP_IS_EMBEDDED(bp))
842 metaslab_check_free(spa, bp);
845 * Frees that are for the currently-syncing txg, are not going to be
846 * deferred, and which will not need to do a read (i.e. not GANG or
847 * DEDUP), can be processed immediately. Otherwise, put them on the
848 * in-memory list for later processing.
850 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
851 txg != spa->spa_syncing_txg ||
852 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
853 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
855 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
856 BP_GET_PSIZE(bp), 0)));
861 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
862 uint64_t size, enum zio_flag flags)
865 enum zio_stage stage = ZIO_FREE_PIPELINE;
867 ASSERT(!BP_IS_HOLE(bp));
868 ASSERT(spa_syncing_txg(spa) == txg);
869 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
871 if (BP_IS_EMBEDDED(bp))
872 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
874 metaslab_check_free(spa, bp);
877 if (zfs_trim_enabled)
878 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
879 ZIO_STAGE_VDEV_IO_ASSESS;
881 * GANG and DEDUP blocks can induce a read (for the gang block header,
882 * or the DDT), so issue them asynchronously so that this thread is
885 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
886 stage |= ZIO_STAGE_ISSUE_ASYNC;
888 flags |= ZIO_FLAG_DONT_QUEUE;
890 zio = zio_create(pio, spa, txg, bp, NULL, size,
891 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
892 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
898 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
899 zio_done_func_t *done, void *private, enum zio_flag flags)
903 dprintf_bp(bp, "claiming in txg %llu", txg);
905 if (BP_IS_EMBEDDED(bp))
906 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
909 * A claim is an allocation of a specific block. Claims are needed
910 * to support immediate writes in the intent log. The issue is that
911 * immediate writes contain committed data, but in a txg that was
912 * *not* committed. Upon opening the pool after an unclean shutdown,
913 * the intent log claims all blocks that contain immediate write data
914 * so that the SPA knows they're in use.
916 * All claims *must* be resolved in the first txg -- before the SPA
917 * starts allocating blocks -- so that nothing is allocated twice.
918 * If txg == 0 we just verify that the block is claimable.
920 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
921 ASSERT(txg == spa_first_txg(spa) || txg == 0);
922 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
924 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
925 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
926 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
932 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
933 uint64_t size, zio_done_func_t *done, void *private,
934 zio_priority_t priority, enum zio_flag flags)
939 if (vd->vdev_children == 0) {
940 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
941 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
942 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
946 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
948 for (c = 0; c < vd->vdev_children; c++)
949 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
950 offset, size, done, private, priority, flags));
957 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
958 void *data, int checksum, zio_done_func_t *done, void *private,
959 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
963 ASSERT(vd->vdev_children == 0);
964 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
965 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
966 ASSERT3U(offset + size, <=, vd->vdev_psize);
968 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
969 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
970 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
972 zio->io_prop.zp_checksum = checksum;
978 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
979 void *data, int checksum, zio_done_func_t *done, void *private,
980 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
984 ASSERT(vd->vdev_children == 0);
985 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
986 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
987 ASSERT3U(offset + size, <=, vd->vdev_psize);
989 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
990 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
991 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
993 zio->io_prop.zp_checksum = checksum;
995 if (zio_checksum_table[checksum].ci_eck) {
997 * zec checksums are necessarily destructive -- they modify
998 * the end of the write buffer to hold the verifier/checksum.
999 * Therefore, we must make a local copy in case the data is
1000 * being written to multiple places in parallel.
1002 void *wbuf = zio_buf_alloc(size);
1003 bcopy(data, wbuf, size);
1004 zio_push_transform(zio, wbuf, size, size, NULL);
1011 * Create a child I/O to do some work for us.
1014 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1015 void *data, uint64_t size, int type, zio_priority_t priority,
1016 enum zio_flag flags, zio_done_func_t *done, void *private)
1018 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1021 ASSERT(vd->vdev_parent ==
1022 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1024 if (type == ZIO_TYPE_READ && bp != NULL) {
1026 * If we have the bp, then the child should perform the
1027 * checksum and the parent need not. This pushes error
1028 * detection as close to the leaves as possible and
1029 * eliminates redundant checksums in the interior nodes.
1031 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1032 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1035 /* Not all IO types require vdev io done stage e.g. free */
1036 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1037 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1039 if (vd->vdev_children == 0)
1040 offset += VDEV_LABEL_START_SIZE;
1042 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1045 * If we've decided to do a repair, the write is not speculative --
1046 * even if the original read was.
1048 if (flags & ZIO_FLAG_IO_REPAIR)
1049 flags &= ~ZIO_FLAG_SPECULATIVE;
1051 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1052 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1053 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1055 zio->io_physdone = pio->io_physdone;
1056 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1057 zio->io_logical->io_phys_children++;
1063 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1064 int type, zio_priority_t priority, enum zio_flag flags,
1065 zio_done_func_t *done, void *private)
1069 ASSERT(vd->vdev_ops->vdev_op_leaf);
1071 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1072 data, size, done, private, type, priority,
1073 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1075 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1081 zio_flush(zio_t *zio, vdev_t *vd)
1083 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1084 NULL, NULL, ZIO_PRIORITY_NOW,
1085 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1089 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1092 ASSERT(vd->vdev_ops->vdev_op_leaf);
1094 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1095 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1096 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1097 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1101 zio_shrink(zio_t *zio, uint64_t size)
1103 ASSERT(zio->io_executor == NULL);
1104 ASSERT(zio->io_orig_size == zio->io_size);
1105 ASSERT(size <= zio->io_size);
1108 * We don't shrink for raidz because of problems with the
1109 * reconstruction when reading back less than the block size.
1110 * Note, BP_IS_RAIDZ() assumes no compression.
1112 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1113 if (!BP_IS_RAIDZ(zio->io_bp))
1114 zio->io_orig_size = zio->io_size = size;
1118 * ==========================================================================
1119 * Prepare to read and write logical blocks
1120 * ==========================================================================
1124 zio_read_bp_init(zio_t *zio)
1126 blkptr_t *bp = zio->io_bp;
1128 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1129 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1130 !(zio->io_flags & ZIO_FLAG_RAW)) {
1132 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1133 void *cbuf = zio_buf_alloc(psize);
1135 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1138 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1139 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1140 decode_embedded_bp_compressed(bp, zio->io_data);
1142 ASSERT(!BP_IS_EMBEDDED(bp));
1145 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1146 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1148 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1149 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1151 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1152 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1154 return (ZIO_PIPELINE_CONTINUE);
1158 zio_write_bp_init(zio_t *zio)
1160 spa_t *spa = zio->io_spa;
1161 zio_prop_t *zp = &zio->io_prop;
1162 enum zio_compress compress = zp->zp_compress;
1163 blkptr_t *bp = zio->io_bp;
1164 uint64_t lsize = zio->io_size;
1165 uint64_t psize = lsize;
1169 * If our children haven't all reached the ready stage,
1170 * wait for them and then repeat this pipeline stage.
1172 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1173 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1174 return (ZIO_PIPELINE_STOP);
1176 if (!IO_IS_ALLOCATING(zio))
1177 return (ZIO_PIPELINE_CONTINUE);
1179 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1181 if (zio->io_bp_override) {
1182 ASSERT(bp->blk_birth != zio->io_txg);
1183 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1185 *bp = *zio->io_bp_override;
1186 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1188 if (BP_IS_EMBEDDED(bp))
1189 return (ZIO_PIPELINE_CONTINUE);
1192 * If we've been overridden and nopwrite is set then
1193 * set the flag accordingly to indicate that a nopwrite
1194 * has already occurred.
