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, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/metaslab_impl.h>
47 SYSCTL_DECL(_vfs_zfs);
48 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
49 #if defined(__amd64__)
50 static int zio_use_uma = 1;
52 static int zio_use_uma = 0;
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
55 "Use uma(9) for ZIO allocations");
56 static int zio_exclude_metadata = 0;
57 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
58 "Exclude metadata buffers from dumps as well");
60 zio_trim_stats_t zio_trim_stats = {
61 { "bytes", KSTAT_DATA_UINT64,
62 "Number of bytes successfully TRIMmed" },
63 { "success", KSTAT_DATA_UINT64,
64 "Number of successful TRIM requests" },
65 { "unsupported", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed because TRIM is not supported" },
67 { "failed", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed for reasons other than not supported" },
71 static kstat_t *zio_trim_ksp;
74 * ==========================================================================
75 * I/O type descriptions
76 * ==========================================================================
78 const char *zio_type_name[ZIO_TYPES] = {
79 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
83 boolean_t zio_dva_throttle_enabled = B_TRUE;
84 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
85 &zio_dva_throttle_enabled, 0, "");
88 * ==========================================================================
90 * ==========================================================================
92 kmem_cache_t *zio_cache;
93 kmem_cache_t *zio_link_cache;
94 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
95 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
98 extern vmem_t *zio_alloc_arena;
101 #define ZIO_PIPELINE_CONTINUE 0x100
102 #define ZIO_PIPELINE_STOP 0x101
104 #define BP_SPANB(indblkshift, level) \
105 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
106 #define COMPARE_META_LEVEL 0x80000000ul
108 * The following actions directly effect the spa's sync-to-convergence logic.
109 * The values below define the sync pass when we start performing the action.
110 * Care should be taken when changing these values as they directly impact
111 * spa_sync() performance. Tuning these values may introduce subtle performance
112 * pathologies and should only be done in the context of performance analysis.
113 * These tunables will eventually be removed and replaced with #defines once
114 * enough analysis has been done to determine optimal values.
116 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
117 * regular blocks are not deferred.
119 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
120 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
121 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
122 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
123 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
124 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
125 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
126 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
127 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
130 * An allocating zio is one that either currently has the DVA allocate
131 * stage set or will have it later in its lifetime.
133 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
139 int zio_buf_debug_limit = 16384;
141 int zio_buf_debug_limit = 0;
145 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
151 zio_cache = kmem_cache_create("zio_cache",
152 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
153 zio_link_cache = kmem_cache_create("zio_link_cache",
154 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
159 * For small buffers, we want a cache for each multiple of
160 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
161 * for each quarter-power of 2.
163 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
164 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
167 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
175 * If we are using watchpoints, put each buffer on its own page,
176 * to eliminate the performance overhead of trapping to the
177 * kernel when modifying a non-watched buffer that shares the
178 * page with a watched buffer.
180 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
184 if (size <= 4 * SPA_MINBLOCKSIZE) {
185 align = SPA_MINBLOCKSIZE;
186 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
187 align = MIN(p2 >> 2, PAGESIZE);
192 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
193 zio_buf_cache[c] = kmem_cache_create(name, size,
194 align, NULL, NULL, NULL, NULL, NULL, cflags);
197 * Since zio_data bufs do not appear in crash dumps, we
198 * pass KMC_NOTOUCH so that no allocator metadata is
199 * stored with the buffers.
201 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
202 zio_data_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL,
204 cflags | KMC_NOTOUCH | KMC_NODEBUG);
209 ASSERT(zio_buf_cache[c] != NULL);
210 if (zio_buf_cache[c - 1] == NULL)
211 zio_buf_cache[c - 1] = zio_buf_cache[c];
213 ASSERT(zio_data_buf_cache[c] != NULL);
214 if (zio_data_buf_cache[c - 1] == NULL)
215 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
221 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
223 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
226 if (zio_trim_ksp != NULL) {
227 zio_trim_ksp->ks_data = &zio_trim_stats;
228 kstat_install(zio_trim_ksp);
236 kmem_cache_t *last_cache = NULL;
237 kmem_cache_t *last_data_cache = NULL;
239 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
240 if (zio_buf_cache[c] != last_cache) {
241 last_cache = zio_buf_cache[c];
242 kmem_cache_destroy(zio_buf_cache[c]);
244 zio_buf_cache[c] = NULL;
246 if (zio_data_buf_cache[c] != last_data_cache) {
247 last_data_cache = zio_data_buf_cache[c];
248 kmem_cache_destroy(zio_data_buf_cache[c]);
250 zio_data_buf_cache[c] = NULL;
253 kmem_cache_destroy(zio_link_cache);
254 kmem_cache_destroy(zio_cache);
258 if (zio_trim_ksp != NULL) {
259 kstat_delete(zio_trim_ksp);
265 * ==========================================================================
266 * Allocate and free I/O buffers
267 * ==========================================================================
271 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
272 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
273 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
274 * excess / transient data in-core during a crashdump.
277 zio_buf_alloc(size_t size)
279 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
282 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
285 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
287 return (kmem_alloc(size, KM_SLEEP|flags));
291 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
292 * crashdump if the kernel panics. This exists so that we will limit the amount
293 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
294 * of kernel heap dumped to disk when the kernel panics)
297 zio_data_buf_alloc(size_t size)
299 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
301 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
304 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
306 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
310 zio_buf_free(void *buf, size_t size)
312 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
314 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
317 kmem_cache_free(zio_buf_cache[c], buf);
319 kmem_free(buf, size);
323 zio_data_buf_free(void *buf, size_t size)
325 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
327 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
330 kmem_cache_free(zio_data_buf_cache[c], buf);
332 kmem_free(buf, size);
336 * ==========================================================================
337 * Push and pop I/O transform buffers
338 * ==========================================================================
341 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
342 zio_transform_func_t *transform)
344 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
347 * Ensure that anyone expecting this zio to contain a linear ABD isn't
348 * going to get a nasty surprise when they try to access the data.
351 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
353 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
354 abd_is_linear(data));
357 zt->zt_orig_abd = zio->io_abd;
358 zt->zt_orig_size = zio->io_size;
359 zt->zt_bufsize = bufsize;
360 zt->zt_transform = transform;
362 zt->zt_next = zio->io_transform_stack;
363 zio->io_transform_stack = zt;
370 zio_pop_transforms(zio_t *zio)
374 while ((zt = zio->io_transform_stack) != NULL) {
375 if (zt->zt_transform != NULL)
376 zt->zt_transform(zio,
377 zt->zt_orig_abd, zt->zt_orig_size);
379 if (zt->zt_bufsize != 0)
380 abd_free(zio->io_abd);
382 zio->io_abd = zt->zt_orig_abd;
383 zio->io_size = zt->zt_orig_size;
384 zio->io_transform_stack = zt->zt_next;
386 kmem_free(zt, sizeof (zio_transform_t));
391 * ==========================================================================
392 * I/O transform callbacks for subblocks and decompression
393 * ==========================================================================
396 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
398 ASSERT(zio->io_size > size);
400 if (zio->io_type == ZIO_TYPE_READ)
401 abd_copy(data, zio->io_abd, size);
405 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
407 if (zio->io_error == 0) {
408 void *tmp = abd_borrow_buf(data, size);
409 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
410 zio->io_abd, tmp, zio->io_size, size);
411 abd_return_buf_copy(data, tmp, size);
414 zio->io_error = SET_ERROR(EIO);
419 * ==========================================================================
420 * I/O parent/child relationships and pipeline interlocks
421 * ==========================================================================
424 zio_walk_parents(zio_t *cio, zio_link_t **zl)
426 list_t *pl = &cio->io_parent_list;
428 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
432 ASSERT((*zl)->zl_child == cio);
433 return ((*zl)->zl_parent);
437 zio_walk_children(zio_t *pio, zio_link_t **zl)
439 list_t *cl = &pio->io_child_list;
441 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
445 ASSERT((*zl)->zl_parent == pio);
446 return ((*zl)->zl_child);
450 zio_unique_parent(zio_t *cio)
452 zio_link_t *zl = NULL;
453 zio_t *pio = zio_walk_parents(cio, &zl);
455 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
460 zio_add_child(zio_t *pio, zio_t *cio)
462 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
465 * Logical I/Os can have logical, gang, or vdev children.
466 * Gang I/Os can have gang or vdev children.
467 * Vdev I/Os can only have vdev children.
468 * The following ASSERT captures all of these constraints.
470 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
475 mutex_enter(&cio->io_lock);
476 mutex_enter(&pio->io_lock);
478 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
480 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
481 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
483 list_insert_head(&pio->io_child_list, zl);
484 list_insert_head(&cio->io_parent_list, zl);
486 pio->io_child_count++;
487 cio->io_parent_count++;
489 mutex_exit(&pio->io_lock);
490 mutex_exit(&cio->io_lock);
494 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
496 ASSERT(zl->zl_parent == pio);
497 ASSERT(zl->zl_child == cio);
499 mutex_enter(&cio->io_lock);
500 mutex_enter(&pio->io_lock);
502 list_remove(&pio->io_child_list, zl);
503 list_remove(&cio->io_parent_list, zl);
505 pio->io_child_count--;
506 cio->io_parent_count--;
508 mutex_exit(&pio->io_lock);
509 mutex_exit(&cio->io_lock);
511 kmem_cache_free(zio_link_cache, zl);
515 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
517 boolean_t waiting = B_FALSE;
519 mutex_enter(&zio->io_lock);
520 ASSERT(zio->io_stall == NULL);
521 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
522 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
525 uint64_t *countp = &zio->io_children[c][wait];
528 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
529 zio->io_stall = countp;
534 mutex_exit(&zio->io_lock);
539 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
541 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
542 int *errorp = &pio->io_child_error[zio->io_child_type];
544 mutex_enter(&pio->io_lock);
545 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
546 *errorp = zio_worst_error(*errorp, zio->io_error);
547 pio->io_reexecute |= zio->io_reexecute;
548 ASSERT3U(*countp, >, 0);
552 if (*countp == 0 && pio->io_stall == countp) {
553 zio_taskq_type_t type =
554 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
556 pio->io_stall = NULL;
557 mutex_exit(&pio->io_lock);
559 * Dispatch the parent zio in its own taskq so that
560 * the child can continue to make progress. This also
561 * prevents overflowing the stack when we have deeply nested
562 * parent-child relationships.
