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_impl(size_t size, boolean_t canwait)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c],
285 canwait ? KM_PUSHPAGE : KM_NOSLEEP));
287 return (kmem_alloc(size,
288 (canwait ? KM_SLEEP : KM_NOSLEEP) | flags));
293 zio_buf_alloc(size_t size)
295 return (zio_buf_alloc_impl(size, B_TRUE));
299 zio_buf_alloc_nowait(size_t size)
301 return (zio_buf_alloc_impl(size, B_FALSE));
305 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
306 * crashdump if the kernel panics. This exists so that we will limit the amount
307 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
308 * of kernel heap dumped to disk when the kernel panics)
311 zio_data_buf_alloc(size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
320 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
324 zio_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_buf_cache[c], buf);
333 kmem_free(buf, size);
337 zio_data_buf_free(void *buf, size_t size)
339 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
341 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
344 kmem_cache_free(zio_data_buf_cache[c], buf);
346 kmem_free(buf, size);
350 * ==========================================================================
351 * Push and pop I/O transform buffers
352 * ==========================================================================
355 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
356 zio_transform_func_t *transform)
358 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
360 zt->zt_orig_data = zio->io_data;
361 zt->zt_orig_size = zio->io_size;
362 zt->zt_bufsize = bufsize;
363 zt->zt_transform = transform;
365 zt->zt_next = zio->io_transform_stack;
366 zio->io_transform_stack = zt;
373 zio_pop_transforms(zio_t *zio)
377 while ((zt = zio->io_transform_stack) != NULL) {
378 if (zt->zt_transform != NULL)
379 zt->zt_transform(zio,
380 zt->zt_orig_data, zt->zt_orig_size);
382 if (zt->zt_bufsize != 0)
383 zio_buf_free(zio->io_data, zt->zt_bufsize);
385 zio->io_data = zt->zt_orig_data;
386 zio->io_size = zt->zt_orig_size;
387 zio->io_transform_stack = zt->zt_next;
389 kmem_free(zt, sizeof (zio_transform_t));
394 * ==========================================================================
395 * I/O transform callbacks for subblocks and decompression
396 * ==========================================================================
399 zio_subblock(zio_t *zio, void *data, uint64_t size)
401 ASSERT(zio->io_size > size);
403 if (zio->io_type == ZIO_TYPE_READ)
404 bcopy(zio->io_data, data, size);
408 zio_decompress(zio_t *zio, void *data, uint64_t size)
410 if (zio->io_error == 0 &&
411 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
412 zio->io_data, data, zio->io_size, size) != 0)
413 zio->io_error = SET_ERROR(EIO);
417 * ==========================================================================
418 * I/O parent/child relationships and pipeline interlocks
419 * ==========================================================================
422 zio_walk_parents(zio_t *cio, zio_link_t **zl)
424 list_t *pl = &cio->io_parent_list;
426 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
430 ASSERT((*zl)->zl_child == cio);
431 return ((*zl)->zl_parent);
435 zio_walk_children(zio_t *pio, zio_link_t **zl)
437 list_t *cl = &pio->io_child_list;
439 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
443 ASSERT((*zl)->zl_parent == pio);
444 return ((*zl)->zl_child);
448 zio_unique_parent(zio_t *cio)
450 zio_link_t *zl = NULL;
451 zio_t *pio = zio_walk_parents(cio, &zl);
453 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
458 zio_add_child(zio_t *pio, zio_t *cio)
460 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
463 * Logical I/Os can have logical, gang, or vdev children.
464 * Gang I/Os can have gang or vdev children.
465 * Vdev I/Os can only have vdev children.
466 * The following ASSERT captures all of these constraints.
468 ASSERT(cio->io_child_type <= pio->io_child_type);
473 mutex_enter(&cio->io_lock);
474 mutex_enter(&pio->io_lock);
476 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
478 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
479 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
481 list_insert_head(&pio->io_child_list, zl);
482 list_insert_head(&cio->io_parent_list, zl);
484 pio->io_child_count++;
485 cio->io_parent_count++;
487 mutex_exit(&pio->io_lock);
488 mutex_exit(&cio->io_lock);
492 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
494 ASSERT(zl->zl_parent == pio);
495 ASSERT(zl->zl_child == cio);
497 mutex_enter(&cio->io_lock);
498 mutex_enter(&pio->io_lock);
500 list_remove(&pio->io_child_list, zl);
501 list_remove(&cio->io_parent_list, zl);
503 pio->io_child_count--;
504 cio->io_parent_count--;
506 mutex_exit(&pio->io_lock);
507 mutex_exit(&cio->io_lock);
509 kmem_cache_free(zio_link_cache, zl);
513 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
515 uint64_t *countp = &zio->io_children[child][wait];
516 boolean_t waiting = B_FALSE;
518 mutex_enter(&zio->io_lock);
519 ASSERT(zio->io_stall == NULL);
522 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
523 zio->io_stall = countp;
526 mutex_exit(&zio->io_lock);
532 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
534 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
535 int *errorp = &pio->io_child_error[zio->io_child_type];
537 mutex_enter(&pio->io_lock);
538 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
539 *errorp = zio_worst_error(*errorp, zio->io_error);
540 pio->io_reexecute |= zio->io_reexecute;
541 ASSERT3U(*countp, >, 0);
545 if (*countp == 0 && pio->io_stall == countp) {
546 zio_taskq_type_t type =
547 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
549 pio->io_stall = NULL;
550 mutex_exit(&pio->io_lock);
552 * Dispatch the parent zio in its own taskq so that
553 * the child can continue to make progress. This also
554 * prevents overflowing the stack when we have deeply nested
555 * parent-child relationships.
557 zio_taskq_dispatch(pio, type, B_FALSE);
559 mutex_exit(&pio->io_lock);
564 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
566 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
567 zio->io_error = zio->io_child_error[c];
571 zio_timestamp_compare(const void *x1, const void *x2)
573 const zio_t *z1 = x1;
574 const zio_t *z2 = x2;
576 if (z1->io_queued_timestamp < z2->io_queued_timestamp)
578 if (z1->io_queued_timestamp > z2->io_queued_timestamp)
581 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
583 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
586 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
588 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
591 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
593 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
596 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
598 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
610 * ==========================================================================
611 * Create the various types of I/O (read, write, free, etc)
612 * ==========================================================================
615 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
616 void *data, uint64_t size, zio_done_func_t *done, void *private,
617 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
618 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
619 enum zio_stage stage, enum zio_stage pipeline)
623 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
624 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
625 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
627 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
628 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
629 ASSERT(vd || stage == ZIO_STAGE_OPEN);
631 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
632 bzero(zio, sizeof (zio_t));
634 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
635 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
637 list_create(&zio->io_parent_list, sizeof (zio_link_t),
638 offsetof(zio_link_t, zl_parent_node));
639 list_create(&zio->io_child_list, sizeof (zio_link_t),
640 offsetof(zio_link_t, zl_child_node));
643 zio->io_child_type = ZIO_CHILD_VDEV;
644 else if (flags & ZIO_FLAG_GANG_CHILD)
645 zio->io_child_type = ZIO_CHILD_GANG;
646 else if (flags & ZIO_FLAG_DDT_CHILD)
647 zio->io_child_type = ZIO_CHILD_DDT;
649 zio->io_child_type = ZIO_CHILD_LOGICAL;
652 zio->io_bp = (blkptr_t *)bp;
653 zio->io_bp_copy = *bp;
654 zio->io_bp_orig = *bp;
655 if (type != ZIO_TYPE_WRITE ||
656 zio->io_child_type == ZIO_CHILD_DDT)
657 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
658 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
659 zio->io_logical = zio;
660 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
661 pipeline |= ZIO_GANG_STAGES;
667 zio->io_private = private;
669 zio->io_priority = priority;
671 zio->io_offset = offset;
672 zio->io_orig_data = zio->io_data = data;
673 zio->io_orig_size = zio->io_size = size;
674 zio->io_orig_flags = zio->io_flags = flags;
675 zio->io_orig_stage = zio->io_stage = stage;
676 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
677 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
679 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
680 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
683 zio->io_bookmark = *zb;
686 if (zio->io_logical == NULL)
687 zio->io_logical = pio->io_logical;
688 if (zio->io_child_type == ZIO_CHILD_GANG)
689 zio->io_gang_leader = pio->io_gang_leader;
690 zio_add_child(pio, zio);
697 zio_destroy(zio_t *zio)
699 list_destroy(&zio->io_parent_list);
700 list_destroy(&zio->io_child_list);
701 mutex_destroy(&zio->io_lock);
702 cv_destroy(&zio->io_cv);
703 kmem_cache_free(zio_cache, zio);
707 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
708 void *private, enum zio_flag flags)
712 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
713 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
714 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
720 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
722 return (zio_null(NULL, spa, NULL, done, private, flags));
726 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
728 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
729 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
730 bp, (longlong_t)BP_GET_TYPE(bp));
732 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
733 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
734 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
735 bp, (longlong_t)BP_GET_CHECKSUM(bp));
737 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
738 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
739 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
740 bp, (longlong_t)BP_GET_COMPRESS(bp));
742 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
743 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
744 bp, (longlong_t)BP_GET_LSIZE(bp));
746 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
747 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
748 bp, (longlong_t)BP_GET_PSIZE(bp));
751 if (BP_IS_EMBEDDED(bp)) {
752 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
753 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
754 bp, (longlong_t)BPE_GET_ETYPE(bp));
759 * Pool-specific checks.
761 * Note: it would be nice to verify that the blk_birth and
762 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
763 * allows the birth time of log blocks (and dmu_sync()-ed blocks
764 * that are in the log) to be arbitrarily large.
766 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
767 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
768 if (vdevid >= spa->spa_root_vdev->vdev_children) {
769 zfs_panic_recover("blkptr at %p DVA %u has invalid "
771 bp, i, (longlong_t)vdevid);
774 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
776 zfs_panic_recover("blkptr at %p DVA %u has invalid "
778 bp, i, (longlong_t)vdevid);
781 if (vd->vdev_ops == &vdev_hole_ops) {
782 zfs_panic_recover("blkptr at %p DVA %u has hole "
784 bp, i, (longlong_t)vdevid);
787 if (vd->vdev_ops == &vdev_missing_ops) {
789 * "missing" vdevs are valid during import, but we
790 * don't have their detailed info (e.g. asize), so
791 * we can't perform any more checks on them.
