1 /*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa.h>
32 #include <sys/txg.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>
39 #include <sys/arc.h>
40 #include <sys/ddt.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
43 #include <sys/metaslab_impl.h>
44 #include <sys/abd.h>
45
46 /*
47 * ==========================================================================
48 * I/O type descriptions
49 * ==========================================================================
50 */
51 const char *zio_type_name[ZIO_TYPES] = {
52 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
53 "zio_ioctl"
54 };
55
56 boolean_t zio_dva_throttle_enabled = B_TRUE;
57
58 /*
59 * ==========================================================================
60 * I/O kmem caches
61 * ==========================================================================
62 */
63 kmem_cache_t *zio_cache;
64 kmem_cache_t *zio_link_cache;
65 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
66 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
67
68 #ifdef _KERNEL
69 extern vmem_t *zio_alloc_arena;
70 #endif
71
72 #define ZIO_PIPELINE_CONTINUE 0x100
73 #define ZIO_PIPELINE_STOP 0x101
74
75 #define BP_SPANB(indblkshift, level) \
76 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
77 #define COMPARE_META_LEVEL 0x80000000ul
78 /*
79 * The following actions directly effect the spa's sync-to-convergence logic.
80 * The values below define the sync pass when we start performing the action.
81 * Care should be taken when changing these values as they directly impact
82 * spa_sync() performance. Tuning these values may introduce subtle performance
83 * pathologies and should only be done in the context of performance analysis.
84 * These tunables will eventually be removed and replaced with #defines once
85 * enough analysis has been done to determine optimal values.
86 *
87 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
88 * regular blocks are not deferred.
89 */
90 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
91 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
92 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
93
94 /*
95 * An allocating zio is one that either currently has the DVA allocate
96 * stage set or will have it later in its lifetime.
97 */
98 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
99
100 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
101
102 #ifdef ZFS_DEBUG
103 int zio_buf_debug_limit = 16384;
104 #else
105 int zio_buf_debug_limit = 0;
106 #endif
107
108 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
109
110 void
111 zio_init(void)
112 {
113 size_t c;
114 vmem_t *data_alloc_arena = NULL;
115
116 #ifdef _KERNEL
117 data_alloc_arena = zio_alloc_arena;
118 #endif
119 zio_cache = kmem_cache_create("zio_cache",
120 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
121 zio_link_cache = kmem_cache_create("zio_link_cache",
122 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
123
124 /*
125 * For small buffers, we want a cache for each multiple of
126 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
127 * for each quarter-power of 2.
163 * stored with the buffers.
164 */
165 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
166 zio_data_buf_cache[c] = kmem_cache_create(name, size,
167 align, NULL, NULL, NULL, NULL, data_alloc_arena,
168 cflags | KMC_NOTOUCH);
169 }
170 }
171
172 while (--c != 0) {
173 ASSERT(zio_buf_cache[c] != NULL);
174 if (zio_buf_cache[c - 1] == NULL)
175 zio_buf_cache[c - 1] = zio_buf_cache[c];
176
177 ASSERT(zio_data_buf_cache[c] != NULL);
178 if (zio_data_buf_cache[c - 1] == NULL)
179 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
180 }
181
182 zio_inject_init();
183 }
184
185 void
186 zio_fini(void)
187 {
188 size_t c;
189 kmem_cache_t *last_cache = NULL;
190 kmem_cache_t *last_data_cache = NULL;
191
192 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
193 if (zio_buf_cache[c] != last_cache) {
194 last_cache = zio_buf_cache[c];
195 kmem_cache_destroy(zio_buf_cache[c]);
196 }
197 zio_buf_cache[c] = NULL;
198
199 if (zio_data_buf_cache[c] != last_data_cache) {
200 last_data_cache = zio_data_buf_cache[c];
201 kmem_cache_destroy(zio_data_buf_cache[c]);
202 }
425 {
426 ASSERT(zl->zl_parent == pio);
427 ASSERT(zl->zl_child == cio);
428
429 mutex_enter(&cio->io_lock);
430 mutex_enter(&pio->io_lock);
431
432 list_remove(&pio->io_child_list, zl);
433 list_remove(&cio->io_parent_list, zl);
434
435 pio->io_child_count--;
436 cio->io_parent_count--;
437
438 mutex_exit(&pio->io_lock);
439 mutex_exit(&cio->io_lock);
440
441 kmem_cache_free(zio_link_cache, zl);
442 }
443
444 static boolean_t
445 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
446 {
447 boolean_t waiting = B_FALSE;
448
449 mutex_enter(&zio->io_lock);
450 ASSERT(zio->io_stall == NULL);
451 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
452 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
453 continue;
454
455 uint64_t *countp = &zio->io_children[c][wait];
456 if (*countp != 0) {
457 zio->io_stage >>= 1;
458 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
459 zio->io_stall = countp;
460 waiting = B_TRUE;
461 break;
462 }
463 }
464 mutex_exit(&zio->io_lock);
465 return (waiting);
466 }
467
468 static void
469 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
470 {
471 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
472 int *errorp = &pio->io_child_error[zio->io_child_type];
473
474 mutex_enter(&pio->io_lock);
475 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
476 *errorp = zio_worst_error(*errorp, zio->io_error);
477 pio->io_reexecute |= zio->io_reexecute;
478 ASSERT3U(*countp, >, 0);
479
480 (*countp)--;
481
482 if (*countp == 0 && pio->io_stall == countp) {
483 zio_taskq_type_t type =
484 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
602 zio->io_private = private;
603 zio->io_type = type;
604 zio->io_priority = priority;
605 zio->io_vd = vd;
606 zio->io_offset = offset;
607 zio->io_orig_abd = zio->io_abd = data;
608 zio->io_orig_size = zio->io_size = psize;
609 zio->io_lsize = lsize;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
613 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
614
615 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
616 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
617
618 if (zb != NULL)
619 zio->io_bookmark = *zb;
620
621 if (pio != NULL) {
622 if (zio->io_logical == NULL)
623 zio->io_logical = pio->io_logical;
624 if (zio->io_child_type == ZIO_CHILD_GANG)
625 zio->io_gang_leader = pio->io_gang_leader;
626 zio_add_child(pio, zio);
627 }
628
629 return (zio);
630 }
631
632 static void
633 zio_destroy(zio_t *zio)
634 {
635 metaslab_trace_fini(&zio->io_alloc_list);
636 list_destroy(&zio->io_parent_list);
637 list_destroy(&zio->io_child_list);
638 mutex_destroy(&zio->io_lock);
639 cv_destroy(&zio->io_cv);
640 kmem_cache_free(zio_cache, zio);
641 }
642
643 zio_t *
644 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
645 void *private, enum zio_flag flags)
646 {
647 zio_t *zio;
648
649 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
650 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
651 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
652
653 return (zio);
654 }
655
656 zio_t *
657 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
658 {
659 return (zio_null(NULL, spa, NULL, done, private, flags));
660 }
661
662 void
663 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
664 {
665 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
666 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
667 bp, (longlong_t)BP_GET_TYPE(bp));
668 }
669 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
670 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
671 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
672 bp, (longlong_t)BP_GET_CHECKSUM(bp));
673 }
674 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
675 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
676 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
677 bp, (longlong_t)BP_GET_COMPRESS(bp));
678 }
679 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
680 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
681 bp, (longlong_t)BP_GET_LSIZE(bp));
682 }
683 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
684 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
685 bp, (longlong_t)BP_GET_PSIZE(bp));
686 }
687
688 if (BP_IS_EMBEDDED(bp)) {
689 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
690 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
691 bp, (longlong_t)BPE_GET_ETYPE(bp));
692 }
693 }
694
695 /*
696 * Do not verify individual DVAs if the config is not trusted. This
697 * will be done once the zio is executed in vdev_mirror_map_alloc.
698 */
699 if (!spa->spa_trust_config)
700 return;
701
702 /*
703 * Pool-specific checks.
704 *
705 * Note: it would be nice to verify that the blk_birth and
706 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
707 * allows the birth time of log blocks (and dmu_sync()-ed blocks
708 * that are in the log) to be arbitrarily large.
709 */
710 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
711 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
712 if (vdevid >= spa->spa_root_vdev->vdev_children) {
713 zfs_panic_recover("blkptr at %p DVA %u has invalid "
714 "VDEV %llu",
715 bp, i, (longlong_t)vdevid);
716 continue;
717 }
718 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
719 if (vd == NULL) {
720 zfs_panic_recover("blkptr at %p DVA %u has invalid "
721 "VDEV %llu",
722 bp, i, (longlong_t)vdevid);
723 continue;
724 }
725 if (vd->vdev_ops == &vdev_hole_ops) {
726 zfs_panic_recover("blkptr at %p DVA %u has hole "
727 "VDEV %llu",
728 bp, i, (longlong_t)vdevid);
729 continue;
730 }
731 if (vd->vdev_ops == &vdev_missing_ops) {
732 /*
733 * "missing" vdevs are valid during import, but we
734 * don't have their detailed info (e.g. asize), so
735 * we can't perform any more checks on them.
736 */
737 continue;
738 }
739 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
740 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
741 if (BP_IS_GANG(bp))
742 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
743 if (offset + asize > vd->vdev_asize) {
744 zfs_panic_recover("blkptr at %p DVA %u has invalid "
745 "OFFSET %llu",
746 bp, i, (longlong_t)offset);
747 }
748 }
749 }
750
751 boolean_t
752 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
753 {
754 uint64_t vdevid = DVA_GET_VDEV(dva);
755
756 if (vdevid >= spa->spa_root_vdev->vdev_children)
757 return (B_FALSE);
758
759 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
760 if (vd == NULL)
761 return (B_FALSE);
762
763 if (vd->vdev_ops == &vdev_hole_ops)
764 return (B_FALSE);
765
766 if (vd->vdev_ops == &vdev_missing_ops) {
767 return (B_FALSE);
768 }
769
770 uint64_t offset = DVA_GET_OFFSET(dva);
771 uint64_t asize = DVA_GET_ASIZE(dva);
772
773 if (BP_IS_GANG(bp))
774 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
775 if (offset + asize > vd->vdev_asize)
776 return (B_FALSE);
777
778 return (B_TRUE);
779 }
780
781 zio_t *
782 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
783 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
784 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
785 {
786 zio_t *zio;
787
788 zfs_blkptr_verify(spa, bp);
789
790 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
791 data, size, size, done, private,
792 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
793 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
794 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
795
796 return (zio);
797 }
798
799 zio_t *
800 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
801 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
802 zio_done_func_t *ready, zio_done_func_t *children_ready,
803 zio_done_func_t *physdone, zio_done_func_t *done,
804 void *private, zio_priority_t priority, enum zio_flag flags,
805 const zbookmark_phys_t *zb)
806 {
807 zio_t *zio;
808
809 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
810 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
811 zp->zp_compress >= ZIO_COMPRESS_OFF &&
812 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
813 DMU_OT_IS_VALID(zp->zp_type) &&
814 zp->zp_level < 32 &&
815 zp->zp_copies > 0 &&
816 zp->zp_copies <= spa_max_replication(spa));
817
818 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
819 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
820 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
821 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
822
823 zio->io_ready = ready;
824 zio->io_children_ready = children_ready;
825 zio->io_physdone = physdone;
826 zio->io_prop = *zp;
827
828 /*
829 * Data can be NULL if we are going to call zio_write_override() to
830 * provide the already-allocated BP. But we may need the data to
831 * verify a dedup hit (if requested). In this case, don't try to
832 * dedup (just take the already-allocated BP verbatim).
833 */
834 if (data == NULL && zio->io_prop.zp_dedup_verify) {
835 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
836 }
837
838 return (zio);
839 }
840
841 zio_t *
842 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
843 uint64_t size, zio_done_func_t *done, void *private,
844 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
845 {
846 zio_t *zio;
858 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
859 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
860 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
861 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
862
863 /*
864 * We must reset the io_prop to match the values that existed
865 * when the bp was first written by dmu_sync() keeping in mind
866 * that nopwrite and dedup are mutually exclusive.
867 */
868 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
869 zio->io_prop.zp_nopwrite = nopwrite;
870 zio->io_prop.zp_copies = copies;
871 zio->io_bp_override = bp;
872 }
873
874 void
875 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
876 {
877
878 zfs_blkptr_verify(spa, bp);
879
880 /*
881 * The check for EMBEDDED is a performance optimization. We
882 * process the free here (by ignoring it) rather than
883 * putting it on the list and then processing it in zio_free_sync().
884 */
885 if (BP_IS_EMBEDDED(bp))
886 return;
887 metaslab_check_free(spa, bp);
888
889 /*
890 * Frees that are for the currently-syncing txg, are not going to be
891 * deferred, and which will not need to do a read (i.e. not GANG or
892 * DEDUP), can be processed immediately. Otherwise, put them on the
893 * in-memory list for later processing.
894 */
895 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
896 txg != spa->spa_syncing_txg ||
897 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
898 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
899 } else {
900 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
901 }
902 }
903
904 zio_t *
905 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
906 enum zio_flag flags)
907 {
908 zio_t *zio;
909 enum zio_stage stage = ZIO_FREE_PIPELINE;
910
911 ASSERT(!BP_IS_HOLE(bp));
912 ASSERT(spa_syncing_txg(spa) == txg);
913 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
914
915 if (BP_IS_EMBEDDED(bp))
916 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
917
918 metaslab_check_free(spa, bp);
919 arc_freed(spa, bp);
920
921 /*
922 * GANG and DEDUP blocks can induce a read (for the gang block header,
923 * or the DDT), so issue them asynchronously so that this thread is
924 * not tied up.
925 */
926 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
927 stage |= ZIO_STAGE_ISSUE_ASYNC;
928
929 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
930 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
931 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
932
933 return (zio);
934 }
935
936 zio_t *
937 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
938 zio_done_func_t *done, void *private, enum zio_flag flags)
939 {
940 zio_t *zio;
941
942 zfs_blkptr_verify(spa, bp);
943
944 if (BP_IS_EMBEDDED(bp))
945 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
946
947 /*
948 * A claim is an allocation of a specific block. Claims are needed
949 * to support immediate writes in the intent log. The issue is that
950 * immediate writes contain committed data, but in a txg that was
951 * *not* committed. Upon opening the pool after an unclean shutdown,
952 * the intent log claims all blocks that contain immediate write data
953 * so that the SPA knows they're in use.
954 *
955 * All claims *must* be resolved in the first txg -- before the SPA
956 * starts allocating blocks -- so that nothing is allocated twice.
957 * If txg == 0 we just verify that the block is claimable.
958 */
959 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
960 ASSERT(txg == spa_first_txg(spa) || txg == 0);
961 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
962
963 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
964 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
965 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
966 ASSERT0(zio->io_queued_timestamp);
967
968 return (zio);
969 }
970
971 zio_t *
972 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
973 zio_done_func_t *done, void *private, enum zio_flag flags)
974 {
975 zio_t *zio;
976 int c;
977
978 if (vd->vdev_children == 0) {
979 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
980 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
981 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
982
983 zio->io_cmd = cmd;
984 } else {
985 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
986
987 for (c = 0; c < vd->vdev_children; c++)
988 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
989 done, private, flags));
990 }
991
992 return (zio);
993 }
994
995 zio_t *
996 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
997 abd_t *data, int checksum, zio_done_func_t *done, void *private,
998 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
999 {
1000 zio_t *zio;
1001
1002 ASSERT(vd->vdev_children == 0);
1003 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1004 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1005 ASSERT3U(offset + size, <=, vd->vdev_psize);
1006
1007 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1008 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1009 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1010
1011 zio->io_prop.zp_checksum = checksum;
1012
1013 return (zio);
1014 }
1015
1041 abd_t *wbuf = abd_alloc_sametype(data, size);
1042 abd_copy(wbuf, data, size);
1043
1044 zio_push_transform(zio, wbuf, size, size, NULL);
1045 }
1046
1047 return (zio);
1048 }
1049
1050 /*
1051 * Create a child I/O to do some work for us.
1052 */
1053 zio_t *
1054 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1055 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1056 enum zio_flag flags, zio_done_func_t *done, void *private)
1057 {
1058 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1059 zio_t *zio;
1060
1061 /*
1062 * vdev child I/Os do not propagate their error to the parent.
1063 * Therefore, for correct operation the caller *must* check for
1064 * and handle the error in the child i/o's done callback.
1065 * The only exceptions are i/os that we don't care about
1066 * (OPTIONAL or REPAIR).
1067 */
1068 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1069 done != NULL);
1070
1071 /*
1072 * In the common case, where the parent zio was to a normal vdev,
1073 * the child zio must be to a child vdev of that vdev. Otherwise,
1074 * the child zio must be to a top-level vdev.
1075 */
1076 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
1077 ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
1078 } else {
1079 ASSERT3P(vd, ==, vd->vdev_top);
1080 }
1081
1082 if (type == ZIO_TYPE_READ && bp != NULL) {
1083 /*
1084 * If we have the bp, then the child should perform the
1085 * checksum and the parent need not. This pushes error
1086 * detection as close to the leaves as possible and
1087 * eliminates redundant checksums in the interior nodes.