1196 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1197 ASSERT(!zp->zp_dedup);
1198 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1199 return (ZIO_PIPELINE_CONTINUE);
1202 ASSERT(!zp->zp_nopwrite);
1204 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1205 return (ZIO_PIPELINE_CONTINUE);
1207 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1208 zp->zp_dedup_verify);
1210 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1211 BP_SET_DEDUP(bp, 1);
1212 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1213 return (ZIO_PIPELINE_CONTINUE);
1215 zio->io_bp_override = NULL;
1219 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1221 * We're rewriting an existing block, which means we're
1222 * working on behalf of spa_sync(). For spa_sync() to
1223 * converge, it must eventually be the case that we don't
1224 * have to allocate new blocks. But compression changes
1225 * the blocksize, which forces a reallocate, and makes
1226 * convergence take longer. Therefore, after the first
1227 * few passes, stop compressing to ensure convergence.
1229 pass = spa_sync_pass(spa);
1231 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1232 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1233 ASSERT(!BP_GET_DEDUP(bp));
1235 if (pass >= zfs_sync_pass_dont_compress)
1236 compress = ZIO_COMPRESS_OFF;
1238 /* Make sure someone doesn't change their mind on overwrites */
1239 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1240 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1243 if (compress != ZIO_COMPRESS_OFF) {
1244 void *cbuf = zio_buf_alloc(lsize);
1245 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1246 if (psize == 0 || psize == lsize) {
1247 compress = ZIO_COMPRESS_OFF;
1248 zio_buf_free(cbuf, lsize);
1249 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1250 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1251 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1252 encode_embedded_bp_compressed(bp,
1253 cbuf, compress, lsize, psize);
1254 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1255 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1256 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1257 zio_buf_free(cbuf, lsize);
1258 bp->blk_birth = zio->io_txg;
1259 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1260 ASSERT(spa_feature_is_active(spa,
1261 SPA_FEATURE_EMBEDDED_DATA));
1262 return (ZIO_PIPELINE_CONTINUE);
1265 * Round up compressed size to MINBLOCKSIZE and
1269 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1270 if (rounded > psize) {
1271 bzero((char *)cbuf + psize, rounded - psize);
1274 if (psize == lsize) {
1275 compress = ZIO_COMPRESS_OFF;
1276 zio_buf_free(cbuf, lsize);
1278 zio_push_transform(zio, cbuf,
1279 psize, lsize, NULL);
1285 * The final pass of spa_sync() must be all rewrites, but the first
1286 * few passes offer a trade-off: allocating blocks defers convergence,
1287 * but newly allocated blocks are sequential, so they can be written
1288 * to disk faster. Therefore, we allow the first few passes of
1289 * spa_sync() to allocate new blocks, but force rewrites after that.
1290 * There should only be a handful of blocks after pass 1 in any case.
1292 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1293 BP_GET_PSIZE(bp) == psize &&
1294 pass >= zfs_sync_pass_rewrite) {
1296 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1297 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1298 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1301 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1305 if (zio->io_bp_orig.blk_birth != 0 &&
1306 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1307 BP_SET_LSIZE(bp, lsize);
1308 BP_SET_TYPE(bp, zp->zp_type);
1309 BP_SET_LEVEL(bp, zp->zp_level);
1310 BP_SET_BIRTH(bp, zio->io_txg, 0);
1312 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1314 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1315 BP_SET_LSIZE(bp, lsize);
1316 BP_SET_TYPE(bp, zp->zp_type);
1317 BP_SET_LEVEL(bp, zp->zp_level);
1318 BP_SET_PSIZE(bp, psize);
1319 BP_SET_COMPRESS(bp, compress);
1320 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1321 BP_SET_DEDUP(bp, zp->zp_dedup);
1322 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1324 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1325 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1326 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1328 if (zp->zp_nopwrite) {
1329 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1330 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1331 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1335 return (ZIO_PIPELINE_CONTINUE);
1339 zio_free_bp_init(zio_t *zio)
1341 blkptr_t *bp = zio->io_bp;
1343 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1344 if (BP_GET_DEDUP(bp))
1345 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1348 return (ZIO_PIPELINE_CONTINUE);
1352 * ==========================================================================
1353 * Execute the I/O pipeline
1354 * ==========================================================================
1358 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1360 spa_t *spa = zio->io_spa;
1361 zio_type_t t = zio->io_type;
1362 int flags = (cutinline ? TQ_FRONT : 0);
1364 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1367 * If we're a config writer or a probe, the normal issue and
1368 * interrupt threads may all be blocked waiting for the config lock.
1369 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1371 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1375 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1377 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1381 * If this is a high priority I/O, then use the high priority taskq if
1384 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1385 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1388 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1391 * NB: We are assuming that the zio can only be dispatched
1392 * to a single taskq at a time. It would be a grievous error
1393 * to dispatch the zio to another taskq at the same time.
1395 #if defined(illumos) || !defined(_KERNEL)
1396 ASSERT(zio->io_tqent.tqent_next == NULL);
1398 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1400 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1401 flags, &zio->io_tqent);
1405 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1407 kthread_t *executor = zio->io_executor;
1408 spa_t *spa = zio->io_spa;
1410 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1411 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1413 for (i = 0; i < tqs->stqs_count; i++) {
1414 if (taskq_member(tqs->stqs_taskq[i], executor))
1423 zio_issue_async(zio_t *zio)
1425 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1427 return (ZIO_PIPELINE_STOP);
1431 zio_interrupt(zio_t *zio)
1433 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1437 * Execute the I/O pipeline until one of the following occurs:
1439 * (1) the I/O completes
1440 * (2) the pipeline stalls waiting for dependent child I/Os
1441 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1442 * (4) the I/O is delegated by vdev-level caching or aggregation
1443 * (5) the I/O is deferred due to vdev-level queueing
1444 * (6) the I/O is handed off to another thread.
1446 * In all cases, the pipeline stops whenever there's no CPU work; it never
1447 * burns a thread in cv_wait().
1449 * There's no locking on io_stage because there's no legitimate way
1450 * for multiple threads to be attempting to process the same I/O.
1452 static zio_pipe_stage_t *zio_pipeline[];
1455 zio_execute(zio_t *zio)
1457 zio->io_executor = curthread;
1459 while (zio->io_stage < ZIO_STAGE_DONE) {
1460 enum zio_stage pipeline = zio->io_pipeline;
1461 enum zio_stage stage = zio->io_stage;
1464 ASSERT(!MUTEX_HELD(&zio->io_lock));
1465 ASSERT(ISP2(stage));
1466 ASSERT(zio->io_stall == NULL);
1470 } while ((stage & pipeline) == 0);
1472 ASSERT(stage <= ZIO_STAGE_DONE);
1475 * If we are in interrupt context and this pipeline stage
1476 * will grab a config lock that is held across I/O,
1477 * or may wait for an I/O that needs an interrupt thread
1478 * to complete, issue async to avoid deadlock.
1480 * For VDEV_IO_START, we cut in line so that the io will
1481 * be sent to disk promptly.
1483 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1484 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1485 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1486 zio_requeue_io_start_cut_in_line : B_FALSE;
1487 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1491 zio->io_stage = stage;
1492 rv = zio_pipeline[highbit64(stage) - 1](zio);
1494 if (rv == ZIO_PIPELINE_STOP)
1497 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1502 * ==========================================================================
1503 * Initiate I/O, either sync or async
1504 * ==========================================================================
1507 zio_wait(zio_t *zio)
1511 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1512 ASSERT(zio->io_executor == NULL);
1514 zio->io_waiter = curthread;
1518 mutex_enter(&zio->io_lock);
1519 while (zio->io_executor != NULL)
1520 cv_wait(&zio->io_cv, &zio->io_lock);
1521 mutex_exit(&zio->io_lock);
1523 error = zio->io_error;
1530 zio_nowait(zio_t *zio)
1532 ASSERT(zio->io_executor == NULL);
1534 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1535 zio_unique_parent(zio) == NULL) {
1537 * This is a logical async I/O with no parent to wait for it.
