ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/cvsroot/libeio/eio.c
(Generate patch)

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.37 by root, Fri Jun 12 16:48:08 2009 UTC vs.
Revision 1.44 by root, Thu Nov 26 05:25:35 2009 UTC

… …
196 196
197/*****************************************************************************/ 197/*****************************************************************************/
198 198
199#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 199#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
200 200
201/* calculcate time difference in ~1/EIO_TICKS of a second */ 201/* calculate time difference in ~1/EIO_TICKS of a second */
202static int tvdiff (struct timeval *tv1, struct timeval *tv2) 202static int tvdiff (struct timeval *tv1, struct timeval *tv2)
203{ 203{
204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
205 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 205 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
206} 206}
… …
598} 598}
599 599
600static void etp_set_max_poll_time (double nseconds) 600static void etp_set_max_poll_time (double nseconds)
601{ 601{
602 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 602 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
603 max_poll_time = nseconds; 603 max_poll_time = nseconds * EIO_TICKS;
604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
605} 605}
606 606
607static void etp_set_max_poll_reqs (unsigned int maxreqs) 607static void etp_set_max_poll_reqs (unsigned int maxreqs)
608{ 608{
… …
867 if (!res || errno != ENOSYS) 867 if (!res || errno != ENOSYS)
868 return res; 868 return res;
869#endif 869#endif
870 870
871 /* even though we could play tricks with the flags, it's better to always 871 /* even though we could play tricks with the flags, it's better to always
872 * call fdatasync, as thta matches the expectation of it's users best */ 872 * call fdatasync, as that matches the expectation of its users best */
873 return fdatasync (fd); 873 return fdatasync (fd);
874} 874}
875 875
876#if !HAVE_READAHEAD 876#if !HAVE_READAHEAD
877# undef readahead 877# undef readahead
… …
993} 993}
994 994
995static signed char 995static signed char
996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) 996eio_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
1003
1002#define EIO_QSORT_CUTOFF 20 /* quite high, but performs well on many filesystems */ 1004#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1005#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1003 1006
1004static void 1007static void
1005eio_dent_sort (eio_dirent *dents, int size) 1008eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1006{ 1009{
1007 /* should be good for 2**31 entries */ 1010 unsigned char bits [9 + sizeof (ino_t) * 8];
1008 struct rng { int l, r; } rng [32]; 1011 unsigned char *bit = bits;
1012
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 */
1025 {
1026 ino_t endianness;
1009 int i, j; 1027 int i, j;
1010 1028
1011 i = 0; 1029 /* we store the byte offset of byte n into byte n of "endianness" */
1012 rng[0].l = 0; 1030 for (i = 0; i < sizeof (ino_t); ++i)
1013 rng[0].r = size; 1031 ((unsigned char *)&endianness)[i] = i;
1014 1032
1015 while (expect_true (i >= 0)) 1033 *bit++ = 0;
1034
1035 for (i = 0; i < sizeof (ino_t); ++i)
1036 {
1037 /* shifting off the byte offsets out of "endianness" */
1038 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1039 endianness >>= 8;
1040
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 (;;)
1069 {
1070 unsigned char O = *bit >> 3;
1071 unsigned char M = 1 << (*bit & 7);
1072
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))
1086 {
1087 eio_dirent tmp = *a; *a = *b; *b = tmp;
1088 ++a;
1089 }
1090 while (b > a);
1091
1092 /* next bit, or stop, if no bits left in this path */
1093 if (!*--bit)
1094 break;
1095
1096 base_stk [stk_idx] = a;
1097 end_stk [stk_idx] = end;
1098 bit_stk [stk_idx] = bit;
1099 ++stk_idx;
1100
1101 end = a;
1102 }
1103 }
1104 while (stk_idx--);
1105 }
1106}
1107
1108static void
1109eio_dent_insertion_sort (eio_dirent *dents, int size)
1110{
