28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/zfeature.h>
49 #include <sys/abd.h>
50 #ifdef _KERNEL
51 #include <sys/vmsystm.h>
52 #include <sys/zfs_znode.h>
53 #endif
54
55 /*
56 * Enable/disable nopwrite feature.
57 */
58 int zfs_nopwrite_enabled = 1;
59
60 /*
61 * Tunable to control percentage of dirtied blocks from frees in one TXG.
62 * After this threshold is crossed, additional dirty blocks from frees
63 * wait until the next TXG.
64 * A value of zero will disable this throttle.
65 */
66 uint32_t zfs_per_txg_dirty_frees_percent = 30;
67
68 /*
69 * This can be used for testing, to ensure that certain actions happen
70 * while in the middle of a remap (which might otherwise complete too
71 * quickly).
72 */
73 int zfs_object_remap_one_indirect_delay_ticks = 0;
74
75 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
76 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
77 { DMU_BSWAP_ZAP, TRUE, "object directory" },
78 { DMU_BSWAP_UINT64, TRUE, "object array" },
79 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
80 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
81 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
82 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
83 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
84 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
85 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
86 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
87 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
88 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
90 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
91 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
92 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
93 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
94 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
95 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
98 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
99 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
100 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
101 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
102 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
103 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
104 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
105 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
106 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
107 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
108 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
109 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
110 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
111 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
112 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
113 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
114 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
115 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
116 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
117 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
118 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
119 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
120 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
121 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
122 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
123 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
124 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
125 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
126 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
127 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
128 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
129 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
130 };
131
132 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
133 { byteswap_uint8_array, "uint8" },
134 { byteswap_uint16_array, "uint16" },
135 { byteswap_uint32_array, "uint32" },
136 { byteswap_uint64_array, "uint64" },
137 { zap_byteswap, "zap" },
138 { dnode_buf_byteswap, "dnode" },
139 { dmu_objset_byteswap, "objset" },
140 { zfs_znode_byteswap, "znode" },
141 { zfs_oldacl_byteswap, "oldacl" },
142 { zfs_acl_byteswap, "acl" }
143 };
144
145 int
146 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
147 void *tag, dmu_buf_t **dbp)
148 {
149 uint64_t blkid;
695 err = dnode_next_offset(dn,
696 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
697
698 /* if there are no indirect blocks before start, we are done */
699 if (err == ESRCH) {
700 *start = minimum;
701 break;
702 } else if (err != 0) {
703 return (err);
704 }
705
706 /* set start to the beginning of this L1 indirect */
707 *start = P2ALIGN(*start, iblkrange);
708 }
709 if (*start < minimum)
710 *start = minimum;
711 return (0);
712 }
713
714 /*
715 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
716 * otherwise return false.
717 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
718 */
719 /*ARGSUSED*/
720 static boolean_t
721 dmu_objset_zfs_unmounting(objset_t *os)
722 {
723 #ifdef _KERNEL
724 if (dmu_objset_type(os) == DMU_OST_ZFS)
725 return (zfs_get_vfs_flag_unmounted(os));
726 #endif
727 return (B_FALSE);
728 }
729
730 static int
731 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
732 uint64_t length)
733 {
734 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
735 int err;
736 uint64_t dirty_frees_threshold;
737 dsl_pool_t *dp = dmu_objset_pool(os);
738
739 if (offset >= object_size)
740 return (0);
741
742 if (zfs_per_txg_dirty_frees_percent <= 100)
743 dirty_frees_threshold =
744 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
745 else
746 dirty_frees_threshold = zfs_dirty_data_max / 4;
747
748 if (length == DMU_OBJECT_END || offset + length > object_size)
749 length = object_size - offset;
750
751 while (length != 0) {
752 uint64_t chunk_end, chunk_begin, chunk_len;
753 uint64_t long_free_dirty_all_txgs = 0;
754 dmu_tx_t *tx;
755
756 if (dmu_objset_zfs_unmounting(dn->dn_objset))
757 return (SET_ERROR(EINTR));
758
759 chunk_end = chunk_begin = offset + length;
760
761 /* move chunk_begin backwards to the beginning of this chunk */
762 err = get_next_chunk(dn, &chunk_begin, offset);
763 if (err)
764 return (err);
765 ASSERT3U(chunk_begin, >=, offset);
766 ASSERT3U(chunk_begin, <=, chunk_end);
767
768 chunk_len = chunk_end - chunk_begin;
769
770 mutex_enter(&dp->dp_lock);
771 for (int t = 0; t < TXG_SIZE; t++) {
772 long_free_dirty_all_txgs +=
773 dp->dp_long_free_dirty_pertxg[t];
774 }
775 mutex_exit(&dp->dp_lock);
776
777 /*
779 * the next TXG to open before freeing more chunks if
780 * we have reached the threshold of frees
781 */
782 if (dirty_frees_threshold != 0 &&
783 long_free_dirty_all_txgs >= dirty_frees_threshold) {
784 txg_wait_open(dp, 0);
785 continue;
786 }
787
788 tx = dmu_tx_create(os);
789 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
790
791 /*
792 * Mark this transaction as typically resulting in a net
793 * reduction in space used.
