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, 2015 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 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
52 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
53 "Use uma(9) for ZIO allocations");
54 static int zio_exclude_metadata = 0;
55 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
57 "Exclude metadata buffers from dumps as well");
59 zio_trim_stats_t zio_trim_stats = {
60 { "bytes", KSTAT_DATA_UINT64,
61 "Number of bytes successfully TRIMmed" },
62 { "success", KSTAT_DATA_UINT64,
63 "Number of successful TRIM requests" },
64 { "unsupported", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed because TRIM is not supported" },
66 { "failed", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed for reasons other than not supported" },
70 static kstat_t *zio_trim_ksp;
73 * ==========================================================================
74 * I/O type descriptions
75 * ==========================================================================
77 const char *zio_type_name[ZIO_TYPES] = {
78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
83 * ==========================================================================
85 * ==========================================================================
87 kmem_cache_t *zio_cache;
88 kmem_cache_t *zio_link_cache;
89 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
90 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
93 extern vmem_t *zio_alloc_arena;
97 * The following actions directly effect the spa's sync-to-convergence logic.
98 * The values below define the sync pass when we start performing the action.
99 * Care should be taken when changing these values as they directly impact
100 * spa_sync() performance. Tuning these values may introduce subtle performance
101 * pathologies and should only be done in the context of performance analysis.
102 * These tunables will eventually be removed and replaced with #defines once
103 * enough analysis has been done to determine optimal values.
105 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
106 * regular blocks are not deferred.
108 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
109 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
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 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
114 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
115 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
116 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
117 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
118 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
119 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
122 * An allocating zio is one that either currently has the DVA allocate
123 * stage set or will have it later in its lifetime.
125 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
127 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
130 int zio_buf_debug_limit = 16384;
132 int zio_buf_debug_limit = 0;
139 zio_cache = kmem_cache_create("zio_cache",
140 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
141 zio_link_cache = kmem_cache_create("zio_link_cache",
142 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
147 * For small buffers, we want a cache for each multiple of
148 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
149 * for each quarter-power of 2.
151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
155 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
163 * If we are using watchpoints, put each buffer on its own page,
164 * to eliminate the performance overhead of trapping to the
165 * kernel when modifying a non-watched buffer that shares the
166 * page with a watched buffer.
168 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
172 if (size <= 4 * SPA_MINBLOCKSIZE) {
173 align = SPA_MINBLOCKSIZE;
174 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
175 align = MIN(p2 >> 2, PAGESIZE);
180 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
181 zio_buf_cache[c] = kmem_cache_create(name, size,
182 align, NULL, NULL, NULL, NULL, NULL, cflags);
185 * Since zio_data bufs do not appear in crash dumps, we
186 * pass KMC_NOTOUCH so that no allocator metadata is
187 * stored with the buffers.
189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
190 zio_data_buf_cache[c] = kmem_cache_create(name, size,
191 align, NULL, NULL, NULL, NULL, NULL,
192 cflags | KMC_NOTOUCH | KMC_NODEBUG);
197 ASSERT(zio_buf_cache[c] != NULL);
198 if (zio_buf_cache[c - 1] == NULL)
199 zio_buf_cache[c - 1] = zio_buf_cache[c];
201 ASSERT(zio_data_buf_cache[c] != NULL);
202 if (zio_data_buf_cache[c - 1] == NULL)
203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
209 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
211 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
214 if (zio_trim_ksp != NULL) {
215 zio_trim_ksp->ks_data = &zio_trim_stats;
216 kstat_install(zio_trim_ksp);
224 kmem_cache_t *last_cache = NULL;
225 kmem_cache_t *last_data_cache = NULL;
227 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
228 if (zio_buf_cache[c] != last_cache) {
229 last_cache = zio_buf_cache[c];
230 kmem_cache_destroy(zio_buf_cache[c]);
232 zio_buf_cache[c] = NULL;
234 if (zio_data_buf_cache[c] != last_data_cache) {
235 last_data_cache = zio_data_buf_cache[c];
236 kmem_cache_destroy(zio_data_buf_cache[c]);
238 zio_data_buf_cache[c] = NULL;
241 kmem_cache_destroy(zio_link_cache);
242 kmem_cache_destroy(zio_cache);
246 if (zio_trim_ksp != NULL) {
247 kstat_delete(zio_trim_ksp);
253 * ==========================================================================
254 * Allocate and free I/O buffers
255 * ==========================================================================
259 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
260 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
261 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
262 * excess / transient data in-core during a crashdump.
265 zio_buf_alloc(size_t size)
267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
270 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
273 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
275 return (kmem_alloc(size, KM_SLEEP|flags));
279 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
280 * crashdump if the kernel panics. This exists so that we will limit the amount
281 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
282 * of kernel heap dumped to disk when the kernel panics)
285 zio_data_buf_alloc(size_t size)
287 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
289 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
292 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
294 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
298 zio_buf_free(void *buf, size_t size)
300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
302 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
305 kmem_cache_free(zio_buf_cache[c], buf);
307 kmem_free(buf, size);
311 zio_data_buf_free(void *buf, size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 kmem_cache_free(zio_data_buf_cache[c], buf);
320 kmem_free(buf, size);
324 * ==========================================================================
325 * Push and pop I/O transform buffers
326 * ==========================================================================
329 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
330 zio_transform_func_t *transform)
332 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
334 zt->zt_orig_data = zio->io_data;
335 zt->zt_orig_size = zio->io_size;
336 zt->zt_bufsize = bufsize;
337 zt->zt_transform = transform;
339 zt->zt_next = zio->io_transform_stack;
340 zio->io_transform_stack = zt;
347 zio_pop_transforms(zio_t *zio)
351 while ((zt = zio->io_transform_stack) != NULL) {
352 if (zt->zt_transform != NULL)
353 zt->zt_transform(zio,
354 zt->zt_orig_data, zt->zt_orig_size);
356 if (zt->zt_bufsize != 0)
357 zio_buf_free(zio->io_data, zt->zt_bufsize);
359 zio->io_data = zt->zt_orig_data;
360 zio->io_size = zt->zt_orig_size;
361 zio->io_transform_stack = zt->zt_next;
363 kmem_free(zt, sizeof (zio_transform_t));
368 * ==========================================================================
369 * I/O transform callbacks for subblocks and decompression
370 * ==========================================================================
373 zio_subblock(zio_t *zio, void *data, uint64_t size)
375 ASSERT(zio->io_size > size);
377 if (zio->io_type == ZIO_TYPE_READ)
378 bcopy(zio->io_data, data, size);
382 zio_decompress(zio_t *zio, void *data, uint64_t size)
384 if (zio->io_error == 0 &&
385 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
386 zio->io_data, data, zio->io_size, size) != 0)
387 zio->io_error = SET_ERROR(EIO);
391 * ==========================================================================
392 * I/O parent/child relationships and pipeline interlocks
393 * ==========================================================================
396 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
397 * continue calling these functions until they return NULL.
398 * Otherwise, the next caller will pick up the list walk in
399 * some indeterminate state. (Otherwise every caller would
400 * have to pass in a cookie to keep the state represented by
401 * io_walk_link, which gets annoying.)
404 zio_walk_parents(zio_t *cio)
406 zio_link_t *zl = cio->io_walk_link;
407 list_t *pl = &cio->io_parent_list;
409 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
410 cio->io_walk_link = zl;
415 ASSERT(zl->zl_child == cio);
416 return (zl->zl_parent);
420 zio_walk_children(zio_t *pio)
422 zio_link_t *zl = pio->io_walk_link;
423 list_t *cl = &pio->io_child_list;
425 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
426 pio->io_walk_link = zl;
431 ASSERT(zl->zl_parent == pio);
432 return (zl->zl_child);
436 zio_unique_parent(zio_t *cio)
438 zio_t *pio = zio_walk_parents(cio);
440 VERIFY(zio_walk_parents(cio) == NULL);
445 zio_add_child(zio_t *pio, zio_t *cio)
447 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
450 * Logical I/Os can have logical, gang, or vdev children.
451 * Gang I/Os can have gang or vdev children.
452 * Vdev I/Os can only have vdev children.
453 * The following ASSERT captures all of these constraints.
