719{
720 zio_t *zio;
721
722 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
723 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
724 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
725
726 return (zio);
727}
728
729void
730zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
731{
732 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
733 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
734 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
735 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
736
737 /*
738 * We must reset the io_prop to match the values that existed
739 * when the bp was first written by dmu_sync() keeping in mind
740 * that nopwrite and dedup are mutually exclusive.
741 */
742 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
743 zio->io_prop.zp_nopwrite = nopwrite;
744 zio->io_prop.zp_copies = copies;
745 zio->io_bp_override = bp;
746}
747
748void
749zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
750{
751
752 /*
753 * The check for EMBEDDED is a performance optimization. We
754 * process the free here (by ignoring it) rather than
755 * putting it on the list and then processing it in zio_free_sync().
756 */
757 if (BP_IS_EMBEDDED(bp))
758 return;
759 metaslab_check_free(spa, bp);
760
761 /*
762 * Frees that are for the currently-syncing txg, are not going to be
763 * deferred, and which will not need to do a read (i.e. not GANG or
764 * DEDUP), can be processed immediately. Otherwise, put them on the
765 * in-memory list for later processing.
766 */
767 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
768 txg != spa->spa_syncing_txg ||
769 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
771 } else {
772 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
773 BP_GET_PSIZE(bp), 0)));
774 }
775}
776
777zio_t *
778zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 uint64_t size, enum zio_flag flags)
780{
781 zio_t *zio;
782 enum zio_stage stage = ZIO_FREE_PIPELINE;
783
784 ASSERT(!BP_IS_HOLE(bp));
785 ASSERT(spa_syncing_txg(spa) == txg);
786 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
787
788 if (BP_IS_EMBEDDED(bp))
789 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
790
791 metaslab_check_free(spa, bp);
792 arc_freed(spa, bp);
793
794 if (zfs_trim_enabled)
795 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
796 ZIO_STAGE_VDEV_IO_ASSESS;
797 /*
798 * GANG and DEDUP blocks can induce a read (for the gang block header,
799 * or the DDT), so issue them asynchronously so that this thread is
800 * not tied up.
801 */
802 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
803 stage |= ZIO_STAGE_ISSUE_ASYNC;
804
805 flags |= ZIO_FLAG_DONT_QUEUE;
806
807 zio = zio_create(pio, spa, txg, bp, NULL, size,
808 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
809 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
810
811 return (zio);
812}
813
814zio_t *
815zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
816 zio_done_func_t *done, void *private, enum zio_flag flags)
817{
818 zio_t *zio;
819
820 dprintf_bp(bp, "claiming in txg %llu", txg);
821
822 if (BP_IS_EMBEDDED(bp))
823 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
824
825 /*
826 * A claim is an allocation of a specific block. Claims are needed
827 * to support immediate writes in the intent log. The issue is that
828 * immediate writes contain committed data, but in a txg that was
829 * *not* committed. Upon opening the pool after an unclean shutdown,
830 * the intent log claims all blocks that contain immediate write data
831 * so that the SPA knows they're in use.
832 *
833 * All claims *must* be resolved in the first txg -- before the SPA
834 * starts allocating blocks -- so that nothing is allocated twice.
835 * If txg == 0 we just verify that the block is claimable.
836 */
837 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
838 ASSERT(txg == spa_first_txg(spa) || txg == 0);
839 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
840
841 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
842 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
843 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
844
845 return (zio);
846}
847
848zio_t *
849zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
850 uint64_t size, zio_done_func_t *done, void *private,
851 zio_priority_t priority, enum zio_flag flags)
852{
853 zio_t *zio;
854 int c;
855
856 if (vd->vdev_children == 0) {
857 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
858 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
859 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
860
861 zio->io_cmd = cmd;
862 } else {
863 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
864
865 for (c = 0; c < vd->vdev_children; c++)
866 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
867 offset, size, done, private, priority, flags));
868 }
869
870 return (zio);
871}
872
873zio_t *
874zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
875 void *data, int checksum, zio_done_func_t *done, void *private,
876 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
877{
878 zio_t *zio;
879
880 ASSERT(vd->vdev_children == 0);
881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
883 ASSERT3U(offset + size, <=, vd->vdev_psize);
884
885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
886 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
887 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
888
889 zio->io_prop.zp_checksum = checksum;
890
891 return (zio);
892}
893
894zio_t *
895zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
896 void *data, int checksum, zio_done_func_t *done, void *private,
897 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
898{
899 zio_t *zio;
900
901 ASSERT(vd->vdev_children == 0);
902 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
903 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
904 ASSERT3U(offset + size, <=, vd->vdev_psize);
905
906 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
907 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
908 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
909
910 zio->io_prop.zp_checksum = checksum;
911
912 if (zio_checksum_table[checksum].ci_eck) {
913 /*
914 * zec checksums are necessarily destructive -- they modify
915 * the end of the write buffer to hold the verifier/checksum.
916 * Therefore, we must make a local copy in case the data is
917 * being written to multiple places in parallel.
918 */
919 void *wbuf = zio_buf_alloc(size);
920 bcopy(data, wbuf, size);
921 zio_push_transform(zio, wbuf, size, size, NULL);
922 }
923
924 return (zio);
925}
926
927/*
928 * Create a child I/O to do some work for us.
929 */
930zio_t *
931zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
932 void *data, uint64_t size, int type, zio_priority_t priority,
933 enum zio_flag flags, zio_done_func_t *done, void *private)
934{
935 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
936 zio_t *zio;
937
938 ASSERT(vd->vdev_parent ==
939 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
940
941 if (type == ZIO_TYPE_READ && bp != NULL) {
942 /*
943 * If we have the bp, then the child should perform the
944 * checksum and the parent need not. This pushes error
945 * detection as close to the leaves as possible and
946 * eliminates redundant checksums in the interior nodes.
947 */
948 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
949 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
950 }
951
952 /* Not all IO types require vdev io done stage e.g. free */
953 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
954 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
955
956 if (vd->vdev_children == 0)
957 offset += VDEV_LABEL_START_SIZE;
958
959 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
960
961 /*
962 * If we've decided to do a repair, the write is not speculative --
963 * even if the original read was.
964 */
965 if (flags & ZIO_FLAG_IO_REPAIR)
966 flags &= ~ZIO_FLAG_SPECULATIVE;
967
968 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
969 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
970 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
971
972 zio->io_physdone = pio->io_physdone;
973 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
974 zio->io_logical->io_phys_children++;
975
976 return (zio);
977}
978
979zio_t *
980zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
981 int type, zio_priority_t priority, enum zio_flag flags,
982 zio_done_func_t *done, void *private)
983{
984 zio_t *zio;
985
986 ASSERT(vd->vdev_ops->vdev_op_leaf);
987
988 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
989 data, size, done, private, type, priority,
990 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
991 vd, offset, NULL,
992 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
993
994 return (zio);
995}
996
997void
998zio_flush(zio_t *zio, vdev_t *vd)
999{
1000 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1001 NULL, NULL, ZIO_PRIORITY_NOW,
1002 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1003}
1004
1005zio_t *
1006zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1007{
1008
1009 ASSERT(vd->vdev_ops->vdev_op_leaf);
1010
1011 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1012 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1013 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1014 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1015}
1016
1017void
1018zio_shrink(zio_t *zio, uint64_t size)
1019{
1020 ASSERT(zio->io_executor == NULL);
1021 ASSERT(zio->io_orig_size == zio->io_size);
1022 ASSERT(size <= zio->io_size);
1023
1024 /*
1025 * We don't shrink for raidz because of problems with the
1026 * reconstruction when reading back less than the block size.
1027 * Note, BP_IS_RAIDZ() assumes no compression.
1028 */
1029 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1030 if (!BP_IS_RAIDZ(zio->io_bp))
1031 zio->io_orig_size = zio->io_size = size;
1032}
1033
1034/*
1035 * ==========================================================================
1036 * Prepare to read and write logical blocks
1037 * ==========================================================================
1038 */
1039
1040static int
1041zio_read_bp_init(zio_t **ziop)
1042{
1043 zio_t *zio = *ziop;
1044 blkptr_t *bp = zio->io_bp;
1045
1046 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1047 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1048 !(zio->io_flags & ZIO_FLAG_RAW)) {
1049 uint64_t psize =
1050 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1051 void *cbuf = zio_buf_alloc(psize);
1052
1053 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1054 }
1055
1056 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1057 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1058 decode_embedded_bp_compressed(bp, zio->io_data);
1059 } else {
1060 ASSERT(!BP_IS_EMBEDDED(bp));
1061 }
1062
1063 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1064 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1065
1066 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1067 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1068
1069 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1070 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1071
1072 return (ZIO_PIPELINE_CONTINUE);
1073}
1074
1075static int
1076zio_write_bp_init(zio_t **ziop)
1077{
1078 zio_t *zio = *ziop;
1079 spa_t *spa = zio->io_spa;
1080 zio_prop_t *zp = &zio->io_prop;
1081 enum zio_compress compress = zp->zp_compress;
1082 blkptr_t *bp = zio->io_bp;
1083 uint64_t lsize = zio->io_size;
1084 uint64_t psize = lsize;
1085 int pass = 1;
1086
1087 /*
1088 * If our children haven't all reached the ready stage,
1089 * wait for them and then repeat this pipeline stage.
