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/dsl_scan.h>
45 #include <sys/metaslab_impl.h>
48 SYSCTL_DECL(_vfs_zfs);
49 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
50 #if defined(__amd64__)
51 static int zio_use_uma = 1;
53 static int zio_use_uma = 0;
55 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
56 "Use uma(9) for ZIO allocations");
57 static int zio_exclude_metadata = 0;
58 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
59 "Exclude metadata buffers from dumps as well");
61 zio_trim_stats_t zio_trim_stats = {
62 { "bytes", KSTAT_DATA_UINT64,
63 "Number of bytes successfully TRIMmed" },
64 { "success", KSTAT_DATA_UINT64,
65 "Number of successful TRIM requests" },
66 { "unsupported", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed because TRIM is not supported" },
68 { "failed", KSTAT_DATA_UINT64,
69 "Number of TRIM requests that failed for reasons other than not supported" },
72 static kstat_t *zio_trim_ksp;
75 * ==========================================================================
76 * I/O type descriptions
77 * ==========================================================================
79 const char *zio_type_name[ZIO_TYPES] = {
80 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
84 boolean_t zio_dva_throttle_enabled = B_TRUE;
85 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
86 &zio_dva_throttle_enabled, 0, "");
89 * ==========================================================================
91 * ==========================================================================
93 kmem_cache_t *zio_cache;
94 kmem_cache_t *zio_link_cache;
95 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
96 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
99 extern vmem_t *zio_alloc_arena;
102 #define ZIO_PIPELINE_CONTINUE 0x100
103 #define ZIO_PIPELINE_STOP 0x101
105 #define BP_SPANB(indblkshift, level) \
106 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
107 #define COMPARE_META_LEVEL 0x80000000ul
109 * The following actions directly effect the spa's sync-to-convergence logic.
110 * The values below define the sync pass when we start performing the action.
111 * Care should be taken when changing these values as they directly impact
112 * spa_sync() performance. Tuning these values may introduce subtle performance
113 * pathologies and should only be done in the context of performance analysis.
114 * These tunables will eventually be removed and replaced with #defines once
115 * enough analysis has been done to determine optimal values.
117 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
118 * regular blocks are not deferred.
120 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
122 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
123 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
124 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
125 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
126 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
127 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
128 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
131 * An allocating zio is one that either currently has the DVA allocate
132 * stage set or will have it later in its lifetime.
134 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
136 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
140 int zio_buf_debug_limit = 16384;
142 int zio_buf_debug_limit = 0;
146 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
152 zio_cache = kmem_cache_create("zio_cache",
153 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
154 zio_link_cache = kmem_cache_create("zio_link_cache",
155 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
160 * For small buffers, we want a cache for each multiple of
161 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
162 * for each quarter-power of 2.
164 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
165 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
168 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
176 * If we are using watchpoints, put each buffer on its own page,
177 * to eliminate the performance overhead of trapping to the
178 * kernel when modifying a non-watched buffer that shares the
179 * page with a watched buffer.
181 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
185 if (size <= 4 * SPA_MINBLOCKSIZE) {
186 align = SPA_MINBLOCKSIZE;
187 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
188 align = MIN(p2 >> 2, PAGESIZE);
193 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
194 zio_buf_cache[c] = kmem_cache_create(name, size,
195 align, NULL, NULL, NULL, NULL, NULL, cflags);
198 * Since zio_data bufs do not appear in crash dumps, we
199 * pass KMC_NOTOUCH so that no allocator metadata is
200 * stored with the buffers.
202 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
203 zio_data_buf_cache[c] = kmem_cache_create(name, size,
204 align, NULL, NULL, NULL, NULL, NULL,
205 cflags | KMC_NOTOUCH | KMC_NODEBUG);
210 ASSERT(zio_buf_cache[c] != NULL);
211 if (zio_buf_cache[c - 1] == NULL)
212 zio_buf_cache[c - 1] = zio_buf_cache[c];
214 ASSERT(zio_data_buf_cache[c] != NULL);
215 if (zio_data_buf_cache[c - 1] == NULL)
216 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
222 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
224 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
227 if (zio_trim_ksp != NULL) {
228 zio_trim_ksp->ks_data = &zio_trim_stats;
229 kstat_install(zio_trim_ksp);
237 kmem_cache_t *last_cache = NULL;
238 kmem_cache_t *last_data_cache = NULL;
240 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
241 if (zio_buf_cache[c] != last_cache) {
242 last_cache = zio_buf_cache[c];
243 kmem_cache_destroy(zio_buf_cache[c]);
245 zio_buf_cache[c] = NULL;
247 if (zio_data_buf_cache[c] != last_data_cache) {
248 last_data_cache = zio_data_buf_cache[c];
249 kmem_cache_destroy(zio_data_buf_cache[c]);
251 zio_data_buf_cache[c] = NULL;
254 kmem_cache_destroy(zio_link_cache);
255 kmem_cache_destroy(zio_cache);
259 if (zio_trim_ksp != NULL) {
260 kstat_delete(zio_trim_ksp);
266 * ==========================================================================
267 * Allocate and free I/O buffers
268 * ==========================================================================
272 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
273 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
274 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
275 * excess / transient data in-core during a crashdump.
278 zio_buf_alloc(size_t size)
280 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
281 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
283 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
286 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
288 return (kmem_alloc(size, KM_SLEEP|flags));
292 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
293 * crashdump if the kernel panics. This exists so that we will limit the amount
294 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
295 * of kernel heap dumped to disk when the kernel panics)
298 zio_data_buf_alloc(size_t size)
300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
302 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
305 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
307 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
311 zio_buf_free(void *buf, size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 kmem_cache_free(zio_buf_cache[c], buf);
320 kmem_free(buf, size);
324 zio_data_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_data_buf_cache[c], buf);
333 kmem_free(buf, size);
337 * ==========================================================================
338 * Push and pop I/O transform buffers
339 * ==========================================================================
342 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
343 zio_transform_func_t *transform)
345 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
348 * Ensure that anyone expecting this zio to contain a linear ABD isn't
349 * going to get a nasty surprise when they try to access the data.
352 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
354 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
355 abd_is_linear(data));
358 zt->zt_orig_abd = zio->io_abd;
359 zt->zt_orig_size = zio->io_size;
360 zt->zt_bufsize = bufsize;
361 zt->zt_transform = transform;
363 zt->zt_next = zio->io_transform_stack;
364 zio->io_transform_stack = zt;
371 zio_pop_transforms(zio_t *zio)
375 while ((zt = zio->io_transform_stack) != NULL) {
376 if (zt->zt_transform != NULL)
377 zt->zt_transform(zio,
378 zt->zt_orig_abd, zt->zt_orig_size);
380 if (zt->zt_bufsize != 0)
381 abd_free(zio->io_abd);
383 zio->io_abd = zt->zt_orig_abd;
384 zio->io_size = zt->zt_orig_size;
385 zio->io_transform_stack = zt->zt_next;
387 kmem_free(zt, sizeof (zio_transform_t));
392 * ==========================================================================
393 * I/O transform callbacks for subblocks and decompression
394 * ==========================================================================
397 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
399 ASSERT(zio->io_size > size);
401 if (zio->io_type == ZIO_TYPE_READ)
402 abd_copy(data, zio->io_abd, size);
406 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
408 if (zio->io_error == 0) {
409 void *tmp = abd_borrow_buf(data, size);
410 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
411 zio->io_abd, tmp, zio->io_size, size);
412 abd_return_buf_copy(data, tmp, size);
415 zio->io_error = SET_ERROR(EIO);
420 * ==========================================================================
421 * I/O parent/child relationships and pipeline interlocks
422 * ==========================================================================
425 zio_walk_parents(zio_t *cio, zio_link_t **zl)
427 list_t *pl = &cio->io_parent_list;
429 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
433 ASSERT((*zl)->zl_child == cio);
434 return ((*zl)->zl_parent);
438 zio_walk_children(zio_t *pio, zio_link_t **zl)
440 list_t *cl = &pio->io_child_list;
442 ASSERT(MUTEX_HELD(&pio->io_lock));
444 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
448 ASSERT((*zl)->zl_parent == pio);
449 return ((*zl)->zl_child);
453 zio_unique_parent(zio_t *cio)
455 zio_link_t *zl = NULL;
456 zio_t *pio = zio_walk_parents(cio, &zl);
458 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
463 zio_add_child(zio_t *pio, zio_t *cio)
465 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
478 mutex_enter(&pio->io_lock);
479 mutex_enter(&cio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&cio->io_lock);
493 mutex_exit(&pio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&pio->io_lock);
503 mutex_enter(&cio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&cio->io_lock);
512 mutex_exit(&pio->io_lock);
513 kmem_cache_free(zio_link_cache, zl);
517 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
519 boolean_t waiting = B_FALSE;
521 mutex_enter(&zio->io_lock);
522 ASSERT(zio->io_stall == NULL);
523 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
524 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
527 uint64_t *countp = &zio->io_children[c][wait];
530 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
531 zio->io_stall = countp;
536 mutex_exit(&zio->io_lock);
541 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
543 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
544 int *errorp = &pio->io_child_error[zio->io_child_type];
546 mutex_enter(&pio->io_lock);
547 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
548 *errorp = zio_worst_error(*errorp, zio->io_error);
549 pio->io_reexecute |= zio->io_reexecute;
550 ASSERT3U(*countp, >, 0);
554 if (*countp == 0 && pio->io_stall == countp) {
555 zio_taskq_type_t type =
556 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
558 pio->io_stall = NULL;
559 mutex_exit(&pio->io_lock);
561 * Dispatch the parent zio in its own taskq so that
562 * the child can continue to make progress. This also
563 * prevents overflowing the stack when we have deeply nested
564 * parent-child relationships.
566 zio_taskq_dispatch(pio, type, B_FALSE);
568 mutex_exit(&pio->io_lock);
573 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
575 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
576 zio->io_error = zio->io_child_error[c];
580 zio_bookmark_compare(const void *x1, const void *x2)
582 const zio_t *z1 = x1;
583 const zio_t *z2 = x2;
585 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
587 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
590 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
592 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
595 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
597 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
600 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
602 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
614 * ==========================================================================
615 * Create the various types of I/O (read, write, free, etc)
616 * ==========================================================================
619 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
620 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
621 void *private, zio_type_t type, zio_priority_t priority,
622 enum zio_flag flags, vdev_t *vd, uint64_t offset,
623 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
627 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
628 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
629 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
631 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
632 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
633 ASSERT(vd || stage == ZIO_STAGE_OPEN);
635 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
637 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
638 bzero(zio, sizeof (zio_t));
640 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
641 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
643 list_create(&zio->io_parent_list, sizeof (zio_link_t),
644 offsetof(zio_link_t, zl_parent_node));
645 list_create(&zio->io_child_list, sizeof (zio_link_t),
646 offsetof(zio_link_t, zl_child_node));
647 metaslab_trace_init(&zio->io_alloc_list);
650 zio->io_child_type = ZIO_CHILD_VDEV;
651 else if (flags & ZIO_FLAG_GANG_CHILD)
652 zio->io_child_type = ZIO_CHILD_GANG;
653 else if (flags & ZIO_FLAG_DDT_CHILD)
654 zio->io_child_type = ZIO_CHILD_DDT;
656 zio->io_child_type = ZIO_CHILD_LOGICAL;
659 zio->io_bp = (blkptr_t *)bp;
660 zio->io_bp_copy = *bp;
661 zio->io_bp_orig = *bp;
662 if (type != ZIO_TYPE_WRITE ||
663 zio->io_child_type == ZIO_CHILD_DDT)
664 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
665 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
666 zio->io_logical = zio;
667 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
668 pipeline |= ZIO_GANG_STAGES;
674 zio->io_private = private;
676 zio->io_priority = priority;
678 zio->io_offset = offset;
679 zio->io_orig_abd = zio->io_abd = data;
680 zio->io_orig_size = zio->io_size = psize;
681 zio->io_lsize = lsize;
682 zio->io_orig_flags = zio->io_flags = flags;
683 zio->io_orig_stage = zio->io_stage = stage;
684 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
685 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
687 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
688 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
691 zio->io_bookmark = *zb;
694 if (zio->io_logical == NULL)
695 zio->io_logical = pio->io_logical;
696 if (zio->io_child_type == ZIO_CHILD_GANG)
697 zio->io_gang_leader = pio->io_gang_leader;
698 zio_add_child(pio, zio);
705 zio_destroy(zio_t *zio)
707 metaslab_trace_fini(&zio->io_alloc_list);
708 list_destroy(&zio->io_parent_list);
709 list_destroy(&zio->io_child_list);
710 mutex_destroy(&zio->io_lock);
711 cv_destroy(&zio->io_cv);
712 kmem_cache_free(zio_cache, zio);
716 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
717 void *private, enum zio_flag flags)
721 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
722 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
723 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
729 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
731 return (zio_null(NULL, spa, NULL, done, private, flags));
735 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
737 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
738 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
739 bp, (longlong_t)BP_GET_TYPE(bp));
741 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
742 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
743 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
744 bp, (longlong_t)BP_GET_CHECKSUM(bp));
746 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
747 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
748 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
749 bp, (longlong_t)BP_GET_COMPRESS(bp));
751 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
752 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
753 bp, (longlong_t)BP_GET_LSIZE(bp));
755 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
756 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
757 bp, (longlong_t)BP_GET_PSIZE(bp));
760 if (BP_IS_EMBEDDED(bp)) {
761 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
762 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
763 bp, (longlong_t)BPE_GET_ETYPE(bp));
768 * Do not verify individual DVAs if the config is not trusted. This
769 * will be done once the zio is executed in vdev_mirror_map_alloc.
