… | |
… | |
993 | } |
993 | } |
994 | |
994 | |
995 | static signed char |
995 | static signed char |
996 | eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) |
996 | eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) |
997 | { |
997 | { |
998 | return b->score - a->score ? b->score - a->score /* works because our signed char is always 0..100 */ |
998 | return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */ |
999 | : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; |
999 | : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; |
1000 | } |
1000 | } |
1001 | |
1001 | |
1002 | #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 |
1002 | #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 |
1003 | #define EIO_DENT_NUM_SCORES 9 |
|
|
1004 | |
1003 | |
1005 | #define EIO_QSORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ |
1004 | #define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ |
1006 | #define EIO_QSORT_SKIP 60 /* when to skip qsort completely */ |
1005 | #define EIO_SORT_FAST 60 /* when to only use insertion sort */ |
1007 | |
1006 | |
1008 | static void |
1007 | static void |
1009 | eio_dent_sort (eio_dirent *dents, int size) |
1008 | eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) |
1010 | { |
1009 | { |
1011 | if (size <= 1) |
1010 | unsigned char bits [9 + sizeof (ino_t) * 8]; |
1012 | return; /* our insertion sort relies on size > 0 */ |
1011 | unsigned char *bit = bits; |
1013 | |
1012 | |
1014 | if (size > EIO_QSORT_SKIP) /* skip quicksort for small directories */ |
1013 | assert (CHAR_BIT == 8); |
|
|
1014 | assert (sizeof (eio_dirent) * 8 < 256); |
|
|
1015 | assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */ |
|
|
1016 | assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */ |
|
|
1017 | |
|
|
1018 | if (size <= EIO_SORT_FAST) |
|
|
1019 | return; |
|
|
1020 | |
|
|
1021 | /* first prepare an array of bits to test in our radix sort */ |
|
|
1022 | /* try to take endianness into account, as well as differences in ino_t sizes */ |
|
|
1023 | /* inode_bits must contain all inodes ORed together */ |
|
|
1024 | /* which is used to skip bits that are 0 everywhere, which is very common */ |
1015 | { |
1025 | { |
1016 | /* first, use quicksort */ |
1026 | ino_t endianness; |
1017 | /* should be good for 2**31 entries */ |
|
|
1018 | int i; |
1027 | int i, j; |
1019 | struct rng { eio_dirent *l, *r; } rng [32]; |
|
|
1020 | |
1028 | |
1021 | i = 0; |
1029 | /* we store the byte offset of byte n into byte n of "endianness" */ |
1022 | rng[0].l = dents; |
1030 | for (i = 0; i < sizeof (ino_t); ++i) |
1023 | rng[0].r = dents + size; |
1031 | ((unsigned char *)&endianness)[i] = i; |
1024 | |
1032 | |
1025 | while (expect_true (i >= 0)) |
1033 | *bit++ = 0; |
|
|
1034 | |
|
|
1035 | for (i = 0; i < sizeof (ino_t); ++i) |
1026 | { |
1036 | { |
1027 | eio_dirent *L = rng [i].l; |
1037 | /* shifting off the byte offsets out of "endianness" */ |
1028 | eio_dirent *R = rng [i].r - 1; |
1038 | int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8; |
|
|
1039 | endianness >>= 8; |
1029 | |
1040 | |
1030 | if (expect_false (L + EIO_QSORT_CUTOFF < R)) |
1041 | for (j = 0; j < 8; ++j) |
|
|
1042 | if (inode_bits & (((ino_t)1) << (i * 8 + j))) |
|
|
1043 | *bit++ = offs + j; |
|
|
1044 | } |
|
|
1045 | |
|
|