1538 * We add it to the spa_async_root_zio "Godfather" I/O which
1539 * will ensure they complete prior to unloading the pool.
1541 spa_t *spa = zio->io_spa;
1543 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1550 * ==========================================================================
1551 * Reexecute or suspend/resume failed I/O
1552 * ==========================================================================
1556 zio_reexecute(zio_t *pio)
1558 zio_t *cio, *cio_next;
1560 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1561 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1562 ASSERT(pio->io_gang_leader == NULL);
1563 ASSERT(pio->io_gang_tree == NULL);
1565 pio->io_flags = pio->io_orig_flags;
1566 pio->io_stage = pio->io_orig_stage;
1567 pio->io_pipeline = pio->io_orig_pipeline;
1568 pio->io_reexecute = 0;
1569 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1571 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1572 pio->io_state[w] = 0;
1573 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1574 pio->io_child_error[c] = 0;
1576 if (IO_IS_ALLOCATING(pio))
1577 BP_ZERO(pio->io_bp);
1580 * As we reexecute pio's children, new children could be created.
1581 * New children go to the head of pio's io_child_list, however,
1582 * so we will (correctly) not reexecute them. The key is that
1583 * the remainder of pio's io_child_list, from 'cio_next' onward,
1584 * cannot be affected by any side effects of reexecuting 'cio'.
1586 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1587 cio_next = zio_walk_children(pio);
1588 mutex_enter(&pio->io_lock);
1589 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1590 pio->io_children[cio->io_child_type][w]++;
1591 mutex_exit(&pio->io_lock);
1596 * Now that all children have been reexecuted, execute the parent.
1597 * We don't reexecute "The Godfather" I/O here as it's the
1598 * responsibility of the caller to wait on him.
1600 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1605 zio_suspend(spa_t *spa, zio_t *zio)
1607 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1608 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1609 "failure and the failure mode property for this pool "
1610 "is set to panic.", spa_name(spa));
1612 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1614 mutex_enter(&spa->spa_suspend_lock);
1616 if (spa->spa_suspend_zio_root == NULL)
1617 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1618 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1619 ZIO_FLAG_GODFATHER);
1621 spa->spa_suspended = B_TRUE;
1624 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1625 ASSERT(zio != spa->spa_suspend_zio_root);
1626 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1627 ASSERT(zio_unique_parent(zio) == NULL);
1628 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1629 zio_add_child(spa->spa_suspend_zio_root, zio);
1632 mutex_exit(&spa->spa_suspend_lock);
1636 zio_resume(spa_t *spa)
1641 * Reexecute all previously suspended i/o.
1643 mutex_enter(&spa->spa_suspend_lock);
1644 spa->spa_suspended = B_FALSE;
1645 cv_broadcast(&spa->spa_suspend_cv);
1646 pio = spa->spa_suspend_zio_root;
1647 spa->spa_suspend_zio_root = NULL;
1648 mutex_exit(&spa->spa_suspend_lock);
1654 return (zio_wait(pio));
1658 zio_resume_wait(spa_t *spa)
1660 mutex_enter(&spa->spa_suspend_lock);
1661 while (spa_suspended(spa))
1662 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1663 mutex_exit(&spa->spa_suspend_lock);
1667 * ==========================================================================
1670 * A gang block is a collection of small blocks that looks to the DMU
1671 * like one large block. When zio_dva_allocate() cannot find a block
1672 * of the requested size, due to either severe fragmentation or the pool
1673 * being nearly full, it calls zio_write_gang_block() to construct the
1674 * block from smaller fragments.
1676 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1677 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1678 * an indirect block: it's an array of block pointers. It consumes
1679 * only one sector and hence is allocatable regardless of fragmentation.
1680 * The gang header's bps point to its gang members, which hold the data.
1682 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1683 * as the verifier to ensure uniqueness of the SHA256 checksum.
1684 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1685 * not the gang header. This ensures that data block signatures (needed for
1686 * deduplication) are independent of how the block is physically stored.
1688 * Gang blocks can be nested: a gang member may itself be a gang block.
1689 * Thus every gang block is a tree in which root and all interior nodes are
1690 * gang headers, and the leaves are normal blocks that contain user data.
1691 * The root of the gang tree is called the gang leader.
1693 * To perform any operation (read, rewrite, free, claim) on a gang block,
1694 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1695 * in the io_gang_tree field of the original logical i/o by recursively
1696 * reading the gang leader and all gang headers below it. This yields
1697 * an in-core tree containing the contents of every gang header and the
1698 * bps for every constituent of the gang block.
1700 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1701 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1702 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1703 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1704 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1705 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1706 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1707 * of the gang header plus zio_checksum_compute() of the data to update the
1708 * gang header's blk_cksum as described above.
1710 * The two-phase assemble/issue model solves the problem of partial failure --
1711 * what if you'd freed part of a gang block but then couldn't read the
1712 * gang header for another part? Assembling the entire gang tree first
1713 * ensures that all the necessary gang header I/O has succeeded before
1714 * starting the actual work of free, claim, or write. Once the gang tree
1715 * is assembled, free and claim are in-memory operations that cannot fail.
1717 * In the event that a gang write fails, zio_dva_unallocate() walks the
1718 * gang tree to immediately free (i.e. insert back into the space map)
1719 * everything we've allocated. This ensures that we don't get ENOSPC
1720 * errors during repeated suspend/resume cycles due to a flaky device.
1722 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1723 * the gang tree, we won't modify the block, so we can safely defer the free
1724 * (knowing that the block is still intact). If we *can* assemble the gang
1725 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1726 * each constituent bp and we can allocate a new block on the next sync pass.
1728 * In all cases, the gang tree allows complete recovery from partial failure.
1729 * ==========================================================================
1733 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1738 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1739 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1740 &pio->io_bookmark));
1744 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1749 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1750 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1751 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1753 * As we rewrite each gang header, the pipeline will compute
1754 * a new gang block header checksum for it; but no one will
1755 * compute a new data checksum, so we do that here. The one
1756 * exception is the gang leader: the pipeline already computed
1757 * its data checksum because that stage precedes gang assembly.
1758 * (Presently, nothing actually uses interior data checksums;
1759 * this is just good hygiene.)
1761 if (gn != pio->io_gang_leader->io_gang_tree) {
1762 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1763 data, BP_GET_PSIZE(bp));
1766 * If we are here to damage data for testing purposes,
1767 * leave the GBH alone so that we can detect the damage.