794 */
795 dmu_tx_mark_netfree(tx);
796 err = dmu_tx_assign(tx, TXG_WAIT);
797 if (err) {
798 dmu_tx_abort(tx);
799 return (err);
800 }
801
802 mutex_enter(&dp->dp_lock);
803 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
804 chunk_len;
805 mutex_exit(&dp->dp_lock);
806 DTRACE_PROBE3(free__long__range,
807 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
808 uint64_t, dmu_tx_get_txg(tx));
809 dnode_free_range(dn, chunk_begin, chunk_len, tx);
810 dmu_tx_commit(tx);
811
812 length -= chunk_len;
813 }
814 return (0);
815 }
816
817 int
818 dmu_free_long_range(objset_t *os, uint64_t object,
1004 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1005 dmu_buf_rele_array(dbp, numbufs, FTAG);
1006 }
1007
1008 void
1009 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1010 const void *buf, dmu_tx_t *tx)
1011 {
1012 dmu_buf_t **dbp;
1013 int numbufs;
1014
1015 if (size == 0)
1016 return;
1017
1018 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1019 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1020 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1021 dmu_buf_rele_array(dbp, numbufs, FTAG);
1022 }
1023
1024 static int
1025 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1026 uint64_t last_removal_txg, uint64_t offset)
1027 {
1028 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1029 int err = 0;
1030
1031 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1032 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1033 ASSERT3P(dbuf, !=, NULL);
1034
1035 /*
1036 * If the block hasn't been written yet, this default will ensure
1037 * we don't try to remap it.
1038 */
1039 uint64_t birth = UINT64_MAX;
1040 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1041 if (dbuf->db_blkptr != NULL)
1042 birth = dbuf->db_blkptr->blk_birth;
1043 rw_exit(&dn->dn_struct_rwlock);
1044
1045 /*
1046 * If this L1 was already written after the last removal, then we've
1047 * already tried to remap it.
1048 */
1049 if (birth <= last_removal_txg &&
1050 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1051 dbuf_can_remap(dbuf)) {
1052 dmu_tx_t *tx = dmu_tx_create(os);
1053 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1054 err = dmu_tx_assign(tx, TXG_WAIT);
1055 if (err == 0) {
1056 (void) dbuf_dirty(dbuf, tx);
1057 dmu_tx_commit(tx);
1058 } else {
1059 dmu_tx_abort(tx);
1060 }
1061 }
1062
1063 dbuf_rele(dbuf, FTAG);
1064
1065 delay(zfs_object_remap_one_indirect_delay_ticks);
1066
1067 return (err);
1068 }
1069
1070 /*
1071 * Remap all blockpointers in the object, if possible, so that they reference
1072 * only concrete vdevs.
1073 *
1074 * To do this, iterate over the L0 blockpointers and remap any that reference
1075 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1076 * cannot guarantee that we can remap all blockpointer anyways (due to split
1077 * blocks), we do not want to make the code unnecessarily complicated to
1078 * catch the unlikely case that there is an L1 block on an indirect vdev that
1079 * contains no indirect blockpointers.
1080 */
1081 int
1082 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1083 uint64_t last_removal_txg)
1084 {
1085 uint64_t offset, l1span;
1086 int err;
1087 dnode_t *dn;
1088
1089 err = dnode_hold(os, object, FTAG, &dn);
1090 if (err != 0) {
1091 return (err);
1092 }
1093
1094 if (dn->dn_nlevels <= 1) {
1095 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1096 err = SET_ERROR(EINTR);
1097 }
1098
1099 /*
1100 * If the dnode has no indirect blocks, we cannot dirty them.
1101 * We still want to remap the blkptr(s) in the dnode if
1102 * appropriate, so mark it as dirty.
1103 */
1104 if (err == 0 && dnode_needs_remap(dn)) {
1105 dmu_tx_t *tx = dmu_tx_create(os);
1106 dmu_tx_hold_bonus(tx, dn->dn_object);
1107 if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1108 dnode_setdirty(dn, tx);
1109 dmu_tx_commit(tx);
1110 } else {
1111 dmu_tx_abort(tx);
1112 }
1113 }
1114
1115 dnode_rele(dn, FTAG);
1116 return (err);
1117 }
1118
1119 offset = 0;
1120 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1121 dn->dn_datablkshift);
1122 /*
1123 * Find the next L1 indirect that is not a hole.
