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);
1225 zio->io_bp_override = NULL;
1229 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1231 * We're rewriting an existing block, which means we're
1232 * working on behalf of spa_sync(). For spa_sync() to
1233 * converge, it must eventually be the case that we don't
1234 * have to allocate new blocks. But compression changes
1235 * the blocksize, which forces a reallocate, and makes
1236 * convergence take longer. Therefore, after the first
1237 * few passes, stop compressing to ensure convergence.
1239 pass = spa_sync_pass(spa);
1241 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1243 ASSERT(!BP_GET_DEDUP(bp));
1245 if (pass >= zfs_sync_pass_dont_compress)
1246 compress = ZIO_COMPRESS_OFF;
1248 /* Make sure someone doesn't change their mind on overwrites */
1249 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1250 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1253 if (compress != ZIO_COMPRESS_OFF) {
1254 void *cbuf = zio_buf_alloc(lsize);
1255 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1256 if (psize == 0 || psize == lsize) {
1257 compress = ZIO_COMPRESS_OFF;
1258 zio_buf_free(cbuf, lsize);
1259 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1260 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1261 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1262 encode_embedded_bp_compressed(bp,
1263 cbuf, compress, lsize, psize);
1264 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1265 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1266 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1267 zio_buf_free(cbuf, lsize);
1268 bp->blk_birth = zio->io_txg;
1269 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1270 ASSERT(spa_feature_is_active(spa,
1271 SPA_FEATURE_EMBEDDED_DATA));
1272 return (ZIO_PIPELINE_CONTINUE);
1275 * Round up compressed size up to the ashift
1276 * of the smallest-ashift device, and zero the tail.
1277 * This ensures that the compressed size of the BP
1278 * (and thus compressratio property) are correct,
1279 * in that we charge for the padding used to fill out
1282 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1283 size_t rounded = (size_t)P2ROUNDUP(psize,
1284 1ULL << spa->spa_min_ashift);
1285 if (rounded >= lsize) {
1286 compress = ZIO_COMPRESS_OFF;
1287 zio_buf_free(cbuf, lsize);
1290 bzero((char *)cbuf + psize, rounded - psize);
1292 zio_push_transform(zio, cbuf,
1293 psize, lsize, NULL);
1299 * The final pass of spa_sync() must be all rewrites, but the first
1300 * few passes offer a trade-off: allocating blocks defers convergence,
1301 * but newly allocated blocks are sequential, so they can be written
1302 * to disk faster. Therefore, we allow the first few passes of
1303 * spa_sync() to allocate new blocks, but force rewrites after that.
1304 * There should only be a handful of blocks after pass 1 in any case.
1306 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1307 BP_GET_PSIZE(bp) == psize &&
1308 pass >= zfs_sync_pass_rewrite) {
1310 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1311 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1312 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1315 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1319 if (zio->io_bp_orig.blk_birth != 0 &&
1320 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1321 BP_SET_LSIZE(bp, lsize);
1322 BP_SET_TYPE(bp, zp->zp_type);
1323 BP_SET_LEVEL(bp, zp->zp_level);
1324 BP_SET_BIRTH(bp, zio->io_txg, 0);
1326 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1328 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1329 BP_SET_LSIZE(bp, lsize);
1330 BP_SET_TYPE(bp, zp->zp_type);
1331 BP_SET_LEVEL(bp, zp->zp_level);
1332 BP_SET_PSIZE(bp, psize);
1333 BP_SET_COMPRESS(bp, compress);
1334 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1335 BP_SET_DEDUP(bp, zp->zp_dedup);
1336 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1338 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1339 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1340 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1342 if (zp->zp_nopwrite) {
1343 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1344 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1345 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1349 return (ZIO_PIPELINE_CONTINUE);
1353 zio_free_bp_init(zio_t *zio)
1355 blkptr_t *bp = zio->io_bp;
1357 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1358 if (BP_GET_DEDUP(bp))
1359 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1362 return (ZIO_PIPELINE_CONTINUE);
1366 * ==========================================================================
1367 * Execute the I/O pipeline
1368 * ==========================================================================
1372 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1374 spa_t *spa = zio->io_spa;
1375 zio_type_t t = zio->io_type;
1376 int flags = (cutinline ? TQ_FRONT : 0);
1378 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1381 * If we're a config writer or a probe, the normal issue and
1382 * interrupt threads may all be blocked waiting for the config lock.
1383 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1385 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1389 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1391 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1395 * If this is a high priority I/O, then use the high priority taskq if
1398 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1399 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1402 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1405 * NB: We are assuming that the zio can only be dispatched
1406 * to a single taskq at a time. It would be a grievous error
1407 * to dispatch the zio to another taskq at the same time.
1409 #if defined(illumos) || !defined(_KERNEL)
1410 ASSERT(zio->io_tqent.tqent_next == NULL);
1412 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1414 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1415 flags, &zio->io_tqent);
1419 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1421 kthread_t *executor = zio->io_executor;
1422 spa_t *spa = zio->io_spa;
1424 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1425 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1427 for (i = 0; i < tqs->stqs_count; i++) {
1428 if (taskq_member(tqs->stqs_taskq[i], executor))
1437 zio_issue_async(zio_t *zio)
1439 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1441 return (ZIO_PIPELINE_STOP);
1445 zio_interrupt(zio_t *zio)
1447 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1451 zio_delay_interrupt(zio_t *zio)
1454 * The timeout_generic() function isn't defined in userspace, so
1455 * rather than trying to implement the function, the zio delay
1456 * functionality has been disabled for userspace builds.
1461 * If io_target_timestamp is zero, then no delay has been registered
1462 * for this IO, thus jump to the end of this function and "skip" the
1463 * delay; issuing it directly to the zio layer.
1465 if (zio->io_target_timestamp != 0) {
1466 hrtime_t now = gethrtime();
1468 if (now >= zio->io_target_timestamp) {
1470 * This IO has already taken longer than the target
1471 * delay to complete, so we don't want to delay it
1472 * any longer; we "miss" the delay and issue it
1473 * directly to the zio layer. This is likely due to
1474 * the target latency being set to a value less than
1475 * the underlying hardware can satisfy (e.g. delay
1476 * set to 1ms, but the disks take 10ms to complete an
1480 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1485 hrtime_t diff = zio->io_target_timestamp - now;
1487 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1488 hrtime_t, now, hrtime_t, diff);
1490 (void) timeout_generic(CALLOUT_NORMAL,
1491 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1498 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1503 * Execute the I/O pipeline until one of the following occurs:
1505 * (1) the I/O completes
1506 * (2) the pipeline stalls waiting for dependent child I/Os
1507 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1508 * (4) the I/O is delegated by vdev-level caching or aggregation
1509 * (5) the I/O is deferred due to vdev-level queueing
1510 * (6) the I/O is handed off to another thread.
1512 * In all cases, the pipeline stops whenever there's no CPU work; it never
1513 * burns a thread in cv_wait().
1515 * There's no locking on io_stage because there's no legitimate way
1516 * for multiple threads to be attempting to process the same I/O.
