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

Comparing libeio/eio.c (file contents):
Revision 1.41 by root, Sun Jun 14 20:36:59 2009 UTC vs.
Revision 1.55 by root, Wed Mar 31 00:46:18 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
… …
80# include <utime.h> 81# include <utime.h>
81# include <signal.h> 82# include <signal.h>
82# include <dirent.h> 83# include <dirent.h>
83 84
84/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ 85/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
85# if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) 86# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
86# define _DIRENT_HAVE_D_TYPE /* sigh */ 87# define _DIRENT_HAVE_D_TYPE /* sigh */
87# define D_INO(de) (de)->d_fileno 88# define D_INO(de) (de)->d_fileno
88# define D_NAMLEN(de) (de)->d_namlen 89# define D_NAMLEN(de) (de)->d_namlen
89# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 90# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
90# define D_INO(de) (de)->d_ino 91# define D_INO(de) (de)->d_ino
91# endif 92# endif
92 93
93#ifdef _D_EXACT_NAMLEN 94#ifdef _D_EXACT_NAMLEN
94# undef D_NAMLEN 95# undef D_NAMLEN
… …
106#endif 107#endif
107 108
108#if HAVE_SENDFILE 109#if HAVE_SENDFILE
109# if __linux 110# if __linux
110# include <sys/sendfile.h> 111# include <sys/sendfile.h>
111# elif __freebsd 112# elif __FreeBSD__ || defined __APPLE__
112# include <sys/socket.h> 113# include <sys/socket.h>
113# include <sys/uio.h> 114# include <sys/uio.h>
114# elif __hpux 115# elif __hpux
115# include <sys/socket.h> 116# include <sys/socket.h>
116# elif __solaris /* not yet */ 117# elif __solaris
117# include <sys/sendfile.h> 118# include <sys/sendfile.h>
118# else 119# else
119# error sendfile support requested but not available 120# error sendfile support requested but not available
120# endif 121# endif
121#endif 122#endif
… …
134#define IDLE_TIMEOUT 10 135#define IDLE_TIMEOUT 10
135 136
136/* used for struct dirent, AIX doesn't provide it */ 137/* used for struct dirent, AIX doesn't provide it */
137#ifndef NAME_MAX 138#ifndef NAME_MAX
138# define NAME_MAX 4096 139# define NAME_MAX 4096
140#endif
141
142/* used for readlink etc. */
143#ifndef PATH_MAX
144# define PATH_MAX 4096
139#endif 145#endif
140 146
141/* buffer size for various temporary buffers */ 147/* buffer size for various temporary buffers */
142#define EIO_BUFSIZE 65536 148#define EIO_BUFSIZE 65536
143 149
… …
196 202
197/*****************************************************************************/ 203/*****************************************************************************/
198 204
199#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 205#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
200 206
201/* calculcate time difference in ~1/EIO_TICKS of a second */ 207/* calculate time difference in ~1/EIO_TICKS of a second */
202static int tvdiff (struct timeval *tv1, struct timeval *tv2) 208static int tvdiff (struct timeval *tv1, struct timeval *tv2)
203{ 209{
204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 210 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
205 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 211 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
206} 212}
… …
216static volatile unsigned int nreqs; /* reqlock */ 222static volatile unsigned int nreqs; /* reqlock */
217static volatile unsigned int nready; /* reqlock */ 223static volatile unsigned int nready; /* reqlock */
218static volatile unsigned int npending; /* reqlock */ 224static volatile unsigned int npending; /* reqlock */
219static volatile unsigned int max_idle = 4; 225static volatile unsigned int max_idle = 4;
220 226
221static mutex_t wrklock = X_MUTEX_INIT; 227static xmutex_t wrklock = X_MUTEX_INIT;
222static mutex_t reslock = X_MUTEX_INIT; 228static xmutex_t reslock = X_MUTEX_INIT;
223static mutex_t reqlock = X_MUTEX_INIT; 229static xmutex_t reqlock = X_MUTEX_INIT;
224static cond_t reqwait = X_COND_INIT; 230static xcond_t reqwait = X_COND_INIT;
225 231
226#if !HAVE_PREADWRITE 232#if !HAVE_PREADWRITE
227/* 233/*
228 * make our pread/pwrite emulation safe against themselves, but not against 234 * make our pread/pwrite emulation safe against themselves, but not against
229 * normal read/write by using a mutex. slows down execution a lot, 235 * normal read/write by using a mutex. slows down execution a lot,
230 * but that's your problem, not mine. 236 * but that's your problem, not mine.
