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;
96 #define ZIO_PIPELINE_CONTINUE 0x100
97 #define ZIO_PIPELINE_STOP 0x101
99 #define BP_SPANB(indblkshift, level) \
100 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
101 #define COMPARE_META_LEVEL 0x80000000ul
103 * The following actions directly effect the spa's sync-to-convergence logic.
104 * The values below define the sync pass when we start performing the action.
105 * Care should be taken when changing these values as they directly impact
106 * spa_sync() performance. Tuning these values may introduce subtle performance
107 * pathologies and should only be done in the context of performance analysis.
108 * These tunables will eventually be removed and replaced with #defines once
109 * enough analysis has been done to determine optimal values.
111 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
112 * regular blocks are not deferred.
114 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
115 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
116 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
117 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
118 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
119 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
120 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
121 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
122 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
123 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
124 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
125 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
128 * An allocating zio is one that either currently has the DVA allocate
129 * stage set or will have it later in its lifetime.
131 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
133 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
136 int zio_buf_debug_limit = 16384;
138 int zio_buf_debug_limit = 0;
145 zio_cache = kmem_cache_create("zio_cache",
146 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
147 zio_link_cache = kmem_cache_create("zio_link_cache",
148 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
153 * For small buffers, we want a cache for each multiple of
154 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
155 * for each quarter-power of 2.
157 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
158 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
161 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
169 * If we are using watchpoints, put each buffer on its own page,
170 * to eliminate the performance overhead of trapping to the
171 * kernel when modifying a non-watched buffer that shares the
172 * page with a watched buffer.
174 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
178 if (size <= 4 * SPA_MINBLOCKSIZE) {
179 align = SPA_MINBLOCKSIZE;
180 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
181 align = MIN(p2 >> 2, PAGESIZE);
186 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
187 zio_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL, NULL, cflags);
191 * Since zio_data bufs do not appear in crash dumps, we
192 * pass KMC_NOTOUCH so that no allocator metadata is
193 * stored with the buffers.
195 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
196 zio_data_buf_cache[c] = kmem_cache_create(name, size,
197 align, NULL, NULL, NULL, NULL, NULL,
198 cflags | KMC_NOTOUCH | KMC_NODEBUG);
203 ASSERT(zio_buf_cache[c] != NULL);
204 if (zio_buf_cache[c - 1] == NULL)
205 zio_buf_cache[c - 1] = zio_buf_cache[c];
207 ASSERT(zio_data_buf_cache[c] != NULL);
208 if (zio_data_buf_cache[c - 1] == NULL)
209 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
215 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
217 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
220 if (zio_trim_ksp != NULL) {
221 zio_trim_ksp->ks_data = &zio_trim_stats;
222 kstat_install(zio_trim_ksp);
230 kmem_cache_t *last_cache = NULL;
231 kmem_cache_t *last_data_cache = NULL;
233 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
234 if (zio_buf_cache[c] != last_cache) {
235 last_cache = zio_buf_cache[c];
236 kmem_cache_destroy(zio_buf_cache[c]);
238 zio_buf_cache[c] = NULL;
240 if (zio_data_buf_cache[c] != last_data_cache) {
241 last_data_cache = zio_data_buf_cache[c];
242 kmem_cache_destroy(zio_data_buf_cache[c]);
244 zio_data_buf_cache[c] = NULL;
247 kmem_cache_destroy(zio_link_cache);
248 kmem_cache_destroy(zio_cache);
252 if (zio_trim_ksp != NULL) {
253 kstat_delete(zio_trim_ksp);
259 * ==========================================================================
260 * Allocate and free I/O buffers
261 * ==========================================================================
265 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
266 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
267 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
268 * excess / transient data in-core during a crashdump.
271 zio_buf_alloc(size_t size)
273 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
274 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
276 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
279 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
281 return (kmem_alloc(size, KM_SLEEP|flags));
285 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
286 * crashdump if the kernel panics. This exists so that we will limit the amount
287 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
288 * of kernel heap dumped to disk when the kernel panics)
291 zio_data_buf_alloc(size_t size)
293 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
295 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
300 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
304 zio_buf_free(void *buf, size_t size)
306 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
308 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
311 kmem_cache_free(zio_buf_cache[c], buf);
313 kmem_free(buf, size);
317 zio_data_buf_free(void *buf, size_t size)
319 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
321 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
324 kmem_cache_free(zio_data_buf_cache[c], buf);
326 kmem_free(buf, size);
330 * ==========================================================================
331 * Push and pop I/O transform buffers
332 * ==========================================================================
335 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
336 zio_transform_func_t *transform)
338 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
340 zt->zt_orig_data = zio->io_data;
341 zt->zt_orig_size = zio->io_size;
342 zt->zt_bufsize = bufsize;
343 zt->zt_transform = transform;
345 zt->zt_next = zio->io_transform_stack;
346 zio->io_transform_stack = zt;
353 zio_pop_transforms(zio_t *zio)
357 while ((zt = zio->io_transform_stack) != NULL) {
358 if (zt->zt_transform != NULL)
359 zt->zt_transform(zio,
360 zt->zt_orig_data, zt->zt_orig_size);
362 if (zt->zt_bufsize != 0)
363 zio_buf_free(zio->io_data, zt->zt_bufsize);
365 zio->io_data = zt->zt_orig_data;
366 zio->io_size = zt->zt_orig_size;
367 zio->io_transform_stack = zt->zt_next;
369 kmem_free(zt, sizeof (zio_transform_t));
374 * ==========================================================================
375 * I/O transform callbacks for subblocks and decompression
376 * ==========================================================================
379 zio_subblock(zio_t *zio, void *data, uint64_t size)
381 ASSERT(zio->io_size > size);
383 if (zio->io_type == ZIO_TYPE_READ)
384 bcopy(zio->io_data, data, size);
388 zio_decompress(zio_t *zio, void *data, uint64_t size)
390 if (zio->io_error == 0 &&
391 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
392 zio->io_data, data, zio->io_size, size) != 0)
393 zio->io_error = SET_ERROR(EIO);
397 * ==========================================================================
398 * I/O parent/child relationships and pipeline interlocks
399 * ==========================================================================
402 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
403 * continue calling these functions until they return NULL.
404 * Otherwise, the next caller will pick up the list walk in
405 * some indeterminate state. (Otherwise every caller would
406 * have to pass in a cookie to keep the state represented by
407 * io_walk_link, which gets annoying.)
410 zio_walk_parents(zio_t *cio)
412 zio_link_t *zl = cio->io_walk_link;
413 list_t *pl = &cio->io_parent_list;
415 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
416 cio->io_walk_link = zl;
421 ASSERT(zl->zl_child == cio);
422 return (zl->zl_parent);
426 zio_walk_children(zio_t *pio)
428 zio_link_t *zl = pio->io_walk_link;
429 list_t *cl = &pio->io_child_list;
431 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
432 pio->io_walk_link = zl;
437 ASSERT(zl->zl_parent == pio);
438 return (zl->zl_child);
442 zio_unique_parent(zio_t *cio)
444 zio_t *pio = zio_walk_parents(cio);
446 VERIFY(zio_walk_parents(cio) == NULL);
451 zio_add_child(zio_t *pio, zio_t *cio)
453 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
456 * Logical I/Os can have logical, gang, or vdev children.
457 * Gang I/Os can have gang or vdev children.
458 * Vdev I/Os can only have vdev children.
459 * The following ASSERT captures all of these constraints.
461 ASSERT(cio->io_child_type <= pio->io_child_type);
466 mutex_enter(&cio->io_lock);
467 mutex_enter(&pio->io_lock);
469 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
471 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
472 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
474 list_insert_head(&pio->io_child_list, zl);
475 list_insert_head(&cio->io_parent_list, zl);
477 pio->io_child_count++;
478 cio->io_parent_count++;
480 mutex_exit(&pio->io_lock);
481 mutex_exit(&cio->io_lock);
485 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
487 ASSERT(zl->zl_parent == pio);
488 ASSERT(zl->zl_child == cio);
490 mutex_enter(&cio->io_lock);
491 mutex_enter(&pio->io_lock);
493 list_remove(&pio->io_child_list, zl);
494 list_remove(&cio->io_parent_list, zl);
496 pio->io_child_count--;
497 cio->io_parent_count--;
499 mutex_exit(&pio->io_lock);
500 mutex_exit(&cio->io_lock);
502 kmem_cache_free(zio_link_cache, zl);
506 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
508 uint64_t *countp = &zio->io_children[child][wait];
509 boolean_t waiting = B_FALSE;
511 mutex_enter(&zio->io_lock);
512 ASSERT(zio->io_stall == NULL);
515 zio->io_stall = countp;
518 mutex_exit(&zio->io_lock);
524 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
526 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
527 int *errorp = &pio->io_child_error[zio->io_child_type];
529 mutex_enter(&pio->io_lock);
530 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
531 *errorp = zio_worst_error(*errorp, zio->io_error);
532 pio->io_reexecute |= zio->io_reexecute;
533 ASSERT3U(*countp, >, 0);
537 if (*countp == 0 && pio->io_stall == countp) {
538 pio->io_stall = NULL;
539 mutex_exit(&pio->io_lock);
542 mutex_exit(&pio->io_lock);
547 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
549 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
550 zio->io_error = zio->io_child_error[c];
554 * ==========================================================================
555 * Create the various types of I/O (read, write, free, etc)
556 * ==========================================================================
559 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
560 void *data, uint64_t size, zio_done_func_t *done, void *private,
561 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
562 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
563 enum zio_stage stage, enum zio_stage pipeline)
567 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
568 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
569 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
571 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
572 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
573 ASSERT(vd || stage == ZIO_STAGE_OPEN);
575 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
576 bzero(zio, sizeof (zio_t));
578 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
579 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
581 list_create(&zio->io_parent_list, sizeof (zio_link_t),
582 offsetof(zio_link_t, zl_parent_node));
583 list_create(&zio->io_child_list, sizeof (zio_link_t),
584 offsetof(zio_link_t, zl_child_node));
587 zio->io_child_type = ZIO_CHILD_VDEV;
588 else if (flags & ZIO_FLAG_GANG_CHILD)
589 zio->io_child_type = ZIO_CHILD_GANG;
590 else if (flags & ZIO_FLAG_DDT_CHILD)
591 zio->io_child_type = ZIO_CHILD_DDT;
593 zio->io_child_type = ZIO_CHILD_LOGICAL;
596 zio->io_bp = (blkptr_t *)bp;
597 zio->io_bp_copy = *bp;
598 zio->io_bp_orig = *bp;
599 if (type != ZIO_TYPE_WRITE ||
600 zio->io_child_type == ZIO_CHILD_DDT)
601 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
602 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
603 zio->io_logical = zio;
604 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
605 pipeline |= ZIO_GANG_STAGES;
611 zio->io_private = private;
613 zio->io_priority = priority;
615 zio->io_offset = offset;
616 zio->io_orig_data = zio->io_data = data;
617 zio->io_orig_size = zio->io_size = size;
618 zio->io_orig_flags = zio->io_flags = flags;
619 zio->io_orig_stage = zio->io_stage = stage;
620 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
622 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
623 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
626 zio->io_bookmark = *zb;
629 if (zio->io_logical == NULL)
630 zio->io_logical = pio->io_logical;
631 if (zio->io_child_type == ZIO_CHILD_GANG)
632 zio->io_gang_leader = pio->io_gang_leader;
633 zio_add_child(pio, zio);
640 zio_destroy(zio_t *zio)
642 list_destroy(&zio->io_parent_list);
643 list_destroy(&zio->io_child_list);
644 mutex_destroy(&zio->io_lock);
645 cv_destroy(&zio->io_cv);
646 kmem_cache_free(zio_cache, zio);
650 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
651 void *private, enum zio_flag flags)
655 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
656 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
657 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
663 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
665 return (zio_null(NULL, spa, NULL, done, private, flags));
669 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
671 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
672 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
673 bp, (longlong_t)BP_GET_TYPE(bp));
675 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
676 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
677 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
678 bp, (longlong_t)BP_GET_CHECKSUM(bp));
680 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
681 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
682 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
683 bp, (longlong_t)BP_GET_COMPRESS(bp));
685 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
686 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
687 bp, (longlong_t)BP_GET_LSIZE(bp));
689 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
690 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
691 bp, (longlong_t)BP_GET_PSIZE(bp));
694 if (BP_IS_EMBEDDED(bp)) {
695 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
696 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
697 bp, (longlong_t)BPE_GET_ETYPE(bp));
702 * Pool-specific checks.