564 zio_taskq_dispatch(pio, type, B_FALSE);
566 mutex_exit(&pio->io_lock);
571 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
573 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
574 zio->io_error = zio->io_child_error[c];
578 zio_bookmark_compare(const void *x1, const void *x2)
580 const zio_t *z1 = x1;
581 const zio_t *z2 = x2;
583 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
585 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
588 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
590 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
593 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
595 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
598 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
600 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
612 * ==========================================================================
613 * Create the various types of I/O (read, write, free, etc)
614 * ==========================================================================
617 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
618 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
619 void *private, zio_type_t type, zio_priority_t priority,
620 enum zio_flag flags, vdev_t *vd, uint64_t offset,
621 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
625 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
626 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
627 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
629 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
630 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
631 ASSERT(vd || stage == ZIO_STAGE_OPEN);
633 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
635 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
636 bzero(zio, sizeof (zio_t));
638 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
639 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
641 list_create(&zio->io_parent_list, sizeof (zio_link_t),
642 offsetof(zio_link_t, zl_parent_node));
643 list_create(&zio->io_child_list, sizeof (zio_link_t),
644 offsetof(zio_link_t, zl_child_node));
645 metaslab_trace_init(&zio->io_alloc_list);
648 zio->io_child_type = ZIO_CHILD_VDEV;
649 else if (flags & ZIO_FLAG_GANG_CHILD)
650 zio->io_child_type = ZIO_CHILD_GANG;
651 else if (flags & ZIO_FLAG_DDT_CHILD)
652 zio->io_child_type = ZIO_CHILD_DDT;
654 zio->io_child_type = ZIO_CHILD_LOGICAL;
657 zio->io_bp = (blkptr_t *)bp;
658 zio->io_bp_copy = *bp;
659 zio->io_bp_orig = *bp;
660 if (type != ZIO_TYPE_WRITE ||
661 zio->io_child_type == ZIO_CHILD_DDT)
662 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
663 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
664 zio->io_logical = zio;
665 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
666 pipeline |= ZIO_GANG_STAGES;
672 zio->io_private = private;
674 zio->io_priority = priority;
676 zio->io_offset = offset;
677 zio->io_orig_abd = zio->io_abd = data;
678 zio->io_orig_size = zio->io_size = psize;
679 zio->io_lsize = lsize;
680 zio->io_orig_flags = zio->io_flags = flags;
681 zio->io_orig_stage = zio->io_stage = stage;
682 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
683 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
685 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
686 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
689 zio->io_bookmark = *zb;
692 if (zio->io_logical == NULL)
693 zio->io_logical = pio->io_logical;
694 if (zio->io_child_type == ZIO_CHILD_GANG)
695 zio->io_gang_leader = pio->io_gang_leader;
696 zio_add_child(pio, zio);
703 zio_destroy(zio_t *zio)
705 metaslab_trace_fini(&zio->io_alloc_list);
706 list_destroy(&zio->io_parent_list);
707 list_destroy(&zio->io_child_list);
708 mutex_destroy(&zio->io_lock);
709 cv_destroy(&zio->io_cv);
710 kmem_cache_free(zio_cache, zio);
714 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
715 void *private, enum zio_flag flags)
719 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
720 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
721 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
727 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
729 return (zio_null(NULL, spa, NULL, done, private, flags));
733 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
735 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
736 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
737 bp, (longlong_t)BP_GET_TYPE(bp));
739 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
740 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
741 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
742 bp, (longlong_t)BP_GET_CHECKSUM(bp));
744 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
745 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
746 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
747 bp, (longlong_t)BP_GET_COMPRESS(bp));
749 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
750 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
751 bp, (longlong_t)BP_GET_LSIZE(bp));
753 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
754 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
755 bp, (longlong_t)BP_GET_PSIZE(bp));
758 if (BP_IS_EMBEDDED(bp)) {
759 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
760 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
761 bp, (longlong_t)BPE_GET_ETYPE(bp));
766 * Pool-specific checks.
768 * Note: it would be nice to verify that the blk_birth and
769 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
770 * allows the birth time of log blocks (and dmu_sync()-ed blocks
771 * that are in the log) to be arbitrarily large.
773 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
774 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
775 if (vdevid >= spa->spa_root_vdev->vdev_children) {
776 zfs_panic_recover("blkptr at %p DVA %u has invalid "
778 bp, i, (longlong_t)vdevid);
781 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
783 zfs_panic_recover("blkptr at %p DVA %u has invalid "
785 bp, i, (longlong_t)vdevid);
788 if (vd->vdev_ops == &vdev_hole_ops) {
789 zfs_panic_recover("blkptr at %p DVA %u has hole "
791 bp, i, (longlong_t)vdevid);
794 if (vd->vdev_ops == &vdev_missing_ops) {
796 * "missing" vdevs are valid during import, but we
797 * don't have their detailed info (e.g. asize), so
798 * we can't perform any more checks on them.
802 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
803 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
805 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
806 if (offset + asize > vd->vdev_asize) {
807 zfs_panic_recover("blkptr at %p DVA %u has invalid "
809 bp, i, (longlong_t)offset);
815 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
816 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
817 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
821 zfs_blkptr_verify(spa, bp);
823 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
824 data, size, size, done, private,
825 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
826 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
827 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
833 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
834 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
835 zio_done_func_t *ready, zio_done_func_t *children_ready,
836 zio_done_func_t *physdone, zio_done_func_t *done,
837 void *private, zio_priority_t priority, enum zio_flag flags,
838 const zbookmark_phys_t *zb)
842 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
843 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
844 zp->zp_compress >= ZIO_COMPRESS_OFF &&
845 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
846 DMU_OT_IS_VALID(zp->zp_type) &&
849 zp->zp_copies <= spa_max_replication(spa));
851 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
852 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
853 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
854 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
856 zio->io_ready = ready;
857 zio->io_children_ready = children_ready;
858 zio->io_physdone = physdone;
862 * Data can be NULL if we are going to call zio_write_override() to
863 * provide the already-allocated BP. But we may need the data to
864 * verify a dedup hit (if requested). In this case, don't try to
865 * dedup (just take the already-allocated BP verbatim).
867 if (data == NULL && zio->io_prop.zp_dedup_verify) {
868 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
875 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
876 uint64_t size, zio_done_func_t *done, void *private,
877 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
881 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
882 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
883 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
889 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
891 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
892 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
893 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
894 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
897 * We must reset the io_prop to match the values that existed
898 * when the bp was first written by dmu_sync() keeping in mind
899 * that nopwrite and dedup are mutually exclusive.
901 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
902 zio->io_prop.zp_nopwrite = nopwrite;
903 zio->io_prop.zp_copies = copies;
904 zio->io_bp_override = bp;
908 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
911 zfs_blkptr_verify(spa, bp);
914 * The check for EMBEDDED is a performance optimization. We
915 * process the free here (by ignoring it) rather than
916 * putting it on the list and then processing it in zio_free_sync().
918 if (BP_IS_EMBEDDED(bp))
920 metaslab_check_free(spa, bp);
923 * Frees that are for the currently-syncing txg, are not going to be
924 * deferred, and which will not need to do a read (i.e. not GANG or
925 * DEDUP), can be processed immediately. Otherwise, put them on the
926 * in-memory list for later processing.
928 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
929 txg != spa->spa_syncing_txg ||
930 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
931 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
933 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
934 BP_GET_PSIZE(bp), 0)));
939 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
940 uint64_t size, enum zio_flag flags)
943 enum zio_stage stage = ZIO_FREE_PIPELINE;
945 ASSERT(!BP_IS_HOLE(bp));
946 ASSERT(spa_syncing_txg(spa) == txg);
947 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
949 if (BP_IS_EMBEDDED(bp))
950 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
952 metaslab_check_free(spa, bp);
955 if (zfs_trim_enabled)
956 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
957 ZIO_STAGE_VDEV_IO_ASSESS;
959 * GANG and DEDUP blocks can induce a read (for the gang block header,
960 * or the DDT), so issue them asynchronously so that this thread is
963 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
964 stage |= ZIO_STAGE_ISSUE_ASYNC;
966 flags |= ZIO_FLAG_DONT_QUEUE;
968 zio = zio_create(pio, spa, txg, bp, NULL, size,
969 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
970 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
976 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
977 zio_done_func_t *done, void *private, enum zio_flag flags)
981 zfs_blkptr_verify(spa, bp);
983 if (BP_IS_EMBEDDED(bp))
984 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
987 * A claim is an allocation of a specific block. Claims are needed
988 * to support immediate writes in the intent log. The issue is that
989 * immediate writes contain committed data, but in a txg that was
990 * *not* committed. Upon opening the pool after an unclean shutdown,
991 * the intent log claims all blocks that contain immediate write data
992 * so that the SPA knows they're in use.
994 * All claims *must* be resolved in the first txg -- before the SPA
995 * starts allocating blocks -- so that nothing is allocated twice.
996 * If txg == 0 we just verify that the block is claimable.
998 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
999 ASSERT(txg == spa_first_txg(spa) || txg == 0);
1000 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1002 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1003 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1004 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1005 ASSERT0(zio->io_queued_timestamp);
1011 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1012 uint64_t size, zio_done_func_t *done, void *private,
1013 zio_priority_t priority, enum zio_flag flags)
1018 if (vd->vdev_children == 0) {
1019 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1020 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1021 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1025 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1027 for (c = 0; c < vd->vdev_children; c++)
1028 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1029 offset, size, done, private, priority, flags));
1036 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1037 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1038 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1042 ASSERT(vd->vdev_children == 0);
1043 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1044 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1045 ASSERT3U(offset + size, <=, vd->vdev_psize);
1047 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1048 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1049 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1051 zio->io_prop.zp_checksum = checksum;
1057 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1058 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1059 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1063 ASSERT(vd->vdev_children == 0);
1064 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1065 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1066 ASSERT3U(offset + size, <=, vd->vdev_psize);
1068 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1069 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1070 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1072 zio->io_prop.zp_checksum = checksum;
1074 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1076 * zec checksums are necessarily destructive -- they modify
1077 * the end of the write buffer to hold the verifier/checksum.
1078 * Therefore, we must make a local copy in case the data is
1079 * being written to multiple places in parallel.
1081 abd_t *wbuf = abd_alloc_sametype(data, size);
1082 abd_copy(wbuf, data, size);
1084 zio_push_transform(zio, wbuf, size, size, NULL);
1091 * Create a child I/O to do some work for us.
1094 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1095 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1096 enum zio_flag flags, zio_done_func_t *done, void *private)
1098 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1102 * vdev child I/Os do not propagate their error to the parent.
1103 * Therefore, for correct operation the caller *must* check for
1104 * and handle the error in the child i/o's done callback.
1105 * The only exceptions are i/os that we don't care about
1106 * (OPTIONAL or REPAIR).
1108 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1112 * In the common case, where the parent zio was to a normal vdev,
1113 * the child zio must be to a child vdev of that vdev. Otherwise,
1114 * the child zio must be to a top-level vdev.
1116 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1117 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1119 ASSERT3P(vd, ==, vd->vdev_top);
1122 if (type == ZIO_TYPE_READ && bp != NULL) {
1124 * If we have the bp, then the child should perform the
1125 * checksum and the parent need not. This pushes error
1126 * detection as close to the leaves as possible and
1127 * eliminates redundant checksums in the interior nodes.
1129 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1130 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1133 /* Not all IO types require vdev io done stage e.g. free */
1134 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1135 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1137 if (vd->vdev_ops->vdev_op_leaf) {
1138 ASSERT0(vd->vdev_children);
1139 offset += VDEV_LABEL_START_SIZE;
1142 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1145 * If we've decided to do a repair, the write is not speculative --
1146 * even if the original read was.
1148 if (flags & ZIO_FLAG_IO_REPAIR)
1149 flags &= ~ZIO_FLAG_SPECULATIVE;
1152 * If we're creating a child I/O that is not associated with a
1153 * top-level vdev, then the child zio is not an allocating I/O.
1154 * If this is a retried I/O then we ignore it since we will
1155 * have already processed the original allocating I/O.
1157 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1158 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1159 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1161 ASSERT(mc->mc_alloc_throttle_enabled);
1162 ASSERT(type == ZIO_TYPE_WRITE);
1163 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1164 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1165 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1166 pio->io_child_type == ZIO_CHILD_GANG);
1168 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1171 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1172 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1173 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1174 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1176 zio->io_physdone = pio->io_physdone;
1177 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1178 zio->io_logical->io_phys_children++;
1184 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1185 int type, zio_priority_t priority, enum zio_flag flags,
1186 zio_done_func_t *done, void *private)
1190 ASSERT(vd->vdev_ops->vdev_op_leaf);
1192 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1193 data, size, size, done, private, type, priority,
1194 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1196 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1202 zio_flush(zio_t *zio, vdev_t *vd)
1204 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1205 NULL, NULL, ZIO_PRIORITY_NOW,
1206 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1210 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1213 ASSERT(vd->vdev_ops->vdev_op_leaf);
1215 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1216 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1217 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1218 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1222 zio_shrink(zio_t *zio, uint64_t size)
1224 ASSERT3P(zio->io_executor, ==, NULL);
1225 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1226 ASSERT3U(size, <=, zio->io_size);
1229 * We don't shrink for raidz because of problems with the
1230 * reconstruction when reading back less than the block size.
