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

Comparing libeio/eio.c (file contents):
Revision 1.33 by root, Sat Jun 6 19:44:17 2009 UTC vs.
Revision 1.60 by root, Fri Jan 7 21:45:42 2011 UTC

1/* 1/*
2 * libeio implementation 2 * libeio implementation
3 * 3 *
4 * Copyright (c) 2007,2008,2009 Marc Alexander Lehmann <libeio@schmorp.de> 4 * Copyright (c) 2007,2008,2009,2010 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 *
… …
49#include <stdlib.h> 49#include <stdlib.h>
50#include <string.h> 50#include <string.h>
51#include <errno.h> 51#include <errno.h>
52#include <sys/types.h> 52#include <sys/types.h>
53#include <sys/stat.h> 53#include <sys/stat.h>
54#include <sys/statvfs.h>
54#include <limits.h> 55#include <limits.h>
55#include <fcntl.h> 56#include <fcntl.h>
56#include <assert.h> 57#include <assert.h>
57 58
58#ifndef EIO_FINISH 59#ifndef EIO_FINISH
… …
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
84#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
85# include <sys/mman.h>
86#endif
87
83/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ 88/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
84# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) 89# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
85# define _DIRENT_HAVE_D_TYPE /* sigh */ 90# define _DIRENT_HAVE_D_TYPE /* sigh */
86# define D_INO(de) (de)->d_fileno 91# define D_INO(de) (de)->d_fileno
87# define D_NAMLEN(de) (de)->d_namlen 92# define D_NAMLEN(de) (de)->d_namlen
88# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 93# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
89# define D_INO(de) (de)->d_ino 94# define D_INO(de) (de)->d_ino
90# endif 95# endif
91 96
92#ifdef _D_EXACT_NAMLEN 97#ifdef _D_EXACT_NAMLEN
93# undef D_NAMLEN 98# undef D_NAMLEN
… …
105#endif 110#endif
106 111
107#if HAVE_SENDFILE 112#if HAVE_SENDFILE
108# if __linux 113# if __linux
109# include <sys/sendfile.h> 114# include <sys/sendfile.h>
110# elif __freebsd 115# elif __FreeBSD__ || defined __APPLE__
111# include <sys/socket.h> 116# include <sys/socket.h>
112# include <sys/uio.h> 117# include <sys/uio.h>
113# elif __hpux 118# elif __hpux
114# include <sys/socket.h> 119# include <sys/socket.h>
115# elif __solaris /* not yet */ 120# elif __solaris
116# include <sys/sendfile.h> 121# include <sys/sendfile.h>
117# else 122# else
118# error sendfile support requested but not available 123# error sendfile support requested but not available
119# endif 124# endif
120#endif 125#endif
… …
133#define IDLE_TIMEOUT 10 138#define IDLE_TIMEOUT 10
134 139
135/* used for struct dirent, AIX doesn't provide it */ 140/* used for struct dirent, AIX doesn't provide it */
136#ifndef NAME_MAX 141#ifndef NAME_MAX
137# define NAME_MAX 4096 142# define NAME_MAX 4096
143#endif
144
145/* used for readlink etc. */
146#ifndef PATH_MAX
147# define PATH_MAX 4096
138#endif 148#endif
139 149
140/* buffer size for various temporary buffers */ 150/* buffer size for various temporary buffers */
141#define EIO_BUFSIZE 65536 151#define EIO_BUFSIZE 65536
142 152
… …
195 205
196/*****************************************************************************/ 206/*****************************************************************************/
197 207
198#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 208#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
199 209
200/* calculcate time difference in ~1/EIO_TICKS of a second */ 210/* calculate time difference in ~1/EIO_TICKS of a second */
201static int tvdiff (struct timeval *tv1, struct timeval *tv2) 211static int tvdiff (struct timeval *tv1, struct timeval *tv2)
202{ 212{
203 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 213 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
204 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 214 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
205} 215}
… …
215static volatile unsigned int nreqs; /* reqlock */ 225static volatile unsigned int nreqs; /* reqlock */
216static volatile unsigned int nready; /* reqlock */ 226static volatile unsigned int nready; /* reqlock */
217static volatile unsigned int npending; /* reqlock */ 227static volatile unsigned int npending; /* reqlock */
218static volatile unsigned int max_idle = 4; 228static volatile unsigned int max_idle = 4;
219 229
220static mutex_t wrklock = X_MUTEX_INIT; 230static xmutex_t wrklock = X_MUTEX_INIT;
221static mutex_t reslock = X_MUTEX_INIT; 231static xmutex_t reslock = X_MUTEX_INIT;
222static mutex_t reqlock = X_MUTEX_INIT; 232static xmutex_t reqlock = X_MUTEX_INIT;
223static cond_t reqwait = X_COND_INIT; 233static xcond_t reqwait = X_COND_INIT;
224 234
225#if !HAVE_PREADWRITE 235#if !HAVE_PREADWRITE
226/* 236/*
227 * make our pread/pwrite emulation safe against themselves, but not against 237 * make our pread/pwrite emulation safe against themselves, but not against
228 * normal read/write by using a mutex. slows down execution a lot, 238 * normal read/write by using a mutex. slows down execution a lot,
229 * but that's your problem, not mine. 239 * but that's your problem, not mine.