1518 static zio_pipe_stage_t *zio_pipeline[];
1521 zio_execute(zio_t *zio)
1523 zio->io_executor = curthread;
1525 while (zio->io_stage < ZIO_STAGE_DONE) {
1526 enum zio_stage pipeline = zio->io_pipeline;
1527 enum zio_stage stage = zio->io_stage;
1530 ASSERT(!MUTEX_HELD(&zio->io_lock));
1531 ASSERT(ISP2(stage));
1532 ASSERT(zio->io_stall == NULL);
1536 } while ((stage & pipeline) == 0);
1538 ASSERT(stage <= ZIO_STAGE_DONE);
1541 * If we are in interrupt context and this pipeline stage
1542 * will grab a config lock that is held across I/O,
1543 * or may wait for an I/O that needs an interrupt thread
1544 * to complete, issue async to avoid deadlock.
1546 * For VDEV_IO_START, we cut in line so that the io will
1547 * be sent to disk promptly.
1549 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1550 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1551 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1552 zio_requeue_io_start_cut_in_line : B_FALSE;
1553 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1557 zio->io_stage = stage;
1558 rv = zio_pipeline[highbit64(stage) - 1](zio);
1560 if (rv == ZIO_PIPELINE_STOP)
1563 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1568 * ==========================================================================
1569 * Initiate I/O, either sync or async
1570 * ==========================================================================
1573 zio_wait(zio_t *zio)
1577 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1578 ASSERT(zio->io_executor == NULL);
1580 zio->io_waiter = curthread;
1584 mutex_enter(&zio->io_lock);
1585 while (zio->io_executor != NULL)
1586 cv_wait(&zio->io_cv, &zio->io_lock);
1587 mutex_exit(&zio->io_lock);
1589 error = zio->io_error;
1596 zio_nowait(zio_t *zio)
1598 ASSERT(zio->io_executor == NULL);
1600 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1601 zio_unique_parent(zio) == NULL) {
1603 * This is a logical async I/O with no parent to wait for it.
1604 * We add it to the spa_async_root_zio "Godfather" I/O which
1605 * will ensure they complete prior to unloading the pool.
1607 spa_t *spa = zio->io_spa;
1609 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1616 * ==========================================================================
1617 * Reexecute or suspend/resume failed I/O
1618 * ==========================================================================
1622 zio_reexecute(zio_t *pio)
1624 zio_t *cio, *cio_next;
1626 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1627 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1628 ASSERT(pio->io_gang_leader == NULL);
1629 ASSERT(pio->io_gang_tree == NULL);
1631 pio->io_flags = pio->io_orig_flags;
1632 pio->io_stage = pio->io_orig_stage;
1633 pio->io_pipeline = pio->io_orig_pipeline;
1634 pio->io_reexecute = 0;
1635 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1637 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1638 pio->io_state[w] = 0;
1639 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1640 pio->io_child_error[c] = 0;
1642 if (IO_IS_ALLOCATING(pio))
1643 BP_ZERO(pio->io_bp);
1646 * As we reexecute pio's children, new children could be created.
1647 * New children go to the head of pio's io_child_list, however,
1648 * so we will (correctly) not reexecute them. The key is that
1649 * the remainder of pio's io_child_list, from 'cio_next' onward,
1650 * cannot be affected by any side effects of reexecuting 'cio'.
1652 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1653 cio_next = zio_walk_children(pio);
1654 mutex_enter(&pio->io_lock);
1655 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1656 pio->io_children[cio->io_child_type][w]++;
1657 mutex_exit(&pio->io_lock);
1662 * Now that all children have been reexecuted, execute the parent.
1663 * We don't reexecute "The Godfather" I/O here as it's the
1664 * responsibility of the caller to wait on him.
1666 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1671 zio_suspend(spa_t *spa, zio_t *zio)
1673 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1674 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1675 "failure and the failure mode property for this pool "
1676 "is set to panic.", spa_name(spa));
1678 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1680 mutex_enter(&spa->spa_suspend_lock);
1682 if (spa->spa_suspend_zio_root == NULL)
1683 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1684 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1685 ZIO_FLAG_GODFATHER);
1687 spa->spa_suspended = B_TRUE;
1690 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1691 ASSERT(zio != spa->spa_suspend_zio_root);
1692 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1693 ASSERT(zio_unique_parent(zio) == NULL);
1694 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1695 zio_add_child(spa->spa_suspend_zio_root, zio);
1698 mutex_exit(&spa->spa_suspend_lock);
1702 zio_resume(spa_t *spa)
1707 * Reexecute all previously suspended i/o.
1709 mutex_enter(&spa->spa_suspend_lock);
1710 spa->spa_suspended = B_FALSE;
1711 cv_broadcast(&spa->spa_suspend_cv);
1712 pio = spa->spa_suspend_zio_root;
1713 spa->spa_suspend_zio_root = NULL;
1714 mutex_exit(&spa->spa_suspend_lock);
1720 return (zio_wait(pio));
1724 zio_resume_wait(spa_t *spa)
1726 mutex_enter(&spa->spa_suspend_lock);
1727 while (spa_suspended(spa))
1728 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1729 mutex_exit(&spa->spa_suspend_lock);
1733 * ==========================================================================
1736 * A gang block is a collection of small blocks that looks to the DMU
1737 * like one large block. When zio_dva_allocate() cannot find a block
1738 * of the requested size, due to either severe fragmentation or the pool
1739 * being nearly full, it calls zio_write_gang_block() to construct the
1740 * block from smaller fragments.
1742 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1743 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1744 * an indirect block: it's an array of block pointers. It consumes
1745 * only one sector and hence is allocatable regardless of fragmentation.
1746 * The gang header's bps point to its gang members, which hold the data.
1748 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1749 * as the verifier to ensure uniqueness of the SHA256 checksum.
1750 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1751 * not the gang header. This ensures that data block signatures (needed for
1752 * deduplication) are independent of how the block is physically stored.
1754 * Gang blocks can be nested: a gang member may itself be a gang block.
1755 * Thus every gang block is a tree in which root and all interior nodes are
1756 * gang headers, and the leaves are normal blocks that contain user data.
1757 * The root of the gang tree is called the gang leader.
1759 * To perform any operation (read, rewrite, free, claim) on a gang block,
1760 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1761 * in the io_gang_tree field of the original logical i/o by recursively
1762 * reading the gang leader and all gang headers below it. This yields
1763 * an in-core tree containing the contents of every gang header and the
1764 * bps for every constituent of the gang block.
1766 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1767 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1768 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1769 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1770 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1771 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1772 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1773 * of the gang header plus zio_checksum_compute() of the data to update the
1774 * gang header's blk_cksum as described above.
1776 * The two-phase assemble/issue model solves the problem of partial failure --
1777 * what if you'd freed part of a gang block but then couldn't read the
1778 * gang header for another part? Assembling the entire gang tree first
1779 * ensures that all the necessary gang header I/O has succeeded before
1780 * starting the actual work of free, claim, or write. Once the gang tree
1781 * is assembled, free and claim are in-memory operations that cannot fail.
1783 * In the event that a gang write fails, zio_dva_unallocate() walks the
1784 * gang tree to immediately free (i.e. insert back into the space map)
1785 * everything we've allocated. This ensures that we don't get ENOSPC
1786 * errors during repeated suspend/resume cycles due to a flaky device.
1788 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1789 * the gang tree, we won't modify the block, so we can safely defer the free
1790 * (knowing that the block is still intact). If we *can* assemble the gang
1791 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1792 * each constituent bp and we can allocate a new block on the next sync pass.