1046 | for (j = 0; j < 8; ++j) |
|
|
1047 | if (score_bits & (1 << j)) |
|
|
1048 | *bit++ = offsetof (eio_dirent, score) * 8 + j; |
|
|
1049 | } |
|
|
1050 | |
|
|
1051 | /* now actually do the sorting (a variant of MSD radix sort) */ |
|
|
1052 | { |
|
|
1053 | eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base; |
|
|
1054 | eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end; |
|
|
1055 | unsigned char *bit_stk [9 + sizeof (ino_t) * 8]; |
|
|
1056 | int stk_idx = 0; |
|
|
1057 | |
|
|
1058 | base_stk [stk_idx] = dents; |
|
|
1059 | end_stk [stk_idx] = dents + size; |
|
|
1060 | bit_stk [stk_idx] = bit - 1; |
|
|
1061 | |
|
|
1062 | do |
|
|
1063 | { |
|
|
1064 | base = base_stk [stk_idx]; |
|
|
1065 | end = end_stk [stk_idx]; |
|
|
1066 | bit = bit_stk [stk_idx]; |
|
|
1067 | |
|
|
1068 | for (;;) |
1031 | { |
1069 | { |
1032 | eio_dirent *mid = &dents [((L - dents) + (R - dents)) >> 1]; |
1070 | unsigned char O = *bit >> 3; |
1033 | eio_dirent piv = *mid; *mid = *L; *L = piv; |
1071 | unsigned char M = 1 << (*bit & 7); |
1034 | |
1072 | |
1035 | while (L < R) |
1073 | eio_dirent *a = base; |
|
|
1074 | eio_dirent *b = end; |
|
|
1075 | |
|
|
1076 | if (b - a < EIO_SORT_CUTOFF) |
|
|
1077 | break; |
|
|
1078 | |
|
|
1079 | /* now bit-partition the array on the bit */ |
|
|
1080 | /* this ugly asymmetric loop seems to perform much better than typical */ |
|
|
1081 | /* partition algos found in the literature */ |
|
|
1082 | do |
|
|
1083 | if (!(((unsigned char *)a)[O] & M)) |
|
|
1084 | ++a; |
|
|
1085 | else if (!(((unsigned char *)--b)[O] & M)) |
1036 | { |
1086 | { |
1037 | while (EIO_DENT_CMP (*R, >=, piv) && L < R) |
1087 | eio_dirent tmp = *a; *a = *b; *b = tmp; |
1038 | --R; |
|
|
1039 | |
|
|
1040 | if (L < R) |
|
|
1041 | *L++ = *R; |
|
|
1042 | |
|
|
1043 | while (EIO_DENT_CMP (*L, <=, piv) && L < R) |
|
|
1044 | ++L; |
1088 | ++a; |
1045 | |
|
|
1046 | if (L < R) |
|
|
1047 | *R-- = *L; |
|
|
1048 | } |
1089 | } |
|
|
1090 | while (b > a); |
1049 | |
1091 | |
1050 | *L = piv; |
1092 | /* next bit, or stop, if no bits left in this path */ |
|
|
1093 | if (!*--bit) |
|
|
1094 | break; |
1051 | |
1095 | |
|
|
1096 | base_stk [stk_idx] = a; |
|
|
1097 | end_stk [stk_idx] = end; |
|
|
1098 | bit_stk [stk_idx] = bit; |
1052 | ++i; |
1099 | ++stk_idx; |
1053 | rng [i].l = L + 1; |
|
|
1054 | rng [i].r = rng [i - 1].r; |
|
|
1055 | rng [i - 1].r = L; |
|
|
1056 | |
1100 | |
1057 | if (rng [i].r - rng [i].l > rng [i - 1].r - rng [i - 1].l) |
1101 | end = a; |
1058 | { |
|
|
1059 | struct rng t; |
|
|
1060 | |
|
|
1061 | t = rng [i]; rng [i] = rng [i - 1]; rng [i - 1] = t; |
|
|
1062 | } |
|
|
1063 | } |
1102 | } |
1064 | else |
|
|
1065 | --i; |
|
|
1066 | } |
1103 | } |
|
|
1104 | while (stk_idx--); |
1067 | } |
1105 | } |
|
|
1106 | } |
1068 | |
1107 | |
1069 | /* use an insertion sort after qsort, or for small arrays */ |
1108 | static void |
|
|
1109 | eio_dent_insertion_sort (eio_dirent *dents, int size) |
|
|
1110 | { |
1070 | /* first move the smallest element to the front, to act as a sentinel */ |
1111 | /* first move the smallest element to the front, to act as a sentinel */ |
1071 | { |
1112 | { |
1072 | int i; |
1113 | int i; |
1073 | eio_dirent *min = dents; |
1114 | eio_dirent *min = dents; |
1074 | |
1115 | |
|
|