771 if (!spa->spa_trust_config)
775 * Pool-specific checks.
777 * Note: it would be nice to verify that the blk_birth and
778 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
779 * allows the birth time of log blocks (and dmu_sync()-ed blocks
780 * that are in the log) to be arbitrarily large.
782 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
783 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
784 if (vdevid >= spa->spa_root_vdev->vdev_children) {
785 zfs_panic_recover("blkptr at %p DVA %u has invalid "
787 bp, i, (longlong_t)vdevid);
790 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
792 zfs_panic_recover("blkptr at %p DVA %u has invalid "
794 bp, i, (longlong_t)vdevid);
797 if (vd->vdev_ops == &vdev_hole_ops) {
798 zfs_panic_recover("blkptr at %p DVA %u has hole "
800 bp, i, (longlong_t)vdevid);
803 if (vd->vdev_ops == &vdev_missing_ops) {
805 * "missing" vdevs are valid during import, but we
806 * don't have their detailed info (e.g. asize), so
807 * we can't perform any more checks on them.
811 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
812 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
814 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
815 if (offset + asize > vd->vdev_asize) {
816 zfs_panic_recover("blkptr at %p DVA %u has invalid "
818 bp, i, (longlong_t)offset);
824 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
826 uint64_t vdevid = DVA_GET_VDEV(dva);
828 if (vdevid >= spa->spa_root_vdev->vdev_children)
831 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
835 if (vd->vdev_ops == &vdev_hole_ops)
838 if (vd->vdev_ops == &vdev_missing_ops) {
842 uint64_t offset = DVA_GET_OFFSET(dva);
843 uint64_t asize = DVA_GET_ASIZE(dva);
846 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
847 if (offset + asize > vd->vdev_asize)
854 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
855 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
856 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
860 zfs_blkptr_verify(spa, bp);
862 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
863 data, size, size, done, private,
864 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
865 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
866 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
872 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
873 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
874 zio_done_func_t *ready, zio_done_func_t *children_ready,
875 zio_done_func_t *physdone, zio_done_func_t *done,
876 void *private, zio_priority_t priority, enum zio_flag flags,
877 const zbookmark_phys_t *zb)
881 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
882 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
883 zp->zp_compress >= ZIO_COMPRESS_OFF &&
884 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
885 DMU_OT_IS_VALID(zp->zp_type) &&
888 zp->zp_copies <= spa_max_replication(spa));
890 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
891 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
892 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
893 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
895 zio->io_ready = ready;
896 zio->io_children_ready = children_ready;
897 zio->io_physdone = physdone;
901 * Data can be NULL if we are going to call zio_write_override() to
902 * provide the already-allocated BP. But we may need the data to
903 * verify a dedup hit (if requested). In this case, don't try to
904 * dedup (just take the already-allocated BP verbatim).
906 if (data == NULL && zio->io_prop.zp_dedup_verify) {
907 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
914 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
915 uint64_t size, zio_done_func_t *done, void *private,
916 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
920 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
921 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
922 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
928 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
930 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
931 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
932 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
933 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
936 * We must reset the io_prop to match the values that existed
937 * when the bp was first written by dmu_sync() keeping in mind
938 * that nopwrite and dedup are mutually exclusive.
940 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
941 zio->io_prop.zp_nopwrite = nopwrite;
942 zio->io_prop.zp_copies = copies;
943 zio->io_bp_override = bp;
947 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
950 zfs_blkptr_verify(spa, bp);
953 * The check for EMBEDDED is a performance optimization. We
954 * process the free here (by ignoring it) rather than
955 * putting it on the list and then processing it in zio_free_sync().
957 if (BP_IS_EMBEDDED(bp))
959 metaslab_check_free(spa, bp);
962 * Frees that are for the currently-syncing txg, are not going to be
963 * deferred, and which will not need to do a read (i.e. not GANG or
964 * DEDUP), can be processed immediately. Otherwise, put them on the
965 * in-memory list for later processing.
967 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
968 txg != spa->spa_syncing_txg ||
969 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
970 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
972 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
973 BP_GET_PSIZE(bp), 0)));
978 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
979 uint64_t size, enum zio_flag flags)
982 enum zio_stage stage = ZIO_FREE_PIPELINE;
984 ASSERT(!BP_IS_HOLE(bp));
985 ASSERT(spa_syncing_txg(spa) == txg);
986 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
988 if (BP_IS_EMBEDDED(bp))
989 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
991 metaslab_check_free(spa, bp);
993 dsl_scan_freed(spa, bp);
995 if (zfs_trim_enabled)
996 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
997 ZIO_STAGE_VDEV_IO_ASSESS;
999 * GANG and DEDUP blocks can induce a read (for the gang block header,
1000 * or the DDT), so issue them asynchronously so that this thread is
1003 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1004 stage |= ZIO_STAGE_ISSUE_ASYNC;
1006 flags |= ZIO_FLAG_DONT_QUEUE;
1008 zio = zio_create(pio, spa, txg, bp, NULL, size,
1009 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1010 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1016 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1017 zio_done_func_t *done, void *private, enum zio_flag flags)
1021 zfs_blkptr_verify(spa, bp);
1023 if (BP_IS_EMBEDDED(bp))
1024 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1027 * A claim is an allocation of a specific block. Claims are needed
1028 * to support immediate writes in the intent log. The issue is that
1029 * immediate writes contain committed data, but in a txg that was
1030 * *not* committed. Upon opening the pool after an unclean shutdown,
1031 * the intent log claims all blocks that contain immediate write data
1032 * so that the SPA knows they're in use.
1034 * All claims *must* be resolved in the first txg -- before the SPA
1035 * starts allocating blocks -- so that nothing is allocated twice.
1036 * If txg == 0 we just verify that the block is claimable.
1038 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1039 spa_min_claim_txg(spa));
1040 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1041 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1043 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1044 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1045 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1046 ASSERT0(zio->io_queued_timestamp);
1052 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1053 uint64_t size, zio_done_func_t *done, void *private,
1054 zio_priority_t priority, enum zio_flag flags)
1059 if (vd->vdev_children == 0) {
1060 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1061 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1062 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1066 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1068 for (c = 0; c < vd->vdev_children; c++)
1069 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1070 offset, size, done, private, priority, flags));
1077 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1078 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1079 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1083 ASSERT(vd->vdev_children == 0);
1084 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1085 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1086 ASSERT3U(offset + size, <=, vd->vdev_psize);
1088 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1089 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1090 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1092 zio->io_prop.zp_checksum = checksum;
1098 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1099 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1100 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1104 ASSERT(vd->vdev_children == 0);
1105 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1106 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1107 ASSERT3U(offset + size, <=, vd->vdev_psize);
1109 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1110 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1111 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1113 zio->io_prop.zp_checksum = checksum;
1115 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1117 * zec checksums are necessarily destructive -- they modify
1118 * the end of the write buffer to hold the verifier/checksum.
1119 * Therefore, we must make a local copy in case the data is
1120 * being written to multiple places in parallel.
1122 abd_t *wbuf = abd_alloc_sametype(data, size);
1123 abd_copy(wbuf, data, size);
1125 zio_push_transform(zio, wbuf, size, size, NULL);
1132 * Create a child I/O to do some work for us.
1135 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1136 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1137 enum zio_flag flags, zio_done_func_t *done, void *private)
1139 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1143 * vdev child I/Os do not propagate their error to the parent.
1144 * Therefore, for correct operation the caller *must* check for
1145 * and handle the error in the child i/o's done callback.
1146 * The only exceptions are i/os that we don't care about
1147 * (OPTIONAL or REPAIR).
1149 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1153 * In the common case, where the parent zio was to a normal vdev,
1154 * the child zio must be to a child vdev of that vdev. Otherwise,
1155 * the child zio must be to a top-level vdev.
1157 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1158 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1160 ASSERT3P(vd, ==, vd->vdev_top);
1163 if (type == ZIO_TYPE_READ && bp != NULL) {
1165 * If we have the bp, then the child should perform the
1166 * checksum and the parent need not. This pushes error
1167 * detection as close to the leaves as possible and
1168 * eliminates redundant checksums in the interior nodes.
1170 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1171 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1174 /* Not all IO types require vdev io done stage e.g. free */
1175 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1176 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1178 if (vd->vdev_ops->vdev_op_leaf) {
1179 ASSERT0(vd->vdev_children);
1180 offset += VDEV_LABEL_START_SIZE;
1183 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1186 * If we've decided to do a repair, the write is not speculative --
1187 * even if the original read was.
1189 if (flags & ZIO_FLAG_IO_REPAIR)
1190 flags &= ~ZIO_FLAG_SPECULATIVE;
1193 * If we're creating a child I/O that is not associated with a
1194 * top-level vdev, then the child zio is not an allocating I/O.
1195 * If this is a retried I/O then we ignore it since we will
1196 * have already processed the original allocating I/O.
1198 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1199 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1200 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1202 ASSERT(mc->mc_alloc_throttle_enabled);
1203 ASSERT(type == ZIO_TYPE_WRITE);
1204 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1205 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1206 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1207 pio->io_child_type == ZIO_CHILD_GANG);
1209 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1212 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1213 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1214 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1215 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1217 zio->io_physdone = pio->io_physdone;
1218 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1219 zio->io_logical->io_phys_children++;
1225 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1226 int type, zio_priority_t priority, enum zio_flag flags,
1227 zio_done_func_t *done, void *private)
1231 ASSERT(vd->vdev_ops->vdev_op_leaf);
1233 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1234 data, size, size, done, private, type, priority,
1235 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1237 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1243 zio_flush(zio_t *zio, vdev_t *vd)
1245 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1246 NULL, NULL, ZIO_PRIORITY_NOW,
1247 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1251 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1254 ASSERT(vd->vdev_ops->vdev_op_leaf);
1256 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1257 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1258 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1259 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1263 zio_shrink(zio_t *zio, uint64_t size)
1265 ASSERT3P(zio->io_executor, ==, NULL);
1266 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1267 ASSERT3U(size, <=, zio->io_size);
1270 * We don't shrink for raidz because of problems with the
1271 * reconstruction when reading back less than the block size.
