1 | /* |
1 | /* |
2 | * libeio implementation |
2 | * libeio implementation |
3 | * |
3 | * |
4 | * Copyright (c) 2007,2008 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 | * |
… | |
… | |
36 | * provisions above, a recipient may use your version of this file under |
36 | * provisions above, a recipient may use your version of this file under |
37 | * either the BSD or the GPL. |
37 | * either the BSD or the GPL. |
38 | */ |
38 | */ |
39 | |
39 | |
40 | #include "eio.h" |
40 | #include "eio.h" |
|
|
41 | |
|
|
42 | #ifdef EIO_STACKSIZE |
|
|
43 | # define XTHREAD_STACKSIZE EIO_STACKSIZE |
|
|
44 | #endif |
41 | #include "xthread.h" |
45 | #include "xthread.h" |
42 | |
46 | |
43 | #include <errno.h> |
47 | #include <errno.h> |
44 | #include <stddef.h> |
48 | #include <stddef.h> |
45 | #include <stdlib.h> |
49 | #include <stdlib.h> |
… | |
… | |
64 | #endif |
68 | #endif |
65 | |
69 | |
66 | #ifdef _WIN32 |
70 | #ifdef _WIN32 |
67 | |
71 | |
68 | /*doh*/ |
72 | /*doh*/ |
69 | |
|
|
70 | #else |
73 | #else |
71 | |
74 | |
72 | # include "config.h" |
75 | # include "config.h" |
73 | # include <sys/time.h> |
76 | # include <sys/time.h> |
74 | # include <sys/select.h> |
77 | # include <sys/select.h> |
|
|
78 | # include <sys/mman.h> |
75 | # include <unistd.h> |
79 | # include <unistd.h> |
76 | # include <utime.h> |
80 | # include <utime.h> |
77 | # include <signal.h> |
81 | # include <signal.h> |
78 | # include <dirent.h> |
82 | # include <dirent.h> |
79 | |
83 | |
|
|
84 | /* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */ |
|
|
85 | # if __freebsd || defined __NetBSD__ || defined __OpenBSD__ |
|
|
86 | # define _DIRENT_HAVE_D_TYPE /* sigh */ |
|
|
87 | # define D_INO(de) (de)->d_fileno |
|
|
88 | # define D_NAMLEN(de) (de)->d_namlen |
|
|
89 | # elif __linux || defined d_ino || _XOPEN_SOURCE >= 600 |
|
|
90 | # define D_INO(de) (de)->d_ino |
|
|
91 | # endif |
|
|
92 | |
|
|
93 | #ifdef _D_EXACT_NAMLEN |
|
|
94 | # undef D_NAMLEN |
|
|
95 | # define D_NAMLEN(de) _D_EXACT_NAMLEN (de) |
|
|
96 | #endif |
|
|
97 | |
|
|
98 | # ifdef _DIRENT_HAVE_D_TYPE |
|
|
99 | # define D_TYPE(de) (de)->d_type |
|
|
100 | # endif |
|
|
101 | |
80 | # ifndef EIO_STRUCT_DIRENT |
102 | # ifndef EIO_STRUCT_DIRENT |
81 | # define EIO_STRUCT_DIRENT struct dirent |
103 | # define EIO_STRUCT_DIRENT struct dirent |
82 | # endif |
104 | # endif |
83 | |
105 | |
84 | #endif |
106 | #endif |
85 | |
107 | |
86 | #if HAVE_SENDFILE |
108 | #if HAVE_SENDFILE |
87 | # if __linux |
109 | # if __linux |
88 | # include <sys/sendfile.h> |
110 | # include <sys/sendfile.h> |
89 | # elif __freebsd |
111 | # elif __freebsd || defined __APPLE__ |
90 | # include <sys/socket.h> |
112 | # include <sys/socket.h> |
91 | # include <sys/uio.h> |
113 | # include <sys/uio.h> |
92 | # elif __hpux |
114 | # elif __hpux |
93 | # include <sys/socket.h> |
115 | # include <sys/socket.h> |
94 | # elif __solaris /* not yet */ |
116 | # elif __solaris |
95 | # include <sys/sendfile.h> |
117 | # include <sys/sendfile.h> |
96 | # else |
118 | # else |
97 | # error sendfile support requested but not available |
119 | # error sendfile support requested but not available |
98 | # endif |
120 | # endif |
|
|
121 | #endif |
|
|
122 | |
|
|
123 | #ifndef D_TYPE |
|
|
124 | # define D_TYPE(de) 0 |
|
|
125 | #endif |
|
|
126 | #ifndef D_INO |
|
|
127 | # define D_INO(de) 0 |
|
|
128 | #endif |
|
|
129 | #ifndef D_NAMLEN |
|
|
130 | # define D_NAMLEN(de) strlen ((de)->d_name) |
99 | #endif |
131 | #endif |
100 | |
132 | |
101 | /* number of seconds after which an idle threads exit */ |
133 | /* number of seconds after which an idle threads exit */ |
102 | #define IDLE_TIMEOUT 10 |
134 | #define IDLE_TIMEOUT 10 |
103 | |
135 | |
… | |
… | |
155 | if (wrk->dirp) \ |
187 | if (wrk->dirp) \ |
156 | { \ |
188 | { \ |
157 | closedir (wrk->dirp); \ |
189 | closedir (wrk->dirp); \ |
158 | wrk->dirp = 0; \ |
190 | wrk->dirp = 0; \ |
159 | } |
191 | } |
|
|
192 | |
160 | #define ETP_WORKER_COMMON \ |
193 | #define ETP_WORKER_COMMON \ |