1124 */
1125 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1126 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1127 err = SET_ERROR(EINTR);
1128 break;
1129 }
1130 if ((err = dmu_object_remap_one_indirect(os, dn,
1131 last_removal_txg, offset)) != 0) {
1132 break;
1133 }
1134 offset += l1span;
1135 }
1136
1137 dnode_rele(dn, FTAG);
1138 return (err);
1139 }
1140
1141 void
1142 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1143 dmu_tx_t *tx)
1144 {
1145 dmu_buf_t **dbp;
1146 int numbufs, i;
1147
1148 if (size == 0)
1149 return;
1150
1151 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1152 FALSE, FTAG, &numbufs, &dbp));
1153
1154 for (i = 0; i < numbufs; i++) {
1155 dmu_buf_t *db = dbp[i];
1156
1157 dmu_buf_will_not_fill(db, tx);
1158 }
1159 dmu_buf_rele_array(dbp, numbufs, FTAG);
1160 }
1670 } else if (!BP_IS_EMBEDDED(bp)) {
1671 ASSERT(BP_GET_LEVEL(bp) == 0);
1672 bp->blk_fill = 1;
1673 }
1674 }
1675 }
1676
1677 static void
1678 dmu_sync_late_arrival_ready(zio_t *zio)
1679 {
1680 dmu_sync_ready(zio, NULL, zio->io_private);
1681 }
1682
1683 /* ARGSUSED */
1684 static void
1685 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1686 {
1687 dmu_sync_arg_t *dsa = varg;
1688 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1689 dmu_buf_impl_t *db = dr->dr_dbuf;
1690
1691 mutex_enter(&db->db_mtx);
1692 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1693 if (zio->io_error == 0) {
1694 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1695 if (dr->dt.dl.dr_nopwrite) {
1696 blkptr_t *bp = zio->io_bp;
1697 blkptr_t *bp_orig = &zio->io_bp_orig;
1698 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1699
1700 ASSERT(BP_EQUAL(bp, bp_orig));
1701 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1702 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1703 ASSERT(zio_checksum_table[chksum].ci_flags &
1704 ZCHECKSUM_FLAG_NOPWRITE);
1705 }
1706 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1707 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1708 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1709
1710 /*
1720 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1721 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1722 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1723 } else {
1724 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1725 }
1726 cv_broadcast(&db->db_changed);
1727 mutex_exit(&db->db_mtx);
1728
1729 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1730
1731 kmem_free(dsa, sizeof (*dsa));
1732 }
1733
1734 static void
1735 dmu_sync_late_arrival_done(zio_t *zio)
1736 {
1737 blkptr_t *bp = zio->io_bp;
1738 dmu_sync_arg_t *dsa = zio->io_private;
1739 blkptr_t *bp_orig = &zio->io_bp_orig;
1740
1741 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1742 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1743 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1744 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1745 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1746 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1747 }
1748
1749 dmu_tx_commit(dsa->dsa_tx);
1750
1751 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1752
1753 abd_put(zio->io_abd);
1754 kmem_free(dsa, sizeof (*dsa));
1755 }
1756
1757 static int
1758 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1759 zio_prop_t *zp, zbookmark_phys_t *zb)
1760 {
1761 dmu_sync_arg_t *dsa;
1762 dmu_tx_t *tx;
1763
1764 tx = dmu_tx_create(os);
1765 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1766 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1767 dmu_tx_abort(tx);
1768 /* Make zl_get_data do txg_waited_synced() */
1769 return (SET_ERROR(EIO));
1770 }
1771
1772 /*
1773 * In order to prevent the zgd's lwb from being free'd prior to
1774 * dmu_sync_late_arrival_done() being called, we have to ensure
1775 * the lwb's "max txg" takes this tx's txg into account.
1776 */
1777 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1778
1779 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1792 * dirty in a future txg).
1793 *
1794 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1795 * We can not nopwrite against it because although the BP will not
1796 * (typically) be changed, the data has not yet been persisted to this
1797 * location.
1798 *
1799 * Finally, when dbuf_write_done() is called, it is theoretically
1800 * possible to always nopwrite, because the data that was written in
1801 * this txg is the same data that we are trying to write. However we
1802 * would need to check that this dbuf is not dirty in any future
1803 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1804 * don't nopwrite in this case.
1805 */
1806 zp->zp_nopwrite = B_FALSE;
1807
1808 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1809 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1810 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1811 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1812 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1813
1814 return (0);
1815 }
1816
1817 /*
1818 * Intent log support: sync the block associated with db to disk.
1819 * N.B. and XXX: the caller is responsible for making sure that the
1820 * data isn't changing while dmu_sync() is writing it.
1821 *
1822 * Return values:
1823 *
1824 * EEXIST: this txg has already been synced, so there's nothing to do.
1825 * The caller should not log the write.
1826 *
1827 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1828 * The caller should not log the write.
1829 *
1830 * EALREADY: this block is already in the process of being synced.
1831 * The caller should track its progress (somehow).
1832 *
1833 * EIO: could not do the I/O.
1834 * The caller should do a txg_wait_synced().
1835 *
1836 * 0: the I/O has been initiated.