1272 * Note, BP_IS_RAIDZ() assumes no compression.
1274 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1275 if (!BP_IS_RAIDZ(zio->io_bp)) {
1276 /* we are not doing a raw write */
1277 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1278 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1283 * ==========================================================================
1284 * Prepare to read and write logical blocks
1285 * ==========================================================================
1289 zio_read_bp_init(zio_t *zio)
1291 blkptr_t *bp = zio->io_bp;
1293 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1295 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1296 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1297 !(zio->io_flags & ZIO_FLAG_RAW)) {
1299 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1300 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1301 psize, psize, zio_decompress);
1304 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1305 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1307 int psize = BPE_GET_PSIZE(bp);
1308 void *data = abd_borrow_buf(zio->io_abd, psize);
1309 decode_embedded_bp_compressed(bp, data);
1310 abd_return_buf_copy(zio->io_abd, data, psize);
1312 ASSERT(!BP_IS_EMBEDDED(bp));
1313 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1316 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1317 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1319 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1320 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1322 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1323 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1325 return (ZIO_PIPELINE_CONTINUE);
1329 zio_write_bp_init(zio_t *zio)
1331 if (!IO_IS_ALLOCATING(zio))
1332 return (ZIO_PIPELINE_CONTINUE);
1334 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1336 if (zio->io_bp_override) {
1337 blkptr_t *bp = zio->io_bp;
1338 zio_prop_t *zp = &zio->io_prop;
1340 ASSERT(bp->blk_birth != zio->io_txg);
1341 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1343 *bp = *zio->io_bp_override;
1344 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1346 if (BP_IS_EMBEDDED(bp))
1347 return (ZIO_PIPELINE_CONTINUE);
1350 * If we've been overridden and nopwrite is set then
1351 * set the flag accordingly to indicate that a nopwrite
1352 * has already occurred.
1354 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1355 ASSERT(!zp->zp_dedup);
1356 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1357 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1358 return (ZIO_PIPELINE_CONTINUE);
1361 ASSERT(!zp->zp_nopwrite);
1363 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1364 return (ZIO_PIPELINE_CONTINUE);
1366 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1367 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1369 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1370 BP_SET_DEDUP(bp, 1);
1371 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1372 return (ZIO_PIPELINE_CONTINUE);
1376 * We were unable to handle this as an override bp, treat
1377 * it as a regular write I/O.
1379 zio->io_bp_override = NULL;
1380 *bp = zio->io_bp_orig;
1381 zio->io_pipeline = zio->io_orig_pipeline;
1384 return (ZIO_PIPELINE_CONTINUE);
1388 zio_write_compress(zio_t *zio)
1390 spa_t *spa = zio->io_spa;
1391 zio_prop_t *zp = &zio->io_prop;
1392 enum zio_compress compress = zp->zp_compress;
1393 blkptr_t *bp = zio->io_bp;
1394 uint64_t lsize = zio->io_lsize;
1395 uint64_t psize = zio->io_size;
1398 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1401 * If our children haven't all reached the ready stage,
1402 * wait for them and then repeat this pipeline stage.
1404 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1405 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1406 return (ZIO_PIPELINE_STOP);
1409 if (!IO_IS_ALLOCATING(zio))
1410 return (ZIO_PIPELINE_CONTINUE);
1412 if (zio->io_children_ready != NULL) {
1414 * Now that all our children are ready, run the callback
1415 * associated with this zio in case it wants to modify the
1416 * data to be written.
1418 ASSERT3U(zp->zp_level, >, 0);
1419 zio->io_children_ready(zio);
1422 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1423 ASSERT(zio->io_bp_override == NULL);
1425 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1427 * We're rewriting an existing block, which means we're
1428 * working on behalf of spa_sync(). For spa_sync() to
1429 * converge, it must eventually be the case that we don't
1430 * have to allocate new blocks. But compression changes
1431 * the blocksize, which forces a reallocate, and makes
1432 * convergence take longer. Therefore, after the first
1433 * few passes, stop compressing to ensure convergence.
1435 pass = spa_sync_pass(spa);
1437 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1438 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1439 ASSERT(!BP_GET_DEDUP(bp));
1441 if (pass >= zfs_sync_pass_dont_compress)
1442 compress = ZIO_COMPRESS_OFF;
1444 /* Make sure someone doesn't change their mind on overwrites */
1445 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1446 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1449 /* If it's a compressed write that is not raw, compress the buffer. */
1450 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1451 void *cbuf = zio_buf_alloc(lsize);
1452 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1453 if (psize == 0 || psize == lsize) {
1454 compress = ZIO_COMPRESS_OFF;
1455 zio_buf_free(cbuf, lsize);
1456 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1457 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1458 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1459 encode_embedded_bp_compressed(bp,
1460 cbuf, compress, lsize, psize);
1461 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1462 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1463 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1464 zio_buf_free(cbuf, lsize);
1465 bp->blk_birth = zio->io_txg;
1466 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1467 ASSERT(spa_feature_is_active(spa,
1468 SPA_FEATURE_EMBEDDED_DATA));
1469 return (ZIO_PIPELINE_CONTINUE);
1472 * Round up compressed size up to the ashift
1473 * of the smallest-ashift device, and zero the tail.
1474 * This ensures that the compressed size of the BP
1475 * (and thus compressratio property) are correct,
1476 * in that we charge for the padding used to fill out
1479 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1480 size_t rounded = (size_t)P2ROUNDUP(psize,
1481 1ULL << spa->spa_min_ashift);
1482 if (rounded >= lsize) {
1483 compress = ZIO_COMPRESS_OFF;
1484 zio_buf_free(cbuf, lsize);
1487 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1488 abd_take_ownership_of_buf(cdata, B_TRUE);
1489 abd_zero_off(cdata, psize, rounded - psize);
1491 zio_push_transform(zio, cdata,
1492 psize, lsize, NULL);
1497 * We were unable to handle this as an override bp, treat
1498 * it as a regular write I/O.
1500 zio->io_bp_override = NULL;
1501 *bp = zio->io_bp_orig;
1502 zio->io_pipeline = zio->io_orig_pipeline;
1504 ASSERT3U(psize, !=, 0);
1508 * The final pass of spa_sync() must be all rewrites, but the first
1509 * few passes offer a trade-off: allocating blocks defers convergence,
1510 * but newly allocated blocks are sequential, so they can be written
1511 * to disk faster. Therefore, we allow the first few passes of
1512 * spa_sync() to allocate new blocks, but force rewrites after that.
1513 * There should only be a handful of blocks after pass 1 in any case.
1515 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1516 BP_GET_PSIZE(bp) == psize &&
1517 pass >= zfs_sync_pass_rewrite) {
1519 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1520 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1521 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1524 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1528 if (zio->io_bp_orig.blk_birth != 0 &&
1529 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1530 BP_SET_LSIZE(bp, lsize);
1531 BP_SET_TYPE(bp, zp->zp_type);
1532 BP_SET_LEVEL(bp, zp->zp_level);
1533 BP_SET_BIRTH(bp, zio->io_txg, 0);
1535 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1537 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1538 BP_SET_LSIZE(bp, lsize);
1539 BP_SET_TYPE(bp, zp->zp_type);
1540 BP_SET_LEVEL(bp, zp->zp_level);
1541 BP_SET_PSIZE(bp, psize);
1542 BP_SET_COMPRESS(bp, compress);
1543 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1544 BP_SET_DEDUP(bp, zp->zp_dedup);
1545 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1547 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1548 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1549 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1551 if (zp->zp_nopwrite) {
1552 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1553 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1554 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1557 return (ZIO_PIPELINE_CONTINUE);
1561 zio_free_bp_init(zio_t *zio)
1563 blkptr_t *bp = zio->io_bp;
1565 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1566 if (BP_GET_DEDUP(bp))
1567 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1570 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1572 return (ZIO_PIPELINE_CONTINUE);
1576 * ==========================================================================
1577 * Execute the I/O pipeline
1578 * ==========================================================================
1582 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1584 spa_t *spa = zio->io_spa;
1585 zio_type_t t = zio->io_type;
1586 int flags = (cutinline ? TQ_FRONT : 0);
1588 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1591 * If we're a config writer or a probe, the normal issue and
1592 * interrupt threads may all be blocked waiting for the config lock.
1593 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1595 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1599 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1601 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1605 * If this is a high priority I/O, then use the high priority taskq if
1608 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1609 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1612 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1615 * NB: We are assuming that the zio can only be dispatched
1616 * to a single taskq at a time. It would be a grievous error
1617 * to dispatch the zio to another taskq at the same time.
1619 #if defined(illumos) || !defined(_KERNEL)
1620 ASSERT(zio->io_tqent.tqent_next == NULL);
1622 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1624 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1625 flags, &zio->io_tqent);
1629 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1631 kthread_t *executor = zio->io_executor;
1632 spa_t *spa = zio->io_spa;
1634 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1635 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1637 for (i = 0; i < tqs->stqs_count; i++) {
1638 if (taskq_member(tqs->stqs_taskq[i], executor))
1647 zio_issue_async(zio_t *zio)
1649 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1651 return (ZIO_PIPELINE_STOP);
1655 zio_interrupt(zio_t *zio)
1657 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1661 zio_delay_interrupt(zio_t *zio)
1664 * The timeout_generic() function isn't defined in userspace, so
1665 * rather than trying to implement the function, the zio delay
1666 * functionality has been disabled for userspace builds.
1671 * If io_target_timestamp is zero, then no delay has been registered
1672 * for this IO, thus jump to the end of this function and "skip" the
1673 * delay; issuing it directly to the zio layer.
1675 if (zio->io_target_timestamp != 0) {
1676 hrtime_t now = gethrtime();
1678 if (now >= zio->io_target_timestamp) {
1680 * This IO has already taken longer than the target
1681 * delay to complete, so we don't want to delay it
1682 * any longer; we "miss" the delay and issue it
1683 * directly to the zio layer. This is likely due to
1684 * the target latency being set to a value less than
1685 * the underlying hardware can satisfy (e.g. delay
1686 * set to 1ms, but the disks take 10ms to complete an
1690 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1695 hrtime_t diff = zio->io_target_timestamp - now;
1697 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1698 hrtime_t, now, hrtime_t, diff);
1700 (void) timeout_generic(CALLOUT_NORMAL,
1701 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1708 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1713 * Execute the I/O pipeline until one of the following occurs:
1715 * (1) the I/O completes
1716 * (2) the pipeline stalls waiting for dependent child I/Os
1717 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1718 * (4) the I/O is delegated by vdev-level caching or aggregation
1719 * (5) the I/O is deferred due to vdev-level queueing
1720 * (6) the I/O is handed off to another thread.
1722 * In all cases, the pipeline stops whenever there's no CPU work; it never
1723 * burns a thread in cv_wait().
1725 * There's no locking on io_stage because there's no legitimate way
1726 * for multiple threads to be attempting to process the same I/O.
1728 static zio_pipe_stage_t *zio_pipeline[];
1731 zio_execute(zio_t *zio)
1733 zio->io_executor = curthread;
1735 ASSERT3U(zio->io_queued_timestamp, >, 0);
1737 while (zio->io_stage < ZIO_STAGE_DONE) {
1738 enum zio_stage pipeline = zio->io_pipeline;
1739 enum zio_stage stage = zio->io_stage;
1742 ASSERT(!MUTEX_HELD(&zio->io_lock));
1743 ASSERT(ISP2(stage));
1744 ASSERT(zio->io_stall == NULL);
1748 } while ((stage & pipeline) == 0);
1750 ASSERT(stage <= ZIO_STAGE_DONE);
1753 * If we are in interrupt context and this pipeline stage
1754 * will grab a config lock that is held across I/O,
1755 * or may wait for an I/O that needs an interrupt thread
1756 * to complete, issue async to avoid deadlock.