1769 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1770 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1772 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1773 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1774 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1782 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1784 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1785 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1786 ZIO_GANG_CHILD_FLAGS(pio)));
1791 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1793 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1794 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1797 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1806 static void zio_gang_tree_assemble_done(zio_t *zio);
1808 static zio_gang_node_t *
1809 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1811 zio_gang_node_t *gn;
1813 ASSERT(*gnpp == NULL);
1815 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1816 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1823 zio_gang_node_free(zio_gang_node_t **gnpp)
1825 zio_gang_node_t *gn = *gnpp;
1827 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1828 ASSERT(gn->gn_child[g] == NULL);
1830 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1831 kmem_free(gn, sizeof (*gn));
1836 zio_gang_tree_free(zio_gang_node_t **gnpp)
1838 zio_gang_node_t *gn = *gnpp;
1843 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1844 zio_gang_tree_free(&gn->gn_child[g]);
1846 zio_gang_node_free(gnpp);
1850 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1852 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1854 ASSERT(gio->io_gang_leader == gio);
1855 ASSERT(BP_IS_GANG(bp));
1857 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1858 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1859 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1863 zio_gang_tree_assemble_done(zio_t *zio)
1865 zio_t *gio = zio->io_gang_leader;
1866 zio_gang_node_t *gn = zio->io_private;
1867 blkptr_t *bp = zio->io_bp;
1869 ASSERT(gio == zio_unique_parent(zio));
1870 ASSERT(zio->io_child_count == 0);
1875 if (BP_SHOULD_BYTESWAP(bp))
1876 byteswap_uint64_array(zio->io_data, zio->io_size);
1878 ASSERT(zio->io_data == gn->gn_gbh);
1879 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1880 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1882 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1883 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1884 if (!BP_IS_GANG(gbp))
1886 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1891 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1893 zio_t *gio = pio->io_gang_leader;
1896 ASSERT(BP_IS_GANG(bp) == !!gn);
1897 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1898 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1901 * If you're a gang header, your data is in gn->gn_gbh.
1902 * If you're a gang member, your data is in 'data' and gn == NULL.
1904 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1907 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1909 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1910 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1911 if (BP_IS_HOLE(gbp))
1913 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1914 data = (char *)data + BP_GET_PSIZE(gbp);
1918 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1919 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1926 zio_gang_assemble(zio_t *zio)
1928 blkptr_t *bp = zio->io_bp;
1930 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1931 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1933 zio->io_gang_leader = zio;
1935 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1937 return (ZIO_PIPELINE_CONTINUE);
1941 zio_gang_issue(zio_t *zio)
1943 blkptr_t *bp = zio->io_bp;
1945 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1946 return (ZIO_PIPELINE_STOP);
1948 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1949 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1951 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1952 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1954 zio_gang_tree_free(&zio->io_gang_tree);
1956 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1958 return (ZIO_PIPELINE_CONTINUE);
1962 zio_write_gang_member_ready(zio_t *zio)
1964 zio_t *pio = zio_unique_parent(zio);
1965 zio_t *gio = zio->io_gang_leader;
1966 dva_t *cdva = zio->io_bp->blk_dva;
1967 dva_t *pdva = pio->io_bp->blk_dva;
1970 if (BP_IS_HOLE(zio->io_bp))
1973 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1975 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1976 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1977 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1978 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1979 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1981 mutex_enter(&pio->io_lock);
1982 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1983 ASSERT(DVA_GET_GANG(&pdva[d]));
1984 asize = DVA_GET_ASIZE(&pdva[d]);
1985 asize += DVA_GET_ASIZE(&cdva[d]);
1986 DVA_SET_ASIZE(&pdva[d], asize);
1988 mutex_exit(&pio->io_lock);
1992 zio_write_gang_block(zio_t *pio)
1994 spa_t *spa = pio->io_spa;
1995 blkptr_t *bp = pio->io_bp;
1996 zio_t *gio = pio->io_gang_leader;
1998 zio_gang_node_t *gn, **gnpp;
1999 zio_gbh_phys_t *gbh;
2000 uint64_t txg = pio->io_txg;
2001 uint64_t resid = pio->io_size;
2003 int copies = gio->io_prop.zp_copies;
2004 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2008 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2009 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2010 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2012 pio->io_error = error;
2013 return (ZIO_PIPELINE_CONTINUE);
2017 gnpp = &gio->io_gang_tree;
2019 gnpp = pio->io_private;
2020 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2023 gn = zio_gang_node_alloc(gnpp);
2025 bzero(gbh, SPA_GANGBLOCKSIZE);
2028 * Create the gang header.
2030 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2031 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2034 * Create and nowait the gang children.
2036 for (int g = 0; resid != 0; resid -= lsize, g++) {
2037 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2039 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2041 zp.zp_checksum = gio->io_prop.zp_checksum;
2042 zp.zp_compress = ZIO_COMPRESS_OFF;
2043 zp.zp_type = DMU_OT_NONE;
2045 zp.zp_copies = gio->io_prop.zp_copies;
2046 zp.zp_dedup = B_FALSE;
2047 zp.zp_dedup_verify = B_FALSE;
2048 zp.zp_nopwrite = B_FALSE;
2050 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2051 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2052 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2053 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2054 &pio->io_bookmark));
2058 * Set pio's pipeline to just wait for zio to finish.
2060 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2064 return (ZIO_PIPELINE_CONTINUE);
2068 * The zio_nop_write stage in the pipeline determines if allocating
2069 * a new bp is necessary. By leveraging a cryptographically secure checksum,
2070 * such as SHA256, we can compare the checksums of the new data and the old
2071 * to determine if allocating a new block is required. The nopwrite
2072 * feature can handle writes in either syncing or open context (i.e. zil
2073 * writes) and as a result is mutually exclusive with dedup.
2076 zio_nop_write(zio_t *zio)
2078 blkptr_t *bp = zio->io_bp;
2079 blkptr_t *bp_orig = &zio->io_bp_orig;
2080 zio_prop_t *zp = &zio->io_prop;
2082 ASSERT(BP_GET_LEVEL(bp) == 0);
2083 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2084 ASSERT(zp->zp_nopwrite);
2085 ASSERT(!zp->zp_dedup);
2086 ASSERT(zio->io_bp_override == NULL);
2087 ASSERT(IO_IS_ALLOCATING(zio));
2090 * Check to see if the original bp and the new bp have matching
2091 * characteristics (i.e. same checksum, compression algorithms, etc).
2092 * If they don't then just continue with the pipeline which will
2093 * allocate a new bp.
2095 if (BP_IS_HOLE(bp_orig) ||
2096 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2097 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2098 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2099 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2100 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2101 return (ZIO_PIPELINE_CONTINUE);
2104 * If the checksums match then reset the pipeline so that we
2105 * avoid allocating a new bp and issuing any I/O.
