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> |
|
|
79 | # include <sys/mman.h> |
78 | # include <unistd.h> |
80 | # include <unistd.h> |
79 | # include <utime.h> |
81 | # include <utime.h> |
80 | # include <signal.h> |
82 | # include <signal.h> |
81 | # include <dirent.h> |
83 | # include <dirent.h> |
82 | |
84 | |
83 | /* 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 */ |
84 | # if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) |
86 | # if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__ |
85 | # define _DIRENT_HAVE_D_TYPE /* sigh */ |
87 | # define _DIRENT_HAVE_D_TYPE /* sigh */ |
86 | # define D_INO(de) (de)->d_fileno |
88 | # define D_INO(de) (de)->d_fileno |
87 | # define D_NAMLEN(de) (de)->d_namlen |
89 | # define D_NAMLEN(de) (de)->d_namlen |
88 | # elif defined(__linux) || defined(d_ino) || _XOPEN_SOURCE >= 600 |
90 | # elif __linux || defined d_ino || _XOPEN_SOURCE >= 600 |
89 | # define D_INO(de) (de)->d_ino |
91 | # define D_INO(de) (de)->d_ino |
90 | # endif |
92 | # endif |
91 | |
93 | |
92 | #ifdef _D_EXACT_NAMLEN |
94 | #ifdef _D_EXACT_NAMLEN |
93 | # undef D_NAMLEN |
95 | # undef D_NAMLEN |
… | |
… | |
105 | #endif |
107 | #endif |
106 | |
108 | |
107 | #if HAVE_SENDFILE |
109 | #if HAVE_SENDFILE |
108 | # if __linux |
110 | # if __linux |
109 | # include <sys/sendfile.h> |
111 | # include <sys/sendfile.h> |
110 | # elif __freebsd |
112 | # elif __FreeBSD__ || defined __APPLE__ |
111 | # include <sys/socket.h> |
113 | # include <sys/socket.h> |
112 | # include <sys/uio.h> |
114 | # include <sys/uio.h> |
113 | # elif __hpux |
115 | # elif __hpux |
114 | # include <sys/socket.h> |
116 | # include <sys/socket.h> |
115 | # elif __solaris /* not yet */ |
117 | # elif __solaris |
116 | # include <sys/sendfile.h> |
118 | # include <sys/sendfile.h> |
117 | # else |
119 | # else |
118 | # error sendfile support requested but not available |
120 | # error sendfile support requested but not available |
119 | # endif |
121 | # endif |
120 | #endif |
122 | #endif |
… | |
… | |
133 | #define IDLE_TIMEOUT 10 |
135 | #define IDLE_TIMEOUT 10 |
134 | |
136 | |
135 | /* used for struct dirent, AIX doesn't provide it */ |
137 | /* used for struct dirent, AIX doesn't provide it */ |
136 | #ifndef NAME_MAX |
138 | #ifndef NAME_MAX |
137 | # 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 |
138 | #endif |
145 | #endif |
139 | |
146 | |
140 | /* buffer size for various temporary buffers */ |
147 | /* buffer size for various temporary buffers */ |
141 | #define EIO_BUFSIZE 65536 |
148 | #define EIO_BUFSIZE 65536 |
142 | |
149 | |
… | |
… | |
195 | |
202 | |
196 | /*****************************************************************************/ |
203 | /*****************************************************************************/ |
197 | |
204 | |
198 | #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) |
205 | #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) |
199 | |
206 | |
200 | /* calculcate time difference in ~1/EIO_TICKS of a second */ |
207 | /* calculate time difference in ~1/EIO_TICKS of a second */ |
201 | static int tvdiff (struct timeval *tv1, struct timeval *tv2) |
208 | static int tvdiff (struct timeval *tv1, struct timeval *tv2) |
202 | { |
209 | { |
203 | return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS |
210 | return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS |
204 | + ((tv2->tv_usec - tv1->tv_usec) >> 10); |
211 | + ((tv2->tv_usec - tv1->tv_usec) >> 10); |
205 | } |
212 | } |
… | |
… | |
215 | static volatile unsigned int nreqs; /* reqlock */ |
222 | static volatile unsigned int nreqs; /* reqlock */ |
216 | static volatile unsigned int nready; /* reqlock */ |
223 | static volatile unsigned int nready; /* reqlock */ |
217 | static volatile unsigned int npending; /* reqlock */ |
