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 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
54 "Use uma(9) for ZIO allocations");
55 static int zio_exclude_metadata = 0;
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
57 "Exclude metadata buffers from dumps as well");
59 zio_trim_stats_t zio_trim_stats = {
60 { "bytes", KSTAT_DATA_UINT64,
61 "Number of bytes successfully TRIMmed" },
62 { "success", KSTAT_DATA_UINT64,
63 "Number of successful TRIM requests" },
64 { "unsupported", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed because TRIM is not supported" },
66 { "failed", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed for reasons other than not supported" },
70 static kstat_t *zio_trim_ksp;
73 * ==========================================================================
74 * I/O type descriptions
75 * ==========================================================================
77 const char *zio_type_name[ZIO_TYPES] = {
78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
82 boolean_t zio_dva_throttle_enabled = B_TRUE;
83 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
84 &zio_dva_throttle_enabled, 0, "");
87 * ==========================================================================
89 * ==========================================================================
91 kmem_cache_t *zio_cache;
92 kmem_cache_t *zio_link_cache;
93 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
94 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97 extern vmem_t *zio_alloc_arena;
100 #define ZIO_PIPELINE_CONTINUE 0x100
101 #define ZIO_PIPELINE_STOP 0x101
103 #define BP_SPANB(indblkshift, level) \
104 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
105 #define COMPARE_META_LEVEL 0x80000000ul
107 * The following actions directly effect the spa's sync-to-convergence logic.
108 * The values below define the sync pass when we start performing the action.
109 * Care should be taken when changing these values as they directly impact
110 * spa_sync() performance. Tuning these values may introduce subtle performance
111 * pathologies and should only be done in the context of performance analysis.
112 * These tunables will eventually be removed and replaced with #defines once
113 * enough analysis has been done to determine optimal values.
115 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
116 * regular blocks are not deferred.
118 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
119 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
120 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
121 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
123 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
124 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
126 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
138 int zio_buf_debug_limit = 16384;
140 int zio_buf_debug_limit = 0;
144 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
150 zio_cache = kmem_cache_create("zio_cache",
151 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
152 zio_link_cache = kmem_cache_create("zio_link_cache",
153 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 * For small buffers, we want a cache for each multiple of
159 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
160 * for each quarter-power of 2.
162 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
163 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
166 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
174 * If we are using watchpoints, put each buffer on its own page,
175 * to eliminate the performance overhead of trapping to the
176 * kernel when modifying a non-watched buffer that shares the
177 * page with a watched buffer.
179 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
183 if (size <= 4 * SPA_MINBLOCKSIZE) {
184 align = SPA_MINBLOCKSIZE;
185 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
186 align = MIN(p2 >> 2, PAGESIZE);
191 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
192 zio_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL, cflags);
196 * Since zio_data bufs do not appear in crash dumps, we
197 * pass KMC_NOTOUCH so that no allocator metadata is
198 * stored with the buffers.
200 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
201 zio_data_buf_cache[c] = kmem_cache_create(name, size,
202 align, NULL, NULL, NULL, NULL, NULL,
203 cflags | KMC_NOTOUCH | KMC_NODEBUG);
208 ASSERT(zio_buf_cache[c] != NULL);
209 if (zio_buf_cache[c - 1] == NULL)
210 zio_buf_cache[c - 1] = zio_buf_cache[c];
212 ASSERT(zio_data_buf_cache[c] != NULL);
213 if (zio_data_buf_cache[c - 1] == NULL)
214 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
220 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 if (zio_trim_ksp != NULL) {
226 zio_trim_ksp->ks_data = &zio_trim_stats;
227 kstat_install(zio_trim_ksp);
235 kmem_cache_t *last_cache = NULL;
236 kmem_cache_t *last_data_cache = NULL;
238 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
239 if (zio_buf_cache[c] != last_cache) {
240 last_cache = zio_buf_cache[c];
241 kmem_cache_destroy(zio_buf_cache[c]);
243 zio_buf_cache[c] = NULL;
245 if (zio_data_buf_cache[c] != last_data_cache) {
246 last_data_cache = zio_data_buf_cache[c];
247 kmem_cache_destroy(zio_data_buf_cache[c]);
249 zio_data_buf_cache[c] = NULL;
252 kmem_cache_destroy(zio_link_cache);
253 kmem_cache_destroy(zio_cache);
257 if (zio_trim_ksp != NULL) {
258 kstat_delete(zio_trim_ksp);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
286 return (kmem_alloc(size, KM_SLEEP|flags));
290 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
291 * crashdump if the kernel panics. This exists so that we will limit the amount
292 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
293 * of kernel heap dumped to disk when the kernel panics)
296 zio_data_buf_alloc(size_t size)
298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
305 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
309 zio_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316 kmem_cache_free(zio_buf_cache[c], buf);
318 kmem_free(buf, size);
322 zio_data_buf_free(void *buf, size_t size)
324 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
329 kmem_cache_free(zio_data_buf_cache[c], buf);
331 kmem_free(buf, size);
335 * ==========================================================================
336 * Push and pop I/O transform buffers
337 * ==========================================================================
340 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
341 zio_transform_func_t *transform)
343 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
345 zt->zt_orig_data = zio->io_data;
346 zt->zt_orig_size = zio->io_size;
347 zt->zt_bufsize = bufsize;
348 zt->zt_transform = transform;
350 zt->zt_next = zio->io_transform_stack;
351 zio->io_transform_stack = zt;
358 zio_pop_transforms(zio_t *zio)
362 while ((zt = zio->io_transform_stack) != NULL) {
363 if (zt->zt_transform != NULL)
364 zt->zt_transform(zio,
365 zt->zt_orig_data, zt->zt_orig_size);
367 if (zt->zt_bufsize != 0)
368 zio_buf_free(zio->io_data, zt->zt_bufsize);
370 zio->io_data = zt->zt_orig_data;
371 zio->io_size = zt->zt_orig_size;
372 zio->io_transform_stack = zt->zt_next;
374 kmem_free(zt, sizeof (zio_transform_t));
379 * ==========================================================================
380 * I/O transform callbacks for subblocks and decompression
381 * ==========================================================================
384 zio_subblock(zio_t *zio, void *data, uint64_t size)
386 ASSERT(zio->io_size > size);
388 if (zio->io_type == ZIO_TYPE_READ)
389 bcopy(zio->io_data, data, size);
393 zio_decompress(zio_t *zio, void *data, uint64_t size)
395 if (zio->io_error == 0 &&
396 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
397 zio->io_data, data, zio->io_size, size) != 0)
398 zio->io_error = SET_ERROR(EIO);
402 * ==========================================================================
403 * I/O parent/child relationships and pipeline interlocks
404 * ==========================================================================
407 zio_walk_parents(zio_t *cio, zio_link_t **zl)
409 list_t *pl = &cio->io_parent_list;
411 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
415 ASSERT((*zl)->zl_child == cio);
416 return ((*zl)->zl_parent);
420 zio_walk_children(zio_t *pio, zio_link_t **zl)
422 list_t *cl = &pio->io_child_list;
424 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
428 ASSERT((*zl)->zl_parent == pio);
429 return ((*zl)->zl_child);
433 zio_unique_parent(zio_t *cio)
435 zio_link_t *zl = NULL;
436 zio_t *pio = zio_walk_parents(cio, &zl);
438 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
443 zio_add_child(zio_t *pio, zio_t *cio)
445 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
453 ASSERT(cio->io_child_type <= pio->io_child_type);
458 mutex_enter(&cio->io_lock);
459 mutex_enter(&pio->io_lock);
461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
463 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
466 list_insert_head(&pio->io_child_list, zl);
467 list_insert_head(&cio->io_parent_list, zl);
469 pio->io_child_count++;
470 cio->io_parent_count++;
472 mutex_exit(&pio->io_lock);
473 mutex_exit(&cio->io_lock);
477 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
479 ASSERT(zl->zl_parent == pio);
480 ASSERT(zl->zl_child == cio);
482 mutex_enter(&cio->io_lock);
483 mutex_enter(&pio->io_lock);
485 list_remove(&pio->io_child_list, zl);
486 list_remove(&cio->io_parent_list, zl);
488 pio->io_child_count--;
489 cio->io_parent_count--;
491 mutex_exit(&pio->io_lock);
492 mutex_exit(&cio->io_lock);
494 kmem_cache_free(zio_link_cache, zl);
498 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
500 uint64_t *countp = &zio->io_children[child][wait];
501 boolean_t waiting = B_FALSE;
503 mutex_enter(&zio->io_lock);
504 ASSERT(zio->io_stall == NULL);
507 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
508 zio->io_stall = countp;
511 mutex_exit(&zio->io_lock);
517 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
519 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
520 int *errorp = &pio->io_child_error[zio->io_child_type];
522 mutex_enter(&pio->io_lock);
523 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
524 *errorp = zio_worst_error(*errorp, zio->io_error);
525 pio->io_reexecute |= zio->io_reexecute;
526 ASSERT3U(*countp, >, 0);
530 if (*countp == 0 && pio->io_stall == countp) {
531 zio_taskq_type_t type =
532 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
534 pio->io_stall = NULL;
535 mutex_exit(&pio->io_lock);
537 * Dispatch the parent zio in its own taskq so that
538 * the child can continue to make progress. This also
539 * prevents overflowing the stack when we have deeply nested
540 * parent-child relationships.