231 */ 237 */
232static mutex_t preadwritelock = X_MUTEX_INIT; 238static xmutex_t preadwritelock = X_MUTEX_INIT;
233#endif 239#endif
234 240
235typedef struct etp_worker 241typedef struct etp_worker
236{ 242{
237 /* locked by wrklock */ 243 /* locked by wrklock */
238 struct etp_worker *prev, *next; 244 struct etp_worker *prev, *next;
239 245
240 thread_t tid; 246 xthread_t tid;
241 247
242 /* locked by reslock, reqlock or wrklock */ 248 /* locked by reslock, reqlock or wrklock */
243 ETP_REQ *req; /* currently processed request */ 249 ETP_REQ *req; /* currently processed request */
244 250
245 ETP_WORKER_COMMON 251 ETP_WORKER_COMMON
… …
598} 604}
599 605
600static void etp_set_max_poll_time (double nseconds) 606static void etp_set_max_poll_time (double nseconds)
601{ 607{
602 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 608 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
603 max_poll_time = nseconds; 609 max_poll_time = nseconds * EIO_TICKS;
604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 610 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
605} 611}
606 612
607static void etp_set_max_poll_reqs (unsigned int maxreqs) 613static void etp_set_max_poll_reqs (unsigned int maxreqs)
608{ 614{
… …
867 if (!res || errno != ENOSYS) 873 if (!res || errno != ENOSYS)
868 return res; 874 return res;
869#endif 875#endif
870 876
871 /* even though we could play tricks with the flags, it's better to always 877 /* 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 */ 878 * call fdatasync, as that matches the expectation of its users best */
873 return fdatasync (fd); 879 return fdatasync (fd);
874} 880}
875 881
876#if !HAVE_READAHEAD 882#if !HAVE_READAHEAD
877# undef readahead 883# undef readahead
… …
909 915
910#if HAVE_SENDFILE 916#if HAVE_SENDFILE
911# if __linux 917# if __linux
912 res = sendfile (ofd, ifd, &offset, count); 918 res = sendfile (ofd, ifd, &offset, count);
913 919
914# elif __freebsd 920# elif __FreeBSD__
915 /* 921 /*
916 * Of course, the freebsd sendfile is a dire hack with no thoughts 922 * Of course, the freebsd sendfile is a dire hack with no thoughts
917 * wasted on making it similar to other I/O functions. 923 * wasted on making it similar to other I/O functions.