795 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
796 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
798 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
799 if (offset + asize > vd->vdev_asize) {
800 zfs_panic_recover("blkptr at %p DVA %u has invalid "
802 bp, i, (longlong_t)offset);
808 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
809 void *data, uint64_t size, zio_done_func_t *done, void *private,
810 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
814 zfs_blkptr_verify(spa, bp);
816 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
817 data, size, done, private,
818 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
819 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
820 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
826 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
827 void *data, uint64_t size, const zio_prop_t *zp,
828 zio_done_func_t *ready, zio_done_func_t *children_ready,
829 zio_done_func_t *physdone, zio_done_func_t *done,
830 void *private, zio_priority_t priority, enum zio_flag flags,
831 const zbookmark_phys_t *zb)
835 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
836 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
837 zp->zp_compress >= ZIO_COMPRESS_OFF &&
838 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
839 DMU_OT_IS_VALID(zp->zp_type) &&
842 zp->zp_copies <= spa_max_replication(spa));
844 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
845 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
846 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
847 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
849 zio->io_ready = ready;
850 zio->io_children_ready = children_ready;
851 zio->io_physdone = physdone;
855 * Data can be NULL if we are going to call zio_write_override() to
856 * provide the already-allocated BP. But we may need the data to
857 * verify a dedup hit (if requested). In this case, don't try to
858 * dedup (just take the already-allocated BP verbatim).
860 if (data == NULL && zio->io_prop.zp_dedup_verify) {
861 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
868 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
869 uint64_t size, zio_done_func_t *done, void *private,
870 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
874 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
875 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
876 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
882 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
884 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
885 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
886 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
887 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
890 * We must reset the io_prop to match the values that existed
891 * when the bp was first written by dmu_sync() keeping in mind
892 * that nopwrite and dedup are mutually exclusive.
894 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
895 zio->io_prop.zp_nopwrite = nopwrite;
896 zio->io_prop.zp_copies = copies;
897 zio->io_bp_override = bp;
901 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
905 * The check for EMBEDDED is a performance optimization. We
906 * process the free here (by ignoring it) rather than
907 * putting it on the list and then processing it in zio_free_sync().
909 if (BP_IS_EMBEDDED(bp))
911 metaslab_check_free(spa, bp);
914 * Frees that are for the currently-syncing txg, are not going to be
915 * deferred, and which will not need to do a read (i.e. not GANG or
916 * DEDUP), can be processed immediately. Otherwise, put them on the
917 * in-memory list for later processing.
919 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
920 txg != spa->spa_syncing_txg ||
921 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
922 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
924 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
925 BP_GET_PSIZE(bp), 0)));
930 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
931 uint64_t size, enum zio_flag flags)
934 enum zio_stage stage = ZIO_FREE_PIPELINE;
936 ASSERT(!BP_IS_HOLE(bp));
937 ASSERT(spa_syncing_txg(spa) == txg);
938 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
940 if (BP_IS_EMBEDDED(bp))
941 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
943 metaslab_check_free(spa, bp);
946 if (zfs_trim_enabled)
947 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
948 ZIO_STAGE_VDEV_IO_ASSESS;
950 * GANG and DEDUP blocks can induce a read (for the gang block header,
951 * or the DDT), so issue them asynchronously so that this thread is
954 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
955 stage |= ZIO_STAGE_ISSUE_ASYNC;
957 flags |= ZIO_FLAG_DONT_QUEUE;
959 zio = zio_create(pio, spa, txg, bp, NULL, size,
960 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
961 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
967 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
968 zio_done_func_t *done, void *private, enum zio_flag flags)
972 dprintf_bp(bp, "claiming in txg %llu", txg);
974 if (BP_IS_EMBEDDED(bp))
975 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
978 * A claim is an allocation of a specific block. Claims are needed
979 * to support immediate writes in the intent log. The issue is that
980 * immediate writes contain committed data, but in a txg that was
981 * *not* committed. Upon opening the pool after an unclean shutdown,
982 * the intent log claims all blocks that contain immediate write data
983 * so that the SPA knows they're in use.
985 * All claims *must* be resolved in the first txg -- before the SPA
986 * starts allocating blocks -- so that nothing is allocated twice.
987 * If txg == 0 we just verify that the block is claimable.
989 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
990 ASSERT(txg == spa_first_txg(spa) || txg == 0);
991 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
993 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
994 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
995 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
996 ASSERT0(zio->io_queued_timestamp);
1002 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1003 uint64_t size, zio_done_func_t *done, void *private,
1004 zio_priority_t priority, enum zio_flag flags)
1009 if (vd->vdev_children == 0) {
1010 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
1011 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
1012 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1016 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1018 for (c = 0; c < vd->vdev_children; c++)
1019 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1020 offset, size, done, private, priority, flags));
1027 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1028 void *data, int checksum, zio_done_func_t *done, void *private,
1029 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1033 ASSERT(vd->vdev_children == 0);
1034 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1035 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1036 ASSERT3U(offset + size, <=, vd->vdev_psize);
1038 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1039 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1040 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1042 zio->io_prop.zp_checksum = checksum;
1048 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1049 void *data, int checksum, zio_done_func_t *done, void *private,
1050 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1054 ASSERT(vd->vdev_children == 0);
1055 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1056 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1057 ASSERT3U(offset + size, <=, vd->vdev_psize);
1059 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1060 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1061 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1063 zio->io_prop.zp_checksum = checksum;
1065 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1067 * zec checksums are necessarily destructive -- they modify
1068 * the end of the write buffer to hold the verifier/checksum.
1069 * Therefore, we must make a local copy in case the data is
1070 * being written to multiple places in parallel.
1072 void *wbuf = zio_buf_alloc(size);
1073 bcopy(data, wbuf, size);
1074 zio_push_transform(zio, wbuf, size, size, NULL);
1081 * Create a child I/O to do some work for us.
1084 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1085 void *data, uint64_t size, int type, zio_priority_t priority,
1086 enum zio_flag flags, zio_done_func_t *done, void *private)
1088 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1091 ASSERT(vd->vdev_parent ==
1092 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1094 if (type == ZIO_TYPE_READ && bp != NULL) {
1096 * If we have the bp, then the child should perform the
1097 * checksum and the parent need not. This pushes error
1098 * detection as close to the leaves as possible and
1099 * eliminates redundant checksums in the interior nodes.
1101 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1102 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1105 /* Not all IO types require vdev io done stage e.g. free */
1106 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1107 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1109 if (vd->vdev_children == 0)
1110 offset += VDEV_LABEL_START_SIZE;
1112 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1115 * If we've decided to do a repair, the write is not speculative --
1116 * even if the original read was.
1118 if (flags & ZIO_FLAG_IO_REPAIR)
1119 flags &= ~ZIO_FLAG_SPECULATIVE;
1122 * If we're creating a child I/O that is not associated with a
1123 * top-level vdev, then the child zio is not an allocating I/O.
1124 * If this is a retried I/O then we ignore it since we will
1125 * have already processed the original allocating I/O.
1127 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1128 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1129 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1131 ASSERT(mc->mc_alloc_throttle_enabled);
1132 ASSERT(type == ZIO_TYPE_WRITE);
1133 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1134 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1135 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1136 pio->io_child_type == ZIO_CHILD_GANG);
1138 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1141 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1142 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1143 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1144 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1146 zio->io_physdone = pio->io_physdone;
1147 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1148 zio->io_logical->io_phys_children++;
1154 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1155 int type, zio_priority_t priority, enum zio_flag flags,
1156 zio_done_func_t *done, void *private)
1160 ASSERT(vd->vdev_ops->vdev_op_leaf);
1162 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1163 data, size, done, private, type, priority,
1164 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1166 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1172 zio_flush(zio_t *zio, vdev_t *vd)
1174 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1175 NULL, NULL, ZIO_PRIORITY_NOW,
1176 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1180 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1183 ASSERT(vd->vdev_ops->vdev_op_leaf);
1185 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1186 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1187 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1188 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1192 zio_shrink(zio_t *zio, uint64_t size)
1194 ASSERT(zio->io_executor == NULL);
1195 ASSERT(zio->io_orig_size == zio->io_size);
1196 ASSERT(size <= zio->io_size);
1199 * We don't shrink for raidz because of problems with the
1200 * reconstruction when reading back less than the block size.
1201 * Note, BP_IS_RAIDZ() assumes no compression.
1203 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1204 if (!BP_IS_RAIDZ(zio->io_bp))
1205 zio->io_orig_size = zio->io_size = size;
1209 * ==========================================================================
1210 * Prepare to read and write logical blocks
1211 * ==========================================================================
1215 zio_read_bp_init(zio_t *zio)
1217 blkptr_t *bp = zio->io_bp;
1219 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1220 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1221 !(zio->io_flags & ZIO_FLAG_RAW)) {
1223 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1224 void *cbuf = zio_buf_alloc(psize);
1226 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1229 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1230 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1231 decode_embedded_bp_compressed(bp, zio->io_data);
1233 ASSERT(!BP_IS_EMBEDDED(bp));
1236 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1237 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1239 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1240 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1242 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1243 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1245 return (ZIO_PIPELINE_CONTINUE);
1249 zio_write_bp_init(zio_t *zio)
1251 if (!IO_IS_ALLOCATING(zio))
1252 return (ZIO_PIPELINE_CONTINUE);
1254 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1256 if (zio->io_bp_override) {
1257 blkptr_t *bp = zio->io_bp;
1258 zio_prop_t *zp = &zio->io_prop;
1260 ASSERT(bp->blk_birth != zio->io_txg);
1261 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1263 *bp = *zio->io_bp_override;
1264 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1266 if (BP_IS_EMBEDDED(bp))
1267 return (ZIO_PIPELINE_CONTINUE);
1270 * If we've been overridden and nopwrite is set then
1271 * set the flag accordingly to indicate that a nopwrite
1272 * has already occurred.
1274 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1275 ASSERT(!zp->zp_dedup);
1276 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1277 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1278 return (ZIO_PIPELINE_CONTINUE);
1281 ASSERT(!zp->zp_nopwrite);
1283 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1284 return (ZIO_PIPELINE_CONTINUE);
1286 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1287 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1289 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1290 BP_SET_DEDUP(bp, 1);
1291 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1292 return (ZIO_PIPELINE_CONTINUE);
1296 * We were unable to handle this as an override bp, treat
1297 * it as a regular write I/O.
1299 zio->io_bp_override = NULL;
1300 *bp = zio->io_bp_orig;
1301 zio->io_pipeline = zio->io_orig_pipeline;
1304 return (ZIO_PIPELINE_CONTINUE);
1308 zio_write_compress(zio_t *zio)
1310 spa_t *spa = zio->io_spa;
1311 zio_prop_t *zp = &zio->io_prop;
1312 enum zio_compress compress = zp->zp_compress;
1313 blkptr_t *bp = zio->io_bp;
1314 uint64_t lsize = zio->io_size;
1315 uint64_t psize = lsize;
1319 * If our children haven't all reached the ready stage,
1320 * wait for them and then repeat this pipeline stage.
1322 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1323 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1324 return (ZIO_PIPELINE_STOP);
1326 if (!IO_IS_ALLOCATING(zio))
1327 return (ZIO_PIPELINE_CONTINUE);
1329 if (zio->io_children_ready != NULL) {
1331 * Now that all our children are ready, run the callback
1332 * associated with this zio in case it wants to modify the
1333 * data to be written.