224 | static volatile unsigned int npending; /* reqlock */ |
218 | static volatile unsigned int max_idle = 4; |
225 | static volatile unsigned int max_idle = 4; |
219 | |
226 | |
220 | static mutex_t wrklock = X_MUTEX_INIT; |
227 | static xmutex_t wrklock = X_MUTEX_INIT; |
221 | static mutex_t reslock = X_MUTEX_INIT; |
228 | static xmutex_t reslock = X_MUTEX_INIT; |
222 | static mutex_t reqlock = X_MUTEX_INIT; |
229 | static xmutex_t reqlock = X_MUTEX_INIT; |
223 | static cond_t reqwait = X_COND_INIT; |
230 | static xcond_t reqwait = X_COND_INIT; |
224 | |
231 | |
225 | #if !HAVE_PREADWRITE |
232 | #if !HAVE_PREADWRITE |
226 | /* |
233 | /* |
227 | * make our pread/pwrite emulation safe against themselves, but not against |
234 | * make our pread/pwrite emulation safe against themselves, but not against |
228 | * 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, |
229 | * but that's your problem, not mine. |
236 | * but that's your problem, not mine. |
230 | */ |
237 | */ |
231 | static mutex_t preadwritelock = X_MUTEX_INIT; |
238 | static xmutex_t preadwritelock = X_MUTEX_INIT; |
232 | #endif |
239 | #endif |
233 | |
240 | |
234 | typedef struct etp_worker |
241 | typedef struct etp_worker |
235 | { |
242 | { |
236 | /* locked by wrklock */ |
243 | /* locked by wrklock */ |
237 | struct etp_worker *prev, *next; |
244 | struct etp_worker *prev, *next; |
238 | |
245 | |
239 | thread_t tid; |
246 | xthread_t tid; |
240 | |
247 | |
241 | /* locked by reslock, reqlock or wrklock */ |
248 | /* locked by reslock, reqlock or wrklock */ |
242 | ETP_REQ *req; /* currently processed request */ |
249 | ETP_REQ *req; /* currently processed request */ |
243 | |
250 | |
244 | ETP_WORKER_COMMON |
251 | ETP_WORKER_COMMON |
… | |
… | |
597 | } |
604 | } |
598 | |
605 | |
599 | static void etp_set_max_poll_time (double nseconds) |
606 | static void etp_set_max_poll_time (double nseconds) |
600 | { |
607 | { |
601 | if (WORDACCESS_UNSAFE) X_LOCK (reslock); |
608 | if (WORDACCESS_UNSAFE) X_LOCK (reslock); |
602 | max_poll_time = nseconds; |
609 | max_poll_time = nseconds * EIO_TICKS; |
603 | if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); |
610 | if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); |
604 | } |
611 | } |
605 | |
612 | |
606 | static void etp_set_max_poll_reqs (unsigned int maxreqs) |
613 | static void etp_set_max_poll_reqs (unsigned int maxreqs) |
607 | { |
614 | { |
… | |
… | |
866 | if (!res || errno != ENOSYS) |
873 | if (!res || errno != ENOSYS) |
867 | return res; |
874 | return res; |
868 | #endif |
875 | #endif |
869 | |
876 | |
870 | /* 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 |
871 | * call fdatasync, as thta matches the expectation of it's users best */ |
878 | * call fdatasync, as that matches the expectation of its users best */ |
872 | return fdatasync (fd); |
879 | return fdatasync (fd); |
873 | } |
880 | } |
874 | |
881 | |
875 | #if !HAVE_READAHEAD |
882 | #if !HAVE_READAHEAD |
876 | # undef readahead |
883 | # undef readahead |
… | |
… | |
908 | |
915 | |
909 | #if HAVE_SENDFILE |
916 | #if HAVE_SENDFILE |
910 | # if __linux |
917 | # if __linux |
911 | res = sendfile (ofd, ifd, &offset, count); |
918 | res = sendfile (ofd, ifd, &offset, count); |
912 | |
919 | |
913 | # elif __freebsd |
920 | # elif __FreeBSD__ |
914 | /* |
921 | /* |
915 | * 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 |
916 | * wasted on making it similar to other I/O functions. |
923 | * wasted on making it similar to other I/O functions. |
917 | */ |
924 | */ |
918 | { |
925 | { |
919 | off_t sbytes; |
926 | off_t sbytes; |
920 | res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); |
927 | res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); |
921 | |
928 | |
922 | if (res < 0 && sbytes) |