1088 */
1089 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1090 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1091 }
1092
1093 if (vd->vdev_ops->vdev_op_leaf) {
1094 ASSERT0(vd->vdev_children);
1095 offset += VDEV_LABEL_START_SIZE;
1096 }
1097
1098 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1099
1100 /*
1101 * If we've decided to do a repair, the write is not speculative --
1102 * even if the original read was.
1103 */
1104 if (flags & ZIO_FLAG_IO_REPAIR)
1105 flags &= ~ZIO_FLAG_SPECULATIVE;
1106
1107 /*
1108 * If we're creating a child I/O that is not associated with a
1109 * top-level vdev, then the child zio is not an allocating I/O.
1110 * If this is a retried I/O then we ignore it since we will
1111 * have already processed the original allocating I/O.
1112 */
1113 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1114 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1115 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1116
1117 ASSERT(mc->mc_alloc_throttle_enabled);
1118 ASSERT(type == ZIO_TYPE_WRITE);
1119 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1120 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1121 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1122 pio->io_child_type == ZIO_CHILD_GANG);
1123
1124 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1125 }
1126
1127 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1128 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1129 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1130 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1131
1132 zio->io_physdone = pio->io_physdone;
1133 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1134 zio->io_logical->io_phys_children++;
1135
1176 */
1177 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1178 if (!BP_IS_RAIDZ(zio->io_bp)) {
1179 /* we are not doing a raw write */
1180 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1181 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1182 }
1183 }
1184
1185 /*
1186 * ==========================================================================
1187 * Prepare to read and write logical blocks
1188 * ==========================================================================
1189 */
1190
1191 static int
1192 zio_read_bp_init(zio_t *zio)
1193 {
1194 blkptr_t *bp = zio->io_bp;
1195
1196 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1197
1198 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1199 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1200 !(zio->io_flags & ZIO_FLAG_RAW)) {
1201 uint64_t psize =
1202 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1203 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1204 psize, psize, zio_decompress);
1205 }
1206
1207 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1208 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1209
1210 int psize = BPE_GET_PSIZE(bp);
1211 void *data = abd_borrow_buf(zio->io_abd, psize);
1212 decode_embedded_bp_compressed(bp, data);
1213 abd_return_buf_copy(zio->io_abd, data, psize);
1214 } else {
1215 ASSERT(!BP_IS_EMBEDDED(bp));
1216 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1217 }
1218
1219 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1220 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1221
1222 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1223 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1224
1225 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1226 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1227
1228 return (ZIO_PIPELINE_CONTINUE);
1229 }
1230
1231 static int
1232 zio_write_bp_init(zio_t *zio)
1233 {
1234 if (!IO_IS_ALLOCATING(zio))
1235 return (ZIO_PIPELINE_CONTINUE);
1236
1237 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1238
1239 if (zio->io_bp_override) {
1287 return (ZIO_PIPELINE_CONTINUE);
1288 }
1289
1290 static int
1291 zio_write_compress(zio_t *zio)
1292 {
1293 spa_t *spa = zio->io_spa;
1294 zio_prop_t *zp = &zio->io_prop;
1295 enum zio_compress compress = zp->zp_compress;
1296 blkptr_t *bp = zio->io_bp;
1297 uint64_t lsize = zio->io_lsize;
1298 uint64_t psize = zio->io_size;
1299 int pass = 1;
1300
1301 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1302
1303 /*
1304 * If our children haven't all reached the ready stage,
1305 * wait for them and then repeat this pipeline stage.
1306 */
1307 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1308 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1309 return (ZIO_PIPELINE_STOP);
1310 }
1311
1312 if (!IO_IS_ALLOCATING(zio))
1313 return (ZIO_PIPELINE_CONTINUE);
1314
1315 if (zio->io_children_ready != NULL) {
1316 /*
1317 * Now that all our children are ready, run the callback
1318 * associated with this zio in case it wants to modify the
1319 * data to be written.
1320 */
1321 ASSERT3U(zp->zp_level, >, 0);
1322 zio->io_children_ready(zio);
1323 }
1324
1325 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1326 ASSERT(zio->io_bp_override == NULL);
1327
1328 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1329 /*
1330 * We're rewriting an existing block, which means we're
1332 * converge, it must eventually be the case that we don't
1333 * have to allocate new blocks. But compression changes
1334 * the blocksize, which forces a reallocate, and makes
1335 * convergence take longer. Therefore, after the first
1336 * few passes, stop compressing to ensure convergence.
1337 */
1338 pass = spa_sync_pass(spa);
1339
1340 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1342 ASSERT(!BP_GET_DEDUP(bp));
1343
1344 if (pass >= zfs_sync_pass_dont_compress)
1345 compress = ZIO_COMPRESS_OFF;
1346
1347 /* Make sure someone doesn't change their mind on overwrites */
1348 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1349 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1350 }
1351
1352 /* If it's a compressed write that is not raw, compress the buffer. */
1353 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1354 void *cbuf = zio_buf_alloc(lsize);
1355 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1356 if (psize == 0 || psize == lsize) {
1357 compress = ZIO_COMPRESS_OFF;
1358 zio_buf_free(cbuf, lsize);
1359 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1360 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1361 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1362 encode_embedded_bp_compressed(bp,
1363 cbuf, compress, lsize, psize);
1364 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1365 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1366 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1367 zio_buf_free(cbuf, lsize);
1368 bp->blk_birth = zio->io_txg;
1369 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1370 ASSERT(spa_feature_is_active(spa,
1371 SPA_FEATURE_EMBEDDED_DATA));
1372 return (ZIO_PIPELINE_CONTINUE);
1373 } else {
1374 /*
1375 * Round up compressed size up to the ashift
1376 * of the smallest-ashift device, and zero the tail.
1377 * This ensures that the compressed size of the BP
1378 * (and thus compressratio property) are correct,
1379 * in that we charge for the padding used to fill out
1380 * the last sector.
1381 */
1382 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1383 size_t rounded = (size_t)P2ROUNDUP(psize,
1384 1ULL << spa->spa_min_ashift);
1385 if (rounded >= lsize) {
1386 compress = ZIO_COMPRESS_OFF;
1387 zio_buf_free(cbuf, lsize);
1388 psize = lsize;
1389 } else {
1390 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1391 abd_take_ownership_of_buf(cdata, B_TRUE);
1392 abd_zero_off(cdata, psize, rounded - psize);
1393 psize = rounded;
1394 zio_push_transform(zio, cdata,
1395 psize, lsize, NULL);
1396 }
1397 }
1398
1399 /*
1400 * We were unable to handle this as an override bp, treat
1401 * it as a regular write I/O.
1402 */
1403 zio->io_bp_override = NULL;
1404 *bp = zio->io_bp_orig;
1405 zio->io_pipeline = zio->io_orig_pipeline;
1406 } else {
1407 ASSERT3U(psize, !=, 0);
1408 }
1409
1410 /*
1411 * The final pass of spa_sync() must be all rewrites, but the first
1412 * few passes offer a trade-off: allocating blocks defers convergence,
1413 * but newly allocated blocks are sequential, so they can be written
1414 * to disk faster. Therefore, we allow the first few passes of
1415 * spa_sync() to allocate new blocks, but force rewrites after that.
1416 * There should only be a handful of blocks after pass 1 in any case.
1417 */
1418 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1419 BP_GET_PSIZE(bp) == psize &&
1420 pass >= zfs_sync_pass_rewrite) {
1421 ASSERT(psize != 0);
1422 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1423 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1424 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1425 } else {
1426 BP_ZERO(bp);
1427 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1428 }
1429
1430 if (psize == 0) {
1431 if (zio->io_bp_orig.blk_birth != 0 &&
1432 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1433 BP_SET_LSIZE(bp, lsize);
1434 BP_SET_TYPE(bp, zp->zp_type);
1435 BP_SET_LEVEL(bp, zp->zp_level);
1436 BP_SET_BIRTH(bp, zio->io_txg, 0);
1437 }
1438 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1439 } else {
1440 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1441 BP_SET_LSIZE(bp, lsize);
1442 BP_SET_TYPE(bp, zp->zp_type);
1443 BP_SET_LEVEL(bp, zp->zp_level);
1444 BP_SET_PSIZE(bp, psize);
1445 BP_SET_COMPRESS(bp, compress);
1446 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1447 BP_SET_DEDUP(bp, zp->zp_dedup);
1448 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1449 if (zp->zp_dedup) {
1450 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1451 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1452 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1453 }
1454 if (zp->zp_nopwrite) {
1455 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1456 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1457 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1458 }
1459 }
1460 return (ZIO_PIPELINE_CONTINUE);
1461 }
1462
1463 static int
1464 zio_free_bp_init(zio_t *zio)
1465 {
1466 blkptr_t *bp = zio->io_bp;
1467
1468 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1469 if (BP_GET_DEDUP(bp))
1470 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1471 }
1472
1473 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1474
1475 return (ZIO_PIPELINE_CONTINUE);
1476 }
1477
1478 /*
1479 * ==========================================================================
1480 * Execute the I/O pipeline
1481 * ==========================================================================
1482 */
1483
1484 static void
1485 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1486 {
1487 spa_t *spa = zio->io_spa;
1488 zio_type_t t = zio->io_type;
1489 int flags = (cutinline ? TQ_FRONT : 0);
1490
1491 /*
1492 * If we're a config writer or a probe, the normal issue and
1493 * interrupt threads may all be blocked waiting for the config lock.
1494 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1495 */
1496 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1497 t = ZIO_TYPE_NULL;
1498
1499 /*
1500 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1501 */
1502 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1503 t = ZIO_TYPE_NULL;
1504
1505 /*
1506 * If this is a high priority I/O, then use the high priority taskq if
1507 * available.
1508 */
1509 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1510 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1511 q++;
1512
1513 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1514
1515 /*
1516 * NB: We are assuming that the zio can only be dispatched
1517 * to a single taskq at a time. It would be a grievous error
1518 * to dispatch the zio to another taskq at the same time.
1519 */
1520 ASSERT(zio->io_tqent.tqent_next == NULL);
1521 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1522 flags, &zio->io_tqent);
1523 }
1524
1525 static boolean_t
1526 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1527 {
1528 kthread_t *executor = zio->io_executor;
1529 spa_t *spa = zio->io_spa;
1616 * (5) the I/O is deferred due to vdev-level queueing
1617 * (6) the I/O is handed off to another thread.
1618 *
1619 * In all cases, the pipeline stops whenever there's no CPU work; it never
1620 * burns a thread in cv_wait().
1621 *
1622 * There's no locking on io_stage because there's no legitimate way
1623 * for multiple threads to be attempting to process the same I/O.
1624 */
1625 static zio_pipe_stage_t *zio_pipeline[];
1626
1627 void
1628 zio_execute(zio_t *zio)
1629 {
1630 zio->io_executor = curthread;
1631
1632 ASSERT3U(zio->io_queued_timestamp, >, 0);
1633
1634 while (zio->io_stage < ZIO_STAGE_DONE) {
1635 enum zio_stage pipeline = zio->io_pipeline;
1636 enum zio_stage stage = zio->io_stage;
1637 int rv;
1638
1639 ASSERT(!MUTEX_HELD(&zio->io_lock));
1640 ASSERT(ISP2(stage));
1641 ASSERT(zio->io_stall == NULL);
1642
1643 do {
1644 stage <<= 1;
1645 } while ((stage & pipeline) == 0);
1646
1647 ASSERT(stage <= ZIO_STAGE_DONE);
1648
1649 /*
1650 * If we are in interrupt context and this pipeline stage
1651 * will grab a config lock that is held across I/O,
1652 * or may wait for an I/O that needs an interrupt thread
1653 * to complete, issue async to avoid deadlock.
1654 *
1655 * For VDEV_IO_START, we cut in line so that the io will
1656 * be sent to disk promptly.
1657 */
1658 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1659 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1660 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1661 zio_requeue_io_start_cut_in_line : B_FALSE;
1662 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1663 return;
1664 }
1665
1666 zio->io_stage = stage;
1667 zio->io_pipeline_trace |= zio->io_stage;
1668 rv = zio_pipeline[highbit64(stage) - 1](zio);
1669
1670 if (rv == ZIO_PIPELINE_STOP)
1671 return;
1672
1673 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1674 }
1675 }
1676
1677 /*
1678 * ==========================================================================
1679 * Initiate I/O, either sync or async
1680 * ==========================================================================
1681 */
1682 int
1683 zio_wait(zio_t *zio)
1684 {
1685 int error;
1686
1687 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1688 ASSERT3P(zio->io_executor, ==, NULL);
1689
1690 zio->io_waiter = curthread;
1691 ASSERT0(zio->io_queued_timestamp);
1692 zio->io_queued_timestamp = gethrtime();
2133 static int
2134 zio_gang_assemble(zio_t *zio)
2135 {
2136 blkptr_t *bp = zio->io_bp;
2137
2138 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2139 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2140
2141 zio->io_gang_leader = zio;
2142
2143 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2144
2145 return (ZIO_PIPELINE_CONTINUE);
2146 }
2147
2148 static int
2149 zio_gang_issue(zio_t *zio)
2150 {
2151 blkptr_t *bp = zio->io_bp;
2152
2153 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2154 return (ZIO_PIPELINE_STOP);
2155 }
2156
2157 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2158 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2159
2160 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2161 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2162 0);
2163 else
2164 zio_gang_tree_free(&zio->io_gang_tree);
2165
2166 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2167
2168 return (ZIO_PIPELINE_CONTINUE);
2169 }
2170
2171 static void
2172 zio_write_gang_member_ready(zio_t *zio)
2173 {
2174 zio_t *pio = zio_unique_parent(zio);
2175 zio_t *gio = zio->io_gang_leader;
2191 mutex_enter(&pio->io_lock);
2192 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2193 ASSERT(DVA_GET_GANG(&pdva[d]));
2194 asize = DVA_GET_ASIZE(&pdva[d]);
2195 asize += DVA_GET_ASIZE(&cdva[d]);
2196 DVA_SET_ASIZE(&pdva[d], asize);
2197 }
2198 mutex_exit(&pio->io_lock);
2199 }
2200
2201 static void
2202 zio_write_gang_done(zio_t *zio)
2203 {
2204 abd_put(zio->io_abd);
2205 }
2206
2207 static int
2208 zio_write_gang_block(zio_t *pio)
2209 {
2210 spa_t *spa = pio->io_spa;
2211 metaslab_class_t *mc = spa_normal_class(spa);
2212 blkptr_t *bp = pio->io_bp;
2213 zio_t *gio = pio->io_gang_leader;
2214 zio_t *zio;
2215 zio_gang_node_t *gn, **gnpp;
2216 zio_gbh_phys_t *gbh;
2217 abd_t *gbh_abd;
2218 uint64_t txg = pio->io_txg;
2219 uint64_t resid = pio->io_size;
2220 uint64_t lsize;
2221 int copies = gio->io_prop.zp_copies;
2222 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2223 zio_prop_t zp;
2224 int error;
2225
2226 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2227 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2228 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2229 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2230
2231 flags |= METASLAB_ASYNC_ALLOC;
2288 * Create and nowait the gang children.
2289 */
2290 for (int g = 0; resid != 0; resid -= lsize, g++) {
2291 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2292 SPA_MINBLOCKSIZE);
2293 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2294
2295 zp.zp_checksum = gio->io_prop.zp_checksum;
2296 zp.zp_compress = ZIO_COMPRESS_OFF;
2297 zp.zp_type = DMU_OT_NONE;
2298 zp.zp_level = 0;
2299 zp.zp_copies = gio->io_prop.zp_copies;
2300 zp.zp_dedup = B_FALSE;
2301 zp.zp_dedup_verify = B_FALSE;
2302 zp.zp_nopwrite = B_FALSE;
2303
2304 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2305 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2306 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2307 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2308 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2309
2310 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2311 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2312 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2313
2314 /*
2315 * Gang children won't throttle but we should
2316 * account for their work, so reserve an allocation
2317 * slot for them here.
2318 */
2319 VERIFY(metaslab_class_throttle_reserve(mc,
2320 zp.zp_copies, cio, flags));
2321 }
2322 zio_nowait(cio);
2323 }
2324
2325 /*
2326 * Set pio's pipeline to just wait for zio to finish.
