4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/metaslab_impl.h>
46 SYSCTL_DECL(_vfs_zfs);
47 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
48 #if defined(__amd64__)
49 static int zio_use_uma = 1;
51 static int zio_use_uma = 0;
53 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
54 "Use uma(9) for ZIO allocations");
55 static int zio_exclude_metadata = 0;
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
57 "Exclude metadata buffers from dumps as well");
59 zio_trim_stats_t zio_trim_stats = {
60 { "bytes", KSTAT_DATA_UINT64,
61 "Number of bytes successfully TRIMmed" },
62 { "success", KSTAT_DATA_UINT64,
63 "Number of successful TRIM requests" },
64 { "unsupported", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed because TRIM is not supported" },
66 { "failed", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed for reasons other than not supported" },
70 static kstat_t *zio_trim_ksp;
73 * ==========================================================================
74 * I/O type descriptions
75 * ==========================================================================
77 const char *zio_type_name[ZIO_TYPES] = {
78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
82 boolean_t zio_dva_throttle_enabled = B_TRUE;
83 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
84 &zio_dva_throttle_enabled, 0, "");
87 * ==========================================================================
89 * ==========================================================================
91 kmem_cache_t *zio_cache;
92 kmem_cache_t *zio_link_cache;
93 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
94 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97 extern vmem_t *zio_alloc_arena;
100 #define ZIO_PIPELINE_CONTINUE 0x100
101 #define ZIO_PIPELINE_STOP 0x101
103 #define BP_SPANB(indblkshift, level) \
104 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
105 #define COMPARE_META_LEVEL 0x80000000ul
107 * The following actions directly effect the spa's sync-to-convergence logic.
108 * The values below define the sync pass when we start performing the action.
109 * Care should be taken when changing these values as they directly impact
110 * spa_sync() performance. Tuning these values may introduce subtle performance
111 * pathologies and should only be done in the context of performance analysis.
112 * These tunables will eventually be removed and replaced with #defines once
113 * enough analysis has been done to determine optimal values.
115 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
116 * regular blocks are not deferred.
118 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
119 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
120 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
121 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
123 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
124 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
126 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
138 int zio_buf_debug_limit = 16384;
140 int zio_buf_debug_limit = 0;
144 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
150 zio_cache = kmem_cache_create("zio_cache",
151 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
152 zio_link_cache = kmem_cache_create("zio_link_cache",
153 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 * For small buffers, we want a cache for each multiple of
159 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
160 * for each quarter-power of 2.
162 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
163 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
166 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
174 * If we are using watchpoints, put each buffer on its own page,
175 * to eliminate the performance overhead of trapping to the
176 * kernel when modifying a non-watched buffer that shares the
177 * page with a watched buffer.
179 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
183 if (size <= 4 * SPA_MINBLOCKSIZE) {
184 align = SPA_MINBLOCKSIZE;
185 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
186 align = MIN(p2 >> 2, PAGESIZE);
191 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
192 zio_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL, cflags);
196 * Since zio_data bufs do not appear in crash dumps, we
197 * pass KMC_NOTOUCH so that no allocator metadata is
198 * stored with the buffers.
200 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
201 zio_data_buf_cache[c] = kmem_cache_create(name, size,
202 align, NULL, NULL, NULL, NULL, NULL,
203 cflags | KMC_NOTOUCH | KMC_NODEBUG);
208 ASSERT(zio_buf_cache[c] != NULL);
209 if (zio_buf_cache[c - 1] == NULL)
210 zio_buf_cache[c - 1] = zio_buf_cache[c];
212 ASSERT(zio_data_buf_cache[c] != NULL);
213 if (zio_data_buf_cache[c - 1] == NULL)
214 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
220 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 if (zio_trim_ksp != NULL) {
226 zio_trim_ksp->ks_data = &zio_trim_stats;
227 kstat_install(zio_trim_ksp);
235 kmem_cache_t *last_cache = NULL;
236 kmem_cache_t *last_data_cache = NULL;
238 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
239 if (zio_buf_cache[c] != last_cache) {
240 last_cache = zio_buf_cache[c];
241 kmem_cache_destroy(zio_buf_cache[c]);
243 zio_buf_cache[c] = NULL;
245 if (zio_data_buf_cache[c] != last_data_cache) {
246 last_data_cache = zio_data_buf_cache[c];
247 kmem_cache_destroy(zio_data_buf_cache[c]);
249 zio_data_buf_cache[c] = NULL;
252 kmem_cache_destroy(zio_link_cache);
253 kmem_cache_destroy(zio_cache);
257 if (zio_trim_ksp != NULL) {
258 kstat_delete(zio_trim_ksp);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc_impl(size_t size, boolean_t canwait)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c],
285 canwait ? KM_PUSHPAGE : KM_NOSLEEP));
287 return (kmem_alloc(size,
288 (canwait ? KM_SLEEP : KM_NOSLEEP) | flags));
293 zio_buf_alloc(size_t size)
295 return (zio_buf_alloc_impl(size, B_TRUE));
299 zio_buf_alloc_nowait(size_t size)
301 return (zio_buf_alloc_impl(size, B_FALSE));
305 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
306 * crashdump if the kernel panics. This exists so that we will limit the amount
307 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
308 * of kernel heap dumped to disk when the kernel panics)
311 zio_data_buf_alloc(size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
320 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
324 zio_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_buf_cache[c], buf);
333 kmem_free(buf, size);
337 zio_data_buf_free(void *buf, size_t size)
339 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
341 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
344 kmem_cache_free(zio_data_buf_cache[c], buf);
346 kmem_free(buf, size);
350 * ==========================================================================
351 * Push and pop I/O transform buffers
352 * ==========================================================================
355 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
356 zio_transform_func_t *transform)
358 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
360 zt->zt_orig_data = zio->io_data;
361 zt->zt_orig_size = zio->io_size;
362 zt->zt_bufsize = bufsize;
363 zt->zt_transform = transform;
365 zt->zt_next = zio->io_transform_stack;
366 zio->io_transform_stack = zt;
373 zio_pop_transforms(zio_t *zio)
377 while ((zt = zio->io_transform_stack) != NULL) {
378 if (zt->zt_transform != NULL)
379 zt->zt_transform(zio,
380 zt->zt_orig_data, zt->zt_orig_size);
382 if (zt->zt_bufsize != 0)
383 zio_buf_free(zio->io_data, zt->zt_bufsize);
385 zio->io_data = zt->zt_orig_data;
386 zio->io_size = zt->zt_orig_size;
387 zio->io_transform_stack = zt->zt_next;
389 kmem_free(zt, sizeof (zio_transform_t));
394 * ==========================================================================
395 * I/O transform callbacks for subblocks and decompression
396 * ==========================================================================
399 zio_subblock(zio_t *zio, void *data, uint64_t size)
401 ASSERT(zio->io_size > size);
403 if (zio->io_type == ZIO_TYPE_READ)
404 bcopy(zio->io_data, data, size);
408 zio_decompress(zio_t *zio, void *data, uint64_t size)
410 if (zio->io_error == 0 &&
411 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
412 zio->io_data, data, zio->io_size, size) != 0)
413 zio->io_error = SET_ERROR(EIO);
417 * ==========================================================================
418 * I/O parent/child relationships and pipeline interlocks
419 * ==========================================================================
422 zio_walk_parents(zio_t *cio, zio_link_t **zl)
424 list_t *pl = &cio->io_parent_list;
426 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
430 ASSERT((*zl)->zl_child == cio);
431 return ((*zl)->zl_parent);
435 zio_walk_children(zio_t *pio, zio_link_t **zl)
437 list_t *cl = &pio->io_child_list;
439 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
443 ASSERT((*zl)->zl_parent == pio);
444 return ((*zl)->zl_child);
448 zio_unique_parent(zio_t *cio)
450 zio_link_t *zl = NULL;
451 zio_t *pio = zio_walk_parents(cio, &zl);
453 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
458 zio_add_child(zio_t *pio, zio_t *cio)
460 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
463 * Logical I/Os can have logical, gang, or vdev children.
464 * Gang I/Os can have gang or vdev children.
465 * Vdev I/Os can only have vdev children.
466 * The following ASSERT captures all of these constraints.
468 ASSERT(cio->io_child_type <= pio->io_child_type);
473 mutex_enter(&cio->io_lock);
474 mutex_enter(&pio->io_lock);
476 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
478 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
479 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
481 list_insert_head(&pio->io_child_list, zl);
482 list_insert_head(&cio->io_parent_list, zl);
484 pio->io_child_count++;
485 cio->io_parent_count++;
487 mutex_exit(&pio->io_lock);
488 mutex_exit(&cio->io_lock);
492 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
494 ASSERT(zl->zl_parent == pio);
495 ASSERT(zl->zl_child == cio);
497 mutex_enter(&cio->io_lock);
498 mutex_enter(&pio->io_lock);
500 list_remove(&pio->io_child_list, zl);
501 list_remove(&cio->io_parent_list, zl);
503 pio->io_child_count--;
504 cio->io_parent_count--;
506 mutex_exit(&pio->io_lock);
507 mutex_exit(&cio->io_lock);
509 kmem_cache_free(zio_link_cache, zl);
513 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
515 uint64_t *countp = &zio->io_children[child][wait];
516 boolean_t waiting = B_FALSE;
518 mutex_enter(&zio->io_lock);
519 ASSERT(zio->io_stall == NULL);
522 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
523 zio->io_stall = countp;
526 mutex_exit(&zio->io_lock);
532 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
534 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
535 int *errorp = &pio->io_child_error[zio->io_child_type];
537 mutex_enter(&pio->io_lock);
538 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
539 *errorp = zio_worst_error(*errorp, zio->io_error);
540 pio->io_reexecute |= zio->io_reexecute;
541 ASSERT3U(*countp, >, 0);
545 if (*countp == 0 && pio->io_stall == countp) {
546 zio_taskq_type_t type =
547 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
549 pio->io_stall = NULL;
550 mutex_exit(&pio->io_lock);
552 * Dispatch the parent zio in its own taskq so that
553 * the child can continue to make progress. This also
554 * prevents overflowing the stack when we have deeply nested
555 * parent-child relationships.