1090 */
1091 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1092 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1093 return (ZIO_PIPELINE_STOP);
1094
1095 if (!IO_IS_ALLOCATING(zio))
1096 return (ZIO_PIPELINE_CONTINUE);
1097
1098 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1099
1100 if (zio->io_bp_override) {
1101 ASSERT(bp->blk_birth != zio->io_txg);
1102 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1103
1104 *bp = *zio->io_bp_override;
1105 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1106
1107 if (BP_IS_EMBEDDED(bp))
1108 return (ZIO_PIPELINE_CONTINUE);
1109
1110 /*
1111 * If we've been overridden and nopwrite is set then
1112 * set the flag accordingly to indicate that a nopwrite
1113 * has already occurred.
1114 */
1115 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1116 ASSERT(!zp->zp_dedup);
1117 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1118 return (ZIO_PIPELINE_CONTINUE);
1119 }
1120
1121 ASSERT(!zp->zp_nopwrite);
1122
1123 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1124 return (ZIO_PIPELINE_CONTINUE);
1125
1126 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1127 zp->zp_dedup_verify);
1128
1129 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1130 BP_SET_DEDUP(bp, 1);
1131 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1132 return (ZIO_PIPELINE_CONTINUE);
1133 }
1134 zio->io_bp_override = NULL;
1135 BP_ZERO(bp);
1136 }
1137
1138 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1139 /*
1140 * We're rewriting an existing block, which means we're
1141 * working on behalf of spa_sync(). For spa_sync() to
1142 * converge, it must eventually be the case that we don't
1143 * have to allocate new blocks. But compression changes
1144 * the blocksize, which forces a reallocate, and makes
1145 * convergence take longer. Therefore, after the first
1146 * few passes, stop compressing to ensure convergence.
1147 */
1148 pass = spa_sync_pass(spa);
1149
1150 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1151 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1152 ASSERT(!BP_GET_DEDUP(bp));
1153
1154 if (pass >= zfs_sync_pass_dont_compress)
1155 compress = ZIO_COMPRESS_OFF;
1156
1157 /* Make sure someone doesn't change their mind on overwrites */
1158 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1159 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1160 }
1161
1162 if (compress != ZIO_COMPRESS_OFF) {
1163 metaslab_class_t *mc = spa_normal_class(spa);
1164 void *cbuf = zio_buf_alloc(lsize);
1165 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1166 (size_t)metaslab_class_get_minblocksize(mc));
1167 if (psize == 0 || psize == lsize) {
1168 compress = ZIO_COMPRESS_OFF;
1169 zio_buf_free(cbuf, lsize);
1170 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1171 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1172 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1173 encode_embedded_bp_compressed(bp,
1174 cbuf, compress, lsize, psize);
1175 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1176 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1177 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1178 zio_buf_free(cbuf, lsize);
1179 bp->blk_birth = zio->io_txg;
1180 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1181 ASSERT(spa_feature_is_active(spa,
1182 SPA_FEATURE_EMBEDDED_DATA));
1183 return (ZIO_PIPELINE_CONTINUE);
1184 } else {
1185 /*
1186 * Round up compressed size to MINBLOCKSIZE and
1187 * zero the tail.
1188 */
1189 size_t rounded =
1190 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1191 if (rounded > psize) {
1192 bzero((char *)cbuf + psize, rounded - psize);
1193 psize = rounded;
1194 }
1195 if (psize == lsize) {
1196 compress = ZIO_COMPRESS_OFF;
1197 zio_buf_free(cbuf, lsize);
1198 } else {
1199 zio_push_transform(zio, cbuf,
1200 psize, lsize, NULL);
1201 }
1202 }
1203 }
1204
1205 /*
1206 * The final pass of spa_sync() must be all rewrites, but the first
1207 * few passes offer a trade-off: allocating blocks defers convergence,
1208 * but newly allocated blocks are sequential, so they can be written
1209 * to disk faster. Therefore, we allow the first few passes of
1210 * spa_sync() to allocate new blocks, but force rewrites after that.
1211 * There should only be a handful of blocks after pass 1 in any case.
1212 */
1213 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1214 BP_GET_PSIZE(bp) == psize &&
1215 pass >= zfs_sync_pass_rewrite) {
1216 ASSERT(psize != 0);
1217 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1218 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1219 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1220 } else {
1221 BP_ZERO(bp);
1222 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1223 }
1224
1225 if (psize == 0) {
1226 if (zio->io_bp_orig.blk_birth != 0 &&
1227 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1228 BP_SET_LSIZE(bp, lsize);
1229 BP_SET_TYPE(bp, zp->zp_type);
1230 BP_SET_LEVEL(bp, zp->zp_level);
1231 BP_SET_BIRTH(bp, zio->io_txg, 0);
1232 }
1233 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1234 } else {
1235 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1236 BP_SET_LSIZE(bp, lsize);
1237 BP_SET_TYPE(bp, zp->zp_type);
1238 BP_SET_LEVEL(bp, zp->zp_level);
1239 BP_SET_PSIZE(bp, psize);
1240 BP_SET_COMPRESS(bp, compress);
1241 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1242 BP_SET_DEDUP(bp, zp->zp_dedup);
1243 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1244 if (zp->zp_dedup) {
1245 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1246 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1247 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1248 }
1249 if (zp->zp_nopwrite) {
1250 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1251 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1252 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1253 }
1254 }
1255
1256 return (ZIO_PIPELINE_CONTINUE);
1257}
1258
1259static int
1260zio_free_bp_init(zio_t **ziop)
1261{
1262 zio_t *zio = *ziop;
1263 blkptr_t *bp = zio->io_bp;
1264
1265 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1266 if (BP_GET_DEDUP(bp))
1267 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1268 }
1269
1270 return (ZIO_PIPELINE_CONTINUE);
1271}
1272
1273/*
1274 * ==========================================================================
1275 * Execute the I/O pipeline
1276 * ==========================================================================
1277 */
1278
1279static void
1280zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1281{
1282 spa_t *spa = zio->io_spa;
1283 zio_type_t t = zio->io_type;
1284 int flags = (cutinline ? TQ_FRONT : 0);
1285
1286 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1287
1288 /*
1289 * If we're a config writer or a probe, the normal issue and
1290 * interrupt threads may all be blocked waiting for the config lock.
1291 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1292 */
1293 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1294 t = ZIO_TYPE_NULL;
1295
1296 /*
1297 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1298 */
1299 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1300 t = ZIO_TYPE_NULL;
1301
1302 /*
1303 * If this is a high priority I/O, then use the high priority taskq if
1304 * available.
1305 */
1306 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1307 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1308 q++;
1309
1310 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1311
1312 /*
1313 * NB: We are assuming that the zio can only be dispatched
1314 * to a single taskq at a time. It would be a grievous error
1315 * to dispatch the zio to another taskq at the same time.
1316 */
1317#if defined(illumos) || !defined(_KERNEL)
1318 ASSERT(zio->io_tqent.tqent_next == NULL);
1319#else
1320 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1321#endif
1322 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1323 flags, &zio->io_tqent);
1324}
1325
1326static boolean_t
1327zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1328{
1329 kthread_t *executor = zio->io_executor;
1330 spa_t *spa = zio->io_spa;
1331
1332 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1333 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1334 uint_t i;
1335 for (i = 0; i < tqs->stqs_count; i++) {
1336 if (taskq_member(tqs->stqs_taskq[i], executor))
1337 return (B_TRUE);
1338 }
1339 }
1340
1341 return (B_FALSE);
1342}
1343
1344static int
1345zio_issue_async(zio_t **ziop)
1346{
1347 zio_t *zio = *ziop;
1348
1349 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1350
1351 return (ZIO_PIPELINE_STOP);
1352}
1353
1354void
1355zio_interrupt(zio_t *zio)
1356{
1357 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1358}
1359
1360/*
1361 * Execute the I/O pipeline until one of the following occurs:
1362 *
1363 * (1) the I/O completes
1364 * (2) the pipeline stalls waiting for dependent child I/Os
1365 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1366 * (4) the I/O is delegated by vdev-level caching or aggregation
1367 * (5) the I/O is deferred due to vdev-level queueing
1368 * (6) the I/O is handed off to another thread.
1369 *
1370 * In all cases, the pipeline stops whenever there's no CPU work; it never
1371 * burns a thread in cv_wait().
1372 *
1373 * There's no locking on io_stage because there's no legitimate way
1374 * for multiple threads to be attempting to process the same I/O.
1375 */
1376static zio_pipe_stage_t *zio_pipeline[];
1377
1378void
1379zio_execute(zio_t *zio)
1380{
1381 zio->io_executor = curthread;
1382
1383 while (zio->io_stage < ZIO_STAGE_DONE) {
1384 enum zio_stage pipeline = zio->io_pipeline;
1385 enum zio_stage stage = zio->io_stage;
1386 int rv;
1387
1388 ASSERT(!MUTEX_HELD(&zio->io_lock));
1389 ASSERT(ISP2(stage));
1390 ASSERT(zio->io_stall == NULL);
1391
1392 do {
1393 stage <<= 1;
1394 } while ((stage & pipeline) == 0);
1395
1396 ASSERT(stage <= ZIO_STAGE_DONE);
1397
1398 /*
1399 * If we are in interrupt context and this pipeline stage
1400 * will grab a config lock that is held across I/O,
1401 * or may wait for an I/O that needs an interrupt thread
1402 * to complete, issue async to avoid deadlock.
1403 *
1404 * For VDEV_IO_START, we cut in line so that the io will
1405 * be sent to disk promptly.
1406 */
1407 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1408 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1409 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1410 zio_requeue_io_start_cut_in_line : B_FALSE;
1411 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1412 return;
1413 }
1414
1415 zio->io_stage = stage;
1416 rv = zio_pipeline[highbit64(stage) - 1](&zio);
1417
1418 if (rv == ZIO_PIPELINE_STOP)
1419 return;
1420
1421 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1422 }
1423}
1424
1425/*
1426 * ==========================================================================
1427 * Initiate I/O, either sync or async
1428 * ==========================================================================
1429 */
1430int
1431zio_wait(zio_t *zio)
1432{
1433 int error;
1434
1435 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1436 ASSERT(zio->io_executor == NULL);
1437
1438 zio->io_waiter = curthread;
1439
1440 zio_execute(zio);
1441
1442 mutex_enter(&zio->io_lock);
1443 while (zio->io_executor != NULL)
1444 cv_wait(&zio->io_cv, &zio->io_lock);
1445 mutex_exit(&zio->io_lock);
1446
1447 error = zio->io_error;
1448 zio_destroy(zio);
1449
1450 return (error);
1451}
1452
1453void
1454zio_nowait(zio_t *zio)
1455{
1456 ASSERT(zio->io_executor == NULL);
1457
1458 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1459 zio_unique_parent(zio) == NULL) {
1460 /*
1461 * This is a logical async I/O with no parent to wait for it.
1462 * We add it to the spa_async_root_zio "Godfather" I/O which
1463 * will ensure they complete prior to unloading the pool.
1464 */
1465 spa_t *spa = zio->io_spa;
1466
1467 zio_add_child(spa->spa_async_zio_root, zio);
1468 }
1469
1470 zio_execute(zio);
1471}
1472
1473/*
1474 * ==========================================================================
1475 * Reexecute or suspend/resume failed I/O
1476 * ==========================================================================
1477 */
1478
1479static void
1480zio_reexecute(zio_t *pio)
1481{
1482 zio_t *cio, *cio_next;
1483
1484 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1485 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1486 ASSERT(pio->io_gang_leader == NULL);
1487 ASSERT(pio->io_gang_tree == NULL);
1488
1489 pio->io_flags = pio->io_orig_flags;
1490 pio->io_stage = pio->io_orig_stage;
1491 pio->io_pipeline = pio->io_orig_pipeline;
1492 pio->io_reexecute = 0;
1493 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1494 pio->io_error = 0;
1495 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1496 pio->io_state[w] = 0;
1497 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1498 pio->io_child_error[c] = 0;
1499
1500 if (IO_IS_ALLOCATING(pio))
1501 BP_ZERO(pio->io_bp);
1502
1503 /*
1504 * As we reexecute pio's children, new children could be created.
1505 * New children go to the head of pio's io_child_list, however,
1506 * so we will (correctly) not reexecute them. The key is that
1507 * the remainder of pio's io_child_list, from 'cio_next' onward,
1508 * cannot be affected by any side effects of reexecuting 'cio'.
1509 */
1510 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1511 cio_next = zio_walk_children(pio);
1512 mutex_enter(&pio->io_lock);
1513 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1514 pio->io_children[cio->io_child_type][w]++;
1515 mutex_exit(&pio->io_lock);
1516 zio_reexecute(cio);
1517 }
1518
1519 /*
1520 * Now that all children have been reexecuted, execute the parent.
1521 * We don't reexecute "The Godfather" I/O here as it's the
1522 * responsibility of the caller to wait on him.
1523 */
1524 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1525 zio_execute(pio);
1526}
1527
1528void
1529zio_suspend(spa_t *spa, zio_t *zio)
1530{
1531 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1532 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1533 "failure and the failure mode property for this pool "
1534 "is set to panic.", spa_name(spa));
1535
1536 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1537
1538 mutex_enter(&spa->spa_suspend_lock);
1539
1540 if (spa->spa_suspend_zio_root == NULL)
1541 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1542 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1543 ZIO_FLAG_GODFATHER);
1544
1545 spa->spa_suspended = B_TRUE;
1546
1547 if (zio != NULL) {
1548 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1549 ASSERT(zio != spa->spa_suspend_zio_root);
1550 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1551 ASSERT(zio_unique_parent(zio) == NULL);
1552 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1553 zio_add_child(spa->spa_suspend_zio_root, zio);
1554 }
1555
1556 mutex_exit(&spa->spa_suspend_lock);
1557}
1558
1559int
1560zio_resume(spa_t *spa)
1561{
1562 zio_t *pio;
1563
1564 /*
1565 * Reexecute all previously suspended i/o.
1566 */
1567 mutex_enter(&spa->spa_suspend_lock);
1568 spa->spa_suspended = B_FALSE;
1569 cv_broadcast(&spa->spa_suspend_cv);
1570 pio = spa->spa_suspend_zio_root;
1571 spa->spa_suspend_zio_root = NULL;
1572 mutex_exit(&spa->spa_suspend_lock);
1573
1574 if (pio == NULL)
1575 return (0);
1576
1577 zio_reexecute(pio);
1578 return (zio_wait(pio));
1579}
1580
1581void
1582zio_resume_wait(spa_t *spa)
1583{
1584 mutex_enter(&spa->spa_suspend_lock);
1585 while (spa_suspended(spa))
1586 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1587 mutex_exit(&spa->spa_suspend_lock);
1588}
1589
1590/*
1591 * ==========================================================================
1592 * Gang blocks.
1593 *
1594 * A gang block is a collection of small blocks that looks to the DMU
1595 * like one large block. When zio_dva_allocate() cannot find a block
1596 * of the requested size, due to either severe fragmentation or the pool
1597 * being nearly full, it calls zio_write_gang_block() to construct the
1598 * block from smaller fragments.
1599 *
1600 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1601 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1602 * an indirect block: it's an array of block pointers. It consumes
1603 * only one sector and hence is allocatable regardless of fragmentation.
1604 * The gang header's bps point to its gang members, which hold the data.
1605 *
1606 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1607 * as the verifier to ensure uniqueness of the SHA256 checksum.
1608 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1609 * not the gang header. This ensures that data block signatures (needed for
1610 * deduplication) are independent of how the block is physically stored.
1611 *
1612 * Gang blocks can be nested: a gang member may itself be a gang block.
1613 * Thus every gang block is a tree in which root and all interior nodes are
1614 * gang headers, and the leaves are normal blocks that contain user data.
1615 * The root of the gang tree is called the gang leader.
1616 *
1617 * To perform any operation (read, rewrite, free, claim) on a gang block,
1618 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1619 * in the io_gang_tree field of the original logical i/o by recursively
1620 * reading the gang leader and all gang headers below it. This yields
1621 * an in-core tree containing the contents of every gang header and the
1622 * bps for every constituent of the gang block.
1623 *
1624 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1625 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1626 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1627 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1628 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1629 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1630 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1631 * of the gang header plus zio_checksum_compute() of the data to update the
1632 * gang header's blk_cksum as described above.
1633 *
1634 * The two-phase assemble/issue model solves the problem of partial failure --
1635 * what if you'd freed part of a gang block but then couldn't read the
1636 * gang header for another part? Assembling the entire gang tree first
1637 * ensures that all the necessary gang header I/O has succeeded before
1638 * starting the actual work of free, claim, or write. Once the gang tree
1639 * is assembled, free and claim are in-memory operations that cannot fail.
1640 *
1641 * In the event that a gang write fails, zio_dva_unallocate() walks the
1642 * gang tree to immediately free (i.e. insert back into the space map)
1643 * everything we've allocated. This ensures that we don't get ENOSPC
1644 * errors during repeated suspend/resume cycles due to a flaky device.
1645 *
1646 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1647 * the gang tree, we won't modify the block, so we can safely defer the free
1648 * (knowing that the block is still intact). If we *can* assemble the gang
1649 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1650 * each constituent bp and we can allocate a new block on the next sync pass.
1651 *
1652 * In all cases, the gang tree allows complete recovery from partial failure.
1653 * ==========================================================================
1654 */
1655
1656static zio_t *
1657zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1658{
1659 if (gn != NULL)
1660 return (pio);
1661
1662 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1663 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1664 &pio->io_bookmark));
1665}
1666
1667zio_t *
1668zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1669{
1670 zio_t *zio;
1671
1672 if (gn != NULL) {
1673 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1674 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1675 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1676 /*
1677 * As we rewrite each gang header, the pipeline will compute
1678 * a new gang block header checksum for it; but no one will
1679 * compute a new data checksum, so we do that here. The one
1680 * exception is the gang leader: the pipeline already computed
1681 * its data checksum because that stage precedes gang assembly.
1682 * (Presently, nothing actually uses interior data checksums;
1683 * this is just good hygiene.)
1684 */
1685 if (gn != pio->io_gang_leader->io_gang_tree) {
1686 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1687 data, BP_GET_PSIZE(bp));
1688 }
1689 /*
1690 * If we are here to damage data for testing purposes,
1691 * leave the GBH alone so that we can detect the damage.
1692 */
1693 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1694 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1695 } else {
1696 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1697 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1698 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1699 }
1700
1701 return (zio);
1702}
1703
1704/* ARGSUSED */
1705zio_t *
1706zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1707{
1708 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1709 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1710 ZIO_GANG_CHILD_FLAGS(pio)));
1711}
1712
1713/* ARGSUSED */
1714zio_t *
1715zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1716{
1717 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1718 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1719}
1720
1721static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1722 NULL,
1723 zio_read_gang,
1724 zio_rewrite_gang,
1725 zio_free_gang,
1726 zio_claim_gang,
1727 NULL
1728};
1729
1730static void zio_gang_tree_assemble_done(zio_t *zio);
1731
1732static zio_gang_node_t *
1733zio_gang_node_alloc(zio_gang_node_t **gnpp)
1734{
1735 zio_gang_node_t *gn;
1736
1737 ASSERT(*gnpp == NULL);
1738
1739 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1740 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1741 *gnpp = gn;
1742
1743 return (gn);
1744}
1745
1746static void
1747zio_gang_node_free(zio_gang_node_t **gnpp)
1748{
1749 zio_gang_node_t *gn = *gnpp;
1750
1751 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1752 ASSERT(gn->gn_child[g] == NULL);
1753
1754 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1755 kmem_free(gn, sizeof (*gn));
1756 *gnpp = NULL;
1757}
1758
1759static void
1760zio_gang_tree_free(zio_gang_node_t **gnpp)
1761{
1762 zio_gang_node_t *gn = *gnpp;
1763
1764 if (gn == NULL)
1765 return;
1766
1767 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1768 zio_gang_tree_free(&gn->gn_child[g]);
1769
1770 zio_gang_node_free(gnpp);
1771}
1772
1773static void
1774zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1775{
1776 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1777
1778 ASSERT(gio->io_gang_leader == gio);
1779 ASSERT(BP_IS_GANG(bp));
1780
1781 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1782 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1783 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1784}
1785
1786static void
1787zio_gang_tree_assemble_done(zio_t *zio)
1788{
1789 zio_t *gio = zio->io_gang_leader;
1790 zio_gang_node_t *gn = zio->io_private;
1791 blkptr_t *bp = zio->io_bp;
1792
1793 ASSERT(gio == zio_unique_parent(zio));
1794 ASSERT(zio->io_child_count == 0);
1795
1796 if (zio->io_error)
1797 return;
1798
1799 if (BP_SHOULD_BYTESWAP(bp))
1800 byteswap_uint64_array(zio->io_data, zio->io_size);
1801
1802 ASSERT(zio->io_data == gn->gn_gbh);
1803 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1804 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1805
1806 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1807 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1808 if (!BP_IS_GANG(gbp))
1809 continue;
1810 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1811 }
1812}
1813
1814static void
1815zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1816{
1817 zio_t *gio = pio->io_gang_leader;
1818 zio_t *zio;
1819
1820 ASSERT(BP_IS_GANG(bp) == !!gn);
1821 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1822 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1823
1824 /*
1825 * If you're a gang header, your data is in gn->gn_gbh.
1826 * If you're a gang member, your data is in 'data' and gn == NULL.
1827 */
1828 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1829
1830 if (gn != NULL) {
1831 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1832
1833 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1834 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1835 if (BP_IS_HOLE(gbp))
1836 continue;
1837 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1838 data = (char *)data + BP_GET_PSIZE(gbp);
1839 }
1840 }
1841
1842 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1843 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1844
1845 if (zio != pio)
1846 zio_nowait(zio);
1847}
1848
1849static int
1850zio_gang_assemble(zio_t **ziop)
1851{
1852 zio_t *zio = *ziop;
1853 blkptr_t *bp = zio->io_bp;
1854
1855 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1856 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1857
1858 zio->io_gang_leader = zio;
1859
1860 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1861
1862 return (ZIO_PIPELINE_CONTINUE);
1863}
1864
1865static int
1866zio_gang_issue(zio_t **ziop)
1867{
1868 zio_t *zio = *ziop;
1869 blkptr_t *bp = zio->io_bp;
1870
1871 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1872 return (ZIO_PIPELINE_STOP);
1873
1874 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1875 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1876
1877 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1878 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1879 else
1880 zio_gang_tree_free(&zio->io_gang_tree);
1881
1882 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1883
1884 return (ZIO_PIPELINE_CONTINUE);
1885}
1886
1887static void
1888zio_write_gang_member_ready(zio_t *zio)
1889{
1890 zio_t *pio = zio_unique_parent(zio);
1891 zio_t *gio = zio->io_gang_leader;
1892 dva_t *cdva = zio->io_bp->blk_dva;
1893 dva_t *pdva = pio->io_bp->blk_dva;
1894 uint64_t asize;
1895
1896 if (BP_IS_HOLE(zio->io_bp))
1897 return;
1898
1899 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1900
1901 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1902 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1903 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1904 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1905 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1906
1907 mutex_enter(&pio->io_lock);
1908 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1909 ASSERT(DVA_GET_GANG(&pdva[d]));
1910 asize = DVA_GET_ASIZE(&pdva[d]);
1911 asize += DVA_GET_ASIZE(&cdva[d]);
1912 DVA_SET_ASIZE(&pdva[d], asize);
1913 }
1914 mutex_exit(&pio->io_lock);
1915}
1916
1917static int
1918zio_write_gang_block(zio_t *pio)
1919{
1920 spa_t *spa = pio->io_spa;
1921 blkptr_t *bp = pio->io_bp;
1922 zio_t *gio = pio->io_gang_leader;
1923 zio_t *zio;
1924 zio_gang_node_t *gn, **gnpp;
1925 zio_gbh_phys_t *gbh;
1926 uint64_t txg = pio->io_txg;
1927 uint64_t resid = pio->io_size;
1928 uint64_t lsize;
1929 int copies = gio->io_prop.zp_copies;
1930 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1931 zio_prop_t zp;
1932 int error;
1933
1934 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1935 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1936 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1937 if (error) {
1938 pio->io_error = error;
1939 return (ZIO_PIPELINE_CONTINUE);
1940 }
1941
1942 if (pio == gio) {
1943 gnpp = &gio->io_gang_tree;
1944 } else {
1945 gnpp = pio->io_private;
1946 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1947 }
1948
1949 gn = zio_gang_node_alloc(gnpp);
1950 gbh = gn->gn_gbh;
1951 bzero(gbh, SPA_GANGBLOCKSIZE);
1952
1953 /*
1954 * Create the gang header.
1955 */
1956 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1957 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1958
1959 /*
1960 * Create and nowait the gang children.
1961 */
1962 for (int g = 0; resid != 0; resid -= lsize, g++) {
1963 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1964 SPA_MINBLOCKSIZE);
1965 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1966
1967 zp.zp_checksum = gio->io_prop.zp_checksum;
1968 zp.zp_compress = ZIO_COMPRESS_OFF;
1969 zp.zp_type = DMU_OT_NONE;
1970 zp.zp_level = 0;
1971 zp.zp_copies = gio->io_prop.zp_copies;
1972 zp.zp_dedup = B_FALSE;
1973 zp.zp_dedup_verify = B_FALSE;
1974 zp.zp_nopwrite = B_FALSE;
1975
1976 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1977 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1978 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1979 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1980 &pio->io_bookmark));
1981 }
1982
1983 /*
1984 * Set pio's pipeline to just wait for zio to finish.
1985 */
1986 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1987
1988 zio_nowait(zio);
1989
1990 return (ZIO_PIPELINE_CONTINUE);
1991}
1992
1993/*
1994 * The zio_nop_write stage in the pipeline determines if allocating
1995 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1996 * such as SHA256, we can compare the checksums of the new data and the old
1997 * to determine if allocating a new block is required. The nopwrite
1998 * feature can handle writes in either syncing or open context (i.e. zil
1999 * writes) and as a result is mutually exclusive with dedup.
2000 */
2001static int
2002zio_nop_write(zio_t **ziop)
2003{
2004 zio_t *zio = *ziop;
2005 blkptr_t *bp = zio->io_bp;
2006 blkptr_t *bp_orig = &zio->io_bp_orig;
2007 zio_prop_t *zp = &zio->io_prop;
2008
2009 ASSERT(BP_GET_LEVEL(bp) == 0);
2010 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2011 ASSERT(zp->zp_nopwrite);
2012 ASSERT(!zp->zp_dedup);
2013 ASSERT(zio->io_bp_override == NULL);
2014 ASSERT(IO_IS_ALLOCATING(zio));
2015
2016 /*
2017 * Check to see if the original bp and the new bp have matching
2018 * characteristics (i.e. same checksum, compression algorithms, etc).
2019 * If they don't then just continue with the pipeline which will
2020 * allocate a new bp.
2021 */
2022 if (BP_IS_HOLE(bp_orig) ||
2023 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2024 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2025 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2026 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2027 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2028 return (ZIO_PIPELINE_CONTINUE);
2029
2030 /*
2031 * If the checksums match then reset the pipeline so that we
2032 * avoid allocating a new bp and issuing any I/O.
2033 */
2034 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2035 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2036 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2037 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2038 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2039 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2040 sizeof (uint64_t)) == 0);
2041
2042 *bp = *bp_orig;
2043 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2044 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2045 }
2046
2047 return (ZIO_PIPELINE_CONTINUE);
2048}
2049
2050/*
2051 * ==========================================================================
2052 * Dedup
2053 * ==========================================================================
2054 */
2055static void
2056zio_ddt_child_read_done(zio_t *zio)
2057{
2058 blkptr_t *bp = zio->io_bp;
2059 ddt_entry_t *dde = zio->io_private;
2060 ddt_phys_t *ddp;
2061 zio_t *pio = zio_unique_parent(zio);
2062
2063 mutex_enter(&pio->io_lock);
2064 ddp = ddt_phys_select(dde, bp);
2065 if (zio->io_error == 0)
2066 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2067 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2068 dde->dde_repair_data = zio->io_data;
2069 else
2070 zio_buf_free(zio->io_data, zio->io_size);
2071 mutex_exit(&pio->io_lock);
2072}
2073
2074static int
2075zio_ddt_read_start(zio_t **ziop)
2076{
2077 zio_t *zio = *ziop;
2078 blkptr_t *bp = zio->io_bp;
2079
2080 ASSERT(BP_GET_DEDUP(bp));
2081 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2082 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2083
2084 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2085 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2086 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2087 ddt_phys_t *ddp = dde->dde_phys;
2088 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2089 blkptr_t blk;
2090
2091 ASSERT(zio->io_vsd == NULL);
2092 zio->io_vsd = dde;
2093
2094 if (ddp_self == NULL)
2095 return (ZIO_PIPELINE_CONTINUE);
2096
2097 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2098 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2099 continue;
2100 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2101 &blk);
2102 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2103 zio_buf_alloc(zio->io_size), zio->io_size,
2104 zio_ddt_child_read_done, dde, zio->io_priority,
2105 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2106 &zio->io_bookmark));
2107 }
2108 return (ZIO_PIPELINE_CONTINUE);
2109 }
2110
2111 zio_nowait(zio_read(zio, zio->io_spa, bp,
2112 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2113 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2114
2115 return (ZIO_PIPELINE_CONTINUE);
2116}
2117
2118static int
2119zio_ddt_read_done(zio_t **ziop)
2120{
2121 zio_t *zio = *ziop;
2122 blkptr_t *bp = zio->io_bp;
2123
2124 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2125 return (ZIO_PIPELINE_STOP);
2126
2127 ASSERT(BP_GET_DEDUP(bp));
2128 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2129 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2130
2131 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2132 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2133 ddt_entry_t *dde = zio->io_vsd;
2134 if (ddt == NULL) {
2135 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2136 return (ZIO_PIPELINE_CONTINUE);
2137 }
2138 if (dde == NULL) {
2139 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2140 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2141 return (ZIO_PIPELINE_STOP);
2142 }
2143 if (dde->dde_repair_data != NULL) {
2144 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2145 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2146 }
2147 ddt_repair_done(ddt, dde);
2148 zio->io_vsd = NULL;
2149 }
2150
2151 ASSERT(zio->io_vsd == NULL);
2152
2153 return (ZIO_PIPELINE_CONTINUE);
2154}
2155
2156static boolean_t
2157zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2158{
2159 spa_t *spa = zio->io_spa;
2160
2161 /*
2162 * Note: we compare the original data, not the transformed data,
2163 * because when zio->io_bp is an override bp, we will not have
2164 * pushed the I/O transforms. That's an important optimization
2165 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2166 */
2167 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2168 zio_t *lio = dde->dde_lead_zio[p];
2169
2170 if (lio != NULL) {
2171 return (lio->io_orig_size != zio->io_orig_size ||
2172 bcmp(zio->io_orig_data, lio->io_orig_data,
2173 zio->io_orig_size) != 0);
2174 }
2175 }
2176
2177 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2178 ddt_phys_t *ddp = &dde->dde_phys[p];
2179
2180 if (ddp->ddp_phys_birth != 0) {
2181 arc_buf_t *abuf = NULL;
2182 uint32_t aflags = ARC_WAIT;
2183 blkptr_t blk = *zio->io_bp;
2184 int error;
2185
2186 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2187
2188 ddt_exit(ddt);
2189
2190 error = arc_read(NULL, spa, &blk,
2191 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2192 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2193 &aflags, &zio->io_bookmark);
2194
2195 if (error == 0) {
2196 if (arc_buf_size(abuf) != zio->io_orig_size ||
2197 bcmp(abuf->b_data, zio->io_orig_data,
2198 zio->io_orig_size) != 0)
2199 error = SET_ERROR(EEXIST);
2200 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2201 }
2202
2203 ddt_enter(ddt);
2204 return (error != 0);
2205 }
2206 }
2207
2208 return (B_FALSE);
2209}
2210
2211static void
2212zio_ddt_child_write_ready(zio_t *zio)
2213{
2214 int p = zio->io_prop.zp_copies;
2215 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2216 ddt_entry_t *dde = zio->io_private;
2217 ddt_phys_t *ddp = &dde->dde_phys[p];
2218 zio_t *pio;
2219
2220 if (zio->io_error)
2221 return;
2222
2223 ddt_enter(ddt);
2224
2225 ASSERT(dde->dde_lead_zio[p] == zio);
2226
2227 ddt_phys_fill(ddp, zio->io_bp);
2228
2229 while ((pio = zio_walk_parents(zio)) != NULL)
2230 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2231
2232 ddt_exit(ddt);
2233}
2234
2235static void
2236zio_ddt_child_write_done(zio_t *zio)
2237{
2238 int p = zio->io_prop.zp_copies;
2239 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2240 ddt_entry_t *dde = zio->io_private;
2241 ddt_phys_t *ddp = &dde->dde_phys[p];
2242
2243 ddt_enter(ddt);
2244
2245 ASSERT(ddp->ddp_refcnt == 0);
2246 ASSERT(dde->dde_lead_zio[p] == zio);
2247 dde->dde_lead_zio[p] = NULL;
2248
2249 if (zio->io_error == 0) {
2250 while (zio_walk_parents(zio) != NULL)
2251 ddt_phys_addref(ddp);
2252 } else {
2253 ddt_phys_clear(ddp);
2254 }
2255
2256 ddt_exit(ddt);
2257}
2258
2259static void
2260zio_ddt_ditto_write_done(zio_t *zio)
2261{
2262 int p = DDT_PHYS_DITTO;
2263 zio_prop_t *zp = &zio->io_prop;
2264 blkptr_t *bp = zio->io_bp;
2265 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2266 ddt_entry_t *dde = zio->io_private;
2267 ddt_phys_t *ddp = &dde->dde_phys[p];
2268 ddt_key_t *ddk = &dde->dde_key;
2269
2270 ddt_enter(ddt);
2271
2272 ASSERT(ddp->ddp_refcnt == 0);
2273 ASSERT(dde->dde_lead_zio[p] == zio);
2274 dde->dde_lead_zio[p] = NULL;
2275
2276 if (zio->io_error == 0) {
2277 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2278 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2279 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2280 if (ddp->ddp_phys_birth != 0)
2281 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2282 ddt_phys_fill(ddp, bp);
2283 }
2284
2285 ddt_exit(ddt);
2286}
2287
2288static int
2289zio_ddt_write(zio_t **ziop)
2290{
2291 zio_t *zio = *ziop;
2292 spa_t *spa = zio->io_spa;
2293 blkptr_t *bp = zio->io_bp;
2294 uint64_t txg = zio->io_txg;
2295 zio_prop_t *zp = &zio->io_prop;
2296 int p = zp->zp_copies;
2297 int ditto_copies;
2298 zio_t *cio = NULL;
2299 zio_t *dio = NULL;
2300 ddt_t *ddt = ddt_select(spa, bp);
2301 ddt_entry_t *dde;
2302 ddt_phys_t *ddp;
2303
2304 ASSERT(BP_GET_DEDUP(bp));
2305 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2306 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2307
2308 ddt_enter(ddt);
2309 dde = ddt_lookup(ddt, bp, B_TRUE);
2310 ddp = &dde->dde_phys[p];
2311
2312 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2313 /*
2314 * If we're using a weak checksum, upgrade to a strong checksum
2315 * and try again. If we're already using a strong checksum,
2316 * we can't resolve it, so just convert to an ordinary write.
2317 * (And automatically e-mail a paper to Nature?)
2318 */
2319 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2320 zp->zp_checksum = spa_dedup_checksum(spa);
2321 zio_pop_transforms(zio);
2322 zio->io_stage = ZIO_STAGE_OPEN;
2323 BP_ZERO(bp);
2324 } else {
2325 zp->zp_dedup = B_FALSE;
2326 }
2327 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2328 ddt_exit(ddt);
2329 return (ZIO_PIPELINE_CONTINUE);
2330 }
2331
2332 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2333 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2334
2335 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2336 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2337 zio_prop_t czp = *zp;
2338
2339 czp.zp_copies = ditto_copies;
2340
2341 /*
2342 * If we arrived here with an override bp, we won't have run
2343 * the transform stack, so we won't have the data we need to
2344 * generate a child i/o. So, toss the override bp and restart.
2345 * This is safe, because using the override bp is just an
2346 * optimization; and it's rare, so the cost doesn't matter.