1231 * Note, BP_IS_RAIDZ() assumes no compression.
1233 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1234 if (!BP_IS_RAIDZ(zio->io_bp)) {
1235 /* we are not doing a raw write */
1236 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1237 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1242 * ==========================================================================
1243 * Prepare to read and write logical blocks
1244 * ==========================================================================
1248 zio_read_bp_init(zio_t *zio)
1250 blkptr_t *bp = zio->io_bp;
1252 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1254 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1255 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1256 !(zio->io_flags & ZIO_FLAG_RAW)) {
1258 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1259 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1260 psize, psize, zio_decompress);
1263 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1264 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1266 int psize = BPE_GET_PSIZE(bp);
1267 void *data = abd_borrow_buf(zio->io_abd, psize);
1268 decode_embedded_bp_compressed(bp, data);
1269 abd_return_buf_copy(zio->io_abd, data, psize);
1271 ASSERT(!BP_IS_EMBEDDED(bp));
1272 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1275 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1276 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1278 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1279 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1281 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1282 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1284 return (ZIO_PIPELINE_CONTINUE);
1288 zio_write_bp_init(zio_t *zio)
1290 if (!IO_IS_ALLOCATING(zio))
1291 return (ZIO_PIPELINE_CONTINUE);
1293 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1295 if (zio->io_bp_override) {
1296 blkptr_t *bp = zio->io_bp;
1297 zio_prop_t *zp = &zio->io_prop;
1299 ASSERT(bp->blk_birth != zio->io_txg);
1300 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1302 *bp = *zio->io_bp_override;
1303 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1305 if (BP_IS_EMBEDDED(bp))
1306 return (ZIO_PIPELINE_CONTINUE);
1309 * If we've been overridden and nopwrite is set then
1310 * set the flag accordingly to indicate that a nopwrite
1311 * has already occurred.
1313 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1314 ASSERT(!zp->zp_dedup);
1315 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1316 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1317 return (ZIO_PIPELINE_CONTINUE);
1320 ASSERT(!zp->zp_nopwrite);
1322 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1323 return (ZIO_PIPELINE_CONTINUE);
1325 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1326 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1328 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1329 BP_SET_DEDUP(bp, 1);
1330 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1331 return (ZIO_PIPELINE_CONTINUE);
1335 * We were unable to handle this as an override bp, treat
1336 * it as a regular write I/O.
1338 zio->io_bp_override = NULL;
1339 *bp = zio->io_bp_orig;
1340 zio->io_pipeline = zio->io_orig_pipeline;
1343 return (ZIO_PIPELINE_CONTINUE);
1347 zio_write_compress(zio_t *zio)
1349 spa_t *spa = zio->io_spa;
1350 zio_prop_t *zp = &zio->io_prop;
1351 enum zio_compress compress = zp->zp_compress;
1352 blkptr_t *bp = zio->io_bp;
1353 uint64_t lsize = zio->io_lsize;
1354 uint64_t psize = zio->io_size;
1357 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1360 * If our children haven't all reached the ready stage,
1361 * wait for them and then repeat this pipeline stage.
1363 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1364 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1365 return (ZIO_PIPELINE_STOP);
1368 if (!IO_IS_ALLOCATING(zio))
1369 return (ZIO_PIPELINE_CONTINUE);
1371 if (zio->io_children_ready != NULL) {
1373 * Now that all our children are ready, run the callback
1374 * associated with this zio in case it wants to modify the
1375 * data to be written.
1377 ASSERT3U(zp->zp_level, >, 0);
1378 zio->io_children_ready(zio);
1381 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1382 ASSERT(zio->io_bp_override == NULL);
1384 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1386 * We're rewriting an existing block, which means we're
1387 * working on behalf of spa_sync(). For spa_sync() to
1388 * converge, it must eventually be the case that we don't
1389 * have to allocate new blocks. But compression changes
1390 * the blocksize, which forces a reallocate, and makes
1391 * convergence take longer. Therefore, after the first
1392 * few passes, stop compressing to ensure convergence.
1394 pass = spa_sync_pass(spa);
1396 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1397 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1398 ASSERT(!BP_GET_DEDUP(bp));
1400 if (pass >= zfs_sync_pass_dont_compress)
1401 compress = ZIO_COMPRESS_OFF;
1403 /* Make sure someone doesn't change their mind on overwrites */
1404 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1405 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1408 /* If it's a compressed write that is not raw, compress the buffer. */
1409 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1410 void *cbuf = zio_buf_alloc(lsize);
1411 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1412 if (psize == 0 || psize == lsize) {
1413 compress = ZIO_COMPRESS_OFF;
1414 zio_buf_free(cbuf, lsize);
1415 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1416 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1417 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1418 encode_embedded_bp_compressed(bp,
1419 cbuf, compress, lsize, psize);
1420 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1421 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1422 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1423 zio_buf_free(cbuf, lsize);
1424 bp->blk_birth = zio->io_txg;
1425 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1426 ASSERT(spa_feature_is_active(spa,
1427 SPA_FEATURE_EMBEDDED_DATA));
1428 return (ZIO_PIPELINE_CONTINUE);
1431 * Round up compressed size up to the ashift
1432 * of the smallest-ashift device, and zero the tail.
1433 * This ensures that the compressed size of the BP
1434 * (and thus compressratio property) are correct,
1435 * in that we charge for the padding used to fill out
1438 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1439 size_t rounded = (size_t)P2ROUNDUP(psize,
1440 1ULL << spa->spa_min_ashift);
1441 if (rounded >= lsize) {
1442 compress = ZIO_COMPRESS_OFF;
1443 zio_buf_free(cbuf, lsize);
1446 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1447 abd_take_ownership_of_buf(cdata, B_TRUE);
1448 abd_zero_off(cdata, psize, rounded - psize);
1450 zio_push_transform(zio, cdata,
1451 psize, lsize, NULL);
1456 * We were unable to handle this as an override bp, treat
1457 * it as a regular write I/O.
1459 zio->io_bp_override = NULL;
1460 *bp = zio->io_bp_orig;
1461 zio->io_pipeline = zio->io_orig_pipeline;
1463 ASSERT3U(psize, !=, 0);
1467 * The final pass of spa_sync() must be all rewrites, but the first
1468 * few passes offer a trade-off: allocating blocks defers convergence,
1469 * but newly allocated blocks are sequential, so they can be written
1470 * to disk faster. Therefore, we allow the first few passes of
1471 * spa_sync() to allocate new blocks, but force rewrites after that.
1472 * There should only be a handful of blocks after pass 1 in any case.
1474 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1475 BP_GET_PSIZE(bp) == psize &&
1476 pass >= zfs_sync_pass_rewrite) {
1478 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1479 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1480 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1483 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1487 if (zio->io_bp_orig.blk_birth != 0 &&
1488 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1489 BP_SET_LSIZE(bp, lsize);
1490 BP_SET_TYPE(bp, zp->zp_type);
1491 BP_SET_LEVEL(bp, zp->zp_level);
1492 BP_SET_BIRTH(bp, zio->io_txg, 0);
1494 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1496 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1497 BP_SET_LSIZE(bp, lsize);
1498 BP_SET_TYPE(bp, zp->zp_type);
1499 BP_SET_LEVEL(bp, zp->zp_level);
1500 BP_SET_PSIZE(bp, psize);
1501 BP_SET_COMPRESS(bp, compress);
1502 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1503 BP_SET_DEDUP(bp, zp->zp_dedup);
1504 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1506 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1507 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1508 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1510 if (zp->zp_nopwrite) {
1511 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1512 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1513 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1516 return (ZIO_PIPELINE_CONTINUE);
1520 zio_free_bp_init(zio_t *zio)
1522 blkptr_t *bp = zio->io_bp;
1524 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1525 if (BP_GET_DEDUP(bp))
1526 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1529 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1531 return (ZIO_PIPELINE_CONTINUE);
1535 * ==========================================================================
1536 * Execute the I/O pipeline
1537 * ==========================================================================
1541 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1543 spa_t *spa = zio->io_spa;
1544 zio_type_t t = zio->io_type;
1545 int flags = (cutinline ? TQ_FRONT : 0);
1547 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1550 * If we're a config writer or a probe, the normal issue and
1551 * interrupt threads may all be blocked waiting for the config lock.
1552 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1554 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1558 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1560 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1564 * If this is a high priority I/O, then use the high priority taskq if
1567 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1568 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1571 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1574 * NB: We are assuming that the zio can only be dispatched
1575 * to a single taskq at a time. It would be a grievous error
1576 * to dispatch the zio to another taskq at the same time.
1578 #if defined(illumos) || !defined(_KERNEL)
1579 ASSERT(zio->io_tqent.tqent_next == NULL);
1581 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1583 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1584 flags, &zio->io_tqent);
1588 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1590 kthread_t *executor = zio->io_executor;
1591 spa_t *spa = zio->io_spa;
1593 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1594 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1596 for (i = 0; i < tqs->stqs_count; i++) {
1597 if (taskq_member(tqs->stqs_taskq[i], executor))
1606 zio_issue_async(zio_t *zio)
1608 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1610 return (ZIO_PIPELINE_STOP);
1614 zio_interrupt(zio_t *zio)
1616 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1620 zio_delay_interrupt(zio_t *zio)
1623 * The timeout_generic() function isn't defined in userspace, so
1624 * rather than trying to implement the function, the zio delay
1625 * functionality has been disabled for userspace builds.
1630 * If io_target_timestamp is zero, then no delay has been registered
1631 * for this IO, thus jump to the end of this function and "skip" the
1632 * delay; issuing it directly to the zio layer.
1634 if (zio->io_target_timestamp != 0) {
1635 hrtime_t now = gethrtime();
1637 if (now >= zio->io_target_timestamp) {
1639 * This IO has already taken longer than the target
1640 * delay to complete, so we don't want to delay it
1641 * any longer; we "miss" the delay and issue it
1642 * directly to the zio layer. This is likely due to
1643 * the target latency being set to a value less than
1644 * the underlying hardware can satisfy (e.g. delay
1645 * set to 1ms, but the disks take 10ms to complete an
1649 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1654 hrtime_t diff = zio->io_target_timestamp - now;
1656 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1657 hrtime_t, now, hrtime_t, diff);
1659 (void) timeout_generic(CALLOUT_NORMAL,
1660 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1667 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1672 * Execute the I/O pipeline until one of the following occurs:
1674 * (1) the I/O completes
1675 * (2) the pipeline stalls waiting for dependent child I/Os
1676 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1677 * (4) the I/O is delegated by vdev-level caching or aggregation
1678 * (5) the I/O is deferred due to vdev-level queueing
1679 * (6) the I/O is handed off to another thread.
1681 * In all cases, the pipeline stops whenever there's no CPU work; it never
1682 * burns a thread in cv_wait().
1684 * There's no locking on io_stage because there's no legitimate way
1685 * for multiple threads to be attempting to process the same I/O.
1687 static zio_pipe_stage_t *zio_pipeline[];
1690 zio_execute(zio_t *zio)
1692 zio->io_executor = curthread;
1694 ASSERT3U(zio->io_queued_timestamp, >, 0);
1696 while (zio->io_stage < ZIO_STAGE_DONE) {
1697 enum zio_stage pipeline = zio->io_pipeline;
1698 enum zio_stage stage = zio->io_stage;
1701 ASSERT(!MUTEX_HELD(&zio->io_lock));
1702 ASSERT(ISP2(stage));
1703 ASSERT(zio->io_stall == NULL);
1707 } while ((stage & pipeline) == 0);
1709 ASSERT(stage <= ZIO_STAGE_DONE);
1712 * If we are in interrupt context and this pipeline stage
1713 * will grab a config lock that is held across I/O,
1714 * or may wait for an I/O that needs an interrupt thread
1715 * to complete, issue async to avoid deadlock.