1335 ASSERT3U(zp->zp_level, >, 0);
1336 zio->io_children_ready(zio);
1339 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1340 ASSERT(zio->io_bp_override == NULL);
1342 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1344 * We're rewriting an existing block, which means we're
1345 * working on behalf of spa_sync(). For spa_sync() to
1346 * converge, it must eventually be the case that we don't
1347 * have to allocate new blocks. But compression changes
1348 * the blocksize, which forces a reallocate, and makes
1349 * convergence take longer. Therefore, after the first
1350 * few passes, stop compressing to ensure convergence.
1352 pass = spa_sync_pass(spa);
1354 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1355 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1356 ASSERT(!BP_GET_DEDUP(bp));
1358 if (pass >= zfs_sync_pass_dont_compress)
1359 compress = ZIO_COMPRESS_OFF;
1361 /* Make sure someone doesn't change their mind on overwrites */
1362 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1363 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1366 if (compress != ZIO_COMPRESS_OFF) {
1367 void *cbuf = zio_buf_alloc(lsize);
1368 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1369 if (psize == 0 || psize == lsize) {
1370 compress = ZIO_COMPRESS_OFF;
1371 zio_buf_free(cbuf, lsize);
1372 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1373 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1374 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1375 encode_embedded_bp_compressed(bp,
1376 cbuf, compress, lsize, psize);
1377 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1378 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1379 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1380 zio_buf_free(cbuf, lsize);
1381 bp->blk_birth = zio->io_txg;
1382 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1383 ASSERT(spa_feature_is_active(spa,
1384 SPA_FEATURE_EMBEDDED_DATA));
1385 return (ZIO_PIPELINE_CONTINUE);
1388 * Round up compressed size up to the ashift
1389 * of the smallest-ashift device, and zero the tail.
1390 * This ensures that the compressed size of the BP
1391 * (and thus compressratio property) are correct,
1392 * in that we charge for the padding used to fill out
1395 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1396 size_t rounded = (size_t)P2ROUNDUP(psize,
1397 1ULL << spa->spa_min_ashift);
1398 if (rounded >= lsize) {
1399 compress = ZIO_COMPRESS_OFF;
1400 zio_buf_free(cbuf, lsize);
1403 bzero((char *)cbuf + psize, rounded - psize);
1405 zio_push_transform(zio, cbuf,
1406 psize, lsize, NULL);
1411 * We were unable to handle this as an override bp, treat
1412 * it as a regular write I/O.
1414 zio->io_bp_override = NULL;
1415 *bp = zio->io_bp_orig;
1416 zio->io_pipeline = zio->io_orig_pipeline;
1420 * The final pass of spa_sync() must be all rewrites, but the first
1421 * few passes offer a trade-off: allocating blocks defers convergence,
1422 * but newly allocated blocks are sequential, so they can be written
1423 * to disk faster. Therefore, we allow the first few passes of
1424 * spa_sync() to allocate new blocks, but force rewrites after that.
1425 * There should only be a handful of blocks after pass 1 in any case.
1427 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1428 BP_GET_PSIZE(bp) == psize &&
1429 pass >= zfs_sync_pass_rewrite) {
1431 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1432 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1433 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1436 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1440 if (zio->io_bp_orig.blk_birth != 0 &&
1441 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1442 BP_SET_LSIZE(bp, lsize);
1443 BP_SET_TYPE(bp, zp->zp_type);
1444 BP_SET_LEVEL(bp, zp->zp_level);
1445 BP_SET_BIRTH(bp, zio->io_txg, 0);
1447 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1449 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1450 BP_SET_LSIZE(bp, lsize);
1451 BP_SET_TYPE(bp, zp->zp_type);
1452 BP_SET_LEVEL(bp, zp->zp_level);
1453 BP_SET_PSIZE(bp, psize);
1454 BP_SET_COMPRESS(bp, compress);
1455 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1456 BP_SET_DEDUP(bp, zp->zp_dedup);
1457 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1459 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1460 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1461 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1463 if (zp->zp_nopwrite) {
1464 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1465 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1466 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1469 return (ZIO_PIPELINE_CONTINUE);
1473 zio_free_bp_init(zio_t *zio)
1475 blkptr_t *bp = zio->io_bp;
1477 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1478 if (BP_GET_DEDUP(bp))
1479 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1482 return (ZIO_PIPELINE_CONTINUE);
1486 * ==========================================================================
1487 * Execute the I/O pipeline
1488 * ==========================================================================
1492 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1494 spa_t *spa = zio->io_spa;
1495 zio_type_t t = zio->io_type;
1496 int flags = (cutinline ? TQ_FRONT : 0);
1498 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1501 * If we're a config writer or a probe, the normal issue and
1502 * interrupt threads may all be blocked waiting for the config lock.
1503 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1505 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1509 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1511 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1515 * If this is a high priority I/O, then use the high priority taskq if
1518 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1519 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1522 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1525 * NB: We are assuming that the zio can only be dispatched
1526 * to a single taskq at a time. It would be a grievous error
1527 * to dispatch the zio to another taskq at the same time.
1529 #if defined(illumos) || !defined(_KERNEL)
1530 ASSERT(zio->io_tqent.tqent_next == NULL);
1532 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1534 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1535 flags, &zio->io_tqent);
1539 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1541 kthread_t *executor = zio->io_executor;
1542 spa_t *spa = zio->io_spa;
1544 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1545 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1547 for (i = 0; i < tqs->stqs_count; i++) {
1548 if (taskq_member(tqs->stqs_taskq[i], executor))
1557 zio_issue_async(zio_t *zio)
1559 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1561 return (ZIO_PIPELINE_STOP);
1565 zio_interrupt(zio_t *zio)
1567 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1571 zio_delay_interrupt(zio_t *zio)
1574 * The timeout_generic() function isn't defined in userspace, so
1575 * rather than trying to implement the function, the zio delay
1576 * functionality has been disabled for userspace builds.
1581 * If io_target_timestamp is zero, then no delay has been registered
1582 * for this IO, thus jump to the end of this function and "skip" the
1583 * delay; issuing it directly to the zio layer.
1585 if (zio->io_target_timestamp != 0) {
1586 hrtime_t now = gethrtime();
1588 if (now >= zio->io_target_timestamp) {
1590 * This IO has already taken longer than the target
1591 * delay to complete, so we don't want to delay it
1592 * any longer; we "miss" the delay and issue it
1593 * directly to the zio layer. This is likely due to
1594 * the target latency being set to a value less than
1595 * the underlying hardware can satisfy (e.g. delay
1596 * set to 1ms, but the disks take 10ms to complete an
1600 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1605 hrtime_t diff = zio->io_target_timestamp - now;
1607 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1608 hrtime_t, now, hrtime_t, diff);
1610 (void) timeout_generic(CALLOUT_NORMAL,
1611 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1618 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1623 * Execute the I/O pipeline until one of the following occurs:
1625 * (1) the I/O completes
1626 * (2) the pipeline stalls waiting for dependent child I/Os
1627 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1628 * (4) the I/O is delegated by vdev-level caching or aggregation
1629 * (5) the I/O is deferred due to vdev-level queueing
1630 * (6) the I/O is handed off to another thread.
1632 * In all cases, the pipeline stops whenever there's no CPU work; it never
1633 * burns a thread in cv_wait().
1635 * There's no locking on io_stage because there's no legitimate way
1636 * for multiple threads to be attempting to process the same I/O.
1638 static zio_pipe_stage_t *zio_pipeline[];
1641 zio_execute(zio_t *zio)
1643 zio->io_executor = curthread;
1645 ASSERT3U(zio->io_queued_timestamp, >, 0);
1647 while (zio->io_stage < ZIO_STAGE_DONE) {
1648 enum zio_stage pipeline = zio->io_pipeline;
1649 enum zio_stage stage = zio->io_stage;
1652 ASSERT(!MUTEX_HELD(&zio->io_lock));
1653 ASSERT(ISP2(stage));
1654 ASSERT(zio->io_stall == NULL);
1658 } while ((stage & pipeline) == 0);
1660 ASSERT(stage <= ZIO_STAGE_DONE);
1663 * If we are in interrupt context and this pipeline stage
1664 * will grab a config lock that is held across I/O,
1665 * or may wait for an I/O that needs an interrupt thread
1666 * to complete, issue async to avoid deadlock.
1668 * For VDEV_IO_START, we cut in line so that the io will
1669 * be sent to disk promptly.
1671 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1672 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1673 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1674 zio_requeue_io_start_cut_in_line : B_FALSE;
1675 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1679 zio->io_stage = stage;
1680 zio->io_pipeline_trace |= zio->io_stage;
1681 rv = zio_pipeline[highbit64(stage) - 1](zio);
1683 if (rv == ZIO_PIPELINE_STOP)
1686 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1691 * ==========================================================================
1692 * Initiate I/O, either sync or async
1693 * ==========================================================================
1696 zio_wait(zio_t *zio)
1700 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1701 ASSERT(zio->io_executor == NULL);
1703 zio->io_waiter = curthread;
1704 ASSERT0(zio->io_queued_timestamp);
1705 zio->io_queued_timestamp = gethrtime();
1709 mutex_enter(&zio->io_lock);
1710 while (zio->io_executor != NULL)
1711 cv_wait(&zio->io_cv, &zio->io_lock);
1712 mutex_exit(&zio->io_lock);
1714 error = zio->io_error;
1721 zio_nowait(zio_t *zio)
1723 ASSERT(zio->io_executor == NULL);
1725 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1726 zio_unique_parent(zio) == NULL) {
1728 * This is a logical async I/O with no parent to wait for it.
1729 * We add it to the spa_async_root_zio "Godfather" I/O which
1730 * will ensure they complete prior to unloading the pool.
1732 spa_t *spa = zio->io_spa;
1734 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1737 ASSERT0(zio->io_queued_timestamp);
1738 zio->io_queued_timestamp = gethrtime();
1743 * ==========================================================================
1744 * Reexecute or suspend/resume failed I/O
1745 * ==========================================================================
1749 zio_reexecute(zio_t *pio)
1751 zio_t *cio, *cio_next;
1753 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1754 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1755 ASSERT(pio->io_gang_leader == NULL);
1756 ASSERT(pio->io_gang_tree == NULL);
1758 pio->io_flags = pio->io_orig_flags;
1759 pio->io_stage = pio->io_orig_stage;
1760 pio->io_pipeline = pio->io_orig_pipeline;
1761 pio->io_reexecute = 0;
1762 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1763 pio->io_pipeline_trace = 0;
1765 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1766 pio->io_state[w] = 0;
1767 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1768 pio->io_child_error[c] = 0;
1770 if (IO_IS_ALLOCATING(pio))
1771 BP_ZERO(pio->io_bp);
1774 * As we reexecute pio's children, new children could be created.