1794 * In all cases, the gang tree allows complete recovery from partial failure.
1795 * ==========================================================================
1799 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1804 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1805 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1806 &pio->io_bookmark));
1810 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1815 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1816 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1817 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1819 * As we rewrite each gang header, the pipeline will compute
1820 * a new gang block header checksum for it; but no one will
1821 * compute a new data checksum, so we do that here. The one
1822 * exception is the gang leader: the pipeline already computed
1823 * its data checksum because that stage precedes gang assembly.
1824 * (Presently, nothing actually uses interior data checksums;
1825 * this is just good hygiene.)
1827 if (gn != pio->io_gang_leader->io_gang_tree) {
1828 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1829 data, BP_GET_PSIZE(bp));
1832 * If we are here to damage data for testing purposes,
1833 * leave the GBH alone so that we can detect the damage.
1835 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1836 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1838 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1839 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1840 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1848 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1850 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1851 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1852 ZIO_GANG_CHILD_FLAGS(pio)));
1857 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1859 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1860 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1863 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1872 static void zio_gang_tree_assemble_done(zio_t *zio);
1874 static zio_gang_node_t *
1875 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1877 zio_gang_node_t *gn;
1879 ASSERT(*gnpp == NULL);
1881 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1882 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1889 zio_gang_node_free(zio_gang_node_t **gnpp)
1891 zio_gang_node_t *gn = *gnpp;
1893 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1894 ASSERT(gn->gn_child[g] == NULL);
1896 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1897 kmem_free(gn, sizeof (*gn));
1902 zio_gang_tree_free(zio_gang_node_t **gnpp)
1904 zio_gang_node_t *gn = *gnpp;
1909 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1910 zio_gang_tree_free(&gn->gn_child[g]);
1912 zio_gang_node_free(gnpp);
1916 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1918 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1920 ASSERT(gio->io_gang_leader == gio);
1921 ASSERT(BP_IS_GANG(bp));
1923 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1924 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1925 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1929 zio_gang_tree_assemble_done(zio_t *zio)
1931 zio_t *gio = zio->io_gang_leader;
1932 zio_gang_node_t *gn = zio->io_private;
1933 blkptr_t *bp = zio->io_bp;
1935 ASSERT(gio == zio_unique_parent(zio));
1936 ASSERT(zio->io_child_count == 0);
1941 if (BP_SHOULD_BYTESWAP(bp))
1942 byteswap_uint64_array(zio->io_data, zio->io_size);
1944 ASSERT(zio->io_data == gn->gn_gbh);
1945 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1946 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1948 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1949 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1950 if (!BP_IS_GANG(gbp))
1952 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1957 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1959 zio_t *gio = pio->io_gang_leader;
1962 ASSERT(BP_IS_GANG(bp) == !!gn);
1963 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1964 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1967 * If you're a gang header, your data is in gn->gn_gbh.
1968 * If you're a gang member, your data is in 'data' and gn == NULL.
1970 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1973 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1975 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1976 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1977 if (BP_IS_HOLE(gbp))
1979 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1980 data = (char *)data + BP_GET_PSIZE(gbp);
1984 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1985 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1992 zio_gang_assemble(zio_t *zio)
1994 blkptr_t *bp = zio->io_bp;
1996 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1997 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1999 zio->io_gang_leader = zio;
2001 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2003 return (ZIO_PIPELINE_CONTINUE);
2007 zio_gang_issue(zio_t *zio)
2009 blkptr_t *bp = zio->io_bp;
2011 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2012 return (ZIO_PIPELINE_STOP);
2014 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2015 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2017 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2018 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2020 zio_gang_tree_free(&zio->io_gang_tree);
2022 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2024 return (ZIO_PIPELINE_CONTINUE);
2028 zio_write_gang_member_ready(zio_t *zio)
2030 zio_t *pio = zio_unique_parent(zio);
2031 zio_t *gio = zio->io_gang_leader;
2032 dva_t *cdva = zio->io_bp->blk_dva;
2033 dva_t *pdva = pio->io_bp->blk_dva;
2036 if (BP_IS_HOLE(zio->io_bp))
2039 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2041 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2042 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2043 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2044 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2045 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2047 mutex_enter(&pio->io_lock);
2048 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2049 ASSERT(DVA_GET_GANG(&pdva[d]));
2050 asize = DVA_GET_ASIZE(&pdva[d]);
2051 asize += DVA_GET_ASIZE(&cdva[d]);
2052 DVA_SET_ASIZE(&pdva[d], asize);
2054 mutex_exit(&pio->io_lock);
2058 zio_write_gang_block(zio_t *pio)
2060 spa_t *spa = pio->io_spa;
2061 blkptr_t *bp = pio->io_bp;
2062 zio_t *gio = pio->io_gang_leader;
2064 zio_gang_node_t *gn, **gnpp;
2065 zio_gbh_phys_t *gbh;
2066 uint64_t txg = pio->io_txg;
2067 uint64_t resid = pio->io_size;
2069 int copies = gio->io_prop.zp_copies;
2070 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2074 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2075 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2076 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2078 pio->io_error = error;
2079 return (ZIO_PIPELINE_CONTINUE);
2083 gnpp = &gio->io_gang_tree;
2085 gnpp = pio->io_private;
2086 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2089 gn = zio_gang_node_alloc(gnpp);
2091 bzero(gbh, SPA_GANGBLOCKSIZE);
2094 * Create the gang header.
2096 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2097 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2100 * Create and nowait the gang children.
2102 for (int g = 0; resid != 0; resid -= lsize, g++) {
2103 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2105 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2107 zp.zp_checksum = gio->io_prop.zp_checksum;
2108 zp.zp_compress = ZIO_COMPRESS_OFF;
2109 zp.zp_type = DMU_OT_NONE;
2111 zp.zp_copies = gio->io_prop.zp_copies;
2112 zp.zp_dedup = B_FALSE;
2113 zp.zp_dedup_verify = B_FALSE;
2114 zp.zp_nopwrite = B_FALSE;
2116 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2117 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2118 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2119 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2120 &pio->io_bookmark));
2124 * Set pio's pipeline to just wait for zio to finish.
2126 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2130 return (ZIO_PIPELINE_CONTINUE);
2134 * The zio_nop_write stage in the pipeline determines if allocating a
2135 * new bp is necessary. The nopwrite feature can handle writes in
2136 * either syncing or open context (i.e. zil writes) and as a result is
2137 * mutually exclusive with dedup.
2139 * By leveraging a cryptographically secure checksum, such as SHA256, we
2140 * can compare the checksums of the new data and the old to determine if
2141 * allocating a new block is required. Note that our requirements for
2142 * cryptographic strength are fairly weak: there can't be any accidental
2143 * hash collisions, but we don't need to be secure against intentional
2144 * (malicious) collisions. To trigger a nopwrite, you have to be able
2145 * to write the file to begin with, and triggering an incorrect (hash
2146 * collision) nopwrite is no worse than simply writing to the file.
2147 * That said, there are no known attacks against the checksum algorithms
2148 * used for nopwrite, assuming that the salt and the checksums
2149 * themselves remain secret.
