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

Comparing libeio/eio.c (file contents):
Revision 1.27 by root, Wed Oct 22 18:15:36 2008 UTC vs.
Revision 1.42 by root, Mon Jun 15 05:34:49 2009 UTC

1/* 1/*
2 * libeio implementation 2 * libeio implementation
3 * 3 *
4 * Copyright (c) 2007,2008 Marc Alexander Lehmann <libeio@schmorp.de> 4 * Copyright (c) 2007,2008,2009 Marc Alexander Lehmann <libeio@schmorp.de>
5 * All rights reserved. 5 * All rights reserved.
6 * 6 *
7 * Redistribution and use in source and binary forms, with or without modifica- 7 * Redistribution and use in source and binary forms, with or without modifica-
8 * tion, are permitted provided that the following conditions are met: 8 * tion, are permitted provided that the following conditions are met:
9 * 9 *
… …
36 * provisions above, a recipient may use your version of this file under 36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL. 37 * either the BSD or the GPL.
38 */ 38 */
39 39
40#include "eio.h" 40#include "eio.h"
41
42#ifdef EIO_STACKSIZE
43# define XTHREAD_STACKSIZE EIO_STACKSIZE
44#endif
41#include "xthread.h" 45#include "xthread.h"
42 46
43#include <errno.h> 47#include <errno.h>
44#include <stddef.h> 48#include <stddef.h>
45#include <stdlib.h> 49#include <stdlib.h>
… …
64#endif 68#endif
65 69
66#ifdef _WIN32 70#ifdef _WIN32
67 71
68 /*doh*/ 72 /*doh*/
69
70#else 73#else
71 74
72# include "config.h" 75# include "config.h"
73# include <sys/time.h> 76# include <sys/time.h>
74# include <sys/select.h> 77# include <sys/select.h>
78# include <sys/mman.h>
75# include <unistd.h> 79# include <unistd.h>
76# include <utime.h> 80# include <utime.h>
77# include <signal.h> 81# include <signal.h>
78# include <dirent.h> 82# include <dirent.h>
83
84/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
85# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)
86# define _DIRENT_HAVE_D_TYPE /* sigh */
87# define D_INO(de) (de)->d_fileno
88# define D_NAMLEN(de) (de)->d_namlen
89# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600
90# define D_INO(de) (de)->d_ino
91# endif
92
93#ifdef _D_EXACT_NAMLEN
94# undef D_NAMLEN
95# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
96#endif
97
98# ifdef _DIRENT_HAVE_D_TYPE
99# define D_TYPE(de) (de)->d_type
100# endif
79 101
80# ifndef EIO_STRUCT_DIRENT 102# ifndef EIO_STRUCT_DIRENT
81# define EIO_STRUCT_DIRENT struct dirent 103# define EIO_STRUCT_DIRENT struct dirent
82# endif 104# endif
83 105
… …
96# else 118# else
97# error sendfile support requested but not available 119# error sendfile support requested but not available
98# endif 120# endif
99#endif 121#endif
100 122
123#ifndef D_TYPE
124# define D_TYPE(de) 0
125#endif
126#ifndef D_INO
127# define D_INO(de) 0
128#endif
129#ifndef D_NAMLEN
130# define D_NAMLEN(de) strlen ((de)->d_name)
131#endif
132
101/* number of seconds after which an idle threads exit */ 133/* number of seconds after which an idle threads exit */
102#define IDLE_TIMEOUT 10 134#define IDLE_TIMEOUT 10
103 135
104/* used for struct dirent, AIX doesn't provide it */ 136/* used for struct dirent, AIX doesn't provide it */
105#ifndef NAME_MAX 137#ifndef NAME_MAX
… …
155 if (wrk->dirp) \ 187 if (wrk->dirp) \
156 { \ 188 { \
157 closedir (wrk->dirp); \ 189 closedir (wrk->dirp); \
158 wrk->dirp = 0; \ 190 wrk->dirp = 0; \
159 } 191 }
192
160#define ETP_WORKER_COMMON \ 193#define ETP_WORKER_COMMON \
161 void *dbuf; \ 194 void *dbuf; \
162 DIR *dirp; 195 DIR *dirp;
163 196
164/*****************************************************************************/ 197/*****************************************************************************/
… …
767 800
768 X_LOCK (preadwritelock); 801 X_LOCK (preadwritelock);
769 ooffset = lseek (fd, 0, SEEK_CUR); 802 ooffset = lseek (fd, 0, SEEK_CUR);
770 lseek (fd, offset, SEEK_SET); 803 lseek (fd, offset, SEEK_SET);
771 res = write (fd, buf, count); 804 res = write (fd, buf, count);
772 lseek (fd, offset, SEEK_SET); 805 lseek (fd, ooffset, SEEK_SET);
773 X_UNLOCK (preadwritelock); 806 X_UNLOCK (preadwritelock);
774 807
775 return res; 808 return res;
