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

Comparing libeio/eio.c (file contents):
Revision 1.39 by root, Sat Jun 13 03:59:04 2009 UTC vs.
Revision 1.56 by root, Sun Sep 12 03:36:28 2010 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
… …
73#else 74#else
74 75
75# include "config.h" 76# include "config.h"
76# include <sys/time.h> 77# include <sys/time.h>
77# include <sys/select.h> 78# include <sys/select.h>
78# include <sys/mman.h>
79# include <unistd.h> 79# include <unistd.h>
80# include <utime.h> 80# include <utime.h>
81# include <signal.h> 81# include <signal.h>
82# include <dirent.h> 82# include <dirent.h>
83 83
84#if _POSIX_MEMLOCK || _POSIX_MAPPED_FILES
85# include <sys/mman.h>
86#endif
87
84/* 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 */
85# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) 89# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
86# define _DIRENT_HAVE_D_TYPE /* sigh */ 90# define _DIRENT_HAVE_D_TYPE /* sigh */
87# define D_INO(de) (de)->d_fileno 91# define D_INO(de) (de)->d_fileno
88# define D_NAMLEN(de) (de)->d_namlen 92# define D_NAMLEN(de) (de)->d_namlen
89# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 93# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
90# define D_INO(de) (de)->d_ino 94# define D_INO(de) (de)->d_ino
91# endif 95# endif
92 96
93#ifdef _D_EXACT_NAMLEN 97#ifdef _D_EXACT_NAMLEN
94# undef D_NAMLEN 98# undef D_NAMLEN
… …
106#endif 110#endif
107 111
108#if HAVE_SENDFILE 112#if HAVE_SENDFILE
109# if __linux 113# if __linux
110# include <sys/sendfile.h> 114# include <sys/sendfile.h>
111# elif __freebsd 115# elif __FreeBSD__ || defined __APPLE__
112# include <sys/socket.h> 116# include <sys/socket.h>
113# include <sys/uio.h> 117# include <sys/uio.h>
114# elif __hpux 118# elif __hpux
115# include <sys/socket.h> 119# include <sys/socket.h>
116# elif __solaris /* not yet */ 120# elif __solaris
117# include <sys/sendfile.h> 121# include <sys/sendfile.h>
118# else 122# else
119# error sendfile support requested but not available 123# error sendfile support requested but not available
120# endif 124# endif
121#endif 125#endif
… …
134#define IDLE_TIMEOUT 10 138#define IDLE_TIMEOUT 10
135 139
136/* used for struct dirent, AIX doesn't provide it */ 140/* used for struct dirent, AIX doesn't provide it */
137#ifndef NAME_MAX 141#ifndef NAME_MAX
138# 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
139#endif 148#endif
140 149
141/* buffer size for various temporary buffers */ 150/* buffer size for various temporary buffers */
142#define EIO_BUFSIZE 65536 151#define EIO_BUFSIZE 65536
143 152
… …
196 205
197/*****************************************************************************/ 206/*****************************************************************************/
198 207
199#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 208#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
200 209
201/* calculcate time difference in ~1/EIO_TICKS of a second */ 210/* calculate time difference in ~1/EIO_TICKS of a second */
202static int tvdiff (struct timeval *tv1, struct timeval *tv2) 211static int tvdiff (struct timeval *tv1, struct timeval *tv2)
203{ 212{
204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 213 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
205 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 214 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
206} 215}
… …
216static volatile unsigned int nreqs; /* reqlock */ 225static volatile unsigned int nreqs; /* reqlock */
217static volatile unsigned int nready; /* reqlock */ 226static volatile unsigned int nready; /* reqlock */
218static volatile unsigned int npending; /* reqlock */ 227static volatile unsigned int npending; /* reqlock */
219static volatile unsigned int max_idle = 4; 228static volatile unsigned int max_idle = 4;
220 229
221static mutex_t wrklock = X_MUTEX_INIT; 230static xmutex_t wrklock = X_MUTEX_INIT;
222static mutex_t reslock = X_MUTEX_INIT; 231static xmutex_t reslock = X_MUTEX_INIT;
223static mutex_t reqlock = X_MUTEX_INIT; 232static xmutex_t reqlock = X_MUTEX_INIT;
224static cond_t reqwait = X_COND_INIT; 233static xcond_t reqwait = X_COND_INIT;
225 234
226#if !HAVE_PREADWRITE 235#if !HAVE_PREADWRITE
227/* 236/*
228 * make our pread/pwrite emulation safe against themselves, but not against 237 * make our pread/pwrite emulation safe against themselves, but not against
229 * 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,
230 * but that's your problem, not mine. 239 * but that's your problem, not mine.