704 * Note: it would be nice to verify that the blk_birth and
705 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
706 * allows the birth time of log blocks (and dmu_sync()-ed blocks
707 * that are in the log) to be arbitrarily large.
709 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
710 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
711 if (vdevid >= spa->spa_root_vdev->vdev_children) {
712 zfs_panic_recover("blkptr at %p DVA %u has invalid "
714 bp, i, (longlong_t)vdevid);
717 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
719 zfs_panic_recover("blkptr at %p DVA %u has invalid "
721 bp, i, (longlong_t)vdevid);
724 if (vd->vdev_ops == &vdev_hole_ops) {
725 zfs_panic_recover("blkptr at %p DVA %u has hole "
727 bp, i, (longlong_t)vdevid);
730 if (vd->vdev_ops == &vdev_missing_ops) {
732 * "missing" vdevs are valid during import, but we
733 * don't have their detailed info (e.g. asize), so
734 * we can't perform any more checks on them.
738 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
739 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
741 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
742 if (offset + asize > vd->vdev_asize) {
743 zfs_panic_recover("blkptr at %p DVA %u has invalid "
745 bp, i, (longlong_t)offset);
751 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
752 void *data, uint64_t size, zio_done_func_t *done, void *private,
753 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
757 zfs_blkptr_verify(spa, bp);
759 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
760 data, size, done, private,
761 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
762 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
763 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
769 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
770 void *data, uint64_t size, const zio_prop_t *zp,
771 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
773 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
777 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
778 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
779 zp->zp_compress >= ZIO_COMPRESS_OFF &&
780 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
781 DMU_OT_IS_VALID(zp->zp_type) &&
784 zp->zp_copies <= spa_max_replication(spa));
786 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
787 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
788 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
789 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
791 zio->io_ready = ready;
792 zio->io_physdone = physdone;
796 * Data can be NULL if we are going to call zio_write_override() to
797 * provide the already-allocated BP. But we may need the data to
798 * verify a dedup hit (if requested). In this case, don't try to
799 * dedup (just take the already-allocated BP verbatim).
801 if (data == NULL && zio->io_prop.zp_dedup_verify) {
802 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
809 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
810 uint64_t size, zio_done_func_t *done, void *private,
811 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
815 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
816 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
817 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
823 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
825 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
826 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
827 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
828 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
831 * We must reset the io_prop to match the values that existed
832 * when the bp was first written by dmu_sync() keeping in mind
833 * that nopwrite and dedup are mutually exclusive.
835 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
836 zio->io_prop.zp_nopwrite = nopwrite;
837 zio->io_prop.zp_copies = copies;
838 zio->io_bp_override = bp;
842 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
846 * The check for EMBEDDED is a performance optimization. We
847 * process the free here (by ignoring it) rather than
848 * putting it on the list and then processing it in zio_free_sync().
850 if (BP_IS_EMBEDDED(bp))
852 metaslab_check_free(spa, bp);
855 * Frees that are for the currently-syncing txg, are not going to be
856 * deferred, and which will not need to do a read (i.e. not GANG or
857 * DEDUP), can be processed immediately. Otherwise, put them on the
858 * in-memory list for later processing.
860 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
861 txg != spa->spa_syncing_txg ||
862 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
863 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
865 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
866 BP_GET_PSIZE(bp), 0)));
871 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
872 uint64_t size, enum zio_flag flags)
875 enum zio_stage stage = ZIO_FREE_PIPELINE;
877 ASSERT(!BP_IS_HOLE(bp));
878 ASSERT(spa_syncing_txg(spa) == txg);
879 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
881 if (BP_IS_EMBEDDED(bp))
882 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
884 metaslab_check_free(spa, bp);
887 if (zfs_trim_enabled)
888 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
889 ZIO_STAGE_VDEV_IO_ASSESS;
891 * GANG and DEDUP blocks can induce a read (for the gang block header,
892 * or the DDT), so issue them asynchronously so that this thread is
895 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
896 stage |= ZIO_STAGE_ISSUE_ASYNC;
898 flags |= ZIO_FLAG_DONT_QUEUE;
900 zio = zio_create(pio, spa, txg, bp, NULL, size,
901 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
902 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
908 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
909 zio_done_func_t *done, void *private, enum zio_flag flags)
913 dprintf_bp(bp, "claiming in txg %llu", txg);
915 if (BP_IS_EMBEDDED(bp))
916 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
919 * A claim is an allocation of a specific block. Claims are needed
920 * to support immediate writes in the intent log. The issue is that
921 * immediate writes contain committed data, but in a txg that was
922 * *not* committed. Upon opening the pool after an unclean shutdown,
923 * the intent log claims all blocks that contain immediate write data
924 * so that the SPA knows they're in use.
926 * All claims *must* be resolved in the first txg -- before the SPA
927 * starts allocating blocks -- so that nothing is allocated twice.
928 * If txg == 0 we just verify that the block is claimable.
930 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
931 ASSERT(txg == spa_first_txg(spa) || txg == 0);
932 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
934 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
935 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
936 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
942 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
943 uint64_t size, zio_done_func_t *done, void *private,
944 zio_priority_t priority, enum zio_flag flags)
949 if (vd->vdev_children == 0) {
950 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
951 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
952 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
956 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
958 for (c = 0; c < vd->vdev_children; c++)
959 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
960 offset, size, done, private, priority, flags));
967 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
968 void *data, int checksum, zio_done_func_t *done, void *private,
969 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
973 ASSERT(vd->vdev_children == 0);
974 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
975 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
976 ASSERT3U(offset + size, <=, vd->vdev_psize);
978 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
979 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
980 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
982 zio->io_prop.zp_checksum = checksum;
988 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
989 void *data, int checksum, zio_done_func_t *done, void *private,
990 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
994 ASSERT(vd->vdev_children == 0);
995 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
996 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
997 ASSERT3U(offset + size, <=, vd->vdev_psize);
999 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1000 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1001 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1003 zio->io_prop.zp_checksum = checksum;
1005 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1007 * zec checksums are necessarily destructive -- they modify
1008 * the end of the write buffer to hold the verifier/checksum.
1009 * Therefore, we must make a local copy in case the data is
1010 * being written to multiple places in parallel.
1012 void *wbuf = zio_buf_alloc(size);
1013 bcopy(data, wbuf, size);
1014 zio_push_transform(zio, wbuf, size, size, NULL);
1021 * Create a child I/O to do some work for us.
1024 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1025 void *data, uint64_t size, int type, zio_priority_t priority,
1026 enum zio_flag flags, zio_done_func_t *done, void *private)
1028 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1031 ASSERT(vd->vdev_parent ==
1032 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1034 if (type == ZIO_TYPE_READ && bp != NULL) {
1036 * If we have the bp, then the child should perform the
1037 * checksum and the parent need not. This pushes error
1038 * detection as close to the leaves as possible and
1039 * eliminates redundant checksums in the interior nodes.
1041 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1042 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1045 /* Not all IO types require vdev io done stage e.g. free */
1046 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1047 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1049 if (vd->vdev_children == 0)
1050 offset += VDEV_LABEL_START_SIZE;
1052 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1055 * If we've decided to do a repair, the write is not speculative --
1056 * even if the original read was.