2152 zio_nop_write(zio_t *zio)
2154 blkptr_t *bp = zio->io_bp;
2155 blkptr_t *bp_orig = &zio->io_bp_orig;
2156 zio_prop_t *zp = &zio->io_prop;
2158 ASSERT(BP_GET_LEVEL(bp) == 0);
2159 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2160 ASSERT(zp->zp_nopwrite);
2161 ASSERT(!zp->zp_dedup);
2162 ASSERT(zio->io_bp_override == NULL);
2163 ASSERT(IO_IS_ALLOCATING(zio));
2166 * Check to see if the original bp and the new bp have matching
2167 * characteristics (i.e. same checksum, compression algorithms, etc).
2168 * If they don't then just continue with the pipeline which will
2169 * allocate a new bp.
2171 if (BP_IS_HOLE(bp_orig) ||
2172 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2173 ZCHECKSUM_FLAG_NOPWRITE) ||
2174 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2175 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2176 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2177 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2178 return (ZIO_PIPELINE_CONTINUE);
2181 * If the checksums match then reset the pipeline so that we
2182 * avoid allocating a new bp and issuing any I/O.
2184 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2185 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2186 ZCHECKSUM_FLAG_NOPWRITE);
2187 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2188 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2189 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2190 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2191 sizeof (uint64_t)) == 0);
2194 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2195 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2198 return (ZIO_PIPELINE_CONTINUE);
2202 * ==========================================================================
2204 * ==========================================================================
2207 zio_ddt_child_read_done(zio_t *zio)
2209 blkptr_t *bp = zio->io_bp;
2210 ddt_entry_t *dde = zio->io_private;
2212 zio_t *pio = zio_unique_parent(zio);
2214 mutex_enter(&pio->io_lock);
2215 ddp = ddt_phys_select(dde, bp);
2216 if (zio->io_error == 0)
2217 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2218 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2219 dde->dde_repair_data = zio->io_data;
2221 zio_buf_free(zio->io_data, zio->io_size);
2222 mutex_exit(&pio->io_lock);
2226 zio_ddt_read_start(zio_t *zio)
2228 blkptr_t *bp = zio->io_bp;
2230 ASSERT(BP_GET_DEDUP(bp));
2231 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2232 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2234 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2235 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2236 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2237 ddt_phys_t *ddp = dde->dde_phys;
2238 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2241 ASSERT(zio->io_vsd == NULL);
2244 if (ddp_self == NULL)
2245 return (ZIO_PIPELINE_CONTINUE);
2247 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2248 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2250 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2252 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2253 zio_buf_alloc(zio->io_size), zio->io_size,
2254 zio_ddt_child_read_done, dde, zio->io_priority,
2255 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2256 &zio->io_bookmark));
2258 return (ZIO_PIPELINE_CONTINUE);
2261 zio_nowait(zio_read(zio, zio->io_spa, bp,
2262 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2263 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2265 return (ZIO_PIPELINE_CONTINUE);
2269 zio_ddt_read_done(zio_t *zio)
2271 blkptr_t *bp = zio->io_bp;
2273 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2274 return (ZIO_PIPELINE_STOP);
2276 ASSERT(BP_GET_DEDUP(bp));
2277 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2278 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2280 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2281 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2282 ddt_entry_t *dde = zio->io_vsd;
2284 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2285 return (ZIO_PIPELINE_CONTINUE);
2288 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2289 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2290 return (ZIO_PIPELINE_STOP);
2292 if (dde->dde_repair_data != NULL) {
2293 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2294 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2296 ddt_repair_done(ddt, dde);
2300 ASSERT(zio->io_vsd == NULL);
2302 return (ZIO_PIPELINE_CONTINUE);
2306 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2308 spa_t *spa = zio->io_spa;
2311 * Note: we compare the original data, not the transformed data,
2312 * because when zio->io_bp is an override bp, we will not have
2313 * pushed the I/O transforms. That's an important optimization
2314 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2316 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2317 zio_t *lio = dde->dde_lead_zio[p];
2320 return (lio->io_orig_size != zio->io_orig_size ||
2321 bcmp(zio->io_orig_data, lio->io_orig_data,
2322 zio->io_orig_size) != 0);
2326 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2327 ddt_phys_t *ddp = &dde->dde_phys[p];
2329 if (ddp->ddp_phys_birth != 0) {
2330 arc_buf_t *abuf = NULL;
2331 arc_flags_t aflags = ARC_FLAG_WAIT;
2332 blkptr_t blk = *zio->io_bp;
2335 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2339 error = arc_read(NULL, spa, &blk,
2340 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2341 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2342 &aflags, &zio->io_bookmark);
2345 if (arc_buf_size(abuf) != zio->io_orig_size ||
2346 bcmp(abuf->b_data, zio->io_orig_data,
2347 zio->io_orig_size) != 0)
2348 error = SET_ERROR(EEXIST);
2349 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2353 return (error != 0);
2361 zio_ddt_child_write_ready(zio_t *zio)
2363 int p = zio->io_prop.zp_copies;
2364 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2365 ddt_entry_t *dde = zio->io_private;
2366 ddt_phys_t *ddp = &dde->dde_phys[p];
2374 ASSERT(dde->dde_lead_zio[p] == zio);
2376 ddt_phys_fill(ddp, zio->io_bp);
2378 while ((pio = zio_walk_parents(zio)) != NULL)
2379 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2385 zio_ddt_child_write_done(zio_t *zio)
2387 int p = zio->io_prop.zp_copies;
2388 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2389 ddt_entry_t *dde = zio->io_private;
2390 ddt_phys_t *ddp = &dde->dde_phys[p];
2394 ASSERT(ddp->ddp_refcnt == 0);
2395 ASSERT(dde->dde_lead_zio[p] == zio);
2396 dde->dde_lead_zio[p] = NULL;
2398 if (zio->io_error == 0) {
2399 while (zio_walk_parents(zio) != NULL)
2400 ddt_phys_addref(ddp);
2402 ddt_phys_clear(ddp);
2409 zio_ddt_ditto_write_done(zio_t *zio)
2411 int p = DDT_PHYS_DITTO;
2412 zio_prop_t *zp = &zio->io_prop;
2413 blkptr_t *bp = zio->io_bp;
2414 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2415 ddt_entry_t *dde = zio->io_private;
2416 ddt_phys_t *ddp = &dde->dde_phys[p];
2417 ddt_key_t *ddk = &dde->dde_key;
2421 ASSERT(ddp->ddp_refcnt == 0);
2422 ASSERT(dde->dde_lead_zio[p] == zio);
2423 dde->dde_lead_zio[p] = NULL;
2425 if (zio->io_error == 0) {
2426 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2427 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2428 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2429 if (ddp->ddp_phys_birth != 0)
2430 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2431 ddt_phys_fill(ddp, bp);
2438 zio_ddt_write(zio_t *zio)
2440 spa_t *spa = zio->io_spa;
2441 blkptr_t *bp = zio->io_bp;
2442 uint64_t txg = zio->io_txg;
2443 zio_prop_t *zp = &zio->io_prop;
2444 int p = zp->zp_copies;
2448 ddt_t *ddt = ddt_select(spa, bp);
2452 ASSERT(BP_GET_DEDUP(bp));
2453 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2454 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2457 dde = ddt_lookup(ddt, bp, B_TRUE);
2458 ddp = &dde->dde_phys[p];
2460 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2462 * If we're using a weak checksum, upgrade to a strong checksum
2463 * and try again. If we're already using a strong checksum,
2464 * we can't resolve it, so just convert to an ordinary write.