2107 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2108 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2109 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2110 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2111 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2112 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2113 sizeof (uint64_t)) == 0);
2116 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2117 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2120 return (ZIO_PIPELINE_CONTINUE);
2124 * ==========================================================================
2126 * ==========================================================================
2129 zio_ddt_child_read_done(zio_t *zio)
2131 blkptr_t *bp = zio->io_bp;
2132 ddt_entry_t *dde = zio->io_private;
2134 zio_t *pio = zio_unique_parent(zio);
2136 mutex_enter(&pio->io_lock);
2137 ddp = ddt_phys_select(dde, bp);
2138 if (zio->io_error == 0)
2139 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2140 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2141 dde->dde_repair_data = zio->io_data;
2143 zio_buf_free(zio->io_data, zio->io_size);
2144 mutex_exit(&pio->io_lock);
2148 zio_ddt_read_start(zio_t *zio)
2150 blkptr_t *bp = zio->io_bp;
2152 ASSERT(BP_GET_DEDUP(bp));
2153 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2154 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2156 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2157 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2158 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2159 ddt_phys_t *ddp = dde->dde_phys;
2160 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2163 ASSERT(zio->io_vsd == NULL);
2166 if (ddp_self == NULL)
2167 return (ZIO_PIPELINE_CONTINUE);
2169 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2170 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2172 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2174 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2175 zio_buf_alloc(zio->io_size), zio->io_size,
2176 zio_ddt_child_read_done, dde, zio->io_priority,
2177 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2178 &zio->io_bookmark));
2180 return (ZIO_PIPELINE_CONTINUE);
2183 zio_nowait(zio_read(zio, zio->io_spa, bp,
2184 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2185 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2187 return (ZIO_PIPELINE_CONTINUE);
2191 zio_ddt_read_done(zio_t *zio)
2193 blkptr_t *bp = zio->io_bp;
2195 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2196 return (ZIO_PIPELINE_STOP);
2198 ASSERT(BP_GET_DEDUP(bp));
2199 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2200 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2202 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2203 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2204 ddt_entry_t *dde = zio->io_vsd;
2206 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2207 return (ZIO_PIPELINE_CONTINUE);
2210 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2211 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2212 return (ZIO_PIPELINE_STOP);
2214 if (dde->dde_repair_data != NULL) {
2215 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2216 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2218 ddt_repair_done(ddt, dde);
2222 ASSERT(zio->io_vsd == NULL);
2224 return (ZIO_PIPELINE_CONTINUE);
2228 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2230 spa_t *spa = zio->io_spa;
2233 * Note: we compare the original data, not the transformed data,
2234 * because when zio->io_bp is an override bp, we will not have
2235 * pushed the I/O transforms. That's an important optimization
2236 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2238 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2239 zio_t *lio = dde->dde_lead_zio[p];
2242 return (lio->io_orig_size != zio->io_orig_size ||
2243 bcmp(zio->io_orig_data, lio->io_orig_data,
2244 zio->io_orig_size) != 0);
2248 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2249 ddt_phys_t *ddp = &dde->dde_phys[p];
2251 if (ddp->ddp_phys_birth != 0) {
2252 arc_buf_t *abuf = NULL;
2253 arc_flags_t aflags = ARC_FLAG_WAIT;
2254 blkptr_t blk = *zio->io_bp;
2257 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2261 error = arc_read(NULL, spa, &blk,
2262 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2263 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2264 &aflags, &zio->io_bookmark);
2267 if (arc_buf_size(abuf) != zio->io_orig_size ||
2268 bcmp(abuf->b_data, zio->io_orig_data,
2269 zio->io_orig_size) != 0)
2270 error = SET_ERROR(EEXIST);
2271 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2275 return (error != 0);
2283 zio_ddt_child_write_ready(zio_t *zio)
2285 int p = zio->io_prop.zp_copies;
2286 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2287 ddt_entry_t *dde = zio->io_private;
2288 ddt_phys_t *ddp = &dde->dde_phys[p];
2296 ASSERT(dde->dde_lead_zio[p] == zio);
2298 ddt_phys_fill(ddp, zio->io_bp);
2300 while ((pio = zio_walk_parents(zio)) != NULL)
2301 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2307 zio_ddt_child_write_done(zio_t *zio)
2309 int p = zio->io_prop.zp_copies;
2310 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2311 ddt_entry_t *dde = zio->io_private;
2312 ddt_phys_t *ddp = &dde->dde_phys[p];
2316 ASSERT(ddp->ddp_refcnt == 0);
2317 ASSERT(dde->dde_lead_zio[p] == zio);
2318 dde->dde_lead_zio[p] = NULL;
2320 if (zio->io_error == 0) {
2321 while (zio_walk_parents(zio) != NULL)
2322 ddt_phys_addref(ddp);
2324 ddt_phys_clear(ddp);
2331 zio_ddt_ditto_write_done(zio_t *zio)
2333 int p = DDT_PHYS_DITTO;
2334 zio_prop_t *zp = &zio->io_prop;
2335 blkptr_t *bp = zio->io_bp;
2336 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2337 ddt_entry_t *dde = zio->io_private;
2338 ddt_phys_t *ddp = &dde->dde_phys[p];
2339 ddt_key_t *ddk = &dde->dde_key;
2343 ASSERT(ddp->ddp_refcnt == 0);
2344 ASSERT(dde->dde_lead_zio[p] == zio);
2345 dde->dde_lead_zio[p] = NULL;
2347 if (zio->io_error == 0) {
2348 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2349 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2350 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2351 if (ddp->ddp_phys_birth != 0)
2352 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2353 ddt_phys_fill(ddp, bp);
2360 zio_ddt_write(zio_t *zio)
2362 spa_t *spa = zio->io_spa;
2363 blkptr_t *bp = zio->io_bp;
2364 uint64_t txg = zio->io_txg;
2365 zio_prop_t *zp = &zio->io_prop;
2366 int p = zp->zp_copies;
2370 ddt_t *ddt = ddt_select(spa, bp);
2374 ASSERT(BP_GET_DEDUP(bp));
2375 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2376 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2379 dde = ddt_lookup(ddt, bp, B_TRUE);
2380 ddp = &dde->dde_phys[p];
2382 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2384 * If we're using a weak checksum, upgrade to a strong checksum
2385 * and try again. If we're already using a strong checksum,
2386 * we can't resolve it, so just convert to an ordinary write.
2387 * (And automatically e-mail a paper to Nature?)
2389 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2390 zp->zp_checksum = spa_dedup_checksum(spa);
2391 zio_pop_transforms(zio);
2392 zio->io_stage = ZIO_STAGE_OPEN;
2395 zp->zp_dedup = B_FALSE;
2397 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2399 return (ZIO_PIPELINE_CONTINUE);
2402 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2403 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2405 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2406 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2407 zio_prop_t czp = *zp;
2409 czp.zp_copies = ditto_copies;
2412 * If we arrived here with an override bp, we won't have run
2413 * the transform stack, so we won't have the data we need to
2414 * generate a child i/o. So, toss the override bp and restart.
2415 * This is safe, because using the override bp is just an
2416 * optimization; and it's rare, so the cost doesn't matter.
2418 if (zio->io_bp_override) {
2419 zio_pop_transforms(zio);
2420 zio->io_stage = ZIO_STAGE_OPEN;
2421 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2422 zio->io_bp_override = NULL;
2425 return (ZIO_PIPELINE_CONTINUE);
2428 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2429 zio->io_orig_size, &czp, NULL, NULL,
2430 zio_ddt_ditto_write_done, dde, zio->io_priority,
2431 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2433 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2434 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2437 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2438 if (ddp->ddp_phys_birth != 0)
2439 ddt_bp_fill(ddp, bp, txg);
2440 if (dde->dde_lead_zio[p] != NULL)
2441 zio_add_child(zio, dde->dde_lead_zio[p]);
2443 ddt_phys_addref(ddp);
2444 } else if (zio->io_bp_override) {
2445 ASSERT(bp->blk_birth == txg);
2446 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2447 ddt_phys_fill(ddp, bp);
2448 ddt_phys_addref(ddp);
2450 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2451 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2452 zio_ddt_child_write_done, dde, zio->io_priority,
2453 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2455 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2456 dde->dde_lead_zio[p] = cio;
2466 return (ZIO_PIPELINE_CONTINUE);
2469 ddt_entry_t *freedde; /* for debugging */
2472 zio_ddt_free(zio_t *zio)
2474 spa_t *spa = zio->io_spa;
2475 blkptr_t *bp = zio->io_bp;
2476 ddt_t *ddt = ddt_select(spa, bp);
2480 ASSERT(BP_GET_DEDUP(bp));
2481 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2484 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2485 ddp = ddt_phys_select(dde, bp);
2486 ddt_phys_decref(ddp);
2489 return (ZIO_PIPELINE_CONTINUE);
2493 * ==========================================================================
2494 * Allocate and free blocks
2495 * ==========================================================================
2498 zio_dva_allocate(zio_t *zio)
2500 spa_t *spa = zio->io_spa;
2501 metaslab_class_t *mc = spa_normal_class(spa);
2502 blkptr_t *bp = zio->io_bp;
2506 if (zio->io_gang_leader == NULL) {
2507 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2508 zio->io_gang_leader = zio;
2511 ASSERT(BP_IS_HOLE(bp));
2512 ASSERT0(BP_GET_NDVAS(bp));
2513 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2514 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2515 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2518 * The dump device does not support gang blocks so allocation on
2519 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2520 * the "fast" gang feature.