557 zio_taskq_dispatch(pio, type, B_FALSE);
559 mutex_exit(&pio->io_lock);
564 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
566 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
567 zio->io_error = zio->io_child_error[c];
571 zio_bookmark_compare(const void *x1, const void *x2)
573 const zio_t *z1 = x1;
574 const zio_t *z2 = x2;
576 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
578 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
581 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
583 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
586 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
588 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
591 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
593 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
605 * ==========================================================================
606 * Create the various types of I/O (read, write, free, etc)
607 * ==========================================================================
610 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
611 void *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
612 void *private, zio_type_t type, zio_priority_t priority,
613 enum zio_flag flags, vdev_t *vd, uint64_t offset,
614 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
618 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
619 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
620 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
622 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
623 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
624 ASSERT(vd || stage == ZIO_STAGE_OPEN);
626 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
628 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
629 bzero(zio, sizeof (zio_t));
631 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
632 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
634 list_create(&zio->io_parent_list, sizeof (zio_link_t),
635 offsetof(zio_link_t, zl_parent_node));
636 list_create(&zio->io_child_list, sizeof (zio_link_t),
637 offsetof(zio_link_t, zl_child_node));
638 metaslab_trace_init(&zio->io_alloc_list);
641 zio->io_child_type = ZIO_CHILD_VDEV;
642 else if (flags & ZIO_FLAG_GANG_CHILD)
643 zio->io_child_type = ZIO_CHILD_GANG;
644 else if (flags & ZIO_FLAG_DDT_CHILD)
645 zio->io_child_type = ZIO_CHILD_DDT;
647 zio->io_child_type = ZIO_CHILD_LOGICAL;
650 zio->io_bp = (blkptr_t *)bp;
651 zio->io_bp_copy = *bp;
652 zio->io_bp_orig = *bp;
653 if (type != ZIO_TYPE_WRITE ||
654 zio->io_child_type == ZIO_CHILD_DDT)
655 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
656 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
657 zio->io_logical = zio;
658 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
659 pipeline |= ZIO_GANG_STAGES;
665 zio->io_private = private;
667 zio->io_priority = priority;
669 zio->io_offset = offset;
670 zio->io_orig_data = zio->io_data = data;
671 zio->io_orig_size = zio->io_size = psize;
672 zio->io_lsize = lsize;
673 zio->io_orig_flags = zio->io_flags = flags;
674 zio->io_orig_stage = zio->io_stage = stage;
675 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
676 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
678 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
679 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
682 zio->io_bookmark = *zb;
685 if (zio->io_logical == NULL)
686 zio->io_logical = pio->io_logical;
687 if (zio->io_child_type == ZIO_CHILD_GANG)
688 zio->io_gang_leader = pio->io_gang_leader;
689 zio_add_child(pio, zio);
696 zio_destroy(zio_t *zio)
698 metaslab_trace_fini(&zio->io_alloc_list);
699 list_destroy(&zio->io_parent_list);
700 list_destroy(&zio->io_child_list);
701 mutex_destroy(&zio->io_lock);
702 cv_destroy(&zio->io_cv);
703 kmem_cache_free(zio_cache, zio);
707 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
708 void *private, enum zio_flag flags)
712 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
713 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
714 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
720 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
722 return (zio_null(NULL, spa, NULL, done, private, flags));
726 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
728 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
729 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
730 bp, (longlong_t)BP_GET_TYPE(bp));
732 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
733 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
734 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
735 bp, (longlong_t)BP_GET_CHECKSUM(bp));
737 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
738 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
739 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
740 bp, (longlong_t)BP_GET_COMPRESS(bp));
742 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
743 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
744 bp, (longlong_t)BP_GET_LSIZE(bp));
746 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
747 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
748 bp, (longlong_t)BP_GET_PSIZE(bp));
751 if (BP_IS_EMBEDDED(bp)) {
752 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
753 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
754 bp, (longlong_t)BPE_GET_ETYPE(bp));
759 * Pool-specific checks.
761 * Note: it would be nice to verify that the blk_birth and
762 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
763 * allows the birth time of log blocks (and dmu_sync()-ed blocks
764 * that are in the log) to be arbitrarily large.
766 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
767 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
768 if (vdevid >= spa->spa_root_vdev->vdev_children) {
769 zfs_panic_recover("blkptr at %p DVA %u has invalid "
771 bp, i, (longlong_t)vdevid);
774 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
776 zfs_panic_recover("blkptr at %p DVA %u has invalid "
778 bp, i, (longlong_t)vdevid);
781 if (vd->vdev_ops == &vdev_hole_ops) {
782 zfs_panic_recover("blkptr at %p DVA %u has hole "
784 bp, i, (longlong_t)vdevid);
787 if (vd->vdev_ops == &vdev_missing_ops) {
789 * "missing" vdevs are valid during import, but we
790 * don't have their detailed info (e.g. asize), so
791 * we can't perform any more checks on them.
795 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
796 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
798 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
799 if (offset + asize > vd->vdev_asize) {
800 zfs_panic_recover("blkptr at %p DVA %u has invalid "
802 bp, i, (longlong_t)offset);
808 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
809 void *data, uint64_t size, zio_done_func_t *done, void *private,
810 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
814 zfs_blkptr_verify(spa, bp);
816 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
817 data, size, size, done, private,
818 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
819 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
820 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
826 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
827 void *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
828 zio_done_func_t *ready, zio_done_func_t *children_ready,
829 zio_done_func_t *physdone, zio_done_func_t *done,
830 void *private, zio_priority_t priority, enum zio_flag flags,
831 const zbookmark_phys_t *zb)
835 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
836 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
837 zp->zp_compress >= ZIO_COMPRESS_OFF &&
838 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
839 DMU_OT_IS_VALID(zp->zp_type) &&
842 zp->zp_copies <= spa_max_replication(spa));
844 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
845 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
846 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
847 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
849 zio->io_ready = ready;
850 zio->io_children_ready = children_ready;
851 zio->io_physdone = physdone;
855 * Data can be NULL if we are going to call zio_write_override() to
856 * provide the already-allocated BP. But we may need the data to
857 * verify a dedup hit (if requested). In this case, don't try to
858 * dedup (just take the already-allocated BP verbatim).
860 if (data == NULL && zio->io_prop.zp_dedup_verify) {
861 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
868 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
869 uint64_t size, zio_done_func_t *done, void *private,
870 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
874 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
875 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
876 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
882 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
884 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
885 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
886 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
887 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
890 * We must reset the io_prop to match the values that existed
891 * when the bp was first written by dmu_sync() keeping in mind
892 * that nopwrite and dedup are mutually exclusive.
894 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
895 zio->io_prop.zp_nopwrite = nopwrite;
896 zio->io_prop.zp_copies = copies;
897 zio->io_bp_override = bp;
901 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
905 * The check for EMBEDDED is a performance optimization. We
906 * process the free here (by ignoring it) rather than
907 * putting it on the list and then processing it in zio_free_sync().
909 if (BP_IS_EMBEDDED(bp))
911 metaslab_check_free(spa, bp);
914 * Frees that are for the currently-syncing txg, are not going to be
915 * deferred, and which will not need to do a read (i.e. not GANG or
916 * DEDUP), can be processed immediately. Otherwise, put them on the
917 * in-memory list for later processing.
919 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
920 txg != spa->spa_syncing_txg ||
921 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
922 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
924 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
925 BP_GET_PSIZE(bp), 0)));
930 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
931 uint64_t size, enum zio_flag flags)
934 enum zio_stage stage = ZIO_FREE_PIPELINE;
936 ASSERT(!BP_IS_HOLE(bp));
937 ASSERT(spa_syncing_txg(spa) == txg);
938 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
940 if (BP_IS_EMBEDDED(bp))
941 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
943 metaslab_check_free(spa, bp);
946 if (zfs_trim_enabled)
947 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
948 ZIO_STAGE_VDEV_IO_ASSESS;
950 * GANG and DEDUP blocks can induce a read (for the gang block header,
951 * or the DDT), so issue them asynchronously so that this thread is
954 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
955 stage |= ZIO_STAGE_ISSUE_ASYNC;
957 flags |= ZIO_FLAG_DONT_QUEUE;
959 zio = zio_create(pio, spa, txg, bp, NULL, size,
960 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
961 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
967 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
968 zio_done_func_t *done, void *private, enum zio_flag flags)
972 dprintf_bp(bp, "claiming in txg %llu", txg);
974 if (BP_IS_EMBEDDED(bp))
975 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
978 * A claim is an allocation of a specific block. Claims are needed
979 * to support immediate writes in the intent log. The issue is that
980 * immediate writes contain committed data, but in a txg that was
981 * *not* committed. Upon opening the pool after an unclean shutdown,
982 * the intent log claims all blocks that contain immediate write data
983 * so that the SPA knows they're in use.
985 * All claims *must* be resolved in the first txg -- before the SPA
986 * starts allocating blocks -- so that nothing is allocated twice.
987 * If txg == 0 we just verify that the block is claimable.
989 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
990 ASSERT(txg == spa_first_txg(spa) || txg == 0);
991 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
993 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
994 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
995 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
996 ASSERT0(zio->io_queued_timestamp);
1002 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1003 uint64_t size, zio_done_func_t *done, void *private,
1004 zio_priority_t priority, enum zio_flag flags)
1009 if (vd->vdev_children == 0) {
1010 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1011 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1012 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1016 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1018 for (c = 0; c < vd->vdev_children; c++)
1019 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1020 offset, size, done, private, priority, flags));
1027 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1028 void *data, int checksum, zio_done_func_t *done, void *private,
1029 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1033 ASSERT(vd->vdev_children == 0);
1034 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1035 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1036 ASSERT3U(offset + size, <=, vd->vdev_psize);
1038 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1039 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1040 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1042 zio->io_prop.zp_checksum = checksum;
1048 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1049 void *data, int checksum, zio_done_func_t *done, void *private,
1050 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1054 ASSERT(vd->vdev_children == 0);
1055 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1056 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1057 ASSERT3U(offset + size, <=, vd->vdev_psize);
1059 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1060 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1061 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1063 zio->io_prop.zp_checksum = checksum;
1065 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1067 * zec checksums are necessarily destructive -- they modify
1068 * the end of the write buffer to hold the verifier/checksum.
1069 * Therefore, we must make a local copy in case the data is
1070 * being written to multiple places in parallel.
1072 void *wbuf = zio_buf_alloc(size);
1073 bcopy(data, wbuf, size);
1074 zio_push_transform(zio, wbuf, size, size, NULL);
1081 * Create a child I/O to do some work for us.
1084 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1085 void *data, uint64_t size, int type, zio_priority_t priority,
1086 enum zio_flag flags, zio_done_func_t *done, void *private)
1088 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1091 ASSERT(vd->vdev_parent ==
1092 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1094 if (type == ZIO_TYPE_READ && bp != NULL) {
1096 * If we have the bp, then the child should perform the
1097 * checksum and the parent need not. This pushes error
1098 * detection as close to the leaves as possible and
1099 * eliminates redundant checksums in the interior nodes.
1101 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1102 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1105 /* Not all IO types require vdev io done stage e.g. free */
1106 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1107 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1109 if (vd->vdev_children == 0)
1110 offset += VDEV_LABEL_START_SIZE;
1112 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1115 * If we've decided to do a repair, the write is not speculative --
1116 * even if the original read was.
1118 if (flags & ZIO_FLAG_IO_REPAIR)
1119 flags &= ~ZIO_FLAG_SPECULATIVE;
1122 * If we're creating a child I/O that is not associated with a
1123 * top-level vdev, then the child zio is not an allocating I/O.
1124 * If this is a retried I/O then we ignore it since we will
1125 * have already processed the original allocating I/O.
1127 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1128 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1129 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1131 ASSERT(mc->mc_alloc_throttle_enabled);
1132 ASSERT(type == ZIO_TYPE_WRITE);
1133 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1134 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1135 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1136 pio->io_child_type == ZIO_CHILD_GANG);
1138 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1141 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1142 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1143 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1144 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1146 zio->io_physdone = pio->io_physdone;
1147 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1148 zio->io_logical->io_phys_children++;
1154 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1155 int type, zio_priority_t priority, enum zio_flag flags,
1156 zio_done_func_t *done, void *private)
1160 ASSERT(vd->vdev_ops->vdev_op_leaf);
1162 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1163 data, size, size, done, private, type, priority,
1164 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1166 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1172 zio_flush(zio_t *zio, vdev_t *vd)
1174 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1175 NULL, NULL, ZIO_PRIORITY_NOW,
1176 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1180 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1183 ASSERT(vd->vdev_ops->vdev_op_leaf);
1185 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1186 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1187 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1188 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1192 zio_shrink(zio_t *zio, uint64_t size)
1194 ASSERT(zio->io_executor == NULL);
1195 ASSERT(zio->io_orig_size == zio->io_size);
1196 ASSERT(size <= zio->io_size);
1199 * We don't shrink for raidz because of problems with the
1200 * reconstruction when reading back less than the block size.
1201 * Note, BP_IS_RAIDZ() assumes no compression.
1203 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1204 if (!BP_IS_RAIDZ(zio->io_bp)) {
1205 /* we are not doing a raw write */
1206 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1207 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1212 * ==========================================================================
1213 * Prepare to read and write logical blocks
1214 * ==========================================================================
1218 zio_read_bp_init(zio_t *zio)
1220 blkptr_t *bp = zio->io_bp;
1222 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1223 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1224 !(zio->io_flags & ZIO_FLAG_RAW)) {
1226 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1227 void *cbuf = zio_buf_alloc(psize);
1229 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1232 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1233 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1234 decode_embedded_bp_compressed(bp, zio->io_data);
1236 ASSERT(!BP_IS_EMBEDDED(bp));
1239 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1240 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1242 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1243 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1245 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1246 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1248 return (ZIO_PIPELINE_CONTINUE);
1252 zio_write_bp_init(zio_t *zio)
1254 if (!IO_IS_ALLOCATING(zio))
1255 return (ZIO_PIPELINE_CONTINUE);
1257 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1259 if (zio->io_bp_override) {
1260 blkptr_t *bp = zio->io_bp;
1261 zio_prop_t *zp = &zio->io_prop;
1263 ASSERT(bp->blk_birth != zio->io_txg);
1264 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1266 *bp = *zio->io_bp_override;
1267 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1269 if (BP_IS_EMBEDDED(bp))
1270 return (ZIO_PIPELINE_CONTINUE);
1273 * If we've been overridden and nopwrite is set then
1274 * set the flag accordingly to indicate that a nopwrite
1275 * has already occurred.
1277 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1278 ASSERT(!zp->zp_dedup);
1279 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1280 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1281 return (ZIO_PIPELINE_CONTINUE);
1284 ASSERT(!zp->zp_nopwrite);
1286 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1287 return (ZIO_PIPELINE_CONTINUE);
1289 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1290 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1292 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1293 BP_SET_DEDUP(bp, 1);
1294 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1295 return (ZIO_PIPELINE_CONTINUE);
1299 * We were unable to handle this as an override bp, treat
1300 * it as a regular write I/O.
1302 zio->io_bp_override = NULL;
1303 *bp = zio->io_bp_orig;
1304 zio->io_pipeline = zio->io_orig_pipeline;
1307 return (ZIO_PIPELINE_CONTINUE);
1311 zio_write_compress(zio_t *zio)
1313 spa_t *spa = zio->io_spa;
1314 zio_prop_t *zp = &zio->io_prop;
1315 enum zio_compress compress = zp->zp_compress;
1316 blkptr_t *bp = zio->io_bp;
1317 uint64_t lsize = zio->io_lsize;
1318 uint64_t psize = zio->io_size;
1321 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1324 * If our children haven't all reached the ready stage,
1325 * wait for them and then repeat this pipeline stage.
1327 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1328 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1329 return (ZIO_PIPELINE_STOP);
1331 if (!IO_IS_ALLOCATING(zio))
1332 return (ZIO_PIPELINE_CONTINUE);
1334 if (zio->io_children_ready != NULL) {
1336 * Now that all our children are ready, run the callback
1337 * associated with this zio in case it wants to modify the
1338 * data to be written.
1340 ASSERT3U(zp->zp_level, >, 0);
1341 zio->io_children_ready(zio);
1344 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1345 ASSERT(zio->io_bp_override == NULL);
1347 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1349 * We're rewriting an existing block, which means we're
1350 * working on behalf of spa_sync(). For spa_sync() to
1351 * converge, it must eventually be the case that we don't
1352 * have to allocate new blocks. But compression changes
1353 * the blocksize, which forces a reallocate, and makes
1354 * convergence take longer. Therefore, after the first
1355 * few passes, stop compressing to ensure convergence.
1357 pass = spa_sync_pass(spa);
1359 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1360 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1361 ASSERT(!BP_GET_DEDUP(bp));
1363 if (pass >= zfs_sync_pass_dont_compress)
1364 compress = ZIO_COMPRESS_OFF;
1366 /* Make sure someone doesn't change their mind on overwrites */
1367 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1368 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1371 /* If it's a compressed write that is not raw, compress the buffer. */
1372 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1373 void *cbuf = zio_buf_alloc(lsize);
1374 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1375 if (psize == 0 || psize == lsize) {
1376 compress = ZIO_COMPRESS_OFF;
1377 zio_buf_free(cbuf, lsize);
1378 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1379 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1380 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1381 encode_embedded_bp_compressed(bp,
1382 cbuf, compress, lsize, psize);
1383 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1384 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1385 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1386 zio_buf_free(cbuf, lsize);
1387 bp->blk_birth = zio->io_txg;
1388 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1389 ASSERT(spa_feature_is_active(spa,
1390 SPA_FEATURE_EMBEDDED_DATA));
1391 return (ZIO_PIPELINE_CONTINUE);
1394 * Round up compressed size up to the ashift
1395 * of the smallest-ashift device, and zero the tail.
1396 * This ensures that the compressed size of the BP
1397 * (and thus compressratio property) are correct,
1398 * in that we charge for the padding used to fill out
1401 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1402 size_t rounded = (size_t)P2ROUNDUP(psize,
1403 1ULL << spa->spa_min_ashift);
1404 if (rounded >= lsize) {
1405 compress = ZIO_COMPRESS_OFF;
1406 zio_buf_free(cbuf, lsize);
1409 bzero((char *)cbuf + psize, rounded - psize);
1411 zio_push_transform(zio, cbuf,
1412 psize, lsize, NULL);
1417 * We were unable to handle this as an override bp, treat
1418 * it as a regular write I/O.
1420 zio->io_bp_override = NULL;
1421 *bp = zio->io_bp_orig;
1422 zio->io_pipeline = zio->io_orig_pipeline;
1424 ASSERT3U(psize, !=, 0);
1428 * The final pass of spa_sync() must be all rewrites, but the first
1429 * few passes offer a trade-off: allocating blocks defers convergence,
1430 * but newly allocated blocks are sequential, so they can be written
1431 * to disk faster. Therefore, we allow the first few passes of
1432 * spa_sync() to allocate new blocks, but force rewrites after that.
1433 * There should only be a handful of blocks after pass 1 in any case.
1435 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1436 BP_GET_PSIZE(bp) == psize &&
1437 pass >= zfs_sync_pass_rewrite) {
1439 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1440 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1441 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1444 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1448 if (zio->io_bp_orig.blk_birth != 0 &&
1449 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1450 BP_SET_LSIZE(bp, lsize);
1451 BP_SET_TYPE(bp, zp->zp_type);
1452 BP_SET_LEVEL(bp, zp->zp_level);
1453 BP_SET_BIRTH(bp, zio->io_txg, 0);
1455 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1457 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1458 BP_SET_LSIZE(bp, lsize);
1459 BP_SET_TYPE(bp, zp->zp_type);
1460 BP_SET_LEVEL(bp, zp->zp_level);
1461 BP_SET_PSIZE(bp, psize);
1462 BP_SET_COMPRESS(bp, compress);
1463 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1464 BP_SET_DEDUP(bp, zp->zp_dedup);
1465 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1467 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1468 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1469 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1471 if (zp->zp_nopwrite) {
1472 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1473 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1474 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1477 return (ZIO_PIPELINE_CONTINUE);
1481 zio_free_bp_init(zio_t *zio)
1483 blkptr_t *bp = zio->io_bp;
1485 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1486 if (BP_GET_DEDUP(bp))
1487 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1490 return (ZIO_PIPELINE_CONTINUE);
1494 * ==========================================================================
1495 * Execute the I/O pipeline
1496 * ==========================================================================
1500 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1502 spa_t *spa = zio->io_spa;
1503 zio_type_t t = zio->io_type;
1504 int flags = (cutinline ? TQ_FRONT : 0);
1506 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1509 * If we're a config writer or a probe, the normal issue and
1510 * interrupt threads may all be blocked waiting for the config lock.
1511 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1513 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1517 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1519 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1523 * If this is a high priority I/O, then use the high priority taskq if
1526 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1527 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1530 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1533 * NB: We are assuming that the zio can only be dispatched
1534 * to a single taskq at a time. It would be a grievous error
1535 * to dispatch the zio to another taskq at the same time.
1537 #if defined(illumos) || !defined(_KERNEL)
1538 ASSERT(zio->io_tqent.tqent_next == NULL);
1540 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1542 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1543 flags, &zio->io_tqent);
1547 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1549 kthread_t *executor = zio->io_executor;
1550 spa_t *spa = zio->io_spa;
1552 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1553 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1555 for (i = 0; i < tqs->stqs_count; i++) {
1556 if (taskq_member(tqs->stqs_taskq[i], executor))
1565 zio_issue_async(zio_t *zio)
1567 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1569 return (ZIO_PIPELINE_STOP);
1573 zio_interrupt(zio_t *zio)
1575 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1579 zio_delay_interrupt(zio_t *zio)
1582 * The timeout_generic() function isn't defined in userspace, so
1583 * rather than trying to implement the function, the zio delay
1584 * functionality has been disabled for userspace builds.
1589 * If io_target_timestamp is zero, then no delay has been registered
1590 * for this IO, thus jump to the end of this function and "skip" the
1591 * delay; issuing it directly to the zio layer.
1593 if (zio->io_target_timestamp != 0) {
1594 hrtime_t now = gethrtime();
1596 if (now >= zio->io_target_timestamp) {
1598 * This IO has already taken longer than the target
1599 * delay to complete, so we don't want to delay it
1600 * any longer; we "miss" the delay and issue it
1601 * directly to the zio layer. This is likely due to
1602 * the target latency being set to a value less than
1603 * the underlying hardware can satisfy (e.g. delay
1604 * set to 1ms, but the disks take 10ms to complete an
1608 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1613 hrtime_t diff = zio->io_target_timestamp - now;
1615 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1616 hrtime_t, now, hrtime_t, diff);
1618 (void) timeout_generic(CALLOUT_NORMAL,
1619 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1626 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1631 * Execute the I/O pipeline until one of the following occurs:
1633 * (1) the I/O completes
1634 * (2) the pipeline stalls waiting for dependent child I/Os
1635 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1636 * (4) the I/O is delegated by vdev-level caching or aggregation
1637 * (5) the I/O is deferred due to vdev-level queueing
1638 * (6) the I/O is handed off to another thread.
1640 * In all cases, the pipeline stops whenever there's no CPU work; it never
1641 * burns a thread in cv_wait().
1643 * There's no locking on io_stage because there's no legitimate way
1644 * for multiple threads to be attempting to process the same I/O.
1646 static zio_pipe_stage_t *zio_pipeline[];
1649 zio_execute(zio_t *zio)
1651 zio->io_executor = curthread;
1653 ASSERT3U(zio->io_queued_timestamp, >, 0);
1655 while (zio->io_stage < ZIO_STAGE_DONE) {
1656 enum zio_stage pipeline = zio->io_pipeline;
1657 enum zio_stage stage = zio->io_stage;
1660 ASSERT(!MUTEX_HELD(&zio->io_lock));
1661 ASSERT(ISP2(stage));
1662 ASSERT(zio->io_stall == NULL);
1666 } while ((stage & pipeline) == 0);
1668 ASSERT(stage <= ZIO_STAGE_DONE);
1671 * If we are in interrupt context and this pipeline stage
1672 * will grab a config lock that is held across I/O,
1673 * or may wait for an I/O that needs an interrupt thread
1674 * to complete, issue async to avoid deadlock.
1676 * For VDEV_IO_START, we cut in line so that the io will
1677 * be sent to disk promptly.
1679 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1680 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1681 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1682 zio_requeue_io_start_cut_in_line : B_FALSE;
1683 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1687 zio->io_stage = stage;
1688 zio->io_pipeline_trace |= zio->io_stage;
1689 rv = zio_pipeline[highbit64(stage) - 1](zio);
1691 if (rv == ZIO_PIPELINE_STOP)
1694 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1699 * ==========================================================================
1700 * Initiate I/O, either sync or async
1701 * ==========================================================================
1704 zio_wait(zio_t *zio)
1708 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1709 ASSERT(zio->io_executor == NULL);
1711 zio->io_waiter = curthread;
1712 ASSERT0(zio->io_queued_timestamp);
1713 zio->io_queued_timestamp = gethrtime();
1717 mutex_enter(&zio->io_lock);
1718 while (zio->io_executor != NULL)
1719 cv_wait(&zio->io_cv, &zio->io_lock);
1720 mutex_exit(&zio->io_lock);
1722 error = zio->io_error;
1729 zio_nowait(zio_t *zio)
1731 ASSERT(zio->io_executor == NULL);
1733 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1734 zio_unique_parent(zio) == NULL) {
1736 * This is a logical async I/O with no parent to wait for it.
1737 * We add it to the spa_async_root_zio "Godfather" I/O which
1738 * will ensure they complete prior to unloading the pool.
1740 spa_t *spa = zio->io_spa;
1742 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1745 ASSERT0(zio->io_queued_timestamp);
1746 zio->io_queued_timestamp = gethrtime();
1751 * ==========================================================================
1752 * Reexecute or suspend/resume failed I/O
1753 * ==========================================================================
1757 zio_reexecute(zio_t *pio)
1759 zio_t *cio, *cio_next;
1761 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1762 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1763 ASSERT(pio->io_gang_leader == NULL);
1764 ASSERT(pio->io_gang_tree == NULL);
1766 pio->io_flags = pio->io_orig_flags;
1767 pio->io_stage = pio->io_orig_stage;
1768 pio->io_pipeline = pio->io_orig_pipeline;
1769 pio->io_reexecute = 0;
1770 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1771 pio->io_pipeline_trace = 0;
1773 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1774 pio->io_state[w] = 0;
1775 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1776 pio->io_child_error[c] = 0;
1778 if (IO_IS_ALLOCATING(pio))
1779 BP_ZERO(pio->io_bp);
1782 * As we reexecute pio's children, new children could be created.
1783 * New children go to the head of pio's io_child_list, however,
1784 * so we will (correctly) not reexecute them. The key is that
1785 * the remainder of pio's io_child_list, from 'cio_next' onward,
1786 * cannot be affected by any side effects of reexecuting 'cio'.
1788 zio_link_t *zl = NULL;
1789 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1790 cio_next = zio_walk_children(pio, &zl);
1791 mutex_enter(&pio->io_lock);
1792 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1793 pio->io_children[cio->io_child_type][w]++;
1794 mutex_exit(&pio->io_lock);
1799 * Now that all children have been reexecuted, execute the parent.
1800 * We don't reexecute "The Godfather" I/O here as it's the
1801 * responsibility of the caller to wait on him.
1803 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1804 pio->io_queued_timestamp = gethrtime();
1810 zio_suspend(spa_t *spa, zio_t *zio)
1812 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1813 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1814 "failure and the failure mode property for this pool "
1815 "is set to panic.", spa_name(spa));
1817 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1819 mutex_enter(&spa->spa_suspend_lock);
1821 if (spa->spa_suspend_zio_root == NULL)
1822 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1823 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1824 ZIO_FLAG_GODFATHER);
1826 spa->spa_suspended = B_TRUE;
1829 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1830 ASSERT(zio != spa->spa_suspend_zio_root);
1831 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1832 ASSERT(zio_unique_parent(zio) == NULL);
1833 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1834 zio_add_child(spa->spa_suspend_zio_root, zio);
1837 mutex_exit(&spa->spa_suspend_lock);
1841 zio_resume(spa_t *spa)
1846 * Reexecute all previously suspended i/o.
1848 mutex_enter(&spa->spa_suspend_lock);
1849 spa->spa_suspended = B_FALSE;
1850 cv_broadcast(&spa->spa_suspend_cv);
1851 pio = spa->spa_suspend_zio_root;
1852 spa->spa_suspend_zio_root = NULL;
1853 mutex_exit(&spa->spa_suspend_lock);
1859 return (zio_wait(pio));
1863 zio_resume_wait(spa_t *spa)
1865 mutex_enter(&spa->spa_suspend_lock);
1866 while (spa_suspended(spa))
1867 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1868 mutex_exit(&spa->spa_suspend_lock);
1872 * ==========================================================================
1875 * A gang block is a collection of small blocks that looks to the DMU
1876 * like one large block. When zio_dva_allocate() cannot find a block
1877 * of the requested size, due to either severe fragmentation or the pool
1878 * being nearly full, it calls zio_write_gang_block() to construct the
1879 * block from smaller fragments.
1881 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1882 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1883 * an indirect block: it's an array of block pointers. It consumes
1884 * only one sector and hence is allocatable regardless of fragmentation.
1885 * The gang header's bps point to its gang members, which hold the data.
1887 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1888 * as the verifier to ensure uniqueness of the SHA256 checksum.
1889 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1890 * not the gang header. This ensures that data block signatures (needed for
1891 * deduplication) are independent of how the block is physically stored.
1893 * Gang blocks can be nested: a gang member may itself be a gang block.
1894 * Thus every gang block is a tree in which root and all interior nodes are
1895 * gang headers, and the leaves are normal blocks that contain user data.
1896 * The root of the gang tree is called the gang leader.
1898 * To perform any operation (read, rewrite, free, claim) on a gang block,
1899 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1900 * in the io_gang_tree field of the original logical i/o by recursively
1901 * reading the gang leader and all gang headers below it. This yields
1902 * an in-core tree containing the contents of every gang header and the
1903 * bps for every constituent of the gang block.
1905 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1906 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1907 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1908 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1909 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1910 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1911 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1912 * of the gang header plus zio_checksum_compute() of the data to update the
1913 * gang header's blk_cksum as described above.
1915 * The two-phase assemble/issue model solves the problem of partial failure --
1916 * what if you'd freed part of a gang block but then couldn't read the
1917 * gang header for another part? Assembling the entire gang tree first
1918 * ensures that all the necessary gang header I/O has succeeded before
1919 * starting the actual work of free, claim, or write. Once the gang tree
1920 * is assembled, free and claim are in-memory operations that cannot fail.
1922 * In the event that a gang write fails, zio_dva_unallocate() walks the
1923 * gang tree to immediately free (i.e. insert back into the space map)
1924 * everything we've allocated. This ensures that we don't get ENOSPC
1925 * errors during repeated suspend/resume cycles due to a flaky device.
1927 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1928 * the gang tree, we won't modify the block, so we can safely defer the free
1929 * (knowing that the block is still intact). If we *can* assemble the gang
1930 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1931 * each constituent bp and we can allocate a new block on the next sync pass.
1933 * In all cases, the gang tree allows complete recovery from partial failure.
1934 * ==========================================================================
1938 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1943 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1944 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1945 &pio->io_bookmark));
1949 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1954 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1955 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1956 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1958 * As we rewrite each gang header, the pipeline will compute
1959 * a new gang block header checksum for it; but no one will
1960 * compute a new data checksum, so we do that here. The one
1961 * exception is the gang leader: the pipeline already computed
1962 * its data checksum because that stage precedes gang assembly.
1963 * (Presently, nothing actually uses interior data checksums;
1964 * this is just good hygiene.)
1966 if (gn != pio->io_gang_leader->io_gang_tree) {
1967 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1968 data, BP_GET_PSIZE(bp));
1971 * If we are here to damage data for testing purposes,
1972 * leave the GBH alone so that we can detect the damage.
1974 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1975 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1977 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1978 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1979 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1987 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1989 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1990 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1991 ZIO_GANG_CHILD_FLAGS(pio)));
1996 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1998 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1999 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2002 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2011 static void zio_gang_tree_assemble_done(zio_t *zio);
2013 static zio_gang_node_t *
2014 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2016 zio_gang_node_t *gn;
2018 ASSERT(*gnpp == NULL);
2020 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2021 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2028 zio_gang_node_free(zio_gang_node_t **gnpp)
2030 zio_gang_node_t *gn = *gnpp;
2032 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2033 ASSERT(gn->gn_child[g] == NULL);
2035 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2036 kmem_free(gn, sizeof (*gn));
2041 zio_gang_tree_free(zio_gang_node_t **gnpp)
2043 zio_gang_node_t *gn = *gnpp;
2048 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2049 zio_gang_tree_free(&gn->gn_child[g]);
2051 zio_gang_node_free(gnpp);
2055 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2057 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2059 ASSERT(gio->io_gang_leader == gio);
2060 ASSERT(BP_IS_GANG(bp));
2062 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
2063 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2064 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2068 zio_gang_tree_assemble_done(zio_t *zio)
2070 zio_t *gio = zio->io_gang_leader;
2071 zio_gang_node_t *gn = zio->io_private;
2072 blkptr_t *bp = zio->io_bp;
2074 ASSERT(gio == zio_unique_parent(zio));
2075 ASSERT(zio->io_child_count == 0);
2080 if (BP_SHOULD_BYTESWAP(bp))
2081 byteswap_uint64_array(zio->io_data, zio->io_size);
2083 ASSERT(zio->io_data == gn->gn_gbh);
2084 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2085 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2087 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2088 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2089 if (!BP_IS_GANG(gbp))
2091 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2096 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2098 zio_t *gio = pio->io_gang_leader;
2101 ASSERT(BP_IS_GANG(bp) == !!gn);
2102 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2103 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2106 * If you're a gang header, your data is in gn->gn_gbh.
2107 * If you're a gang member, your data is in 'data' and gn == NULL.
2109 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2112 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2114 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2115 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2116 if (BP_IS_HOLE(gbp))
2118 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
2119 data = (char *)data + BP_GET_PSIZE(gbp);
2123 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2124 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2131 zio_gang_assemble(zio_t *zio)
2133 blkptr_t *bp = zio->io_bp;
2135 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2136 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2138 zio->io_gang_leader = zio;
2140 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2142 return (ZIO_PIPELINE_CONTINUE);
2146 zio_gang_issue(zio_t *zio)
2148 blkptr_t *bp = zio->io_bp;
2150 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2151 return (ZIO_PIPELINE_STOP);
2153 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2154 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2156 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2157 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2159 zio_gang_tree_free(&zio->io_gang_tree);
2161 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2163 return (ZIO_PIPELINE_CONTINUE);
2167 zio_write_gang_member_ready(zio_t *zio)
2169 zio_t *pio = zio_unique_parent(zio);
2170 zio_t *gio = zio->io_gang_leader;
2171 dva_t *cdva = zio->io_bp->blk_dva;
2172 dva_t *pdva = pio->io_bp->blk_dva;
2175 if (BP_IS_HOLE(zio->io_bp))
2178 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2180 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2181 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2182 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2183 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2184 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2186 mutex_enter(&pio->io_lock);
2187 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2188 ASSERT(DVA_GET_GANG(&pdva[d]));
2189 asize = DVA_GET_ASIZE(&pdva[d]);
2190 asize += DVA_GET_ASIZE(&cdva[d]);
2191 DVA_SET_ASIZE(&pdva[d], asize);
2193 mutex_exit(&pio->io_lock);
2197 zio_write_gang_block(zio_t *pio)
2199 spa_t *spa = pio->io_spa;
2200 metaslab_class_t *mc = spa_normal_class(spa);
2201 blkptr_t *bp = pio->io_bp;
2202 zio_t *gio = pio->io_gang_leader;
2204 zio_gang_node_t *gn, **gnpp;
2205 zio_gbh_phys_t *gbh;
2206 uint64_t txg = pio->io_txg;
2207 uint64_t resid = pio->io_size;
2209 int copies = gio->io_prop.zp_copies;
2210 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2214 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2215 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2216 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2217 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2219 flags |= METASLAB_ASYNC_ALLOC;
2220 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2223 * The logical zio has already placed a reservation for
2224 * 'copies' allocation slots but gang blocks may require
2225 * additional copies. These additional copies
2226 * (i.e. gbh_copies - copies) are guaranteed to succeed
2227 * since metaslab_class_throttle_reserve() always allows
2228 * additional reservations for gang blocks.
2230 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2234 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2235 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2236 &pio->io_alloc_list, pio);
2238 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2239 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2240 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2243 * If we failed to allocate the gang block header then
2244 * we remove any additional allocation reservations that
2245 * we placed here. The original reservation will
2246 * be removed when the logical I/O goes to the ready
2249 metaslab_class_throttle_unreserve(mc,
2250 gbh_copies - copies, pio);
2252 pio->io_error = error;
2253 return (ZIO_PIPELINE_CONTINUE);
2257 gnpp = &gio->io_gang_tree;
2259 gnpp = pio->io_private;
2260 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2263 gn = zio_gang_node_alloc(gnpp);
2265 bzero(gbh, SPA_GANGBLOCKSIZE);
2268 * Create the gang header.
2270 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2271 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2274 * Create and nowait the gang children.
2276 for (int g = 0; resid != 0; resid -= lsize, g++) {
2277 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2279 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2281 zp.zp_checksum = gio->io_prop.zp_checksum;
2282 zp.zp_compress = ZIO_COMPRESS_OFF;
2283 zp.zp_type = DMU_OT_NONE;
2285 zp.zp_copies = gio->io_prop.zp_copies;
2286 zp.zp_dedup = B_FALSE;
2287 zp.zp_dedup_verify = B_FALSE;
2288 zp.zp_nopwrite = B_FALSE;
2290 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2291 (char *)pio->io_data + (pio->io_size - resid), lsize, lsize,
2292 &zp, zio_write_gang_member_ready, NULL, NULL, NULL,
2293 &gn->gn_child[g], pio->io_priority,
2294 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2296 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2297 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2298 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2301 * Gang children won't throttle but we should
2302 * account for their work, so reserve an allocation
2303 * slot for them here.
2305 VERIFY(metaslab_class_throttle_reserve(mc,
2306 zp.zp_copies, cio, flags));
2312 * Set pio's pipeline to just wait for zio to finish.
2314 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2318 return (ZIO_PIPELINE_CONTINUE);
2322 * The zio_nop_write stage in the pipeline determines if allocating a
2323 * new bp is necessary. The nopwrite feature can handle writes in
2324 * either syncing or open context (i.e. zil writes) and as a result is
2325 * mutually exclusive with dedup.
2327 * By leveraging a cryptographically secure checksum, such as SHA256, we
2328 * can compare the checksums of the new data and the old to determine if
2329 * allocating a new block is required. Note that our requirements for
2330 * cryptographic strength are fairly weak: there can't be any accidental
2331 * hash collisions, but we don't need to be secure against intentional
2332 * (malicious) collisions. To trigger a nopwrite, you have to be able
2333 * to write the file to begin with, and triggering an incorrect (hash
2334 * collision) nopwrite is no worse than simply writing to the file.
2335 * That said, there are no known attacks against the checksum algorithms
2336 * used for nopwrite, assuming that the salt and the checksums
2337 * themselves remain secret.
2340 zio_nop_write(zio_t *zio)
2342 blkptr_t *bp = zio->io_bp;
2343 blkptr_t *bp_orig = &zio->io_bp_orig;
2344 zio_prop_t *zp = &zio->io_prop;
2346 ASSERT(BP_GET_LEVEL(bp) == 0);
2347 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2348 ASSERT(zp->zp_nopwrite);
2349 ASSERT(!zp->zp_dedup);
2350 ASSERT(zio->io_bp_override == NULL);
2351 ASSERT(IO_IS_ALLOCATING(zio));
2354 * Check to see if the original bp and the new bp have matching
2355 * characteristics (i.e. same checksum, compression algorithms, etc).
2356 * If they don't then just continue with the pipeline which will
2357 * allocate a new bp.
2359 if (BP_IS_HOLE(bp_orig) ||
2360 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2361 ZCHECKSUM_FLAG_NOPWRITE) ||
2362 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2363 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2364 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2365 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2366 return (ZIO_PIPELINE_CONTINUE);
2369 * If the checksums match then reset the pipeline so that we
2370 * avoid allocating a new bp and issuing any I/O.
2372 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2373 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2374 ZCHECKSUM_FLAG_NOPWRITE);
2375 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2376 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2377 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2378 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2379 sizeof (uint64_t)) == 0);
2382 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2383 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2386 return (ZIO_PIPELINE_CONTINUE);
2390 * ==========================================================================
2392 * ==========================================================================
2395 zio_ddt_child_read_done(zio_t *zio)
2397 blkptr_t *bp = zio->io_bp;
2398 ddt_entry_t *dde = zio->io_private;
2400 zio_t *pio = zio_unique_parent(zio);
2402 mutex_enter(&pio->io_lock);
2403 ddp = ddt_phys_select(dde, bp);
2404 if (zio->io_error == 0)
2405 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2406 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2407 dde->dde_repair_data = zio->io_data;
2409 zio_buf_free(zio->io_data, zio->io_size);
2410 mutex_exit(&pio->io_lock);
2414 zio_ddt_read_start(zio_t *zio)
2416 blkptr_t *bp = zio->io_bp;
2418 ASSERT(BP_GET_DEDUP(bp));
2419 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2420 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2422 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2423 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2424 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2425 ddt_phys_t *ddp = dde->dde_phys;
2426 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2429 ASSERT(zio->io_vsd == NULL);
2432 if (ddp_self == NULL)
2433 return (ZIO_PIPELINE_CONTINUE);
2435 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2436 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2438 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2440 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2441 zio_buf_alloc(zio->io_size), zio->io_size,
2442 zio_ddt_child_read_done, dde, zio->io_priority,
2443 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2444 &zio->io_bookmark));
2446 return (ZIO_PIPELINE_CONTINUE);
2449 zio_nowait(zio_read(zio, zio->io_spa, bp,
2450 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2451 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2453 return (ZIO_PIPELINE_CONTINUE);
2457 zio_ddt_read_done(zio_t *zio)
2459 blkptr_t *bp = zio->io_bp;
2461 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2462 return (ZIO_PIPELINE_STOP);
2464 ASSERT(BP_GET_DEDUP(bp));
2465 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2466 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2468 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2469 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2470 ddt_entry_t *dde = zio->io_vsd;
2472 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2473 return (ZIO_PIPELINE_CONTINUE);
2476 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2477 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2478 return (ZIO_PIPELINE_STOP);
2480 if (dde->dde_repair_data != NULL) {
2481 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2482 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2484 ddt_repair_done(ddt, dde);
2488 ASSERT(zio->io_vsd == NULL);
2490 return (ZIO_PIPELINE_CONTINUE);
2494 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2496 spa_t *spa = zio->io_spa;
2497 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2499 /* We should never get a raw, override zio */
2500 ASSERT(!(zio->io_bp_override && do_raw));
2503 * Note: we compare the original data, not the transformed data,
2504 * because when zio->io_bp is an override bp, we will not have
2505 * pushed the I/O transforms. That's an important optimization
2506 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2508 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2509 zio_t *lio = dde->dde_lead_zio[p];
2512 return (lio->io_orig_size != zio->io_orig_size ||
2513 bcmp(zio->io_orig_data, lio->io_orig_data,
2514 zio->io_orig_size) != 0);
2518 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2519 ddt_phys_t *ddp = &dde->dde_phys[p];
2521 if (ddp->ddp_phys_birth != 0) {
2522 arc_buf_t *abuf = NULL;
2523 arc_flags_t aflags = ARC_FLAG_WAIT;
2524 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2525 blkptr_t blk = *zio->io_bp;
2528 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2533 * Intuitively, it would make more sense to compare
2534 * io_data than io_orig_data in the raw case since you
2535 * don't want to look at any transformations that have
2536 * happened to the data. However, for raw I/Os the
2537 * data will actually be the same in io_data and
2538 * io_orig_data, so all we have to do is issue this as
2542 zio_flags |= ZIO_FLAG_RAW;
2543 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2544 ASSERT0(bcmp(zio->io_data, zio->io_orig_data,
2546 ASSERT3P(zio->io_transform_stack, ==, NULL);
2549 error = arc_read(NULL, spa, &blk,
2550 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2551 zio_flags, &aflags, &zio->io_bookmark);
2554 if (arc_buf_size(abuf) != zio->io_orig_size ||
2555 bcmp(abuf->b_data, zio->io_orig_data,
2556 zio->io_orig_size) != 0)
2557 error = SET_ERROR(EEXIST);
2558 arc_buf_destroy(abuf, &abuf);
2562 return (error != 0);
2570 zio_ddt_child_write_ready(zio_t *zio)
2572 int p = zio->io_prop.zp_copies;
2573 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2574 ddt_entry_t *dde = zio->io_private;
2575 ddt_phys_t *ddp = &dde->dde_phys[p];
2583 ASSERT(dde->dde_lead_zio[p] == zio);
2585 ddt_phys_fill(ddp, zio->io_bp);
2587 zio_link_t *zl = NULL;
2588 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2589 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2595 zio_ddt_child_write_done(zio_t *zio)
2597 int p = zio->io_prop.zp_copies;
2598 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2599 ddt_entry_t *dde = zio->io_private;
2600 ddt_phys_t *ddp = &dde->dde_phys[p];
2604 ASSERT(ddp->ddp_refcnt == 0);
2605 ASSERT(dde->dde_lead_zio[p] == zio);
2606 dde->dde_lead_zio[p] = NULL;
2608 if (zio->io_error == 0) {
2609 zio_link_t *zl = NULL;
2610 while (zio_walk_parents(zio, &zl) != NULL)
2611 ddt_phys_addref(ddp);
2613 ddt_phys_clear(ddp);
2620 zio_ddt_ditto_write_done(zio_t *zio)
2622 int p = DDT_PHYS_DITTO;
2623 zio_prop_t *zp = &zio->io_prop;
2624 blkptr_t *bp = zio->io_bp;
2625 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2626 ddt_entry_t *dde = zio->io_private;
2627 ddt_phys_t *ddp = &dde->dde_phys[p];
2628 ddt_key_t *ddk = &dde->dde_key;
2632 ASSERT(ddp->ddp_refcnt == 0);
2633 ASSERT(dde->dde_lead_zio[p] == zio);
2634 dde->dde_lead_zio[p] = NULL;
2636 if (zio->io_error == 0) {
2637 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2638 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2639 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2640 if (ddp->ddp_phys_birth != 0)
2641 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2642 ddt_phys_fill(ddp, bp);
2649 zio_ddt_write(zio_t *zio)
2651 spa_t *spa = zio->io_spa;
2652 blkptr_t *bp = zio->io_bp;
2653 uint64_t txg = zio->io_txg;
2654 zio_prop_t *zp = &zio->io_prop;
2655 int p = zp->zp_copies;
2659 ddt_t *ddt = ddt_select(spa, bp);
2663 ASSERT(BP_GET_DEDUP(bp));
2664 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2665 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2666 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2669 dde = ddt_lookup(ddt, bp, B_TRUE);
2670 ddp = &dde->dde_phys[p];
2672 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2674 * If we're using a weak checksum, upgrade to a strong checksum
2675 * and try again. If we're already using a strong checksum,
2676 * we can't resolve it, so just convert to an ordinary write.
2677 * (And automatically e-mail a paper to Nature?)
2679 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2680 ZCHECKSUM_FLAG_DEDUP)) {
2681 zp->zp_checksum = spa_dedup_checksum(spa);
2682 zio_pop_transforms(zio);
2683 zio->io_stage = ZIO_STAGE_OPEN;
2686 zp->zp_dedup = B_FALSE;
2687 BP_SET_DEDUP(bp, B_FALSE);
2689 ASSERT(!BP_GET_DEDUP(bp));
2690 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2692 return (ZIO_PIPELINE_CONTINUE);
2695 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2696 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2698 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2699 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2700 zio_prop_t czp = *zp;
2702 czp.zp_copies = ditto_copies;
2705 * If we arrived here with an override bp, we won't have run
2706 * the transform stack, so we won't have the data we need to
2707 * generate a child i/o. So, toss the override bp and restart.
2708 * This is safe, because using the override bp is just an
2709 * optimization; and it's rare, so the cost doesn't matter.
2711 if (zio->io_bp_override) {
2712 zio_pop_transforms(zio);
2713 zio->io_stage = ZIO_STAGE_OPEN;
2714 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2715 zio->io_bp_override = NULL;
2718 return (ZIO_PIPELINE_CONTINUE);
2721 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2722 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2723 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2724 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2726 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2727 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2730 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2731 if (ddp->ddp_phys_birth != 0)
2732 ddt_bp_fill(ddp, bp, txg);
2733 if (dde->dde_lead_zio[p] != NULL)
2734 zio_add_child(zio, dde->dde_lead_zio[p]);
2736 ddt_phys_addref(ddp);
2737 } else if (zio->io_bp_override) {
2738 ASSERT(bp->blk_birth == txg);
2739 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2740 ddt_phys_fill(ddp, bp);
2741 ddt_phys_addref(ddp);
2743 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2744 zio->io_orig_size, zio->io_orig_size, zp,
2745 zio_ddt_child_write_ready, NULL, NULL,
2746 zio_ddt_child_write_done, dde, zio->io_priority,
2747 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2749 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2750 dde->dde_lead_zio[p] = cio;
2760 return (ZIO_PIPELINE_CONTINUE);
2763 ddt_entry_t *freedde; /* for debugging */
2766 zio_ddt_free(zio_t *zio)
2768 spa_t *spa = zio->io_spa;
2769 blkptr_t *bp = zio->io_bp;
2770 ddt_t *ddt = ddt_select(spa, bp);
2774 ASSERT(BP_GET_DEDUP(bp));
2775 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2778 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2779 ddp = ddt_phys_select(dde, bp);
2780 ddt_phys_decref(ddp);
2783 return (ZIO_PIPELINE_CONTINUE);
2787 * ==========================================================================
2788 * Allocate and free blocks
2789 * ==========================================================================
2793 zio_io_to_allocate(spa_t *spa)
2797 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2799 zio = avl_first(&spa->spa_alloc_tree);
2803 ASSERT(IO_IS_ALLOCATING(zio));
2806 * Try to place a reservation for this zio. If we're unable to
2807 * reserve then we throttle.
2809 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2810 zio->io_prop.zp_copies, zio, 0)) {
2814 avl_remove(&spa->spa_alloc_tree, zio);
2815 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2821 zio_dva_throttle(zio_t *zio)
2823 spa_t *spa = zio->io_spa;
2826 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2827 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2828 zio->io_child_type == ZIO_CHILD_GANG ||
2829 zio->io_flags & ZIO_FLAG_NODATA) {
2830 return (ZIO_PIPELINE_CONTINUE);
2833 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2835 ASSERT3U(zio->io_queued_timestamp, >, 0);
2836 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2838 mutex_enter(&spa->spa_alloc_lock);
2840 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2841 avl_add(&spa->spa_alloc_tree, zio);
2843 nio = zio_io_to_allocate(zio->io_spa);
2844 mutex_exit(&spa->spa_alloc_lock);
2847 return (ZIO_PIPELINE_CONTINUE);
2850 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2852 * We are passing control to a new zio so make sure that
2853 * it is processed by a different thread. We do this to
2854 * avoid stack overflows that can occur when parents are
2855 * throttled and children are making progress. We allow
2856 * it to go to the head of the taskq since it's already
2859 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2861 return (ZIO_PIPELINE_STOP);
2865 zio_allocate_dispatch(spa_t *spa)
2869 mutex_enter(&spa->spa_alloc_lock);
2870 zio = zio_io_to_allocate(spa);
2871 mutex_exit(&spa->spa_alloc_lock);
2875 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2876 ASSERT0(zio->io_error);
2877 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2881 zio_dva_allocate(zio_t *zio)
2883 spa_t *spa = zio->io_spa;
2884 metaslab_class_t *mc = spa_normal_class(spa);
2885 blkptr_t *bp = zio->io_bp;
2889 if (zio->io_gang_leader == NULL) {
2890 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2891 zio->io_gang_leader = zio;
2894 ASSERT(BP_IS_HOLE(bp));
2895 ASSERT0(BP_GET_NDVAS(bp));
2896 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2897 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2898 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2900 if (zio->io_flags & ZIO_FLAG_NODATA) {
2901 flags |= METASLAB_DONT_THROTTLE;
2903 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2904 flags |= METASLAB_GANG_CHILD;
2906 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2907 flags |= METASLAB_ASYNC_ALLOC;
2910 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2911 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2912 &zio->io_alloc_list, zio);
2915 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2916 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2918 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2919 return (zio_write_gang_block(zio));
2920 zio->io_error = error;
2923 return (ZIO_PIPELINE_CONTINUE);
2927 zio_dva_free(zio_t *zio)
2929 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2931 return (ZIO_PIPELINE_CONTINUE);
2935 zio_dva_claim(zio_t *zio)
2939 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2941 zio->io_error = error;
2943 return (ZIO_PIPELINE_CONTINUE);
2947 * Undo an allocation. This is used by zio_done() when an I/O fails
2948 * and we want to give back the block we just allocated.
2949 * This handles both normal blocks and gang blocks.
2952 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2954 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2955 ASSERT(zio->io_bp_override == NULL);
2957 if (!BP_IS_HOLE(bp))
2958 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2961 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2962 zio_dva_unallocate(zio, gn->gn_child[g],
2963 &gn->gn_gbh->zg_blkptr[g]);
2969 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2972 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2973 uint64_t size, boolean_t *slog)
2976 zio_alloc_list_t io_alloc_list;
2978 ASSERT(txg > spa_syncing_txg(spa));
2980 metaslab_trace_init(&io_alloc_list);
2981 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
2982 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
2986 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2987 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
2988 &io_alloc_list, NULL);
2992 metaslab_trace_fini(&io_alloc_list);
2995 BP_SET_LSIZE(new_bp, size);
2996 BP_SET_PSIZE(new_bp, size);
2997 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2998 BP_SET_CHECKSUM(new_bp,
2999 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3000 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3001 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3002 BP_SET_LEVEL(new_bp, 0);
3003 BP_SET_DEDUP(new_bp, 0);
3004 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3011 * Free an intent log block.
3014 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3016 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3017 ASSERT(!BP_IS_GANG(bp));
3019 zio_free(spa, txg, bp);
3023 * ==========================================================================
3024 * Read, write and delete to physical devices
3025 * ==========================================================================
3030 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3031 * stops after this stage and will resume upon I/O completion.
3032 * However, there are instances where the vdev layer may need to
3033 * continue the pipeline when an I/O was not issued. Since the I/O
3034 * that was sent to the vdev layer might be different than the one
3035 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3036 * force the underlying vdev layers to call either zio_execute() or
3037 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3040 zio_vdev_io_start(zio_t *zio)
3042 vdev_t *vd = zio->io_vd;
3044 spa_t *spa = zio->io_spa;
3047 ASSERT(zio->io_error == 0);
3048 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3051 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3052 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3055 * The mirror_ops handle multiple DVAs in a single BP.
3057 vdev_mirror_ops.vdev_op_io_start(zio);
3058 return (ZIO_PIPELINE_STOP);
3061 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3062 zio->io_priority == ZIO_PRIORITY_NOW) {
3063 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3064 return (ZIO_PIPELINE_CONTINUE);
3067 ASSERT3P(zio->io_logical, !=, zio);
3070 * We keep track of time-sensitive I/Os so that the scan thread
3071 * can quickly react to certain workloads. In particular, we care
3072 * about non-scrubbing, top-level reads and writes with the following
3074 * - synchronous writes of user data to non-slog devices
3075 * - any reads of user data
3076 * When these conditions are met, adjust the timestamp of spa_last_io
3077 * which allows the scan thread to adjust its workload accordingly.
3079 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3080 vd == vd->vdev_top && !vd->vdev_islog &&
3081 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3082 zio->io_txg != spa_syncing_txg(spa)) {
3083 uint64_t old = spa->spa_last_io;
3084 uint64_t new = ddi_get_lbolt64();
3086 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3089 align = 1ULL << vd->vdev_top->vdev_ashift;
3091 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3092 P2PHASE(zio->io_size, align) != 0) {
3093 /* Transform logical writes to be a full physical block size. */
3094 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3096 if (zio->io_type == ZIO_TYPE_READ ||
3097 zio->io_type == ZIO_TYPE_WRITE)
3098 abuf = zio_buf_alloc(asize);
3099 ASSERT(vd == vd->vdev_top);
3100 if (zio->io_type == ZIO_TYPE_WRITE) {
3101 bcopy(zio->io_data, abuf, zio->io_size);
3102 bzero(abuf + zio->io_size, asize - zio->io_size);
3104 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3109 * If this is not a physical io, make sure that it is properly aligned
3110 * before proceeding.
3112 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3113 ASSERT0(P2PHASE(zio->io_offset, align));
3114 ASSERT0(P2PHASE(zio->io_size, align));
3117 * For the physical io we allow alignment
3118 * to a logical block size.
3120 uint64_t log_align =
3121 1ULL << vd->vdev_top->vdev_logical_ashift;
3122 ASSERT0(P2PHASE(zio->io_offset, log_align));
3123 ASSERT0(P2PHASE(zio->io_size, log_align));
3126 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3129 * If this is a repair I/O, and there's no self-healing involved --
3130 * that is, we're just resilvering what we expect to resilver --
3131 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3132 * This prevents spurious resilvering with nested replication.
3133 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3134 * A is out of date, we'll read from C+D, then use the data to
3135 * resilver A+B -- but we don't actually want to resilver B, just A.
3136 * The top-level mirror has no way to know this, so instead we just
3137 * discard unnecessary repairs as we work our way down the vdev tree.