929 | #if 0 /* according to the manpage, this is correct, but broken behaviour */ |
923 | /* 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) |
924 | res = sbytes; |
950 | res = sbytes; |
925 | } |
951 | } |
926 | |
952 | |
927 | # elif __hpux |
953 | # elif __hpux |
928 | res = sendfile (ofd, ifd, offset, count, 0, 0); |
954 | res = sendfile (ofd, ifd, offset, count, 0, 0); |
… | |
… | |
942 | if (res < 0 && sbytes) |
968 | if (res < 0 && sbytes) |
943 | res = sbytes; |
969 | res = sbytes; |
944 | } |
970 | } |
945 | |
971 | |
946 | # 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 | |
947 | #else |
983 | #else |
948 | res = -1; |
984 | res = -1; |
949 | errno = ENOSYS; |
985 | errno = ENOSYS; |
950 | #endif |
986 | #endif |
951 | |
987 | |
952 | if (res < 0 |
988 | if (res < 0 |
953 | && (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 */ |
954 | #if __solaris |
995 | #if __solaris |
955 | || errno == EAFNOSUPPORT || errno == EPROTOTYPE |
996 | || errno == EAFNOSUPPORT || errno == EPROTOTYPE |
956 | #endif |
997 | #endif |
957 | ) |
998 | ) |
958 | ) |
999 | ) |
… | |
… | |
989 | } |
1030 | } |
990 | |
1031 | |
991 | return res; |
1032 | return res; |
992 | } |
1033 | } |
993 | |
1034 | |
994 | static int |
1035 | static signed char |
995 | eio_dent_cmp (const void *a_, const void *b_) |
1036 | eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) |
996 | { |
1037 | { |
997 | const eio_dirent *a = (const eio_dirent *)a_; |
1038 | return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */ |
998 | const eio_dirent *b = (const eio_dirent *)b_; |
|
|
999 | |
|
|
1000 | return (int)b->score - (int)a->score ? (int)b->score - (int)a->score |
|
|
1001 | : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; /* int might be < ino_t */ |
1039 | : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; |
|
|
1040 | } |
|
|
1041 | |
|
|
1042 | #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 |
|
|
1043 | |
|
|
1044 | #define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ |
|
|
1045 | #define EIO_SORT_FAST 60 /* when to only use insertion sort */ |
|
|
1046 | |
|
|
1047 | static void |
|
|
1048 | eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) |
|
|
1049 | { |
|
|
1050 | unsigned char bits [9 + sizeof (ino_t) * 8]; |
|
|
1051 | unsigned char *bit = bits; |
|
|
1052 | |
|
|
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 */ |
|
|
1065 | { |
|
|
1066 | ino_t endianness; |
|
|
1067 | int i, j; |
|
|
1068 | |
|
|
1069 | /* we store the byte offset of byte n into byte n of "endianness" */ |
|
|
1070 | for (i = 0; i < sizeof (ino_t); ++i) |
|
|
1071 | ((unsigned char *)&endianness)[i] = i; |
|
|
1072 | |
|
|
1073 | *bit++ = 0; |
|
|
1074 | |
|
|
1075 | for (i = 0; i < sizeof (ino_t); ++i) |
|
|
1076 | { |
|
|
1077 | /* shifting off the byte offsets out of "endianness" */ |
|
|
1078 | int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8; |
|
|
1079 | endianness >>= 8; |
|
|
1080 | |
|
|
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 (;;) |
|
|
1109 | { |
|
|
1110 | unsigned char O = *bit >> 3; |
|
|
1111 | unsigned char M = 1 << (*bit & 7); |
|
|
1112 | |
|
|
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)) |
|
|
1126 | { |
|
|
1127 | eio_dirent tmp = *a; *a = *b; *b = tmp; |
|
|
1128 | ++a; |
|
|
1129 | } |
|
|
1130 | while (b > a); |
|
|
1131 | |
|
|
1132 | /* next bit, or stop, if no bits left in this path */ |
|
|
1133 | if (!*--bit) |
|
|
1134 | break; |
|
|
1135 | |
|
|
1136 | base_stk [stk_idx] = a; |
|
|
1137 | end_stk [stk_idx] = end; |
|
|
1138 | bit_stk [stk_idx] = bit; |
|
|
1139 | ++stk_idx; |
|
|
1140 | |
|
|
1141 | end = a; |
|
|
1142 | } |
|
|
1143 | } |
|
|
1144 | while (stk_idx--); |
|
|
1145 | } |
|
|
1146 | } |
|
|
1147 | |
|
|
1148 | static void |
|
|
