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, enum zio_child child, enum zio_wait_type wait)
517 uint64_t *countp = &zio->io_children[child][wait];
518 boolean_t waiting = B_FALSE;
520 mutex_enter(&zio->io_lock);
521 ASSERT(zio->io_stall == NULL);
524 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
525 zio->io_stall = countp;
528 mutex_exit(&zio->io_lock);
534 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
536 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
537 int *errorp = &pio->io_child_error[zio->io_child_type];
539 mutex_enter(&pio->io_lock);
540 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
541 *errorp = zio_worst_error(*errorp, zio->io_error);
542 pio->io_reexecute |= zio->io_reexecute;
543 ASSERT3U(*countp, >, 0);
547 if (*countp == 0 && pio->io_stall == countp) {
548 zio_taskq_type_t type =
549 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
551 pio->io_stall = NULL;
552 mutex_exit(&pio->io_lock);
554 * Dispatch the parent zio in its own taskq so that
555 * the child can continue to make progress. This also
556 * prevents overflowing the stack when we have deeply nested
557 * parent-child relationships.
559 zio_taskq_dispatch(pio, type, B_FALSE);
561 mutex_exit(&pio->io_lock);
566 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
568 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
569 zio->io_error = zio->io_child_error[c];
573 zio_bookmark_compare(const void *x1, const void *x2)
575 const zio_t *z1 = x1;
576 const zio_t *z2 = x2;
578 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
580 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
583 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
585 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
588 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
590 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
593 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
595 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
607 * ==========================================================================
608 * Create the various types of I/O (read, write, free, etc)
609 * ==========================================================================
612 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
613 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
614 void *private, zio_type_t type, zio_priority_t priority,
615 enum zio_flag flags, vdev_t *vd, uint64_t offset,
616 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
620 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
621 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
622 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
624 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
625 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
626 ASSERT(vd || stage == ZIO_STAGE_OPEN);
628 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
630 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
631 bzero(zio, sizeof (zio_t));
633 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
634 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
636 list_create(&zio->io_parent_list, sizeof (zio_link_t),
637 offsetof(zio_link_t, zl_parent_node));
638 list_create(&zio->io_child_list, sizeof (zio_link_t),
639 offsetof(zio_link_t, zl_child_node));
640 metaslab_trace_init(&zio->io_alloc_list);
643 zio->io_child_type = ZIO_CHILD_VDEV;
644 else if (flags & ZIO_FLAG_GANG_CHILD)
645 zio->io_child_type = ZIO_CHILD_GANG;
646 else if (flags & ZIO_FLAG_DDT_CHILD)
647 zio->io_child_type = ZIO_CHILD_DDT;
649 zio->io_child_type = ZIO_CHILD_LOGICAL;
652 zio->io_bp = (blkptr_t *)bp;
653 zio->io_bp_copy = *bp;
654 zio->io_bp_orig = *bp;
655 if (type != ZIO_TYPE_WRITE ||
656 zio->io_child_type == ZIO_CHILD_DDT)
657 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
658 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
659 zio->io_logical = zio;
660 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
661 pipeline |= ZIO_GANG_STAGES;
667 zio->io_private = private;
669 zio->io_priority = priority;
671 zio->io_offset = offset;
672 zio->io_orig_abd = zio->io_abd = data;
673 zio->io_orig_size = zio->io_size = psize;
674 zio->io_lsize = lsize;
675 zio->io_orig_flags = zio->io_flags = flags;
676 zio->io_orig_stage = zio->io_stage = stage;
677 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
678 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
680 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
681 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
684 zio->io_bookmark = *zb;
687 if (zio->io_logical == NULL)
688 zio->io_logical = pio->io_logical;
689 if (zio->io_child_type == ZIO_CHILD_GANG)
690 zio->io_gang_leader = pio->io_gang_leader;
691 zio_add_child(pio, zio);
698 zio_destroy(zio_t *zio)
700 metaslab_trace_fini(&zio->io_alloc_list);
701 list_destroy(&zio->io_parent_list);
702 list_destroy(&zio->io_child_list);
703 mutex_destroy(&zio->io_lock);
704 cv_destroy(&zio->io_cv);
705 kmem_cache_free(zio_cache, zio);
709 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
710 void *private, enum zio_flag flags)
714 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
715 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
716 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
722 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
724 return (zio_null(NULL, spa, NULL, done, private, flags));
728 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
730 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
731 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
732 bp, (longlong_t)BP_GET_TYPE(bp));
734 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
735 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
736 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
737 bp, (longlong_t)BP_GET_CHECKSUM(bp));
739 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
740 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
741 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
742 bp, (longlong_t)BP_GET_COMPRESS(bp));
744 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
745 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
746 bp, (longlong_t)BP_GET_LSIZE(bp));
748 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
749 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
750 bp, (longlong_t)BP_GET_PSIZE(bp));
753 if (BP_IS_EMBEDDED(bp)) {
754 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
755 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
756 bp, (longlong_t)BPE_GET_ETYPE(bp));
761 * Pool-specific checks.
763 * Note: it would be nice to verify that the blk_birth and
764 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
765 * allows the birth time of log blocks (and dmu_sync()-ed blocks
766 * that are in the log) to be arbitrarily large.
768 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
769 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
770 if (vdevid >= spa->spa_root_vdev->vdev_children) {
771 zfs_panic_recover("blkptr at %p DVA %u has invalid "
773 bp, i, (longlong_t)vdevid);
776 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
778 zfs_panic_recover("blkptr at %p DVA %u has invalid "
780 bp, i, (longlong_t)vdevid);
783 if (vd->vdev_ops == &vdev_hole_ops) {
784 zfs_panic_recover("blkptr at %p DVA %u has hole "
786 bp, i, (longlong_t)vdevid);
789 if (vd->vdev_ops == &vdev_missing_ops) {
791 * "missing" vdevs are valid during import, but we
792 * don't have their detailed info (e.g. asize), so
793 * we can't perform any more checks on them.
797 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
798 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
800 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
801 if (offset + asize > vd->vdev_asize) {
802 zfs_panic_recover("blkptr at %p DVA %u has invalid "
804 bp, i, (longlong_t)offset);
810 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
811 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
812 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
816 zfs_blkptr_verify(spa, bp);
818 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
819 data, size, size, done, private,
820 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
821 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
822 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
828 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
829 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
830 zio_done_func_t *ready, zio_done_func_t *children_ready,
831 zio_done_func_t *physdone, zio_done_func_t *done,
832 void *private, zio_priority_t priority, enum zio_flag flags,
833 const zbookmark_phys_t *zb)
837 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
838 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
839 zp->zp_compress >= ZIO_COMPRESS_OFF &&
840 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
841 DMU_OT_IS_VALID(zp->zp_type) &&
844 zp->zp_copies <= spa_max_replication(spa));
846 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
847 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
848 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
849 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
851 zio->io_ready = ready;
852 zio->io_children_ready = children_ready;
853 zio->io_physdone = physdone;
857 * Data can be NULL if we are going to call zio_write_override() to
858 * provide the already-allocated BP. But we may need the data to
859 * verify a dedup hit (if requested). In this case, don't try to
860 * dedup (just take the already-allocated BP verbatim).
862 if (data == NULL && zio->io_prop.zp_dedup_verify) {
863 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
870 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
871 uint64_t size, zio_done_func_t *done, void *private,
872 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
876 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
877 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
878 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
884 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
886 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
887 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
888 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
889 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
892 * We must reset the io_prop to match the values that existed
893 * when the bp was first written by dmu_sync() keeping in mind
894 * that nopwrite and dedup are mutually exclusive.
896 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
897 zio->io_prop.zp_nopwrite = nopwrite;
898 zio->io_prop.zp_copies = copies;
899 zio->io_bp_override = bp;
903 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
907 * The check for EMBEDDED is a performance optimization. We
908 * process the free here (by ignoring it) rather than
909 * putting it on the list and then processing it in zio_free_sync().
911 if (BP_IS_EMBEDDED(bp))
913 metaslab_check_free(spa, bp);
916 * Frees that are for the currently-syncing txg, are not going to be
917 * deferred, and which will not need to do a read (i.e. not GANG or
918 * DEDUP), can be processed immediately. Otherwise, put them on the
919 * in-memory list for later processing.
921 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
922 txg != spa->spa_syncing_txg ||
923 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
924 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
926 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
927 BP_GET_PSIZE(bp), 0)));
932 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
933 uint64_t size, enum zio_flag flags)
936 enum zio_stage stage = ZIO_FREE_PIPELINE;
938 ASSERT(!BP_IS_HOLE(bp));
939 ASSERT(spa_syncing_txg(spa) == txg);
940 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
942 if (BP_IS_EMBEDDED(bp))
943 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
945 metaslab_check_free(spa, bp);
948 if (zfs_trim_enabled)
949 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
950 ZIO_STAGE_VDEV_IO_ASSESS;
952 * GANG and DEDUP blocks can induce a read (for the gang block header,
953 * or the DDT), so issue them asynchronously so that this thread is
956 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
957 stage |= ZIO_STAGE_ISSUE_ASYNC;
959 flags |= ZIO_FLAG_DONT_QUEUE;
961 zio = zio_create(pio, spa, txg, bp, NULL, size,
962 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
963 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
969 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
970 zio_done_func_t *done, void *private, enum zio_flag flags)
974 dprintf_bp(bp, "claiming in txg %llu", txg);
976 if (BP_IS_EMBEDDED(bp))
977 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
980 * A claim is an allocation of a specific block. Claims are needed
981 * to support immediate writes in the intent log. The issue is that
982 * immediate writes contain committed data, but in a txg that was
983 * *not* committed. Upon opening the pool after an unclean shutdown,
984 * the intent log claims all blocks that contain immediate write data
985 * so that the SPA knows they're in use.
987 * All claims *must* be resolved in the first txg -- before the SPA
988 * starts allocating blocks -- so that nothing is allocated twice.
989 * If txg == 0 we just verify that the block is claimable.