231 */ 240 */
232static mutex_t preadwritelock = X_MUTEX_INIT; 241static xmutex_t preadwritelock = X_MUTEX_INIT;
233#endif 242#endif
234 243
235typedef struct etp_worker 244typedef struct etp_worker
236{ 245{
237 /* locked by wrklock */ 246 /* locked by wrklock */
238 struct etp_worker *prev, *next; 247 struct etp_worker *prev, *next;
239 248
240 thread_t tid; 249 xthread_t tid;
241 250
242 /* locked by reslock, reqlock or wrklock */ 251 /* locked by reslock, reqlock or wrklock */
243 ETP_REQ *req; /* currently processed request */ 252 ETP_REQ *req; /* currently processed request */
244 253
245 ETP_WORKER_COMMON 254 ETP_WORKER_COMMON
… …
598} 607}
599 608
600static void etp_set_max_poll_time (double nseconds) 609static void etp_set_max_poll_time (double nseconds)
601{ 610{
602 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 611 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
603 max_poll_time = nseconds; 612 max_poll_time = nseconds * EIO_TICKS;
604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 613 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
605} 614}
606 615
607static void etp_set_max_poll_reqs (unsigned int maxreqs) 616static void etp_set_max_poll_reqs (unsigned int maxreqs)
608{ 617{
… …
867 if (!res || errno != ENOSYS) 876 if (!res || errno != ENOSYS)
868 return res; 877 return res;
869#endif 878#endif
870 879
871 /* even though we could play tricks with the flags, it's better to always 880 /* 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 */ 881 * call fdatasync, as that matches the expectation of its users best */
873 return fdatasync (fd); 882 return fdatasync (fd);
874} 883}
875 884
876#if !HAVE_READAHEAD 885#if !HAVE_READAHEAD
877# undef readahead 886# undef readahead
… …
909 918
910#if HAVE_SENDFILE 919#if HAVE_SENDFILE
911# if __linux 920# if __linux
912 res = sendfile (ofd, ifd, &offset, count); 921 res = sendfile (ofd, ifd, &offset, count);
913 922
914# elif __freebsd 923# elif __FreeBSD__
915 /* 924 /*
916 * Of course, the freebsd sendfile is a dire hack with no thoughts 925 * Of course, the freebsd sendfile is a dire hack with no thoughts
917 * wasted on making it similar to other I/O functions. 926 * wasted on making it similar to other I/O functions.
918 */ 927 */
919 { 928 {
920 off_t sbytes; 929 off_t sbytes;
921 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 930 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
922 931
923 if (res < 0 && sbytes) 932 #if 0 /* according to the manpage, this is correct, but broken behaviour */
924 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 933 /* freebsd' sendfile will return 0 on success */
934 /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
935 /* not on e.g. EIO or EPIPE - sounds broken */
936 if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
937 res = sbytes;
938 #endif
939
940 /* according to source inspection, this is correct, and useful behaviour */
941 if (sbytes)
942 res = sbytes;
943 }
944
945# elif defined (__APPLE__)
946
947 {
948 off_t sbytes = count;
949 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
950
951 /* according to the manpage, sbytes is always valid */
952 if (sbytes)
925 res = sbytes; 953 res = sbytes;
926 } 954 }
927 955
928# elif __hpux 956# elif __hpux
929 res = sendfile (ofd, ifd, offset, count, 0, 0); 957 res = sendfile (ofd, ifd, offset, count, 0, 0);
… …
943 if (res < 0 && sbytes) 971 if (res < 0 && sbytes)
944 res = sbytes; 972 res = sbytes;
945 } 973 }
946 974
947# endif 975# endif
976
977#elif defined (_WIN32)
978
979 /* does not work, just for documentation of what would need to be done */
980 {
981 HANDLE h = TO_SOCKET (ifd);
982 SetFilePointer (h, offset, 0, FILE_BEGIN);
983 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
984 }
985
948#else 986#else
949 res = -1; 987 res = -1;
950 errno = ENOSYS; 988 errno = ENOSYS;
951#endif 989#endif
952 990
953 if (res < 0 991 if (res < 0
954 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 992 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
993 /* BSDs */
994#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
995 || errno == ENOTSUP
996#endif
997 || errno == EOPNOTSUPP /* BSDs */
955#if __solaris 998#if __solaris
956 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 999 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
957#endif 1000#endif
958 ) 1001 )
959 ) 1002 )
… …
993} 1036}
994 1037
995static signed char 1038static signed char