1149 | eio_dent_insertion_sort (eio_dirent *dents, int size) |
|
|
1150 | { |
|
|
1151 | /* first move the smallest element to the front, to act as a sentinel */ |
|
|
1152 | { |
|
|
1153 | int i; |
|
|
1154 | eio_dirent *min = dents; |
|
|
1155 | |
|
|
1156 | /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */ |
|
|
1157 | for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; ) |
|
|
1158 | if (EIO_DENT_CMP (dents [i], <, *min)) |
|
|
1159 | min = &dents [i]; |
|
|
1160 | |
|
|
1161 | /* swap elements 0 and j (minimum) */ |
|
|
1162 | { |
|
|
1163 | eio_dirent tmp = *dents; *dents = *min; *min = tmp; |
|
|
1164 | } |
|
|
1165 | } |
|
|
1166 | |
|
|
1167 | /* then do standard insertion sort, assuming that all elements are >= dents [0] */ |
|
|
1168 | { |
|
|
1169 | eio_dirent *i, *j; |
|
|
1170 | |
|
|
1171 | for (i = dents + 1; i < dents + size; ++i) |
|
|
1172 | { |
|
|
1173 | eio_dirent value = *i; |
|
|
1174 | |
|
|
1175 | for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j) |
|
|
1176 | j [1] = j [0]; |
|
|
1177 | |
|
|
1178 | j [1] = value; |
|
|
1179 | } |
|
|
1180 | } |
|
|
1181 | } |
|
|
1182 | |
|
|
1183 | static void |
|
|
1184 | eio_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); |
1002 | } |
1196 | } |
1003 | |
1197 | |
1004 | /* read a full directory */ |
1198 | /* read a full directory */ |
1005 | static void |
1199 | static void |
1006 | eio__scandir (eio_req *req, etp_worker *self) |
1200 | eio__scandir (eio_req *req, etp_worker *self) |
1007 | { |
1201 | { |
1008 | DIR *dirp; |
1202 | DIR *dirp; |
1009 | EIO_STRUCT_DIRENT *entp; |
1203 | EIO_STRUCT_DIRENT *entp; |
1010 | unsigned char *name, *names; |
1204 | char *name, *names; |
1011 | int namesalloc = 4096; |
1205 | int namesalloc = 4096; |
1012 | int namesoffs = 0; |
1206 | int namesoffs = 0; |
1013 | int flags = req->int1; |
1207 | int flags = req->int1; |
1014 | eio_dirent *dents = 0; |
1208 | eio_dirent *dents = 0; |
1015 | int dentalloc = 128; |
1209 | int dentalloc = 128; |
1016 | int dentoffs = 0; |
1210 | int dentoffs = 0; |
|
|
1211 | ino_t inode_bits = 0; |
1017 | |
1212 | |
1018 | req->result = -1; |
1213 | req->result = -1; |
1019 | |
1214 | |
1020 | if (!(flags & EIO_READDIR_DENTS)) |
1215 | if (!(flags & EIO_READDIR_DENTS)) |
1021 | flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); |
1216 | flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); |
1022 | |
1217 | |
1023 | X_LOCK (wrklock); |
1218 | X_LOCK (wrklock); |
1024 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
1219 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
1025 | self->dirp = dirp = opendir (req->ptr1); |
1220 | self->dirp = dirp = opendir (req->ptr1); |
1026 | req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; |
1221 | req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; |
1027 | req->ptr1 = names = malloc (namesalloc); |
|
|
1028 | req->ptr2 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; |
1222 | req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; |
|
|
1223 | req->ptr2 = names = malloc (namesalloc); |
1029 | X_UNLOCK (wrklock); |
1224 | X_UNLOCK (wrklock); |
1030 | |
1225 | |
1031 | if (dirp && names && (!flags || dents)) |
1226 | if (dirp && names && (!flags || dents)) |
1032 | for (;;) |
1227 | for (;;) |
1033 | { |
1228 | { |
… | |
… | |
1041 | |
1236 | |
1042 | /* sort etc. */ |
1237 | /* sort etc. */ |
1043 | req->int1 = flags; |
1238 | req->int1 = flags; |
1044 | req->result = dentoffs; |
1239 | req->result = dentoffs; |
1045 | |
1240 | |
1046 | if (dents) |
|
|
1047 | { |
|
|
1048 | eio_dirent *ent = dents + dentoffs; |
|
|
1049 | |
|
|
1050 | while (ent > dents) |
|
|
1051 | (--ent)->name = names + (size_t)ent->name; |
|
|
1052 | } |
|
|
1053 | |
|
|
1054 | if (flags & EIO_READDIR_STAT_ORDER |
1241 | if (flags & EIO_READDIR_STAT_ORDER) |
1055 | || !(~flags & (EIO_READDIR_DIRS_FIRST | EIO_READDIR_FOUND_UNKNOWN))) |
1242 | eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */ |
1056 | { |
|
|
1057 | /* pray your qsort doesn't use quicksort */ |
|
|
1058 | qsort (dents, dentoffs, sizeof (*dents), eio_dent_cmp); /* score depends of DIRS_FIRST */ |
|
|
1059 | } |
|
|
1060 | else if (flags & EIO_READDIR_DIRS_FIRST) |
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 |
1061 | { |
1247 | { |
1062 | /* 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 */ |
1063 | eio_dirent *ent = dents + dentoffs; |
1249 | eio_dirent *oth = dents + dentoffs; |
1064 | eio_dirent *dir = dents; |
1250 | eio_dirent *dir = dents; |
1065 | |
1251 | |
1066 | /* 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 */ |
1067 | /* by walkign 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 */ |
1068 | while (ent > dir) |
1255 | while (oth > dir) |
1069 | { |
1256 | { |
1070 | if (dir->type == DT_DIR) |
1257 | if (dir->type == EIO_DT_DIR) |
1071 | ++dir; |
1258 | ++dir; |
1072 | else |
|
|
1073 | { |
|
|
1074 | --ent; |
|
|
1075 | |
|
|
1076 | if (ent->type == DT_DIR) |
1259 | else if ((--oth)->type == EIO_DT_DIR) |
1077 | { |
1260 | { |
1078 | eio_dirent tmp = *dir; |
1261 | eio_dirent tmp = *dir; *dir = *oth; *oth = tmp; |
1079 | *dir = *ent; |
|
|
1080 | *ent = tmp; |
|
|
1081 | |
1262 | |
1082 | ++dir; |
1263 | ++dir; |
1083 | } |
1264 | } |
1084 | } |
1265 | } |
1085 | } |
|
|
1086 | |
1266 | |
1087 | /* now sort the dirs only */ |
1267 | /* now sort the dirs only */ |
1088 | qsort (dents, dir - dents, sizeof (*dents), eio_dent_cmp); |
1268 | eio_dent_sort (dents, dir - dents, 0, inode_bits); |
1089 | } |
1269 | } |
1090 | |
|
|
1091 | /* only provide the names array unless DENTS is specified */ |
|
|
1092 | if (!(flags & EIO_READDIR_DENTS)) |
|
|
1093 | { |
|
|
1094 | X_LOCK (wrklock); |
|
|
1095 | assert (!dents); |
|
|
1096 | req->ptr1 = 0; |
|
|
1097 | req->ptr2 = names; |
|
|
1098 | X_UNLOCK (wrklock); |
|
|
1099 | } |
|
|
1100 | |
1270 | |
1101 | break; |
1271 | break; |
1102 | } |
1272 | } |
1103 | |
1273 | |
1104 | /* now add the entry to our list(s) */ |
1274 | /* now add the entry to our list(s) */ |
… | |
… | |
1111 | |
1281 | |
1112 | while (expect_false (namesoffs + len > namesalloc)) |
1282 | while (expect_false (namesoffs + len > namesalloc)) |
1113 | { |
1283 | { |
1114 | namesalloc *= 2; |
1284 | namesalloc *= 2; |
1115 | X_LOCK (wrklock); |
1285 | X_LOCK (wrklock); |
1116 | req->ptr1 = names = realloc (names, namesalloc); |
1286 | req->ptr2 = names = realloc (names, namesalloc); |
1117 | X_UNLOCK (wrklock); |
1287 | X_UNLOCK (wrklock); |
1118 | |
1288 | |
1119 | if (!names) |
1289 | if (!names) |
1120 | break; |
1290 | break; |
1121 | } |
1291 | } |
… | |
… | |
1128 | |
1298 | |
1129 | if (expect_false (dentoffs == dentalloc)) |
1299 | if (expect_false (dentoffs == dentalloc)) |
1130 | { |
1300 | { |
1131 | dentalloc *= 2; |
1301 | dentalloc *= 2; |
1132 | X_LOCK (wrklock); |
1302 | X_LOCK (wrklock); |
1133 | req->ptr2 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); |
1303 | req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); |
1134 | X_UNLOCK (wrklock); |
1304 | X_UNLOCK (wrklock); |
1135 | |
1305 | |
1136 | if (!dents) |
1306 | if (!dents) |
1137 | break; |
1307 | break; |
1138 | } |
1308 | } |
1139 | |
1309 | |
1140 | ent = dents + dentoffs; |
1310 | ent = dents + dentoffs; |
1141 | |
1311 | |
1142 | 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 */ |
1143 | ent->namelen = len - 1; |
1313 | ent->namelen = len - 1; |
1144 | ent->inode = D_INO (entp); |
1314 | ent->inode = D_INO (entp); |
|
|
1315 | |
|
|
1316 | inode_bits |= ent->inode; |
1145 | |
1317 | |
1146 | switch (D_TYPE (entp)) |
1318 | switch (D_TYPE (entp)) |
1147 | { |
1319 | { |
1148 | default: |