1758 * For VDEV_IO_START, we cut in line so that the io will
1759 * be sent to disk promptly.
1761 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1762 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1763 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1764 zio_requeue_io_start_cut_in_line : B_FALSE;
1765 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1769 zio->io_stage = stage;
1770 zio->io_pipeline_trace |= zio->io_stage;
1771 rv = zio_pipeline[highbit64(stage) - 1](zio);
1773 if (rv == ZIO_PIPELINE_STOP)
1776 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1781 * ==========================================================================
1782 * Initiate I/O, either sync or async
1783 * ==========================================================================
1786 zio_wait(zio_t *zio)
1790 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1791 ASSERT3P(zio->io_executor, ==, NULL);
1793 zio->io_waiter = curthread;
1794 ASSERT0(zio->io_queued_timestamp);
1795 zio->io_queued_timestamp = gethrtime();
1799 mutex_enter(&zio->io_lock);
1800 while (zio->io_executor != NULL)
1801 cv_wait(&zio->io_cv, &zio->io_lock);
1802 mutex_exit(&zio->io_lock);
1804 error = zio->io_error;
1811 zio_nowait(zio_t *zio)
1813 ASSERT3P(zio->io_executor, ==, NULL);
1815 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1816 zio_unique_parent(zio) == NULL) {
1818 * This is a logical async I/O with no parent to wait for it.
1819 * We add it to the spa_async_root_zio "Godfather" I/O which
1820 * will ensure they complete prior to unloading the pool.
1822 spa_t *spa = zio->io_spa;
1824 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1827 ASSERT0(zio->io_queued_timestamp);
1828 zio->io_queued_timestamp = gethrtime();
1833 * ==========================================================================
1834 * Reexecute, cancel, or suspend/resume failed I/O
1835 * ==========================================================================
1839 zio_reexecute(zio_t *pio)
1841 zio_t *cio, *cio_next;
1843 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1844 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1845 ASSERT(pio->io_gang_leader == NULL);
1846 ASSERT(pio->io_gang_tree == NULL);
1848 pio->io_flags = pio->io_orig_flags;
1849 pio->io_stage = pio->io_orig_stage;
1850 pio->io_pipeline = pio->io_orig_pipeline;
1851 pio->io_reexecute = 0;
1852 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1853 pio->io_pipeline_trace = 0;
1855 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1856 pio->io_state[w] = 0;
1857 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1858 pio->io_child_error[c] = 0;
1860 if (IO_IS_ALLOCATING(pio))
1861 BP_ZERO(pio->io_bp);
1864 * As we reexecute pio's children, new children could be created.
1865 * New children go to the head of pio's io_child_list, however,
1866 * so we will (correctly) not reexecute them. The key is that
1867 * the remainder of pio's io_child_list, from 'cio_next' onward,
1868 * cannot be affected by any side effects of reexecuting 'cio'.
1870 zio_link_t *zl = NULL;
1871 mutex_enter(&pio->io_lock);
1872 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1873 cio_next = zio_walk_children(pio, &zl);
1874 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1875 pio->io_children[cio->io_child_type][w]++;
1876 mutex_exit(&pio->io_lock);
1878 mutex_enter(&pio->io_lock);
1880 mutex_exit(&pio->io_lock);
1883 * Now that all children have been reexecuted, execute the parent.
1884 * We don't reexecute "The Godfather" I/O here as it's the
1885 * responsibility of the caller to wait on it.
1887 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1888 pio->io_queued_timestamp = gethrtime();
1894 zio_suspend(spa_t *spa, zio_t *zio)
1896 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1897 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1898 "failure and the failure mode property for this pool "
1899 "is set to panic.", spa_name(spa));
1901 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1903 mutex_enter(&spa->spa_suspend_lock);
1905 if (spa->spa_suspend_zio_root == NULL)
1906 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1907 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1908 ZIO_FLAG_GODFATHER);
1910 spa->spa_suspended = B_TRUE;
1913 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1914 ASSERT(zio != spa->spa_suspend_zio_root);
1915 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1916 ASSERT(zio_unique_parent(zio) == NULL);
1917 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1918 zio_add_child(spa->spa_suspend_zio_root, zio);
1921 mutex_exit(&spa->spa_suspend_lock);
1925 zio_resume(spa_t *spa)
1930 * Reexecute all previously suspended i/o.
1932 mutex_enter(&spa->spa_suspend_lock);
1933 spa->spa_suspended = B_FALSE;
1934 cv_broadcast(&spa->spa_suspend_cv);
1935 pio = spa->spa_suspend_zio_root;
1936 spa->spa_suspend_zio_root = NULL;
1937 mutex_exit(&spa->spa_suspend_lock);
1943 return (zio_wait(pio));
1947 zio_resume_wait(spa_t *spa)
1949 mutex_enter(&spa->spa_suspend_lock);
1950 while (spa_suspended(spa))
1951 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1952 mutex_exit(&spa->spa_suspend_lock);
1956 * ==========================================================================
1959 * A gang block is a collection of small blocks that looks to the DMU
1960 * like one large block. When zio_dva_allocate() cannot find a block
1961 * of the requested size, due to either severe fragmentation or the pool
1962 * being nearly full, it calls zio_write_gang_block() to construct the
1963 * block from smaller fragments.
1965 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1966 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1967 * an indirect block: it's an array of block pointers. It consumes
1968 * only one sector and hence is allocatable regardless of fragmentation.
1969 * The gang header's bps point to its gang members, which hold the data.
1971 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1972 * as the verifier to ensure uniqueness of the SHA256 checksum.
1973 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1974 * not the gang header. This ensures that data block signatures (needed for
1975 * deduplication) are independent of how the block is physically stored.
1977 * Gang blocks can be nested: a gang member may itself be a gang block.
1978 * Thus every gang block is a tree in which root and all interior nodes are
1979 * gang headers, and the leaves are normal blocks that contain user data.
1980 * The root of the gang tree is called the gang leader.
1982 * To perform any operation (read, rewrite, free, claim) on a gang block,
1983 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1984 * in the io_gang_tree field of the original logical i/o by recursively
1985 * reading the gang leader and all gang headers below it. This yields
1986 * an in-core tree containing the contents of every gang header and the
1987 * bps for every constituent of the gang block.
1989 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1990 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1991 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1992 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1993 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1994 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1995 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1996 * of the gang header plus zio_checksum_compute() of the data to update the
1997 * gang header's blk_cksum as described above.
1999 * The two-phase assemble/issue model solves the problem of partial failure --
2000 * what if you'd freed part of a gang block but then couldn't read the
2001 * gang header for another part? Assembling the entire gang tree first
2002 * ensures that all the necessary gang header I/O has succeeded before
2003 * starting the actual work of free, claim, or write. Once the gang tree
2004 * is assembled, free and claim are in-memory operations that cannot fail.
2006 * In the event that a gang write fails, zio_dva_unallocate() walks the
2007 * gang tree to immediately free (i.e. insert back into the space map)
2008 * everything we've allocated. This ensures that we don't get ENOSPC
2009 * errors during repeated suspend/resume cycles due to a flaky device.
2011 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2012 * the gang tree, we won't modify the block, so we can safely defer the free
2013 * (knowing that the block is still intact). If we *can* assemble the gang
2014 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2015 * each constituent bp and we can allocate a new block on the next sync pass.
2017 * In all cases, the gang tree allows complete recovery from partial failure.
2018 * ==========================================================================
2022 zio_gang_issue_func_done(zio_t *zio)
2024 abd_put(zio->io_abd);
2028 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2034 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2035 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2036 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2037 &pio->io_bookmark));
2041 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2048 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2049 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2050 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2051 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2054 * As we rewrite each gang header, the pipeline will compute
2055 * a new gang block header checksum for it; but no one will
2056 * compute a new data checksum, so we do that here. The one
2057 * exception is the gang leader: the pipeline already computed
2058 * its data checksum because that stage precedes gang assembly.
2059 * (Presently, nothing actually uses interior data checksums;
2060 * this is just good hygiene.)
2062 if (gn != pio->io_gang_leader->io_gang_tree) {
2063 abd_t *buf = abd_get_offset(data, offset);
2065 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2066 buf, BP_GET_PSIZE(bp));
2071 * If we are here to damage data for testing purposes,
2072 * leave the GBH alone so that we can detect the damage.
2074 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2075 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2077 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2078 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2079 zio_gang_issue_func_done, NULL, pio->io_priority,
2080 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2088 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2091 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2092 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2093 ZIO_GANG_CHILD_FLAGS(pio)));
2098 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2101 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2102 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2105 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2114 static void zio_gang_tree_assemble_done(zio_t *zio);
2116 static zio_gang_node_t *
2117 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2119 zio_gang_node_t *gn;
2121 ASSERT(*gnpp == NULL);
2123 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2124 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2131 zio_gang_node_free(zio_gang_node_t **gnpp)
2133 zio_gang_node_t *gn = *gnpp;
2135 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2136 ASSERT(gn->gn_child[g] == NULL);
2138 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2139 kmem_free(gn, sizeof (*gn));
2144 zio_gang_tree_free(zio_gang_node_t **gnpp)
2146 zio_gang_node_t *gn = *gnpp;
2151 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2152 zio_gang_tree_free(&gn->gn_child[g]);
2154 zio_gang_node_free(gnpp);
2158 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2160 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2161 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2163 ASSERT(gio->io_gang_leader == gio);
2164 ASSERT(BP_IS_GANG(bp));
2166 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2167 zio_gang_tree_assemble_done, gn, gio->io_priority,
2168 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2172 zio_gang_tree_assemble_done(zio_t *zio)
2174 zio_t *gio = zio->io_gang_leader;
2175 zio_gang_node_t *gn = zio->io_private;
2176 blkptr_t *bp = zio->io_bp;
2178 ASSERT(gio == zio_unique_parent(zio));
2179 ASSERT(zio->io_child_count == 0);
2184 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2185 if (BP_SHOULD_BYTESWAP(bp))
2186 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2188 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2189 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2190 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2192 abd_put(zio->io_abd);
2194 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2195 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2196 if (!BP_IS_GANG(gbp))
2198 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2203 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2206 zio_t *gio = pio->io_gang_leader;
2209 ASSERT(BP_IS_GANG(bp) == !!gn);
2210 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2211 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2214 * If you're a gang header, your data is in gn->gn_gbh.
2215 * If you're a gang member, your data is in 'data' and gn == NULL.