230 */ 240 */
231static mutex_t preadwritelock = X_MUTEX_INIT; 241static xmutex_t preadwritelock = X_MUTEX_INIT;
232#endif 242#endif
233 243
234typedef struct etp_worker 244typedef struct etp_worker
235{ 245{
236 /* locked by wrklock */ 246 /* locked by wrklock */
237 struct etp_worker *prev, *next; 247 struct etp_worker *prev, *next;
238 248
239 thread_t tid; 249 xthread_t tid;
240 250
241 /* locked by reslock, reqlock or wrklock */ 251 /* locked by reslock, reqlock or wrklock */
242 ETP_REQ *req; /* currently processed request */ 252 ETP_REQ *req; /* currently processed request */
243 253
244 ETP_WORKER_COMMON 254 ETP_WORKER_COMMON
… …
597} 607}
598 608
599static void etp_set_max_poll_time (double nseconds) 609static void etp_set_max_poll_time (double nseconds)
600{ 610{
601 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 611 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
602 max_poll_time = nseconds; 612 max_poll_time = nseconds * EIO_TICKS;
603 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 613 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
604} 614}
605 615
606static void etp_set_max_poll_reqs (unsigned int maxreqs) 616static void etp_set_max_poll_reqs (unsigned int maxreqs)
607{ 617{
… …
806 816
807 return res; 817 return res;
808} 818}
809#endif 819#endif
810 820
811#ifndef HAVE_FUTIMES 821#ifndef HAVE_UTIMES
812 822
813# undef utimes 823# undef utimes
814# undef futimes
815# define utimes(path,times) eio__utimes (path, times) 824# define utimes(path,times) eio__utimes (path, times)
816# define futimes(fd,times) eio__futimes (fd, times)
817 825
818static int 826static int
819eio__utimes (const char *filename, const struct timeval times[2]) 827eio__utimes (const char *filename, const struct timeval times[2])
820{ 828{
821 if (times) 829 if (times)
… …
828 return utime (filename, &buf); 836 return utime (filename, &buf);
829 } 837 }
830 else 838 else
831 return utime (filename, 0); 839 return utime (filename, 0);
832} 840}
841
842#endif
843
844#ifndef HAVE_FUTIMES
845
846# undef futimes
847# define futimes(fd,times) eio__futimes (fd, times)
833 848
834static int eio__futimes (int fd, const struct timeval tv[2]) 849static int eio__futimes (int fd, const struct timeval tv[2])
835{ 850{
836 errno = ENOSYS; 851 errno = ENOSYS;
837 return -1; 852 return -1;
… …
866 if (!res || errno != ENOSYS) 881 if (!res || errno != ENOSYS)
867 return res; 882 return res;
868#endif 883#endif
869 884
870 /* even though we could play tricks with the flags, it's better to always 885 /* even though we could play tricks with the flags, it's better to always
871 * call fdatasync, as thta matches the expectation of it's users best */ 886 * call fdatasync, as that matches the expectation of its users best */
872 return fdatasync (fd); 887 return fdatasync (fd);
873} 888}
874 889
875#if !HAVE_READAHEAD 890#if !HAVE_READAHEAD
876# undef readahead 891# undef readahead
… …
908 923
909#if HAVE_SENDFILE 924#if HAVE_SENDFILE
910# if __linux 925# if __linux
911 res = sendfile (ofd, ifd, &offset, count); 926 res = sendfile (ofd, ifd, &offset, count);
912 927
913# elif __freebsd 928# elif __FreeBSD__
914 /* 929 /*
915 * Of course, the freebsd sendfile is a dire hack with no thoughts 930 * Of course, the freebsd sendfile is a dire hack with no thoughts
916 * wasted on making it similar to other I/O functions. 931 * wasted on making it similar to other I/O functions.
