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, 2018 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>
47 #include <sys/cityhash.h>
49 SYSCTL_DECL(_vfs_zfs);
50 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
51 #if defined(__amd64__)
52 static int zio_use_uma = 1;
54 static int zio_use_uma = 0;
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
57 "Use uma(9) for ZIO allocations");
58 static int zio_exclude_metadata = 0;
59 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
60 "Exclude metadata buffers from dumps as well");
62 zio_trim_stats_t zio_trim_stats = {
63 { "bytes", KSTAT_DATA_UINT64,
64 "Number of bytes successfully TRIMmed" },
65 { "success", KSTAT_DATA_UINT64,
66 "Number of successful TRIM requests" },
67 { "unsupported", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed because TRIM is not supported" },
69 { "failed", KSTAT_DATA_UINT64,
70 "Number of TRIM requests that failed for reasons other than not supported" },
73 static kstat_t *zio_trim_ksp;
76 * ==========================================================================
77 * I/O type descriptions
78 * ==========================================================================
80 const char *zio_type_name[ZIO_TYPES] = {
81 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
85 boolean_t zio_dva_throttle_enabled = B_TRUE;
86 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
87 &zio_dva_throttle_enabled, 0, "");
90 * ==========================================================================
92 * ==========================================================================
94 kmem_cache_t *zio_cache;
95 kmem_cache_t *zio_link_cache;
96 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
100 extern vmem_t *zio_alloc_arena;
103 #define ZIO_PIPELINE_CONTINUE 0x100
104 #define ZIO_PIPELINE_STOP 0x101
106 #define BP_SPANB(indblkshift, level) \
107 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
108 #define COMPARE_META_LEVEL 0x80000000ul
110 * The following actions directly effect the spa's sync-to-convergence logic.
111 * The values below define the sync pass when we start performing the action.
112 * Care should be taken when changing these values as they directly impact
113 * spa_sync() performance. Tuning these values may introduce subtle performance
114 * pathologies and should only be done in the context of performance analysis.
115 * These tunables will eventually be removed and replaced with #defines once
116 * enough analysis has been done to determine optimal values.
118 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
119 * regular blocks are not deferred.
121 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
123 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
124 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
126 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
127 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
128 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
129 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
132 * An allocating zio is one that either currently has the DVA allocate
133 * stage set or will have it later in its lifetime.
135 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
137 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
141 int zio_buf_debug_limit = 16384;
143 int zio_buf_debug_limit = 0;
147 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
153 zio_cache = kmem_cache_create("zio_cache",
154 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
155 zio_link_cache = kmem_cache_create("zio_link_cache",
156 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
161 * For small buffers, we want a cache for each multiple of
162 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
163 * for each quarter-power of 2.
165 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
166 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
169 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
177 * If we are using watchpoints, put each buffer on its own page,
178 * to eliminate the performance overhead of trapping to the
179 * kernel when modifying a non-watched buffer that shares the
180 * page with a watched buffer.
182 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
186 if (size <= 4 * SPA_MINBLOCKSIZE) {
187 align = SPA_MINBLOCKSIZE;
188 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
189 align = MIN(p2 >> 2, PAGESIZE);
194 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
195 zio_buf_cache[c] = kmem_cache_create(name, size,
196 align, NULL, NULL, NULL, NULL, NULL, cflags);
199 * Since zio_data bufs do not appear in crash dumps, we
200 * pass KMC_NOTOUCH so that no allocator metadata is
201 * stored with the buffers.
203 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
204 zio_data_buf_cache[c] = kmem_cache_create(name, size,
205 align, NULL, NULL, NULL, NULL, NULL,
206 cflags | KMC_NOTOUCH | KMC_NODEBUG);
211 ASSERT(zio_buf_cache[c] != NULL);
212 if (zio_buf_cache[c - 1] == NULL)
213 zio_buf_cache[c - 1] = zio_buf_cache[c];
215 ASSERT(zio_data_buf_cache[c] != NULL);
216 if (zio_data_buf_cache[c - 1] == NULL)
217 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
223 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
225 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
228 if (zio_trim_ksp != NULL) {
229 zio_trim_ksp->ks_data = &zio_trim_stats;
230 kstat_install(zio_trim_ksp);
238 kmem_cache_t *last_cache = NULL;
239 kmem_cache_t *last_data_cache = NULL;
241 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
242 if (zio_buf_cache[c] != last_cache) {
243 last_cache = zio_buf_cache[c];
244 kmem_cache_destroy(zio_buf_cache[c]);
246 zio_buf_cache[c] = NULL;
248 if (zio_data_buf_cache[c] != last_data_cache) {
249 last_data_cache = zio_data_buf_cache[c];
250 kmem_cache_destroy(zio_data_buf_cache[c]);
252 zio_data_buf_cache[c] = NULL;
255 kmem_cache_destroy(zio_link_cache);
256 kmem_cache_destroy(zio_cache);
260 if (zio_trim_ksp != NULL) {
261 kstat_delete(zio_trim_ksp);
267 * ==========================================================================
268 * Allocate and free I/O buffers
269 * ==========================================================================
273 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
274 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
275 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
276 * excess / transient data in-core during a crashdump.
279 zio_buf_alloc(size_t size)
281 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
282 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
284 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
287 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
289 return (kmem_alloc(size, KM_SLEEP|flags));
293 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
294 * crashdump if the kernel panics. This exists so that we will limit the amount
295 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
296 * of kernel heap dumped to disk when the kernel panics)
299 zio_data_buf_alloc(size_t size)
301 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
303 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
306 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
308 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
312 zio_buf_free(void *buf, size_t size)
314 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
316 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
319 kmem_cache_free(zio_buf_cache[c], buf);
321 kmem_free(buf, size);
325 zio_data_buf_free(void *buf, size_t size)
327 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
329 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
332 kmem_cache_free(zio_data_buf_cache[c], buf);
334 kmem_free(buf, size);
338 * ==========================================================================
339 * Push and pop I/O transform buffers
340 * ==========================================================================
343 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
344 zio_transform_func_t *transform)
346 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
349 * Ensure that anyone expecting this zio to contain a linear ABD isn't
350 * going to get a nasty surprise when they try to access the data.
353 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
355 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
356 abd_is_linear(data));
359 zt->zt_orig_abd = zio->io_abd;
360 zt->zt_orig_size = zio->io_size;
361 zt->zt_bufsize = bufsize;
362 zt->zt_transform = transform;
364 zt->zt_next = zio->io_transform_stack;
365 zio->io_transform_stack = zt;
372 zio_pop_transforms(zio_t *zio)
376 while ((zt = zio->io_transform_stack) != NULL) {
377 if (zt->zt_transform != NULL)
378 zt->zt_transform(zio,
379 zt->zt_orig_abd, zt->zt_orig_size);
381 if (zt->zt_bufsize != 0)
382 abd_free(zio->io_abd);
384 zio->io_abd = zt->zt_orig_abd;
385 zio->io_size = zt->zt_orig_size;
386 zio->io_transform_stack = zt->zt_next;
388 kmem_free(zt, sizeof (zio_transform_t));
393 * ==========================================================================
394 * I/O transform callbacks for subblocks and decompression
395 * ==========================================================================
398 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
400 ASSERT(zio->io_size > size);
402 if (zio->io_type == ZIO_TYPE_READ)
403 abd_copy(data, zio->io_abd, size);
407 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
409 if (zio->io_error == 0) {
410 void *tmp = abd_borrow_buf(data, size);
411 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
412 zio->io_abd, tmp, zio->io_size, size);
413 abd_return_buf_copy(data, tmp, size);
416 zio->io_error = SET_ERROR(EIO);
421 * ==========================================================================
422 * I/O parent/child relationships and pipeline interlocks
423 * ==========================================================================
426 zio_walk_parents(zio_t *cio, zio_link_t **zl)
428 list_t *pl = &cio->io_parent_list;
430 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
434 ASSERT((*zl)->zl_child == cio);
435 return ((*zl)->zl_parent);
439 zio_walk_children(zio_t *pio, zio_link_t **zl)
441 list_t *cl = &pio->io_child_list;
443 ASSERT(MUTEX_HELD(&pio->io_lock));
445 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
449 ASSERT((*zl)->zl_parent == pio);
450 return ((*zl)->zl_child);
454 zio_unique_parent(zio_t *cio)
456 zio_link_t *zl = NULL;
457 zio_t *pio = zio_walk_parents(cio, &zl);
459 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
464 zio_add_child(zio_t *pio, zio_t *cio)
466 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
469 * Logical I/Os can have logical, gang, or vdev children.
470 * Gang I/Os can have gang or vdev children.
471 * Vdev I/Os can only have vdev children.
472 * The following ASSERT captures all of these constraints.
474 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
479 mutex_enter(&pio->io_lock);
480 mutex_enter(&cio->io_lock);
482 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
484 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
485 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
487 list_insert_head(&pio->io_child_list, zl);
488 list_insert_head(&cio->io_parent_list, zl);
490 pio->io_child_count++;
491 cio->io_parent_count++;
493 mutex_exit(&cio->io_lock);
494 mutex_exit(&pio->io_lock);
498 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
500 ASSERT(zl->zl_parent == pio);
501 ASSERT(zl->zl_child == cio);
503 mutex_enter(&pio->io_lock);
504 mutex_enter(&cio->io_lock);
506 list_remove(&pio->io_child_list, zl);
507 list_remove(&cio->io_parent_list, zl);
509 pio->io_child_count--;
510 cio->io_parent_count--;
512 mutex_exit(&cio->io_lock);
513 mutex_exit(&pio->io_lock);
514 kmem_cache_free(zio_link_cache, zl);
518 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
520 boolean_t waiting = B_FALSE;
522 mutex_enter(&zio->io_lock);
523 ASSERT(zio->io_stall == NULL);
524 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
525 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
528 uint64_t *countp = &zio->io_children[c][wait];
531 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
532 zio->io_stall = countp;
537 mutex_exit(&zio->io_lock);
542 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
544 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
545 int *errorp = &pio->io_child_error[zio->io_child_type];
547 mutex_enter(&pio->io_lock);
548 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
549 *errorp = zio_worst_error(*errorp, zio->io_error);
550 pio->io_reexecute |= zio->io_reexecute;
551 ASSERT3U(*countp, >, 0);
555 if (*countp == 0 && pio->io_stall == countp) {
556 zio_taskq_type_t type =
557 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
559 pio->io_stall = NULL;
560 mutex_exit(&pio->io_lock);
562 * Dispatch the parent zio in its own taskq so that
563 * the child can continue to make progress. This also
564 * prevents overflowing the stack when we have deeply nested
565 * parent-child relationships.