455 ASSERT(cio->io_child_type <= pio->io_child_type);
460 mutex_enter(&cio->io_lock);
461 mutex_enter(&pio->io_lock);
463 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
465 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
466 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
468 list_insert_head(&pio->io_child_list, zl);
469 list_insert_head(&cio->io_parent_list, zl);
471 pio->io_child_count++;
472 cio->io_parent_count++;
474 mutex_exit(&pio->io_lock);
475 mutex_exit(&cio->io_lock);
479 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
481 ASSERT(zl->zl_parent == pio);
482 ASSERT(zl->zl_child == cio);
484 mutex_enter(&cio->io_lock);
485 mutex_enter(&pio->io_lock);
487 list_remove(&pio->io_child_list, zl);
488 list_remove(&cio->io_parent_list, zl);
490 pio->io_child_count--;
491 cio->io_parent_count--;
493 mutex_exit(&pio->io_lock);
494 mutex_exit(&cio->io_lock);
496 kmem_cache_free(zio_link_cache, zl);
500 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
502 uint64_t *countp = &zio->io_children[child][wait];
503 boolean_t waiting = B_FALSE;
505 mutex_enter(&zio->io_lock);
506 ASSERT(zio->io_stall == NULL);
509 zio->io_stall = countp;
512 mutex_exit(&zio->io_lock);
518 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
520 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
521 int *errorp = &pio->io_child_error[zio->io_child_type];
523 mutex_enter(&pio->io_lock);
524 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
525 *errorp = zio_worst_error(*errorp, zio->io_error);
526 pio->io_reexecute |= zio->io_reexecute;
527 ASSERT3U(*countp, >, 0);
531 if (*countp == 0 && pio->io_stall == countp) {
532 pio->io_stall = NULL;
533 mutex_exit(&pio->io_lock);
536 mutex_exit(&pio->io_lock);
541 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
543 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
544 zio->io_error = zio->io_child_error[c];
548 * ==========================================================================
549 * Create the various types of I/O (read, write, free, etc)
550 * ==========================================================================
553 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
554 void *data, uint64_t size, zio_done_func_t *done, void *private,
555 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
556 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
557 enum zio_stage stage, enum zio_stage pipeline)
561 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
562 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
563 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
565 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
566 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
567 ASSERT(vd || stage == ZIO_STAGE_OPEN);
569 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
570 bzero(zio, sizeof (zio_t));
572 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
573 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
575 list_create(&zio->io_parent_list, sizeof (zio_link_t),
576 offsetof(zio_link_t, zl_parent_node));
577 list_create(&zio->io_child_list, sizeof (zio_link_t),
578 offsetof(zio_link_t, zl_child_node));
581 zio->io_child_type = ZIO_CHILD_VDEV;
582 else if (flags & ZIO_FLAG_GANG_CHILD)
583 zio->io_child_type = ZIO_CHILD_GANG;
584 else if (flags & ZIO_FLAG_DDT_CHILD)
585 zio->io_child_type = ZIO_CHILD_DDT;
587 zio->io_child_type = ZIO_CHILD_LOGICAL;
590 zio->io_bp = (blkptr_t *)bp;
591 zio->io_bp_copy = *bp;
592 zio->io_bp_orig = *bp;
593 if (type != ZIO_TYPE_WRITE ||
594 zio->io_child_type == ZIO_CHILD_DDT)
595 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
596 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
597 zio->io_logical = zio;
598 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
599 pipeline |= ZIO_GANG_STAGES;
605 zio->io_private = private;
607 zio->io_priority = priority;
609 zio->io_offset = offset;
610 zio->io_orig_data = zio->io_data = data;
611 zio->io_orig_size = zio->io_size = size;
612 zio->io_orig_flags = zio->io_flags = flags;
613 zio->io_orig_stage = zio->io_stage = stage;
614 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
616 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
617 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
620 zio->io_bookmark = *zb;
623 if (zio->io_logical == NULL)
624 zio->io_logical = pio->io_logical;
625 if (zio->io_child_type == ZIO_CHILD_GANG)
626 zio->io_gang_leader = pio->io_gang_leader;
627 zio_add_child(pio, zio);
634 zio_destroy(zio_t *zio)
636 list_destroy(&zio->io_parent_list);
637 list_destroy(&zio->io_child_list);
638 mutex_destroy(&zio->io_lock);
639 cv_destroy(&zio->io_cv);
640 kmem_cache_free(zio_cache, zio);
644 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
645 void *private, enum zio_flag flags)
649 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
650 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
651 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
657 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
659 return (zio_null(NULL, spa, NULL, done, private, flags));
663 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
665 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
666 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
667 bp, (longlong_t)BP_GET_TYPE(bp));
669 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
670 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
671 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
672 bp, (longlong_t)BP_GET_CHECKSUM(bp));
674 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
675 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
676 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
677 bp, (longlong_t)BP_GET_COMPRESS(bp));
679 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
680 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
681 bp, (longlong_t)BP_GET_LSIZE(bp));
683 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
684 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
685 bp, (longlong_t)BP_GET_PSIZE(bp));
688 if (BP_IS_EMBEDDED(bp)) {
689 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
690 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
691 bp, (longlong_t)BPE_GET_ETYPE(bp));
696 * Pool-specific checks.
698 * Note: it would be nice to verify that the blk_birth and
699 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
700 * allows the birth time of log blocks (and dmu_sync()-ed blocks
701 * that are in the log) to be arbitrarily large.
703 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
704 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
705 if (vdevid >= spa->spa_root_vdev->vdev_children) {
706 zfs_panic_recover("blkptr at %p DVA %u has invalid "
708 bp, i, (longlong_t)vdevid);
711 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
713 zfs_panic_recover("blkptr at %p DVA %u has invalid "
715 bp, i, (longlong_t)vdevid);
718 if (vd->vdev_ops == &vdev_hole_ops) {
719 zfs_panic_recover("blkptr at %p DVA %u has hole "
721 bp, i, (longlong_t)vdevid);
724 if (vd->vdev_ops == &vdev_missing_ops) {
726 * "missing" vdevs are valid during import, but we
727 * don't have their detailed info (e.g. asize), so
728 * we can't perform any more checks on them.
732 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
733 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
735 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
736 if (offset + asize > vd->vdev_asize) {
737 zfs_panic_recover("blkptr at %p DVA %u has invalid "
739 bp, i, (longlong_t)offset);
745 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
746 void *data, uint64_t size, zio_done_func_t *done, void *private,
747 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
751 zfs_blkptr_verify(spa, bp);
753 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
754 data, size, done, private,
755 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
756 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
757 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
763 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
764 void *data, uint64_t size, const zio_prop_t *zp,
765 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
767 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
771 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
772 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
773 zp->zp_compress >= ZIO_COMPRESS_OFF &&
774 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
775 DMU_OT_IS_VALID(zp->zp_type) &&
778 zp->zp_copies <= spa_max_replication(spa));
780 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
781 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
782 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
783 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
785 zio->io_ready = ready;
786 zio->io_physdone = physdone;
790 * Data can be NULL if we are going to call zio_write_override() to
791 * provide the already-allocated BP. But we may need the data to
792 * verify a dedup hit (if requested). In this case, don't try to
793 * dedup (just take the already-allocated BP verbatim).
795 if (data == NULL && zio->io_prop.zp_dedup_verify) {
796 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
803 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
804 uint64_t size, zio_done_func_t *done, void *private,
805 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
809 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
810 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
811 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
817 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
819 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
820 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
821 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
822 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
825 * We must reset the io_prop to match the values that existed
826 * when the bp was first written by dmu_sync() keeping in mind
827 * that nopwrite and dedup are mutually exclusive.
829 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
830 zio->io_prop.zp_nopwrite = nopwrite;
831 zio->io_prop.zp_copies = copies;
832 zio->io_bp_override = bp;
836 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
840 * The check for EMBEDDED is a performance optimization. We
841 * process the free here (by ignoring it) rather than
842 * putting it on the list and then processing it in zio_free_sync().
844 if (BP_IS_EMBEDDED(bp))
846 metaslab_check_free(spa, bp);
849 * Frees that are for the currently-syncing txg, are not going to be
850 * deferred, and which will not need to do a read (i.e. not GANG or
851 * DEDUP), can be processed immediately. Otherwise, put them on the
852 * in-memory list for later processing.
854 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
855 txg != spa->spa_syncing_txg ||
856 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
857 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
859 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
860 BP_GET_PSIZE(bp), 0)));
865 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
866 uint64_t size, enum zio_flag flags)
869 enum zio_stage stage = ZIO_FREE_PIPELINE;
871 ASSERT(!BP_IS_HOLE(bp));
872 ASSERT(spa_syncing_txg(spa) == txg);
873 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
875 if (BP_IS_EMBEDDED(bp))
876 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
878 metaslab_check_free(spa, bp);
881 if (zfs_trim_enabled)
882 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
883 ZIO_STAGE_VDEV_IO_ASSESS;
885 * GANG and DEDUP blocks can induce a read (for the gang block header,
886 * or the DDT), so issue them asynchronously so that this thread is
889 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
890 stage |= ZIO_STAGE_ISSUE_ASYNC;
892 flags |= ZIO_FLAG_DONT_QUEUE;
894 zio = zio_create(pio, spa, txg, bp, NULL, size,
895 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
896 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
902 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
903 zio_done_func_t *done, void *private, enum zio_flag flags)
907 dprintf_bp(bp, "claiming in txg %llu", txg);
909 if (BP_IS_EMBEDDED(bp))
910 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
913 * A claim is an allocation of a specific block. Claims are needed
914 * to support immediate writes in the intent log. The issue is that
915 * immediate writes contain committed data, but in a txg that was
916 * *not* committed. Upon opening the pool after an unclean shutdown,
917 * the intent log claims all blocks that contain immediate write data
918 * so that the SPA knows they're in use.
920 * All claims *must* be resolved in the first txg -- before the SPA
921 * starts allocating blocks -- so that nothing is allocated twice.
922 * If txg == 0 we just verify that the block is claimable.
924 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
925 ASSERT(txg == spa_first_txg(spa) || txg == 0);
926 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
928 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
929 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
930 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
936 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
937 uint64_t size, zio_done_func_t *done, void *private,
938 zio_priority_t priority, enum zio_flag flags)
943 if (vd->vdev_children == 0) {
944 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
945 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
946 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
950 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
952 for (c = 0; c < vd->vdev_children; c++)
953 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
954 offset, size, done, private, priority, flags));
961 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
962 void *data, int checksum, zio_done_func_t *done, void *private,
963 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
967 ASSERT(vd->vdev_children == 0);
968 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
969 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
970 ASSERT3U(offset + size, <=, vd->vdev_psize);
972 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
973 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
974 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
976 zio->io_prop.zp_checksum = checksum;
982 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
983 void *data, int checksum, zio_done_func_t *done, void *private,
984 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
988 ASSERT(vd->vdev_children == 0);
989 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
990 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
991 ASSERT3U(offset + size, <=, vd->vdev_psize);
993 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
994 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
995 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
997 zio->io_prop.zp_checksum = checksum;
999 if (zio_checksum_table[checksum].ci_eck) {
1001 * zec checksums are necessarily destructive -- they modify
1002 * the end of the write buffer to hold the verifier/checksum.