1116 | /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */ |
1075 | for (i = size > EIO_QSORT_SKIP ? EIO_QSORT_CUTOFF + 1 : size; --i; ) |
1117 | for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; ) |
1076 | if (EIO_DENT_CMP (dents [i], <, *min)) |
1118 | if (EIO_DENT_CMP (dents [i], <, *min)) |
1077 | min = &dents [i]; |
1119 | min = &dents [i]; |
1078 | |
1120 | |
1079 | /* swap elements 0 and j (minimum) */ |
1121 | /* swap elements 0 and j (minimum) */ |
1080 | { |
1122 | { |
1081 | eio_dirent tmp = *dents; *dents = *min; *min = tmp; |
1123 | eio_dirent tmp = *dents; *dents = *min; *min = tmp; |
1082 | } |
1124 | } |
1083 | } |
1125 | } |
1084 | |
1126 | |
|
|
1127 | /* then do standard insertion sort, assuming that all elements are >= dents [0] */ |
1085 | { |
1128 | { |
1086 | /* then do standard insertion sort */ |
|
|
1087 | eio_dirent *i, *j; |
1129 | eio_dirent *i, *j; |
1088 | |
1130 | |
1089 | for (i = dents + 1; i < dents + size; ++i) |
1131 | for (i = dents + 1; i < dents + size; ++i) |
1090 | { |
1132 | { |
1091 | eio_dirent value = *i; |
1133 | eio_dirent value = *i; |
… | |
… | |
1094 | j [1] = j [0]; |
1136 | j [1] = j [0]; |
1095 | |
1137 | |
1096 | j [1] = value; |
1138 | j [1] = value; |
1097 | } |
1139 | } |
1098 | } |
1140 | } |
|
|
1141 | } |
|
|
1142 | |
|
|
1143 | static void |
|
|
1144 | eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) |
|
|
1145 | { |
|
|
1146 | if (size <= 1) |
|
|
1147 | return; /* our insertion sort relies on size > 0 */ |
|
|
1148 | |
|
|
1149 | /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */ |
|
|
1150 | /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */ |
|
|
1151 | eio_dent_radix_sort (dents, size, score_bits, inode_bits); |
|
|
1152 | |
|
|
1153 | /* use an insertion sort at the end, or for small arrays, */ |
|
|
1154 | /* as insertion sort is more efficient for small partitions */ |
|
|
1155 | eio_dent_insertion_sort (dents, size); |
1099 | } |
1156 | } |
1100 | |
1157 | |
1101 | /* read a full directory */ |
1158 | /* read a full directory */ |
1102 | static void |
1159 | static void |
1103 | eio__scandir (eio_req *req, etp_worker *self) |
1160 | eio__scandir (eio_req *req, etp_worker *self) |
… | |
… | |
1109 | int namesoffs = 0; |
1166 | int namesoffs = 0; |
1110 | int flags = req->int1; |
1167 | int flags = req->int1; |
1111 | eio_dirent *dents = 0; |
1168 | eio_dirent *dents = 0; |
1112 | int dentalloc = 128; |
1169 | int dentalloc = 128; |
1113 | int dentoffs = 0; |
1170 | int dentoffs = 0; |
|
|
1171 | ino_t inode_bits = 0; |
1114 | |
1172 | |
1115 | req->result = -1; |
1173 | req->result = -1; |
1116 | |
1174 | |
1117 | if (!(flags & EIO_READDIR_DENTS)) |
1175 | if (!(flags & EIO_READDIR_DENTS)) |
1118 | flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); |
1176 | flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); |
1119 | |
1177 | |
1120 | X_LOCK (wrklock); |
1178 | X_LOCK (wrklock); |
1121 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
1179 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
1122 | self->dirp = dirp = opendir (req->ptr1); |
1180 | self->dirp = dirp = opendir (req->ptr1); |
1123 | req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; |
1181 | req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; |
1124 | req->ptr1 = names = malloc (namesalloc); |
|
|