1775 * New children go to the head of pio's io_child_list, however,
1776 * so we will (correctly) not reexecute them. The key is that
1777 * the remainder of pio's io_child_list, from 'cio_next' onward,
1778 * cannot be affected by any side effects of reexecuting 'cio'.
1780 zio_link_t *zl = NULL;
1781 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1782 cio_next = zio_walk_children(pio, &zl);
1783 mutex_enter(&pio->io_lock);
1784 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1785 pio->io_children[cio->io_child_type][w]++;
1786 mutex_exit(&pio->io_lock);
1791 * Now that all children have been reexecuted, execute the parent.
1792 * We don't reexecute "The Godfather" I/O here as it's the
1793 * responsibility of the caller to wait on him.
1795 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1796 pio->io_queued_timestamp = gethrtime();
1802 zio_suspend(spa_t *spa, zio_t *zio)
1804 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1805 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1806 "failure and the failure mode property for this pool "
1807 "is set to panic.", spa_name(spa));
1809 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1811 mutex_enter(&spa->spa_suspend_lock);
1813 if (spa->spa_suspend_zio_root == NULL)
1814 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1815 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1816 ZIO_FLAG_GODFATHER);
1818 spa->spa_suspended = B_TRUE;
1821 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1822 ASSERT(zio != spa->spa_suspend_zio_root);
1823 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1824 ASSERT(zio_unique_parent(zio) == NULL);
1825 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1826 zio_add_child(spa->spa_suspend_zio_root, zio);
1829 mutex_exit(&spa->spa_suspend_lock);
1833 zio_resume(spa_t *spa)
1838 * Reexecute all previously suspended i/o.
1840 mutex_enter(&spa->spa_suspend_lock);
1841 spa->spa_suspended = B_FALSE;
1842 cv_broadcast(&spa->spa_suspend_cv);
1843 pio = spa->spa_suspend_zio_root;
1844 spa->spa_suspend_zio_root = NULL;
1845 mutex_exit(&spa->spa_suspend_lock);
1851 return (zio_wait(pio));
1855 zio_resume_wait(spa_t *spa)
1857 mutex_enter(&spa->spa_suspend_lock);
1858 while (spa_suspended(spa))
1859 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1860 mutex_exit(&spa->spa_suspend_lock);
1864 * ==========================================================================
1867 * A gang block is a collection of small blocks that looks to the DMU
1868 * like one large block. When zio_dva_allocate() cannot find a block
1869 * of the requested size, due to either severe fragmentation or the pool
1870 * being nearly full, it calls zio_write_gang_block() to construct the
1871 * block from smaller fragments.
1873 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1874 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1875 * an indirect block: it's an array of block pointers. It consumes
1876 * only one sector and hence is allocatable regardless of fragmentation.
1877 * The gang header's bps point to its gang members, which hold the data.
1879 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1880 * as the verifier to ensure uniqueness of the SHA256 checksum.
1881 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1882 * not the gang header. This ensures that data block signatures (needed for
1883 * deduplication) are independent of how the block is physically stored.
1885 * Gang blocks can be nested: a gang member may itself be a gang block.
1886 * Thus every gang block is a tree in which root and all interior nodes are
1887 * gang headers, and the leaves are normal blocks that contain user data.
1888 * The root of the gang tree is called the gang leader.
1890 * To perform any operation (read, rewrite, free, claim) on a gang block,
1891 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1892 * in the io_gang_tree field of the original logical i/o by recursively
1893 * reading the gang leader and all gang headers below it. This yields
1894 * an in-core tree containing the contents of every gang header and the
1895 * bps for every constituent of the gang block.
1897 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1898 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1899 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1900 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1901 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1902 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1903 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1904 * of the gang header plus zio_checksum_compute() of the data to update the
1905 * gang header's blk_cksum as described above.
1907 * The two-phase assemble/issue model solves the problem of partial failure --
1908 * what if you'd freed part of a gang block but then couldn't read the
1909 * gang header for another part? Assembling the entire gang tree first
1910 * ensures that all the necessary gang header I/O has succeeded before
1911 * starting the actual work of free, claim, or write. Once the gang tree
1912 * is assembled, free and claim are in-memory operations that cannot fail.
1914 * In the event that a gang write fails, zio_dva_unallocate() walks the
1915 * gang tree to immediately free (i.e. insert back into the space map)
1916 * everything we've allocated. This ensures that we don't get ENOSPC
1917 * errors during repeated suspend/resume cycles due to a flaky device.
1919 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1920 * the gang tree, we won't modify the block, so we can safely defer the free
1921 * (knowing that the block is still intact). If we *can* assemble the gang
1922 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1923 * each constituent bp and we can allocate a new block on the next sync pass.
1925 * In all cases, the gang tree allows complete recovery from partial failure.
1926 * ==========================================================================
1930 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1935 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1936 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1937 &pio->io_bookmark));
1941 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1946 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1947 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1948 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1950 * As we rewrite each gang header, the pipeline will compute
1951 * a new gang block header checksum for it; but no one will
1952 * compute a new data checksum, so we do that here. The one
1953 * exception is the gang leader: the pipeline already computed
1954 * its data checksum because that stage precedes gang assembly.
1955 * (Presently, nothing actually uses interior data checksums;
1956 * this is just good hygiene.)
1958 if (gn != pio->io_gang_leader->io_gang_tree) {
1959 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1960 data, BP_GET_PSIZE(bp));
1963 * If we are here to damage data for testing purposes,
1964 * leave the GBH alone so that we can detect the damage.
1966 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1967 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1969 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1970 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1971 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1979 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1981 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1982 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1983 ZIO_GANG_CHILD_FLAGS(pio)));
1988 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1990 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1991 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1994 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2003 static void zio_gang_tree_assemble_done(zio_t *zio);
2005 static zio_gang_node_t *
2006 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2008 zio_gang_node_t *gn;
2010 ASSERT(*gnpp == NULL);
2012 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2013 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2020 zio_gang_node_free(zio_gang_node_t **gnpp)
2022 zio_gang_node_t *gn = *gnpp;
2024 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2025 ASSERT(gn->gn_child[g] == NULL);
2027 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2028 kmem_free(gn, sizeof (*gn));
2033 zio_gang_tree_free(zio_gang_node_t **gnpp)
2035 zio_gang_node_t *gn = *gnpp;
2040 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2041 zio_gang_tree_free(&gn->gn_child[g]);
2043 zio_gang_node_free(gnpp);
2047 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2049 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2051 ASSERT(gio->io_gang_leader == gio);
2052 ASSERT(BP_IS_GANG(bp));
2054 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
2055 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2056 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2060 zio_gang_tree_assemble_done(zio_t *zio)
2062 zio_t *gio = zio->io_gang_leader;
2063 zio_gang_node_t *gn = zio->io_private;
2064 blkptr_t *bp = zio->io_bp;
2066 ASSERT(gio == zio_unique_parent(zio));
2067 ASSERT(zio->io_child_count == 0);
2072 if (BP_SHOULD_BYTESWAP(bp))
2073 byteswap_uint64_array(zio->io_data, zio->io_size);
2075 ASSERT(zio->io_data == gn->gn_gbh);
2076 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2077 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2079 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2080 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2081 if (!BP_IS_GANG(gbp))
2083 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2088 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2090 zio_t *gio = pio->io_gang_leader;
2093 ASSERT(BP_IS_GANG(bp) == !!gn);
2094 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2095 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2098 * If you're a gang header, your data is in gn->gn_gbh.
2099 * If you're a gang member, your data is in 'data' and gn == NULL.
2101 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2104 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2106 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2107 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2108 if (BP_IS_HOLE(gbp))
2110 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
2111 data = (char *)data + BP_GET_PSIZE(gbp);
2115 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2116 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2123 zio_gang_assemble(zio_t *zio)
2125 blkptr_t *bp = zio->io_bp;
2127 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2128 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2130 zio->io_gang_leader = zio;
2132 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2134 return (ZIO_PIPELINE_CONTINUE);
2138 zio_gang_issue(zio_t *zio)
2140 blkptr_t *bp = zio->io_bp;
2142 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2143 return (ZIO_PIPELINE_STOP);
2145 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2146 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2148 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2149 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2151 zio_gang_tree_free(&zio->io_gang_tree);
2153 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2155 return (ZIO_PIPELINE_CONTINUE);
2159 zio_write_gang_member_ready(zio_t *zio)
2161 zio_t *pio = zio_unique_parent(zio);
2162 zio_t *gio = zio->io_gang_leader;
2163 dva_t *cdva = zio->io_bp->blk_dva;
2164 dva_t *pdva = pio->io_bp->blk_dva;
2167 if (BP_IS_HOLE(zio->io_bp))
2170 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2172 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2173 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2174 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2175 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2176 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2178 mutex_enter(&pio->io_lock);
2179 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2180 ASSERT(DVA_GET_GANG(&pdva[d]));
2181 asize = DVA_GET_ASIZE(&pdva[d]);
2182 asize += DVA_GET_ASIZE(&cdva[d]);
2183 DVA_SET_ASIZE(&pdva[d], asize);
2185 mutex_exit(&pio->io_lock);
2189 zio_write_gang_block(zio_t *pio)
2191 spa_t *spa = pio->io_spa;
2192 metaslab_class_t *mc = spa_normal_class(spa);
2193 blkptr_t *bp = pio->io_bp;
2194 zio_t *gio = pio->io_gang_leader;
2196 zio_gang_node_t *gn, **gnpp;
2197 zio_gbh_phys_t *gbh;
2198 uint64_t txg = pio->io_txg;
2199 uint64_t resid = pio->io_size;
2201 int copies = gio->io_prop.zp_copies;
2202 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2206 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2207 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2208 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2209 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2211 flags |= METASLAB_ASYNC_ALLOC;
2212 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2215 * The logical zio has already placed a reservation for
2216 * 'copies' allocation slots but gang blocks may require
2217 * additional copies. These additional copies
2218 * (i.e. gbh_copies - copies) are guaranteed to succeed
2219 * since metaslab_class_throttle_reserve() always allows
2220 * additional reservations for gang blocks.
2222 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2226 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2227 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio);
2229 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2230 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2231 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2234 * If we failed to allocate the gang block header then
2235 * we remove any additional allocation reservations that
2236 * we placed here. The original reservation will
2237 * be removed when the logical I/O goes to the ready
2240 metaslab_class_throttle_unreserve(mc,
2241 gbh_copies - copies, pio);
2243 pio->io_error = error;
2244 return (ZIO_PIPELINE_CONTINUE);
2248 gnpp = &gio->io_gang_tree;
2250 gnpp = pio->io_private;
2251 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2254 gn = zio_gang_node_alloc(gnpp);
2256 bzero(gbh, SPA_GANGBLOCKSIZE);
2259 * Create the gang header.