2327 */
2328 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2329
2456 abd_alloc_for_io(zio->io_size, B_TRUE),
2457 zio->io_size, zio_ddt_child_read_done, dde,
2458 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2459 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2460 }
2461 return (ZIO_PIPELINE_CONTINUE);
2462 }
2463
2464 zio_nowait(zio_read(zio, zio->io_spa, bp,
2465 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2466 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2467
2468 return (ZIO_PIPELINE_CONTINUE);
2469 }
2470
2471 static int
2472 zio_ddt_read_done(zio_t *zio)
2473 {
2474 blkptr_t *bp = zio->io_bp;
2475
2476 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2477 return (ZIO_PIPELINE_STOP);
2478 }
2479
2480 ASSERT(BP_GET_DEDUP(bp));
2481 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2482 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2483
2484 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2485 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2486 ddt_entry_t *dde = zio->io_vsd;
2487 if (ddt == NULL) {
2488 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2489 return (ZIO_PIPELINE_CONTINUE);
2490 }
2491 if (dde == NULL) {
2492 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2493 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2494 return (ZIO_PIPELINE_STOP);
2495 }
2496 if (dde->dde_repair_abd != NULL) {
2497 abd_copy(zio->io_abd, dde->dde_repair_abd,
2498 zio->io_size);
2499 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2500 }
2501 ddt_repair_done(ddt, dde);
2502 zio->io_vsd = NULL;
2503 }
2504
2505 ASSERT(zio->io_vsd == NULL);
2506
2507 return (ZIO_PIPELINE_CONTINUE);
2508 }
2509
2510 static boolean_t
2511 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2512 {
2513 spa_t *spa = zio->io_spa;
2514 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2515
2516 /* We should never get a raw, override zio */
2517 ASSERT(!(zio->io_bp_override && do_raw));
2518
2519 /*
2520 * Note: we compare the original data, not the transformed data,
2521 * because when zio->io_bp is an override bp, we will not have
2522 * pushed the I/O transforms. That's an important optimization
2523 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2524 */
2525 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2526 zio_t *lio = dde->dde_lead_zio[p];
2527
2528 if (lio != NULL) {
2529 return (lio->io_orig_size != zio->io_orig_size ||
2530 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2531 zio->io_orig_size) != 0);
2532 }
2533 }
2534
2535 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2536 ddt_phys_t *ddp = &dde->dde_phys[p];
2537
2538 if (ddp->ddp_phys_birth != 0) {
2539 arc_buf_t *abuf = NULL;
2540 arc_flags_t aflags = ARC_FLAG_WAIT;
2541 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2542 blkptr_t blk = *zio->io_bp;
2543 int error;
2544
2545 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2546
2547 ddt_exit(ddt);
2548
2549 /*
2550 * Intuitively, it would make more sense to compare
2551 * io_abd than io_orig_abd in the raw case since you
2552 * don't want to look at any transformations that have
2553 * happened to the data. However, for raw I/Os the
2554 * data will actually be the same in io_abd and
2555 * io_orig_abd, so all we have to do is issue this as
2556 * a raw ARC read.
2557 */
2558 if (do_raw) {
2559 zio_flags |= ZIO_FLAG_RAW;
2560 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2561 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2562 zio->io_size));
2563 ASSERT3P(zio->io_transform_stack, ==, NULL);
2564 }
2565
2566 error = arc_read(NULL, spa, &blk,
2567 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2568 zio_flags, &aflags, &zio->io_bookmark);
2569
2570 if (error == 0) {
2571 if (arc_buf_size(abuf) != zio->io_orig_size ||
2572 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2573 zio->io_orig_size) != 0)
2574 error = SET_ERROR(EEXIST);
2575 arc_buf_destroy(abuf, &abuf);
2576 }
2577
2578 ddt_enter(ddt);
2579 return (error != 0);
2580 }
2581 }
2582
2583 return (B_FALSE);
2584 }
2585
2586 static void
2587 zio_ddt_child_write_ready(zio_t *zio)
2588 {
2589 int p = zio->io_prop.zp_copies;
2590 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2591 ddt_entry_t *dde = zio->io_private;
2592 ddt_phys_t *ddp = &dde->dde_phys[p];
2593 zio_t *pio;
2594
2595 if (zio->io_error)
2596 return;
2597
2598 ddt_enter(ddt);
2599
2600 ASSERT(dde->dde_lead_zio[p] == zio);
2601
2602 ddt_phys_fill(ddp, zio->io_bp);
2603
2604 zio_link_t *zl = NULL;
2605 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2606 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2607
2608 ddt_exit(ddt);
2609 }
2610
2611 static void
2612 zio_ddt_child_write_done(zio_t *zio)
2613 {
2614 int p = zio->io_prop.zp_copies;
2615 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2616 ddt_entry_t *dde = zio->io_private;
2617 ddt_phys_t *ddp = &dde->dde_phys[p];
2618
2619 ddt_enter(ddt);
2620
2621 ASSERT(ddp->ddp_refcnt == 0);
2622 ASSERT(dde->dde_lead_zio[p] == zio);
2623 dde->dde_lead_zio[p] = NULL;
2624
2625 if (zio->io_error == 0) {
2626 zio_link_t *zl = NULL;
2627 while (zio_walk_parents(zio, &zl) != NULL)
2628 ddt_phys_addref(ddp);
2629 } else {
2630 ddt_phys_clear(ddp);
2631 }
2632
2633 ddt_exit(ddt);
2634 }
2635
2636 static void
2637 zio_ddt_ditto_write_done(zio_t *zio)
2638 {
2639 int p = DDT_PHYS_DITTO;
2640 zio_prop_t *zp = &zio->io_prop;
2641 blkptr_t *bp = zio->io_bp;
2642 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2643 ddt_entry_t *dde = zio->io_private;
2644 ddt_phys_t *ddp = &dde->dde_phys[p];
2645 ddt_key_t *ddk = &dde->dde_key;
2646
2647 ddt_enter(ddt);
2648
2649 ASSERT(ddp->ddp_refcnt == 0);
2650 ASSERT(dde->dde_lead_zio[p] == zio);
2651 dde->dde_lead_zio[p] = NULL;
2652
2653 if (zio->io_error == 0) {
2654 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2655 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2656 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2657 if (ddp->ddp_phys_birth != 0)
2658 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2659 ddt_phys_fill(ddp, bp);
2660 }
2661
2662 ddt_exit(ddt);
2663 }
2664
2665 static int
2666 zio_ddt_write(zio_t *zio)
2667 {
2668 spa_t *spa = zio->io_spa;
2669 blkptr_t *bp = zio->io_bp;
2670 uint64_t txg = zio->io_txg;
2671 zio_prop_t *zp = &zio->io_prop;
2672 int p = zp->zp_copies;
2673 int ditto_copies;
2674 zio_t *cio = NULL;
2675 zio_t *dio = NULL;
2676 ddt_t *ddt = ddt_select(spa, bp);
2677 ddt_entry_t *dde;
2678 ddt_phys_t *ddp;
2679
2680 ASSERT(BP_GET_DEDUP(bp));
2681 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2682 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2683 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2684
2685 ddt_enter(ddt);
2686 dde = ddt_lookup(ddt, bp, B_TRUE);
2687 ddp = &dde->dde_phys[p];
2688
2689 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2690 /*
2691 * If we're using a weak checksum, upgrade to a strong checksum
2692 * and try again. If we're already using a strong checksum,
2693 * we can't resolve it, so just convert to an ordinary write.
2694 * (And automatically e-mail a paper to Nature?)
2695 */
2696 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2697 ZCHECKSUM_FLAG_DEDUP)) {
2698 zp->zp_checksum = spa_dedup_checksum(spa);
2699 zio_pop_transforms(zio);
2700 zio->io_stage = ZIO_STAGE_OPEN;
2701 BP_ZERO(bp);
2702 } else {
2703 zp->zp_dedup = B_FALSE;
2704 BP_SET_DEDUP(bp, B_FALSE);
2705 }
2706 ASSERT(!BP_GET_DEDUP(bp));
2707 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2708 ddt_exit(ddt);
2709 return (ZIO_PIPELINE_CONTINUE);
2710 }
2711
2712 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2713 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2714
2715 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2716 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2717 zio_prop_t czp = *zp;
2718
2719 czp.zp_copies = ditto_copies;
2720
2721 /*
2722 * If we arrived here with an override bp, we won't have run
2723 * the transform stack, so we won't have the data we need to
2724 * generate a child i/o. So, toss the override bp and restart.
2725 * This is safe, because using the override bp is just an
2726 * optimization; and it's rare, so the cost doesn't matter.
2727 */
2728 if (zio->io_bp_override) {
2729 zio_pop_transforms(zio);
2730 zio->io_stage = ZIO_STAGE_OPEN;
2731 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2732 zio->io_bp_override = NULL;
2733 BP_ZERO(bp);
2734 ddt_exit(ddt);
2735 return (ZIO_PIPELINE_CONTINUE);
2736 }
2737
2738 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2739 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2740 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2741 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2742
2743 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2744 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2745 }
2746
2747 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2748 if (ddp->ddp_phys_birth != 0)
2749 ddt_bp_fill(ddp, bp, txg);
2750 if (dde->dde_lead_zio[p] != NULL)
2751 zio_add_child(zio, dde->dde_lead_zio[p]);
2752 else
2753 ddt_phys_addref(ddp);
2754 } else if (zio->io_bp_override) {
2755 ASSERT(bp->blk_birth == txg);
2756 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2757 ddt_phys_fill(ddp, bp);
2758 ddt_phys_addref(ddp);
2759 } else {
2760 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2761 zio->io_orig_size, zio->io_orig_size, zp,
2762 zio_ddt_child_write_ready, NULL, NULL,
2763 zio_ddt_child_write_done, dde, zio->io_priority,
2764 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2765
2766 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2767 dde->dde_lead_zio[p] = cio;
2768 }
2769
2770 ddt_exit(ddt);
2771
2772 if (cio)
2773 zio_nowait(cio);
2774 if (dio)
2775 zio_nowait(dio);
2776
2777 return (ZIO_PIPELINE_CONTINUE);
2778 }
2779
2780 ddt_entry_t *freedde; /* for debugging */
2781
2782 static int
2783 zio_ddt_free(zio_t *zio)
2784 {
2785 spa_t *spa = zio->io_spa;
2786 blkptr_t *bp = zio->io_bp;
2787 ddt_t *ddt = ddt_select(spa, bp);
2788 ddt_entry_t *dde;
2789 ddt_phys_t *ddp;
2790
2791 ASSERT(BP_GET_DEDUP(bp));
2792 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2793
2794 ddt_enter(ddt);
2795 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2796 ddp = ddt_phys_select(dde, bp);
2797 ddt_phys_decref(ddp);
2798 ddt_exit(ddt);
2799
2800 return (ZIO_PIPELINE_CONTINUE);
2801 }
2802
2803 /*
2804 * ==========================================================================
2805 * Allocate and free blocks
2806 * ==========================================================================
2807 */
2808
2809 static zio_t *
2810 zio_io_to_allocate(spa_t *spa)
2811 {
2812 zio_t *zio;
2813
2814 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2815
2816 zio = avl_first(&spa->spa_alloc_tree);
2817 if (zio == NULL)
2818 return (NULL);
2819
2820 ASSERT(IO_IS_ALLOCATING(zio));
2821
2822 /*
2823 * Try to place a reservation for this zio. If we're unable to
2824 * reserve then we throttle.
2825 */
2826 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2827 zio->io_prop.zp_copies, zio, 0)) {
2828 return (NULL);
2829 }
2830
2831 avl_remove(&spa->spa_alloc_tree, zio);
2832 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2833
2834 return (zio);
2835 }
2836
2837 static int
2838 zio_dva_throttle(zio_t *zio)
2839 {
2840 spa_t *spa = zio->io_spa;
2841 zio_t *nio;
2842
2843 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2844 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2845 zio->io_child_type == ZIO_CHILD_GANG ||
2846 zio->io_flags & ZIO_FLAG_NODATA) {
2847 return (ZIO_PIPELINE_CONTINUE);
2848 }
2849
2850 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2851
2852 ASSERT3U(zio->io_queued_timestamp, >, 0);
2853 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2854
2855 mutex_enter(&spa->spa_alloc_lock);
2856
2857 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2858 avl_add(&spa->spa_alloc_tree, zio);
2859
2860 nio = zio_io_to_allocate(zio->io_spa);
2861 mutex_exit(&spa->spa_alloc_lock);
2862
2863 if (nio == zio)
2864 return (ZIO_PIPELINE_CONTINUE);
2865
2866 if (nio != NULL) {
2867 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2868 /*
2869 * We are passing control to a new zio so make sure that
2870 * it is processed by a different thread. We do this to
2871 * avoid stack overflows that can occur when parents are
2872 * throttled and children are making progress. We allow
2873 * it to go to the head of the taskq since it's already
2874 * been waiting.
2875 */
2876 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2877 }
2878 return (ZIO_PIPELINE_STOP);
2879 }
2880
2881 void
2882 zio_allocate_dispatch(spa_t *spa)
2883 {
2884 zio_t *zio;
2885
2886 mutex_enter(&spa->spa_alloc_lock);
2887 zio = zio_io_to_allocate(spa);
2888 mutex_exit(&spa->spa_alloc_lock);
2889 if (zio == NULL)
2890 return;
2891
2892 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2893 ASSERT0(zio->io_error);
2894 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2895 }
2896
2897 static int
2898 zio_dva_allocate(zio_t *zio)
2899 {
2900 spa_t *spa = zio->io_spa;
2901 metaslab_class_t *mc = spa_normal_class(spa);
2902 blkptr_t *bp = zio->io_bp;
2903 int error;
2904 int flags = 0;
2905
2906 if (zio->io_gang_leader == NULL) {
2907 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2908 zio->io_gang_leader = zio;
2909 }
2910
2911 ASSERT(BP_IS_HOLE(bp));
2912 ASSERT0(BP_GET_NDVAS(bp));
2913 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2914 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2915 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2916
2917 if (zio->io_flags & ZIO_FLAG_NODATA) {
2918 flags |= METASLAB_DONT_THROTTLE;
2919 }
2920 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2921 flags |= METASLAB_GANG_CHILD;
2922 }
2923 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2924 flags |= METASLAB_ASYNC_ALLOC;
2925 }
2926
2927 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2928 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2929 &zio->io_alloc_list, zio);
2930
2931 if (error != 0) {
2932 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2933 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2934 error);
2935 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2936 return (zio_write_gang_block(zio));
2937 zio->io_error = error;
2938 }
2939
2940 return (ZIO_PIPELINE_CONTINUE);
2941 }
2942
2943 static int
2944 zio_dva_free(zio_t *zio)
2945 {
2946 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2947
2948 return (ZIO_PIPELINE_CONTINUE);
2949 }
2950
2951 static int
2952 zio_dva_claim(zio_t *zio)
2953 {
2954 int error;
2955
2956 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2974 if (!BP_IS_HOLE(bp))
2975 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2976
2977 if (gn != NULL) {
2978 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2979 zio_dva_unallocate(zio, gn->gn_child[g],
2980 &gn->gn_gbh->zg_blkptr[g]);
2981 }
2982 }
2983 }
2984
2985 /*
2986 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2987 */
2988 int
2989 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2990 uint64_t size, boolean_t *slog)
2991 {
2992 int error = 1;
2993 zio_alloc_list_t io_alloc_list;
2994
2995 ASSERT(txg > spa_syncing_txg(spa));
2996
2997 metaslab_trace_init(&io_alloc_list);
2998 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
2999 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3000 if (error == 0) {
3001 *slog = TRUE;
3002 } else {
3003 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3004 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3005 &io_alloc_list, NULL);
3006 if (error == 0)
3007 *slog = FALSE;
3008 }
3009 metaslab_trace_fini(&io_alloc_list);
3010
3011 if (error == 0) {
3012 BP_SET_LSIZE(new_bp, size);
3013 BP_SET_PSIZE(new_bp, size);
3014 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3015 BP_SET_CHECKSUM(new_bp,
3016 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3017 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3018 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3019 BP_SET_LEVEL(new_bp, 0);
3020 BP_SET_DEDUP(new_bp, 0);
3021 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3022 } else {
3023 zfs_dbgmsg("%s: zil block allocation failure: "
3024 "size %llu, error %d", spa_name(spa), size, error);
3025 }
3026
3027 return (error);
3028 }
3045 * ==========================================================================
3046 */
3047
3048
3049 /*
3050 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3051 * stops after this stage and will resume upon I/O completion.
3052 * However, there are instances where the vdev layer may need to
3053 * continue the pipeline when an I/O was not issued. Since the I/O
3054 * that was sent to the vdev layer might be different than the one
3055 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3056 * force the underlying vdev layers to call either zio_execute() or
3057 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3058 */
3059 static int
3060 zio_vdev_io_start(zio_t *zio)
3061 {
3062 vdev_t *vd = zio->io_vd;
3063 uint64_t align;
3064 spa_t *spa = zio->io_spa;
3065
3066 ASSERT(zio->io_error == 0);
3067 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3068
3069 if (vd == NULL) {
3070 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3071 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3072
3073 /*
3074 * The mirror_ops handle multiple DVAs in a single BP.
3075 */
3076 vdev_mirror_ops.vdev_op_io_start(zio);
3077 return (ZIO_PIPELINE_STOP);
3078 }
3079
3080 ASSERT3P(zio->io_logical, !=, zio);
3081 if (zio->io_type == ZIO_TYPE_WRITE) {
3082 ASSERT(spa->spa_trust_config);
3083
3084 if (zio->io_vd->vdev_removing) {
3085 ASSERT(zio->io_flags &
3086 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3087 ZIO_FLAG_INDUCE_DAMAGE));
3088 }
3089 }
3090
3091 /*
3092 * We keep track of time-sensitive I/Os so that the scan thread
3093 * can quickly react to certain workloads. In particular, we care
3094 * about non-scrubbing, top-level reads and writes with the following
3095 * characteristics:
3096 * - synchronous writes of user data to non-slog devices
3097 * - any reads of user data
3098 * When these conditions are met, adjust the timestamp of spa_last_io
3099 * which allows the scan thread to adjust its workload accordingly.