567 zio_taskq_dispatch(pio, type, B_FALSE);
569 mutex_exit(&pio->io_lock);
574 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
576 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
577 zio->io_error = zio->io_child_error[c];
581 zio_bookmark_compare(const void *x1, const void *x2)
583 const zio_t *z1 = x1;
584 const zio_t *z2 = x2;
586 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
588 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
591 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
593 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
596 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
598 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
601 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
603 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
615 * ==========================================================================
616 * Create the various types of I/O (read, write, free, etc)
617 * ==========================================================================
620 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
621 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
622 void *private, zio_type_t type, zio_priority_t priority,
623 enum zio_flag flags, vdev_t *vd, uint64_t offset,
624 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
628 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
629 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
630 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
632 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
633 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
634 ASSERT(vd || stage == ZIO_STAGE_OPEN);
636 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
638 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
639 bzero(zio, sizeof (zio_t));
641 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
642 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
644 list_create(&zio->io_parent_list, sizeof (zio_link_t),
645 offsetof(zio_link_t, zl_parent_node));
646 list_create(&zio->io_child_list, sizeof (zio_link_t),
647 offsetof(zio_link_t, zl_child_node));
648 metaslab_trace_init(&zio->io_alloc_list);
651 zio->io_child_type = ZIO_CHILD_VDEV;
652 else if (flags & ZIO_FLAG_GANG_CHILD)
653 zio->io_child_type = ZIO_CHILD_GANG;
654 else if (flags & ZIO_FLAG_DDT_CHILD)
655 zio->io_child_type = ZIO_CHILD_DDT;
657 zio->io_child_type = ZIO_CHILD_LOGICAL;
660 zio->io_bp = (blkptr_t *)bp;
661 zio->io_bp_copy = *bp;
662 zio->io_bp_orig = *bp;
663 if (type != ZIO_TYPE_WRITE ||
664 zio->io_child_type == ZIO_CHILD_DDT)
665 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
666 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
667 zio->io_logical = zio;
668 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
669 pipeline |= ZIO_GANG_STAGES;
675 zio->io_private = private;
677 zio->io_priority = priority;
679 zio->io_offset = offset;
680 zio->io_orig_abd = zio->io_abd = data;
681 zio->io_orig_size = zio->io_size = psize;
682 zio->io_lsize = lsize;
683 zio->io_orig_flags = zio->io_flags = flags;
684 zio->io_orig_stage = zio->io_stage = stage;
685 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
686 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
688 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
689 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
692 zio->io_bookmark = *zb;
695 if (zio->io_logical == NULL)
696 zio->io_logical = pio->io_logical;
697 if (zio->io_child_type == ZIO_CHILD_GANG)
698 zio->io_gang_leader = pio->io_gang_leader;
699 zio_add_child(pio, zio);
706 zio_destroy(zio_t *zio)
708 metaslab_trace_fini(&zio->io_alloc_list);
709 list_destroy(&zio->io_parent_list);
710 list_destroy(&zio->io_child_list);
711 mutex_destroy(&zio->io_lock);
712 cv_destroy(&zio->io_cv);
713 kmem_cache_free(zio_cache, zio);
717 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
718 void *private, enum zio_flag flags)
722 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
723 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
724 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
730 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
732 return (zio_null(NULL, spa, NULL, done, private, flags));
736 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
738 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
739 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
740 bp, (longlong_t)BP_GET_TYPE(bp));
742 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
743 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
744 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
745 bp, (longlong_t)BP_GET_CHECKSUM(bp));
747 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
748 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
749 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
750 bp, (longlong_t)BP_GET_COMPRESS(bp));
752 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
753 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
754 bp, (longlong_t)BP_GET_LSIZE(bp));
756 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
757 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
758 bp, (longlong_t)BP_GET_PSIZE(bp));
761 if (BP_IS_EMBEDDED(bp)) {
762 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
763 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
764 bp, (longlong_t)BPE_GET_ETYPE(bp));
769 * Do not verify individual DVAs if the config is not trusted. This
770 * will be done once the zio is executed in vdev_mirror_map_alloc.
772 if (!spa->spa_trust_config)
776 * Pool-specific checks.
778 * Note: it would be nice to verify that the blk_birth and
779 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
780 * allows the birth time of log blocks (and dmu_sync()-ed blocks
781 * that are in the log) to be arbitrarily large.
783 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
784 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
785 if (vdevid >= spa->spa_root_vdev->vdev_children) {
786 zfs_panic_recover("blkptr at %p DVA %u has invalid "
788 bp, i, (longlong_t)vdevid);
791 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
793 zfs_panic_recover("blkptr at %p DVA %u has invalid "
795 bp, i, (longlong_t)vdevid);
798 if (vd->vdev_ops == &vdev_hole_ops) {
799 zfs_panic_recover("blkptr at %p DVA %u has hole "
801 bp, i, (longlong_t)vdevid);
804 if (vd->vdev_ops == &vdev_missing_ops) {
806 * "missing" vdevs are valid during import, but we
807 * don't have their detailed info (e.g. asize), so
808 * we can't perform any more checks on them.
812 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
813 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
815 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
816 if (offset + asize > vd->vdev_asize) {
817 zfs_panic_recover("blkptr at %p DVA %u has invalid "
819 bp, i, (longlong_t)offset);
825 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
827 uint64_t vdevid = DVA_GET_VDEV(dva);
829 if (vdevid >= spa->spa_root_vdev->vdev_children)
832 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
836 if (vd->vdev_ops == &vdev_hole_ops)
839 if (vd->vdev_ops == &vdev_missing_ops) {
843 uint64_t offset = DVA_GET_OFFSET(dva);
844 uint64_t asize = DVA_GET_ASIZE(dva);
847 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
848 if (offset + asize > vd->vdev_asize)
855 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
856 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
857 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
861 zfs_blkptr_verify(spa, bp);
863 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
864 data, size, size, done, private,
865 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
866 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
867 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
873 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
874 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
875 zio_done_func_t *ready, zio_done_func_t *children_ready,
876 zio_done_func_t *physdone, zio_done_func_t *done,
877 void *private, zio_priority_t priority, enum zio_flag flags,
878 const zbookmark_phys_t *zb)
882 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
883 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
884 zp->zp_compress >= ZIO_COMPRESS_OFF &&
885 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
886 DMU_OT_IS_VALID(zp->zp_type) &&
889 zp->zp_copies <= spa_max_replication(spa));
891 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
892 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
893 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
894 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
896 zio->io_ready = ready;
897 zio->io_children_ready = children_ready;
898 zio->io_physdone = physdone;
902 * Data can be NULL if we are going to call zio_write_override() to
903 * provide the already-allocated BP. But we may need the data to
904 * verify a dedup hit (if requested). In this case, don't try to
905 * dedup (just take the already-allocated BP verbatim).
907 if (data == NULL && zio->io_prop.zp_dedup_verify) {
908 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
915 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
916 uint64_t size, zio_done_func_t *done, void *private,
917 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
921 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
922 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
923 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
929 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
931 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
932 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
933 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
934 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
937 * We must reset the io_prop to match the values that existed
938 * when the bp was first written by dmu_sync() keeping in mind
939 * that nopwrite and dedup are mutually exclusive.
941 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
942 zio->io_prop.zp_nopwrite = nopwrite;
943 zio->io_prop.zp_copies = copies;
944 zio->io_bp_override = bp;
948 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
951 zfs_blkptr_verify(spa, bp);
954 * The check for EMBEDDED is a performance optimization. We
955 * process the free here (by ignoring it) rather than
956 * putting it on the list and then processing it in zio_free_sync().
958 if (BP_IS_EMBEDDED(bp))
960 metaslab_check_free(spa, bp);
963 * Frees that are for the currently-syncing txg, are not going to be
964 * deferred, and which will not need to do a read (i.e. not GANG or
965 * DEDUP), can be processed immediately. Otherwise, put them on the
966 * in-memory list for later processing.
968 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
969 txg != spa->spa_syncing_txg ||
970 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
971 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
973 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
974 BP_GET_PSIZE(bp), 0)));
979 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
980 uint64_t size, enum zio_flag flags)
983 enum zio_stage stage = ZIO_FREE_PIPELINE;
985 ASSERT(!BP_IS_HOLE(bp));
986 ASSERT(spa_syncing_txg(spa) == txg);
987 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
989 if (BP_IS_EMBEDDED(bp))
990 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
992 metaslab_check_free(spa, bp);
994 dsl_scan_freed(spa, bp);
996 if (zfs_trim_enabled)
997 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
998 ZIO_STAGE_VDEV_IO_ASSESS;
1000 * GANG and DEDUP blocks can induce a read (for the gang block header,
1001 * or the DDT), so issue them asynchronously so that this thread is
1004 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1005 stage |= ZIO_STAGE_ISSUE_ASYNC;
1007 flags |= ZIO_FLAG_DONT_QUEUE;
1009 zio = zio_create(pio, spa, txg, bp, NULL, size,
1010 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1011 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1017 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1018 zio_done_func_t *done, void *private, enum zio_flag flags)
1022 zfs_blkptr_verify(spa, bp);
1024 if (BP_IS_EMBEDDED(bp))
1025 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1028 * A claim is an allocation of a specific block. Claims are needed
1029 * to support immediate writes in the intent log. The issue is that
1030 * immediate writes contain committed data, but in a txg that was
1031 * *not* committed. Upon opening the pool after an unclean shutdown,
1032 * the intent log claims all blocks that contain immediate write data
1033 * so that the SPA knows they're in use.
1035 * All claims *must* be resolved in the first txg -- before the SPA
1036 * starts allocating blocks -- so that nothing is allocated twice.
1037 * If txg == 0 we just verify that the block is claimable.
1039 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1040 spa_min_claim_txg(spa));
1041 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1042 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1044 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1045 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1046 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1047 ASSERT0(zio->io_queued_timestamp);
1053 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1054 uint64_t size, zio_done_func_t *done, void *private,
1055 zio_priority_t priority, enum zio_flag flags)
1060 if (vd->vdev_children == 0) {
1061 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1062 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1063 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1067 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1069 for (c = 0; c < vd->vdev_children; c++)
1070 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1071 offset, size, done, private, priority, flags));
1078 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1079 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1080 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1084 ASSERT(vd->vdev_children == 0);
1085 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1086 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1087 ASSERT3U(offset + size, <=, vd->vdev_psize);
1089 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1090 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1091 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1093 zio->io_prop.zp_checksum = checksum;
1099 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1100 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1101 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1105 ASSERT(vd->vdev_children == 0);
1106 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1107 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1108 ASSERT3U(offset + size, <=, vd->vdev_psize);
1110 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1111 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1112 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1114 zio->io_prop.zp_checksum = checksum;
1116 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1118 * zec checksums are necessarily destructive -- they modify
1119 * the end of the write buffer to hold the verifier/checksum.
1120 * Therefore, we must make a local copy in case the data is
1121 * being written to multiple places in parallel.
1123 abd_t *wbuf = abd_alloc_sametype(data, size);
1124 abd_copy(wbuf, data, size);
1126 zio_push_transform(zio, wbuf, size, size, NULL);
1133 * Create a child I/O to do some work for us.
1136 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1137 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1138 enum zio_flag flags, zio_done_func_t *done, void *private)
1140 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1144 * vdev child I/Os do not propagate their error to the parent.
1145 * Therefore, for correct operation the caller *must* check for
1146 * and handle the error in the child i/o's done callback.
1147 * The only exceptions are i/os that we don't care about
1148 * (OPTIONAL or REPAIR).
1150 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1153 if (type == ZIO_TYPE_READ && bp != NULL) {
1155 * If we have the bp, then the child should perform the
1156 * checksum and the parent need not. This pushes error
1157 * detection as close to the leaves as possible and
1158 * eliminates redundant checksums in the interior nodes.
1160 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1161 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1164 /* Not all IO types require vdev io done stage e.g. free */
1165 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1166 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1168 if (vd->vdev_ops->vdev_op_leaf) {
1169 ASSERT0(vd->vdev_children);
1170 offset += VDEV_LABEL_START_SIZE;
1173 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1176 * If we've decided to do a repair, the write is not speculative --
1177 * even if the original read was.
1179 if (flags & ZIO_FLAG_IO_REPAIR)
1180 flags &= ~ZIO_FLAG_SPECULATIVE;
1183 * If we're creating a child I/O that is not associated with a
1184 * top-level vdev, then the child zio is not an allocating I/O.
1185 * If this is a retried I/O then we ignore it since we will
1186 * have already processed the original allocating I/O.
1188 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1189 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1190 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1192 ASSERT(mc->mc_alloc_throttle_enabled);
1193 ASSERT(type == ZIO_TYPE_WRITE);
1194 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1195 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1196 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1197 pio->io_child_type == ZIO_CHILD_GANG);
1199 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1202 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1203 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1204 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1205 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1207 zio->io_physdone = pio->io_physdone;
1208 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1209 zio->io_logical->io_phys_children++;
1215 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1216 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1217 zio_done_func_t *done, void *private)
1221 ASSERT(vd->vdev_ops->vdev_op_leaf);
1223 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1224 data, size, size, done, private, type, priority,
1225 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1227 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1233 zio_flush(zio_t *zio, vdev_t *vd)
1235 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1236 NULL, NULL, ZIO_PRIORITY_NOW,
1237 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1241 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1244 ASSERT(vd->vdev_ops->vdev_op_leaf);
1246 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1247 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1248 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1249 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1253 zio_shrink(zio_t *zio, uint64_t size)
1255 ASSERT3P(zio->io_executor, ==, NULL);
1256 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1257 ASSERT3U(size, <=, zio->io_size);
1260 * We don't shrink for raidz because of problems with the
1261 * reconstruction when reading back less than the block size.