991 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
992 ASSERT(txg == spa_first_txg(spa) || txg == 0);
993 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
995 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
996 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
997 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
998 ASSERT0(zio->io_queued_timestamp);
1004 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1005 uint64_t size, zio_done_func_t *done, void *private,
1006 zio_priority_t priority, enum zio_flag flags)
1011 if (vd->vdev_children == 0) {
1012 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1013 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1014 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1018 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1020 for (c = 0; c < vd->vdev_children; c++)
1021 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1022 offset, size, done, private, priority, flags));
1029 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1030 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1031 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1035 ASSERT(vd->vdev_children == 0);
1036 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1037 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1038 ASSERT3U(offset + size, <=, vd->vdev_psize);
1040 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1041 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1042 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1044 zio->io_prop.zp_checksum = checksum;
1050 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1051 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1052 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1056 ASSERT(vd->vdev_children == 0);
1057 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1058 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1059 ASSERT3U(offset + size, <=, vd->vdev_psize);
1061 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1062 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1063 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1065 zio->io_prop.zp_checksum = checksum;
1067 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1069 * zec checksums are necessarily destructive -- they modify
1070 * the end of the write buffer to hold the verifier/checksum.
1071 * Therefore, we must make a local copy in case the data is
1072 * being written to multiple places in parallel.
1074 abd_t *wbuf = abd_alloc_sametype(data, size);
1075 abd_copy(wbuf, data, size);
1077 zio_push_transform(zio, wbuf, size, size, NULL);
1084 * Create a child I/O to do some work for us.
1087 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1088 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1089 enum zio_flag flags, zio_done_func_t *done, void *private)
1091 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1094 ASSERT(vd->vdev_parent ==
1095 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1097 if (type == ZIO_TYPE_READ && bp != NULL) {
1099 * If we have the bp, then the child should perform the
1100 * checksum and the parent need not. This pushes error
1101 * detection as close to the leaves as possible and
1102 * eliminates redundant checksums in the interior nodes.
1104 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1105 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1108 /* Not all IO types require vdev io done stage e.g. free */
1109 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1110 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1112 if (vd->vdev_children == 0)
1113 offset += VDEV_LABEL_START_SIZE;
1115 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1118 * If we've decided to do a repair, the write is not speculative --
1119 * even if the original read was.
1121 if (flags & ZIO_FLAG_IO_REPAIR)
1122 flags &= ~ZIO_FLAG_SPECULATIVE;
1125 * If we're creating a child I/O that is not associated with a
1126 * top-level vdev, then the child zio is not an allocating I/O.
1127 * If this is a retried I/O then we ignore it since we will
1128 * have already processed the original allocating I/O.
1130 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1131 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1132 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1134 ASSERT(mc->mc_alloc_throttle_enabled);
1135 ASSERT(type == ZIO_TYPE_WRITE);
1136 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1137 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1138 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1139 pio->io_child_type == ZIO_CHILD_GANG);
1141 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1144 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1145 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1146 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1147 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1149 zio->io_physdone = pio->io_physdone;
1150 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1151 zio->io_logical->io_phys_children++;
1157 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1158 int type, zio_priority_t priority, enum zio_flag flags,
1159 zio_done_func_t *done, void *private)
1163 ASSERT(vd->vdev_ops->vdev_op_leaf);
1165 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1166 data, size, size, done, private, type, priority,
1167 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1169 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1175 zio_flush(zio_t *zio, vdev_t *vd)
1177 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1178 NULL, NULL, ZIO_PRIORITY_NOW,
1179 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1183 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1186 ASSERT(vd->vdev_ops->vdev_op_leaf);
1188 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1189 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1190 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1191 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1195 zio_shrink(zio_t *zio, uint64_t size)
1197 ASSERT3P(zio->io_executor, ==, NULL);
1198 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1199 ASSERT3U(size, <=, zio->io_size);
1202 * We don't shrink for raidz because of problems with the
1203 * reconstruction when reading back less than the block size.
1204 * Note, BP_IS_RAIDZ() assumes no compression.
1206 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1207 if (!BP_IS_RAIDZ(zio->io_bp)) {
1208 /* we are not doing a raw write */
1209 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1210 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1215 * ==========================================================================
1216 * Prepare to read and write logical blocks
1217 * ==========================================================================
1221 zio_read_bp_init(zio_t *zio)
1223 blkptr_t *bp = zio->io_bp;
1225 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1226 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1227 !(zio->io_flags & ZIO_FLAG_RAW)) {
1229 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1230 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1231 psize, psize, zio_decompress);
1234 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1235 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1237 int psize = BPE_GET_PSIZE(bp);
1238 void *data = abd_borrow_buf(zio->io_abd, psize);
1239 decode_embedded_bp_compressed(bp, data);
1240 abd_return_buf_copy(zio->io_abd, data, psize);
1242 ASSERT(!BP_IS_EMBEDDED(bp));
1245 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1246 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1248 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1249 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1251 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1252 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1254 return (ZIO_PIPELINE_CONTINUE);
1258 zio_write_bp_init(zio_t *zio)
1260 if (!IO_IS_ALLOCATING(zio))
1261 return (ZIO_PIPELINE_CONTINUE);
1263 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1265 if (zio->io_bp_override) {
1266 blkptr_t *bp = zio->io_bp;
1267 zio_prop_t *zp = &zio->io_prop;
1269 ASSERT(bp->blk_birth != zio->io_txg);
1270 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1272 *bp = *zio->io_bp_override;
1273 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1275 if (BP_IS_EMBEDDED(bp))
1276 return (ZIO_PIPELINE_CONTINUE);
1279 * If we've been overridden and nopwrite is set then
1280 * set the flag accordingly to indicate that a nopwrite
1281 * has already occurred.
1283 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1284 ASSERT(!zp->zp_dedup);
1285 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1286 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1287 return (ZIO_PIPELINE_CONTINUE);
1290 ASSERT(!zp->zp_nopwrite);
1292 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1293 return (ZIO_PIPELINE_CONTINUE);
1295 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1296 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1298 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1299 BP_SET_DEDUP(bp, 1);
1300 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1301 return (ZIO_PIPELINE_CONTINUE);
1305 * We were unable to handle this as an override bp, treat
1306 * it as a regular write I/O.
1308 zio->io_bp_override = NULL;
1309 *bp = zio->io_bp_orig;
1310 zio->io_pipeline = zio->io_orig_pipeline;
1313 return (ZIO_PIPELINE_CONTINUE);
1317 zio_write_compress(zio_t *zio)
1319 spa_t *spa = zio->io_spa;
1320 zio_prop_t *zp = &zio->io_prop;
1321 enum zio_compress compress = zp->zp_compress;
1322 blkptr_t *bp = zio->io_bp;
1323 uint64_t lsize = zio->io_lsize;
1324 uint64_t psize = zio->io_size;
1327 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1330 * If our children haven't all reached the ready stage,
1331 * wait for them and then repeat this pipeline stage.
1333 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1334 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1335 return (ZIO_PIPELINE_STOP);
1337 if (!IO_IS_ALLOCATING(zio))
1338 return (ZIO_PIPELINE_CONTINUE);
1340 if (zio->io_children_ready != NULL) {
1342 * Now that all our children are ready, run the callback
1343 * associated with this zio in case it wants to modify the
1344 * data to be written.
1346 ASSERT3U(zp->zp_level, >, 0);
1347 zio->io_children_ready(zio);
1350 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1351 ASSERT(zio->io_bp_override == NULL);
1353 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1355 * We're rewriting an existing block, which means we're
1356 * working on behalf of spa_sync(). For spa_sync() to
1357 * converge, it must eventually be the case that we don't
1358 * have to allocate new blocks. But compression changes
1359 * the blocksize, which forces a reallocate, and makes
1360 * convergence take longer. Therefore, after the first
1361 * few passes, stop compressing to ensure convergence.
1363 pass = spa_sync_pass(spa);
1365 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1366 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1367 ASSERT(!BP_GET_DEDUP(bp));
1369 if (pass >= zfs_sync_pass_dont_compress)
1370 compress = ZIO_COMPRESS_OFF;
1372 /* Make sure someone doesn't change their mind on overwrites */
1373 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1374 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1377 /* If it's a compressed write that is not raw, compress the buffer. */
1378 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1379 void *cbuf = zio_buf_alloc(lsize);
1380 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1381 if (psize == 0 || psize == lsize) {
1382 compress = ZIO_COMPRESS_OFF;
1383 zio_buf_free(cbuf, lsize);
1384 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1385 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1386 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1387 encode_embedded_bp_compressed(bp,
1388 cbuf, compress, lsize, psize);
1389 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1390 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1391 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1392 zio_buf_free(cbuf, lsize);
1393 bp->blk_birth = zio->io_txg;
1394 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1395 ASSERT(spa_feature_is_active(spa,
1396 SPA_FEATURE_EMBEDDED_DATA));
1397 return (ZIO_PIPELINE_CONTINUE);
1400 * Round up compressed size up to the ashift
1401 * of the smallest-ashift device, and zero the tail.
1402 * This ensures that the compressed size of the BP
1403 * (and thus compressratio property) are correct,
1404 * in that we charge for the padding used to fill out
1407 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1408 size_t rounded = (size_t)P2ROUNDUP(psize,
1409 1ULL << spa->spa_min_ashift);
1410 if (rounded >= lsize) {
1411 compress = ZIO_COMPRESS_OFF;
1412 zio_buf_free(cbuf, lsize);
1415 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1416 abd_take_ownership_of_buf(cdata, B_TRUE);
1417 abd_zero_off(cdata, psize, rounded - psize);
1419 zio_push_transform(zio, cdata,
1420 psize, lsize, NULL);
1425 * We were unable to handle this as an override bp, treat
1426 * it as a regular write I/O.
1428 zio->io_bp_override = NULL;
1429 *bp = zio->io_bp_orig;
1430 zio->io_pipeline = zio->io_orig_pipeline;
1432 ASSERT3U(psize, !=, 0);
1436 * The final pass of spa_sync() must be all rewrites, but the first
1437 * few passes offer a trade-off: allocating blocks defers convergence,
1438 * but newly allocated blocks are sequential, so they can be written
1439 * to disk faster. Therefore, we allow the first few passes of
1440 * spa_sync() to allocate new blocks, but force rewrites after that.
1441 * There should only be a handful of blocks after pass 1 in any case.
