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

Comparing libeio/eio.c (file contents):
Revision 1.31 by root, Sat Jun 6 17:25:13 2009 UTC vs.
Revision 1.42 by root, Mon Jun 15 05:34:49 2009 UTC

… …
73#else 73#else
74 74
75# include "config.h" 75# include "config.h"
76# include <sys/time.h> 76# include <sys/time.h>
77# include <sys/select.h> 77# include <sys/select.h>
78# include <sys/mman.h>
78# include <unistd.h> 79# include <unistd.h>
79# include <utime.h> 80# include <utime.h>
80# include <signal.h> 81# include <signal.h>
81# include <dirent.h> 82# include <dirent.h>
82 83
83/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ 84/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
84# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) 85# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)
86# define _DIRENT_HAVE_D_TYPE /* sigh */
85# define D_INO(de) (de)->d_fileno 87# define D_INO(de) (de)->d_fileno
86# define _DIRENT_HAVE_D_TYPE /* sigh */ 88# define D_NAMLEN(de) (de)->d_namlen
87# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 89# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600
88# define D_INO(de) (de)->d_ino 90# define D_INO(de) (de)->d_ino
89# endif 91# endif
92
93#ifdef _D_EXACT_NAMLEN
94# undef D_NAMLEN
95# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
96#endif
90 97
91# ifdef _DIRENT_HAVE_D_TYPE 98# ifdef _DIRENT_HAVE_D_TYPE
92# define D_TYPE(de) (de)->d_type 99# define D_TYPE(de) (de)->d_type
93# endif 100# endif
94 101
… …
114#endif 121#endif
115 122
116#ifndef D_TYPE 123#ifndef D_TYPE
117# define D_TYPE(de) 0 124# define D_TYPE(de) 0
118#endif 125#endif
119
120#ifndef D_INO 126#ifndef D_INO
121# define D_INO(de) 0 127# define D_INO(de) 0
128#endif
129#ifndef D_NAMLEN
130# define D_NAMLEN(de) strlen ((de)->d_name)
122#endif 131#endif
123 132
124/* number of seconds after which an idle threads exit */ 133/* number of seconds after which an idle threads exit */
125#define IDLE_TIMEOUT 10 134#define IDLE_TIMEOUT 10
126 135
… …
178 if (wrk->dirp) \ 187 if (wrk->dirp) \
179 { \ 188 { \
180 closedir (wrk->dirp); \ 189 closedir (wrk->dirp); \
181 wrk->dirp = 0; \ 190 wrk->dirp = 0; \
182 } 191 }
192
183#define ETP_WORKER_COMMON \ 193#define ETP_WORKER_COMMON \
184 void *dbuf; \ 194 void *dbuf; \
185 DIR *dirp; 195 DIR *dirp;
186 196
187/*****************************************************************************/ 197/*****************************************************************************/
… …
980 } 990 }
981 991
982 return res; 992 return res;
983} 993}
984 994
985static int 995static signed char
986eio_dent_cmp (const void *a_, const void *b_) 996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
987{ 997{
988 const eio_dirent *a = (const eio_dirent *)a_; 998 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
989 const eio_dirent *b = (const eio_dirent *)b_;
990
991 return (int)b->score - (int)a->score ? (int)b->score - (int)a->score
992 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; /* int might be < ino_t */ 999 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1000}
1001
1002#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1003
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 */
1006
1007static void
1008eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1009{
1010 unsigned char bits [9 + sizeof (ino_t) * 8];