2217 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2220 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2222 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2223 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2224 if (BP_IS_HOLE(gbp))
2226 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2228 offset += BP_GET_PSIZE(gbp);
2232 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2233 ASSERT3U(gio->io_size, ==, offset);
2240 zio_gang_assemble(zio_t *zio)
2242 blkptr_t *bp = zio->io_bp;
2244 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2245 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2247 zio->io_gang_leader = zio;
2249 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2251 return (ZIO_PIPELINE_CONTINUE);
2255 zio_gang_issue(zio_t *zio)
2257 blkptr_t *bp = zio->io_bp;
2259 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2260 return (ZIO_PIPELINE_STOP);
2263 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2264 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2266 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2267 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2270 zio_gang_tree_free(&zio->io_gang_tree);
2272 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2274 return (ZIO_PIPELINE_CONTINUE);
2278 zio_write_gang_member_ready(zio_t *zio)
2280 zio_t *pio = zio_unique_parent(zio);
2281 zio_t *gio = zio->io_gang_leader;
2282 dva_t *cdva = zio->io_bp->blk_dva;
2283 dva_t *pdva = pio->io_bp->blk_dva;
2286 if (BP_IS_HOLE(zio->io_bp))
2289 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2291 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2292 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2293 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2294 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2295 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2297 mutex_enter(&pio->io_lock);
2298 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2299 ASSERT(DVA_GET_GANG(&pdva[d]));
2300 asize = DVA_GET_ASIZE(&pdva[d]);
2301 asize += DVA_GET_ASIZE(&cdva[d]);
2302 DVA_SET_ASIZE(&pdva[d], asize);
2304 mutex_exit(&pio->io_lock);
2308 zio_write_gang_done(zio_t *zio)
2310 abd_put(zio->io_abd);
2314 zio_write_gang_block(zio_t *pio)
2316 spa_t *spa = pio->io_spa;
2317 metaslab_class_t *mc = spa_normal_class(spa);
2318 blkptr_t *bp = pio->io_bp;
2319 zio_t *gio = pio->io_gang_leader;
2321 zio_gang_node_t *gn, **gnpp;
2322 zio_gbh_phys_t *gbh;
2324 uint64_t txg = pio->io_txg;
2325 uint64_t resid = pio->io_size;
2327 int copies = gio->io_prop.zp_copies;
2328 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2332 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2333 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2334 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2335 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2337 flags |= METASLAB_ASYNC_ALLOC;
2338 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2341 * The logical zio has already placed a reservation for
2342 * 'copies' allocation slots but gang blocks may require
2343 * additional copies. These additional copies
2344 * (i.e. gbh_copies - copies) are guaranteed to succeed
2345 * since metaslab_class_throttle_reserve() always allows
2346 * additional reservations for gang blocks.
2348 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2352 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2353 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2354 &pio->io_alloc_list, pio);
2356 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2357 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2358 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2361 * If we failed to allocate the gang block header then
2362 * we remove any additional allocation reservations that
2363 * we placed here. The original reservation will
2364 * be removed when the logical I/O goes to the ready
2367 metaslab_class_throttle_unreserve(mc,
2368 gbh_copies - copies, pio);
2370 pio->io_error = error;
2371 return (ZIO_PIPELINE_CONTINUE);
2375 gnpp = &gio->io_gang_tree;
2377 gnpp = pio->io_private;
2378 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2381 gn = zio_gang_node_alloc(gnpp);
2383 bzero(gbh, SPA_GANGBLOCKSIZE);
2384 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2387 * Create the gang header.
2389 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2390 zio_write_gang_done, NULL, pio->io_priority,
2391 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2394 * Create and nowait the gang children.
2396 for (int g = 0; resid != 0; resid -= lsize, g++) {
2397 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2399 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2401 zp.zp_checksum = gio->io_prop.zp_checksum;
2402 zp.zp_compress = ZIO_COMPRESS_OFF;
2403 zp.zp_type = DMU_OT_NONE;
2405 zp.zp_copies = gio->io_prop.zp_copies;
2406 zp.zp_dedup = B_FALSE;
2407 zp.zp_dedup_verify = B_FALSE;
2408 zp.zp_nopwrite = B_FALSE;
2410 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2411 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2412 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2413 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2414 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2416 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2417 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2418 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2421 * Gang children won't throttle but we should
2422 * account for their work, so reserve an allocation
2423 * slot for them here.
2425 VERIFY(metaslab_class_throttle_reserve(mc,
2426 zp.zp_copies, cio, flags));
2432 * Set pio's pipeline to just wait for zio to finish.
2434 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2438 return (ZIO_PIPELINE_CONTINUE);
2442 * The zio_nop_write stage in the pipeline determines if allocating a
2443 * new bp is necessary. The nopwrite feature can handle writes in
2444 * either syncing or open context (i.e. zil writes) and as a result is
2445 * mutually exclusive with dedup.
2447 * By leveraging a cryptographically secure checksum, such as SHA256, we
2448 * can compare the checksums of the new data and the old to determine if
2449 * allocating a new block is required. Note that our requirements for
2450 * cryptographic strength are fairly weak: there can't be any accidental
2451 * hash collisions, but we don't need to be secure against intentional
2452 * (malicious) collisions. To trigger a nopwrite, you have to be able
2453 * to write the file to begin with, and triggering an incorrect (hash
2454 * collision) nopwrite is no worse than simply writing to the file.
2455 * That said, there are no known attacks against the checksum algorithms
2456 * used for nopwrite, assuming that the salt and the checksums
2457 * themselves remain secret.
2460 zio_nop_write(zio_t *zio)
2462 blkptr_t *bp = zio->io_bp;
2463 blkptr_t *bp_orig = &zio->io_bp_orig;
2464 zio_prop_t *zp = &zio->io_prop;
2466 ASSERT(BP_GET_LEVEL(bp) == 0);
2467 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2468 ASSERT(zp->zp_nopwrite);
2469 ASSERT(!zp->zp_dedup);
2470 ASSERT(zio->io_bp_override == NULL);
2471 ASSERT(IO_IS_ALLOCATING(zio));
2474 * Check to see if the original bp and the new bp have matching
2475 * characteristics (i.e. same checksum, compression algorithms, etc).
2476 * If they don't then just continue with the pipeline which will
2477 * allocate a new bp.
2479 if (BP_IS_HOLE(bp_orig) ||
2480 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2481 ZCHECKSUM_FLAG_NOPWRITE) ||
2482 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2483 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2484 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2485 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2486 return (ZIO_PIPELINE_CONTINUE);
2489 * If the checksums match then reset the pipeline so that we
2490 * avoid allocating a new bp and issuing any I/O.
2492 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2493 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2494 ZCHECKSUM_FLAG_NOPWRITE);
2495 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2496 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2497 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2498 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2499 sizeof (uint64_t)) == 0);
2502 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2503 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2506 return (ZIO_PIPELINE_CONTINUE);
2510 * ==========================================================================
2512 * ==========================================================================
2515 zio_ddt_child_read_done(zio_t *zio)
2517 blkptr_t *bp = zio->io_bp;
2518 ddt_entry_t *dde = zio->io_private;
2520 zio_t *pio = zio_unique_parent(zio);
2522 mutex_enter(&pio->io_lock);
2523 ddp = ddt_phys_select(dde, bp);
2524 if (zio->io_error == 0)
2525 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2527 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2528 dde->dde_repair_abd = zio->io_abd;
2530 abd_free(zio->io_abd);
2531 mutex_exit(&pio->io_lock);
2535 zio_ddt_read_start(zio_t *zio)
2537 blkptr_t *bp = zio->io_bp;
2539 ASSERT(BP_GET_DEDUP(bp));
2540 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2541 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2543 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2544 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2545 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2546 ddt_phys_t *ddp = dde->dde_phys;
2547 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2550 ASSERT(zio->io_vsd == NULL);
2553 if (ddp_self == NULL)
2554 return (ZIO_PIPELINE_CONTINUE);
2556 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2557 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2559 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2561 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2562 abd_alloc_for_io(zio->io_size, B_TRUE),
2563 zio->io_size, zio_ddt_child_read_done, dde,
2564 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2565 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2567 return (ZIO_PIPELINE_CONTINUE);
2570 zio_nowait(zio_read(zio, zio->io_spa, bp,
2571 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2572 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2574 return (ZIO_PIPELINE_CONTINUE);
2578 zio_ddt_read_done(zio_t *zio)
2580 blkptr_t *bp = zio->io_bp;
2582 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2583 return (ZIO_PIPELINE_STOP);
2586 ASSERT(BP_GET_DEDUP(bp));
2587 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2588 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2590 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2591 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2592 ddt_entry_t *dde = zio->io_vsd;
2594 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2595 return (ZIO_PIPELINE_CONTINUE);
2598 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2599 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2600 return (ZIO_PIPELINE_STOP);
2602 if (dde->dde_repair_abd != NULL) {
2603 abd_copy(zio->io_abd, dde->dde_repair_abd,
2605 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2607 ddt_repair_done(ddt, dde);
2611 ASSERT(zio->io_vsd == NULL);
2613 return (ZIO_PIPELINE_CONTINUE);
2617 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2619 spa_t *spa = zio->io_spa;
2620 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2622 /* We should never get a raw, override zio */
2623 ASSERT(!(zio->io_bp_override && do_raw));
2626 * Note: we compare the original data, not the transformed data,
2627 * because when zio->io_bp is an override bp, we will not have
2628 * pushed the I/O transforms. That's an important optimization
2629 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2631 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2632 zio_t *lio = dde->dde_lead_zio[p];
2635 return (lio->io_orig_size != zio->io_orig_size ||
2636 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2637 zio->io_orig_size) != 0);
2641 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2642 ddt_phys_t *ddp = &dde->dde_phys[p];
2644 if (ddp->ddp_phys_birth != 0) {
2645 arc_buf_t *abuf = NULL;
2646 arc_flags_t aflags = ARC_FLAG_WAIT;
2647 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2648 blkptr_t blk = *zio->io_bp;
2651 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2656 * Intuitively, it would make more sense to compare
2657 * io_abd than io_orig_abd in the raw case since you
2658 * don't want to look at any transformations that have
2659 * happened to the data. However, for raw I/Os the
2660 * data will actually be the same in io_abd and
2661 * io_orig_abd, so all we have to do is issue this as
2665 zio_flags |= ZIO_FLAG_RAW;
2666 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2667 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2669 ASSERT3P(zio->io_transform_stack, ==, NULL);
2672 error = arc_read(NULL, spa, &blk,
2673 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2674 zio_flags, &aflags, &zio->io_bookmark);
2677 if (arc_buf_size(abuf) != zio->io_orig_size ||
2678 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2679 zio->io_orig_size) != 0)
2680 error = SET_ERROR(EEXIST);
2681 arc_buf_destroy(abuf, &abuf);
2685 return (error != 0);
2693 zio_ddt_child_write_ready(zio_t *zio)
2695 int p = zio->io_prop.zp_copies;
2696 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2697 ddt_entry_t *dde = zio->io_private;
2698 ddt_phys_t *ddp = &dde->dde_phys[p];
2706 ASSERT(dde->dde_lead_zio[p] == zio);
2708 ddt_phys_fill(ddp, zio->io_bp);
2710 zio_link_t *zl = NULL;
2711 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2712 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2718 zio_ddt_child_write_done(zio_t *zio)
2720 int p = zio->io_prop.zp_copies;
2721 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2722 ddt_entry_t *dde = zio->io_private;
2723 ddt_phys_t *ddp = &dde->dde_phys[p];
2727 ASSERT(ddp->ddp_refcnt == 0);
2728 ASSERT(dde->dde_lead_zio[p] == zio);
2729 dde->dde_lead_zio[p] = NULL;
2731 if (zio->io_error == 0) {
2732 zio_link_t *zl = NULL;
2733 while (zio_walk_parents(zio, &zl) != NULL)
2734 ddt_phys_addref(ddp);
2736 ddt_phys_clear(ddp);
2743 zio_ddt_ditto_write_done(zio_t *zio)
2745 int p = DDT_PHYS_DITTO;
2746 zio_prop_t *zp = &zio->io_prop;
2747 blkptr_t *bp = zio->io_bp;
2748 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2749 ddt_entry_t *dde = zio->io_private;
2750 ddt_phys_t *ddp = &dde->dde_phys[p];
2751 ddt_key_t *ddk = &dde->dde_key;
2755 ASSERT(ddp->ddp_refcnt == 0);
2756 ASSERT(dde->dde_lead_zio[p] == zio);
2757 dde->dde_lead_zio[p] = NULL;
2759 if (zio->io_error == 0) {
2760 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2761 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2762 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2763 if (ddp->ddp_phys_birth != 0)
2764 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2765 ddt_phys_fill(ddp, bp);
2772 zio_ddt_write(zio_t *zio)
2774 spa_t *spa = zio->io_spa;
2775 blkptr_t *bp = zio->io_bp;
2776 uint64_t txg = zio->io_txg;
2777 zio_prop_t *zp = &zio->io_prop;
2778 int p = zp->zp_copies;
2782 ddt_t *ddt = ddt_select(spa, bp);
2786 ASSERT(BP_GET_DEDUP(bp));
2787 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2788 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2789 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2792 dde = ddt_lookup(ddt, bp, B_TRUE);
2793 ddp = &dde->dde_phys[p];
2795 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2797 * If we're using a weak checksum, upgrade to a strong checksum
2798 * and try again. If we're already using a strong checksum,
2799 * we can't resolve it, so just convert to an ordinary write.