1003 * Therefore, we must make a local copy in case the data is
1004 * being written to multiple places in parallel.
1006 void *wbuf = zio_buf_alloc(size);
1007 bcopy(data, wbuf, size);
1008 zio_push_transform(zio, wbuf, size, size, NULL);
1015 * Create a child I/O to do some work for us.
1018 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1019 void *data, uint64_t size, int type, zio_priority_t priority,
1020 enum zio_flag flags, zio_done_func_t *done, void *private)
1022 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1025 ASSERT(vd->vdev_parent ==
1026 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1028 if (type == ZIO_TYPE_READ && bp != NULL) {
1030 * If we have the bp, then the child should perform the
1031 * checksum and the parent need not. This pushes error
1032 * detection as close to the leaves as possible and
1033 * eliminates redundant checksums in the interior nodes.
1035 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1036 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1039 /* Not all IO types require vdev io done stage e.g. free */
1040 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1041 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1043 if (vd->vdev_children == 0)
1044 offset += VDEV_LABEL_START_SIZE;
1046 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1049 * If we've decided to do a repair, the write is not speculative --
1050 * even if the original read was.
1052 if (flags & ZIO_FLAG_IO_REPAIR)
1053 flags &= ~ZIO_FLAG_SPECULATIVE;
1055 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1056 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1057 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1059 zio->io_physdone = pio->io_physdone;
1060 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1061 zio->io_logical->io_phys_children++;
1067 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1068 int type, zio_priority_t priority, enum zio_flag flags,
1069 zio_done_func_t *done, void *private)
1073 ASSERT(vd->vdev_ops->vdev_op_leaf);
1075 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1076 data, size, done, private, type, priority,
1077 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1079 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1085 zio_flush(zio_t *zio, vdev_t *vd)
1087 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1088 NULL, NULL, ZIO_PRIORITY_NOW,
1089 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1093 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1096 ASSERT(vd->vdev_ops->vdev_op_leaf);
1098 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1099 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1100 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1101 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1105 zio_shrink(zio_t *zio, uint64_t size)
1107 ASSERT(zio->io_executor == NULL);
1108 ASSERT(zio->io_orig_size == zio->io_size);
1109 ASSERT(size <= zio->io_size);
1112 * We don't shrink for raidz because of problems with the
1113 * reconstruction when reading back less than the block size.
1114 * Note, BP_IS_RAIDZ() assumes no compression.
1116 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1117 if (!BP_IS_RAIDZ(zio->io_bp))
1118 zio->io_orig_size = zio->io_size = size;
1122 * ==========================================================================
1123 * Prepare to read and write logical blocks
1124 * ==========================================================================
1128 zio_read_bp_init(zio_t *zio)
1130 blkptr_t *bp = zio->io_bp;
1132 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1133 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1134 !(zio->io_flags & ZIO_FLAG_RAW)) {
1136 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1137 void *cbuf = zio_buf_alloc(psize);
1139 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1142 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1143 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1144 decode_embedded_bp_compressed(bp, zio->io_data);
1146 ASSERT(!BP_IS_EMBEDDED(bp));
1149 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1150 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1152 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1153 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1155 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1156 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1158 return (ZIO_PIPELINE_CONTINUE);
1162 zio_write_bp_init(zio_t *zio)
1164 spa_t *spa = zio->io_spa;
1165 zio_prop_t *zp = &zio->io_prop;
1166 enum zio_compress compress = zp->zp_compress;
1167 blkptr_t *bp = zio->io_bp;
1168 uint64_t lsize = zio->io_size;
1169 uint64_t psize = lsize;
1173 * If our children haven't all reached the ready stage,
1174 * wait for them and then repeat this pipeline stage.
1176 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1177 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1178 return (ZIO_PIPELINE_STOP);
1180 if (!IO_IS_ALLOCATING(zio))
1181 return (ZIO_PIPELINE_CONTINUE);
1183 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1185 if (zio->io_bp_override) {
1186 ASSERT(bp->blk_birth != zio->io_txg);
1187 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1189 *bp = *zio->io_bp_override;
1190 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1192 if (BP_IS_EMBEDDED(bp))
1193 return (ZIO_PIPELINE_CONTINUE);
1196 * If we've been overridden and nopwrite is set then
1197 * set the flag accordingly to indicate that a nopwrite
1198 * has already occurred.
1200 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1201 ASSERT(!zp->zp_dedup);
1202 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1203 return (ZIO_PIPELINE_CONTINUE);
1206 ASSERT(!zp->zp_nopwrite);
1208 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1209 return (ZIO_PIPELINE_CONTINUE);
1211 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1212 zp->zp_dedup_verify);
1214 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1215 BP_SET_DEDUP(bp, 1);
1216 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1217 return (ZIO_PIPELINE_CONTINUE);
1219 zio->io_bp_override = NULL;
1223 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1225 * We're rewriting an existing block, which means we're
1226 * working on behalf of spa_sync(). For spa_sync() to
1227 * converge, it must eventually be the case that we don't
1228 * have to allocate new blocks. But compression changes
1229 * the blocksize, which forces a reallocate, and makes
1230 * convergence take longer. Therefore, after the first
1231 * few passes, stop compressing to ensure convergence.
1233 pass = spa_sync_pass(spa);
1235 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1236 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1237 ASSERT(!BP_GET_DEDUP(bp));
1239 if (pass >= zfs_sync_pass_dont_compress)
1240 compress = ZIO_COMPRESS_OFF;
1242 /* Make sure someone doesn't change their mind on overwrites */
1243 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1244 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1247 if (compress != ZIO_COMPRESS_OFF) {
1248 void *cbuf = zio_buf_alloc(lsize);
1249 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1250 if (psize == 0 || psize == lsize) {
1251 compress = ZIO_COMPRESS_OFF;
1252 zio_buf_free(cbuf, lsize);
1253 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1254 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1255 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1256 encode_embedded_bp_compressed(bp,
1257 cbuf, compress, lsize, psize);
1258 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1259 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1260 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1261 zio_buf_free(cbuf, lsize);
1262 bp->blk_birth = zio->io_txg;
1263 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1264 ASSERT(spa_feature_is_active(spa,
1265 SPA_FEATURE_EMBEDDED_DATA));
1266 return (ZIO_PIPELINE_CONTINUE);
1269 * Round up compressed size up to the ashift
1270 * of the smallest-ashift device, and zero the tail.
1271 * This ensures that the compressed size of the BP
1272 * (and thus compressratio property) are correct,
1273 * in that we charge for the padding used to fill out
1276 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1277 size_t rounded = (size_t)P2ROUNDUP(psize,
1278 1ULL << spa->spa_min_ashift);
1279 if (rounded >= lsize) {
1280 compress = ZIO_COMPRESS_OFF;
1281 zio_buf_free(cbuf, lsize);
1284 bzero((char *)cbuf + psize, rounded - psize);
1286 zio_push_transform(zio, cbuf,
1287 psize, lsize, NULL);
1293 * The final pass of spa_sync() must be all rewrites, but the first
1294 * few passes offer a trade-off: allocating blocks defers convergence,
1295 * but newly allocated blocks are sequential, so they can be written
1296 * to disk faster. Therefore, we allow the first few passes of
1297 * spa_sync() to allocate new blocks, but force rewrites after that.
1298 * There should only be a handful of blocks after pass 1 in any case.
1300 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1301 BP_GET_PSIZE(bp) == psize &&
1302 pass >= zfs_sync_pass_rewrite) {
1304 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1305 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1306 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1309 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1313 if (zio->io_bp_orig.blk_birth != 0 &&
1314 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
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_BIRTH(bp, zio->io_txg, 0);
1320 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1322 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1323 BP_SET_LSIZE(bp, lsize);
1324 BP_SET_TYPE(bp, zp->zp_type);
1325 BP_SET_LEVEL(bp, zp->zp_level);
1326 BP_SET_PSIZE(bp, psize);
1327 BP_SET_COMPRESS(bp, compress);
1328 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1329 BP_SET_DEDUP(bp, zp->zp_dedup);
1330 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1332 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1333 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1334 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1336 if (zp->zp_nopwrite) {
1337 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1338 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1339 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1343 return (ZIO_PIPELINE_CONTINUE);
1347 zio_free_bp_init(zio_t *zio)
1349 blkptr_t *bp = zio->io_bp;
1351 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1352 if (BP_GET_DEDUP(bp))
1353 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1356 return (ZIO_PIPELINE_CONTINUE);
1360 * ==========================================================================
1361 * Execute the I/O pipeline
1362 * ==========================================================================
1366 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1368 spa_t *spa = zio->io_spa;
1369 zio_type_t t = zio->io_type;
1370 int flags = (cutinline ? TQ_FRONT : 0);
1372 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1375 * If we're a config writer or a probe, the normal issue and
1376 * interrupt threads may all be blocked waiting for the config lock.
1377 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1379 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1383 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1385 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1389 * If this is a high priority I/O, then use the high priority taskq if
1392 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1393 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1396 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1399 * NB: We are assuming that the zio can only be dispatched
1400 * to a single taskq at a time. It would be a grievous error
1401 * to dispatch the zio to another taskq at the same time.