918 */ 924 */
919 { 925 {
920 off_t sbytes; 926 off_t sbytes;
921 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 927 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
922 928
923 if (res < 0 && sbytes) 929 #if 0 /* according to the manpage, this is correct, but broken behaviour */
924 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 930 /* freebsd' sendfile will return 0 on success */
931 /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
932 /* not on e.g. EIO or EPIPE - sounds broken */
933 if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
934 res = sbytes;
935 #endif
936
937 /* according to source inspection, this is correct, and useful behaviour */
938 if (sbytes)
939 res = sbytes;
940 }
941
942# elif defined (__APPLE__)
943
944 {
945 off_t sbytes = count;
946 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
947
948 /* according to the manpage, sbytes is always valid */
949 if (sbytes)
925 res = sbytes; 950 res = sbytes;
926 } 951 }
927 952
928# elif __hpux 953# elif __hpux
929 res = sendfile (ofd, ifd, offset, count, 0, 0); 954 res = sendfile (ofd, ifd, offset, count, 0, 0);
… …
943 if (res < 0 && sbytes) 968 if (res < 0 && sbytes)
944 res = sbytes; 969 res = sbytes;
945 } 970 }
946 971
947# endif 972# endif
973
974#elif defined (_WIN32)
975
976 /* does not work, just for documentation of what would need to be done */
977 {
978 HANDLE h = TO_SOCKET (ifd);
979 SetFilePointer (h, offset, 0, FILE_BEGIN);
980 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
981 }
982
948#else 983#else
949 res = -1; 984 res = -1;
950 errno = ENOSYS; 985 errno = ENOSYS;
951#endif 986#endif
952 987
953 if (res < 0 988 if (res < 0
954 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 989 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
990 /* BSDs */
991#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
992 || errno == ENOTSUP
993#endif
994 || errno == EOPNOTSUPP /* BSDs */
955#if __solaris 995#if __solaris
956 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 996 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
957#endif 997#endif
958 ) 998 )
959 ) 999 )
… …
993} 1033}
994 1034
995static signed char 1035static signed char
996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) 1036eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
997{ 1037{
998 return b->score - a->score ? b->score - a->score /* works because our signed char is always 0..100 */ 1038 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; 1039 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1000} 1040}
1001 1041
1002#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 1042#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1003#define EIO_DENT_NUM_SCORES 9
1004 1043
1005#define EIO_QSORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ 1044#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1006#define EIO_QSORT_SKIP 60 /* when to skip qsort completely */ 1045#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1007 1046
1008static void 1047static void
1009eio_dent_sort (eio_dirent *dents, int size) 1048eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1010{ 1049{
1011 if (size <= 1) 1050 unsigned char bits [9 + sizeof (ino_t) * 8];
1012 return; /* our insertion sort relies on size > 0 */ 1051 unsigned char *bit = bits;
1013 1052
1014 if (size > EIO_QSORT_SKIP) /* skip quicksort for small directories */ 1053 assert (CHAR_BIT == 8);
1054 assert (sizeof (eio_dirent) * 8 < 256);
1055 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1056 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1057
1058 if (size <= EIO_SORT_FAST)
1059 return;
1060
1061 /* first prepare an array of bits to test in our radix sort */
1062 /* try to take endianness into account, as well as differences in ino_t sizes */
1063 /* inode_bits must contain all inodes ORed together */
1064 /* which is used to skip bits that are 0 everywhere, which is very common */
1015 { 1065 {
1016 /* first, use quicksort */ 1066 ino_t endianness;
1017 /* should be good for 2**31 entries */
1018 int i; 1067 int i, j;
1019 struct rng { eio_dirent *l, *r; } rng [32];
1020 1068
1021 i = 0; 1069 /* we store the byte offset of byte n into byte n of "endianness" */
1022 rng[0].l = dents; 1070 for (i = 0; i < sizeof (ino_t); ++i)
1023 rng[0].r = dents + size; 1071 ((unsigned char *)&endianness)[i] = i;
1024 1072
1025 while (expect_true (i >= 0)) 1073 *bit++ = 0;
1074
1075 for (i = 0; i < sizeof (ino_t); ++i)
1026 { 1076 {
1027 eio_dirent *L = rng [i].l; 1077 /* shifting off the byte offsets out of "endianness" */
1028 eio_dirent *R = rng [i].r - 1; 1078 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1079 endianness >>= 8;
1029 1080
1030 if (expect_false (L + EIO_QSORT_CUTOFF < R)) 1081 for (j = 0; j < 8; ++j)
1082 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1083 *bit++ = offs + j;
1084 }
1085
1086 for (j = 0; j < 8; ++j)
1087 if (score_bits & (1 << j))
1088 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1089 }
1090
1091 /* now actually do the sorting (a variant of MSD radix sort) */
1092 {
1093 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1094 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1095 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1096 int stk_idx = 0;
1097
1098 base_stk [stk_idx] = dents;
1099 end_stk [stk_idx] = dents + size;
1100 bit_stk [stk_idx] = bit - 1;
1101
1102 do
1103 {
1104 base = base_stk [stk_idx];
1105 end = end_stk [stk_idx];
1106 bit = bit_stk [stk_idx];
1107
1108 for (;;)
1031 { 1109 {
1032 eio_dirent *mid = &dents [((L - dents) + (R - dents)) >> 1]; 1110 unsigned char O = *bit >> 3;
1033 eio_dirent piv = *mid; *mid = *L; *L = piv; 1111 unsigned char M = 1 << (*bit & 7);
1034 1112
1035 while (L < R) 1113 eio_dirent *a = base;
1114 eio_dirent *b = end;
1115
1116 if (b - a < EIO_SORT_CUTOFF)
1117 break;
1118
1119 /* now bit-partition the array on the bit */
1120 /* this ugly asymmetric loop seems to perform much better than typical */
1121 /* partition algos found in the literature */
1122 do
1123 if (!(((unsigned char *)a)[O] & M))
1124 ++a;
1125 else if (!(((unsigned char *)--b)[O] & M))
1036 { 1126 {
1037 while (EIO_DENT_CMP (*R, >=, piv) && L < R) 1127 eio_dirent tmp = *a; *a = *b; *b = tmp;
1038 --R;
1039
1040 if (L < R)
1041 *L++ = *R;
1042
1043 while (EIO_DENT_CMP (*L, <=, piv) && L < R)
1044 ++L; 1128 ++a;
1045
1046 if (L < R)
1047 *R-- = *L;
1048 } 1129 }
1130 while (b > a);
1049 1131
1050 *L = piv; 1132 /* next bit, or stop, if no bits left in this path */
1133 if (!*--bit)
1134 break;
1051 1135
1136 base_stk [stk_idx] = a;
1137 end_stk [stk_idx] = end;
1138 bit_stk [stk_idx] = bit;
1052 ++i; 1139 ++stk_idx;
1053 rng [i].l = L + 1;
1054 rng [i].r = rng [i - 1].r;
1055 rng [i - 1].r = L;
1056 1140
1057 if (rng [i].r - rng [i].l > rng [i - 1].r - rng [i - 1].l) 1141 end = a;
1058 {
1059 struct rng t;
1060
1061 t = rng [i]; rng [i] = rng [i - 1]; rng [i - 1] = t;
1062 }
1063 } 1142 }
1064 else
1065 --i;
1066 } 1143 }
1144 while (stk_idx--);
1067 } 1145 }
1146}
1068 1147
1069 /* use an insertion sort after qsort, or for small arrays */ 1148static void
1149eio_dent_insertion_sort (eio_dirent *dents, int size)
1150{
1070 /* first move the smallest element to the front, to act as a sentinel */ 1151 /* first move the smallest element to the front, to act as a sentinel */
1071 { 1152 {
1072 int i; 1153 int i;
1073 eio_dirent *min = dents; 1154 eio_dirent *min = dents;
1074 1155
1156 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1075 for (i = size > EIO_QSORT_SKIP ? EIO_QSORT_CUTOFF + 1 : size; --i; ) 1157 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1076 if (EIO_DENT_CMP (dents [i], <, *min)) 1158 if (EIO_DENT_CMP (dents [i], <, *min))
1077 min = &dents [i]; 1159 min = &dents [i];
1078 1160
1079 /* swap elements 0 and j (minimum) */ 1161 /* swap elements 0 and j (minimum) */
1080 { 1162 {
1081 eio_dirent tmp = *dents; *dents = *min; *min = tmp; 1163 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1082 } 1164 }
1083 } 1165 }
1084 1166
1167 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1085 { 1168 {
1086 /* then do standard insertion sort */
1087 eio_dirent *i, *j; 1169 eio_dirent *i, *j;
1088 1170
1089 for (i = dents + 1; i < dents + size; ++i) 1171 for (i = dents + 1; i < dents + size; ++i)
1090 { 1172 {
1091 eio_dirent value = *i; 1173 eio_dirent value = *i;
… …
1094 j [1] = j [0]; 1176 j [1] = j [0];
1095 1177
1096 j [1] = value; 1178 j [1] = value;
1097 } 1179 }
1098 } 1180 }
1181}
1182
1183static void
1184eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1185{
1186 if (size <= 1)
1187 return; /* our insertion sort relies on size > 0 */