1443 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1444 BP_GET_PSIZE(bp) == psize &&
1445 pass >= zfs_sync_pass_rewrite) {
1447 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1448 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1449 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1452 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1456 if (zio->io_bp_orig.blk_birth != 0 &&
1457 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1458 BP_SET_LSIZE(bp, lsize);
1459 BP_SET_TYPE(bp, zp->zp_type);
1460 BP_SET_LEVEL(bp, zp->zp_level);
1461 BP_SET_BIRTH(bp, zio->io_txg, 0);
1463 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1465 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1466 BP_SET_LSIZE(bp, lsize);
1467 BP_SET_TYPE(bp, zp->zp_type);
1468 BP_SET_LEVEL(bp, zp->zp_level);
1469 BP_SET_PSIZE(bp, psize);
1470 BP_SET_COMPRESS(bp, compress);
1471 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1472 BP_SET_DEDUP(bp, zp->zp_dedup);
1473 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1475 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1477 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1479 if (zp->zp_nopwrite) {
1480 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1481 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1482 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1485 return (ZIO_PIPELINE_CONTINUE);
1489 zio_free_bp_init(zio_t *zio)
1491 blkptr_t *bp = zio->io_bp;
1493 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1494 if (BP_GET_DEDUP(bp))
1495 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1498 return (ZIO_PIPELINE_CONTINUE);
1502 * ==========================================================================
1503 * Execute the I/O pipeline
1504 * ==========================================================================
1508 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1510 spa_t *spa = zio->io_spa;
1511 zio_type_t t = zio->io_type;
1512 int flags = (cutinline ? TQ_FRONT : 0);
1514 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1517 * If we're a config writer or a probe, the normal issue and
1518 * interrupt threads may all be blocked waiting for the config lock.
1519 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1521 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1525 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1527 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1531 * If this is a high priority I/O, then use the high priority taskq if
1534 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1535 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1538 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1541 * NB: We are assuming that the zio can only be dispatched
1542 * to a single taskq at a time. It would be a grievous error
1543 * to dispatch the zio to another taskq at the same time.
1545 #if defined(illumos) || !defined(_KERNEL)
1546 ASSERT(zio->io_tqent.tqent_next == NULL);
1548 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1550 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1551 flags, &zio->io_tqent);
1555 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1557 kthread_t *executor = zio->io_executor;
1558 spa_t *spa = zio->io_spa;
1560 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1561 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1563 for (i = 0; i < tqs->stqs_count; i++) {
1564 if (taskq_member(tqs->stqs_taskq[i], executor))
1573 zio_issue_async(zio_t *zio)
1575 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1577 return (ZIO_PIPELINE_STOP);
1581 zio_interrupt(zio_t *zio)
1583 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1587 zio_delay_interrupt(zio_t *zio)
1590 * The timeout_generic() function isn't defined in userspace, so
1591 * rather than trying to implement the function, the zio delay
1592 * functionality has been disabled for userspace builds.
1597 * If io_target_timestamp is zero, then no delay has been registered
1598 * for this IO, thus jump to the end of this function and "skip" the
1599 * delay; issuing it directly to the zio layer.
1601 if (zio->io_target_timestamp != 0) {
1602 hrtime_t now = gethrtime();
1604 if (now >= zio->io_target_timestamp) {
1606 * This IO has already taken longer than the target
1607 * delay to complete, so we don't want to delay it
1608 * any longer; we "miss" the delay and issue it
1609 * directly to the zio layer. This is likely due to
1610 * the target latency being set to a value less than
1611 * the underlying hardware can satisfy (e.g. delay
1612 * set to 1ms, but the disks take 10ms to complete an
1616 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1621 hrtime_t diff = zio->io_target_timestamp - now;
1623 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1624 hrtime_t, now, hrtime_t, diff);
1626 (void) timeout_generic(CALLOUT_NORMAL,
1627 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1634 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1639 * Execute the I/O pipeline until one of the following occurs:
1641 * (1) the I/O completes
1642 * (2) the pipeline stalls waiting for dependent child I/Os
1643 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1644 * (4) the I/O is delegated by vdev-level caching or aggregation
1645 * (5) the I/O is deferred due to vdev-level queueing
1646 * (6) the I/O is handed off to another thread.
1648 * In all cases, the pipeline stops whenever there's no CPU work; it never
1649 * burns a thread in cv_wait().
1651 * There's no locking on io_stage because there's no legitimate way
1652 * for multiple threads to be attempting to process the same I/O.
1654 static zio_pipe_stage_t *zio_pipeline[];
1657 zio_execute(zio_t *zio)
1659 zio->io_executor = curthread;
1661 ASSERT3U(zio->io_queued_timestamp, >, 0);
1663 while (zio->io_stage < ZIO_STAGE_DONE) {
1664 enum zio_stage pipeline = zio->io_pipeline;
1665 enum zio_stage stage = zio->io_stage;
1668 ASSERT(!MUTEX_HELD(&zio->io_lock));
1669 ASSERT(ISP2(stage));
1670 ASSERT(zio->io_stall == NULL);
1674 } while ((stage & pipeline) == 0);
1676 ASSERT(stage <= ZIO_STAGE_DONE);
1679 * If we are in interrupt context and this pipeline stage
1680 * will grab a config lock that is held across I/O,
1681 * or may wait for an I/O that needs an interrupt thread
1682 * to complete, issue async to avoid deadlock.
1684 * For VDEV_IO_START, we cut in line so that the io will
1685 * be sent to disk promptly.
1687 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1688 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1689 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1690 zio_requeue_io_start_cut_in_line : B_FALSE;
1691 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1695 zio->io_stage = stage;
1696 zio->io_pipeline_trace |= zio->io_stage;
1697 rv = zio_pipeline[highbit64(stage) - 1](zio);
1699 if (rv == ZIO_PIPELINE_STOP)
1702 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1707 * ==========================================================================
1708 * Initiate I/O, either sync or async
1709 * ==========================================================================
1712 zio_wait(zio_t *zio)
1716 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1717 ASSERT3P(zio->io_executor, ==, NULL);
1719 zio->io_waiter = curthread;
1720 ASSERT0(zio->io_queued_timestamp);
1721 zio->io_queued_timestamp = gethrtime();
1725 mutex_enter(&zio->io_lock);
1726 while (zio->io_executor != NULL)
1727 cv_wait(&zio->io_cv, &zio->io_lock);
1728 mutex_exit(&zio->io_lock);
1730 error = zio->io_error;
1737 zio_nowait(zio_t *zio)
1739 ASSERT3P(zio->io_executor, ==, NULL);
1741 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1742 zio_unique_parent(zio) == NULL) {
1744 * This is a logical async I/O with no parent to wait for it.
1745 * We add it to the spa_async_root_zio "Godfather" I/O which
1746 * will ensure they complete prior to unloading the pool.
1748 spa_t *spa = zio->io_spa;
1750 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1753 ASSERT0(zio->io_queued_timestamp);
1754 zio->io_queued_timestamp = gethrtime();
1759 * ==========================================================================
1760 * Reexecute, cancel, or suspend/resume failed I/O
1761 * ==========================================================================
1765 zio_reexecute(zio_t *pio)
1767 zio_t *cio, *cio_next;
1769 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1770 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1771 ASSERT(pio->io_gang_leader == NULL);
1772 ASSERT(pio->io_gang_tree == NULL);
1774 pio->io_flags = pio->io_orig_flags;
1775 pio->io_stage = pio->io_orig_stage;
1776 pio->io_pipeline = pio->io_orig_pipeline;
1777 pio->io_reexecute = 0;
1778 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1779 pio->io_pipeline_trace = 0;
1781 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1782 pio->io_state[w] = 0;
1783 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1784 pio->io_child_error[c] = 0;
1786 if (IO_IS_ALLOCATING(pio))
1787 BP_ZERO(pio->io_bp);
1790 * As we reexecute pio's children, new children could be created.
1791 * New children go to the head of pio's io_child_list, however,
1792 * so we will (correctly) not reexecute them. The key is that
1793 * the remainder of pio's io_child_list, from 'cio_next' onward,
1794 * cannot be affected by any side effects of reexecuting 'cio'.
1796 zio_link_t *zl = NULL;
1797 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1798 cio_next = zio_walk_children(pio, &zl);
1799 mutex_enter(&pio->io_lock);
1800 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1801 pio->io_children[cio->io_child_type][w]++;
1802 mutex_exit(&pio->io_lock);
1807 * Now that all children have been reexecuted, execute the parent.
1808 * We don't reexecute "The Godfather" I/O here as it's the
1809 * responsibility of the caller to wait on him.
1811 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1812 pio->io_queued_timestamp = gethrtime();
1818 zio_cancel(zio_t *zio)
1821 * Disallow cancellation of a zio that's already been issued.
1823 VERIFY3P(zio->io_executor, ==, NULL);
1825 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1826 zio->io_done = NULL;
1832 zio_suspend(spa_t *spa, zio_t *zio)
1834 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1835 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1836 "failure and the failure mode property for this pool "
1837 "is set to panic.", spa_name(spa));
1839 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1841 mutex_enter(&spa->spa_suspend_lock);
1843 if (spa->spa_suspend_zio_root == NULL)
1844 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1845 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1846 ZIO_FLAG_GODFATHER);
1848 spa->spa_suspended = B_TRUE;
1851 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1852 ASSERT(zio != spa->spa_suspend_zio_root);
1853 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1854 ASSERT(zio_unique_parent(zio) == NULL);
1855 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1856 zio_add_child(spa->spa_suspend_zio_root, zio);
1859 mutex_exit(&spa->spa_suspend_lock);
1863 zio_resume(spa_t *spa)
1868 * Reexecute all previously suspended i/o.
1870 mutex_enter(&spa->spa_suspend_lock);
1871 spa->spa_suspended = B_FALSE;
1872 cv_broadcast(&spa->spa_suspend_cv);
1873 pio = spa->spa_suspend_zio_root;
1874 spa->spa_suspend_zio_root = NULL;
1875 mutex_exit(&spa->spa_suspend_lock);
1881 return (zio_wait(pio));
1885 zio_resume_wait(spa_t *spa)
1887 mutex_enter(&spa->spa_suspend_lock);
1888 while (spa_suspended(spa))
1889 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1890 mutex_exit(&spa->spa_suspend_lock);
1894 * ==========================================================================
1897 * A gang block is a collection of small blocks that looks to the DMU
1898 * like one large block. When zio_dva_allocate() cannot find a block
1899 * of the requested size, due to either severe fragmentation or the pool
1900 * being nearly full, it calls zio_write_gang_block() to construct the
1901 * block from smaller fragments.
1903 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1904 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1905 * an indirect block: it's an array of block pointers. It consumes
1906 * only one sector and hence is allocatable regardless of fragmentation.
1907 * The gang header's bps point to its gang members, which hold the data.
1909 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1910 * as the verifier to ensure uniqueness of the SHA256 checksum.
1911 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1912 * not the gang header. This ensures that data block signatures (needed for
1913 * deduplication) are independent of how the block is physically stored.
1915 * Gang blocks can be nested: a gang member may itself be a gang block.