1717 * For VDEV_IO_START, we cut in line so that the io will
1718 * be sent to disk promptly.
1720 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1721 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1722 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1723 zio_requeue_io_start_cut_in_line : B_FALSE;
1724 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1728 zio->io_stage = stage;
1729 zio->io_pipeline_trace |= zio->io_stage;
1730 rv = zio_pipeline[highbit64(stage) - 1](zio);
1732 if (rv == ZIO_PIPELINE_STOP)
1735 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1740 * ==========================================================================
1741 * Initiate I/O, either sync or async
1742 * ==========================================================================
1745 zio_wait(zio_t *zio)
1749 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1750 ASSERT3P(zio->io_executor, ==, NULL);
1752 zio->io_waiter = curthread;
1753 ASSERT0(zio->io_queued_timestamp);
1754 zio->io_queued_timestamp = gethrtime();
1758 mutex_enter(&zio->io_lock);
1759 while (zio->io_executor != NULL)
1760 cv_wait(&zio->io_cv, &zio->io_lock);
1761 mutex_exit(&zio->io_lock);
1763 error = zio->io_error;
1770 zio_nowait(zio_t *zio)
1772 ASSERT3P(zio->io_executor, ==, NULL);
1774 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1775 zio_unique_parent(zio) == NULL) {
1777 * This is a logical async I/O with no parent to wait for it.
1778 * We add it to the spa_async_root_zio "Godfather" I/O which
1779 * will ensure they complete prior to unloading the pool.
1781 spa_t *spa = zio->io_spa;
1783 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1786 ASSERT0(zio->io_queued_timestamp);
1787 zio->io_queued_timestamp = gethrtime();
1792 * ==========================================================================
1793 * Reexecute, cancel, or suspend/resume failed I/O
1794 * ==========================================================================
1798 zio_reexecute(zio_t *pio)
1800 zio_t *cio, *cio_next;
1802 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1803 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1804 ASSERT(pio->io_gang_leader == NULL);
1805 ASSERT(pio->io_gang_tree == NULL);
1807 pio->io_flags = pio->io_orig_flags;
1808 pio->io_stage = pio->io_orig_stage;
1809 pio->io_pipeline = pio->io_orig_pipeline;
1810 pio->io_reexecute = 0;
1811 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1812 pio->io_pipeline_trace = 0;
1814 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1815 pio->io_state[w] = 0;
1816 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1817 pio->io_child_error[c] = 0;
1819 if (IO_IS_ALLOCATING(pio))
1820 BP_ZERO(pio->io_bp);
1823 * As we reexecute pio's children, new children could be created.
1824 * New children go to the head of pio's io_child_list, however,
1825 * so we will (correctly) not reexecute them. The key is that
1826 * the remainder of pio's io_child_list, from 'cio_next' onward,
1827 * cannot be affected by any side effects of reexecuting 'cio'.
1829 zio_link_t *zl = NULL;
1830 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1831 cio_next = zio_walk_children(pio, &zl);
1832 mutex_enter(&pio->io_lock);
1833 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1834 pio->io_children[cio->io_child_type][w]++;
1835 mutex_exit(&pio->io_lock);
1840 * Now that all children have been reexecuted, execute the parent.
1841 * We don't reexecute "The Godfather" I/O here as it's the
1842 * responsibility of the caller to wait on it.
1844 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1845 pio->io_queued_timestamp = gethrtime();
1851 zio_suspend(spa_t *spa, zio_t *zio)
1853 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1854 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1855 "failure and the failure mode property for this pool "
1856 "is set to panic.", spa_name(spa));
1858 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1860 mutex_enter(&spa->spa_suspend_lock);
1862 if (spa->spa_suspend_zio_root == NULL)
1863 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1864 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1865 ZIO_FLAG_GODFATHER);
1867 spa->spa_suspended = B_TRUE;
1870 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1871 ASSERT(zio != spa->spa_suspend_zio_root);
1872 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1873 ASSERT(zio_unique_parent(zio) == NULL);
1874 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1875 zio_add_child(spa->spa_suspend_zio_root, zio);
1878 mutex_exit(&spa->spa_suspend_lock);
1882 zio_resume(spa_t *spa)
1887 * Reexecute all previously suspended i/o.
1889 mutex_enter(&spa->spa_suspend_lock);
1890 spa->spa_suspended = B_FALSE;
1891 cv_broadcast(&spa->spa_suspend_cv);
1892 pio = spa->spa_suspend_zio_root;
1893 spa->spa_suspend_zio_root = NULL;
1894 mutex_exit(&spa->spa_suspend_lock);
1900 return (zio_wait(pio));
1904 zio_resume_wait(spa_t *spa)
1906 mutex_enter(&spa->spa_suspend_lock);
1907 while (spa_suspended(spa))
1908 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1909 mutex_exit(&spa->spa_suspend_lock);
1913 * ==========================================================================
1916 * A gang block is a collection of small blocks that looks to the DMU
1917 * like one large block. When zio_dva_allocate() cannot find a block
1918 * of the requested size, due to either severe fragmentation or the pool
1919 * being nearly full, it calls zio_write_gang_block() to construct the
1920 * block from smaller fragments.
1922 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1923 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1924 * an indirect block: it's an array of block pointers. It consumes
1925 * only one sector and hence is allocatable regardless of fragmentation.
1926 * The gang header's bps point to its gang members, which hold the data.
1928 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1929 * as the verifier to ensure uniqueness of the SHA256 checksum.
1930 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1931 * not the gang header. This ensures that data block signatures (needed for
1932 * deduplication) are independent of how the block is physically stored.
1934 * Gang blocks can be nested: a gang member may itself be a gang block.
1935 * Thus every gang block is a tree in which root and all interior nodes are
1936 * gang headers, and the leaves are normal blocks that contain user data.
1937 * The root of the gang tree is called the gang leader.
1939 * To perform any operation (read, rewrite, free, claim) on a gang block,
1940 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1941 * in the io_gang_tree field of the original logical i/o by recursively
1942 * reading the gang leader and all gang headers below it. This yields
1943 * an in-core tree containing the contents of every gang header and the
1944 * bps for every constituent of the gang block.
1946 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1947 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1948 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1949 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1950 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1951 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1952 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1953 * of the gang header plus zio_checksum_compute() of the data to update the
1954 * gang header's blk_cksum as described above.
1956 * The two-phase assemble/issue model solves the problem of partial failure --
1957 * what if you'd freed part of a gang block but then couldn't read the
1958 * gang header for another part? Assembling the entire gang tree first
1959 * ensures that all the necessary gang header I/O has succeeded before
1960 * starting the actual work of free, claim, or write. Once the gang tree
1961 * is assembled, free and claim are in-memory operations that cannot fail.
1963 * In the event that a gang write fails, zio_dva_unallocate() walks the
1964 * gang tree to immediately free (i.e. insert back into the space map)
1965 * everything we've allocated. This ensures that we don't get ENOSPC
1966 * errors during repeated suspend/resume cycles due to a flaky device.
1968 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1969 * the gang tree, we won't modify the block, so we can safely defer the free
1970 * (knowing that the block is still intact). If we *can* assemble the gang
1971 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1972 * each constituent bp and we can allocate a new block on the next sync pass.
1974 * In all cases, the gang tree allows complete recovery from partial failure.
1975 * ==========================================================================
1979 zio_gang_issue_func_done(zio_t *zio)
1981 abd_put(zio->io_abd);
1985 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1991 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
1992 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
1993 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1994 &pio->io_bookmark));
1998 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2005 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2006 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2007 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2008 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2011 * As we rewrite each gang header, the pipeline will compute
2012 * a new gang block header checksum for it; but no one will
2013 * compute a new data checksum, so we do that here. The one
2014 * exception is the gang leader: the pipeline already computed
2015 * its data checksum because that stage precedes gang assembly.
2016 * (Presently, nothing actually uses interior data checksums;
2017 * this is just good hygiene.)
2019 if (gn != pio->io_gang_leader->io_gang_tree) {
2020 abd_t *buf = abd_get_offset(data, offset);
2022 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2023 buf, BP_GET_PSIZE(bp));
2028 * If we are here to damage data for testing purposes,
2029 * leave the GBH alone so that we can detect the damage.
2031 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2032 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2034 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2035 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2036 zio_gang_issue_func_done, NULL, pio->io_priority,
2037 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2045 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2048 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2049 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2050 ZIO_GANG_CHILD_FLAGS(pio)));
2055 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2058 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2059 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2062 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2071 static void zio_gang_tree_assemble_done(zio_t *zio);
2073 static zio_gang_node_t *
2074 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2076 zio_gang_node_t *gn;
2078 ASSERT(*gnpp == NULL);
2080 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2081 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2088 zio_gang_node_free(zio_gang_node_t **gnpp)
2090 zio_gang_node_t *gn = *gnpp;
2092 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2093 ASSERT(gn->gn_child[g] == NULL);
2095 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2096 kmem_free(gn, sizeof (*gn));
2101 zio_gang_tree_free(zio_gang_node_t **gnpp)
2103 zio_gang_node_t *gn = *gnpp;
2108 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2109 zio_gang_tree_free(&gn->gn_child[g]);
2111 zio_gang_node_free(gnpp);
2115 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2117 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2118 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2120 ASSERT(gio->io_gang_leader == gio);
2121 ASSERT(BP_IS_GANG(bp));
2123 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2124 zio_gang_tree_assemble_done, gn, gio->io_priority,
2125 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2129 zio_gang_tree_assemble_done(zio_t *zio)
2131 zio_t *gio = zio->io_gang_leader;
2132 zio_gang_node_t *gn = zio->io_private;
2133 blkptr_t *bp = zio->io_bp;
2135 ASSERT(gio == zio_unique_parent(zio));
2136 ASSERT(zio->io_child_count == 0);
2141 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2142 if (BP_SHOULD_BYTESWAP(bp))
2143 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2145 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2146 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2147 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2149 abd_put(zio->io_abd);
2151 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2152 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2153 if (!BP_IS_GANG(gbp))
2155 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2160 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2163 zio_t *gio = pio->io_gang_leader;
2166 ASSERT(BP_IS_GANG(bp) == !!gn);
2167 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2168 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2171 * If you're a gang header, your data is in gn->gn_gbh.
2172 * If you're a gang member, your data is in 'data' and gn == NULL.