2465 * (And automatically e-mail a paper to Nature?)
2467 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2468 ZCHECKSUM_FLAG_DEDUP)) {
2469 zp->zp_checksum = spa_dedup_checksum(spa);
2470 zio_pop_transforms(zio);
2471 zio->io_stage = ZIO_STAGE_OPEN;
2474 zp->zp_dedup = B_FALSE;
2476 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2478 return (ZIO_PIPELINE_CONTINUE);
2481 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2482 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2484 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2485 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2486 zio_prop_t czp = *zp;
2488 czp.zp_copies = ditto_copies;
2491 * If we arrived here with an override bp, we won't have run
2492 * the transform stack, so we won't have the data we need to
2493 * generate a child i/o. So, toss the override bp and restart.
2494 * This is safe, because using the override bp is just an
2495 * optimization; and it's rare, so the cost doesn't matter.
2497 if (zio->io_bp_override) {
2498 zio_pop_transforms(zio);
2499 zio->io_stage = ZIO_STAGE_OPEN;
2500 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2501 zio->io_bp_override = NULL;
2504 return (ZIO_PIPELINE_CONTINUE);
2507 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2508 zio->io_orig_size, &czp, NULL, NULL,
2509 zio_ddt_ditto_write_done, dde, zio->io_priority,
2510 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2512 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2513 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2516 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2517 if (ddp->ddp_phys_birth != 0)
2518 ddt_bp_fill(ddp, bp, txg);
2519 if (dde->dde_lead_zio[p] != NULL)
2520 zio_add_child(zio, dde->dde_lead_zio[p]);
2522 ddt_phys_addref(ddp);
2523 } else if (zio->io_bp_override) {
2524 ASSERT(bp->blk_birth == txg);
2525 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2526 ddt_phys_fill(ddp, bp);
2527 ddt_phys_addref(ddp);
2529 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2530 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2531 zio_ddt_child_write_done, dde, zio->io_priority,
2532 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2534 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2535 dde->dde_lead_zio[p] = cio;
2545 return (ZIO_PIPELINE_CONTINUE);
2548 ddt_entry_t *freedde; /* for debugging */
2551 zio_ddt_free(zio_t *zio)
2553 spa_t *spa = zio->io_spa;
2554 blkptr_t *bp = zio->io_bp;
2555 ddt_t *ddt = ddt_select(spa, bp);
2559 ASSERT(BP_GET_DEDUP(bp));
2560 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2563 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2564 ddp = ddt_phys_select(dde, bp);
2565 ddt_phys_decref(ddp);
2568 return (ZIO_PIPELINE_CONTINUE);
2572 * ==========================================================================
2573 * Allocate and free blocks
2574 * ==========================================================================
2577 zio_dva_allocate(zio_t *zio)
2579 spa_t *spa = zio->io_spa;
2580 metaslab_class_t *mc = spa_normal_class(spa);
2581 blkptr_t *bp = zio->io_bp;
2585 if (zio->io_gang_leader == NULL) {
2586 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2587 zio->io_gang_leader = zio;
2590 ASSERT(BP_IS_HOLE(bp));
2591 ASSERT0(BP_GET_NDVAS(bp));
2592 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2593 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2594 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2597 * The dump device does not support gang blocks so allocation on
2598 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2599 * the "fast" gang feature.
2601 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2602 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2603 METASLAB_GANG_CHILD : 0;
2604 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2605 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2608 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2609 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2611 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2612 return (zio_write_gang_block(zio));
2613 zio->io_error = error;
2616 return (ZIO_PIPELINE_CONTINUE);
2620 zio_dva_free(zio_t *zio)
2622 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2624 return (ZIO_PIPELINE_CONTINUE);
2628 zio_dva_claim(zio_t *zio)
2632 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2634 zio->io_error = error;
2636 return (ZIO_PIPELINE_CONTINUE);
2640 * Undo an allocation. This is used by zio_done() when an I/O fails
2641 * and we want to give back the block we just allocated.
2642 * This handles both normal blocks and gang blocks.
2645 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2647 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2648 ASSERT(zio->io_bp_override == NULL);
2650 if (!BP_IS_HOLE(bp))
2651 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2654 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2655 zio_dva_unallocate(zio, gn->gn_child[g],
2656 &gn->gn_gbh->zg_blkptr[g]);
2662 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2665 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2666 uint64_t size, boolean_t use_slog)
2670 ASSERT(txg > spa_syncing_txg(spa));
2673 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2674 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2675 * when allocating them.
2678 error = metaslab_alloc(spa, spa_log_class(spa), size,
2679 new_bp, 1, txg, old_bp,
2680 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2684 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2685 new_bp, 1, txg, old_bp,
2686 METASLAB_HINTBP_AVOID);
2690 BP_SET_LSIZE(new_bp, size);
2691 BP_SET_PSIZE(new_bp, size);
2692 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2693 BP_SET_CHECKSUM(new_bp,
2694 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2695 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2696 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2697 BP_SET_LEVEL(new_bp, 0);
2698 BP_SET_DEDUP(new_bp, 0);
2699 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2706 * Free an intent log block.
2709 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2711 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2712 ASSERT(!BP_IS_GANG(bp));
2714 zio_free(spa, txg, bp);
2718 * ==========================================================================
2719 * Read, write and delete to physical devices
2720 * ==========================================================================
2725 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2726 * stops after this stage and will resume upon I/O completion.
2727 * However, there are instances where the vdev layer may need to
2728 * continue the pipeline when an I/O was not issued. Since the I/O
2729 * that was sent to the vdev layer might be different than the one
2730 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2731 * force the underlying vdev layers to call either zio_execute() or
2732 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2735 zio_vdev_io_start(zio_t *zio)
2737 vdev_t *vd = zio->io_vd;
2739 spa_t *spa = zio->io_spa;
2742 ASSERT(zio->io_error == 0);
2743 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2746 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2747 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2750 * The mirror_ops handle multiple DVAs in a single BP.
2752 vdev_mirror_ops.vdev_op_io_start(zio);
2753 return (ZIO_PIPELINE_STOP);
2756 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2757 zio->io_priority == ZIO_PRIORITY_NOW) {
2758 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2759 return (ZIO_PIPELINE_CONTINUE);
2763 * We keep track of time-sensitive I/Os so that the scan thread
2764 * can quickly react to certain workloads. In particular, we care
2765 * about non-scrubbing, top-level reads and writes with the following
2767 * - synchronous writes of user data to non-slog devices
2768 * - any reads of user data
2769 * When these conditions are met, adjust the timestamp of spa_last_io
2770 * which allows the scan thread to adjust its workload accordingly.
2772 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2773 vd == vd->vdev_top && !vd->vdev_islog &&
2774 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2775 zio->io_txg != spa_syncing_txg(spa)) {
2776 uint64_t old = spa->spa_last_io;
2777 uint64_t new = ddi_get_lbolt64();
2779 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2782 align = 1ULL << vd->vdev_top->vdev_ashift;
2784 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2785 P2PHASE(zio->io_size, align) != 0) {
2786 /* Transform logical writes to be a full physical block size. */
2787 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2789 if (zio->io_type == ZIO_TYPE_READ ||
2790 zio->io_type == ZIO_TYPE_WRITE)
2791 abuf = zio_buf_alloc(asize);
2792 ASSERT(vd == vd->vdev_top);
2793 if (zio->io_type == ZIO_TYPE_WRITE) {
2794 bcopy(zio->io_data, abuf, zio->io_size);
2795 bzero(abuf + zio->io_size, asize - zio->io_size);
2797 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2802 * If this is not a physical io, make sure that it is properly aligned
2803 * before proceeding.