3100 */
3101 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3102 vd == vd->vdev_top && !vd->vdev_islog &&
3103 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3104 zio->io_txg != spa_syncing_txg(spa)) {
3105 uint64_t old = spa->spa_last_io;
3106 uint64_t new = ddi_get_lbolt64();
3107 if (old != new)
3108 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3109 }
3110
3111 align = 1ULL << vd->vdev_top->vdev_ashift;
3112
3113 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3114 P2PHASE(zio->io_size, align) != 0) {
3115 /* Transform logical writes to be a full physical block size. */
3116 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3117 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3118 ASSERT(vd == vd->vdev_top);
3119 if (zio->io_type == ZIO_TYPE_WRITE) {
3120 abd_copy(abuf, zio->io_abd, zio->io_size);
3121 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3122 }
3123 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3124 }
3125
3126 /*
3127 * If this is not a physical io, make sure that it is properly aligned
3128 * before proceeding.
3129 */
3130 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3131 ASSERT0(P2PHASE(zio->io_offset, align));
3132 ASSERT0(P2PHASE(zio->io_size, align));
3133 } else {
3134 /*
3135 * For physical writes, we allow 512b aligned writes and assume
3136 * the device will perform a read-modify-write as necessary.
3137 */
3138 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3139 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3140 }
3141
3142 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3143
3144 /*
3145 * If this is a repair I/O, and there's no self-healing involved --
3146 * that is, we're just resilvering what we expect to resilver --
3147 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3148 * This prevents spurious resilvering with nested replication.
3149 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3150 * A is out of date, we'll read from C+D, then use the data to
3151 * resilver A+B -- but we don't actually want to resilver B, just A.
3152 * The top-level mirror has no way to know this, so instead we just
3153 * discard unnecessary repairs as we work our way down the vdev tree.
3154 * The same logic applies to any form of nested replication:
3155 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3156 */
3157 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3158 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3159 zio->io_txg != 0 && /* not a delegated i/o */
3160 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3161 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3162 zio_vdev_io_bypass(zio);
3163 return (ZIO_PIPELINE_CONTINUE);
3164 }
3165
3166 if (vd->vdev_ops->vdev_op_leaf &&
3167 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
3168
3169 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3170 return (ZIO_PIPELINE_CONTINUE);
3171
3172 if ((zio = vdev_queue_io(zio)) == NULL)
3173 return (ZIO_PIPELINE_STOP);
3174
3175 if (!vdev_accessible(vd, zio)) {
3176 zio->io_error = SET_ERROR(ENXIO);
3177 zio_interrupt(zio);
3178 return (ZIO_PIPELINE_STOP);
3179 }
3180 }
3181
3182 vd->vdev_ops->vdev_op_io_start(zio);
3183 return (ZIO_PIPELINE_STOP);
3184 }
3185
3186 static int
3187 zio_vdev_io_done(zio_t *zio)
3188 {
3189 vdev_t *vd = zio->io_vd;
3190 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3191 boolean_t unexpected_error = B_FALSE;
3192
3193 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3194 return (ZIO_PIPELINE_STOP);
3195 }
3196
3197 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3198
3199 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3200
3201 vdev_queue_io_done(zio);
3202
3203 if (zio->io_type == ZIO_TYPE_WRITE)
3204 vdev_cache_write(zio);
3205
3206 if (zio_injection_enabled && zio->io_error == 0)
3207 zio->io_error = zio_handle_device_injection(vd,
3208 zio, EIO);
3209
3210 if (zio_injection_enabled && zio->io_error == 0)
3211 zio->io_error = zio_handle_label_injection(zio, EIO);
3212
3213 if (zio->io_error) {
3214 if (!vdev_accessible(vd, zio)) {
3215 zio->io_error = SET_ERROR(ENXIO);
3216 } else {
3217 unexpected_error = B_TRUE;
3218 }
3219 }
3220 }
3221
3222 ops->vdev_op_io_done(zio);
3223
3224 if (unexpected_error)
3225 VERIFY(vdev_probe(vd, zio) == NULL);
3226
3227 return (ZIO_PIPELINE_CONTINUE);
3228 }
3229
3230 /*
3231 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3232 * disk, and use that to finish the checksum ereport later.
3233 */
3234 static void
3235 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3236 const void *good_buf)
3237 {
3238 /* no processing needed */
3239 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3240 }
3241
3242 /*ARGSUSED*/
3243 void
3244 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3245 {
3246 void *buf = zio_buf_alloc(zio->io_size);
3247
3248 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3249
3250 zcr->zcr_cbinfo = zio->io_size;
3251 zcr->zcr_cbdata = buf;
3252 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3253 zcr->zcr_free = zio_buf_free;
3254 }
3255
3256 static int
3257 zio_vdev_io_assess(zio_t *zio)
3258 {
3259 vdev_t *vd = zio->io_vd;
3260
3261 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3262 return (ZIO_PIPELINE_STOP);
3263 }
3264
3265 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3266 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3267
3268 if (zio->io_vsd != NULL) {
3269 zio->io_vsd_ops->vsd_free(zio);
3270 zio->io_vsd = NULL;
3271 }
3272
3273 if (zio_injection_enabled && zio->io_error == 0)
3274 zio->io_error = zio_handle_fault_injection(zio, EIO);
3275
3276 /*
3277 * If the I/O failed, determine whether we should attempt to retry it.
3278 *
3279 * On retry, we cut in line in the issue queue, since we don't want
3280 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3281 */
3282 if (zio->io_error && vd == NULL &&
3283 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3458
3459 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3460 if (e2 == zio_error_rank[r2])
3461 break;
3462
3463 return (r1 > r2 ? e1 : e2);
3464 }
3465
3466 /*
3467 * ==========================================================================
3468 * I/O completion
3469 * ==========================================================================
3470 */
3471 static int
3472 zio_ready(zio_t *zio)
3473 {
3474 blkptr_t *bp = zio->io_bp;
3475 zio_t *pio, *pio_next;
3476 zio_link_t *zl = NULL;
3477
3478 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3479 ZIO_WAIT_READY)) {
3480 return (ZIO_PIPELINE_STOP);
3481 }
3482
3483 if (zio->io_ready) {
3484 ASSERT(IO_IS_ALLOCATING(zio));
3485 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3486 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3487 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3488
3489 zio->io_ready(zio);
3490 }
3491
3492 if (bp != NULL && bp != &zio->io_bp_copy)
3493 zio->io_bp_copy = *bp;
3494
3495 if (zio->io_error != 0) {
3496 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3497
3498 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3499 ASSERT(IO_IS_ALLOCATING(zio));
3500 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3501 /*
3502 * We were unable to allocate anything, unreserve and
3503 * issue the next I/O to allocate.
3504 */
3505 metaslab_class_throttle_unreserve(
3506 spa_normal_class(zio->io_spa),
3507 zio->io_prop.zp_copies, zio);
3508 zio_allocate_dispatch(zio->io_spa);
3509 }
3510 }
3511
3512 mutex_enter(&zio->io_lock);
3513 zio->io_state[ZIO_WAIT_READY] = 1;
3514 pio = zio_walk_parents(zio, &zl);
3515 mutex_exit(&zio->io_lock);
3516
3517 /*
3518 * As we notify zio's parents, new parents could be added.
3519 * New parents go to the head of zio's io_parent_list, however,
3520 * so we will (correctly) not notify them. The remainder of zio's
3521 * io_parent_list, from 'pio_next' onward, cannot change because
3522 * all parents must wait for us to be done before they can be done.
3523 */
3524 for (; pio != NULL; pio = pio_next) {
3525 pio_next = zio_walk_parents(zio, &zl);
3526 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3527 }
3528
3574 if (pio->io_child_type == ZIO_CHILD_GANG) {
3575 /*
3576 * If our parent is a rewrite gang child then our grandparent
3577 * would have been the one that performed the allocation.
3578 */
3579 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3580 pio = zio_unique_parent(pio);
3581 flags |= METASLAB_GANG_CHILD;
3582 }
3583
3584 ASSERT(IO_IS_ALLOCATING(pio));
3585 ASSERT3P(zio, !=, zio->io_logical);
3586 ASSERT(zio->io_logical != NULL);
3587 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3588 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3589
3590 mutex_enter(&pio->io_lock);
3591 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3592 mutex_exit(&pio->io_lock);
3593
3594 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3595 1, pio);
3596
3597 /*
3598 * Call into the pipeline to see if there is more work that
3599 * needs to be done. If there is work to be done it will be
3600 * dispatched to another taskq thread.
3601 */
3602 zio_allocate_dispatch(zio->io_spa);
3603 }
3604
3605 static int
3606 zio_done(zio_t *zio)
3607 {
3608 spa_t *spa = zio->io_spa;
3609 zio_t *lio = zio->io_logical;
3610 blkptr_t *bp = zio->io_bp;
3611 vdev_t *vd = zio->io_vd;
3612 uint64_t psize = zio->io_size;
3613 zio_t *pio, *pio_next;
3614 metaslab_class_t *mc = spa_normal_class(spa);
3615 zio_link_t *zl = NULL;
3616
3617 /*
3618 * If our children haven't all completed,
3619 * wait for them and then repeat this pipeline stage.
3620 */
3621 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3622 return (ZIO_PIPELINE_STOP);
3623 }
3624
3625 /*
3626 * If the allocation throttle is enabled, then update the accounting.
3627 * We only track child I/Os that are part of an allocating async
3628 * write. We must do this since the allocation is performed
3629 * by the logical I/O but the actual write is done by child I/Os.
3630 */
3631 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3632 zio->io_child_type == ZIO_CHILD_VDEV) {
3633 ASSERT(mc->mc_alloc_throttle_enabled);
3634 zio_dva_throttle_done(zio);
3635 }
3636
3637 /*
3638 * If the allocation throttle is enabled, verify that
3639 * we have decremented the refcounts for every I/O that was throttled.
3640 */
3641 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3642 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3643 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3893 zl = NULL;
3894 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3895 zio_link_t *remove_zl = zl;
3896 pio_next = zio_walk_parents(zio, &zl);
3897 zio_remove_child(pio, zio, remove_zl);
3898 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3899 }
3900
3901 if (zio->io_waiter != NULL) {
3902 mutex_enter(&zio->io_lock);
3903 zio->io_executor = NULL;
3904 cv_broadcast(&zio->io_cv);
3905 mutex_exit(&zio->io_lock);
3906 } else {
3907 zio_destroy(zio);
3908 }
3909
3910 return (ZIO_PIPELINE_STOP);
3911 }
3912
3913 /*
3914 * ==========================================================================
3915 * I/O pipeline definition
3916 * ==========================================================================
3917 */
3918 static zio_pipe_stage_t *zio_pipeline[] = {
3919 NULL,
3920 zio_read_bp_init,
3921 zio_write_bp_init,
3922 zio_free_bp_init,
3923 zio_issue_async,
3924 zio_write_compress,
3925 zio_checksum_generate,
3926 zio_nop_write,
3927 zio_ddt_read_start,
3928 zio_ddt_read_done,
3929 zio_ddt_write,
3930 zio_ddt_free,
3931 zio_gang_assemble,
3932 zio_gang_issue,
|
1 /*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 #include <sys/sysmacros.h>
30 #include <sys/zfs_context.h>
31 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa.h>
33 #include <sys/txg.h>
34 #include <sys/spa_impl.h>
35 #include <sys/vdev_impl.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
40 #include <sys/arc.h>
41 #include <sys/ddt.h>
42 #include <sys/blkptr.h>
43 #include <sys/special.h>
44 #include <sys/blkptr.h>
45 #include <sys/zfeature.h>
46 #include <sys/dkioc_free_util.h>
47 #include <sys/dsl_scan.h>
48
49 #include <sys/metaslab_impl.h>
50 #include <sys/abd.h>
51
52 extern int zfs_txg_timeout;
53
54 /*
55 * ==========================================================================
56 * I/O type descriptions
57 * ==========================================================================
58 */
59 const char *zio_type_name[ZIO_TYPES] = {
60 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
61 "zio_ioctl"
62 };
63
64 boolean_t zio_dva_throttle_enabled = B_TRUE;
65
66 /*
67 * ==========================================================================
68 * I/O kmem caches
69 * ==========================================================================
70 */
71 kmem_cache_t *zio_cache;
72 kmem_cache_t *zio_link_cache;
73 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75
76 #ifdef _KERNEL
77 extern vmem_t *zio_alloc_arena;
78 #endif
79
80 #define BP_SPANB(indblkshift, level) \
81 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
82 #define COMPARE_META_LEVEL 0x80000000ul
83
84 /*
85 * The following actions directly effect the spa's sync-to-convergence logic.
86 * The values below define the sync pass when we start performing the action.
87 * Care should be taken when changing these values as they directly impact
88 * spa_sync() performance. Tuning these values may introduce subtle performance
89 * pathologies and should only be done in the context of performance analysis.
90 * These tunables will eventually be removed and replaced with #defines once
91 * enough analysis has been done to determine optimal values.
92 *
93 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
94 * regular blocks are not deferred.
95 */
96 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
97 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
98 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
99
100 /*
101 * An allocating zio is one that either currently has the DVA allocate
102 * stage set or will have it later in its lifetime.
103 */
104 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
105
106 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
107
108 #ifdef ZFS_DEBUG
109 int zio_buf_debug_limit = 16384;
110 #else
111 int zio_buf_debug_limit = 0;
112 #endif
113
114 /*
115 * Fault insertion for stress testing
116 */
117 int zio_faulty_vdev_enabled = 0;
118 uint64_t zio_faulty_vdev_guid;
119 uint64_t zio_faulty_vdev_delay_us = 1000000; /* 1 second */
120
121 /*
122 * Tunable to allow for debugging SCSI UNMAP/SATA TRIM calls. Disabling
123 * it will prevent ZFS from attempting to issue DKIOCFREE ioctls to the
124 * underlying storage.
125 */
126 boolean_t zfs_trim = B_TRUE;
127 uint64_t zfs_trim_min_ext_sz = 1 << 20; /* 1 MB */
128
129 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
130
131 void
132 zio_init(void)
133 {
134 size_t c;
135 vmem_t *data_alloc_arena = NULL;
136
137 #ifdef _KERNEL
138 data_alloc_arena = zio_alloc_arena;
139 #endif
140 zio_cache = kmem_cache_create("zio_cache",
141 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
142 zio_link_cache = kmem_cache_create("zio_link_cache",
143 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
144
145 /*
146 * For small buffers, we want a cache for each multiple of
147 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
148 * for each quarter-power of 2.
184 * stored with the buffers.
185 */
186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
187 zio_data_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL, data_alloc_arena,
189 cflags | KMC_NOTOUCH);
190 }
191 }
192
193 while (--c != 0) {
194 ASSERT(zio_buf_cache[c] != NULL);
195 if (zio_buf_cache[c - 1] == NULL)
196 zio_buf_cache[c - 1] = zio_buf_cache[c];
197
198 ASSERT(zio_data_buf_cache[c] != NULL);
199 if (zio_data_buf_cache[c - 1] == NULL)
200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
201 }
202
203 zio_inject_init();
204
205 }
206
207 void
208 zio_fini(void)
209 {
210 size_t c;
211 kmem_cache_t *last_cache = NULL;
212 kmem_cache_t *last_data_cache = NULL;
213
214 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
215 if (zio_buf_cache[c] != last_cache) {
216 last_cache = zio_buf_cache[c];
217 kmem_cache_destroy(zio_buf_cache[c]);
218 }
219 zio_buf_cache[c] = NULL;
220
221 if (zio_data_buf_cache[c] != last_data_cache) {
222 last_data_cache = zio_data_buf_cache[c];
223 kmem_cache_destroy(zio_data_buf_cache[c]);
224 }
447 {
448 ASSERT(zl->zl_parent == pio);
449 ASSERT(zl->zl_child == cio);
450
451 mutex_enter(&cio->io_lock);
452 mutex_enter(&pio->io_lock);
453
454 list_remove(&pio->io_child_list, zl);
455 list_remove(&cio->io_parent_list, zl);
456
457 pio->io_child_count--;
458 cio->io_parent_count--;
459
460 mutex_exit(&pio->io_lock);
461 mutex_exit(&cio->io_lock);
462
463 kmem_cache_free(zio_link_cache, zl);
464 }
465
466 static boolean_t
467 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
468 {
469 uint64_t *countp = &zio->io_children[child][wait];
470 boolean_t waiting = B_FALSE;
471
472 mutex_enter(&zio->io_lock);
473 ASSERT(zio->io_stall == NULL);
474 if (*countp != 0) {
475 zio->io_stage >>= 1;
476 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
477 zio->io_stall = countp;
478 waiting = B_TRUE;
479 }
480 mutex_exit(&zio->io_lock);
481
482 return (waiting);
483 }
484
485 static void
486 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
487 {
488 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
489 int *errorp = &pio->io_child_error[zio->io_child_type];
490
491 mutex_enter(&pio->io_lock);
492 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
493 *errorp = zio_worst_error(*errorp, zio->io_error);
494 pio->io_reexecute |= zio->io_reexecute;
495 ASSERT3U(*countp, >, 0);
496
497 (*countp)--;
498
499 if (*countp == 0 && pio->io_stall == countp) {
500 zio_taskq_type_t type =
501 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
619 zio->io_private = private;
620 zio->io_type = type;
621 zio->io_priority = priority;
622 zio->io_vd = vd;
623 zio->io_offset = offset;
624 zio->io_orig_abd = zio->io_abd = data;
625 zio->io_orig_size = zio->io_size = psize;
626 zio->io_lsize = lsize;
627 zio->io_orig_flags = zio->io_flags = flags;
628 zio->io_orig_stage = zio->io_stage = stage;
629 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
630 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
631
632 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
633 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
634
635 if (zb != NULL)
636 zio->io_bookmark = *zb;
637
638 if (pio != NULL) {
639 zio->io_mc = pio->io_mc;
640 if (zio->io_logical == NULL)
641 zio->io_logical = pio->io_logical;
642 if (zio->io_child_type == ZIO_CHILD_GANG)
643 zio->io_gang_leader = pio->io_gang_leader;
644 zio_add_child(pio, zio);
645
646 /* copy the smartcomp setting when creating child zio's */
647 bcopy(&pio->io_smartcomp, &zio->io_smartcomp,
648 sizeof (zio->io_smartcomp));
649 }
650
651 return (zio);
652 }
653
654 static void
655 zio_destroy(zio_t *zio)
656 {
657 metaslab_trace_fini(&zio->io_alloc_list);
658 list_destroy(&zio->io_parent_list);
659 list_destroy(&zio->io_child_list);
660 mutex_destroy(&zio->io_lock);
661 cv_destroy(&zio->io_cv);
662 kmem_cache_free(zio_cache, zio);
663 }
664
665 zio_t *
666 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
667 void *private, enum zio_flag flags)
668 {
669 zio_t *zio;
670
671 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
672 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
673 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
674
675 return (zio);
676 }
677
678 zio_t *
679 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
680 {
681 return (zio_null(NULL, spa, NULL, done, private, flags));
682 }
683
684 void
685 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
686 {
687 /*
688 * SPECIAL-BP has two DVAs, but DVA[0] in this case is a
689 * temporary DVA, and after migration only the DVA[1]
690 * contains valid data. Therefore, we start walking for
691 * these BPs from DVA[1].