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;
1027 int i, j;
1028
1029 /* we store the byte offset of byte n into byte n of "endianness" */
1030 for (i = 0; i < sizeof (ino_t); ++i)
1031 ((unsigned char *)&endianness)[i] = i;
1032
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) */
1122 {
1123 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1124 }
1125 }
1126
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)
1132 {
1133 eio_dirent value = *i;
1134
1135 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1136 j [1] = j [0];
1137
1138 j [1] = value;
1139 }
1140 }
1141}
1142
1143static void
1144eio_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);
993} 1156}
994 1157
995/* read a full directory */ 1158/* read a full directory */
996static void 1159static void
997eio__scandir (eio_req *req, etp_worker *self) 1160eio__scandir (eio_req *req, etp_worker *self)
998{ 1161{
999 DIR *dirp; 1162 DIR *dirp;
1000 EIO_STRUCT_DIRENT *entp; 1163 EIO_STRUCT_DIRENT *entp;
1001 unsigned char *name, *names; 1164 char *name, *names;
1002 int namesalloc = 4096; 1165 int namesalloc = 4096;
1003 int namesoffs = 0; 1166 int namesoffs = 0;
1004 int flags = req->int1; 1167 int flags = req->int1;
1005 eio_dirent *dents = 0; 1168 eio_dirent *dents = 0;
1006 int dentalloc = 128; 1169 int dentalloc = 128;
1007 int dentoffs = 0; 1170 int dentoffs = 0;
1171 ino_t inode_bits = 0;
1008 1172
1009 req->result = -1; 1173 req->result = -1;
1010 1174
1011 if (!(flags & EIO_READDIR_DENTS)) 1175 if (!(flags & EIO_READDIR_DENTS))
1012 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1176 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1013 1177
1014 X_LOCK (wrklock); 1178 X_LOCK (wrklock);
1015
1016 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1179 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1017 self->dirp = dirp = opendir (req->ptr1); 1180 self->dirp = dirp = opendir (req->ptr1);
1018 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1181 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1019 req->ptr1 = names = malloc (namesalloc);
1020 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1182 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1021 1183 req->ptr2 = names = malloc (namesalloc);
1022 X_UNLOCK (wrklock); 1184 X_UNLOCK (wrklock);
1023 1185
1024 if (dirp && names && (!flags || dents)) 1186 if (dirp && names && (!flags || dents))
1025 for (;;) 1187 for (;;)
1026 { 1188 {
… …
1034 1196
1035 /* sort etc. */ 1197 /* sort etc. */
1036 req->int1 = flags; 1198 req->int1 = flags;
1037 req->result = dentoffs; 1199 req->result = dentoffs;
1038 1200
1039 if (flags & EIO_READDIR_STAT_ORDER || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN)) 1201 if (flags & EIO_READDIR_STAT_ORDER)
1202 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
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
1040 { 1207 {
1041 /* pray your qsort doesn't use quicksort */
1042 qsort (dents, dentoffs, sizeof (*dents), eio_dent_cmp); /* score depends of DIRS_FIRST */
1043 }
1044 else if (flags & EIO_READDIR_DIRS_FIRST && !(flags & EIO_READDIR_FOUND_UNKNOWN))
1045 {
1046 /* 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 */
1047 eio_dirent *ent = dents + dentoffs; 1209 eio_dirent *oth = dents + dentoffs;
1048 eio_dirent *dir = dents; 1210 eio_dirent *dir = dents;