2805 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2806 ASSERT0(P2PHASE(zio->io_offset, align));
2807 ASSERT0(P2PHASE(zio->io_size, align));
2810 * For physical writes, we allow 512b aligned writes and assume
2811 * the device will perform a read-modify-write as necessary.
2813 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2814 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2817 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2820 * If this is a repair I/O, and there's no self-healing involved --
2821 * that is, we're just resilvering what we expect to resilver --
2822 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2823 * This prevents spurious resilvering with nested replication.
2824 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2825 * A is out of date, we'll read from C+D, then use the data to
2826 * resilver A+B -- but we don't actually want to resilver B, just A.
2827 * The top-level mirror has no way to know this, so instead we just
2828 * discard unnecessary repairs as we work our way down the vdev tree.
2829 * The same logic applies to any form of nested replication:
2830 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2832 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2833 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2834 zio->io_txg != 0 && /* not a delegated i/o */
2835 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2836 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2837 zio_vdev_io_bypass(zio);
2838 return (ZIO_PIPELINE_CONTINUE);
2841 if (vd->vdev_ops->vdev_op_leaf) {
2842 switch (zio->io_type) {
2844 if (vdev_cache_read(zio))
2845 return (ZIO_PIPELINE_CONTINUE);
2847 case ZIO_TYPE_WRITE:
2849 if ((zio = vdev_queue_io(zio)) == NULL)
2850 return (ZIO_PIPELINE_STOP);
2852 if (!vdev_accessible(vd, zio)) {
2853 zio->io_error = SET_ERROR(ENXIO);
2855 return (ZIO_PIPELINE_STOP);
2860 * Note that we ignore repair writes for TRIM because they can
2861 * conflict with normal writes. This isn't an issue because, by
2862 * definition, we only repair blocks that aren't freed.
2864 if (zio->io_type == ZIO_TYPE_WRITE &&
2865 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2866 !trim_map_write_start(zio))
2867 return (ZIO_PIPELINE_STOP);
2870 vd->vdev_ops->vdev_op_io_start(zio);
2871 return (ZIO_PIPELINE_STOP);
2875 zio_vdev_io_done(zio_t *zio)
2877 vdev_t *vd = zio->io_vd;
2878 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2879 boolean_t unexpected_error = B_FALSE;
2881 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2882 return (ZIO_PIPELINE_STOP);
2884 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2885 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2887 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2888 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2889 zio->io_type == ZIO_TYPE_FREE)) {
2891 if (zio->io_type == ZIO_TYPE_WRITE &&
2892 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2893 trim_map_write_done(zio);
2895 vdev_queue_io_done(zio);
2897 if (zio->io_type == ZIO_TYPE_WRITE)
2898 vdev_cache_write(zio);
2900 if (zio_injection_enabled && zio->io_error == 0)
2901 zio->io_error = zio_handle_device_injection(vd,
2904 if (zio_injection_enabled && zio->io_error == 0)
2905 zio->io_error = zio_handle_label_injection(zio, EIO);
2907 if (zio->io_error) {
2908 if (zio->io_error == ENOTSUP &&
2909 zio->io_type == ZIO_TYPE_FREE) {
2910 /* Not all devices support TRIM. */
2911 } else if (!vdev_accessible(vd, zio)) {
2912 zio->io_error = SET_ERROR(ENXIO);
2914 unexpected_error = B_TRUE;
2919 ops->vdev_op_io_done(zio);
2921 if (unexpected_error)
2922 VERIFY(vdev_probe(vd, zio) == NULL);
2924 return (ZIO_PIPELINE_CONTINUE);
2928 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2929 * disk, and use that to finish the checksum ereport later.
2932 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2933 const void *good_buf)
2935 /* no processing needed */
2936 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2941 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2943 void *buf = zio_buf_alloc(zio->io_size);
2945 bcopy(zio->io_data, buf, zio->io_size);
2947 zcr->zcr_cbinfo = zio->io_size;
2948 zcr->zcr_cbdata = buf;
2949 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2950 zcr->zcr_free = zio_buf_free;
2954 zio_vdev_io_assess(zio_t *zio)
2956 vdev_t *vd = zio->io_vd;
2958 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2959 return (ZIO_PIPELINE_STOP);
2961 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2962 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2964 if (zio->io_vsd != NULL) {
2965 zio->io_vsd_ops->vsd_free(zio);
2969 if (zio_injection_enabled && zio->io_error == 0)
2970 zio->io_error = zio_handle_fault_injection(zio, EIO);
2972 if (zio->io_type == ZIO_TYPE_FREE &&
2973 zio->io_priority != ZIO_PRIORITY_NOW) {
2974 switch (zio->io_error) {
2976 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2977 ZIO_TRIM_STAT_BUMP(success);
2980 ZIO_TRIM_STAT_BUMP(unsupported);
2983 ZIO_TRIM_STAT_BUMP(failed);
2989 * If the I/O failed, determine whether we should attempt to retry it.
2991 * On retry, we cut in line in the issue queue, since we don't want
2992 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2994 if (zio->io_error && vd == NULL &&
2995 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2996 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2997 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2999 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3000 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3001 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3002 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3003 zio_requeue_io_start_cut_in_line);
3004 return (ZIO_PIPELINE_STOP);
3008 * If we got an error on a leaf device, convert it to ENXIO
3009 * if the device is not accessible at all.
3011 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3012 !vdev_accessible(vd, zio))
3013 zio->io_error = SET_ERROR(ENXIO);
3016 * If we can't write to an interior vdev (mirror or RAID-Z),
3017 * set vdev_cant_write so that we stop trying to allocate from it.
3019 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3020 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3021 vd->vdev_cant_write = B_TRUE;
3025 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3027 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3028 zio->io_physdone != NULL) {
3029 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3030 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3031 zio->io_physdone(zio->io_logical);
3034 return (ZIO_PIPELINE_CONTINUE);
3038 zio_vdev_io_reissue(zio_t *zio)
3040 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3041 ASSERT(zio->io_error == 0);
3043 zio->io_stage >>= 1;
3047 zio_vdev_io_redone(zio_t *zio)
3049 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3051 zio->io_stage >>= 1;
3055 zio_vdev_io_bypass(zio_t *zio)
3057 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3058 ASSERT(zio->io_error == 0);
3060 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3061 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3065 * ==========================================================================
3066 * Generate and verify checksums
3067 * ==========================================================================
3070 zio_checksum_generate(zio_t *zio)
3072 blkptr_t *bp = zio->io_bp;
3073 enum zio_checksum checksum;
3077 * This is zio_write_phys().
3078 * We're either generating a label checksum, or none at all.