1916 * Thus every gang block is a tree in which root and all interior nodes are
1917 * gang headers, and the leaves are normal blocks that contain user data.
1918 * The root of the gang tree is called the gang leader.
1920 * To perform any operation (read, rewrite, free, claim) on a gang block,
1921 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1922 * in the io_gang_tree field of the original logical i/o by recursively
1923 * reading the gang leader and all gang headers below it. This yields
1924 * an in-core tree containing the contents of every gang header and the
1925 * bps for every constituent of the gang block.
1927 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1928 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1929 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1930 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1931 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1932 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1933 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1934 * of the gang header plus zio_checksum_compute() of the data to update the
1935 * gang header's blk_cksum as described above.
1937 * The two-phase assemble/issue model solves the problem of partial failure --
1938 * what if you'd freed part of a gang block but then couldn't read the
1939 * gang header for another part? Assembling the entire gang tree first
1940 * ensures that all the necessary gang header I/O has succeeded before
1941 * starting the actual work of free, claim, or write. Once the gang tree
1942 * is assembled, free and claim are in-memory operations that cannot fail.
1944 * In the event that a gang write fails, zio_dva_unallocate() walks the
1945 * gang tree to immediately free (i.e. insert back into the space map)
1946 * everything we've allocated. This ensures that we don't get ENOSPC
1947 * errors during repeated suspend/resume cycles due to a flaky device.
1949 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1950 * the gang tree, we won't modify the block, so we can safely defer the free
1951 * (knowing that the block is still intact). If we *can* assemble the gang
1952 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1953 * each constituent bp and we can allocate a new block on the next sync pass.
1955 * In all cases, the gang tree allows complete recovery from partial failure.
1956 * ==========================================================================
1960 zio_gang_issue_func_done(zio_t *zio)
1962 abd_put(zio->io_abd);
1966 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1972 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
1973 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
1974 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1975 &pio->io_bookmark));
1979 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1986 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1987 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1988 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
1989 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1992 * As we rewrite each gang header, the pipeline will compute
1993 * a new gang block header checksum for it; but no one will
1994 * compute a new data checksum, so we do that here. The one
1995 * exception is the gang leader: the pipeline already computed
1996 * its data checksum because that stage precedes gang assembly.
1997 * (Presently, nothing actually uses interior data checksums;
1998 * this is just good hygiene.)
2000 if (gn != pio->io_gang_leader->io_gang_tree) {
2001 abd_t *buf = abd_get_offset(data, offset);
2003 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2004 buf, BP_GET_PSIZE(bp));
2009 * If we are here to damage data for testing purposes,
2010 * leave the GBH alone so that we can detect the damage.
2012 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2013 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2015 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2016 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2017 zio_gang_issue_func_done, NULL, pio->io_priority,
2018 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2026 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2029 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2030 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2031 ZIO_GANG_CHILD_FLAGS(pio)));
2036 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2039 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2040 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2043 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2052 static void zio_gang_tree_assemble_done(zio_t *zio);
2054 static zio_gang_node_t *
2055 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2057 zio_gang_node_t *gn;
2059 ASSERT(*gnpp == NULL);
2061 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2062 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2069 zio_gang_node_free(zio_gang_node_t **gnpp)
2071 zio_gang_node_t *gn = *gnpp;
2073 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2074 ASSERT(gn->gn_child[g] == NULL);
2076 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2077 kmem_free(gn, sizeof (*gn));
2082 zio_gang_tree_free(zio_gang_node_t **gnpp)
2084 zio_gang_node_t *gn = *gnpp;
2089 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2090 zio_gang_tree_free(&gn->gn_child[g]);
2092 zio_gang_node_free(gnpp);
2096 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2098 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2099 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2101 ASSERT(gio->io_gang_leader == gio);
2102 ASSERT(BP_IS_GANG(bp));
2104 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2105 zio_gang_tree_assemble_done, gn, gio->io_priority,
2106 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2110 zio_gang_tree_assemble_done(zio_t *zio)
2112 zio_t *gio = zio->io_gang_leader;
2113 zio_gang_node_t *gn = zio->io_private;
2114 blkptr_t *bp = zio->io_bp;
2116 ASSERT(gio == zio_unique_parent(zio));
2117 ASSERT(zio->io_child_count == 0);
2122 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2123 if (BP_SHOULD_BYTESWAP(bp))
2124 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2126 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2127 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2128 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2130 abd_put(zio->io_abd);
2132 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2133 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2134 if (!BP_IS_GANG(gbp))
2136 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2141 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2144 zio_t *gio = pio->io_gang_leader;
2147 ASSERT(BP_IS_GANG(bp) == !!gn);
2148 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2149 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2152 * If you're a gang header, your data is in gn->gn_gbh.
2153 * If you're a gang member, your data is in 'data' and gn == NULL.
2155 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2158 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2160 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2161 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2162 if (BP_IS_HOLE(gbp))
2164 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2166 offset += BP_GET_PSIZE(gbp);
2170 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2171 ASSERT3U(gio->io_size, ==, offset);
2178 zio_gang_assemble(zio_t *zio)
2180 blkptr_t *bp = zio->io_bp;
2182 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2183 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2185 zio->io_gang_leader = zio;
2187 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2189 return (ZIO_PIPELINE_CONTINUE);
2193 zio_gang_issue(zio_t *zio)
2195 blkptr_t *bp = zio->io_bp;
2197 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2198 return (ZIO_PIPELINE_STOP);
2200 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2201 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2203 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2204 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2207 zio_gang_tree_free(&zio->io_gang_tree);
2209 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2211 return (ZIO_PIPELINE_CONTINUE);
2215 zio_write_gang_member_ready(zio_t *zio)
2217 zio_t *pio = zio_unique_parent(zio);
2218 zio_t *gio = zio->io_gang_leader;
2219 dva_t *cdva = zio->io_bp->blk_dva;
2220 dva_t *pdva = pio->io_bp->blk_dva;
2223 if (BP_IS_HOLE(zio->io_bp))
2226 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2228 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2229 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2230 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2231 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2232 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2234 mutex_enter(&pio->io_lock);
2235 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2236 ASSERT(DVA_GET_GANG(&pdva[d]));
2237 asize = DVA_GET_ASIZE(&pdva[d]);
2238 asize += DVA_GET_ASIZE(&cdva[d]);
2239 DVA_SET_ASIZE(&pdva[d], asize);
2241 mutex_exit(&pio->io_lock);
2245 zio_write_gang_done(zio_t *zio)
2247 abd_put(zio->io_abd);
2251 zio_write_gang_block(zio_t *pio)
2253 spa_t *spa = pio->io_spa;
2254 metaslab_class_t *mc = spa_normal_class(spa);
2255 blkptr_t *bp = pio->io_bp;
2256 zio_t *gio = pio->io_gang_leader;
2258 zio_gang_node_t *gn, **gnpp;
2259 zio_gbh_phys_t *gbh;
2261 uint64_t txg = pio->io_txg;
2262 uint64_t resid = pio->io_size;
2264 int copies = gio->io_prop.zp_copies;
2265 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2269 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2270 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2271 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2272 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2274 flags |= METASLAB_ASYNC_ALLOC;
2275 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2278 * The logical zio has already placed a reservation for
2279 * 'copies' allocation slots but gang blocks may require
2280 * additional copies. These additional copies
2281 * (i.e. gbh_copies - copies) are guaranteed to succeed
2282 * since metaslab_class_throttle_reserve() always allows
2283 * additional reservations for gang blocks.
2285 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2289 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2290 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2291 &pio->io_alloc_list, pio);
2293 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2294 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2295 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2298 * If we failed to allocate the gang block header then
2299 * we remove any additional allocation reservations that
2300 * we placed here. The original reservation will
2301 * be removed when the logical I/O goes to the ready
2304 metaslab_class_throttle_unreserve(mc,
2305 gbh_copies - copies, pio);
2307 pio->io_error = error;
2308 return (ZIO_PIPELINE_CONTINUE);
2312 gnpp = &gio->io_gang_tree;
2314 gnpp = pio->io_private;
2315 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2318 gn = zio_gang_node_alloc(gnpp);
2320 bzero(gbh, SPA_GANGBLOCKSIZE);
2321 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2324 * Create the gang header.
2326 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2327 zio_write_gang_done, NULL, pio->io_priority,
2328 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2331 * Create and nowait the gang children.
2333 for (int g = 0; resid != 0; resid -= lsize, g++) {
2334 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2336 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2338 zp.zp_checksum = gio->io_prop.zp_checksum;
2339 zp.zp_compress = ZIO_COMPRESS_OFF;
2340 zp.zp_type = DMU_OT_NONE;
2342 zp.zp_copies = gio->io_prop.zp_copies;
2343 zp.zp_dedup = B_FALSE;
2344 zp.zp_dedup_verify = B_FALSE;
2345 zp.zp_nopwrite = B_FALSE;
2347 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2348 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2349 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2350 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2351 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2353 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2354 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2355 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2358 * Gang children won't throttle but we should
2359 * account for their work, so reserve an allocation
2360 * slot for them here.
2362 VERIFY(metaslab_class_throttle_reserve(mc,
2363 zp.zp_copies, cio, flags));
2369 * Set pio's pipeline to just wait for zio to finish.
2371 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2375 return (ZIO_PIPELINE_CONTINUE);
2379 * The zio_nop_write stage in the pipeline determines if allocating a
2380 * new bp is necessary. The nopwrite feature can handle writes in
2381 * either syncing or open context (i.e. zil writes) and as a result is
2382 * mutually exclusive with dedup.
2384 * By leveraging a cryptographically secure checksum, such as SHA256, we
2385 * can compare the checksums of the new data and the old to determine if
2386 * allocating a new block is required. Note that our requirements for
2387 * cryptographic strength are fairly weak: there can't be any accidental
2388 * hash collisions, but we don't need to be secure against intentional
2389 * (malicious) collisions. To trigger a nopwrite, you have to be able
2390 * to write the file to begin with, and triggering an incorrect (hash
2391 * collision) nopwrite is no worse than simply writing to the file.
2392 * That said, there are no known attacks against the checksum algorithms
2393 * used for nopwrite, assuming that the salt and the checksums
2394 * themselves remain secret.
2397 zio_nop_write(zio_t *zio)
2399 blkptr_t *bp = zio->io_bp;
2400 blkptr_t *bp_orig = &zio->io_bp_orig;
2401 zio_prop_t *zp = &zio->io_prop;
2403 ASSERT(BP_GET_LEVEL(bp) == 0);
2404 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2405 ASSERT(zp->zp_nopwrite);
2406 ASSERT(!zp->zp_dedup);
2407 ASSERT(zio->io_bp_override == NULL);
2408 ASSERT(IO_IS_ALLOCATING(zio));
2411 * Check to see if the original bp and the new bp have matching
2412 * characteristics (i.e. same checksum, compression algorithms, etc).