1837 * The caller should log this blkptr in the done callback.
1838 * It is possible that the I/O will fail, in which case
1839 * the error will be reported to the done callback and
1840 * propagated to pio from zio_done().
1841 */
1842 int
1843 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1844 {
1845 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1846 objset_t *os = db->db_objset;
1847 dsl_dataset_t *ds = os->os_dsl_dataset;
1848 dbuf_dirty_record_t *dr;
1849 dmu_sync_arg_t *dsa;
1850 zbookmark_phys_t zb;
1851 zio_prop_t zp;
1852 dnode_t *dn;
1853
1854 ASSERT(pio != NULL);
1855 ASSERT(txg != 0);
1856
1857 SET_BOOKMARK(&zb, ds->ds_object,
1858 db->db.db_object, db->db_level, db->db_blkid);
1859
1860 DB_DNODE_ENTER(db);
1861 dn = DB_DNODE(db);
1862 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1863 DB_DNODE_EXIT(db);
1864
1865 /*
1866 * If we're frozen (running ziltest), we always need to generate a bp.
1867 */
1868 if (txg > spa_freeze_txg(os->os_spa))
1869 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1870
1871 /*
1872 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1873 * and us. If we determine that this txg is not yet syncing,
1874 * but it begins to sync a moment later, that's OK because the
1875 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1876 */
1877 mutex_enter(&db->db_mtx);
1878
1879 if (txg <= spa_last_synced_txg(os->os_spa)) {
1880 /*
1881 * This txg has already synced. There's nothing to do.
1882 */
1883 mutex_exit(&db->db_mtx);
1884 return (SET_ERROR(EEXIST));
1885 }
1886
1887 if (txg <= spa_syncing_txg(os->os_spa)) {
1888 /*
1889 * This txg is currently syncing, so we can't mess with
1890 * the dirty record anymore; just write a new log block.
1891 */
1892 mutex_exit(&db->db_mtx);
1893 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1894 }
1895
1896 dr = db->db_last_dirty;
1897 while (dr && dr->dr_txg != txg)
1898 dr = dr->dr_next;
1899
1900 if (dr == NULL) {
1901 /*
1902 * There's no dr for this dbuf, so it must have been freed.
1903 * There's no need to log writes to freed blocks, so we're done.
1904 */
1905 mutex_exit(&db->db_mtx);
1906 return (SET_ERROR(ENOENT));
1907 }
1908
1909 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1910
1911 if (db->db_blkptr != NULL) {
1912 /*
1913 * We need to fill in zgd_bp with the current blkptr so that
1959 * or this buffer has already been synced. It could not
1960 * have been dirtied since, or we would have cleared the state.
1961 */
1962 mutex_exit(&db->db_mtx);
1963 return (SET_ERROR(EALREADY));
1964 }
1965
1966 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1967 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1968 mutex_exit(&db->db_mtx);
1969
1970 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1971 dsa->dsa_dr = dr;
1972 dsa->dsa_done = done;
1973 dsa->dsa_zgd = zgd;
1974 dsa->dsa_tx = NULL;
1975
1976 zio_nowait(arc_write(pio, os->os_spa, txg,
1977 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1978 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1979 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1980
1981 return (0);
1982 }
1983
1984 int
1985 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1986 dmu_tx_t *tx)
1987 {
1988 dnode_t *dn;
1989 int err;
1990
1991 err = dnode_hold(os, object, FTAG, &dn);
1992 if (err)
1993 return (err);
1994 err = dnode_set_blksz(dn, size, ibs, tx);
1995 dnode_rele(dn, FTAG);
1996 return (err);
1997 }
1998
1999 void
2125 */
2126 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2127 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2128 if (!(zio_checksum_table[checksum].ci_flags &
2129 ZCHECKSUM_FLAG_DEDUP))
2130 dedup_verify = B_TRUE;
2131 }
2132
2133 /*
2134 * Enable nopwrite if we have secure enough checksum
2135 * algorithm (see comment in zio_nop_write) and
2136 * compression is enabled. We don't enable nopwrite if
2137 * dedup is enabled as the two features are mutually
2138 * exclusive.
2139 */
2140 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2141 ZCHECKSUM_FLAG_NOPWRITE) &&
2142 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2143 }
2144
2145 zp->zp_checksum = checksum;
2146 zp->zp_compress = compress;
2147 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2148
2149 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2150 zp->zp_level = level;
2151 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2152 zp->zp_dedup = dedup;
2153 zp->zp_dedup_verify = dedup && dedup_verify;
2154 zp->zp_nopwrite = nopwrite;
2155 }
2156
2157 int
2158 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2159 {
2160 dnode_t *dn;
2161 int err;
2162
2163 /*
2164 * Sync any current changes before
2165 * we go trundling through the block pointers.