161 | void *dbuf; \ |
194 | void *dbuf; \ |
162 | DIR *dirp; |
195 | DIR *dirp; |
163 | |
196 | |
164 | /*****************************************************************************/ |
197 | /*****************************************************************************/ |
165 | |
198 | |
166 | #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) |
199 | #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) |
167 | |
200 | |
168 | /* calculcate time difference in ~1/EIO_TICKS of a second */ |
201 | /* calculate time difference in ~1/EIO_TICKS of a second */ |
169 | static int tvdiff (struct timeval *tv1, struct timeval *tv2) |
202 | static int tvdiff (struct timeval *tv1, struct timeval *tv2) |
170 | { |
203 | { |
171 | return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS |
204 | return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS |
172 | + ((tv2->tv_usec - tv1->tv_usec) >> 10); |
205 | + ((tv2->tv_usec - tv1->tv_usec) >> 10); |
173 | } |
206 | } |
… | |
… | |
565 | } |
598 | } |
566 | |
599 | |
567 | static void etp_set_max_poll_time (double nseconds) |
600 | static void etp_set_max_poll_time (double nseconds) |
568 | { |
601 | { |
569 | if (WORDACCESS_UNSAFE) X_LOCK (reslock); |
602 | if (WORDACCESS_UNSAFE) X_LOCK (reslock); |
570 | max_poll_time = nseconds; |
603 | max_poll_time = nseconds * EIO_TICKS; |
571 | if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); |
604 | if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); |
572 | } |
605 | } |
573 | |
606 | |
574 | static void etp_set_max_poll_reqs (unsigned int maxreqs) |
607 | static void etp_set_max_poll_reqs (unsigned int maxreqs) |
575 | { |
608 | { |
… | |
… | |
604 | |
637 | |
605 | static void grp_try_feed (eio_req *grp) |
638 | static void grp_try_feed (eio_req *grp) |
606 | { |
639 | { |
607 | while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) |
640 | while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) |
608 | { |
641 | { |
609 | int old_len = grp->size; |
642 | grp->flags &= ~EIO_FLAG_GROUPADD; |
610 | |
643 | |
611 | EIO_FEED (grp); |
644 | EIO_FEED (grp); |
612 | |
645 | |
613 | /* stop if no progress has been made */ |
646 | /* stop if no progress has been made */ |
614 | if (old_len == grp->size) |
647 | if (!(grp->flags & EIO_FLAG_GROUPADD)) |
615 | { |
648 | { |
616 | grp->feed = 0; |
649 | grp->feed = 0; |
617 | break; |
650 | break; |
618 | } |
651 | } |
619 | } |
652 | } |
… | |
… | |
736 | |
769 | |
737 | /*****************************************************************************/ |
770 | /*****************************************************************************/ |
738 | /* work around various missing functions */ |
771 | /* work around various missing functions */ |
739 | |
772 | |
740 | #if !HAVE_PREADWRITE |
773 | #if !HAVE_PREADWRITE |
|
|
774 | # undef pread |
|
|
775 | # undef pwrite |
741 | # define pread eio__pread |
776 | # define pread eio__pread |
742 | # define pwrite eio__pwrite |
777 | # define pwrite eio__pwrite |
743 | |
778 | |
744 | static ssize_t |
779 | static ssize_t |
745 | eio__pread (int fd, void *buf, size_t count, off_t offset) |
780 | eio__pread (int fd, void *buf, size_t count, off_t offset) |
… | |
… | |
765 | |
800 | |
766 | X_LOCK (preadwritelock); |
801 | X_LOCK (preadwritelock); |
767 | ooffset = lseek (fd, 0, SEEK_CUR); |
802 | ooffset = lseek (fd, 0, SEEK_CUR); |
768 | lseek (fd, offset, SEEK_SET); |
803 | lseek (fd, offset, SEEK_SET); |
769 | res = write (fd, buf, count); |
804 | res = write (fd, buf, count); |
770 | lseek (fd, offset, SEEK_SET); |
805 | lseek (fd, ooffset, SEEK_SET); |
771 | X_UNLOCK (preadwritelock); |
806 | X_UNLOCK (preadwritelock); |
772 | |
807 | |
773 | return res; |
808 | return res; |
774 | } |
809 | } |
775 | #endif |
810 | #endif |
776 | |
811 | |
777 | #ifndef HAVE_FUTIMES |
812 | #ifndef HAVE_FUTIMES |
778 | |
813 | |
|
|
814 | # undef utimes |
|
|
815 | # undef futimes |
779 | # define utimes(path,times) eio__utimes (path, times) |