2413 * If they don't then just continue with the pipeline which will
2414 * allocate a new bp.
2416 if (BP_IS_HOLE(bp_orig) ||
2417 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2418 ZCHECKSUM_FLAG_NOPWRITE) ||
2419 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2420 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2421 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2422 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2423 return (ZIO_PIPELINE_CONTINUE);
2426 * If the checksums match then reset the pipeline so that we
2427 * avoid allocating a new bp and issuing any I/O.
2429 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2430 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2431 ZCHECKSUM_FLAG_NOPWRITE);
2432 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2433 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2434 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2435 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2436 sizeof (uint64_t)) == 0);
2439 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2440 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2443 return (ZIO_PIPELINE_CONTINUE);
2447 * ==========================================================================
2449 * ==========================================================================
2452 zio_ddt_child_read_done(zio_t *zio)
2454 blkptr_t *bp = zio->io_bp;
2455 ddt_entry_t *dde = zio->io_private;
2457 zio_t *pio = zio_unique_parent(zio);
2459 mutex_enter(&pio->io_lock);
2460 ddp = ddt_phys_select(dde, bp);
2461 if (zio->io_error == 0)
2462 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2464 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2465 dde->dde_repair_abd = zio->io_abd;
2467 abd_free(zio->io_abd);
2468 mutex_exit(&pio->io_lock);
2472 zio_ddt_read_start(zio_t *zio)
2474 blkptr_t *bp = zio->io_bp;
2476 ASSERT(BP_GET_DEDUP(bp));
2477 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2478 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2480 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2481 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2482 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2483 ddt_phys_t *ddp = dde->dde_phys;
2484 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2487 ASSERT(zio->io_vsd == NULL);
2490 if (ddp_self == NULL)
2491 return (ZIO_PIPELINE_CONTINUE);
2493 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2494 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2496 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2498 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2499 abd_alloc_for_io(zio->io_size, B_TRUE),
2500 zio->io_size, zio_ddt_child_read_done, dde,
2501 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2502 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2504 return (ZIO_PIPELINE_CONTINUE);
2507 zio_nowait(zio_read(zio, zio->io_spa, bp,
2508 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2509 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2511 return (ZIO_PIPELINE_CONTINUE);
2515 zio_ddt_read_done(zio_t *zio)
2517 blkptr_t *bp = zio->io_bp;
2519 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2520 return (ZIO_PIPELINE_STOP);
2522 ASSERT(BP_GET_DEDUP(bp));
2523 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2524 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2526 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2527 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2528 ddt_entry_t *dde = zio->io_vsd;
2530 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2531 return (ZIO_PIPELINE_CONTINUE);
2534 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2535 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2536 return (ZIO_PIPELINE_STOP);
2538 if (dde->dde_repair_abd != NULL) {
2539 abd_copy(zio->io_abd, dde->dde_repair_abd,
2541 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2543 ddt_repair_done(ddt, dde);
2547 ASSERT(zio->io_vsd == NULL);
2549 return (ZIO_PIPELINE_CONTINUE);
2553 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2555 spa_t *spa = zio->io_spa;
2556 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2558 /* We should never get a raw, override zio */
2559 ASSERT(!(zio->io_bp_override && do_raw));
2562 * Note: we compare the original data, not the transformed data,
2563 * because when zio->io_bp is an override bp, we will not have
2564 * pushed the I/O transforms. That's an important optimization
2565 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2567 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2568 zio_t *lio = dde->dde_lead_zio[p];
2571 return (lio->io_orig_size != zio->io_orig_size ||
2572 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2573 zio->io_orig_size) != 0);
2577 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2578 ddt_phys_t *ddp = &dde->dde_phys[p];
2580 if (ddp->ddp_phys_birth != 0) {
2581 arc_buf_t *abuf = NULL;
2582 arc_flags_t aflags = ARC_FLAG_WAIT;
2583 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2584 blkptr_t blk = *zio->io_bp;
2587 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2592 * Intuitively, it would make more sense to compare
2593 * io_abd than io_orig_abd in the raw case since you
2594 * don't want to look at any transformations that have
2595 * happened to the data. However, for raw I/Os the
2596 * data will actually be the same in io_abd and
2597 * io_orig_abd, so all we have to do is issue this as
2601 zio_flags |= ZIO_FLAG_RAW;
2602 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2603 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2605 ASSERT3P(zio->io_transform_stack, ==, NULL);
2608 error = arc_read(NULL, spa, &blk,
2609 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2610 zio_flags, &aflags, &zio->io_bookmark);
2613 if (arc_buf_size(abuf) != zio->io_orig_size ||
2614 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2615 zio->io_orig_size) != 0)
2616 error = SET_ERROR(EEXIST);
2617 arc_buf_destroy(abuf, &abuf);
2621 return (error != 0);
2629 zio_ddt_child_write_ready(zio_t *zio)
2631 int p = zio->io_prop.zp_copies;
2632 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2633 ddt_entry_t *dde = zio->io_private;
2634 ddt_phys_t *ddp = &dde->dde_phys[p];
2642 ASSERT(dde->dde_lead_zio[p] == zio);
2644 ddt_phys_fill(ddp, zio->io_bp);
2646 zio_link_t *zl = NULL;
2647 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2648 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2654 zio_ddt_child_write_done(zio_t *zio)
2656 int p = zio->io_prop.zp_copies;
2657 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2658 ddt_entry_t *dde = zio->io_private;
2659 ddt_phys_t *ddp = &dde->dde_phys[p];
2663 ASSERT(ddp->ddp_refcnt == 0);
2664 ASSERT(dde->dde_lead_zio[p] == zio);
2665 dde->dde_lead_zio[p] = NULL;
2667 if (zio->io_error == 0) {
2668 zio_link_t *zl = NULL;
2669 while (zio_walk_parents(zio, &zl) != NULL)
2670 ddt_phys_addref(ddp);
2672 ddt_phys_clear(ddp);
2679 zio_ddt_ditto_write_done(zio_t *zio)
2681 int p = DDT_PHYS_DITTO;
2682 zio_prop_t *zp = &zio->io_prop;
2683 blkptr_t *bp = zio->io_bp;
2684 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2685 ddt_entry_t *dde = zio->io_private;
2686 ddt_phys_t *ddp = &dde->dde_phys[p];
2687 ddt_key_t *ddk = &dde->dde_key;
2691 ASSERT(ddp->ddp_refcnt == 0);
2692 ASSERT(dde->dde_lead_zio[p] == zio);
2693 dde->dde_lead_zio[p] = NULL;
2695 if (zio->io_error == 0) {
2696 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2697 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2698 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2699 if (ddp->ddp_phys_birth != 0)
2700 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2701 ddt_phys_fill(ddp, bp);
2708 zio_ddt_write(zio_t *zio)
2710 spa_t *spa = zio->io_spa;
2711 blkptr_t *bp = zio->io_bp;
2712 uint64_t txg = zio->io_txg;
2713 zio_prop_t *zp = &zio->io_prop;
2714 int p = zp->zp_copies;
2718 ddt_t *ddt = ddt_select(spa, bp);
2722 ASSERT(BP_GET_DEDUP(bp));
2723 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2724 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2725 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2728 dde = ddt_lookup(ddt, bp, B_TRUE);
2729 ddp = &dde->dde_phys[p];
2731 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2733 * If we're using a weak checksum, upgrade to a strong checksum
2734 * and try again. If we're already using a strong checksum,
2735 * we can't resolve it, so just convert to an ordinary write.
2736 * (And automatically e-mail a paper to Nature?)
2738 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2739 ZCHECKSUM_FLAG_DEDUP)) {
2740 zp->zp_checksum = spa_dedup_checksum(spa);
2741 zio_pop_transforms(zio);
2742 zio->io_stage = ZIO_STAGE_OPEN;
2745 zp->zp_dedup = B_FALSE;
2746 BP_SET_DEDUP(bp, B_FALSE);
2748 ASSERT(!BP_GET_DEDUP(bp));
2749 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2751 return (ZIO_PIPELINE_CONTINUE);
2754 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2755 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2757 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2758 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2759 zio_prop_t czp = *zp;
2761 czp.zp_copies = ditto_copies;
2764 * If we arrived here with an override bp, we won't have run
2765 * the transform stack, so we won't have the data we need to
2766 * generate a child i/o. So, toss the override bp and restart.
2767 * This is safe, because using the override bp is just an
2768 * optimization; and it's rare, so the cost doesn't matter.
2770 if (zio->io_bp_override) {
2771 zio_pop_transforms(zio);
2772 zio->io_stage = ZIO_STAGE_OPEN;
2773 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2774 zio->io_bp_override = NULL;
2777 return (ZIO_PIPELINE_CONTINUE);
2780 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2781 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2782 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2783 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2785 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2786 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2789 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2790 if (ddp->ddp_phys_birth != 0)
2791 ddt_bp_fill(ddp, bp, txg);
2792 if (dde->dde_lead_zio[p] != NULL)
2793 zio_add_child(zio, dde->dde_lead_zio[p]);
2795 ddt_phys_addref(ddp);
2796 } else if (zio->io_bp_override) {
2797 ASSERT(bp->blk_birth == txg);
2798 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2799 ddt_phys_fill(ddp, bp);
2800 ddt_phys_addref(ddp);
2802 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2803 zio->io_orig_size, zio->io_orig_size, zp,
2804 zio_ddt_child_write_ready, NULL, NULL,
2805 zio_ddt_child_write_done, dde, zio->io_priority,
2806 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2808 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2809 dde->dde_lead_zio[p] = cio;
2819 return (ZIO_PIPELINE_CONTINUE);
2822 ddt_entry_t *freedde; /* for debugging */
2825 zio_ddt_free(zio_t *zio)
2827 spa_t *spa = zio->io_spa;
2828 blkptr_t *bp = zio->io_bp;
2829 ddt_t *ddt = ddt_select(spa, bp);
2833 ASSERT(BP_GET_DEDUP(bp));
2834 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2837 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2838 ddp = ddt_phys_select(dde, bp);
2839 ddt_phys_decref(ddp);
2842 return (ZIO_PIPELINE_CONTINUE);
2846 * ==========================================================================
2847 * Allocate and free blocks
2848 * ==========================================================================
2852 zio_io_to_allocate(spa_t *spa)
2856 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2858 zio = avl_first(&spa->spa_alloc_tree);
2862 ASSERT(IO_IS_ALLOCATING(zio));
2865 * Try to place a reservation for this zio. If we're unable to
2866 * reserve then we throttle.