542 zio_taskq_dispatch(pio, type, B_FALSE);
544 mutex_exit(&pio->io_lock);
549 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
551 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
552 zio->io_error = zio->io_child_error[c];
556 zio_timestamp_compare(const void *x1, const void *x2)
558 const zio_t *z1 = x1;
559 const zio_t *z2 = x2;
561 if (z1->io_queued_timestamp < z2->io_queued_timestamp)
563 if (z1->io_queued_timestamp > z2->io_queued_timestamp)
566 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
568 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
571 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
573 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
576 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
578 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
581 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
583 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
595 * ==========================================================================
596 * Create the various types of I/O (read, write, free, etc)
597 * ==========================================================================
600 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
601 void *data, uint64_t size, zio_done_func_t *done, void *private,
602 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
603 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
604 enum zio_stage stage, enum zio_stage pipeline)
608 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
609 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
610 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
612 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
613 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
614 ASSERT(vd || stage == ZIO_STAGE_OPEN);
616 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
617 bzero(zio, sizeof (zio_t));
619 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
620 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
622 list_create(&zio->io_parent_list, sizeof (zio_link_t),
623 offsetof(zio_link_t, zl_parent_node));
624 list_create(&zio->io_child_list, sizeof (zio_link_t),
625 offsetof(zio_link_t, zl_child_node));
628 zio->io_child_type = ZIO_CHILD_VDEV;
629 else if (flags & ZIO_FLAG_GANG_CHILD)
630 zio->io_child_type = ZIO_CHILD_GANG;
631 else if (flags & ZIO_FLAG_DDT_CHILD)
632 zio->io_child_type = ZIO_CHILD_DDT;
634 zio->io_child_type = ZIO_CHILD_LOGICAL;
637 zio->io_bp = (blkptr_t *)bp;
638 zio->io_bp_copy = *bp;
639 zio->io_bp_orig = *bp;
640 if (type != ZIO_TYPE_WRITE ||
641 zio->io_child_type == ZIO_CHILD_DDT)
642 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
643 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
644 zio->io_logical = zio;
645 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
646 pipeline |= ZIO_GANG_STAGES;
652 zio->io_private = private;
654 zio->io_priority = priority;
656 zio->io_offset = offset;
657 zio->io_orig_data = zio->io_data = data;
658 zio->io_orig_size = zio->io_size = size;
659 zio->io_orig_flags = zio->io_flags = flags;
660 zio->io_orig_stage = zio->io_stage = stage;
661 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
662 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
664 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
665 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
668 zio->io_bookmark = *zb;
671 if (zio->io_logical == NULL)
672 zio->io_logical = pio->io_logical;
673 if (zio->io_child_type == ZIO_CHILD_GANG)
674 zio->io_gang_leader = pio->io_gang_leader;
675 zio_add_child(pio, zio);
682 zio_destroy(zio_t *zio)
684 list_destroy(&zio->io_parent_list);
685 list_destroy(&zio->io_child_list);
686 mutex_destroy(&zio->io_lock);
687 cv_destroy(&zio->io_cv);
688 kmem_cache_free(zio_cache, zio);
692 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
693 void *private, enum zio_flag flags)
697 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
698 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
699 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
705 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
707 return (zio_null(NULL, spa, NULL, done, private, flags));
711 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
713 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
714 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
715 bp, (longlong_t)BP_GET_TYPE(bp));
717 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
718 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
719 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
720 bp, (longlong_t)BP_GET_CHECKSUM(bp));
722 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
723 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
724 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
725 bp, (longlong_t)BP_GET_COMPRESS(bp));
727 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
728 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
729 bp, (longlong_t)BP_GET_LSIZE(bp));
731 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
732 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
733 bp, (longlong_t)BP_GET_PSIZE(bp));
736 if (BP_IS_EMBEDDED(bp)) {
737 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
738 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
739 bp, (longlong_t)BPE_GET_ETYPE(bp));
744 * Pool-specific checks.
746 * Note: it would be nice to verify that the blk_birth and
747 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
748 * allows the birth time of log blocks (and dmu_sync()-ed blocks
749 * that are in the log) to be arbitrarily large.
751 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
752 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
753 if (vdevid >= spa->spa_root_vdev->vdev_children) {
754 zfs_panic_recover("blkptr at %p DVA %u has invalid "
756 bp, i, (longlong_t)vdevid);
759 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
761 zfs_panic_recover("blkptr at %p DVA %u has invalid "
763 bp, i, (longlong_t)vdevid);
766 if (vd->vdev_ops == &vdev_hole_ops) {
767 zfs_panic_recover("blkptr at %p DVA %u has hole "
769 bp, i, (longlong_t)vdevid);
772 if (vd->vdev_ops == &vdev_missing_ops) {
774 * "missing" vdevs are valid during import, but we
775 * don't have their detailed info (e.g. asize), so
776 * we can't perform any more checks on them.
780 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
781 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
783 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
784 if (offset + asize > vd->vdev_asize) {
785 zfs_panic_recover("blkptr at %p DVA %u has invalid "
787 bp, i, (longlong_t)offset);
793 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
794 void *data, uint64_t size, zio_done_func_t *done, void *private,
795 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
799 zfs_blkptr_verify(spa, bp);
801 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
802 data, size, done, private,
803 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
804 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
805 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
811 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
812 void *data, uint64_t size, const zio_prop_t *zp,
813 zio_done_func_t *ready, zio_done_func_t *children_ready,
814 zio_done_func_t *physdone, zio_done_func_t *done,
815 void *private, zio_priority_t priority, enum zio_flag flags,
816 const zbookmark_phys_t *zb)
820 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
821 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
822 zp->zp_compress >= ZIO_COMPRESS_OFF &&
823 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
824 DMU_OT_IS_VALID(zp->zp_type) &&
827 zp->zp_copies <= spa_max_replication(spa));
829 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
830 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
831 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
832 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
834 zio->io_ready = ready;
835 zio->io_children_ready = children_ready;
836 zio->io_physdone = physdone;
840 * Data can be NULL if we are going to call zio_write_override() to
841 * provide the already-allocated BP. But we may need the data to
842 * verify a dedup hit (if requested). In this case, don't try to
843 * dedup (just take the already-allocated BP verbatim).
845 if (data == NULL && zio->io_prop.zp_dedup_verify) {
846 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
853 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
854 uint64_t size, zio_done_func_t *done, void *private,
855 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
859 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
860 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
861 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
867 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
869 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
870 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
871 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
872 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
875 * We must reset the io_prop to match the values that existed
876 * when the bp was first written by dmu_sync() keeping in mind
877 * that nopwrite and dedup are mutually exclusive.
879 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
880 zio->io_prop.zp_nopwrite = nopwrite;
881 zio->io_prop.zp_copies = copies;
882 zio->io_bp_override = bp;
886 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
890 * The check for EMBEDDED is a performance optimization. We
891 * process the free here (by ignoring it) rather than
892 * putting it on the list and then processing it in zio_free_sync().
894 if (BP_IS_EMBEDDED(bp))
896 metaslab_check_free(spa, bp);
899 * Frees that are for the currently-syncing txg, are not going to be
900 * deferred, and which will not need to do a read (i.e. not GANG or
901 * DEDUP), can be processed immediately. Otherwise, put them on the
902 * in-memory list for later processing.
904 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
905 txg != spa->spa_syncing_txg ||
906 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
907 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
909 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
910 BP_GET_PSIZE(bp), 0)));
915 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
916 uint64_t size, enum zio_flag flags)
919 enum zio_stage stage = ZIO_FREE_PIPELINE;
921 ASSERT(!BP_IS_HOLE(bp));
922 ASSERT(spa_syncing_txg(spa) == txg);
923 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
925 if (BP_IS_EMBEDDED(bp))
926 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
928 metaslab_check_free(spa, bp);
931 if (zfs_trim_enabled)
932 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
933 ZIO_STAGE_VDEV_IO_ASSESS;
935 * GANG and DEDUP blocks can induce a read (for the gang block header,
936 * or the DDT), so issue them asynchronously so that this thread is
939 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
940 stage |= ZIO_STAGE_ISSUE_ASYNC;
942 flags |= ZIO_FLAG_DONT_QUEUE;
944 zio = zio_create(pio, spa, txg, bp, NULL, size,
945 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
946 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
952 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
953 zio_done_func_t *done, void *private, enum zio_flag flags)
957 dprintf_bp(bp, "claiming in txg %llu", txg);
959 if (BP_IS_EMBEDDED(bp))
960 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
963 * A claim is an allocation of a specific block. Claims are needed
964 * to support immediate writes in the intent log. The issue is that
965 * immediate writes contain committed data, but in a txg that was
966 * *not* committed. Upon opening the pool after an unclean shutdown,
967 * the intent log claims all blocks that contain immediate write data
968 * so that the SPA knows they're in use.
970 * All claims *must* be resolved in the first txg -- before the SPA
971 * starts allocating blocks -- so that nothing is allocated twice.
972 * If txg == 0 we just verify that the block is claimable.
974 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
975 ASSERT(txg == spa_first_txg(spa) || txg == 0);
976 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
978 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
979 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
980 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
981 ASSERT0(zio->io_queued_timestamp);
987 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
988 uint64_t size, zio_done_func_t *done, void *private,
989 zio_priority_t priority, enum zio_flag flags)
994 if (vd->vdev_children == 0) {
995 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
996 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
997 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1001 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1003 for (c = 0; c < vd->vdev_children; c++)
1004 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1005 offset, size, done, private, priority, flags));
1012 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1013 void *data, int checksum, zio_done_func_t *done, void *private,
1014 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1018 ASSERT(vd->vdev_children == 0);
1019 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1020 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1021 ASSERT3U(offset + size, <=, vd->vdev_psize);
1023 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1024 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1025 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1027 zio->io_prop.zp_checksum = checksum;
1033 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1034 void *data, int checksum, zio_done_func_t *done, void *private,
1035 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1039 ASSERT(vd->vdev_children == 0);
1040 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1041 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1042 ASSERT3U(offset + size, <=, vd->vdev_psize);
1044 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1045 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1046 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1048 zio->io_prop.zp_checksum = checksum;
1050 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1052 * zec checksums are necessarily destructive -- they modify
1053 * the end of the write buffer to hold the verifier/checksum.
1054 * Therefore, we must make a local copy in case the data is
1055 * being written to multiple places in parallel.
1057 void *wbuf = zio_buf_alloc(size);
1058 bcopy(data, wbuf, size);
1059 zio_push_transform(zio, wbuf, size, size, NULL);
1066 * Create a child I/O to do some work for us.
1069 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1070 void *data, uint64_t size, int type, zio_priority_t priority,
1071 enum zio_flag flags, zio_done_func_t *done, void *private)
1073 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1076 ASSERT(vd->vdev_parent ==
1077 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1079 if (type == ZIO_TYPE_READ && bp != NULL) {
1081 * If we have the bp, then the child should perform the
1082 * checksum and the parent need not. This pushes error
1083 * detection as close to the leaves as possible and
1084 * eliminates redundant checksums in the interior nodes.
1086 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1087 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1090 /* Not all IO types require vdev io done stage e.g. free */
1091 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1092 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1094 if (vd->vdev_children == 0)
1095 offset += VDEV_LABEL_START_SIZE;
1097 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1100 * If we've decided to do a repair, the write is not speculative --
1101 * even if the original read was.
1103 if (flags & ZIO_FLAG_IO_REPAIR)
1104 flags &= ~ZIO_FLAG_SPECULATIVE;
1107 * If we're creating a child I/O that is not associated with a
1108 * top-level vdev, then the child zio is not an allocating I/O.
1109 * If this is a retried I/O then we ignore it since we will
1110 * have already processed the original allocating I/O.
1112 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1113 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1114 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1116 ASSERT(mc->mc_alloc_throttle_enabled);
1117 ASSERT(type == ZIO_TYPE_WRITE);
1118 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1119 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1120 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1121 pio->io_child_type == ZIO_CHILD_GANG);
1123 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1126 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1127 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1128 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1129 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1131 zio->io_physdone = pio->io_physdone;
1132 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1133 zio->io_logical->io_phys_children++;
1139 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1140 int type, zio_priority_t priority, enum zio_flag flags,
1141 zio_done_func_t *done, void *private)
1145 ASSERT(vd->vdev_ops->vdev_op_leaf);
1147 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1148 data, size, done, private, type, priority,
1149 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1151 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1157 zio_flush(zio_t *zio, vdev_t *vd)
1159 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1160 NULL, NULL, ZIO_PRIORITY_NOW,
1161 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1165 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1168 ASSERT(vd->vdev_ops->vdev_op_leaf);
1170 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1171 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1172 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1173 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1177 zio_shrink(zio_t *zio, uint64_t size)
1179 ASSERT(zio->io_executor == NULL);
1180 ASSERT(zio->io_orig_size == zio->io_size);
1181 ASSERT(size <= zio->io_size);
1184 * We don't shrink for raidz because of problems with the
1185 * reconstruction when reading back less than the block size.
1186 * Note, BP_IS_RAIDZ() assumes no compression.