2522 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2523 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2524 METASLAB_GANG_CHILD : 0;
2525 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2526 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2529 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2530 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2532 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2533 return (zio_write_gang_block(zio));
2534 zio->io_error = error;
2537 return (ZIO_PIPELINE_CONTINUE);
2541 zio_dva_free(zio_t *zio)
2543 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2545 return (ZIO_PIPELINE_CONTINUE);
2549 zio_dva_claim(zio_t *zio)
2553 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2555 zio->io_error = error;
2557 return (ZIO_PIPELINE_CONTINUE);
2561 * Undo an allocation. This is used by zio_done() when an I/O fails
2562 * and we want to give back the block we just allocated.
2563 * This handles both normal blocks and gang blocks.
2566 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2568 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2569 ASSERT(zio->io_bp_override == NULL);
2571 if (!BP_IS_HOLE(bp))
2572 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2575 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2576 zio_dva_unallocate(zio, gn->gn_child[g],
2577 &gn->gn_gbh->zg_blkptr[g]);
2583 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2586 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2587 uint64_t size, boolean_t use_slog)
2591 ASSERT(txg > spa_syncing_txg(spa));
2594 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2595 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2596 * when allocating them.
2599 error = metaslab_alloc(spa, spa_log_class(spa), size,
2600 new_bp, 1, txg, old_bp,
2601 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2605 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2606 new_bp, 1, txg, old_bp,
2607 METASLAB_HINTBP_AVOID);
2611 BP_SET_LSIZE(new_bp, size);
2612 BP_SET_PSIZE(new_bp, size);
2613 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2614 BP_SET_CHECKSUM(new_bp,
2615 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2616 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2617 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2618 BP_SET_LEVEL(new_bp, 0);
2619 BP_SET_DEDUP(new_bp, 0);
2620 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2627 * Free an intent log block.
2630 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2632 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2633 ASSERT(!BP_IS_GANG(bp));
2635 zio_free(spa, txg, bp);
2639 * ==========================================================================
2640 * Read, write and delete to physical devices
2641 * ==========================================================================
2646 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2647 * stops after this stage and will resume upon I/O completion.
2648 * However, there are instances where the vdev layer may need to
2649 * continue the pipeline when an I/O was not issued. Since the I/O
2650 * that was sent to the vdev layer might be different than the one
2651 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2652 * force the underlying vdev layers to call either zio_execute() or
2653 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2656 zio_vdev_io_start(zio_t *zio)
2658 vdev_t *vd = zio->io_vd;
2660 spa_t *spa = zio->io_spa;
2663 ASSERT(zio->io_error == 0);
2664 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2667 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2668 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2671 * The mirror_ops handle multiple DVAs in a single BP.
2673 vdev_mirror_ops.vdev_op_io_start(zio);
2674 return (ZIO_PIPELINE_STOP);
2677 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2678 zio->io_priority == ZIO_PRIORITY_NOW) {
2679 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2680 return (ZIO_PIPELINE_CONTINUE);
2684 * We keep track of time-sensitive I/Os so that the scan thread
2685 * can quickly react to certain workloads. In particular, we care
2686 * about non-scrubbing, top-level reads and writes with the following
2688 * - synchronous writes of user data to non-slog devices
2689 * - any reads of user data
2690 * When these conditions are met, adjust the timestamp of spa_last_io
2691 * which allows the scan thread to adjust its workload accordingly.
2693 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2694 vd == vd->vdev_top && !vd->vdev_islog &&
2695 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2696 zio->io_txg != spa_syncing_txg(spa)) {
2697 uint64_t old = spa->spa_last_io;
2698 uint64_t new = ddi_get_lbolt64();
2700 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2703 align = 1ULL << vd->vdev_top->vdev_ashift;
2705 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2706 P2PHASE(zio->io_size, align) != 0) {
2707 /* Transform logical writes to be a full physical block size. */
2708 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2710 if (zio->io_type == ZIO_TYPE_READ ||
2711 zio->io_type == ZIO_TYPE_WRITE)
2712 abuf = zio_buf_alloc(asize);
2713 ASSERT(vd == vd->vdev_top);
2714 if (zio->io_type == ZIO_TYPE_WRITE) {
2715 bcopy(zio->io_data, abuf, zio->io_size);
2716 bzero(abuf + zio->io_size, asize - zio->io_size);
2718 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2723 * If this is not a physical io, make sure that it is properly aligned
2724 * before proceeding.
2726 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2727 ASSERT0(P2PHASE(zio->io_offset, align));
2728 ASSERT0(P2PHASE(zio->io_size, align));
2731 * For physical writes, we allow 512b aligned writes and assume
2732 * the device will perform a read-modify-write as necessary.
2734 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2735 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2738 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2741 * If this is a repair I/O, and there's no self-healing involved --
2742 * that is, we're just resilvering what we expect to resilver --
2743 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2744 * This prevents spurious resilvering with nested replication.
2745 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2746 * A is out of date, we'll read from C+D, then use the data to
2747 * resilver A+B -- but we don't actually want to resilver B, just A.
2748 * The top-level mirror has no way to know this, so instead we just
2749 * discard unnecessary repairs as we work our way down the vdev tree.
2750 * The same logic applies to any form of nested replication:
2751 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2753 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2754 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2755 zio->io_txg != 0 && /* not a delegated i/o */
2756 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2757 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2758 zio_vdev_io_bypass(zio);
2759 return (ZIO_PIPELINE_CONTINUE);
2762 if (vd->vdev_ops->vdev_op_leaf) {
2763 switch (zio->io_type) {
2765 if (vdev_cache_read(zio))
2766 return (ZIO_PIPELINE_CONTINUE);
2768 case ZIO_TYPE_WRITE:
2770 if ((zio = vdev_queue_io(zio)) == NULL)
2771 return (ZIO_PIPELINE_STOP);
2773 if (!vdev_accessible(vd, zio)) {
2774 zio->io_error = SET_ERROR(ENXIO);
2776 return (ZIO_PIPELINE_STOP);
2781 * Note that we ignore repair writes for TRIM because they can
2782 * conflict with normal writes. This isn't an issue because, by
2783 * definition, we only repair blocks that aren't freed.
2785 if (zio->io_type == ZIO_TYPE_WRITE &&
2786 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2787 !trim_map_write_start(zio))
2788 return (ZIO_PIPELINE_STOP);
2791 vd->vdev_ops->vdev_op_io_start(zio);
2792 return (ZIO_PIPELINE_STOP);
2796 zio_vdev_io_done(zio_t *zio)
2798 vdev_t *vd = zio->io_vd;
2799 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2800 boolean_t unexpected_error = B_FALSE;
2802 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2803 return (ZIO_PIPELINE_STOP);
2805 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2806 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2808 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2809 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2810 zio->io_type == ZIO_TYPE_FREE)) {
2812 if (zio->io_type == ZIO_TYPE_WRITE &&
2813 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2814 trim_map_write_done(zio);
2816 vdev_queue_io_done(zio);
2818 if (zio->io_type == ZIO_TYPE_WRITE)
2819 vdev_cache_write(zio);
2821 if (zio_injection_enabled && zio->io_error == 0)
2822 zio->io_error = zio_handle_device_injection(vd,
2825 if (zio_injection_enabled && zio->io_error == 0)
2826 zio->io_error = zio_handle_label_injection(zio, EIO);
2828 if (zio->io_error) {
2829 if (zio->io_error == ENOTSUP &&
2830 zio->io_type == ZIO_TYPE_FREE) {
2831 /* Not all devices support TRIM. */
2832 } else if (!vdev_accessible(vd, zio)) {
2833 zio->io_error = SET_ERROR(ENXIO);
2835 unexpected_error = B_TRUE;
2840 ops->vdev_op_io_done(zio);
2842 if (unexpected_error)
2843 VERIFY(vdev_probe(vd, zio) == NULL);
2845 return (ZIO_PIPELINE_CONTINUE);
2849 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2850 * disk, and use that to finish the checksum ereport later.