3138 * The same logic applies to any form of nested replication:
3139 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3141 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3142 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3143 zio->io_txg != 0 && /* not a delegated i/o */
3144 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3145 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3146 zio_vdev_io_bypass(zio);
3147 return (ZIO_PIPELINE_CONTINUE);
3150 if (vd->vdev_ops->vdev_op_leaf) {
3151 switch (zio->io_type) {
3153 if (vdev_cache_read(zio))
3154 return (ZIO_PIPELINE_CONTINUE);
3156 case ZIO_TYPE_WRITE:
3158 if ((zio = vdev_queue_io(zio)) == NULL)
3159 return (ZIO_PIPELINE_STOP);
3161 if (!vdev_accessible(vd, zio)) {
3162 zio->io_error = SET_ERROR(ENXIO);
3164 return (ZIO_PIPELINE_STOP);
3169 * Note that we ignore repair writes for TRIM because they can
3170 * conflict with normal writes. This isn't an issue because, by
3171 * definition, we only repair blocks that aren't freed.
3173 if (zio->io_type == ZIO_TYPE_WRITE &&
3174 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3175 !trim_map_write_start(zio))
3176 return (ZIO_PIPELINE_STOP);
3179 vd->vdev_ops->vdev_op_io_start(zio);
3180 return (ZIO_PIPELINE_STOP);
3184 zio_vdev_io_done(zio_t *zio)
3186 vdev_t *vd = zio->io_vd;
3187 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3188 boolean_t unexpected_error = B_FALSE;
3190 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3191 return (ZIO_PIPELINE_STOP);
3193 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3194 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3196 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3197 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3198 zio->io_type == ZIO_TYPE_FREE)) {
3200 if (zio->io_type == ZIO_TYPE_WRITE &&
3201 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3202 trim_map_write_done(zio);
3204 vdev_queue_io_done(zio);
3206 if (zio->io_type == ZIO_TYPE_WRITE)
3207 vdev_cache_write(zio);
3209 if (zio_injection_enabled && zio->io_error == 0)
3210 zio->io_error = zio_handle_device_injection(vd,
3213 if (zio_injection_enabled && zio->io_error == 0)
3214 zio->io_error = zio_handle_label_injection(zio, EIO);
3216 if (zio->io_error) {
3217 if (zio->io_error == ENOTSUP &&
3218 zio->io_type == ZIO_TYPE_FREE) {
3219 /* Not all devices support TRIM. */
3220 } else if (!vdev_accessible(vd, zio)) {
3221 zio->io_error = SET_ERROR(ENXIO);
3223 unexpected_error = B_TRUE;
3228 ops->vdev_op_io_done(zio);
3230 if (unexpected_error)
3231 VERIFY(vdev_probe(vd, zio) == NULL);
3233 return (ZIO_PIPELINE_CONTINUE);
3237 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3238 * disk, and use that to finish the checksum ereport later.
3241 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3242 const void *good_buf)
3244 /* no processing needed */
3245 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3250 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3252 void *buf = zio_buf_alloc(zio->io_size);
3254 bcopy(zio->io_data, buf, zio->io_size);
3256 zcr->zcr_cbinfo = zio->io_size;
3257 zcr->zcr_cbdata = buf;
3258 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3259 zcr->zcr_free = zio_buf_free;
3263 zio_vdev_io_assess(zio_t *zio)
3265 vdev_t *vd = zio->io_vd;
3267 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3268 return (ZIO_PIPELINE_STOP);
3270 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3271 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3273 if (zio->io_vsd != NULL) {
3274 zio->io_vsd_ops->vsd_free(zio);
3278 if (zio_injection_enabled && zio->io_error == 0)
3279 zio->io_error = zio_handle_fault_injection(zio, EIO);
3281 if (zio->io_type == ZIO_TYPE_FREE &&
3282 zio->io_priority != ZIO_PRIORITY_NOW) {
3283 switch (zio->io_error) {
3285 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3286 ZIO_TRIM_STAT_BUMP(success);
3289 ZIO_TRIM_STAT_BUMP(unsupported);
3292 ZIO_TRIM_STAT_BUMP(failed);
3298 * If the I/O failed, determine whether we should attempt to retry it.
3300 * On retry, we cut in line in the issue queue, since we don't want
3301 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3303 if (zio->io_error && vd == NULL &&
3304 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3305 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3306 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3308 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3309 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3310 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3311 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3312 zio_requeue_io_start_cut_in_line);
3313 return (ZIO_PIPELINE_STOP);
3317 * If we got an error on a leaf device, convert it to ENXIO
3318 * if the device is not accessible at all.
3320 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3321 !vdev_accessible(vd, zio))
3322 zio->io_error = SET_ERROR(ENXIO);
3325 * If we can't write to an interior vdev (mirror or RAID-Z),
3326 * set vdev_cant_write so that we stop trying to allocate from it.
3328 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3329 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3330 vd->vdev_cant_write = B_TRUE;
3334 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3335 * attempts will ever succeed. In this case we set a persistent bit so
3336 * that we don't bother with it in the future.
3338 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3339 zio->io_type == ZIO_TYPE_IOCTL &&
3340 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3341 vd->vdev_nowritecache = B_TRUE;
3344 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3346 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3347 zio->io_physdone != NULL) {
3348 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3349 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3350 zio->io_physdone(zio->io_logical);
3353 return (ZIO_PIPELINE_CONTINUE);
3357 zio_vdev_io_reissue(zio_t *zio)
3359 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3360 ASSERT(zio->io_error == 0);
3362 zio->io_stage >>= 1;
3366 zio_vdev_io_redone(zio_t *zio)
3368 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3370 zio->io_stage >>= 1;
3374 zio_vdev_io_bypass(zio_t *zio)
3376 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3377 ASSERT(zio->io_error == 0);
3379 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3380 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3384 * ==========================================================================
3385 * Generate and verify checksums
3386 * ==========================================================================
3389 zio_checksum_generate(zio_t *zio)
3391 blkptr_t *bp = zio->io_bp;
3392 enum zio_checksum checksum;
3396 * This is zio_write_phys().
3397 * We're either generating a label checksum, or none at all.
3399 checksum = zio->io_prop.zp_checksum;
3401 if (checksum == ZIO_CHECKSUM_OFF)
3402 return (ZIO_PIPELINE_CONTINUE);
3404 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3406 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3407 ASSERT(!IO_IS_ALLOCATING(zio));
3408 checksum = ZIO_CHECKSUM_GANG_HEADER;
3410 checksum = BP_GET_CHECKSUM(bp);
3414 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3416 return (ZIO_PIPELINE_CONTINUE);
3420 zio_checksum_verify(zio_t *zio)
3422 zio_bad_cksum_t info;
3423 blkptr_t *bp = zio->io_bp;
3426 ASSERT(zio->io_vd != NULL);
3430 * This is zio_read_phys().
3431 * We're either verifying a label checksum, or nothing at all.
3433 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3434 return (ZIO_PIPELINE_CONTINUE);
3436 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3439 if ((error = zio_checksum_error(zio, &info)) != 0) {
3440 zio->io_error = error;
3441 if (error == ECKSUM &&
3442 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3443 zfs_ereport_start_checksum(zio->io_spa,
3444 zio->io_vd, zio, zio->io_offset,
3445 zio->io_size, NULL, &info);
3449 return (ZIO_PIPELINE_CONTINUE);
3453 * Called by RAID-Z to ensure we don't compute the checksum twice.
3456 zio_checksum_verified(zio_t *zio)
3458 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3462 * ==========================================================================
3463 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3464 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3465 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3466 * indicate errors that are specific to one I/O, and most likely permanent.
3467 * Any other error is presumed to be worse because we weren't expecting it.
3468 * ==========================================================================
3471 zio_worst_error(int e1, int e2)
3473 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3476 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3477 if (e1 == zio_error_rank[r1])
3480 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3481 if (e2 == zio_error_rank[r2])
3484 return (r1 > r2 ? e1 : e2);
3488 * ==========================================================================
3490 * ==========================================================================
3493 zio_ready(zio_t *zio)
3495 blkptr_t *bp = zio->io_bp;
3496 zio_t *pio, *pio_next;
3497 zio_link_t *zl = NULL;
3499 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3500 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3501 return (ZIO_PIPELINE_STOP);
3503 if (zio->io_ready) {
3504 ASSERT(IO_IS_ALLOCATING(zio));
3505 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3506 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3507 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3512 if (bp != NULL && bp != &zio->io_bp_copy)
3513 zio->io_bp_copy = *bp;
3515 if (zio->io_error != 0) {
3516 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3518 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3519 ASSERT(IO_IS_ALLOCATING(zio));
3520 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3522 * We were unable to allocate anything, unreserve and
3523 * issue the next I/O to allocate.
3525 metaslab_class_throttle_unreserve(
3526 spa_normal_class(zio->io_spa),
3527 zio->io_prop.zp_copies, zio);
3528 zio_allocate_dispatch(zio->io_spa);
3532 mutex_enter(&zio->io_lock);
3533 zio->io_state[ZIO_WAIT_READY] = 1;
3534 pio = zio_walk_parents(zio, &zl);
3535 mutex_exit(&zio->io_lock);
3538 * As we notify zio's parents, new parents could be added.
3539 * New parents go to the head of zio's io_parent_list, however,
3540 * so we will (correctly) not notify them. The remainder of zio's
3541 * io_parent_list, from 'pio_next' onward, cannot change because
3542 * all parents must wait for us to be done before they can be done.
3544 for (; pio != NULL; pio = pio_next) {
3545 pio_next = zio_walk_parents(zio, &zl);
3546 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3549 if (zio->io_flags & ZIO_FLAG_NODATA) {
3550 if (BP_IS_GANG(bp)) {
3551 zio->io_flags &= ~ZIO_FLAG_NODATA;
3553 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3554 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3558 if (zio_injection_enabled &&
3559 zio->io_spa->spa_syncing_txg == zio->io_txg)
3560 zio_handle_ignored_writes(zio);
3562 return (ZIO_PIPELINE_CONTINUE);
3566 * Update the allocation throttle accounting.
3569 zio_dva_throttle_done(zio_t *zio)
3571 zio_t *lio = zio->io_logical;
3572 zio_t *pio = zio_unique_parent(zio);
3573 vdev_t *vd = zio->io_vd;
3574 int flags = METASLAB_ASYNC_ALLOC;
3576 ASSERT3P(zio->io_bp, !=, NULL);
3577 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3578 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3579 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3581 ASSERT3P(vd, ==, vd->vdev_top);
3582 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3583 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3584 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3585 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3588 * Parents of gang children can have two flavors -- ones that
3589 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3590 * and ones that allocated the constituent blocks. The allocation
3591 * throttle needs to know the allocating parent zio so we must find
3594 if (pio->io_child_type == ZIO_CHILD_GANG) {
3596 * If our parent is a rewrite gang child then our grandparent
3597 * would have been the one that performed the allocation.
3599 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3600 pio = zio_unique_parent(pio);
3601 flags |= METASLAB_GANG_CHILD;
3604 ASSERT(IO_IS_ALLOCATING(pio));
3605 ASSERT3P(zio, !=, zio->io_logical);
3606 ASSERT(zio->io_logical != NULL);
3607 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3608 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3610 mutex_enter(&pio->io_lock);
3611 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3612 mutex_exit(&pio->io_lock);
3614 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3618 * Call into the pipeline to see if there is more work that
3619 * needs to be done. If there is work to be done it will be
3620 * dispatched to another taskq thread.
3622 zio_allocate_dispatch(zio->io_spa);
3626 zio_done(zio_t *zio)
3628 spa_t *spa = zio->io_spa;
3629 zio_t *lio = zio->io_logical;
3630 blkptr_t *bp = zio->io_bp;
3631 vdev_t *vd = zio->io_vd;
3632 uint64_t psize = zio->io_size;
3633 zio_t *pio, *pio_next;
3634 metaslab_class_t *mc = spa_normal_class(spa);
3635 zio_link_t *zl = NULL;
3638 * If our children haven't all completed,
3639 * wait for them and then repeat this pipeline stage.