1403 #if defined(illumos) || !defined(_KERNEL)
1404 ASSERT(zio->io_tqent.tqent_next == NULL);
1406 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1408 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1409 flags, &zio->io_tqent);
1413 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1415 kthread_t *executor = zio->io_executor;
1416 spa_t *spa = zio->io_spa;
1418 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1419 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1421 for (i = 0; i < tqs->stqs_count; i++) {
1422 if (taskq_member(tqs->stqs_taskq[i], executor))
1431 zio_issue_async(zio_t *zio)
1433 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1435 return (ZIO_PIPELINE_STOP);
1439 zio_interrupt(zio_t *zio)
1441 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1445 * Execute the I/O pipeline until one of the following occurs:
1447 * (1) the I/O completes
1448 * (2) the pipeline stalls waiting for dependent child I/Os
1449 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1450 * (4) the I/O is delegated by vdev-level caching or aggregation
1451 * (5) the I/O is deferred due to vdev-level queueing
1452 * (6) the I/O is handed off to another thread.
1454 * In all cases, the pipeline stops whenever there's no CPU work; it never
1455 * burns a thread in cv_wait().
1457 * There's no locking on io_stage because there's no legitimate way
1458 * for multiple threads to be attempting to process the same I/O.
1460 static zio_pipe_stage_t *zio_pipeline[];
1463 zio_execute(zio_t *zio)
1465 zio->io_executor = curthread;
1467 while (zio->io_stage < ZIO_STAGE_DONE) {
1468 enum zio_stage pipeline = zio->io_pipeline;
1469 enum zio_stage stage = zio->io_stage;
1472 ASSERT(!MUTEX_HELD(&zio->io_lock));
1473 ASSERT(ISP2(stage));
1474 ASSERT(zio->io_stall == NULL);
1478 } while ((stage & pipeline) == 0);
1480 ASSERT(stage <= ZIO_STAGE_DONE);
1483 * If we are in interrupt context and this pipeline stage
1484 * will grab a config lock that is held across I/O,
1485 * or may wait for an I/O that needs an interrupt thread
1486 * to complete, issue async to avoid deadlock.
1488 * For VDEV_IO_START, we cut in line so that the io will
1489 * be sent to disk promptly.
1491 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1492 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1493 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1494 zio_requeue_io_start_cut_in_line : B_FALSE;
1495 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1499 zio->io_stage = stage;
1500 rv = zio_pipeline[highbit64(stage) - 1](zio);
1502 if (rv == ZIO_PIPELINE_STOP)
1505 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1510 * ==========================================================================
1511 * Initiate I/O, either sync or async
1512 * ==========================================================================
1515 zio_wait(zio_t *zio)
1519 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1520 ASSERT(zio->io_executor == NULL);
1522 zio->io_waiter = curthread;
1526 mutex_enter(&zio->io_lock);
1527 while (zio->io_executor != NULL)
1528 cv_wait(&zio->io_cv, &zio->io_lock);
1529 mutex_exit(&zio->io_lock);
1531 error = zio->io_error;
1538 zio_nowait(zio_t *zio)
1540 ASSERT(zio->io_executor == NULL);
1542 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1543 zio_unique_parent(zio) == NULL) {
1545 * This is a logical async I/O with no parent to wait for it.
1546 * We add it to the spa_async_root_zio "Godfather" I/O which
1547 * will ensure they complete prior to unloading the pool.
1549 spa_t *spa = zio->io_spa;
1551 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1558 * ==========================================================================
1559 * Reexecute or suspend/resume failed I/O
1560 * ==========================================================================
1564 zio_reexecute(zio_t *pio)
1566 zio_t *cio, *cio_next;
1568 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1569 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1570 ASSERT(pio->io_gang_leader == NULL);
1571 ASSERT(pio->io_gang_tree == NULL);
1573 pio->io_flags = pio->io_orig_flags;
1574 pio->io_stage = pio->io_orig_stage;
1575 pio->io_pipeline = pio->io_orig_pipeline;
1576 pio->io_reexecute = 0;
1577 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1579 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1580 pio->io_state[w] = 0;
1581 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1582 pio->io_child_error[c] = 0;
1584 if (IO_IS_ALLOCATING(pio))
1585 BP_ZERO(pio->io_bp);
1588 * As we reexecute pio's children, new children could be created.
1589 * New children go to the head of pio's io_child_list, however,
1590 * so we will (correctly) not reexecute them. The key is that
1591 * the remainder of pio's io_child_list, from 'cio_next' onward,
1592 * cannot be affected by any side effects of reexecuting 'cio'.
1594 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1595 cio_next = zio_walk_children(pio);
1596 mutex_enter(&pio->io_lock);
1597 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1598 pio->io_children[cio->io_child_type][w]++;
1599 mutex_exit(&pio->io_lock);
1604 * Now that all children have been reexecuted, execute the parent.
1605 * We don't reexecute "The Godfather" I/O here as it's the
1606 * responsibility of the caller to wait on him.
1608 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1613 zio_suspend(spa_t *spa, zio_t *zio)
1615 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1616 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1617 "failure and the failure mode property for this pool "
1618 "is set to panic.", spa_name(spa));
1620 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1622 mutex_enter(&spa->spa_suspend_lock);
1624 if (spa->spa_suspend_zio_root == NULL)
1625 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1626 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1627 ZIO_FLAG_GODFATHER);
1629 spa->spa_suspended = B_TRUE;
1632 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1633 ASSERT(zio != spa->spa_suspend_zio_root);
1634 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1635 ASSERT(zio_unique_parent(zio) == NULL);
1636 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1637 zio_add_child(spa->spa_suspend_zio_root, zio);
1640 mutex_exit(&spa->spa_suspend_lock);
1644 zio_resume(spa_t *spa)
1649 * Reexecute all previously suspended i/o.
1651 mutex_enter(&spa->spa_suspend_lock);
1652 spa->spa_suspended = B_FALSE;
1653 cv_broadcast(&spa->spa_suspend_cv);
1654 pio = spa->spa_suspend_zio_root;
1655 spa->spa_suspend_zio_root = NULL;
1656 mutex_exit(&spa->spa_suspend_lock);
1662 return (zio_wait(pio));
1666 zio_resume_wait(spa_t *spa)
1668 mutex_enter(&spa->spa_suspend_lock);
1669 while (spa_suspended(spa))
1670 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1671 mutex_exit(&spa->spa_suspend_lock);
1675 * ==========================================================================
1678 * A gang block is a collection of small blocks that looks to the DMU
1679 * like one large block. When zio_dva_allocate() cannot find a block
1680 * of the requested size, due to either severe fragmentation or the pool
1681 * being nearly full, it calls zio_write_gang_block() to construct the
1682 * block from smaller fragments.
1684 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1685 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1686 * an indirect block: it's an array of block pointers. It consumes
1687 * only one sector and hence is allocatable regardless of fragmentation.
1688 * The gang header's bps point to its gang members, which hold the data.
1690 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1691 * as the verifier to ensure uniqueness of the SHA256 checksum.
1692 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1693 * not the gang header. This ensures that data block signatures (needed for
1694 * deduplication) are independent of how the block is physically stored.
1696 * Gang blocks can be nested: a gang member may itself be a gang block.
1697 * Thus every gang block is a tree in which root and all interior nodes are
1698 * gang headers, and the leaves are normal blocks that contain user data.
1699 * The root of the gang tree is called the gang leader.
1701 * To perform any operation (read, rewrite, free, claim) on a gang block,
1702 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1703 * in the io_gang_tree field of the original logical i/o by recursively
1704 * reading the gang leader and all gang headers below it. This yields
1705 * an in-core tree containing the contents of every gang header and the
1706 * bps for every constituent of the gang block.
1708 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1709 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1710 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1711 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1712 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1713 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1714 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1715 * of the gang header plus zio_checksum_compute() of the data to update the
1716 * gang header's blk_cksum as described above.
1718 * The two-phase assemble/issue model solves the problem of partial failure --
1719 * what if you'd freed part of a gang block but then couldn't read the
1720 * gang header for another part? Assembling the entire gang tree first
1721 * ensures that all the necessary gang header I/O has succeeded before
1722 * starting the actual work of free, claim, or write. Once the gang tree
1723 * is assembled, free and claim are in-memory operations that cannot fail.
1725 * In the event that a gang write fails, zio_dva_unallocate() walks the
1726 * gang tree to immediately free (i.e. insert back into the space map)
1727 * everything we've allocated. This ensures that we don't get ENOSPC
1728 * errors during repeated suspend/resume cycles due to a flaky device.
1730 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1731 * the gang tree, we won't modify the block, so we can safely defer the free
1732 * (knowing that the block is still intact). If we *can* assemble the gang
1733 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1734 * each constituent bp and we can allocate a new block on the next sync pass.
1736 * In all cases, the gang tree allows complete recovery from partial failure.
1737 * ==========================================================================
1741 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1746 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1747 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1748 &pio->io_bookmark));
1752 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1757 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1758 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1759 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1761 * As we rewrite each gang header, the pipeline will compute
1762 * a new gang block header checksum for it; but no one will
1763 * compute a new data checksum, so we do that here. The one
1764 * exception is the gang leader: the pipeline already computed
1765 * its data checksum because that stage precedes gang assembly.
1766 * (Presently, nothing actually uses interior data checksums;
1767 * this is just good hygiene.)
1769 if (gn != pio->io_gang_leader->io_gang_tree) {
1770 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1771 data, BP_GET_PSIZE(bp));
1774 * If we are here to damage data for testing purposes,
1775 * leave the GBH alone so that we can detect the damage.