2853 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2854 const void *good_buf)
2856 /* no processing needed */
2857 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2862 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2864 void *buf = zio_buf_alloc(zio->io_size);
2866 bcopy(zio->io_data, buf, zio->io_size);
2868 zcr->zcr_cbinfo = zio->io_size;
2869 zcr->zcr_cbdata = buf;
2870 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2871 zcr->zcr_free = zio_buf_free;
2875 zio_vdev_io_assess(zio_t *zio)
2877 vdev_t *vd = zio->io_vd;
2879 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2880 return (ZIO_PIPELINE_STOP);
2882 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2883 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2885 if (zio->io_vsd != NULL) {
2886 zio->io_vsd_ops->vsd_free(zio);
2890 if (zio_injection_enabled && zio->io_error == 0)
2891 zio->io_error = zio_handle_fault_injection(zio, EIO);
2893 if (zio->io_type == ZIO_TYPE_FREE &&
2894 zio->io_priority != ZIO_PRIORITY_NOW) {
2895 switch (zio->io_error) {
2897 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2898 ZIO_TRIM_STAT_BUMP(success);
2901 ZIO_TRIM_STAT_BUMP(unsupported);
2904 ZIO_TRIM_STAT_BUMP(failed);
2910 * If the I/O failed, determine whether we should attempt to retry it.
2912 * On retry, we cut in line in the issue queue, since we don't want
2913 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2915 if (zio->io_error && vd == NULL &&
2916 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2917 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2918 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2920 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2921 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2922 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2923 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2924 zio_requeue_io_start_cut_in_line);
2925 return (ZIO_PIPELINE_STOP);
2929 * If we got an error on a leaf device, convert it to ENXIO
2930 * if the device is not accessible at all.
2932 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2933 !vdev_accessible(vd, zio))
2934 zio->io_error = SET_ERROR(ENXIO);
2937 * If we can't write to an interior vdev (mirror or RAID-Z),
2938 * set vdev_cant_write so that we stop trying to allocate from it.
2940 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2941 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2942 vd->vdev_cant_write = B_TRUE;
2946 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2948 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2949 zio->io_physdone != NULL) {
2950 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2951 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2952 zio->io_physdone(zio->io_logical);
2955 return (ZIO_PIPELINE_CONTINUE);
2959 zio_vdev_io_reissue(zio_t *zio)
2961 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2962 ASSERT(zio->io_error == 0);
2964 zio->io_stage >>= 1;
2968 zio_vdev_io_redone(zio_t *zio)
2970 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2972 zio->io_stage >>= 1;
2976 zio_vdev_io_bypass(zio_t *zio)
2978 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2979 ASSERT(zio->io_error == 0);
2981 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2982 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2986 * ==========================================================================
2987 * Generate and verify checksums
2988 * ==========================================================================
2991 zio_checksum_generate(zio_t *zio)
2993 blkptr_t *bp = zio->io_bp;
2994 enum zio_checksum checksum;
2998 * This is zio_write_phys().
2999 * We're either generating a label checksum, or none at all.
3001 checksum = zio->io_prop.zp_checksum;
3003 if (checksum == ZIO_CHECKSUM_OFF)
3004 return (ZIO_PIPELINE_CONTINUE);
3006 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3008 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3009 ASSERT(!IO_IS_ALLOCATING(zio));
3010 checksum = ZIO_CHECKSUM_GANG_HEADER;
3012 checksum = BP_GET_CHECKSUM(bp);
3016 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3018 return (ZIO_PIPELINE_CONTINUE);
3022 zio_checksum_verify(zio_t *zio)
3024 zio_bad_cksum_t info;
3025 blkptr_t *bp = zio->io_bp;
3028 ASSERT(zio->io_vd != NULL);
3032 * This is zio_read_phys().
3033 * We're either verifying a label checksum, or nothing at all.
3035 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3036 return (ZIO_PIPELINE_CONTINUE);
3038 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3041 if ((error = zio_checksum_error(zio, &info)) != 0) {
3042 zio->io_error = error;
3043 if (error == ECKSUM &&
3044 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3045 zfs_ereport_start_checksum(zio->io_spa,
3046 zio->io_vd, zio, zio->io_offset,
3047 zio->io_size, NULL, &info);
3051 return (ZIO_PIPELINE_CONTINUE);
3055 * Called by RAID-Z to ensure we don't compute the checksum twice.
3058 zio_checksum_verified(zio_t *zio)
3060 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3064 * ==========================================================================
3065 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3066 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3067 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3068 * indicate errors that are specific to one I/O, and most likely permanent.
3069 * Any other error is presumed to be worse because we weren't expecting it.
3070 * ==========================================================================
3073 zio_worst_error(int e1, int e2)
3075 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3078 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3079 if (e1 == zio_error_rank[r1])
3082 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3083 if (e2 == zio_error_rank[r2])
3086 return (r1 > r2 ? e1 : e2);
3090 * ==========================================================================
3092 * ==========================================================================
3095 zio_ready(zio_t *zio)
3097 blkptr_t *bp = zio->io_bp;
3098 zio_t *pio, *pio_next;
3100 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3101 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3102 return (ZIO_PIPELINE_STOP);
3104 if (zio->io_ready) {
3105 ASSERT(IO_IS_ALLOCATING(zio));
3106 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3107 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3108 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3113 if (bp != NULL && bp != &zio->io_bp_copy)
3114 zio->io_bp_copy = *bp;
3117 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3119 mutex_enter(&zio->io_lock);
3120 zio->io_state[ZIO_WAIT_READY] = 1;
3121 pio = zio_walk_parents(zio);
3122 mutex_exit(&zio->io_lock);
3125 * As we notify zio's parents, new parents could be added.
3126 * New parents go to the head of zio's io_parent_list, however,
3127 * so we will (correctly) not notify them. The remainder of zio's
3128 * io_parent_list, from 'pio_next' onward, cannot change because
3129 * all parents must wait for us to be done before they can be done.
3131 for (; pio != NULL; pio = pio_next) {
3132 pio_next = zio_walk_parents(zio);
3133 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3136 if (zio->io_flags & ZIO_FLAG_NODATA) {
3137 if (BP_IS_GANG(bp)) {
3138 zio->io_flags &= ~ZIO_FLAG_NODATA;
3140 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3141 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3145 if (zio_injection_enabled &&
3146 zio->io_spa->spa_syncing_txg == zio->io_txg)
3147 zio_handle_ignored_writes(zio);
3149 return (ZIO_PIPELINE_CONTINUE);
3153 zio_done(zio_t *zio)
3155 spa_t *spa = zio->io_spa;
3156 zio_t *lio = zio->io_logical;
3157 blkptr_t *bp = zio->io_bp;
3158 vdev_t *vd = zio->io_vd;
3159 uint64_t psize = zio->io_size;
3160 zio_t *pio, *pio_next;
3163 * If our children haven't all completed,
3164 * wait for them and then repeat this pipeline stage.