2166 */
2167 err = dmu_object_wait_synced(os, object);
2168 if (err) {
2169 return (err);
2170 }
2171
2172 err = dnode_hold(os, object, FTAG, &dn);
2173 if (err) {
2174 return (err);
|
28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/spa_impl.h>
49 #include <sys/zfeature.h>
50 #include <sys/abd.h>
51 #ifdef _KERNEL
52 #include <sys/vmsystm.h>
53 #include <sys/zfs_znode.h>
54 #include <sys/zfs_vfsops.h>
55 #endif
56 #include <sys/special.h>
57
58 /*
59 * Enable/disable nopwrite feature.
60 */
61 int zfs_nopwrite_enabled = 1;
62
63 /*
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
68 */
69 uint32_t zfs_per_txg_dirty_frees_percent = 30;
70
71 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
72 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
73 { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" },
74 { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" },
75 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
76 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
77 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
78 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
79 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
80 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
81 { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" },
82 { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" },
83 { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" },
84 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" },
85 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" },
86 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" },
87 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" },
88 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" },
89 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
90 { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" },
91 { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" },
92 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" },
93 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
94 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" },
95 { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" },
96 { DMU_BSWAP_ZAP, TRUE, TRUE, "zvol prop" },
97 { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" },
98 { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" },
99 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
100 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
101 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
102 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
103 { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" },
104 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" },
105 { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" },
106 { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" },
107 { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" },
108 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
109 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" },
110 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
111 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" },
112 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" },
113 { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" },
114 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
115 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
116 { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" },
117 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" },
118 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" },
119 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" },
120 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
121 { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" },
122 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" },
123 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" },
124 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" },
125 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
126 };
127
128 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
129 { byteswap_uint8_array, "uint8" },
130 { byteswap_uint16_array, "uint16" },
131 { byteswap_uint32_array, "uint32" },
132 { byteswap_uint64_array, "uint64" },
133 { zap_byteswap, "zap" },
134 { dnode_buf_byteswap, "dnode" },
135 { dmu_objset_byteswap, "objset" },
136 { zfs_znode_byteswap, "znode" },
137 { zfs_oldacl_byteswap, "oldacl" },
138 { zfs_acl_byteswap, "acl" }
139 };
140
141 int
142 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
143 void *tag, dmu_buf_t **dbp)
144 {
145 uint64_t blkid;
691 err = dnode_next_offset(dn,
692 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
693
694 /* if there are no indirect blocks before start, we are done */
695 if (err == ESRCH) {
696 *start = minimum;
697 break;
698 } else if (err != 0) {
699 return (err);
700 }
701
702 /* set start to the beginning of this L1 indirect */
703 *start = P2ALIGN(*start, iblkrange);
704 }
705 if (*start < minimum)
706 *start = minimum;
707 return (0);
708 }
709
710 /*
711 * If this dnode is in the ZFS object set
712 * return true if vfs's unmounted flag is set or the