1777 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1778 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1780 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1781 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1782 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1790 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1792 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1793 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1794 ZIO_GANG_CHILD_FLAGS(pio)));
1799 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1801 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1802 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1805 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1814 static void zio_gang_tree_assemble_done(zio_t *zio);
1816 static zio_gang_node_t *
1817 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1819 zio_gang_node_t *gn;
1821 ASSERT(*gnpp == NULL);
1823 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1824 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1831 zio_gang_node_free(zio_gang_node_t **gnpp)
1833 zio_gang_node_t *gn = *gnpp;
1835 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1836 ASSERT(gn->gn_child[g] == NULL);
1838 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1839 kmem_free(gn, sizeof (*gn));
1844 zio_gang_tree_free(zio_gang_node_t **gnpp)
1846 zio_gang_node_t *gn = *gnpp;
1851 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1852 zio_gang_tree_free(&gn->gn_child[g]);
1854 zio_gang_node_free(gnpp);
1858 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1860 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1862 ASSERT(gio->io_gang_leader == gio);
1863 ASSERT(BP_IS_GANG(bp));
1865 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1866 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1867 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1871 zio_gang_tree_assemble_done(zio_t *zio)
1873 zio_t *gio = zio->io_gang_leader;
1874 zio_gang_node_t *gn = zio->io_private;
1875 blkptr_t *bp = zio->io_bp;
1877 ASSERT(gio == zio_unique_parent(zio));
1878 ASSERT(zio->io_child_count == 0);
1883 if (BP_SHOULD_BYTESWAP(bp))
1884 byteswap_uint64_array(zio->io_data, zio->io_size);
1886 ASSERT(zio->io_data == gn->gn_gbh);
1887 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1888 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1890 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1891 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1892 if (!BP_IS_GANG(gbp))
1894 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1899 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1901 zio_t *gio = pio->io_gang_leader;
1904 ASSERT(BP_IS_GANG(bp) == !!gn);
1905 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1906 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1909 * If you're a gang header, your data is in gn->gn_gbh.
1910 * If you're a gang member, your data is in 'data' and gn == NULL.
1912 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1915 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1917 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1918 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1919 if (BP_IS_HOLE(gbp))
1921 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1922 data = (char *)data + BP_GET_PSIZE(gbp);
1926 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1927 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1934 zio_gang_assemble(zio_t *zio)
1936 blkptr_t *bp = zio->io_bp;
1938 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1939 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1941 zio->io_gang_leader = zio;
1943 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1945 return (ZIO_PIPELINE_CONTINUE);
1949 zio_gang_issue(zio_t *zio)
1951 blkptr_t *bp = zio->io_bp;
1953 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1954 return (ZIO_PIPELINE_STOP);
1956 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1957 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1959 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1960 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1962 zio_gang_tree_free(&zio->io_gang_tree);
1964 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1966 return (ZIO_PIPELINE_CONTINUE);
1970 zio_write_gang_member_ready(zio_t *zio)
1972 zio_t *pio = zio_unique_parent(zio);
1973 zio_t *gio = zio->io_gang_leader;
1974 dva_t *cdva = zio->io_bp->blk_dva;
1975 dva_t *pdva = pio->io_bp->blk_dva;
1978 if (BP_IS_HOLE(zio->io_bp))
1981 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1983 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1984 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1985 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1986 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1987 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1989 mutex_enter(&pio->io_lock);
1990 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1991 ASSERT(DVA_GET_GANG(&pdva[d]));
1992 asize = DVA_GET_ASIZE(&pdva[d]);
1993 asize += DVA_GET_ASIZE(&cdva[d]);
1994 DVA_SET_ASIZE(&pdva[d], asize);
1996 mutex_exit(&pio->io_lock);
2000 zio_write_gang_block(zio_t *pio)
2002 spa_t *spa = pio->io_spa;
2003 blkptr_t *bp = pio->io_bp;
2004 zio_t *gio = pio->io_gang_leader;
2006 zio_gang_node_t *gn, **gnpp;
2007 zio_gbh_phys_t *gbh;
2008 uint64_t txg = pio->io_txg;
2009 uint64_t resid = pio->io_size;
2011 int copies = gio->io_prop.zp_copies;
2012 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2016 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2017 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2018 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2020 pio->io_error = error;
2021 return (ZIO_PIPELINE_CONTINUE);
2025 gnpp = &gio->io_gang_tree;
2027 gnpp = pio->io_private;
2028 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2031 gn = zio_gang_node_alloc(gnpp);
2033 bzero(gbh, SPA_GANGBLOCKSIZE);
2036 * Create the gang header.
2038 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2039 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2042 * Create and nowait the gang children.
2044 for (int g = 0; resid != 0; resid -= lsize, g++) {
2045 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2047 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2049 zp.zp_checksum = gio->io_prop.zp_checksum;
2050 zp.zp_compress = ZIO_COMPRESS_OFF;
2051 zp.zp_type = DMU_OT_NONE;
2053 zp.zp_copies = gio->io_prop.zp_copies;
2054 zp.zp_dedup = B_FALSE;
2055 zp.zp_dedup_verify = B_FALSE;
2056 zp.zp_nopwrite = B_FALSE;
2058 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2059 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2060 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2061 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2062 &pio->io_bookmark));
2066 * Set pio's pipeline to just wait for zio to finish.
2068 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2072 return (ZIO_PIPELINE_CONTINUE);
2076 * The zio_nop_write stage in the pipeline determines if allocating
2077 * a new bp is necessary. By leveraging a cryptographically secure checksum,
2078 * such as SHA256, we can compare the checksums of the new data and the old
2079 * to determine if allocating a new block is required. The nopwrite
2080 * feature can handle writes in either syncing or open context (i.e. zil
2081 * writes) and as a result is mutually exclusive with dedup.
2084 zio_nop_write(zio_t *zio)
2086 blkptr_t *bp = zio->io_bp;
2087 blkptr_t *bp_orig = &zio->io_bp_orig;
2088 zio_prop_t *zp = &zio->io_prop;
2090 ASSERT(BP_GET_LEVEL(bp) == 0);
2091 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2092 ASSERT(zp->zp_nopwrite);
2093 ASSERT(!zp->zp_dedup);
2094 ASSERT(zio->io_bp_override == NULL);
2095 ASSERT(IO_IS_ALLOCATING(zio));
2098 * Check to see if the original bp and the new bp have matching
2099 * characteristics (i.e. same checksum, compression algorithms, etc).
2100 * If they don't then just continue with the pipeline which will
2101 * allocate a new bp.
2103 if (BP_IS_HOLE(bp_orig) ||
2104 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2105 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2106 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2107 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2108 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2109 return (ZIO_PIPELINE_CONTINUE);
2112 * If the checksums match then reset the pipeline so that we
2113 * avoid allocating a new bp and issuing any I/O.
2115 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2116 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2117 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2118 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2119 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2120 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2121 sizeof (uint64_t)) == 0);
2124 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2125 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2128 return (ZIO_PIPELINE_CONTINUE);
2132 * ==========================================================================
2134 * ==========================================================================
2137 zio_ddt_child_read_done(zio_t *zio)
2139 blkptr_t *bp = zio->io_bp;
2140 ddt_entry_t *dde = zio->io_private;
2142 zio_t *pio = zio_unique_parent(zio);
2144 mutex_enter(&pio->io_lock);
2145 ddp = ddt_phys_select(dde, bp);
2146 if (zio->io_error == 0)
2147 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2148 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2149 dde->dde_repair_data = zio->io_data;
2151 zio_buf_free(zio->io_data, zio->io_size);
2152 mutex_exit(&pio->io_lock);
2156 zio_ddt_read_start(zio_t *zio)
2158 blkptr_t *bp = zio->io_bp;
2160 ASSERT(BP_GET_DEDUP(bp));
2161 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2162 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2164 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2165 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2166 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2167 ddt_phys_t *ddp = dde->dde_phys;
2168 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2171 ASSERT(zio->io_vsd == NULL);
2174 if (ddp_self == NULL)
2175 return (ZIO_PIPELINE_CONTINUE);
2177 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2178 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2180 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2182 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2183 zio_buf_alloc(zio->io_size), zio->io_size,
2184 zio_ddt_child_read_done, dde, zio->io_priority,
2185 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2186 &zio->io_bookmark));
2188 return (ZIO_PIPELINE_CONTINUE);
2191 zio_nowait(zio_read(zio, zio->io_spa, bp,
2192 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2193 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2195 return (ZIO_PIPELINE_CONTINUE);
2199 zio_ddt_read_done(zio_t *zio)
2201 blkptr_t *bp = zio->io_bp;
2203 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2204 return (ZIO_PIPELINE_STOP);
2206 ASSERT(BP_GET_DEDUP(bp));
2207 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2208 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2210 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2211 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2212 ddt_entry_t *dde = zio->io_vsd;
2214 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2215 return (ZIO_PIPELINE_CONTINUE);
2218 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2219 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2220 return (ZIO_PIPELINE_STOP);
2222 if (dde->dde_repair_data != NULL) {
2223 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2224 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2226 ddt_repair_done(ddt, dde);
2230 ASSERT(zio->io_vsd == NULL);
2232 return (ZIO_PIPELINE_CONTINUE);
2236 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2238 spa_t *spa = zio->io_spa;
2241 * Note: we compare the original data, not the transformed data,
2242 * because when zio->io_bp is an override bp, we will not have
2243 * pushed the I/O transforms. That's an important optimization
2244 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2246 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2247 zio_t *lio = dde->dde_lead_zio[p];
2250 return (lio->io_orig_size != zio->io_orig_size ||
2251 bcmp(zio->io_orig_data, lio->io_orig_data,
2252 zio->io_orig_size) != 0);
2256 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2257 ddt_phys_t *ddp = &dde->dde_phys[p];
2259 if (ddp->ddp_phys_birth != 0) {
2260 arc_buf_t *abuf = NULL;
2261 arc_flags_t aflags = ARC_FLAG_WAIT;
2262 blkptr_t blk = *zio->io_bp;
2265 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2269 error = arc_read(NULL, spa, &blk,
2270 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2271 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2272 &aflags, &zio->io_bookmark);
2275 if (arc_buf_size(abuf) != zio->io_orig_size ||
2276 bcmp(abuf->b_data, zio->io_orig_data,
2277 zio->io_orig_size) != 0)
2278 error = SET_ERROR(EEXIST);
2279 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2283 return (error != 0);
2291 zio_ddt_child_write_ready(zio_t *zio)
2293 int p = zio->io_prop.zp_copies;
2294 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2295 ddt_entry_t *dde = zio->io_private;
2296 ddt_phys_t *ddp = &dde->dde_phys[p];
2304 ASSERT(dde->dde_lead_zio[p] == zio);
2306 ddt_phys_fill(ddp, zio->io_bp);
2308 while ((pio = zio_walk_parents(zio)) != NULL)
2309 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2315 zio_ddt_child_write_done(zio_t *zio)
2317 int p = zio->io_prop.zp_copies;
2318 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2319 ddt_entry_t *dde = zio->io_private;
2320 ddt_phys_t *ddp = &dde->dde_phys[p];
2324 ASSERT(ddp->ddp_refcnt == 0);
2325 ASSERT(dde->dde_lead_zio[p] == zio);
2326 dde->dde_lead_zio[p] = NULL;
2328 if (zio->io_error == 0) {
2329 while (zio_walk_parents(zio) != NULL)
2330 ddt_phys_addref(ddp);
2332 ddt_phys_clear(ddp);
2339 zio_ddt_ditto_write_done(zio_t *zio)
2341 int p = DDT_PHYS_DITTO;
2342 zio_prop_t *zp = &zio->io_prop;
2343 blkptr_t *bp = zio->io_bp;
2344 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2345 ddt_entry_t *dde = zio->io_private;
2346 ddt_phys_t *ddp = &dde->dde_phys[p];
2347 ddt_key_t *ddk = &dde->dde_key;
2351 ASSERT(ddp->ddp_refcnt == 0);
2352 ASSERT(dde->dde_lead_zio[p] == zio);
2353 dde->dde_lead_zio[p] = NULL;
2355 if (zio->io_error == 0) {
2356 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2357 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2358 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2359 if (ddp->ddp_phys_birth != 0)
2360 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2361 ddt_phys_fill(ddp, bp);
2368 zio_ddt_write(zio_t *zio)
2370 spa_t *spa = zio->io_spa;
2371 blkptr_t *bp = zio->io_bp;
2372 uint64_t txg = zio->io_txg;
2373 zio_prop_t *zp = &zio->io_prop;
2374 int p = zp->zp_copies;
2378 ddt_t *ddt = ddt_select(spa, bp);
2382 ASSERT(BP_GET_DEDUP(bp));
2383 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2384 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2387 dde = ddt_lookup(ddt, bp, B_TRUE);
2388 ddp = &dde->dde_phys[p];
2390 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2392 * If we're using a weak checksum, upgrade to a strong checksum
2393 * and try again. If we're already using a strong checksum,
2394 * we can't resolve it, so just convert to an ordinary write.
2395 * (And automatically e-mail a paper to Nature?)
2397 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2398 zp->zp_checksum = spa_dedup_checksum(spa);
2399 zio_pop_transforms(zio);
2400 zio->io_stage = ZIO_STAGE_OPEN;
2403 zp->zp_dedup = B_FALSE;
2405 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2407 return (ZIO_PIPELINE_CONTINUE);
2410 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2411 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2413 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2414 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2415 zio_prop_t czp = *zp;
2417 czp.zp_copies = ditto_copies;
2420 * If we arrived here with an override bp, we won't have run
2421 * the transform stack, so we won't have the data we need to
2422 * generate a child i/o. So, toss the override bp and restart.
2423 * This is safe, because using the override bp is just an
2424 * optimization; and it's rare, so the cost doesn't matter.
2426 if (zio->io_bp_override) {
2427 zio_pop_transforms(zio);
2428 zio->io_stage = ZIO_STAGE_OPEN;
2429 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2430 zio->io_bp_override = NULL;
2433 return (ZIO_PIPELINE_CONTINUE);
2436 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2437 zio->io_orig_size, &czp, NULL, NULL,
2438 zio_ddt_ditto_write_done, dde, zio->io_priority,
2439 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2441 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2442 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2445 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2446 if (ddp->ddp_phys_birth != 0)
2447 ddt_bp_fill(ddp, bp, txg);
2448 if (dde->dde_lead_zio[p] != NULL)
2449 zio_add_child(zio, dde->dde_lead_zio[p]);
2451 ddt_phys_addref(ddp);
2452 } else if (zio->io_bp_override) {
2453 ASSERT(bp->blk_birth == txg);
2454 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2455 ddt_phys_fill(ddp, bp);
2456 ddt_phys_addref(ddp);
2458 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2459 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2460 zio_ddt_child_write_done, dde, zio->io_priority,
2461 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2463 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2464 dde->dde_lead_zio[p] = cio;
2474 return (ZIO_PIPELINE_CONTINUE);
2477 ddt_entry_t *freedde; /* for debugging */
2480 zio_ddt_free(zio_t *zio)
2482 spa_t *spa = zio->io_spa;
2483 blkptr_t *bp = zio->io_bp;
2484 ddt_t *ddt = ddt_select(spa, bp);
2488 ASSERT(BP_GET_DEDUP(bp));
2489 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2492 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2493 ddp = ddt_phys_select(dde, bp);
2494 ddt_phys_decref(ddp);
2497 return (ZIO_PIPELINE_CONTINUE);
2501 * ==========================================================================
2502 * Allocate and free blocks
2503 * ==========================================================================
2506 zio_dva_allocate(zio_t *zio)
2508 spa_t *spa = zio->io_spa;
2509 metaslab_class_t *mc = spa_normal_class(spa);
2510 blkptr_t *bp = zio->io_bp;
2514 if (zio->io_gang_leader == NULL) {
2515 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2516 zio->io_gang_leader = zio;
2519 ASSERT(BP_IS_HOLE(bp));
2520 ASSERT0(BP_GET_NDVAS(bp));
2521 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2522 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2523 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2526 * The dump device does not support gang blocks so allocation on
2527 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2528 * the "fast" gang feature.
2530 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2531 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2532 METASLAB_GANG_CHILD : 0;
2533 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2534 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2537 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2538 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2540 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2541 return (zio_write_gang_block(zio));
2542 zio->io_error = error;
2545 return (ZIO_PIPELINE_CONTINUE);
2549 zio_dva_free(zio_t *zio)
2551 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2553 return (ZIO_PIPELINE_CONTINUE);
2557 zio_dva_claim(zio_t *zio)
2561 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2563 zio->io_error = error;
2565 return (ZIO_PIPELINE_CONTINUE);
2569 * Undo an allocation. This is used by zio_done() when an I/O fails
2570 * and we want to give back the block we just allocated.
2571 * This handles both normal blocks and gang blocks.
2574 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2576 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2577 ASSERT(zio->io_bp_override == NULL);
2579 if (!BP_IS_HOLE(bp))
2580 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2583 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2584 zio_dva_unallocate(zio, gn->gn_child[g],
2585 &gn->gn_gbh->zg_blkptr[g]);
2591 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2594 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2595 uint64_t size, boolean_t use_slog)
2599 ASSERT(txg > spa_syncing_txg(spa));
2602 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2603 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2604 * when allocating them.
2607 error = metaslab_alloc(spa, spa_log_class(spa), size,
2608 new_bp, 1, txg, old_bp,
2609 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2613 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2614 new_bp, 1, txg, old_bp,
2615 METASLAB_HINTBP_AVOID);
2619 BP_SET_LSIZE(new_bp, size);
2620 BP_SET_PSIZE(new_bp, size);
2621 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2622 BP_SET_CHECKSUM(new_bp,
2623 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2624 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2625 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2626 BP_SET_LEVEL(new_bp, 0);
2627 BP_SET_DEDUP(new_bp, 0);
2628 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2635 * Free an intent log block.
2638 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2640 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2641 ASSERT(!BP_IS_GANG(bp));
2643 zio_free(spa, txg, bp);
2647 * ==========================================================================
2648 * Read, write and delete to physical devices
2649 * ==========================================================================
2652 zio_vdev_io_start(zio_t *zio)
2654 vdev_t *vd = zio->io_vd;
2656 spa_t *spa = zio->io_spa;
2659 ASSERT(zio->io_error == 0);
2660 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2663 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2664 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2667 * The mirror_ops handle multiple DVAs in a single BP.
2669 return (vdev_mirror_ops.vdev_op_io_start(zio));
2672 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2673 zio->io_priority == ZIO_PRIORITY_NOW) {
2674 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2675 return (ZIO_PIPELINE_CONTINUE);
2679 * We keep track of time-sensitive I/Os so that the scan thread
2680 * can quickly react to certain workloads. In particular, we care
2681 * about non-scrubbing, top-level reads and writes with the following
2683 * - synchronous writes of user data to non-slog devices
2684 * - any reads of user data
2685 * When these conditions are met, adjust the timestamp of spa_last_io
2686 * which allows the scan thread to adjust its workload accordingly.