816 | # define utimes(path,times) eio__utimes (path, times) |
780 | # define futimes(fd,times) eio__futimes (fd, times) |
817 | # define futimes(fd,times) eio__futimes (fd, times) |
781 | |
818 | |
782 | static int |
819 | static int |
783 | eio__utimes (const char *filename, const struct timeval times[2]) |
820 | eio__utimes (const char *filename, const struct timeval times[2]) |
… | |
… | |
802 | } |
839 | } |
803 | |
840 | |
804 | #endif |
841 | #endif |
805 | |
842 | |
806 | #if !HAVE_FDATASYNC |
843 | #if !HAVE_FDATASYNC |
|
|
844 | # undef fdatasync |
807 | # define fdatasync fsync |
845 | # define fdatasync(fd) fsync (fd) |
808 | #endif |
846 | #endif |
|
|
847 | |
|
|
848 | /* sync_file_range always needs emulation */ |
|
|
849 | int |
|
|
850 | eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags) |
|
|
851 | { |
|
|
852 | #if HAVE_SYNC_FILE_RANGE |
|
|
853 | int res; |
|
|
854 | |
|
|
855 | if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE |
|
|
856 | || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE |
|
|
857 | || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER) |
|
|
858 | { |
|
|
859 | flags = 0 |
|
|
860 | | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0) |
|
|
861 | | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0) |
|
|
862 | | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0); |
|
|
863 | } |
|
|
864 | |
|
|
865 | res = sync_file_range (fd, offset, nbytes, flags); |
|
|
866 | |
|
|
867 | if (!res || errno != ENOSYS) |
|
|
868 | return res; |
|
|
869 | #endif |
|
|
870 | |
|
|
871 | /* even though we could play tricks with the flags, it's better to always |
|
|
872 | * call fdatasync, as that matches the expectation of its users best */ |
|
|
873 | return fdatasync (fd); |
|
|
874 | } |
809 | |
875 | |
810 | #if !HAVE_READAHEAD |
876 | #if !HAVE_READAHEAD |
|
|
877 | # undef readahead |
811 | # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) |
878 | # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) |
812 | |
879 | |
813 | static ssize_t |
880 | static ssize_t |
814 | eio__readahead (int fd, off_t offset, size_t count, etp_worker *self) |
881 | eio__readahead (int fd, off_t offset, size_t count, etp_worker *self) |
815 | { |
882 | { |
… | |
… | |
856 | if (res < 0 && sbytes) |
923 | if (res < 0 && sbytes) |
857 | /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ |
924 | /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ |
858 | res = sbytes; |
925 | res = sbytes; |
859 | } |
926 | } |
860 | |
927 | |
|
|
928 | # elif defined (__APPLE__) |
|
|
929 | |
|
|
930 | { |
|
|
931 | off_t sbytes = count; |
|
|
932 | res = sendfile (ifd, ofd, offset, &sbytes, 0, 0); |
|
|
933 | |
|
|
934 | if (res < 0 && errno == EAGAIN && sbytes) |
|
|
935 | res = sbytes; |
|
|
936 | } |
|
|
937 | |
861 | # elif __hpux |
938 | # elif __hpux |
862 | res = sendfile (ofd, ifd, offset, count, 0, 0); |
939 | res = sendfile (ofd, ifd, offset, count, 0, 0); |
863 | |
940 | |
864 | # elif __solaris |
941 | # elif __solaris |
865 | { |
942 | { |
… | |
… | |
876 | if (res < 0 && sbytes) |
953 | if (res < 0 && sbytes) |
877 | res = sbytes; |
954 | res = sbytes; |
878 | } |
955 | } |
879 | |
956 | |
880 | # endif |
957 | # endif |
|
|
958 | |
|
|
959 | #elif defined (_WIN32) |
|
|
960 | |
|
|
961 | /* does not work, just for documentation of what would need to be done */ |
|
|
962 | { |
|
|
963 | HANDLE h = TO_SOCKET (ifd); |
|
|
964 | SetFilePointer (h, offset, 0, FILE_BEGIN); |
|
|
965 | res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0); |
|
|
966 | } |
|
|
967 | |
881 | #else |
968 | #else |
882 | res = -1; |
969 | res = -1; |
883 | errno = ENOSYS; |
970 | errno = ENOSYS; |
884 | #endif |
971 | #endif |
885 | |
972 | |
886 | if (res < 0 |
973 | if (res < 0 |
887 | && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK |
974 | && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK |
|
|
975 | || errno == ENOTSUP || errno == EOPNOTSUPP /* BSDs */ |
888 | #if __solaris |
976 | #if __solaris |
889 | || errno == EAFNOSUPPORT || errno == EPROTOTYPE |
977 | || errno == EAFNOSUPPORT || errno == EPROTOTYPE |
890 | #endif |
978 | #endif |
891 | ) |
979 | ) |
892 | ) |
980 | ) |
… | |
… | |
923 | } |
1011 | } |
924 | |
1012 | |
925 | return res; |
1013 | return res; |
926 | } |
1014 | } |
927 | |
1015 | |
|
|
1016 | static signed char |
|
|
1017 | eio_dent_cmp (const eio_dirent *a, const eio_dirent *b) |
|
|
1018 | { |
|
|
1019 | return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */ |
|
|
1020 | : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0; |
|
|
1021 | } |
|
|
1022 | |
|
|
1023 | #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0 |
|
|
1024 | |
|
|
1025 | #define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */ |
|
|
1026 | #define EIO_SORT_FAST 60 /* when to only use insertion sort */ |
|
|
1027 | |
|
|
1028 | static void |
|
|
1029 | eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) |
|
|
1030 | { |
|
|
1031 | unsigned char bits [9 + sizeof (ino_t) * 8]; |
|
|
1032 | unsigned char *bit = bits; |
|
|
1033 | |
|
|
1034 | assert (CHAR_BIT == 8); |
|
|
1035 | assert (sizeof (eio_dirent) * 8 < 256); |
|
|
1036 | assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */ |
|
|
1037 | assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */ |
|
|
1038 | |
|
|
1039 | if (size <= EIO_SORT_FAST) |
|
|
1040 | return; |
|
|
1041 | |
|
|
1042 | /* first prepare an array of bits to test in our radix sort */ |
|
|
1043 | /* try to take endianness into account, as well as differences in ino_t sizes */ |
|
|
1044 | /* inode_bits must contain all inodes ORed together */ |
|
|
1045 | /* which is used to skip bits that are 0 everywhere, which is very common */ |
|
|
1046 | { |
|
|
1047 | ino_t endianness; |
|
|
1048 | int i, j; |
|
|
1049 | |
|
|
1050 | /* we store the byte offset of byte n into byte n of "endianness" */ |
|
|
1051 | for (i = 0; i < sizeof (ino_t); ++i) |
|
|
1052 | ((unsigned char *)&endianness)[i] = i; |
|
|
1053 | |
|
|
1054 | *bit++ = 0; |
|
|
1055 | |
|
|
1056 | for (i = 0; i < sizeof (ino_t); ++i) |
|
|
1057 | { |
|
|
1058 | /* shifting off the byte offsets out of "endianness" */ |
|
|
1059 | int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8; |
|
|
1060 | endianness >>= 8; |
|
|
1061 | |
|
|
1062 | for (j = 0; j < 8; ++j) |
|
|
1063 | if (inode_bits & (((ino_t)1) << (i * 8 + j))) |
|
|
1064 | *bit++ = offs + j; |
|
|
1065 | } |
|
|
1066 | |
|
|
1067 | for (j = 0; j < 8; ++j) |
|
|
1068 | if (score_bits & (1 << j)) |
|
|
1069 | *bit++ = offsetof (eio_dirent, score) * 8 + j; |
|
|
1070 | } |
|
|
1071 | |
|
|
1072 | /* now actually do the sorting (a variant of MSD radix sort) */ |
|
|
1073 | { |
|
|
1074 | eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base; |
|
|
1075 | eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end; |
|
|
1076 | unsigned char *bit_stk [9 + sizeof (ino_t) * 8]; |
|
|
1077 | int stk_idx = 0; |
|
|
1078 | |
|
|
1079 | base_stk [stk_idx] = dents; |
|
|
1080 | end_stk [stk_idx] = dents + size; |
|
|
1081 | bit_stk [stk_idx] = bit - 1; |
|
|
1082 | |
|
|
1083 | do |
|
|
1084 | { |
|
|
1085 | base = base_stk [stk_idx]; |
|
|
1086 | end = end_stk [stk_idx]; |
|
|
1087 | bit = bit_stk [stk_idx]; |
|
|
1088 | |
|
|
1089 | for (;;) |
|
|
1090 | { |
|
|
1091 | unsigned char O = *bit >> 3; |
|
|
1092 | unsigned char M = 1 << (*bit & 7); |
|
|
1093 | |
|
|
1094 | eio_dirent *a = base; |
|
|
1095 | eio_dirent *b = end; |
|
|
1096 | |
|
|
1097 | if (b - a < EIO_SORT_CUTOFF) |
|
|
1098 | break; |
|
|
1099 | |
|
|
1100 | /* now bit-partition the array on the bit */ |
|
|
1101 | /* this ugly asymmetric loop seems to perform much better than typical */ |
|
|
1102 | /* partition algos found in the literature */ |
|
|
1103 | do |
|
|
1104 | if (!(((unsigned char *)a)[O] & M)) |
|
|
1105 | ++a; |
|
|
1106 | else if (!(((unsigned char *)--b)[O] & M)) |
|
|
1107 | { |
|
|
1108 | eio_dirent tmp = *a; *a = *b; *b = tmp; |
|
|
1109 | ++a; |
|
|
1110 | } |
|
|
1111 | while (b > a); |
|
|
1112 | |
|
|
1113 | /* next bit, or stop, if no bits left in this path */ |
|
|
1114 | if (!*--bit) |
|
|
1115 | break; |
|
|
1116 | |
|
|
1117 | base_stk [stk_idx] = a; |
|
|
1118 | end_stk [stk_idx] = end; |
|
|
1119 | bit_stk [stk_idx] = bit; |
|
|
1120 | ++stk_idx; |
|
|
1121 | |
|
|
1122 | end = a; |
|
|
1123 | } |
|
|
1124 | } |
|
|
1125 | while (stk_idx--); |
|
|
1126 | } |
|
|
1127 | } |
|
|
1128 | |
|
|
1129 | static void |
|
|
1130 | eio_dent_insertion_sort (eio_dirent *dents, int size) |
|
|
1131 | { |
|
|
1132 | /* first move the smallest element to the front, to act as a sentinel */ |
|
|
1133 | { |
|
|
1134 | int i; |
|
|
1135 | eio_dirent *min = dents; |
|
|
1136 | |
|
|
1137 | /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */ |
|
|
1138 | for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; ) |
|
|
1139 | if (EIO_DENT_CMP (dents [i], <, *min)) |
|
|
1140 | min = &dents [i]; |
|
|
1141 | |
|
|
1142 | /* swap elements 0 and j (minimum) */ |
|
|
1143 | { |
|
|
1144 | eio_dirent tmp = *dents; *dents = *min; *min = tmp; |
|
|
1145 | } |
|
|
1146 | } |
|
|
1147 | |
|
|
1148 | /* then do standard insertion sort, assuming that all elements are >= dents [0] */ |
|
|
1149 | { |
|
|
1150 | eio_dirent *i, *j; |
|
|
1151 | |
|
|
1152 | for (i = dents + 1; i < dents + size; ++i) |
|
|
1153 | { |
|
|
1154 | eio_dirent value = *i; |
|
|
1155 | |
|
|
1156 | for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j) |
|
|
1157 | j [1] = j [0]; |
|
|
1158 | |
|
|
1159 | j [1] = value; |
|
|
1160 | } |
|
|
1161 | } |
|
|
1162 | } |
|
|
1163 | |
|
|
1164 | static void |
|
|
1165 | eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits) |
|
|
1166 | { |
|
|
1167 | if (size <= 1) |
|
|
1168 | return; /* our insertion sort relies on size > 0 */ |
|
|
1169 | |
|
|
1170 | /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */ |
|
|
1171 | /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */ |
|
|
1172 | eio_dent_radix_sort (dents, size, score_bits, inode_bits); |
|
|
1173 | |
|
|
1174 | /* use an insertion sort at the end, or for small arrays, */ |
|
|
1175 | /* as insertion sort is more efficient for small partitions */ |
|
|
1176 | eio_dent_insertion_sort (dents, size); |
|
|
1177 | } |
|
|
1178 | |
928 | /* read a full directory */ |
1179 | /* read a full directory */ |
929 | static void |
1180 | static void |
930 | eio__scandir (eio_req *req, etp_worker *self) |
1181 | eio__scandir (eio_req *req, etp_worker *self) |
931 | { |
1182 | { |
932 | DIR *dirp; |
1183 | DIR *dirp; |
933 | EIO_STRUCT_DIRENT *entp; |
1184 | EIO_STRUCT_DIRENT *entp; |
934 | char *name, *names; |
1185 | char *name, *names; |
935 | int memlen = 4096; |
1186 | int namesalloc = 4096; |
936 | int memofs = 0; |
1187 | int namesoffs = 0; |
|
|
1188 | int flags = req->int1; |
|
|
1189 | eio_dirent *dents = 0; |
|
|
1190 | int dentalloc = 128; |
937 | int res = 0; |
1191 | int dentoffs = 0; |
|
|
1192 | ino_t inode_bits = 0; |
|
|
1193 | |
|
|
1194 | req->result = -1; |
|
|
1195 | |
|
|
1196 | if (!(flags & EIO_READDIR_DENTS)) |
|
|
1197 | flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER); |
938 | |
1198 | |
939 | X_LOCK (wrklock); |
1199 | X_LOCK (wrklock); |
940 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
1200 | /* the corresponding closedir is in ETP_WORKER_CLEAR */ |
941 | self->dirp = dirp = opendir (req->ptr1); |
1201 | self->dirp = dirp = opendir (req->ptr1); |
942 | req->flags |= EIO_FLAG_PTR2_FREE; |
1202 | req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE; |
|
|
1203 | req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0; |
943 | req->ptr2 = names = malloc (memlen); |
1204 | req->ptr2 = names = malloc (namesalloc); |
944 | X_UNLOCK (wrklock); |
1205 | X_UNLOCK (wrklock); |
945 | |
1206 | |
946 | if (dirp && names) |
1207 | if (dirp && names && (!flags || dents)) |
947 | for (;;) |
1208 | for (;;) |
948 | { |
1209 | { |
949 | errno = 0; |
1210 | errno = 0; |
950 | entp = readdir (dirp); |
1211 | entp = readdir (dirp); |
951 | |
1212 | |