2347 */
2348 if (zio->io_bp_override) {
2349 zio_pop_transforms(zio);
2350 zio->io_stage = ZIO_STAGE_OPEN;
2351 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2352 zio->io_bp_override = NULL;
2353 BP_ZERO(bp);
2354 ddt_exit(ddt);
2355 return (ZIO_PIPELINE_CONTINUE);
2356 }
2357
2358 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2359 zio->io_orig_size, &czp, NULL, NULL,
2360 zio_ddt_ditto_write_done, dde, zio->io_priority,
2361 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2362
2363 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2364 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2365 }
2366
2367 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2368 if (ddp->ddp_phys_birth != 0)
2369 ddt_bp_fill(ddp, bp, txg);
2370 if (dde->dde_lead_zio[p] != NULL)
2371 zio_add_child(zio, dde->dde_lead_zio[p]);
2372 else
2373 ddt_phys_addref(ddp);
2374 } else if (zio->io_bp_override) {
2375 ASSERT(bp->blk_birth == txg);
2376 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2377 ddt_phys_fill(ddp, bp);
2378 ddt_phys_addref(ddp);
2379 } else {
2380 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2381 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2382 zio_ddt_child_write_done, dde, zio->io_priority,
2383 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2384
2385 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2386 dde->dde_lead_zio[p] = cio;
2387 }
2388
2389 ddt_exit(ddt);
2390
2391 if (cio)
2392 zio_nowait(cio);
2393 if (dio)
2394 zio_nowait(dio);
2395
2396 return (ZIO_PIPELINE_CONTINUE);
2397}
2398
2399ddt_entry_t *freedde; /* for debugging */
2400
2401static int
2402zio_ddt_free(zio_t **ziop)
2403{
2404 zio_t *zio = *ziop;
2405 spa_t *spa = zio->io_spa;
2406 blkptr_t *bp = zio->io_bp;
2407 ddt_t *ddt = ddt_select(spa, bp);
2408 ddt_entry_t *dde;
2409 ddt_phys_t *ddp;
2410
2411 ASSERT(BP_GET_DEDUP(bp));
2412 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2413
2414 ddt_enter(ddt);
2415 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2416 ddp = ddt_phys_select(dde, bp);
2417 ddt_phys_decref(ddp);
2418 ddt_exit(ddt);
2419
2420 return (ZIO_PIPELINE_CONTINUE);
2421}
2422
2423/*
2424 * ==========================================================================
2425 * Allocate and free blocks
2426 * ==========================================================================
2427 */
2428static int
2429zio_dva_allocate(zio_t **ziop)
2430{
2431 zio_t *zio = *ziop;
2432 spa_t *spa = zio->io_spa;
2433 metaslab_class_t *mc = spa_normal_class(spa);
2434 blkptr_t *bp = zio->io_bp;
2435 int error;
2436 int flags = 0;
2437
2438 if (zio->io_gang_leader == NULL) {
2439 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2440 zio->io_gang_leader = zio;
2441 }
2442
2443 ASSERT(BP_IS_HOLE(bp));
2444 ASSERT0(BP_GET_NDVAS(bp));
2445 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2446 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2447 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2448
2449 /*
2450 * The dump device does not support gang blocks so allocation on
2451 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2452 * the "fast" gang feature.
2453 */
2454 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2455 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2456 METASLAB_GANG_CHILD : 0;
2457 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2458 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2459
2460 if (error) {
2461 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2462 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2463 error);
2464 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2465 return (zio_write_gang_block(zio));
2466 zio->io_error = error;
2467 }
2468
2469 return (ZIO_PIPELINE_CONTINUE);
2470}
2471
2472static int
2473zio_dva_free(zio_t **ziop)
2474{
2475 zio_t *zio = *ziop;
2476
2477 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2478
2479 return (ZIO_PIPELINE_CONTINUE);
2480}
2481
2482static int
2483zio_dva_claim(zio_t **ziop)
2484{
2485 zio_t *zio = *ziop;
2486 int error;
2487
2488 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2489 if (error)
2490 zio->io_error = error;
2491
2492 return (ZIO_PIPELINE_CONTINUE);
2493}
2494
2495/*
2496 * Undo an allocation. This is used by zio_done() when an I/O fails
2497 * and we want to give back the block we just allocated.
2498 * This handles both normal blocks and gang blocks.
2499 */
2500static void
2501zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2502{
2503 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2504 ASSERT(zio->io_bp_override == NULL);
2505
2506 if (!BP_IS_HOLE(bp))
2507 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2508
2509 if (gn != NULL) {
2510 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2511 zio_dva_unallocate(zio, gn->gn_child[g],
2512 &gn->gn_gbh->zg_blkptr[g]);
2513 }
2514 }
2515}
2516
2517/*
2518 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2519 */
2520int
2521zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2522 uint64_t size, boolean_t use_slog)
2523{
2524 int error = 1;
2525
2526 ASSERT(txg > spa_syncing_txg(spa));
2527
2528 /*
2529 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2530 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2531 * when allocating them.
2532 */
2533 if (use_slog) {
2534 error = metaslab_alloc(spa, spa_log_class(spa), size,
2535 new_bp, 1, txg, old_bp,
2536 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2537 }
2538
2539 if (error) {
2540 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2541 new_bp, 1, txg, old_bp,
2542 METASLAB_HINTBP_AVOID);
2543 }
2544
2545 if (error == 0) {
2546 BP_SET_LSIZE(new_bp, size);
2547 BP_SET_PSIZE(new_bp, size);
2548 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2549 BP_SET_CHECKSUM(new_bp,
2550 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2551 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2552 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2553 BP_SET_LEVEL(new_bp, 0);
2554 BP_SET_DEDUP(new_bp, 0);
2555 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2556 }
2557
2558 return (error);
2559}
2560
2561/*
2562 * Free an intent log block.
2563 */
2564void
2565zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2566{
2567 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2568 ASSERT(!BP_IS_GANG(bp));
2569
2570 zio_free(spa, txg, bp);
2571}
2572
2573/*
2574 * ==========================================================================
2575 * Read, write and delete to physical devices
2576 * ==========================================================================
2577 */
2578static int
2579zio_vdev_io_start(zio_t **ziop)
2580{
2581 zio_t *zio = *ziop;
2582 vdev_t *vd = zio->io_vd;
2583 uint64_t align;
2584 spa_t *spa = zio->io_spa;
2585
2586 ASSERT(zio->io_error == 0);
2587 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2588
2589 if (vd == NULL) {
2590 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2591 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2592
2593 /*
2594 * The mirror_ops handle multiple DVAs in a single BP.
2595 */
2596 return (vdev_mirror_ops.vdev_op_io_start(zio));
2597 }
2598
2599 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2600 zio->io_priority == ZIO_PRIORITY_NOW) {
2601 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2602 return (ZIO_PIPELINE_CONTINUE);
2603 }
2604
2605 /*
2606 * We keep track of time-sensitive I/Os so that the scan thread
2607 * can quickly react to certain workloads. In particular, we care
2608 * about non-scrubbing, top-level reads and writes with the following
2609 * characteristics:
2610 * - synchronous writes of user data to non-slog devices
2611 * - any reads of user data
2612 * When these conditions are met, adjust the timestamp of spa_last_io
2613 * which allows the scan thread to adjust its workload accordingly.
2614 */
2615 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2616 vd == vd->vdev_top && !vd->vdev_islog &&
2617 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2618 zio->io_txg != spa_syncing_txg(spa)) {
2619 uint64_t old = spa->spa_last_io;
2620 uint64_t new = ddi_get_lbolt64();
2621 if (old != new)
2622 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2623 }
2624
2625 align = 1ULL << vd->vdev_top->vdev_ashift;
2626
2627 if (P2PHASE(zio->io_size, align) != 0) {
2628 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2629 char *abuf = NULL;
2630 if (zio->io_type == ZIO_TYPE_READ ||
2631 zio->io_type == ZIO_TYPE_WRITE)
2632 abuf = zio_buf_alloc(asize);
2633 ASSERT(vd == vd->vdev_top);
2634 if (zio->io_type == ZIO_TYPE_WRITE) {
2635 bcopy(zio->io_data, abuf, zio->io_size);
2636 bzero(abuf + zio->io_size, asize - zio->io_size);
2637 }
2638 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2639 zio_subblock);
2640 }
2641
2642 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2643 ASSERT(P2PHASE(zio->io_size, align) == 0);
2644 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2645
2646 /*
2647 * If this is a repair I/O, and there's no self-healing involved --
2648 * that is, we're just resilvering what we expect to resilver --
2649 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2650 * This prevents spurious resilvering with nested replication.
2651 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2652 * A is out of date, we'll read from C+D, then use the data to
2653 * resilver A+B -- but we don't actually want to resilver B, just A.
2654 * The top-level mirror has no way to know this, so instead we just
2655 * discard unnecessary repairs as we work our way down the vdev tree.
2656 * The same logic applies to any form of nested replication:
2657 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2658 */
2659 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2660 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2661 zio->io_txg != 0 && /* not a delegated i/o */
2662 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2663 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2664 zio_vdev_io_bypass(zio);
2665 return (ZIO_PIPELINE_CONTINUE);
2666 }
2667
2668 if (vd->vdev_ops->vdev_op_leaf) {
2669 switch (zio->io_type) {
2670 case ZIO_TYPE_READ:
2671 if (vdev_cache_read(zio))
2672 return (ZIO_PIPELINE_CONTINUE);
2673 /* FALLTHROUGH */
2674 case ZIO_TYPE_WRITE:
2675 case ZIO_TYPE_FREE:
2676 if ((zio = vdev_queue_io(zio)) == NULL)
2677 return (ZIO_PIPELINE_STOP);
2678 *ziop = zio;
2679
2680 if (!vdev_accessible(vd, zio)) {
2681 zio->io_error = SET_ERROR(ENXIO);
2682 zio_interrupt(zio);
2683 return (ZIO_PIPELINE_STOP);
2684 }
2685 break;
2686 }
2687 /*
2688 * Note that we ignore repair writes for TRIM because they can
2689 * conflict with normal writes. This isn't an issue because, by
2690 * definition, we only repair blocks that aren't freed.