917 */ 932 */
918 { 933 {
919 off_t sbytes; 934 off_t sbytes;
920 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 935 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
921 936
922 if (res < 0 && sbytes) 937 #if 0 /* according to the manpage, this is correct, but broken behaviour */
923 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 938 /* freebsd' sendfile will return 0 on success */
939 /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
940 /* not on e.g. EIO or EPIPE - sounds broken */
941 if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
942 res = sbytes;
943 #endif
944
945 /* according to source inspection, this is correct, and useful behaviour */
946 if (sbytes)
947 res = sbytes;
948 }
949
950# elif defined (__APPLE__)
951
952 {
953 off_t sbytes = count;
954 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
955
956 /* according to the manpage, sbytes is always valid */
957 if (sbytes)
924 res = sbytes; 958 res = sbytes;
925 } 959 }
926 960
927# elif __hpux 961# elif __hpux
928 res = sendfile (ofd, ifd, offset, count, 0, 0); 962 res = sendfile (ofd, ifd, offset, count, 0, 0);
… …
942 if (res < 0 && sbytes) 976 if (res < 0 && sbytes)
943 res = sbytes; 977 res = sbytes;
944 } 978 }
945 979
946# endif 980# endif
981
982#elif defined (_WIN32)
983
984 /* does not work, just for documentation of what would need to be done */
985 {
986 HANDLE h = TO_SOCKET (ifd);
987 SetFilePointer (h, offset, 0, FILE_BEGIN);
988 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
989 }
990
947#else 991#else
948 res = -1; 992 res = -1;
949 errno = ENOSYS; 993 errno = ENOSYS;
950#endif 994#endif
951 995
952 if (res < 0 996 if (res < 0
953 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 997 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
998 /* BSDs */
999#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
1000 || errno == ENOTSUP
1001#endif
1002 || errno == EOPNOTSUPP /* BSDs */
954#if __solaris 1003#if __solaris
955 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 1004 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
956#endif 1005#endif
957 ) 1006 )
958 ) 1007 )
… …
989 } 1038 }
990 1039
991 return res; 1040 return res;
992} 1041}
993 1042
994static int 1043static signed char
995eio_dent_cmp (const void *a_, const void *b_) 1044eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
996{ 1045{
997 const eio_dirent *a = (const eio_dirent *)a_; 1046 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
998 const eio_dirent *b = (const eio_dirent *)b_;
999
1000 return (int)b->score - (int)a->score ? (int)b->score - (int)a->score
1001 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; /* int might be < ino_t */ 1047 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1048}
1049
1050#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1051
1052#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1053#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1054
1055static void
1056eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1057{
1058 unsigned char bits [9 + sizeof (ino_t) * 8];
1059 unsigned char *bit = bits;
1060
1061 assert (CHAR_BIT == 8);
1062 assert (sizeof (eio_dirent) * 8 < 256);
1063 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1064 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1065
1066 if (size <= EIO_SORT_FAST)
1067 return;
1068
1069 /* first prepare an array of bits to test in our radix sort */
1070 /* try to take endianness into account, as well as differences in ino_t sizes */
1071 /* inode_bits must contain all inodes ORed together */
1072 /* which is used to skip bits that are 0 everywhere, which is very common */
1073 {
1074 ino_t endianness;
1075 int i, j;
1076
1077 /* we store the byte offset of byte n into byte n of "endianness" */
1078 for (i = 0; i < sizeof (ino_t); ++i)
1079 ((unsigned char *)&endianness)[i] = i;
1080
1081 *bit++ = 0;
1082
1083 for (i = 0; i < sizeof (ino_t); ++i)
1084 {
1085 /* shifting off the byte offsets out of "endianness" */
1086 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1087 endianness >>= 8;
1088
1089 for (j = 0; j < 8; ++j)
1090 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1091 *bit++ = offs + j;