692 */
693 int start_dva = BP_IS_SPECIAL(bp) ? 1 : 0;
694
695 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
696 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
697 bp, (longlong_t)BP_GET_TYPE(bp));
698 }
699 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
700 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
701 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
702 bp, (longlong_t)BP_GET_CHECKSUM(bp));
703 }
704 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
705 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
706 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
707 bp, (longlong_t)BP_GET_COMPRESS(bp));
708 }
709 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
710 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
711 bp, (longlong_t)BP_GET_LSIZE(bp));
712 }
713 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
714 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
715 bp, (longlong_t)BP_GET_PSIZE(bp));
716 }
717
718 if (BP_IS_EMBEDDED(bp)) {
719 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
720 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
721 bp, (longlong_t)BPE_GET_ETYPE(bp));
722 }
723 }
724
725 /*
726 * Pool-specific checks.
727 *
728 * Note: it would be nice to verify that the blk_birth and
729 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
730 * allows the birth time of log blocks (and dmu_sync()-ed blocks
731 * that are in the log) to be arbitrarily large.
732 */
733 for (int i = start_dva; i < BP_GET_NDVAS(bp); i++) {
734 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
735 if (vdevid >= spa->spa_root_vdev->vdev_children) {
736 zfs_panic_recover("blkptr at %p DVA %u has invalid "
737 "VDEV %llu",
738 bp, i, (longlong_t)vdevid);
739 continue;
740 }
741 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
742 if (vd == NULL) {
743 zfs_panic_recover("blkptr at %p DVA %u has invalid "
744 "VDEV %llu",
745 bp, i, (longlong_t)vdevid);
746 continue;
747 }
748 if (vd->vdev_ops == &vdev_hole_ops) {
749 zfs_panic_recover("blkptr at %p DVA %u has hole "
750 "VDEV %llu",
751 bp, i, (longlong_t)vdevid);
752 continue;
753 }
754 if (vd->vdev_ops == &vdev_missing_ops) {
755 /*
756 * "missing" vdevs are valid during import, but we
757 * don't have their detailed info (e.g. asize), so
758 * we can't perform any more checks on them.
759 */
760 continue;
761 }
762 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
763 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
764 if (BP_IS_GANG(bp))
765 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
766 if (offset + asize > vd->vdev_asize) {
767 zfs_panic_recover("blkptr at %p DVA %u has invalid "
768 "OFFSET %llu",
769 bp, i, (longlong_t)offset);
770 }
771 }
772 }
773
774 zio_t *
775 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
776 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
777 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
778 {
779 zio_t *zio;
780
781 zfs_blkptr_verify(spa, bp);
782
783 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
784 data, size, size, done, private,
785 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
786 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
787 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
788
789 return (zio);
790 }
791
792 zio_t *
793 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
794 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
795 zio_done_func_t *ready, zio_done_func_t *children_ready,
796 zio_done_func_t *physdone, zio_done_func_t *done,
797 void *private, zio_priority_t priority, enum zio_flag flags,
798 const zbookmark_phys_t *zb,
799 const zio_smartcomp_info_t *smartcomp)
800 {
801 zio_t *zio;
802
803 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
804 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
805 zp->zp_compress >= ZIO_COMPRESS_OFF &&
806 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
807 DMU_OT_IS_VALID(zp->zp_type) &&
808 zp->zp_level < 32 &&
809 zp->zp_copies > 0 &&
810 zp->zp_copies <= spa_max_replication(spa));
811
812 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
813 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
814 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
815 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
816
817 zio->io_ready = ready;
818 zio->io_children_ready = children_ready;
819 zio->io_physdone = physdone;
820 zio->io_prop = *zp;
821 if (smartcomp != NULL)
822 bcopy(smartcomp, &zio->io_smartcomp, sizeof (*smartcomp));
823
824 /*
825 * Data can be NULL if we are going to call zio_write_override() to
826 * provide the already-allocated BP. But we may need the data to
827 * verify a dedup hit (if requested). In this case, don't try to
828 * dedup (just take the already-allocated BP verbatim).
829 */
830 if (data == NULL && zio->io_prop.zp_dedup_verify) {
831 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
832 }
833
834 return (zio);
835 }
836
837 zio_t *
838 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
839 uint64_t size, zio_done_func_t *done, void *private,
840 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
841 {
842 zio_t *zio;
854 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
855 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
856 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
857 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
858
859 /*
860 * We must reset the io_prop to match the values that existed
861 * when the bp was first written by dmu_sync() keeping in mind
862 * that nopwrite and dedup are mutually exclusive.
863 */
864 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
865 zio->io_prop.zp_nopwrite = nopwrite;
866 zio->io_prop.zp_copies = copies;
867 zio->io_bp_override = bp;
868 }
869
870 void
871 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
872 {
873
874 /*
875 * The check for EMBEDDED is a performance optimization. We
876 * process the free here (by ignoring it) rather than
877 * putting it on the list and then processing it in zio_free_sync().
878 */
879 if (BP_IS_EMBEDDED(bp))
880 return;
881 metaslab_check_free(spa, bp);
882
883 /*
884 * Frees that are for the currently-syncing txg, are not going to be
885 * deferred, and which will not need to do a read (i.e. not GANG or
886 * DEDUP), can be processed immediately. Otherwise, put them on the
887 * in-memory list for later processing.
888 */
889 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
890 txg != spa->spa_syncing_txg ||
891 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
892 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
893 } else {
894 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
895 }
896 }
897
898 zio_t *
899 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
900 enum zio_flag flags)
901 {
902 zio_t *zio;
903 enum zio_stage stage = ZIO_FREE_PIPELINE;
904
905 ASSERT(!BP_IS_HOLE(bp));
906 ASSERT(spa_syncing_txg(spa) == txg);
907 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
908
909 if (BP_IS_EMBEDDED(bp))
910 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
911
912 metaslab_check_free(spa, bp);
913 arc_freed(spa, bp);
914 dsl_scan_freed(spa, bp);
915
916 /*
917 * GANG and DEDUP blocks can induce a read (for the gang block header,
918 * or the DDT), so issue them asynchronously so that this thread is
919 * not tied up.
920 */
921 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
922 stage |= ZIO_STAGE_ISSUE_ASYNC;
923
924 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
925 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
926 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
927
928 return (zio);
929 }
930
931 zio_t *
932 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
933 zio_done_func_t *done, void *private, enum zio_flag flags)
934 {
935 zio_t *zio;
936
937 dprintf_bp(bp, "claiming in txg %llu", txg);
938
939 if (BP_IS_EMBEDDED(bp))
940 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
941
942 /*
943 * A claim is an allocation of a specific block. Claims are needed
944 * to support immediate writes in the intent log. The issue is that
945 * immediate writes contain committed data, but in a txg that was
946 * *not* committed. Upon opening the pool after an unclean shutdown,
947 * the intent log claims all blocks that contain immediate write data
948 * so that the SPA knows they're in use.
949 *
950 * All claims *must* be resolved in the first txg -- before the SPA
951 * starts allocating blocks -- so that nothing is allocated twice.
952 * If txg == 0 we just verify that the block is claimable.
953 */
954 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
955 ASSERT(txg == spa_first_txg(spa) || txg == 0);
956 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
957
958 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
959 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
960 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
961 ASSERT0(zio->io_queued_timestamp);
962
963 return (zio);
964 }
965
966 static zio_t *
967 zio_ioctl_with_pipeline(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
968 zio_done_func_t *done, void *private, enum zio_flag flags,
969 enum zio_stage pipeline)
970 {
971 zio_t *zio;
972 int c;
973
974 if (vd->vdev_children == 0) {
975 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
976 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
977 ZIO_STAGE_OPEN, pipeline);
978
979 zio->io_cmd = cmd;
980 } else {
981 zio = zio_null(pio, spa, vd, done, private, flags);
982 /*
983 * DKIOCFREE ioctl's need some special handling on interior
984 * vdevs. If the device provides an ops function to handle
985 * recomputing dkioc_free extents, then we call it.
986 * Otherwise the default behavior applies, which simply fans
987 * out the ioctl to all component vdevs.
988 */
989 if (cmd == DKIOCFREE && vd->vdev_ops->vdev_op_trim != NULL) {
990 vd->vdev_ops->vdev_op_trim(vd, zio, private);
991 } else {
992 for (c = 0; c < vd->vdev_children; c++)
993 zio_nowait(zio_ioctl_with_pipeline(zio,
994 spa, vd->vdev_child[c], cmd, NULL,
995 private, flags, pipeline));
996 }
997 }
998
999 return (zio);
1000 }
1001
1002 zio_t *
1003 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1004 zio_done_func_t *done, void *private, enum zio_flag flags)
1005 {
1006 return (zio_ioctl_with_pipeline(pio, spa, vd, cmd, done,
1007 private, flags, ZIO_IOCTL_PIPELINE));
1008 }
1009
1010 /*
1011 * Callback for when a trim zio has completed. This simply frees the
1012 * dkioc_free_list_t extent list of the DKIOCFREE ioctl.
1013 */
1014 static void
1015 zio_trim_done(zio_t *zio)
1016 {
1017 VERIFY(zio->io_private != NULL);
1018 dfl_free(zio->io_private);
1019 }
1020
1021 static void
1022 zio_trim_check(uint64_t start, uint64_t len, void *msp)
1023 {
1024 metaslab_t *ms = msp;
1025 boolean_t held = MUTEX_HELD(&ms->ms_lock);
1026 if (!held)
1027 mutex_enter(&ms->ms_lock);
1028 ASSERT(ms->ms_trimming_ts != NULL);
1029 ASSERT(range_tree_contains(ms->ms_trimming_ts->ts_tree,
1030 start - VDEV_LABEL_START_SIZE, len));
1031 if (!held)
1032 mutex_exit(&ms->ms_lock);
1033 }
1034
1035 /*
1036 * Takes a bunch of freed extents and tells the underlying vdevs that the
1037 * space associated with these extents can be released.
1038 * This is used by flash storage to pre-erase blocks for rapid reuse later
1039 * and thin-provisioned block storage to reclaim unused blocks.
1040 */
1041 zio_t *
1042 zio_trim(spa_t *spa, vdev_t *vd, struct range_tree *tree,
1043 zio_done_func_t *done, void *private, enum zio_flag flags,
1044 int trim_flags, metaslab_t *msp)
1045 {
1046 dkioc_free_list_t *dfl = NULL;
1047 range_seg_t *rs;
1048 uint64_t rs_idx;
1049 uint64_t num_exts;
1050 uint64_t bytes_issued = 0, bytes_skipped = 0, exts_skipped = 0;
1051 /*
1052 * We need this to invoke the caller's `done' callback with the
1053 * correct io_private (not the dkioc_free_list_t, which is needed
1054 * by the underlying DKIOCFREE ioctl).
1055 */
1056 zio_t *sub_pio = zio_root(spa, done, private, flags);
1057
1058 ASSERT(range_tree_space(tree) != 0);
1059
1060 if (!zfs_trim)
1061 return (sub_pio);
1062
1063 num_exts = avl_numnodes(&tree->rt_root);
1064 dfl = kmem_zalloc(DFL_SZ(num_exts), KM_SLEEP);
1065 dfl->dfl_flags = trim_flags;
1066 dfl->dfl_num_exts = num_exts;
1067 dfl->dfl_offset = VDEV_LABEL_START_SIZE;
1068 if (msp) {
1069 dfl->dfl_ck_func = zio_trim_check;
1070 dfl->dfl_ck_arg = msp;
1071 }
1072
1073 for (rs = avl_first(&tree->rt_root), rs_idx = 0; rs != NULL;
1074 rs = AVL_NEXT(&tree->rt_root, rs)) {
1075 uint64_t len = rs->rs_end - rs->rs_start;
1076
1077 if (len < zfs_trim_min_ext_sz) {
1078 bytes_skipped += len;
1079 exts_skipped++;
1080 continue;
1081 }
1082
1083 dfl->dfl_exts[rs_idx].dfle_start = rs->rs_start;
1084 dfl->dfl_exts[rs_idx].dfle_length = len;
1085
1086 // check we're a multiple of the vdev ashift
1087 ASSERT0(dfl->dfl_exts[rs_idx].dfle_start &
1088 ((1 << vd->vdev_ashift) - 1));
1089 ASSERT0(dfl->dfl_exts[rs_idx].dfle_length &
1090 ((1 << vd->vdev_ashift) - 1));
1091
1092 rs_idx++;
1093 bytes_issued += len;
1094 }
1095
1096 spa_trimstats_update(spa, rs_idx, bytes_issued, exts_skipped,
1097 bytes_skipped);
1098
1099 /* the zfs_trim_min_ext_sz filter may have shortened the list */
1100 if (dfl->dfl_num_exts != rs_idx) {
1101 dkioc_free_list_t *dfl2 = kmem_zalloc(DFL_SZ(rs_idx), KM_SLEEP);
1102 bcopy(dfl, dfl2, DFL_SZ(rs_idx));
1103 dfl2->dfl_num_exts = rs_idx;
1104 dfl_free(dfl);
1105 dfl = dfl2;
1106 }
1107
1108 zio_nowait(zio_ioctl_with_pipeline(sub_pio, spa, vd, DKIOCFREE,
1109 zio_trim_done, dfl, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
1110 ZIO_FLAG_DONT_RETRY, ZIO_TRIM_PIPELINE));
1111 return (sub_pio);
1112 }
1113
1114 zio_t *
1115 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1116 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1117 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1118 {
1119 zio_t *zio;
1120
1121 ASSERT(vd->vdev_children == 0);
1122 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1123 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1124 ASSERT3U(offset + size, <=, vd->vdev_psize);
1125
1126 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1127 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1128 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1129
1130 zio->io_prop.zp_checksum = checksum;
1131
1132 return (zio);
1133 }
1134
1160 abd_t *wbuf = abd_alloc_sametype(data, size);
1161 abd_copy(wbuf, data, size);
1162
1163 zio_push_transform(zio, wbuf, size, size, NULL);
1164 }
1165
1166 return (zio);
1167 }
1168
1169 /*
1170 * Create a child I/O to do some work for us.
1171 */
1172 zio_t *
1173 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1174 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1175 enum zio_flag flags, zio_done_func_t *done, void *private)
1176 {
1177 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1178 zio_t *zio;
1179
1180 ASSERT(vd->vdev_parent ==
1181 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1182
1183 if (type == ZIO_TYPE_READ && bp != NULL) {
1184 /*
1185 * If we have the bp, then the child should perform the
1186 * checksum and the parent need not. This pushes error
1187 * detection as close to the leaves as possible and
1188 * eliminates redundant checksums in the interior nodes.
1189 */
1190 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1191 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1192 }
1193
1194 if (vd->vdev_children == 0)
1195 offset += VDEV_LABEL_START_SIZE;
1196
1197 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1198
1199 /*
1200 * If we've decided to do a repair, the write is not speculative --
1201 * even if the original read was.
