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

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.37 by root, Fri Jun 12 16:48:08 2009 UTC vs.
Revision 1.59 by root, Wed Dec 1 07:30:39 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{
… …
807 816
808 return res; 817 return res;
809} 818}
810#endif 819#endif
811 820
812#ifndef HAVE_FUTIMES 821#ifndef HAVE_UTIMES
813 822
814# undef utimes 823# undef utimes
815# undef futimes
816# define utimes(path,times) eio__utimes (path, times) 824# define utimes(path,times) eio__utimes (path, times)
817# define futimes(fd,times) eio__futimes (fd, times)
818 825
819static int 826static int
820eio__utimes (const char *filename, const struct timeval times[2]) 827eio__utimes (const char *filename, const struct timeval times[2])
821{ 828{
822 if (times) 829 if (times)
… …
829 return utime (filename, &buf); 836 return utime (filename, &buf);
830 } 837 }
831 else 838 else
832 return utime (filename, 0); 839 return utime (filename, 0);
833} 840}
841
842#endif
843
844#ifndef HAVE_FUTIMES
845
846# undef futimes
847# define futimes(fd,times) eio__futimes (fd, times)
834 848
835static int eio__futimes (int fd, const struct timeval tv[2]) 849static int eio__futimes (int fd, const struct timeval tv[2])
836{ 850{
837 errno = ENOSYS; 851 errno = ENOSYS;
838 return -1; 852 return -1;
… …
867 if (!res || errno != ENOSYS) 881 if (!res || errno != ENOSYS)
868 return res; 882 return res;
869#endif 883#endif
870 884
871 /* even though we could play tricks with the flags, it's better to always 885 /* even though we could play tricks with the flags, it's better to always
872 * call fdatasync, as thta matches the expectation of it's users best */ 886 * call fdatasync, as that matches the expectation of its users best */
873 return fdatasync (fd); 887 return fdatasync (fd);
874} 888}
875 889
876#if !HAVE_READAHEAD 890#if !HAVE_READAHEAD
877# undef readahead 891# undef readahead
… …
909 923
910#if HAVE_SENDFILE 924#if HAVE_SENDFILE
911# if __linux 925# if __linux
912 res = sendfile (ofd, ifd, &offset, count); 926 res = sendfile (ofd, ifd, &offset, count);
913 927
914# elif __freebsd 928# elif __FreeBSD__
915 /* 929 /*
916 * Of course, the freebsd sendfile is a dire hack with no thoughts 930 * Of course, the freebsd sendfile is a dire hack with no thoughts
917 * wasted on making it similar to other I/O functions. 931 * wasted on making it similar to other I/O functions.
918 */ 932 */
919 { 933 {
920 off_t sbytes; 934 off_t sbytes;
921 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 935 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
922 936
923 if (res < 0 && sbytes) 937 #if 0 /* according to the manpage, this is correct, but broken behaviour */
924 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 938 /* freebsd' sendfile will return 0 on success */
939 /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
940 /* not on e.g. EIO or EPIPE - sounds broken */
941 if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
942 res = sbytes;
943 #endif
944
945 /* according to source inspection, this is correct, and useful behaviour */
946 if (sbytes)
947 res = sbytes;
948 }
949
950# elif defined (__APPLE__)
951
952 {
953 off_t sbytes = count;
954 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
955
956 /* according to the manpage, sbytes is always valid */
957 if (sbytes)
925 res = sbytes; 958 res = sbytes;
926 } 959 }
927 960
928# elif __hpux 961# elif __hpux
929 res = sendfile (ofd, ifd, offset, count, 0, 0); 962 res = sendfile (ofd, ifd, offset, count, 0, 0);
… …
943 if (res < 0 && sbytes) 976 if (res < 0 && sbytes)
944 res = sbytes; 977 res = sbytes;
945 } 978 }
946 979
947# endif 980# endif
981
982#elif defined (_WIN32)
983
984 /* does not work, just for documentation of what would need to be done */
985 {
986 HANDLE h = TO_SOCKET (ifd);
987 SetFilePointer (h, offset, 0, FILE_BEGIN);
988 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
989 }
990
948#else 991#else
949 res = -1; 992 res = -1;
950 errno = ENOSYS; 993 errno = ENOSYS;
951#endif 994#endif
952 995
953 if (res < 0 996 if (res < 0
954 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 997 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
998 /* BSDs */
999#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
1000 || errno == ENOTSUP
1001#endif
1002 || errno == EOPNOTSUPP /* BSDs */
955#if __solaris 1003#if __solaris
956 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 1004 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
957#endif 1005#endif
958 ) 1006 )
959 ) 1007 )
… …
993} 1041}
994 1042
995static signed char 1043static signed char
996eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) 1044eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
997{ 1045{
998 return b->score - a->score ? b->score - a->score /* works because our signed char is always 0..100 */ 1046 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; 1047 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1000} 1048}
1001 1049
1050#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1051
1002#define EIO_QSORT_CUTOFF 20 /* quite high, but performs well on many filesystems */ 1052#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1053#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1003 1054
1004static void 1055static void
1005eio_dent_sort (eio_dirent *dents, int size) 1056eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1006{ 1057{
1007 /* should be good for 2**31 entries */ 1058 unsigned char bits [9 + sizeof (ino_t) * 8];
1008 struct rng { int l, r; } rng [32]; 1059 unsigned char *bit = bits;
1060
1061 assert (CHAR_BIT == 8);
1062 assert (sizeof (eio_dirent) * 8 < 256);
1063 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1064 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1065
1066 if (size <= EIO_SORT_FAST)
1067 return;
1068
1069 /* first prepare an array of bits to test in our radix sort */
1070 /* try to take endianness into account, as well as differences in ino_t sizes */
1071 /* inode_bits must contain all inodes ORed together */
1072 /* which is used to skip bits that are 0 everywhere, which is very common */
1073 {
1074 ino_t endianness;
1009 int i, j; 1075 int i, j;
1010 1076
1011 i = 0; 1077 /* we store the byte offset of byte n into byte n of "endianness" */
1012 rng[0].l = 0; 1078 for (i = 0; i < sizeof (ino_t); ++i)
1013 rng[0].r = size; 1079 ((unsigned char *)&endianness)[i] = i;
1014 1080
1015 while (expect_true (i >= 0)) 1081 *bit++ = 0;
1082
1083 for (i = 0; i < sizeof (ino_t); ++i)
1084 {
1085 /* shifting off the byte offsets out of "endianness" */
1086 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1087 endianness >>= 8;
1088
1089 for (j = 0; j < 8; ++j)
1090 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1091 *bit++ = offs + j;
1092 }
1093
1094 for (j = 0; j < 8; ++j)
1095 if (score_bits & (1 << j))
1096 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1097 }
1098
1099 /* now actually do the sorting (a variant of MSD radix sort) */
1100 {
1101 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1102 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1103 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1104 int stk_idx = 0;
1105
1106 base_stk [stk_idx] = dents;
1107 end_stk [stk_idx] = dents + size;
1108 bit_stk [stk_idx] = bit - 1;
1109
1110 do
1111 {
1112 base = base_stk [stk_idx];
1113 end = end_stk [stk_idx];
1114 bit = bit_stk [stk_idx];
1115
1116 for (;;)
1117 {
1118 unsigned char O = *bit >> 3;
1119 unsigned char M = 1 << (*bit & 7);
1120
1121 eio_dirent *a = base;
1122 eio_dirent *b = end;
1123
1124 if (b - a < EIO_SORT_CUTOFF)
1125 break;
1126
1127 /* now bit-partition the array on the bit */
1128 /* this ugly asymmetric loop seems to perform much better than typical */
1129 /* partition algos found in the literature */
1130 do
1131 if (!(((unsigned char *)a)[O] & M))
1132 ++a;
1133 else if (!(((unsigned char *)--b)[O] & M))
1134 {
1135 eio_dirent tmp = *a; *a = *b; *b = tmp;
1136 ++a;
1137 }
1138 while (b > a);
1139
1140 /* next bit, or stop, if no bits left in this path */
1141 if (!*--bit)
1142 break;
1143
1144 base_stk [stk_idx] = a;
1145 end_stk [stk_idx] = end;
1146 bit_stk [stk_idx] = bit;
1147 ++stk_idx;
1148
1149 end = a;
1150 }
1151 }
1152 while (stk_idx--);
1153 }
1154}
1155
1156static void
1157eio_dent_insertion_sort (eio_dirent *dents, int size)
1158{
1159 /* first move the smallest element to the front, to act as a sentinel */
1160 {
1161 int i;
1162 eio_dirent *min = dents;
1163
1164 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1165 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1166 if (EIO_DENT_CMP (dents [i], <, *min))
1167 min = &dents [i];
1168
1169 /* swap elements 0 and j (minimum) */
1016 { 1170 {
1017 int L = rng [i].l; 1171 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1018 int R = rng [i].r - 1; 1172 }
1173 }
1019 1174
1020 if (expect_false (L + EIO_QSORT_CUTOFF < R)) 1175 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1176 {
1177 eio_dirent *i, *j;
1178
1179 for (i = dents + 1; i < dents + size; ++i)
1021 { 1180 {
1022 eio_dirent piv = dents [L]; 1181 eio_dirent value = *i;
1023 1182
1024 while (L < R) 1183 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1025 { 1184 j [1] = j [0];
1026 while (eio_dent_cmp (&dents [R], &piv) >= 0 && L < R)
1027 --R;
1028 1185
1029 if (L < R) 1186 j [1] = value;
1030 dents [L++] = dents [R];
1031
1032 while (eio_dent_cmp (&dents [L], &piv) <= 0 && L < R)
1033 ++L;
1034
1035 if (L < R)
1036 dents [R--] = dents [L];
1037 }
1038
1039 dents [L] = piv;
1040
1041 ++i;
1042 rng [i].l = L + 1;
1043 rng [i].r = rng [i - 1].r;
1044 rng [i - 1].r = L;
1045
1046 if (rng [i].r - rng [i].l > rng [i - 1].r - rng [i - 1].l)
1047 {
1048 struct rng t;
1049
1050 t = rng [i]; rng [i] = rng [i - 1]; rng [i - 1] = t;
1051 }
1052 } 1187 }
1053 else
1054 --i;
1055 } 1188 }
1189}
1056 1190
1057 /* use a simple insertion sort at the end */ 1191static void
1058 for (i = 1; i < size; ++i) 1192eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1059 { 1193{
1060 eio_dirent value = dents [i]; 1194 if (size <= 1)
1195 return; /* our insertion sort relies on size > 0 */
1061 1196
1062 for (j = i - 1; j >= 0 && eio_dent_cmp (&dents [j], &value) > 0; --j) 1197 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1063 dents [j + 1] = dents [j]; 1198 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1199 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1064 1200
1065 dents [j + 1] = value; 1201 /* use an insertion sort at the end, or for small arrays, */
1066 } 1202 /* as insertion sort is more efficient for small partitions */
1203 eio_dent_insertion_sort (dents, size);
1067} 1204}
1068 1205
1069/* read a full directory */ 1206/* read a full directory */
1070static void 1207static void
1071eio__scandir (eio_req *req, etp_worker *self) 1208eio__scandir (eio_req *req, etp_worker *self)
… …
1077 int namesoffs = 0; 1214 int namesoffs = 0;
1078 int flags = req->int1; 1215 int flags = req->int1;
1079 eio_dirent *dents = 0; 1216 eio_dirent *dents = 0;
1080 int dentalloc = 128; 1217 int dentalloc = 128;
1081 int dentoffs = 0; 1218 int dentoffs = 0;
1219 ino_t inode_bits = 0;
1082 1220
1083 req->result = -1; 1221 req->result = -1;
1084 1222
1085 if (!(flags & EIO_READDIR_DENTS)) 1223 if (!(flags & EIO_READDIR_DENTS))
1086 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); 1224 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1087 1225
1088 X_LOCK (wrklock); 1226 X_LOCK (wrklock);
1089 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1227 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1090 self->dirp = dirp = opendir (req->ptr1); 1228 self->dirp = dirp = opendir (req->ptr1);
1091 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; 1229 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1092 req->ptr1 = names = malloc (namesalloc);
1093 req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; 1230 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1231 req->ptr2 = names = malloc (namesalloc);
1094 X_UNLOCK (wrklock); 1232 X_UNLOCK (wrklock);
1095 1233
1096 if (dirp && names && (!flags || dents)) 1234 if (dirp && names && (!flags || dents))
1097 for (;;) 1235 for (;;)
1098 { 1236 {
… …
1106 1244
1107 /* sort etc. */ 1245 /* sort etc. */
1108 req->int1 = flags; 1246 req->int1 = flags;
1109 req->result = dentoffs; 1247 req->result = dentoffs;
1110 1248
1111 if (dents)
1112 {
1113 eio_dirent *ent = dents + dentoffs;
1114
1115 while (ent > dents)
1116 (--ent)->name = names + (size_t)ent->name;
1117 }
1118
1119 if (flags & EIO_READDIR_STAT_ORDER 1249 if (flags & EIO_READDIR_STAT_ORDER)
1120 || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) 1250 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1121 eio_dent_sort (dents, dentoffs); /* score depends of DIRS_FIRST */
1122 else if (flags & EIO_READDIR_DIRS_FIRST) 1251 else if (flags & EIO_READDIR_DIRS_FIRST)
1252 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1253 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1254 else
1123 { 1255 {
1124 /* in this case, all is known, and we just put dirs first and sort them */ 1256 /* in this case, all is known, and we just put dirs first and sort them */
1125 eio_dirent *ent = dents + dentoffs; 1257 eio_dirent *oth = dents + dentoffs;
1126 eio_dirent *dir = dents; 1258 eio_dirent *dir = dents;
1127 1259
1128 /* now move dirs to the front, and non-dirs to the back */ 1260 /* now partition dirs to the front, and non-dirs to the back */
1129 /* by walking from both sides and swapping if necessary */ 1261 /* by walking from both sides and swapping if necessary */
1262 /* also clear score, so it doesn't influence sorting */
1130 while (ent > dir) 1263 while (oth > dir)
1131 { 1264 {
1132 if (dir->type == DT_DIR) 1265 if (dir->type == EIO_DT_DIR)
1133 ++dir; 1266 ++dir;
1134 else
1135 {
1136 --ent;
1137
1138 if (ent->type == DT_DIR) 1267 else if ((--oth)->type == EIO_DT_DIR)
1139 { 1268 {
1140 eio_dirent tmp = *dir; 1269 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1141 *dir = *ent;
1142 *ent = tmp;
1143 1270
1144 ++dir; 1271 ++dir;
1145 } 1272 }
1146 } 1273 }
1147 }
1148 1274
1149 /* now sort the dirs only */ 1275 /* now sort the dirs only */
1150 eio_dent_sort (dents, dir - dents); 1276 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1151 } 1277 }
1152
1153 /* only provide the names array unless DENTS is specified */
1154 if (!(flags & EIO_READDIR_DENTS))
1155 {
1156 X_LOCK (wrklock);
1157 assert (!dents);
1158 req->ptr1 = 0;
1159 req->ptr2 = names;
1160 X_UNLOCK (wrklock);
1161 }
1162 1278
1163 break; 1279 break;
1164 } 1280 }
1165 1281
1166 /* now add the entry to our list(s) */ 1282 /* now add the entry to our list(s) */
… …
1173 1289
1174 while (expect_false (namesoffs + len > namesalloc)) 1290 while (expect_false (namesoffs + len > namesalloc))
1175 { 1291 {
1176 namesalloc *= 2; 1292 namesalloc *= 2;
1177 X_LOCK (wrklock); 1293 X_LOCK (wrklock);
1178 req->ptr1 = names = realloc (names, namesalloc); 1294 req->ptr2 = names = realloc (names, namesalloc);
1179 X_UNLOCK (wrklock); 1295 X_UNLOCK (wrklock);
1180 1296
1181 if (!names) 1297 if (!names)
1182 break; 1298 break;
1183 } 1299 }
… …
1190 1306
1191 if (expect_false (dentoffs == dentalloc)) 1307 if (expect_false (dentoffs == dentalloc))
1192 { 1308 {
1193 dentalloc *= 2; 1309 dentalloc *= 2;
1194 X_LOCK (wrklock); 1310 X_LOCK (wrklock);
1195 req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); 1311 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1196 X_UNLOCK (wrklock); 1312 X_UNLOCK (wrklock);
1197 1313
1198 if (!dents) 1314 if (!dents)
1199 break; 1315 break;
1200 } 1316 }
1201 1317
1202 ent = dents + dentoffs; 1318 ent = dents + dentoffs;
1203 1319
1204 ent->name = (char *)(size_t)namesoffs; /* rather dirtily we store the offset in the pointer */ 1320 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1205 ent->namelen = len - 1; 1321 ent->namelen = len - 1;
1206 ent->inode = D_INO (entp); 1322 ent->inode = D_INO (entp);
1323
1324 inode_bits |= ent->inode;
1207 1325
1208 switch (D_TYPE (entp)) 1326 switch (D_TYPE (entp))
1209 { 1327 {
1210 default: 1328 default:
1211 ent->type = EIO_DT_UNKNOWN; 1329 ent->type = EIO_DT_UNKNOWN;
… …
1254 #ifdef DT_WHT 1372 #ifdef DT_WHT
1255 case DT_WHT: ent->type = EIO_DT_WHT; break; 1373 case DT_WHT: ent->type = EIO_DT_WHT; break;
1256 #endif 1374 #endif
1257 } 1375 }
1258 1376
1259 ent->score = 0; 1377 ent->score = 7;
1260 1378
1261 if (flags & EIO_READDIR_DIRS_FIRST) 1379 if (flags & EIO_READDIR_DIRS_FIRST)
1262 { 1380 {
1263 if (ent->type == EIO_DT_UNKNOWN) 1381 if (ent->type == EIO_DT_UNKNOWN)
1264 { 1382 {
1265 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ 1383 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1266 ent->score = 98; 1384 ent->score = 1;
1267 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ 1385 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1268 ent->score = len <= 2 ? len + 6 : len <= 4 ? 5 : len <= 7 ? 4 : 1; /* shorter == more likely dir, but avoid too many classes */ 1386 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1269 } 1387 }
1270 else if (ent->type == EIO_DT_DIR) 1388 else if (ent->type == EIO_DT_DIR)
1271 ent->score = 100; 1389 ent->score = 0;
1272 } 1390 }
1273 } 1391 }
1274 1392
1275 namesoffs += len; 1393 namesoffs += len;
1276 ++dentoffs; 1394 ++dentoffs;
… …
1282 break; 1400 break;
1283 } 1401 }
1284 } 1402 }
1285} 1403}
1286 1404
1405#ifdef PAGESIZE
1406# define eio_pagesize() PAGESIZE
1407#else
1408static intptr_t
1409eio_pagesize (void)
1410{
1411 static intptr_t page;
1412
1413 if (!page)
1414 page = sysconf (_SC_PAGESIZE);
1415
1416 return page;
1417}
1418#endif
1419
1420static void
1421eio_page_align (void **addr, size_t *length)
1422{
1423 intptr_t mask = eio_pagesize () - 1;
1424
1425 /* round down addr */
1426 intptr_t adj = mask & (intptr_t)*addr;
1427
1428 *addr = (void *)((intptr_t)*addr - adj);
1429 *length += adj;
1430
1431 /* round up length */
1432 *length = (*length + mask) & ~mask;
1433}
1434
1435#if !_POSIX_MEMLOCK
1436# define eio__mlock(a,b) ((errno = ENOSYS), -1)
1437# define eio__mlockall(a) ((errno = ENOSYS), -1)
1438#else
1439
1440static int
1441eio__mlock (void *addr, size_t length)
1442{
1443 eio_page_align (&addr, &length);
1444
1445 return mlock (addr, length);
1446}
1447
1448static int
1449eio__mlockall (int flags)
1450{
1451 #if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
1452 extern int mallopt (int, int);
1453 mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
1454 #endif
1455
1456 if (EIO_MCL_CURRENT != MCL_CURRENT
1457 || EIO_MCL_FUTURE != MCL_FUTURE)
1458 {
1459 flags = 0
1460 | (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
1461 | (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
1462 }
1463
1464 return mlockall (flags);
1465}
1466#endif
1467
1287#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) 1468#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1288# undef msync
1289# define msync(a,b,c) ((errno = ENOSYS), -1) 1469# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1470#else
1471
1472int
1473eio__msync (void *mem, size_t len, int flags)
1474{
1475 eio_page_align (&mem, &len);
1476
1477 if (EIO_MS_ASYNC != MS_SYNC
1478 || EIO_MS_INVALIDATE != MS_INVALIDATE
1479 || EIO_MS_SYNC != MS_SYNC)
1480 {
1481 flags = 0
1482 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1483 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1484 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1485 }
1486
1487 return msync (mem, len, flags);
1488}
1489
1290#endif 1490#endif
1291 1491
1292int 1492int
1293eio__mtouch (void *mem, size_t len, int flags) 1493eio__mtouch (void *mem, size_t len, int flags)
1294{ 1494{
1495 eio_page_align (&mem, &len);
1496
1497 {
1295 intptr_t addr = (intptr_t)mem; 1498 intptr_t addr = (intptr_t)mem;
1296 intptr_t end = addr + len; 1499 intptr_t end = addr + len;
1297#ifdef PAGESIZE 1500 intptr_t page = eio_pagesize ();
1298 const intptr_t page = PAGESIZE;
1299#else
1300 static intptr_t page;
1301 1501
1302 if (!page)
1303 page = sysconf (_SC_PAGESIZE);
1304#endif
1305
1306 addr &= ~(page - 1); /* assume page size is always a power of two */
1307
1308 if (addr < end) 1502 if (addr < end)
1309 if (flags) /* modify */ 1503 if (flags & EIO_MT_MODIFY) /* modify */
1310 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); 1504 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1311 else 1505 else
1312 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); 1506 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1507 }
1313 1508
1314 return 0; 1509 return 0;
1315} 1510}
1316 1511
1317/*****************************************************************************/ 1512/*****************************************************************************/
… …
1440 return 0; \ 1635 return 0; \
1441 } 1636 }
1442 1637
1443static void eio_execute (etp_worker *self, eio_req *req) 1638static void eio_execute (etp_worker *self, eio_req *req)
1444{ 1639{
1445 errno = 0;
1446
1447 switch (req->type) 1640 switch (req->type)
1448 { 1641 {
1449 case EIO_READ: ALLOC (req->size); 1642 case EIO_READ: ALLOC (req->size);
1450 req->result = req->offs >= 0 1643 req->result = req->offs >= 0
1451 ? pread (req->int1, req->ptr2, req->size, req->offs) 1644 ? pread (req->int1, req->ptr2, req->size, req->offs)
… …
1461 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1654 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1462 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1655 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1463 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1656 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1464 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1657 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1465 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1658 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1659
1660 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1661 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1662 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1663 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1466 1664
1467 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; 1665 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1468 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; 1666 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1469 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; 1667 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1470 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; 1668 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
… …
1480 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1678 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1481 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1679 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1482 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1680 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1483 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break; 1681 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1484 1682
1485 case EIO_READLINK: ALLOC (NAME_MAX); 1683 case EIO_READLINK: ALLOC (PATH_MAX);
1486 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1684 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1487 1685
1488 case EIO_SYNC: req->result = 0; sync (); break; 1686 case EIO_SYNC: req->result = 0; sync (); break;
1489 case EIO_FSYNC: req->result = fsync (req->int1); break; 1687 case EIO_FSYNC: req->result = fsync (req->int1); break;
1490 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1688 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1491 case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; 1689 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1492 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; 1690 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1691 case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
1692 case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
1493 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; 1693 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1494 1694
1495 case EIO_READDIR: eio__scandir (req, self); break; 1695 case EIO_READDIR: eio__scandir (req, self); break;
1496 1696
1497 case EIO_BUSY: 1697 case EIO_BUSY:
1498#ifdef _WIN32 1698#ifdef _WIN32
1499 Sleep (req->nv1 * 1000.); 1699 Sleep (req->nv1 * 1e3);
1500#else 1700#else
1501 { 1701 {
1502 struct timeval tv; 1702 struct timeval tv;
1503 1703
1504 tv.tv_sec = req->nv1; 1704 tv.tv_sec = req->nv1;
1505 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.; 1705 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
1506 1706
1507 req->result = select (0, 0, 0, 0, &tv); 1707 req->result = select (0, 0, 0, 0, &tv);
1508 } 1708 }
1509#endif 1709#endif
1510 break; 1710 break;
… …
1525 times = tv; 1725 times = tv;
1526 } 1726 }
1527 else 1727 else
1528 times = 0; 1728 times = 0;
1529 1729
1530
1531 req->result = req->type == EIO_FUTIME 1730 req->result = req->type == EIO_FUTIME
1532 ? futimes (req->int1, times) 1731 ? futimes (req->int1, times)
1533 : utimes (req->ptr1, times); 1732 : utimes (req->ptr1, times);
1534 } 1733 }
1535 break; 1734 break;
… …
1544 case EIO_CUSTOM: 1743 case EIO_CUSTOM:
1545 ((void (*)(eio_req *))req->feed) (req); 1744 ((void (*)(eio_req *))req->feed) (req);
1546 break; 1745 break;
1547 1746
1548 default: 1747 default:
1748 errno = ENOSYS;
1549 req->result = -1; 1749 req->result = -1;
1550 break; 1750 break;
1551 } 1751 }
1552 1752
1553 req->errorno = errno; 1753 req->errorno = errno;
… …
1583eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) 1783eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1584{ 1784{
1585 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; 1785 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1586} 1786}
1587 1787
1788eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
1789{
1790 REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
1791}
1792
1793eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
1794{
1795 REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
1796}
1797
1588eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data) 1798eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1589{ 1799{
1590 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND; 1800 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1591} 1801}
1592 1802
… …
1616} 1826}
1617 1827
1618eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) 1828eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1619{ 1829{
1620 REQ (EIO_FSTAT); req->int1 = fd; SEND; 1830 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1831}
1832
1833eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1834{
1835 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1621} 1836}
1622 1837
1623eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) 1838eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1624{ 1839{
1625 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; 1840 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
… …
1697} 1912}
1698 1913
1699eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) 1914eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1700{ 1915{
1701 return eio__1path (EIO_LSTAT, path, pri, cb, data); 1916 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1917}
1918
1919eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
1920{
1921 return eio__1path (EIO_STATVFS, path, pri, cb, data);
1702} 1922}
1703 1923
1704eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) 1924eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1705{ 1925{
1706 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1926 return eio__1path (EIO_UNLINK, path, pri, cb, data);

Diff Legend

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