1092 }
1093
1094 for (j = 0; j < 8; ++j)
1095 if (score_bits & (1 << j))
1096 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1097 }
1098
1099 /* now actually do the sorting (a variant of MSD radix sort) */
1100 {
1101 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1102 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1103 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1104 int stk_idx = 0;
1105
1106 base_stk [stk_idx] = dents;
1107 end_stk [stk_idx] = dents + size;
1108 bit_stk [stk_idx] = bit - 1;
1109
1110 do
1111 {
1112 base = base_stk [stk_idx];
1113 end = end_stk [stk_idx];
1114 bit = bit_stk [stk_idx];
1115
1116 for (;;)
1117 {
1118 unsigned char O = *bit >> 3;
1119 unsigned char M = 1 << (*bit & 7);
1120
1121 eio_dirent *a = base;
1122 eio_dirent *b = end;
1123
1124 if (b - a < EIO_SORT_CUTOFF)
1125 break;
1126
1127 /* now bit-partition the array on the bit */
1128 /* this ugly asymmetric loop seems to perform much better than typical */
1129 /* partition algos found in the literature */
1130 do
1131 if (!(((unsigned char *)a)[O] & M))
1132 ++a;
1133 else if (!(((unsigned char *)--b)[O] & M))
1134 {
1135 eio_dirent tmp = *a; *a = *b; *b = tmp;
1136 ++a;
1137 }
1138 while (b > a);
1139
1140 /* next bit, or stop, if no bits left in this path */
1141 if (!*--bit)
1142 break;
1143
1144 base_stk [stk_idx] = a;
1145 end_stk [stk_idx] = end;
1146 bit_stk [stk_idx] = bit;
1147 ++stk_idx;
1148
1149 end = a;
1150 }
1151 }
1152 while (stk_idx--);
1153 }
1154}
1155
1156static void
1157eio_dent_insertion_sort (eio_dirent *dents, int size)
1158{
1159 /* first move the smallest element to the front, to act as a sentinel */
1160 {
1161 int i;
1162 eio_dirent *min = dents;
1163
1164 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1165 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1166 if (EIO_DENT_CMP (dents [i], <, *min))
1167 min = &dents [i];
1168
1169 /* swap elements 0 and j (minimum) */
1170 {
1171 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1172 }
1173 }
1174
1175 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1176 {
1177 eio_dirent *i, *j;
1178
1179 for (i = dents + 1; i < dents + size; ++i)
1180 {
1181 eio_dirent value = *i;
1182
1183 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1184 j [1] = j [0];
1185
1186 j [1] = value;
1187 }
1188 }
1189}
1190
1191static void
1192eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1193{
1194 if (size <= 1)
1195 return; /* our insertion sort relies on size > 0 */
1196
1197 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1198 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1199 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1200
1201 /* use an insertion sort at the end, or for small arrays, */
1202 /* as insertion sort is more efficient for small partitions */
1203 eio_dent_insertion_sort (dents, size);
1002} 1204}
1003 1205
1004/* read a full directory */ 1206/* read a full directory */
1005static void 1207static void
1006eio__scandir (eio_req *req, etp_worker *self) 1208eio__scandir (eio_req *req, etp_worker *self)
1007{ 1209{
1008 DIR *dirp; 1210 DIR *dirp;
1009 EIO_STRUCT_DIRENT *entp; 1211 EIO_STRUCT_DIRENT *entp;
1010 unsigned char *name, *names; 1212 char *name, *names;
1011 int namesalloc = 4096; 1213 int namesalloc = 4096;
1012 int namesoffs = 0; 1214 int namesoffs = 0;
1013 int flags = req->int1; 1215 int flags = req->int1;
1014 eio_dirent *dents = 0; 1216 eio_dirent *dents = 0;
1015 int dentalloc = 128; 1217 int dentalloc = 128;
1016 int dentoffs = 0; 1218 int dentoffs = 0;
1219 ino_t inode_bits = 0;
1017 1220
1018 req->result = -1; 1221 req->result = -1;
1019 1222
1020 if (!(flags & EIO_READDIR_DENTS)) 1223 if (!(flags & EIO_READDIR_DENTS))
1021 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1224 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1022 1225
1023 X_LOCK (wrklock); 1226 X_LOCK (wrklock);
1024 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1227 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1025 self->dirp = dirp = opendir (req->ptr1); 1228 self->dirp = dirp = opendir (req->ptr1);