1125 | req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; |
1182 | req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; |
|
|
1183 | req->ptr2 = names = malloc (namesalloc); |
1126 | X_UNLOCK (wrklock); |
1184 | X_UNLOCK (wrklock); |
1127 | |
1185 | |
1128 | if (dirp && names && (!flags || dents)) |
1186 | if (dirp && names && (!flags || dents)) |
1129 | for (;;) |
1187 | for (;;) |
1130 | { |
1188 | { |
… | |
… | |
1138 | |
1196 | |
1139 | /* sort etc. */ |
1197 | /* sort etc. */ |
1140 | req->int1 = flags; |
1198 | req->int1 = flags; |
1141 | req->result = dentoffs; |
1199 | req->result = dentoffs; |
1142 | |
1200 | |
1143 | if (dents) |
|
|
1144 | { |
|
|
1145 | eio_dirent *ent = dents + dentoffs; |
|
|
1146 | |
|
|
1147 | while (ent > dents) |
|
|
1148 | (--ent)->name = names + (size_t)ent->name; |
|
|
1149 | } |
|
|
1150 | |
|
|
1151 | if (flags & EIO_READDIR_STAT_ORDER |
1201 | if (flags & EIO_READDIR_STAT_ORDER) |
1152 | || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) |
1202 | eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */ |
1153 | eio_dent_sort (dents, dentoffs); /* score depends of DIRS_FIRST */ |
|
|
1154 | else if (flags & EIO_READDIR_DIRS_FIRST) |
1203 | else if (flags & EIO_READDIR_DIRS_FIRST) |
|
|
1204 | if (flags & EIO_READDIR_FOUND_UNKNOWN) |
|
|
1205 | eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */ |
|
|
1206 | else |
1155 | { |
1207 | { |
1156 | /* in this case, all is known, and we just put dirs first and sort them */ |
1208 | /* in this case, all is known, and we just put dirs first and sort them */ |
1157 | eio_dirent *ent = dents + dentoffs; |
1209 | eio_dirent *oth = dents + dentoffs; |
1158 | eio_dirent *dir = dents; |
1210 | eio_dirent *dir = dents; |
1159 | |
1211 | |
1160 | /* now move dirs to the front, and non-dirs to the back */ |
1212 | /* now partition dirs to the front, and non-dirs to the back */ |
1161 | /* by walking from both sides and swapping if necessary */ |
1213 | /* by walking from both sides and swapping if necessary */ |
|
|
1214 | /* also clear score, so it doesn't influence sorting */ |
1162 | while (ent > dir) |
1215 | while (oth > dir) |
1163 | { |
|
|
1164 | if (dir->type == DT_DIR) |
|
|
1165 | { |
1216 | { |
1166 | dir->score = 0; |
1217 | if (dir->type == EIO_DT_DIR) |
1167 | ++dir; |
1218 | ++dir; |
1168 | } |
|
|
1169 | else |
|
|
1170 | { |
|
|
1171 | --ent; |
|
|
1172 | |
|
|
1173 | if (ent->type == DT_DIR) |
1219 | else if ((--oth)->type == EIO_DT_DIR) |
1174 | { |
1220 | { |
1175 | eio_dirent tmp = *dir; |
1221 | eio_dirent tmp = *dir; *dir = *oth; *oth = tmp; |
1176 | *dir = *ent; |
|
|
1177 | *ent = tmp; |
|
|
1178 | |
1222 | |
1179 | dir->score = 0; |
|
|
1180 | ++dir; |
1223 | ++dir; |
1181 | } |
1224 | } |
1182 | } |
1225 | } |
1183 | } |
|
|
1184 | |
1226 | |
1185 | /* now sort the dirs only */ |
1227 | /* now sort the dirs only */ |
1186 | eio_dent_sort (dents, dir - dents); |
1228 | eio_dent_sort (dents, dir - dents, 0, inode_bits); |
1187 | } |
1229 | } |
1188 | |
|
|
1189 | /* only provide the names array unless DENTS is specified */ |
|
|
1190 | if (!(flags & EIO_READDIR_DENTS)) |
|
|
1191 | { |
|
|
1192 | X_LOCK (wrklock); |
|
|
1193 | assert (!dents); |
|
|
1194 | req->ptr1 = 0; |
|
|
1195 | req->ptr2 = names; |
|
|
1196 | X_UNLOCK (wrklock); |
|
|
1197 | } |
|
|