713 * zfsvfs is currently suspended, otherwise return false.
714 */
715 /*ARGSUSED*/
716 static boolean_t
717 dmu_dnode_fs_unmounting_or_suspended(dnode_t *freeing_dn)
718 {
719 #ifdef _KERNEL
720 boolean_t busy = B_FALSE;
721 objset_t *os = freeing_dn->dn_objset;
722 zfsvfs_t *zfsvfs;
723
724 if (dmu_objset_type(os) == DMU_OST_ZFS) {
725 mutex_enter(&os->os_user_ptr_lock);
726 zfsvfs = dmu_objset_get_user(os);
727 if (zfsvfs != NULL && zfsvfs->z_vfs != NULL &&
728 ((zfsvfs->z_vfs->vfs_flag & VFS_UNMOUNTED) ||
729 zfsvfs->z_busy))
730 busy = B_TRUE;
731 mutex_exit(&os->os_user_ptr_lock);
732 }
733
734 return (busy);
735 #else
736 return (B_FALSE);
737 #endif
738 }
739
740 static int
741 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
742 uint64_t length)
743 {
744 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
745 int err;
746 uint64_t dirty_frees_threshold;
747 dsl_pool_t *dp = dmu_objset_pool(os);
748
749 if (offset >= object_size)
750 return (0);
751
752 if (zfs_per_txg_dirty_frees_percent <= 100)
753 dirty_frees_threshold =
754 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
755 else
756 dirty_frees_threshold = zfs_dirty_data_max / 4;
757
758 if (length == DMU_OBJECT_END && offset == 0)
759 dnode_evict_dbufs(dn, 0);
760
761 if (length == DMU_OBJECT_END || offset + length > object_size)
762 length = object_size - offset;
763
764 mutex_enter(&dp->dp_lock);
765 dp->dp_long_freeing_total += length;
766 mutex_exit(&dp->dp_lock);
767
768 while (length != 0) {
769 uint64_t chunk_end, chunk_begin, chunk_len;
770 uint64_t long_free_dirty_all_txgs = 0;
771 dmu_tx_t *tx;
772
773 if (dmu_dnode_fs_unmounting_or_suspended(dn)) {
774 mutex_enter(&dp->dp_lock);
775 dp->dp_long_freeing_total -= length;
776 mutex_exit(&dp->dp_lock);
777
778 return (SET_ERROR(EINTR));
779 }
780
781 chunk_end = chunk_begin = offset + length;
782
783 /* move chunk_begin backwards to the beginning of this chunk */
784 err = get_next_chunk(dn, &chunk_begin, offset);
785 if (err)
786 return (err);
787 ASSERT3U(chunk_begin, >=, offset);
788 ASSERT3U(chunk_begin, <=, chunk_end);
789
790 chunk_len = chunk_end - chunk_begin;
791
792 mutex_enter(&dp->dp_lock);
793 for (int t = 0; t < TXG_SIZE; t++) {
794 long_free_dirty_all_txgs +=
795 dp->dp_long_free_dirty_pertxg[t];
796 }
797 mutex_exit(&dp->dp_lock);
798
799 /*
801 * the next TXG to open before freeing more chunks if
802 * we have reached the threshold of frees
803 */
804 if (dirty_frees_threshold != 0 &&
805 long_free_dirty_all_txgs >= dirty_frees_threshold) {
806 txg_wait_open(dp, 0);
807 continue;
808 }
809
810 tx = dmu_tx_create(os);
811 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
812
813 /*
814 * Mark this transaction as typically resulting in a net
815 * reduction in space used.
816 */
817 dmu_tx_mark_netfree(tx);
818 err = dmu_tx_assign(tx, TXG_WAIT);
819 if (err) {
820 dmu_tx_abort(tx);
821 mutex_enter(&dp->dp_lock);
822 dp->dp_long_freeing_total -= length - chunk_len;
823 mutex_exit(&dp->dp_lock);
824 return (err);
825 }
826
827 mutex_enter(&dp->dp_lock);
828 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
829 chunk_len;
830 mutex_exit(&dp->dp_lock);
831 DTRACE_PROBE3(free__long__range,
832 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
833 uint64_t, dmu_tx_get_txg(tx));
834 dnode_free_range(dn, chunk_begin, chunk_len, tx);
835 dmu_tx_commit(tx);
836
837 length -= chunk_len;
838 }
839 return (0);
840 }
841
842 int
843 dmu_free_long_range(objset_t *os, uint64_t object,
1029 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1030 dmu_buf_rele_array(dbp, numbufs, FTAG);
1031 }
1032
1033 void
1034 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1035 const void *buf, dmu_tx_t *tx)
1036 {
1037 dmu_buf_t **dbp;
1038 int numbufs;
1039
1040 if (size == 0)
1041 return;
1042
1043 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1044 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1045 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1046 dmu_buf_rele_array(dbp, numbufs, FTAG);
1047 }
1048
1049 void
1050 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1051 dmu_tx_t *tx)
1052 {
1053 dmu_buf_t **dbp;
1054 int numbufs, i;
1055
1056 if (size == 0)
1057 return;
1058
1059 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1060 FALSE, FTAG, &numbufs, &dbp));
1061
1062 for (i = 0; i < numbufs; i++) {
1063 dmu_buf_t *db = dbp[i];
1064
1065 dmu_buf_will_not_fill(db, tx);
1066 }
1067 dmu_buf_rele_array(dbp, numbufs, FTAG);
1068 }
1578 } else if (!BP_IS_EMBEDDED(bp)) {
1579 ASSERT(BP_GET_LEVEL(bp) == 0);
1580 bp->blk_fill = 1;
1581 }
1582 }
1583 }
1584
1585 static void
1586 dmu_sync_late_arrival_ready(zio_t *zio)
1587 {
1588 dmu_sync_ready(zio, NULL, zio->io_private);
1589 }
1590
1591 /* ARGSUSED */
1592 static void
1593 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1594 {
1595 dmu_sync_arg_t *dsa = varg;
1596 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1597 dmu_buf_impl_t *db = dr->dr_dbuf;
1598 zgd_t *zgd = dsa->dsa_zgd;
1599
1600 /*
1601 * Record the vdev(s) backing this blkptr so they can be flushed after
1602 * the writes for the lwb have completed.