2688 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2689 vd == vd->vdev_top && !vd->vdev_islog &&
2690 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2691 zio->io_txg != spa_syncing_txg(spa)) {
2692 uint64_t old = spa->spa_last_io;
2693 uint64_t new = ddi_get_lbolt64();
2695 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2698 align = 1ULL << vd->vdev_top->vdev_ashift;
2700 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2701 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2702 P2PHASE(zio->io_size, align) != 0) {
2703 /* Transform logical writes to be a full physical block size. */
2704 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2706 if (zio->io_type == ZIO_TYPE_READ ||
2707 zio->io_type == ZIO_TYPE_WRITE)
2708 abuf = zio_buf_alloc(asize);
2709 ASSERT(vd == vd->vdev_top);
2710 if (zio->io_type == ZIO_TYPE_WRITE) {
2711 bcopy(zio->io_data, abuf, zio->io_size);
2712 bzero(abuf + zio->io_size, asize - zio->io_size);
2714 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2719 * If this is not a physical io, make sure that it is properly aligned
2720 * before proceeding.
2722 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2723 ASSERT0(P2PHASE(zio->io_offset, align));
2724 ASSERT0(P2PHASE(zio->io_size, align));
2727 * For physical writes, we allow 512b aligned writes and assume
2728 * the device will perform a read-modify-write as necessary.
2730 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2731 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2734 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2737 * If this is a repair I/O, and there's no self-healing involved --
2738 * that is, we're just resilvering what we expect to resilver --
2739 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2740 * This prevents spurious resilvering with nested replication.
2741 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2742 * A is out of date, we'll read from C+D, then use the data to
2743 * resilver A+B -- but we don't actually want to resilver B, just A.
2744 * The top-level mirror has no way to know this, so instead we just
2745 * discard unnecessary repairs as we work our way down the vdev tree.
2746 * The same logic applies to any form of nested replication:
2747 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2749 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2750 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2751 zio->io_txg != 0 && /* not a delegated i/o */
2752 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2753 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2754 zio_vdev_io_bypass(zio);
2755 return (ZIO_PIPELINE_CONTINUE);
2758 if (vd->vdev_ops->vdev_op_leaf) {
2759 switch (zio->io_type) {
2761 if (vdev_cache_read(zio))
2762 return (ZIO_PIPELINE_CONTINUE);
2764 case ZIO_TYPE_WRITE:
2766 if ((zio = vdev_queue_io(zio)) == NULL)
2767 return (ZIO_PIPELINE_STOP);
2769 if (!vdev_accessible(vd, zio)) {
2770 zio->io_error = SET_ERROR(ENXIO);
2772 return (ZIO_PIPELINE_STOP);
2777 * Note that we ignore repair writes for TRIM because they can
2778 * conflict with normal writes. This isn't an issue because, by
2779 * definition, we only repair blocks that aren't freed.
2781 if (zio->io_type == ZIO_TYPE_WRITE &&
2782 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2783 !trim_map_write_start(zio))
2784 return (ZIO_PIPELINE_STOP);
2787 ret = vd->vdev_ops->vdev_op_io_start(zio);
2788 ASSERT(ret == ZIO_PIPELINE_STOP);
2794 zio_vdev_io_done(zio_t *zio)
2796 vdev_t *vd = zio->io_vd;
2797 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2798 boolean_t unexpected_error = B_FALSE;
2800 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2801 return (ZIO_PIPELINE_STOP);
2803 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2804 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2806 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2807 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2808 zio->io_type == ZIO_TYPE_FREE)) {
2810 if (zio->io_type == ZIO_TYPE_WRITE &&
2811 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2812 trim_map_write_done(zio);
2814 vdev_queue_io_done(zio);
2816 if (zio->io_type == ZIO_TYPE_WRITE)
2817 vdev_cache_write(zio);
2819 if (zio_injection_enabled && zio->io_error == 0)
2820 zio->io_error = zio_handle_device_injection(vd,
2823 if (zio_injection_enabled && zio->io_error == 0)
2824 zio->io_error = zio_handle_label_injection(zio, EIO);
2826 if (zio->io_error) {
2827 if (zio->io_error == ENOTSUP &&
2828 zio->io_type == ZIO_TYPE_FREE) {
2829 /* Not all devices support TRIM. */
2830 } else if (!vdev_accessible(vd, zio)) {
2831 zio->io_error = SET_ERROR(ENXIO);
2833 unexpected_error = B_TRUE;
2838 ops->vdev_op_io_done(zio);
2840 if (unexpected_error)
2841 VERIFY(vdev_probe(vd, zio) == NULL);
2843 return (ZIO_PIPELINE_CONTINUE);
2847 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2848 * disk, and use that to finish the checksum ereport later.
2851 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2852 const void *good_buf)
2854 /* no processing needed */
2855 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2860 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2862 void *buf = zio_buf_alloc(zio->io_size);
2864 bcopy(zio->io_data, buf, zio->io_size);
2866 zcr->zcr_cbinfo = zio->io_size;
2867 zcr->zcr_cbdata = buf;
2868 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2869 zcr->zcr_free = zio_buf_free;
2873 zio_vdev_io_assess(zio_t *zio)
2875 vdev_t *vd = zio->io_vd;
2877 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2878 return (ZIO_PIPELINE_STOP);
2880 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2881 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2883 if (zio->io_vsd != NULL) {
2884 zio->io_vsd_ops->vsd_free(zio);
2888 if (zio_injection_enabled && zio->io_error == 0)
2889 zio->io_error = zio_handle_fault_injection(zio, EIO);
2891 if (zio->io_type == ZIO_TYPE_FREE &&
2892 zio->io_priority != ZIO_PRIORITY_NOW) {
2893 switch (zio->io_error) {
2895 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2896 ZIO_TRIM_STAT_BUMP(success);
2899 ZIO_TRIM_STAT_BUMP(unsupported);
2902 ZIO_TRIM_STAT_BUMP(failed);
2908 * If the I/O failed, determine whether we should attempt to retry it.
2910 * On retry, we cut in line in the issue queue, since we don't want
2911 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2913 if (zio->io_error && vd == NULL &&
2914 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2915 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2916 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2918 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2919 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2920 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2921 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2922 zio_requeue_io_start_cut_in_line);
2923 return (ZIO_PIPELINE_STOP);
2927 * If we got an error on a leaf device, convert it to ENXIO
2928 * if the device is not accessible at all.
2930 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2931 !vdev_accessible(vd, zio))
2932 zio->io_error = SET_ERROR(ENXIO);
2935 * If we can't write to an interior vdev (mirror or RAID-Z),
2936 * set vdev_cant_write so that we stop trying to allocate from it.
2938 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2939 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2940 vd->vdev_cant_write = B_TRUE;
2944 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2946 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2947 zio->io_physdone != NULL) {
2948 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2949 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2950 zio->io_physdone(zio->io_logical);
2953 return (ZIO_PIPELINE_CONTINUE);
2957 zio_vdev_io_reissue(zio_t *zio)
2959 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2960 ASSERT(zio->io_error == 0);
2962 zio->io_stage >>= 1;
2966 zio_vdev_io_redone(zio_t *zio)
2968 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2970 zio->io_stage >>= 1;
2974 zio_vdev_io_bypass(zio_t *zio)
2976 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2977 ASSERT(zio->io_error == 0);
2979 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2980 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2984 * ==========================================================================
2985 * Generate and verify checksums
2986 * ==========================================================================
2989 zio_checksum_generate(zio_t *zio)
2991 blkptr_t *bp = zio->io_bp;
2992 enum zio_checksum checksum;
2996 * This is zio_write_phys().
2997 * We're either generating a label checksum, or none at all.
2999 checksum = zio->io_prop.zp_checksum;
3001 if (checksum == ZIO_CHECKSUM_OFF)
3002 return (ZIO_PIPELINE_CONTINUE);
3004 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3006 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3007 ASSERT(!IO_IS_ALLOCATING(zio));
3008 checksum = ZIO_CHECKSUM_GANG_HEADER;
3010 checksum = BP_GET_CHECKSUM(bp);
3014 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3016 return (ZIO_PIPELINE_CONTINUE);
3020 zio_checksum_verify(zio_t *zio)
3022 zio_bad_cksum_t info;
3023 blkptr_t *bp = zio->io_bp;
3026 ASSERT(zio->io_vd != NULL);
3030 * This is zio_read_phys().
3031 * We're either verifying a label checksum, or nothing at all.
3033 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3034 return (ZIO_PIPELINE_CONTINUE);
3036 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3039 if ((error = zio_checksum_error(zio, &info)) != 0) {
3040 zio->io_error = error;
3041 if (error == ECKSUM &&
3042 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3043 zfs_ereport_start_checksum(zio->io_spa,
3044 zio->io_vd, zio, zio->io_offset,
3045 zio->io_size, NULL, &info);
3049 return (ZIO_PIPELINE_CONTINUE);
3053 * Called by RAID-Z to ensure we don't compute the checksum twice.
3056 zio_checksum_verified(zio_t *zio)
3058 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3062 * ==========================================================================
3063 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3064 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3065 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3066 * indicate errors that are specific to one I/O, and most likely permanent.
3067 * Any other error is presumed to be worse because we weren't expecting it.