952 | if (!entp) |
1213 | if (!entp) |
|
|
1214 | { |
|
|
1215 | if (errno) |
|
|
1216 | break; |
|
|
1217 | |
|
|
1218 | /* sort etc. */ |
|
|
1219 | req->int1 = flags; |
|
|
1220 | req->result = dentoffs; |
|
|
1221 | |
|
|
1222 | if (flags & EIO_READDIR_STAT_ORDER) |
|
|
1223 | eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */ |
|
|
1224 | else if (flags & EIO_READDIR_DIRS_FIRST) |
|
|
1225 | if (flags & EIO_READDIR_FOUND_UNKNOWN) |
|
|
1226 | eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */ |
|
|
1227 | else |
|
|
1228 | { |
|
|
1229 | /* in this case, all is known, and we just put dirs first and sort them */ |
|
|
1230 | eio_dirent *oth = dents + dentoffs; |
|
|
1231 | eio_dirent *dir = dents; |
|
|
1232 | |
|
|
1233 | /* now partition dirs to the front, and non-dirs to the back */ |
|
|
1234 | /* by walking from both sides and swapping if necessary */ |
|
|
1235 | /* also clear score, so it doesn't influence sorting */ |
|
|
1236 | while (oth > dir) |
|
|
1237 | { |
|
|
1238 | if (dir->type == EIO_DT_DIR) |
|
|
1239 | ++dir; |
|
|
1240 | else if ((--oth)->type == EIO_DT_DIR) |
|
|
1241 | { |
|
|
1242 | eio_dirent tmp = *dir; *dir = *oth; *oth = tmp; |
|
|
1243 | |
|
|
1244 | ++dir; |
|
|
1245 | } |
|
|
1246 | } |
|
|
1247 | |
|
|
1248 | /* now sort the dirs only */ |
|
|
1249 | eio_dent_sort (dents, dir - dents, 0, inode_bits); |
|
|
1250 | } |
|
|
1251 | |
953 | break; |
1252 | break; |
|
|
1253 | } |
954 | |
1254 | |
|
|
1255 | /* now add the entry to our list(s) */ |
955 | name = entp->d_name; |
1256 | name = entp->d_name; |
956 | |
1257 | |
|
|
1258 | /* skip . and .. entries */ |
957 | if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) |
1259 | if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) |
958 | { |
1260 | { |
959 | int len = strlen (name) + 1; |
1261 | int len = D_NAMLEN (entp) + 1; |
960 | |
1262 | |
961 | res++; |
1263 | while (expect_false (namesoffs + len > namesalloc)) |
962 | |
|
|
963 | while (memofs + len > memlen) |
|
|
964 | { |
1264 | { |
965 | memlen *= 2; |
1265 | namesalloc *= 2; |
966 | X_LOCK (wrklock); |
1266 | X_LOCK (wrklock); |
967 | req->ptr2 = names = realloc (names, memlen); |
1267 | req->ptr2 = names = realloc (names, namesalloc); |
968 | X_UNLOCK (wrklock); |
1268 | X_UNLOCK (wrklock); |
969 | |
1269 | |
970 | if (!names) |
1270 | if (!names) |
971 | break; |
1271 | break; |
972 | } |
1272 | } |
973 | |
1273 | |
974 | memcpy (names + memofs, name, len); |
1274 | memcpy (names + namesoffs, name, len); |
|
|
1275 | |
|
|
1276 | if (dents) |
|
|
1277 | { |
|
|
1278 | struct eio_dirent *ent; |
|
|
1279 | |
|
|
1280 | if (expect_false (dentoffs == dentalloc)) |
|
|
1281 | { |
|
|
1282 | dentalloc *= 2; |
|
|
1283 | X_LOCK (wrklock); |
|
|
1284 | req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent)); |
|
|
1285 | X_UNLOCK (wrklock); |
|
|
1286 | |
|
|
1287 | if (!dents) |
|
|
1288 | break; |
|
|
1289 | } |
|
|
1290 | |
|
|
1291 | ent = dents + dentoffs; |
|
|
1292 | |
|
|
1293 | ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */ |
|
|
1294 | ent->namelen = len - 1; |
|
|
1295 | ent->inode = D_INO (entp); |
|
|
1296 | |
|
|
1297 | inode_bits |= ent->inode; |
|
|
1298 | |
|
|
1299 | switch (D_TYPE (entp)) |
|
|
1300 | { |
|
|
1301 | default: |
|
|
1302 | ent->type = EIO_DT_UNKNOWN; |
|
|
1303 | flags |= EIO_READDIR_FOUND_UNKNOWN; |
|
|
1304 | break; |
|
|
1305 | |
|
|
1306 | #ifdef DT_FIFO |
|
|
1307 | case DT_FIFO: ent->type = EIO_DT_FIFO; break; |
|
|
1308 | #endif |
|
|
1309 | #ifdef DT_CHR |
|
|
1310 | case DT_CHR: ent->type = EIO_DT_CHR; break; |
|
|
1311 | #endif |
|
|
1312 | #ifdef DT_MPC |
|
|
1313 | case DT_MPC: ent->type = EIO_DT_MPC; break; |
|
|
1314 | #endif |
|
|
1315 | #ifdef DT_DIR |
|
|
1316 | case DT_DIR: ent->type = EIO_DT_DIR; break; |
|
|
1317 | #endif |
|
|
1318 | #ifdef DT_NAM |
|
|
1319 | case DT_NAM: ent->type = EIO_DT_NAM; break; |
|
|
1320 | #endif |
|
|
1321 | #ifdef DT_BLK |
|
|
1322 | case DT_BLK: ent->type = EIO_DT_BLK; break; |
|
|
1323 | #endif |
|