1262 * Note, BP_IS_RAIDZ() assumes no compression.
1264 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1265 if (!BP_IS_RAIDZ(zio->io_bp)) {
1266 /* we are not doing a raw write */
1267 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1268 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1273 * ==========================================================================
1274 * Prepare to read and write logical blocks
1275 * ==========================================================================
1279 zio_read_bp_init(zio_t *zio)
1281 blkptr_t *bp = zio->io_bp;
1283 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1285 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1286 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1287 !(zio->io_flags & ZIO_FLAG_RAW)) {
1289 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1290 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1291 psize, psize, zio_decompress);
1294 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1295 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1297 int psize = BPE_GET_PSIZE(bp);
1298 void *data = abd_borrow_buf(zio->io_abd, psize);
1299 decode_embedded_bp_compressed(bp, data);
1300 abd_return_buf_copy(zio->io_abd, data, psize);
1302 ASSERT(!BP_IS_EMBEDDED(bp));
1303 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1306 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1307 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1309 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1310 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1312 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1313 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1315 return (ZIO_PIPELINE_CONTINUE);
1319 zio_write_bp_init(zio_t *zio)
1321 if (!IO_IS_ALLOCATING(zio))
1322 return (ZIO_PIPELINE_CONTINUE);
1324 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1326 if (zio->io_bp_override) {
1327 blkptr_t *bp = zio->io_bp;
1328 zio_prop_t *zp = &zio->io_prop;
1330 ASSERT(bp->blk_birth != zio->io_txg);
1331 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1333 *bp = *zio->io_bp_override;
1334 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1336 if (BP_IS_EMBEDDED(bp))
1337 return (ZIO_PIPELINE_CONTINUE);
1340 * If we've been overridden and nopwrite is set then
1341 * set the flag accordingly to indicate that a nopwrite
1342 * has already occurred.
1344 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1345 ASSERT(!zp->zp_dedup);
1346 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1347 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1348 return (ZIO_PIPELINE_CONTINUE);
1351 ASSERT(!zp->zp_nopwrite);
1353 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1354 return (ZIO_PIPELINE_CONTINUE);
1356 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1357 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1359 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1360 BP_SET_DEDUP(bp, 1);
1361 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1362 return (ZIO_PIPELINE_CONTINUE);
1366 * We were unable to handle this as an override bp, treat
1367 * it as a regular write I/O.
1369 zio->io_bp_override = NULL;
1370 *bp = zio->io_bp_orig;
1371 zio->io_pipeline = zio->io_orig_pipeline;
1374 return (ZIO_PIPELINE_CONTINUE);
1378 zio_write_compress(zio_t *zio)
1380 spa_t *spa = zio->io_spa;
1381 zio_prop_t *zp = &zio->io_prop;
1382 enum zio_compress compress = zp->zp_compress;
1383 blkptr_t *bp = zio->io_bp;
1384 uint64_t lsize = zio->io_lsize;
1385 uint64_t psize = zio->io_size;
1388 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1391 * If our children haven't all reached the ready stage,
1392 * wait for them and then repeat this pipeline stage.
1394 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1395 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1396 return (ZIO_PIPELINE_STOP);
1399 if (!IO_IS_ALLOCATING(zio))
1400 return (ZIO_PIPELINE_CONTINUE);
1402 if (zio->io_children_ready != NULL) {
1404 * Now that all our children are ready, run the callback
1405 * associated with this zio in case it wants to modify the
1406 * data to be written.
1408 ASSERT3U(zp->zp_level, >, 0);
1409 zio->io_children_ready(zio);
1412 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1413 ASSERT(zio->io_bp_override == NULL);
1415 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1417 * We're rewriting an existing block, which means we're
1418 * working on behalf of spa_sync(). For spa_sync() to
1419 * converge, it must eventually be the case that we don't
1420 * have to allocate new blocks. But compression changes
1421 * the blocksize, which forces a reallocate, and makes
1422 * convergence take longer. Therefore, after the first
1423 * few passes, stop compressing to ensure convergence.
1425 pass = spa_sync_pass(spa);
1427 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1428 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1429 ASSERT(!BP_GET_DEDUP(bp));
1431 if (pass >= zfs_sync_pass_dont_compress)
1432 compress = ZIO_COMPRESS_OFF;
1434 /* Make sure someone doesn't change their mind on overwrites */
1435 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1436 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1439 /* If it's a compressed write that is not raw, compress the buffer. */
1440 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1441 void *cbuf = zio_buf_alloc(lsize);
1442 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1443 if (psize == 0 || psize == lsize) {
1444 compress = ZIO_COMPRESS_OFF;
1445 zio_buf_free(cbuf, lsize);
1446 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1447 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1448 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1449 encode_embedded_bp_compressed(bp,
1450 cbuf, compress, lsize, psize);
1451 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1452 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1453 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1454 zio_buf_free(cbuf, lsize);
1455 bp->blk_birth = zio->io_txg;
1456 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1457 ASSERT(spa_feature_is_active(spa,
1458 SPA_FEATURE_EMBEDDED_DATA));
1459 return (ZIO_PIPELINE_CONTINUE);
1462 * Round up compressed size up to the ashift
1463 * of the smallest-ashift device, and zero the tail.
1464 * This ensures that the compressed size of the BP
1465 * (and thus compressratio property) are correct,
1466 * in that we charge for the padding used to fill out
1469 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1470 size_t rounded = (size_t)P2ROUNDUP(psize,
1471 1ULL << spa->spa_min_ashift);
1472 if (rounded >= lsize) {
1473 compress = ZIO_COMPRESS_OFF;
1474 zio_buf_free(cbuf, lsize);
1477 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1478 abd_take_ownership_of_buf(cdata, B_TRUE);
1479 abd_zero_off(cdata, psize, rounded - psize);
1481 zio_push_transform(zio, cdata,
1482 psize, lsize, NULL);
1487 * We were unable to handle this as an override bp, treat
1488 * it as a regular write I/O.
1490 zio->io_bp_override = NULL;
1491 *bp = zio->io_bp_orig;
1492 zio->io_pipeline = zio->io_orig_pipeline;
1494 ASSERT3U(psize, !=, 0);
1498 * The final pass of spa_sync() must be all rewrites, but the first
1499 * few passes offer a trade-off: allocating blocks defers convergence,
1500 * but newly allocated blocks are sequential, so they can be written
1501 * to disk faster. Therefore, we allow the first few passes of
1502 * spa_sync() to allocate new blocks, but force rewrites after that.
1503 * There should only be a handful of blocks after pass 1 in any case.
1505 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1506 BP_GET_PSIZE(bp) == psize &&
1507 pass >= zfs_sync_pass_rewrite) {
1509 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1510 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1511 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1514 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1518 if (zio->io_bp_orig.blk_birth != 0 &&
1519 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1520 BP_SET_LSIZE(bp, lsize);
1521 BP_SET_TYPE(bp, zp->zp_type);
1522 BP_SET_LEVEL(bp, zp->zp_level);
1523 BP_SET_BIRTH(bp, zio->io_txg, 0);
1525 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1527 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1528 BP_SET_LSIZE(bp, lsize);
1529 BP_SET_TYPE(bp, zp->zp_type);
1530 BP_SET_LEVEL(bp, zp->zp_level);
1531 BP_SET_PSIZE(bp, psize);
1532 BP_SET_COMPRESS(bp, compress);
1533 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1534 BP_SET_DEDUP(bp, zp->zp_dedup);
1535 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1537 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1538 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1539 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1541 if (zp->zp_nopwrite) {
1542 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1543 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1544 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1547 return (ZIO_PIPELINE_CONTINUE);
1551 zio_free_bp_init(zio_t *zio)
1553 blkptr_t *bp = zio->io_bp;
1555 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1556 if (BP_GET_DEDUP(bp))
1557 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1560 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1562 return (ZIO_PIPELINE_CONTINUE);
1566 * ==========================================================================
1567 * Execute the I/O pipeline
1568 * ==========================================================================
1572 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1574 spa_t *spa = zio->io_spa;
1575 zio_type_t t = zio->io_type;
1576 int flags = (cutinline ? TQ_FRONT : 0);
1578 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1581 * If we're a config writer or a probe, the normal issue and
1582 * interrupt threads may all be blocked waiting for the config lock.
1583 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1585 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1589 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1591 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1595 * If this is a high priority I/O, then use the high priority taskq if
1598 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1599 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1602 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1605 * NB: We are assuming that the zio can only be dispatched
1606 * to a single taskq at a time. It would be a grievous error
1607 * to dispatch the zio to another taskq at the same time.
1609 #if defined(illumos) || !defined(_KERNEL)
1610 ASSERT(zio->io_tqent.tqent_next == NULL);
1612 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1614 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1615 flags, &zio->io_tqent);
1619 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1621 kthread_t *executor = zio->io_executor;
1622 spa_t *spa = zio->io_spa;
1624 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1625 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1627 for (i = 0; i < tqs->stqs_count; i++) {
1628 if (taskq_member(tqs->stqs_taskq[i], executor))
1637 zio_issue_async(zio_t *zio)
1639 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1641 return (ZIO_PIPELINE_STOP);
1645 zio_interrupt(zio_t *zio)
1647 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1651 zio_delay_interrupt(zio_t *zio)
1654 * The timeout_generic() function isn't defined in userspace, so
1655 * rather than trying to implement the function, the zio delay
1656 * functionality has been disabled for userspace builds.
1661 * If io_target_timestamp is zero, then no delay has been registered
1662 * for this IO, thus jump to the end of this function and "skip" the
1663 * delay; issuing it directly to the zio layer.
1665 if (zio->io_target_timestamp != 0) {
1666 hrtime_t now = gethrtime();
1668 if (now >= zio->io_target_timestamp) {
1670 * This IO has already taken longer than the target
1671 * delay to complete, so we don't want to delay it
1672 * any longer; we "miss" the delay and issue it
1673 * directly to the zio layer. This is likely due to
1674 * the target latency being set to a value less than
1675 * the underlying hardware can satisfy (e.g. delay
1676 * set to 1ms, but the disks take 10ms to complete an
1680 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1685 hrtime_t diff = zio->io_target_timestamp - now;
1687 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1688 hrtime_t, now, hrtime_t, diff);
1690 (void) timeout_generic(CALLOUT_NORMAL,
1691 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1698 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1703 * Execute the I/O pipeline until one of the following occurs:
1705 * (1) the I/O completes
1706 * (2) the pipeline stalls waiting for dependent child I/Os
1707 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1708 * (4) the I/O is delegated by vdev-level caching or aggregation
1709 * (5) the I/O is deferred due to vdev-level queueing
1710 * (6) the I/O is handed off to another thread.
1712 * In all cases, the pipeline stops whenever there's no CPU work; it never
1713 * burns a thread in cv_wait().
1715 * There's no locking on io_stage because there's no legitimate way
1716 * for multiple threads to be attempting to process the same I/O.
1718 static zio_pipe_stage_t *zio_pipeline[];
1721 zio_execute(zio_t *zio)
1723 zio->io_executor = curthread;
1725 ASSERT3U(zio->io_queued_timestamp, >, 0);
1727 while (zio->io_stage < ZIO_STAGE_DONE) {
1728 enum zio_stage pipeline = zio->io_pipeline;
1729 enum zio_stage stage = zio->io_stage;
1732 ASSERT(!MUTEX_HELD(&zio->io_lock));
1733 ASSERT(ISP2(stage));
1734 ASSERT(zio->io_stall == NULL);
1738 } while ((stage & pipeline) == 0);
1740 ASSERT(stage <= ZIO_STAGE_DONE);
1743 * If we are in interrupt context and this pipeline stage
1744 * will grab a config lock that is held across I/O,
1745 * or may wait for an I/O that needs an interrupt thread
1746 * to complete, issue async to avoid deadlock.
1748 * For VDEV_IO_START, we cut in line so that the io will
1749 * be sent to disk promptly.