1026 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1229 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1027 req->ptr1 = names = malloc (namesalloc);
1028 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1230 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1231 req->ptr2 = names = malloc (namesalloc);
1029 X_UNLOCK (wrklock); 1232 X_UNLOCK (wrklock);
1030 1233
1031 if (dirp && names && (!flags || dents)) 1234 if (dirp && names && (!flags || dents))
1032 for (;;) 1235 for (;;)
1033 { 1236 {
… …
1041 1244
1042 /* sort etc. */ 1245 /* sort etc. */
1043 req->int1 = flags; 1246 req->int1 = flags;
1044 req->result = dentoffs; 1247 req->result = dentoffs;
1045 1248
1046 if (dents)
1047 {
1048 eio_dirent *ent = dents + dentoffs;
1049
1050 while (ent > dents)
1051 (--ent)->name = names + (size_t)ent->name;
1052 }
1053
1054 if (flags & EIO_READDIR_STAT_ORDER 1249 if (flags & EIO_READDIR_STAT_ORDER)
1055 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1250 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1056 {
1057 /* pray your qsort doesn't use quicksort */
1058 qsort (dents, dentoffs, sizeof (*dents), eio_dent_cmp); /* score depends of DIRS_FIRST */
1059 }
1060 else if (flags & EIO_READDIR_DIRS_FIRST) 1251 else if (flags & EIO_READDIR_DIRS_FIRST)
1252 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1253 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1254 else
1061 { 1255 {
1062 /* in this case, all is known, and we just put dirs first and sort them */ 1256 /* in this case, all is known, and we just put dirs first and sort them */
1063 eio_dirent *ent = dents + dentoffs; 1257 eio_dirent *oth = dents + dentoffs;
1064 eio_dirent *dir = dents; 1258 eio_dirent *dir = dents;
1065 1259
1260 /* now partition dirs to the front, and non-dirs to the back */
1261 /* by walking from both sides and swapping if necessary */
1262 /* also clear score, so it doesn't influence sorting */
1066 while (ent > dir) 1263 while (oth > dir)
1067 { 1264 {
1068 if (dir->type == DT_DIR) 1265 if (dir->type == EIO_DT_DIR)
1069 ++dir; 1266 ++dir;
1070 else
1071 {
1072 --ent;
1073
1074 if (ent->type == DT_DIR) 1267 else if ((--oth)->type == EIO_DT_DIR)
1075 { 1268 {
1076 eio_dirent tmp = *dir; 1269 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1077 *dir = *ent;
1078 *ent = tmp;
1079 1270
1080 ++dir; 1271 ++dir;
1081 } 1272 }
1082 } 1273 }
1083 }
1084 1274
1085 /* now sort the dirs only */ 1275 /* now sort the dirs only */
1086 qsort (dents, dir - dents, sizeof (*dents), eio_dent_cmp); 1276 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1087 } 1277 }
1088
1089 /* only provide the names array unless DENTS is specified */
1090 if (!(flags & EIO_READDIR_DENTS))
1091 {
1092 X_LOCK (wrklock);
1093 assert (!dents);
1094 req->ptr1 = 0;
1095 req->ptr2 = names;
1096 X_UNLOCK (wrklock);
1097 }
1098 1278
1099 break; 1279 break;
1100 } 1280 }
1101 1281
1102 /* now add the entry to our list(s) */ 1282 /* now add the entry to our list(s) */
… …
1109 1289
1110 while (expect_false (namesoffs + len > namesalloc)) 1290 while (expect_false (namesoffs + len > namesalloc))
1111 { 1291 {
1112 namesalloc *= 2; 1292 namesalloc *= 2;
1113 X_LOCK (wrklock); 1293 X_LOCK (wrklock);
1114 req->ptr1 = names = realloc (names, namesalloc); 1294 req->ptr2 = names = realloc (names, namesalloc);
1115 X_UNLOCK (wrklock); 1295 X_UNLOCK (wrklock);
1116 1296
1117 if (!names) 1297 if (!names)
1118 break; 1298 break;
1119 } 1299 }
… …
1126 1306
1127 if (expect_false (dentoffs == dentalloc)) 1307 if (expect_false (dentoffs == dentalloc))
1128 { 1308 {
1129 dentalloc *= 2; 1309 dentalloc *= 2;
1130 X_LOCK (wrklock); 1310 X_LOCK (wrklock);
1131 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1311 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1132 X_UNLOCK (wrklock); 1312 X_UNLOCK (wrklock);
1133 1313
1134 if (!dents) 1314 if (!dents)
1135 break; 1315 break;
1136 } 1316 }
1137 1317
1138 ent = dents + dentoffs; 1318 ent = dents + dentoffs;
1139 1319
1140 ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ 1320 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1141 ent->namelen = len - 1; 1321 ent->namelen = len - 1;
1142 ent->inode = D_INO (entp); 1322 ent->inode = D_INO (entp);