2174 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2177 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2179 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2180 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2181 if (BP_IS_HOLE(gbp))
2183 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2185 offset += BP_GET_PSIZE(gbp);
2189 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2190 ASSERT3U(gio->io_size, ==, offset);
2197 zio_gang_assemble(zio_t *zio)
2199 blkptr_t *bp = zio->io_bp;
2201 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2202 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2204 zio->io_gang_leader = zio;
2206 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2208 return (ZIO_PIPELINE_CONTINUE);
2212 zio_gang_issue(zio_t *zio)
2214 blkptr_t *bp = zio->io_bp;
2216 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2217 return (ZIO_PIPELINE_STOP);
2220 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2221 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2223 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2224 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2227 zio_gang_tree_free(&zio->io_gang_tree);
2229 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2231 return (ZIO_PIPELINE_CONTINUE);
2235 zio_write_gang_member_ready(zio_t *zio)
2237 zio_t *pio = zio_unique_parent(zio);
2238 zio_t *gio = zio->io_gang_leader;
2239 dva_t *cdva = zio->io_bp->blk_dva;
2240 dva_t *pdva = pio->io_bp->blk_dva;
2243 if (BP_IS_HOLE(zio->io_bp))
2246 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2248 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2249 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2250 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2251 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2252 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2254 mutex_enter(&pio->io_lock);
2255 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2256 ASSERT(DVA_GET_GANG(&pdva[d]));
2257 asize = DVA_GET_ASIZE(&pdva[d]);
2258 asize += DVA_GET_ASIZE(&cdva[d]);
2259 DVA_SET_ASIZE(&pdva[d], asize);
2261 mutex_exit(&pio->io_lock);
2265 zio_write_gang_done(zio_t *zio)
2267 abd_put(zio->io_abd);
2271 zio_write_gang_block(zio_t *pio)
2273 spa_t *spa = pio->io_spa;
2274 metaslab_class_t *mc = spa_normal_class(spa);
2275 blkptr_t *bp = pio->io_bp;
2276 zio_t *gio = pio->io_gang_leader;
2278 zio_gang_node_t *gn, **gnpp;
2279 zio_gbh_phys_t *gbh;
2281 uint64_t txg = pio->io_txg;
2282 uint64_t resid = pio->io_size;
2284 int copies = gio->io_prop.zp_copies;
2285 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2289 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2290 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2291 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2292 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2294 flags |= METASLAB_ASYNC_ALLOC;
2295 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2298 * The logical zio has already placed a reservation for
2299 * 'copies' allocation slots but gang blocks may require
2300 * additional copies. These additional copies
2301 * (i.e. gbh_copies - copies) are guaranteed to succeed
2302 * since metaslab_class_throttle_reserve() always allows
2303 * additional reservations for gang blocks.
2305 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2309 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2310 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2311 &pio->io_alloc_list, pio);
2313 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2314 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2315 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2318 * If we failed to allocate the gang block header then
2319 * we remove any additional allocation reservations that
2320 * we placed here. The original reservation will
2321 * be removed when the logical I/O goes to the ready
2324 metaslab_class_throttle_unreserve(mc,
2325 gbh_copies - copies, pio);
2327 pio->io_error = error;
2328 return (ZIO_PIPELINE_CONTINUE);
2332 gnpp = &gio->io_gang_tree;
2334 gnpp = pio->io_private;
2335 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2338 gn = zio_gang_node_alloc(gnpp);
2340 bzero(gbh, SPA_GANGBLOCKSIZE);
2341 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2344 * Create the gang header.
2346 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2347 zio_write_gang_done, NULL, pio->io_priority,
2348 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2351 * Create and nowait the gang children.
2353 for (int g = 0; resid != 0; resid -= lsize, g++) {
2354 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2356 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2358 zp.zp_checksum = gio->io_prop.zp_checksum;
2359 zp.zp_compress = ZIO_COMPRESS_OFF;
2360 zp.zp_type = DMU_OT_NONE;
2362 zp.zp_copies = gio->io_prop.zp_copies;
2363 zp.zp_dedup = B_FALSE;
2364 zp.zp_dedup_verify = B_FALSE;
2365 zp.zp_nopwrite = B_FALSE;
2367 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2368 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2369 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2370 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2371 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2373 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2374 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2375 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2378 * Gang children won't throttle but we should
2379 * account for their work, so reserve an allocation
2380 * slot for them here.
2382 VERIFY(metaslab_class_throttle_reserve(mc,
2383 zp.zp_copies, cio, flags));
2389 * Set pio's pipeline to just wait for zio to finish.
2391 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2395 return (ZIO_PIPELINE_CONTINUE);
2399 * The zio_nop_write stage in the pipeline determines if allocating a
2400 * new bp is necessary. The nopwrite feature can handle writes in
2401 * either syncing or open context (i.e. zil writes) and as a result is
2402 * mutually exclusive with dedup.
2404 * By leveraging a cryptographically secure checksum, such as SHA256, we
2405 * can compare the checksums of the new data and the old to determine if
2406 * allocating a new block is required. Note that our requirements for
2407 * cryptographic strength are fairly weak: there can't be any accidental
2408 * hash collisions, but we don't need to be secure against intentional
2409 * (malicious) collisions. To trigger a nopwrite, you have to be able
2410 * to write the file to begin with, and triggering an incorrect (hash
2411 * collision) nopwrite is no worse than simply writing to the file.
2412 * That said, there are no known attacks against the checksum algorithms
2413 * used for nopwrite, assuming that the salt and the checksums
2414 * themselves remain secret.
2417 zio_nop_write(zio_t *zio)
2419 blkptr_t *bp = zio->io_bp;
2420 blkptr_t *bp_orig = &zio->io_bp_orig;
2421 zio_prop_t *zp = &zio->io_prop;
2423 ASSERT(BP_GET_LEVEL(bp) == 0);
2424 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2425 ASSERT(zp->zp_nopwrite);
2426 ASSERT(!zp->zp_dedup);
2427 ASSERT(zio->io_bp_override == NULL);
2428 ASSERT(IO_IS_ALLOCATING(zio));
2431 * Check to see if the original bp and the new bp have matching
2432 * characteristics (i.e. same checksum, compression algorithms, etc).
2433 * If they don't then just continue with the pipeline which will
2434 * allocate a new bp.
2436 if (BP_IS_HOLE(bp_orig) ||
2437 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2438 ZCHECKSUM_FLAG_NOPWRITE) ||
2439 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2440 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2441 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2442 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2443 return (ZIO_PIPELINE_CONTINUE);
2446 * If the checksums match then reset the pipeline so that we
2447 * avoid allocating a new bp and issuing any I/O.
2449 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2450 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2451 ZCHECKSUM_FLAG_NOPWRITE);
2452 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2453 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2454 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2455 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2456 sizeof (uint64_t)) == 0);
2459 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2460 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2463 return (ZIO_PIPELINE_CONTINUE);
2467 * ==========================================================================
2469 * ==========================================================================
2472 zio_ddt_child_read_done(zio_t *zio)
2474 blkptr_t *bp = zio->io_bp;
2475 ddt_entry_t *dde = zio->io_private;
2477 zio_t *pio = zio_unique_parent(zio);
2479 mutex_enter(&pio->io_lock);
2480 ddp = ddt_phys_select(dde, bp);
2481 if (zio->io_error == 0)
2482 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2484 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2485 dde->dde_repair_abd = zio->io_abd;
2487 abd_free(zio->io_abd);
2488 mutex_exit(&pio->io_lock);
2492 zio_ddt_read_start(zio_t *zio)
2494 blkptr_t *bp = zio->io_bp;
2496 ASSERT(BP_GET_DEDUP(bp));
2497 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2498 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2500 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2501 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2502 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2503 ddt_phys_t *ddp = dde->dde_phys;
2504 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2507 ASSERT(zio->io_vsd == NULL);
2510 if (ddp_self == NULL)
2511 return (ZIO_PIPELINE_CONTINUE);
2513 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2514 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2516 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2518 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2519 abd_alloc_for_io(zio->io_size, B_TRUE),
2520 zio->io_size, zio_ddt_child_read_done, dde,
2521 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2522 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2524 return (ZIO_PIPELINE_CONTINUE);
2527 zio_nowait(zio_read(zio, zio->io_spa, bp,
2528 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2529 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2531 return (ZIO_PIPELINE_CONTINUE);
2535 zio_ddt_read_done(zio_t *zio)
2537 blkptr_t *bp = zio->io_bp;
2539 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2540 return (ZIO_PIPELINE_STOP);
2543 ASSERT(BP_GET_DEDUP(bp));
2544 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2545 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2547 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2548 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2549 ddt_entry_t *dde = zio->io_vsd;
2551 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2552 return (ZIO_PIPELINE_CONTINUE);
2555 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2556 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2557 return (ZIO_PIPELINE_STOP);
2559 if (dde->dde_repair_abd != NULL) {
2560 abd_copy(zio->io_abd, dde->dde_repair_abd,
2562 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2564 ddt_repair_done(ddt, dde);
2568 ASSERT(zio->io_vsd == NULL);
2570 return (ZIO_PIPELINE_CONTINUE);
2574 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2576 spa_t *spa = zio->io_spa;
2577 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2579 /* We should never get a raw, override zio */
2580 ASSERT(!(zio->io_bp_override && do_raw));
2583 * Note: we compare the original data, not the transformed data,
2584 * because when zio->io_bp is an override bp, we will not have
2585 * pushed the I/O transforms. That's an important optimization
2586 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2588 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2589 zio_t *lio = dde->dde_lead_zio[p];
2592 return (lio->io_orig_size != zio->io_orig_size ||
2593 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2594 zio->io_orig_size) != 0);
2598 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2599 ddt_phys_t *ddp = &dde->dde_phys[p];
2601 if (ddp->ddp_phys_birth != 0) {
2602 arc_buf_t *abuf = NULL;
2603 arc_flags_t aflags = ARC_FLAG_WAIT;
2604 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2605 blkptr_t blk = *zio->io_bp;
2608 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2613 * Intuitively, it would make more sense to compare
2614 * io_abd than io_orig_abd in the raw case since you
2615 * don't want to look at any transformations that have
2616 * happened to the data. However, for raw I/Os the
2617 * data will actually be the same in io_abd and
2618 * io_orig_abd, so all we have to do is issue this as
2622 zio_flags |= ZIO_FLAG_RAW;
2623 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2624 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2626 ASSERT3P(zio->io_transform_stack, ==, NULL);
2629 error = arc_read(NULL, spa, &blk,
2630 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2631 zio_flags, &aflags, &zio->io_bookmark);
2634 if (arc_buf_size(abuf) != zio->io_orig_size ||
2635 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2636 zio->io_orig_size) != 0)
2637 error = SET_ERROR(EEXIST);
2638 arc_buf_destroy(abuf, &abuf);
2642 return (error != 0);
2650 zio_ddt_child_write_ready(zio_t *zio)
2652 int p = zio->io_prop.zp_copies;
2653 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2654 ddt_entry_t *dde = zio->io_private;
2655 ddt_phys_t *ddp = &dde->dde_phys[p];
2663 ASSERT(dde->dde_lead_zio[p] == zio);
2665 ddt_phys_fill(ddp, zio->io_bp);
2667 zio_link_t *zl = NULL;
2668 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2669 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2675 zio_ddt_child_write_done(zio_t *zio)
2677 int p = zio->io_prop.zp_copies;
2678 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2679 ddt_entry_t *dde = zio->io_private;
2680 ddt_phys_t *ddp = &dde->dde_phys[p];
2684 ASSERT(ddp->ddp_refcnt == 0);
2685 ASSERT(dde->dde_lead_zio[p] == zio);
2686 dde->dde_lead_zio[p] = NULL;
2688 if (zio->io_error == 0) {
2689 zio_link_t *zl = NULL;
2690 while (zio_walk_parents(zio, &zl) != NULL)
2691 ddt_phys_addref(ddp);
2693 ddt_phys_clear(ddp);
2700 zio_ddt_ditto_write_done(zio_t *zio)
2702 int p = DDT_PHYS_DITTO;
2703 zio_prop_t *zp = &zio->io_prop;
2704 blkptr_t *bp = zio->io_bp;
2705 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2706 ddt_entry_t *dde = zio->io_private;
2707 ddt_phys_t *ddp = &dde->dde_phys[p];
2708 ddt_key_t *ddk = &dde->dde_key;
2712 ASSERT(ddp->ddp_refcnt == 0);
2713 ASSERT(dde->dde_lead_zio[p] == zio);
2714 dde->dde_lead_zio[p] = NULL;
2716 if (zio->io_error == 0) {
2717 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2718 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2719 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2720 if (ddp->ddp_phys_birth != 0)
2721 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2722 ddt_phys_fill(ddp, bp);
2729 zio_ddt_write(zio_t *zio)
2731 spa_t *spa = zio->io_spa;
2732 blkptr_t *bp = zio->io_bp;
2733 uint64_t txg = zio->io_txg;
2734 zio_prop_t *zp = &zio->io_prop;
2735 int p = zp->zp_copies;
2739 ddt_t *ddt = ddt_select(spa, bp);
2743 ASSERT(BP_GET_DEDUP(bp));
2744 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2745 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2746 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2749 dde = ddt_lookup(ddt, bp, B_TRUE);
2750 ddp = &dde->dde_phys[p];
2752 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2754 * If we're using a weak checksum, upgrade to a strong checksum
2755 * and try again. If we're already using a strong checksum,
2756 * we can't resolve it, so just convert to an ordinary write.