1751 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1752 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1753 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1754 zio_requeue_io_start_cut_in_line : B_FALSE;
1755 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1759 zio->io_stage = stage;
1760 zio->io_pipeline_trace |= zio->io_stage;
1761 rv = zio_pipeline[highbit64(stage) - 1](zio);
1763 if (rv == ZIO_PIPELINE_STOP)
1766 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1771 * ==========================================================================
1772 * Initiate I/O, either sync or async
1773 * ==========================================================================
1776 zio_wait(zio_t *zio)
1780 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1781 ASSERT3P(zio->io_executor, ==, NULL);
1783 zio->io_waiter = curthread;
1784 ASSERT0(zio->io_queued_timestamp);
1785 zio->io_queued_timestamp = gethrtime();
1789 mutex_enter(&zio->io_lock);
1790 while (zio->io_executor != NULL)
1791 cv_wait(&zio->io_cv, &zio->io_lock);
1792 mutex_exit(&zio->io_lock);
1794 error = zio->io_error;
1801 zio_nowait(zio_t *zio)
1803 ASSERT3P(zio->io_executor, ==, NULL);
1805 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1806 zio_unique_parent(zio) == NULL) {
1808 * This is a logical async I/O with no parent to wait for it.
1809 * We add it to the spa_async_root_zio "Godfather" I/O which
1810 * will ensure they complete prior to unloading the pool.
1812 spa_t *spa = zio->io_spa;
1814 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1817 ASSERT0(zio->io_queued_timestamp);
1818 zio->io_queued_timestamp = gethrtime();
1823 * ==========================================================================
1824 * Reexecute, cancel, or suspend/resume failed I/O
1825 * ==========================================================================
1829 zio_reexecute(zio_t *pio)
1831 zio_t *cio, *cio_next;
1833 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1834 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1835 ASSERT(pio->io_gang_leader == NULL);
1836 ASSERT(pio->io_gang_tree == NULL);
1838 pio->io_flags = pio->io_orig_flags;
1839 pio->io_stage = pio->io_orig_stage;
1840 pio->io_pipeline = pio->io_orig_pipeline;
1841 pio->io_reexecute = 0;
1842 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1843 pio->io_pipeline_trace = 0;
1845 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1846 pio->io_state[w] = 0;
1847 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1848 pio->io_child_error[c] = 0;
1850 if (IO_IS_ALLOCATING(pio))
1851 BP_ZERO(pio->io_bp);
1854 * As we reexecute pio's children, new children could be created.
1855 * New children go to the head of pio's io_child_list, however,
1856 * so we will (correctly) not reexecute them. The key is that
1857 * the remainder of pio's io_child_list, from 'cio_next' onward,
1858 * cannot be affected by any side effects of reexecuting 'cio'.
1860 zio_link_t *zl = NULL;
1861 mutex_enter(&pio->io_lock);
1862 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1863 cio_next = zio_walk_children(pio, &zl);
1864 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1865 pio->io_children[cio->io_child_type][w]++;
1866 mutex_exit(&pio->io_lock);
1868 mutex_enter(&pio->io_lock);
1870 mutex_exit(&pio->io_lock);
1873 * Now that all children have been reexecuted, execute the parent.
1874 * We don't reexecute "The Godfather" I/O here as it's the
1875 * responsibility of the caller to wait on it.
1877 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1878 pio->io_queued_timestamp = gethrtime();
1884 zio_suspend(spa_t *spa, zio_t *zio)
1886 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1887 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1888 "failure and the failure mode property for this pool "
1889 "is set to panic.", spa_name(spa));
1891 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1893 mutex_enter(&spa->spa_suspend_lock);
1895 if (spa->spa_suspend_zio_root == NULL)
1896 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1897 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1898 ZIO_FLAG_GODFATHER);
1900 spa->spa_suspended = B_TRUE;
1903 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1904 ASSERT(zio != spa->spa_suspend_zio_root);
1905 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1906 ASSERT(zio_unique_parent(zio) == NULL);
1907 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1908 zio_add_child(spa->spa_suspend_zio_root, zio);
1911 mutex_exit(&spa->spa_suspend_lock);
1915 zio_resume(spa_t *spa)
1920 * Reexecute all previously suspended i/o.
1922 mutex_enter(&spa->spa_suspend_lock);
1923 spa->spa_suspended = B_FALSE;
1924 cv_broadcast(&spa->spa_suspend_cv);
1925 pio = spa->spa_suspend_zio_root;
1926 spa->spa_suspend_zio_root = NULL;
1927 mutex_exit(&spa->spa_suspend_lock);
1933 return (zio_wait(pio));
1937 zio_resume_wait(spa_t *spa)
1939 mutex_enter(&spa->spa_suspend_lock);
1940 while (spa_suspended(spa))
1941 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1942 mutex_exit(&spa->spa_suspend_lock);
1946 * ==========================================================================
1949 * A gang block is a collection of small blocks that looks to the DMU
1950 * like one large block. When zio_dva_allocate() cannot find a block
1951 * of the requested size, due to either severe fragmentation or the pool
1952 * being nearly full, it calls zio_write_gang_block() to construct the
1953 * block from smaller fragments.
1955 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1956 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1957 * an indirect block: it's an array of block pointers. It consumes
1958 * only one sector and hence is allocatable regardless of fragmentation.
1959 * The gang header's bps point to its gang members, which hold the data.
1961 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1962 * as the verifier to ensure uniqueness of the SHA256 checksum.
1963 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1964 * not the gang header. This ensures that data block signatures (needed for
1965 * deduplication) are independent of how the block is physically stored.
1967 * Gang blocks can be nested: a gang member may itself be a gang block.
1968 * Thus every gang block is a tree in which root and all interior nodes are
1969 * gang headers, and the leaves are normal blocks that contain user data.
1970 * The root of the gang tree is called the gang leader.
1972 * To perform any operation (read, rewrite, free, claim) on a gang block,
1973 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1974 * in the io_gang_tree field of the original logical i/o by recursively
1975 * reading the gang leader and all gang headers below it. This yields
1976 * an in-core tree containing the contents of every gang header and the
1977 * bps for every constituent of the gang block.
1979 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1980 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1981 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1982 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1983 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1984 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1985 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1986 * of the gang header plus zio_checksum_compute() of the data to update the
1987 * gang header's blk_cksum as described above.
1989 * The two-phase assemble/issue model solves the problem of partial failure --
1990 * what if you'd freed part of a gang block but then couldn't read the
1991 * gang header for another part? Assembling the entire gang tree first
1992 * ensures that all the necessary gang header I/O has succeeded before
1993 * starting the actual work of free, claim, or write. Once the gang tree
1994 * is assembled, free and claim are in-memory operations that cannot fail.
1996 * In the event that a gang write fails, zio_dva_unallocate() walks the
1997 * gang tree to immediately free (i.e. insert back into the space map)
1998 * everything we've allocated. This ensures that we don't get ENOSPC
1999 * errors during repeated suspend/resume cycles due to a flaky device.
2001 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2002 * the gang tree, we won't modify the block, so we can safely defer the free
2003 * (knowing that the block is still intact). If we *can* assemble the gang
2004 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2005 * each constituent bp and we can allocate a new block on the next sync pass.
2007 * In all cases, the gang tree allows complete recovery from partial failure.
2008 * ==========================================================================
2012 zio_gang_issue_func_done(zio_t *zio)
2014 abd_put(zio->io_abd);
2018 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2024 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2025 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2026 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2027 &pio->io_bookmark));
2031 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2038 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2039 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2040 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2041 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2044 * As we rewrite each gang header, the pipeline will compute
2045 * a new gang block header checksum for it; but no one will
2046 * compute a new data checksum, so we do that here. The one
2047 * exception is the gang leader: the pipeline already computed
2048 * its data checksum because that stage precedes gang assembly.
2049 * (Presently, nothing actually uses interior data checksums;
2050 * this is just good hygiene.)
2052 if (gn != pio->io_gang_leader->io_gang_tree) {
2053 abd_t *buf = abd_get_offset(data, offset);
2055 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2056 buf, BP_GET_PSIZE(bp));
2061 * If we are here to damage data for testing purposes,
2062 * leave the GBH alone so that we can detect the damage.
2064 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2065 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2067 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2068 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2069 zio_gang_issue_func_done, NULL, pio->io_priority,
2070 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2078 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2081 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2082 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2083 ZIO_GANG_CHILD_FLAGS(pio)));
2088 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2091 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2092 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2095 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2104 static void zio_gang_tree_assemble_done(zio_t *zio);
2106 static zio_gang_node_t *
2107 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2109 zio_gang_node_t *gn;
2111 ASSERT(*gnpp == NULL);
2113 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2114 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2121 zio_gang_node_free(zio_gang_node_t **gnpp)
2123 zio_gang_node_t *gn = *gnpp;
2125 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2126 ASSERT(gn->gn_child[g] == NULL);
2128 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2129 kmem_free(gn, sizeof (*gn));
2134 zio_gang_tree_free(zio_gang_node_t **gnpp)
2136 zio_gang_node_t *gn = *gnpp;
2141 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2142 zio_gang_tree_free(&gn->gn_child[g]);
2144 zio_gang_node_free(gnpp);
2148 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2150 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2151 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2153 ASSERT(gio->io_gang_leader == gio);
2154 ASSERT(BP_IS_GANG(bp));
2156 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2157 zio_gang_tree_assemble_done, gn, gio->io_priority,
2158 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2162 zio_gang_tree_assemble_done(zio_t *zio)
2164 zio_t *gio = zio->io_gang_leader;
2165 zio_gang_node_t *gn = zio->io_private;
2166 blkptr_t *bp = zio->io_bp;
2168 ASSERT(gio == zio_unique_parent(zio));
2169 ASSERT(zio->io_child_count == 0);
2174 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2175 if (BP_SHOULD_BYTESWAP(bp))
2176 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2178 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2179 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2180 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2182 abd_put(zio->io_abd);
2184 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2185 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2186 if (!BP_IS_GANG(gbp))
2188 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2193 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2196 zio_t *gio = pio->io_gang_leader;
2199 ASSERT(BP_IS_GANG(bp) == !!gn);
2200 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2201 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2204 * If you're a gang header, your data is in gn->gn_gbh.
2205 * If you're a gang member, your data is in 'data' and gn == NULL.
2207 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2210 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2212 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2213 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2214 if (BP_IS_HOLE(gbp))
2216 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2218 offset += BP_GET_PSIZE(gbp);
2222 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2223 ASSERT3U(gio->io_size, ==, offset);
2230 zio_gang_assemble(zio_t *zio)
2232 blkptr_t *bp = zio->io_bp;
2234 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2235 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2237 zio->io_gang_leader = zio;
2239 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2241 return (ZIO_PIPELINE_CONTINUE);
2245 zio_gang_issue(zio_t *zio)
2247 blkptr_t *bp = zio->io_bp;
2249 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2250 return (ZIO_PIPELINE_STOP);
2253 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2254 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2256 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2257 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2260 zio_gang_tree_free(&zio->io_gang_tree);
2262 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2264 return (ZIO_PIPELINE_CONTINUE);
2268 zio_write_gang_member_ready(zio_t *zio)
2270 zio_t *pio = zio_unique_parent(zio);
2271 zio_t *gio = zio->io_gang_leader;
2272 dva_t *cdva = zio->io_bp->blk_dva;
2273 dva_t *pdva = pio->io_bp->blk_dva;
2276 if (BP_IS_HOLE(zio->io_bp))
2279 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2281 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2282 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2283 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2284 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2285 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2287 mutex_enter(&pio->io_lock);
2288 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2289 ASSERT(DVA_GET_GANG(&pdva[d]));
2290 asize = DVA_GET_ASIZE(&pdva[d]);
2291 asize += DVA_GET_ASIZE(&cdva[d]);
2292 DVA_SET_ASIZE(&pdva[d], asize);
2294 mutex_exit(&pio->io_lock);
2298 zio_write_gang_done(zio_t *zio)
2300 abd_put(zio->io_abd);
2304 zio_write_gang_block(zio_t *pio)
2306 spa_t *spa = pio->io_spa;
2307 metaslab_class_t *mc = spa_normal_class(spa);
2308 blkptr_t *bp = pio->io_bp;
2309 zio_t *gio = pio->io_gang_leader;
2311 zio_gang_node_t *gn, **gnpp;
2312 zio_gbh_phys_t *gbh;
2314 uint64_t txg = pio->io_txg;
2315 uint64_t resid = pio->io_size;
2317 int copies = gio->io_prop.zp_copies;
2318 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2322 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2323 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2324 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2325 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2327 flags |= METASLAB_ASYNC_ALLOC;
2328 VERIFY(refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2332 * The logical zio has already placed a reservation for
2333 * 'copies' allocation slots but gang blocks may require
2334 * additional copies. These additional copies
2335 * (i.e. gbh_copies - copies) are guaranteed to succeed
2336 * since metaslab_class_throttle_reserve() always allows
2337 * additional reservations for gang blocks.