1188
1189 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1190 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1191 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1192
1193 /* use an insertion sort at the end, or for small arrays, */
1194 /* as insertion sort is more efficient for small partitions */
1195 eio_dent_insertion_sort (dents, size);
1099} 1196}
1100 1197
1101/* read a full directory */ 1198/* read a full directory */
1102static void 1199static void
1103eio__scandir (eio_req *req, etp_worker *self) 1200eio__scandir (eio_req *req, etp_worker *self)
… …
1109 int namesoffs = 0; 1206 int namesoffs = 0;
1110 int flags = req->int1; 1207 int flags = req->int1;
1111 eio_dirent *dents = 0; 1208 eio_dirent *dents = 0;
1112 int dentalloc = 128; 1209 int dentalloc = 128;
1113 int dentoffs = 0; 1210 int dentoffs = 0;
1211 ino_t inode_bits = 0;
1114 1212
1115 req->result = -1; 1213 req->result = -1;
1116 1214
1117 if (!(flags & EIO_READDIR_DENTS)) 1215 if (!(flags & EIO_READDIR_DENTS))
1118 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1216 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1119 1217
1120 X_LOCK (wrklock); 1218 X_LOCK (wrklock);
1121 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1219 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1122 self->dirp = dirp = opendir (req->ptr1); 1220 self->dirp = dirp = opendir (req->ptr1);
1123 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1221 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1124 req->ptr1 = names = malloc (namesalloc);
1125 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1222 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1223 req->ptr2 = names = malloc (namesalloc);
1126 X_UNLOCK (wrklock); 1224 X_UNLOCK (wrklock);
1127 1225
1128 if (dirp && names && (!flags || dents)) 1226 if (dirp && names && (!flags || dents))
1129 for (;;) 1227 for (;;)
1130 { 1228 {
… …
1138 1236
1139 /* sort etc. */ 1237 /* sort etc. */
1140 req->int1 = flags; 1238 req->int1 = flags;
1141 req->result = dentoffs; 1239 req->result = dentoffs;
1142 1240
1143 if (dents)
1144 {
1145 eio_dirent *ent = dents + dentoffs;
1146
1147 while (ent > dents)
1148 (--ent)->name = names + (size_t)ent->name;
1149 }
1150
1151 if (flags & EIO_READDIR_STAT_ORDER 1241 if (flags & EIO_READDIR_STAT_ORDER)
1152 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1242 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1153 eio_dent_sort (dents, dentoffs); /* score depends of DIRS_FIRST */
1154 else if (flags & EIO_READDIR_DIRS_FIRST) 1243 else if (flags & EIO_READDIR_DIRS_FIRST)
1244 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1245 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1246 else
1155 { 1247 {
1156 /* in this case, all is known, and we just put dirs first and sort them */ 1248 /* in this case, all is known, and we just put dirs first and sort them */
1157 eio_dirent *ent = dents + dentoffs; 1249 eio_dirent *oth = dents + dentoffs;
1158 eio_dirent *dir = dents; 1250 eio_dirent *dir = dents;
1159 1251
1160 /* now move dirs to the front, and non-dirs to the back */ 1252 /* now partition dirs to the front, and non-dirs to the back */
1161 /* by walking from both sides and swapping if necessary */ 1253 /* by walking from both sides and swapping if necessary */
1254 /* also clear score, so it doesn't influence sorting */
1162 while (ent > dir) 1255 while (oth > dir)
1163 {
1164 if (dir->type == DT_DIR)
1165 { 1256 {
1166 dir->score = 0; 1257 if (dir->type == EIO_DT_DIR)
1167 ++dir; 1258 ++dir;
1168 }
1169 else
1170 {
1171 --ent;
1172
1173 if (ent->type == DT_DIR) 1259 else if ((--oth)->type == EIO_DT_DIR)
1174 { 1260 {
1175 eio_dirent tmp = *dir; 1261 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1176 *dir = *ent;
1177 *ent = tmp;
1178 1262
1179 dir->score = 0;
1180 ++dir; 1263 ++dir;
1181 } 1264 }
1182 } 1265 }
1183 }
1184 1266
1185 /* now sort the dirs only */ 1267 /* now sort the dirs only */
1186 eio_dent_sort (dents, dir - dents); 1268 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1187 } 1269 }
1188
1189 /* only provide the names array unless DENTS is specified */
1190 if (!(flags & EIO_READDIR_DENTS))
1191 {
1192 X_LOCK (wrklock);
1193 assert (!dents);
1194 req->ptr1 = 0;
1195 req->ptr2 = names;
1196 X_UNLOCK (wrklock);
1197 }
1198 1270
1199 break; 1271 break;