2261 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2262 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2265 * Create and nowait the gang children.
2267 for (int g = 0; resid != 0; resid -= lsize, g++) {
2268 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2270 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2272 zp.zp_checksum = gio->io_prop.zp_checksum;
2273 zp.zp_compress = ZIO_COMPRESS_OFF;
2274 zp.zp_type = DMU_OT_NONE;
2276 zp.zp_copies = gio->io_prop.zp_copies;
2277 zp.zp_dedup = B_FALSE;
2278 zp.zp_dedup_verify = B_FALSE;
2279 zp.zp_nopwrite = B_FALSE;
2281 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2282 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2283 zio_write_gang_member_ready, NULL, NULL, NULL,
2284 &gn->gn_child[g], pio->io_priority,
2285 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2287 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2288 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2289 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2292 * Gang children won't throttle but we should
2293 * account for their work, so reserve an allocation
2294 * slot for them here.
2296 VERIFY(metaslab_class_throttle_reserve(mc,
2297 zp.zp_copies, cio, flags));
2303 * Set pio's pipeline to just wait for zio to finish.
2305 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2309 return (ZIO_PIPELINE_CONTINUE);
2313 * The zio_nop_write stage in the pipeline determines if allocating a
2314 * new bp is necessary. The nopwrite feature can handle writes in
2315 * either syncing or open context (i.e. zil writes) and as a result is
2316 * mutually exclusive with dedup.
2318 * By leveraging a cryptographically secure checksum, such as SHA256, we
2319 * can compare the checksums of the new data and the old to determine if
2320 * allocating a new block is required. Note that our requirements for
2321 * cryptographic strength are fairly weak: there can't be any accidental
2322 * hash collisions, but we don't need to be secure against intentional
2323 * (malicious) collisions. To trigger a nopwrite, you have to be able
2324 * to write the file to begin with, and triggering an incorrect (hash
2325 * collision) nopwrite is no worse than simply writing to the file.
2326 * That said, there are no known attacks against the checksum algorithms
2327 * used for nopwrite, assuming that the salt and the checksums
2328 * themselves remain secret.
2331 zio_nop_write(zio_t *zio)
2333 blkptr_t *bp = zio->io_bp;
2334 blkptr_t *bp_orig = &zio->io_bp_orig;
2335 zio_prop_t *zp = &zio->io_prop;
2337 ASSERT(BP_GET_LEVEL(bp) == 0);
2338 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2339 ASSERT(zp->zp_nopwrite);
2340 ASSERT(!zp->zp_dedup);
2341 ASSERT(zio->io_bp_override == NULL);
2342 ASSERT(IO_IS_ALLOCATING(zio));
2345 * Check to see if the original bp and the new bp have matching
2346 * characteristics (i.e. same checksum, compression algorithms, etc).
2347 * If they don't then just continue with the pipeline which will
2348 * allocate a new bp.
2350 if (BP_IS_HOLE(bp_orig) ||
2351 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2352 ZCHECKSUM_FLAG_NOPWRITE) ||
2353 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2354 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2355 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2356 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2357 return (ZIO_PIPELINE_CONTINUE);
2360 * If the checksums match then reset the pipeline so that we
2361 * avoid allocating a new bp and issuing any I/O.
2363 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2364 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2365 ZCHECKSUM_FLAG_NOPWRITE);
2366 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2367 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2368 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2369 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2370 sizeof (uint64_t)) == 0);
2373 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2374 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2377 return (ZIO_PIPELINE_CONTINUE);
2381 * ==========================================================================
2383 * ==========================================================================
2386 zio_ddt_child_read_done(zio_t *zio)
2388 blkptr_t *bp = zio->io_bp;
2389 ddt_entry_t *dde = zio->io_private;
2391 zio_t *pio = zio_unique_parent(zio);
2393 mutex_enter(&pio->io_lock);
2394 ddp = ddt_phys_select(dde, bp);
2395 if (zio->io_error == 0)
2396 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2397 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2398 dde->dde_repair_data = zio->io_data;
2400 zio_buf_free(zio->io_data, zio->io_size);
2401 mutex_exit(&pio->io_lock);
2405 zio_ddt_read_start(zio_t *zio)
2407 blkptr_t *bp = zio->io_bp;
2409 ASSERT(BP_GET_DEDUP(bp));
2410 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2411 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2413 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2414 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2415 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2416 ddt_phys_t *ddp = dde->dde_phys;
2417 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2420 ASSERT(zio->io_vsd == NULL);
2423 if (ddp_self == NULL)
2424 return (ZIO_PIPELINE_CONTINUE);
2426 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2427 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2429 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2431 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2432 zio_buf_alloc(zio->io_size), zio->io_size,
2433 zio_ddt_child_read_done, dde, zio->io_priority,
2434 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2435 &zio->io_bookmark));
2437 return (ZIO_PIPELINE_CONTINUE);
2440 zio_nowait(zio_read(zio, zio->io_spa, bp,
2441 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2442 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2444 return (ZIO_PIPELINE_CONTINUE);
2448 zio_ddt_read_done(zio_t *zio)
2450 blkptr_t *bp = zio->io_bp;
2452 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2453 return (ZIO_PIPELINE_STOP);
2455 ASSERT(BP_GET_DEDUP(bp));
2456 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2457 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2459 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2460 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2461 ddt_entry_t *dde = zio->io_vsd;
2463 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2464 return (ZIO_PIPELINE_CONTINUE);
2467 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2468 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2469 return (ZIO_PIPELINE_STOP);
2471 if (dde->dde_repair_data != NULL) {
2472 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2473 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2475 ddt_repair_done(ddt, dde);
2479 ASSERT(zio->io_vsd == NULL);
2481 return (ZIO_PIPELINE_CONTINUE);
2485 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2487 spa_t *spa = zio->io_spa;
2490 * Note: we compare the original data, not the transformed data,
2491 * because when zio->io_bp is an override bp, we will not have
2492 * pushed the I/O transforms. That's an important optimization
2493 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2495 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2496 zio_t *lio = dde->dde_lead_zio[p];
2499 return (lio->io_orig_size != zio->io_orig_size ||
2500 bcmp(zio->io_orig_data, lio->io_orig_data,
2501 zio->io_orig_size) != 0);
2505 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2506 ddt_phys_t *ddp = &dde->dde_phys[p];
2508 if (ddp->ddp_phys_birth != 0) {
2509 arc_buf_t *abuf = NULL;
2510 arc_flags_t aflags = ARC_FLAG_WAIT;
2511 blkptr_t blk = *zio->io_bp;
2514 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2518 error = arc_read(NULL, spa, &blk,
2519 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2520 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2521 &aflags, &zio->io_bookmark);
2524 if (arc_buf_size(abuf) != zio->io_orig_size ||
2525 bcmp(abuf->b_data, zio->io_orig_data,
2526 zio->io_orig_size) != 0)
2527 error = SET_ERROR(EEXIST);
2528 arc_buf_destroy(abuf, &abuf);
2532 return (error != 0);
2540 zio_ddt_child_write_ready(zio_t *zio)
2542 int p = zio->io_prop.zp_copies;
2543 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2544 ddt_entry_t *dde = zio->io_private;
2545 ddt_phys_t *ddp = &dde->dde_phys[p];
2553 ASSERT(dde->dde_lead_zio[p] == zio);
2555 ddt_phys_fill(ddp, zio->io_bp);
2557 zio_link_t *zl = NULL;
2558 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2559 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2565 zio_ddt_child_write_done(zio_t *zio)
2567 int p = zio->io_prop.zp_copies;
2568 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2569 ddt_entry_t *dde = zio->io_private;
2570 ddt_phys_t *ddp = &dde->dde_phys[p];
2574 ASSERT(ddp->ddp_refcnt == 0);
2575 ASSERT(dde->dde_lead_zio[p] == zio);
2576 dde->dde_lead_zio[p] = NULL;
2578 if (zio->io_error == 0) {
2579 zio_link_t *zl = NULL;
2580 while (zio_walk_parents(zio, &zl) != NULL)
2581 ddt_phys_addref(ddp);
2583 ddt_phys_clear(ddp);
2590 zio_ddt_ditto_write_done(zio_t *zio)
2592 int p = DDT_PHYS_DITTO;
2593 zio_prop_t *zp = &zio->io_prop;
2594 blkptr_t *bp = zio->io_bp;
2595 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2596 ddt_entry_t *dde = zio->io_private;
2597 ddt_phys_t *ddp = &dde->dde_phys[p];
2598 ddt_key_t *ddk = &dde->dde_key;
2602 ASSERT(ddp->ddp_refcnt == 0);
2603 ASSERT(dde->dde_lead_zio[p] == zio);
2604 dde->dde_lead_zio[p] = NULL;
2606 if (zio->io_error == 0) {
2607 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2608 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2609 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2610 if (ddp->ddp_phys_birth != 0)
2611 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2612 ddt_phys_fill(ddp, bp);
2619 zio_ddt_write(zio_t *zio)
2621 spa_t *spa = zio->io_spa;
2622 blkptr_t *bp = zio->io_bp;
2623 uint64_t txg = zio->io_txg;
2624 zio_prop_t *zp = &zio->io_prop;
2625 int p = zp->zp_copies;
2629 ddt_t *ddt = ddt_select(spa, bp);
2633 ASSERT(BP_GET_DEDUP(bp));
2634 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2635 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2638 dde = ddt_lookup(ddt, bp, B_TRUE);
2639 ddp = &dde->dde_phys[p];
2641 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2643 * If we're using a weak checksum, upgrade to a strong checksum
2644 * and try again. If we're already using a strong checksum,
2645 * we can't resolve it, so just convert to an ordinary write.
2646 * (And automatically e-mail a paper to Nature?)
2648 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2649 ZCHECKSUM_FLAG_DEDUP)) {
2650 zp->zp_checksum = spa_dedup_checksum(spa);
2651 zio_pop_transforms(zio);
2652 zio->io_stage = ZIO_STAGE_OPEN;
2655 zp->zp_dedup = B_FALSE;
2657 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2659 return (ZIO_PIPELINE_CONTINUE);
2662 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2663 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2665 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2666 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2667 zio_prop_t czp = *zp;
2669 czp.zp_copies = ditto_copies;
2672 * If we arrived here with an override bp, we won't have run
2673 * the transform stack, so we won't have the data we need to
2674 * generate a child i/o. So, toss the override bp and restart.
2675 * This is safe, because using the override bp is just an
2676 * optimization; and it's rare, so the cost doesn't matter.