2339 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2340 pio->io_allocator, pio, flags));
2343 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2344 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2345 &pio->io_alloc_list, pio, pio->io_allocator);
2347 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2348 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2349 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2352 * If we failed to allocate the gang block header then
2353 * we remove any additional allocation reservations that
2354 * we placed here. The original reservation will
2355 * be removed when the logical I/O goes to the ready
2358 metaslab_class_throttle_unreserve(mc,
2359 gbh_copies - copies, pio->io_allocator, pio);
2361 pio->io_error = error;
2362 return (ZIO_PIPELINE_CONTINUE);
2366 gnpp = &gio->io_gang_tree;
2368 gnpp = pio->io_private;
2369 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2372 gn = zio_gang_node_alloc(gnpp);
2374 bzero(gbh, SPA_GANGBLOCKSIZE);
2375 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2378 * Create the gang header.
2380 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2381 zio_write_gang_done, NULL, pio->io_priority,
2382 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2385 * Create and nowait the gang children.
2387 for (int g = 0; resid != 0; resid -= lsize, g++) {
2388 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2390 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2392 zp.zp_checksum = gio->io_prop.zp_checksum;
2393 zp.zp_compress = ZIO_COMPRESS_OFF;
2394 zp.zp_type = DMU_OT_NONE;
2396 zp.zp_copies = gio->io_prop.zp_copies;
2397 zp.zp_dedup = B_FALSE;
2398 zp.zp_dedup_verify = B_FALSE;
2399 zp.zp_nopwrite = B_FALSE;
2401 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2402 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2403 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2404 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2405 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2407 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2408 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2409 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2412 * Gang children won't throttle but we should
2413 * account for their work, so reserve an allocation
2414 * slot for them here.
2416 VERIFY(metaslab_class_throttle_reserve(mc,
2417 zp.zp_copies, cio->io_allocator, cio, flags));
2423 * Set pio's pipeline to just wait for zio to finish.
2425 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2429 return (ZIO_PIPELINE_CONTINUE);
2433 * The zio_nop_write stage in the pipeline determines if allocating a
2434 * new bp is necessary. The nopwrite feature can handle writes in
2435 * either syncing or open context (i.e. zil writes) and as a result is
2436 * mutually exclusive with dedup.
2438 * By leveraging a cryptographically secure checksum, such as SHA256, we
2439 * can compare the checksums of the new data and the old to determine if
2440 * allocating a new block is required. Note that our requirements for
2441 * cryptographic strength are fairly weak: there can't be any accidental
2442 * hash collisions, but we don't need to be secure against intentional
2443 * (malicious) collisions. To trigger a nopwrite, you have to be able
2444 * to write the file to begin with, and triggering an incorrect (hash
2445 * collision) nopwrite is no worse than simply writing to the file.
2446 * That said, there are no known attacks against the checksum algorithms
2447 * used for nopwrite, assuming that the salt and the checksums
2448 * themselves remain secret.
2451 zio_nop_write(zio_t *zio)
2453 blkptr_t *bp = zio->io_bp;
2454 blkptr_t *bp_orig = &zio->io_bp_orig;
2455 zio_prop_t *zp = &zio->io_prop;
2457 ASSERT(BP_GET_LEVEL(bp) == 0);
2458 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2459 ASSERT(zp->zp_nopwrite);
2460 ASSERT(!zp->zp_dedup);
2461 ASSERT(zio->io_bp_override == NULL);
2462 ASSERT(IO_IS_ALLOCATING(zio));
2465 * Check to see if the original bp and the new bp have matching
2466 * characteristics (i.e. same checksum, compression algorithms, etc).
2467 * If they don't then just continue with the pipeline which will
2468 * allocate a new bp.
2470 if (BP_IS_HOLE(bp_orig) ||
2471 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2472 ZCHECKSUM_FLAG_NOPWRITE) ||
2473 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2474 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2475 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2476 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2477 return (ZIO_PIPELINE_CONTINUE);
2480 * If the checksums match then reset the pipeline so that we
2481 * avoid allocating a new bp and issuing any I/O.
2483 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2484 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2485 ZCHECKSUM_FLAG_NOPWRITE);
2486 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2487 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2488 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2489 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2490 sizeof (uint64_t)) == 0);
2493 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2494 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2497 return (ZIO_PIPELINE_CONTINUE);
2501 * ==========================================================================
2503 * ==========================================================================
2506 zio_ddt_child_read_done(zio_t *zio)
2508 blkptr_t *bp = zio->io_bp;
2509 ddt_entry_t *dde = zio->io_private;
2511 zio_t *pio = zio_unique_parent(zio);
2513 mutex_enter(&pio->io_lock);
2514 ddp = ddt_phys_select(dde, bp);
2515 if (zio->io_error == 0)
2516 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2518 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2519 dde->dde_repair_abd = zio->io_abd;
2521 abd_free(zio->io_abd);
2522 mutex_exit(&pio->io_lock);
2526 zio_ddt_read_start(zio_t *zio)
2528 blkptr_t *bp = zio->io_bp;
2530 ASSERT(BP_GET_DEDUP(bp));
2531 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2532 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2534 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2535 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2536 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2537 ddt_phys_t *ddp = dde->dde_phys;
2538 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2541 ASSERT(zio->io_vsd == NULL);
2544 if (ddp_self == NULL)
2545 return (ZIO_PIPELINE_CONTINUE);
2547 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2548 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2550 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2552 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2553 abd_alloc_for_io(zio->io_size, B_TRUE),
2554 zio->io_size, zio_ddt_child_read_done, dde,
2555 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2556 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2558 return (ZIO_PIPELINE_CONTINUE);
2561 zio_nowait(zio_read(zio, zio->io_spa, bp,
2562 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2563 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2565 return (ZIO_PIPELINE_CONTINUE);
2569 zio_ddt_read_done(zio_t *zio)
2571 blkptr_t *bp = zio->io_bp;
2573 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2574 return (ZIO_PIPELINE_STOP);
2577 ASSERT(BP_GET_DEDUP(bp));
2578 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2579 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2581 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2582 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2583 ddt_entry_t *dde = zio->io_vsd;
2585 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2586 return (ZIO_PIPELINE_CONTINUE);
2589 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2590 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2591 return (ZIO_PIPELINE_STOP);
2593 if (dde->dde_repair_abd != NULL) {
2594 abd_copy(zio->io_abd, dde->dde_repair_abd,
2596 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2598 ddt_repair_done(ddt, dde);
2602 ASSERT(zio->io_vsd == NULL);
2604 return (ZIO_PIPELINE_CONTINUE);
2608 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2610 spa_t *spa = zio->io_spa;
2611 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2613 /* We should never get a raw, override zio */
2614 ASSERT(!(zio->io_bp_override && do_raw));
2617 * Note: we compare the original data, not the transformed data,
2618 * because when zio->io_bp is an override bp, we will not have
2619 * pushed the I/O transforms. That's an important optimization
2620 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2622 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2623 zio_t *lio = dde->dde_lead_zio[p];
2626 return (lio->io_orig_size != zio->io_orig_size ||
2627 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2628 zio->io_orig_size) != 0);
2632 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2633 ddt_phys_t *ddp = &dde->dde_phys[p];
2635 if (ddp->ddp_phys_birth != 0) {
2636 arc_buf_t *abuf = NULL;
2637 arc_flags_t aflags = ARC_FLAG_WAIT;
2638 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2639 blkptr_t blk = *zio->io_bp;
2642 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2647 * Intuitively, it would make more sense to compare
2648 * io_abd than io_orig_abd in the raw case since you
2649 * don't want to look at any transformations that have
2650 * happened to the data. However, for raw I/Os the
2651 * data will actually be the same in io_abd and
2652 * io_orig_abd, so all we have to do is issue this as
2656 zio_flags |= ZIO_FLAG_RAW;
2657 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2658 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2660 ASSERT3P(zio->io_transform_stack, ==, NULL);
2663 error = arc_read(NULL, spa, &blk,
2664 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2665 zio_flags, &aflags, &zio->io_bookmark);
2668 if (arc_buf_size(abuf) != zio->io_orig_size ||
2669 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2670 zio->io_orig_size) != 0)
2671 error = SET_ERROR(EEXIST);
2672 arc_buf_destroy(abuf, &abuf);
2676 return (error != 0);
2684 zio_ddt_child_write_ready(zio_t *zio)
2686 int p = zio->io_prop.zp_copies;
2687 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2688 ddt_entry_t *dde = zio->io_private;
2689 ddt_phys_t *ddp = &dde->dde_phys[p];
2697 ASSERT(dde->dde_lead_zio[p] == zio);
2699 ddt_phys_fill(ddp, zio->io_bp);
2701 zio_link_t *zl = NULL;
2702 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2703 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2709 zio_ddt_child_write_done(zio_t *zio)
2711 int p = zio->io_prop.zp_copies;
2712 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2713 ddt_entry_t *dde = zio->io_private;
2714 ddt_phys_t *ddp = &dde->dde_phys[p];
2718 ASSERT(ddp->ddp_refcnt == 0);
2719 ASSERT(dde->dde_lead_zio[p] == zio);
2720 dde->dde_lead_zio[p] = NULL;
2722 if (zio->io_error == 0) {
2723 zio_link_t *zl = NULL;
2724 while (zio_walk_parents(zio, &zl) != NULL)
2725 ddt_phys_addref(ddp);
2727 ddt_phys_clear(ddp);
2734 zio_ddt_ditto_write_done(zio_t *zio)
2736 int p = DDT_PHYS_DITTO;
2737 zio_prop_t *zp = &zio->io_prop;
2738 blkptr_t *bp = zio->io_bp;
2739 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2740 ddt_entry_t *dde = zio->io_private;
2741 ddt_phys_t *ddp = &dde->dde_phys[p];
2742 ddt_key_t *ddk = &dde->dde_key;
2746 ASSERT(ddp->ddp_refcnt == 0);
2747 ASSERT(dde->dde_lead_zio[p] == zio);
2748 dde->dde_lead_zio[p] = NULL;
2750 if (zio->io_error == 0) {
2751 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2752 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2753 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2754 if (ddp->ddp_phys_birth != 0)
2755 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2756 ddt_phys_fill(ddp, bp);
2763 zio_ddt_write(zio_t *zio)
2765 spa_t *spa = zio->io_spa;
2766 blkptr_t *bp = zio->io_bp;
2767 uint64_t txg = zio->io_txg;
2768 zio_prop_t *zp = &zio->io_prop;
2769 int p = zp->zp_copies;
2773 ddt_t *ddt = ddt_select(spa, bp);
2777 ASSERT(BP_GET_DEDUP(bp));
2778 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2779 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2780 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2783 dde = ddt_lookup(ddt, bp, B_TRUE);
2784 ddp = &dde->dde_phys[p];
2786 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2788 * If we're using a weak checksum, upgrade to a strong checksum
2789 * and try again. If we're already using a strong checksum,
2790 * we can't resolve it, so just convert to an ordinary write.
2791 * (And automatically e-mail a paper to Nature?)