1603 */
1604 if (zio->io_error == 0) {
1605 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1606 }
1607
1608 mutex_enter(&db->db_mtx);
1609 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1610 if (zio->io_error == 0) {
1611 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1612 if (dr->dt.dl.dr_nopwrite) {
1613 blkptr_t *bp = zio->io_bp;
1614 blkptr_t *bp_orig = &zio->io_bp_orig;
1615 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1616
1617 ASSERT(BP_EQUAL(bp, bp_orig));
1618 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1619 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1620 ASSERT(zio_checksum_table[chksum].ci_flags &
1621 ZCHECKSUM_FLAG_NOPWRITE);
1622 }
1623 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1624 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1625 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1626
1627 /*
1637 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1638 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1639 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1640 } else {
1641 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1642 }
1643 cv_broadcast(&db->db_changed);
1644 mutex_exit(&db->db_mtx);
1645
1646 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1647
1648 kmem_free(dsa, sizeof (*dsa));
1649 }
1650
1651 static void
1652 dmu_sync_late_arrival_done(zio_t *zio)
1653 {
1654 blkptr_t *bp = zio->io_bp;
1655 dmu_sync_arg_t *dsa = zio->io_private;
1656 blkptr_t *bp_orig = &zio->io_bp_orig;
1657 zgd_t *zgd = dsa->dsa_zgd;
1658
1659 if (zio->io_error == 0) {
1660 /*
1661 * Record the vdev(s) backing this blkptr so they can be
1662 * flushed after the writes for the lwb have completed.
1663 */
1664 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1665
1666 if (!BP_IS_HOLE(bp)) {
1667 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1668 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1669 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1670 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1671 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1672 }
1673 }
1674
1675 dmu_tx_commit(dsa->dsa_tx);
1676
1677 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1678
1679 abd_put(zio->io_abd);
1680 kmem_free(dsa, sizeof (*dsa));
1681 }
1682
1683 static int
1684 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1685 zio_prop_t *zp, zbookmark_phys_t *zb, const zio_smartcomp_info_t *sc)
1686 {
1687 dmu_sync_arg_t *dsa;
1688 dmu_tx_t *tx;
1689
1690 tx = dmu_tx_create(os);
1691 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1692 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1693 dmu_tx_abort(tx);
1694 /* Make zl_get_data do txg_waited_synced() */
1695 return (SET_ERROR(EIO));
1696 }
1697
1698 /*
1699 * In order to prevent the zgd's lwb from being free'd prior to
1700 * dmu_sync_late_arrival_done() being called, we have to ensure
1701 * the lwb's "max txg" takes this tx's txg into account.
1702 */
1703 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1704
1705 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1718 * dirty in a future txg).
1719 *
1720 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1721 * We can not nopwrite against it because although the BP will not
1722 * (typically) be changed, the data has not yet been persisted to this
1723 * location.
1724 *
1725 * Finally, when dbuf_write_done() is called, it is theoretically
1726 * possible to always nopwrite, because the data that was written in
1727 * this txg is the same data that we are trying to write. However we
1728 * would need to check that this dbuf is not dirty in any future
1729 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1730 * don't nopwrite in this case.
1731 */
1732 zp->zp_nopwrite = B_FALSE;
1733
1734 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1735 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1736 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1737 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1738 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb, sc));
1739
1740 return (0);
1741 }
1742
1743 /*
1744 * Intent log support: sync the block associated with db to disk.
1745 * N.B. and XXX: the caller is responsible for making sure that the
1746 * data isn't changing while dmu_sync() is writing it.
1747 *
1748 * Return values:
1749 *
1750 * EEXIST: this txg has already been synced, so there's nothing to do.
1751 * The caller should not log the write.
1752 *
1753 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1754 * The caller should not log the write.
1755 *
1756 * EALREADY: this block is already in the process of being synced.
1757 * The caller should track its progress (somehow).
1758 *
1759 * EIO: could not do the I/O.
1760 * The caller should do a txg_wait_synced().
1761 *
1762 * 0: the I/O has been initiated.
1763 * The caller should log this blkptr in the done callback.
1764 * It is possible that the I/O will fail, in which case
1765 * the error will be reported to the done callback and
1766 * propagated to pio from zio_done().
1767 */
1768
1769 int
1770 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1771 {
1772 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1773 objset_t *os = db->db_objset;
1774 dsl_dataset_t *ds = os->os_dsl_dataset;
1775 dbuf_dirty_record_t *dr;
1776 dmu_sync_arg_t *dsa;
1777 zbookmark_phys_t zb;
1778 zio_prop_t zp;
1779 dnode_t *dn;
1780 int flags = 0;
1781 zio_smartcomp_info_t sc;
1782
1783 ASSERT(pio != NULL);
1784 ASSERT(txg != 0);
1785
1786 SET_BOOKMARK(&zb, ds->ds_object,
1787 db->db.db_object, db->db_level, db->db_blkid);
1788
1789 /* write to special only if proper conditions hold */
1790 if (spa_write_data_to_special(os->os_spa, os))
1791 WP_SET_SPECIALCLASS(flags, B_TRUE);
1792
1793 DB_DNODE_ENTER(db);
1794 dn = DB_DNODE(db);
1795 dmu_write_policy(os, dn, db->db_level, flags | WP_DMU_SYNC, &zp);
1796 dnode_setup_zio_smartcomp(db, &sc);
1797 DB_DNODE_EXIT(db);
1798
1799 /*
1800 * If we're frozen (running ziltest), we always need to generate a bp.