3080 checksum = zio->io_prop.zp_checksum;
3082 if (checksum == ZIO_CHECKSUM_OFF)
3083 return (ZIO_PIPELINE_CONTINUE);
3085 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3087 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3088 ASSERT(!IO_IS_ALLOCATING(zio));
3089 checksum = ZIO_CHECKSUM_GANG_HEADER;
3091 checksum = BP_GET_CHECKSUM(bp);
3095 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3097 return (ZIO_PIPELINE_CONTINUE);
3101 zio_checksum_verify(zio_t *zio)
3103 zio_bad_cksum_t info;
3104 blkptr_t *bp = zio->io_bp;
3107 ASSERT(zio->io_vd != NULL);
3111 * This is zio_read_phys().
3112 * We're either verifying a label checksum, or nothing at all.
3114 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3115 return (ZIO_PIPELINE_CONTINUE);
3117 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3120 if ((error = zio_checksum_error(zio, &info)) != 0) {
3121 zio->io_error = error;
3122 if (error == ECKSUM &&
3123 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3124 zfs_ereport_start_checksum(zio->io_spa,
3125 zio->io_vd, zio, zio->io_offset,
3126 zio->io_size, NULL, &info);
3130 return (ZIO_PIPELINE_CONTINUE);
3134 * Called by RAID-Z to ensure we don't compute the checksum twice.
3137 zio_checksum_verified(zio_t *zio)
3139 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3143 * ==========================================================================
3144 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3145 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3146 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3147 * indicate errors that are specific to one I/O, and most likely permanent.
3148 * Any other error is presumed to be worse because we weren't expecting it.
3149 * ==========================================================================
3152 zio_worst_error(int e1, int e2)
3154 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3157 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3158 if (e1 == zio_error_rank[r1])
3161 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3162 if (e2 == zio_error_rank[r2])
3165 return (r1 > r2 ? e1 : e2);
3169 * ==========================================================================
3171 * ==========================================================================
3174 zio_ready(zio_t *zio)
3176 blkptr_t *bp = zio->io_bp;
3177 zio_t *pio, *pio_next;
3179 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3180 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3181 return (ZIO_PIPELINE_STOP);
3183 if (zio->io_ready) {
3184 ASSERT(IO_IS_ALLOCATING(zio));
3185 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3186 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3187 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3192 if (bp != NULL && bp != &zio->io_bp_copy)
3193 zio->io_bp_copy = *bp;
3196 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3198 mutex_enter(&zio->io_lock);
3199 zio->io_state[ZIO_WAIT_READY] = 1;
3200 pio = zio_walk_parents(zio);
3201 mutex_exit(&zio->io_lock);
3204 * As we notify zio's parents, new parents could be added.
3205 * New parents go to the head of zio's io_parent_list, however,
3206 * so we will (correctly) not notify them. The remainder of zio's
3207 * io_parent_list, from 'pio_next' onward, cannot change because
3208 * all parents must wait for us to be done before they can be done.
3210 for (; pio != NULL; pio = pio_next) {
3211 pio_next = zio_walk_parents(zio);
3212 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3215 if (zio->io_flags & ZIO_FLAG_NODATA) {
3216 if (BP_IS_GANG(bp)) {
3217 zio->io_flags &= ~ZIO_FLAG_NODATA;
3219 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3220 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3224 if (zio_injection_enabled &&
3225 zio->io_spa->spa_syncing_txg == zio->io_txg)
3226 zio_handle_ignored_writes(zio);
3228 return (ZIO_PIPELINE_CONTINUE);
3232 zio_done(zio_t *zio)
3234 spa_t *spa = zio->io_spa;
3235 zio_t *lio = zio->io_logical;
3236 blkptr_t *bp = zio->io_bp;
3237 vdev_t *vd = zio->io_vd;
3238 uint64_t psize = zio->io_size;
3239 zio_t *pio, *pio_next;
3242 * If our children haven't all completed,
3243 * wait for them and then repeat this pipeline stage.
3245 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3246 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3247 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3248 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3249 return (ZIO_PIPELINE_STOP);
3251 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3252 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3253 ASSERT(zio->io_children[c][w] == 0);
3255 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3256 ASSERT(bp->blk_pad[0] == 0);
3257 ASSERT(bp->blk_pad[1] == 0);
3258 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3259 (bp == zio_unique_parent(zio)->io_bp));
3260 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3261 zio->io_bp_override == NULL &&
3262 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3263 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3264 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3265 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3266 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3268 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3269 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3273 * If there were child vdev/gang/ddt errors, they apply to us now.
3275 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3276 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3277 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3280 * If the I/O on the transformed data was successful, generate any
3281 * checksum reports now while we still have the transformed data.
3283 if (zio->io_error == 0) {
3284 while (zio->io_cksum_report != NULL) {
3285 zio_cksum_report_t *zcr = zio->io_cksum_report;
3286 uint64_t align = zcr->zcr_align;
3287 uint64_t asize = P2ROUNDUP(psize, align);
3288 char *abuf = zio->io_data;
3290 if (asize != psize) {
3291 abuf = zio_buf_alloc(asize);
3292 bcopy(zio->io_data, abuf, psize);
3293 bzero(abuf + psize, asize - psize);
3296 zio->io_cksum_report = zcr->zcr_next;
3297 zcr->zcr_next = NULL;
3298 zcr->zcr_finish(zcr, abuf);
3299 zfs_ereport_free_checksum(zcr);
3302 zio_buf_free(abuf, asize);
3306 zio_pop_transforms(zio); /* note: may set zio->io_error */
3308 vdev_stat_update(zio, psize);
3310 if (zio->io_error) {
3312 * If this I/O is attached to a particular vdev,
3313 * generate an error message describing the I/O failure
3314 * at the block level. We ignore these errors if the
3315 * device is currently unavailable.
3317 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3318 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3320 if ((zio->io_error == EIO || !(zio->io_flags &
3321 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3324 * For logical I/O requests, tell the SPA to log the
3325 * error and generate a logical data ereport.
3327 spa_log_error(spa, zio);
3328 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3333 if (zio->io_error && zio == lio) {
3335 * Determine whether zio should be reexecuted. This will
3336 * propagate all the way to the root via zio_notify_parent().
3338 ASSERT(vd == NULL && bp != NULL);
3339 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3341 if (IO_IS_ALLOCATING(zio) &&
3342 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3343 if (zio->io_error != ENOSPC)
3344 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3346 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3349 if ((zio->io_type == ZIO_TYPE_READ ||
3350 zio->io_type == ZIO_TYPE_FREE) &&
3351 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3352 zio->io_error == ENXIO &&
3353 spa_load_state(spa) == SPA_LOAD_NONE &&
3354 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3355 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3357 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3358 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3361 * Here is a possibly good place to attempt to do
3362 * either combinatorial reconstruction or error correction
3363 * based on checksums. It also might be a good place
3364 * to send out preliminary ereports before we suspend
3370 * If there were logical child errors, they apply to us now.
3371 * We defer this until now to avoid conflating logical child
3372 * errors with errors that happened to the zio itself when
3373 * updating vdev stats and reporting FMA events above.
3375 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3377 if ((zio->io_error || zio->io_reexecute) &&
3378 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3379 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3380 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3382 zio_gang_tree_free(&zio->io_gang_tree);
3385 * Godfather I/Os should never suspend.
3387 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3388 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3389 zio->io_reexecute = 0;
3391 if (zio->io_reexecute) {
3393 * This is a logical I/O that wants to reexecute.
