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 1.22 by root, Fri Jul 11 10:54:50 2008 UTC vs.
Revision 1.51 by root, Thu Jan 7 18:08:21 2010 UTC

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>
46#include <string.h> 50#include <string.h>
47#include <errno.h> 51#include <errno.h>
48#include <sys/types.h> 52#include <sys/types.h>
49#include <sys/stat.h> 53#include <sys/stat.h>
54#include <sys/statvfs.h>
50#include <limits.h> 55#include <limits.h>
51#include <fcntl.h> 56#include <fcntl.h>
52#include <assert.h> 57#include <assert.h>
53 58
54#ifndef EIO_FINISH 59#ifndef EIO_FINISH
64#endif 69#endif
65 70
66#ifdef _WIN32 71#ifdef _WIN32
67 72
68 /*doh*/ 73 /*doh*/
69
70#else 74#else
71 75
72# include "config.h" 76# include "config.h"
73# include <sys/time.h> 77# include <sys/time.h>
74# include <sys/select.h> 78# include <sys/select.h>
79# include <sys/mman.h>
75# include <unistd.h> 80# include <unistd.h>
76# include <utime.h> 81# include <utime.h>
77# include <signal.h> 82# include <signal.h>
78# include <dirent.h> 83# include <dirent.h>
79 84
85/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
86# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
87# define _DIRENT_HAVE_D_TYPE /* sigh */
88# define D_INO(de) (de)->d_fileno
89# define D_NAMLEN(de) (de)->d_namlen
90# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
91# define D_INO(de) (de)->d_ino
92# endif
93
94#ifdef _D_EXACT_NAMLEN
95# undef D_NAMLEN
96# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
97#endif
98
99# ifdef _DIRENT_HAVE_D_TYPE
100# define D_TYPE(de) (de)->d_type
101# endif
102
80# ifndef EIO_STRUCT_DIRENT 103# ifndef EIO_STRUCT_DIRENT
81# define EIO_STRUCT_DIRENT struct dirent 104# define EIO_STRUCT_DIRENT struct dirent
82# endif 105# endif
83 106
84#endif 107#endif
85 108
86#if HAVE_SENDFILE 109#if HAVE_SENDFILE
87# if __linux 110# if __linux
88# include <sys/sendfile.h> 111# include <sys/sendfile.h>
89# elif __freebsd 112# elif __FreeBSD__ || defined __APPLE__
90# include <sys/socket.h> 113# include <sys/socket.h>
91# include <sys/uio.h> 114# include <sys/uio.h>
92# elif __hpux 115# elif __hpux
93# include <sys/socket.h> 116# include <sys/socket.h>
94# elif __solaris /* not yet */ 117# elif __solaris
95# include <sys/sendfile.h> 118# include <sys/sendfile.h>
96# else 119# else
97# error sendfile support requested but not available 120# error sendfile support requested but not available
98# endif 121# endif
122#endif
123
124#ifndef D_TYPE
125# define D_TYPE(de) 0
126#endif
127#ifndef D_INO
128# define D_INO(de) 0
129#endif
130#ifndef D_NAMLEN
131# define D_NAMLEN(de) strlen ((de)->d_name)
99#endif 132#endif
100 133
101/* number of seconds after which an idle threads exit */ 134/* number of seconds after which an idle threads exit */
102#define IDLE_TIMEOUT 10 135#define IDLE_TIMEOUT 10
103 136
117 errno = ENOMEM; \ 150 errno = ENOMEM; \
118 if (!eio_buf) \ 151 if (!eio_buf) \
119 return -1; 152 return -1;
120 153
121#define EIO_TICKS ((1000000 + 1023) >> 10) 154#define EIO_TICKS ((1000000 + 1023) >> 10)
155
156/*****************************************************************************/
157
158#if __GNUC__ >= 3
159# define expect(expr,value) __builtin_expect ((expr),(value))
160#else
161# define expect(expr,value) (expr)
162#endif
163
164#define expect_false(expr) expect ((expr) != 0, 0)
165#define expect_true(expr) expect ((expr) != 0, 1)
122 166
123/*****************************************************************************/ 167/*****************************************************************************/
124 168
125#define ETP_PRI_MIN EIO_PRI_MIN 169#define ETP_PRI_MIN EIO_PRI_MIN
126#define ETP_PRI_MAX EIO_PRI_MAX 170#define ETP_PRI_MAX EIO_PRI_MAX
144 if (wrk->dirp) \ 188 if (wrk->dirp) \
145 { \ 189 { \
146 closedir (wrk->dirp); \ 190 closedir (wrk->dirp); \
147 wrk->dirp = 0; \ 191 wrk->dirp = 0; \
148 } 192 }
193
149#define ETP_WORKER_COMMON \ 194#define ETP_WORKER_COMMON \
150 void *dbuf; \ 195 void *dbuf; \
151 DIR *dirp; 196 DIR *dirp;
152 197