1323
1324 inode_bits |= ent->inode;
1143 1325
1144 switch (D_TYPE (entp)) 1326 switch (D_TYPE (entp))
1145 { 1327 {
1146 default: 1328 default:
1147 ent->type = EIO_DT_UNKNOWN; 1329 ent->type = EIO_DT_UNKNOWN;
… …
1190 #ifdef DT_WHT 1372 #ifdef DT_WHT
1191 case DT_WHT: ent->type = EIO_DT_WHT; break; 1373 case DT_WHT: ent->type = EIO_DT_WHT; break;
1192 #endif 1374 #endif
1193 } 1375 }
1194 1376
1195 ent->score = 0; 1377 ent->score = 7;
1196 1378
1197 if (flags & EIO_READDIR_DIRS_FIRST) 1379 if (flags & EIO_READDIR_DIRS_FIRST)
1198 { 1380 {
1199 if (ent->type == EIO_DT_UNKNOWN) 1381 if (ent->type == EIO_DT_UNKNOWN)
1200 { 1382 {
1201 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ 1383 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1202 ent->score = 98; 1384 ent->score = 1;
1203 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1385 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1204 ent->score = len <= 2 ? len + 60 : len <= 4 ? 50 : len <= 7 ? 40 : 10; /* shorter == more likely dir, but avoid too many classes */ 1386 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1205 } 1387 }
1206 else if (ent->type == EIO_DT_DIR) 1388 else if (ent->type == EIO_DT_DIR)
1207 ent->score = 100; 1389 ent->score = 0;
1208 } 1390 }
1209 } 1391 }
1210 1392
1211 namesoffs += len; 1393 namesoffs += len;
1212 ++dentoffs; 1394 ++dentoffs;
1213 } 1395 }
1396
1397 if (EIO_CANCELLED (req))
1398 {
1399 errno = ECANCELED;
1400 break;
1401 }
1214 } 1402 }
1215 else
1216 req->result = -1;
1217} 1403}
1404
1405#ifdef PAGESIZE
1406# define eio_pagesize() PAGESIZE
1407#else
1408static intptr_t
1409eio_pagesize (void)
1410{
1411 static intptr_t page;
1412
1413 if (!page)
1414 page = sysconf (_SC_PAGESIZE);
1415
1416 return page;
1417}
1418#endif
1419
1420static void
1421eio_page_align (void **addr, size_t *length)
1422{
1423 intptr_t mask = eio_pagesize () - 1;
1424
1425 /* round down addr */
1426 intptr_t adj = mask & (intptr_t)*addr;
1427
1428 *addr = (void *)((intptr_t)*addr - adj);
1429 *length += adj;
1430
1431 /* round up length */
1432 *length = (*length + mask) & ~mask;
1433}
1434
1435#if !_POSIX_MEMLOCK
1436# define eio__mlockall(a) ((errno = ENOSYS), -1)
1437#else
1438
1439static int
1440eio__mlockall (int flags)
1441{
1442 #if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
1443 extern int mallopt (int, int);
1444 mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
1445 #endif
1446
1447 if (EIO_MCL_CURRENT != MCL_CURRENT
1448 || EIO_MCL_FUTURE != MCL_FUTURE)
1449 {
1450 flags = 0
1451 | (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
1452 | (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
1453 }
1454
1455 return mlockall (flags);
1456}
1457#endif
1458
1459#if !_POSIX_MEMLOCK_RANGE
1460# define eio__mlock(a,b) ((errno = ENOSYS), -1)
1461#else
1462
1463static int
1464eio__mlock (void *addr, size_t length)
1465{
1466 eio_page_align (&addr, &length);
1467
1468 return mlock (addr, length);
1469}
1470
1471#endif
1218 1472
1219#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1473#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1220# undef msync
1221# define msync(a,b,c) ((errno = ENOSYS), -1) 1474# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1475#else
1476
1477int
1478eio__msync (void *mem, size_t len, int flags)
1479{
1480 eio_page_align (&mem, &len);
1481
1482 if (EIO_MS_ASYNC != MS_SYNC
1483 || EIO_MS_INVALIDATE != MS_INVALIDATE
1484 || EIO_MS_SYNC != MS_SYNC)
1485 {
1486 flags = 0
1487 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1488 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1489 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1490 }
1491
1492 return msync (mem, len, flags);
1493}
1494
1222#endif 1495#endif
1223 1496
1224int 1497int
1225eio__mtouch (void *mem, size_t len, int flags) 1498eio__mtouch (void *mem, size_t len, int flags)
1226{ 1499{
1500 eio_page_align (&mem, &len);
1501
1502 {
1227 intptr_t addr = (intptr_t)mem; 1503 intptr_t addr = (intptr_t)mem;
1228 intptr_t end = addr + len; 1504 intptr_t end = addr + len;
1229#ifdef PAGESIZE 1505 intptr_t page = eio_pagesize ();
1230 const intptr_t page = PAGESIZE;
1231#else
1232 static intptr_t page;