|
1324 | #ifdef DT_MPB |
|
|
1325 | case DT_MPB: ent->type = EIO_DT_MPB; break; |
|
|
1326 | #endif |
|
|
1327 | #ifdef DT_REG |
|
|
1328 | case DT_REG: ent->type = EIO_DT_REG; break; |
|
|
1329 | #endif |
|
|
1330 | #ifdef DT_NWK |
|
|
1331 | case DT_NWK: ent->type = EIO_DT_NWK; break; |
|
|
1332 | #endif |
|
|
1333 | #ifdef DT_CMP |
|
|
1334 | case DT_CMP: ent->type = EIO_DT_CMP; break; |
|
|
1335 | #endif |
|
|
1336 | #ifdef DT_LNK |
|
|
1337 | case DT_LNK: ent->type = EIO_DT_LNK; break; |
|
|
1338 | #endif |
|
|
1339 | #ifdef DT_SOCK |
|
|
1340 | case DT_SOCK: ent->type = EIO_DT_SOCK; break; |
|
|
1341 | #endif |
|
|
1342 | #ifdef DT_DOOR |
|
|
1343 | case DT_DOOR: ent->type = EIO_DT_DOOR; break; |
|
|
1344 | #endif |
|
|
1345 | #ifdef DT_WHT |
|
|
1346 | case DT_WHT: ent->type = EIO_DT_WHT; break; |
|
|
1347 | #endif |
|
|
1348 | } |
|
|
1349 | |
|
|
1350 | ent->score = 7; |
|
|
1351 | |
|
|
1352 | if (flags & EIO_READDIR_DIRS_FIRST) |
|
|
1353 | { |
|
|
1354 | if (ent->type == EIO_DT_UNKNOWN) |
|
|
1355 | { |
|
|
1356 | if (*name == '.') /* leading dots are likely directories, and, in any case, rare */ |
|
|
1357 | ent->score = 1; |
|
|
1358 | else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */ |
|
|
1359 | ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */ |
|
|
1360 | } |
|
|
1361 | else if (ent->type == EIO_DT_DIR) |
|
|
1362 | ent->score = 0; |
|
|
1363 | } |
|
|
1364 | } |
|
|
1365 | |
975 | memofs += len; |
1366 | namesoffs += len; |
|
|
1367 | ++dentoffs; |
|
|
1368 | } |
|
|
1369 | |
|
|
1370 | if (EIO_CANCELLED (req)) |
|
|
1371 | { |
|
|
1372 | errno = ECANCELED; |
|
|
1373 | break; |
976 | } |
1374 | } |
977 | } |
1375 | } |
978 | |
|
|
979 | if (errno) |
|
|
980 | res = -1; |
|
|
981 | |
|
|
982 | req->result = res; |
|
|
983 | } |
1376 | } |
984 | |
1377 | |
985 | #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) |
1378 | #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO) |
986 | # define msync(a,b,c) ENOSYS |
1379 | # define eio__msync(a,b,c) ((errno = ENOSYS), -1) |
|
|
1380 | #else |
|
|
1381 | |
|
|
1382 | int |
|
|
1383 | eio__msync (void *mem, size_t len, int flags) |
|
|
1384 | { |
|
|
1385 | if (EIO_MS_ASYNC != MS_SYNC |
|
|
1386 | || EIO_MS_INVALIDATE != MS_INVALIDATE |
|
|
1387 | || EIO_MS_SYNC != MS_SYNC) |
|
|
1388 | { |
|
|
1389 | flags = 0 |
|
|
1390 | | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0) |
|
|
1391 | | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0) |
|
|
1392 | | (flags & EIO_MS_SYNC ? MS_SYNC : 0); |
|
|
1393 | } |
|
|
1394 | |
|
|
1395 | return msync (mem, len, flags); |
|
|
1396 | } |
|
|
1397 | |
987 | #endif |
1398 | #endif |
988 | |
1399 | |
989 | int |
1400 | int |
990 | eio__mtouch (void *mem, size_t len, int flags) |
1401 | eio__mtouch (void *mem, size_t len, int flags) |
991 | { |
1402 | { |
… | |
… | |
998 | |
1409 | |
999 | if (!page) |
1410 | if (!page) |
1000 | page = sysconf (_SC_PAGESIZE); |
1411 | page = sysconf (_SC_PAGESIZE); |
1001 | #endif |
1412 | #endif |
1002 | |
1413 | |
|
|
1414 | /* round down to start of page, although this is probably useless */ |
1003 | addr &= ~(page - 1); /* assume page size is always a power of two */ |
1415 | addr &= ~(page - 1); /* assume page size is always a power of two */ |
1004 | |
1416 | |
1005 | if (addr < end) |
1417 | if (addr < end) |
1006 | if (flags) /* modify */ |
1418 | if (flags & EIO_MT_MODIFY) /* modify */ |
1007 | do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); |
1419 | do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len); |
1008 | else |
1420 | else |
1009 | do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); |
1421 | do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len); |
1010 | |
1422 | |
1011 | return 0; |
1423 | return 0; |
… | |
… | |
1183 | req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; |
1595 | req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; |
1184 | |
1596 | |
1185 | case EIO_SYNC: req->result = 0; sync (); break; |