2757 * (And automatically e-mail a paper to Nature?)
2759 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2760 ZCHECKSUM_FLAG_DEDUP)) {
2761 zp->zp_checksum = spa_dedup_checksum(spa);
2762 zio_pop_transforms(zio);
2763 zio->io_stage = ZIO_STAGE_OPEN;
2766 zp->zp_dedup = B_FALSE;
2767 BP_SET_DEDUP(bp, B_FALSE);
2769 ASSERT(!BP_GET_DEDUP(bp));
2770 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2772 return (ZIO_PIPELINE_CONTINUE);
2775 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2776 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2778 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2779 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2780 zio_prop_t czp = *zp;
2782 czp.zp_copies = ditto_copies;
2785 * If we arrived here with an override bp, we won't have run
2786 * the transform stack, so we won't have the data we need to
2787 * generate a child i/o. So, toss the override bp and restart.
2788 * This is safe, because using the override bp is just an
2789 * optimization; and it's rare, so the cost doesn't matter.
2791 if (zio->io_bp_override) {
2792 zio_pop_transforms(zio);
2793 zio->io_stage = ZIO_STAGE_OPEN;
2794 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2795 zio->io_bp_override = NULL;
2798 return (ZIO_PIPELINE_CONTINUE);
2801 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2802 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2803 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2804 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2806 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2807 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2810 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2811 if (ddp->ddp_phys_birth != 0)
2812 ddt_bp_fill(ddp, bp, txg);
2813 if (dde->dde_lead_zio[p] != NULL)
2814 zio_add_child(zio, dde->dde_lead_zio[p]);
2816 ddt_phys_addref(ddp);
2817 } else if (zio->io_bp_override) {
2818 ASSERT(bp->blk_birth == txg);
2819 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2820 ddt_phys_fill(ddp, bp);
2821 ddt_phys_addref(ddp);
2823 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2824 zio->io_orig_size, zio->io_orig_size, zp,
2825 zio_ddt_child_write_ready, NULL, NULL,
2826 zio_ddt_child_write_done, dde, zio->io_priority,
2827 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2829 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2830 dde->dde_lead_zio[p] = cio;
2840 return (ZIO_PIPELINE_CONTINUE);
2843 ddt_entry_t *freedde; /* for debugging */
2846 zio_ddt_free(zio_t *zio)
2848 spa_t *spa = zio->io_spa;
2849 blkptr_t *bp = zio->io_bp;
2850 ddt_t *ddt = ddt_select(spa, bp);
2854 ASSERT(BP_GET_DEDUP(bp));
2855 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2858 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2859 ddp = ddt_phys_select(dde, bp);
2860 ddt_phys_decref(ddp);
2863 return (ZIO_PIPELINE_CONTINUE);
2867 * ==========================================================================
2868 * Allocate and free blocks
2869 * ==========================================================================
2873 zio_io_to_allocate(spa_t *spa)
2877 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2879 zio = avl_first(&spa->spa_alloc_tree);
2883 ASSERT(IO_IS_ALLOCATING(zio));
2886 * Try to place a reservation for this zio. If we're unable to
2887 * reserve then we throttle.
2889 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2890 zio->io_prop.zp_copies, zio, 0)) {
2894 avl_remove(&spa->spa_alloc_tree, zio);
2895 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2901 zio_dva_throttle(zio_t *zio)
2903 spa_t *spa = zio->io_spa;
2906 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2907 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2908 zio->io_child_type == ZIO_CHILD_GANG ||
2909 zio->io_flags & ZIO_FLAG_NODATA) {
2910 return (ZIO_PIPELINE_CONTINUE);
2913 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2915 ASSERT3U(zio->io_queued_timestamp, >, 0);
2916 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2918 mutex_enter(&spa->spa_alloc_lock);
2920 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2921 avl_add(&spa->spa_alloc_tree, zio);
2923 nio = zio_io_to_allocate(zio->io_spa);
2924 mutex_exit(&spa->spa_alloc_lock);
2927 return (ZIO_PIPELINE_CONTINUE);
2930 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2932 * We are passing control to a new zio so make sure that
2933 * it is processed by a different thread. We do this to
2934 * avoid stack overflows that can occur when parents are
2935 * throttled and children are making progress. We allow
2936 * it to go to the head of the taskq since it's already
2939 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2941 return (ZIO_PIPELINE_STOP);
2945 zio_allocate_dispatch(spa_t *spa)
2949 mutex_enter(&spa->spa_alloc_lock);
2950 zio = zio_io_to_allocate(spa);
2951 mutex_exit(&spa->spa_alloc_lock);
2955 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2956 ASSERT0(zio->io_error);
2957 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2961 zio_dva_allocate(zio_t *zio)
2963 spa_t *spa = zio->io_spa;
2964 metaslab_class_t *mc = spa_normal_class(spa);
2965 blkptr_t *bp = zio->io_bp;
2969 if (zio->io_gang_leader == NULL) {
2970 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2971 zio->io_gang_leader = zio;
2974 ASSERT(BP_IS_HOLE(bp));
2975 ASSERT0(BP_GET_NDVAS(bp));
2976 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2977 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2978 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2980 if (zio->io_flags & ZIO_FLAG_NODATA) {
2981 flags |= METASLAB_DONT_THROTTLE;
2983 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2984 flags |= METASLAB_GANG_CHILD;
2986 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2987 flags |= METASLAB_ASYNC_ALLOC;
2990 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2991 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2992 &zio->io_alloc_list, zio);
2995 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2996 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2998 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2999 return (zio_write_gang_block(zio));
3000 zio->io_error = error;
3003 return (ZIO_PIPELINE_CONTINUE);
3007 zio_dva_free(zio_t *zio)
3009 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3011 return (ZIO_PIPELINE_CONTINUE);
3015 zio_dva_claim(zio_t *zio)
3019 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3021 zio->io_error = error;
3023 return (ZIO_PIPELINE_CONTINUE);
3027 * Undo an allocation. This is used by zio_done() when an I/O fails
3028 * and we want to give back the block we just allocated.
3029 * This handles both normal blocks and gang blocks.
3032 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3034 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3035 ASSERT(zio->io_bp_override == NULL);
3037 if (!BP_IS_HOLE(bp))
3038 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3041 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3042 zio_dva_unallocate(zio, gn->gn_child[g],
3043 &gn->gn_gbh->zg_blkptr[g]);
3049 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3052 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3053 uint64_t size, boolean_t *slog)
3056 zio_alloc_list_t io_alloc_list;
3058 ASSERT(txg > spa_syncing_txg(spa));
3060 metaslab_trace_init(&io_alloc_list);
3061 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3062 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3066 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3067 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3068 &io_alloc_list, NULL);
3072 metaslab_trace_fini(&io_alloc_list);
3075 BP_SET_LSIZE(new_bp, size);
3076 BP_SET_PSIZE(new_bp, size);
3077 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3078 BP_SET_CHECKSUM(new_bp,
3079 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3080 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3081 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3082 BP_SET_LEVEL(new_bp, 0);
3083 BP_SET_DEDUP(new_bp, 0);
3084 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3086 zfs_dbgmsg("%s: zil block allocation failure: "
3087 "size %llu, error %d", spa_name(spa), size, error);
3094 * Free an intent log block.
3097 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3099 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3100 ASSERT(!BP_IS_GANG(bp));
3102 zio_free(spa, txg, bp);
3106 * ==========================================================================
3107 * Read, write and delete to physical devices
3108 * ==========================================================================
3113 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3114 * stops after this stage and will resume upon I/O completion.
3115 * However, there are instances where the vdev layer may need to
3116 * continue the pipeline when an I/O was not issued. Since the I/O
3117 * that was sent to the vdev layer might be different than the one
3118 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3119 * force the underlying vdev layers to call either zio_execute() or
3120 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3123 zio_vdev_io_start(zio_t *zio)
3125 vdev_t *vd = zio->io_vd;
3127 spa_t *spa = zio->io_spa;
3130 ASSERT(zio->io_error == 0);
3131 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3134 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3135 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3138 * The mirror_ops handle multiple DVAs in a single BP.
3140 vdev_mirror_ops.vdev_op_io_start(zio);
3141 return (ZIO_PIPELINE_STOP);
3144 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3145 zio->io_priority == ZIO_PRIORITY_NOW) {
3146 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3147 return (ZIO_PIPELINE_CONTINUE);
3150 ASSERT3P(zio->io_logical, !=, zio);
3151 if (zio->io_type == ZIO_TYPE_WRITE && zio->io_vd->vdev_removing) {
3152 ASSERT(zio->io_flags &
3153 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3154 ZIO_FLAG_INDUCE_DAMAGE));
3158 * We keep track of time-sensitive I/Os so that the scan thread
3159 * can quickly react to certain workloads. In particular, we care
3160 * about non-scrubbing, top-level reads and writes with the following
3162 * - synchronous writes of user data to non-slog devices
3163 * - any reads of user data
3164 * When these conditions are met, adjust the timestamp of spa_last_io
3165 * which allows the scan thread to adjust its workload accordingly.
3167 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3168 vd == vd->vdev_top && !vd->vdev_islog &&
3169 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3170 zio->io_txg != spa_syncing_txg(spa)) {
3171 uint64_t old = spa->spa_last_io;
3172 uint64_t new = ddi_get_lbolt64();
3174 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3177 align = 1ULL << vd->vdev_top->vdev_ashift;
3179 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3180 P2PHASE(zio->io_size, align) != 0) {
3181 /* Transform logical writes to be a full physical block size. */
3182 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3184 if (zio->io_type == ZIO_TYPE_READ ||
3185 zio->io_type == ZIO_TYPE_WRITE)
3186 abuf = abd_alloc_sametype(zio->io_abd, asize);
3187 ASSERT(vd == vd->vdev_top);
3188 if (zio->io_type == ZIO_TYPE_WRITE) {
3189 abd_copy(abuf, zio->io_abd, zio->io_size);
3190 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3192 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3197 * If this is not a physical io, make sure that it is properly aligned
3198 * before proceeding.
3200 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3201 ASSERT0(P2PHASE(zio->io_offset, align));
3202 ASSERT0(P2PHASE(zio->io_size, align));
3205 * For the physical io we allow alignment
3206 * to a logical block size.
3208 uint64_t log_align =
3209 1ULL << vd->vdev_top->vdev_logical_ashift;
3210 ASSERT0(P2PHASE(zio->io_offset, log_align));
3211 ASSERT0(P2PHASE(zio->io_size, log_align));
3214 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3217 * If this is a repair I/O, and there's no self-healing involved --
3218 * that is, we're just resilvering what we expect to resilver --
3219 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3220 * This prevents spurious resilvering with nested replication.
3221 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3222 * A is out of date, we'll read from C+D, then use the data to
3223 * resilver A+B -- but we don't actually want to resilver B, just A.
3224 * The top-level mirror has no way to know this, so instead we just
3225 * discard unnecessary repairs as we work our way down the vdev tree.