996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) 1039eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
997{ 1040{
998 return b->score - a->score ? b->score - a->score /* works because our signed char is always 0..100 */ 1041 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; 1042 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1000} 1043}
1001 1044
1002#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 1045#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1003 1046
1004#define EIO_QSORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ 1047#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1005#define EIO_QSORT_SKIP 60 /* when to skip qsort completely */ 1048#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1006 1049
1007static void 1050static void
1008eio_dent_sort (eio_dirent *dents, int size) 1051eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1009{ 1052{
1053 unsigned char bits [9 + sizeof (ino_t) * 8];
1054 unsigned char *bit = bits;
1055
1056 assert (CHAR_BIT == 8);
1057 assert (sizeof (eio_dirent) * 8 < 256);
1058 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1059 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1060
1061 if (size <= EIO_SORT_FAST)
1062 return;
1063
1064 /* first prepare an array of bits to test in our radix sort */
1065 /* try to take endianness into account, as well as differences in ino_t sizes */
1066 /* inode_bits must contain all inodes ORed together */
1067 /* which is used to skip bits that are 0 everywhere, which is very common */
1068 {
1069 ino_t endianness;
1010 int i, j; 1070 int i, j;
1011 1071
1072 /* we store the byte offset of byte n into byte n of "endianness" */
1073 for (i = 0; i < sizeof (ino_t); ++i)
1074 ((unsigned char *)&endianness)[i] = i;
1075
1076 *bit++ = 0;
1077
1078 for (i = 0; i < sizeof (ino_t); ++i)
1079 {
1080 /* shifting off the byte offsets out of "endianness" */
1081 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1082 endianness >>= 8;
1083
1084 for (j = 0; j < 8; ++j)
1085 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1086 *bit++ = offs + j;
1087 }
1088
1089 for (j = 0; j < 8; ++j)
1090 if (score_bits & (1 << j))
1091 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1092 }
1093
1094 /* now actually do the sorting (a variant of MSD radix sort) */
1095 {
1096 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1097 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1098 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1099 int stk_idx = 0;
1100
1101 base_stk [stk_idx] = dents;
1102 end_stk [stk_idx] = dents + size;
1103 bit_stk [stk_idx] = bit - 1;
1104
1105 do
1106 {
1107 base = base_stk [stk_idx];
1108 end = end_stk [stk_idx];
1109 bit = bit_stk [stk_idx];
1110
1111 for (;;)
1112 {
1113 unsigned char O = *bit >> 3;
1114 unsigned char M = 1 << (*bit & 7);
1115
1116 eio_dirent *a = base;
1117 eio_dirent *b = end;
1118
1119 if (b - a < EIO_SORT_CUTOFF)
1120 break;
1121
1122 /* now bit-partition the array on the bit */
1123 /* this ugly asymmetric loop seems to perform much better than typical */
1124 /* partition algos found in the literature */
1125 do
1126 if (!(((unsigned char *)a)[O] & M))
1127 ++a;
1128 else if (!(((unsigned char *)--b)[O] & M))
1129 {
1130 eio_dirent tmp = *a; *a = *b; *b = tmp;
1131 ++a;
1132 }
1133 while (b > a);
1134
1135 /* next bit, or stop, if no bits left in this path */
1136 if (!*--bit)
1137 break;
1138
1139 base_stk [stk_idx] = a;
1140 end_stk [stk_idx] = end;
1141 bit_stk [stk_idx] = bit;
1142 ++stk_idx;
1143
1144 end = a;
1145 }
1146 }
1147 while (stk_idx--);
1148 }
1149}
1150
1151static void
1152eio_dent_insertion_sort (eio_dirent *dents, int size)
1153{
1154 /* first move the smallest element to the front, to act as a sentinel */
1155 {
1156 int i;
1157 eio_dirent *min = dents;
1158
1159 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1160 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1161 if (EIO_DENT_CMP (dents [i], <, *min))
1162 min = &dents [i];
1163
1164 /* swap elements 0 and j (minimum) */
1165 {
1166 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1167 }
1168 }
1169
1170 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1171 {
1172 eio_dirent *i, *j;
1173
1174 for (i = dents + 1; i < dents + size; ++i)
1175 {
1176 eio_dirent value = *i;
1177
1178 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1179 j [1] = j [0];
1180
1181 j [1] = value;
1182 }
1183 }
1184}
1185
1186static void
1187eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1188{
1012 if (size <= 1) 1189 if (size <= 1)
1013 return; /* our insertion sort relies on size > 0 */ 1190 return; /* our insertion sort relies on size > 0 */
1014 1191
1015 if (size > EIO_QSORT_SKIP) /* skip quicksort for small directories */ 1192 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1016 { 1193 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1017 /* first, use quicksort */ 1194 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1018 /* should be good for 2**31 entries */
1019 struct rng { int l, r; } rng [32];
1020 1195
1021 i = 0;
1022 rng[0].l = 0;
1023 rng[0].r = size;
1024
1025 while (expect_true (i >= 0))
1026 {
1027 int L = rng [i].l;
1028 int R = rng [i].r - 1;
1029
1030 if (expect_false (L + EIO_QSORT_CUTOFF < R))
1031 {
1032 eio_dirent piv = dents [L];
1033
1034 while (L < R)
1035 {
1036 while (EIO_DENT_CMP (dents [R], >=, piv) && L < R)
1037 --R;
1038
1039 if (L < R)
1040 dents [L++] = dents [R];
1041
1042 while (EIO_DENT_CMP (dents [L], <=, piv) && L < R)
1043 ++L;
1044
1045 if (L < R)
1046 dents [R--] = dents [L];
1047 }
1048
1049 dents [L] = piv;
1050
1051 ++i;
1052 rng [i].l = L + 1;
1053 rng [i].r = rng [i - 1].r;
1054 rng [i - 1].r = L;
1055
1056 if (rng [i].r - rng [i].l > rng [i - 1].r - rng [i - 1].l)
1057 {
1058 struct rng t;
1059
1060 t = rng [i]; rng [i] = rng [i - 1]; rng [i - 1] = t;
1061 }
1062 }
1063 else
1064 --i;
1065 }
1066 }
1067
1068 /* use an insertion sort after qsort, or for small arrays */ 1196 /* use an insertion sort at the end, or for small arrays, */
1069 /* first move the smallest element to the front, to act as a sentinel */ 1197 /* as insertion sort is more efficient for small partitions */
1070 { 1198 eio_dent_insertion_sort (dents, size);
1071 int min = 0;
1072
1073 for (i = size > EIO_QSORT_SKIP ? EIO_QSORT_CUTOFF : size; --i; )
1074 if (EIO_DENT_CMP (dents [i], <, dents [min]))
1075 min = i;
1076
1077 /* swap elements 0 and j (minimum) */
1078 {
1079 eio_dirent tmp = dents [0]; dents [0] = dents [min]; dents [min] = tmp;
1080 }
1081 }
1082
1083 /* then do standard insertion sort */
1084 for (i = 1; i < size; ++i)
1085 {
1086 eio_dirent value = dents [i];
1087
1088 for (j = i - 1; EIO_DENT_CMP (dents [j], >, value); --j)
1089 dents [j + 1] = dents [j];
1090
1091 dents [j + 1] = value;
1092 }
1093} 1199}
1094 1200
1095/* read a full directory */ 1201/* read a full directory */
1096static void 1202static void
1097eio__scandir (eio_req *req, etp_worker *self) 1203eio__scandir (eio_req *req, etp_worker *self)
… …
1103 int namesoffs = 0; 1209 int namesoffs = 0;
1104 int flags = req->int1; 1210 int flags = req->int1;
1105 eio_dirent *dents = 0; 1211 eio_dirent *dents = 0;
1106 int dentalloc = 128; 1212 int dentalloc = 128;
1107 int dentoffs = 0; 1213 int dentoffs = 0;
1214 ino_t inode_bits = 0;
1108 1215
1109 req->result = -1; 1216 req->result = -1;
1110 1217
1111 if (!(flags & EIO_READDIR_DENTS)) 1218 if (!(flags & EIO_READDIR_DENTS))
1112 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1219 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1113 1220
1114 X_LOCK (wrklock); 1221 X_LOCK (wrklock);
1115 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1222 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1116 self->dirp = dirp = opendir (req->ptr1); 1223 self->dirp = dirp = opendir (req->ptr1);
1117 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1224 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1118 req->ptr1 = names = malloc (namesalloc);
1119 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1225 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1226 req->ptr2 = names = malloc (namesalloc);
1120 X_UNLOCK (wrklock); 1227 X_UNLOCK (wrklock);
1121 1228
1122 if (dirp && names && (!flags || dents)) 1229 if (dirp && names && (!flags || dents))
1123 for (;;) 1230 for (;;)
1124 { 1231 {
… …
1132 1239
1133 /* sort etc. */ 1240 /* sort etc. */
1134 req->int1 = flags; 1241 req->int1 = flags;
1135 req->result = dentoffs; 1242 req->result = dentoffs;
1136 1243
1137 if (dents)
1138 {
1139 eio_dirent *ent = dents + dentoffs;
1140
1141 while (ent > dents)
1142 (--ent)->name = names + (size_t)ent->name;
1143 }
1144
1145 if (flags & EIO_READDIR_STAT_ORDER 1244 if (flags & EIO_READDIR_STAT_ORDER)
1146 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1245 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1147 eio_dent_sort (dents, dentoffs); /* score depends of DIRS_FIRST */
1148 else if (flags & EIO_READDIR_DIRS_FIRST) 1246 else if (flags & EIO_READDIR_DIRS_FIRST)
1247 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1248 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1249 else
1149 { 1250 {
1150 /* in this case, all is known, and we just put dirs first and sort them */ 1251 /* in this case, all is known, and we just put dirs first and sort them */
1151 eio_dirent *ent = dents + dentoffs; 1252 eio_dirent *oth = dents + dentoffs;
1152 eio_dirent *dir = dents; 1253 eio_dirent *dir = dents;
1153 1254
1154 /* now move dirs to the front, and non-dirs to the back */ 1255 /* now partition dirs to the front, and non-dirs to the back */
1155 /* by walking from both sides and swapping if necessary */ 1256 /* by walking from both sides and swapping if necessary */
1257 /* also clear score, so it doesn't influence sorting */
1156 while (ent > dir) 1258 while (oth > dir)
1157 { 1259 {
1158 if (dir->type == DT_DIR) 1260 if (dir->type == EIO_DT_DIR)
1159 ++dir; 1261 ++dir;
1160 else
1161 {
1162 --ent;
1163
1164 if (ent->type == DT_DIR) 1262 else if ((--oth)->type == EIO_DT_DIR)
1165 { 1263 {
1166 eio_dirent tmp = *dir; 1264 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1167 *dir = *ent;
1168 *ent = tmp;
1169 1265
1170 ++dir; 1266 ++dir;
1171 } 1267 }
1172 } 1268 }
1173 }
1174 1269
1175 /* now sort the dirs only */ 1270 /* now sort the dirs only */
1176 eio_dent_sort (dents, dir - dents); 1271 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1177 } 1272 }
1178
1179 /* only provide the names array unless DENTS is specified */
1180 if (!(flags & EIO_READDIR_DENTS))
1181 {
1182 X_LOCK (wrklock);
1183 assert (!dents);
1184 req->ptr1 = 0;
1185 req->ptr2 = names;
1186 X_UNLOCK (wrklock);
1187 }
1188 1273
1189 break; 1274 break;
1190 } 1275 }
1191 1276
1192 /* now add the entry to our list(s) */ 1277 /* now add the entry to our list(s) */
… …
1199 1284
1200 while (expect_false (namesoffs + len > namesalloc)) 1285 while (expect_false (namesoffs + len > namesalloc))
1201 { 1286 {
1202 namesalloc *= 2; 1287 namesalloc *= 2;
1203 X_LOCK (wrklock); 1288 X_LOCK (wrklock);
1204 req->ptr1 = names = realloc (names, namesalloc); 1289 req->ptr2 = names = realloc (names, namesalloc);
1205 X_UNLOCK (wrklock); 1290 X_UNLOCK (wrklock);
1206 1291
1207 if (!names) 1292 if (!names)
1208 break; 1293 break;
1209 } 1294 }
… …
1216 1301
1217 if (expect_false (dentoffs == dentalloc)) 1302 if (expect_false (dentoffs == dentalloc))
1218 { 1303 {
1219 dentalloc *= 2; 1304 dentalloc *= 2;
1220 X_LOCK (wrklock); 1305 X_LOCK (wrklock);
1221 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1306 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1222 X_UNLOCK (wrklock); 1307 X_UNLOCK (wrklock);
1223 1308
1224 if (!dents) 1309 if (!dents)
1225 break; 1310 break;
1226 } 1311 }
1227 1312
1228 ent = dents + dentoffs; 1313 ent = dents + dentoffs;
1229 1314
1230 ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ 1315 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1231 ent->namelen = len - 1; 1316 ent->namelen = len - 1;
1232 ent->inode = D_INO (entp); 1317 ent->inode = D_INO (entp);
1318
1319 inode_bits |= ent->inode;
1233 1320
1234 switch (D_TYPE (entp)) 1321 switch (D_TYPE (entp))
1235 { 1322 {
1236 default: 1323 default:
1237 ent->type = EIO_DT_UNKNOWN; 1324 ent->type = EIO_DT_UNKNOWN;
… …
1280 #ifdef DT_WHT 1367 #ifdef DT_WHT
1281 case DT_WHT: ent->type = EIO_DT_WHT; break; 1368 case DT_WHT: ent->type = EIO_DT_WHT; break;
1282 #endif 1369 #endif
1283 } 1370 }
1284 1371
1285 ent->score = 0; 1372 ent->score = 7;
1286 1373
1287 if (flags & EIO_READDIR_DIRS_FIRST) 1374 if (flags & EIO_READDIR_DIRS_FIRST)
1288 { 1375 {
1289 if (ent->type == EIO_DT_UNKNOWN) 1376 if (ent->type == EIO_DT_UNKNOWN)
1290 { 1377 {
1291 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ 1378 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1292 ent->score = 98; 1379 ent->score = 1;
1293 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1380 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1294 ent->score = len <= 2 ? len + 6 : len <= 4 ? 5 : len <= 7 ? 4 : 1; /* shorter == more likely dir, but avoid too many classes */ 1381 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1295 } 1382 }
1296 else if (ent->type == EIO_DT_DIR) 1383 else if (ent->type == EIO_DT_DIR)
1297 ent->score = 100; 1384 ent->score = 0;
1298 } 1385 }
1299 } 1386 }
1300 1387
1301 namesoffs += len; 1388 namesoffs += len;
1302 ++dentoffs; 1389 ++dentoffs;
… …
1308 break; 1395 break;
1309 } 1396 }
1310 } 1397 }
1311} 1398}
1312 1399
1400#ifdef PAGESIZE
1401# define eio_pagesize() PAGESIZE
1402#else
1403static intptr_t
1404eio_pagesize (void)
1405{
1406 static intptr_t page;
1407
1408 if (!page)
1409 page = sysconf (_SC_PAGESIZE);
1410
1411 return page;
1412}
1413#endif
1414
1415static void
1416eio_page_align (void **addr, size_t *length)
1417{
1418 intptr_t mask = eio_pagesize () - 1;
1419
1420 /* round down addr */
1421 intptr_t adj = mask & (intptr_t)*addr;
1422
1423 *addr = (void *)((intptr_t)*addr - adj);
1424 *length += adj;
1425
1426 /* round up length */
1427 *length = (*length + mask) & ~mask;
1428}
1429
1430#if !_POSIX_MEMLOCK
1431# define eio__mlock(a,b) ((errno = ENOSYS), -1)
1432# define eio__mlockall(a) ((errno = ENOSYS), -1)
1433#else
1434
1435static int
1436eio__mlock (void *addr, size_t length)
1437{
1438 eio_page_align (&addr, &length);
1439
1440 mlock (addr, length);
1441}
1442
1443static int
1444eio__mlockall (int flags)
1445{
1446 #if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
1447 extern int mallopt (int, int);
1448 mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
1449 #endif
1450
1451 if (EIO_MCL_CURRENT != MCL_CURRENT
1452 || EIO_MCL_FUTURE != MCL_FUTURE)
1453 {
1454 flags = 0
1455 | (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
1456 | (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
1457 }
1458
1459 mlockall (flags);
1460}
1461#endif
1462
1313#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1463#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1314# undef msync
1315# define msync(a,b,c) ((errno = ENOSYS), -1) 1464# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1465#else
1466
1467int
1468eio__msync (void *mem, size_t len, int flags)
1469{
1470 eio_page_align (&mem, &len);
1471
1472 if (EIO_MS_ASYNC != MS_SYNC
1473 || EIO_MS_INVALIDATE != MS_INVALIDATE
1474 || EIO_MS_SYNC != MS_SYNC)
1475 {
1476 flags = 0
1477 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1478 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1479 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1480 }
1481
1482 return msync (mem, len, flags);
1483}
1484
1316#endif 1485#endif
1317 1486
1318int 1487int
1319eio__mtouch (void *mem, size_t len, int flags) 1488eio__mtouch (void *mem, size_t len, int flags)
1320{ 1489{
1490 eio_page_align (&mem, &len);
1491
1492 {
1321 intptr_t addr = (intptr_t)mem; 1493 intptr_t addr = (intptr_t)mem;
1322 intptr_t end = addr + len; 1494 intptr_t end = addr + len;
1323#ifdef PAGESIZE 1495 intptr_t page = eio_pagesize ();
1324 const intptr_t page = PAGESIZE;
1325#else
1326 static intptr_t page;
1327 1496
1328 if (!page)
1329 page = sysconf (_SC_PAGESIZE);
1330#endif
1331
1332 addr &= ~(page - 1); /* assume page size is always a power of two */
1333
1334 if (addr < end) 1497 if (addr < end)
1335 if (flags) /* modify */ 1498 if (flags & EIO_MT_MODIFY) /* modify */
1336 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); 1499 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1337 else 1500 else
1338 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); 1501 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1502 }
1339 1503
1340 return 0; 1504 return 0;
1341} 1505}
1342 1506
1343/*****************************************************************************/ 1507/*****************************************************************************/
… …
1488 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1652 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1489 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1653 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1490 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1654 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1491 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1655 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1492 1656
1657 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1658 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1659 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1660 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1661