776} 809}
777#endif 810#endif
… …
829 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0); 862 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
830 } 863 }
831 864
832 res = sync_file_range (fd, offset, nbytes, flags); 865 res = sync_file_range (fd, offset, nbytes, flags);
833 866
834 if (res != ENOSYS) 867 if (!res || errno != ENOSYS)
835 return res; 868 return res;
836#endif 869#endif
837 870
838 /* 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
839 * call fdatasync, as thta matches the expectation of it's users best */ 872 * call fdatasync, as thta matches the expectation of it's users best */
… …
957 } 990 }
958 991
959 return res; 992 return res;
960} 993}
961 994
995static signed char
996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
997{
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;
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);
1156}
1157
962/* read a full directory */ 1158/* read a full directory */
963static void 1159static void
964eio__scandir (eio_req *req, etp_worker *self) 1160eio__scandir (eio_req *req, etp_worker *self)
965{ 1161{
966 DIR *dirp; 1162 DIR *dirp;
967 EIO_STRUCT_DIRENT *entp; 1163 EIO_STRUCT_DIRENT *entp;
968 char *name, *names; 1164 char *name, *names;
969 int memlen = 4096; 1165 int namesalloc = 4096;
970 int memofs = 0; 1166 int namesoffs = 0;
1167 int flags = req->int1;
1168 eio_dirent *dents = 0;
1169 int dentalloc = 128;
971 int res = 0; 1170 int dentoffs = 0;
1171 ino_t inode_bits = 0;
1172
1173 req->result = -1;
1174
1175 if (!(flags & EIO_READDIR_DENTS))
1176 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
972 1177
973 X_LOCK (wrklock); 1178 X_LOCK (wrklock);
974 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1179 /* the corresponding closedir is in ETP_WORKER_CLEAR */
975 self->dirp = dirp = opendir (req->ptr1); 1180 self->dirp = dirp = opendir (req->ptr1);
976 req->flags |= EIO_FLAG_PTR2_FREE; 1181 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1182 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
977 req->ptr2 = names = malloc (memlen); 1183 req->ptr2 = names = malloc (namesalloc);
978 X_UNLOCK (wrklock); 1184 X_UNLOCK (wrklock);
979 1185
980 if (dirp && names) 1186 if (dirp && names && (!flags || dents))
981 for (;;) 1187 for (;;)
982 { 1188 {
983 errno = 0; 1189 errno = 0;
984 entp = readdir (dirp); 1190 entp = readdir (dirp);
985 1191
986 if (!entp) 1192 if (!entp)
1193 {
1194 if (errno)
1195 break;
1196
1197 /* sort etc. */
1198 req->int1 = flags;
1199 req->result = dentoffs;
1200
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
1207 {
1208 /* in this case, all is known, and we just put dirs first and sort them */
1209 eio_dirent *oth = dents + dentoffs;
1210 eio_dirent *dir = dents;
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 */
1215 while (oth > dir)
1216 {
1217 if (dir->type == EIO_DT_DIR)
1218 ++dir;
1219 else if ((--oth)->type == EIO_DT_DIR)
1220 {
1221 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1222
1223 ++dir;
1224 }
1225 }
1226
1227 /* now sort the dirs only */
1228 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1229 }
1230
987 break; 1231 break;
1232 }
988 1233
1234 /* now add the entry to our list(s) */
989 name = entp->d_name; 1235 name = entp->d_name;
990 1236
1237 /* skip . and .. entries */
991 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) 1238 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
992 { 1239 {
993 int len = strlen (name) + 1; 1240 int len = D_NAMLEN (entp) + 1;
994 1241
995 res++; 1242 while (expect_false (namesoffs + len > namesalloc))
996
997 while (memofs + len > memlen)
998 { 1243 {
999 memlen *= 2; 1244 namesalloc *= 2;
1000 X_LOCK (wrklock); 1245 X_LOCK (wrklock);
1001 req->ptr2 = names = realloc (names, memlen); 1246 req->ptr2 = names = realloc (names, namesalloc);
1002 X_UNLOCK (wrklock); 1247 X_UNLOCK (wrklock);
1003 1248
1004 if (!names) 1249 if (!names)
1005 break; 1250 break;
1006 } 1251 }
1007 1252
1008 memcpy (names + memofs, name, len); 1253 memcpy (names + namesoffs, name, len);
1254
1255 if (dents)