1801 */
1802 if (txg > spa_freeze_txg(os->os_spa))
1803 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb,
1804 &sc));
1805
1806 /*
1807 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1808 * and us. If we determine that this txg is not yet syncing,
1809 * but it begins to sync a moment later, that's OK because the
1810 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1811 */
1812 mutex_enter(&db->db_mtx);
1813
1814 if (txg <= spa_last_synced_txg(os->os_spa)) {
1815 /*
1816 * This txg has already synced. There's nothing to do.
1817 */
1818 mutex_exit(&db->db_mtx);
1819 return (SET_ERROR(EEXIST));
1820 }
1821
1822 if (txg <= spa_syncing_txg(os->os_spa)) {
1823 /*
1824 * This txg is currently syncing, so we can't mess with
1825 * the dirty record anymore; just write a new log block.
1826 */
1827 mutex_exit(&db->db_mtx);
1828 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb,
1829 &sc));
1830 }
1831
1832 dr = db->db_last_dirty;
1833 while (dr && dr->dr_txg != txg)
1834 dr = dr->dr_next;
1835
1836 if (dr == NULL) {
1837 /*
1838 * There's no dr for this dbuf, so it must have been freed.
1839 * There's no need to log writes to freed blocks, so we're done.
1840 */
1841 mutex_exit(&db->db_mtx);
1842 return (SET_ERROR(ENOENT));
1843 }
1844
1845 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1846
1847 if (db->db_blkptr != NULL) {
1848 /*
1849 * We need to fill in zgd_bp with the current blkptr so that
1895 * or this buffer has already been synced. It could not
1896 * have been dirtied since, or we would have cleared the state.
1897 */
1898 mutex_exit(&db->db_mtx);
1899 return (SET_ERROR(EALREADY));
1900 }
1901
1902 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1903 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1904 mutex_exit(&db->db_mtx);
1905
1906 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1907 dsa->dsa_dr = dr;
1908 dsa->dsa_done = done;
1909 dsa->dsa_zgd = zgd;
1910 dsa->dsa_tx = NULL;
1911
1912 zio_nowait(arc_write(pio, os->os_spa, txg,
1913 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1914 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1915 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb, &sc));
1916
1917 return (0);
1918 }
1919
1920 int
1921 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1922 dmu_tx_t *tx)
1923 {
1924 dnode_t *dn;
1925 int err;
1926
1927 err = dnode_hold(os, object, FTAG, &dn);
1928 if (err)
1929 return (err);
1930 err = dnode_set_blksz(dn, size, ibs, tx);
1931 dnode_rele(dn, FTAG);
1932 return (err);
1933 }
1934
1935 void
2061 */
2062 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2063 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2064 if (!(zio_checksum_table[checksum].ci_flags &
2065 ZCHECKSUM_FLAG_DEDUP))
2066 dedup_verify = B_TRUE;
2067 }
2068
2069 /*
2070 * Enable nopwrite if we have secure enough checksum
2071 * algorithm (see comment in zio_nop_write) and
2072 * compression is enabled. We don't enable nopwrite if
2073 * dedup is enabled as the two features are mutually
2074 * exclusive.
2075 */
2076 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2077 ZCHECKSUM_FLAG_NOPWRITE) &&
2078 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2079 }
2080
2081 zp->zp_usesc = WP_GET_SPECIALCLASS(wp);
2082 zp->zp_checksum = checksum;
2083 zp->zp_compress = compress;
2084 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2085
2086 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2087 zp->zp_level = level;
2088 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2089 zp->zp_dedup = dedup;
2090 zp->zp_dedup_verify = dedup && dedup_verify;
2091 zp->zp_metadata = ismd;
2092 zp->zp_nopwrite = nopwrite;
2093 zp->zp_zpl_meta_to_special = os->os_zpl_meta_to_special;
2094 zp->zp_usewbc = (zp->zp_usesc &&
2095 os->os_wbc_mode == ZFS_WBC_MODE_ON && !ismd);
2096
2097 /* explicitly control the number for copies for DDT */
2098 if (DMU_OT_IS_DDT_META(type) &&
2099 os->os_spa->spa_ddt_meta_copies > 0) {
2100 zp->zp_copies =
2101 MIN(os->os_spa->spa_ddt_meta_copies,
2102 spa_max_replication(os->os_spa));
2103 }
2104
2105 DTRACE_PROBE2(dmu_wp, boolean_t, zp->zp_metadata,
2106 boolean_t, zp->zp_usesc);
2107 }
2108
2109 int
2110 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2111 {
2112 dnode_t *dn;
2113 int err;
2114
2115 /*
2116 * Sync any current changes before
2117 * we go trundling through the block pointers.
2118 */
2119 err = dmu_object_wait_synced(os, object);
2120 if (err) {
2121 return (err);
2122 }
2123
2124 err = dnode_hold(os, object, FTAG, &dn);
2125 if (err) {
2126 return (err);
|