1493 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; 1662 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1494 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; 1663 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1495 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; 1664 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1496 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; 1665 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
1497 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; 1666 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
… …
1506 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1675 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1507 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1676 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1508 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1677 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1509 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break; 1678 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1510 1679
1511 case EIO_READLINK: ALLOC (NAME_MAX); 1680 case EIO_READLINK: ALLOC (PATH_MAX);
1512 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1681 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1513 1682
1514 case EIO_SYNC: req->result = 0; sync (); break; 1683 case EIO_SYNC: req->result = 0; sync (); break;
1515 case EIO_FSYNC: req->result = fsync (req->int1); break; 1684 case EIO_FSYNC: req->result = fsync (req->int1); break;
1516 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1685 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1517 case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; 1686 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1518 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; 1687 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1688 case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
1689 case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
1519 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; 1690 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1520 1691
1521 case EIO_READDIR: eio__scandir (req, self); break; 1692 case EIO_READDIR: eio__scandir (req, self); break;
1522 1693
1523 case EIO_BUSY: 1694 case EIO_BUSY:
… …
1609eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) 1780eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1610{ 1781{
1611 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; 1782 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1612} 1783}
1613 1784
1785eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
1786{
1787 REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
1788}
1789
1790eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
1791{
1792 REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
1793}
1794
1614eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data) 1795eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1615{ 1796{
1616 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND; 1797 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1617} 1798}
1618 1799
… …
1642} 1823}
1643 1824
1644eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) 1825eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1645{ 1826{
1646 REQ (EIO_FSTAT); req->int1 = fd; SEND; 1827 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1828}
1829
1830eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1831{
1832 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1647} 1833}
1648 1834
1649eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) 1835eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1650{ 1836{
1651 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; 1837 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
… …
1723} 1909}
1724 1910
1725eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) 1911eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1726{ 1912{
1727 return eio__1path (EIO_LSTAT, path, pri, cb, data); 1913 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1914}
1915
1916eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
1917{
1918 return eio__1path (EIO_STATVFS, path, pri, cb, data);
1728} 1919}
1729 1920
1730eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) 1921eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1731{ 1922{
1732 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1923 return eio__1path (EIO_UNLINK, path, pri, cb, data);

Diff Legend

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