2793 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2794 ZCHECKSUM_FLAG_DEDUP)) {
2795 zp->zp_checksum = spa_dedup_checksum(spa);
2796 zio_pop_transforms(zio);
2797 zio->io_stage = ZIO_STAGE_OPEN;
2800 zp->zp_dedup = B_FALSE;
2801 BP_SET_DEDUP(bp, B_FALSE);
2803 ASSERT(!BP_GET_DEDUP(bp));
2804 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2806 return (ZIO_PIPELINE_CONTINUE);
2809 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2810 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2812 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2813 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2814 zio_prop_t czp = *zp;
2816 czp.zp_copies = ditto_copies;
2819 * If we arrived here with an override bp, we won't have run
2820 * the transform stack, so we won't have the data we need to
2821 * generate a child i/o. So, toss the override bp and restart.
2822 * This is safe, because using the override bp is just an
2823 * optimization; and it's rare, so the cost doesn't matter.
2825 if (zio->io_bp_override) {
2826 zio_pop_transforms(zio);
2827 zio->io_stage = ZIO_STAGE_OPEN;
2828 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2829 zio->io_bp_override = NULL;
2832 return (ZIO_PIPELINE_CONTINUE);
2835 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2836 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2837 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2838 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2840 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2841 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2844 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2845 if (ddp->ddp_phys_birth != 0)
2846 ddt_bp_fill(ddp, bp, txg);
2847 if (dde->dde_lead_zio[p] != NULL)
2848 zio_add_child(zio, dde->dde_lead_zio[p]);
2850 ddt_phys_addref(ddp);
2851 } else if (zio->io_bp_override) {
2852 ASSERT(bp->blk_birth == txg);
2853 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2854 ddt_phys_fill(ddp, bp);
2855 ddt_phys_addref(ddp);
2857 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2858 zio->io_orig_size, zio->io_orig_size, zp,
2859 zio_ddt_child_write_ready, NULL, NULL,
2860 zio_ddt_child_write_done, dde, zio->io_priority,
2861 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2863 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2864 dde->dde_lead_zio[p] = cio;
2874 return (ZIO_PIPELINE_CONTINUE);
2877 ddt_entry_t *freedde; /* for debugging */
2880 zio_ddt_free(zio_t *zio)
2882 spa_t *spa = zio->io_spa;
2883 blkptr_t *bp = zio->io_bp;
2884 ddt_t *ddt = ddt_select(spa, bp);
2888 ASSERT(BP_GET_DEDUP(bp));
2889 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2892 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2893 ddp = ddt_phys_select(dde, bp);
2894 ddt_phys_decref(ddp);
2897 return (ZIO_PIPELINE_CONTINUE);
2901 * ==========================================================================
2902 * Allocate and free blocks
2903 * ==========================================================================
2907 zio_io_to_allocate(spa_t *spa, int allocator)
2911 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
2913 zio = avl_first(&spa->spa_alloc_trees[allocator]);
2917 ASSERT(IO_IS_ALLOCATING(zio));
2920 * Try to place a reservation for this zio. If we're unable to
2921 * reserve then we throttle.
2923 ASSERT3U(zio->io_allocator, ==, allocator);
2924 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2925 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
2929 avl_remove(&spa->spa_alloc_trees[allocator], zio);
2930 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2936 zio_dva_throttle(zio_t *zio)
2938 spa_t *spa = zio->io_spa;
2941 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2942 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2943 zio->io_child_type == ZIO_CHILD_GANG ||
2944 zio->io_flags & ZIO_FLAG_NODATA) {
2945 return (ZIO_PIPELINE_CONTINUE);
2948 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2950 ASSERT3U(zio->io_queued_timestamp, >, 0);
2951 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2953 zbookmark_phys_t *bm = &zio->io_bookmark;
2955 * We want to try to use as many allocators as possible to help improve
2956 * performance, but we also want logically adjacent IOs to be physically
2957 * adjacent to improve sequential read performance. We chunk each object
2958 * into 2^20 block regions, and then hash based on the objset, object,
2959 * level, and region to accomplish both of these goals.
2961 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
2962 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
2963 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
2965 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2966 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
2968 nio = zio_io_to_allocate(zio->io_spa, zio->io_allocator);
2969 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
2972 return (ZIO_PIPELINE_CONTINUE);
2975 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2977 * We are passing control to a new zio so make sure that
2978 * it is processed by a different thread. We do this to
2979 * avoid stack overflows that can occur when parents are
2980 * throttled and children are making progress. We allow
2981 * it to go to the head of the taskq since it's already
2984 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2986 return (ZIO_PIPELINE_STOP);
2990 zio_allocate_dispatch(spa_t *spa, int allocator)
2994 mutex_enter(&spa->spa_alloc_locks[allocator]);
2995 zio = zio_io_to_allocate(spa, allocator);
2996 mutex_exit(&spa->spa_alloc_locks[allocator]);
3000 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3001 ASSERT0(zio->io_error);
3002 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3006 zio_dva_allocate(zio_t *zio)
3008 spa_t *spa = zio->io_spa;
3009 metaslab_class_t *mc = spa_normal_class(spa);
3010 blkptr_t *bp = zio->io_bp;
3014 if (zio->io_gang_leader == NULL) {
3015 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3016 zio->io_gang_leader = zio;
3019 ASSERT(BP_IS_HOLE(bp));
3020 ASSERT0(BP_GET_NDVAS(bp));
3021 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3022 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3023 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3025 if (zio->io_flags & ZIO_FLAG_NODATA) {
3026 flags |= METASLAB_DONT_THROTTLE;
3028 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3029 flags |= METASLAB_GANG_CHILD;
3031 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
3032 flags |= METASLAB_ASYNC_ALLOC;
3035 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3036 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3037 &zio->io_alloc_list, zio, zio->io_allocator);
3040 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3041 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3043 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3044 return (zio_write_gang_block(zio));
3045 zio->io_error = error;
3048 return (ZIO_PIPELINE_CONTINUE);
3052 zio_dva_free(zio_t *zio)
3054 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3056 return (ZIO_PIPELINE_CONTINUE);
3060 zio_dva_claim(zio_t *zio)
3064 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3066 zio->io_error = error;
3068 return (ZIO_PIPELINE_CONTINUE);
3072 * Undo an allocation. This is used by zio_done() when an I/O fails
3073 * and we want to give back the block we just allocated.
3074 * This handles both normal blocks and gang blocks.
3077 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3079 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3080 ASSERT(zio->io_bp_override == NULL);
3082 if (!BP_IS_HOLE(bp))
3083 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3086 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3087 zio_dva_unallocate(zio, gn->gn_child[g],
3088 &gn->gn_gbh->zg_blkptr[g]);
3094 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3097 zio_alloc_zil(spa_t *spa, uint64_t objset, uint64_t txg, blkptr_t *new_bp,
3098 blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3101 zio_alloc_list_t io_alloc_list;
3103 ASSERT(txg > spa_syncing_txg(spa));
3105 metaslab_trace_init(&io_alloc_list);
3107 * When allocating a zil block, we don't have information about
3108 * the final destination of the block except the objset it's part
3109 * of, so we just hash the objset ID to pick the allocator to get
3112 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3113 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3114 cityhash4(0, 0, 0, objset) % spa->spa_alloc_count);
3118 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3119 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3120 &io_alloc_list, NULL, cityhash4(0, 0, 0, objset) %
3121 spa->spa_alloc_count);
3125 metaslab_trace_fini(&io_alloc_list);
3128 BP_SET_LSIZE(new_bp, size);
3129 BP_SET_PSIZE(new_bp, size);
3130 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3131 BP_SET_CHECKSUM(new_bp,
3132 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3133 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3134 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3135 BP_SET_LEVEL(new_bp, 0);
3136 BP_SET_DEDUP(new_bp, 0);
3137 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3139 zfs_dbgmsg("%s: zil block allocation failure: "
3140 "size %llu, error %d", spa_name(spa), size, error);
3147 * ==========================================================================
3148 * Read, write and delete to physical devices
3149 * ==========================================================================
3154 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3155 * stops after this stage and will resume upon I/O completion.
3156 * However, there are instances where the vdev layer may need to
3157 * continue the pipeline when an I/O was not issued. Since the I/O
3158 * that was sent to the vdev layer might be different than the one
3159 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3160 * force the underlying vdev layers to call either zio_execute() or
3161 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3164 zio_vdev_io_start(zio_t *zio)
3166 vdev_t *vd = zio->io_vd;
3168 spa_t *spa = zio->io_spa;
3171 ASSERT(zio->io_error == 0);
3172 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3175 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3176 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3179 * The mirror_ops handle multiple DVAs in a single BP.
3181 vdev_mirror_ops.vdev_op_io_start(zio);
3182 return (ZIO_PIPELINE_STOP);
3185 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3186 zio->io_priority == ZIO_PRIORITY_NOW) {
3187 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3188 return (ZIO_PIPELINE_CONTINUE);
3191 ASSERT3P(zio->io_logical, !=, zio);
3192 if (zio->io_type == ZIO_TYPE_WRITE) {
3193 ASSERT(spa->spa_trust_config);
3195 if (zio->io_vd->vdev_removing) {
3197 * Note: the code can handle other kinds of writes,
3198 * but we don't expect them.
3200 ASSERT(zio->io_flags &
3201 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3202 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3207 * We keep track of time-sensitive I/Os so that the scan thread
3208 * can quickly react to certain workloads. In particular, we care
3209 * about non-scrubbing, top-level reads and writes with the following
3211 * - synchronous writes of user data to non-slog devices
3212 * - any reads of user data
3213 * When these conditions are met, adjust the timestamp of spa_last_io
3214 * which allows the scan thread to adjust its workload accordingly.
3216 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3217 vd == vd->vdev_top && !vd->vdev_islog &&
3218 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3219 zio->io_txg != spa_syncing_txg(spa)) {
3220 uint64_t old = spa->spa_last_io;
3221 uint64_t new = ddi_get_lbolt64();
3223 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3225 align = 1ULL << vd->vdev_top->vdev_ashift;
3227 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3228 P2PHASE(zio->io_size, align) != 0) {
3229 /* Transform logical writes to be a full physical block size. */
3230 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3232 if (zio->io_type == ZIO_TYPE_READ ||
3233 zio->io_type == ZIO_TYPE_WRITE)
3234 abuf = abd_alloc_sametype(zio->io_abd, asize);
3235 ASSERT(vd == vd->vdev_top);
3236 if (zio->io_type == ZIO_TYPE_WRITE) {
3237 abd_copy(abuf, zio->io_abd, zio->io_size);
3238 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3240 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3245 * If this is not a physical io, make sure that it is properly aligned
3246 * before proceeding.
3248 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3249 ASSERT0(P2PHASE(zio->io_offset, align));
3250 ASSERT0(P2PHASE(zio->io_size, align));
3253 * For the physical io we allow alignment
3254 * to a logical block size.
3256 uint64_t log_align =
3257 1ULL << vd->vdev_top->vdev_logical_ashift;
3258 ASSERT0(P2PHASE(zio->io_offset, log_align));
3259 ASSERT0(P2PHASE(zio->io_size, log_align));
3262 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3265 * If this is a repair I/O, and there's no self-healing involved --
3266 * that is, we're just resilvering what we expect to resilver --
3267 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3268 * This prevents spurious resilvering.
3270 * There are a few ways that we can end up creating these spurious
3273 * 1. A resilver i/o will be issued if any DVA in the BP has a
3274 * dirty DTL. The mirror code will issue resilver writes to
3275 * each DVA, including the one(s) that are not on vdevs with dirty
3278 * 2. With nested replication, which happens when we have a
3279 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3280 * For example, given mirror(replacing(A+B), C), it's likely that
3281 * only A is out of date (it's the new device). In this case, we'll
3282 * read from C, then use the data to resilver A+B -- but we don't
3283 * actually want to resilver B, just A. The top-level mirror has no
3284 * way to know this, so instead we just discard unnecessary repairs
3285 * as we work our way down the vdev tree.
3287 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3288 * The same logic applies to any form of nested replication: ditto
3289 * + mirror, RAID-Z + replacing, etc.
3291 * However, indirect vdevs point off to other vdevs which may have
3292 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3293 * will be properly bypassed instead.