2678 if (zio->io_bp_override) {
2679 zio_pop_transforms(zio);
2680 zio->io_stage = ZIO_STAGE_OPEN;
2681 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2682 zio->io_bp_override = NULL;
2685 return (ZIO_PIPELINE_CONTINUE);
2688 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2689 zio->io_orig_size, &czp, NULL, NULL,
2690 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2691 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2693 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2694 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2697 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2698 if (ddp->ddp_phys_birth != 0)
2699 ddt_bp_fill(ddp, bp, txg);
2700 if (dde->dde_lead_zio[p] != NULL)
2701 zio_add_child(zio, dde->dde_lead_zio[p]);
2703 ddt_phys_addref(ddp);
2704 } else if (zio->io_bp_override) {
2705 ASSERT(bp->blk_birth == txg);
2706 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2707 ddt_phys_fill(ddp, bp);
2708 ddt_phys_addref(ddp);
2710 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2711 zio->io_orig_size, zp,
2712 zio_ddt_child_write_ready, NULL, NULL,
2713 zio_ddt_child_write_done, dde, zio->io_priority,
2714 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2716 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2717 dde->dde_lead_zio[p] = cio;
2727 return (ZIO_PIPELINE_CONTINUE);
2730 ddt_entry_t *freedde; /* for debugging */
2733 zio_ddt_free(zio_t *zio)
2735 spa_t *spa = zio->io_spa;
2736 blkptr_t *bp = zio->io_bp;
2737 ddt_t *ddt = ddt_select(spa, bp);
2741 ASSERT(BP_GET_DEDUP(bp));
2742 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2745 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2746 ddp = ddt_phys_select(dde, bp);
2747 ddt_phys_decref(ddp);
2750 return (ZIO_PIPELINE_CONTINUE);
2754 * ==========================================================================
2755 * Allocate and free blocks
2756 * ==========================================================================
2760 zio_io_to_allocate(spa_t *spa)
2764 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2766 zio = avl_first(&spa->spa_alloc_tree);
2770 ASSERT(IO_IS_ALLOCATING(zio));
2773 * Try to place a reservation for this zio. If we're unable to
2774 * reserve then we throttle.
2776 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2777 zio->io_prop.zp_copies, zio, 0)) {
2781 avl_remove(&spa->spa_alloc_tree, zio);
2782 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2788 zio_dva_throttle(zio_t *zio)
2790 spa_t *spa = zio->io_spa;
2793 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2794 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2795 zio->io_child_type == ZIO_CHILD_GANG ||
2796 zio->io_flags & ZIO_FLAG_NODATA) {
2797 return (ZIO_PIPELINE_CONTINUE);
2800 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2802 ASSERT3U(zio->io_queued_timestamp, >, 0);
2803 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2805 mutex_enter(&spa->spa_alloc_lock);
2807 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2808 avl_add(&spa->spa_alloc_tree, zio);
2810 nio = zio_io_to_allocate(zio->io_spa);
2811 mutex_exit(&spa->spa_alloc_lock);
2814 return (ZIO_PIPELINE_CONTINUE);
2817 ASSERT3U(nio->io_queued_timestamp, <=,
2818 zio->io_queued_timestamp);
2819 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2821 * We are passing control to a new zio so make sure that
2822 * it is processed by a different thread. We do this to
2823 * avoid stack overflows that can occur when parents are
2824 * throttled and children are making progress. We allow
2825 * it to go to the head of the taskq since it's already
2828 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2830 return (ZIO_PIPELINE_STOP);
2834 zio_allocate_dispatch(spa_t *spa)
2838 mutex_enter(&spa->spa_alloc_lock);
2839 zio = zio_io_to_allocate(spa);
2840 mutex_exit(&spa->spa_alloc_lock);
2844 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2845 ASSERT0(zio->io_error);
2846 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2850 zio_dva_allocate(zio_t *zio)
2852 spa_t *spa = zio->io_spa;
2853 metaslab_class_t *mc = spa_normal_class(spa);
2854 blkptr_t *bp = zio->io_bp;
2858 if (zio->io_gang_leader == NULL) {
2859 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2860 zio->io_gang_leader = zio;
2863 ASSERT(BP_IS_HOLE(bp));
2864 ASSERT0(BP_GET_NDVAS(bp));
2865 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2866 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2867 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2869 if (zio->io_flags & ZIO_FLAG_NODATA) {
2870 flags |= METASLAB_DONT_THROTTLE;
2872 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2873 flags |= METASLAB_GANG_CHILD;
2875 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2876 flags |= METASLAB_ASYNC_ALLOC;
2879 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2880 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio);
2883 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2884 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2886 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2887 return (zio_write_gang_block(zio));
2888 zio->io_error = error;
2891 return (ZIO_PIPELINE_CONTINUE);
2895 zio_dva_free(zio_t *zio)
2897 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2899 return (ZIO_PIPELINE_CONTINUE);
2903 zio_dva_claim(zio_t *zio)
2907 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2909 zio->io_error = error;
2911 return (ZIO_PIPELINE_CONTINUE);
2915 * Undo an allocation. This is used by zio_done() when an I/O fails
2916 * and we want to give back the block we just allocated.
2917 * This handles both normal blocks and gang blocks.
2920 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2922 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2923 ASSERT(zio->io_bp_override == NULL);
2925 if (!BP_IS_HOLE(bp))
2926 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2929 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2930 zio_dva_unallocate(zio, gn->gn_child[g],
2931 &gn->gn_gbh->zg_blkptr[g]);
2937 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2940 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2941 uint64_t size, boolean_t *slog)
2945 ASSERT(txg > spa_syncing_txg(spa));
2947 error = metaslab_alloc(spa, spa_log_class(spa), size,
2948 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2952 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2953 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL);
2959 BP_SET_LSIZE(new_bp, size);
2960 BP_SET_PSIZE(new_bp, size);
2961 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2962 BP_SET_CHECKSUM(new_bp,
2963 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2964 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2965 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2966 BP_SET_LEVEL(new_bp, 0);
2967 BP_SET_DEDUP(new_bp, 0);
2968 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2975 * Free an intent log block.
2978 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2980 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2981 ASSERT(!BP_IS_GANG(bp));
2983 zio_free(spa, txg, bp);
2987 * ==========================================================================
2988 * Read, write and delete to physical devices
2989 * ==========================================================================
2994 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2995 * stops after this stage and will resume upon I/O completion.
2996 * However, there are instances where the vdev layer may need to
2997 * continue the pipeline when an I/O was not issued. Since the I/O
2998 * that was sent to the vdev layer might be different than the one
2999 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3000 * force the underlying vdev layers to call either zio_execute() or
3001 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3004 zio_vdev_io_start(zio_t *zio)
3006 vdev_t *vd = zio->io_vd;
3008 spa_t *spa = zio->io_spa;
3011 ASSERT(zio->io_error == 0);
3012 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3015 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3016 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3019 * The mirror_ops handle multiple DVAs in a single BP.
3021 vdev_mirror_ops.vdev_op_io_start(zio);
3022 return (ZIO_PIPELINE_STOP);
3025 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3026 zio->io_priority == ZIO_PRIORITY_NOW) {
3027 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3028 return (ZIO_PIPELINE_CONTINUE);
3031 ASSERT3P(zio->io_logical, !=, zio);
3034 * We keep track of time-sensitive I/Os so that the scan thread
3035 * can quickly react to certain workloads. In particular, we care
3036 * about non-scrubbing, top-level reads and writes with the following
3038 * - synchronous writes of user data to non-slog devices
3039 * - any reads of user data
3040 * When these conditions are met, adjust the timestamp of spa_last_io
3041 * which allows the scan thread to adjust its workload accordingly.
3043 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3044 vd == vd->vdev_top && !vd->vdev_islog &&
3045 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3046 zio->io_txg != spa_syncing_txg(spa)) {
3047 uint64_t old = spa->spa_last_io;
3048 uint64_t new = ddi_get_lbolt64();
3050 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3053 align = 1ULL << vd->vdev_top->vdev_ashift;
3055 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3056 P2PHASE(zio->io_size, align) != 0) {
3057 /* Transform logical writes to be a full physical block size. */
3058 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3060 if (zio->io_type == ZIO_TYPE_READ ||
3061 zio->io_type == ZIO_TYPE_WRITE)
3062 abuf = zio_buf_alloc(asize);
3063 ASSERT(vd == vd->vdev_top);
3064 if (zio->io_type == ZIO_TYPE_WRITE) {
3065 bcopy(zio->io_data, abuf, zio->io_size);
3066 bzero(abuf + zio->io_size, asize - zio->io_size);
3068 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3073 * If this is not a physical io, make sure that it is properly aligned
3074 * before proceeding.
3076 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3077 ASSERT0(P2PHASE(zio->io_offset, align));
3078 ASSERT0(P2PHASE(zio->io_size, align));
3081 * For the physical io we allow alignment
3082 * to a logical block size.
3084 uint64_t log_align =
3085 1ULL << vd->vdev_top->vdev_logical_ashift;
3086 ASSERT0(P2PHASE(zio->io_offset, log_align));
3087 ASSERT0(P2PHASE(zio->io_size, log_align));
3090 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3093 * If this is a repair I/O, and there's no self-healing involved --
3094 * that is, we're just resilvering what we expect to resilver --
3095 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3096 * This prevents spurious resilvering with nested replication.
3097 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3098 * A is out of date, we'll read from C+D, then use the data to
3099 * resilver A+B -- but we don't actually want to resilver B, just A.
3100 * The top-level mirror has no way to know this, so instead we just
3101 * discard unnecessary repairs as we work our way down the vdev tree.
3102 * The same logic applies to any form of nested replication:
3103 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3105 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3106 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3107 zio->io_txg != 0 && /* not a delegated i/o */
3108 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3109 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3110 zio_vdev_io_bypass(zio);
3111 return (ZIO_PIPELINE_CONTINUE);
3114 if (vd->vdev_ops->vdev_op_leaf) {
3115 switch (zio->io_type) {
3117 if (vdev_cache_read(zio))
3118 return (ZIO_PIPELINE_CONTINUE);
3120 case ZIO_TYPE_WRITE:
3122 if ((zio = vdev_queue_io(zio)) == NULL)
3123 return (ZIO_PIPELINE_STOP);
3125 if (!vdev_accessible(vd, zio)) {
3126 zio->io_error = SET_ERROR(ENXIO);
3128 return (ZIO_PIPELINE_STOP);
3133 * Note that we ignore repair writes for TRIM because they can
3134 * conflict with normal writes. This isn't an issue because, by
3135 * definition, we only repair blocks that aren't freed.
3137 if (zio->io_type == ZIO_TYPE_WRITE &&
3138 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3139 !trim_map_write_start(zio))
3140 return (ZIO_PIPELINE_STOP);
3143 vd->vdev_ops->vdev_op_io_start(zio);
3144 return (ZIO_PIPELINE_STOP);
3148 zio_vdev_io_done(zio_t *zio)
3150 vdev_t *vd = zio->io_vd;
3151 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3152 boolean_t unexpected_error = B_FALSE;
3154 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3155 return (ZIO_PIPELINE_STOP);
3157 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3158 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3160 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3161 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3162 zio->io_type == ZIO_TYPE_FREE)) {
3164 if (zio->io_type == ZIO_TYPE_WRITE &&
3165 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3166 trim_map_write_done(zio);
3168 vdev_queue_io_done(zio);
3170 if (zio->io_type == ZIO_TYPE_WRITE)
3171 vdev_cache_write(zio);
3173 if (zio_injection_enabled && zio->io_error == 0)
3174 zio->io_error = zio_handle_device_injection(vd,
3177 if (zio_injection_enabled && zio->io_error == 0)
3178 zio->io_error = zio_handle_label_injection(zio, EIO);
3180 if (zio->io_error) {
3181 if (zio->io_error == ENOTSUP &&
3182 zio->io_type == ZIO_TYPE_FREE) {
3183 /* Not all devices support TRIM. */
3184 } else if (!vdev_accessible(vd, zio)) {
3185 zio->io_error = SET_ERROR(ENXIO);
3187 unexpected_error = B_TRUE;
3192 ops->vdev_op_io_done(zio);
3194 if (unexpected_error)
3195 VERIFY(vdev_probe(vd, zio) == NULL);
3197 return (ZIO_PIPELINE_CONTINUE);
3201 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3202 * disk, and use that to finish the checksum ereport later.
3205 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3206 const void *good_buf)
3208 /* no processing needed */
3209 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3214 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3216 void *buf = zio_buf_alloc(zio->io_size);
3218 bcopy(zio->io_data, buf, zio->io_size);
3220 zcr->zcr_cbinfo = zio->io_size;
3221 zcr->zcr_cbdata = buf;
3222 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3223 zcr->zcr_free = zio_buf_free;
3227 zio_vdev_io_assess(zio_t *zio)
3229 vdev_t *vd = zio->io_vd;
3231 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3232 return (ZIO_PIPELINE_STOP);
3234 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3235 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3237 if (zio->io_vsd != NULL) {
3238 zio->io_vsd_ops->vsd_free(zio);
3242 if (zio_injection_enabled && zio->io_error == 0)
3243 zio->io_error = zio_handle_fault_injection(zio, EIO);
3245 if (zio->io_type == ZIO_TYPE_FREE &&
3246 zio->io_priority != ZIO_PRIORITY_NOW) {
3247 switch (zio->io_error) {
3249 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3250 ZIO_TRIM_STAT_BUMP(success);
3253 ZIO_TRIM_STAT_BUMP(unsupported);
3256 ZIO_TRIM_STAT_BUMP(failed);
3262 * If the I/O failed, determine whether we should attempt to retry it.
3264 * On retry, we cut in line in the issue queue, since we don't want
3265 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3267 if (zio->io_error && vd == NULL &&
3268 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3269 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3270 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3272 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3273 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3274 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3275 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3276 zio_requeue_io_start_cut_in_line);
3277 return (ZIO_PIPELINE_STOP);
3281 * If we got an error on a leaf device, convert it to ENXIO
3282 * if the device is not accessible at all.
3284 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3285 !vdev_accessible(vd, zio))
3286 zio->io_error = SET_ERROR(ENXIO);
3289 * If we can't write to an interior vdev (mirror or RAID-Z),
3290 * set vdev_cant_write so that we stop trying to allocate from it.
3292 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3293 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3294 vd->vdev_cant_write = B_TRUE;
3298 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3300 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3301 zio->io_physdone != NULL) {
3302 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3303 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3304 zio->io_physdone(zio->io_logical);
3307 return (ZIO_PIPELINE_CONTINUE);
3311 zio_vdev_io_reissue(zio_t *zio)
3313 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3314 ASSERT(zio->io_error == 0);
3316 zio->io_stage >>= 1;
3320 zio_vdev_io_redone(zio_t *zio)
3322 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3324 zio->io_stage >>= 1;
3328 zio_vdev_io_bypass(zio_t *zio)
3330 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3331 ASSERT(zio->io_error == 0);
3333 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3334 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3338 * ==========================================================================
3339 * Generate and verify checksums
3340 * ==========================================================================
3343 zio_checksum_generate(zio_t *zio)
3345 blkptr_t *bp = zio->io_bp;
3346 enum zio_checksum checksum;
3350 * This is zio_write_phys().
3351 * We're either generating a label checksum, or none at all.
3353 checksum = zio->io_prop.zp_checksum;
3355 if (checksum == ZIO_CHECKSUM_OFF)
3356 return (ZIO_PIPELINE_CONTINUE);
3358 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3360 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3361 ASSERT(!IO_IS_ALLOCATING(zio));
3362 checksum = ZIO_CHECKSUM_GANG_HEADER;
3364 checksum = BP_GET_CHECKSUM(bp);
3368 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3370 return (ZIO_PIPELINE_CONTINUE);
3374 zio_checksum_verify(zio_t *zio)
3376 zio_bad_cksum_t info;
3377 blkptr_t *bp = zio->io_bp;
3380 ASSERT(zio->io_vd != NULL);
3384 * This is zio_read_phys().
3385 * We're either verifying a label checksum, or nothing at all.
3387 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3388 return (ZIO_PIPELINE_CONTINUE);
3390 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3393 if ((error = zio_checksum_error(zio, &info)) != 0) {
3394 zio->io_error = error;
3395 if (error == ECKSUM &&
3396 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3397 zfs_ereport_start_checksum(zio->io_spa,
3398 zio->io_vd, zio, zio->io_offset,
3399 zio->io_size, NULL, &info);
3403 return (ZIO_PIPELINE_CONTINUE);
3407 * Called by RAID-Z to ensure we don't compute the checksum twice.
3410 zio_checksum_verified(zio_t *zio)
3412 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3416 * ==========================================================================
3417 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3418 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3419 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3420 * indicate errors that are specific to one I/O, and most likely permanent.
3421 * Any other error is presumed to be worse because we weren't expecting it.
3422 * ==========================================================================
3425 zio_worst_error(int e1, int e2)
3427 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3430 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3431 if (e1 == zio_error_rank[r1])
3434 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3435 if (e2 == zio_error_rank[r2])
3438 return (r1 > r2 ? e1 : e2);
3442 * ==========================================================================
3444 * ==========================================================================
3447 zio_ready(zio_t *zio)
3449 blkptr_t *bp = zio->io_bp;
3450 zio_t *pio, *pio_next;
3451 zio_link_t *zl = NULL;
3453 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3454 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3455 return (ZIO_PIPELINE_STOP);
3457 if (zio->io_ready) {
3458 ASSERT(IO_IS_ALLOCATING(zio));
3459 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3460 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3461 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3466 if (bp != NULL && bp != &zio->io_bp_copy)
3467 zio->io_bp_copy = *bp;
3469 if (zio->io_error != 0) {
3470 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3472 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3473 ASSERT(IO_IS_ALLOCATING(zio));
3474 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3476 * We were unable to allocate anything, unreserve and
3477 * issue the next I/O to allocate.
3479 metaslab_class_throttle_unreserve(
3480 spa_normal_class(zio->io_spa),
3481 zio->io_prop.zp_copies, zio);
3482 zio_allocate_dispatch(zio->io_spa);
3486 mutex_enter(&zio->io_lock);
3487 zio->io_state[ZIO_WAIT_READY] = 1;
3488 pio = zio_walk_parents(zio, &zl);
3489 mutex_exit(&zio->io_lock);
3492 * As we notify zio's parents, new parents could be added.
3493 * New parents go to the head of zio's io_parent_list, however,
3494 * so we will (correctly) not notify them. The remainder of zio's
3495 * io_parent_list, from 'pio_next' onward, cannot change because
3496 * all parents must wait for us to be done before they can be done.
3498 for (; pio != NULL; pio = pio_next) {
3499 pio_next = zio_walk_parents(zio, &zl);
3500 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3503 if (zio->io_flags & ZIO_FLAG_NODATA) {
3504 if (BP_IS_GANG(bp)) {
3505 zio->io_flags &= ~ZIO_FLAG_NODATA;
3507 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3508 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3512 if (zio_injection_enabled &&
3513 zio->io_spa->spa_syncing_txg == zio->io_txg)
3514 zio_handle_ignored_writes(zio);
3516 return (ZIO_PIPELINE_CONTINUE);
3520 * Update the allocation throttle accounting.
3523 zio_dva_throttle_done(zio_t *zio)
3525 zio_t *lio = zio->io_logical;
3526 zio_t *pio = zio_unique_parent(zio);
3527 vdev_t *vd = zio->io_vd;
3528 int flags = METASLAB_ASYNC_ALLOC;
3530 ASSERT3P(zio->io_bp, !=, NULL);
3531 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3532 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3533 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3535 ASSERT3P(vd, ==, vd->vdev_top);
3536 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3537 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3538 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3539 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3542 * Parents of gang children can have two flavors -- ones that
3543 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3544 * and ones that allocated the constituent blocks. The allocation
3545 * throttle needs to know the allocating parent zio so we must find
3548 if (pio->io_child_type == ZIO_CHILD_GANG) {
3550 * If our parent is a rewrite gang child then our grandparent
3551 * would have been the one that performed the allocation.
3553 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3554 pio = zio_unique_parent(pio);
3555 flags |= METASLAB_GANG_CHILD;
3558 ASSERT(IO_IS_ALLOCATING(pio));
3559 ASSERT3P(zio, !=, zio->io_logical);
3560 ASSERT(zio->io_logical != NULL);
3561 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3562 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3564 mutex_enter(&pio->io_lock);
3565 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3566 mutex_exit(&pio->io_lock);
3568 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3572 * Call into the pipeline to see if there is more work that
3573 * needs to be done. If there is work to be done it will be
3574 * dispatched to another taskq thread.
3576 zio_allocate_dispatch(zio->io_spa);
3580 zio_done(zio_t *zio)
3582 spa_t *spa = zio->io_spa;
3583 zio_t *lio = zio->io_logical;
3584 blkptr_t *bp = zio->io_bp;
3585 vdev_t *vd = zio->io_vd;
3586 uint64_t psize = zio->io_size;
3587 zio_t *pio, *pio_next;
3588 metaslab_class_t *mc = spa_normal_class(spa);
3589 zio_link_t *zl = NULL;
3592 * If our children haven't all completed,
3593 * wait for them and then repeat this pipeline stage.
3595 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3596 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3597 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3598 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3599 return (ZIO_PIPELINE_STOP);
3602 * If the allocation throttle is enabled, then update the accounting.
3603 * We only track child I/Os that are part of an allocating async
3604 * write. We must do this since the allocation is performed
3605 * by the logical I/O but the actual write is done by child I/Os.
3607 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3608 zio->io_child_type == ZIO_CHILD_VDEV) {
3609 ASSERT(mc->mc_alloc_throttle_enabled);
3610 zio_dva_throttle_done(zio);
3614 * If the allocation throttle is enabled, verify that
3615 * we have decremented the refcounts for every I/O that was throttled.