1200 } 1272 }
1201 1273
1202 /* now add the entry to our list(s) */ 1274 /* now add the entry to our list(s) */
… …
1209 1281
1210 while (expect_false (namesoffs + len > namesalloc)) 1282 while (expect_false (namesoffs + len > namesalloc))
1211 { 1283 {
1212 namesalloc *= 2; 1284 namesalloc *= 2;
1213 X_LOCK (wrklock); 1285 X_LOCK (wrklock);
1214 req->ptr1 = names = realloc (names, namesalloc); 1286 req->ptr2 = names = realloc (names, namesalloc);
1215 X_UNLOCK (wrklock); 1287 X_UNLOCK (wrklock);
1216 1288
1217 if (!names) 1289 if (!names)
1218 break; 1290 break;
1219 } 1291 }
… …
1226 1298
1227 if (expect_false (dentoffs == dentalloc)) 1299 if (expect_false (dentoffs == dentalloc))
1228 { 1300 {
1229 dentalloc *= 2; 1301 dentalloc *= 2;
1230 X_LOCK (wrklock); 1302 X_LOCK (wrklock);
1231 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1303 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1232 X_UNLOCK (wrklock); 1304 X_UNLOCK (wrklock);
1233 1305
1234 if (!dents) 1306 if (!dents)
1235 break; 1307 break;
1236 } 1308 }
1237 1309
1238 ent = dents + dentoffs; 1310 ent = dents + dentoffs;
1239 1311
1240 ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ 1312 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1241 ent->namelen = len - 1; 1313 ent->namelen = len - 1;
1242 ent->inode = D_INO (entp) ? D_INO (entp) : dentoffs; 1314 ent->inode = D_INO (entp);
1315
1316 inode_bits |= ent->inode;
1243 1317
1244 switch (D_TYPE (entp)) 1318 switch (D_TYPE (entp))
1245 { 1319 {
1246 default: 1320 default:
1247 ent->type = EIO_DT_UNKNOWN; 1321 ent->type = EIO_DT_UNKNOWN;
… …
1290 #ifdef DT_WHT 1364 #ifdef DT_WHT
1291 case DT_WHT: ent->type = EIO_DT_WHT; break; 1365 case DT_WHT: ent->type = EIO_DT_WHT; break;
1292 #endif 1366 #endif
1293 } 1367 }
1294 1368
1295 ent->score = 0; 1369 ent->score = 7;
1296 1370
1297 if (flags & EIO_READDIR_DIRS_FIRST) 1371 if (flags & EIO_READDIR_DIRS_FIRST)
1298 { 1372 {
1299 if (ent->type == EIO_DT_UNKNOWN) 1373 if (ent->type == EIO_DT_UNKNOWN)
1300 { 1374 {
1301 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ 1375 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1302 ent->score = 7; 1376 ent->score = 1;
1303 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1377 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1304 ent->score = len <= 2 ? len + 4 : len <= 4 ? 3 : len <= 7 ? 2 : 1; /* shorter == more likely dir, but avoid too many classes */ 1378 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1305 } 1379 }
1306 else if (ent->type == EIO_DT_DIR) 1380 else if (ent->type == EIO_DT_DIR)
1307 ent->score = 8; 1381 ent->score = 0;
1308 } 1382 }
1309 } 1383 }
1310 1384
1311 namesoffs += len; 1385 namesoffs += len;
1312 ++dentoffs; 1386 ++dentoffs;
… …
1319 } 1393 }
1320 } 1394 }
1321} 1395}
1322 1396
1323#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1397#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1324# undef msync
1325# define msync(a,b,c) ((errno = ENOSYS), -1) 1398# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1399#else
1400
1401int
1402eio__msync (void *mem, size_t len, int flags)
1403{
1404 if (EIO_MS_ASYNC != MS_SYNC
1405 || EIO_MS_INVALIDATE != MS_INVALIDATE
1406 || EIO_MS_SYNC != MS_SYNC)
1407 {
1408 flags = 0
1409 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1410 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1411 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1412 }
1413
1414 return msync (mem, len, flags);
1415}
1416
1326#endif 1417#endif
1327 1418
1328int 1419int
1329eio__mtouch (void *mem, size_t len, int flags) 1420eio__mtouch (void *mem, size_t len, int flags)
1330{ 1421{
… …
1337 1428
1338 if (!page) 1429 if (!page)
1339 page = sysconf (_SC_PAGESIZE); 1430 page = sysconf (_SC_PAGESIZE);
1340#endif 1431#endif
1341 1432
1433 /* round down to start of page, although this is probably useless */
1342 addr &= ~(page - 1); /* assume page size is always a power of two */ 1434 addr &= ~(page - 1); /* assume page size is always a power of two */
1343 1435
1344 if (addr < end) 1436 if (addr < end)
1345 if (flags) /* modify */ 1437 if (flags & EIO_MT_MODIFY) /* modify */
1346 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); 1438 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1347 else 1439 else