153/*****************************************************************************/ 198/*****************************************************************************/
154 199
155#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 200#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
156 201
157/* calculcate time difference in ~1/EIO_TICKS of a second */ 202/* calculate time difference in ~1/EIO_TICKS of a second */
158static int tvdiff (struct timeval *tv1, struct timeval *tv2) 203static int tvdiff (struct timeval *tv1, struct timeval *tv2)
159{ 204{
160 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 205 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
161 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 206 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
162} 207}
411 X_UNLOCK (wrklock); 456 X_UNLOCK (wrklock);
412} 457}
413 458
414static void etp_maybe_start_thread (void) 459static void etp_maybe_start_thread (void)
415{ 460{
416 if (etp_nthreads () >= wanted) 461 if (expect_true (etp_nthreads () >= wanted))
417 return; 462 return;
418 463
419 /* todo: maybe use idle here, but might be less exact */ 464 /* todo: maybe use idle here, but might be less exact */
420 if (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()) 465 if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
421 return; 466 return;
422 467
423 etp_start_thread (); 468 etp_start_thread ();
424} 469}
425 470
478 523
479 X_LOCK (reqlock); 524 X_LOCK (reqlock);
480 --nreqs; 525 --nreqs;
481 X_UNLOCK (reqlock); 526 X_UNLOCK (reqlock);
482 527
483 if (req->type == EIO_GROUP && req->size) 528 if (expect_false (req->type == EIO_GROUP && req->size))
484 { 529 {
485 req->int1 = 1; /* mark request as delayed */ 530 req->int1 = 1; /* mark request as delayed */
486 continue; 531 continue;
487 } 532 }
488 else 533 else
489 { 534 {
490 int res = ETP_FINISH (req); 535 int res = ETP_FINISH (req);
491 if (res) 536 if (expect_false (res))
492 return res; 537 return res;
493 } 538 }
494 539
495 if (maxreqs && !--maxreqs) 540 if (expect_false (maxreqs && !--maxreqs))
496 break; 541 break;
497 542
498 if (maxtime) 543 if (maxtime)
499 { 544 {
500 gettimeofday (&tv_now, 0); 545 gettimeofday (&tv_now, 0);
519 564
520static void etp_submit (ETP_REQ *req) 565static void etp_submit (ETP_REQ *req)
521{ 566{
522 req->pri -= ETP_PRI_MIN; 567 req->pri -= ETP_PRI_MIN;
523 568
524 if (req->pri < ETP_PRI_MIN - ETP_PRI_MIN) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; 569 if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
525 if (req->pri > ETP_PRI_MAX - ETP_PRI_MIN) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; 570 if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
526 571
572 if (expect_false (req->type == EIO_GROUP))
573 {
574 /* I hope this is worth it :/ */
527 X_LOCK (reqlock); 575 X_LOCK (reqlock);
528 ++nreqs; 576 ++nreqs;
577 X_UNLOCK (reqlock);
578
579 X_LOCK (reslock);
580
581 ++npending;
582
583 if (!reqq_push (&res_queue, req) && want_poll_cb)
584 want_poll_cb ();
585
586 X_UNLOCK (reslock);
587 }
588 else
589 {
590 X_LOCK (reqlock);
591 ++nreqs;
529 ++nready; 592 ++nready;
530 reqq_push (&req_queue, req); 593 reqq_push (&req_queue, req);
531 X_COND_SIGNAL (reqwait); 594 X_COND_SIGNAL (reqwait);
532 X_UNLOCK (reqlock); 595 X_UNLOCK (reqlock);
533 596
534 etp_maybe_start_thread (); 597 etp_maybe_start_thread ();
598 }
535} 599}
536 600
537static void etp_set_max_poll_time (double nseconds) 601static void etp_set_max_poll_time (double nseconds)
538{ 602{
539 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 603 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
540 max_poll_time = nseconds; 604 max_poll_time = nseconds * EIO_TICKS;
541 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 605 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
542} 606}
543 607
544static void etp_set_max_poll_reqs (unsigned int maxreqs) 608static void etp_set_max_poll_reqs (unsigned int maxreqs)
545{ 609{
574 638
575static void grp_try_feed (eio_req *grp) 639static void grp_try_feed (eio_req *grp)
576{ 640{