1111 /* first move the smallest element to the front, to act as a sentinel */
1112 {
1113 int i;
1114 eio_dirent *min = dents;
1115
1116 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1117 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1118 if (EIO_DENT_CMP (dents [i], <, *min))
1119 min = &dents [i];
1120
1121 /* swap elements 0 and j (minimum) */
1016 { 1122 {
1017 int L = rng [i].l; 1123 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1018 int R = rng [i].r - 1; 1124 }
1125 }
1019 1126
1020 if (expect_false (L + EIO_QSORT_CUTOFF < R)) 1127 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1128 {
1129 eio_dirent *i, *j;
1130
1131 for (i = dents + 1; i < dents + size; ++i)
1021 { 1132 {
1022 eio_dirent piv = dents [L]; 1133 eio_dirent value = *i;
1023 1134
1024 while (L < R) 1135 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1025 { 1136 j [1] = j [0];
1026 while (eio_dent_cmp (&dents [R], &piv) >= 0 && L < R)
1027 --R;
1028 1137
1029 if (L < R) 1138 j [1] = value;
1030 dents [L++] = dents [R];
1031
1032 while (eio_dent_cmp (&dents [L], &piv) <= 0 && L < R)
1033 ++L;
1034
1035 if (L < R)
1036 dents [R--] = dents [L];
1037 }
1038
1039 dents [L] = piv;
1040
1041 ++i;
1042 rng [i].l = L + 1;
1043 rng [i].r = rng [i - 1].r;
1044 rng [i - 1].r = L;
1045
1046 if (rng [i].r - rng [i].l > rng [i - 1].r - rng [i - 1].l)
1047 {
1048 struct rng t;
1049
1050 t = rng [i]; rng [i] = rng [i - 1]; rng [i - 1] = t;
1051 }
1052 } 1139 }
1053 else
1054 --i;
1055 } 1140 }
1141}
1056 1142
1057 /* use a simple insertion sort at the end */ 1143static void
1058 for (i = 1; i < size; ++i) 1144eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1059 { 1145{
1060 eio_dirent value = dents [i]; 1146 if (size <= 1)
1147 return; /* our insertion sort relies on size > 0 */
1061 1148
1062 for (j = i - 1; j >= 0 && eio_dent_cmp (&dents [j], &value) > 0; --j) 1149 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1063 dents [j + 1] = dents [j]; 1150 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1151 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1064 1152
1065 dents [j + 1] = value; 1153 /* use an insertion sort at the end, or for small arrays, */
1066 } 1154 /* as insertion sort is more efficient for small partitions */
1155 eio_dent_insertion_sort (dents, size);
1067} 1156}
1068 1157
1069/* read a full directory */ 1158/* read a full directory */
1070static void 1159static void
1071eio__scandir (eio_req *req, etp_worker *self) 1160eio__scandir (eio_req *req, etp_worker *self)
… …
1077 int namesoffs = 0; 1166 int namesoffs = 0;
1078 int flags = req->int1; 1167 int flags = req->int1;
1079 eio_dirent *dents = 0; 1168 eio_dirent *dents = 0;
1080 int dentalloc = 128; 1169 int dentalloc = 128;
1081 int dentoffs = 0; 1170 int dentoffs = 0;
1171 ino_t inode_bits = 0;
1082 1172
1083 req->result = -1; 1173 req->result = -1;
1084 1174
1085 if (!(flags & EIO_READDIR_DENTS)) 1175 if (!(flags & EIO_READDIR_DENTS))
1086 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1176 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1087 1177
1088 X_LOCK (wrklock); 1178 X_LOCK (wrklock);
1089 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1179 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1090 self->dirp = dirp = opendir (req->ptr1); 1180 self->dirp = dirp = opendir (req->ptr1);
1091 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1181 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1092 req->ptr1 = names = malloc (namesalloc);
1093 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);
1094 X_UNLOCK (wrklock); 1184 X_UNLOCK (wrklock);
1095 1185
1096 if (dirp && names && (!flags || dents)) 1186 if (dirp && names && (!flags || dents))
1097 for (;;) 1187 for (;;)
1098 { 1188 {
… …
1106 1196
1107 /* sort etc. */ 1197 /* sort etc. */
1108 req->int1 = flags; 1198 req->int1 = flags;
1109 req->result = dentoffs; 1199 req->result = dentoffs;
1110 1200
1111 if (dents)
1112 {
1113 eio_dirent *ent = dents + dentoffs;
1114
1115 while (ent > dents)
1116 (--ent)->name = names + (size_t)ent->name;
1117 }
1118
1119 if (flags & EIO_READDIR_STAT_ORDER 1201 if (flags & EIO_READDIR_STAT_ORDER)