1188 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1189 if (!BP_IS_RAIDZ(zio->io_bp))
1190 zio->io_orig_size = zio->io_size = size;
1194 * ==========================================================================
1195 * Prepare to read and write logical blocks
1196 * ==========================================================================
1200 zio_read_bp_init(zio_t *zio)
1202 blkptr_t *bp = zio->io_bp;
1204 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1205 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1206 !(zio->io_flags & ZIO_FLAG_RAW)) {
1208 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1209 void *cbuf = zio_buf_alloc(psize);
1211 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1214 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1215 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1216 decode_embedded_bp_compressed(bp, zio->io_data);
1218 ASSERT(!BP_IS_EMBEDDED(bp));
1221 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1222 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1224 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1225 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1227 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1228 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1230 return (ZIO_PIPELINE_CONTINUE);
1234 zio_write_bp_init(zio_t *zio)
1236 if (!IO_IS_ALLOCATING(zio))
1237 return (ZIO_PIPELINE_CONTINUE);
1239 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1241 if (zio->io_bp_override) {
1242 blkptr_t *bp = zio->io_bp;
1243 zio_prop_t *zp = &zio->io_prop;
1245 ASSERT(bp->blk_birth != zio->io_txg);
1246 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1248 *bp = *zio->io_bp_override;
1249 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1251 if (BP_IS_EMBEDDED(bp))
1252 return (ZIO_PIPELINE_CONTINUE);
1255 * If we've been overridden and nopwrite is set then
1256 * set the flag accordingly to indicate that a nopwrite
1257 * has already occurred.
1259 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1260 ASSERT(!zp->zp_dedup);
1261 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1262 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1263 return (ZIO_PIPELINE_CONTINUE);
1266 ASSERT(!zp->zp_nopwrite);
1268 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1269 return (ZIO_PIPELINE_CONTINUE);
1271 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1272 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1274 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1275 BP_SET_DEDUP(bp, 1);
1276 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1277 return (ZIO_PIPELINE_CONTINUE);
1281 * We were unable to handle this as an override bp, treat
1282 * it as a regular write I/O.
1284 zio->io_bp_override = NULL;
1285 *bp = zio->io_bp_orig;
1286 zio->io_pipeline = zio->io_orig_pipeline;
1289 return (ZIO_PIPELINE_CONTINUE);
1293 zio_write_compress(zio_t *zio)
1295 spa_t *spa = zio->io_spa;
1296 zio_prop_t *zp = &zio->io_prop;
1297 enum zio_compress compress = zp->zp_compress;
1298 blkptr_t *bp = zio->io_bp;
1299 uint64_t lsize = zio->io_size;
1300 uint64_t psize = lsize;
1304 * If our children haven't all reached the ready stage,
1305 * wait for them and then repeat this pipeline stage.
1307 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1308 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1309 return (ZIO_PIPELINE_STOP);
1311 if (!IO_IS_ALLOCATING(zio))
1312 return (ZIO_PIPELINE_CONTINUE);
1314 if (zio->io_children_ready != NULL) {
1316 * Now that all our children are ready, run the callback
1317 * associated with this zio in case it wants to modify the
1318 * data to be written.
1320 ASSERT3U(zp->zp_level, >, 0);
1321 zio->io_children_ready(zio);
1324 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1325 ASSERT(zio->io_bp_override == NULL);
1327 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1329 * We're rewriting an existing block, which means we're
1330 * working on behalf of spa_sync(). For spa_sync() to
1331 * converge, it must eventually be the case that we don't
1332 * have to allocate new blocks. But compression changes
1333 * the blocksize, which forces a reallocate, and makes
1334 * convergence take longer. Therefore, after the first
1335 * few passes, stop compressing to ensure convergence.
1337 pass = spa_sync_pass(spa);
1339 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1340 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1341 ASSERT(!BP_GET_DEDUP(bp));
1343 if (pass >= zfs_sync_pass_dont_compress)
1344 compress = ZIO_COMPRESS_OFF;
1346 /* Make sure someone doesn't change their mind on overwrites */
1347 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1348 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1351 if (compress != ZIO_COMPRESS_OFF) {
1352 void *cbuf = zio_buf_alloc(lsize);
1353 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1354 if (psize == 0 || psize == lsize) {
1355 compress = ZIO_COMPRESS_OFF;
1356 zio_buf_free(cbuf, lsize);
1357 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1358 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1359 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1360 encode_embedded_bp_compressed(bp,
1361 cbuf, compress, lsize, psize);
1362 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1363 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1364 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1365 zio_buf_free(cbuf, lsize);
1366 bp->blk_birth = zio->io_txg;
1367 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1368 ASSERT(spa_feature_is_active(spa,
1369 SPA_FEATURE_EMBEDDED_DATA));
1370 return (ZIO_PIPELINE_CONTINUE);
1373 * Round up compressed size up to the ashift
1374 * of the smallest-ashift device, and zero the tail.
1375 * This ensures that the compressed size of the BP
1376 * (and thus compressratio property) are correct,
1377 * in that we charge for the padding used to fill out
1380 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1381 size_t rounded = (size_t)P2ROUNDUP(psize,
1382 1ULL << spa->spa_min_ashift);
1383 if (rounded >= lsize) {
1384 compress = ZIO_COMPRESS_OFF;
1385 zio_buf_free(cbuf, lsize);
1388 bzero((char *)cbuf + psize, rounded - psize);
1390 zio_push_transform(zio, cbuf,
1391 psize, lsize, NULL);
1396 * We were unable to handle this as an override bp, treat
1397 * it as a regular write I/O.
1399 zio->io_bp_override = NULL;
1400 *bp = zio->io_bp_orig;
1401 zio->io_pipeline = zio->io_orig_pipeline;
1405 * The final pass of spa_sync() must be all rewrites, but the first
1406 * few passes offer a trade-off: allocating blocks defers convergence,
1407 * but newly allocated blocks are sequential, so they can be written
1408 * to disk faster. Therefore, we allow the first few passes of
1409 * spa_sync() to allocate new blocks, but force rewrites after that.
1410 * There should only be a handful of blocks after pass 1 in any case.
1412 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1413 BP_GET_PSIZE(bp) == psize &&
1414 pass >= zfs_sync_pass_rewrite) {
1416 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1417 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1418 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1421 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1425 if (zio->io_bp_orig.blk_birth != 0 &&
1426 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1427 BP_SET_LSIZE(bp, lsize);
1428 BP_SET_TYPE(bp, zp->zp_type);
1429 BP_SET_LEVEL(bp, zp->zp_level);
1430 BP_SET_BIRTH(bp, zio->io_txg, 0);
1432 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1434 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1435 BP_SET_LSIZE(bp, lsize);
1436 BP_SET_TYPE(bp, zp->zp_type);
1437 BP_SET_LEVEL(bp, zp->zp_level);
1438 BP_SET_PSIZE(bp, psize);
1439 BP_SET_COMPRESS(bp, compress);
1440 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1441 BP_SET_DEDUP(bp, zp->zp_dedup);
1442 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1444 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1445 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1446 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1448 if (zp->zp_nopwrite) {
1449 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1450 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1451 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1454 return (ZIO_PIPELINE_CONTINUE);
1458 zio_free_bp_init(zio_t *zio)
1460 blkptr_t *bp = zio->io_bp;
1462 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1463 if (BP_GET_DEDUP(bp))
1464 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1467 return (ZIO_PIPELINE_CONTINUE);
1471 * ==========================================================================
1472 * Execute the I/O pipeline
1473 * ==========================================================================
1477 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1479 spa_t *spa = zio->io_spa;
1480 zio_type_t t = zio->io_type;
1481 int flags = (cutinline ? TQ_FRONT : 0);
1483 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1486 * If we're a config writer or a probe, the normal issue and
1487 * interrupt threads may all be blocked waiting for the config lock.
1488 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1490 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1494 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1496 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1500 * If this is a high priority I/O, then use the high priority taskq if
1503 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1504 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1507 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1510 * NB: We are assuming that the zio can only be dispatched
1511 * to a single taskq at a time. It would be a grievous error
1512 * to dispatch the zio to another taskq at the same time.
1514 #if defined(illumos) || !defined(_KERNEL)
1515 ASSERT(zio->io_tqent.tqent_next == NULL);
1517 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1519 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1520 flags, &zio->io_tqent);
1524 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1526 kthread_t *executor = zio->io_executor;
1527 spa_t *spa = zio->io_spa;
1529 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1530 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1532 for (i = 0; i < tqs->stqs_count; i++) {
1533 if (taskq_member(tqs->stqs_taskq[i], executor))
1542 zio_issue_async(zio_t *zio)
1544 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1546 return (ZIO_PIPELINE_STOP);
1550 zio_interrupt(zio_t *zio)
1552 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1556 zio_delay_interrupt(zio_t *zio)
1559 * The timeout_generic() function isn't defined in userspace, so
1560 * rather than trying to implement the function, the zio delay
1561 * functionality has been disabled for userspace builds.
1566 * If io_target_timestamp is zero, then no delay has been registered
1567 * for this IO, thus jump to the end of this function and "skip" the
1568 * delay; issuing it directly to the zio layer.
1570 if (zio->io_target_timestamp != 0) {
1571 hrtime_t now = gethrtime();
1573 if (now >= zio->io_target_timestamp) {
1575 * This IO has already taken longer than the target
1576 * delay to complete, so we don't want to delay it
1577 * any longer; we "miss" the delay and issue it
1578 * directly to the zio layer. This is likely due to
1579 * the target latency being set to a value less than
1580 * the underlying hardware can satisfy (e.g. delay
1581 * set to 1ms, but the disks take 10ms to complete an
1585 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1590 hrtime_t diff = zio->io_target_timestamp - now;
1592 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1593 hrtime_t, now, hrtime_t, diff);
1595 (void) timeout_generic(CALLOUT_NORMAL,
1596 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1603 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1608 * Execute the I/O pipeline until one of the following occurs:
1610 * (1) the I/O completes
1611 * (2) the pipeline stalls waiting for dependent child I/Os
1612 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1613 * (4) the I/O is delegated by vdev-level caching or aggregation
1614 * (5) the I/O is deferred due to vdev-level queueing
1615 * (6) the I/O is handed off to another thread.
1617 * In all cases, the pipeline stops whenever there's no CPU work; it never
1618 * burns a thread in cv_wait().
1620 * There's no locking on io_stage because there's no legitimate way
1621 * for multiple threads to be attempting to process the same I/O.
1623 static zio_pipe_stage_t *zio_pipeline[];
1626 zio_execute(zio_t *zio)
1628 zio->io_executor = curthread;
1630 ASSERT3U(zio->io_queued_timestamp, >, 0);
1632 while (zio->io_stage < ZIO_STAGE_DONE) {
1633 enum zio_stage pipeline = zio->io_pipeline;
1634 enum zio_stage stage = zio->io_stage;
1637 ASSERT(!MUTEX_HELD(&zio->io_lock));
1638 ASSERT(ISP2(stage));
1639 ASSERT(zio->io_stall == NULL);
1643 } while ((stage & pipeline) == 0);
1645 ASSERT(stage <= ZIO_STAGE_DONE);
1648 * If we are in interrupt context and this pipeline stage
1649 * will grab a config lock that is held across I/O,
1650 * or may wait for an I/O that needs an interrupt thread
1651 * to complete, issue async to avoid deadlock.