1320 | default: |
1149 | ent->type = EIO_DT_UNKNOWN; |
1321 | ent->type = EIO_DT_UNKNOWN; |
… | |
… | |
1192 | #ifdef DT_WHT |
1364 | #ifdef DT_WHT |
1193 | case DT_WHT: ent->type = EIO_DT_WHT; break; |
1365 | case DT_WHT: ent->type = EIO_DT_WHT; break; |
1194 | #endif |
1366 | #endif |
1195 | } |
1367 | } |
1196 | |
1368 | |
1197 | ent->score = 0; |
1369 | ent->score = 7; |
1198 | |
1370 | |
1199 | if (flags & EIO_READDIR_DIRS_FIRST) |
1371 | if (flags & EIO_READDIR_DIRS_FIRST) |
1200 | { |
1372 | { |
1201 | if (ent->type == EIO_DT_UNKNOWN) |
1373 | if (ent->type == EIO_DT_UNKNOWN) |
1202 | { |
1374 | { |
1203 | 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 */ |
1204 | ent->score = 98; |
1376 | ent->score = 1; |
1205 | else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ |
1377 | else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ |
1206 | ent->score = len <= 2 ? len + 6 : len <= 4 ? 5 : len <= 7 ? 4 : 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 */ |
1207 | } |
1379 | } |
1208 | else if (ent->type == EIO_DT_DIR) |
1380 | else if (ent->type == EIO_DT_DIR) |
1209 | ent->score = 100; |
1381 | ent->score = 0; |
1210 | } |
1382 | } |
1211 | } |
1383 | } |
1212 | |
1384 | |
1213 | namesoffs += len; |
1385 | namesoffs += len; |
1214 | ++dentoffs; |
1386 | ++dentoffs; |
… | |
… | |
1221 | } |
1393 | } |
1222 | } |
1394 | } |
1223 | } |
1395 | } |
1224 | |
1396 | |
1225 | #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) |
1397 | #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) |
1226 | # undef msync |
|
|
1227 | # define msync(a,b,c) ((errno = ENOSYS), -1) |
1398 | # define eio__msync(a,b,c) ((errno = ENOSYS), -1) |
|
|
1399 | #else |
|
|
1400 | |
|
|
1401 | int |
|
|
1402 | eio__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 | |
1228 | #endif |
1417 | #endif |
1229 | |
1418 | |
1230 | int |
1419 | int |
1231 | eio__mtouch (void *mem, size_t len, int flags) |
1420 | eio__mtouch (void *mem, size_t len, int flags) |
1232 | { |
1421 | { |
… | |
… | |
1239 | |
1428 | |
1240 | if (!page) |
1429 | if (!page) |
1241 | page = sysconf (_SC_PAGESIZE); |
1430 | page = sysconf (_SC_PAGESIZE); |
1242 | #endif |
1431 | #endif |
1243 | |
1432 | |
|
|
1433 | /* round down to start of page, although this is probably useless */ |
1244 | 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 */ |
1245 | |
1435 | |
1246 | if (addr < end) |
1436 | if (addr < end) |
1247 | if (flags) /* modify */ |
1437 | if (flags & EIO_MT_MODIFY) /* modify */ |
1248 | do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); |
1438 | do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); |
1249 | else |
1439 | else |
1250 | do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); |
1440 | do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); |
1251 | |
1441 | |
1252 | return 0; |
1442 | return 0; |
… | |
… | |
1400 | case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); |
1590 | case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); |
1401 | req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; |
1591 | req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; |
1402 | case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); |
1592 | case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); |
1403 | req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; |
1593 | req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; |
1404 | |
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 | |
1405 | 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; |
1406 | 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; |
1407 | 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; |
1408 | 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; |
1409 | case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; |