3166 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3167 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3168 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3169 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3170 return (ZIO_PIPELINE_STOP);
3172 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3173 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3174 ASSERT(zio->io_children[c][w] == 0);
3176 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3177 ASSERT(bp->blk_pad[0] == 0);
3178 ASSERT(bp->blk_pad[1] == 0);
3179 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3180 (bp == zio_unique_parent(zio)->io_bp));
3181 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3182 zio->io_bp_override == NULL &&
3183 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3184 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3185 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3186 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3187 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3189 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3190 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3194 * If there were child vdev/gang/ddt errors, they apply to us now.
3196 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3197 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3198 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3201 * If the I/O on the transformed data was successful, generate any
3202 * checksum reports now while we still have the transformed data.
3204 if (zio->io_error == 0) {
3205 while (zio->io_cksum_report != NULL) {
3206 zio_cksum_report_t *zcr = zio->io_cksum_report;
3207 uint64_t align = zcr->zcr_align;
3208 uint64_t asize = P2ROUNDUP(psize, align);
3209 char *abuf = zio->io_data;
3211 if (asize != psize) {
3212 abuf = zio_buf_alloc(asize);
3213 bcopy(zio->io_data, abuf, psize);
3214 bzero(abuf + psize, asize - psize);
3217 zio->io_cksum_report = zcr->zcr_next;
3218 zcr->zcr_next = NULL;
3219 zcr->zcr_finish(zcr, abuf);
3220 zfs_ereport_free_checksum(zcr);
3223 zio_buf_free(abuf, asize);
3227 zio_pop_transforms(zio); /* note: may set zio->io_error */
3229 vdev_stat_update(zio, psize);
3231 if (zio->io_error) {
3233 * If this I/O is attached to a particular vdev,
3234 * generate an error message describing the I/O failure
3235 * at the block level. We ignore these errors if the
3236 * device is currently unavailable.
3238 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3239 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3241 if ((zio->io_error == EIO || !(zio->io_flags &
3242 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3245 * For logical I/O requests, tell the SPA to log the
3246 * error and generate a logical data ereport.
3248 spa_log_error(spa, zio);
3249 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3254 if (zio->io_error && zio == lio) {
3256 * Determine whether zio should be reexecuted. This will
3257 * propagate all the way to the root via zio_notify_parent().
3259 ASSERT(vd == NULL && bp != NULL);
3260 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3262 if (IO_IS_ALLOCATING(zio) &&
3263 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3264 if (zio->io_error != ENOSPC)
3265 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3267 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3270 if ((zio->io_type == ZIO_TYPE_READ ||
3271 zio->io_type == ZIO_TYPE_FREE) &&
3272 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3273 zio->io_error == ENXIO &&
3274 spa_load_state(spa) == SPA_LOAD_NONE &&
3275 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3276 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3278 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3279 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3282 * Here is a possibly good place to attempt to do
3283 * either combinatorial reconstruction or error correction
3284 * based on checksums. It also might be a good place
3285 * to send out preliminary ereports before we suspend
3291 * If there were logical child errors, they apply to us now.
3292 * We defer this until now to avoid conflating logical child
3293 * errors with errors that happened to the zio itself when
3294 * updating vdev stats and reporting FMA events above.
3296 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3298 if ((zio->io_error || zio->io_reexecute) &&
3299 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3300 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3301 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3303 zio_gang_tree_free(&zio->io_gang_tree);
3306 * Godfather I/Os should never suspend.
3308 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3309 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3310 zio->io_reexecute = 0;
3312 if (zio->io_reexecute) {
3314 * This is a logical I/O that wants to reexecute.
3316 * Reexecute is top-down. When an i/o fails, if it's not
3317 * the root, it simply notifies its parent and sticks around.
3318 * The parent, seeing that it still has children in zio_done(),
3319 * does the same. This percolates all the way up to the root.
3320 * The root i/o will reexecute or suspend the entire tree.
3322 * This approach ensures that zio_reexecute() honors
3323 * all the original i/o dependency relationships, e.g.
3324 * parents not executing until children are ready.
3326 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3328 zio->io_gang_leader = NULL;
3330 mutex_enter(&zio->io_lock);
3331 zio->io_state[ZIO_WAIT_DONE] = 1;
3332 mutex_exit(&zio->io_lock);
3335 * "The Godfather" I/O monitors its children but is
3336 * not a true parent to them. It will track them through
3337 * the pipeline but severs its ties whenever they get into
3338 * trouble (e.g. suspended). This allows "The Godfather"
3339 * I/O to return status without blocking.
3341 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3342 zio_link_t *zl = zio->io_walk_link;
3343 pio_next = zio_walk_parents(zio);
3345 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3346 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3347 zio_remove_child(pio, zio, zl);
3348 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3352 if ((pio = zio_unique_parent(zio)) != NULL) {
3354 * We're not a root i/o, so there's nothing to do
3355 * but notify our parent. Don't propagate errors
3356 * upward since we haven't permanently failed yet.
3358 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3359 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3360 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3361 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3363 * We'd fail again if we reexecuted now, so suspend
3364 * until conditions improve (e.g. device comes online).
3366 zio_suspend(spa, zio);
3369 * Reexecution is potentially a huge amount of work.
3370 * Hand it off to the otherwise-unused claim taskq.
3372 #if defined(illumos) || !defined(_KERNEL)
3373 ASSERT(zio->io_tqent.tqent_next == NULL);
3375 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3377 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3378 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3381 return (ZIO_PIPELINE_STOP);
3384 ASSERT(zio->io_child_count == 0);
3385 ASSERT(zio->io_reexecute == 0);
3386 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3389 * Report any checksum errors, since the I/O is complete.
3391 while (zio->io_cksum_report != NULL) {
3392 zio_cksum_report_t *zcr = zio->io_cksum_report;
3393 zio->io_cksum_report = zcr->zcr_next;
3394 zcr->zcr_next = NULL;
3395 zcr->zcr_finish(zcr, NULL);
3396 zfs_ereport_free_checksum(zcr);
3400 * It is the responsibility of the done callback to ensure that this
3401 * particular zio is no longer discoverable for adoption, and as
3402 * such, cannot acquire any new parents.
3407 mutex_enter(&zio->io_lock);
3408 zio->io_state[ZIO_WAIT_DONE] = 1;
3409 mutex_exit(&zio->io_lock);
3411 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3412 zio_link_t *zl = zio->io_walk_link;
3413 pio_next = zio_walk_parents(zio);
3414 zio_remove_child(pio, zio, zl);
3415 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3418 if (zio->io_waiter != NULL) {
3419 mutex_enter(&zio->io_lock);
3420 zio->io_executor = NULL;
3421 cv_broadcast(&zio->io_cv);
3422 mutex_exit(&zio->io_lock);
3427 return (ZIO_PIPELINE_STOP);
3431 * ==========================================================================
3432 * I/O pipeline definition
3433 * ==========================================================================
3435 static zio_pipe_stage_t *zio_pipeline[] = {
3441 zio_checksum_generate,
3456 zio_checksum_verify,
3460 /* dnp is the dnode for zb1->zb_object */
3462 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3463 const zbookmark_phys_t *zb2)
3465 uint64_t zb1nextL0, zb2thisobj;
3467 ASSERT(zb1->zb_objset == zb2->zb_objset);
3468 ASSERT(zb2->zb_level == 0);
3470 /* The objset_phys_t isn't before anything. */
3474 zb1nextL0 = (zb1->zb_blkid + 1) <<
3475 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3477 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3478 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3480 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3481 uint64_t nextobj = zb1nextL0 *
3482 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3483 return (nextobj <= zb2thisobj);
3486 if (zb1->zb_object < zb2thisobj)
3488 if (zb1->zb_object > zb2thisobj)
3490 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3492 return (zb1nextL0 <= zb2->zb_blkid);