1653 * For VDEV_IO_START, we cut in line so that the io will
1654 * be sent to disk promptly.
1656 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1657 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1658 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1659 zio_requeue_io_start_cut_in_line : B_FALSE;
1660 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1664 zio->io_stage = stage;
1665 zio->io_pipeline_trace |= zio->io_stage;
1666 rv = zio_pipeline[highbit64(stage) - 1](zio);
1668 if (rv == ZIO_PIPELINE_STOP)
1671 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1676 * ==========================================================================
1677 * Initiate I/O, either sync or async
1678 * ==========================================================================
1681 zio_wait(zio_t *zio)
1685 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1686 ASSERT(zio->io_executor == NULL);
1688 zio->io_waiter = curthread;
1689 ASSERT0(zio->io_queued_timestamp);
1690 zio->io_queued_timestamp = gethrtime();
1694 mutex_enter(&zio->io_lock);
1695 while (zio->io_executor != NULL)
1696 cv_wait(&zio->io_cv, &zio->io_lock);
1697 mutex_exit(&zio->io_lock);
1699 error = zio->io_error;
1706 zio_nowait(zio_t *zio)
1708 ASSERT(zio->io_executor == NULL);
1710 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1711 zio_unique_parent(zio) == NULL) {
1713 * This is a logical async I/O with no parent to wait for it.
1714 * We add it to the spa_async_root_zio "Godfather" I/O which
1715 * will ensure they complete prior to unloading the pool.
1717 spa_t *spa = zio->io_spa;
1719 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1722 ASSERT0(zio->io_queued_timestamp);
1723 zio->io_queued_timestamp = gethrtime();
1728 * ==========================================================================
1729 * Reexecute or suspend/resume failed I/O
1730 * ==========================================================================
1734 zio_reexecute(zio_t *pio)
1736 zio_t *cio, *cio_next;
1738 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1739 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1740 ASSERT(pio->io_gang_leader == NULL);
1741 ASSERT(pio->io_gang_tree == NULL);
1743 pio->io_flags = pio->io_orig_flags;
1744 pio->io_stage = pio->io_orig_stage;
1745 pio->io_pipeline = pio->io_orig_pipeline;
1746 pio->io_reexecute = 0;
1747 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1748 pio->io_pipeline_trace = 0;
1750 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1751 pio->io_state[w] = 0;
1752 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1753 pio->io_child_error[c] = 0;
1755 if (IO_IS_ALLOCATING(pio))
1756 BP_ZERO(pio->io_bp);
1759 * As we reexecute pio's children, new children could be created.
1760 * New children go to the head of pio's io_child_list, however,
1761 * so we will (correctly) not reexecute them. The key is that
1762 * the remainder of pio's io_child_list, from 'cio_next' onward,
1763 * cannot be affected by any side effects of reexecuting 'cio'.
1765 zio_link_t *zl = NULL;
1766 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1767 cio_next = zio_walk_children(pio, &zl);
1768 mutex_enter(&pio->io_lock);
1769 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1770 pio->io_children[cio->io_child_type][w]++;
1771 mutex_exit(&pio->io_lock);
1776 * Now that all children have been reexecuted, execute the parent.
1777 * We don't reexecute "The Godfather" I/O here as it's the
1778 * responsibility of the caller to wait on him.
1780 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1781 pio->io_queued_timestamp = gethrtime();
1787 zio_suspend(spa_t *spa, zio_t *zio)
1789 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1790 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1791 "failure and the failure mode property for this pool "
1792 "is set to panic.", spa_name(spa));
1794 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1796 mutex_enter(&spa->spa_suspend_lock);
1798 if (spa->spa_suspend_zio_root == NULL)
1799 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1800 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1801 ZIO_FLAG_GODFATHER);
1803 spa->spa_suspended = B_TRUE;
1806 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1807 ASSERT(zio != spa->spa_suspend_zio_root);
1808 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1809 ASSERT(zio_unique_parent(zio) == NULL);
1810 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1811 zio_add_child(spa->spa_suspend_zio_root, zio);
1814 mutex_exit(&spa->spa_suspend_lock);
1818 zio_resume(spa_t *spa)
1823 * Reexecute all previously suspended i/o.
1825 mutex_enter(&spa->spa_suspend_lock);
1826 spa->spa_suspended = B_FALSE;
1827 cv_broadcast(&spa->spa_suspend_cv);
1828 pio = spa->spa_suspend_zio_root;
1829 spa->spa_suspend_zio_root = NULL;
1830 mutex_exit(&spa->spa_suspend_lock);
1836 return (zio_wait(pio));
1840 zio_resume_wait(spa_t *spa)
1842 mutex_enter(&spa->spa_suspend_lock);
1843 while (spa_suspended(spa))
1844 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1845 mutex_exit(&spa->spa_suspend_lock);
1849 * ==========================================================================
1852 * A gang block is a collection of small blocks that looks to the DMU
1853 * like one large block. When zio_dva_allocate() cannot find a block
1854 * of the requested size, due to either severe fragmentation or the pool
1855 * being nearly full, it calls zio_write_gang_block() to construct the
1856 * block from smaller fragments.
1858 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1859 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1860 * an indirect block: it's an array of block pointers. It consumes
1861 * only one sector and hence is allocatable regardless of fragmentation.
1862 * The gang header's bps point to its gang members, which hold the data.
1864 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1865 * as the verifier to ensure uniqueness of the SHA256 checksum.
1866 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1867 * not the gang header. This ensures that data block signatures (needed for
1868 * deduplication) are independent of how the block is physically stored.
1870 * Gang blocks can be nested: a gang member may itself be a gang block.
1871 * Thus every gang block is a tree in which root and all interior nodes are
1872 * gang headers, and the leaves are normal blocks that contain user data.
1873 * The root of the gang tree is called the gang leader.
1875 * To perform any operation (read, rewrite, free, claim) on a gang block,
1876 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1877 * in the io_gang_tree field of the original logical i/o by recursively
1878 * reading the gang leader and all gang headers below it. This yields
1879 * an in-core tree containing the contents of every gang header and the
1880 * bps for every constituent of the gang block.
1882 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1883 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1884 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1885 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1886 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1887 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1888 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1889 * of the gang header plus zio_checksum_compute() of the data to update the
1890 * gang header's blk_cksum as described above.
1892 * The two-phase assemble/issue model solves the problem of partial failure --
1893 * what if you'd freed part of a gang block but then couldn't read the
1894 * gang header for another part? Assembling the entire gang tree first
1895 * ensures that all the necessary gang header I/O has succeeded before
1896 * starting the actual work of free, claim, or write. Once the gang tree
1897 * is assembled, free and claim are in-memory operations that cannot fail.
1899 * In the event that a gang write fails, zio_dva_unallocate() walks the
1900 * gang tree to immediately free (i.e. insert back into the space map)
1901 * everything we've allocated. This ensures that we don't get ENOSPC
1902 * errors during repeated suspend/resume cycles due to a flaky device.
1904 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1905 * the gang tree, we won't modify the block, so we can safely defer the free
1906 * (knowing that the block is still intact). If we *can* assemble the gang
1907 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1908 * each constituent bp and we can allocate a new block on the next sync pass.
1910 * In all cases, the gang tree allows complete recovery from partial failure.
1911 * ==========================================================================
1915 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1920 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1921 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1922 &pio->io_bookmark));
1926 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1931 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1932 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1933 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1935 * As we rewrite each gang header, the pipeline will compute
1936 * a new gang block header checksum for it; but no one will
1937 * compute a new data checksum, so we do that here. The one
1938 * exception is the gang leader: the pipeline already computed
1939 * its data checksum because that stage precedes gang assembly.
1940 * (Presently, nothing actually uses interior data checksums;
1941 * this is just good hygiene.)
1943 if (gn != pio->io_gang_leader->io_gang_tree) {
1944 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1945 data, BP_GET_PSIZE(bp));
1948 * If we are here to damage data for testing purposes,
1949 * leave the GBH alone so that we can detect the damage.
1951 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1952 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1954 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1955 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1956 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1964 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1966 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1967 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1968 ZIO_GANG_CHILD_FLAGS(pio)));
1973 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1975 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1976 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1979 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1988 static void zio_gang_tree_assemble_done(zio_t *zio);
1990 static zio_gang_node_t *
1991 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1993 zio_gang_node_t *gn;
1995 ASSERT(*gnpp == NULL);
1997 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1998 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2005 zio_gang_node_free(zio_gang_node_t **gnpp)
2007 zio_gang_node_t *gn = *gnpp;
2009 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2010 ASSERT(gn->gn_child[g] == NULL);
2012 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2013 kmem_free(gn, sizeof (*gn));
2018 zio_gang_tree_free(zio_gang_node_t **gnpp)
2020 zio_gang_node_t *gn = *gnpp;
2025 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2026 zio_gang_tree_free(&gn->gn_child[g]);
2028 zio_gang_node_free(gnpp);
2032 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2034 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2036 ASSERT(gio->io_gang_leader == gio);
2037 ASSERT(BP_IS_GANG(bp));
2039 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
2040 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2041 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2045 zio_gang_tree_assemble_done(zio_t *zio)
2047 zio_t *gio = zio->io_gang_leader;
2048 zio_gang_node_t *gn = zio->io_private;
2049 blkptr_t *bp = zio->io_bp;
2051 ASSERT(gio == zio_unique_parent(zio));
2052 ASSERT(zio->io_child_count == 0);
2057 if (BP_SHOULD_BYTESWAP(bp))
2058 byteswap_uint64_array(zio->io_data, zio->io_size);
2060 ASSERT(zio->io_data == gn->gn_gbh);
2061 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2062 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2064 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2065 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2066 if (!BP_IS_GANG(gbp))
2068 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2073 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2075 zio_t *gio = pio->io_gang_leader;
2078 ASSERT(BP_IS_GANG(bp) == !!gn);
2079 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2080 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2083 * If you're a gang header, your data is in gn->gn_gbh.
2084 * If you're a gang member, your data is in 'data' and gn == NULL.