577 while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) 641 while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
578 { 642 {
579 int old_len = grp->size; 643 grp->flags &= ~EIO_FLAG_GROUPADD;
580 644
581 EIO_FEED (grp); 645 EIO_FEED (grp);
582 646
583 /* stop if no progress has been made */ 647 /* stop if no progress has been made */
584 if (old_len == grp->size) 648 if (!(grp->flags & EIO_FLAG_GROUPADD))
585 { 649 {
586 grp->feed = 0; 650 grp->feed = 0;
587 break; 651 break;
588 } 652 }
589 } 653 }
706 770
707/*****************************************************************************/ 771/*****************************************************************************/
708/* work around various missing functions */ 772/* work around various missing functions */
709 773
710#if !HAVE_PREADWRITE 774#if !HAVE_PREADWRITE
775# undef pread
776# undef pwrite
711# define pread eio__pread 777# define pread eio__pread
712# define pwrite eio__pwrite 778# define pwrite eio__pwrite
713 779
714static ssize_t 780static ssize_t
715eio__pread (int fd, void *buf, size_t count, off_t offset) 781eio__pread (int fd, void *buf, size_t count, off_t offset)
735 801
736 X_LOCK (preadwritelock); 802 X_LOCK (preadwritelock);
737 ooffset = lseek (fd, 0, SEEK_CUR); 803 ooffset = lseek (fd, 0, SEEK_CUR);
738 lseek (fd, offset, SEEK_SET); 804 lseek (fd, offset, SEEK_SET);
739 res = write (fd, buf, count); 805 res = write (fd, buf, count);
740 lseek (fd, offset, SEEK_SET); 806 lseek (fd, ooffset, SEEK_SET);
741 X_UNLOCK (preadwritelock); 807 X_UNLOCK (preadwritelock);
742 808
743 return res; 809 return res;
744} 810}
745#endif 811#endif
746 812
747#ifndef HAVE_FUTIMES 813#ifndef HAVE_FUTIMES
748 814
815# undef utimes
816# undef futimes
749# define utimes(path,times) eio__utimes (path, times) 817# define utimes(path,times) eio__utimes (path, times)
750# define futimes(fd,times) eio__futimes (fd, times) 818# define futimes(fd,times) eio__futimes (fd, times)
751 819
752static int 820static int
753eio__utimes (const char *filename, const struct timeval times[2]) 821eio__utimes (const char *filename, const struct timeval times[2])
772} 840}
773 841
774#endif 842#endif
775 843
776#if !HAVE_FDATASYNC 844#if !HAVE_FDATASYNC
845# undef fdatasync
777# define fdatasync fsync 846# define fdatasync(fd) fsync (fd)
778#endif 847#endif
848
849/* sync_file_range always needs emulation */
850int
851eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
852{
853#if HAVE_SYNC_FILE_RANGE
854 int res;
855
856 if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
857 || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
858 || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
859 {
860 flags = 0
861 | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
862 | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
863 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
864 }
865
866 res = sync_file_range (fd, offset, nbytes, flags);
867
868 if (!res || errno != ENOSYS)
869 return res;
870#endif
871
872 /* even though we could play tricks with the flags, it's better to always
873 * call fdatasync, as that matches the expectation of its users best */
874 return fdatasync (fd);
875}
779 876
780#if !HAVE_READAHEAD 877#if !HAVE_READAHEAD
878# undef readahead
781# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) 879# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
782 880
783static ssize_t 881static ssize_t
784eio__readahead (int fd, off_t offset, size_t count, etp_worker *self) 882eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
785{ 883{
812 910
813#if HAVE_SENDFILE 911#if HAVE_SENDFILE
814# if __linux 912# if __linux
815 res = sendfile (ofd, ifd, &offset, count); 913 res = sendfile (ofd, ifd, &offset, count);
816 914
817# elif __freebsd 915# elif __FreeBSD__
818 /* 916 /*
819 * Of course, the freebsd sendfile is a dire hack with no thoughts 917 * Of course, the freebsd sendfile is a dire hack with no thoughts
820 * wasted on making it similar to other I/O functions. 918 * wasted on making it similar to other I/O functions.