1348 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); 1440 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1349 1441
1350 return 0; 1442 return 0;
… …
1498 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1590 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1499 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1591 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1500 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1592 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1501 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1593 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1502 1594
1595 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1596 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1597 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1598 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1599
1503 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; 1600 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1504 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; 1601 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1505 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; 1602 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1506 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; 1603 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
1507 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; 1604 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
… …
1516 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1613 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1517 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1614 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1518 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1615 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1519 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break; 1616 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1520 1617
1521 case EIO_READLINK: ALLOC (NAME_MAX); 1618 case EIO_READLINK: ALLOC (PATH_MAX);
1522 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1619 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1523 1620
1524 case EIO_SYNC: req->result = 0; sync (); break; 1621 case EIO_SYNC: req->result = 0; sync (); break;
1525 case EIO_FSYNC: req->result = fsync (req->int1); break; 1622 case EIO_FSYNC: req->result = fsync (req->int1); break;
1526 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1623 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1527 case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; 1624 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1528 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; 1625 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1529 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; 1626 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1530 1627
1531 case EIO_READDIR: eio__scandir (req, self); break; 1628 case EIO_READDIR: eio__scandir (req, self); break;
1532 1629
… …
1654eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) 1751eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1655{ 1752{
1656 REQ (EIO_FSTAT); req->int1 = fd; SEND; 1753 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1657} 1754}
1658 1755
1756eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1757{
1758 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1759}
1760
1659eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) 1761eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1660{ 1762{
1661 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; 1763 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
1662} 1764}
1663 1765
… …
1733} 1835}
1734 1836
1735eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) 1837eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1736{ 1838{
1737 return eio__1path (EIO_LSTAT, path, pri, cb, data); 1839 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1840}
1841
1842eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
1843{
1844 return eio__1path (EIO_STATVFS, path, pri, cb, data);
1738} 1845}
1739 1846
1740eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) 1847eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1741{ 1848{
1742 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1849 return eio__1path (EIO_UNLINK, path, pri, cb, data);

Diff Legend

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