3617 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3618 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3619 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3621 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3622 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3625 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3626 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3627 ASSERT(zio->io_children[c][w] == 0);
3629 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3630 ASSERT(bp->blk_pad[0] == 0);
3631 ASSERT(bp->blk_pad[1] == 0);
3632 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3633 (bp == zio_unique_parent(zio)->io_bp));
3634 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3635 zio->io_bp_override == NULL &&
3636 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3637 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3638 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3639 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3640 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3642 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3643 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3647 * If there were child vdev/gang/ddt errors, they apply to us now.
3649 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3650 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3651 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3654 * If the I/O on the transformed data was successful, generate any
3655 * checksum reports now while we still have the transformed data.
3657 if (zio->io_error == 0) {
3658 while (zio->io_cksum_report != NULL) {
3659 zio_cksum_report_t *zcr = zio->io_cksum_report;
3660 uint64_t align = zcr->zcr_align;
3661 uint64_t asize = P2ROUNDUP(psize, align);
3662 char *abuf = zio->io_data;
3664 if (asize != psize) {
3665 abuf = zio_buf_alloc(asize);
3666 bcopy(zio->io_data, abuf, psize);
3667 bzero(abuf + psize, asize - psize);
3670 zio->io_cksum_report = zcr->zcr_next;
3671 zcr->zcr_next = NULL;
3672 zcr->zcr_finish(zcr, abuf);
3673 zfs_ereport_free_checksum(zcr);
3676 zio_buf_free(abuf, asize);
3680 zio_pop_transforms(zio); /* note: may set zio->io_error */
3682 vdev_stat_update(zio, psize);
3684 if (zio->io_error) {
3686 * If this I/O is attached to a particular vdev,
3687 * generate an error message describing the I/O failure
3688 * at the block level. We ignore these errors if the
3689 * device is currently unavailable.
3691 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3692 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3694 if ((zio->io_error == EIO || !(zio->io_flags &
3695 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3698 * For logical I/O requests, tell the SPA to log the
3699 * error and generate a logical data ereport.
3701 spa_log_error(spa, zio);
3702 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3707 if (zio->io_error && zio == lio) {
3709 * Determine whether zio should be reexecuted. This will
3710 * propagate all the way to the root via zio_notify_parent().
3712 ASSERT(vd == NULL && bp != NULL);
3713 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3715 if (IO_IS_ALLOCATING(zio) &&
3716 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3717 if (zio->io_error != ENOSPC)
3718 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3720 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3723 if ((zio->io_type == ZIO_TYPE_READ ||
3724 zio->io_type == ZIO_TYPE_FREE) &&
3725 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3726 zio->io_error == ENXIO &&
3727 spa_load_state(spa) == SPA_LOAD_NONE &&
3728 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3729 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3731 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3732 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3735 * Here is a possibly good place to attempt to do
3736 * either combinatorial reconstruction or error correction
3737 * based on checksums. It also might be a good place
3738 * to send out preliminary ereports before we suspend
3744 * If there were logical child errors, they apply to us now.
3745 * We defer this until now to avoid conflating logical child
3746 * errors with errors that happened to the zio itself when
3747 * updating vdev stats and reporting FMA events above.
3749 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3751 if ((zio->io_error || zio->io_reexecute) &&
3752 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3753 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3754 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3756 zio_gang_tree_free(&zio->io_gang_tree);
3759 * Godfather I/Os should never suspend.
3761 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3762 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3763 zio->io_reexecute = 0;
3765 if (zio->io_reexecute) {
3767 * This is a logical I/O that wants to reexecute.
3769 * Reexecute is top-down. When an i/o fails, if it's not
3770 * the root, it simply notifies its parent and sticks around.
3771 * The parent, seeing that it still has children in zio_done(),
3772 * does the same. This percolates all the way up to the root.
3773 * The root i/o will reexecute or suspend the entire tree.
3775 * This approach ensures that zio_reexecute() honors
3776 * all the original i/o dependency relationships, e.g.
3777 * parents not executing until children are ready.
3779 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3781 zio->io_gang_leader = NULL;
3783 mutex_enter(&zio->io_lock);
3784 zio->io_state[ZIO_WAIT_DONE] = 1;
3785 mutex_exit(&zio->io_lock);
3788 * "The Godfather" I/O monitors its children but is
3789 * not a true parent to them. It will track them through
3790 * the pipeline but severs its ties whenever they get into
3791 * trouble (e.g. suspended). This allows "The Godfather"
3792 * I/O to return status without blocking.
3795 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3797 zio_link_t *remove_zl = zl;
3798 pio_next = zio_walk_parents(zio, &zl);
3800 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3801 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3802 zio_remove_child(pio, zio, remove_zl);
3803 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3807 if ((pio = zio_unique_parent(zio)) != NULL) {
3809 * We're not a root i/o, so there's nothing to do
3810 * but notify our parent. Don't propagate errors
3811 * upward since we haven't permanently failed yet.
3813 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3814 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3815 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3816 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3818 * We'd fail again if we reexecuted now, so suspend
3819 * until conditions improve (e.g. device comes online).
3821 zio_suspend(spa, zio);
3824 * Reexecution is potentially a huge amount of work.
3825 * Hand it off to the otherwise-unused claim taskq.
3827 #if defined(illumos) || !defined(_KERNEL)
3828 ASSERT(zio->io_tqent.tqent_next == NULL);
3830 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3832 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3833 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3836 return (ZIO_PIPELINE_STOP);
3839 ASSERT(zio->io_child_count == 0);
3840 ASSERT(zio->io_reexecute == 0);
3841 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3844 * Report any checksum errors, since the I/O is complete.
3846 while (zio->io_cksum_report != NULL) {
3847 zio_cksum_report_t *zcr = zio->io_cksum_report;
3848 zio->io_cksum_report = zcr->zcr_next;
3849 zcr->zcr_next = NULL;
3850 zcr->zcr_finish(zcr, NULL);
3851 zfs_ereport_free_checksum(zcr);
3855 * It is the responsibility of the done callback to ensure that this
3856 * particular zio is no longer discoverable for adoption, and as
3857 * such, cannot acquire any new parents.
3862 mutex_enter(&zio->io_lock);
3863 zio->io_state[ZIO_WAIT_DONE] = 1;
3864 mutex_exit(&zio->io_lock);
3867 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3868 zio_link_t *remove_zl = zl;
3869 pio_next = zio_walk_parents(zio, &zl);
3870 zio_remove_child(pio, zio, remove_zl);
3871 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3874 if (zio->io_waiter != NULL) {
3875 mutex_enter(&zio->io_lock);
3876 zio->io_executor = NULL;
3877 cv_broadcast(&zio->io_cv);
3878 mutex_exit(&zio->io_lock);
3883 return (ZIO_PIPELINE_STOP);
3887 * ==========================================================================
3888 * I/O pipeline definition
3889 * ==========================================================================
3891 static zio_pipe_stage_t *zio_pipeline[] = {
3898 zio_checksum_generate,
3914 zio_checksum_verify,
3922 * Compare two zbookmark_phys_t's to see which we would reach first in a
3923 * pre-order traversal of the object tree.
3925 * This is simple in every case aside from the meta-dnode object. For all other
3926 * objects, we traverse them in order (object 1 before object 2, and so on).
3927 * However, all of these objects are traversed while traversing object 0, since
3928 * the data it points to is the list of objects. Thus, we need to convert to a
3929 * canonical representation so we can compare meta-dnode bookmarks to
3930 * non-meta-dnode bookmarks.
3932 * We do this by calculating "equivalents" for each field of the zbookmark.
3933 * zbookmarks outside of the meta-dnode use their own object and level, and
3934 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3935 * blocks this bookmark refers to) by multiplying their blkid by their span
3936 * (the number of L0 blocks contained within one block at their level).
3937 * zbookmarks inside the meta-dnode calculate their object equivalent
3938 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3939 * level + 1<<31 (any value larger than a level could ever be) for their level.
3940 * This causes them to always compare before a bookmark in their object
3941 * equivalent, compare appropriately to bookmarks in other objects, and to
3942 * compare appropriately to other bookmarks in the meta-dnode.
3945 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3946 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3949 * These variables represent the "equivalent" values for the zbookmark,
3950 * after converting zbookmarks inside the meta dnode to their
3951 * normal-object equivalents.
3953 uint64_t zb1obj, zb2obj;
3954 uint64_t zb1L0, zb2L0;
3955 uint64_t zb1level, zb2level;
3957 if (zb1->zb_object == zb2->zb_object &&
3958 zb1->zb_level == zb2->zb_level &&
3959 zb1->zb_blkid == zb2->zb_blkid)
3963 * BP_SPANB calculates the span in blocks.
3965 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3966 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3968 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3969 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3971 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3973 zb1obj = zb1->zb_object;
3974 zb1level = zb1->zb_level;
3977 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3978 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3980 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3982 zb2obj = zb2->zb_object;
3983 zb2level = zb2->zb_level;
3986 /* Now that we have a canonical representation, do the comparison. */
3987 if (zb1obj != zb2obj)
3988 return (zb1obj < zb2obj ? -1 : 1);
3989 else if (zb1L0 != zb2L0)
3990 return (zb1L0 < zb2L0 ? -1 : 1);
3991 else if (zb1level != zb2level)
3992 return (zb1level > zb2level ? -1 : 1);
3994 * This can (theoretically) happen if the bookmarks have the same object
3995 * and level, but different blkids, if the block sizes are not the same.
3996 * There is presently no way to change the indirect block sizes
4002 * This function checks the following: given that last_block is the place that
4003 * our traversal stopped last time, does that guarantee that we've visited
4004 * every node under subtree_root? Therefore, we can't just use the raw output
4005 * of zbookmark_compare. We have to pass in a modified version of
4006 * subtree_root; by incrementing the block id, and then checking whether
4007 * last_block is before or equal to that, we can tell whether or not having
4008 * visited last_block implies that all of subtree_root's children have been
4012 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4013 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4015 zbookmark_phys_t mod_zb = *subtree_root;
4017 ASSERT(last_block->zb_level == 0);
4019 /* The objset_phys_t isn't before anything. */
4024 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4025 * data block size in sectors, because that variable is only used if
4026 * the bookmark refers to a block in the meta-dnode. Since we don't
4027 * know without examining it what object it refers to, and there's no
4028 * harm in passing in this value in other cases, we always pass it in.
4030 * We pass in 0 for the indirect block size shift because zb2 must be
4031 * level 0. The indirect block size is only used to calculate the span
4032 * of the bookmark, but since the bookmark must be level 0, the span is
4033 * always 1, so the math works out.
4035 * If you make changes to how the zbookmark_compare code works, be sure
4036 * to make sure that this code still works afterwards.
4038 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4039 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,