821 */ 919 */
822 { 920 {
823 off_t sbytes; 921 off_t sbytes;
824 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0); 922 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
825 923
924 /* freebsd' sendfile will return 0 when success */
826 if (res < 0 && sbytes) 925 if (res == 0 && sbytes)
827 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 926 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
927 res = sbytes;
928 }
929
930# elif defined (__APPLE__)
931
932 {
933 off_t sbytes = count;
934 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
935
936 if (res == 0 && errno == EAGAIN && sbytes)
828 res = sbytes; 937 res = sbytes;
829 } 938 }
830 939
831# elif __hpux 940# elif __hpux
832 res = sendfile (ofd, ifd, offset, count, 0, 0); 941 res = sendfile (ofd, ifd, offset, count, 0, 0);
846 if (res < 0 && sbytes) 955 if (res < 0 && sbytes)
847 res = sbytes; 956 res = sbytes;
848 } 957 }
849 958
850# endif 959# endif
960
961#elif defined (_WIN32)
962
963 /* does not work, just for documentation of what would need to be done */
964 {
965 HANDLE h = TO_SOCKET (ifd);
966 SetFilePointer (h, offset, 0, FILE_BEGIN);
967 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
968 }
969
851#else 970#else
852 res = -1; 971 res = -1;
853 errno = ENOSYS; 972 errno = ENOSYS;
854#endif 973#endif
855 974
856 if (res < 0 975 if (res < 0
857 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 976 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
977 /* BSDs */
978#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
979 || errno == ENOTSUP
980#endif
981 || errno == EOPNOTSUPP /* BSDs */
858#if __solaris 982#if __solaris
859 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 983 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
860#endif 984#endif
861 ) 985 )
862 ) 986 )
893 } 1017 }
894 1018
895 return res; 1019 return res;
896} 1020}
897 1021
1022static signed char
1023eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
1024{
1025 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
1026 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1027}
1028
1029#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1030
1031#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1032#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1033
1034static void
1035eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1036{
1037 unsigned char bits [9 + sizeof (ino_t) * 8];
1038 unsigned char *bit = bits;
1039
1040 assert (CHAR_BIT == 8);
1041 assert (sizeof (eio_dirent) * 8 < 256);
1042 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1043 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1044
1045 if (size <= EIO_SORT_FAST)
1046 return;
1047
1048 /* first prepare an array of bits to test in our radix sort */
1049 /* try to take endianness into account, as well as differences in ino_t sizes */
1050 /* inode_bits must contain all inodes ORed together */
1051 /* which is used to skip bits that are 0 everywhere, which is very common */
1052 {
1053 ino_t endianness;
1054 int i, j;
1055
1056 /* we store the byte offset of byte n into byte n of "endianness" */
1057 for (i = 0; i < sizeof (ino_t); ++i)
1058 ((unsigned char *)&endianness)[i] = i;
1059
1060 *bit++ = 0;
1061
1062 for (i = 0; i < sizeof (ino_t); ++i)
1063 {
1064 /* shifting off the byte offsets out of "endianness" */
1065 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1066 endianness >>= 8;
1067
1068 for (j = 0; j < 8; ++j)
1069 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1070 *bit++ = offs + j;
1071 }
1072
1073 for (j = 0; j < 8; ++j)
1074 if (score_bits & (1 << j))
1075 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1076 }
1077
1078 /* now actually do the sorting (a variant of MSD radix sort) */
1079 {
1080 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1081 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1082 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1083 int stk_idx = 0;
1084
1085 base_stk [stk_idx] = dents;
1086 end_stk [stk_idx] = dents + size;
1087 bit_stk [stk_idx] = bit - 1;
1088
1089 do
1090 {
1091 base = base_stk [stk_idx];
1092 end = end_stk [stk_idx];
1093 bit = bit_stk [stk_idx];
1094
1095 for (;;)
1096 {
1097 unsigned char O = *bit >> 3;
1098 unsigned char M = 1 << (*bit & 7);
1099
1100 eio_dirent *a = base;
1101 eio_dirent *b = end;
1102
1103 if (b - a < EIO_SORT_CUTOFF)
1104 break;
1105
1106 /* now bit-partition the array on the bit */
1107 /* this ugly asymmetric loop seems to perform much better than typical */
1108 /* partition algos found in the literature */
1109 do
1110 if (!(((unsigned char *)a)[O] & M))
1111 ++a;
1112 else if (!(((unsigned char *)--b)[O] & M))
1113 {
1114 eio_dirent tmp = *a; *a = *b; *b = tmp;
1115 ++a;
1116 }
1117 while (b > a);
1118
1119 /* next bit, or stop, if no bits left in this path */
1120 if (!*--bit)
1121 break;
1122
1123 base_stk [stk_idx] = a;
1124 end_stk [stk_idx] = end;
1125 bit_stk [stk_idx] = bit;
1126 ++stk_idx;
1127
1128 end = a;
1129 }
1130 }
1131 while (stk_idx--);
1132 }
1133}
1134
1135static void
1136eio_dent_insertion_sort (eio_dirent *dents, int size)
1137{
1138 /* first move the smallest element to the front, to act as a sentinel */
1139 {
1140 int i;
1141 eio_dirent *min = dents;
1142
1143 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1144 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1145 if (EIO_DENT_CMP (dents [i], <, *min))
1146 min = &dents [i];
1147
1148 /* swap elements 0 and j (minimum) */
1149 {
1150 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1151 }
1152 }
1153
1154 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1155 {
1156 eio_dirent *i, *j;
1157
1158 for (i = dents + 1; i < dents + size; ++i)
1159 {
1160 eio_dirent value = *i;
1161
1162 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1163 j [1] = j [0];
1164
1165 j [1] = value;
1166 }
1167 }
1168}
1169
1170static void
1171eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1172{
1173 if (size <= 1)
1174 return; /* our insertion sort relies on size > 0 */
1175
1176 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1177 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1178 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1179
1180 /* use an insertion sort at the end, or for small arrays, */
1181 /* as insertion sort is more efficient for small partitions */
1182 eio_dent_insertion_sort (dents, size);
1183}
1184
898/* read a full directory */ 1185/* read a full directory */
899static void 1186static void
900eio__scandir (eio_req *req, etp_worker *self) 1187eio__scandir (eio_req *req, etp_worker *self)
901{ 1188{
902 DIR *dirp; 1189 DIR *dirp;
903 EIO_STRUCT_DIRENT *entp; 1190 EIO_STRUCT_DIRENT *entp;
904 char *name, *names; 1191 char *name, *names;
905 int memlen = 4096; 1192 int namesalloc = 4096;
906 int memofs = 0; 1193 int namesoffs = 0;
1194 int flags = req->int1;
1195 eio_dirent *dents = 0;
1196 int dentalloc = 128;
907 int res = 0; 1197 int dentoffs = 0;
1198 ino_t inode_bits = 0;
1199
1200 req->result = -1;
1201
1202 if (!(flags & EIO_READDIR_DENTS))
1203 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
908 1204
909 X_LOCK (wrklock); 1205 X_LOCK (wrklock);
910 /* the corresponding closedir is in ETP_WORKER_CLEAR */ 1206 /* the corresponding closedir is in ETP_WORKER_CLEAR */
911 self->dirp = dirp = opendir (req->ptr1); 1207 self->dirp = dirp = opendir (req->ptr1);
912 req->flags |= EIO_FLAG_PTR2_FREE; 1208 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1209 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
913 req->ptr2 = names = malloc (memlen); 1210 req->ptr2 = names = malloc (namesalloc);
914 X_UNLOCK (wrklock); 1211 X_UNLOCK (wrklock);
915 1212
916 if (dirp && names) 1213 if (dirp && names && (!flags || dents))
917 for (;;) 1214 for (;;)
918 { 1215 {
919 errno = 0; 1216 errno = 0;
920 entp = readdir (dirp); 1217 entp = readdir (dirp);
921 1218
922 if (!entp) 1219 if (!entp)
1220 {
1221 if (errno)
1222 break;
1223
1224 /* sort etc. */