1058 if (flags & ZIO_FLAG_IO_REPAIR)
1059 flags &= ~ZIO_FLAG_SPECULATIVE;
1061 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1062 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1063 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1065 zio->io_physdone = pio->io_physdone;
1066 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1067 zio->io_logical->io_phys_children++;
1073 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1074 int type, zio_priority_t priority, enum zio_flag flags,
1075 zio_done_func_t *done, void *private)
1079 ASSERT(vd->vdev_ops->vdev_op_leaf);
1081 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1082 data, size, done, private, type, priority,
1083 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1085 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1091 zio_flush(zio_t *zio, vdev_t *vd)
1093 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1094 NULL, NULL, ZIO_PRIORITY_NOW,
1095 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1099 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1102 ASSERT(vd->vdev_ops->vdev_op_leaf);
1104 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1105 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1106 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1107 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1111 zio_shrink(zio_t *zio, uint64_t size)
1113 ASSERT(zio->io_executor == NULL);
1114 ASSERT(zio->io_orig_size == zio->io_size);
1115 ASSERT(size <= zio->io_size);
1118 * We don't shrink for raidz because of problems with the
1119 * reconstruction when reading back less than the block size.
1120 * Note, BP_IS_RAIDZ() assumes no compression.
1122 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1123 if (!BP_IS_RAIDZ(zio->io_bp))
1124 zio->io_orig_size = zio->io_size = size;
1128 * ==========================================================================
1129 * Prepare to read and write logical blocks
1130 * ==========================================================================
1134 zio_read_bp_init(zio_t *zio)
1136 blkptr_t *bp = zio->io_bp;
1138 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1139 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1140 !(zio->io_flags & ZIO_FLAG_RAW)) {
1142 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1143 void *cbuf = zio_buf_alloc(psize);
1145 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1148 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1149 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1150 decode_embedded_bp_compressed(bp, zio->io_data);
1152 ASSERT(!BP_IS_EMBEDDED(bp));
1155 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1156 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1158 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1159 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1161 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1162 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1164 return (ZIO_PIPELINE_CONTINUE);
1168 zio_write_bp_init(zio_t *zio)
1170 spa_t *spa = zio->io_spa;
1171 zio_prop_t *zp = &zio->io_prop;
1172 enum zio_compress compress = zp->zp_compress;
1173 blkptr_t *bp = zio->io_bp;
1174 uint64_t lsize = zio->io_size;
1175 uint64_t psize = lsize;
1179 * If our children haven't all reached the ready stage,
1180 * wait for them and then repeat this pipeline stage.
1182 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1183 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1184 return (ZIO_PIPELINE_STOP);
1186 if (!IO_IS_ALLOCATING(zio))
1187 return (ZIO_PIPELINE_CONTINUE);
1189 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1191 if (zio->io_bp_override) {
1192 ASSERT(bp->blk_birth != zio->io_txg);
1193 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1195 *bp = *zio->io_bp_override;
1196 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1198 if (BP_IS_EMBEDDED(bp))
1199 return (ZIO_PIPELINE_CONTINUE);
1202 * If we've been overridden and nopwrite is set then
1203 * set the flag accordingly to indicate that a nopwrite
1204 * has already occurred.
1206 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1207 ASSERT(!zp->zp_dedup);
1208 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1209 return (ZIO_PIPELINE_CONTINUE);
1212 ASSERT(!zp->zp_nopwrite);
1214 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1215 return (ZIO_PIPELINE_CONTINUE);
1217 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1218 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1220 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1221 BP_SET_DEDUP(bp, 1);
1222 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1223 return (ZIO_PIPELINE_CONTINUE);
1227 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1229 * We're rewriting an existing block, which means we're
1230 * working on behalf of spa_sync(). For spa_sync() to
1231 * converge, it must eventually be the case that we don't
1232 * have to allocate new blocks. But compression changes
1233 * the blocksize, which forces a reallocate, and makes
1234 * convergence take longer. Therefore, after the first
1235 * few passes, stop compressing to ensure convergence.
1237 pass = spa_sync_pass(spa);
1239 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1240 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1241 ASSERT(!BP_GET_DEDUP(bp));
1243 if (pass >= zfs_sync_pass_dont_compress)
1244 compress = ZIO_COMPRESS_OFF;
1246 /* Make sure someone doesn't change their mind on overwrites */
1247 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1248 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1251 if (compress != ZIO_COMPRESS_OFF) {
1252 void *cbuf = zio_buf_alloc(lsize);
1253 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1254 if (psize == 0 || psize == lsize) {
1255 compress = ZIO_COMPRESS_OFF;
1256 zio_buf_free(cbuf, lsize);
1257 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1258 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1259 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1260 encode_embedded_bp_compressed(bp,
1261 cbuf, compress, lsize, psize);
1262 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1263 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1264 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1265 zio_buf_free(cbuf, lsize);
1266 bp->blk_birth = zio->io_txg;
1267 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1268 ASSERT(spa_feature_is_active(spa,
1269 SPA_FEATURE_EMBEDDED_DATA));
1270 return (ZIO_PIPELINE_CONTINUE);
1273 * Round up compressed size up to the ashift
1274 * of the smallest-ashift device, and zero the tail.
1275 * This ensures that the compressed size of the BP
1276 * (and thus compressratio property) are correct,
1277 * in that we charge for the padding used to fill out
1280 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1281 size_t rounded = (size_t)P2ROUNDUP(psize,
1282 1ULL << spa->spa_min_ashift);
1283 if (rounded >= lsize) {
1284 compress = ZIO_COMPRESS_OFF;
1285 zio_buf_free(cbuf, lsize);
1288 bzero((char *)cbuf + psize, rounded - psize);
1290 zio_push_transform(zio, cbuf,
1291 psize, lsize, NULL);
1297 * The final pass of spa_sync() must be all rewrites, but the first
1298 * few passes offer a trade-off: allocating blocks defers convergence,
1299 * but newly allocated blocks are sequential, so they can be written
1300 * to disk faster. Therefore, we allow the first few passes of
1301 * spa_sync() to allocate new blocks, but force rewrites after that.
1302 * There should only be a handful of blocks after pass 1 in any case.
1304 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1305 BP_GET_PSIZE(bp) == psize &&
1306 pass >= zfs_sync_pass_rewrite) {
1308 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1309 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1310 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1313 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1317 if (zio->io_bp_orig.blk_birth != 0 &&
1318 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1319 BP_SET_LSIZE(bp, lsize);
1320 BP_SET_TYPE(bp, zp->zp_type);
1321 BP_SET_LEVEL(bp, zp->zp_level);
1322 BP_SET_BIRTH(bp, zio->io_txg, 0);
1324 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1326 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1327 BP_SET_LSIZE(bp, lsize);
1328 BP_SET_TYPE(bp, zp->zp_type);
1329 BP_SET_LEVEL(bp, zp->zp_level);
1330 BP_SET_PSIZE(bp, psize);
1331 BP_SET_COMPRESS(bp, compress);
1332 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1333 BP_SET_DEDUP(bp, zp->zp_dedup);
1334 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1336 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1337 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1338 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1340 if (zp->zp_nopwrite) {
1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1342 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1343 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1347 return (ZIO_PIPELINE_CONTINUE);
1351 zio_free_bp_init(zio_t *zio)
1353 blkptr_t *bp = zio->io_bp;
1355 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1356 if (BP_GET_DEDUP(bp))
1357 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1360 return (ZIO_PIPELINE_CONTINUE);
1364 * ==========================================================================
1365 * Execute the I/O pipeline
1366 * ==========================================================================
1370 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1372 spa_t *spa = zio->io_spa;
1373 zio_type_t t = zio->io_type;
1374 int flags = (cutinline ? TQ_FRONT : 0);
1376 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1379 * If we're a config writer or a probe, the normal issue and
1380 * interrupt threads may all be blocked waiting for the config lock.
1381 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1383 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1387 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1389 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1393 * If this is a high priority I/O, then use the high priority taskq if
1396 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1397 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1400 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1403 * NB: We are assuming that the zio can only be dispatched
1404 * to a single taskq at a time. It would be a grievous error
1405 * to dispatch the zio to another taskq at the same time.
1407 #if defined(illumos) || !defined(_KERNEL)
1408 ASSERT(zio->io_tqent.tqent_next == NULL);
1410 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1412 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1413 flags, &zio->io_tqent);
1417 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1419 kthread_t *executor = zio->io_executor;
1420 spa_t *spa = zio->io_spa;
1422 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1423 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1425 for (i = 0; i < tqs->stqs_count; i++) {
1426 if (taskq_member(tqs->stqs_taskq[i], executor))
1435 zio_issue_async(zio_t *zio)
1437 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1439 return (ZIO_PIPELINE_STOP);
1443 zio_interrupt(zio_t *zio)
1445 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1449 * Execute the I/O pipeline until one of the following occurs:
1451 * (1) the I/O completes
1452 * (2) the pipeline stalls waiting for dependent child I/Os
1453 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1454 * (4) the I/O is delegated by vdev-level caching or aggregation
1455 * (5) the I/O is deferred due to vdev-level queueing
1456 * (6) the I/O is handed off to another thread.
1458 * In all cases, the pipeline stops whenever there's no CPU work; it never
1459 * burns a thread in cv_wait().
1461 * There's no locking on io_stage because there's no legitimate way
1462 * for multiple threads to be attempting to process the same I/O.
1464 static zio_pipe_stage_t *zio_pipeline[];
1467 zio_execute(zio_t *zio)
1469 zio->io_executor = curthread;
1471 while (zio->io_stage < ZIO_STAGE_DONE) {
1472 enum zio_stage pipeline = zio->io_pipeline;
1473 enum zio_stage stage = zio->io_stage;
1476 ASSERT(!MUTEX_HELD(&zio->io_lock));
1477 ASSERT(ISP2(stage));
1478 ASSERT(zio->io_stall == NULL);
1482 } while ((stage & pipeline) == 0);
1484 ASSERT(stage <= ZIO_STAGE_DONE);
1487 * If we are in interrupt context and this pipeline stage
1488 * will grab a config lock that is held across I/O,
1489 * or may wait for an I/O that needs an interrupt thread
1490 * to complete, issue async to avoid deadlock.
1492 * For VDEV_IO_START, we cut in line so that the io will
1493 * be sent to disk promptly.