2691 */
2692 if (zio->io_type == ZIO_TYPE_WRITE &&
2693 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2694 !trim_map_write_start(zio))
2695 return (ZIO_PIPELINE_STOP);
2696 }
2697
2698 return (vd->vdev_ops->vdev_op_io_start(zio));
2699}
2700
2701static int
2702zio_vdev_io_done(zio_t **ziop)
2703{
2704 zio_t *zio = *ziop;
2705 vdev_t *vd = zio->io_vd;
2706 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2707 boolean_t unexpected_error = B_FALSE;
2708
2709 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2710 return (ZIO_PIPELINE_STOP);
2711
2712 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2713 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2714
2715 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2716 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2717 zio->io_type == ZIO_TYPE_FREE)) {
2718
2719 if (zio->io_type == ZIO_TYPE_WRITE &&
2720 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2721 trim_map_write_done(zio);
2722
2723 vdev_queue_io_done(zio);
2724
2725 if (zio->io_type == ZIO_TYPE_WRITE)
2726 vdev_cache_write(zio);
2727
2728 if (zio_injection_enabled && zio->io_error == 0)
2729 zio->io_error = zio_handle_device_injection(vd,
2730 zio, EIO);
2731
2732 if (zio_injection_enabled && zio->io_error == 0)
2733 zio->io_error = zio_handle_label_injection(zio, EIO);
2734
2735 if (zio->io_error) {
2736 if (zio->io_error == ENOTSUP &&
2737 zio->io_type == ZIO_TYPE_FREE) {
2738 /* Not all devices support TRIM. */
2739 } else if (!vdev_accessible(vd, zio)) {
2740 zio->io_error = SET_ERROR(ENXIO);
2741 } else {
2742 unexpected_error = B_TRUE;
2743 }
2744 }
2745 }
2746
2747 ops->vdev_op_io_done(zio);
2748
2749 if (unexpected_error)
2750 VERIFY(vdev_probe(vd, zio) == NULL);
2751
2752 return (ZIO_PIPELINE_CONTINUE);
2753}
2754
2755/*
2756 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2757 * disk, and use that to finish the checksum ereport later.
2758 */
2759static void
2760zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2761 const void *good_buf)
2762{
2763 /* no processing needed */
2764 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2765}
2766
2767/*ARGSUSED*/
2768void
2769zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2770{
2771 void *buf = zio_buf_alloc(zio->io_size);
2772
2773 bcopy(zio->io_data, buf, zio->io_size);
2774
2775 zcr->zcr_cbinfo = zio->io_size;
2776 zcr->zcr_cbdata = buf;
2777 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2778 zcr->zcr_free = zio_buf_free;
2779}
2780
2781static int
2782zio_vdev_io_assess(zio_t **ziop)
2783{
2784 zio_t *zio = *ziop;
2785 vdev_t *vd = zio->io_vd;
2786
2787 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2788 return (ZIO_PIPELINE_STOP);
2789
2790 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2791 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2792
2793 if (zio->io_vsd != NULL) {
2794 zio->io_vsd_ops->vsd_free(zio);
2795 zio->io_vsd = NULL;
2796 }
2797
2798 if (zio_injection_enabled && zio->io_error == 0)
2799 zio->io_error = zio_handle_fault_injection(zio, EIO);
2800
2801 if (zio->io_type == ZIO_TYPE_FREE &&
2802 zio->io_priority != ZIO_PRIORITY_NOW) {
2803 switch (zio->io_error) {
2804 case 0:
2805 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2806 ZIO_TRIM_STAT_BUMP(success);
2807 break;
2808 case EOPNOTSUPP:
2809 ZIO_TRIM_STAT_BUMP(unsupported);
2810 break;
2811 default:
2812 ZIO_TRIM_STAT_BUMP(failed);
2813 break;
2814 }
2815 }
2816
2817 /*
2818 * If the I/O failed, determine whether we should attempt to retry it.
2819 *
2820 * On retry, we cut in line in the issue queue, since we don't want
2821 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2822 */
2823 if (zio->io_error && vd == NULL &&
2824 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2825 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2826 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2827 zio->io_error = 0;
2828 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2829 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2830 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2831 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2832 zio_requeue_io_start_cut_in_line);
2833 return (ZIO_PIPELINE_STOP);
2834 }
2835
2836 /*
2837 * If we got an error on a leaf device, convert it to ENXIO
2838 * if the device is not accessible at all.
2839 */
2840 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2841 !vdev_accessible(vd, zio))
2842 zio->io_error = SET_ERROR(ENXIO);
2843
2844 /*
2845 * If we can't write to an interior vdev (mirror or RAID-Z),
2846 * set vdev_cant_write so that we stop trying to allocate from it.
2847 */
2848 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2849 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2850 vd->vdev_cant_write = B_TRUE;
2851 }
2852
2853 if (zio->io_error)
2854 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2855
2856 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2857 zio->io_physdone != NULL) {
2858 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2859 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2860 zio->io_physdone(zio->io_logical);
2861 }
2862
2863 return (ZIO_PIPELINE_CONTINUE);
2864}
2865
2866void
2867zio_vdev_io_reissue(zio_t *zio)
2868{
2869 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2870 ASSERT(zio->io_error == 0);
2871
2872 zio->io_stage >>= 1;
2873}
2874
2875void
2876zio_vdev_io_redone(zio_t *zio)
2877{
2878 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2879
2880 zio->io_stage >>= 1;
2881}
2882
2883void
2884zio_vdev_io_bypass(zio_t *zio)
2885{
2886 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2887 ASSERT(zio->io_error == 0);
2888
2889 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2890 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2891}
2892
2893/*
2894 * ==========================================================================
2895 * Generate and verify checksums
2896 * ==========================================================================
2897 */
2898static int
2899zio_checksum_generate(zio_t **ziop)
2900{
2901 zio_t *zio = *ziop;
2902 blkptr_t *bp = zio->io_bp;
2903 enum zio_checksum checksum;
2904
2905 if (bp == NULL) {
2906 /*
2907 * This is zio_write_phys().
2908 * We're either generating a label checksum, or none at all.
2909 */
2910 checksum = zio->io_prop.zp_checksum;
2911
2912 if (checksum == ZIO_CHECKSUM_OFF)
2913 return (ZIO_PIPELINE_CONTINUE);
2914
2915 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2916 } else {
2917 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2918 ASSERT(!IO_IS_ALLOCATING(zio));
2919 checksum = ZIO_CHECKSUM_GANG_HEADER;
2920 } else {
2921 checksum = BP_GET_CHECKSUM(bp);
2922 }
2923 }
2924
2925 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2926
2927 return (ZIO_PIPELINE_CONTINUE);
2928}
2929
2930static int
2931zio_checksum_verify(zio_t **ziop)
2932{
2933 zio_t *zio = *ziop;
2934 zio_bad_cksum_t info;
2935 blkptr_t *bp = zio->io_bp;
2936 int error;
2937
2938 ASSERT(zio->io_vd != NULL);
2939
2940 if (bp == NULL) {
2941 /*
2942 * This is zio_read_phys().
2943 * We're either verifying a label checksum, or nothing at all.
2944 */
2945 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2946 return (ZIO_PIPELINE_CONTINUE);
2947
2948 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2949 }
2950
2951 if ((error = zio_checksum_error(zio, &info)) != 0) {
2952 zio->io_error = error;
2953 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2954 zfs_ereport_start_checksum(zio->io_spa,
2955 zio->io_vd, zio, zio->io_offset,
2956 zio->io_size, NULL, &info);
2957 }
2958 }
2959
2960 return (ZIO_PIPELINE_CONTINUE);
2961}
2962
2963/*
2964 * Called by RAID-Z to ensure we don't compute the checksum twice.
2965 */
2966void
2967zio_checksum_verified(zio_t *zio)
2968{
2969 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2970}
2971
2972/*
2973 * ==========================================================================
2974 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2975 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2976 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2977 * indicate errors that are specific to one I/O, and most likely permanent.
2978 * Any other error is presumed to be worse because we weren't expecting it.
2979 * ==========================================================================
2980 */
2981int
2982zio_worst_error(int e1, int e2)
2983{
2984 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2985 int r1, r2;
2986
2987 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2988 if (e1 == zio_error_rank[r1])
2989 break;
2990
2991 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2992 if (e2 == zio_error_rank[r2])
2993 break;
2994
2995 return (r1 > r2 ? e1 : e2);
2996}
2997
2998/*
2999 * ==========================================================================
3000 * I/O completion
3001 * ==========================================================================
3002 */
3003static int
3004zio_ready(zio_t **ziop)
3005{
3006 zio_t *zio = *ziop;
3007 blkptr_t *bp = zio->io_bp;
3008 zio_t *pio, *pio_next;
3009
3010 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3011 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3012 return (ZIO_PIPELINE_STOP);
3013
3014 if (zio->io_ready) {
3015 ASSERT(IO_IS_ALLOCATING(zio));
3016 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3017 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3018 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3019
3020 zio->io_ready(zio);
3021 }
3022
3023 if (bp != NULL && bp != &zio->io_bp_copy)
3024 zio->io_bp_copy = *bp;
3025
3026 if (zio->io_error)
3027 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3028
3029 mutex_enter(&zio->io_lock);
3030 zio->io_state[ZIO_WAIT_READY] = 1;
3031 pio = zio_walk_parents(zio);
3032 mutex_exit(&zio->io_lock);
3033
3034 /*
3035 * As we notify zio's parents, new parents could be added.
3036 * New parents go to the head of zio's io_parent_list, however,
3037 * so we will (correctly) not notify them. The remainder of zio's
3038 * io_parent_list, from 'pio_next' onward, cannot change because
3039 * all parents must wait for us to be done before they can be done.
3040 */
3041 for (; pio != NULL; pio = pio_next) {
3042 pio_next = zio_walk_parents(zio);
3043 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3044 }
3045
3046 if (zio->io_flags & ZIO_FLAG_NODATA) {
3047 if (BP_IS_GANG(bp)) {
3048 zio->io_flags &= ~ZIO_FLAG_NODATA;
3049 } else {
3050 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3051 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3052 }
3053 }
3054
3055 if (zio_injection_enabled &&
3056 zio->io_spa->spa_syncing_txg == zio->io_txg)
3057 zio_handle_ignored_writes(zio);
3058
3059 return (ZIO_PIPELINE_CONTINUE);
3060}
3061
3062static int
3063zio_done(zio_t **ziop)
3064{
3065 zio_t *zio = *ziop;
3066 spa_t *spa = zio->io_spa;
3067 zio_t *lio = zio->io_logical;
3068 blkptr_t *bp = zio->io_bp;
3069 vdev_t *vd = zio->io_vd;
3070 uint64_t psize = zio->io_size;
3071 zio_t *pio, *pio_next;
3072
3073 /*
3074 * If our children haven't all completed,
3075 * wait for them and then repeat this pipeline stage.
3076 */
3077 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3078 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3079 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3080 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3081 return (ZIO_PIPELINE_STOP);
3082
3083 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3084 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3085 ASSERT(zio->io_children[c][w] == 0);
3086
3087 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3088 ASSERT(bp->blk_pad[0] == 0);
3089 ASSERT(bp->blk_pad[1] == 0);
3090 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3091 (bp == zio_unique_parent(zio)->io_bp));
3092 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3093 zio->io_bp_override == NULL &&
3094 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3095 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3096 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3097 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3098 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3099 }
3100 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3101 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3102 }
3103
3104 /*
3105 * If there were child vdev/gang/ddt errors, they apply to us now.
3106 */
3107 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3108 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3109 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3110
3111 /*
3112 * If the I/O on the transformed data was successful, generate any
3113 * checksum reports now while we still have the transformed data.
3114 */
3115 if (zio->io_error == 0) {
3116 while (zio->io_cksum_report != NULL) {
3117 zio_cksum_report_t *zcr = zio->io_cksum_report;
3118 uint64_t align = zcr->zcr_align;
3119 uint64_t asize = P2ROUNDUP(psize, align);
3120 char *abuf = zio->io_data;
3121
3122 if (asize != psize) {
3123 abuf = zio_buf_alloc(asize);
3124 bcopy(zio->io_data, abuf, psize);
3125 bzero(abuf + psize, asize - psize);
3126 }
3127
3128 zio->io_cksum_report = zcr->zcr_next;
3129 zcr->zcr_next = NULL;
3130 zcr->zcr_finish(zcr, abuf);
3131 zfs_ereport_free_checksum(zcr);
3132
3133 if (asize != psize)
3134 zio_buf_free(abuf, asize);
3135 }
3136 }
3137
3138 zio_pop_transforms(zio); /* note: may set zio->io_error */
3139
3140 vdev_stat_update(zio, psize);
3141
3142 if (zio->io_error) {
3143 /*
3144 * If this I/O is attached to a particular vdev,
3145 * generate an error message describing the I/O failure
3146 * at the block level. We ignore these errors if the
3147 * device is currently unavailable.
3148 */
3149 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3150 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3151
3152 if ((zio->io_error == EIO || !(zio->io_flags &
3153 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3154 zio == lio) {
3155 /*
3156 * For logical I/O requests, tell the SPA to log the
3157 * error and generate a logical data ereport.
3158 */
3159 spa_log_error(spa, zio);
3160 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3161 0, 0);
3162 }
3163 }
3164
3165 if (zio->io_error && zio == lio) {
3166 /*
3167 * Determine whether zio should be reexecuted. This will
3168 * propagate all the way to the root via zio_notify_parent().
3169 */
3170 ASSERT(vd == NULL && bp != NULL);
3171 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3172
3173 if (IO_IS_ALLOCATING(zio) &&
3174 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3175 if (zio->io_error != ENOSPC)
3176 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3177 else
3178 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3179 }
3180
3181 if ((zio->io_type == ZIO_TYPE_READ ||
3182 zio->io_type == ZIO_TYPE_FREE) &&
3183 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3184 zio->io_error == ENXIO &&
3185 spa_load_state(spa) == SPA_LOAD_NONE &&
3186 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3187 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3188
3189 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3190 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3191
3192 /*
3193 * Here is a possibly good place to attempt to do
3194 * either combinatorial reconstruction or error correction
3195 * based on checksums. It also might be a good place
3196 * to send out preliminary ereports before we suspend
3197 * processing.
3198 */
3199 }
3200
3201 /*
3202 * If there were logical child errors, they apply to us now.
3203 * We defer this until now to avoid conflating logical child
3204 * errors with errors that happened to the zio itself when
3205 * updating vdev stats and reporting FMA events above.
3206 */
3207 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3208
3209 if ((zio->io_error || zio->io_reexecute) &&
3210 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3211 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3212 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3213
3214 zio_gang_tree_free(&zio->io_gang_tree);
3215
3216 /*
3217 * Godfather I/Os should never suspend.
3218 */
3219 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3220 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3221 zio->io_reexecute = 0;
3222
3223 if (zio->io_reexecute) {
3224 /*
3225 * This is a logical I/O that wants to reexecute.
3226 *
3227 * Reexecute is top-down. When an i/o fails, if it's not
3228 * the root, it simply notifies its parent and sticks around.
3229 * The parent, seeing that it still has children in zio_done(),
3230 * does the same. This percolates all the way up to the root.
3231 * The root i/o will reexecute or suspend the entire tree.
3232 *
3233 * This approach ensures that zio_reexecute() honors
3234 * all the original i/o dependency relationships, e.g.
3235 * parents not executing until children are ready.
3236 */
3237 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3238
3239 zio->io_gang_leader = NULL;
3240
3241 mutex_enter(&zio->io_lock);
3242 zio->io_state[ZIO_WAIT_DONE] = 1;
3243 mutex_exit(&zio->io_lock);
3244
3245 /*
3246 * "The Godfather" I/O monitors its children but is
3247 * not a true parent to them. It will track them through
3248 * the pipeline but severs its ties whenever they get into
3249 * trouble (e.g. suspended). This allows "The Godfather"
3250 * I/O to return status without blocking.
3251 */
3252 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3253 zio_link_t *zl = zio->io_walk_link;
3254 pio_next = zio_walk_parents(zio);
3255
3256 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3257 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3258 zio_remove_child(pio, zio, zl);
3259 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3260 }
3261 }
3262
3263 if ((pio = zio_unique_parent(zio)) != NULL) {
3264 /*
3265 * We're not a root i/o, so there's nothing to do
3266 * but notify our parent. Don't propagate errors
3267 * upward since we haven't permanently failed yet.
3268 */
3269 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3270 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3271 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3272 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3273 /*
3274 * We'd fail again if we reexecuted now, so suspend
3275 * until conditions improve (e.g. device comes online).
3276 */
3277 zio_suspend(spa, zio);
3278 } else {
3279 /*
3280 * Reexecution is potentially a huge amount of work.
3281 * Hand it off to the otherwise-unused claim taskq.
3282 */
3283#if defined(illumos) || !defined(_KERNEL)
3284 ASSERT(zio->io_tqent.tqent_next == NULL);
3285#else
3286 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3287#endif
3288 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3289 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3290 0, &zio->io_tqent);
3291 }
3292 return (ZIO_PIPELINE_STOP);
3293 }
3294
3295 ASSERT(zio->io_child_count == 0);
3296 ASSERT(zio->io_reexecute == 0);
3297 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3298
3299 /*
3300 * Report any checksum errors, since the I/O is complete.
3301 */
3302 while (zio->io_cksum_report != NULL) {
3303 zio_cksum_report_t *zcr = zio->io_cksum_report;
3304 zio->io_cksum_report = zcr->zcr_next;
3305 zcr->zcr_next = NULL;
3306 zcr->zcr_finish(zcr, NULL);
3307 zfs_ereport_free_checksum(zcr);
3308 }
3309
3310 /*
3311 * It is the responsibility of the done callback to ensure that this
3312 * particular zio is no longer discoverable for adoption, and as
3313 * such, cannot acquire any new parents.
3314 */
3315 if (zio->io_done)
3316 zio->io_done(zio);
3317
3318 mutex_enter(&zio->io_lock);
3319 zio->io_state[ZIO_WAIT_DONE] = 1;
3320 mutex_exit(&zio->io_lock);
3321
3322 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3323 zio_link_t *zl = zio->io_walk_link;
3324 pio_next = zio_walk_parents(zio);
3325 zio_remove_child(pio, zio, zl);
3326 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3327 }
3328
3329 if (zio->io_waiter != NULL) {
3330 mutex_enter(&zio->io_lock);
3331 zio->io_executor = NULL;
3332 cv_broadcast(&zio->io_cv);
3333 mutex_exit(&zio->io_lock);
3334 } else {
3335 zio_destroy(zio);
3336 }
3337
3338 return (ZIO_PIPELINE_STOP);
3339}
3340
3341/*
3342 * ==========================================================================
3343 * I/O pipeline definition
3344 * ==========================================================================
3345 */
3346static zio_pipe_stage_t *zio_pipeline[] = {
3347 NULL,
3348 zio_read_bp_init,
3349 zio_free_bp_init,
3350 zio_issue_async,
3351 zio_write_bp_init,
3352 zio_checksum_generate,
3353 zio_nop_write,
3354 zio_ddt_read_start,
3355 zio_ddt_read_done,
3356 zio_ddt_write,
3357 zio_ddt_free,
3358 zio_gang_assemble,
3359 zio_gang_issue,
3360 zio_dva_allocate,
3361 zio_dva_free,
3362 zio_dva_claim,
3363 zio_ready,
3364 zio_vdev_io_start,
3365 zio_vdev_io_done,
3366 zio_vdev_io_assess,
3367 zio_checksum_verify,
3368 zio_done
3369};
3370
3371/* dnp is the dnode for zb1->zb_object */
3372boolean_t
|