1233 1506
1234 if (!page)
1235 page = sysconf (_SC_PAGESIZE);
1236#endif
1237
1238 addr &= ~(page - 1); /* assume page size is always a power of two */
1239
1240 if (addr < end) 1507 if (addr < end)
1241 if (flags) /* modify */ 1508 if (flags & EIO_MT_MODIFY) /* modify */
1242 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); 1509 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1243 else 1510 else
1244 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); 1511 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1512 }
1245 1513
1246 return 0; 1514 return 0;
1247} 1515}
1248 1516
1249/*****************************************************************************/ 1517/*****************************************************************************/
… …
1372 return 0; \ 1640 return 0; \
1373 } 1641 }
1374 1642
1375static void eio_execute (etp_worker *self, eio_req *req) 1643static void eio_execute (etp_worker *self, eio_req *req)
1376{ 1644{
1377 errno = 0;
1378
1379 switch (req->type) 1645 switch (req->type)
1380 { 1646 {
1381 case EIO_READ: ALLOC (req->size); 1647 case EIO_READ: ALLOC (req->size);
1382 req->result = req->offs >= 0 1648 req->result = req->offs >= 0
1383 ? pread (req->int1, req->ptr2, req->size, req->offs) 1649 ? pread (req->int1, req->ptr2, req->size, req->offs)
… …
1393 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1659 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1394 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1660 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1395 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1661 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1396 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1662 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1397 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1663 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1664
1665 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1666 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1667 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1668 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1398 1669
1399 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; 1670 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1400 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; 1671 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1401 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; 1672 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1402 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; 1673 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
… …
1412 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1683 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1413 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1684 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1414 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1685 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1415 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break; 1686 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1416 1687
1417 case EIO_READLINK: ALLOC (NAME_MAX); 1688 case EIO_READLINK: ALLOC (PATH_MAX);
1418 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1689 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1419 1690
1420 case EIO_SYNC: req->result = 0; sync (); break; 1691 case EIO_SYNC: req->result = 0; sync (); break;
1421 case EIO_FSYNC: req->result = fsync (req->int1); break; 1692 case EIO_FSYNC: req->result = fsync (req->int1); break;
1422 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1693 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1423 case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; 1694 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1424 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; 1695 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1696 case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
1697 case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
1425 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; 1698 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1426 1699