2086 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2089 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2091 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2092 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2093 if (BP_IS_HOLE(gbp))
2095 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
2096 data = (char *)data + BP_GET_PSIZE(gbp);
2100 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2101 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2108 zio_gang_assemble(zio_t *zio)
2110 blkptr_t *bp = zio->io_bp;
2112 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2113 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2115 zio->io_gang_leader = zio;
2117 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2119 return (ZIO_PIPELINE_CONTINUE);
2123 zio_gang_issue(zio_t *zio)
2125 blkptr_t *bp = zio->io_bp;
2127 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2128 return (ZIO_PIPELINE_STOP);
2130 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2131 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2133 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2134 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2136 zio_gang_tree_free(&zio->io_gang_tree);
2138 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2140 return (ZIO_PIPELINE_CONTINUE);
2144 zio_write_gang_member_ready(zio_t *zio)
2146 zio_t *pio = zio_unique_parent(zio);
2147 zio_t *gio = zio->io_gang_leader;
2148 dva_t *cdva = zio->io_bp->blk_dva;
2149 dva_t *pdva = pio->io_bp->blk_dva;
2152 if (BP_IS_HOLE(zio->io_bp))
2155 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2157 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2158 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2159 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2160 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2161 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2163 mutex_enter(&pio->io_lock);
2164 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2165 ASSERT(DVA_GET_GANG(&pdva[d]));
2166 asize = DVA_GET_ASIZE(&pdva[d]);
2167 asize += DVA_GET_ASIZE(&cdva[d]);
2168 DVA_SET_ASIZE(&pdva[d], asize);
2170 mutex_exit(&pio->io_lock);
2174 zio_write_gang_block(zio_t *pio)
2176 spa_t *spa = pio->io_spa;
2177 metaslab_class_t *mc = spa_normal_class(spa);
2178 blkptr_t *bp = pio->io_bp;
2179 zio_t *gio = pio->io_gang_leader;
2181 zio_gang_node_t *gn, **gnpp;
2182 zio_gbh_phys_t *gbh;
2183 uint64_t txg = pio->io_txg;
2184 uint64_t resid = pio->io_size;
2186 int copies = gio->io_prop.zp_copies;
2187 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2191 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2192 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2193 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2194 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2196 flags |= METASLAB_ASYNC_ALLOC;
2197 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2200 * The logical zio has already placed a reservation for
2201 * 'copies' allocation slots but gang blocks may require
2202 * additional copies. These additional copies
2203 * (i.e. gbh_copies - copies) are guaranteed to succeed
2204 * since metaslab_class_throttle_reserve() always allows
2205 * additional reservations for gang blocks.
2207 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2211 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2212 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio);
2214 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2215 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2216 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2219 * If we failed to allocate the gang block header then
2220 * we remove any additional allocation reservations that
2221 * we placed here. The original reservation will
2222 * be removed when the logical I/O goes to the ready
2225 metaslab_class_throttle_unreserve(mc,
2226 gbh_copies - copies, pio);
2228 pio->io_error = error;
2229 return (ZIO_PIPELINE_CONTINUE);
2233 gnpp = &gio->io_gang_tree;
2235 gnpp = pio->io_private;
2236 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2239 gn = zio_gang_node_alloc(gnpp);
2241 bzero(gbh, SPA_GANGBLOCKSIZE);
2244 * Create the gang header.
2246 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2247 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2250 * Create and nowait the gang children.
2252 for (int g = 0; resid != 0; resid -= lsize, g++) {
2253 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2255 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2257 zp.zp_checksum = gio->io_prop.zp_checksum;
2258 zp.zp_compress = ZIO_COMPRESS_OFF;
2259 zp.zp_type = DMU_OT_NONE;
2261 zp.zp_copies = gio->io_prop.zp_copies;
2262 zp.zp_dedup = B_FALSE;
2263 zp.zp_dedup_verify = B_FALSE;
2264 zp.zp_nopwrite = B_FALSE;
2266 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2267 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2268 zio_write_gang_member_ready, NULL, NULL, NULL,
2269 &gn->gn_child[g], pio->io_priority,
2270 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2272 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2273 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2274 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2277 * Gang children won't throttle but we should
2278 * account for their work, so reserve an allocation
2279 * slot for them here.
2281 VERIFY(metaslab_class_throttle_reserve(mc,
2282 zp.zp_copies, cio, flags));
2288 * Set pio's pipeline to just wait for zio to finish.
2290 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2294 return (ZIO_PIPELINE_CONTINUE);
2298 * The zio_nop_write stage in the pipeline determines if allocating a
2299 * new bp is necessary. The nopwrite feature can handle writes in
2300 * either syncing or open context (i.e. zil writes) and as a result is
2301 * mutually exclusive with dedup.
2303 * By leveraging a cryptographically secure checksum, such as SHA256, we
2304 * can compare the checksums of the new data and the old to determine if
2305 * allocating a new block is required. Note that our requirements for
2306 * cryptographic strength are fairly weak: there can't be any accidental
2307 * hash collisions, but we don't need to be secure against intentional
2308 * (malicious) collisions. To trigger a nopwrite, you have to be able
2309 * to write the file to begin with, and triggering an incorrect (hash
2310 * collision) nopwrite is no worse than simply writing to the file.
2311 * That said, there are no known attacks against the checksum algorithms
2312 * used for nopwrite, assuming that the salt and the checksums
2313 * themselves remain secret.
2316 zio_nop_write(zio_t *zio)
2318 blkptr_t *bp = zio->io_bp;
2319 blkptr_t *bp_orig = &zio->io_bp_orig;
2320 zio_prop_t *zp = &zio->io_prop;
2322 ASSERT(BP_GET_LEVEL(bp) == 0);
2323 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2324 ASSERT(zp->zp_nopwrite);
2325 ASSERT(!zp->zp_dedup);
2326 ASSERT(zio->io_bp_override == NULL);
2327 ASSERT(IO_IS_ALLOCATING(zio));
2330 * Check to see if the original bp and the new bp have matching
2331 * characteristics (i.e. same checksum, compression algorithms, etc).
2332 * If they don't then just continue with the pipeline which will
2333 * allocate a new bp.
2335 if (BP_IS_HOLE(bp_orig) ||
2336 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2337 ZCHECKSUM_FLAG_NOPWRITE) ||
2338 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2339 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2340 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2341 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2342 return (ZIO_PIPELINE_CONTINUE);
2345 * If the checksums match then reset the pipeline so that we
2346 * avoid allocating a new bp and issuing any I/O.
2348 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2349 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2350 ZCHECKSUM_FLAG_NOPWRITE);
2351 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2352 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2353 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2354 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2355 sizeof (uint64_t)) == 0);
2358 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2359 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2362 return (ZIO_PIPELINE_CONTINUE);
2366 * ==========================================================================
2368 * ==========================================================================
2371 zio_ddt_child_read_done(zio_t *zio)
2373 blkptr_t *bp = zio->io_bp;
2374 ddt_entry_t *dde = zio->io_private;
2376 zio_t *pio = zio_unique_parent(zio);
2378 mutex_enter(&pio->io_lock);
2379 ddp = ddt_phys_select(dde, bp);
2380 if (zio->io_error == 0)
2381 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2382 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2383 dde->dde_repair_data = zio->io_data;
2385 zio_buf_free(zio->io_data, zio->io_size);
2386 mutex_exit(&pio->io_lock);
2390 zio_ddt_read_start(zio_t *zio)
2392 blkptr_t *bp = zio->io_bp;
2394 ASSERT(BP_GET_DEDUP(bp));
2395 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2396 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2398 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2399 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2400 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2401 ddt_phys_t *ddp = dde->dde_phys;
2402 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2405 ASSERT(zio->io_vsd == NULL);
2408 if (ddp_self == NULL)
2409 return (ZIO_PIPELINE_CONTINUE);
2411 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2412 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2414 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2416 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2417 zio_buf_alloc(zio->io_size), zio->io_size,
2418 zio_ddt_child_read_done, dde, zio->io_priority,
2419 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2420 &zio->io_bookmark));
2422 return (ZIO_PIPELINE_CONTINUE);
2425 zio_nowait(zio_read(zio, zio->io_spa, bp,
2426 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2427 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2429 return (ZIO_PIPELINE_CONTINUE);
2433 zio_ddt_read_done(zio_t *zio)
2435 blkptr_t *bp = zio->io_bp;
2437 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2438 return (ZIO_PIPELINE_STOP);
2440 ASSERT(BP_GET_DEDUP(bp));
2441 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2442 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2444 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2445 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2446 ddt_entry_t *dde = zio->io_vsd;
2448 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2449 return (ZIO_PIPELINE_CONTINUE);
2452 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2453 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2454 return (ZIO_PIPELINE_STOP);
2456 if (dde->dde_repair_data != NULL) {
2457 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2458 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2460 ddt_repair_done(ddt, dde);
2464 ASSERT(zio->io_vsd == NULL);
2466 return (ZIO_PIPELINE_CONTINUE);
2470 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2472 spa_t *spa = zio->io_spa;
2475 * Note: we compare the original data, not the transformed data,
2476 * because when zio->io_bp is an override bp, we will not have
2477 * pushed the I/O transforms. That's an important optimization
2478 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2480 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2481 zio_t *lio = dde->dde_lead_zio[p];
2484 return (lio->io_orig_size != zio->io_orig_size ||
2485 bcmp(zio->io_orig_data, lio->io_orig_data,
2486 zio->io_orig_size) != 0);
2490 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2491 ddt_phys_t *ddp = &dde->dde_phys[p];
2493 if (ddp->ddp_phys_birth != 0) {
2494 arc_buf_t *abuf = NULL;
2495 arc_flags_t aflags = ARC_FLAG_WAIT;
2496 blkptr_t blk = *zio->io_bp;
2499 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2503 error = arc_read(NULL, spa, &blk,
2504 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2505 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2506 &aflags, &zio->io_bookmark);
2509 if (arc_buf_size(abuf) != zio->io_orig_size ||
2510 bcmp(abuf->b_data, zio->io_orig_data,
2511 zio->io_orig_size) != 0)
2512 error = SET_ERROR(EEXIST);
2513 arc_buf_destroy(abuf, &abuf);
2517 return (error != 0);
2525 zio_ddt_child_write_ready(zio_t *zio)
2527 int p = zio->io_prop.zp_copies;
2528 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2529 ddt_entry_t *dde = zio->io_private;
2530 ddt_phys_t *ddp = &dde->dde_phys[p];
2538 ASSERT(dde->dde_lead_zio[p] == zio);
2540 ddt_phys_fill(ddp, zio->io_bp);
2542 zio_link_t *zl = NULL;
2543 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2544 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2550 zio_ddt_child_write_done(zio_t *zio)
2552 int p = zio->io_prop.zp_copies;
2553 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2554 ddt_entry_t *dde = zio->io_private;
2555 ddt_phys_t *ddp = &dde->dde_phys[p];
2559 ASSERT(ddp->ddp_refcnt == 0);
2560 ASSERT(dde->dde_lead_zio[p] == zio);
2561 dde->dde_lead_zio[p] = NULL;
2563 if (zio->io_error == 0) {
2564 zio_link_t *zl = NULL;
2565 while (zio_walk_parents(zio, &zl) != NULL)
2566 ddt_phys_addref(ddp);
2568 ddt_phys_clear(ddp);
2575 zio_ddt_ditto_write_done(zio_t *zio)
2577 int p = DDT_PHYS_DITTO;
2578 zio_prop_t *zp = &zio->io_prop;
2579 blkptr_t *bp = zio->io_bp;
2580 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2581 ddt_entry_t *dde = zio->io_private;
2582 ddt_phys_t *ddp = &dde->dde_phys[p];
2583 ddt_key_t *ddk = &dde->dde_key;
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 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2593 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2594 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2595 if (ddp->ddp_phys_birth != 0)
2596 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2597 ddt_phys_fill(ddp, bp);
2604 zio_ddt_write(zio_t *zio)
2606 spa_t *spa = zio->io_spa;
2607 blkptr_t *bp = zio->io_bp;
2608 uint64_t txg = zio->io_txg;
2609 zio_prop_t *zp = &zio->io_prop;
2610 int p = zp->zp_copies;
2614 ddt_t *ddt = ddt_select(spa, bp);
2618 ASSERT(BP_GET_DEDUP(bp));
2619 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2620 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2623 dde = ddt_lookup(ddt, bp, B_TRUE);
2624 ddp = &dde->dde_phys[p];
2626 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2628 * If we're using a weak checksum, upgrade to a strong checksum
2629 * and try again. If we're already using a strong checksum,
2630 * we can't resolve it, so just convert to an ordinary write.