1597 | case EIO_SYNC: req->result = 0; sync (); break; |
1186 | case EIO_FSYNC: req->result = fsync (req->int1); break; |
1598 | case EIO_FSYNC: req->result = fsync (req->int1); break; |
1187 | case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; |
1599 | case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; |
1188 | case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break; |
1600 | case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break; |
1189 | case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; |
1601 | case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break; |
|
|
1602 | case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break; |
1190 | |
1603 | |
1191 | case EIO_READDIR: eio__scandir (req, self); break; |
1604 | case EIO_READDIR: eio__scandir (req, self); break; |
1192 | |
1605 | |
1193 | case EIO_BUSY: |
1606 | case EIO_BUSY: |
1194 | #ifdef _WIN32 |
1607 | #ifdef _WIN32 |
… | |
… | |
1279 | eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) |
1692 | eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data) |
1280 | { |
1693 | { |
1281 | REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; |
1694 | REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND; |
1282 | } |
1695 | } |
1283 | |
1696 | |
|
|
1697 | eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data) |
|
|
1698 | { |
|
|
1699 | REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND; |
|
|
1700 | } |
|
|
1701 | |
1284 | eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) |
1702 | eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) |
1285 | { |
1703 | { |
1286 | REQ (EIO_FDATASYNC); req->int1 = fd; SEND; |
1704 | REQ (EIO_FDATASYNC); req->int1 = fd; SEND; |
1287 | } |
1705 | } |
1288 | |
1706 | |
… | |
… | |
1400 | eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data) |
1818 | eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data) |
1401 | { |
1819 | { |
1402 | return eio__1path (EIO_RMDIR, path, pri, cb, data); |
1820 | return eio__1path (EIO_RMDIR, path, pri, cb, data); |
1403 | } |
1821 | } |
1404 | |
1822 | |
1405 | eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data) |
1823 | eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data) |
1406 | { |
1824 | { |
1407 | return eio__1path (EIO_READDIR, path, pri, cb, data); |
1825 | REQ (EIO_READDIR); PATH; req->int1 = flags; SEND; |
1408 | } |
1826 | } |
1409 | |
1827 | |
1410 | eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) |
1828 | eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) |
1411 | { |
1829 | { |
1412 | REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND; |
1830 | REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND; |
… | |
… | |
1481 | |
1899 | |
1482 | void eio_grp_add (eio_req *grp, eio_req *req) |
1900 | void eio_grp_add (eio_req *grp, eio_req *req) |
1483 | { |
1901 | { |
1484 | assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); |
1902 | assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); |
1485 | |
1903 | |
|
|
1904 | grp->flags |= EIO_FLAG_GROUPADD; |
|
|
1905 | |
1486 | ++grp->size; |
1906 | ++grp->size; |
1487 | req->grp = grp; |
1907 | req->grp = grp; |
1488 | |
1908 | |
1489 | req->grp_prev = 0; |
1909 | req->grp_prev = 0; |
1490 | req->grp_next = grp->grp_first; |
1910 | req->grp_next = grp->grp_first; |
… | |
… | |
1499 | /* misc garbage */ |
1919 | /* misc garbage */ |
1500 | |
1920 | |
1501 | ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) |
1921 | ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) |
1502 | { |
1922 | { |
1503 | etp_worker wrk; |
1923 | etp_worker wrk; |
|
|
1924 | ssize_t ret; |
1504 | |
1925 | |
1505 | wrk.dbuf = 0; |
1926 | wrk.dbuf = 0; |
1506 | |
1927 | |
1507 | eio__sendfile (ofd, ifd, offset, count, &wrk); |
1928 | ret = eio__sendfile (ofd, ifd, offset, count, &wrk); |
1508 | |
1929 | |
1509 | if (wrk.dbuf) |
1930 | if (wrk.dbuf) |
1510 | free (wrk.dbuf); |
1931 | free (wrk.dbuf); |
1511 | } |
|
|
1512 | |
1932 | |
|
|
1933 | return ret; |
|
|
1934 | } |
|
|
1935 | |