3641 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3642 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3643 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3644 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3645 return (ZIO_PIPELINE_STOP);
3648 * If the allocation throttle is enabled, then update the accounting.
3649 * We only track child I/Os that are part of an allocating async
3650 * write. We must do this since the allocation is performed
3651 * by the logical I/O but the actual write is done by child I/Os.
3653 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3654 zio->io_child_type == ZIO_CHILD_VDEV) {
3655 ASSERT(mc->mc_alloc_throttle_enabled);
3656 zio_dva_throttle_done(zio);
3660 * If the allocation throttle is enabled, verify that
3661 * we have decremented the refcounts for every I/O that was throttled.
3663 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3664 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3665 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3667 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3668 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3671 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3672 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3673 ASSERT(zio->io_children[c][w] == 0);
3675 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3676 ASSERT(bp->blk_pad[0] == 0);
3677 ASSERT(bp->blk_pad[1] == 0);
3678 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3679 (bp == zio_unique_parent(zio)->io_bp));
3680 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3681 zio->io_bp_override == NULL &&
3682 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3683 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3684 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3685 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3686 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3688 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3689 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3693 * If there were child vdev/gang/ddt errors, they apply to us now.
3695 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3696 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3697 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3700 * If the I/O on the transformed data was successful, generate any
3701 * checksum reports now while we still have the transformed data.
3703 if (zio->io_error == 0) {
3704 while (zio->io_cksum_report != NULL) {
3705 zio_cksum_report_t *zcr = zio->io_cksum_report;
3706 uint64_t align = zcr->zcr_align;
3707 uint64_t asize = P2ROUNDUP(psize, align);
3708 char *abuf = zio->io_data;
3710 if (asize != psize) {
3711 abuf = zio_buf_alloc(asize);
3712 bcopy(zio->io_data, abuf, psize);
3713 bzero(abuf + psize, asize - psize);
3716 zio->io_cksum_report = zcr->zcr_next;
3717 zcr->zcr_next = NULL;
3718 zcr->zcr_finish(zcr, abuf);
3719 zfs_ereport_free_checksum(zcr);
3722 zio_buf_free(abuf, asize);
3726 zio_pop_transforms(zio); /* note: may set zio->io_error */
3728 vdev_stat_update(zio, psize);
3730 if (zio->io_error) {
3732 * If this I/O is attached to a particular vdev,
3733 * generate an error message describing the I/O failure
3734 * at the block level. We ignore these errors if the
3735 * device is currently unavailable.
3737 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3738 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3740 if ((zio->io_error == EIO || !(zio->io_flags &
3741 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3744 * For logical I/O requests, tell the SPA to log the
3745 * error and generate a logical data ereport.
3747 spa_log_error(spa, zio);
3748 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3753 if (zio->io_error && zio == lio) {
3755 * Determine whether zio should be reexecuted. This will
3756 * propagate all the way to the root via zio_notify_parent().
3758 ASSERT(vd == NULL && bp != NULL);
3759 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3761 if (IO_IS_ALLOCATING(zio) &&
3762 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3763 if (zio->io_error != ENOSPC)
3764 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3766 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3769 if ((zio->io_type == ZIO_TYPE_READ ||
3770 zio->io_type == ZIO_TYPE_FREE) &&
3771 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3772 zio->io_error == ENXIO &&
3773 spa_load_state(spa) == SPA_LOAD_NONE &&
3774 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3775 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3777 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3778 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3781 * Here is a possibly good place to attempt to do
3782 * either combinatorial reconstruction or error correction
3783 * based on checksums. It also might be a good place
3784 * to send out preliminary ereports before we suspend
3790 * If there were logical child errors, they apply to us now.
3791 * We defer this until now to avoid conflating logical child
3792 * errors with errors that happened to the zio itself when
3793 * updating vdev stats and reporting FMA events above.
3795 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3797 if ((zio->io_error || zio->io_reexecute) &&
3798 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3799 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3800 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3802 zio_gang_tree_free(&zio->io_gang_tree);
3805 * Godfather I/Os should never suspend.
3807 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3808 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3809 zio->io_reexecute = 0;
3811 if (zio->io_reexecute) {
3813 * This is a logical I/O that wants to reexecute.
3815 * Reexecute is top-down. When an i/o fails, if it's not
3816 * the root, it simply notifies its parent and sticks around.
3817 * The parent, seeing that it still has children in zio_done(),
3818 * does the same. This percolates all the way up to the root.
3819 * The root i/o will reexecute or suspend the entire tree.
3821 * This approach ensures that zio_reexecute() honors
3822 * all the original i/o dependency relationships, e.g.
3823 * parents not executing until children are ready.
3825 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3827 zio->io_gang_leader = NULL;
3829 mutex_enter(&zio->io_lock);
3830 zio->io_state[ZIO_WAIT_DONE] = 1;
3831 mutex_exit(&zio->io_lock);
3834 * "The Godfather" I/O monitors its children but is
3835 * not a true parent to them. It will track them through
3836 * the pipeline but severs its ties whenever they get into
3837 * trouble (e.g. suspended). This allows "The Godfather"
3838 * I/O to return status without blocking.
3841 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3843 zio_link_t *remove_zl = zl;
3844 pio_next = zio_walk_parents(zio, &zl);
3846 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3847 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3848 zio_remove_child(pio, zio, remove_zl);
3849 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3853 if ((pio = zio_unique_parent(zio)) != NULL) {
3855 * We're not a root i/o, so there's nothing to do
3856 * but notify our parent. Don't propagate errors
3857 * upward since we haven't permanently failed yet.
3859 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3860 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3861 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3862 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3864 * We'd fail again if we reexecuted now, so suspend
3865 * until conditions improve (e.g. device comes online).
3867 zio_suspend(spa, zio);
3870 * Reexecution is potentially a huge amount of work.
3871 * Hand it off to the otherwise-unused claim taskq.
3873 #if defined(illumos) || !defined(_KERNEL)
3874 ASSERT(zio->io_tqent.tqent_next == NULL);
3876 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3878 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3879 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3882 return (ZIO_PIPELINE_STOP);
3885 ASSERT(zio->io_child_count == 0);
3886 ASSERT(zio->io_reexecute == 0);
3887 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3890 * Report any checksum errors, since the I/O is complete.
3892 while (zio->io_cksum_report != NULL) {
3893 zio_cksum_report_t *zcr = zio->io_cksum_report;
3894 zio->io_cksum_report = zcr->zcr_next;
3895 zcr->zcr_next = NULL;
3896 zcr->zcr_finish(zcr, NULL);
3897 zfs_ereport_free_checksum(zcr);
3901 * It is the responsibility of the done callback to ensure that this
3902 * particular zio is no longer discoverable for adoption, and as
3903 * such, cannot acquire any new parents.
3908 mutex_enter(&zio->io_lock);
3909 zio->io_state[ZIO_WAIT_DONE] = 1;
3910 mutex_exit(&zio->io_lock);
3913 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3914 zio_link_t *remove_zl = zl;
3915 pio_next = zio_walk_parents(zio, &zl);
3916 zio_remove_child(pio, zio, remove_zl);
3917 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3920 if (zio->io_waiter != NULL) {
3921 mutex_enter(&zio->io_lock);
3922 zio->io_executor = NULL;
3923 cv_broadcast(&zio->io_cv);
3924 mutex_exit(&zio->io_lock);
3929 return (ZIO_PIPELINE_STOP);
3933 * ==========================================================================
3934 * I/O pipeline definition
3935 * ==========================================================================
3937 static zio_pipe_stage_t *zio_pipeline[] = {
3944 zio_checksum_generate,
3960 zio_checksum_verify,
3968 * Compare two zbookmark_phys_t's to see which we would reach first in a
3969 * pre-order traversal of the object tree.
3971 * This is simple in every case aside from the meta-dnode object. For all other
3972 * objects, we traverse them in order (object 1 before object 2, and so on).
3973 * However, all of these objects are traversed while traversing object 0, since
3974 * the data it points to is the list of objects. Thus, we need to convert to a
3975 * canonical representation so we can compare meta-dnode bookmarks to
3976 * non-meta-dnode bookmarks.
3978 * We do this by calculating "equivalents" for each field of the zbookmark.
3979 * zbookmarks outside of the meta-dnode use their own object and level, and
3980 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3981 * blocks this bookmark refers to) by multiplying their blkid by their span
3982 * (the number of L0 blocks contained within one block at their level).
3983 * zbookmarks inside the meta-dnode calculate their object equivalent
3984 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3985 * level + 1<<31 (any value larger than a level could ever be) for their level.
3986 * This causes them to always compare before a bookmark in their object
3987 * equivalent, compare appropriately to bookmarks in other objects, and to
3988 * compare appropriately to other bookmarks in the meta-dnode.
3991 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3992 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3995 * These variables represent the "equivalent" values for the zbookmark,
3996 * after converting zbookmarks inside the meta dnode to their
3997 * normal-object equivalents.
3999 uint64_t zb1obj, zb2obj;
4000 uint64_t zb1L0, zb2L0;
4001 uint64_t zb1level, zb2level;
4003 if (zb1->zb_object == zb2->zb_object &&
4004 zb1->zb_level == zb2->zb_level &&
4005 zb1->zb_blkid == zb2->zb_blkid)
4009 * BP_SPANB calculates the span in blocks.
4011 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4012 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4014 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4015 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4017 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4019 zb1obj = zb1->zb_object;
4020 zb1level = zb1->zb_level;
4023 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4024 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4026 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4028 zb2obj = zb2->zb_object;
4029 zb2level = zb2->zb_level;
4032 /* Now that we have a canonical representation, do the comparison. */
4033 if (zb1obj != zb2obj)
4034 return (zb1obj < zb2obj ? -1 : 1);
4035 else if (zb1L0 != zb2L0)
4036 return (zb1L0 < zb2L0 ? -1 : 1);
4037 else if (zb1level != zb2level)
4038 return (zb1level > zb2level ? -1 : 1);
4040 * This can (theoretically) happen if the bookmarks have the same object
4041 * and level, but different blkids, if the block sizes are not the same.
4042 * There is presently no way to change the indirect block sizes
4048 * This function checks the following: given that last_block is the place that
4049 * our traversal stopped last time, does that guarantee that we've visited
4050 * every node under subtree_root? Therefore, we can't just use the raw output
4051 * of zbookmark_compare. We have to pass in a modified version of
4052 * subtree_root; by incrementing the block id, and then checking whether
4053 * last_block is before or equal to that, we can tell whether or not having
4054 * visited last_block implies that all of subtree_root's children have been
4058 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4059 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4061 zbookmark_phys_t mod_zb = *subtree_root;
4063 ASSERT(last_block->zb_level == 0);
4065 /* The objset_phys_t isn't before anything. */
4070 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4071 * data block size in sectors, because that variable is only used if
4072 * the bookmark refers to a block in the meta-dnode. Since we don't
4073 * know without examining it what object it refers to, and there's no
4074 * harm in passing in this value in other cases, we always pass it in.
4076 * We pass in 0 for the indirect block size shift because zb2 must be
4077 * level 0. The indirect block size is only used to calculate the span
4078 * of the bookmark, but since the bookmark must be level 0, the span is
4079 * always 1, so the math works out.
4081 * If you make changes to how the zbookmark_compare code works, be sure
4082 * to make sure that this code still works afterwards.
4084 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4085 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,