1049 1211
1212 /* now partition dirs to the front, and non-dirs to the back */
1213 /* by walking from both sides and swapping if necessary */
1214 /* also clear score, so it doesn't influence sorting */
1050 while (ent > dir) 1215 while (oth > dir)
1051 { 1216 {
1052 if (dir->type == DT_DIR) 1217 if (dir->type == EIO_DT_DIR)
1053 ++dir; 1218 ++dir;
1054 else
1055 {
1056 --ent;
1057
1058 if (ent->type == DT_DIR) 1219 else if ((--oth)->type == EIO_DT_DIR)
1059 { 1220 {
1060 eio_dirent tmp = *dir; 1221 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1061 *dir = *ent;
1062 *ent = tmp;
1063 1222
1064 ++dir; 1223 ++dir;
1065 } 1224 }
1066 } 1225 }
1067 }
1068 1226
1069 /* now sort the dirs only */ 1227 /* now sort the dirs only */
1070 qsort (dents, dir - dents, sizeof (*dents), eio_dent_cmp); 1228 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1071 } 1229 }
1072
1073 {int i; for(i=0;i<dentoffs;++i){eio_dirent *e=dents+i; printf ("%9ld %3d %s\n", e->inode,e->score,e->name);}}//D
1074 1230
1075 break; 1231 break;
1076 } 1232 }
1077 1233
1234 /* now add the entry to our list(s) */
1078 name = entp->d_name; 1235 name = entp->d_name;
1079 1236
1237 /* skip . and .. entries */
1080 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) 1238 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
1081 { 1239 {
1082 int len = strlen (name) + 1; 1240 int len = D_NAMLEN (entp) + 1;
1083 1241
1084 while (expect_false (namesoffs + len > namesalloc)) 1242 while (expect_false (namesoffs + len > namesalloc))
1085 { 1243 {
1086 namesalloc *= 2; 1244 namesalloc *= 2;
1087 X_LOCK (wrklock); 1245 X_LOCK (wrklock);
1088 req->ptr1 = names = realloc (names, namesalloc); 1246 req->ptr2 = names = realloc (names, namesalloc);
1089 X_UNLOCK (wrklock); 1247 X_UNLOCK (wrklock);
1090 1248
1091 if (!names) 1249 if (!names)
1092 break; 1250 break;
1093 } 1251 }
… …
1100 1258
1101 if (expect_false (dentoffs == dentalloc)) 1259 if (expect_false (dentoffs == dentalloc))
1102 { 1260 {
1103 dentalloc *= 2; 1261 dentalloc *= 2;
1104 X_LOCK (wrklock); 1262 X_LOCK (wrklock);
1105 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1263 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1106 X_UNLOCK (wrklock); 1264 X_UNLOCK (wrklock);
1107 1265
1108 if (!dents) 1266 if (!dents)
1109 break; 1267 break;
1110 } 1268 }
1111 1269
1112 ent = dents + dentoffs; 1270 ent = dents + dentoffs;
1113 1271
1114 ent->name = names + namesoffs; 1272 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1115 ent->namelen = len - 1; 1273 ent->namelen = len - 1;
1116 ent->inode = D_INO (entp); 1274 ent->inode = D_INO (entp);
1275
1276 inode_bits |= ent->inode;
1117 1277
1118 switch (D_TYPE (entp)) 1278 switch (D_TYPE (entp))
1119 { 1279 {
1120 default: 1280 default:
1121 ent->type = EIO_DT_UNKNOWN; 1281 ent->type = EIO_DT_UNKNOWN;
… …
1126 case DT_FIFO: ent->type = EIO_DT_FIFO; break; 1286 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1127 #endif 1287 #endif
1128 #ifdef DT_CHR 1288 #ifdef DT_CHR
1129 case DT_CHR: ent->type = EIO_DT_CHR; break; 1289 case DT_CHR: ent->type = EIO_DT_CHR; break;
1130 #endif 1290 #endif
1291 #ifdef DT_MPC
1292 case DT_MPC: ent->type = EIO_DT_MPC; break;
1293 #endif
1131 #ifdef DT_DIR 1294 #ifdef DT_DIR
1132 case DT_DIR: ent->type = EIO_DT_DIR; break; 1295 case DT_DIR: ent->type = EIO_DT_DIR; break;