1202 */
1203 if (flags & ZIO_FLAG_IO_REPAIR)
1204 flags &= ~ZIO_FLAG_SPECULATIVE;
1205
1206 /*
1207 * If we're creating a child I/O that is not associated with a
1208 * top-level vdev, then the child zio is not an allocating I/O.
1209 * If this is a retried I/O then we ignore it since we will
1210 * have already processed the original allocating I/O.
1211 */
1212 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1213 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1214 metaslab_class_t *mc = pio->io_mc;
1215
1216 ASSERT(mc->mc_alloc_throttle_enabled);
1217 ASSERT(type == ZIO_TYPE_WRITE);
1218 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1219 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1220 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1221 pio->io_child_type == ZIO_CHILD_GANG);
1222
1223 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1224 }
1225
1226 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1227 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1228 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1229 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1230
1231 zio->io_physdone = pio->io_physdone;
1232 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1233 zio->io_logical->io_phys_children++;
1234
1275 */
1276 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1277 if (!BP_IS_RAIDZ(zio->io_bp)) {
1278 /* we are not doing a raw write */
1279 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1280 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1281 }
1282 }
1283
1284 /*
1285 * ==========================================================================
1286 * Prepare to read and write logical blocks
1287 * ==========================================================================
1288 */
1289
1290 static int
1291 zio_read_bp_init(zio_t *zio)
1292 {
1293 blkptr_t *bp = zio->io_bp;
1294
1295 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1296 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1297 !(zio->io_flags & ZIO_FLAG_RAW)) {
1298 uint64_t psize =
1299 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1300 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1301 psize, psize, zio_decompress);
1302 }
1303
1304 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1305 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1306
1307 int psize = BPE_GET_PSIZE(bp);
1308 void *data = abd_borrow_buf(zio->io_abd, psize);
1309 decode_embedded_bp_compressed(bp, data);
1310 abd_return_buf_copy(zio->io_abd, data, psize);
1311 } else {
1312 ASSERT(!BP_IS_EMBEDDED(bp));
1313 }
1314
1315 if (!BP_IS_METADATA(bp))
1316 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1317
1318 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1319 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1320
1321 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1322 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1323
1324 return (ZIO_PIPELINE_CONTINUE);
1325 }
1326
1327 static int
1328 zio_write_bp_init(zio_t *zio)
1329 {
1330 if (!IO_IS_ALLOCATING(zio))
1331 return (ZIO_PIPELINE_CONTINUE);
1332
1333 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1334
1335 if (zio->io_bp_override) {
1383 return (ZIO_PIPELINE_CONTINUE);
1384 }
1385
1386 static int
1387 zio_write_compress(zio_t *zio)
1388 {
1389 spa_t *spa = zio->io_spa;
1390 zio_prop_t *zp = &zio->io_prop;
1391 enum zio_compress compress = zp->zp_compress;
1392 blkptr_t *bp = zio->io_bp;
1393 uint64_t lsize = zio->io_lsize;
1394 uint64_t psize = zio->io_size;
1395 int pass = 1;
1396
1397 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1398
1399 /*
1400 * If our children haven't all reached the ready stage,
1401 * wait for them and then repeat this pipeline stage.
1402 */
1403 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1404 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1405 return (ZIO_PIPELINE_STOP);
1406
1407 if (!IO_IS_ALLOCATING(zio))
1408 return (ZIO_PIPELINE_CONTINUE);
1409
1410 if (zio->io_children_ready != NULL) {
1411 /*
1412 * Now that all our children are ready, run the callback
1413 * associated with this zio in case it wants to modify the
1414 * data to be written.
1415 */
1416 ASSERT3U(zp->zp_level, >, 0);
1417 zio->io_children_ready(zio);
1418 }
1419
1420 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1421 ASSERT(zio->io_bp_override == NULL);
1422
1423 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1424 /*
1425 * We're rewriting an existing block, which means we're
1427 * converge, it must eventually be the case that we don't
1428 * have to allocate new blocks. But compression changes
1429 * the blocksize, which forces a reallocate, and makes
1430 * convergence take longer. Therefore, after the first
1431 * few passes, stop compressing to ensure convergence.
1432 */
1433 pass = spa_sync_pass(spa);
1434
1435 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1436 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1437 ASSERT(!BP_GET_DEDUP(bp));
1438
1439 if (pass >= zfs_sync_pass_dont_compress)
1440 compress = ZIO_COMPRESS_OFF;
1441
1442 /* Make sure someone doesn't change their mind on overwrites */
1443 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1444 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1445 }
1446
1447 DTRACE_PROBE1(zio_compress_ready, zio_t *, zio);
1448 /* If it's a compressed write that is not raw, compress the buffer. */
1449 if (compress != ZIO_COMPRESS_OFF && psize == lsize &&
1450 ZIO_SHOULD_COMPRESS(zio)) {
1451 void *cbuf = zio_buf_alloc(lsize);
1452 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1453 if (psize == 0 || psize == lsize) {
1454 compress = ZIO_COMPRESS_OFF;
1455 zio_buf_free(cbuf, lsize);
1456 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1457 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1458 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1459 encode_embedded_bp_compressed(bp,
1460 cbuf, compress, lsize, psize);
1461 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1462 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1463 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1464 zio_buf_free(cbuf, lsize);
1465 bp->blk_birth = zio->io_txg;
1466 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1467 ASSERT(spa_feature_is_active(spa,
1468 SPA_FEATURE_EMBEDDED_DATA));
1469 if (zio->io_smartcomp.sc_result != NULL) {
1470 zio->io_smartcomp.sc_result(
1471 zio->io_smartcomp.sc_userinfo, zio);
1472 } else {
1473 ASSERT(zio->io_smartcomp.sc_ask == NULL);
1474 }
1475 return (ZIO_PIPELINE_CONTINUE);
1476 } else {
1477 /*
1478 * Round up compressed size up to the ashift
1479 * of the smallest-ashift device, and zero the tail.
1480 * This ensures that the compressed size of the BP
1481 * (and thus compressratio property) are correct,
1482 * in that we charge for the padding used to fill out
1483 * the last sector.
1484 */
1485 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1486 size_t rounded = (size_t)P2ROUNDUP(psize,
1487 1ULL << spa->spa_min_ashift);
1488 if (rounded >= lsize) {
1489 compress = ZIO_COMPRESS_OFF;
1490 zio_buf_free(cbuf, lsize);
1491 psize = lsize;
1492 } else {
1493 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1494 abd_take_ownership_of_buf(cdata, B_TRUE);
1495 abd_zero_off(cdata, psize, rounded - psize);
1496 psize = rounded;
1497 zio_push_transform(zio, cdata,
1498 psize, lsize, NULL);
1499 }
1500 }
1501
1502 if (zio->io_smartcomp.sc_result != NULL) {
1503 zio->io_smartcomp.sc_result(
1504 zio->io_smartcomp.sc_userinfo, zio);
1505 } else {
1506 ASSERT(zio->io_smartcomp.sc_ask == NULL);
1507 }
1508
1509 /*
1510 * We were unable to handle this as an override bp, treat
1511 * it as a regular write I/O.
1512 */
1513 zio->io_bp_override = NULL;
1514 *bp = zio->io_bp_orig;
1515 zio->io_pipeline = zio->io_orig_pipeline;
1516 } else {
1517 ASSERT3U(psize, !=, 0);
1518
1519 /*
1520 * We are here because of:
1521 * - compress == ZIO_COMPRESS_OFF
1522 * - SmartCompression decides don't compress this data
1523 * - this is a RAW-write
1524 *
1525 * In case of RAW-write we should not override "compress"
1526 */
1527 if ((zio->io_flags & ZIO_FLAG_RAW) == 0)
1528 compress = ZIO_COMPRESS_OFF;
1529 }
1530
1531 /*
1532 * The final pass of spa_sync() must be all rewrites, but the first
1533 * few passes offer a trade-off: allocating blocks defers convergence,
1534 * but newly allocated blocks are sequential, so they can be written
1535 * to disk faster. Therefore, we allow the first few passes of
1536 * spa_sync() to allocate new blocks, but force rewrites after that.
1537 * There should only be a handful of blocks after pass 1 in any case.
1538 */
1539 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1540 BP_GET_PSIZE(bp) == psize &&
1541 pass >= zfs_sync_pass_rewrite) {
1542 ASSERT(psize != 0);
1543 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1544 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1545 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1546 } else {
1547 BP_ZERO(bp);
1548 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1549 }
1550
1551 if (psize == 0) {
1552 if (zio->io_bp_orig.blk_birth != 0 &&
1553 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1554 BP_SET_LSIZE(bp, lsize);
1555 BP_SET_TYPE(bp, zp->zp_type);
1556 BP_SET_LEVEL(bp, zp->zp_level);
1557 BP_SET_BIRTH(bp, zio->io_txg, 0);
1558 }
1559 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1560 } else {
1561 if (zp->zp_dedup) {
1562 /* check the best-effort dedup setting */
1563 zio_best_effort_dedup(zio);
1564 }
1565 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1566 BP_SET_LSIZE(bp, lsize);
1567 BP_SET_TYPE(bp, zp->zp_type);
1568 BP_SET_LEVEL(bp, zp->zp_level);
1569 BP_SET_PSIZE(bp, psize);
1570 BP_SET_COMPRESS(bp, compress);
1571 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1572 BP_SET_DEDUP(bp, zp->zp_dedup);
1573 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1574 if (zp->zp_dedup) {
1575 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1576 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1577 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1578 }
1579 if (zp->zp_nopwrite) {
1580 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1581 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1582 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1583 }
1584 }
1585 return (ZIO_PIPELINE_CONTINUE);
1586 }
1587
1588 static int
1589 zio_free_bp_init(zio_t *zio)
1590 {
1591 blkptr_t *bp = zio->io_bp;
1592
1593 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1594 if (BP_GET_DEDUP(bp))
1595 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1596 }
1597
1598 return (ZIO_PIPELINE_CONTINUE);
1599 }
1600
1601 /*
1602 * ==========================================================================
1603 * Execute the I/O pipeline
1604 * ==========================================================================
1605 */
1606
1607 static void
1608 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1609 {
1610 spa_t *spa = zio->io_spa;
1611 zio_type_t t = zio->io_type;
1612 int flags = (cutinline ? TQ_FRONT : 0);
1613
1614 /*
1615 * If we're a config writer or a probe, the normal issue and
1616 * interrupt threads may all be blocked waiting for the config lock.
1617 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1618 */
1619 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1620 t = ZIO_TYPE_NULL;
1621
1622 /*
1623 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1624 */
1625 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1626 t = ZIO_TYPE_NULL;
1627
1628 /*
1629 * If this is a high priority I/O, then use the high priority taskq if
1630 * available.
1631 */
1632 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1633 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1634 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1635 q++;
1636
1637 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1638
1639 /*
1640 * NB: We are assuming that the zio can only be dispatched
1641 * to a single taskq at a time. It would be a grievous error
1642 * to dispatch the zio to another taskq at the same time.
1643 */
1644 ASSERT(zio->io_tqent.tqent_next == NULL);
1645 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1646 flags, &zio->io_tqent);
1647 }
1648
1649 static boolean_t
1650 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1651 {
1652 kthread_t *executor = zio->io_executor;
1653 spa_t *spa = zio->io_spa;
1740 * (5) the I/O is deferred due to vdev-level queueing
1741 * (6) the I/O is handed off to another thread.
1742 *
1743 * In all cases, the pipeline stops whenever there's no CPU work; it never
1744 * burns a thread in cv_wait().
1745 *
1746 * There's no locking on io_stage because there's no legitimate way
1747 * for multiple threads to be attempting to process the same I/O.
1748 */
1749 static zio_pipe_stage_t *zio_pipeline[];
1750
1751 void
1752 zio_execute(zio_t *zio)
1753 {
1754 zio->io_executor = curthread;
1755
1756 ASSERT3U(zio->io_queued_timestamp, >, 0);
1757
1758 while (zio->io_stage < ZIO_STAGE_DONE) {
1759 enum zio_stage pipeline = zio->io_pipeline;
1760 enum zio_stage old_stage = zio->io_stage;
1761 enum zio_stage stage = zio->io_stage;
1762 int rv;
1763
1764 ASSERT(!MUTEX_HELD(&zio->io_lock));
1765 ASSERT(ISP2(stage));
1766 ASSERT(zio->io_stall == NULL);
1767
1768 do {
1769 stage <<= 1;
1770 } while ((stage & pipeline) == 0);
1771
1772 ASSERT(stage <= ZIO_STAGE_DONE);
1773
1774 /*
1775 * If we are in interrupt context and this pipeline stage
1776 * will grab a config lock that is held across I/O,
1777 * or may wait for an I/O that needs an interrupt thread
1778 * to complete, issue async to avoid deadlock.
1779 *
1780 * For VDEV_IO_START, we cut in line so that the io will
1781 * be sent to disk promptly.
1782 */
1783 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1784 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1785 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1786 zio_requeue_io_start_cut_in_line : B_FALSE;
1787 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1788 return;
1789 }
1790
1791 zio->io_stage = stage;
1792 zio->io_pipeline_trace |= zio->io_stage;
1793 rv = zio_pipeline[highbit64(stage) - 1](zio);
1794
1795 if (rv == ZIO_PIPELINE_STOP)
1796 return;
1797
1798 if (rv == ZIO_PIPELINE_RESTART_STAGE) {
1799 zio->io_stage = old_stage;
1800 (void) zio_issue_async(zio);
1801 return;
1802 }
1803
1804 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1805 }
1806 }
1807
1808 /*
1809 * ==========================================================================
1810 * Initiate I/O, either sync or async
1811 * ==========================================================================
1812 */
1813 int
1814 zio_wait(zio_t *zio)
1815 {
1816 int error;
1817
1818 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1819 ASSERT3P(zio->io_executor, ==, NULL);
1820
1821 zio->io_waiter = curthread;
1822 ASSERT0(zio->io_queued_timestamp);
1823 zio->io_queued_timestamp = gethrtime();
2264 static int
2265 zio_gang_assemble(zio_t *zio)
2266 {
2267 blkptr_t *bp = zio->io_bp;
2268
2269 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2270 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2271
2272 zio->io_gang_leader = zio;
2273
2274 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2275
2276 return (ZIO_PIPELINE_CONTINUE);
2277 }
2278
2279 static int
2280 zio_gang_issue(zio_t *zio)
2281 {
2282 blkptr_t *bp = zio->io_bp;
2283
2284 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2285 return (ZIO_PIPELINE_STOP);
2286
2287 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2288 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2289
2290 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2291 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2292 0);
2293 else
2294 zio_gang_tree_free(&zio->io_gang_tree);
2295
2296 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2297
2298 return (ZIO_PIPELINE_CONTINUE);
2299 }
2300
2301 static void
2302 zio_write_gang_member_ready(zio_t *zio)
2303 {
2304 zio_t *pio = zio_unique_parent(zio);
2305 zio_t *gio = zio->io_gang_leader;
2321 mutex_enter(&pio->io_lock);
2322 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2323 ASSERT(DVA_GET_GANG(&pdva[d]));
2324 asize = DVA_GET_ASIZE(&pdva[d]);
2325 asize += DVA_GET_ASIZE(&cdva[d]);
2326 DVA_SET_ASIZE(&pdva[d], asize);
2327 }
2328 mutex_exit(&pio->io_lock);
2329 }
2330
2331 static void
2332 zio_write_gang_done(zio_t *zio)
2333 {
2334 abd_put(zio->io_abd);
2335 }
2336
2337 static int
2338 zio_write_gang_block(zio_t *pio)
2339 {
2340 spa_t *spa = pio->io_spa;
2341 metaslab_class_t *mc = pio->io_mc;
2342 blkptr_t *bp = pio->io_bp;
2343 zio_t *gio = pio->io_gang_leader;
2344 zio_t *zio;
2345 zio_gang_node_t *gn, **gnpp;
2346 zio_gbh_phys_t *gbh;
2347 abd_t *gbh_abd;
2348 uint64_t txg = pio->io_txg;
2349 uint64_t resid = pio->io_size;
2350 uint64_t lsize;
2351 int copies = gio->io_prop.zp_copies;
2352 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2353 zio_prop_t zp;
2354 int error;
2355
2356 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2357 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2358 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2359 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2360
2361 flags |= METASLAB_ASYNC_ALLOC;
2418 * Create and nowait the gang children.
2419 */
2420 for (int g = 0; resid != 0; resid -= lsize, g++) {
2421 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2422 SPA_MINBLOCKSIZE);
2423 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2424
2425 zp.zp_checksum = gio->io_prop.zp_checksum;
2426 zp.zp_compress = ZIO_COMPRESS_OFF;
2427 zp.zp_type = DMU_OT_NONE;
2428 zp.zp_level = 0;
2429 zp.zp_copies = gio->io_prop.zp_copies;
2430 zp.zp_dedup = B_FALSE;
2431 zp.zp_dedup_verify = B_FALSE;
2432 zp.zp_nopwrite = B_FALSE;
2433
2434 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2435 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2436 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2437 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2438 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark,
2439 &pio->io_smartcomp);
2440
2441 cio->io_mc = mc;
2442
2443 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2444 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2445 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2446
2447 /*
2448 * Gang children won't throttle but we should
2449 * account for their work, so reserve an allocation
2450 * slot for them here.
2451 */
2452 VERIFY(metaslab_class_throttle_reserve(mc,
2453 zp.zp_copies, cio, flags));
2454 }
2455 zio_nowait(cio);
2456 }
2457
2458 /*
2459 * Set pio's pipeline to just wait for zio to finish.
2460 */
2461 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2462
2589 abd_alloc_for_io(zio->io_size, B_TRUE),
2590 zio->io_size, zio_ddt_child_read_done, dde,
2591 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2592 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2593 }
2594 return (ZIO_PIPELINE_CONTINUE);
2595 }
2596
2597 zio_nowait(zio_read(zio, zio->io_spa, bp,
2598 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2599 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2600
2601 return (ZIO_PIPELINE_CONTINUE);
2602 }
2603
2604 static int
2605 zio_ddt_read_done(zio_t *zio)
2606 {
2607 blkptr_t *bp = zio->io_bp;
2608
2609 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2610 return (ZIO_PIPELINE_STOP);
2611
2612 ASSERT(BP_GET_DEDUP(bp));
2613 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2614 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2615
2616 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2617 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2618 ddt_entry_t *dde = zio->io_vsd;
2619 if (ddt == NULL) {
2620 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2621 return (ZIO_PIPELINE_CONTINUE);
2622 }
2623 if (dde == NULL) {
2624 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2625 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2626 return (ZIO_PIPELINE_STOP);
2627 }
2628 if (dde->dde_repair_abd != NULL) {
2629 abd_copy(zio->io_abd, dde->dde_repair_abd,
2630 zio->io_size);
2631 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2632 }
2633 ddt_repair_done(ddt, dde);
2634 zio->io_vsd = NULL;
2635 }
2636
2637 ASSERT(zio->io_vsd == NULL);
2638
2639 return (ZIO_PIPELINE_CONTINUE);
2640 }
2641
2642 /* ARGSUSED */
2643 static boolean_t
2644 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2645 {
2646 spa_t *spa = zio->io_spa;
2647 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2648
2649 /* We should never get a raw, override zio */
2650 ASSERT(!(zio->io_bp_override && do_raw));
2651
2652 /*
2653 * Note: we compare the original data, not the transformed data,
2654 * because when zio->io_bp is an override bp, we will not have
2655 * pushed the I/O transforms. That's an important optimization
2656 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2657 */
2658 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2659 zio_t *lio = dde->dde_lead_zio[p];
2660
2661 if (lio != NULL) {
2662 return (lio->io_orig_size != zio->io_orig_size ||
2663 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2664 zio->io_orig_size) != 0);
2665 }
2666 }
2667
2668 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2669 ddt_phys_t *ddp = &dde->dde_phys[p];
2670
2671 if (ddp->ddp_phys_birth != 0) {
2672 arc_buf_t *abuf = NULL;
2673 arc_flags_t aflags = ARC_FLAG_WAIT;
2674 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2675 blkptr_t blk = *zio->io_bp;
2676 int error;
2677
2678 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2679
2680 dde_exit(dde);
2681
2682 /*
2683 * Intuitively, it would make more sense to compare
2684 * io_abd than io_orig_abd in the raw case since you
2685 * don't want to look at any transformations that have
2686 * happened to the data. However, for raw I/Os the
2687 * data will actually be the same in io_abd and
2688 * io_orig_abd, so all we have to do is issue this as
2689 * a raw ARC read.
2690 */
2691 if (do_raw) {
2692 zio_flags |= ZIO_FLAG_RAW;
2693 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2694 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2695 zio->io_size));
2696 ASSERT3P(zio->io_transform_stack, ==, NULL);
2697 }
2698
2699 error = arc_read(NULL, spa, &blk,
2700 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2701 zio_flags, &aflags, &zio->io_bookmark);
2702
2703 if (error == 0) {
2704 if (arc_buf_size(abuf) != zio->io_orig_size ||
2705 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2706 zio->io_orig_size) != 0)
2707 error = SET_ERROR(EEXIST);
2708 arc_buf_destroy(abuf, &abuf);
2709 }
2710
2711 dde_enter(dde);
2712 return (error != 0);
2713 }
2714 }
2715
2716 return (B_FALSE);
2717 }
2718
2719 static void
2720 zio_ddt_child_write_ready(zio_t *zio)
2721 {
2722 int p = zio->io_prop.zp_copies;
2723 ddt_entry_t *dde = zio->io_private;
2724 ddt_phys_t *ddp = &dde->dde_phys[p];
2725 zio_t *pio;
2726
2727 if (zio->io_error)
2728 return;
2729
2730 dde_enter(dde);
2731
2732 ASSERT(dde->dde_lead_zio[p] == zio);
2733
2734 ddt_phys_fill(ddp, zio->io_bp);
2735
2736 zio_link_t *zl = NULL;
2737 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2738 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2739
2740 dde_exit(dde);
2741 }
2742
2743 static void
2744 zio_ddt_child_write_done(zio_t *zio)
2745 {
2746 int p = zio->io_prop.zp_copies;
2747 ddt_entry_t *dde = zio->io_private;
2748 ddt_phys_t *ddp = &dde->dde_phys[p];
2749
2750 dde_enter(dde);
2751
2752 ASSERT(ddp->ddp_refcnt == 0);
2753 ASSERT(dde->dde_lead_zio[p] == zio);
2754 dde->dde_lead_zio[p] = NULL;
2755
2756 if (zio->io_error == 0) {
2757 zio_link_t *zl = NULL;
2758 while (zio_walk_parents(zio, &zl) != NULL)
2759 ddt_phys_addref(ddp);
2760 } else {
2761 ddt_phys_clear(ddp);
2762 }
2763
2764 dde_exit(dde);
2765 }
2766
2767 static void
2768 zio_ddt_ditto_write_done(zio_t *zio)
2769 {
2770 int p = DDT_PHYS_DITTO;
2771 zio_prop_t *zp = &zio->io_prop;
2772 blkptr_t *bp = zio->io_bp;
2773 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2774 ddt_entry_t *dde = zio->io_private;
2775 ddt_phys_t *ddp = &dde->dde_phys[p];
2776 ddt_key_t *ddk = &dde->dde_key;
2777
2778 dde_enter(dde);
2779
2780 ASSERT(ddp->ddp_refcnt == 0);
2781 ASSERT(dde->dde_lead_zio[p] == zio);
2782 dde->dde_lead_zio[p] = NULL;
2783
2784 if (zio->io_error == 0) {
2785 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2786 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2787 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2788 if (ddp->ddp_phys_birth != 0)
2789 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2790 ddt_phys_fill(ddp, bp);
2791 }
2792
2793 dde_exit(dde);
2794 }
2795
2796 static int
2797 zio_ddt_write(zio_t *zio)
2798 {
2799 spa_t *spa = zio->io_spa;
2800 blkptr_t *bp = zio->io_bp;
2801 uint64_t txg = zio->io_txg;
2802 zio_prop_t *zp = &zio->io_prop;
2803 int p = zp->zp_copies;
2804 int ditto_copies;
2805 zio_t *cio = NULL;
2806 zio_t *dio = NULL;
2807 ddt_t *ddt = ddt_select(spa, bp);
2808 ddt_entry_t *dde;
2809 ddt_phys_t *ddp;
2810
2811 ASSERT(BP_GET_DEDUP(bp));
2812 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2813 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2814 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2815
2816 dde = ddt_lookup(ddt, bp, B_TRUE);
2817
2818 /*
2819 * If we're not using special tier, for each new DDE that's not on disk:
2820 * disable dedup if we have exhausted "allowed" DDT L2/ARC space
2821 */
2822 if ((dde->dde_state & DDE_NEW) && !spa->spa_usesc &&
2823 (zfs_ddt_limit_type != DDT_NO_LIMIT || zfs_ddt_byte_ceiling != 0)) {
2824 /* turn off dedup if we need to stop DDT growth */
2825 if (spa_enable_dedup_cap(spa)) {
2826 dde->dde_state |= DDE_DONT_SYNC;
2827
2828 /* disable dedup and use the ordinary write pipeline */
2829 zio_pop_transforms(zio);
2830 zp->zp_dedup = zp->zp_dedup_verify = B_FALSE;
2831 zio->io_stage = ZIO_STAGE_OPEN;
2832 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2833 zio->io_bp_override = NULL;
2834 BP_ZERO(bp);
2835 dde_exit(dde);
2836
2837 return (ZIO_PIPELINE_CONTINUE);
2838 }
2839 }
2840 ASSERT(!(dde->dde_state & DDE_DONT_SYNC));
2841
2842 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2843 /*
2844 * If we're using a weak checksum, upgrade to a strong checksum
2845 * and try again. If we're already using a strong checksum,
2846 * we can't resolve it, so just convert to an ordinary write.
2847 * (And automatically e-mail a paper to Nature?)
2848 */
2849 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2850 ZCHECKSUM_FLAG_DEDUP)) {
2851 zp->zp_checksum = spa_dedup_checksum(spa);
2852 zio_pop_transforms(zio);
2853 zio->io_stage = ZIO_STAGE_OPEN;
2854 BP_ZERO(bp);
2855 } else {
2856 zp->zp_dedup = B_FALSE;
2857 BP_SET_DEDUP(bp, B_FALSE);
2858 }
2859 ASSERT(!BP_GET_DEDUP(bp));
2860 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2861 dde_exit(dde);
2862 return (ZIO_PIPELINE_CONTINUE);
2863 }
2864
2865 ddp = &dde->dde_phys[p];
2866 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2867 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2868
2869 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2870 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2871 zio_prop_t czp = *zp;
2872
2873 czp.zp_copies = ditto_copies;
2874
2875 /*
2876 * If we arrived here with an override bp, we won't have run
2877 * the transform stack, so we won't have the data we need to
2878 * generate a child i/o. So, toss the override bp and restart.
2879 * This is safe, because using the override bp is just an
2880 * optimization; and it's rare, so the cost doesn't matter.
2881 */
2882 if (zio->io_bp_override) {
2883 zio_pop_transforms(zio);
2884 zio->io_stage = ZIO_STAGE_OPEN;
2885 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2886 zio->io_bp_override = NULL;
2887 BP_ZERO(bp);
2888 dde_exit(dde);
2889 return (ZIO_PIPELINE_CONTINUE);
2890 }
2891
2892 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2893 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2894 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2895 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark, NULL);
2896
2897 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2898 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2899 }
2900
2901 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2902 if (ddp->ddp_phys_birth != 0)
2903 ddt_bp_fill(ddp, bp, txg);
2904 if (dde->dde_lead_zio[p] != NULL)
2905 zio_add_child(zio, dde->dde_lead_zio[p]);
2906 else
2907 ddt_phys_addref(ddp);
2908 } else if (zio->io_bp_override) {
2909 ASSERT(bp->blk_birth == txg);
2910 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2911 ddt_phys_fill(ddp, bp);
2912 ddt_phys_addref(ddp);
2913 } else {
2914 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2915 zio->io_orig_size, zio->io_orig_size, zp,
2916 zio_ddt_child_write_ready, NULL, NULL,
2917 zio_ddt_child_write_done, dde, zio->io_priority,
2918 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark, NULL);
2919
2920 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2921 dde->dde_lead_zio[p] = cio;
2922 }
2923
2924 dde_exit(dde);
2925
2926 if (cio)
2927 zio_nowait(cio);
2928 if (dio)
2929 zio_nowait(dio);
2930
2931 return (ZIO_PIPELINE_CONTINUE);
2932 }
2933
2934 ddt_entry_t *freedde; /* for debugging */
2935
2936 static int
2937 zio_ddt_free(zio_t *zio)
2938 {
2939 spa_t *spa = zio->io_spa;
2940 blkptr_t *bp = zio->io_bp;
2941 ddt_t *ddt = ddt_select(spa, bp);
2942 ddt_entry_t *dde;
2943 ddt_phys_t *ddp;
2944
2945 ASSERT(BP_GET_DEDUP(bp));
2946 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2947
2948 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2949 ddp = ddt_phys_select(dde, bp);
2950 if (ddp)
2951 ddt_phys_decref(ddp);
2952 dde_exit(dde);
2953
2954 return (ZIO_PIPELINE_CONTINUE);
2955 }
2956
2957 /*
2958 * ==========================================================================
2959 * Allocate and free blocks
2960 * ==========================================================================
2961 */
2962
2963 static zio_t *
2964 zio_io_to_allocate(metaslab_class_t *mc)
2965 {
2966 zio_t *zio;
2967
2968 ASSERT(MUTEX_HELD(&mc->mc_alloc_lock));
2969
2970 zio = avl_first(&mc->mc_alloc_tree);
2971 if (zio == NULL)
2972 return (NULL);
2973
2974 ASSERT(IO_IS_ALLOCATING(zio));
2975
2976 /*
2977 * Try to place a reservation for this zio. If we're unable to
2978 * reserve then we throttle.
2979 */
2980 if (!metaslab_class_throttle_reserve(mc,
2981 zio->io_prop.zp_copies, zio, 0)) {
2982 return (NULL);
2983 }
2984
2985 avl_remove(&mc->mc_alloc_tree, zio);
2986 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2987
2988 return (zio);
2989 }
2990
2991 static int
2992 zio_dva_throttle(zio_t *zio)
2993 {
2994 spa_t *spa = zio->io_spa;
2995 zio_t *nio;
2996
2997 /* We need to use parent's MetaslabClass */
2998 if (zio->io_mc == NULL) {
2999 zio->io_mc = spa_select_class(spa, zio);
3000 if (zio->io_prop.zp_usewbc)
3001 return (ZIO_PIPELINE_CONTINUE);
3002 }
3003
3004 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3005 !zio->io_mc->mc_alloc_throttle_enabled ||
3006 zio->io_child_type == ZIO_CHILD_GANG ||
3007 zio->io_flags & ZIO_FLAG_NODATA) {
3008 return (ZIO_PIPELINE_CONTINUE);
3009 }
3010
3011 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3012
3013 ASSERT3U(zio->io_queued_timestamp, >, 0);
3014 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3015
3016 mutex_enter(&zio->io_mc->mc_alloc_lock);
3017
3018 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3019 avl_add(&zio->io_mc->mc_alloc_tree, zio);
3020
3021 nio = zio_io_to_allocate(zio->io_mc);
3022 mutex_exit(&zio->io_mc->mc_alloc_lock);
3023
3024 if (nio == zio)
3025 return (ZIO_PIPELINE_CONTINUE);
3026
3027 if (nio != NULL) {
3028 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3029 /*
3030 * We are passing control to a new zio so make sure that
3031 * it is processed by a different thread. We do this to
3032 * avoid stack overflows that can occur when parents are
3033 * throttled and children are making progress. We allow
3034 * it to go to the head of the taskq since it's already
3035 * been waiting.
3036 */
3037 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3038 }
3039 return (ZIO_PIPELINE_STOP);
3040 }
3041
3042 void
3043 zio_allocate_dispatch(metaslab_class_t *mc)
3044 {
3045 zio_t *zio;
3046
3047 mutex_enter(&mc->mc_alloc_lock);
3048 zio = zio_io_to_allocate(mc);
3049 mutex_exit(&mc->mc_alloc_lock);
3050 if (zio == NULL)
3051 return;
3052
3053 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3054 ASSERT0(zio->io_error);
3055 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3056 }
3057
3058 static int
3059 zio_dva_allocate(zio_t *zio)
3060 {
3061 spa_t *spa = zio->io_spa;
3062 metaslab_class_t *mc = zio->io_mc;
3063
3064 blkptr_t *bp = zio->io_bp;
3065 int error;
3066 int flags = 0;
3067
3068 if (zio->io_gang_leader == NULL) {
3069 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3070 zio->io_gang_leader = zio;
3071 }
3072
3073 ASSERT(BP_IS_HOLE(bp));
3074 ASSERT0(BP_GET_NDVAS(bp));
3075 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3076 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3077 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3078
3079 if (zio->io_flags & ZIO_FLAG_NODATA || zio->io_prop.zp_usewbc) {
3080 flags |= METASLAB_DONT_THROTTLE;
3081 }
3082 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3083 flags |= METASLAB_GANG_CHILD;
3084 }
3085 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE &&
3086 zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3087 flags |= METASLAB_ASYNC_ALLOC;
3088 }
3089
3090 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3091 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3092 &zio->io_alloc_list, zio);
3093
3094 #ifdef _KERNEL
3095 DTRACE_PROBE6(zio_dva_allocate,
3096 uint64_t, DVA_GET_VDEV(&bp->blk_dva[0]),
3097 uint64_t, DVA_GET_VDEV(&bp->blk_dva[1]),
3098 uint64_t, BP_GET_LEVEL(bp),
3099 boolean_t, BP_IS_SPECIAL(bp),
3100 boolean_t, BP_IS_METADATA(bp),
3101 int, error);
3102 #endif
3103
3104 if (error != 0) {
3105 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
3106 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3107 error);
3108 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3109 if (zio->io_prop.zp_usewbc) {
3110 zio->io_prop.zp_usewbc = B_FALSE;
3111 zio->io_prop.zp_usesc = B_FALSE;
3112 zio->io_mc = spa_normal_class(spa);
3113 }
3114
3115 return (zio_write_gang_block(zio));
3116 }
3117
3118 zio->io_error = error;
3119 }
3120
3121 return (ZIO_PIPELINE_CONTINUE);
3122 }
3123
3124 static int
3125 zio_dva_free(zio_t *zio)
3126 {
3127 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3128
3129 return (ZIO_PIPELINE_CONTINUE);
3130 }
3131
3132 static int
3133 zio_dva_claim(zio_t *zio)
3134 {
3135 int error;
3136
3137 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3155 if (!BP_IS_HOLE(bp))
3156 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3157
3158 if (gn != NULL) {
3159 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3160 zio_dva_unallocate(zio, gn->gn_child[g],
3161 &gn->gn_gbh->zg_blkptr[g]);
3162 }
3163 }
3164 }
3165
3166 /*
3167 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3168 */
3169 int
3170 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3171 uint64_t size, boolean_t *slog)
3172 {
3173 int error = 1;
3174 zio_alloc_list_t io_alloc_list;
3175 spa_meta_placement_t *mp = &spa->spa_meta_policy;
3176
3177 ASSERT(txg > spa_syncing_txg(spa));
3178
3179 metaslab_trace_init(&io_alloc_list);
3180
3181 /*
3182 * ZIL blocks are always contiguous (i.e. not gang blocks)
3183 * so we set the METASLAB_HINTBP_AVOID flag so that they
3184 * don't "fast gang" when allocating them.