2800 * (And automatically e-mail a paper to Nature?)
2802 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2803 ZCHECKSUM_FLAG_DEDUP)) {
2804 zp->zp_checksum = spa_dedup_checksum(spa);
2805 zio_pop_transforms(zio);
2806 zio->io_stage = ZIO_STAGE_OPEN;
2809 zp->zp_dedup = B_FALSE;
2810 BP_SET_DEDUP(bp, B_FALSE);
2812 ASSERT(!BP_GET_DEDUP(bp));
2813 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2815 return (ZIO_PIPELINE_CONTINUE);
2818 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2819 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2821 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2822 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2823 zio_prop_t czp = *zp;
2825 czp.zp_copies = ditto_copies;
2828 * If we arrived here with an override bp, we won't have run
2829 * the transform stack, so we won't have the data we need to
2830 * generate a child i/o. So, toss the override bp and restart.
2831 * This is safe, because using the override bp is just an
2832 * optimization; and it's rare, so the cost doesn't matter.
2834 if (zio->io_bp_override) {
2835 zio_pop_transforms(zio);
2836 zio->io_stage = ZIO_STAGE_OPEN;
2837 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2838 zio->io_bp_override = NULL;
2841 return (ZIO_PIPELINE_CONTINUE);
2844 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2845 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2846 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2847 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2849 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2850 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2853 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2854 if (ddp->ddp_phys_birth != 0)
2855 ddt_bp_fill(ddp, bp, txg);
2856 if (dde->dde_lead_zio[p] != NULL)
2857 zio_add_child(zio, dde->dde_lead_zio[p]);
2859 ddt_phys_addref(ddp);
2860 } else if (zio->io_bp_override) {
2861 ASSERT(bp->blk_birth == txg);
2862 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2863 ddt_phys_fill(ddp, bp);
2864 ddt_phys_addref(ddp);
2866 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2867 zio->io_orig_size, zio->io_orig_size, zp,
2868 zio_ddt_child_write_ready, NULL, NULL,
2869 zio_ddt_child_write_done, dde, zio->io_priority,
2870 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2872 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2873 dde->dde_lead_zio[p] = cio;
2883 return (ZIO_PIPELINE_CONTINUE);
2886 ddt_entry_t *freedde; /* for debugging */
2889 zio_ddt_free(zio_t *zio)
2891 spa_t *spa = zio->io_spa;
2892 blkptr_t *bp = zio->io_bp;
2893 ddt_t *ddt = ddt_select(spa, bp);
2897 ASSERT(BP_GET_DEDUP(bp));
2898 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2901 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2902 ddp = ddt_phys_select(dde, bp);
2903 ddt_phys_decref(ddp);
2906 return (ZIO_PIPELINE_CONTINUE);
2910 * ==========================================================================
2911 * Allocate and free blocks
2912 * ==========================================================================
2916 zio_io_to_allocate(spa_t *spa)
2920 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2922 zio = avl_first(&spa->spa_alloc_tree);
2926 ASSERT(IO_IS_ALLOCATING(zio));
2929 * Try to place a reservation for this zio. If we're unable to
2930 * reserve then we throttle.
2932 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2933 zio->io_prop.zp_copies, zio, 0)) {
2937 avl_remove(&spa->spa_alloc_tree, zio);
2938 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2944 zio_dva_throttle(zio_t *zio)
2946 spa_t *spa = zio->io_spa;
2949 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2950 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2951 zio->io_child_type == ZIO_CHILD_GANG ||
2952 zio->io_flags & ZIO_FLAG_NODATA) {
2953 return (ZIO_PIPELINE_CONTINUE);
2956 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2958 ASSERT3U(zio->io_queued_timestamp, >, 0);
2959 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2961 mutex_enter(&spa->spa_alloc_lock);
2963 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2964 avl_add(&spa->spa_alloc_tree, zio);
2966 nio = zio_io_to_allocate(zio->io_spa);
2967 mutex_exit(&spa->spa_alloc_lock);
2970 return (ZIO_PIPELINE_CONTINUE);
2973 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2975 * We are passing control to a new zio so make sure that
2976 * it is processed by a different thread. We do this to
2977 * avoid stack overflows that can occur when parents are
2978 * throttled and children are making progress. We allow
2979 * it to go to the head of the taskq since it's already
2982 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2984 return (ZIO_PIPELINE_STOP);
2988 zio_allocate_dispatch(spa_t *spa)
2992 mutex_enter(&spa->spa_alloc_lock);
2993 zio = zio_io_to_allocate(spa);
2994 mutex_exit(&spa->spa_alloc_lock);
2998 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2999 ASSERT0(zio->io_error);
3000 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3004 zio_dva_allocate(zio_t *zio)
3006 spa_t *spa = zio->io_spa;
3007 metaslab_class_t *mc = spa_normal_class(spa);
3008 blkptr_t *bp = zio->io_bp;
3012 if (zio->io_gang_leader == NULL) {
3013 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3014 zio->io_gang_leader = zio;
3017 ASSERT(BP_IS_HOLE(bp));
3018 ASSERT0(BP_GET_NDVAS(bp));
3019 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3020 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3021 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3023 if (zio->io_flags & ZIO_FLAG_NODATA) {
3024 flags |= METASLAB_DONT_THROTTLE;
3026 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3027 flags |= METASLAB_GANG_CHILD;
3029 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
3030 flags |= METASLAB_ASYNC_ALLOC;
3033 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3034 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3035 &zio->io_alloc_list, zio);
3038 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3039 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3041 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3042 return (zio_write_gang_block(zio));
3043 zio->io_error = error;
3046 return (ZIO_PIPELINE_CONTINUE);
3050 zio_dva_free(zio_t *zio)
3052 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3054 return (ZIO_PIPELINE_CONTINUE);
3058 zio_dva_claim(zio_t *zio)
3062 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3064 zio->io_error = error;
3066 return (ZIO_PIPELINE_CONTINUE);
3070 * Undo an allocation. This is used by zio_done() when an I/O fails
3071 * and we want to give back the block we just allocated.
3072 * This handles both normal blocks and gang blocks.
3075 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3077 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3078 ASSERT(zio->io_bp_override == NULL);
3080 if (!BP_IS_HOLE(bp))
3081 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3084 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3085 zio_dva_unallocate(zio, gn->gn_child[g],
3086 &gn->gn_gbh->zg_blkptr[g]);
3092 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3095 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3096 uint64_t size, boolean_t *slog)
3099 zio_alloc_list_t io_alloc_list;
3101 ASSERT(txg > spa_syncing_txg(spa));
3103 metaslab_trace_init(&io_alloc_list);
3104 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3105 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3109 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3110 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3111 &io_alloc_list, NULL);
3115 metaslab_trace_fini(&io_alloc_list);
3118 BP_SET_LSIZE(new_bp, size);
3119 BP_SET_PSIZE(new_bp, size);
3120 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3121 BP_SET_CHECKSUM(new_bp,
3122 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3123 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3124 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3125 BP_SET_LEVEL(new_bp, 0);
3126 BP_SET_DEDUP(new_bp, 0);
3127 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3129 zfs_dbgmsg("%s: zil block allocation failure: "
3130 "size %llu, error %d", spa_name(spa), size, error);
3137 * ==========================================================================
3138 * Read, write and delete to physical devices
3139 * ==========================================================================
3144 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3145 * stops after this stage and will resume upon I/O completion.
3146 * However, there are instances where the vdev layer may need to
3147 * continue the pipeline when an I/O was not issued. Since the I/O
3148 * that was sent to the vdev layer might be different than the one
3149 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3150 * force the underlying vdev layers to call either zio_execute() or
3151 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3154 zio_vdev_io_start(zio_t *zio)
3156 vdev_t *vd = zio->io_vd;
3158 spa_t *spa = zio->io_spa;
3161 ASSERT(zio->io_error == 0);
3162 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3165 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3166 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3169 * The mirror_ops handle multiple DVAs in a single BP.
3171 vdev_mirror_ops.vdev_op_io_start(zio);
3172 return (ZIO_PIPELINE_STOP);
3175 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3176 zio->io_priority == ZIO_PRIORITY_NOW) {
3177 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3178 return (ZIO_PIPELINE_CONTINUE);
3181 ASSERT3P(zio->io_logical, !=, zio);
3182 if (zio->io_type == ZIO_TYPE_WRITE) {
3183 ASSERT(spa->spa_trust_config);
3185 if (zio->io_vd->vdev_removing) {
3186 ASSERT(zio->io_flags &
3187 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3188 ZIO_FLAG_INDUCE_DAMAGE));
3193 * We keep track of time-sensitive I/Os so that the scan thread
3194 * can quickly react to certain workloads. In particular, we care
3195 * about non-scrubbing, top-level reads and writes with the following
3197 * - synchronous writes of user data to non-slog devices
3198 * - any reads of user data
3199 * When these conditions are met, adjust the timestamp of spa_last_io
3200 * which allows the scan thread to adjust its workload accordingly.
3202 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3203 vd == vd->vdev_top && !vd->vdev_islog &&
3204 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3205 zio->io_txg != spa_syncing_txg(spa)) {
3206 uint64_t old = spa->spa_last_io;
3207 uint64_t new = ddi_get_lbolt64();
3209 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3211 align = 1ULL << vd->vdev_top->vdev_ashift;
3213 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3214 P2PHASE(zio->io_size, align) != 0) {
3215 /* Transform logical writes to be a full physical block size. */
3216 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3218 if (zio->io_type == ZIO_TYPE_READ ||
3219 zio->io_type == ZIO_TYPE_WRITE)
3220 abuf = abd_alloc_sametype(zio->io_abd, asize);
3221 ASSERT(vd == vd->vdev_top);
3222 if (zio->io_type == ZIO_TYPE_WRITE) {
3223 abd_copy(abuf, zio->io_abd, zio->io_size);
3224 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3226 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3231 * If this is not a physical io, make sure that it is properly aligned
3232 * before proceeding.
3234 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3235 ASSERT0(P2PHASE(zio->io_offset, align));
3236 ASSERT0(P2PHASE(zio->io_size, align));
3239 * For the physical io we allow alignment
3240 * to a logical block size.
3242 uint64_t log_align =
3243 1ULL << vd->vdev_top->vdev_logical_ashift;
3244 ASSERT0(P2PHASE(zio->io_offset, log_align));
3245 ASSERT0(P2PHASE(zio->io_size, log_align));
3248 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3251 * If this is a repair I/O, and there's no self-healing involved --
3252 * that is, we're just resilvering what we expect to resilver --
3253 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3254 * This prevents spurious resilvering with nested replication.
3255 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3256 * A is out of date, we'll read from C+D, then use the data to
3257 * resilver A+B -- but we don't actually want to resilver B, just A.
3258 * The top-level mirror has no way to know this, so instead we just
3259 * discard unnecessary repairs as we work our way down the vdev tree.
3260 * The same logic applies to any form of nested replication:
3261 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3263 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3264 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3265 zio->io_txg != 0 && /* not a delegated i/o */
3266 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3267 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3268 zio_vdev_io_bypass(zio);
3269 return (ZIO_PIPELINE_CONTINUE);
3272 if (vd->vdev_ops->vdev_op_leaf) {
3273 switch (zio->io_type) {
3275 if (vdev_cache_read(zio))
3276 return (ZIO_PIPELINE_CONTINUE);
3278 case ZIO_TYPE_WRITE:
3280 if ((zio = vdev_queue_io(zio)) == NULL)
3281 return (ZIO_PIPELINE_STOP);
3283 if (!vdev_accessible(vd, zio)) {
3284 zio->io_error = SET_ERROR(ENXIO);
3286 return (ZIO_PIPELINE_STOP);
3291 * Note that we ignore repair writes for TRIM because they can
3292 * conflict with normal writes. This isn't an issue because, by
3293 * definition, we only repair blocks that aren't freed.
3295 if (zio->io_type == ZIO_TYPE_WRITE &&
3296 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3297 !trim_map_write_start(zio))
3298 return (ZIO_PIPELINE_STOP);
3301 vd->vdev_ops->vdev_op_io_start(zio);
3302 return (ZIO_PIPELINE_STOP);
3306 zio_vdev_io_done(zio_t *zio)
3308 vdev_t *vd = zio->io_vd;
3309 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3310 boolean_t unexpected_error = B_FALSE;
3312 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3313 return (ZIO_PIPELINE_STOP);
3316 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3317 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3319 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3320 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3321 zio->io_type == ZIO_TYPE_FREE)) {
3323 if (zio->io_type == ZIO_TYPE_WRITE &&
3324 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3325 trim_map_write_done(zio);
3327 vdev_queue_io_done(zio);
3329 if (zio->io_type == ZIO_TYPE_WRITE)
3330 vdev_cache_write(zio);
3332 if (zio_injection_enabled && zio->io_error == 0)
3333 zio->io_error = zio_handle_device_injection(vd,
3336 if (zio_injection_enabled && zio->io_error == 0)
3337 zio->io_error = zio_handle_label_injection(zio, EIO);
3339 if (zio->io_error) {
3340 if (zio->io_error == ENOTSUP &&
3341 zio->io_type == ZIO_TYPE_FREE) {
3342 /* Not all devices support TRIM. */
3343 } else if (!vdev_accessible(vd, zio)) {
3344 zio->io_error = SET_ERROR(ENXIO);
3346 unexpected_error = B_TRUE;
3351 ops->vdev_op_io_done(zio);
3353 if (unexpected_error)
3354 VERIFY(vdev_probe(vd, zio) == NULL);
3356 return (ZIO_PIPELINE_CONTINUE);
3360 * This function is used to change the priority of an existing zio that is
3361 * currently in-flight. This is used by the arc to upgrade priority in the
3362 * event that a demand read is made for a block that is currently queued
3363 * as a scrub or async read IO. Otherwise, the high priority read request
3364 * would end up having to wait for the lower priority IO.
3367 zio_change_priority(zio_t *pio, zio_priority_t priority)
3369 zio_t *cio, *cio_next;
3370 zio_link_t *zl = NULL;
3372 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3374 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3375 vdev_queue_change_io_priority(pio, priority);
3377 pio->io_priority = priority;
3380 mutex_enter(&pio->io_lock);
3381 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3382 cio_next = zio_walk_children(pio, &zl);
3383 zio_change_priority(cio, priority);
3385 mutex_exit(&pio->io_lock);
3389 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3390 * disk, and use that to finish the checksum ereport later.
3393 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3394 const void *good_buf)
3396 /* no processing needed */
3397 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3402 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3404 void *buf = zio_buf_alloc(zio->io_size);
3406 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3408 zcr->zcr_cbinfo = zio->io_size;
3409 zcr->zcr_cbdata = buf;
3410 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3411 zcr->zcr_free = zio_buf_free;
3415 zio_vdev_io_assess(zio_t *zio)
3417 vdev_t *vd = zio->io_vd;
3419 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3420 return (ZIO_PIPELINE_STOP);
3423 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3424 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3426 if (zio->io_vsd != NULL) {
3427 zio->io_vsd_ops->vsd_free(zio);
3431 if (zio_injection_enabled && zio->io_error == 0)
3432 zio->io_error = zio_handle_fault_injection(zio, EIO);
3434 if (zio->io_type == ZIO_TYPE_FREE &&
3435 zio->io_priority != ZIO_PRIORITY_NOW) {
3436 switch (zio->io_error) {
3438 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3439 ZIO_TRIM_STAT_BUMP(success);
3442 ZIO_TRIM_STAT_BUMP(unsupported);
3445 ZIO_TRIM_STAT_BUMP(failed);
3451 * If the I/O failed, determine whether we should attempt to retry it.
3453 * On retry, we cut in line in the issue queue, since we don't want
3454 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3456 if (zio->io_error && vd == NULL &&
3457 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3458 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3459 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3461 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3462 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3463 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3464 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3465 zio_requeue_io_start_cut_in_line);
3466 return (ZIO_PIPELINE_STOP);
3470 * If we got an error on a leaf device, convert it to ENXIO
3471 * if the device is not accessible at all.
3473 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3474 !vdev_accessible(vd, zio))
3475 zio->io_error = SET_ERROR(ENXIO);
3478 * If we can't write to an interior vdev (mirror or RAID-Z),
3479 * set vdev_cant_write so that we stop trying to allocate from it.
3481 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3482 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3483 vd->vdev_cant_write = B_TRUE;
3487 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3488 * attempts will ever succeed. In this case we set a persistent bit so
3489 * that we don't bother with it in the future.
3491 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3492 zio->io_type == ZIO_TYPE_IOCTL &&
3493 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3494 vd->vdev_nowritecache = B_TRUE;
3497 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3499 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3500 zio->io_physdone != NULL) {
3501 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3502 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3503 zio->io_physdone(zio->io_logical);
3506 return (ZIO_PIPELINE_CONTINUE);
3510 zio_vdev_io_reissue(zio_t *zio)
3512 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3513 ASSERT(zio->io_error == 0);
3515 zio->io_stage >>= 1;
3519 zio_vdev_io_redone(zio_t *zio)
3521 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3523 zio->io_stage >>= 1;
3527 zio_vdev_io_bypass(zio_t *zio)
3529 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3530 ASSERT(zio->io_error == 0);
3532 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3533 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3537 * ==========================================================================
3538 * Generate and verify checksums
3539 * ==========================================================================
3542 zio_checksum_generate(zio_t *zio)
3544 blkptr_t *bp = zio->io_bp;
3545 enum zio_checksum checksum;
3549 * This is zio_write_phys().
3550 * We're either generating a label checksum, or none at all.
3552 checksum = zio->io_prop.zp_checksum;
3554 if (checksum == ZIO_CHECKSUM_OFF)
3555 return (ZIO_PIPELINE_CONTINUE);
3557 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3559 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3560 ASSERT(!IO_IS_ALLOCATING(zio));
3561 checksum = ZIO_CHECKSUM_GANG_HEADER;
3563 checksum = BP_GET_CHECKSUM(bp);
3567 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3569 return (ZIO_PIPELINE_CONTINUE);
3573 zio_checksum_verify(zio_t *zio)
3575 zio_bad_cksum_t info;
3576 blkptr_t *bp = zio->io_bp;
3579 ASSERT(zio->io_vd != NULL);
3583 * This is zio_read_phys().
3584 * We're either verifying a label checksum, or nothing at all.
3586 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3587 return (ZIO_PIPELINE_CONTINUE);
3589 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3592 if ((error = zio_checksum_error(zio, &info)) != 0) {
3593 zio->io_error = error;
3594 if (error == ECKSUM &&
3595 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3596 zfs_ereport_start_checksum(zio->io_spa,
3597 zio->io_vd, zio, zio->io_offset,
3598 zio->io_size, NULL, &info);
3602 return (ZIO_PIPELINE_CONTINUE);
3606 * Called by RAID-Z to ensure we don't compute the checksum twice.
3609 zio_checksum_verified(zio_t *zio)
3611 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3615 * ==========================================================================
3616 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3617 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3618 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3619 * indicate errors that are specific to one I/O, and most likely permanent.
3620 * Any other error is presumed to be worse because we weren't expecting it.
3621 * ==========================================================================
3624 zio_worst_error(int e1, int e2)
3626 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3629 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3630 if (e1 == zio_error_rank[r1])
3633 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3634 if (e2 == zio_error_rank[r2])
3637 return (r1 > r2 ? e1 : e2);
3641 * ==========================================================================
3643 * ==========================================================================
3646 zio_ready(zio_t *zio)
3648 blkptr_t *bp = zio->io_bp;
3649 zio_t *pio, *pio_next;
3650 zio_link_t *zl = NULL;
3652 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3654 return (ZIO_PIPELINE_STOP);
3657 if (zio->io_ready) {
3658 ASSERT(IO_IS_ALLOCATING(zio));
3659 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3660 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3661 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3666 if (bp != NULL && bp != &zio->io_bp_copy)
3667 zio->io_bp_copy = *bp;
3669 if (zio->io_error != 0) {
3670 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3672 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3673 ASSERT(IO_IS_ALLOCATING(zio));
3674 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3676 * We were unable to allocate anything, unreserve and
3677 * issue the next I/O to allocate.
3679 metaslab_class_throttle_unreserve(
3680 spa_normal_class(zio->io_spa),
3681 zio->io_prop.zp_copies, zio);
3682 zio_allocate_dispatch(zio->io_spa);
3686 mutex_enter(&zio->io_lock);
3687 zio->io_state[ZIO_WAIT_READY] = 1;
3688 pio = zio_walk_parents(zio, &zl);
3689 mutex_exit(&zio->io_lock);
3692 * As we notify zio's parents, new parents could be added.
3693 * New parents go to the head of zio's io_parent_list, however,
3694 * so we will (correctly) not notify them. The remainder of zio's
3695 * io_parent_list, from 'pio_next' onward, cannot change because
3696 * all parents must wait for us to be done before they can be done.
3698 for (; pio != NULL; pio = pio_next) {
3699 pio_next = zio_walk_parents(zio, &zl);
3700 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3703 if (zio->io_flags & ZIO_FLAG_NODATA) {
3704 if (BP_IS_GANG(bp)) {
3705 zio->io_flags &= ~ZIO_FLAG_NODATA;
3707 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3708 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3712 if (zio_injection_enabled &&
3713 zio->io_spa->spa_syncing_txg == zio->io_txg)
3714 zio_handle_ignored_writes(zio);
3716 return (ZIO_PIPELINE_CONTINUE);
3720 * Update the allocation throttle accounting.
3723 zio_dva_throttle_done(zio_t *zio)
3725 zio_t *lio = zio->io_logical;
3726 zio_t *pio = zio_unique_parent(zio);
3727 vdev_t *vd = zio->io_vd;
3728 int flags = METASLAB_ASYNC_ALLOC;
3730 ASSERT3P(zio->io_bp, !=, NULL);
3731 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3732 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3733 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3735 ASSERT3P(vd, ==, vd->vdev_top);
3736 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3737 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3738 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3739 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3742 * Parents of gang children can have two flavors -- ones that
3743 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3744 * and ones that allocated the constituent blocks. The allocation
3745 * throttle needs to know the allocating parent zio so we must find
3748 if (pio->io_child_type == ZIO_CHILD_GANG) {
3750 * If our parent is a rewrite gang child then our grandparent
3751 * would have been the one that performed the allocation.
3753 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3754 pio = zio_unique_parent(pio);
3755 flags |= METASLAB_GANG_CHILD;
3758 ASSERT(IO_IS_ALLOCATING(pio));
3759 ASSERT3P(zio, !=, zio->io_logical);
3760 ASSERT(zio->io_logical != NULL);
3761 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3762 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3764 mutex_enter(&pio->io_lock);
3765 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3766 mutex_exit(&pio->io_lock);
3768 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3772 * Call into the pipeline to see if there is more work that
3773 * needs to be done. If there is work to be done it will be
3774 * dispatched to another taskq thread.
3776 zio_allocate_dispatch(zio->io_spa);
3780 zio_done(zio_t *zio)
3782 spa_t *spa = zio->io_spa;
3783 zio_t *lio = zio->io_logical;
3784 blkptr_t *bp = zio->io_bp;
3785 vdev_t *vd = zio->io_vd;
3786 uint64_t psize = zio->io_size;
3787 zio_t *pio, *pio_next;
3788 metaslab_class_t *mc = spa_normal_class(spa);
3789 zio_link_t *zl = NULL;
3792 * If our children haven't all completed,
3793 * wait for them and then repeat this pipeline stage.
3795 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3796 return (ZIO_PIPELINE_STOP);
3800 * If the allocation throttle is enabled, then update the accounting.
3801 * We only track child I/Os that are part of an allocating async
3802 * write. We must do this since the allocation is performed
3803 * by the logical I/O but the actual write is done by child I/Os.
3805 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3806 zio->io_child_type == ZIO_CHILD_VDEV) {
3807 ASSERT(mc->mc_alloc_throttle_enabled);
3808 zio_dva_throttle_done(zio);
3812 * If the allocation throttle is enabled, verify that
3813 * we have decremented the refcounts for every I/O that was throttled.
3815 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3816 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3817 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3819 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3820 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3823 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3824 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3825 ASSERT(zio->io_children[c][w] == 0);
3827 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3828 ASSERT(bp->blk_pad[0] == 0);
3829 ASSERT(bp->blk_pad[1] == 0);
3830 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3831 (bp == zio_unique_parent(zio)->io_bp));
3832 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3833 zio->io_bp_override == NULL &&
3834 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3835 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3836 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3837 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3838 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3840 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3841 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3845 * If there were child vdev/gang/ddt errors, they apply to us now.
3847 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3848 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3849 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3852 * If the I/O on the transformed data was successful, generate any
3853 * checksum reports now while we still have the transformed data.
3855 if (zio->io_error == 0) {
3856 while (zio->io_cksum_report != NULL) {
3857 zio_cksum_report_t *zcr = zio->io_cksum_report;
3858 uint64_t align = zcr->zcr_align;
3859 uint64_t asize = P2ROUNDUP(psize, align);
3861 abd_t *adata = zio->io_abd;
3863 if (asize != psize) {
3864 adata = abd_alloc_linear(asize, B_TRUE);
3865 abd_copy(adata, zio->io_abd, psize);
3866 abd_zero_off(adata, psize, asize - psize);
3870 abuf = abd_borrow_buf_copy(adata, asize);
3872 zio->io_cksum_report = zcr->zcr_next;
3873 zcr->zcr_next = NULL;
3874 zcr->zcr_finish(zcr, abuf);
3875 zfs_ereport_free_checksum(zcr);
3878 abd_return_buf(adata, abuf, asize);
3885 zio_pop_transforms(zio); /* note: may set zio->io_error */
3887 vdev_stat_update(zio, psize);
3889 if (zio->io_error) {
3891 * If this I/O is attached to a particular vdev,
3892 * generate an error message describing the I/O failure
3893 * at the block level. We ignore these errors if the
3894 * device is currently unavailable.
3896 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3897 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3899 if ((zio->io_error == EIO || !(zio->io_flags &
3900 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3903 * For logical I/O requests, tell the SPA to log the
3904 * error and generate a logical data ereport.
3906 spa_log_error(spa, zio);
3907 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3912 if (zio->io_error && zio == lio) {
3914 * Determine whether zio should be reexecuted. This will
3915 * propagate all the way to the root via zio_notify_parent().
3917 ASSERT(vd == NULL && bp != NULL);
3918 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3920 if (IO_IS_ALLOCATING(zio) &&
3921 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3922 if (zio->io_error != ENOSPC)
3923 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3925 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3928 if ((zio->io_type == ZIO_TYPE_READ ||
3929 zio->io_type == ZIO_TYPE_FREE) &&
3930 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3931 zio->io_error == ENXIO &&
3932 spa_load_state(spa) == SPA_LOAD_NONE &&
3933 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3934 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3936 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3937 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3940 * Here is a possibly good place to attempt to do
3941 * either combinatorial reconstruction or error correction
3942 * based on checksums. It also might be a good place
3943 * to send out preliminary ereports before we suspend
3949 * If there were logical child errors, they apply to us now.
3950 * We defer this until now to avoid conflating logical child
3951 * errors with errors that happened to the zio itself when
3952 * updating vdev stats and reporting FMA events above.
3954 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3956 if ((zio->io_error || zio->io_reexecute) &&
3957 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3958 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3959 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3961 zio_gang_tree_free(&zio->io_gang_tree);
3964 * Godfather I/Os should never suspend.
3966 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3967 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3968 zio->io_reexecute = 0;
3970 if (zio->io_reexecute) {
3972 * This is a logical I/O that wants to reexecute.
3974 * Reexecute is top-down. When an i/o fails, if it's not
3975 * the root, it simply notifies its parent and sticks around.
3976 * The parent, seeing that it still has children in zio_done(),
3977 * does the same. This percolates all the way up to the root.
3978 * The root i/o will reexecute or suspend the entire tree.
3980 * This approach ensures that zio_reexecute() honors
3981 * all the original i/o dependency relationships, e.g.
3982 * parents not executing until children are ready.
3984 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3986 zio->io_gang_leader = NULL;
3988 mutex_enter(&zio->io_lock);
3989 zio->io_state[ZIO_WAIT_DONE] = 1;
3990 mutex_exit(&zio->io_lock);
3993 * "The Godfather" I/O monitors its children but is
3994 * not a true parent to them. It will track them through
3995 * the pipeline but severs its ties whenever they get into
3996 * trouble (e.g. suspended). This allows "The Godfather"
3997 * I/O to return status without blocking.
4000 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4002 zio_link_t *remove_zl = zl;
4003 pio_next = zio_walk_parents(zio, &zl);
4005 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4006 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4007 zio_remove_child(pio, zio, remove_zl);
4008 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4012 if ((pio = zio_unique_parent(zio)) != NULL) {
4014 * We're not a root i/o, so there's nothing to do
4015 * but notify our parent. Don't propagate errors
4016 * upward since we haven't permanently failed yet.
4018 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4019 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4020 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4021 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4023 * We'd fail again if we reexecuted now, so suspend
4024 * until conditions improve (e.g. device comes online).
4026 zio_suspend(spa, zio);
4029 * Reexecution is potentially a huge amount of work.
4030 * Hand it off to the otherwise-unused claim taskq.
4032 #if defined(illumos) || !defined(_KERNEL)
4033 ASSERT(zio->io_tqent.tqent_next == NULL);
4035 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4037 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4038 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4041 return (ZIO_PIPELINE_STOP);
4044 ASSERT(zio->io_child_count == 0);
4045 ASSERT(zio->io_reexecute == 0);
4046 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4049 * Report any checksum errors, since the I/O is complete.
4051 while (zio->io_cksum_report != NULL) {
4052 zio_cksum_report_t *zcr = zio->io_cksum_report;
4053 zio->io_cksum_report = zcr->zcr_next;
4054 zcr->zcr_next = NULL;
4055 zcr->zcr_finish(zcr, NULL);
4056 zfs_ereport_free_checksum(zcr);
4060 * It is the responsibility of the done callback to ensure that this
4061 * particular zio is no longer discoverable for adoption, and as
4062 * such, cannot acquire any new parents.
4067 mutex_enter(&zio->io_lock);
4068 zio->io_state[ZIO_WAIT_DONE] = 1;
4069 mutex_exit(&zio->io_lock);
4072 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4073 zio_link_t *remove_zl = zl;
4074 pio_next = zio_walk_parents(zio, &zl);
4075 zio_remove_child(pio, zio, remove_zl);
4076 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4079 if (zio->io_waiter != NULL) {
4080 mutex_enter(&zio->io_lock);
4081 zio->io_executor = NULL;
4082 cv_broadcast(&zio->io_cv);
4083 mutex_exit(&zio->io_lock);
4088 return (ZIO_PIPELINE_STOP);
4092 * ==========================================================================
4093 * I/O pipeline definition
4094 * ==========================================================================
4096 static zio_pipe_stage_t *zio_pipeline[] = {
4103 zio_checksum_generate,
4119 zio_checksum_verify,
4127 * Compare two zbookmark_phys_t's to see which we would reach first in a
4128 * pre-order traversal of the object tree.
4130 * This is simple in every case aside from the meta-dnode object. For all other
4131 * objects, we traverse them in order (object 1 before object 2, and so on).
4132 * However, all of these objects are traversed while traversing object 0, since
4133 * the data it points to is the list of objects. Thus, we need to convert to a
4134 * canonical representation so we can compare meta-dnode bookmarks to
4135 * non-meta-dnode bookmarks.
4137 * We do this by calculating "equivalents" for each field of the zbookmark.
4138 * zbookmarks outside of the meta-dnode use their own object and level, and
4139 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4140 * blocks this bookmark refers to) by multiplying their blkid by their span
4141 * (the number of L0 blocks contained within one block at their level).
4142 * zbookmarks inside the meta-dnode calculate their object equivalent
4143 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4144 * level + 1<<31 (any value larger than a level could ever be) for their level.
4145 * This causes them to always compare before a bookmark in their object
4146 * equivalent, compare appropriately to bookmarks in other objects, and to
4147 * compare appropriately to other bookmarks in the meta-dnode.
4150 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4151 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4154 * These variables represent the "equivalent" values for the zbookmark,
4155 * after converting zbookmarks inside the meta dnode to their
4156 * normal-object equivalents.
4158 uint64_t zb1obj, zb2obj;
4159 uint64_t zb1L0, zb2L0;
4160 uint64_t zb1level, zb2level;
4162 if (zb1->zb_object == zb2->zb_object &&
4163 zb1->zb_level == zb2->zb_level &&
4164 zb1->zb_blkid == zb2->zb_blkid)
4168 * BP_SPANB calculates the span in blocks.
4170 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4171 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4173 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4174 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4176 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4178 zb1obj = zb1->zb_object;
4179 zb1level = zb1->zb_level;
4182 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4183 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4185 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4187 zb2obj = zb2->zb_object;
4188 zb2level = zb2->zb_level;
4191 /* Now that we have a canonical representation, do the comparison. */
4192 if (zb1obj != zb2obj)
4193 return (zb1obj < zb2obj ? -1 : 1);
4194 else if (zb1L0 != zb2L0)
4195 return (zb1L0 < zb2L0 ? -1 : 1);
4196 else if (zb1level != zb2level)
4197 return (zb1level > zb2level ? -1 : 1);
4199 * This can (theoretically) happen if the bookmarks have the same object
4200 * and level, but different blkids, if the block sizes are not the same.
4201 * There is presently no way to change the indirect block sizes
4207 * This function checks the following: given that last_block is the place that
4208 * our traversal stopped last time, does that guarantee that we've visited
4209 * every node under subtree_root? Therefore, we can't just use the raw output
4210 * of zbookmark_compare. We have to pass in a modified version of
4211 * subtree_root; by incrementing the block id, and then checking whether
4212 * last_block is before or equal to that, we can tell whether or not having
4213 * visited last_block implies that all of subtree_root's children have been
4217 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4218 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4220 zbookmark_phys_t mod_zb = *subtree_root;
4222 ASSERT(last_block->zb_level == 0);
4224 /* The objset_phys_t isn't before anything. */
4229 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4230 * data block size in sectors, because that variable is only used if
4231 * the bookmark refers to a block in the meta-dnode. Since we don't
4232 * know without examining it what object it refers to, and there's no
4233 * harm in passing in this value in other cases, we always pass it in.
4235 * We pass in 0 for the indirect block size shift because zb2 must be
4236 * level 0. The indirect block size is only used to calculate the span
4237 * of the bookmark, but since the bookmark must be level 0, the span is
4238 * always 1, so the math works out.
4240 * If you make changes to how the zbookmark_compare code works, be sure
4241 * to make sure that this code still works afterwards.
4243 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4244 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,