3395 * Reexecute is top-down. When an i/o fails, if it's not
3396 * the root, it simply notifies its parent and sticks around.
3397 * The parent, seeing that it still has children in zio_done(),
3398 * does the same. This percolates all the way up to the root.
3399 * The root i/o will reexecute or suspend the entire tree.
3401 * This approach ensures that zio_reexecute() honors
3402 * all the original i/o dependency relationships, e.g.
3403 * parents not executing until children are ready.
3405 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3407 zio->io_gang_leader = NULL;
3409 mutex_enter(&zio->io_lock);
3410 zio->io_state[ZIO_WAIT_DONE] = 1;
3411 mutex_exit(&zio->io_lock);
3414 * "The Godfather" I/O monitors its children but is
3415 * not a true parent to them. It will track them through
3416 * the pipeline but severs its ties whenever they get into
3417 * trouble (e.g. suspended). This allows "The Godfather"
3418 * I/O to return status without blocking.
3420 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3421 zio_link_t *zl = zio->io_walk_link;
3422 pio_next = zio_walk_parents(zio);
3424 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3425 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3426 zio_remove_child(pio, zio, zl);
3427 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3431 if ((pio = zio_unique_parent(zio)) != NULL) {
3433 * We're not a root i/o, so there's nothing to do
3434 * but notify our parent. Don't propagate errors
3435 * upward since we haven't permanently failed yet.
3437 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3438 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3439 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3440 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3442 * We'd fail again if we reexecuted now, so suspend
3443 * until conditions improve (e.g. device comes online).
3445 zio_suspend(spa, zio);
3448 * Reexecution is potentially a huge amount of work.
3449 * Hand it off to the otherwise-unused claim taskq.
3451 #if defined(illumos) || !defined(_KERNEL)
3452 ASSERT(zio->io_tqent.tqent_next == NULL);
3454 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3456 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3457 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3460 return (ZIO_PIPELINE_STOP);
3463 ASSERT(zio->io_child_count == 0);
3464 ASSERT(zio->io_reexecute == 0);
3465 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3468 * Report any checksum errors, since the I/O is complete.
3470 while (zio->io_cksum_report != NULL) {
3471 zio_cksum_report_t *zcr = zio->io_cksum_report;
3472 zio->io_cksum_report = zcr->zcr_next;
3473 zcr->zcr_next = NULL;
3474 zcr->zcr_finish(zcr, NULL);
3475 zfs_ereport_free_checksum(zcr);
3479 * It is the responsibility of the done callback to ensure that this
3480 * particular zio is no longer discoverable for adoption, and as
3481 * such, cannot acquire any new parents.
3486 mutex_enter(&zio->io_lock);
3487 zio->io_state[ZIO_WAIT_DONE] = 1;
3488 mutex_exit(&zio->io_lock);
3490 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3491 zio_link_t *zl = zio->io_walk_link;
3492 pio_next = zio_walk_parents(zio);
3493 zio_remove_child(pio, zio, zl);
3494 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3497 if (zio->io_waiter != NULL) {
3498 mutex_enter(&zio->io_lock);
3499 zio->io_executor = NULL;
3500 cv_broadcast(&zio->io_cv);
3501 mutex_exit(&zio->io_lock);
3506 return (ZIO_PIPELINE_STOP);
3510 * ==========================================================================
3511 * I/O pipeline definition
3512 * ==========================================================================
3514 static zio_pipe_stage_t *zio_pipeline[] = {
3520 zio_checksum_generate,
3535 zio_checksum_verify,
3543 * Compare two zbookmark_phys_t's to see which we would reach first in a
3544 * pre-order traversal of the object tree.
3546 * This is simple in every case aside from the meta-dnode object. For all other
3547 * objects, we traverse them in order (object 1 before object 2, and so on).
3548 * However, all of these objects are traversed while traversing object 0, since
3549 * the data it points to is the list of objects. Thus, we need to convert to a
3550 * canonical representation so we can compare meta-dnode bookmarks to
3551 * non-meta-dnode bookmarks.
3553 * We do this by calculating "equivalents" for each field of the zbookmark.
3554 * zbookmarks outside of the meta-dnode use their own object and level, and
3555 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3556 * blocks this bookmark refers to) by multiplying their blkid by their span
3557 * (the number of L0 blocks contained within one block at their level).
3558 * zbookmarks inside the meta-dnode calculate their object equivalent
3559 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3560 * level + 1<<31 (any value larger than a level could ever be) for their level.
3561 * This causes them to always compare before a bookmark in their object
3562 * equivalent, compare appropriately to bookmarks in other objects, and to
3563 * compare appropriately to other bookmarks in the meta-dnode.
3566 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3567 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3570 * These variables represent the "equivalent" values for the zbookmark,
3571 * after converting zbookmarks inside the meta dnode to their
3572 * normal-object equivalents.
3574 uint64_t zb1obj, zb2obj;
3575 uint64_t zb1L0, zb2L0;
3576 uint64_t zb1level, zb2level;
3578 if (zb1->zb_object == zb2->zb_object &&
3579 zb1->zb_level == zb2->zb_level &&
3580 zb1->zb_blkid == zb2->zb_blkid)
3584 * BP_SPANB calculates the span in blocks.
3586 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3587 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3589 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3590 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3592 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3594 zb1obj = zb1->zb_object;
3595 zb1level = zb1->zb_level;
3598 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3599 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3601 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3603 zb2obj = zb2->zb_object;
3604 zb2level = zb2->zb_level;
3607 /* Now that we have a canonical representation, do the comparison. */
3608 if (zb1obj != zb2obj)
3609 return (zb1obj < zb2obj ? -1 : 1);
3610 else if (zb1L0 != zb2L0)
3611 return (zb1L0 < zb2L0 ? -1 : 1);
3612 else if (zb1level != zb2level)
3613 return (zb1level > zb2level ? -1 : 1);
3615 * This can (theoretically) happen if the bookmarks have the same object
3616 * and level, but different blkids, if the block sizes are not the same.
3617 * There is presently no way to change the indirect block sizes
3623 * This function checks the following: given that last_block is the place that
3624 * our traversal stopped last time, does that guarantee that we've visited
3625 * every node under subtree_root? Therefore, we can't just use the raw output
3626 * of zbookmark_compare. We have to pass in a modified version of
3627 * subtree_root; by incrementing the block id, and then checking whether
3628 * last_block is before or equal to that, we can tell whether or not having
3629 * visited last_block implies that all of subtree_root's children have been
3633 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3634 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3636 zbookmark_phys_t mod_zb = *subtree_root;
3638 ASSERT(last_block->zb_level == 0);
3640 /* The objset_phys_t isn't before anything. */
3645 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3646 * data block size in sectors, because that variable is only used if
3647 * the bookmark refers to a block in the meta-dnode. Since we don't
3648 * know without examining it what object it refers to, and there's no
3649 * harm in passing in this value in other cases, we always pass it in.
3651 * We pass in 0 for the indirect block size shift because zb2 must be
3652 * level 0. The indirect block size is only used to calculate the span
3653 * of the bookmark, but since the bookmark must be level 0, the span is
3654 * always 1, so the math works out.
3656 * If you make changes to how the zbookmark_compare code works, be sure
3657 * to make sure that this code still works afterwards.
3659 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3660 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,