1495 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1496 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1497 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1498 zio_requeue_io_start_cut_in_line : B_FALSE;
1499 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1503 zio->io_stage = stage;
1504 rv = zio_pipeline[highbit64(stage) - 1](zio);
1506 if (rv == ZIO_PIPELINE_STOP)
1509 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1514 * ==========================================================================
1515 * Initiate I/O, either sync or async
1516 * ==========================================================================
1519 zio_wait(zio_t *zio)
1523 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1524 ASSERT(zio->io_executor == NULL);
1526 zio->io_waiter = curthread;
1530 mutex_enter(&zio->io_lock);
1531 while (zio->io_executor != NULL)
1532 cv_wait(&zio->io_cv, &zio->io_lock);
1533 mutex_exit(&zio->io_lock);
1535 error = zio->io_error;
1542 zio_nowait(zio_t *zio)
1544 ASSERT(zio->io_executor == NULL);
1546 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1547 zio_unique_parent(zio) == NULL) {
1549 * This is a logical async I/O with no parent to wait for it.
1550 * We add it to the spa_async_root_zio "Godfather" I/O which
1551 * will ensure they complete prior to unloading the pool.
1553 spa_t *spa = zio->io_spa;
1555 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1562 * ==========================================================================
1563 * Reexecute or suspend/resume failed I/O
1564 * ==========================================================================
1568 zio_reexecute(zio_t *pio)
1570 zio_t *cio, *cio_next;
1572 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1573 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1574 ASSERT(pio->io_gang_leader == NULL);
1575 ASSERT(pio->io_gang_tree == NULL);
1577 pio->io_flags = pio->io_orig_flags;
1578 pio->io_stage = pio->io_orig_stage;
1579 pio->io_pipeline = pio->io_orig_pipeline;
1580 pio->io_reexecute = 0;
1581 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1583 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1584 pio->io_state[w] = 0;
1585 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1586 pio->io_child_error[c] = 0;
1588 if (IO_IS_ALLOCATING(pio))
1589 BP_ZERO(pio->io_bp);
1592 * As we reexecute pio's children, new children could be created.
1593 * New children go to the head of pio's io_child_list, however,
1594 * so we will (correctly) not reexecute them. The key is that
1595 * the remainder of pio's io_child_list, from 'cio_next' onward,
1596 * cannot be affected by any side effects of reexecuting 'cio'.
1598 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1599 cio_next = zio_walk_children(pio);
1600 mutex_enter(&pio->io_lock);
1601 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1602 pio->io_children[cio->io_child_type][w]++;
1603 mutex_exit(&pio->io_lock);
1608 * Now that all children have been reexecuted, execute the parent.
1609 * We don't reexecute "The Godfather" I/O here as it's the
1610 * responsibility of the caller to wait on him.
1612 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1617 zio_suspend(spa_t *spa, zio_t *zio)
1619 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1620 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1621 "failure and the failure mode property for this pool "
1622 "is set to panic.", spa_name(spa));
1624 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1626 mutex_enter(&spa->spa_suspend_lock);
1628 if (spa->spa_suspend_zio_root == NULL)
1629 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1630 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1631 ZIO_FLAG_GODFATHER);
1633 spa->spa_suspended = B_TRUE;
1636 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1637 ASSERT(zio != spa->spa_suspend_zio_root);
1638 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1639 ASSERT(zio_unique_parent(zio) == NULL);
1640 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1641 zio_add_child(spa->spa_suspend_zio_root, zio);
1644 mutex_exit(&spa->spa_suspend_lock);
1648 zio_resume(spa_t *spa)
1653 * Reexecute all previously suspended i/o.
1655 mutex_enter(&spa->spa_suspend_lock);
1656 spa->spa_suspended = B_FALSE;
1657 cv_broadcast(&spa->spa_suspend_cv);
1658 pio = spa->spa_suspend_zio_root;
1659 spa->spa_suspend_zio_root = NULL;
1660 mutex_exit(&spa->spa_suspend_lock);
1666 return (zio_wait(pio));
1670 zio_resume_wait(spa_t *spa)
1672 mutex_enter(&spa->spa_suspend_lock);
1673 while (spa_suspended(spa))
1674 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1675 mutex_exit(&spa->spa_suspend_lock);
1679 * ==========================================================================
1682 * A gang block is a collection of small blocks that looks to the DMU
1683 * like one large block. When zio_dva_allocate() cannot find a block
1684 * of the requested size, due to either severe fragmentation or the pool
1685 * being nearly full, it calls zio_write_gang_block() to construct the
1686 * block from smaller fragments.
1688 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1689 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1690 * an indirect block: it's an array of block pointers. It consumes
1691 * only one sector and hence is allocatable regardless of fragmentation.
1692 * The gang header's bps point to its gang members, which hold the data.
1694 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1695 * as the verifier to ensure uniqueness of the SHA256 checksum.
1696 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1697 * not the gang header. This ensures that data block signatures (needed for
1698 * deduplication) are independent of how the block is physically stored.
1700 * Gang blocks can be nested: a gang member may itself be a gang block.
1701 * Thus every gang block is a tree in which root and all interior nodes are
1702 * gang headers, and the leaves are normal blocks that contain user data.
1703 * The root of the gang tree is called the gang leader.
1705 * To perform any operation (read, rewrite, free, claim) on a gang block,
1706 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1707 * in the io_gang_tree field of the original logical i/o by recursively
1708 * reading the gang leader and all gang headers below it. This yields
1709 * an in-core tree containing the contents of every gang header and the
1710 * bps for every constituent of the gang block.
1712 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1713 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1714 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1715 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1716 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1717 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1718 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1719 * of the gang header plus zio_checksum_compute() of the data to update the
1720 * gang header's blk_cksum as described above.
1722 * The two-phase assemble/issue model solves the problem of partial failure --
1723 * what if you'd freed part of a gang block but then couldn't read the
1724 * gang header for another part? Assembling the entire gang tree first
1725 * ensures that all the necessary gang header I/O has succeeded before
1726 * starting the actual work of free, claim, or write. Once the gang tree
1727 * is assembled, free and claim are in-memory operations that cannot fail.
1729 * In the event that a gang write fails, zio_dva_unallocate() walks the
1730 * gang tree to immediately free (i.e. insert back into the space map)
1731 * everything we've allocated. This ensures that we don't get ENOSPC
1732 * errors during repeated suspend/resume cycles due to a flaky device.
1734 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1735 * the gang tree, we won't modify the block, so we can safely defer the free
1736 * (knowing that the block is still intact). If we *can* assemble the gang
1737 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1738 * each constituent bp and we can allocate a new block on the next sync pass.
1740 * In all cases, the gang tree allows complete recovery from partial failure.
1741 * ==========================================================================
1745 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1750 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1751 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1752 &pio->io_bookmark));
1756 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1761 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1762 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1763 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1765 * As we rewrite each gang header, the pipeline will compute
1766 * a new gang block header checksum for it; but no one will
1767 * compute a new data checksum, so we do that here. The one
1768 * exception is the gang leader: the pipeline already computed
1769 * its data checksum because that stage precedes gang assembly.
1770 * (Presently, nothing actually uses interior data checksums;
1771 * this is just good hygiene.)
1773 if (gn != pio->io_gang_leader->io_gang_tree) {
1774 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1775 data, BP_GET_PSIZE(bp));
1778 * If we are here to damage data for testing purposes,
1779 * leave the GBH alone so that we can detect the damage.
1781 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1782 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1784 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1785 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1786 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1794 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1796 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1797 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1798 ZIO_GANG_CHILD_FLAGS(pio)));
1803 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1805 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1806 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1809 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1818 static void zio_gang_tree_assemble_done(zio_t *zio);
1820 static zio_gang_node_t *
1821 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1823 zio_gang_node_t *gn;
1825 ASSERT(*gnpp == NULL);
1827 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1828 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1835 zio_gang_node_free(zio_gang_node_t **gnpp)
1837 zio_gang_node_t *gn = *gnpp;
1839 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1840 ASSERT(gn->gn_child[g] == NULL);
1842 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1843 kmem_free(gn, sizeof (*gn));
1848 zio_gang_tree_free(zio_gang_node_t **gnpp)
1850 zio_gang_node_t *gn = *gnpp;
1855 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1856 zio_gang_tree_free(&gn->gn_child[g]);
1858 zio_gang_node_free(gnpp);
1862 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1864 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1866 ASSERT(gio->io_gang_leader == gio);
1867 ASSERT(BP_IS_GANG(bp));
1869 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1870 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1871 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1875 zio_gang_tree_assemble_done(zio_t *zio)
1877 zio_t *gio = zio->io_gang_leader;
1878 zio_gang_node_t *gn = zio->io_private;
1879 blkptr_t *bp = zio->io_bp;
1881 ASSERT(gio == zio_unique_parent(zio));
1882 ASSERT(zio->io_child_count == 0);
1887 if (BP_SHOULD_BYTESWAP(bp))
1888 byteswap_uint64_array(zio->io_data, zio->io_size);
1890 ASSERT(zio->io_data == gn->gn_gbh);
1891 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1892 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1894 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1895 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1896 if (!BP_IS_GANG(gbp))
1898 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1903 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1905 zio_t *gio = pio->io_gang_leader;
1908 ASSERT(BP_IS_GANG(bp) == !!gn);
1909 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1910 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1913 * If you're a gang header, your data is in gn->gn_gbh.
1914 * If you're a gang member, your data is in 'data' and gn == NULL.