3226 * The same logic applies to any form of nested replication:
3227 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3229 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3230 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3231 zio->io_txg != 0 && /* not a delegated i/o */
3232 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3233 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3234 zio_vdev_io_bypass(zio);
3235 return (ZIO_PIPELINE_CONTINUE);
3238 if (vd->vdev_ops->vdev_op_leaf) {
3239 switch (zio->io_type) {
3241 if (vdev_cache_read(zio))
3242 return (ZIO_PIPELINE_CONTINUE);
3244 case ZIO_TYPE_WRITE:
3246 if ((zio = vdev_queue_io(zio)) == NULL)
3247 return (ZIO_PIPELINE_STOP);
3249 if (!vdev_accessible(vd, zio)) {
3250 zio->io_error = SET_ERROR(ENXIO);
3252 return (ZIO_PIPELINE_STOP);
3257 * Note that we ignore repair writes for TRIM because they can
3258 * conflict with normal writes. This isn't an issue because, by
3259 * definition, we only repair blocks that aren't freed.
3261 if (zio->io_type == ZIO_TYPE_WRITE &&
3262 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3263 !trim_map_write_start(zio))
3264 return (ZIO_PIPELINE_STOP);
3267 vd->vdev_ops->vdev_op_io_start(zio);
3268 return (ZIO_PIPELINE_STOP);
3272 zio_vdev_io_done(zio_t *zio)
3274 vdev_t *vd = zio->io_vd;
3275 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3276 boolean_t unexpected_error = B_FALSE;
3278 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3279 return (ZIO_PIPELINE_STOP);
3282 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3283 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3285 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3286 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3287 zio->io_type == ZIO_TYPE_FREE)) {
3289 if (zio->io_type == ZIO_TYPE_WRITE &&
3290 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3291 trim_map_write_done(zio);
3293 vdev_queue_io_done(zio);
3295 if (zio->io_type == ZIO_TYPE_WRITE)
3296 vdev_cache_write(zio);
3298 if (zio_injection_enabled && zio->io_error == 0)
3299 zio->io_error = zio_handle_device_injection(vd,
3302 if (zio_injection_enabled && zio->io_error == 0)
3303 zio->io_error = zio_handle_label_injection(zio, EIO);
3305 if (zio->io_error) {
3306 if (zio->io_error == ENOTSUP &&
3307 zio->io_type == ZIO_TYPE_FREE) {
3308 /* Not all devices support TRIM. */
3309 } else if (!vdev_accessible(vd, zio)) {
3310 zio->io_error = SET_ERROR(ENXIO);
3312 unexpected_error = B_TRUE;
3317 ops->vdev_op_io_done(zio);
3319 if (unexpected_error)
3320 VERIFY(vdev_probe(vd, zio) == NULL);
3322 return (ZIO_PIPELINE_CONTINUE);
3326 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3327 * disk, and use that to finish the checksum ereport later.
3330 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3331 const void *good_buf)
3333 /* no processing needed */
3334 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3339 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3341 void *buf = zio_buf_alloc(zio->io_size);
3343 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3345 zcr->zcr_cbinfo = zio->io_size;
3346 zcr->zcr_cbdata = buf;
3347 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3348 zcr->zcr_free = zio_buf_free;
3352 zio_vdev_io_assess(zio_t *zio)
3354 vdev_t *vd = zio->io_vd;
3356 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3357 return (ZIO_PIPELINE_STOP);
3360 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3361 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3363 if (zio->io_vsd != NULL) {
3364 zio->io_vsd_ops->vsd_free(zio);
3368 if (zio_injection_enabled && zio->io_error == 0)
3369 zio->io_error = zio_handle_fault_injection(zio, EIO);
3371 if (zio->io_type == ZIO_TYPE_FREE &&
3372 zio->io_priority != ZIO_PRIORITY_NOW) {
3373 switch (zio->io_error) {
3375 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3376 ZIO_TRIM_STAT_BUMP(success);
3379 ZIO_TRIM_STAT_BUMP(unsupported);
3382 ZIO_TRIM_STAT_BUMP(failed);
3388 * If the I/O failed, determine whether we should attempt to retry it.
3390 * On retry, we cut in line in the issue queue, since we don't want
3391 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3393 if (zio->io_error && vd == NULL &&
3394 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3395 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3396 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3398 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3399 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3400 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3401 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3402 zio_requeue_io_start_cut_in_line);
3403 return (ZIO_PIPELINE_STOP);
3407 * If we got an error on a leaf device, convert it to ENXIO
3408 * if the device is not accessible at all.
3410 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3411 !vdev_accessible(vd, zio))
3412 zio->io_error = SET_ERROR(ENXIO);
3415 * If we can't write to an interior vdev (mirror or RAID-Z),
3416 * set vdev_cant_write so that we stop trying to allocate from it.
3418 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3419 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3420 vd->vdev_cant_write = B_TRUE;
3424 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3425 * attempts will ever succeed. In this case we set a persistent bit so
3426 * that we don't bother with it in the future.
3428 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3429 zio->io_type == ZIO_TYPE_IOCTL &&
3430 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3431 vd->vdev_nowritecache = B_TRUE;
3434 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3436 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3437 zio->io_physdone != NULL) {
3438 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3439 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3440 zio->io_physdone(zio->io_logical);
3443 return (ZIO_PIPELINE_CONTINUE);
3447 zio_vdev_io_reissue(zio_t *zio)
3449 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3450 ASSERT(zio->io_error == 0);
3452 zio->io_stage >>= 1;
3456 zio_vdev_io_redone(zio_t *zio)
3458 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3460 zio->io_stage >>= 1;
3464 zio_vdev_io_bypass(zio_t *zio)
3466 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3467 ASSERT(zio->io_error == 0);
3469 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3470 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3474 * ==========================================================================
3475 * Generate and verify checksums
3476 * ==========================================================================
3479 zio_checksum_generate(zio_t *zio)
3481 blkptr_t *bp = zio->io_bp;
3482 enum zio_checksum checksum;
3486 * This is zio_write_phys().
3487 * We're either generating a label checksum, or none at all.
3489 checksum = zio->io_prop.zp_checksum;
3491 if (checksum == ZIO_CHECKSUM_OFF)
3492 return (ZIO_PIPELINE_CONTINUE);
3494 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3496 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3497 ASSERT(!IO_IS_ALLOCATING(zio));
3498 checksum = ZIO_CHECKSUM_GANG_HEADER;
3500 checksum = BP_GET_CHECKSUM(bp);
3504 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3506 return (ZIO_PIPELINE_CONTINUE);
3510 zio_checksum_verify(zio_t *zio)
3512 zio_bad_cksum_t info;
3513 blkptr_t *bp = zio->io_bp;
3516 ASSERT(zio->io_vd != NULL);
3520 * This is zio_read_phys().
3521 * We're either verifying a label checksum, or nothing at all.
3523 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3524 return (ZIO_PIPELINE_CONTINUE);
3526 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3529 if ((error = zio_checksum_error(zio, &info)) != 0) {
3530 zio->io_error = error;
3531 if (error == ECKSUM &&
3532 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3533 zfs_ereport_start_checksum(zio->io_spa,
3534 zio->io_vd, zio, zio->io_offset,
3535 zio->io_size, NULL, &info);
3539 return (ZIO_PIPELINE_CONTINUE);
3543 * Called by RAID-Z to ensure we don't compute the checksum twice.
3546 zio_checksum_verified(zio_t *zio)
3548 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3552 * ==========================================================================
3553 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3554 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3555 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3556 * indicate errors that are specific to one I/O, and most likely permanent.
3557 * Any other error is presumed to be worse because we weren't expecting it.
3558 * ==========================================================================
3561 zio_worst_error(int e1, int e2)
3563 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3566 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3567 if (e1 == zio_error_rank[r1])
3570 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3571 if (e2 == zio_error_rank[r2])
3574 return (r1 > r2 ? e1 : e2);
3578 * ==========================================================================
3580 * ==========================================================================
3583 zio_ready(zio_t *zio)
3585 blkptr_t *bp = zio->io_bp;
3586 zio_t *pio, *pio_next;
3587 zio_link_t *zl = NULL;
3589 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3591 return (ZIO_PIPELINE_STOP);
3594 if (zio->io_ready) {
3595 ASSERT(IO_IS_ALLOCATING(zio));
3596 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3597 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3598 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3603 if (bp != NULL && bp != &zio->io_bp_copy)
3604 zio->io_bp_copy = *bp;
3606 if (zio->io_error != 0) {
3607 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3609 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3610 ASSERT(IO_IS_ALLOCATING(zio));
3611 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3613 * We were unable to allocate anything, unreserve and
3614 * issue the next I/O to allocate.
3616 metaslab_class_throttle_unreserve(
3617 spa_normal_class(zio->io_spa),
3618 zio->io_prop.zp_copies, zio);
3619 zio_allocate_dispatch(zio->io_spa);
3623 mutex_enter(&zio->io_lock);
3624 zio->io_state[ZIO_WAIT_READY] = 1;
3625 pio = zio_walk_parents(zio, &zl);
3626 mutex_exit(&zio->io_lock);
3629 * As we notify zio's parents, new parents could be added.
3630 * New parents go to the head of zio's io_parent_list, however,
3631 * so we will (correctly) not notify them. The remainder of zio's
3632 * io_parent_list, from 'pio_next' onward, cannot change because
3633 * all parents must wait for us to be done before they can be done.
3635 for (; pio != NULL; pio = pio_next) {
3636 pio_next = zio_walk_parents(zio, &zl);
3637 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3640 if (zio->io_flags & ZIO_FLAG_NODATA) {
3641 if (BP_IS_GANG(bp)) {
3642 zio->io_flags &= ~ZIO_FLAG_NODATA;
3644 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3645 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3649 if (zio_injection_enabled &&
3650 zio->io_spa->spa_syncing_txg == zio->io_txg)
3651 zio_handle_ignored_writes(zio);
3653 return (ZIO_PIPELINE_CONTINUE);
3657 * Update the allocation throttle accounting.
3660 zio_dva_throttle_done(zio_t *zio)
3662 zio_t *lio = zio->io_logical;
3663 zio_t *pio = zio_unique_parent(zio);
3664 vdev_t *vd = zio->io_vd;
3665 int flags = METASLAB_ASYNC_ALLOC;
3667 ASSERT3P(zio->io_bp, !=, NULL);
3668 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3669 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3670 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3672 ASSERT3P(vd, ==, vd->vdev_top);
3673 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3674 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3675 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3676 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3679 * Parents of gang children can have two flavors -- ones that
3680 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3681 * and ones that allocated the constituent blocks. The allocation
3682 * throttle needs to know the allocating parent zio so we must find
3685 if (pio->io_child_type == ZIO_CHILD_GANG) {
3687 * If our parent is a rewrite gang child then our grandparent
3688 * would have been the one that performed the allocation.
3690 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3691 pio = zio_unique_parent(pio);
3692 flags |= METASLAB_GANG_CHILD;
3695 ASSERT(IO_IS_ALLOCATING(pio));
3696 ASSERT3P(zio, !=, zio->io_logical);
3697 ASSERT(zio->io_logical != NULL);
3698 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3699 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3701 mutex_enter(&pio->io_lock);
3702 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3703 mutex_exit(&pio->io_lock);
3705 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3709 * Call into the pipeline to see if there is more work that
3710 * needs to be done. If there is work to be done it will be
3711 * dispatched to another taskq thread.
3713 zio_allocate_dispatch(zio->io_spa);
3717 zio_done(zio_t *zio)
3719 spa_t *spa = zio->io_spa;
3720 zio_t *lio = zio->io_logical;
3721 blkptr_t *bp = zio->io_bp;
3722 vdev_t *vd = zio->io_vd;
3723 uint64_t psize = zio->io_size;
3724 zio_t *pio, *pio_next;
3725 metaslab_class_t *mc = spa_normal_class(spa);
3726 zio_link_t *zl = NULL;
3729 * If our children haven't all completed,
3730 * wait for them and then repeat this pipeline stage.