1427 case EIO_READDIR: eio__scandir (req, self); break; 1700 case EIO_READDIR: eio__scandir (req, self); break;
1428 1701
1429 case EIO_BUSY: 1702 case EIO_BUSY:
1430#ifdef _WIN32 1703#ifdef _WIN32
1431 Sleep (req->nv1 * 1000.); 1704 Sleep (req->nv1 * 1e3);
1432#else 1705#else
1433 { 1706 {
1434 struct timeval tv; 1707 struct timeval tv;
1435 1708
1436 tv.tv_sec = req->nv1; 1709 tv.tv_sec = req->nv1;
1437 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.; 1710 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
1438 1711
1439 req->result = select (0, 0, 0, 0, &tv); 1712 req->result = select (0, 0, 0, 0, &tv);
1440 } 1713 }
1441#endif 1714#endif
1442 break; 1715 break;
… …
1457 times = tv; 1730 times = tv;
1458 } 1731 }
1459 else 1732 else
1460 times = 0; 1733 times = 0;
1461 1734
1462
1463 req->result = req->type == EIO_FUTIME 1735 req->result = req->type == EIO_FUTIME
1464 ? futimes (req->int1, times) 1736 ? futimes (req->int1, times)
1465 : utimes (req->ptr1, times); 1737 : utimes (req->ptr1, times);
1466 } 1738 }
1467 break; 1739 break;
… …
1476 case EIO_CUSTOM: 1748 case EIO_CUSTOM:
1477 ((void (*)(eio_req *))req->feed) (req); 1749 ((void (*)(eio_req *))req->feed) (req);
1478 break; 1750 break;
1479 1751
1480 default: 1752 default:
1753 errno = ENOSYS;
1481 req->result = -1; 1754 req->result = -1;
1482 break; 1755 break;
1483 } 1756 }
1484 1757
1485 req->errorno = errno; 1758 req->errorno = errno;
… …
1515eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) 1788eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1516{ 1789{
1517 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; 1790 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1518} 1791}
1519 1792
1793eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
1794{
1795 REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
1796}
1797
1798eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
1799{
1800 REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
1801}
1802
1520eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data) 1803eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1521{ 1804{
1522 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND; 1805 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1523} 1806}
1524 1807
… …
1548} 1831}
1549 1832
1550eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) 1833eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1551{ 1834{
1552 REQ (EIO_FSTAT); req->int1 = fd; SEND; 1835 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1836}
1837
1838eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1839{
1840 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1553} 1841}
1554 1842
1555eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) 1843eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1556{ 1844{
1557 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; 1845 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
… …
1629} 1917}
1630 1918
1631eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) 1919eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1632{ 1920{
1633 return eio__1path (EIO_LSTAT, path, pri, cb, data); 1921 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1922}
1923
1924eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
1925{
1926 return eio__1path (EIO_STATVFS, path, pri, cb, data);
1634} 1927}
1635 1928
1636eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) 1929eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1637{ 1930{
1638 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1931 return eio__1path (EIO_UNLINK, path, pri, cb, data);
… …
1742/* misc garbage */ 2035/* misc garbage */
1743 2036
1744ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) 2037ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1745{ 2038{
1746 etp_worker wrk; 2039 etp_worker wrk;
2040 ssize_t ret;
1747 2041
1748 wrk.dbuf = 0; 2042 wrk.dbuf = 0;
1749 2043
1750 eio__sendfile (ofd, ifd, offset, count, &wrk); 2044 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1751 2045
1752 if (wrk.dbuf) 2046 if (wrk.dbuf)
1753 free (wrk.dbuf); 2047 free (wrk.dbuf);
1754}
1755 2048
2049 return ret;
2050}
2051

Diff Legend

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