3295 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3296 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3297 zio->io_txg != 0 && /* not a delegated i/o */
3298 vd->vdev_ops != &vdev_indirect_ops &&
3299 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3300 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3301 zio_vdev_io_bypass(zio);
3302 return (ZIO_PIPELINE_CONTINUE);
3305 if (vd->vdev_ops->vdev_op_leaf) {
3306 switch (zio->io_type) {
3308 if (vdev_cache_read(zio))
3309 return (ZIO_PIPELINE_CONTINUE);
3311 case ZIO_TYPE_WRITE:
3313 if ((zio = vdev_queue_io(zio)) == NULL)
3314 return (ZIO_PIPELINE_STOP);
3316 if (!vdev_accessible(vd, zio)) {
3317 zio->io_error = SET_ERROR(ENXIO);
3319 return (ZIO_PIPELINE_STOP);
3324 * Note that we ignore repair writes for TRIM because they can
3325 * conflict with normal writes. This isn't an issue because, by
3326 * definition, we only repair blocks that aren't freed.
3328 if (zio->io_type == ZIO_TYPE_WRITE &&
3329 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3330 !trim_map_write_start(zio))
3331 return (ZIO_PIPELINE_STOP);
3334 vd->vdev_ops->vdev_op_io_start(zio);
3335 return (ZIO_PIPELINE_STOP);
3339 zio_vdev_io_done(zio_t *zio)
3341 vdev_t *vd = zio->io_vd;
3342 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3343 boolean_t unexpected_error = B_FALSE;
3345 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3346 return (ZIO_PIPELINE_STOP);
3349 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3350 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3352 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3353 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3354 zio->io_type == ZIO_TYPE_FREE)) {
3356 if (zio->io_type == ZIO_TYPE_WRITE &&
3357 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3358 trim_map_write_done(zio);
3360 vdev_queue_io_done(zio);
3362 if (zio->io_type == ZIO_TYPE_WRITE)
3363 vdev_cache_write(zio);
3365 if (zio_injection_enabled && zio->io_error == 0)
3366 zio->io_error = zio_handle_device_injection(vd,
3369 if (zio_injection_enabled && zio->io_error == 0)
3370 zio->io_error = zio_handle_label_injection(zio, EIO);
3372 if (zio->io_error) {
3373 if (zio->io_error == ENOTSUP &&
3374 zio->io_type == ZIO_TYPE_FREE) {
3375 /* Not all devices support TRIM. */
3376 } else if (!vdev_accessible(vd, zio)) {
3377 zio->io_error = SET_ERROR(ENXIO);
3379 unexpected_error = B_TRUE;
3384 ops->vdev_op_io_done(zio);
3386 if (unexpected_error)
3387 VERIFY(vdev_probe(vd, zio) == NULL);
3389 return (ZIO_PIPELINE_CONTINUE);
3393 * This function is used to change the priority of an existing zio that is
3394 * currently in-flight. This is used by the arc to upgrade priority in the
3395 * event that a demand read is made for a block that is currently queued
3396 * as a scrub or async read IO. Otherwise, the high priority read request
3397 * would end up having to wait for the lower priority IO.
3400 zio_change_priority(zio_t *pio, zio_priority_t priority)
3402 zio_t *cio, *cio_next;
3403 zio_link_t *zl = NULL;
3405 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3407 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3408 vdev_queue_change_io_priority(pio, priority);
3410 pio->io_priority = priority;
3413 mutex_enter(&pio->io_lock);
3414 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3415 cio_next = zio_walk_children(pio, &zl);
3416 zio_change_priority(cio, priority);
3418 mutex_exit(&pio->io_lock);
3422 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3423 * disk, and use that to finish the checksum ereport later.
3426 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3427 const void *good_buf)
3429 /* no processing needed */
3430 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3435 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3437 void *buf = zio_buf_alloc(zio->io_size);
3439 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3441 zcr->zcr_cbinfo = zio->io_size;
3442 zcr->zcr_cbdata = buf;
3443 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3444 zcr->zcr_free = zio_buf_free;
3448 zio_vdev_io_assess(zio_t *zio)
3450 vdev_t *vd = zio->io_vd;
3452 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3453 return (ZIO_PIPELINE_STOP);
3456 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3457 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3459 if (zio->io_vsd != NULL) {
3460 zio->io_vsd_ops->vsd_free(zio);
3464 if (zio_injection_enabled && zio->io_error == 0)
3465 zio->io_error = zio_handle_fault_injection(zio, EIO);
3467 if (zio->io_type == ZIO_TYPE_FREE &&
3468 zio->io_priority != ZIO_PRIORITY_NOW) {
3469 switch (zio->io_error) {
3471 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3472 ZIO_TRIM_STAT_BUMP(success);
3475 ZIO_TRIM_STAT_BUMP(unsupported);
3478 ZIO_TRIM_STAT_BUMP(failed);
3484 * If the I/O failed, determine whether we should attempt to retry it.
3486 * On retry, we cut in line in the issue queue, since we don't want
3487 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3489 if (zio->io_error && vd == NULL &&
3490 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3491 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3492 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3494 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3495 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3496 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3497 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3498 zio_requeue_io_start_cut_in_line);
3499 return (ZIO_PIPELINE_STOP);
3503 * If we got an error on a leaf device, convert it to ENXIO
3504 * if the device is not accessible at all.
3506 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3507 !vdev_accessible(vd, zio))
3508 zio->io_error = SET_ERROR(ENXIO);
3511 * If we can't write to an interior vdev (mirror or RAID-Z),
3512 * set vdev_cant_write so that we stop trying to allocate from it.
3514 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3515 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3516 vd->vdev_cant_write = B_TRUE;
3520 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3521 * attempts will ever succeed. In this case we set a persistent bit so
3522 * that we don't bother with it in the future.
3524 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3525 zio->io_type == ZIO_TYPE_IOCTL &&
3526 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3527 vd->vdev_nowritecache = B_TRUE;
3530 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3532 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3533 zio->io_physdone != NULL) {
3534 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3535 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3536 zio->io_physdone(zio->io_logical);
3539 return (ZIO_PIPELINE_CONTINUE);
3543 zio_vdev_io_reissue(zio_t *zio)
3545 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3546 ASSERT(zio->io_error == 0);
3548 zio->io_stage >>= 1;
3552 zio_vdev_io_redone(zio_t *zio)
3554 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3556 zio->io_stage >>= 1;
3560 zio_vdev_io_bypass(zio_t *zio)
3562 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3563 ASSERT(zio->io_error == 0);
3565 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3566 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3570 * ==========================================================================
3571 * Generate and verify checksums
3572 * ==========================================================================
3575 zio_checksum_generate(zio_t *zio)
3577 blkptr_t *bp = zio->io_bp;
3578 enum zio_checksum checksum;
3582 * This is zio_write_phys().
3583 * We're either generating a label checksum, or none at all.
3585 checksum = zio->io_prop.zp_checksum;
3587 if (checksum == ZIO_CHECKSUM_OFF)
3588 return (ZIO_PIPELINE_CONTINUE);
3590 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3592 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3593 ASSERT(!IO_IS_ALLOCATING(zio));
3594 checksum = ZIO_CHECKSUM_GANG_HEADER;
3596 checksum = BP_GET_CHECKSUM(bp);
3600 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3602 return (ZIO_PIPELINE_CONTINUE);
3606 zio_checksum_verify(zio_t *zio)
3608 zio_bad_cksum_t info;
3609 blkptr_t *bp = zio->io_bp;
3612 ASSERT(zio->io_vd != NULL);
3616 * This is zio_read_phys().
3617 * We're either verifying a label checksum, or nothing at all.
3619 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3620 return (ZIO_PIPELINE_CONTINUE);
3622 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3625 if ((error = zio_checksum_error(zio, &info)) != 0) {
3626 zio->io_error = error;
3627 if (error == ECKSUM &&
3628 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3629 zfs_ereport_start_checksum(zio->io_spa,
3630 zio->io_vd, zio, zio->io_offset,
3631 zio->io_size, NULL, &info);
3635 return (ZIO_PIPELINE_CONTINUE);
3639 * Called by RAID-Z to ensure we don't compute the checksum twice.
3642 zio_checksum_verified(zio_t *zio)
3644 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3648 * ==========================================================================
3649 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3650 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3651 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3652 * indicate errors that are specific to one I/O, and most likely permanent.
3653 * Any other error is presumed to be worse because we weren't expecting it.
3654 * ==========================================================================
3657 zio_worst_error(int e1, int e2)
3659 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3662 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3663 if (e1 == zio_error_rank[r1])
3666 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3667 if (e2 == zio_error_rank[r2])
3670 return (r1 > r2 ? e1 : e2);
3674 * ==========================================================================
3676 * ==========================================================================
3679 zio_ready(zio_t *zio)
3681 blkptr_t *bp = zio->io_bp;
3682 zio_t *pio, *pio_next;
3683 zio_link_t *zl = NULL;
3685 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3687 return (ZIO_PIPELINE_STOP);
3690 if (zio->io_ready) {
3691 ASSERT(IO_IS_ALLOCATING(zio));
3692 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3693 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3694 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3699 if (bp != NULL && bp != &zio->io_bp_copy)
3700 zio->io_bp_copy = *bp;
3702 if (zio->io_error != 0) {
3703 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3705 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3706 ASSERT(IO_IS_ALLOCATING(zio));
3707 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3709 * We were unable to allocate anything, unreserve and
3710 * issue the next I/O to allocate.
3712 metaslab_class_throttle_unreserve(
3713 spa_normal_class(zio->io_spa),
3714 zio->io_prop.zp_copies, zio->io_allocator, zio);
3715 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
3719 mutex_enter(&zio->io_lock);
3720 zio->io_state[ZIO_WAIT_READY] = 1;
3721 pio = zio_walk_parents(zio, &zl);
3722 mutex_exit(&zio->io_lock);
3725 * As we notify zio's parents, new parents could be added.
3726 * New parents go to the head of zio's io_parent_list, however,
3727 * so we will (correctly) not notify them. The remainder of zio's
3728 * io_parent_list, from 'pio_next' onward, cannot change because
3729 * all parents must wait for us to be done before they can be done.
3731 for (; pio != NULL; pio = pio_next) {
3732 pio_next = zio_walk_parents(zio, &zl);
3733 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3736 if (zio->io_flags & ZIO_FLAG_NODATA) {
3737 if (BP_IS_GANG(bp)) {
3738 zio->io_flags &= ~ZIO_FLAG_NODATA;
3740 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3741 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3745 if (zio_injection_enabled &&
3746 zio->io_spa->spa_syncing_txg == zio->io_txg)
3747 zio_handle_ignored_writes(zio);
3749 return (ZIO_PIPELINE_CONTINUE);
3753 * Update the allocation throttle accounting.
3756 zio_dva_throttle_done(zio_t *zio)
3758 zio_t *lio = zio->io_logical;
3759 zio_t *pio = zio_unique_parent(zio);
3760 vdev_t *vd = zio->io_vd;
3761 int flags = METASLAB_ASYNC_ALLOC;
3763 ASSERT3P(zio->io_bp, !=, NULL);
3764 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3765 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3766 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3768 ASSERT3P(vd, ==, vd->vdev_top);
3769 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3770 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3771 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3772 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3775 * Parents of gang children can have two flavors -- ones that
3776 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3777 * and ones that allocated the constituent blocks. The allocation
3778 * throttle needs to know the allocating parent zio so we must find
3781 if (pio->io_child_type == ZIO_CHILD_GANG) {
3783 * If our parent is a rewrite gang child then our grandparent
3784 * would have been the one that performed the allocation.
3786 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3787 pio = zio_unique_parent(pio);
3788 flags |= METASLAB_GANG_CHILD;
3791 ASSERT(IO_IS_ALLOCATING(pio));
3792 ASSERT3P(zio, !=, zio->io_logical);
3793 ASSERT(zio->io_logical != NULL);
3794 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3795 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3797 mutex_enter(&pio->io_lock);
3798 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
3799 pio->io_allocator, B_TRUE);
3800 mutex_exit(&pio->io_lock);
3802 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3803 1, pio->io_allocator, pio);
3806 * Call into the pipeline to see if there is more work that
3807 * needs to be done. If there is work to be done it will be
3808 * dispatched to another taskq thread.
3810 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
3814 zio_done(zio_t *zio)
3816 spa_t *spa = zio->io_spa;
3817 zio_t *lio = zio->io_logical;
3818 blkptr_t *bp = zio->io_bp;
3819 vdev_t *vd = zio->io_vd;
3820 uint64_t psize = zio->io_size;
3821 zio_t *pio, *pio_next;
3822 metaslab_class_t *mc = spa_normal_class(spa);
3823 zio_link_t *zl = NULL;
3826 * If our children haven't all completed,
3827 * wait for them and then repeat this pipeline stage.
3829 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3830 return (ZIO_PIPELINE_STOP);
3834 * If the allocation throttle is enabled, then update the accounting.
3835 * We only track child I/Os that are part of an allocating async
3836 * write. We must do this since the allocation is performed
3837 * by the logical I/O but the actual write is done by child I/Os.
3839 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3840 zio->io_child_type == ZIO_CHILD_VDEV) {
3841 ASSERT(mc->mc_alloc_throttle_enabled);
3842 zio_dva_throttle_done(zio);
3846 * If the allocation throttle is enabled, verify that
3847 * we have decremented the refcounts for every I/O that was throttled.
3849 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3850 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3851 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3853 metaslab_group_alloc_verify(spa, zio->io_bp, zio,
3855 VERIFY(refcount_not_held(&mc->mc_alloc_slots[zio->io_allocator],
3859 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3860 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3861 ASSERT(zio->io_children[c][w] == 0);
3863 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3864 ASSERT(bp->blk_pad[0] == 0);
3865 ASSERT(bp->blk_pad[1] == 0);
3866 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3867 (bp == zio_unique_parent(zio)->io_bp));
3868 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3869 zio->io_bp_override == NULL &&
3870 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3871 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3872 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3873 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3874 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3876 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3877 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3881 * If there were child vdev/gang/ddt errors, they apply to us now.
3883 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3884 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3885 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3888 * If the I/O on the transformed data was successful, generate any
3889 * checksum reports now while we still have the transformed data.
3891 if (zio->io_error == 0) {
3892 while (zio->io_cksum_report != NULL) {
3893 zio_cksum_report_t *zcr = zio->io_cksum_report;
3894 uint64_t align = zcr->zcr_align;
3895 uint64_t asize = P2ROUNDUP(psize, align);
3897 abd_t *adata = zio->io_abd;
3899 if (asize != psize) {
3900 adata = abd_alloc_linear(asize, B_TRUE);
3901 abd_copy(adata, zio->io_abd, psize);
3902 abd_zero_off(adata, psize, asize - psize);
3906 abuf = abd_borrow_buf_copy(adata, asize);
3908 zio->io_cksum_report = zcr->zcr_next;
3909 zcr->zcr_next = NULL;
3910 zcr->zcr_finish(zcr, abuf);
3911 zfs_ereport_free_checksum(zcr);
3914 abd_return_buf(adata, abuf, asize);
3921 zio_pop_transforms(zio); /* note: may set zio->io_error */
3923 vdev_stat_update(zio, psize);
3925 if (zio->io_error) {
3927 * If this I/O is attached to a particular vdev,
3928 * generate an error message describing the I/O failure
3929 * at the block level. We ignore these errors if the
3930 * device is currently unavailable.
3932 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3933 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3935 if ((zio->io_error == EIO || !(zio->io_flags &
3936 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3939 * For logical I/O requests, tell the SPA to log the
3940 * error and generate a logical data ereport.
3942 spa_log_error(spa, zio);
3943 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3948 if (zio->io_error && zio == lio) {
3950 * Determine whether zio should be reexecuted. This will
3951 * propagate all the way to the root via zio_notify_parent().
3953 ASSERT(vd == NULL && bp != NULL);
3954 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3956 if (IO_IS_ALLOCATING(zio) &&
3957 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3958 if (zio->io_error != ENOSPC)
3959 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3961 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3964 if ((zio->io_type == ZIO_TYPE_READ ||
3965 zio->io_type == ZIO_TYPE_FREE) &&
3966 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3967 zio->io_error == ENXIO &&
3968 spa_load_state(spa) == SPA_LOAD_NONE &&
3969 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3970 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3972 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3973 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3976 * Here is a possibly good place to attempt to do
3977 * either combinatorial reconstruction or error correction
3978 * based on checksums. It also might be a good place
3979 * to send out preliminary ereports before we suspend
3985 * If there were logical child errors, they apply to us now.
3986 * We defer this until now to avoid conflating logical child
3987 * errors with errors that happened to the zio itself when
3988 * updating vdev stats and reporting FMA events above.
3990 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3992 if ((zio->io_error || zio->io_reexecute) &&
3993 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3994 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3995 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3997 zio_gang_tree_free(&zio->io_gang_tree);
4000 * Godfather I/Os should never suspend.
4002 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4003 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4004 zio->io_reexecute = 0;
4006 if (zio->io_reexecute) {
4008 * This is a logical I/O that wants to reexecute.
4010 * Reexecute is top-down. When an i/o fails, if it's not
4011 * the root, it simply notifies its parent and sticks around.
4012 * The parent, seeing that it still has children in zio_done(),
4013 * does the same. This percolates all the way up to the root.
4014 * The root i/o will reexecute or suspend the entire tree.
4016 * This approach ensures that zio_reexecute() honors
4017 * all the original i/o dependency relationships, e.g.
4018 * parents not executing until children are ready.
4020 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4022 zio->io_gang_leader = NULL;
4024 mutex_enter(&zio->io_lock);
4025 zio->io_state[ZIO_WAIT_DONE] = 1;
4026 mutex_exit(&zio->io_lock);
4029 * "The Godfather" I/O monitors its children but is
4030 * not a true parent to them. It will track them through
4031 * the pipeline but severs its ties whenever they get into
4032 * trouble (e.g. suspended). This allows "The Godfather"
4033 * I/O to return status without blocking.
4036 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4038 zio_link_t *remove_zl = zl;
4039 pio_next = zio_walk_parents(zio, &zl);
4041 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4042 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4043 zio_remove_child(pio, zio, remove_zl);
4044 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4048 if ((pio = zio_unique_parent(zio)) != NULL) {
4050 * We're not a root i/o, so there's nothing to do
4051 * but notify our parent. Don't propagate errors
4052 * upward since we haven't permanently failed yet.
4054 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4055 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4056 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4057 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4059 * We'd fail again if we reexecuted now, so suspend
4060 * until conditions improve (e.g. device comes online).
4062 zio_suspend(spa, zio);
4065 * Reexecution is potentially a huge amount of work.
4066 * Hand it off to the otherwise-unused claim taskq.
4068 #if defined(illumos) || !defined(_KERNEL)
4069 ASSERT(zio->io_tqent.tqent_next == NULL);
4071 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4073 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4074 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4077 return (ZIO_PIPELINE_STOP);
4080 ASSERT(zio->io_child_count == 0);
4081 ASSERT(zio->io_reexecute == 0);
4082 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4085 * Report any checksum errors, since the I/O is complete.
4087 while (zio->io_cksum_report != NULL) {
4088 zio_cksum_report_t *zcr = zio->io_cksum_report;
4089 zio->io_cksum_report = zcr->zcr_next;
4090 zcr->zcr_next = NULL;
4091 zcr->zcr_finish(zcr, NULL);
4092 zfs_ereport_free_checksum(zcr);
4096 * It is the responsibility of the done callback to ensure that this
4097 * particular zio is no longer discoverable for adoption, and as
4098 * such, cannot acquire any new parents.
4103 mutex_enter(&zio->io_lock);
4104 zio->io_state[ZIO_WAIT_DONE] = 1;
4105 mutex_exit(&zio->io_lock);
4108 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4109 zio_link_t *remove_zl = zl;
4110 pio_next = zio_walk_parents(zio, &zl);
4111 zio_remove_child(pio, zio, remove_zl);
4112 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4115 if (zio->io_waiter != NULL) {
4116 mutex_enter(&zio->io_lock);
4117 zio->io_executor = NULL;
4118 cv_broadcast(&zio->io_cv);
4119 mutex_exit(&zio->io_lock);
4124 return (ZIO_PIPELINE_STOP);
4128 * ==========================================================================
4129 * I/O pipeline definition
4130 * ==========================================================================
4132 static zio_pipe_stage_t *zio_pipeline[] = {
4139 zio_checksum_generate,
4155 zio_checksum_verify,
4163 * Compare two zbookmark_phys_t's to see which we would reach first in a
4164 * pre-order traversal of the object tree.
4166 * This is simple in every case aside from the meta-dnode object. For all other
4167 * objects, we traverse them in order (object 1 before object 2, and so on).
4168 * However, all of these objects are traversed while traversing object 0, since
4169 * the data it points to is the list of objects. Thus, we need to convert to a
4170 * canonical representation so we can compare meta-dnode bookmarks to
4171 * non-meta-dnode bookmarks.
4173 * We do this by calculating "equivalents" for each field of the zbookmark.
4174 * zbookmarks outside of the meta-dnode use their own object and level, and
4175 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4176 * blocks this bookmark refers to) by multiplying their blkid by their span
4177 * (the number of L0 blocks contained within one block at their level).
4178 * zbookmarks inside the meta-dnode calculate their object equivalent
4179 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4180 * level + 1<<31 (any value larger than a level could ever be) for their level.
4181 * This causes them to always compare before a bookmark in their object
4182 * equivalent, compare appropriately to bookmarks in other objects, and to
4183 * compare appropriately to other bookmarks in the meta-dnode.
4186 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4187 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4190 * These variables represent the "equivalent" values for the zbookmark,
4191 * after converting zbookmarks inside the meta dnode to their
4192 * normal-object equivalents.
4194 uint64_t zb1obj, zb2obj;
4195 uint64_t zb1L0, zb2L0;
4196 uint64_t zb1level, zb2level;
4198 if (zb1->zb_object == zb2->zb_object &&
4199 zb1->zb_level == zb2->zb_level &&
4200 zb1->zb_blkid == zb2->zb_blkid)
4204 * BP_SPANB calculates the span in blocks.
4206 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4207 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4209 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4210 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4212 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4214 zb1obj = zb1->zb_object;
4215 zb1level = zb1->zb_level;
4218 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4219 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4221 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4223 zb2obj = zb2->zb_object;
4224 zb2level = zb2->zb_level;
4227 /* Now that we have a canonical representation, do the comparison. */
4228 if (zb1obj != zb2obj)
4229 return (zb1obj < zb2obj ? -1 : 1);
4230 else if (zb1L0 != zb2L0)
4231 return (zb1L0 < zb2L0 ? -1 : 1);
4232 else if (zb1level != zb2level)
4233 return (zb1level > zb2level ? -1 : 1);
4235 * This can (theoretically) happen if the bookmarks have the same object
4236 * and level, but different blkids, if the block sizes are not the same.
4237 * There is presently no way to change the indirect block sizes
4243 * This function checks the following: given that last_block is the place that
4244 * our traversal stopped last time, does that guarantee that we've visited
4245 * every node under subtree_root? Therefore, we can't just use the raw output
4246 * of zbookmark_compare. We have to pass in a modified version of
4247 * subtree_root; by incrementing the block id, and then checking whether
4248 * last_block is before or equal to that, we can tell whether or not having
4249 * visited last_block implies that all of subtree_root's children have been
4253 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4254 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4256 zbookmark_phys_t mod_zb = *subtree_root;
4258 ASSERT(last_block->zb_level == 0);
4260 /* The objset_phys_t isn't before anything. */
4265 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4266 * data block size in sectors, because that variable is only used if
4267 * the bookmark refers to a block in the meta-dnode. Since we don't
4268 * know without examining it what object it refers to, and there's no
4269 * harm in passing in this value in other cases, we always pass it in.
4271 * We pass in 0 for the indirect block size shift because zb2 must be
4272 * level 0. The indirect block size is only used to calculate the span
4273 * of the bookmark, but since the bookmark must be level 0, the span is
4274 * always 1, so the math works out.
4276 * If you make changes to how the zbookmark_compare code works, be sure
4277 * to make sure that this code still works afterwards.
4279 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4280 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,