3732 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3733 return (ZIO_PIPELINE_STOP);
3737 * If the allocation throttle is enabled, then update the accounting.
3738 * We only track child I/Os that are part of an allocating async
3739 * write. We must do this since the allocation is performed
3740 * by the logical I/O but the actual write is done by child I/Os.
3742 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3743 zio->io_child_type == ZIO_CHILD_VDEV) {
3744 ASSERT(mc->mc_alloc_throttle_enabled);
3745 zio_dva_throttle_done(zio);
3749 * If the allocation throttle is enabled, verify that
3750 * we have decremented the refcounts for every I/O that was throttled.
3752 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3753 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3754 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3756 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3757 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3760 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3761 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3762 ASSERT(zio->io_children[c][w] == 0);
3764 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3765 ASSERT(bp->blk_pad[0] == 0);
3766 ASSERT(bp->blk_pad[1] == 0);
3767 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3768 (bp == zio_unique_parent(zio)->io_bp));
3769 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3770 zio->io_bp_override == NULL &&
3771 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3772 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3773 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3774 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3775 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3777 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3778 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3782 * If there were child vdev/gang/ddt errors, they apply to us now.
3784 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3785 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3786 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3789 * If the I/O on the transformed data was successful, generate any
3790 * checksum reports now while we still have the transformed data.
3792 if (zio->io_error == 0) {
3793 while (zio->io_cksum_report != NULL) {
3794 zio_cksum_report_t *zcr = zio->io_cksum_report;
3795 uint64_t align = zcr->zcr_align;
3796 uint64_t asize = P2ROUNDUP(psize, align);
3798 abd_t *adata = zio->io_abd;
3800 if (asize != psize) {
3801 adata = abd_alloc_linear(asize, B_TRUE);
3802 abd_copy(adata, zio->io_abd, psize);
3803 abd_zero_off(adata, psize, asize - psize);
3807 abuf = abd_borrow_buf_copy(adata, asize);
3809 zio->io_cksum_report = zcr->zcr_next;
3810 zcr->zcr_next = NULL;
3811 zcr->zcr_finish(zcr, abuf);
3812 zfs_ereport_free_checksum(zcr);
3815 abd_return_buf(adata, abuf, asize);
3822 zio_pop_transforms(zio); /* note: may set zio->io_error */
3824 vdev_stat_update(zio, psize);
3826 if (zio->io_error) {
3828 * If this I/O is attached to a particular vdev,
3829 * generate an error message describing the I/O failure
3830 * at the block level. We ignore these errors if the
3831 * device is currently unavailable.
3833 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3834 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3836 if ((zio->io_error == EIO || !(zio->io_flags &
3837 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3840 * For logical I/O requests, tell the SPA to log the
3841 * error and generate a logical data ereport.
3843 spa_log_error(spa, zio);
3844 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3849 if (zio->io_error && zio == lio) {
3851 * Determine whether zio should be reexecuted. This will
3852 * propagate all the way to the root via zio_notify_parent().
3854 ASSERT(vd == NULL && bp != NULL);
3855 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3857 if (IO_IS_ALLOCATING(zio) &&
3858 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3859 if (zio->io_error != ENOSPC)
3860 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3862 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3865 if ((zio->io_type == ZIO_TYPE_READ ||
3866 zio->io_type == ZIO_TYPE_FREE) &&
3867 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3868 zio->io_error == ENXIO &&
3869 spa_load_state(spa) == SPA_LOAD_NONE &&
3870 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3871 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3873 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3874 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3877 * Here is a possibly good place to attempt to do
3878 * either combinatorial reconstruction or error correction
3879 * based on checksums. It also might be a good place
3880 * to send out preliminary ereports before we suspend
3886 * If there were logical child errors, they apply to us now.
3887 * We defer this until now to avoid conflating logical child
3888 * errors with errors that happened to the zio itself when
3889 * updating vdev stats and reporting FMA events above.
3891 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3893 if ((zio->io_error || zio->io_reexecute) &&
3894 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3895 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3896 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3898 zio_gang_tree_free(&zio->io_gang_tree);
3901 * Godfather I/Os should never suspend.
3903 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3904 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3905 zio->io_reexecute = 0;
3907 if (zio->io_reexecute) {
3909 * This is a logical I/O that wants to reexecute.
3911 * Reexecute is top-down. When an i/o fails, if it's not
3912 * the root, it simply notifies its parent and sticks around.
3913 * The parent, seeing that it still has children in zio_done(),
3914 * does the same. This percolates all the way up to the root.
3915 * The root i/o will reexecute or suspend the entire tree.
3917 * This approach ensures that zio_reexecute() honors
3918 * all the original i/o dependency relationships, e.g.
3919 * parents not executing until children are ready.
3921 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3923 zio->io_gang_leader = NULL;
3925 mutex_enter(&zio->io_lock);
3926 zio->io_state[ZIO_WAIT_DONE] = 1;
3927 mutex_exit(&zio->io_lock);
3930 * "The Godfather" I/O monitors its children but is
3931 * not a true parent to them. It will track them through
3932 * the pipeline but severs its ties whenever they get into
3933 * trouble (e.g. suspended). This allows "The Godfather"
3934 * I/O to return status without blocking.
3937 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3939 zio_link_t *remove_zl = zl;
3940 pio_next = zio_walk_parents(zio, &zl);
3942 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3943 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3944 zio_remove_child(pio, zio, remove_zl);
3945 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3949 if ((pio = zio_unique_parent(zio)) != NULL) {
3951 * We're not a root i/o, so there's nothing to do
3952 * but notify our parent. Don't propagate errors
3953 * upward since we haven't permanently failed yet.
3955 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3956 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3957 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3958 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3960 * We'd fail again if we reexecuted now, so suspend
3961 * until conditions improve (e.g. device comes online).
3963 zio_suspend(spa, zio);
3966 * Reexecution is potentially a huge amount of work.
3967 * Hand it off to the otherwise-unused claim taskq.
3969 #if defined(illumos) || !defined(_KERNEL)
3970 ASSERT(zio->io_tqent.tqent_next == NULL);
3972 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3974 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3975 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3978 return (ZIO_PIPELINE_STOP);
3981 ASSERT(zio->io_child_count == 0);
3982 ASSERT(zio->io_reexecute == 0);
3983 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3986 * Report any checksum errors, since the I/O is complete.
3988 while (zio->io_cksum_report != NULL) {
3989 zio_cksum_report_t *zcr = zio->io_cksum_report;
3990 zio->io_cksum_report = zcr->zcr_next;
3991 zcr->zcr_next = NULL;
3992 zcr->zcr_finish(zcr, NULL);
3993 zfs_ereport_free_checksum(zcr);
3997 * It is the responsibility of the done callback to ensure that this
3998 * particular zio is no longer discoverable for adoption, and as
3999 * such, cannot acquire any new parents.
4004 mutex_enter(&zio->io_lock);
4005 zio->io_state[ZIO_WAIT_DONE] = 1;
4006 mutex_exit(&zio->io_lock);
4009 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4010 zio_link_t *remove_zl = zl;
4011 pio_next = zio_walk_parents(zio, &zl);
4012 zio_remove_child(pio, zio, remove_zl);
4013 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4016 if (zio->io_waiter != NULL) {
4017 mutex_enter(&zio->io_lock);
4018 zio->io_executor = NULL;
4019 cv_broadcast(&zio->io_cv);
4020 mutex_exit(&zio->io_lock);
4025 return (ZIO_PIPELINE_STOP);
4029 * ==========================================================================
4030 * I/O pipeline definition
4031 * ==========================================================================
4033 static zio_pipe_stage_t *zio_pipeline[] = {
4040 zio_checksum_generate,
4056 zio_checksum_verify,
4064 * Compare two zbookmark_phys_t's to see which we would reach first in a
4065 * pre-order traversal of the object tree.
4067 * This is simple in every case aside from the meta-dnode object. For all other
4068 * objects, we traverse them in order (object 1 before object 2, and so on).
4069 * However, all of these objects are traversed while traversing object 0, since
4070 * the data it points to is the list of objects. Thus, we need to convert to a
4071 * canonical representation so we can compare meta-dnode bookmarks to
4072 * non-meta-dnode bookmarks.
4074 * We do this by calculating "equivalents" for each field of the zbookmark.
4075 * zbookmarks outside of the meta-dnode use their own object and level, and
4076 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4077 * blocks this bookmark refers to) by multiplying their blkid by their span
4078 * (the number of L0 blocks contained within one block at their level).
4079 * zbookmarks inside the meta-dnode calculate their object equivalent
4080 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4081 * level + 1<<31 (any value larger than a level could ever be) for their level.
4082 * This causes them to always compare before a bookmark in their object
4083 * equivalent, compare appropriately to bookmarks in other objects, and to
4084 * compare appropriately to other bookmarks in the meta-dnode.
4087 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4088 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4091 * These variables represent the "equivalent" values for the zbookmark,
4092 * after converting zbookmarks inside the meta dnode to their
4093 * normal-object equivalents.
4095 uint64_t zb1obj, zb2obj;
4096 uint64_t zb1L0, zb2L0;
4097 uint64_t zb1level, zb2level;
4099 if (zb1->zb_object == zb2->zb_object &&
4100 zb1->zb_level == zb2->zb_level &&
4101 zb1->zb_blkid == zb2->zb_blkid)
4105 * BP_SPANB calculates the span in blocks.
4107 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4108 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4110 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4111 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4113 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4115 zb1obj = zb1->zb_object;
4116 zb1level = zb1->zb_level;
4119 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4120 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4122 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4124 zb2obj = zb2->zb_object;
4125 zb2level = zb2->zb_level;
4128 /* Now that we have a canonical representation, do the comparison. */
4129 if (zb1obj != zb2obj)
4130 return (zb1obj < zb2obj ? -1 : 1);
4131 else if (zb1L0 != zb2L0)
4132 return (zb1L0 < zb2L0 ? -1 : 1);
4133 else if (zb1level != zb2level)
4134 return (zb1level > zb2level ? -1 : 1);
4136 * This can (theoretically) happen if the bookmarks have the same object
4137 * and level, but different blkids, if the block sizes are not the same.
4138 * There is presently no way to change the indirect block sizes
4144 * This function checks the following: given that last_block is the place that
4145 * our traversal stopped last time, does that guarantee that we've visited
4146 * every node under subtree_root? Therefore, we can't just use the raw output
4147 * of zbookmark_compare. We have to pass in a modified version of
4148 * subtree_root; by incrementing the block id, and then checking whether
4149 * last_block is before or equal to that, we can tell whether or not having
4150 * visited last_block implies that all of subtree_root's children have been
4154 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4155 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4157 zbookmark_phys_t mod_zb = *subtree_root;
4159 ASSERT(last_block->zb_level == 0);
4161 /* The objset_phys_t isn't before anything. */
4166 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4167 * data block size in sectors, because that variable is only used if
4168 * the bookmark refers to a block in the meta-dnode. Since we don't
4169 * know without examining it what object it refers to, and there's no
4170 * harm in passing in this value in other cases, we always pass it in.
4172 * We pass in 0 for the indirect block size shift because zb2 must be
4173 * level 0. The indirect block size is only used to calculate the span
4174 * of the bookmark, but since the bookmark must be level 0, the span is
4175 * always 1, so the math works out.
4177 * If you make changes to how the zbookmark_compare code works, be sure
4178 * to make sure that this code still works afterwards.
4180 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4181 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,