1916 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1919 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1921 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1922 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1923 if (BP_IS_HOLE(gbp))
1925 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1926 data = (char *)data + BP_GET_PSIZE(gbp);
1930 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1931 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1938 zio_gang_assemble(zio_t *zio)
1940 blkptr_t *bp = zio->io_bp;
1942 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1943 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1945 zio->io_gang_leader = zio;
1947 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1949 return (ZIO_PIPELINE_CONTINUE);
1953 zio_gang_issue(zio_t *zio)
1955 blkptr_t *bp = zio->io_bp;
1957 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1958 return (ZIO_PIPELINE_STOP);
1960 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1961 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1963 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1964 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1966 zio_gang_tree_free(&zio->io_gang_tree);
1968 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1970 return (ZIO_PIPELINE_CONTINUE);
1974 zio_write_gang_member_ready(zio_t *zio)
1976 zio_t *pio = zio_unique_parent(zio);
1977 zio_t *gio = zio->io_gang_leader;
1978 dva_t *cdva = zio->io_bp->blk_dva;
1979 dva_t *pdva = pio->io_bp->blk_dva;
1982 if (BP_IS_HOLE(zio->io_bp))
1985 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1987 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1988 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1989 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1990 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1991 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1993 mutex_enter(&pio->io_lock);
1994 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1995 ASSERT(DVA_GET_GANG(&pdva[d]));
1996 asize = DVA_GET_ASIZE(&pdva[d]);
1997 asize += DVA_GET_ASIZE(&cdva[d]);
1998 DVA_SET_ASIZE(&pdva[d], asize);
2000 mutex_exit(&pio->io_lock);
2004 zio_write_gang_block(zio_t *pio)
2006 spa_t *spa = pio->io_spa;
2007 blkptr_t *bp = pio->io_bp;
2008 zio_t *gio = pio->io_gang_leader;
2010 zio_gang_node_t *gn, **gnpp;
2011 zio_gbh_phys_t *gbh;
2012 uint64_t txg = pio->io_txg;
2013 uint64_t resid = pio->io_size;
2015 int copies = gio->io_prop.zp_copies;
2016 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2020 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2021 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2022 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2024 pio->io_error = error;
2025 return (ZIO_PIPELINE_CONTINUE);
2029 gnpp = &gio->io_gang_tree;
2031 gnpp = pio->io_private;
2032 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2035 gn = zio_gang_node_alloc(gnpp);
2037 bzero(gbh, SPA_GANGBLOCKSIZE);
2040 * Create the gang header.
2042 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2043 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2046 * Create and nowait the gang children.
2048 for (int g = 0; resid != 0; resid -= lsize, g++) {
2049 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2051 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2053 zp.zp_checksum = gio->io_prop.zp_checksum;
2054 zp.zp_compress = ZIO_COMPRESS_OFF;
2055 zp.zp_type = DMU_OT_NONE;
2057 zp.zp_copies = gio->io_prop.zp_copies;
2058 zp.zp_dedup = B_FALSE;
2059 zp.zp_dedup_verify = B_FALSE;
2060 zp.zp_nopwrite = B_FALSE;
2062 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2063 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2064 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2065 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2066 &pio->io_bookmark));
2070 * Set pio's pipeline to just wait for zio to finish.
2072 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2076 return (ZIO_PIPELINE_CONTINUE);
2080 * The zio_nop_write stage in the pipeline determines if allocating a
2081 * new bp is necessary. The nopwrite feature can handle writes in
2082 * either syncing or open context (i.e. zil writes) and as a result is
2083 * mutually exclusive with dedup.
2085 * By leveraging a cryptographically secure checksum, such as SHA256, we
2086 * can compare the checksums of the new data and the old to determine if
2087 * allocating a new block is required. Note that our requirements for
2088 * cryptographic strength are fairly weak: there can't be any accidental
2089 * hash collisions, but we don't need to be secure against intentional
2090 * (malicious) collisions. To trigger a nopwrite, you have to be able
2091 * to write the file to begin with, and triggering an incorrect (hash
2092 * collision) nopwrite is no worse than simply writing to the file.
2093 * That said, there are no known attacks against the checksum algorithms
2094 * used for nopwrite, assuming that the salt and the checksums
2095 * themselves remain secret.
2098 zio_nop_write(zio_t *zio)
2100 blkptr_t *bp = zio->io_bp;
2101 blkptr_t *bp_orig = &zio->io_bp_orig;
2102 zio_prop_t *zp = &zio->io_prop;
2104 ASSERT(BP_GET_LEVEL(bp) == 0);
2105 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2106 ASSERT(zp->zp_nopwrite);
2107 ASSERT(!zp->zp_dedup);
2108 ASSERT(zio->io_bp_override == NULL);
2109 ASSERT(IO_IS_ALLOCATING(zio));
2112 * Check to see if the original bp and the new bp have matching
2113 * characteristics (i.e. same checksum, compression algorithms, etc).
2114 * If they don't then just continue with the pipeline which will
2115 * allocate a new bp.
2117 if (BP_IS_HOLE(bp_orig) ||
2118 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2119 ZCHECKSUM_FLAG_NOPWRITE) ||
2120 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2121 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2122 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2123 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2124 return (ZIO_PIPELINE_CONTINUE);
2127 * If the checksums match then reset the pipeline so that we
2128 * avoid allocating a new bp and issuing any I/O.
2130 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2131 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2132 ZCHECKSUM_FLAG_NOPWRITE);
2133 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2134 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2135 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2136 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2137 sizeof (uint64_t)) == 0);
2140 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2141 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2144 return (ZIO_PIPELINE_CONTINUE);
2148 * ==========================================================================
2150 * ==========================================================================
2153 zio_ddt_child_read_done(zio_t *zio)
2155 blkptr_t *bp = zio->io_bp;
2156 ddt_entry_t *dde = zio->io_private;
2158 zio_t *pio = zio_unique_parent(zio);
2160 mutex_enter(&pio->io_lock);
2161 ddp = ddt_phys_select(dde, bp);
2162 if (zio->io_error == 0)
2163 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2164 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2165 dde->dde_repair_data = zio->io_data;
2167 zio_buf_free(zio->io_data, zio->io_size);
2168 mutex_exit(&pio->io_lock);
2172 zio_ddt_read_start(zio_t *zio)
2174 blkptr_t *bp = zio->io_bp;
2176 ASSERT(BP_GET_DEDUP(bp));
2177 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2178 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2180 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2181 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2182 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2183 ddt_phys_t *ddp = dde->dde_phys;
2184 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2187 ASSERT(zio->io_vsd == NULL);
2190 if (ddp_self == NULL)
2191 return (ZIO_PIPELINE_CONTINUE);
2193 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2194 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2196 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2198 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2199 zio_buf_alloc(zio->io_size), zio->io_size,
2200 zio_ddt_child_read_done, dde, zio->io_priority,
2201 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2202 &zio->io_bookmark));
2204 return (ZIO_PIPELINE_CONTINUE);
2207 zio_nowait(zio_read(zio, zio->io_spa, bp,
2208 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2209 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2211 return (ZIO_PIPELINE_CONTINUE);
2215 zio_ddt_read_done(zio_t *zio)
2217 blkptr_t *bp = zio->io_bp;
2219 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2220 return (ZIO_PIPELINE_STOP);
2222 ASSERT(BP_GET_DEDUP(bp));
2223 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2224 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2226 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2227 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2228 ddt_entry_t *dde = zio->io_vsd;
2230 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2231 return (ZIO_PIPELINE_CONTINUE);
2234 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2235 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2236 return (ZIO_PIPELINE_STOP);
2238 if (dde->dde_repair_data != NULL) {
2239 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2240 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2242 ddt_repair_done(ddt, dde);
2246 ASSERT(zio->io_vsd == NULL);
2248 return (ZIO_PIPELINE_CONTINUE);
2252 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2254 spa_t *spa = zio->io_spa;
2257 * Note: we compare the original data, not the transformed data,
2258 * because when zio->io_bp is an override bp, we will not have
2259 * pushed the I/O transforms. That's an important optimization
2260 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2262 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2263 zio_t *lio = dde->dde_lead_zio[p];
2266 return (lio->io_orig_size != zio->io_orig_size ||
2267 bcmp(zio->io_orig_data, lio->io_orig_data,
2268 zio->io_orig_size) != 0);
2272 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2273 ddt_phys_t *ddp = &dde->dde_phys[p];
2275 if (ddp->ddp_phys_birth != 0) {
2276 arc_buf_t *abuf = NULL;
2277 arc_flags_t aflags = ARC_FLAG_WAIT;
2278 blkptr_t blk = *zio->io_bp;
2281 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2285 error = arc_read(NULL, spa, &blk,
2286 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2287 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2288 &aflags, &zio->io_bookmark);
2291 if (arc_buf_size(abuf) != zio->io_orig_size ||
2292 bcmp(abuf->b_data, zio->io_orig_data,
2293 zio->io_orig_size) != 0)
2294 error = SET_ERROR(EEXIST);
2295 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2299 return (error != 0);
2307 zio_ddt_child_write_ready(zio_t *zio)
2309 int p = zio->io_prop.zp_copies;
2310 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2311 ddt_entry_t *dde = zio->io_private;
2312 ddt_phys_t *ddp = &dde->dde_phys[p];
2320 ASSERT(dde->dde_lead_zio[p] == zio);
2322 ddt_phys_fill(ddp, zio->io_bp);
2324 while ((pio = zio_walk_parents(zio)) != NULL)
2325 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2331 zio_ddt_child_write_done(zio_t *zio)
2333 int p = zio->io_prop.zp_copies;
2334 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2335 ddt_entry_t *dde = zio->io_private;
2336 ddt_phys_t *ddp = &dde->dde_phys[p];
2340 ASSERT(ddp->ddp_refcnt == 0);
2341 ASSERT(dde->dde_lead_zio[p] == zio);
2342 dde->dde_lead_zio[p] = NULL;
2344 if (zio->io_error == 0) {
2345 while (zio_walk_parents(zio) != NULL)
2346 ddt_phys_addref(ddp);
2348 ddt_phys_clear(ddp);
2355 zio_ddt_ditto_write_done(zio_t *zio)
2357 int p = DDT_PHYS_DITTO;
2358 zio_prop_t *zp = &zio->io_prop;
2359 blkptr_t *bp = zio->io_bp;
2360 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2361 ddt_entry_t *dde = zio->io_private;
2362 ddt_phys_t *ddp = &dde->dde_phys[p];
2363 ddt_key_t *ddk = &dde->dde_key;
2367 ASSERT(ddp->ddp_refcnt == 0);
2368 ASSERT(dde->dde_lead_zio[p] == zio);
2369 dde->dde_lead_zio[p] = NULL;
2371 if (zio->io_error == 0) {
2372 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2373 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2374 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2375 if (ddp->ddp_phys_birth != 0)
2376 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2377 ddt_phys_fill(ddp, bp);
2384 zio_ddt_write(zio_t *zio)
2386 spa_t *spa = zio->io_spa;
2387 blkptr_t *bp = zio->io_bp;
2388 uint64_t txg = zio->io_txg;
2389 zio_prop_t *zp = &zio->io_prop;
2390 int p = zp->zp_copies;
2394 ddt_t *ddt = ddt_select(spa, bp);
2398 ASSERT(BP_GET_DEDUP(bp));
2399 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2400 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2403 dde = ddt_lookup(ddt, bp, B_TRUE);
2404 ddp = &dde->dde_phys[p];
2406 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2408 * If we're using a weak checksum, upgrade to a strong checksum
2409 * and try again. If we're already using a strong checksum,
2410 * we can't resolve it, so just convert to an ordinary write.