2868 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2869 zio->io_prop.zp_copies, zio, 0)) {
2873 avl_remove(&spa->spa_alloc_tree, zio);
2874 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2880 zio_dva_throttle(zio_t *zio)
2882 spa_t *spa = zio->io_spa;
2885 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2886 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2887 zio->io_child_type == ZIO_CHILD_GANG ||
2888 zio->io_flags & ZIO_FLAG_NODATA) {
2889 return (ZIO_PIPELINE_CONTINUE);
2892 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2894 ASSERT3U(zio->io_queued_timestamp, >, 0);
2895 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2897 mutex_enter(&spa->spa_alloc_lock);
2899 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2900 avl_add(&spa->spa_alloc_tree, zio);
2902 nio = zio_io_to_allocate(zio->io_spa);
2903 mutex_exit(&spa->spa_alloc_lock);
2906 return (ZIO_PIPELINE_CONTINUE);
2909 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2911 * We are passing control to a new zio so make sure that
2912 * it is processed by a different thread. We do this to
2913 * avoid stack overflows that can occur when parents are
2914 * throttled and children are making progress. We allow
2915 * it to go to the head of the taskq since it's already
2918 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2920 return (ZIO_PIPELINE_STOP);
2924 zio_allocate_dispatch(spa_t *spa)
2928 mutex_enter(&spa->spa_alloc_lock);
2929 zio = zio_io_to_allocate(spa);
2930 mutex_exit(&spa->spa_alloc_lock);
2934 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2935 ASSERT0(zio->io_error);
2936 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2940 zio_dva_allocate(zio_t *zio)
2942 spa_t *spa = zio->io_spa;
2943 metaslab_class_t *mc = spa_normal_class(spa);
2944 blkptr_t *bp = zio->io_bp;
2948 if (zio->io_gang_leader == NULL) {
2949 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2950 zio->io_gang_leader = zio;
2953 ASSERT(BP_IS_HOLE(bp));
2954 ASSERT0(BP_GET_NDVAS(bp));
2955 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2956 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2957 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2959 if (zio->io_flags & ZIO_FLAG_NODATA) {
2960 flags |= METASLAB_DONT_THROTTLE;
2962 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2963 flags |= METASLAB_GANG_CHILD;
2965 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2966 flags |= METASLAB_ASYNC_ALLOC;
2969 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2970 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2971 &zio->io_alloc_list, zio);
2974 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2975 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2977 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2978 return (zio_write_gang_block(zio));
2979 zio->io_error = error;
2982 return (ZIO_PIPELINE_CONTINUE);
2986 zio_dva_free(zio_t *zio)
2988 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2990 return (ZIO_PIPELINE_CONTINUE);
2994 zio_dva_claim(zio_t *zio)
2998 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3000 zio->io_error = error;
3002 return (ZIO_PIPELINE_CONTINUE);
3006 * Undo an allocation. This is used by zio_done() when an I/O fails
3007 * and we want to give back the block we just allocated.
3008 * This handles both normal blocks and gang blocks.
3011 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3013 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3014 ASSERT(zio->io_bp_override == NULL);
3016 if (!BP_IS_HOLE(bp))
3017 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3020 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3021 zio_dva_unallocate(zio, gn->gn_child[g],
3022 &gn->gn_gbh->zg_blkptr[g]);
3028 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3031 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3032 uint64_t size, boolean_t *slog)
3035 zio_alloc_list_t io_alloc_list;
3037 ASSERT(txg > spa_syncing_txg(spa));
3039 metaslab_trace_init(&io_alloc_list);
3040 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3041 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3045 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3046 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3047 &io_alloc_list, NULL);
3051 metaslab_trace_fini(&io_alloc_list);
3054 BP_SET_LSIZE(new_bp, size);
3055 BP_SET_PSIZE(new_bp, size);
3056 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3057 BP_SET_CHECKSUM(new_bp,
3058 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3059 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3060 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3061 BP_SET_LEVEL(new_bp, 0);
3062 BP_SET_DEDUP(new_bp, 0);
3063 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3065 zfs_dbgmsg("%s: zil block allocation failure: "
3066 "size %llu, error %d", spa_name(spa), size, error);
3073 * Free an intent log block.
3076 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3078 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3079 ASSERT(!BP_IS_GANG(bp));
3081 zio_free(spa, txg, bp);
3085 * ==========================================================================
3086 * Read, write and delete to physical devices
3087 * ==========================================================================
3092 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3093 * stops after this stage and will resume upon I/O completion.
3094 * However, there are instances where the vdev layer may need to
3095 * continue the pipeline when an I/O was not issued. Since the I/O
3096 * that was sent to the vdev layer might be different than the one
3097 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3098 * force the underlying vdev layers to call either zio_execute() or
3099 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3102 zio_vdev_io_start(zio_t *zio)
3104 vdev_t *vd = zio->io_vd;
3106 spa_t *spa = zio->io_spa;
3109 ASSERT(zio->io_error == 0);
3110 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3113 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3114 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3117 * The mirror_ops handle multiple DVAs in a single BP.
3119 vdev_mirror_ops.vdev_op_io_start(zio);
3120 return (ZIO_PIPELINE_STOP);
3123 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3124 zio->io_priority == ZIO_PRIORITY_NOW) {
3125 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3126 return (ZIO_PIPELINE_CONTINUE);
3129 ASSERT3P(zio->io_logical, !=, zio);
3132 * We keep track of time-sensitive I/Os so that the scan thread
3133 * can quickly react to certain workloads. In particular, we care
3134 * about non-scrubbing, top-level reads and writes with the following
3136 * - synchronous writes of user data to non-slog devices
3137 * - any reads of user data
3138 * When these conditions are met, adjust the timestamp of spa_last_io
3139 * which allows the scan thread to adjust its workload accordingly.
3141 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3142 vd == vd->vdev_top && !vd->vdev_islog &&
3143 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3144 zio->io_txg != spa_syncing_txg(spa)) {
3145 uint64_t old = spa->spa_last_io;
3146 uint64_t new = ddi_get_lbolt64();
3148 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3151 align = 1ULL << vd->vdev_top->vdev_ashift;
3153 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3154 P2PHASE(zio->io_size, align) != 0) {
3155 /* Transform logical writes to be a full physical block size. */
3156 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3158 if (zio->io_type == ZIO_TYPE_READ ||
3159 zio->io_type == ZIO_TYPE_WRITE)
3160 abuf = abd_alloc_sametype(zio->io_abd, asize);
3161 ASSERT(vd == vd->vdev_top);
3162 if (zio->io_type == ZIO_TYPE_WRITE) {
3163 abd_copy(abuf, zio->io_abd, zio->io_size);
3164 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3166 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3171 * If this is not a physical io, make sure that it is properly aligned
3172 * before proceeding.
3174 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3175 ASSERT0(P2PHASE(zio->io_offset, align));
3176 ASSERT0(P2PHASE(zio->io_size, align));
3179 * For the physical io we allow alignment
3180 * to a logical block size.
3182 uint64_t log_align =
3183 1ULL << vd->vdev_top->vdev_logical_ashift;
3184 ASSERT0(P2PHASE(zio->io_offset, log_align));
3185 ASSERT0(P2PHASE(zio->io_size, log_align));
3188 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3191 * If this is a repair I/O, and there's no self-healing involved --
3192 * that is, we're just resilvering what we expect to resilver --
3193 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3194 * This prevents spurious resilvering with nested replication.
3195 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3196 * A is out of date, we'll read from C+D, then use the data to
3197 * resilver A+B -- but we don't actually want to resilver B, just A.
3198 * The top-level mirror has no way to know this, so instead we just
3199 * discard unnecessary repairs as we work our way down the vdev tree.
3200 * The same logic applies to any form of nested replication:
3201 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3203 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3204 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3205 zio->io_txg != 0 && /* not a delegated i/o */
3206 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3207 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3208 zio_vdev_io_bypass(zio);
3209 return (ZIO_PIPELINE_CONTINUE);
3212 if (vd->vdev_ops->vdev_op_leaf) {
3213 switch (zio->io_type) {
3215 if (vdev_cache_read(zio))
3216 return (ZIO_PIPELINE_CONTINUE);
3218 case ZIO_TYPE_WRITE:
3220 if ((zio = vdev_queue_io(zio)) == NULL)
3221 return (ZIO_PIPELINE_STOP);
3223 if (!vdev_accessible(vd, zio)) {
3224 zio->io_error = SET_ERROR(ENXIO);
3226 return (ZIO_PIPELINE_STOP);
3231 * Note that we ignore repair writes for TRIM because they can
3232 * conflict with normal writes. This isn't an issue because, by
3233 * definition, we only repair blocks that aren't freed.
3235 if (zio->io_type == ZIO_TYPE_WRITE &&
3236 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3237 !trim_map_write_start(zio))
3238 return (ZIO_PIPELINE_STOP);
3241 vd->vdev_ops->vdev_op_io_start(zio);
3242 return (ZIO_PIPELINE_STOP);
3246 zio_vdev_io_done(zio_t *zio)
3248 vdev_t *vd = zio->io_vd;
3249 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3250 boolean_t unexpected_error = B_FALSE;
3252 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3253 return (ZIO_PIPELINE_STOP);
3255 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3256 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3258 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3259 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3260 zio->io_type == ZIO_TYPE_FREE)) {
3262 if (zio->io_type == ZIO_TYPE_WRITE &&
3263 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3264 trim_map_write_done(zio);
3266 vdev_queue_io_done(zio);
3268 if (zio->io_type == ZIO_TYPE_WRITE)
3269 vdev_cache_write(zio);
3271 if (zio_injection_enabled && zio->io_error == 0)
3272 zio->io_error = zio_handle_device_injection(vd,
3275 if (zio_injection_enabled && zio->io_error == 0)
3276 zio->io_error = zio_handle_label_injection(zio, EIO);
3278 if (zio->io_error) {
3279 if (zio->io_error == ENOTSUP &&
3280 zio->io_type == ZIO_TYPE_FREE) {
3281 /* Not all devices support TRIM. */
3282 } else if (!vdev_accessible(vd, zio)) {
3283 zio->io_error = SET_ERROR(ENXIO);
3285 unexpected_error = B_TRUE;
3290 ops->vdev_op_io_done(zio);
3292 if (unexpected_error)
3293 VERIFY(vdev_probe(vd, zio) == NULL);
3295 return (ZIO_PIPELINE_CONTINUE);
3299 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3300 * disk, and use that to finish the checksum ereport later.