1225 req->int1 = flags;
1226 req->result = dentoffs;
1227
1228 if (flags & EIO_READDIR_STAT_ORDER)
1229 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1230 else if (flags & EIO_READDIR_DIRS_FIRST)
1231 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1232 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1233 else
1234 {
1235 /* in this case, all is known, and we just put dirs first and sort them */
1236 eio_dirent *oth = dents + dentoffs;
1237 eio_dirent *dir = dents;
1238
1239 /* now partition dirs to the front, and non-dirs to the back */
1240 /* by walking from both sides and swapping if necessary */
1241 /* also clear score, so it doesn't influence sorting */
1242 while (oth > dir)
1243 {
1244 if (dir->type == EIO_DT_DIR)
1245 ++dir;
1246 else if ((--oth)->type == EIO_DT_DIR)
1247 {
1248 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1249
1250 ++dir;
1251 }
1252 }
1253
1254 /* now sort the dirs only */
1255 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1256 }
1257
923 break; 1258 break;
1259 }
924 1260
1261 /* now add the entry to our list(s) */
925 name = entp->d_name; 1262 name = entp->d_name;
926 1263
1264 /* skip . and .. entries */
927 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) 1265 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
928 { 1266 {
929 int len = strlen (name) + 1; 1267 int len = D_NAMLEN (entp) + 1;
930 1268
931 res++; 1269 while (expect_false (namesoffs + len > namesalloc))
932
933 while (memofs + len > memlen)
934 { 1270 {
935 memlen *= 2; 1271 namesalloc *= 2;
936 X_LOCK (wrklock); 1272 X_LOCK (wrklock);
937 req->ptr2 = names = realloc (names, memlen); 1273 req->ptr2 = names = realloc (names, namesalloc);
938 X_UNLOCK (wrklock); 1274 X_UNLOCK (wrklock);
939 1275
940 if (!names) 1276 if (!names)
941 break; 1277 break;
942 } 1278 }
943 1279
944 memcpy (names + memofs, name, len); 1280 memcpy (names + namesoffs, name, len);
1281
1282 if (dents)
1283 {
1284 struct eio_dirent *ent;
1285
1286 if (expect_false (dentoffs == dentalloc))
1287 {
1288 dentalloc *= 2;
1289 X_LOCK (wrklock);
1290 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1291 X_UNLOCK (wrklock);
1292
1293 if (!dents)
1294 break;
1295 }
1296
1297 ent = dents + dentoffs;
1298
1299 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1300 ent->namelen = len - 1;
1301 ent->inode = D_INO (entp);
1302
1303 inode_bits |= ent->inode;
1304
1305 switch (D_TYPE (entp))
1306 {
1307 default:
1308 ent->type = EIO_DT_UNKNOWN;
1309 flags |= EIO_READDIR_FOUND_UNKNOWN;
1310 break;
1311
1312 #ifdef DT_FIFO
1313 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1314 #endif
1315 #ifdef DT_CHR
1316 case DT_CHR: ent->type = EIO_DT_CHR; break;
1317 #endif
1318 #ifdef DT_MPC
1319 case DT_MPC: ent->type = EIO_DT_MPC; break;
1320 #endif
1321 #ifdef DT_DIR
1322 case DT_DIR: ent->type = EIO_DT_DIR; break;
1323 #endif
1324 #ifdef DT_NAM
1325 case DT_NAM: ent->type = EIO_DT_NAM; break;
1326 #endif
1327 #ifdef DT_BLK
1328 case DT_BLK: ent->type = EIO_DT_BLK; break;
1329 #endif
1330 #ifdef DT_MPB
1331 case DT_MPB: ent->type = EIO_DT_MPB; break;
1332 #endif
1333 #ifdef DT_REG
1334 case DT_REG: ent->type = EIO_DT_REG; break;
1335 #endif
1336 #ifdef DT_NWK
1337 case DT_NWK: ent->type = EIO_DT_NWK; break;
1338 #endif
1339 #ifdef DT_CMP
1340 case DT_CMP: ent->type = EIO_DT_CMP; break;
1341 #endif
1342 #ifdef DT_LNK
1343 case DT_LNK: ent->type = EIO_DT_LNK; break;
1344 #endif
1345 #ifdef DT_SOCK
1346 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1347 #endif
1348 #ifdef DT_DOOR
1349 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1350 #endif
1351 #ifdef DT_WHT
1352 case DT_WHT: ent->type = EIO_DT_WHT; break;
1353 #endif
1354 }
1355
1356 ent->score = 7;
1357
1358 if (flags & EIO_READDIR_DIRS_FIRST)
1359 {
1360 if (ent->type == EIO_DT_UNKNOWN)
1361 {
1362 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1363 ent->score = 1;
1364 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1365 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1366 }