1256 {
1257 struct eio_dirent *ent;
1258
1259 if (expect_false (dentoffs == dentalloc))
1260 {
1261 dentalloc *= 2;
1262 X_LOCK (wrklock);
1263 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1264 X_UNLOCK (wrklock);
1265
1266 if (!dents)
1267 break;
1268 }
1269
1270 ent = dents + dentoffs;
1271
1272 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1273 ent->namelen = len - 1;
1274 ent->inode = D_INO (entp);
1275
1276 inode_bits |= ent->inode;
1277
1278 switch (D_TYPE (entp))
1279 {
1280 default:
1281 ent->type = EIO_DT_UNKNOWN;
1282 flags |= EIO_READDIR_FOUND_UNKNOWN;
1283 break;
1284
1285 #ifdef DT_FIFO
1286 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1287 #endif
1288 #ifdef DT_CHR
1289 case DT_CHR: ent->type = EIO_DT_CHR; break;
1290 #endif
1291 #ifdef DT_MPC
1292 case DT_MPC: ent->type = EIO_DT_MPC; break;
1293 #endif
1294 #ifdef DT_DIR
1295 case DT_DIR: ent->type = EIO_DT_DIR; break;
1296 #endif
1297 #ifdef DT_NAM
1298 case DT_NAM: ent->type = EIO_DT_NAM; break;
1299 #endif
1300 #ifdef DT_BLK
1301 case DT_BLK: ent->type = EIO_DT_BLK; break;
1302 #endif
1303 #ifdef DT_MPB
1304 case DT_MPB: ent->type = EIO_DT_MPB; break;
1305 #endif
1306 #ifdef DT_REG
1307 case DT_REG: ent->type = EIO_DT_REG; break;
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
1315 #ifdef DT_LNK
1316 case DT_LNK: ent->type = EIO_DT_LNK; break;
1317 #endif
1318 #ifdef DT_SOCK
1319 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1320 #endif
1321 #ifdef DT_DOOR
1322 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1323 #endif
1324 #ifdef DT_WHT
1325 case DT_WHT: ent->type = EIO_DT_WHT; break;
1326 #endif
1327 }
1328
1329 ent->score = 7;
1330
1331 if (flags & EIO_READDIR_DIRS_FIRST)
1332 {
1333 if (ent->type == EIO_DT_UNKNOWN)
1334 {
1335 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1336 ent->score = 1;
1337 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1338 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1339 }
1340 else if (ent->type == EIO_DT_DIR)
1341 ent->score = 0;
1342 }
1343 }
1344
1009 memofs += len; 1345 namesoffs += len;
1346 ++dentoffs;
1347 }
1348
1349 if (EIO_CANCELLED (req))
1350 {
1351 errno = ECANCELED;
1352 break;
1010 } 1353 }
1011 } 1354 }
1012
1013 if (errno)
1014 res = -1;
1015
1016 req->result = res;
1017} 1355}
1018 1356
1019#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1357#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1020# undef msync 1358# undef msync
1021# define msync(a,b,c) ENOSYS 1359# define msync(a,b,c) ((errno = ENOSYS), -1)
1022#endif 1360#endif
1023 1361
1024int 1362int
1025eio__mtouch (void *mem, size_t len, int flags) 1363eio__mtouch (void *mem, size_t len, int flags)
1026{ 1364{
… …
1441eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data) 1779eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1442{ 1780{
1443 return eio__1path (EIO_RMDIR, path, pri, cb, data); 1781 return eio__1path (EIO_RMDIR, path, pri, cb, data);
1444} 1782}
1445 1783
1446eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data) 1784eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1447{ 1785{
1448 return eio__1path (EIO_READDIR, path, pri, cb, data); 1786 REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1449} 1787}
1450 1788
1451eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) 1789eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1452{ 1790{
1453 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND; 1791 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
… …
1542/* misc garbage */ 1880/* misc garbage */
1543 1881
1544ssize_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)
1545{ 1883{
1546 etp_worker wrk; 1884 etp_worker wrk;
1885 ssize_t ret;
1547 1886
1548 wrk.dbuf = 0; 1887 wrk.dbuf = 0;
1549 1888
1550 eio__sendfile (ofd, ifd, offset, count, &wrk); 1889 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1551 1890
1552 if (wrk.dbuf) 1891 if (wrk.dbuf)
1553 free (wrk.dbuf); 1892 free (wrk.dbuf);
1554}
1555 1893
1894 return ret;
1895}
1896

Diff Legend

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