2631 * (And automatically e-mail a paper to Nature?)
2633 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2634 ZCHECKSUM_FLAG_DEDUP)) {
2635 zp->zp_checksum = spa_dedup_checksum(spa);
2636 zio_pop_transforms(zio);
2637 zio->io_stage = ZIO_STAGE_OPEN;
2640 zp->zp_dedup = B_FALSE;
2642 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2644 return (ZIO_PIPELINE_CONTINUE);
2647 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2648 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2650 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2651 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2652 zio_prop_t czp = *zp;
2654 czp.zp_copies = ditto_copies;
2657 * If we arrived here with an override bp, we won't have run
2658 * the transform stack, so we won't have the data we need to
2659 * generate a child i/o. So, toss the override bp and restart.
2660 * This is safe, because using the override bp is just an
2661 * optimization; and it's rare, so the cost doesn't matter.
2663 if (zio->io_bp_override) {
2664 zio_pop_transforms(zio);
2665 zio->io_stage = ZIO_STAGE_OPEN;
2666 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2667 zio->io_bp_override = NULL;
2670 return (ZIO_PIPELINE_CONTINUE);
2673 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2674 zio->io_orig_size, &czp, NULL, NULL,
2675 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2676 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2678 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2679 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2682 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2683 if (ddp->ddp_phys_birth != 0)
2684 ddt_bp_fill(ddp, bp, txg);
2685 if (dde->dde_lead_zio[p] != NULL)
2686 zio_add_child(zio, dde->dde_lead_zio[p]);
2688 ddt_phys_addref(ddp);
2689 } else if (zio->io_bp_override) {
2690 ASSERT(bp->blk_birth == txg);
2691 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2692 ddt_phys_fill(ddp, bp);
2693 ddt_phys_addref(ddp);
2695 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2696 zio->io_orig_size, zp,
2697 zio_ddt_child_write_ready, NULL, NULL,
2698 zio_ddt_child_write_done, dde, zio->io_priority,
2699 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2701 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2702 dde->dde_lead_zio[p] = cio;
2712 return (ZIO_PIPELINE_CONTINUE);
2715 ddt_entry_t *freedde; /* for debugging */
2718 zio_ddt_free(zio_t *zio)
2720 spa_t *spa = zio->io_spa;
2721 blkptr_t *bp = zio->io_bp;
2722 ddt_t *ddt = ddt_select(spa, bp);
2726 ASSERT(BP_GET_DEDUP(bp));
2727 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2730 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2731 ddp = ddt_phys_select(dde, bp);
2732 ddt_phys_decref(ddp);
2735 return (ZIO_PIPELINE_CONTINUE);
2739 * ==========================================================================
2740 * Allocate and free blocks
2741 * ==========================================================================
2745 zio_io_to_allocate(spa_t *spa)
2749 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2751 zio = avl_first(&spa->spa_alloc_tree);
2755 ASSERT(IO_IS_ALLOCATING(zio));
2758 * Try to place a reservation for this zio. If we're unable to
2759 * reserve then we throttle.
2761 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2762 zio->io_prop.zp_copies, zio, 0)) {
2766 avl_remove(&spa->spa_alloc_tree, zio);
2767 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2773 zio_dva_throttle(zio_t *zio)
2775 spa_t *spa = zio->io_spa;
2778 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2779 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2780 zio->io_child_type == ZIO_CHILD_GANG ||
2781 zio->io_flags & ZIO_FLAG_NODATA) {
2782 return (ZIO_PIPELINE_CONTINUE);
2785 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2787 ASSERT3U(zio->io_queued_timestamp, >, 0);
2788 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2790 mutex_enter(&spa->spa_alloc_lock);
2792 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2793 avl_add(&spa->spa_alloc_tree, zio);
2795 nio = zio_io_to_allocate(zio->io_spa);
2796 mutex_exit(&spa->spa_alloc_lock);
2799 return (ZIO_PIPELINE_CONTINUE);
2802 ASSERT3U(nio->io_queued_timestamp, <=,
2803 zio->io_queued_timestamp);
2804 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2806 * We are passing control to a new zio so make sure that
2807 * it is processed by a different thread. We do this to
2808 * avoid stack overflows that can occur when parents are
2809 * throttled and children are making progress. We allow
2810 * it to go to the head of the taskq since it's already
2813 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2815 return (ZIO_PIPELINE_STOP);
2819 zio_allocate_dispatch(spa_t *spa)
2823 mutex_enter(&spa->spa_alloc_lock);
2824 zio = zio_io_to_allocate(spa);
2825 mutex_exit(&spa->spa_alloc_lock);
2829 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2830 ASSERT0(zio->io_error);
2831 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2835 zio_dva_allocate(zio_t *zio)
2837 spa_t *spa = zio->io_spa;
2838 metaslab_class_t *mc = spa_normal_class(spa);
2839 blkptr_t *bp = zio->io_bp;
2843 if (zio->io_gang_leader == NULL) {
2844 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2845 zio->io_gang_leader = zio;
2848 ASSERT(BP_IS_HOLE(bp));
2849 ASSERT0(BP_GET_NDVAS(bp));
2850 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2851 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2852 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2854 if (zio->io_flags & ZIO_FLAG_NODATA) {
2855 flags |= METASLAB_DONT_THROTTLE;
2857 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2858 flags |= METASLAB_GANG_CHILD;
2860 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2861 flags |= METASLAB_ASYNC_ALLOC;
2864 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2865 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio);
2868 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2869 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2871 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2872 return (zio_write_gang_block(zio));
2873 zio->io_error = error;
2876 return (ZIO_PIPELINE_CONTINUE);
2880 zio_dva_free(zio_t *zio)
2882 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2884 return (ZIO_PIPELINE_CONTINUE);
2888 zio_dva_claim(zio_t *zio)
2892 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2894 zio->io_error = error;
2896 return (ZIO_PIPELINE_CONTINUE);
2900 * Undo an allocation. This is used by zio_done() when an I/O fails
2901 * and we want to give back the block we just allocated.
2902 * This handles both normal blocks and gang blocks.
2905 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2907 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2908 ASSERT(zio->io_bp_override == NULL);
2910 if (!BP_IS_HOLE(bp))
2911 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2914 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2915 zio_dva_unallocate(zio, gn->gn_child[g],
2916 &gn->gn_gbh->zg_blkptr[g]);
2922 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2925 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2926 uint64_t size, boolean_t *slog)
2930 ASSERT(txg > spa_syncing_txg(spa));
2932 error = metaslab_alloc(spa, spa_log_class(spa), size,
2933 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2937 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2938 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2944 BP_SET_LSIZE(new_bp, size);
2945 BP_SET_PSIZE(new_bp, size);
2946 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2947 BP_SET_CHECKSUM(new_bp,
2948 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2949 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2950 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2951 BP_SET_LEVEL(new_bp, 0);
2952 BP_SET_DEDUP(new_bp, 0);
2953 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2960 * Free an intent log block.
2963 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2965 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2966 ASSERT(!BP_IS_GANG(bp));
2968 zio_free(spa, txg, bp);
2972 * ==========================================================================
2973 * Read, write and delete to physical devices
2974 * ==========================================================================
2979 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2980 * stops after this stage and will resume upon I/O completion.
2981 * However, there are instances where the vdev layer may need to
2982 * continue the pipeline when an I/O was not issued. Since the I/O
2983 * that was sent to the vdev layer might be different than the one
2984 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2985 * force the underlying vdev layers to call either zio_execute() or
2986 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2989 zio_vdev_io_start(zio_t *zio)
2991 vdev_t *vd = zio->io_vd;
2993 spa_t *spa = zio->io_spa;
2996 ASSERT(zio->io_error == 0);
2997 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3000 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3001 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3004 * The mirror_ops handle multiple DVAs in a single BP.
3006 vdev_mirror_ops.vdev_op_io_start(zio);
3007 return (ZIO_PIPELINE_STOP);
3010 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3011 zio->io_priority == ZIO_PRIORITY_NOW) {
3012 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3013 return (ZIO_PIPELINE_CONTINUE);
3016 ASSERT3P(zio->io_logical, !=, zio);
3019 * We keep track of time-sensitive I/Os so that the scan thread
3020 * can quickly react to certain workloads. In particular, we care
3021 * about non-scrubbing, top-level reads and writes with the following
3023 * - synchronous writes of user data to non-slog devices
3024 * - any reads of user data
3025 * When these conditions are met, adjust the timestamp of spa_last_io
3026 * which allows the scan thread to adjust its workload accordingly.
3028 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3029 vd == vd->vdev_top && !vd->vdev_islog &&
3030 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3031 zio->io_txg != spa_syncing_txg(spa)) {
3032 uint64_t old = spa->spa_last_io;
3033 uint64_t new = ddi_get_lbolt64();
3035 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3038 align = 1ULL << vd->vdev_top->vdev_ashift;
3040 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3041 P2PHASE(zio->io_size, align) != 0) {
3042 /* Transform logical writes to be a full physical block size. */
3043 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3045 if (zio->io_type == ZIO_TYPE_READ ||
3046 zio->io_type == ZIO_TYPE_WRITE)
3047 abuf = zio_buf_alloc(asize);
3048 ASSERT(vd == vd->vdev_top);
3049 if (zio->io_type == ZIO_TYPE_WRITE) {
3050 bcopy(zio->io_data, abuf, zio->io_size);
3051 bzero(abuf + zio->io_size, asize - zio->io_size);
3053 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3058 * If this is not a physical io, make sure that it is properly aligned
3059 * before proceeding.
3061 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3062 ASSERT0(P2PHASE(zio->io_offset, align));
3063 ASSERT0(P2PHASE(zio->io_size, align));
3066 * For the physical io we allow alignment
3067 * to a logical block size.
3069 uint64_t log_align =
3070 1ULL << vd->vdev_top->vdev_logical_ashift;
3071 ASSERT0(P2PHASE(zio->io_offset, log_align));
3072 ASSERT0(P2PHASE(zio->io_size, log_align));
3075 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3078 * If this is a repair I/O, and there's no self-healing involved --
3079 * that is, we're just resilvering what we expect to resilver --
3080 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3081 * This prevents spurious resilvering with nested replication.
3082 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3083 * A is out of date, we'll read from C+D, then use the data to
3084 * resilver A+B -- but we don't actually want to resilver B, just A.
3085 * The top-level mirror has no way to know this, so instead we just
3086 * discard unnecessary repairs as we work our way down the vdev tree.
3087 * The same logic applies to any form of nested replication:
3088 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3090 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3091 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3092 zio->io_txg != 0 && /* not a delegated i/o */
3093 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3094 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3095 zio_vdev_io_bypass(zio);
3096 return (ZIO_PIPELINE_CONTINUE);
3099 if (vd->vdev_ops->vdev_op_leaf) {
3100 switch (zio->io_type) {
3102 if (vdev_cache_read(zio))
3103 return (ZIO_PIPELINE_CONTINUE);
3105 case ZIO_TYPE_WRITE:
3107 if ((zio = vdev_queue_io(zio)) == NULL)
3108 return (ZIO_PIPELINE_STOP);
3110 if (!vdev_accessible(vd, zio)) {
3111 zio->io_error = SET_ERROR(ENXIO);
3113 return (ZIO_PIPELINE_STOP);
3118 * Note that we ignore repair writes for TRIM because they can
3119 * conflict with normal writes. This isn't an issue because, by
3120 * definition, we only repair blocks that aren't freed.
3122 if (zio->io_type == ZIO_TYPE_WRITE &&
3123 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3124 !trim_map_write_start(zio))
3125 return (ZIO_PIPELINE_STOP);
3128 vd->vdev_ops->vdev_op_io_start(zio);
3129 return (ZIO_PIPELINE_STOP);
3133 zio_vdev_io_done(zio_t *zio)
3135 vdev_t *vd = zio->io_vd;
3136 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3137 boolean_t unexpected_error = B_FALSE;
3139 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3140 return (ZIO_PIPELINE_STOP);
3142 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3143 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3145 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3146 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3147 zio->io_type == ZIO_TYPE_FREE)) {
3149 if (zio->io_type == ZIO_TYPE_WRITE &&
3150 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3151 trim_map_write_done(zio);
3153 vdev_queue_io_done(zio);
3155 if (zio->io_type == ZIO_TYPE_WRITE)
3156 vdev_cache_write(zio);
3158 if (zio_injection_enabled && zio->io_error == 0)
3159 zio->io_error = zio_handle_device_injection(vd,
3162 if (zio_injection_enabled && zio->io_error == 0)
3163 zio->io_error = zio_handle_label_injection(zio, EIO);
3165 if (zio->io_error) {
3166 if (zio->io_error == ENOTSUP &&
3167 zio->io_type == ZIO_TYPE_FREE) {
3168 /* Not all devices support TRIM. */
3169 } else if (!vdev_accessible(vd, zio)) {
3170 zio->io_error = SET_ERROR(ENXIO);
3172 unexpected_error = B_TRUE;
3177 ops->vdev_op_io_done(zio);
3179 if (unexpected_error)
3180 VERIFY(vdev_probe(vd, zio) == NULL);
3182 return (ZIO_PIPELINE_CONTINUE);
3186 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3187 * disk, and use that to finish the checksum ereport later.
3190 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3191 const void *good_buf)
3193 /* no processing needed */
3194 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3199 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3201 void *buf = zio_buf_alloc(zio->io_size);
3203 bcopy(zio->io_data, buf, zio->io_size);
3205 zcr->zcr_cbinfo = zio->io_size;
3206 zcr->zcr_cbdata = buf;
3207 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3208 zcr->zcr_free = zio_buf_free;
3212 zio_vdev_io_assess(zio_t *zio)
3214 vdev_t *vd = zio->io_vd;
3216 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3217 return (ZIO_PIPELINE_STOP);
3219 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3220 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3222 if (zio->io_vsd != NULL) {
3223 zio->io_vsd_ops->vsd_free(zio);
3227 if (zio_injection_enabled && zio->io_error == 0)
3228 zio->io_error = zio_handle_fault_injection(zio, EIO);
3230 if (zio->io_type == ZIO_TYPE_FREE &&
3231 zio->io_priority != ZIO_PRIORITY_NOW) {
3232 switch (zio->io_error) {
3234 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3235 ZIO_TRIM_STAT_BUMP(success);
3238 ZIO_TRIM_STAT_BUMP(unsupported);
3241 ZIO_TRIM_STAT_BUMP(failed);
3247 * If the I/O failed, determine whether we should attempt to retry it.
3249 * On retry, we cut in line in the issue queue, since we don't want
3250 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3252 if (zio->io_error && vd == NULL &&
3253 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3254 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3255 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3257 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3258 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3259 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3260 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3261 zio_requeue_io_start_cut_in_line);
3262 return (ZIO_PIPELINE_STOP);
3266 * If we got an error on a leaf device, convert it to ENXIO
3267 * if the device is not accessible at all.
3269 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3270 !vdev_accessible(vd, zio))
3271 zio->io_error = SET_ERROR(ENXIO);
3274 * If we can't write to an interior vdev (mirror or RAID-Z),
3275 * set vdev_cant_write so that we stop trying to allocate from it.
3277 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3278 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3279 vd->vdev_cant_write = B_TRUE;
3283 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3285 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3286 zio->io_physdone != NULL) {
3287 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3288 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3289 zio->io_physdone(zio->io_logical);
3292 return (ZIO_PIPELINE_CONTINUE);
3296 zio_vdev_io_reissue(zio_t *zio)
3298 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3299 ASSERT(zio->io_error == 0);
3301 zio->io_stage >>= 1;
3305 zio_vdev_io_redone(zio_t *zio)
3307 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3309 zio->io_stage >>= 1;
3313 zio_vdev_io_bypass(zio_t *zio)
3315 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3316 ASSERT(zio->io_error == 0);
3318 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3319 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3323 * ==========================================================================
3324 * Generate and verify checksums
3325 * ==========================================================================
3328 zio_checksum_generate(zio_t *zio)
3330 blkptr_t *bp = zio->io_bp;
3331 enum zio_checksum checksum;
3335 * This is zio_write_phys().
3336 * We're either generating a label checksum, or none at all.
3338 checksum = zio->io_prop.zp_checksum;
3340 if (checksum == ZIO_CHECKSUM_OFF)
3341 return (ZIO_PIPELINE_CONTINUE);
3343 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3345 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3346 ASSERT(!IO_IS_ALLOCATING(zio));
3347 checksum = ZIO_CHECKSUM_GANG_HEADER;
3349 checksum = BP_GET_CHECKSUM(bp);
3353 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3355 return (ZIO_PIPELINE_CONTINUE);
3359 zio_checksum_verify(zio_t *zio)
3361 zio_bad_cksum_t info;
3362 blkptr_t *bp = zio->io_bp;
3365 ASSERT(zio->io_vd != NULL);
3369 * This is zio_read_phys().
3370 * We're either verifying a label checksum, or nothing at all.
3372 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3373 return (ZIO_PIPELINE_CONTINUE);
3375 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3378 if ((error = zio_checksum_error(zio, &info)) != 0) {
3379 zio->io_error = error;
3380 if (error == ECKSUM &&
3381 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3382 zfs_ereport_start_checksum(zio->io_spa,
3383 zio->io_vd, zio, zio->io_offset,
3384 zio->io_size, NULL, &info);
3388 return (ZIO_PIPELINE_CONTINUE);
3392 * Called by RAID-Z to ensure we don't compute the checksum twice.
3395 zio_checksum_verified(zio_t *zio)
3397 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3401 * ==========================================================================
3402 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3403 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3404 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3405 * indicate errors that are specific to one I/O, and most likely permanent.
3406 * Any other error is presumed to be worse because we weren't expecting it.
3407 * ==========================================================================
3410 zio_worst_error(int e1, int e2)
3412 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3415 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3416 if (e1 == zio_error_rank[r1])
3419 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3420 if (e2 == zio_error_rank[r2])
3423 return (r1 > r2 ? e1 : e2);
3427 * ==========================================================================
3429 * ==========================================================================
3432 zio_ready(zio_t *zio)
3434 blkptr_t *bp = zio->io_bp;
3435 zio_t *pio, *pio_next;
3436 zio_link_t *zl = NULL;
3438 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3439 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3440 return (ZIO_PIPELINE_STOP);
3442 if (zio->io_ready) {
3443 ASSERT(IO_IS_ALLOCATING(zio));
3444 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3445 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3446 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3451 if (bp != NULL && bp != &zio->io_bp_copy)
3452 zio->io_bp_copy = *bp;
3454 if (zio->io_error != 0) {
3455 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3457 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3458 ASSERT(IO_IS_ALLOCATING(zio));
3459 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3461 * We were unable to allocate anything, unreserve and
3462 * issue the next I/O to allocate.
3464 metaslab_class_throttle_unreserve(
3465 spa_normal_class(zio->io_spa),
3466 zio->io_prop.zp_copies, zio);
3467 zio_allocate_dispatch(zio->io_spa);
3471 mutex_enter(&zio->io_lock);
3472 zio->io_state[ZIO_WAIT_READY] = 1;
3473 pio = zio_walk_parents(zio, &zl);
3474 mutex_exit(&zio->io_lock);
3477 * As we notify zio's parents, new parents could be added.
3478 * New parents go to the head of zio's io_parent_list, however,
3479 * so we will (correctly) not notify them. The remainder of zio's
3480 * io_parent_list, from 'pio_next' onward, cannot change because
3481 * all parents must wait for us to be done before they can be done.
3483 for (; pio != NULL; pio = pio_next) {
3484 pio_next = zio_walk_parents(zio, &zl);
3485 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3488 if (zio->io_flags & ZIO_FLAG_NODATA) {
3489 if (BP_IS_GANG(bp)) {
3490 zio->io_flags &= ~ZIO_FLAG_NODATA;
3492 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3493 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3497 if (zio_injection_enabled &&
3498 zio->io_spa->spa_syncing_txg == zio->io_txg)
3499 zio_handle_ignored_writes(zio);
3501 return (ZIO_PIPELINE_CONTINUE);
3505 * Update the allocation throttle accounting.
3508 zio_dva_throttle_done(zio_t *zio)
3510 zio_t *lio = zio->io_logical;
3511 zio_t *pio = zio_unique_parent(zio);
3512 vdev_t *vd = zio->io_vd;
3513 int flags = METASLAB_ASYNC_ALLOC;
3515 ASSERT3P(zio->io_bp, !=, NULL);
3516 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3517 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3518 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3520 ASSERT3P(vd, ==, vd->vdev_top);
3521 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3522 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3523 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3524 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3527 * Parents of gang children can have two flavors -- ones that
3528 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3529 * and ones that allocated the constituent blocks. The allocation
3530 * throttle needs to know the allocating parent zio so we must find
3533 if (pio->io_child_type == ZIO_CHILD_GANG) {
3535 * If our parent is a rewrite gang child then our grandparent
3536 * would have been the one that performed the allocation.
3538 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3539 pio = zio_unique_parent(pio);
3540 flags |= METASLAB_GANG_CHILD;
3543 ASSERT(IO_IS_ALLOCATING(pio));
3544 ASSERT3P(zio, !=, zio->io_logical);
3545 ASSERT(zio->io_logical != NULL);
3546 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3547 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3549 mutex_enter(&pio->io_lock);
3550 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3551 mutex_exit(&pio->io_lock);
3553 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3557 * Call into the pipeline to see if there is more work that
3558 * needs to be done. If there is work to be done it will be
3559 * dispatched to another taskq thread.
3561 zio_allocate_dispatch(zio->io_spa);
3565 zio_done(zio_t *zio)
3567 spa_t *spa = zio->io_spa;
3568 zio_t *lio = zio->io_logical;
3569 blkptr_t *bp = zio->io_bp;
3570 vdev_t *vd = zio->io_vd;
3571 uint64_t psize = zio->io_size;
3572 zio_t *pio, *pio_next;
3573 metaslab_class_t *mc = spa_normal_class(spa);
3574 zio_link_t *zl = NULL;
3577 * If our children haven't all completed,
3578 * wait for them and then repeat this pipeline stage.
3580 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3581 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3582 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3583 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3584 return (ZIO_PIPELINE_STOP);
3587 * If the allocation throttle is enabled, then update the accounting.