3068 * ==========================================================================
3071 zio_worst_error(int e1, int e2)
3073 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3076 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3077 if (e1 == zio_error_rank[r1])
3080 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3081 if (e2 == zio_error_rank[r2])
3084 return (r1 > r2 ? e1 : e2);
3088 * ==========================================================================
3090 * ==========================================================================
3093 zio_ready(zio_t *zio)
3095 blkptr_t *bp = zio->io_bp;
3096 zio_t *pio, *pio_next;
3098 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3099 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3100 return (ZIO_PIPELINE_STOP);
3102 if (zio->io_ready) {
3103 ASSERT(IO_IS_ALLOCATING(zio));
3104 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3105 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3106 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3111 if (bp != NULL && bp != &zio->io_bp_copy)
3112 zio->io_bp_copy = *bp;
3115 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3117 mutex_enter(&zio->io_lock);
3118 zio->io_state[ZIO_WAIT_READY] = 1;
3119 pio = zio_walk_parents(zio);
3120 mutex_exit(&zio->io_lock);
3123 * As we notify zio's parents, new parents could be added.
3124 * New parents go to the head of zio's io_parent_list, however,
3125 * so we will (correctly) not notify them. The remainder of zio's
3126 * io_parent_list, from 'pio_next' onward, cannot change because
3127 * all parents must wait for us to be done before they can be done.
3129 for (; pio != NULL; pio = pio_next) {
3130 pio_next = zio_walk_parents(zio);
3131 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3134 if (zio->io_flags & ZIO_FLAG_NODATA) {
3135 if (BP_IS_GANG(bp)) {
3136 zio->io_flags &= ~ZIO_FLAG_NODATA;
3138 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3139 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3143 if (zio_injection_enabled &&
3144 zio->io_spa->spa_syncing_txg == zio->io_txg)
3145 zio_handle_ignored_writes(zio);
3147 return (ZIO_PIPELINE_CONTINUE);
3151 zio_done(zio_t *zio)
3153 spa_t *spa = zio->io_spa;
3154 zio_t *lio = zio->io_logical;
3155 blkptr_t *bp = zio->io_bp;
3156 vdev_t *vd = zio->io_vd;
3157 uint64_t psize = zio->io_size;
3158 zio_t *pio, *pio_next;
3161 * If our children haven't all completed,
3162 * wait for them and then repeat this pipeline stage.
3164 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3165 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3166 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3167 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3168 return (ZIO_PIPELINE_STOP);
3170 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3171 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3172 ASSERT(zio->io_children[c][w] == 0);
3174 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3175 ASSERT(bp->blk_pad[0] == 0);
3176 ASSERT(bp->blk_pad[1] == 0);
3177 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3178 (bp == zio_unique_parent(zio)->io_bp));
3179 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3180 zio->io_bp_override == NULL &&
3181 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3182 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3183 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3184 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3185 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3187 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3188 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3192 * If there were child vdev/gang/ddt errors, they apply to us now.
3194 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3195 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3196 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3199 * If the I/O on the transformed data was successful, generate any
3200 * checksum reports now while we still have the transformed data.
3202 if (zio->io_error == 0) {
3203 while (zio->io_cksum_report != NULL) {
3204 zio_cksum_report_t *zcr = zio->io_cksum_report;
3205 uint64_t align = zcr->zcr_align;
3206 uint64_t asize = P2ROUNDUP(psize, align);
3207 char *abuf = zio->io_data;
3209 if (asize != psize) {
3210 abuf = zio_buf_alloc(asize);
3211 bcopy(zio->io_data, abuf, psize);
3212 bzero(abuf + psize, asize - psize);
3215 zio->io_cksum_report = zcr->zcr_next;
3216 zcr->zcr_next = NULL;
3217 zcr->zcr_finish(zcr, abuf);
3218 zfs_ereport_free_checksum(zcr);
3221 zio_buf_free(abuf, asize);
3225 zio_pop_transforms(zio); /* note: may set zio->io_error */
3227 vdev_stat_update(zio, psize);
3229 if (zio->io_error) {
3231 * If this I/O is attached to a particular vdev,
3232 * generate an error message describing the I/O failure
3233 * at the block level. We ignore these errors if the
3234 * device is currently unavailable.
3236 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3237 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3239 if ((zio->io_error == EIO || !(zio->io_flags &
3240 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3243 * For logical I/O requests, tell the SPA to log the
3244 * error and generate a logical data ereport.
3246 spa_log_error(spa, zio);
3247 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3252 if (zio->io_error && zio == lio) {
3254 * Determine whether zio should be reexecuted. This will
3255 * propagate all the way to the root via zio_notify_parent().
3257 ASSERT(vd == NULL && bp != NULL);
3258 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3260 if (IO_IS_ALLOCATING(zio) &&
3261 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3262 if (zio->io_error != ENOSPC)
3263 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3265 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3268 if ((zio->io_type == ZIO_TYPE_READ ||
3269 zio->io_type == ZIO_TYPE_FREE) &&
3270 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3271 zio->io_error == ENXIO &&
3272 spa_load_state(spa) == SPA_LOAD_NONE &&
3273 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3274 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3276 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3277 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3280 * Here is a possibly good place to attempt to do
3281 * either combinatorial reconstruction or error correction
3282 * based on checksums. It also might be a good place
3283 * to send out preliminary ereports before we suspend
3289 * If there were logical child errors, they apply to us now.
3290 * We defer this until now to avoid conflating logical child
3291 * errors with errors that happened to the zio itself when
3292 * updating vdev stats and reporting FMA events above.
3294 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3296 if ((zio->io_error || zio->io_reexecute) &&
3297 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3298 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3299 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3301 zio_gang_tree_free(&zio->io_gang_tree);
3304 * Godfather I/Os should never suspend.
3306 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3307 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3308 zio->io_reexecute = 0;
3310 if (zio->io_reexecute) {
3312 * This is a logical I/O that wants to reexecute.
3314 * Reexecute is top-down. When an i/o fails, if it's not
3315 * the root, it simply notifies its parent and sticks around.
3316 * The parent, seeing that it still has children in zio_done(),
3317 * does the same. This percolates all the way up to the root.
3318 * The root i/o will reexecute or suspend the entire tree.
3320 * This approach ensures that zio_reexecute() honors
3321 * all the original i/o dependency relationships, e.g.
3322 * parents not executing until children are ready.
3324 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3326 zio->io_gang_leader = NULL;
3328 mutex_enter(&zio->io_lock);
3329 zio->io_state[ZIO_WAIT_DONE] = 1;
3330 mutex_exit(&zio->io_lock);
3333 * "The Godfather" I/O monitors its children but is
3334 * not a true parent to them. It will track them through
3335 * the pipeline but severs its ties whenever they get into
3336 * trouble (e.g. suspended). This allows "The Godfather"
3337 * I/O to return status without blocking.
3339 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3340 zio_link_t *zl = zio->io_walk_link;
3341 pio_next = zio_walk_parents(zio);
3343 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3344 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3345 zio_remove_child(pio, zio, zl);
3346 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3350 if ((pio = zio_unique_parent(zio)) != NULL) {
3352 * We're not a root i/o, so there's nothing to do
3353 * but notify our parent. Don't propagate errors
3354 * upward since we haven't permanently failed yet.
3356 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3357 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3358 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3359 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3361 * We'd fail again if we reexecuted now, so suspend
3362 * until conditions improve (e.g. device comes online).
3364 zio_suspend(spa, zio);
3367 * Reexecution is potentially a huge amount of work.
3368 * Hand it off to the otherwise-unused claim taskq.
3370 #if defined(illumos) || !defined(_KERNEL)
3371 ASSERT(zio->io_tqent.tqent_next == NULL);
3373 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3375 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3376 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3379 return (ZIO_PIPELINE_STOP);
3382 ASSERT(zio->io_child_count == 0);
3383 ASSERT(zio->io_reexecute == 0);
3384 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3387 * Report any checksum errors, since the I/O is complete.
3389 while (zio->io_cksum_report != NULL) {
3390 zio_cksum_report_t *zcr = zio->io_cksum_report;
3391 zio->io_cksum_report = zcr->zcr_next;
3392 zcr->zcr_next = NULL;
3393 zcr->zcr_finish(zcr, NULL);
3394 zfs_ereport_free_checksum(zcr);
3398 * It is the responsibility of the done callback to ensure that this
3399 * particular zio is no longer discoverable for adoption, and as
3400 * such, cannot acquire any new parents.
3405 mutex_enter(&zio->io_lock);
3406 zio->io_state[ZIO_WAIT_DONE] = 1;
3407 mutex_exit(&zio->io_lock);
3409 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3410 zio_link_t *zl = zio->io_walk_link;
3411 pio_next = zio_walk_parents(zio);
3412 zio_remove_child(pio, zio, zl);
3413 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3416 if (zio->io_waiter != NULL) {
3417 mutex_enter(&zio->io_lock);
3418 zio->io_executor = NULL;
3419 cv_broadcast(&zio->io_cv);
3420 mutex_exit(&zio->io_lock);
3425 return (ZIO_PIPELINE_STOP);
3429 * ==========================================================================
3430 * I/O pipeline definition
3431 * ==========================================================================
3433 static zio_pipe_stage_t *zio_pipeline[] = {
3439 zio_checksum_generate,
3454 zio_checksum_verify,
3458 /* dnp is the dnode for zb1->zb_object */
3460 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3461 const zbookmark_phys_t *zb2)
3463 uint64_t zb1nextL0, zb2thisobj;
3465 ASSERT(zb1->zb_objset == zb2->zb_objset);
3466 ASSERT(zb2->zb_level == 0);
3468 /* The objset_phys_t isn't before anything. */
3472 zb1nextL0 = (zb1->zb_blkid + 1) <<
3473 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3475 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3476 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3478 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3479 uint64_t nextobj = zb1nextL0 *
3480 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3481 return (nextobj <= zb2thisobj);
3484 if (zb1->zb_object < zb2thisobj)
3486 if (zb1->zb_object > zb2thisobj)
3488 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3490 return (zb1nextL0 <= zb2->zb_blkid);