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, 2016 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 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
55 "Use uma(9) for ZIO allocations");
56 static int zio_exclude_metadata = 0;
57 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
58 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
59 "Exclude metadata buffers from dumps as well");
61 zio_trim_stats_t zio_trim_stats = {
62 { "bytes", KSTAT_DATA_UINT64,
63 "Number of bytes successfully TRIMmed" },
64 { "success", KSTAT_DATA_UINT64,
65 "Number of successful TRIM requests" },
66 { "unsupported", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed because TRIM is not supported" },
68 { "failed", KSTAT_DATA_UINT64,
69 "Number of TRIM requests that failed for reasons other than not supported" },
72 static kstat_t *zio_trim_ksp;
75 * ==========================================================================
76 * I/O type descriptions
77 * ==========================================================================
79 const char *zio_type_name[ZIO_TYPES] = {
80 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
84 boolean_t zio_dva_throttle_enabled = B_TRUE;
85 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
86 &zio_dva_throttle_enabled, 0, "");
89 * ==========================================================================
91 * ==========================================================================
93 kmem_cache_t *zio_cache;
94 kmem_cache_t *zio_link_cache;
95 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
96 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
99 extern vmem_t *zio_alloc_arena;
102 #define ZIO_PIPELINE_CONTINUE 0x100
103 #define ZIO_PIPELINE_STOP 0x101
105 #define BP_SPANB(indblkshift, level) \
106 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
107 #define COMPARE_META_LEVEL 0x80000000ul
109 * The following actions directly effect the spa's sync-to-convergence logic.
110 * The values below define the sync pass when we start performing the action.
111 * Care should be taken when changing these values as they directly impact
112 * spa_sync() performance. Tuning these values may introduce subtle performance
113 * pathologies and should only be done in the context of performance analysis.
114 * These tunables will eventually be removed and replaced with #defines once
115 * enough analysis has been done to determine optimal values.
117 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
118 * regular blocks are not deferred.
120 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
121 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
123 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
124 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
125 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
126 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
127 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
128 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
129 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
130 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
131 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
134 * An allocating zio is one that either currently has the DVA allocate
135 * stage set or will have it later in its lifetime.
137 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
139 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
143 int zio_buf_debug_limit = 16384;
145 int zio_buf_debug_limit = 0;
149 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
155 zio_cache = kmem_cache_create("zio_cache",
156 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
157 zio_link_cache = kmem_cache_create("zio_link_cache",
158 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
163 * For small buffers, we want a cache for each multiple of
164 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
165 * for each quarter-power of 2.
167 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
168 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
171 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
179 * If we are using watchpoints, put each buffer on its own page,
180 * to eliminate the performance overhead of trapping to the
181 * kernel when modifying a non-watched buffer that shares the
182 * page with a watched buffer.
184 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
188 if (size <= 4 * SPA_MINBLOCKSIZE) {
189 align = SPA_MINBLOCKSIZE;
190 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
191 align = MIN(p2 >> 2, PAGESIZE);
196 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
197 zio_buf_cache[c] = kmem_cache_create(name, size,
198 align, NULL, NULL, NULL, NULL, NULL, cflags);
201 * Since zio_data bufs do not appear in crash dumps, we
202 * pass KMC_NOTOUCH so that no allocator metadata is
203 * stored with the buffers.
205 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
206 zio_data_buf_cache[c] = kmem_cache_create(name, size,
207 align, NULL, NULL, NULL, NULL, NULL,
208 cflags | KMC_NOTOUCH | KMC_NODEBUG);
213 ASSERT(zio_buf_cache[c] != NULL);
214 if (zio_buf_cache[c - 1] == NULL)
215 zio_buf_cache[c - 1] = zio_buf_cache[c];
217 ASSERT(zio_data_buf_cache[c] != NULL);
218 if (zio_data_buf_cache[c - 1] == NULL)
219 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
225 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
227 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
230 if (zio_trim_ksp != NULL) {
231 zio_trim_ksp->ks_data = &zio_trim_stats;
232 kstat_install(zio_trim_ksp);
240 kmem_cache_t *last_cache = NULL;
241 kmem_cache_t *last_data_cache = NULL;
243 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
244 if (zio_buf_cache[c] != last_cache) {
245 last_cache = zio_buf_cache[c];
246 kmem_cache_destroy(zio_buf_cache[c]);
248 zio_buf_cache[c] = NULL;
250 if (zio_data_buf_cache[c] != last_data_cache) {
251 last_data_cache = zio_data_buf_cache[c];
252 kmem_cache_destroy(zio_data_buf_cache[c]);
254 zio_data_buf_cache[c] = NULL;
257 kmem_cache_destroy(zio_link_cache);
258 kmem_cache_destroy(zio_cache);
262 if (zio_trim_ksp != NULL) {
263 kstat_delete(zio_trim_ksp);
269 * ==========================================================================
270 * Allocate and free I/O buffers
271 * ==========================================================================
275 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
276 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
277 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
278 * excess / transient data in-core during a crashdump.
281 zio_buf_alloc_impl(size_t size, boolean_t canwait)
283 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
284 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
286 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
289 return (kmem_cache_alloc(zio_buf_cache[c],
290 canwait ? KM_PUSHPAGE : KM_NOSLEEP));
292 return (kmem_alloc(size,
293 (canwait ? KM_SLEEP : KM_NOSLEEP) | flags));
298 zio_buf_alloc(size_t size)
300 return (zio_buf_alloc_impl(size, B_TRUE));
304 zio_buf_alloc_nowait(size_t size)
306 return (zio_buf_alloc_impl(size, B_FALSE));
310 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
311 * crashdump if the kernel panics. This exists so that we will limit the amount
312 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
313 * of kernel heap dumped to disk when the kernel panics)
316 zio_data_buf_alloc(size_t size)
318 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
320 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
323 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
325 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
329 zio_buf_free(void *buf, size_t size)
331 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
333 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
336 kmem_cache_free(zio_buf_cache[c], buf);
338 kmem_free(buf, size);
342 zio_data_buf_free(void *buf, size_t size)
344 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
346 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
349 kmem_cache_free(zio_data_buf_cache[c], buf);
351 kmem_free(buf, size);
355 * ==========================================================================
356 * Push and pop I/O transform buffers
357 * ==========================================================================
360 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
361 zio_transform_func_t *transform)
363 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
365 zt->zt_orig_data = zio->io_data;
366 zt->zt_orig_size = zio->io_size;
367 zt->zt_bufsize = bufsize;
368 zt->zt_transform = transform;
370 zt->zt_next = zio->io_transform_stack;
371 zio->io_transform_stack = zt;
378 zio_pop_transforms(zio_t *zio)
382 while ((zt = zio->io_transform_stack) != NULL) {
383 if (zt->zt_transform != NULL)
384 zt->zt_transform(zio,
385 zt->zt_orig_data, zt->zt_orig_size);
387 if (zt->zt_bufsize != 0)
388 zio_buf_free(zio->io_data, zt->zt_bufsize);
390 zio->io_data = zt->zt_orig_data;
391 zio->io_size = zt->zt_orig_size;
392 zio->io_transform_stack = zt->zt_next;
394 kmem_free(zt, sizeof (zio_transform_t));
399 * ==========================================================================
400 * I/O transform callbacks for subblocks and decompression
401 * ==========================================================================
404 zio_subblock(zio_t *zio, void *data, uint64_t size)
406 ASSERT(zio->io_size > size);
408 if (zio->io_type == ZIO_TYPE_READ)
409 bcopy(zio->io_data, data, size);
413 zio_decompress(zio_t *zio, void *data, uint64_t size)
415 if (zio->io_error == 0 &&
416 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
417 zio->io_data, data, zio->io_size, size) != 0)
418 zio->io_error = SET_ERROR(EIO);
422 * ==========================================================================
423 * I/O parent/child relationships and pipeline interlocks
424 * ==========================================================================
427 zio_walk_parents(zio_t *cio, zio_link_t **zl)
429 list_t *pl = &cio->io_parent_list;
431 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
435 ASSERT((*zl)->zl_child == cio);
436 return ((*zl)->zl_parent);
440 zio_walk_children(zio_t *pio, zio_link_t **zl)
442 list_t *cl = &pio->io_child_list;
444 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
448 ASSERT((*zl)->zl_parent == pio);
449 return ((*zl)->zl_child);
453 zio_unique_parent(zio_t *cio)
455 zio_link_t *zl = NULL;
456 zio_t *pio = zio_walk_parents(cio, &zl);
458 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
463 zio_add_child(zio_t *pio, zio_t *cio)
465 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT(cio->io_child_type <= pio->io_child_type);
478 mutex_enter(&cio->io_lock);
479 mutex_enter(&pio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&pio->io_lock);
493 mutex_exit(&cio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&cio->io_lock);
503 mutex_enter(&pio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&pio->io_lock);
512 mutex_exit(&cio->io_lock);
514 kmem_cache_free(zio_link_cache, zl);
518 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
520 boolean_t waiting = B_FALSE;
522 mutex_enter(&zio->io_lock);
523 ASSERT(zio->io_stall == NULL);
524 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
525 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
528 uint64_t *countp = &zio->io_children[c][wait];
531 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
532 zio->io_stall = countp;
537 mutex_exit(&zio->io_lock);
542 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
544 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
545 int *errorp = &pio->io_child_error[zio->io_child_type];
547 mutex_enter(&pio->io_lock);
548 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
549 *errorp = zio_worst_error(*errorp, zio->io_error);
550 pio->io_reexecute |= zio->io_reexecute;
551 ASSERT3U(*countp, >, 0);
555 if (*countp == 0 && pio->io_stall == countp) {
556 zio_taskq_type_t type =
557 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
559 pio->io_stall = NULL;
560 mutex_exit(&pio->io_lock);
562 * Dispatch the parent zio in its own taskq so that
563 * the child can continue to make progress. This also
564 * prevents overflowing the stack when we have deeply nested
565 * parent-child relationships.