2411 * (And automatically e-mail a paper to Nature?)
2413 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2414 ZCHECKSUM_FLAG_DEDUP)) {
2415 zp->zp_checksum = spa_dedup_checksum(spa);
2416 zio_pop_transforms(zio);
2417 zio->io_stage = ZIO_STAGE_OPEN;
2420 zp->zp_dedup = B_FALSE;
2422 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2424 return (ZIO_PIPELINE_CONTINUE);
2427 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2428 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2430 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2431 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2432 zio_prop_t czp = *zp;
2434 czp.zp_copies = ditto_copies;
2437 * If we arrived here with an override bp, we won't have run
2438 * the transform stack, so we won't have the data we need to
2439 * generate a child i/o. So, toss the override bp and restart.
2440 * This is safe, because using the override bp is just an
2441 * optimization; and it's rare, so the cost doesn't matter.
2443 if (zio->io_bp_override) {
2444 zio_pop_transforms(zio);
2445 zio->io_stage = ZIO_STAGE_OPEN;
2446 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2447 zio->io_bp_override = NULL;
2450 return (ZIO_PIPELINE_CONTINUE);
2453 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2454 zio->io_orig_size, &czp, NULL, NULL,
2455 zio_ddt_ditto_write_done, dde, zio->io_priority,
2456 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2458 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2459 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2462 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2463 if (ddp->ddp_phys_birth != 0)
2464 ddt_bp_fill(ddp, bp, txg);
2465 if (dde->dde_lead_zio[p] != NULL)
2466 zio_add_child(zio, dde->dde_lead_zio[p]);
2468 ddt_phys_addref(ddp);
2469 } else if (zio->io_bp_override) {
2470 ASSERT(bp->blk_birth == txg);
2471 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2472 ddt_phys_fill(ddp, bp);
2473 ddt_phys_addref(ddp);
2475 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2476 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2477 zio_ddt_child_write_done, dde, zio->io_priority,
2478 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2480 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2481 dde->dde_lead_zio[p] = cio;
2491 return (ZIO_PIPELINE_CONTINUE);
2494 ddt_entry_t *freedde; /* for debugging */
2497 zio_ddt_free(zio_t *zio)
2499 spa_t *spa = zio->io_spa;
2500 blkptr_t *bp = zio->io_bp;
2501 ddt_t *ddt = ddt_select(spa, bp);
2505 ASSERT(BP_GET_DEDUP(bp));
2506 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2509 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2510 ddp = ddt_phys_select(dde, bp);
2511 ddt_phys_decref(ddp);
2514 return (ZIO_PIPELINE_CONTINUE);
2518 * ==========================================================================
2519 * Allocate and free blocks
2520 * ==========================================================================
2523 zio_dva_allocate(zio_t *zio)
2525 spa_t *spa = zio->io_spa;
2526 metaslab_class_t *mc = spa_normal_class(spa);
2527 blkptr_t *bp = zio->io_bp;
2531 if (zio->io_gang_leader == NULL) {
2532 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2533 zio->io_gang_leader = zio;
2536 ASSERT(BP_IS_HOLE(bp));
2537 ASSERT0(BP_GET_NDVAS(bp));
2538 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2539 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2540 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2543 * The dump device does not support gang blocks so allocation on
2544 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2545 * the "fast" gang feature.
2547 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2548 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2549 METASLAB_GANG_CHILD : 0;
2550 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2551 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2554 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2555 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2557 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2558 return (zio_write_gang_block(zio));
2559 zio->io_error = error;
2562 return (ZIO_PIPELINE_CONTINUE);
2566 zio_dva_free(zio_t *zio)
2568 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2570 return (ZIO_PIPELINE_CONTINUE);
2574 zio_dva_claim(zio_t *zio)
2578 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2580 zio->io_error = error;
2582 return (ZIO_PIPELINE_CONTINUE);
2586 * Undo an allocation. This is used by zio_done() when an I/O fails
2587 * and we want to give back the block we just allocated.
2588 * This handles both normal blocks and gang blocks.
2591 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2593 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2594 ASSERT(zio->io_bp_override == NULL);
2596 if (!BP_IS_HOLE(bp))
2597 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2600 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2601 zio_dva_unallocate(zio, gn->gn_child[g],
2602 &gn->gn_gbh->zg_blkptr[g]);
2608 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2611 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2612 uint64_t size, boolean_t use_slog)
2616 ASSERT(txg > spa_syncing_txg(spa));
2619 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2620 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2621 * when allocating them.
2624 error = metaslab_alloc(spa, spa_log_class(spa), size,
2625 new_bp, 1, txg, old_bp,
2626 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2630 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2631 new_bp, 1, txg, old_bp,
2632 METASLAB_HINTBP_AVOID);
2636 BP_SET_LSIZE(new_bp, size);
2637 BP_SET_PSIZE(new_bp, size);
2638 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2639 BP_SET_CHECKSUM(new_bp,
2640 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2641 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2642 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2643 BP_SET_LEVEL(new_bp, 0);
2644 BP_SET_DEDUP(new_bp, 0);
2645 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2652 * Free an intent log block.
2655 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2657 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2658 ASSERT(!BP_IS_GANG(bp));
2660 zio_free(spa, txg, bp);
2664 * ==========================================================================
2665 * Read, write and delete to physical devices
2666 * ==========================================================================
2671 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2672 * stops after this stage and will resume upon I/O completion.
2673 * However, there are instances where the vdev layer may need to
2674 * continue the pipeline when an I/O was not issued. Since the I/O
2675 * that was sent to the vdev layer might be different than the one
2676 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2677 * force the underlying vdev layers to call either zio_execute() or
2678 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2681 zio_vdev_io_start(zio_t *zio)
2683 vdev_t *vd = zio->io_vd;
2685 spa_t *spa = zio->io_spa;
2688 ASSERT(zio->io_error == 0);
2689 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2692 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2693 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2696 * The mirror_ops handle multiple DVAs in a single BP.
2698 vdev_mirror_ops.vdev_op_io_start(zio);
2699 return (ZIO_PIPELINE_STOP);
2702 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2703 zio->io_priority == ZIO_PRIORITY_NOW) {
2704 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2705 return (ZIO_PIPELINE_CONTINUE);
2709 * We keep track of time-sensitive I/Os so that the scan thread
2710 * can quickly react to certain workloads. In particular, we care
2711 * about non-scrubbing, top-level reads and writes with the following
2713 * - synchronous writes of user data to non-slog devices
2714 * - any reads of user data
2715 * When these conditions are met, adjust the timestamp of spa_last_io
2716 * which allows the scan thread to adjust its workload accordingly.
2718 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2719 vd == vd->vdev_top && !vd->vdev_islog &&
2720 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2721 zio->io_txg != spa_syncing_txg(spa)) {
2722 uint64_t old = spa->spa_last_io;
2723 uint64_t new = ddi_get_lbolt64();
2725 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2728 align = 1ULL << vd->vdev_top->vdev_ashift;
2730 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2731 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2732 P2PHASE(zio->io_size, align) != 0) {
2733 /* Transform logical writes to be a full physical block size. */
2734 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2736 if (zio->io_type == ZIO_TYPE_READ ||
2737 zio->io_type == ZIO_TYPE_WRITE)
2738 abuf = zio_buf_alloc(asize);
2739 ASSERT(vd == vd->vdev_top);
2740 if (zio->io_type == ZIO_TYPE_WRITE) {
2741 bcopy(zio->io_data, abuf, zio->io_size);
2742 bzero(abuf + zio->io_size, asize - zio->io_size);
2744 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2749 * If this is not a physical io, make sure that it is properly aligned
2750 * before proceeding.
2752 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2753 ASSERT0(P2PHASE(zio->io_offset, align));
2754 ASSERT0(P2PHASE(zio->io_size, align));
2757 * For physical writes, we allow 512b aligned writes and assume
2758 * the device will perform a read-modify-write as necessary.
2760 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2761 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2764 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2767 * If this is a repair I/O, and there's no self-healing involved --
2768 * that is, we're just resilvering what we expect to resilver --
2769 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2770 * This prevents spurious resilvering with nested replication.
2771 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2772 * A is out of date, we'll read from C+D, then use the data to
2773 * resilver A+B -- but we don't actually want to resilver B, just A.
2774 * The top-level mirror has no way to know this, so instead we just
2775 * discard unnecessary repairs as we work our way down the vdev tree.