1120 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1202 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1121 eio_dent_sort (dents, dentoffs); /* score depends of DIRS_FIRST */
1122 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
1123 { 1207 {
1124 /* 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 */
1125 eio_dirent *ent = dents + dentoffs; 1209 eio_dirent *oth = dents + dentoffs;
1126 eio_dirent *dir = dents; 1210 eio_dirent *dir = dents;
1127 1211
1128 /* 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 */
1129 /* 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 */
1130 while (ent > dir) 1215 while (oth > dir)
1131 { 1216 {
1132 if (dir->type == DT_DIR) 1217 if (dir->type == EIO_DT_DIR)
1133 ++dir; 1218 ++dir;
1134 else
1135 {
1136 --ent;
1137
1138 if (ent->type == DT_DIR) 1219 else if ((--oth)->type == EIO_DT_DIR)
1139 { 1220 {
1140 eio_dirent tmp = *dir; 1221 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1141 *dir = *ent;
1142 *ent = tmp;
1143 1222
1144 ++dir; 1223 ++dir;
1145 } 1224 }
1146 } 1225 }
1147 }
1148 1226
1149 /* now sort the dirs only */ 1227 /* now sort the dirs only */
1150 eio_dent_sort (dents, dir - dents); 1228 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1151 } 1229 }
1152
1153 /* only provide the names array unless DENTS is specified */
1154 if (!(flags & EIO_READDIR_DENTS))
1155 {
1156 X_LOCK (wrklock);
1157 assert (!dents);
1158 req->ptr1 = 0;
1159 req->ptr2 = names;
1160 X_UNLOCK (wrklock);
1161 }
1162 1230
1163 break; 1231 break;
1164 } 1232 }
1165 1233
1166 /* now add the entry to our list(s) */ 1234 /* now add the entry to our list(s) */
… …
1173 1241
1174 while (expect_false (namesoffs + len > namesalloc)) 1242 while (expect_false (namesoffs + len > namesalloc))
1175 { 1243 {
1176 namesalloc *= 2; 1244 namesalloc *= 2;
1177 X_LOCK (wrklock); 1245 X_LOCK (wrklock);
1178 req->ptr1 = names = realloc (names, namesalloc); 1246 req->ptr2 = names = realloc (names, namesalloc);
1179 X_UNLOCK (wrklock); 1247 X_UNLOCK (wrklock);
1180 1248
1181 if (!names) 1249 if (!names)
1182 break; 1250 break;
1183 } 1251 }
… …
1190 1258
1191 if (expect_false (dentoffs == dentalloc)) 1259 if (expect_false (dentoffs == dentalloc))
1192 { 1260 {
1193 dentalloc *= 2; 1261 dentalloc *= 2;
1194 X_LOCK (wrklock); 1262 X_LOCK (wrklock);
1195 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1263 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1196 X_UNLOCK (wrklock); 1264 X_UNLOCK (wrklock);
1197 1265
1198 if (!dents) 1266 if (!dents)
1199 break; 1267 break;
1200 } 1268 }
1201 1269
1202 ent = dents + dentoffs; 1270 ent = dents + dentoffs;
1203 1271
1204 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 */
1205 ent->namelen = len - 1; 1273 ent->namelen = len - 1;
1206 ent->inode = D_INO (entp); 1274 ent->inode = D_INO (entp);
1275
1276 inode_bits |= ent->inode;
1207 1277
1208 switch (D_TYPE (entp)) 1278 switch (D_TYPE (entp))
1209 { 1279 {
1210 default: 1280 default:
1211 ent->type = EIO_DT_UNKNOWN; 1281 ent->type = EIO_DT_UNKNOWN;
… …
1254 #ifdef DT_WHT 1324 #ifdef DT_WHT
1255 case DT_WHT: ent->type = EIO_DT_WHT; break; 1325 case DT_WHT: ent->type = EIO_DT_WHT; break;
1256 #endif 1326 #endif
1257 } 1327 }
1258 1328
1259 ent->score = 0; 1329 ent->score = 7;
1260 1330
1261 if (flags & EIO_READDIR_DIRS_FIRST) 1331 if (flags & EIO_READDIR_DIRS_FIRST)
1262 { 1332 {
1263 if (ent->type == EIO_DT_UNKNOWN) 1333 if (ent->type == EIO_DT_UNKNOWN)
1264 { 1334 {
1265 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 */
1266 ent->score = 98; 1336 ent->score = 1;
1267 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1337 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1268 ent->score = len <= 2 ? len + 6 : len <= 4 ? 5 : len <= 7 ? 4 : 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 */
1269 } 1339 }
1270 else if (ent->type == EIO_DT_DIR) 1340 else if (ent->type == EIO_DT_DIR)
1271 ent->score = 100; 1341 ent->score = 0;
1272 } 1342 }
1273 } 1343 }
1274 1344
1275 namesoffs += len; 1345 namesoffs += len;
1276 ++dentoffs; 1346 ++dentoffs;

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines