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

Comparing libeio/eio.c (file contents):
Revision 1.32 by root, Sat Jun 6 18:06:55 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
… …
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_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__
90# define _DIRENT_HAVE_D_TYPE /* sigh */
85# define D_INO(de) (de)->d_fileno 91# define D_INO(de) (de)->d_fileno
86# define _DIRENT_HAVE_D_TYPE /* sigh */ 92# define D_NAMLEN(de) (de)->d_namlen
87# elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 93# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
88# define D_INO(de) (de)->d_ino 94# define D_INO(de) (de)->d_ino
89# endif 95# endif
96
97#ifdef _D_EXACT_NAMLEN
98# undef D_NAMLEN
99# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
100#endif
90 101
91# ifdef _DIRENT_HAVE_D_TYPE 102# ifdef _DIRENT_HAVE_D_TYPE
92# define D_TYPE(de) (de)->d_type 103# define D_TYPE(de) (de)->d_type
93# endif 104# endif
94 105
… …
99#endif 110#endif
100 111
101#if HAVE_SENDFILE 112#if HAVE_SENDFILE
102# if __linux 113# if __linux
103# include <sys/sendfile.h> 114# include <sys/sendfile.h>
104# elif __freebsd 115# elif __FreeBSD__ || defined __APPLE__
105# include <sys/socket.h> 116# include <sys/socket.h>
106# include <sys/uio.h> 117# include <sys/uio.h>
107# elif __hpux 118# elif __hpux
108# include <sys/socket.h> 119# include <sys/socket.h>
109# elif __solaris /* not yet */ 120# elif __solaris
110# include <sys/sendfile.h> 121# include <sys/sendfile.h>
111# else 122# else
112# error sendfile support requested but not available 123# error sendfile support requested but not available
113# endif 124# endif
114#endif 125#endif
115 126
116#ifndef D_TYPE 127#ifndef D_TYPE
117# define D_TYPE(de) 0 128# define D_TYPE(de) 0
118#endif 129#endif
119
120#ifndef D_INO 130#ifndef D_INO
121# define D_INO(de) 0 131# define D_INO(de) 0
132#endif
133#ifndef D_NAMLEN
134# define D_NAMLEN(de) strlen ((de)->d_name)
122#endif 135#endif
123 136
124/* number of seconds after which an idle threads exit */ 137/* number of seconds after which an idle threads exit */
125#define IDLE_TIMEOUT 10 138#define IDLE_TIMEOUT 10
126 139
127/* used for struct dirent, AIX doesn't provide it */ 140/* used for struct dirent, AIX doesn't provide it */
128#ifndef NAME_MAX 141#ifndef NAME_MAX
129# 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
130#endif 148#endif
131 149
132/* buffer size for various temporary buffers */ 150/* buffer size for various temporary buffers */
133#define EIO_BUFSIZE 65536 151#define EIO_BUFSIZE 65536
134 152
… …
187 205
188/*****************************************************************************/ 206/*****************************************************************************/
189 207
190#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 208#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
191 209
192/* calculcate time difference in ~1/EIO_TICKS of a second */ 210/* calculate time difference in ~1/EIO_TICKS of a second */
193static int tvdiff (struct timeval *tv1, struct timeval *tv2) 211static int tvdiff (struct timeval *tv1, struct timeval *tv2)
194{ 212{
195 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 213 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
196 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 214 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
197} 215}
… …
207static volatile unsigned int nreqs; /* reqlock */ 225static volatile unsigned int nreqs; /* reqlock */
208static volatile unsigned int nready; /* reqlock */ 226static volatile unsigned int nready; /* reqlock */
209static volatile unsigned int npending; /* reqlock */ 227static volatile unsigned int npending; /* reqlock */
210static volatile unsigned int max_idle = 4; 228static volatile unsigned int max_idle = 4;
211 229
212static mutex_t wrklock = X_MUTEX_INIT; 230static xmutex_t wrklock = X_MUTEX_INIT;
213static mutex_t reslock = X_MUTEX_INIT; 231static xmutex_t reslock = X_MUTEX_INIT;
214static mutex_t reqlock = X_MUTEX_INIT; 232static xmutex_t reqlock = X_MUTEX_INIT;
215static cond_t reqwait = X_COND_INIT; 233static xcond_t reqwait = X_COND_INIT;
216 234
217#if !HAVE_PREADWRITE 235#if !HAVE_PREADWRITE
218/* 236/*
219 * make our pread/pwrite emulation safe against themselves, but not against 237 * make our pread/pwrite emulation safe against themselves, but not against
220 * 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,
221 * but that's your problem, not mine. 239 * but that's your problem, not mine.
222 */ 240 */
223static mutex_t preadwritelock = X_MUTEX_INIT; 241static xmutex_t preadwritelock = X_MUTEX_INIT;
224#endif 242#endif
225 243
226typedef struct etp_worker 244typedef struct etp_worker
227{ 245{
228 /* locked by wrklock */ 246 /* locked by wrklock */
229 struct etp_worker *prev, *next; 247 struct etp_worker *prev, *next;
230 248
231 thread_t tid; 249 xthread_t tid;
232 250
233 /* locked by reslock, reqlock or wrklock */ 251 /* locked by reslock, reqlock or wrklock */
234 ETP_REQ *req; /* currently processed request */ 252 ETP_REQ *req; /* currently processed request */
235 253
236 ETP_WORKER_COMMON 254 ETP_WORKER_COMMON
… …
589} 607}
590 608
591static void etp_set_max_poll_time (double nseconds) 609static void etp_set_max_poll_time (double nseconds)
592{ 610{
593 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 611 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
594 max_poll_time = nseconds; 612 max_poll_time = nseconds * EIO_TICKS;
595 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 613 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
596} 614}
597 615
598static void etp_set_max_poll_reqs (unsigned int maxreqs) 616static void etp_set_max_poll_reqs (unsigned int maxreqs)
599{ 617{
… …
858 if (!res || errno != ENOSYS) 876 if (!res || errno != ENOSYS)
859 return res; 877 return res;
860#endif 878#endif
861 879
862 /* 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
863 * call fdatasync, as thta matches the expectation of it's users best */ 881 * call fdatasync, as that matches the expectation of its users best */
864 return fdatasync (fd); 882 return fdatasync (fd);
865} 883}
866 884
867#if !HAVE_READAHEAD 885#if !HAVE_READAHEAD
868# undef readahead 886# undef readahead
… …
900 918
901#if HAVE_SENDFILE 919#if HAVE_SENDFILE
902# if __linux 920# if __linux
903 res = sendfile (ofd, ifd, &offset, count); 921 res = sendfile (ofd, ifd, &offset, count);
904 922
905# elif __freebsd 923# elif __FreeBSD__
906 /* 924 /*
907 * 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
908 * wasted on making it similar to other I/O functions. 926 * wasted on making it similar to other I/O functions.
909 */ 927 */
910 { 928 {
911 off_t sbytes; 929 off_t sbytes;
912 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 930 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
913 931
914 if (res < 0 && sbytes) 932 #if 0 /* according to the manpage, this is correct, but broken behaviour */
915 /* 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)
916 res = sbytes; 953 res = sbytes;
917 } 954 }
918 955
919# elif __hpux 956# elif __hpux
920 res = sendfile (ofd, ifd, offset, count, 0, 0); 957 res = sendfile (ofd, ifd, offset, count, 0, 0);
… …
934 if (res < 0 && sbytes) 971 if (res < 0 && sbytes)
935 res = sbytes; 972 res = sbytes;
936 } 973 }
937 974
938# 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
939#else 986#else
940 res = -1; 987 res = -1;
941 errno = ENOSYS; 988 errno = ENOSYS;
942#endif 989#endif
943 990
944 if (res < 0 991 if (res < 0
945 && (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 */
946#if __solaris 998#if __solaris
947 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 999 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
948#endif 1000#endif
949 ) 1001 )
950 ) 1002 )
… …
981 } 1033 }
982 1034
983 return res; 1035 return res;
984} 1036}
985 1037
986static int 1038static signed char
987eio_dent_cmp (const void *a_, const void *b_) 1039eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
988{ 1040{
989 const eio_dirent *a = (const eio_dirent *)a_; 1041 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
990 const eio_dirent *b = (const eio_dirent *)b_;
991
992 return (int)b->score - (int)a->score ? (int)b->score - (int)a->score
993 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; /* int might be < ino_t */ 1042 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1043}
1044
1045#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1046
1047#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1048#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1049
1050static void
1051eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
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;
1070 int i, j;
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{
1189 if (size <= 1)
1190 return; /* our insertion sort relies on size > 0 */
1191
1192 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1193 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1194 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1195
1196 /* use an insertion sort at the end, or for small arrays, */
1197 /* as insertion sort is more efficient for small partitions */
1198 eio_dent_insertion_sort (dents, size);
994} 1199}
995 1200
996/* read a full directory */ 1201/* read a full directory */
997static void 1202static void
998eio__scandir (eio_req *req, etp_worker *self) 1203eio__scandir (eio_req *req, etp_worker *self)
999{ 1204{
1000 DIR *dirp; 1205 DIR *dirp;
1001 EIO_STRUCT_DIRENT *entp; 1206 EIO_STRUCT_DIRENT *entp;
1002 unsigned char *name, *names; 1207 char *name, *names;
1003 int namesalloc = 4096; 1208 int namesalloc = 4096;
1004 int namesoffs = 0; 1209 int namesoffs = 0;
1005 int flags = req->int1; 1210 int flags = req->int1;
1006 eio_dirent *dents = 0; 1211 eio_dirent *dents = 0;
1007 int dentalloc = 128; 1212 int dentalloc = 128;
1008 int dentoffs = 0; 1213 int dentoffs = 0;
1214 ino_t inode_bits = 0;
1009 1215
1010 req->result = -1; 1216 req->result = -1;
1011 1217
1012 if (!(flags & EIO_READDIR_DENTS)) 1218 if (!(flags & EIO_READDIR_DENTS))
1013 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1219 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1014 1220
1015 X_LOCK (wrklock); 1221 X_LOCK (wrklock);
1016 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1222 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1017 self->dirp = dirp = opendir (req->ptr1); 1223 self->dirp = dirp = opendir (req->ptr1);
1018 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1224 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1019 req->ptr1 = names = malloc (namesalloc);
1020 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1225 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1226 req->ptr2 = names = malloc (namesalloc);
1021 X_UNLOCK (wrklock); 1227 X_UNLOCK (wrklock);
1022 1228
1023 if (dirp && names && (!flags || dents)) 1229 if (dirp && names && (!flags || dents))
1024 for (;;) 1230 for (;;)
1025 { 1231 {
… …
1033 1239
1034 /* sort etc. */ 1240 /* sort etc. */
1035 req->int1 = flags; 1241 req->int1 = flags;
1036 req->result = dentoffs; 1242 req->result = dentoffs;
1037 1243
1038 if (dents)
1039 {
1040 eio_dirent *ent = dents + dentoffs;
1041
1042 while (ent > dents)
1043 (--ent)->name = names + (size_t)ent->name;
1044 }
1045
1046 if (flags & EIO_READDIR_STAT_ORDER 1244 if (flags & EIO_READDIR_STAT_ORDER)
1047 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1245 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1048 {
1049 /* pray your qsort doesn't use quicksort */
1050 qsort (dents, dentoffs, sizeof (*dents), eio_dent_cmp); /* score depends of DIRS_FIRST */
1051 }
1052 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
1053 { 1250 {
1054 /* 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 */
1055 eio_dirent *ent = dents + dentoffs; 1252 eio_dirent *oth = dents + dentoffs;
1056 eio_dirent *dir = dents; 1253 eio_dirent *dir = dents;
1057 1254
1255 /* now partition dirs to the front, and non-dirs to the back */
1256 /* by walking from both sides and swapping if necessary */
1257 /* also clear score, so it doesn't influence sorting */
1058 while (ent > dir) 1258 while (oth > dir)
1059 { 1259 {
1060 if (dir->type == DT_DIR) 1260 if (dir->type == EIO_DT_DIR)
1061 ++dir; 1261 ++dir;
1062 else
1063 {
1064 --ent;
1065
1066 if (ent->type == DT_DIR) 1262 else if ((--oth)->type == EIO_DT_DIR)
1067 { 1263 {
1068 eio_dirent tmp = *dir; 1264 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1069 *dir = *ent;
1070 *ent = tmp;
1071 1265
1072 ++dir; 1266 ++dir;
1073 } 1267 }
1074 } 1268 }
1075 }
1076 1269
1077 /* now sort the dirs only */ 1270 /* now sort the dirs only */
1078 qsort (dents, dir - dents, sizeof (*dents), eio_dent_cmp); 1271 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1079 } 1272 }
1080
1081 /* only provide the names array unless DENTS is specified */
1082 if (!(flags & EIO_READDIR_DENTS))
1083 {
1084 X_LOCK (wrklock);
1085 assert (!dents);
1086 req->ptr1 = 0;
1087 req->ptr2 = names;
1088 X_UNLOCK (wrklock);
1089 }
1090 1273
1091 break; 1274 break;
1092 } 1275 }
1093 1276
1094 /* now add the entry to our list(s) */ 1277 /* now add the entry to our list(s) */
1095 name = entp->d_name; 1278 name = entp->d_name;
1096 1279
1097 /* skip . and .. entries */ 1280 /* skip . and .. entries */
1098 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) 1281 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
1099 { 1282 {
1100 int len = strlen (name) + 1; 1283 int len = D_NAMLEN (entp) + 1;
1101 1284
1102 while (expect_false (namesoffs + len > namesalloc)) 1285 while (expect_false (namesoffs + len > namesalloc))
1103 { 1286 {
1104 namesalloc *= 2; 1287 namesalloc *= 2;
1105 X_LOCK (wrklock); 1288 X_LOCK (wrklock);
1106 req->ptr1 = names = realloc (names, namesalloc); 1289 req->ptr2 = names = realloc (names, namesalloc);
1107 X_UNLOCK (wrklock); 1290 X_UNLOCK (wrklock);
1108 1291
1109 if (!names) 1292 if (!names)
1110 break; 1293 break;
1111 } 1294 }
… …
1118 1301
1119 if (expect_false (dentoffs == dentalloc)) 1302 if (expect_false (dentoffs == dentalloc))
1120 { 1303 {
1121 dentalloc *= 2; 1304 dentalloc *= 2;
1122 X_LOCK (wrklock); 1305 X_LOCK (wrklock);
1123 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1306 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1124 X_UNLOCK (wrklock); 1307 X_UNLOCK (wrklock);
1125 1308
1126 if (!dents) 1309 if (!dents)
1127 break; 1310 break;
1128 } 1311 }
1129 1312
1130 ent = dents + dentoffs; 1313 ent = dents + dentoffs;
1131 1314
1132 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 */
1133 ent->namelen = len - 1; 1316 ent->namelen = len - 1;
1134 ent->inode = D_INO (entp); 1317 ent->inode = D_INO (entp);
1318
1319 inode_bits |= ent->inode;
1135 1320
1136 switch (D_TYPE (entp)) 1321 switch (D_TYPE (entp))
1137 { 1322 {
1138 default: 1323 default:
1139 ent->type = EIO_DT_UNKNOWN; 1324 ent->type = EIO_DT_UNKNOWN;
… …
1144 case DT_FIFO: ent->type = EIO_DT_FIFO; break; 1329 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1145 #endif 1330 #endif
1146 #ifdef DT_CHR 1331 #ifdef DT_CHR
1147 case DT_CHR: ent->type = EIO_DT_CHR; break; 1332 case DT_CHR: ent->type = EIO_DT_CHR; break;
1148 #endif 1333 #endif
1334 #ifdef DT_MPC
1335 case DT_MPC: ent->type = EIO_DT_MPC; break;
1336 #endif
1149 #ifdef DT_DIR 1337 #ifdef DT_DIR
1150 case DT_DIR: ent->type = EIO_DT_DIR; break; 1338 case DT_DIR: ent->type = EIO_DT_DIR; break;
1339 #endif
1340 #ifdef DT_NAM
1341 case DT_NAM: ent->type = EIO_DT_NAM; break;
1151 #endif 1342 #endif
1152 #ifdef DT_BLK 1343 #ifdef DT_BLK
1153 case DT_BLK: ent->type = EIO_DT_BLK; break; 1344 case DT_BLK: ent->type = EIO_DT_BLK; break;
1154 #endif 1345 #endif
1346 #ifdef DT_MPB
1347 case DT_MPB: ent->type = EIO_DT_MPB; break;
1348 #endif
1155 #ifdef DT_REG 1349 #ifdef DT_REG
1156 case DT_REG: ent->type = EIO_DT_REG; break; 1350 case DT_REG: ent->type = EIO_DT_REG; break;
1351 #endif
1352 #ifdef DT_NWK
1353 case DT_NWK: ent->type = EIO_DT_NWK; break;
1354 #endif
1355 #ifdef DT_CMP
1356 case DT_CMP: ent->type = EIO_DT_CMP; break;
1157 #endif 1357 #endif
1158 #ifdef DT_LNK 1358 #ifdef DT_LNK
1159 case DT_LNK: ent->type = EIO_DT_LNK; break; 1359 case DT_LNK: ent->type = EIO_DT_LNK; break;
1160 #endif 1360 #endif
1161 #ifdef DT_SOCK 1361 #ifdef DT_SOCK
1162 case DT_SOCK: ent->type = EIO_DT_SOCK; break; 1362 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1163 #endif 1363 #endif
1364 #ifdef DT_DOOR
1365 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1366 #endif
1164 #ifdef DT_WHT 1367 #ifdef DT_WHT
1165 case DT_WHT: ent->type = EIO_DT_WHT; break; 1368 case DT_WHT: ent->type = EIO_DT_WHT; break;
1166 #endif 1369 #endif
1167 } 1370 }
1168 1371
1169 ent->score = 0; 1372 ent->score = 7;
1170 1373
1171 if (flags & EIO_READDIR_DIRS_FIRST) 1374 if (flags & EIO_READDIR_DIRS_FIRST)
1172 { 1375 {
1173 if (ent->type == EIO_DT_UNKNOWN) 1376 if (ent->type == EIO_DT_UNKNOWN)
1174 { 1377 {
1175 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 */
1176 ent->score = 98; 1379 ent->score = 1;
1177 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1380 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1178 ent->score = 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 */
1179 } 1382 }
1180 else if (ent->type == DT_DIR) 1383 else if (ent->type == EIO_DT_DIR)
1181 ent->score = 100; 1384 ent->score = 0;
1182 } 1385 }
1183 } 1386 }
1184 1387
1185 namesoffs += len; 1388 namesoffs += len;
1186 ++dentoffs; 1389 ++dentoffs;
1187 } 1390 }
1391
1392 if (EIO_CANCELLED (req))
1393 {
1394 errno = ECANCELED;
1395 break;
1396 }
1188 } 1397 }
1189 else
1190 req->result = -1;
1191} 1398}
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
1192 1462
1193#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1463#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1194# undef msync
1195# 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
1196#endif 1485#endif
1197 1486
1198int 1487int
1199eio__mtouch (void *mem, size_t len, int flags) 1488eio__mtouch (void *mem, size_t len, int flags)
1200{ 1489{
1490 eio_page_align (&mem, &len);
1491
1492 {
1201 intptr_t addr = (intptr_t)mem; 1493 intptr_t addr = (intptr_t)mem;
1202 intptr_t end = addr + len; 1494 intptr_t end = addr + len;
1203#ifdef PAGESIZE 1495 intptr_t page = eio_pagesize ();
1204 const intptr_t page = PAGESIZE;
1205#else
1206 static intptr_t page;
1207 1496
1208 if (!page)
1209 page = sysconf (_SC_PAGESIZE);
1210#endif
1211
1212 addr &= ~(page - 1); /* assume page size is always a power of two */
1213
1214 if (addr < end) 1497 if (addr < end)
1215 if (flags) /* modify */ 1498 if (flags & EIO_MT_MODIFY) /* modify */
1216 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); 1499 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1217 else 1500 else
1218 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); 1501 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1502 }
1219 1503
1220 return 0; 1504 return 0;
1221} 1505}
1222 1506
1223/*****************************************************************************/ 1507/*****************************************************************************/
… …
1368 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1652 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1369 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1653 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1370 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1654 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1371 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1655 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1372 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
1373 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;
1374 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;
1375 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;
1376 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;
1377 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; 1666 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
… …
1386 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1675 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1387 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1676 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1388 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1677 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1389 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;
1390 1679
1391 case EIO_READLINK: ALLOC (NAME_MAX); 1680 case EIO_READLINK: ALLOC (PATH_MAX);
1392 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1681 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1393 1682
1394 case EIO_SYNC: req->result = 0; sync (); break; 1683 case EIO_SYNC: req->result = 0; sync (); break;
1395 case EIO_FSYNC: req->result = fsync (req->int1); break; 1684 case EIO_FSYNC: req->result = fsync (req->int1); break;
1396 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1685 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1397 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;
1398 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;
1399 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;
1400 1691
1401 case EIO_READDIR: eio__scandir (req, self); break; 1692 case EIO_READDIR: eio__scandir (req, self); break;
1402 1693
1403 case EIO_BUSY: 1694 case EIO_BUSY:
… …
1489eio_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)
1490{ 1781{
1491 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;
1492} 1783}
1493 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
1494eio_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)
1495{ 1796{
1496 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;
1497} 1798}
1498 1799
… …
1522} 1823}
1523 1824
1524eio_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)
1525{ 1826{
1526 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;
1527} 1833}
1528 1834
1529eio_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)
1530{ 1836{
1531 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;
… …
1603} 1909}
1604 1910
1605eio_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)
1606{ 1912{
1607 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);
1608} 1919}
1609 1920
1610eio_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)
1611{ 1922{
1612 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1923 return eio__1path (EIO_UNLINK, path, pri, cb, data);
… …
1716/* misc garbage */ 2027/* misc garbage */
1717 2028
1718ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) 2029ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1719{ 2030{
1720 etp_worker wrk; 2031 etp_worker wrk;
2032 ssize_t ret;
1721 2033
1722 wrk.dbuf = 0; 2034 wrk.dbuf = 0;
1723 2035
1724 eio__sendfile (ofd, ifd, offset, count, &wrk); 2036 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1725 2037
1726 if (wrk.dbuf) 2038 if (wrk.dbuf)
1727 free (wrk.dbuf); 2039 free (wrk.dbuf);
1728}
1729 2040
2041 return ret;
2042}
2043

Diff Legend

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