2776 * The same logic applies to any form of nested replication:
2777 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2779 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2780 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2781 zio->io_txg != 0 && /* not a delegated i/o */
2782 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2783 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2784 zio_vdev_io_bypass(zio);
2785 return (ZIO_PIPELINE_CONTINUE);
2788 if (vd->vdev_ops->vdev_op_leaf) {
2789 switch (zio->io_type) {
2791 if (vdev_cache_read(zio))
2792 return (ZIO_PIPELINE_CONTINUE);
2794 case ZIO_TYPE_WRITE:
2796 if ((zio = vdev_queue_io(zio)) == NULL)
2797 return (ZIO_PIPELINE_STOP);
2799 if (!vdev_accessible(vd, zio)) {
2800 zio->io_error = SET_ERROR(ENXIO);
2802 return (ZIO_PIPELINE_STOP);
2807 * Note that we ignore repair writes for TRIM because they can
2808 * conflict with normal writes. This isn't an issue because, by
2809 * definition, we only repair blocks that aren't freed.
2811 if (zio->io_type == ZIO_TYPE_WRITE &&
2812 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2813 !trim_map_write_start(zio))
2814 return (ZIO_PIPELINE_STOP);
2817 vd->vdev_ops->vdev_op_io_start(zio);
2818 return (ZIO_PIPELINE_STOP);
2822 zio_vdev_io_done(zio_t *zio)
2824 vdev_t *vd = zio->io_vd;
2825 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2826 boolean_t unexpected_error = B_FALSE;
2828 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2829 return (ZIO_PIPELINE_STOP);
2831 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2832 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2834 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2835 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2836 zio->io_type == ZIO_TYPE_FREE)) {
2838 if (zio->io_type == ZIO_TYPE_WRITE &&
2839 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2840 trim_map_write_done(zio);
2842 vdev_queue_io_done(zio);
2844 if (zio->io_type == ZIO_TYPE_WRITE)
2845 vdev_cache_write(zio);
2847 if (zio_injection_enabled && zio->io_error == 0)
2848 zio->io_error = zio_handle_device_injection(vd,
2851 if (zio_injection_enabled && zio->io_error == 0)
2852 zio->io_error = zio_handle_label_injection(zio, EIO);
2854 if (zio->io_error) {
2855 if (zio->io_error == ENOTSUP &&
2856 zio->io_type == ZIO_TYPE_FREE) {
2857 /* Not all devices support TRIM. */
2858 } else if (!vdev_accessible(vd, zio)) {
2859 zio->io_error = SET_ERROR(ENXIO);
2861 unexpected_error = B_TRUE;
2866 ops->vdev_op_io_done(zio);
2868 if (unexpected_error)
2869 VERIFY(vdev_probe(vd, zio) == NULL);
2871 return (ZIO_PIPELINE_CONTINUE);
2875 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2876 * disk, and use that to finish the checksum ereport later.
2879 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2880 const void *good_buf)
2882 /* no processing needed */
2883 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2888 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2890 void *buf = zio_buf_alloc(zio->io_size);
2892 bcopy(zio->io_data, buf, zio->io_size);
2894 zcr->zcr_cbinfo = zio->io_size;
2895 zcr->zcr_cbdata = buf;
2896 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2897 zcr->zcr_free = zio_buf_free;
2901 zio_vdev_io_assess(zio_t *zio)
2903 vdev_t *vd = zio->io_vd;
2905 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2906 return (ZIO_PIPELINE_STOP);
2908 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2909 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2911 if (zio->io_vsd != NULL) {
2912 zio->io_vsd_ops->vsd_free(zio);
2916 if (zio_injection_enabled && zio->io_error == 0)
2917 zio->io_error = zio_handle_fault_injection(zio, EIO);
2919 if (zio->io_type == ZIO_TYPE_FREE &&
2920 zio->io_priority != ZIO_PRIORITY_NOW) {
2921 switch (zio->io_error) {
2923 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2924 ZIO_TRIM_STAT_BUMP(success);
2927 ZIO_TRIM_STAT_BUMP(unsupported);
2930 ZIO_TRIM_STAT_BUMP(failed);
2936 * If the I/O failed, determine whether we should attempt to retry it.
2938 * On retry, we cut in line in the issue queue, since we don't want
2939 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2941 if (zio->io_error && vd == NULL &&
2942 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2943 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2944 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2946 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2947 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2948 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2949 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2950 zio_requeue_io_start_cut_in_line);
2951 return (ZIO_PIPELINE_STOP);
2955 * If we got an error on a leaf device, convert it to ENXIO
2956 * if the device is not accessible at all.
2958 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2959 !vdev_accessible(vd, zio))
2960 zio->io_error = SET_ERROR(ENXIO);
2963 * If we can't write to an interior vdev (mirror or RAID-Z),
2964 * set vdev_cant_write so that we stop trying to allocate from it.
2966 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2967 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2968 vd->vdev_cant_write = B_TRUE;
2972 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2974 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2975 zio->io_physdone != NULL) {
2976 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2977 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2978 zio->io_physdone(zio->io_logical);
2981 return (ZIO_PIPELINE_CONTINUE);
2985 zio_vdev_io_reissue(zio_t *zio)
2987 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2988 ASSERT(zio->io_error == 0);
2990 zio->io_stage >>= 1;
2994 zio_vdev_io_redone(zio_t *zio)
2996 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2998 zio->io_stage >>= 1;
3002 zio_vdev_io_bypass(zio_t *zio)
3004 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3005 ASSERT(zio->io_error == 0);
3007 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3008 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3012 * ==========================================================================
3013 * Generate and verify checksums
3014 * ==========================================================================
3017 zio_checksum_generate(zio_t *zio)
3019 blkptr_t *bp = zio->io_bp;
3020 enum zio_checksum checksum;
3024 * This is zio_write_phys().
3025 * We're either generating a label checksum, or none at all.
3027 checksum = zio->io_prop.zp_checksum;
3029 if (checksum == ZIO_CHECKSUM_OFF)
3030 return (ZIO_PIPELINE_CONTINUE);
3032 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3034 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3035 ASSERT(!IO_IS_ALLOCATING(zio));
3036 checksum = ZIO_CHECKSUM_GANG_HEADER;
3038 checksum = BP_GET_CHECKSUM(bp);
3042 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3044 return (ZIO_PIPELINE_CONTINUE);
3048 zio_checksum_verify(zio_t *zio)
3050 zio_bad_cksum_t info;
3051 blkptr_t *bp = zio->io_bp;
3054 ASSERT(zio->io_vd != NULL);
3058 * This is zio_read_phys().
3059 * We're either verifying a label checksum, or nothing at all.
3061 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3062 return (ZIO_PIPELINE_CONTINUE);
3064 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3067 if ((error = zio_checksum_error(zio, &info)) != 0) {
3068 zio->io_error = error;
3069 if (error == ECKSUM &&
3070 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3071 zfs_ereport_start_checksum(zio->io_spa,
3072 zio->io_vd, zio, zio->io_offset,
3073 zio->io_size, NULL, &info);
3077 return (ZIO_PIPELINE_CONTINUE);
3081 * Called by RAID-Z to ensure we don't compute the checksum twice.
3084 zio_checksum_verified(zio_t *zio)
3086 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3090 * ==========================================================================
3091 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3092 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3093 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3094 * indicate errors that are specific to one I/O, and most likely permanent.
3095 * Any other error is presumed to be worse because we weren't expecting it.
3096 * ==========================================================================
3099 zio_worst_error(int e1, int e2)
3101 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3104 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3105 if (e1 == zio_error_rank[r1])
3108 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3109 if (e2 == zio_error_rank[r2])
3112 return (r1 > r2 ? e1 : e2);
3116 * ==========================================================================
3118 * ==========================================================================
3121 zio_ready(zio_t *zio)
3123 blkptr_t *bp = zio->io_bp;
3124 zio_t *pio, *pio_next;
3126 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3127 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3128 return (ZIO_PIPELINE_STOP);
3130 if (zio->io_ready) {
3131 ASSERT(IO_IS_ALLOCATING(zio));
3132 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3133 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3134 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3139 if (bp != NULL && bp != &zio->io_bp_copy)
3140 zio->io_bp_copy = *bp;
3143 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3145 mutex_enter(&zio->io_lock);
3146 zio->io_state[ZIO_WAIT_READY] = 1;
3147 pio = zio_walk_parents(zio);
3148 mutex_exit(&zio->io_lock);
3151 * As we notify zio's parents, new parents could be added.
3152 * New parents go to the head of zio's io_parent_list, however,
3153 * so we will (correctly) not notify them. The remainder of zio's
3154 * io_parent_list, from 'pio_next' onward, cannot change because
3155 * all parents must wait for us to be done before they can be done.
3157 for (; pio != NULL; pio = pio_next) {
3158 pio_next = zio_walk_parents(zio);
3159 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3162 if (zio->io_flags & ZIO_FLAG_NODATA) {
3163 if (BP_IS_GANG(bp)) {
3164 zio->io_flags &= ~ZIO_FLAG_NODATA;
3166 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3167 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3171 if (zio_injection_enabled &&
3172 zio->io_spa->spa_syncing_txg == zio->io_txg)
3173 zio_handle_ignored_writes(zio);
3175 return (ZIO_PIPELINE_CONTINUE);
3179 zio_done(zio_t *zio)
3181 spa_t *spa = zio->io_spa;
3182 zio_t *lio = zio->io_logical;
3183 blkptr_t *bp = zio->io_bp;
3184 vdev_t *vd = zio->io_vd;
3185 uint64_t psize = zio->io_size;
3186 zio_t *pio, *pio_next;
3189 * If our children haven't all completed,
3190 * wait for them and then repeat this pipeline stage.
3192 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3193 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3194 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3195 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3196 return (ZIO_PIPELINE_STOP);
3198 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3199 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3200 ASSERT(zio->io_children[c][w] == 0);
3202 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3203 ASSERT(bp->blk_pad[0] == 0);
3204 ASSERT(bp->blk_pad[1] == 0);
3205 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3206 (bp == zio_unique_parent(zio)->io_bp));
3207 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3208 zio->io_bp_override == NULL &&
3209 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3210 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3211 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3212 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3213 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3215 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3216 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3220 * If there were child vdev/gang/ddt errors, they apply to us now.
3222 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3223 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3224 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3227 * If the I/O on the transformed data was successful, generate any
3228 * checksum reports now while we still have the transformed data.