3185 * If the caller indicates that slog is not to be used
3186 * (via use_slog)
3187 * separate allocation class will not indeed be used,
3188 * independently of whether this is log or special
3189 */
3190
3191 if (spa_has_slogs(spa)) {
3192 error = metaslab_alloc(spa, spa_log_class(spa),
3193 size, new_bp, 1, txg, old_bp,
3194 METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3195
3196 DTRACE_PROBE2(zio_alloc_zil_log,
3197 spa_t *, spa, int, error);
3198
3199 if (error == 0)
3200 *slog = TRUE;
3201 }
3202
3203 /*
3204 * use special when failed to allocate from the regular
3205 * slog, but only if allowed and if the special used
3206 * space is below watermarks
3207 */
3208 if (error != 0 && spa_can_special_be_used(spa) &&
3209 mp->spa_sync_to_special != SYNC_TO_SPECIAL_DISABLED) {
3210 error = metaslab_alloc(spa, spa_special_class(spa),
3211 size, new_bp, 1, txg, old_bp,
3212 METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3213
3214 DTRACE_PROBE2(zio_alloc_zil_special,
3215 spa_t *, spa, int, error);
3216
3217 if (error == 0)
3218 *slog = FALSE;
3219 }
3220
3221 if (error != 0) {
3222 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3223 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3224 &io_alloc_list, NULL);
3225
3226 DTRACE_PROBE2(zio_alloc_zil_normal,
3227 spa_t *, spa, int, error);
3228
3229 if (error == 0)
3230 *slog = FALSE;
3231 }
3232
3233 metaslab_trace_fini(&io_alloc_list);
3234
3235 if (error == 0) {
3236 BP_SET_LSIZE(new_bp, size);
3237 BP_SET_PSIZE(new_bp, size);
3238 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3239 BP_SET_CHECKSUM(new_bp,
3240 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3241 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3242 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3243 BP_SET_LEVEL(new_bp, 0);
3244 BP_SET_DEDUP(new_bp, 0);
3245 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3246 } else {
3247 zfs_dbgmsg("%s: zil block allocation failure: "
3248 "size %llu, error %d", spa_name(spa), size, error);
3249 }
3250
3251 return (error);
3252 }
3269 * ==========================================================================
3270 */
3271
3272
3273 /*
3274 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3275 * stops after this stage and will resume upon I/O completion.
3276 * However, there are instances where the vdev layer may need to
3277 * continue the pipeline when an I/O was not issued. Since the I/O
3278 * that was sent to the vdev layer might be different than the one
3279 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3280 * force the underlying vdev layers to call either zio_execute() or
3281 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3282 */
3283 static int
3284 zio_vdev_io_start(zio_t *zio)
3285 {
3286 vdev_t *vd = zio->io_vd;
3287 uint64_t align;
3288 spa_t *spa = zio->io_spa;
3289 zio_type_t type = zio->io_type;
3290 zio->io_vd_timestamp = gethrtime();
3291
3292 ASSERT(zio->io_error == 0);
3293 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3294
3295 if (vd == NULL) {
3296 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3297 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3298
3299 /*
3300 * The mirror_ops handle multiple DVAs in a single BP.
3301 */
3302 vdev_mirror_ops.vdev_op_io_start(zio);
3303 return (ZIO_PIPELINE_STOP);
3304 }
3305
3306 ASSERT3P(zio->io_logical, !=, zio);
3307
3308 align = 1ULL << vd->vdev_top->vdev_ashift;
3309
3310 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3311 P2PHASE(zio->io_size, align) != 0) {
3312 /* Transform logical writes to be a full physical block size. */
3313 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3314 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3315 ASSERT(vd == vd->vdev_top);
3316 if (type == ZIO_TYPE_WRITE) {
3317 abd_copy(abuf, zio->io_abd, zio->io_size);
3318 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3319 }
3320 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3321 }
3322
3323 /*
3324 * If this is not a physical io, make sure that it is properly aligned
3325 * before proceeding.
3326 */
3327 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3328 ASSERT0(P2PHASE(zio->io_offset, align));
3329 ASSERT0(P2PHASE(zio->io_size, align));
3330 } else {
3331 /*
3332 * For physical writes, we allow 512b aligned writes and assume
3333 * the device will perform a read-modify-write as necessary.
3334 */
3335 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3336 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3337 }
3338
3339 VERIFY(type != ZIO_TYPE_WRITE || spa_writeable(spa));
3340
3341 /*
3342 * If this is a repair I/O, and there's no self-healing involved --
3343 * that is, we're just resilvering what we expect to resilver --
3344 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3345 * This prevents spurious resilvering with nested replication.
3346 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3347 * A is out of date, we'll read from C+D, then use the data to
3348 * resilver A+B -- but we don't actually want to resilver B, just A.
3349 * The top-level mirror has no way to know this, so instead we just
3350 * discard unnecessary repairs as we work our way down the vdev tree.
3351 * The same logic applies to any form of nested replication:
3352 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3353 */
3354 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3355 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3356 zio->io_txg != 0 && /* not a delegated i/o */
3357 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3358 ASSERT(type == ZIO_TYPE_WRITE);
3359 zio_vdev_io_bypass(zio);
3360 return (ZIO_PIPELINE_CONTINUE);
3361 }
3362
3363 if (vd->vdev_ops->vdev_op_leaf &&
3364 (type == ZIO_TYPE_READ || type == ZIO_TYPE_WRITE)) {
3365 if (type == ZIO_TYPE_READ && vdev_cache_read(zio))
3366 return (ZIO_PIPELINE_CONTINUE);
3367
3368 if ((zio = vdev_queue_io(zio)) == NULL)
3369 return (ZIO_PIPELINE_STOP);
3370
3371 if (!vdev_accessible(vd, zio)) {
3372 zio->io_error = SET_ERROR(ENXIO);
3373 zio_interrupt(zio);
3374 return (ZIO_PIPELINE_STOP);
3375 }
3376
3377 /*
3378 * Insert a fault simulation delay for a particular vdev.
3379 */
3380 if (zio_faulty_vdev_enabled &&
3381 (zio->io_vd->vdev_guid == zio_faulty_vdev_guid)) {
3382 delay(NSEC_TO_TICK(zio_faulty_vdev_delay_us *
3383 (NANOSEC / MICROSEC)));
3384 }
3385 }
3386
3387 vd->vdev_ops->vdev_op_io_start(zio);
3388 return (ZIO_PIPELINE_STOP);
3389 }
3390
3391 static int
3392 zio_vdev_io_done(zio_t *zio)
3393 {
3394 vdev_t *vd = zio->io_vd;
3395 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3396 boolean_t unexpected_error = B_FALSE;
3397
3398 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3399 return (ZIO_PIPELINE_STOP);
3400
3401 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3402
3403 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3404 vdev_queue_io_done(zio);
3405
3406 if (zio->io_type == ZIO_TYPE_WRITE)
3407 vdev_cache_write(zio);
3408
3409 if (zio_injection_enabled && zio->io_error == 0)
3410 zio->io_error = zio_handle_device_injection(vd,
3411 zio, EIO);
3412
3413 if (zio_injection_enabled && zio->io_error == 0)
3414 zio->io_error = zio_handle_label_injection(zio, EIO);
3415
3416 if (zio->io_error) {
3417 if (!vdev_accessible(vd, zio)) {
3418 zio->io_error = SET_ERROR(ENXIO);
3419 } else {
3420 unexpected_error = B_TRUE;
3421 }
3422 }
3423 }
3424
3425 ops->vdev_op_io_done(zio);
3426
3427 if (unexpected_error)
3428 VERIFY(vdev_probe(vd, zio) == NULL);
3429
3430 /*
3431 * Measure delta between start and end of the I/O in nanoseconds.
3432 * XXX: Handle overflow.
3433 */
3434 zio->io_vd_timestamp = gethrtime() - zio->io_vd_timestamp;
3435
3436 return (ZIO_PIPELINE_CONTINUE);
3437 }
3438
3439 /*
3440 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3441 * disk, and use that to finish the checksum ereport later.
3442 */
3443 static void
3444 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3445 const void *good_buf)
3446 {
3447 /* no processing needed */
3448 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3449 }
3450
3451 /*ARGSUSED*/
3452 void
3453 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3454 {
3455 void *buf = zio_buf_alloc(zio->io_size);
3456
3457 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3458
3459 zcr->zcr_cbinfo = zio->io_size;
3460 zcr->zcr_cbdata = buf;
3461 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3462 zcr->zcr_free = zio_buf_free;
3463 }
3464
3465 static int
3466 zio_vdev_io_assess(zio_t *zio)
3467 {
3468 vdev_t *vd = zio->io_vd;
3469
3470 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3471 return (ZIO_PIPELINE_STOP);
3472
3473 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3474 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3475
3476 if (zio->io_vsd != NULL) {
3477 zio->io_vsd_ops->vsd_free(zio);
3478 zio->io_vsd = NULL;
3479 }
3480
3481 if (zio_injection_enabled && zio->io_error == 0)
3482 zio->io_error = zio_handle_fault_injection(zio, EIO);
3483
3484 /*
3485 * If the I/O failed, determine whether we should attempt to retry it.
3486 *
3487 * On retry, we cut in line in the issue queue, since we don't want
3488 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3489 */
3490 if (zio->io_error && vd == NULL &&
3491 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3666
3667 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3668 if (e2 == zio_error_rank[r2])
3669 break;
3670
3671 return (r1 > r2 ? e1 : e2);
3672 }
3673
3674 /*
3675 * ==========================================================================
3676 * I/O completion
3677 * ==========================================================================
3678 */
3679 static int
3680 zio_ready(zio_t *zio)
3681 {
3682 blkptr_t *bp = zio->io_bp;
3683 zio_t *pio, *pio_next;
3684 zio_link_t *zl = NULL;
3685
3686 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3687 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3688 return (ZIO_PIPELINE_STOP);
3689
3690 if (zio->io_ready) {
3691 ASSERT(IO_IS_ALLOCATING(zio));
3692 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3693 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3694 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3695
3696 zio->io_ready(zio);
3697 }
3698
3699 if (bp != NULL && bp != &zio->io_bp_copy)
3700 zio->io_bp_copy = *bp;
3701
3702 if (zio->io_error != 0) {
3703 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3704
3705 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3706 ASSERT(IO_IS_ALLOCATING(zio));
3707 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3708 /*
3709 * We were unable to allocate anything, unreserve and
3710 * issue the next I/O to allocate.
3711 */
3712 metaslab_class_throttle_unreserve(zio->io_mc,
3713 zio->io_prop.zp_copies, zio);
3714 zio_allocate_dispatch(zio->io_mc);
3715 }
3716 }
3717
3718 mutex_enter(&zio->io_lock);
3719 zio->io_state[ZIO_WAIT_READY] = 1;
3720 pio = zio_walk_parents(zio, &zl);
3721 mutex_exit(&zio->io_lock);
3722
3723 /*
3724 * As we notify zio's parents, new parents could be added.
3725 * New parents go to the head of zio's io_parent_list, however,
3726 * so we will (correctly) not notify them. The remainder of zio's
3727 * io_parent_list, from 'pio_next' onward, cannot change because
3728 * all parents must wait for us to be done before they can be done.
3729 */
3730 for (; pio != NULL; pio = pio_next) {
3731 pio_next = zio_walk_parents(zio, &zl);
3732 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3733 }
3734
3780 if (pio->io_child_type == ZIO_CHILD_GANG) {
3781 /*
3782 * If our parent is a rewrite gang child then our grandparent
3783 * would have been the one that performed the allocation.
3784 */
3785 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3786 pio = zio_unique_parent(pio);
3787 flags |= METASLAB_GANG_CHILD;
3788 }
3789
3790 ASSERT(IO_IS_ALLOCATING(pio));
3791 ASSERT3P(zio, !=, zio->io_logical);
3792 ASSERT(zio->io_logical != NULL);
3793 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3794 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3795
3796 mutex_enter(&pio->io_lock);
3797 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3798 mutex_exit(&pio->io_lock);
3799
3800 metaslab_class_throttle_unreserve(pio->io_mc, 1, pio);
3801
3802 /*
3803 * Call into the pipeline to see if there is more work that
3804 * needs to be done. If there is work to be done it will be
3805 * dispatched to another taskq thread.
3806 */
3807 zio_allocate_dispatch(pio->io_mc);
3808 }
3809
3810 static int
3811 zio_done(zio_t *zio)
3812 {
3813 spa_t *spa = zio->io_spa;
3814 zio_t *lio = zio->io_logical;
3815 blkptr_t *bp = zio->io_bp;
3816 vdev_t *vd = zio->io_vd;
3817 uint64_t psize = zio->io_size;
3818 zio_t *pio, *pio_next;
3819 metaslab_class_t *mc = zio->io_mc;
3820 zio_link_t *zl = NULL;
3821
3822 /*
3823 * If our children haven't all completed,
3824 * wait for them and then repeat this pipeline stage.
3825 */
3826 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3827 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3828 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3829 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3830 return (ZIO_PIPELINE_STOP);
3831
3832 /*
3833 * If the allocation throttle is enabled, then update the accounting.
3834 * We only track child I/Os that are part of an allocating async
3835 * write. We must do this since the allocation is performed
3836 * by the logical I/O but the actual write is done by child I/Os.
3837 */
3838 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3839 zio->io_child_type == ZIO_CHILD_VDEV) {
3840 ASSERT(mc->mc_alloc_throttle_enabled);
3841 zio_dva_throttle_done(zio);
3842 }
3843
3844 /*
3845 * If the allocation throttle is enabled, verify that
3846 * we have decremented the refcounts for every I/O that was throttled.
3847 */
3848 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3849 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3850 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4100 zl = NULL;
4101 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4102 zio_link_t *remove_zl = zl;
4103 pio_next = zio_walk_parents(zio, &zl);
4104 zio_remove_child(pio, zio, remove_zl);
4105 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4106 }
4107
4108 if (zio->io_waiter != NULL) {
4109 mutex_enter(&zio->io_lock);
4110 zio->io_executor = NULL;
4111 cv_broadcast(&zio->io_cv);
4112 mutex_exit(&zio->io_lock);
4113 } else {
4114 zio_destroy(zio);
4115 }
4116
4117 return (ZIO_PIPELINE_STOP);
4118 }
4119
4120 zio_t *
4121 zio_wbc(zio_type_t type, vdev_t *vd, abd_t *data,
4122 uint64_t size, uint64_t offset)
4123 {
4124 zio_t *zio = NULL;
4125
4126 switch (type) {
4127 case ZIO_TYPE_WRITE:
4128 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, data, size,
4129 size, NULL, NULL, ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
4130 ZIO_FLAG_PHYSICAL, vd, offset,
4131 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
4132 break;
4133 case ZIO_TYPE_READ:
4134 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, data, size,
4135 size, NULL, NULL, ZIO_TYPE_READ, ZIO_PRIORITY_ASYNC_READ,
4136 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_PHYSICAL, vd, offset,
4137 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
4138 break;
4139 default:
4140 ASSERT(0);
4141 }
4142
4143 zio->io_prop.zp_checksum = ZIO_CHECKSUM_OFF;
4144
4145 return (zio);
4146 }
4147
4148 /*
4149 * ==========================================================================
4150 * I/O pipeline definition
4151 * ==========================================================================
4152 */
4153 static zio_pipe_stage_t *zio_pipeline[] = {
4154 NULL,
4155 zio_read_bp_init,
4156 zio_write_bp_init,
4157 zio_free_bp_init,
4158 zio_issue_async,
4159 zio_write_compress,
4160 zio_checksum_generate,
4161 zio_nop_write,
4162 zio_ddt_read_start,
4163 zio_ddt_read_done,
4164 zio_ddt_write,
4165 zio_ddt_free,
4166 zio_gang_assemble,
4167 zio_gang_issue,
|