1133 #endif 1296 #endif
1297 #ifdef DT_NAM
1298 case DT_NAM: ent->type = EIO_DT_NAM; break;
1299 #endif
1134 #ifdef DT_BLK 1300 #ifdef DT_BLK
1135 case DT_BLK: ent->type = EIO_DT_BLK; break; 1301 case DT_BLK: ent->type = EIO_DT_BLK; break;
1136 #endif 1302 #endif
1303 #ifdef DT_MPB
1304 case DT_MPB: ent->type = EIO_DT_MPB; break;
1305 #endif
1137 #ifdef DT_REG 1306 #ifdef DT_REG
1138 case DT_REG: ent->type = EIO_DT_REG; break; 1307 case DT_REG: ent->type = EIO_DT_REG; break;
1139 #endif 1308 #endif
1309 #ifdef DT_NWK
1310 case DT_NWK: ent->type = EIO_DT_NWK; break;
1311 #endif
1312 #ifdef DT_CMP
1313 case DT_CMP: ent->type = EIO_DT_CMP; break;
1314 #endif
1140 #ifdef DT_LNK 1315 #ifdef DT_LNK
1141 case DT_LNK: ent->type = EIO_DT_LNK; break; 1316 case DT_LNK: ent->type = EIO_DT_LNK; break;
1142 #endif 1317 #endif
1143 #ifdef DT_SOCK 1318 #ifdef DT_SOCK
1144 case DT_SOCK: ent->type = EIO_DT_SOCK; break; 1319 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1145 #endif 1320 #endif
1321 #ifdef DT_DOOR
1322 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1323 #endif
1146 #ifdef DT_WHT 1324 #ifdef DT_WHT
1147 case DT_WHT: ent->type = EIO_DT_WHT; break; 1325 case DT_WHT: ent->type = EIO_DT_WHT; break;
1148 #endif 1326 #endif
1149 } 1327 }
1150 1328
1151 ent->score = 0; 1329 ent->score = 7;
1152 1330
1153 if (flags & EIO_READDIR_DIRS_FIRST) 1331 if (flags & EIO_READDIR_DIRS_FIRST)
1154 { 1332 {
1155 if (ent->type == EIO_DT_UNKNOWN) 1333 if (ent->type == EIO_DT_UNKNOWN)
1156 { 1334 {
1157 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 */
1158 ent->score = 98; 1336 ent->score = 1;
1159 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1337 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1160 ent->score = 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 */
1161 } 1339 }
1162 else if (ent->type == DT_DIR) 1340 else if (ent->type == EIO_DT_DIR)
1163 ent->score = 100; 1341 ent->score = 0;
1164 } 1342 }
1165 } 1343 }
1166 1344
1167 namesoffs += len; 1345 namesoffs += len;
1168 ++dentoffs; 1346 ++dentoffs;
1169 } 1347 }
1348
1349 if (EIO_CANCELLED (req))
1350 {
1351 errno = ECANCELED;
1352 break;
1353 }
1170 } 1354 }
1171 else
1172 req->result = -1;
1173
1174 /* if user doesn't want the dents, do not provide it */
1175 if (!(flags & EIO_READDIR_DENTS))
1176 {
1177 X_LOCK (wrklock);
1178 free (dents);
1179 req->ptr2 = req->ptr1;
1180 req->ptr1 = 0;
1181 X_UNLOCK (wrklock);
1182 }
1183} 1355}
1184 1356
1185#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1357#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1186# undef msync 1358# undef msync
1187# define msync(a,b,c) ((errno = ENOSYS), -1) 1359# define msync(a,b,c) ((errno = ENOSYS), -1)
… …
1708/* misc garbage */ 1880/* misc garbage */
1709 1881
1710ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) 1882ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1711{ 1883{
1712 etp_worker wrk; 1884 etp_worker wrk;
1885 ssize_t ret;
1713 1886
1714 wrk.dbuf = 0; 1887 wrk.dbuf = 0;
1715 1888
1716 eio__sendfile (ofd, ifd, offset, count, &wrk); 1889 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1717 1890
1718 if (wrk.dbuf) 1891 if (wrk.dbuf)
1719 free (wrk.dbuf); 1892 free (wrk.dbuf);
1720}
1721 1893
1894 return ret;
1895}
1896

Diff Legend

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