567 zio_taskq_dispatch(pio, type, B_FALSE);
569 mutex_exit(&pio->io_lock);
574 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
576 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
577 zio->io_error = zio->io_child_error[c];
581 zio_timestamp_compare(const void *x1, const void *x2)
583 const zio_t *z1 = x1;
584 const zio_t *z2 = x2;
586 if (z1->io_queued_timestamp < z2->io_queued_timestamp)
588 if (z1->io_queued_timestamp > z2->io_queued_timestamp)
591 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
593 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
596 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
598 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
601 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
603 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
606 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
608 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
620 * ==========================================================================
621 * Create the various types of I/O (read, write, free, etc)
622 * ==========================================================================
625 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
626 void *data, uint64_t size, zio_done_func_t *done, void *private,
627 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
628 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
629 enum zio_stage stage, enum zio_stage pipeline)
633 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
634 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
635 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
637 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
638 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
639 ASSERT(vd || stage == ZIO_STAGE_OPEN);
641 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
642 bzero(zio, sizeof (zio_t));
644 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
645 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
647 list_create(&zio->io_parent_list, sizeof (zio_link_t),
648 offsetof(zio_link_t, zl_parent_node));
649 list_create(&zio->io_child_list, sizeof (zio_link_t),
650 offsetof(zio_link_t, zl_child_node));
653 zio->io_child_type = ZIO_CHILD_VDEV;
654 else if (flags & ZIO_FLAG_GANG_CHILD)
655 zio->io_child_type = ZIO_CHILD_GANG;
656 else if (flags & ZIO_FLAG_DDT_CHILD)
657 zio->io_child_type = ZIO_CHILD_DDT;
659 zio->io_child_type = ZIO_CHILD_LOGICAL;
662 zio->io_bp = (blkptr_t *)bp;
663 zio->io_bp_copy = *bp;
664 zio->io_bp_orig = *bp;
665 if (type != ZIO_TYPE_WRITE ||
666 zio->io_child_type == ZIO_CHILD_DDT)
667 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
668 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
669 zio->io_logical = zio;
670 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
671 pipeline |= ZIO_GANG_STAGES;
677 zio->io_private = private;
679 zio->io_priority = priority;
681 zio->io_offset = offset;
682 zio->io_orig_data = zio->io_data = data;
683 zio->io_orig_size = zio->io_size = size;
684 zio->io_orig_flags = zio->io_flags = flags;
685 zio->io_orig_stage = zio->io_stage = stage;
686 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
687 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
689 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
690 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
693 zio->io_bookmark = *zb;
696 if (zio->io_logical == NULL)
697 zio->io_logical = pio->io_logical;
698 if (zio->io_child_type == ZIO_CHILD_GANG)
699 zio->io_gang_leader = pio->io_gang_leader;
700 zio_add_child(pio, zio);
707 zio_destroy(zio_t *zio)
709 list_destroy(&zio->io_parent_list);
710 list_destroy(&zio->io_child_list);
711 mutex_destroy(&zio->io_lock);
712 cv_destroy(&zio->io_cv);
713 kmem_cache_free(zio_cache, zio);
717 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
718 void *private, enum zio_flag flags)
722 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
723 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
724 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
730 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
732 return (zio_null(NULL, spa, NULL, done, private, flags));
736 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
738 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
739 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
740 bp, (longlong_t)BP_GET_TYPE(bp));
742 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
743 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
744 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
745 bp, (longlong_t)BP_GET_CHECKSUM(bp));
747 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
748 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
749 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
750 bp, (longlong_t)BP_GET_COMPRESS(bp));
752 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
753 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
754 bp, (longlong_t)BP_GET_LSIZE(bp));
756 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
757 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
758 bp, (longlong_t)BP_GET_PSIZE(bp));
761 if (BP_IS_EMBEDDED(bp)) {
762 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
763 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
764 bp, (longlong_t)BPE_GET_ETYPE(bp));
769 * Pool-specific checks.
771 * Note: it would be nice to verify that the blk_birth and
772 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
773 * allows the birth time of log blocks (and dmu_sync()-ed blocks
774 * that are in the log) to be arbitrarily large.
776 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
777 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
778 if (vdevid >= spa->spa_root_vdev->vdev_children) {
779 zfs_panic_recover("blkptr at %p DVA %u has invalid "
781 bp, i, (longlong_t)vdevid);
784 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
786 zfs_panic_recover("blkptr at %p DVA %u has invalid "
788 bp, i, (longlong_t)vdevid);
791 if (vd->vdev_ops == &vdev_hole_ops) {
792 zfs_panic_recover("blkptr at %p DVA %u has hole "
794 bp, i, (longlong_t)vdevid);
797 if (vd->vdev_ops == &vdev_missing_ops) {
799 * "missing" vdevs are valid during import, but we
800 * don't have their detailed info (e.g. asize), so
801 * we can't perform any more checks on them.
805 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
806 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
808 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
809 if (offset + asize > vd->vdev_asize) {
810 zfs_panic_recover("blkptr at %p DVA %u has invalid "
812 bp, i, (longlong_t)offset);
818 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
819 void *data, uint64_t size, zio_done_func_t *done, void *private,
820 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
824 zfs_blkptr_verify(spa, bp);
826 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
827 data, size, done, private,
828 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
829 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
830 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
836 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
837 void *data, uint64_t size, const zio_prop_t *zp,
838 zio_done_func_t *ready, zio_done_func_t *children_ready,
839 zio_done_func_t *physdone, zio_done_func_t *done,
840 void *private, zio_priority_t priority, enum zio_flag flags,
841 const zbookmark_phys_t *zb)
845 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
846 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
847 zp->zp_compress >= ZIO_COMPRESS_OFF &&
848 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
849 DMU_OT_IS_VALID(zp->zp_type) &&
852 zp->zp_copies <= spa_max_replication(spa));
854 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
855 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
856 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
857 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
859 zio->io_ready = ready;
860 zio->io_children_ready = children_ready;
861 zio->io_physdone = physdone;
865 * Data can be NULL if we are going to call zio_write_override() to
866 * provide the already-allocated BP. But we may need the data to
867 * verify a dedup hit (if requested). In this case, don't try to
868 * dedup (just take the already-allocated BP verbatim).
870 if (data == NULL && zio->io_prop.zp_dedup_verify) {
871 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
878 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
879 uint64_t size, zio_done_func_t *done, void *private,
880 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
884 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
885 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
886 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
892 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
894 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
895 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
896 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
897 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
900 * We must reset the io_prop to match the values that existed
901 * when the bp was first written by dmu_sync() keeping in mind
902 * that nopwrite and dedup are mutually exclusive.
904 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
905 zio->io_prop.zp_nopwrite = nopwrite;
906 zio->io_prop.zp_copies = copies;
907 zio->io_bp_override = bp;
911 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
915 * The check for EMBEDDED is a performance optimization. We
916 * process the free here (by ignoring it) rather than
917 * putting it on the list and then processing it in zio_free_sync().
919 if (BP_IS_EMBEDDED(bp))
921 metaslab_check_free(spa, bp);
924 * Frees that are for the currently-syncing txg, are not going to be
925 * deferred, and which will not need to do a read (i.e. not GANG or
926 * DEDUP), can be processed immediately. Otherwise, put them on the
927 * in-memory list for later processing.
929 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
930 txg != spa->spa_syncing_txg ||
931 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
932 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
934 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
935 BP_GET_PSIZE(bp), 0)));
940 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
941 uint64_t size, enum zio_flag flags)
944 enum zio_stage stage = ZIO_FREE_PIPELINE;
946 ASSERT(!BP_IS_HOLE(bp));
947 ASSERT(spa_syncing_txg(spa) == txg);
948 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
950 if (BP_IS_EMBEDDED(bp))
951 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
953 metaslab_check_free(spa, bp);
956 if (zfs_trim_enabled)
957 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
958 ZIO_STAGE_VDEV_IO_ASSESS;
960 * GANG and DEDUP blocks can induce a read (for the gang block header,
961 * or the DDT), so issue them asynchronously so that this thread is
964 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
965 stage |= ZIO_STAGE_ISSUE_ASYNC;
967 flags |= ZIO_FLAG_DONT_QUEUE;
969 zio = zio_create(pio, spa, txg, bp, NULL, size,
970 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
971 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
977 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
978 zio_done_func_t *done, void *private, enum zio_flag flags)
982 dprintf_bp(bp, "claiming in txg %llu", txg);
984 if (BP_IS_EMBEDDED(bp))
985 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
988 * A claim is an allocation of a specific block. Claims are needed
989 * to support immediate writes in the intent log. The issue is that
990 * immediate writes contain committed data, but in a txg that was
991 * *not* committed. Upon opening the pool after an unclean shutdown,
992 * the intent log claims all blocks that contain immediate write data
993 * so that the SPA knows they're in use.
995 * All claims *must* be resolved in the first txg -- before the SPA
996 * starts allocating blocks -- so that nothing is allocated twice.
997 * If txg == 0 we just verify that the block is claimable.
999 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
1000 ASSERT(txg == spa_first_txg(spa) || txg == 0);
1001 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1003 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1004 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
1005 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1006 ASSERT0(zio->io_queued_timestamp);
1012 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1013 uint64_t size, zio_done_func_t *done, void *private,
1014 zio_priority_t priority, enum zio_flag flags)
1019 if (vd->vdev_children == 0) {
1020 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
1021 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
1022 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1026 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1028 for (c = 0; c < vd->vdev_children; c++)
1029 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1030 offset, size, done, private, priority, flags));
1037 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1038 void *data, int checksum, zio_done_func_t *done, void *private,
1039 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1043 ASSERT(vd->vdev_children == 0);
1044 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1045 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1046 ASSERT3U(offset + size, <=, vd->vdev_psize);
1048 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1049 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1050 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1052 zio->io_prop.zp_checksum = checksum;
1058 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1059 void *data, int checksum, zio_done_func_t *done, void *private,
1060 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1064 ASSERT(vd->vdev_children == 0);
1065 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1066 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1067 ASSERT3U(offset + size, <=, vd->vdev_psize);
1069 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1070 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1071 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1073 zio->io_prop.zp_checksum = checksum;
1075 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1077 * zec checksums are necessarily destructive -- they modify
1078 * the end of the write buffer to hold the verifier/checksum.
1079 * Therefore, we must make a local copy in case the data is
1080 * being written to multiple places in parallel.
1082 void *wbuf = zio_buf_alloc(size);
1083 bcopy(data, wbuf, size);
1084 zio_push_transform(zio, wbuf, size, size, NULL);
1091 * Create a child I/O to do some work for us.
1094 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1095 void *data, uint64_t size, int type, zio_priority_t priority,
1096 enum zio_flag flags, zio_done_func_t *done, void *private)
1098 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1101 ASSERT(vd->vdev_parent ==
1102 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1104 if (type == ZIO_TYPE_READ && bp != NULL) {
1106 * If we have the bp, then the child should perform the
1107 * checksum and the parent need not. This pushes error
1108 * detection as close to the leaves as possible and
1109 * eliminates redundant checksums in the interior nodes.
1111 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1112 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1115 /* Not all IO types require vdev io done stage e.g. free */
1116 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1117 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1119 if (vd->vdev_children == 0)
1120 offset += VDEV_LABEL_START_SIZE;
1122 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1125 * If we've decided to do a repair, the write is not speculative --
1126 * even if the original read was.
1128 if (flags & ZIO_FLAG_IO_REPAIR)
1129 flags &= ~ZIO_FLAG_SPECULATIVE;
1132 * If we're creating a child I/O that is not associated with a
1133 * top-level vdev, then the child zio is not an allocating I/O.
1134 * If this is a retried I/O then we ignore it since we will
1135 * have already processed the original allocating I/O.
1137 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1138 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1139 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1141 ASSERT(mc->mc_alloc_throttle_enabled);
1142 ASSERT(type == ZIO_TYPE_WRITE);
1143 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1144 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1145 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1146 pio->io_child_type == ZIO_CHILD_GANG);
1148 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1151 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1152 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1153 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1154 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1156 zio->io_physdone = pio->io_physdone;
1157 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1158 zio->io_logical->io_phys_children++;
1164 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1165 int type, zio_priority_t priority, enum zio_flag flags,
1166 zio_done_func_t *done, void *private)
1170 ASSERT(vd->vdev_ops->vdev_op_leaf);
1172 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1173 data, size, done, private, type, priority,
1174 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1176 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1182 zio_flush(zio_t *zio, vdev_t *vd)
1184 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1185 NULL, NULL, ZIO_PRIORITY_NOW,
1186 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1190 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1193 ASSERT(vd->vdev_ops->vdev_op_leaf);
1195 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1196 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1197 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1198 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1202 zio_shrink(zio_t *zio, uint64_t size)
1204 ASSERT(zio->io_executor == NULL);
1205 ASSERT(zio->io_orig_size == zio->io_size);
1206 ASSERT(size <= zio->io_size);
1209 * We don't shrink for raidz because of problems with the
1210 * reconstruction when reading back less than the block size.
1211 * Note, BP_IS_RAIDZ() assumes no compression.
1213 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1214 if (!BP_IS_RAIDZ(zio->io_bp))
1215 zio->io_orig_size = zio->io_size = size;
1219 * ==========================================================================
1220 * Prepare to read and write logical blocks
1221 * ==========================================================================
1225 zio_read_bp_init(zio_t *zio)
1227 blkptr_t *bp = zio->io_bp;
1229 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1230 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1231 !(zio->io_flags & ZIO_FLAG_RAW)) {
1233 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1234 void *cbuf = zio_buf_alloc(psize);
1236 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1239 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1240 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1241 decode_embedded_bp_compressed(bp, zio->io_data);
1243 ASSERT(!BP_IS_EMBEDDED(bp));
1246 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1247 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1249 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1250 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1252 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1253 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1255 return (ZIO_PIPELINE_CONTINUE);
1259 zio_write_bp_init(zio_t *zio)
1261 if (!IO_IS_ALLOCATING(zio))
1262 return (ZIO_PIPELINE_CONTINUE);
1264 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1266 if (zio->io_bp_override) {
1267 blkptr_t *bp = zio->io_bp;
1268 zio_prop_t *zp = &zio->io_prop;
1270 ASSERT(bp->blk_birth != zio->io_txg);
1271 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1273 *bp = *zio->io_bp_override;
1274 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1276 if (BP_IS_EMBEDDED(bp))
1277 return (ZIO_PIPELINE_CONTINUE);
1280 * If we've been overridden and nopwrite is set then
1281 * set the flag accordingly to indicate that a nopwrite
1282 * has already occurred.
1284 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1285 ASSERT(!zp->zp_dedup);
1286 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1287 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1288 return (ZIO_PIPELINE_CONTINUE);
1291 ASSERT(!zp->zp_nopwrite);
1293 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1294 return (ZIO_PIPELINE_CONTINUE);
1296 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1297 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1299 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1300 BP_SET_DEDUP(bp, 1);
1301 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1302 return (ZIO_PIPELINE_CONTINUE);
1306 * We were unable to handle this as an override bp, treat
1307 * it as a regular write I/O.
1309 zio->io_bp_override = NULL;
1310 *bp = zio->io_bp_orig;
1311 zio->io_pipeline = zio->io_orig_pipeline;
1314 return (ZIO_PIPELINE_CONTINUE);
1318 zio_write_compress(zio_t *zio)
1320 spa_t *spa = zio->io_spa;
1321 zio_prop_t *zp = &zio->io_prop;
1322 enum zio_compress compress = zp->zp_compress;
1323 blkptr_t *bp = zio->io_bp;
1324 uint64_t lsize = zio->io_size;
1325 uint64_t psize = lsize;
1329 * If our children haven't all reached the ready stage,
1330 * wait for them and then repeat this pipeline stage.
1332 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1333 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1334 return (ZIO_PIPELINE_STOP);
1337 if (!IO_IS_ALLOCATING(zio))
1338 return (ZIO_PIPELINE_CONTINUE);
1340 if (zio->io_children_ready != NULL) {
1342 * Now that all our children are ready, run the callback
1343 * associated with this zio in case it wants to modify the
1344 * data to be written.
1346 ASSERT3U(zp->zp_level, >, 0);
1347 zio->io_children_ready(zio);
1350 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1351 ASSERT(zio->io_bp_override == NULL);
1353 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1355 * We're rewriting an existing block, which means we're
1356 * working on behalf of spa_sync(). For spa_sync() to
1357 * converge, it must eventually be the case that we don't
1358 * have to allocate new blocks. But compression changes
1359 * the blocksize, which forces a reallocate, and makes
1360 * convergence take longer. Therefore, after the first
1361 * few passes, stop compressing to ensure convergence.
1363 pass = spa_sync_pass(spa);
1365 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1366 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1367 ASSERT(!BP_GET_DEDUP(bp));
1369 if (pass >= zfs_sync_pass_dont_compress)
1370 compress = ZIO_COMPRESS_OFF;
1372 /* Make sure someone doesn't change their mind on overwrites */
1373 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1374 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1377 if (compress != ZIO_COMPRESS_OFF) {
1378 void *cbuf = zio_buf_alloc(lsize);
1379 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1380 if (psize == 0 || psize == lsize) {
1381 compress = ZIO_COMPRESS_OFF;
1382 zio_buf_free(cbuf, lsize);
1383 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1384 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1385 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1386 encode_embedded_bp_compressed(bp,
1387 cbuf, compress, lsize, psize);
1388 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1389 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1390 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1391 zio_buf_free(cbuf, lsize);
1392 bp->blk_birth = zio->io_txg;
1393 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1394 ASSERT(spa_feature_is_active(spa,
1395 SPA_FEATURE_EMBEDDED_DATA));
1396 return (ZIO_PIPELINE_CONTINUE);
1399 * Round up compressed size up to the ashift
1400 * of the smallest-ashift device, and zero the tail.
1401 * This ensures that the compressed size of the BP
1402 * (and thus compressratio property) are correct,
1403 * in that we charge for the padding used to fill out
1406 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1407 size_t rounded = (size_t)P2ROUNDUP(psize,
1408 1ULL << spa->spa_min_ashift);
1409 if (rounded >= lsize) {
1410 compress = ZIO_COMPRESS_OFF;
1411 zio_buf_free(cbuf, lsize);
1414 bzero((char *)cbuf + psize, rounded - psize);
1416 zio_push_transform(zio, cbuf,
1417 psize, lsize, NULL);
1422 * We were unable to handle this as an override bp, treat
1423 * it as a regular write I/O.
1425 zio->io_bp_override = NULL;
1426 *bp = zio->io_bp_orig;
1427 zio->io_pipeline = zio->io_orig_pipeline;
1431 * The final pass of spa_sync() must be all rewrites, but the first
1432 * few passes offer a trade-off: allocating blocks defers convergence,
1433 * but newly allocated blocks are sequential, so they can be written
1434 * to disk faster. Therefore, we allow the first few passes of
1435 * spa_sync() to allocate new blocks, but force rewrites after that.
1436 * There should only be a handful of blocks after pass 1 in any case.
1438 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1439 BP_GET_PSIZE(bp) == psize &&
1440 pass >= zfs_sync_pass_rewrite) {
1442 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1443 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1444 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1447 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1451 if (zio->io_bp_orig.blk_birth != 0 &&
1452 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1453 BP_SET_LSIZE(bp, lsize);
1454 BP_SET_TYPE(bp, zp->zp_type);
1455 BP_SET_LEVEL(bp, zp->zp_level);
1456 BP_SET_BIRTH(bp, zio->io_txg, 0);
1458 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1460 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1461 BP_SET_LSIZE(bp, lsize);
1462 BP_SET_TYPE(bp, zp->zp_type);
1463 BP_SET_LEVEL(bp, zp->zp_level);
1464 BP_SET_PSIZE(bp, psize);
1465 BP_SET_COMPRESS(bp, compress);
1466 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1467 BP_SET_DEDUP(bp, zp->zp_dedup);
1468 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1470 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1471 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1472 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1474 if (zp->zp_nopwrite) {
1475 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1477 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1480 return (ZIO_PIPELINE_CONTINUE);
1484 zio_free_bp_init(zio_t *zio)
1486 blkptr_t *bp = zio->io_bp;
1488 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1489 if (BP_GET_DEDUP(bp))
1490 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1493 return (ZIO_PIPELINE_CONTINUE);
1497 * ==========================================================================
1498 * Execute the I/O pipeline
1499 * ==========================================================================
1503 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1505 spa_t *spa = zio->io_spa;
1506 zio_type_t t = zio->io_type;
1507 int flags = (cutinline ? TQ_FRONT : 0);
1509 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1512 * If we're a config writer or a probe, the normal issue and
1513 * interrupt threads may all be blocked waiting for the config lock.
1514 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1516 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1520 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1522 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1526 * If this is a high priority I/O, then use the high priority taskq if
1529 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1530 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1533 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1536 * NB: We are assuming that the zio can only be dispatched
1537 * to a single taskq at a time. It would be a grievous error
1538 * to dispatch the zio to another taskq at the same time.
1540 #if defined(illumos) || !defined(_KERNEL)
1541 ASSERT(zio->io_tqent.tqent_next == NULL);
1543 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1545 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1546 flags, &zio->io_tqent);
1550 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1552 kthread_t *executor = zio->io_executor;
1553 spa_t *spa = zio->io_spa;
1555 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1556 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1558 for (i = 0; i < tqs->stqs_count; i++) {
1559 if (taskq_member(tqs->stqs_taskq[i], executor))
1568 zio_issue_async(zio_t *zio)
1570 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1572 return (ZIO_PIPELINE_STOP);
1576 zio_interrupt(zio_t *zio)
1578 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1582 zio_delay_interrupt(zio_t *zio)
1585 * The timeout_generic() function isn't defined in userspace, so
1586 * rather than trying to implement the function, the zio delay
1587 * functionality has been disabled for userspace builds.
1592 * If io_target_timestamp is zero, then no delay has been registered
1593 * for this IO, thus jump to the end of this function and "skip" the
1594 * delay; issuing it directly to the zio layer.
1596 if (zio->io_target_timestamp != 0) {
1597 hrtime_t now = gethrtime();
1599 if (now >= zio->io_target_timestamp) {
1601 * This IO has already taken longer than the target
1602 * delay to complete, so we don't want to delay it
1603 * any longer; we "miss" the delay and issue it
1604 * directly to the zio layer. This is likely due to
1605 * the target latency being set to a value less than
1606 * the underlying hardware can satisfy (e.g. delay
1607 * set to 1ms, but the disks take 10ms to complete an
1611 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1616 hrtime_t diff = zio->io_target_timestamp - now;
1618 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1619 hrtime_t, now, hrtime_t, diff);
1621 (void) timeout_generic(CALLOUT_NORMAL,
1622 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1629 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1634 * Execute the I/O pipeline until one of the following occurs:
1636 * (1) the I/O completes
1637 * (2) the pipeline stalls waiting for dependent child I/Os
1638 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1639 * (4) the I/O is delegated by vdev-level caching or aggregation
1640 * (5) the I/O is deferred due to vdev-level queueing
1641 * (6) the I/O is handed off to another thread.
1643 * In all cases, the pipeline stops whenever there's no CPU work; it never
1644 * burns a thread in cv_wait().
1646 * There's no locking on io_stage because there's no legitimate way
1647 * for multiple threads to be attempting to process the same I/O.
1649 static zio_pipe_stage_t *zio_pipeline[];
1652 zio_execute(zio_t *zio)
1654 zio->io_executor = curthread;
1656 ASSERT3U(zio->io_queued_timestamp, >, 0);
1658 while (zio->io_stage < ZIO_STAGE_DONE) {
1659 enum zio_stage pipeline = zio->io_pipeline;
1660 enum zio_stage stage = zio->io_stage;
1663 ASSERT(!MUTEX_HELD(&zio->io_lock));
1664 ASSERT(ISP2(stage));
1665 ASSERT(zio->io_stall == NULL);
1669 } while ((stage & pipeline) == 0);
1671 ASSERT(stage <= ZIO_STAGE_DONE);
1674 * If we are in interrupt context and this pipeline stage
1675 * will grab a config lock that is held across I/O,
1676 * or may wait for an I/O that needs an interrupt thread
1677 * to complete, issue async to avoid deadlock.
1679 * For VDEV_IO_START, we cut in line so that the io will
1680 * be sent to disk promptly.
1682 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1683 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1684 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1685 zio_requeue_io_start_cut_in_line : B_FALSE;
1686 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1690 zio->io_stage = stage;
1691 zio->io_pipeline_trace |= zio->io_stage;
1692 rv = zio_pipeline[highbit64(stage) - 1](zio);
1694 if (rv == ZIO_PIPELINE_STOP)
1697 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1702 * ==========================================================================
1703 * Initiate I/O, either sync or async
1704 * ==========================================================================
1707 zio_wait(zio_t *zio)
1711 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1712 ASSERT(zio->io_executor == NULL);
1714 zio->io_waiter = curthread;
1715 ASSERT0(zio->io_queued_timestamp);
1716 zio->io_queued_timestamp = gethrtime();
1720 mutex_enter(&zio->io_lock);
1721 while (zio->io_executor != NULL)
1722 cv_wait(&zio->io_cv, &zio->io_lock);
1723 mutex_exit(&zio->io_lock);
1725 error = zio->io_error;
1732 zio_nowait(zio_t *zio)
1734 ASSERT(zio->io_executor == NULL);
1736 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1737 zio_unique_parent(zio) == NULL) {
1739 * This is a logical async I/O with no parent to wait for it.
1740 * We add it to the spa_async_root_zio "Godfather" I/O which
1741 * will ensure they complete prior to unloading the pool.
1743 spa_t *spa = zio->io_spa;
1745 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1748 ASSERT0(zio->io_queued_timestamp);
1749 zio->io_queued_timestamp = gethrtime();
1754 * ==========================================================================
1755 * Reexecute or suspend/resume failed I/O
1756 * ==========================================================================
1760 zio_reexecute(zio_t *pio)
1762 zio_t *cio, *cio_next;
1764 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1765 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1766 ASSERT(pio->io_gang_leader == NULL);
1767 ASSERT(pio->io_gang_tree == NULL);
1769 pio->io_flags = pio->io_orig_flags;
1770 pio->io_stage = pio->io_orig_stage;
1771 pio->io_pipeline = pio->io_orig_pipeline;
1772 pio->io_reexecute = 0;
1773 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1774 pio->io_pipeline_trace = 0;
1776 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1777 pio->io_state[w] = 0;
1778 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1779 pio->io_child_error[c] = 0;
1781 if (IO_IS_ALLOCATING(pio))
1782 BP_ZERO(pio->io_bp);
1785 * As we reexecute pio's children, new children could be created.
1786 * New children go to the head of pio's io_child_list, however,
1787 * so we will (correctly) not reexecute them. The key is that
1788 * the remainder of pio's io_child_list, from 'cio_next' onward,
1789 * cannot be affected by any side effects of reexecuting 'cio'.
1791 zio_link_t *zl = NULL;
1792 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1793 cio_next = zio_walk_children(pio, &zl);
1794 mutex_enter(&pio->io_lock);
1795 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1796 pio->io_children[cio->io_child_type][w]++;
1797 mutex_exit(&pio->io_lock);
1802 * Now that all children have been reexecuted, execute the parent.
1803 * We don't reexecute "The Godfather" I/O here as it's the
1804 * responsibility of the caller to wait on him.
1806 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1807 pio->io_queued_timestamp = gethrtime();
1813 zio_suspend(spa_t *spa, zio_t *zio)
1815 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1816 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1817 "failure and the failure mode property for this pool "
1818 "is set to panic.", spa_name(spa));
1820 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1822 mutex_enter(&spa->spa_suspend_lock);
1824 if (spa->spa_suspend_zio_root == NULL)
1825 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1826 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1827 ZIO_FLAG_GODFATHER);
1829 spa->spa_suspended = B_TRUE;
1832 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1833 ASSERT(zio != spa->spa_suspend_zio_root);
1834 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1835 ASSERT(zio_unique_parent(zio) == NULL);
1836 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1837 zio_add_child(spa->spa_suspend_zio_root, zio);
1840 mutex_exit(&spa->spa_suspend_lock);
1844 zio_resume(spa_t *spa)
1849 * Reexecute all previously suspended i/o.
1851 mutex_enter(&spa->spa_suspend_lock);
1852 spa->spa_suspended = B_FALSE;
1853 cv_broadcast(&spa->spa_suspend_cv);
1854 pio = spa->spa_suspend_zio_root;
1855 spa->spa_suspend_zio_root = NULL;
1856 mutex_exit(&spa->spa_suspend_lock);
1862 return (zio_wait(pio));
1866 zio_resume_wait(spa_t *spa)
1868 mutex_enter(&spa->spa_suspend_lock);
1869 while (spa_suspended(spa))
1870 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1871 mutex_exit(&spa->spa_suspend_lock);
1875 * ==========================================================================
1878 * A gang block is a collection of small blocks that looks to the DMU
1879 * like one large block. When zio_dva_allocate() cannot find a block
1880 * of the requested size, due to either severe fragmentation or the pool
1881 * being nearly full, it calls zio_write_gang_block() to construct the
1882 * block from smaller fragments.
1884 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1885 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1886 * an indirect block: it's an array of block pointers. It consumes
1887 * only one sector and hence is allocatable regardless of fragmentation.
1888 * The gang header's bps point to its gang members, which hold the data.
1890 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1891 * as the verifier to ensure uniqueness of the SHA256 checksum.
1892 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1893 * not the gang header. This ensures that data block signatures (needed for
1894 * deduplication) are independent of how the block is physically stored.
1896 * Gang blocks can be nested: a gang member may itself be a gang block.
1897 * Thus every gang block is a tree in which root and all interior nodes are
1898 * gang headers, and the leaves are normal blocks that contain user data.
1899 * The root of the gang tree is called the gang leader.
1901 * To perform any operation (read, rewrite, free, claim) on a gang block,
1902 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1903 * in the io_gang_tree field of the original logical i/o by recursively
1904 * reading the gang leader and all gang headers below it. This yields
1905 * an in-core tree containing the contents of every gang header and the
1906 * bps for every constituent of the gang block.
1908 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1909 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1910 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1911 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1912 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1913 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1914 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1915 * of the gang header plus zio_checksum_compute() of the data to update the
1916 * gang header's blk_cksum as described above.
1918 * The two-phase assemble/issue model solves the problem of partial failure --
1919 * what if you'd freed part of a gang block but then couldn't read the
1920 * gang header for another part? Assembling the entire gang tree first
1921 * ensures that all the necessary gang header I/O has succeeded before
1922 * starting the actual work of free, claim, or write. Once the gang tree
1923 * is assembled, free and claim are in-memory operations that cannot fail.
1925 * In the event that a gang write fails, zio_dva_unallocate() walks the
1926 * gang tree to immediately free (i.e. insert back into the space map)
1927 * everything we've allocated. This ensures that we don't get ENOSPC
1928 * errors during repeated suspend/resume cycles due to a flaky device.
1930 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1931 * the gang tree, we won't modify the block, so we can safely defer the free
1932 * (knowing that the block is still intact). If we *can* assemble the gang
1933 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1934 * each constituent bp and we can allocate a new block on the next sync pass.
1936 * In all cases, the gang tree allows complete recovery from partial failure.
1937 * ==========================================================================
1941 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1946 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1947 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1948 &pio->io_bookmark));
1952 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1957 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1958 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1959 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1961 * As we rewrite each gang header, the pipeline will compute
1962 * a new gang block header checksum for it; but no one will
1963 * compute a new data checksum, so we do that here. The one
1964 * exception is the gang leader: the pipeline already computed
1965 * its data checksum because that stage precedes gang assembly.
1966 * (Presently, nothing actually uses interior data checksums;
1967 * this is just good hygiene.)
1969 if (gn != pio->io_gang_leader->io_gang_tree) {
1970 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1971 data, BP_GET_PSIZE(bp));
1974 * If we are here to damage data for testing purposes,
1975 * leave the GBH alone so that we can detect the damage.
1977 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1978 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1980 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1981 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1982 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1990 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1992 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1993 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1994 ZIO_GANG_CHILD_FLAGS(pio)));
1999 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
2001 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2002 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2005 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2014 static void zio_gang_tree_assemble_done(zio_t *zio);
2016 static zio_gang_node_t *
2017 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2019 zio_gang_node_t *gn;
2021 ASSERT(*gnpp == NULL);
2023 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2024 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2031 zio_gang_node_free(zio_gang_node_t **gnpp)
2033 zio_gang_node_t *gn = *gnpp;
2035 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2036 ASSERT(gn->gn_child[g] == NULL);
2038 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2039 kmem_free(gn, sizeof (*gn));
2044 zio_gang_tree_free(zio_gang_node_t **gnpp)
2046 zio_gang_node_t *gn = *gnpp;
2051 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2052 zio_gang_tree_free(&gn->gn_child[g]);
2054 zio_gang_node_free(gnpp);
2058 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2060 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2062 ASSERT(gio->io_gang_leader == gio);
2063 ASSERT(BP_IS_GANG(bp));
2065 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
2066 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2067 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2071 zio_gang_tree_assemble_done(zio_t *zio)
2073 zio_t *gio = zio->io_gang_leader;
2074 zio_gang_node_t *gn = zio->io_private;
2075 blkptr_t *bp = zio->io_bp;
2077 ASSERT(gio == zio_unique_parent(zio));
2078 ASSERT(zio->io_child_count == 0);
2083 if (BP_SHOULD_BYTESWAP(bp))
2084 byteswap_uint64_array(zio->io_data, zio->io_size);
2086 ASSERT(zio->io_data == gn->gn_gbh);
2087 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2088 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2090 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2091 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2092 if (!BP_IS_GANG(gbp))
2094 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2099 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2101 zio_t *gio = pio->io_gang_leader;
2104 ASSERT(BP_IS_GANG(bp) == !!gn);
2105 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2106 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2109 * If you're a gang header, your data is in gn->gn_gbh.
2110 * If you're a gang member, your data is in 'data' and gn == NULL.
2112 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2115 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2117 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2118 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2119 if (BP_IS_HOLE(gbp))
2121 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
2122 data = (char *)data + BP_GET_PSIZE(gbp);
2126 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2127 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2134 zio_gang_assemble(zio_t *zio)
2136 blkptr_t *bp = zio->io_bp;
2138 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2139 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2141 zio->io_gang_leader = zio;
2143 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2145 return (ZIO_PIPELINE_CONTINUE);
2149 zio_gang_issue(zio_t *zio)
2151 blkptr_t *bp = zio->io_bp;
2153 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2154 return (ZIO_PIPELINE_STOP);
2157 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2158 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2160 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2161 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2163 zio_gang_tree_free(&zio->io_gang_tree);
2165 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2167 return (ZIO_PIPELINE_CONTINUE);
2171 zio_write_gang_member_ready(zio_t *zio)
2173 zio_t *pio = zio_unique_parent(zio);
2174 zio_t *gio = zio->io_gang_leader;
2175 dva_t *cdva = zio->io_bp->blk_dva;
2176 dva_t *pdva = pio->io_bp->blk_dva;
2179 if (BP_IS_HOLE(zio->io_bp))
2182 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2184 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2185 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2186 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2187 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2188 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2190 mutex_enter(&pio->io_lock);
2191 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2192 ASSERT(DVA_GET_GANG(&pdva[d]));
2193 asize = DVA_GET_ASIZE(&pdva[d]);
2194 asize += DVA_GET_ASIZE(&cdva[d]);
2195 DVA_SET_ASIZE(&pdva[d], asize);
2197 mutex_exit(&pio->io_lock);
2201 zio_write_gang_block(zio_t *pio)
2203 spa_t *spa = pio->io_spa;
2204 metaslab_class_t *mc = spa_normal_class(spa);
2205 blkptr_t *bp = pio->io_bp;
2206 zio_t *gio = pio->io_gang_leader;
2208 zio_gang_node_t *gn, **gnpp;
2209 zio_gbh_phys_t *gbh;
2210 uint64_t txg = pio->io_txg;
2211 uint64_t resid = pio->io_size;
2213 int copies = gio->io_prop.zp_copies;
2214 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2218 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2219 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2220 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2221 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2223 flags |= METASLAB_ASYNC_ALLOC;
2224 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2227 * The logical zio has already placed a reservation for
2228 * 'copies' allocation slots but gang blocks may require
2229 * additional copies. These additional copies
2230 * (i.e. gbh_copies - copies) are guaranteed to succeed
2231 * since metaslab_class_throttle_reserve() always allows
2232 * additional reservations for gang blocks.
2234 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2238 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2239 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio);
2241 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2242 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2243 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2246 * If we failed to allocate the gang block header then
2247 * we remove any additional allocation reservations that
2248 * we placed here. The original reservation will
2249 * be removed when the logical I/O goes to the ready
2252 metaslab_class_throttle_unreserve(mc,
2253 gbh_copies - copies, pio);
2255 pio->io_error = error;
2256 return (ZIO_PIPELINE_CONTINUE);
2260 gnpp = &gio->io_gang_tree;
2262 gnpp = pio->io_private;
2263 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2266 gn = zio_gang_node_alloc(gnpp);
2268 bzero(gbh, SPA_GANGBLOCKSIZE);
2271 * Create the gang header.
2273 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2274 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2277 * Create and nowait the gang children.
2279 for (int g = 0; resid != 0; resid -= lsize, g++) {
2280 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2282 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2284 zp.zp_checksum = gio->io_prop.zp_checksum;
2285 zp.zp_compress = ZIO_COMPRESS_OFF;
2286 zp.zp_type = DMU_OT_NONE;
2288 zp.zp_copies = gio->io_prop.zp_copies;
2289 zp.zp_dedup = B_FALSE;
2290 zp.zp_dedup_verify = B_FALSE;
2291 zp.zp_nopwrite = B_FALSE;
2293 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2294 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2295 zio_write_gang_member_ready, NULL, NULL, NULL,
2296 &gn->gn_child[g], pio->io_priority,
2297 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2299 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2300 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2301 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2304 * Gang children won't throttle but we should
2305 * account for their work, so reserve an allocation
2306 * slot for them here.
2308 VERIFY(metaslab_class_throttle_reserve(mc,
2309 zp.zp_copies, cio, flags));
2315 * Set pio's pipeline to just wait for zio to finish.
2317 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2321 return (ZIO_PIPELINE_CONTINUE);
2325 * The zio_nop_write stage in the pipeline determines if allocating a
2326 * new bp is necessary. The nopwrite feature can handle writes in
2327 * either syncing or open context (i.e. zil writes) and as a result is
2328 * mutually exclusive with dedup.
2330 * By leveraging a cryptographically secure checksum, such as SHA256, we
2331 * can compare the checksums of the new data and the old to determine if
2332 * allocating a new block is required. Note that our requirements for
2333 * cryptographic strength are fairly weak: there can't be any accidental
2334 * hash collisions, but we don't need to be secure against intentional
2335 * (malicious) collisions. To trigger a nopwrite, you have to be able
2336 * to write the file to begin with, and triggering an incorrect (hash
2337 * collision) nopwrite is no worse than simply writing to the file.
2338 * That said, there are no known attacks against the checksum algorithms
2339 * used for nopwrite, assuming that the salt and the checksums
2340 * themselves remain secret.
2343 zio_nop_write(zio_t *zio)
2345 blkptr_t *bp = zio->io_bp;
2346 blkptr_t *bp_orig = &zio->io_bp_orig;
2347 zio_prop_t *zp = &zio->io_prop;
2349 ASSERT(BP_GET_LEVEL(bp) == 0);
2350 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2351 ASSERT(zp->zp_nopwrite);
2352 ASSERT(!zp->zp_dedup);
2353 ASSERT(zio->io_bp_override == NULL);
2354 ASSERT(IO_IS_ALLOCATING(zio));
2357 * Check to see if the original bp and the new bp have matching
2358 * characteristics (i.e. same checksum, compression algorithms, etc).
2359 * If they don't then just continue with the pipeline which will
2360 * allocate a new bp.
2362 if (BP_IS_HOLE(bp_orig) ||
2363 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2364 ZCHECKSUM_FLAG_NOPWRITE) ||
2365 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2366 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2367 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2368 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2369 return (ZIO_PIPELINE_CONTINUE);
2372 * If the checksums match then reset the pipeline so that we
2373 * avoid allocating a new bp and issuing any I/O.
2375 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2376 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2377 ZCHECKSUM_FLAG_NOPWRITE);
2378 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2379 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2380 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2381 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2382 sizeof (uint64_t)) == 0);
2385 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2386 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2389 return (ZIO_PIPELINE_CONTINUE);
2393 * ==========================================================================
2395 * ==========================================================================
2398 zio_ddt_child_read_done(zio_t *zio)
2400 blkptr_t *bp = zio->io_bp;
2401 ddt_entry_t *dde = zio->io_private;
2403 zio_t *pio = zio_unique_parent(zio);
2405 mutex_enter(&pio->io_lock);
2406 ddp = ddt_phys_select(dde, bp);
2407 if (zio->io_error == 0)
2408 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2409 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2410 dde->dde_repair_data = zio->io_data;
2412 zio_buf_free(zio->io_data, zio->io_size);
2413 mutex_exit(&pio->io_lock);
2417 zio_ddt_read_start(zio_t *zio)
2419 blkptr_t *bp = zio->io_bp;
2421 ASSERT(BP_GET_DEDUP(bp));
2422 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2423 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2425 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2426 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2427 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2428 ddt_phys_t *ddp = dde->dde_phys;
2429 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2432 ASSERT(zio->io_vsd == NULL);
2435 if (ddp_self == NULL)
2436 return (ZIO_PIPELINE_CONTINUE);
2438 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2439 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2441 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2443 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2444 zio_buf_alloc(zio->io_size), zio->io_size,
2445 zio_ddt_child_read_done, dde, zio->io_priority,
2446 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2447 &zio->io_bookmark));
2449 return (ZIO_PIPELINE_CONTINUE);
2452 zio_nowait(zio_read(zio, zio->io_spa, bp,
2453 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2454 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2456 return (ZIO_PIPELINE_CONTINUE);
2460 zio_ddt_read_done(zio_t *zio)
2462 blkptr_t *bp = zio->io_bp;
2464 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2465 return (ZIO_PIPELINE_STOP);
2468 ASSERT(BP_GET_DEDUP(bp));
2469 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2470 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2472 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2473 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2474 ddt_entry_t *dde = zio->io_vsd;
2476 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2477 return (ZIO_PIPELINE_CONTINUE);
2480 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2481 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2482 return (ZIO_PIPELINE_STOP);
2484 if (dde->dde_repair_data != NULL) {
2485 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2486 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2488 ddt_repair_done(ddt, dde);
2492 ASSERT(zio->io_vsd == NULL);
2494 return (ZIO_PIPELINE_CONTINUE);
2498 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2500 spa_t *spa = zio->io_spa;
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 blkptr_t blk = *zio->io_bp;
2527 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2531 error = arc_read(NULL, spa, &blk,
2532 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2533 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2534 &aflags, &zio->io_bookmark);
2537 if (arc_buf_size(abuf) != zio->io_orig_size ||
2538 bcmp(abuf->b_data, zio->io_orig_data,
2539 zio->io_orig_size) != 0)
2540 error = SET_ERROR(EEXIST);
2541 arc_buf_destroy(abuf, &abuf);
2545 return (error != 0);
2553 zio_ddt_child_write_ready(zio_t *zio)
2555 int p = zio->io_prop.zp_copies;
2556 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2557 ddt_entry_t *dde = zio->io_private;
2558 ddt_phys_t *ddp = &dde->dde_phys[p];
2566 ASSERT(dde->dde_lead_zio[p] == zio);
2568 ddt_phys_fill(ddp, zio->io_bp);
2570 zio_link_t *zl = NULL;
2571 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2572 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2578 zio_ddt_child_write_done(zio_t *zio)
2580 int p = zio->io_prop.zp_copies;
2581 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2582 ddt_entry_t *dde = zio->io_private;
2583 ddt_phys_t *ddp = &dde->dde_phys[p];
2587 ASSERT(ddp->ddp_refcnt == 0);
2588 ASSERT(dde->dde_lead_zio[p] == zio);
2589 dde->dde_lead_zio[p] = NULL;
2591 if (zio->io_error == 0) {
2592 zio_link_t *zl = NULL;
2593 while (zio_walk_parents(zio, &zl) != NULL)
2594 ddt_phys_addref(ddp);
2596 ddt_phys_clear(ddp);
2603 zio_ddt_ditto_write_done(zio_t *zio)
2605 int p = DDT_PHYS_DITTO;
2606 zio_prop_t *zp = &zio->io_prop;
2607 blkptr_t *bp = zio->io_bp;
2608 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2609 ddt_entry_t *dde = zio->io_private;
2610 ddt_phys_t *ddp = &dde->dde_phys[p];
2611 ddt_key_t *ddk = &dde->dde_key;
2615 ASSERT(ddp->ddp_refcnt == 0);
2616 ASSERT(dde->dde_lead_zio[p] == zio);
2617 dde->dde_lead_zio[p] = NULL;
2619 if (zio->io_error == 0) {
2620 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2621 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2622 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2623 if (ddp->ddp_phys_birth != 0)
2624 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2625 ddt_phys_fill(ddp, bp);
2632 zio_ddt_write(zio_t *zio)
2634 spa_t *spa = zio->io_spa;
2635 blkptr_t *bp = zio->io_bp;
2636 uint64_t txg = zio->io_txg;
2637 zio_prop_t *zp = &zio->io_prop;
2638 int p = zp->zp_copies;
2642 ddt_t *ddt = ddt_select(spa, bp);
2646 ASSERT(BP_GET_DEDUP(bp));
2647 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2648 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2651 dde = ddt_lookup(ddt, bp, B_TRUE);
2652 ddp = &dde->dde_phys[p];
2654 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2656 * If we're using a weak checksum, upgrade to a strong checksum
2657 * and try again. If we're already using a strong checksum,
2658 * we can't resolve it, so just convert to an ordinary write.
2659 * (And automatically e-mail a paper to Nature?)
2661 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2662 ZCHECKSUM_FLAG_DEDUP)) {
2663 zp->zp_checksum = spa_dedup_checksum(spa);
2664 zio_pop_transforms(zio);
2665 zio->io_stage = ZIO_STAGE_OPEN;
2668 zp->zp_dedup = B_FALSE;
2670 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2672 return (ZIO_PIPELINE_CONTINUE);
2675 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2676 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2678 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2679 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2680 zio_prop_t czp = *zp;
2682 czp.zp_copies = ditto_copies;
2685 * If we arrived here with an override bp, we won't have run
2686 * the transform stack, so we won't have the data we need to
2687 * generate a child i/o. So, toss the override bp and restart.
2688 * This is safe, because using the override bp is just an
2689 * optimization; and it's rare, so the cost doesn't matter.
2691 if (zio->io_bp_override) {
2692 zio_pop_transforms(zio);
2693 zio->io_stage = ZIO_STAGE_OPEN;
2694 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2695 zio->io_bp_override = NULL;
2698 return (ZIO_PIPELINE_CONTINUE);
2701 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2702 zio->io_orig_size, &czp, NULL, NULL,
2703 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2704 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2706 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2707 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2710 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2711 if (ddp->ddp_phys_birth != 0)
2712 ddt_bp_fill(ddp, bp, txg);
2713 if (dde->dde_lead_zio[p] != NULL)
2714 zio_add_child(zio, dde->dde_lead_zio[p]);
2716 ddt_phys_addref(ddp);
2717 } else if (zio->io_bp_override) {
2718 ASSERT(bp->blk_birth == txg);
2719 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2720 ddt_phys_fill(ddp, bp);
2721 ddt_phys_addref(ddp);
2723 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2724 zio->io_orig_size, zp,
2725 zio_ddt_child_write_ready, NULL, NULL,
2726 zio_ddt_child_write_done, dde, zio->io_priority,
2727 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2729 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2730 dde->dde_lead_zio[p] = cio;
2740 return (ZIO_PIPELINE_CONTINUE);
2743 ddt_entry_t *freedde; /* for debugging */
2746 zio_ddt_free(zio_t *zio)
2748 spa_t *spa = zio->io_spa;
2749 blkptr_t *bp = zio->io_bp;
2750 ddt_t *ddt = ddt_select(spa, bp);
2754 ASSERT(BP_GET_DEDUP(bp));
2755 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2758 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2759 ddp = ddt_phys_select(dde, bp);
2760 ddt_phys_decref(ddp);
2763 return (ZIO_PIPELINE_CONTINUE);
2767 * ==========================================================================
2768 * Allocate and free blocks
2769 * ==========================================================================
2773 zio_io_to_allocate(spa_t *spa)
2777 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2779 zio = avl_first(&spa->spa_alloc_tree);
2783 ASSERT(IO_IS_ALLOCATING(zio));
2786 * Try to place a reservation for this zio. If we're unable to
2787 * reserve then we throttle.
2789 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2790 zio->io_prop.zp_copies, zio, 0)) {
2794 avl_remove(&spa->spa_alloc_tree, zio);
2795 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2801 zio_dva_throttle(zio_t *zio)
2803 spa_t *spa = zio->io_spa;
2806 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2807 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2808 zio->io_child_type == ZIO_CHILD_GANG ||
2809 zio->io_flags & ZIO_FLAG_NODATA) {
2810 return (ZIO_PIPELINE_CONTINUE);
2813 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2815 ASSERT3U(zio->io_queued_timestamp, >, 0);
2816 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2818 mutex_enter(&spa->spa_alloc_lock);
2820 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2821 avl_add(&spa->spa_alloc_tree, zio);
2823 nio = zio_io_to_allocate(zio->io_spa);
2824 mutex_exit(&spa->spa_alloc_lock);
2827 return (ZIO_PIPELINE_CONTINUE);
2830 ASSERT3U(nio->io_queued_timestamp, <=,
2831 zio->io_queued_timestamp);
2832 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2834 * We are passing control to a new zio so make sure that
2835 * it is processed by a different thread. We do this to
2836 * avoid stack overflows that can occur when parents are
2837 * throttled and children are making progress. We allow
2838 * it to go to the head of the taskq since it's already
2841 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2843 return (ZIO_PIPELINE_STOP);
2847 zio_allocate_dispatch(spa_t *spa)
2851 mutex_enter(&spa->spa_alloc_lock);
2852 zio = zio_io_to_allocate(spa);
2853 mutex_exit(&spa->spa_alloc_lock);
2857 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2858 ASSERT0(zio->io_error);
2859 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2863 zio_dva_allocate(zio_t *zio)
2865 spa_t *spa = zio->io_spa;
2866 metaslab_class_t *mc = spa_normal_class(spa);
2867 blkptr_t *bp = zio->io_bp;
2871 if (zio->io_gang_leader == NULL) {
2872 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2873 zio->io_gang_leader = zio;
2876 ASSERT(BP_IS_HOLE(bp));
2877 ASSERT0(BP_GET_NDVAS(bp));
2878 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2879 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2880 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2882 if (zio->io_flags & ZIO_FLAG_NODATA) {
2883 flags |= METASLAB_DONT_THROTTLE;
2885 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2886 flags |= METASLAB_GANG_CHILD;
2888 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2889 flags |= METASLAB_ASYNC_ALLOC;
2892 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2893 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio);
2896 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2897 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2899 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2900 return (zio_write_gang_block(zio));
2901 zio->io_error = error;
2904 return (ZIO_PIPELINE_CONTINUE);
2908 zio_dva_free(zio_t *zio)
2910 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2912 return (ZIO_PIPELINE_CONTINUE);
2916 zio_dva_claim(zio_t *zio)
2920 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2922 zio->io_error = error;
2924 return (ZIO_PIPELINE_CONTINUE);
2928 * Undo an allocation. This is used by zio_done() when an I/O fails
2929 * and we want to give back the block we just allocated.
2930 * This handles both normal blocks and gang blocks.
2933 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2935 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2936 ASSERT(zio->io_bp_override == NULL);
2938 if (!BP_IS_HOLE(bp))
2939 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2942 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2943 zio_dva_unallocate(zio, gn->gn_child[g],
2944 &gn->gn_gbh->zg_blkptr[g]);
2950 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2953 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2954 uint64_t size, boolean_t *slog)
2958 ASSERT(txg > spa_syncing_txg(spa));
2960 error = metaslab_alloc(spa, spa_log_class(spa), size,
2961 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2965 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2966 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2972 BP_SET_LSIZE(new_bp, size);
2973 BP_SET_PSIZE(new_bp, size);
2974 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2975 BP_SET_CHECKSUM(new_bp,
2976 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2977 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2978 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2979 BP_SET_LEVEL(new_bp, 0);
2980 BP_SET_DEDUP(new_bp, 0);
2981 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2988 * Free an intent log block.
2991 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2993 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2994 ASSERT(!BP_IS_GANG(bp));
2996 zio_free(spa, txg, bp);
3000 * ==========================================================================
3001 * Read, write and delete to physical devices
3002 * ==========================================================================
3007 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3008 * stops after this stage and will resume upon I/O completion.
3009 * However, there are instances where the vdev layer may need to
3010 * continue the pipeline when an I/O was not issued. Since the I/O
3011 * that was sent to the vdev layer might be different than the one
3012 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3013 * force the underlying vdev layers to call either zio_execute() or
3014 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3017 zio_vdev_io_start(zio_t *zio)
3019 vdev_t *vd = zio->io_vd;
3021 spa_t *spa = zio->io_spa;
3024 ASSERT(zio->io_error == 0);
3025 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3028 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3029 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3032 * The mirror_ops handle multiple DVAs in a single BP.
3034 vdev_mirror_ops.vdev_op_io_start(zio);
3035 return (ZIO_PIPELINE_STOP);
3038 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3039 zio->io_priority == ZIO_PRIORITY_NOW) {
3040 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3041 return (ZIO_PIPELINE_CONTINUE);
3044 ASSERT3P(zio->io_logical, !=, zio);
3047 * We keep track of time-sensitive I/Os so that the scan thread
3048 * can quickly react to certain workloads. In particular, we care
3049 * about non-scrubbing, top-level reads and writes with the following
3051 * - synchronous writes of user data to non-slog devices
3052 * - any reads of user data
3053 * When these conditions are met, adjust the timestamp of spa_last_io
3054 * which allows the scan thread to adjust its workload accordingly.
3056 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3057 vd == vd->vdev_top && !vd->vdev_islog &&
3058 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3059 zio->io_txg != spa_syncing_txg(spa)) {
3060 uint64_t old = spa->spa_last_io;
3061 uint64_t new = ddi_get_lbolt64();
3063 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3066 align = 1ULL << vd->vdev_top->vdev_ashift;
3068 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3069 P2PHASE(zio->io_size, align) != 0) {
3070 /* Transform logical writes to be a full physical block size. */
3071 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3073 if (zio->io_type == ZIO_TYPE_READ ||
3074 zio->io_type == ZIO_TYPE_WRITE)
3075 abuf = zio_buf_alloc(asize);
3076 ASSERT(vd == vd->vdev_top);
3077 if (zio->io_type == ZIO_TYPE_WRITE) {
3078 bcopy(zio->io_data, abuf, zio->io_size);
3079 bzero(abuf + zio->io_size, asize - zio->io_size);
3081 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3086 * If this is not a physical io, make sure that it is properly aligned
3087 * before proceeding.
3089 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3090 ASSERT0(P2PHASE(zio->io_offset, align));
3091 ASSERT0(P2PHASE(zio->io_size, align));
3094 * For the physical io we allow alignment
3095 * to a logical block size.
3097 uint64_t log_align =
3098 1ULL << vd->vdev_top->vdev_logical_ashift;
3099 ASSERT0(P2PHASE(zio->io_offset, log_align));
3100 ASSERT0(P2PHASE(zio->io_size, log_align));
3103 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3106 * If this is a repair I/O, and there's no self-healing involved --
3107 * that is, we're just resilvering what we expect to resilver --
3108 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3109 * This prevents spurious resilvering with nested replication.
3110 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3111 * A is out of date, we'll read from C+D, then use the data to
3112 * resilver A+B -- but we don't actually want to resilver B, just A.
3113 * The top-level mirror has no way to know this, so instead we just
3114 * discard unnecessary repairs as we work our way down the vdev tree.
3115 * The same logic applies to any form of nested replication:
3116 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3118 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3119 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3120 zio->io_txg != 0 && /* not a delegated i/o */
3121 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3122 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3123 zio_vdev_io_bypass(zio);
3124 return (ZIO_PIPELINE_CONTINUE);
3127 if (vd->vdev_ops->vdev_op_leaf) {
3128 switch (zio->io_type) {
3130 if (vdev_cache_read(zio))
3131 return (ZIO_PIPELINE_CONTINUE);
3133 case ZIO_TYPE_WRITE:
3135 if ((zio = vdev_queue_io(zio)) == NULL)
3136 return (ZIO_PIPELINE_STOP);
3138 if (!vdev_accessible(vd, zio)) {
3139 zio->io_error = SET_ERROR(ENXIO);
3141 return (ZIO_PIPELINE_STOP);
3146 * Note that we ignore repair writes for TRIM because they can
3147 * conflict with normal writes. This isn't an issue because, by
3148 * definition, we only repair blocks that aren't freed.
3150 if (zio->io_type == ZIO_TYPE_WRITE &&
3151 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3152 !trim_map_write_start(zio))
3153 return (ZIO_PIPELINE_STOP);
3156 vd->vdev_ops->vdev_op_io_start(zio);
3157 return (ZIO_PIPELINE_STOP);
3161 zio_vdev_io_done(zio_t *zio)
3163 vdev_t *vd = zio->io_vd;
3164 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3165 boolean_t unexpected_error = B_FALSE;
3167 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3168 return (ZIO_PIPELINE_STOP);
3171 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3172 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3174 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3175 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3176 zio->io_type == ZIO_TYPE_FREE)) {
3178 if (zio->io_type == ZIO_TYPE_WRITE &&
3179 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3180 trim_map_write_done(zio);
3182 vdev_queue_io_done(zio);
3184 if (zio->io_type == ZIO_TYPE_WRITE)
3185 vdev_cache_write(zio);
3187 if (zio_injection_enabled && zio->io_error == 0)
3188 zio->io_error = zio_handle_device_injection(vd,
3191 if (zio_injection_enabled && zio->io_error == 0)
3192 zio->io_error = zio_handle_label_injection(zio, EIO);
3194 if (zio->io_error) {
3195 if (zio->io_error == ENOTSUP &&
3196 zio->io_type == ZIO_TYPE_FREE) {
3197 /* Not all devices support TRIM. */
3198 } else if (!vdev_accessible(vd, zio)) {
3199 zio->io_error = SET_ERROR(ENXIO);
3201 unexpected_error = B_TRUE;
3206 ops->vdev_op_io_done(zio);
3208 if (unexpected_error)
3209 VERIFY(vdev_probe(vd, zio) == NULL);
3211 return (ZIO_PIPELINE_CONTINUE);
3215 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3216 * disk, and use that to finish the checksum ereport later.
3219 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3220 const void *good_buf)
3222 /* no processing needed */
3223 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3228 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3230 void *buf = zio_buf_alloc(zio->io_size);
3232 bcopy(zio->io_data, buf, zio->io_size);
3234 zcr->zcr_cbinfo = zio->io_size;
3235 zcr->zcr_cbdata = buf;
3236 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3237 zcr->zcr_free = zio_buf_free;
3241 zio_vdev_io_assess(zio_t *zio)
3243 vdev_t *vd = zio->io_vd;
3245 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3246 return (ZIO_PIPELINE_STOP);
3249 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3250 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3252 if (zio->io_vsd != NULL) {
3253 zio->io_vsd_ops->vsd_free(zio);
3257 if (zio_injection_enabled && zio->io_error == 0)
3258 zio->io_error = zio_handle_fault_injection(zio, EIO);
3260 if (zio->io_type == ZIO_TYPE_FREE &&
3261 zio->io_priority != ZIO_PRIORITY_NOW) {
3262 switch (zio->io_error) {
3264 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3265 ZIO_TRIM_STAT_BUMP(success);
3268 ZIO_TRIM_STAT_BUMP(unsupported);
3271 ZIO_TRIM_STAT_BUMP(failed);
3277 * If the I/O failed, determine whether we should attempt to retry it.
3279 * On retry, we cut in line in the issue queue, since we don't want
3280 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3282 if (zio->io_error && vd == NULL &&
3283 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3284 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3285 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3287 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3288 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3289 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3290 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3291 zio_requeue_io_start_cut_in_line);
3292 return (ZIO_PIPELINE_STOP);
3296 * If we got an error on a leaf device, convert it to ENXIO
3297 * if the device is not accessible at all.
3299 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3300 !vdev_accessible(vd, zio))
3301 zio->io_error = SET_ERROR(ENXIO);
3304 * If we can't write to an interior vdev (mirror or RAID-Z),
3305 * set vdev_cant_write so that we stop trying to allocate from it.
3307 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3308 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3309 vd->vdev_cant_write = B_TRUE;
3313 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3315 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3316 zio->io_physdone != NULL) {
3317 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3318 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3319 zio->io_physdone(zio->io_logical);
3322 return (ZIO_PIPELINE_CONTINUE);
3326 zio_vdev_io_reissue(zio_t *zio)
3328 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3329 ASSERT(zio->io_error == 0);
3331 zio->io_stage >>= 1;
3335 zio_vdev_io_redone(zio_t *zio)
3337 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3339 zio->io_stage >>= 1;
3343 zio_vdev_io_bypass(zio_t *zio)
3345 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3346 ASSERT(zio->io_error == 0);
3348 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3349 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3353 * ==========================================================================
3354 * Generate and verify checksums
3355 * ==========================================================================
3358 zio_checksum_generate(zio_t *zio)
3360 blkptr_t *bp = zio->io_bp;
3361 enum zio_checksum checksum;
3365 * This is zio_write_phys().
3366 * We're either generating a label checksum, or none at all.
3368 checksum = zio->io_prop.zp_checksum;
3370 if (checksum == ZIO_CHECKSUM_OFF)
3371 return (ZIO_PIPELINE_CONTINUE);
3373 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3375 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3376 ASSERT(!IO_IS_ALLOCATING(zio));
3377 checksum = ZIO_CHECKSUM_GANG_HEADER;
3379 checksum = BP_GET_CHECKSUM(bp);
3383 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3385 return (ZIO_PIPELINE_CONTINUE);
3389 zio_checksum_verify(zio_t *zio)
3391 zio_bad_cksum_t info;
3392 blkptr_t *bp = zio->io_bp;
3395 ASSERT(zio->io_vd != NULL);
3399 * This is zio_read_phys().
3400 * We're either verifying a label checksum, or nothing at all.
3402 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3403 return (ZIO_PIPELINE_CONTINUE);
3405 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3408 if ((error = zio_checksum_error(zio, &info)) != 0) {
3409 zio->io_error = error;
3410 if (error == ECKSUM &&
3411 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3412 zfs_ereport_start_checksum(zio->io_spa,
3413 zio->io_vd, zio, zio->io_offset,
3414 zio->io_size, NULL, &info);
3418 return (ZIO_PIPELINE_CONTINUE);
3422 * Called by RAID-Z to ensure we don't compute the checksum twice.
3425 zio_checksum_verified(zio_t *zio)
3427 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3431 * ==========================================================================
3432 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3433 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3434 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3435 * indicate errors that are specific to one I/O, and most likely permanent.
3436 * Any other error is presumed to be worse because we weren't expecting it.
3437 * ==========================================================================
3440 zio_worst_error(int e1, int e2)
3442 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3445 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3446 if (e1 == zio_error_rank[r1])
3449 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3450 if (e2 == zio_error_rank[r2])
3453 return (r1 > r2 ? e1 : e2);
3457 * ==========================================================================
3459 * ==========================================================================
3462 zio_ready(zio_t *zio)
3464 blkptr_t *bp = zio->io_bp;
3465 zio_t *pio, *pio_next;
3466 zio_link_t *zl = NULL;
3468 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3470 return (ZIO_PIPELINE_STOP);
3473 if (zio->io_ready) {
3474 ASSERT(IO_IS_ALLOCATING(zio));
3475 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3476 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3477 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3482 if (bp != NULL && bp != &zio->io_bp_copy)
3483 zio->io_bp_copy = *bp;
3485 if (zio->io_error != 0) {
3486 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3488 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3489 ASSERT(IO_IS_ALLOCATING(zio));
3490 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3492 * We were unable to allocate anything, unreserve and
3493 * issue the next I/O to allocate.
3495 metaslab_class_throttle_unreserve(
3496 spa_normal_class(zio->io_spa),
3497 zio->io_prop.zp_copies, zio);
3498 zio_allocate_dispatch(zio->io_spa);
3502 mutex_enter(&zio->io_lock);
3503 zio->io_state[ZIO_WAIT_READY] = 1;
3504 pio = zio_walk_parents(zio, &zl);
3505 mutex_exit(&zio->io_lock);
3508 * As we notify zio's parents, new parents could be added.
3509 * New parents go to the head of zio's io_parent_list, however,
3510 * so we will (correctly) not notify them. The remainder of zio's
3511 * io_parent_list, from 'pio_next' onward, cannot change because
3512 * all parents must wait for us to be done before they can be done.
3514 for (; pio != NULL; pio = pio_next) {
3515 pio_next = zio_walk_parents(zio, &zl);
3516 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3519 if (zio->io_flags & ZIO_FLAG_NODATA) {
3520 if (BP_IS_GANG(bp)) {
3521 zio->io_flags &= ~ZIO_FLAG_NODATA;
3523 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3524 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3528 if (zio_injection_enabled &&
3529 zio->io_spa->spa_syncing_txg == zio->io_txg)
3530 zio_handle_ignored_writes(zio);
3532 return (ZIO_PIPELINE_CONTINUE);
3536 * Update the allocation throttle accounting.
3539 zio_dva_throttle_done(zio_t *zio)
3541 zio_t *lio = zio->io_logical;
3542 zio_t *pio = zio_unique_parent(zio);
3543 vdev_t *vd = zio->io_vd;
3544 int flags = METASLAB_ASYNC_ALLOC;
3546 ASSERT3P(zio->io_bp, !=, NULL);
3547 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3548 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3549 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3551 ASSERT3P(vd, ==, vd->vdev_top);
3552 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3553 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3554 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3555 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3558 * Parents of gang children can have two flavors -- ones that
3559 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3560 * and ones that allocated the constituent blocks. The allocation
3561 * throttle needs to know the allocating parent zio so we must find
3564 if (pio->io_child_type == ZIO_CHILD_GANG) {
3566 * If our parent is a rewrite gang child then our grandparent
3567 * would have been the one that performed the allocation.
3569 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3570 pio = zio_unique_parent(pio);
3571 flags |= METASLAB_GANG_CHILD;
3574 ASSERT(IO_IS_ALLOCATING(pio));
3575 ASSERT3P(zio, !=, zio->io_logical);
3576 ASSERT(zio->io_logical != NULL);
3577 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3578 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3580 mutex_enter(&pio->io_lock);
3581 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3582 mutex_exit(&pio->io_lock);
3584 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3588 * Call into the pipeline to see if there is more work that
3589 * needs to be done. If there is work to be done it will be
3590 * dispatched to another taskq thread.
3592 zio_allocate_dispatch(zio->io_spa);
3596 zio_done(zio_t *zio)
3598 spa_t *spa = zio->io_spa;
3599 zio_t *lio = zio->io_logical;
3600 blkptr_t *bp = zio->io_bp;
3601 vdev_t *vd = zio->io_vd;
3602 uint64_t psize = zio->io_size;
3603 zio_t *pio, *pio_next;
3604 metaslab_class_t *mc = spa_normal_class(spa);
3605 zio_link_t *zl = NULL;
3608 * If our children haven't all completed,
3609 * wait for them and then repeat this pipeline stage.
3611 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3612 return (ZIO_PIPELINE_STOP);
3616 * If the allocation throttle is enabled, then update the accounting.
3617 * We only track child I/Os that are part of an allocating async
3618 * write. We must do this since the allocation is performed
3619 * by the logical I/O but the actual write is done by child I/Os.
3621 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3622 zio->io_child_type == ZIO_CHILD_VDEV) {
3623 ASSERT(mc->mc_alloc_throttle_enabled);
3624 zio_dva_throttle_done(zio);
3628 * If the allocation throttle is enabled, verify that
3629 * we have decremented the refcounts for every I/O that was throttled.
3631 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3632 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3633 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3635 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3636 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3639 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3640 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3641 ASSERT(zio->io_children[c][w] == 0);
3643 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3644 ASSERT(bp->blk_pad[0] == 0);
3645 ASSERT(bp->blk_pad[1] == 0);
3646 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3647 (bp == zio_unique_parent(zio)->io_bp));
3648 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3649 zio->io_bp_override == NULL &&
3650 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3651 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3652 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3653 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3654 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3656 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3657 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3661 * If there were child vdev/gang/ddt errors, they apply to us now.
3663 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3664 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3665 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3668 * If the I/O on the transformed data was successful, generate any
3669 * checksum reports now while we still have the transformed data.
3671 if (zio->io_error == 0) {
3672 while (zio->io_cksum_report != NULL) {
3673 zio_cksum_report_t *zcr = zio->io_cksum_report;
3674 uint64_t align = zcr->zcr_align;
3675 uint64_t asize = P2ROUNDUP(psize, align);
3676 char *abuf = zio->io_data;
3678 if (asize != psize) {
3679 abuf = zio_buf_alloc(asize);
3680 bcopy(zio->io_data, abuf, psize);
3681 bzero(abuf + psize, asize - psize);
3684 zio->io_cksum_report = zcr->zcr_next;
3685 zcr->zcr_next = NULL;
3686 zcr->zcr_finish(zcr, abuf);
3687 zfs_ereport_free_checksum(zcr);
3690 zio_buf_free(abuf, asize);
3694 zio_pop_transforms(zio); /* note: may set zio->io_error */
3696 vdev_stat_update(zio, psize);
3698 if (zio->io_error) {
3700 * If this I/O is attached to a particular vdev,
3701 * generate an error message describing the I/O failure
3702 * at the block level. We ignore these errors if the
3703 * device is currently unavailable.
3705 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3706 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3708 if ((zio->io_error == EIO || !(zio->io_flags &
3709 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3712 * For logical I/O requests, tell the SPA to log the
3713 * error and generate a logical data ereport.
3715 spa_log_error(spa, zio);
3716 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3721 if (zio->io_error && zio == lio) {
3723 * Determine whether zio should be reexecuted. This will
3724 * propagate all the way to the root via zio_notify_parent().
3726 ASSERT(vd == NULL && bp != NULL);
3727 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3729 if (IO_IS_ALLOCATING(zio) &&
3730 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3731 if (zio->io_error != ENOSPC)
3732 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3734 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3737 if ((zio->io_type == ZIO_TYPE_READ ||
3738 zio->io_type == ZIO_TYPE_FREE) &&
3739 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3740 zio->io_error == ENXIO &&
3741 spa_load_state(spa) == SPA_LOAD_NONE &&
3742 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3743 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3745 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3746 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3749 * Here is a possibly good place to attempt to do
3750 * either combinatorial reconstruction or error correction
3751 * based on checksums. It also might be a good place
3752 * to send out preliminary ereports before we suspend
3758 * If there were logical child errors, they apply to us now.
3759 * We defer this until now to avoid conflating logical child
3760 * errors with errors that happened to the zio itself when
3761 * updating vdev stats and reporting FMA events above.
3763 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3765 if ((zio->io_error || zio->io_reexecute) &&
3766 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3767 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3768 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3770 zio_gang_tree_free(&zio->io_gang_tree);
3773 * Godfather I/Os should never suspend.
3775 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3776 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3777 zio->io_reexecute = 0;
3779 if (zio->io_reexecute) {
3781 * This is a logical I/O that wants to reexecute.
3783 * Reexecute is top-down. When an i/o fails, if it's not
3784 * the root, it simply notifies its parent and sticks around.
3785 * The parent, seeing that it still has children in zio_done(),
3786 * does the same. This percolates all the way up to the root.
3787 * The root i/o will reexecute or suspend the entire tree.
3789 * This approach ensures that zio_reexecute() honors
3790 * all the original i/o dependency relationships, e.g.
3791 * parents not executing until children are ready.
3793 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3795 zio->io_gang_leader = NULL;
3797 mutex_enter(&zio->io_lock);
3798 zio->io_state[ZIO_WAIT_DONE] = 1;
3799 mutex_exit(&zio->io_lock);
3802 * "The Godfather" I/O monitors its children but is
3803 * not a true parent to them. It will track them through
3804 * the pipeline but severs its ties whenever they get into
3805 * trouble (e.g. suspended). This allows "The Godfather"
3806 * I/O to return status without blocking.
3809 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3811 zio_link_t *remove_zl = zl;
3812 pio_next = zio_walk_parents(zio, &zl);
3814 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3815 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3816 zio_remove_child(pio, zio, remove_zl);
3817 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3821 if ((pio = zio_unique_parent(zio)) != NULL) {
3823 * We're not a root i/o, so there's nothing to do
3824 * but notify our parent. Don't propagate errors
3825 * upward since we haven't permanently failed yet.
3827 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3828 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3829 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3830 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3832 * We'd fail again if we reexecuted now, so suspend
3833 * until conditions improve (e.g. device comes online).
3835 zio_suspend(spa, zio);
3838 * Reexecution is potentially a huge amount of work.
3839 * Hand it off to the otherwise-unused claim taskq.
3841 #if defined(illumos) || !defined(_KERNEL)
3842 ASSERT(zio->io_tqent.tqent_next == NULL);
3844 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3846 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3847 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3850 return (ZIO_PIPELINE_STOP);
3853 ASSERT(zio->io_child_count == 0);
3854 ASSERT(zio->io_reexecute == 0);
3855 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3858 * Report any checksum errors, since the I/O is complete.
3860 while (zio->io_cksum_report != NULL) {
3861 zio_cksum_report_t *zcr = zio->io_cksum_report;
3862 zio->io_cksum_report = zcr->zcr_next;
3863 zcr->zcr_next = NULL;
3864 zcr->zcr_finish(zcr, NULL);
3865 zfs_ereport_free_checksum(zcr);
3869 * It is the responsibility of the done callback to ensure that this
3870 * particular zio is no longer discoverable for adoption, and as
3871 * such, cannot acquire any new parents.
3876 mutex_enter(&zio->io_lock);
3877 zio->io_state[ZIO_WAIT_DONE] = 1;
3878 mutex_exit(&zio->io_lock);
3881 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3882 zio_link_t *remove_zl = zl;
3883 pio_next = zio_walk_parents(zio, &zl);
3884 zio_remove_child(pio, zio, remove_zl);
3885 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3888 if (zio->io_waiter != NULL) {
3889 mutex_enter(&zio->io_lock);
3890 zio->io_executor = NULL;
3891 cv_broadcast(&zio->io_cv);
3892 mutex_exit(&zio->io_lock);
3897 return (ZIO_PIPELINE_STOP);
3901 * ==========================================================================
3902 * I/O pipeline definition
3903 * ==========================================================================
3905 static zio_pipe_stage_t *zio_pipeline[] = {
3912 zio_checksum_generate,
3928 zio_checksum_verify,
3936 * Compare two zbookmark_phys_t's to see which we would reach first in a
3937 * pre-order traversal of the object tree.
3939 * This is simple in every case aside from the meta-dnode object. For all other
3940 * objects, we traverse them in order (object 1 before object 2, and so on).
3941 * However, all of these objects are traversed while traversing object 0, since
3942 * the data it points to is the list of objects. Thus, we need to convert to a
3943 * canonical representation so we can compare meta-dnode bookmarks to
3944 * non-meta-dnode bookmarks.
3946 * We do this by calculating "equivalents" for each field of the zbookmark.
3947 * zbookmarks outside of the meta-dnode use their own object and level, and
3948 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3949 * blocks this bookmark refers to) by multiplying their blkid by their span
3950 * (the number of L0 blocks contained within one block at their level).
3951 * zbookmarks inside the meta-dnode calculate their object equivalent
3952 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3953 * level + 1<<31 (any value larger than a level could ever be) for their level.
3954 * This causes them to always compare before a bookmark in their object
3955 * equivalent, compare appropriately to bookmarks in other objects, and to
3956 * compare appropriately to other bookmarks in the meta-dnode.
3959 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3960 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3963 * These variables represent the "equivalent" values for the zbookmark,
3964 * after converting zbookmarks inside the meta dnode to their
3965 * normal-object equivalents.
3967 uint64_t zb1obj, zb2obj;
3968 uint64_t zb1L0, zb2L0;
3969 uint64_t zb1level, zb2level;
3971 if (zb1->zb_object == zb2->zb_object &&
3972 zb1->zb_level == zb2->zb_level &&
3973 zb1->zb_blkid == zb2->zb_blkid)
3977 * BP_SPANB calculates the span in blocks.
3979 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3980 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3982 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3983 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3985 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3987 zb1obj = zb1->zb_object;
3988 zb1level = zb1->zb_level;
3991 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3992 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3994 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3996 zb2obj = zb2->zb_object;
3997 zb2level = zb2->zb_level;
4000 /* Now that we have a canonical representation, do the comparison. */
4001 if (zb1obj != zb2obj)
4002 return (zb1obj < zb2obj ? -1 : 1);
4003 else if (zb1L0 != zb2L0)
4004 return (zb1L0 < zb2L0 ? -1 : 1);
4005 else if (zb1level != zb2level)
4006 return (zb1level > zb2level ? -1 : 1);
4008 * This can (theoretically) happen if the bookmarks have the same object
4009 * and level, but different blkids, if the block sizes are not the same.
4010 * There is presently no way to change the indirect block sizes
4016 * This function checks the following: given that last_block is the place that
4017 * our traversal stopped last time, does that guarantee that we've visited
4018 * every node under subtree_root? Therefore, we can't just use the raw output
4019 * of zbookmark_compare. We have to pass in a modified version of
4020 * subtree_root; by incrementing the block id, and then checking whether
4021 * last_block is before or equal to that, we can tell whether or not having
4022 * visited last_block implies that all of subtree_root's children have been
4026 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4027 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4029 zbookmark_phys_t mod_zb = *subtree_root;
4031 ASSERT(last_block->zb_level == 0);
4033 /* The objset_phys_t isn't before anything. */
4038 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4039 * data block size in sectors, because that variable is only used if
4040 * the bookmark refers to a block in the meta-dnode. Since we don't
4041 * know without examining it what object it refers to, and there's no
4042 * harm in passing in this value in other cases, we always pass it in.
4044 * We pass in 0 for the indirect block size shift because zb2 must be
4045 * level 0. The indirect block size is only used to calculate the span
4046 * of the bookmark, but since the bookmark must be level 0, the span is
4047 * always 1, so the math works out.
4049 * If you make changes to how the zbookmark_compare code works, be sure
4050 * to make sure that this code still works afterwards.
4052 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4053 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,