3303 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3304 const void *good_buf)
3306 /* no processing needed */
3307 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3312 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3314 void *buf = zio_buf_alloc(zio->io_size);
3316 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3318 zcr->zcr_cbinfo = zio->io_size;
3319 zcr->zcr_cbdata = buf;
3320 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3321 zcr->zcr_free = zio_buf_free;
3325 zio_vdev_io_assess(zio_t *zio)
3327 vdev_t *vd = zio->io_vd;
3329 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3330 return (ZIO_PIPELINE_STOP);
3332 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3333 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3335 if (zio->io_vsd != NULL) {
3336 zio->io_vsd_ops->vsd_free(zio);
3340 if (zio_injection_enabled && zio->io_error == 0)
3341 zio->io_error = zio_handle_fault_injection(zio, EIO);
3343 if (zio->io_type == ZIO_TYPE_FREE &&
3344 zio->io_priority != ZIO_PRIORITY_NOW) {
3345 switch (zio->io_error) {
3347 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3348 ZIO_TRIM_STAT_BUMP(success);
3351 ZIO_TRIM_STAT_BUMP(unsupported);
3354 ZIO_TRIM_STAT_BUMP(failed);
3360 * If the I/O failed, determine whether we should attempt to retry it.
3362 * On retry, we cut in line in the issue queue, since we don't want
3363 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3365 if (zio->io_error && vd == NULL &&
3366 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3367 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3368 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3370 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3371 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3372 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3373 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3374 zio_requeue_io_start_cut_in_line);
3375 return (ZIO_PIPELINE_STOP);
3379 * If we got an error on a leaf device, convert it to ENXIO
3380 * if the device is not accessible at all.
3382 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3383 !vdev_accessible(vd, zio))
3384 zio->io_error = SET_ERROR(ENXIO);
3387 * If we can't write to an interior vdev (mirror or RAID-Z),
3388 * set vdev_cant_write so that we stop trying to allocate from it.
3390 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3391 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3392 vd->vdev_cant_write = B_TRUE;
3396 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3397 * attempts will ever succeed. In this case we set a persistent bit so
3398 * that we don't bother with it in the future.
3400 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3401 zio->io_type == ZIO_TYPE_IOCTL &&
3402 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3403 vd->vdev_nowritecache = B_TRUE;
3406 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3408 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3409 zio->io_physdone != NULL) {
3410 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3411 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3412 zio->io_physdone(zio->io_logical);
3415 return (ZIO_PIPELINE_CONTINUE);
3419 zio_vdev_io_reissue(zio_t *zio)
3421 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3422 ASSERT(zio->io_error == 0);
3424 zio->io_stage >>= 1;
3428 zio_vdev_io_redone(zio_t *zio)
3430 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3432 zio->io_stage >>= 1;
3436 zio_vdev_io_bypass(zio_t *zio)
3438 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3439 ASSERT(zio->io_error == 0);
3441 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3442 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3446 * ==========================================================================
3447 * Generate and verify checksums
3448 * ==========================================================================
3451 zio_checksum_generate(zio_t *zio)
3453 blkptr_t *bp = zio->io_bp;
3454 enum zio_checksum checksum;
3458 * This is zio_write_phys().
3459 * We're either generating a label checksum, or none at all.
3461 checksum = zio->io_prop.zp_checksum;
3463 if (checksum == ZIO_CHECKSUM_OFF)
3464 return (ZIO_PIPELINE_CONTINUE);
3466 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3468 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3469 ASSERT(!IO_IS_ALLOCATING(zio));
3470 checksum = ZIO_CHECKSUM_GANG_HEADER;
3472 checksum = BP_GET_CHECKSUM(bp);
3476 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3478 return (ZIO_PIPELINE_CONTINUE);
3482 zio_checksum_verify(zio_t *zio)
3484 zio_bad_cksum_t info;
3485 blkptr_t *bp = zio->io_bp;
3488 ASSERT(zio->io_vd != NULL);
3492 * This is zio_read_phys().
3493 * We're either verifying a label checksum, or nothing at all.
3495 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3496 return (ZIO_PIPELINE_CONTINUE);
3498 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3501 if ((error = zio_checksum_error(zio, &info)) != 0) {
3502 zio->io_error = error;
3503 if (error == ECKSUM &&
3504 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3505 zfs_ereport_start_checksum(zio->io_spa,
3506 zio->io_vd, zio, zio->io_offset,
3507 zio->io_size, NULL, &info);
3511 return (ZIO_PIPELINE_CONTINUE);
3515 * Called by RAID-Z to ensure we don't compute the checksum twice.
3518 zio_checksum_verified(zio_t *zio)
3520 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3524 * ==========================================================================
3525 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3526 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3527 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3528 * indicate errors that are specific to one I/O, and most likely permanent.
3529 * Any other error is presumed to be worse because we weren't expecting it.
3530 * ==========================================================================
3533 zio_worst_error(int e1, int e2)
3535 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3538 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3539 if (e1 == zio_error_rank[r1])
3542 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3543 if (e2 == zio_error_rank[r2])
3546 return (r1 > r2 ? e1 : e2);
3550 * ==========================================================================
3552 * ==========================================================================
3555 zio_ready(zio_t *zio)
3557 blkptr_t *bp = zio->io_bp;
3558 zio_t *pio, *pio_next;
3559 zio_link_t *zl = NULL;
3561 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3562 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3563 return (ZIO_PIPELINE_STOP);
3565 if (zio->io_ready) {
3566 ASSERT(IO_IS_ALLOCATING(zio));
3567 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3568 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3569 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3574 if (bp != NULL && bp != &zio->io_bp_copy)
3575 zio->io_bp_copy = *bp;
3577 if (zio->io_error != 0) {
3578 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3580 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3581 ASSERT(IO_IS_ALLOCATING(zio));
3582 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3584 * We were unable to allocate anything, unreserve and
3585 * issue the next I/O to allocate.
3587 metaslab_class_throttle_unreserve(
3588 spa_normal_class(zio->io_spa),
3589 zio->io_prop.zp_copies, zio);
3590 zio_allocate_dispatch(zio->io_spa);
3594 mutex_enter(&zio->io_lock);
3595 zio->io_state[ZIO_WAIT_READY] = 1;
3596 pio = zio_walk_parents(zio, &zl);
3597 mutex_exit(&zio->io_lock);
3600 * As we notify zio's parents, new parents could be added.
3601 * New parents go to the head of zio's io_parent_list, however,
3602 * so we will (correctly) not notify them. The remainder of zio's
3603 * io_parent_list, from 'pio_next' onward, cannot change because
3604 * all parents must wait for us to be done before they can be done.
3606 for (; pio != NULL; pio = pio_next) {
3607 pio_next = zio_walk_parents(zio, &zl);
3608 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3611 if (zio->io_flags & ZIO_FLAG_NODATA) {
3612 if (BP_IS_GANG(bp)) {
3613 zio->io_flags &= ~ZIO_FLAG_NODATA;
3615 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3616 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3620 if (zio_injection_enabled &&
3621 zio->io_spa->spa_syncing_txg == zio->io_txg)
3622 zio_handle_ignored_writes(zio);
3624 return (ZIO_PIPELINE_CONTINUE);
3628 * Update the allocation throttle accounting.
3631 zio_dva_throttle_done(zio_t *zio)
3633 zio_t *lio = zio->io_logical;
3634 zio_t *pio = zio_unique_parent(zio);
3635 vdev_t *vd = zio->io_vd;
3636 int flags = METASLAB_ASYNC_ALLOC;
3638 ASSERT3P(zio->io_bp, !=, NULL);
3639 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3640 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3641 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3643 ASSERT3P(vd, ==, vd->vdev_top);
3644 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3645 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3646 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3647 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3650 * Parents of gang children can have two flavors -- ones that
3651 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3652 * and ones that allocated the constituent blocks. The allocation
3653 * throttle needs to know the allocating parent zio so we must find
3656 if (pio->io_child_type == ZIO_CHILD_GANG) {
3658 * If our parent is a rewrite gang child then our grandparent
3659 * would have been the one that performed the allocation.
3661 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3662 pio = zio_unique_parent(pio);
3663 flags |= METASLAB_GANG_CHILD;
3666 ASSERT(IO_IS_ALLOCATING(pio));
3667 ASSERT3P(zio, !=, zio->io_logical);
3668 ASSERT(zio->io_logical != NULL);
3669 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3670 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3672 mutex_enter(&pio->io_lock);
3673 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3674 mutex_exit(&pio->io_lock);
3676 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3680 * Call into the pipeline to see if there is more work that
3681 * needs to be done. If there is work to be done it will be
3682 * dispatched to another taskq thread.
3684 zio_allocate_dispatch(zio->io_spa);
3688 zio_done(zio_t *zio)
3690 spa_t *spa = zio->io_spa;
3691 zio_t *lio = zio->io_logical;
3692 blkptr_t *bp = zio->io_bp;
3693 vdev_t *vd = zio->io_vd;
3694 uint64_t psize = zio->io_size;
3695 zio_t *pio, *pio_next;
3696 metaslab_class_t *mc = spa_normal_class(spa);
3697 zio_link_t *zl = NULL;
3700 * If our children haven't all completed,
3701 * wait for them and then repeat this pipeline stage.
3703 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3704 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3705 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3706 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3707 return (ZIO_PIPELINE_STOP);
3710 * If the allocation throttle is enabled, then update the accounting.
3711 * We only track child I/Os that are part of an allocating async
3712 * write. We must do this since the allocation is performed
3713 * by the logical I/O but the actual write is done by child I/Os.
3715 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3716 zio->io_child_type == ZIO_CHILD_VDEV) {
3717 ASSERT(mc->mc_alloc_throttle_enabled);
3718 zio_dva_throttle_done(zio);
3722 * If the allocation throttle is enabled, verify that
3723 * we have decremented the refcounts for every I/O that was throttled.
3725 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3726 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3727 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3729 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3730 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3733 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3734 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3735 ASSERT(zio->io_children[c][w] == 0);
3737 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3738 ASSERT(bp->blk_pad[0] == 0);
3739 ASSERT(bp->blk_pad[1] == 0);
3740 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3741 (bp == zio_unique_parent(zio)->io_bp));
3742 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3743 zio->io_bp_override == NULL &&
3744 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3745 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3746 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3747 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3748 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3750 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3751 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3755 * If there were child vdev/gang/ddt errors, they apply to us now.