3588 * We only track child I/Os that are part of an allocating async
3589 * write. We must do this since the allocation is performed
3590 * by the logical I/O but the actual write is done by child I/Os.
3592 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3593 zio->io_child_type == ZIO_CHILD_VDEV) {
3594 ASSERT(mc->mc_alloc_throttle_enabled);
3595 zio_dva_throttle_done(zio);
3599 * If the allocation throttle is enabled, verify that
3600 * we have decremented the refcounts for every I/O that was throttled.
3602 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3603 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3604 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3606 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3607 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3610 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3611 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3612 ASSERT(zio->io_children[c][w] == 0);
3614 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3615 ASSERT(bp->blk_pad[0] == 0);
3616 ASSERT(bp->blk_pad[1] == 0);
3617 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3618 (bp == zio_unique_parent(zio)->io_bp));
3619 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3620 zio->io_bp_override == NULL &&
3621 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3622 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3623 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3624 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3625 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3627 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3628 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3632 * If there were child vdev/gang/ddt errors, they apply to us now.
3634 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3635 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3636 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3639 * If the I/O on the transformed data was successful, generate any
3640 * checksum reports now while we still have the transformed data.
3642 if (zio->io_error == 0) {
3643 while (zio->io_cksum_report != NULL) {
3644 zio_cksum_report_t *zcr = zio->io_cksum_report;
3645 uint64_t align = zcr->zcr_align;
3646 uint64_t asize = P2ROUNDUP(psize, align);
3647 char *abuf = zio->io_data;
3649 if (asize != psize) {
3650 abuf = zio_buf_alloc(asize);
3651 bcopy(zio->io_data, abuf, psize);
3652 bzero(abuf + psize, asize - psize);
3655 zio->io_cksum_report = zcr->zcr_next;
3656 zcr->zcr_next = NULL;
3657 zcr->zcr_finish(zcr, abuf);
3658 zfs_ereport_free_checksum(zcr);
3661 zio_buf_free(abuf, asize);
3665 zio_pop_transforms(zio); /* note: may set zio->io_error */
3667 vdev_stat_update(zio, psize);
3669 if (zio->io_error) {
3671 * If this I/O is attached to a particular vdev,
3672 * generate an error message describing the I/O failure
3673 * at the block level. We ignore these errors if the
3674 * device is currently unavailable.
3676 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3677 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3679 if ((zio->io_error == EIO || !(zio->io_flags &
3680 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3683 * For logical I/O requests, tell the SPA to log the
3684 * error and generate a logical data ereport.
3686 spa_log_error(spa, zio);
3687 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3692 if (zio->io_error && zio == lio) {
3694 * Determine whether zio should be reexecuted. This will
3695 * propagate all the way to the root via zio_notify_parent().
3697 ASSERT(vd == NULL && bp != NULL);
3698 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3700 if (IO_IS_ALLOCATING(zio) &&
3701 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3702 if (zio->io_error != ENOSPC)
3703 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3705 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3708 if ((zio->io_type == ZIO_TYPE_READ ||
3709 zio->io_type == ZIO_TYPE_FREE) &&
3710 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3711 zio->io_error == ENXIO &&
3712 spa_load_state(spa) == SPA_LOAD_NONE &&
3713 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3714 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3716 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3717 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3720 * Here is a possibly good place to attempt to do
3721 * either combinatorial reconstruction or error correction
3722 * based on checksums. It also might be a good place
3723 * to send out preliminary ereports before we suspend
3729 * If there were logical child errors, they apply to us now.
3730 * We defer this until now to avoid conflating logical child
3731 * errors with errors that happened to the zio itself when
3732 * updating vdev stats and reporting FMA events above.
3734 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3736 if ((zio->io_error || zio->io_reexecute) &&
3737 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3738 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3739 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3741 zio_gang_tree_free(&zio->io_gang_tree);
3744 * Godfather I/Os should never suspend.
3746 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3747 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3748 zio->io_reexecute = 0;
3750 if (zio->io_reexecute) {
3752 * This is a logical I/O that wants to reexecute.
3754 * Reexecute is top-down. When an i/o fails, if it's not
3755 * the root, it simply notifies its parent and sticks around.
3756 * The parent, seeing that it still has children in zio_done(),
3757 * does the same. This percolates all the way up to the root.
3758 * The root i/o will reexecute or suspend the entire tree.
3760 * This approach ensures that zio_reexecute() honors
3761 * all the original i/o dependency relationships, e.g.
3762 * parents not executing until children are ready.
3764 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3766 zio->io_gang_leader = NULL;
3768 mutex_enter(&zio->io_lock);
3769 zio->io_state[ZIO_WAIT_DONE] = 1;
3770 mutex_exit(&zio->io_lock);
3773 * "The Godfather" I/O monitors its children but is
3774 * not a true parent to them. It will track them through
3775 * the pipeline but severs its ties whenever they get into
3776 * trouble (e.g. suspended). This allows "The Godfather"
3777 * I/O to return status without blocking.
3780 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3782 zio_link_t *remove_zl = zl;
3783 pio_next = zio_walk_parents(zio, &zl);
3785 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3786 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3787 zio_remove_child(pio, zio, remove_zl);
3788 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3792 if ((pio = zio_unique_parent(zio)) != NULL) {
3794 * We're not a root i/o, so there's nothing to do
3795 * but notify our parent. Don't propagate errors
3796 * upward since we haven't permanently failed yet.
3798 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3799 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3800 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3801 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3803 * We'd fail again if we reexecuted now, so suspend
3804 * until conditions improve (e.g. device comes online).
3806 zio_suspend(spa, zio);
3809 * Reexecution is potentially a huge amount of work.
3810 * Hand it off to the otherwise-unused claim taskq.
3812 #if defined(illumos) || !defined(_KERNEL)
3813 ASSERT(zio->io_tqent.tqent_next == NULL);
3815 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3817 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3818 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3821 return (ZIO_PIPELINE_STOP);
3824 ASSERT(zio->io_child_count == 0);
3825 ASSERT(zio->io_reexecute == 0);
3826 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3829 * Report any checksum errors, since the I/O is complete.
3831 while (zio->io_cksum_report != NULL) {
3832 zio_cksum_report_t *zcr = zio->io_cksum_report;
3833 zio->io_cksum_report = zcr->zcr_next;
3834 zcr->zcr_next = NULL;
3835 zcr->zcr_finish(zcr, NULL);
3836 zfs_ereport_free_checksum(zcr);
3840 * It is the responsibility of the done callback to ensure that this
3841 * particular zio is no longer discoverable for adoption, and as
3842 * such, cannot acquire any new parents.
3847 mutex_enter(&zio->io_lock);
3848 zio->io_state[ZIO_WAIT_DONE] = 1;
3849 mutex_exit(&zio->io_lock);
3852 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3853 zio_link_t *remove_zl = zl;
3854 pio_next = zio_walk_parents(zio, &zl);
3855 zio_remove_child(pio, zio, remove_zl);
3856 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3859 if (zio->io_waiter != NULL) {
3860 mutex_enter(&zio->io_lock);
3861 zio->io_executor = NULL;
3862 cv_broadcast(&zio->io_cv);
3863 mutex_exit(&zio->io_lock);
3868 return (ZIO_PIPELINE_STOP);
3872 * ==========================================================================
3873 * I/O pipeline definition
3874 * ==========================================================================
3876 static zio_pipe_stage_t *zio_pipeline[] = {
3883 zio_checksum_generate,
3899 zio_checksum_verify,
3907 * Compare two zbookmark_phys_t's to see which we would reach first in a
3908 * pre-order traversal of the object tree.
3910 * This is simple in every case aside from the meta-dnode object. For all other
3911 * objects, we traverse them in order (object 1 before object 2, and so on).
3912 * However, all of these objects are traversed while traversing object 0, since
3913 * the data it points to is the list of objects. Thus, we need to convert to a
3914 * canonical representation so we can compare meta-dnode bookmarks to
3915 * non-meta-dnode bookmarks.
3917 * We do this by calculating "equivalents" for each field of the zbookmark.
3918 * zbookmarks outside of the meta-dnode use their own object and level, and
3919 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3920 * blocks this bookmark refers to) by multiplying their blkid by their span
3921 * (the number of L0 blocks contained within one block at their level).
3922 * zbookmarks inside the meta-dnode calculate their object equivalent
3923 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3924 * level + 1<<31 (any value larger than a level could ever be) for their level.
3925 * This causes them to always compare before a bookmark in their object
3926 * equivalent, compare appropriately to bookmarks in other objects, and to
3927 * compare appropriately to other bookmarks in the meta-dnode.
3930 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3931 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3934 * These variables represent the "equivalent" values for the zbookmark,
3935 * after converting zbookmarks inside the meta dnode to their
3936 * normal-object equivalents.
3938 uint64_t zb1obj, zb2obj;
3939 uint64_t zb1L0, zb2L0;
3940 uint64_t zb1level, zb2level;
3942 if (zb1->zb_object == zb2->zb_object &&
3943 zb1->zb_level == zb2->zb_level &&
3944 zb1->zb_blkid == zb2->zb_blkid)
3948 * BP_SPANB calculates the span in blocks.
3950 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3951 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3953 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3954 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3956 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3958 zb1obj = zb1->zb_object;
3959 zb1level = zb1->zb_level;
3962 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3963 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3965 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3967 zb2obj = zb2->zb_object;
3968 zb2level = zb2->zb_level;
3971 /* Now that we have a canonical representation, do the comparison. */
3972 if (zb1obj != zb2obj)
3973 return (zb1obj < zb2obj ? -1 : 1);
3974 else if (zb1L0 != zb2L0)
3975 return (zb1L0 < zb2L0 ? -1 : 1);
3976 else if (zb1level != zb2level)
3977 return (zb1level > zb2level ? -1 : 1);
3979 * This can (theoretically) happen if the bookmarks have the same object
3980 * and level, but different blkids, if the block sizes are not the same.
3981 * There is presently no way to change the indirect block sizes
3987 * This function checks the following: given that last_block is the place that
3988 * our traversal stopped last time, does that guarantee that we've visited
3989 * every node under subtree_root? Therefore, we can't just use the raw output
3990 * of zbookmark_compare. We have to pass in a modified version of
3991 * subtree_root; by incrementing the block id, and then checking whether
3992 * last_block is before or equal to that, we can tell whether or not having
3993 * visited last_block implies that all of subtree_root's children have been
3997 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3998 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4000 zbookmark_phys_t mod_zb = *subtree_root;
4002 ASSERT(last_block->zb_level == 0);
4004 /* The objset_phys_t isn't before anything. */
4009 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4010 * data block size in sectors, because that variable is only used if
4011 * the bookmark refers to a block in the meta-dnode. Since we don't
4012 * know without examining it what object it refers to, and there's no
4013 * harm in passing in this value in other cases, we always pass it in.
4015 * We pass in 0 for the indirect block size shift because zb2 must be
4016 * level 0. The indirect block size is only used to calculate the span
4017 * of the bookmark, but since the bookmark must be level 0, the span is
4018 * always 1, so the math works out.
4020 * If you make changes to how the zbookmark_compare code works, be sure
4021 * to make sure that this code still works afterwards.
4023 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4024 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,