1604 | case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; |
… | |
… | |
1418 | case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; |
1613 | case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; |
1419 | case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; |
1614 | case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; |
1420 | case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; |
1615 | case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; |
1421 | 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; |
1422 | |
1617 | |
1423 | case EIO_READLINK: ALLOC (NAME_MAX); |
1618 | case EIO_READLINK: ALLOC (PATH_MAX); |
1424 | req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; |
1619 | req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break; |
1425 | |
1620 | |
1426 | case EIO_SYNC: req->result = 0; sync (); break; |
1621 | case EIO_SYNC: req->result = 0; sync (); break; |
1427 | case EIO_FSYNC: req->result = fsync (req->int1); break; |
1622 | case EIO_FSYNC: req->result = fsync (req->int1); break; |
1428 | case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; |
1623 | case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; |
1429 | 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; |
1430 | 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; |
1431 | 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; |
1432 | |
1627 | |
1433 | case EIO_READDIR: eio__scandir (req, self); break; |
1628 | case EIO_READDIR: eio__scandir (req, self); break; |
1434 | |
1629 | |
… | |
… | |
1556 | eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) |
1751 | eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) |
1557 | { |
1752 | { |
1558 | REQ (EIO_FSTAT); req->int1 = fd; SEND; |
1753 | REQ (EIO_FSTAT); req->int1 = fd; SEND; |
1559 | } |
1754 | } |
1560 | |
1755 | |
|
|
1756 | eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data) |
|
|
1757 | { |
|
|
1758 | REQ (EIO_FSTATVFS); req->int1 = fd; SEND; |
|
|
1759 | } |
|
|
1760 | |
1561 | eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) |
1761 | eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) |
1562 | { |
1762 | { |
1563 | 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; |
1564 | } |
1764 | } |
1565 | |
1765 | |
… | |
… | |
1635 | } |
1835 | } |
1636 | |
1836 | |
1637 | eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) |
1837 | eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) |
1638 | { |
1838 | { |
1639 | return eio__1path (EIO_LSTAT, path, pri, cb, data); |
1839 | return eio__1path (EIO_LSTAT, path, pri, cb, data); |
|
|
1840 | } |
|
|
1841 | |
|
|
1842 | eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data) |
|
|
1843 | { |
|
|
1844 | return eio__1path (EIO_STATVFS, path, pri, cb, data); |
1640 | } |
1845 | } |
1641 | |
1846 | |
1642 | eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) |
1847 | eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) |
1643 | { |
1848 | { |
1644 | return eio__1path (EIO_UNLINK, path, pri, cb, data); |
1849 | return eio__1path (EIO_UNLINK, path, pri, cb, data); |
… | |
… | |
1748 | /* misc garbage */ |
1953 | /* misc garbage */ |
1749 | |
1954 | |
1750 | ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) |
1955 | ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) |
1751 | { |
1956 | { |
1752 | etp_worker wrk; |
1957 | etp_worker wrk; |
|
|
1958 | ssize_t ret; |
1753 | |
1959 | |
1754 | wrk.dbuf = 0; |
1960 | wrk.dbuf = 0; |
1755 | |
1961 | |
1756 | eio__sendfile (ofd, ifd, offset, count, &wrk); |
1962 | ret = eio__sendfile (ofd, ifd, offset, count, &wrk); |
1757 | |
1963 | |
1758 | if (wrk.dbuf) |
1964 | if (wrk.dbuf) |
1759 | free (wrk.dbuf); |
1965 | free (wrk.dbuf); |
1760 | } |
|
|
1761 | |
1966 | |
|
|
1967 | return ret; |
|
|
1968 | } |
|
|
1969 | |