1198 | |
1230 | |
1199 | break; |
1231 | break; |
1200 | } |
1232 | } |
1201 | |
1233 | |
1202 | /* now add the entry to our list(s) */ |
1234 | /* now add the entry to our list(s) */ |
… | |
… | |
1209 | |
1241 | |
1210 | while (expect_false (namesoffs + len > namesalloc)) |
1242 | while (expect_false (namesoffs + len > namesalloc)) |
1211 | { |
1243 | { |
1212 | namesalloc *= 2; |
1244 | namesalloc *= 2; |
1213 | X_LOCK (wrklock); |
1245 | X_LOCK (wrklock); |
1214 | req->ptr1 = names = realloc (names, namesalloc); |
1246 | req->ptr2 = names = realloc (names, namesalloc); |
1215 | X_UNLOCK (wrklock); |
1247 | X_UNLOCK (wrklock); |
1216 | |
1248 | |
1217 | if (!names) |
1249 | if (!names) |
1218 | break; |
1250 | break; |
1219 | } |
1251 | } |
… | |
… | |
1226 | |
1258 | |
1227 | if (expect_false (dentoffs == dentalloc)) |
1259 | if (expect_false (dentoffs == dentalloc)) |
1228 | { |
1260 | { |
1229 | dentalloc *= 2; |
1261 | dentalloc *= 2; |
1230 | X_LOCK (wrklock); |
1262 | X_LOCK (wrklock); |
1231 | req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); |
1263 | req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); |
1232 | X_UNLOCK (wrklock); |
1264 | X_UNLOCK (wrklock); |
1233 | |
1265 | |
1234 | if (!dents) |
1266 | if (!dents) |
1235 | break; |
1267 | break; |
1236 | } |
1268 | } |
1237 | |
1269 | |
1238 | ent = dents + dentoffs; |
1270 | ent = dents + dentoffs; |
1239 | |
1271 | |
1240 | ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ |
1272 | ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */ |
1241 | ent->namelen = len - 1; |
1273 | ent->namelen = len - 1; |
1242 | ent->inode = D_INO (entp) ? D_INO (entp) : dentoffs; |
1274 | ent->inode = D_INO (entp); |
|
|
1275 | |
|
|
1276 | inode_bits |= ent->inode; |
1243 | |
1277 | |
1244 | switch (D_TYPE (entp)) |
1278 | switch (D_TYPE (entp)) |
1245 | { |
1279 | { |
1246 | default: |
1280 | default: |
1247 | ent->type = EIO_DT_UNKNOWN; |
1281 | ent->type = EIO_DT_UNKNOWN; |
… | |
… | |
1290 | #ifdef DT_WHT |
1324 | #ifdef DT_WHT |
1291 | case DT_WHT: ent->type = EIO_DT_WHT; break; |
1325 | case DT_WHT: ent->type = EIO_DT_WHT; break; |
1292 | #endif |
1326 | #endif |
1293 | } |
1327 | } |
1294 | |
1328 | |
1295 | ent->score = 0; |
1329 | ent->score = 7; |
1296 | |
1330 | |
1297 | if (flags & EIO_READDIR_DIRS_FIRST) |
1331 | if (flags & EIO_READDIR_DIRS_FIRST) |
1298 | { |
1332 | { |
1299 | if (ent->type == EIO_DT_UNKNOWN) |
1333 | if (ent->type == EIO_DT_UNKNOWN) |
1300 | { |
1334 | { |
1301 | if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ |
1335 | if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ |
1302 | ent->score = 7; |
1336 | ent->score = 1; |
1303 | else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ |
1337 | else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ |
1304 | ent->score = len <= 2 ? len + 4 : len <= 4 ? 3 : len <= 7 ? 2 : 1; /* shorter == more likely dir, but avoid too many classes */ |
1338 | ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */ |
1305 | } |
1339 | } |
1306 | else if (ent->type == EIO_DT_DIR) |
1340 | else if (ent->type == EIO_DT_DIR) |
1307 | ent->score = 8; |
1341 | ent->score = 0; |
1308 | } |
1342 | } |
1309 | } |
1343 | } |
1310 | |
1344 | |
1311 | namesoffs += len; |
1345 | namesoffs += len; |
1312 | ++dentoffs; |
1346 | ++dentoffs; |