3230 if (zio->io_error == 0) {
3231 while (zio->io_cksum_report != NULL) {
3232 zio_cksum_report_t *zcr = zio->io_cksum_report;
3233 uint64_t align = zcr->zcr_align;
3234 uint64_t asize = P2ROUNDUP(psize, align);
3235 char *abuf = zio->io_data;
3237 if (asize != psize) {
3238 abuf = zio_buf_alloc(asize);
3239 bcopy(zio->io_data, abuf, psize);
3240 bzero(abuf + psize, asize - psize);
3243 zio->io_cksum_report = zcr->zcr_next;
3244 zcr->zcr_next = NULL;
3245 zcr->zcr_finish(zcr, abuf);
3246 zfs_ereport_free_checksum(zcr);
3249 zio_buf_free(abuf, asize);
3253 zio_pop_transforms(zio); /* note: may set zio->io_error */
3255 vdev_stat_update(zio, psize);
3257 if (zio->io_error) {
3259 * If this I/O is attached to a particular vdev,
3260 * generate an error message describing the I/O failure
3261 * at the block level. We ignore these errors if the
3262 * device is currently unavailable.
3264 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3265 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3267 if ((zio->io_error == EIO || !(zio->io_flags &
3268 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3271 * For logical I/O requests, tell the SPA to log the
3272 * error and generate a logical data ereport.
3274 spa_log_error(spa, zio);
3275 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3280 if (zio->io_error && zio == lio) {
3282 * Determine whether zio should be reexecuted. This will
3283 * propagate all the way to the root via zio_notify_parent().
3285 ASSERT(vd == NULL && bp != NULL);
3286 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3288 if (IO_IS_ALLOCATING(zio) &&
3289 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3290 if (zio->io_error != ENOSPC)
3291 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3293 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3296 if ((zio->io_type == ZIO_TYPE_READ ||
3297 zio->io_type == ZIO_TYPE_FREE) &&
3298 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3299 zio->io_error == ENXIO &&
3300 spa_load_state(spa) == SPA_LOAD_NONE &&
3301 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3302 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3304 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3305 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3308 * Here is a possibly good place to attempt to do
3309 * either combinatorial reconstruction or error correction
3310 * based on checksums. It also might be a good place
3311 * to send out preliminary ereports before we suspend
3317 * If there were logical child errors, they apply to us now.
3318 * We defer this until now to avoid conflating logical child
3319 * errors with errors that happened to the zio itself when
3320 * updating vdev stats and reporting FMA events above.
3322 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3324 if ((zio->io_error || zio->io_reexecute) &&
3325 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3326 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3327 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3329 zio_gang_tree_free(&zio->io_gang_tree);
3332 * Godfather I/Os should never suspend.
3334 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3335 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3336 zio->io_reexecute = 0;
3338 if (zio->io_reexecute) {
3340 * This is a logical I/O that wants to reexecute.
3342 * Reexecute is top-down. When an i/o fails, if it's not
3343 * the root, it simply notifies its parent and sticks around.
3344 * The parent, seeing that it still has children in zio_done(),
3345 * does the same. This percolates all the way up to the root.
3346 * The root i/o will reexecute or suspend the entire tree.
3348 * This approach ensures that zio_reexecute() honors
3349 * all the original i/o dependency relationships, e.g.
3350 * parents not executing until children are ready.
3352 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3354 zio->io_gang_leader = NULL;
3356 mutex_enter(&zio->io_lock);
3357 zio->io_state[ZIO_WAIT_DONE] = 1;
3358 mutex_exit(&zio->io_lock);
3361 * "The Godfather" I/O monitors its children but is
3362 * not a true parent to them. It will track them through
3363 * the pipeline but severs its ties whenever they get into
3364 * trouble (e.g. suspended). This allows "The Godfather"
3365 * I/O to return status without blocking.
3367 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3368 zio_link_t *zl = zio->io_walk_link;
3369 pio_next = zio_walk_parents(zio);
3371 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3372 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3373 zio_remove_child(pio, zio, zl);
3374 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3378 if ((pio = zio_unique_parent(zio)) != NULL) {
3380 * We're not a root i/o, so there's nothing to do
3381 * but notify our parent. Don't propagate errors
3382 * upward since we haven't permanently failed yet.
3384 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3385 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3386 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3387 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3389 * We'd fail again if we reexecuted now, so suspend
3390 * until conditions improve (e.g. device comes online).
3392 zio_suspend(spa, zio);
3395 * Reexecution is potentially a huge amount of work.
3396 * Hand it off to the otherwise-unused claim taskq.
3398 #if defined(illumos) || !defined(_KERNEL)
3399 ASSERT(zio->io_tqent.tqent_next == NULL);
3401 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3403 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3404 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3407 return (ZIO_PIPELINE_STOP);
3410 ASSERT(zio->io_child_count == 0);
3411 ASSERT(zio->io_reexecute == 0);
3412 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3415 * Report any checksum errors, since the I/O is complete.
3417 while (zio->io_cksum_report != NULL) {
3418 zio_cksum_report_t *zcr = zio->io_cksum_report;
3419 zio->io_cksum_report = zcr->zcr_next;
3420 zcr->zcr_next = NULL;
3421 zcr->zcr_finish(zcr, NULL);
3422 zfs_ereport_free_checksum(zcr);
3426 * It is the responsibility of the done callback to ensure that this
3427 * particular zio is no longer discoverable for adoption, and as
3428 * such, cannot acquire any new parents.
3433 mutex_enter(&zio->io_lock);
3434 zio->io_state[ZIO_WAIT_DONE] = 1;
3435 mutex_exit(&zio->io_lock);
3437 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3438 zio_link_t *zl = zio->io_walk_link;
3439 pio_next = zio_walk_parents(zio);
3440 zio_remove_child(pio, zio, zl);
3441 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3444 if (zio->io_waiter != NULL) {
3445 mutex_enter(&zio->io_lock);
3446 zio->io_executor = NULL;
3447 cv_broadcast(&zio->io_cv);
3448 mutex_exit(&zio->io_lock);
3453 return (ZIO_PIPELINE_STOP);
3457 * ==========================================================================
3458 * I/O pipeline definition
3459 * ==========================================================================
3461 static zio_pipe_stage_t *zio_pipeline[] = {
3467 zio_checksum_generate,
3482 zio_checksum_verify,
3490 * Compare two zbookmark_phys_t's to see which we would reach first in a
3491 * pre-order traversal of the object tree.
3493 * This is simple in every case aside from the meta-dnode object. For all other
3494 * objects, we traverse them in order (object 1 before object 2, and so on).
3495 * However, all of these objects are traversed while traversing object 0, since
3496 * the data it points to is the list of objects. Thus, we need to convert to a
3497 * canonical representation so we can compare meta-dnode bookmarks to
3498 * non-meta-dnode bookmarks.
3500 * We do this by calculating "equivalents" for each field of the zbookmark.
3501 * zbookmarks outside of the meta-dnode use their own object and level, and
3502 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3503 * blocks this bookmark refers to) by multiplying their blkid by their span
3504 * (the number of L0 blocks contained within one block at their level).
3505 * zbookmarks inside the meta-dnode calculate their object equivalent
3506 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3507 * level + 1<<31 (any value larger than a level could ever be) for their level.
3508 * This causes them to always compare before a bookmark in their object
3509 * equivalent, compare appropriately to bookmarks in other objects, and to
3510 * compare appropriately to other bookmarks in the meta-dnode.
3513 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3514 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3517 * These variables represent the "equivalent" values for the zbookmark,
3518 * after converting zbookmarks inside the meta dnode to their
3519 * normal-object equivalents.
3521 uint64_t zb1obj, zb2obj;
3522 uint64_t zb1L0, zb2L0;
3523 uint64_t zb1level, zb2level;
3525 if (zb1->zb_object == zb2->zb_object &&
3526 zb1->zb_level == zb2->zb_level &&
3527 zb1->zb_blkid == zb2->zb_blkid)
3531 * BP_SPANB calculates the span in blocks.
3533 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3534 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3536 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3537 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3539 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3541 zb1obj = zb1->zb_object;
3542 zb1level = zb1->zb_level;
3545 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3546 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3548 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3550 zb2obj = zb2->zb_object;
3551 zb2level = zb2->zb_level;
3554 /* Now that we have a canonical representation, do the comparison. */
3555 if (zb1obj != zb2obj)
3556 return (zb1obj < zb2obj ? -1 : 1);
3557 else if (zb1L0 != zb2L0)
3558 return (zb1L0 < zb2L0 ? -1 : 1);
3559 else if (zb1level != zb2level)
3560 return (zb1level > zb2level ? -1 : 1);
3562 * This can (theoretically) happen if the bookmarks have the same object
3563 * and level, but different blkids, if the block sizes are not the same.
3564 * There is presently no way to change the indirect block sizes
3570 * This function checks the following: given that last_block is the place that
3571 * our traversal stopped last time, does that guarantee that we've visited
3572 * every node under subtree_root? Therefore, we can't just use the raw output
3573 * of zbookmark_compare. We have to pass in a modified version of
3574 * subtree_root; by incrementing the block id, and then checking whether
3575 * last_block is before or equal to that, we can tell whether or not having
3576 * visited last_block implies that all of subtree_root's children have been
3580 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3581 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3583 zbookmark_phys_t mod_zb = *subtree_root;
3585 ASSERT(last_block->zb_level == 0);
3587 /* The objset_phys_t isn't before anything. */
3592 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3593 * data block size in sectors, because that variable is only used if
3594 * the bookmark refers to a block in the meta-dnode. Since we don't
3595 * know without examining it what object it refers to, and there's no
3596 * harm in passing in this value in other cases, we always pass it in.
3598 * We pass in 0 for the indirect block size shift because zb2 must be
3599 * level 0. The indirect block size is only used to calculate the span
3600 * of the bookmark, but since the bookmark must be level 0, the span is
3601 * always 1, so the math works out.
3603 * If you make changes to how the zbookmark_compare code works, be sure
3604 * to make sure that this code still works afterwards.
3606 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3607 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,