3757 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3758 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3759 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3762 * If the I/O on the transformed data was successful, generate any
3763 * checksum reports now while we still have the transformed data.
3765 if (zio->io_error == 0) {
3766 while (zio->io_cksum_report != NULL) {
3767 zio_cksum_report_t *zcr = zio->io_cksum_report;
3768 uint64_t align = zcr->zcr_align;
3769 uint64_t asize = P2ROUNDUP(psize, align);
3771 abd_t *adata = zio->io_abd;
3773 if (asize != psize) {
3774 adata = abd_alloc_linear(asize, B_TRUE);
3775 abd_copy(adata, zio->io_abd, psize);
3776 abd_zero_off(adata, psize, asize - psize);
3780 abuf = abd_borrow_buf_copy(adata, asize);
3782 zio->io_cksum_report = zcr->zcr_next;
3783 zcr->zcr_next = NULL;
3784 zcr->zcr_finish(zcr, abuf);
3785 zfs_ereport_free_checksum(zcr);
3788 abd_return_buf(adata, abuf, asize);
3795 zio_pop_transforms(zio); /* note: may set zio->io_error */
3797 vdev_stat_update(zio, psize);
3799 if (zio->io_error) {
3801 * If this I/O is attached to a particular vdev,
3802 * generate an error message describing the I/O failure
3803 * at the block level. We ignore these errors if the
3804 * device is currently unavailable.
3806 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3807 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3809 if ((zio->io_error == EIO || !(zio->io_flags &
3810 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3813 * For logical I/O requests, tell the SPA to log the
3814 * error and generate a logical data ereport.
3816 spa_log_error(spa, zio);
3817 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3822 if (zio->io_error && zio == lio) {
3824 * Determine whether zio should be reexecuted. This will
3825 * propagate all the way to the root via zio_notify_parent().
3827 ASSERT(vd == NULL && bp != NULL);
3828 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3830 if (IO_IS_ALLOCATING(zio) &&
3831 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3832 if (zio->io_error != ENOSPC)
3833 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3835 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3838 if ((zio->io_type == ZIO_TYPE_READ ||
3839 zio->io_type == ZIO_TYPE_FREE) &&
3840 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3841 zio->io_error == ENXIO &&
3842 spa_load_state(spa) == SPA_LOAD_NONE &&
3843 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3844 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3846 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3847 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3850 * Here is a possibly good place to attempt to do
3851 * either combinatorial reconstruction or error correction
3852 * based on checksums. It also might be a good place
3853 * to send out preliminary ereports before we suspend
3859 * If there were logical child errors, they apply to us now.
3860 * We defer this until now to avoid conflating logical child
3861 * errors with errors that happened to the zio itself when
3862 * updating vdev stats and reporting FMA events above.
3864 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3866 if ((zio->io_error || zio->io_reexecute) &&
3867 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3868 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3869 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3871 zio_gang_tree_free(&zio->io_gang_tree);
3874 * Godfather I/Os should never suspend.
3876 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3877 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3878 zio->io_reexecute = 0;
3880 if (zio->io_reexecute) {
3882 * This is a logical I/O that wants to reexecute.
3884 * Reexecute is top-down. When an i/o fails, if it's not
3885 * the root, it simply notifies its parent and sticks around.
3886 * The parent, seeing that it still has children in zio_done(),
3887 * does the same. This percolates all the way up to the root.
3888 * The root i/o will reexecute or suspend the entire tree.
3890 * This approach ensures that zio_reexecute() honors
3891 * all the original i/o dependency relationships, e.g.
3892 * parents not executing until children are ready.
3894 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3896 zio->io_gang_leader = NULL;
3898 mutex_enter(&zio->io_lock);
3899 zio->io_state[ZIO_WAIT_DONE] = 1;
3900 mutex_exit(&zio->io_lock);
3903 * "The Godfather" I/O monitors its children but is
3904 * not a true parent to them. It will track them through
3905 * the pipeline but severs its ties whenever they get into
3906 * trouble (e.g. suspended). This allows "The Godfather"
3907 * I/O to return status without blocking.
3910 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3912 zio_link_t *remove_zl = zl;
3913 pio_next = zio_walk_parents(zio, &zl);
3915 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3916 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3917 zio_remove_child(pio, zio, remove_zl);
3918 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3922 if ((pio = zio_unique_parent(zio)) != NULL) {
3924 * We're not a root i/o, so there's nothing to do
3925 * but notify our parent. Don't propagate errors
3926 * upward since we haven't permanently failed yet.
3928 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3929 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3930 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3931 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3933 * We'd fail again if we reexecuted now, so suspend
3934 * until conditions improve (e.g. device comes online).
3936 zio_suspend(spa, zio);
3939 * Reexecution is potentially a huge amount of work.
3940 * Hand it off to the otherwise-unused claim taskq.
3942 #if defined(illumos) || !defined(_KERNEL)
3943 ASSERT(zio->io_tqent.tqent_next == NULL);
3945 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3947 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3948 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3951 return (ZIO_PIPELINE_STOP);
3954 ASSERT(zio->io_child_count == 0);
3955 ASSERT(zio->io_reexecute == 0);
3956 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3959 * Report any checksum errors, since the I/O is complete.
3961 while (zio->io_cksum_report != NULL) {
3962 zio_cksum_report_t *zcr = zio->io_cksum_report;
3963 zio->io_cksum_report = zcr->zcr_next;
3964 zcr->zcr_next = NULL;
3965 zcr->zcr_finish(zcr, NULL);
3966 zfs_ereport_free_checksum(zcr);
3970 * It is the responsibility of the done callback to ensure that this
3971 * particular zio is no longer discoverable for adoption, and as
3972 * such, cannot acquire any new parents.
3977 mutex_enter(&zio->io_lock);
3978 zio->io_state[ZIO_WAIT_DONE] = 1;
3979 mutex_exit(&zio->io_lock);
3982 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3983 zio_link_t *remove_zl = zl;
3984 pio_next = zio_walk_parents(zio, &zl);
3985 zio_remove_child(pio, zio, remove_zl);
3986 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3989 if (zio->io_waiter != NULL) {
3990 mutex_enter(&zio->io_lock);
3991 zio->io_executor = NULL;
3992 cv_broadcast(&zio->io_cv);
3993 mutex_exit(&zio->io_lock);
3998 return (ZIO_PIPELINE_STOP);
4002 * ==========================================================================
4003 * I/O pipeline definition
4004 * ==========================================================================
4006 static zio_pipe_stage_t *zio_pipeline[] = {
4013 zio_checksum_generate,
4029 zio_checksum_verify,
4037 * Compare two zbookmark_phys_t's to see which we would reach first in a
4038 * pre-order traversal of the object tree.
4040 * This is simple in every case aside from the meta-dnode object. For all other
4041 * objects, we traverse them in order (object 1 before object 2, and so on).
4042 * However, all of these objects are traversed while traversing object 0, since
4043 * the data it points to is the list of objects. Thus, we need to convert to a
4044 * canonical representation so we can compare meta-dnode bookmarks to
4045 * non-meta-dnode bookmarks.
4047 * We do this by calculating "equivalents" for each field of the zbookmark.
4048 * zbookmarks outside of the meta-dnode use their own object and level, and
4049 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4050 * blocks this bookmark refers to) by multiplying their blkid by their span
4051 * (the number of L0 blocks contained within one block at their level).
4052 * zbookmarks inside the meta-dnode calculate their object equivalent
4053 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4054 * level + 1<<31 (any value larger than a level could ever be) for their level.
4055 * This causes them to always compare before a bookmark in their object
4056 * equivalent, compare appropriately to bookmarks in other objects, and to
4057 * compare appropriately to other bookmarks in the meta-dnode.
4060 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4061 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4064 * These variables represent the "equivalent" values for the zbookmark,
4065 * after converting zbookmarks inside the meta dnode to their
4066 * normal-object equivalents.
4068 uint64_t zb1obj, zb2obj;
4069 uint64_t zb1L0, zb2L0;
4070 uint64_t zb1level, zb2level;
4072 if (zb1->zb_object == zb2->zb_object &&
4073 zb1->zb_level == zb2->zb_level &&
4074 zb1->zb_blkid == zb2->zb_blkid)
4078 * BP_SPANB calculates the span in blocks.
4080 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4081 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4083 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4084 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4086 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4088 zb1obj = zb1->zb_object;
4089 zb1level = zb1->zb_level;
4092 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4093 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4095 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4097 zb2obj = zb2->zb_object;
4098 zb2level = zb2->zb_level;
4101 /* Now that we have a canonical representation, do the comparison. */
4102 if (zb1obj != zb2obj)
4103 return (zb1obj < zb2obj ? -1 : 1);
4104 else if (zb1L0 != zb2L0)
4105 return (zb1L0 < zb2L0 ? -1 : 1);
4106 else if (zb1level != zb2level)
4107 return (zb1level > zb2level ? -1 : 1);
4109 * This can (theoretically) happen if the bookmarks have the same object
4110 * and level, but different blkids, if the block sizes are not the same.
4111 * There is presently no way to change the indirect block sizes
4117 * This function checks the following: given that last_block is the place that
4118 * our traversal stopped last time, does that guarantee that we've visited
4119 * every node under subtree_root? Therefore, we can't just use the raw output
4120 * of zbookmark_compare. We have to pass in a modified version of
4121 * subtree_root; by incrementing the block id, and then checking whether
4122 * last_block is before or equal to that, we can tell whether or not having
4123 * visited last_block implies that all of subtree_root's children have been
4127 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4128 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4130 zbookmark_phys_t mod_zb = *subtree_root;
4132 ASSERT(last_block->zb_level == 0);
4134 /* The objset_phys_t isn't before anything. */
4139 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4140 * data block size in sectors, because that variable is only used if
4141 * the bookmark refers to a block in the meta-dnode. Since we don't
4142 * know without examining it what object it refers to, and there's no
4143 * harm in passing in this value in other cases, we always pass it in.
4145 * We pass in 0 for the indirect block size shift because zb2 must be
4146 * level 0. The indirect block size is only used to calculate the span
4147 * of the bookmark, but since the bookmark must be level 0, the span is
4148 * always 1, so the math works out.
4150 * If you make changes to how the zbookmark_compare code works, be sure
4151 * to make sure that this code still works afterwards.
4153 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4154 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,