1367 else if (ent->type == EIO_DT_DIR)
1368 ent->score = 0;
1369 }
1370 }
1371
945 memofs += len; 1372 namesoffs += len;
1373 ++dentoffs;
1374 }
1375
1376 if (EIO_CANCELLED (req))
1377 {
1378 errno = ECANCELED;
1379 break;
946 } 1380 }
947 } 1381 }
1382}
948 1383
949 if (errno) 1384#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
950 res = -1; 1385# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
951 1386#else
952 req->result = res; 1387
1388int
1389eio__msync (void *mem, size_t len, int flags)
1390{
1391 if (EIO_MS_ASYNC != MS_SYNC
1392 || EIO_MS_INVALIDATE != MS_INVALIDATE
1393 || EIO_MS_SYNC != MS_SYNC)
1394 {
1395 flags = 0
1396 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1397 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1398 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1399 }
1400
1401 return msync (mem, len, flags);
1402}
1403
1404#endif
1405
1406int
1407eio__mtouch (void *mem, size_t len, int flags)
1408{
1409 intptr_t addr = (intptr_t)mem;
1410 intptr_t end = addr + len;
1411#ifdef PAGESIZE
1412 const intptr_t page = PAGESIZE;
1413#else
1414 static intptr_t page;
1415
1416 if (!page)
1417 page = sysconf (_SC_PAGESIZE);
1418#endif
1419
1420 /* round down to start of page, although this is probably useless */
1421 addr &= ~(page - 1); /* assume page size is always a power of two */
1422
1423 if (addr < end)
1424 if (flags & EIO_MT_MODIFY) /* modify */
1425 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1426 else
1427 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1428
1429 return 0;
953} 1430}
954 1431
955/*****************************************************************************/ 1432/*****************************************************************************/
956 1433
957#define ALLOC(len) \ 1434#define ALLOC(len) \
1100 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1577 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1101 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break; 1578 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1102 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT)); 1579 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1103 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break; 1580 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1104 1581
1582 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1583 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1584 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1585 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1586
1105 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break; 1587 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1106 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break; 1588 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1107 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break; 1589 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1108 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break; 1590 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
1109 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break; 1591 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
1124 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1606 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1125 1607
1126 case EIO_SYNC: req->result = 0; sync (); break; 1608 case EIO_SYNC: req->result = 0; sync (); break;
1127 case EIO_FSYNC: req->result = fsync (req->int1); break; 1609 case EIO_FSYNC: req->result = fsync (req->int1); break;
1128 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1610 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1611 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1612 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1613 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1129 1614
1130 case EIO_READDIR: eio__scandir (req, self); break; 1615 case EIO_READDIR: eio__scandir (req, self); break;
1131 1616
1132 case EIO_BUSY: 1617 case EIO_BUSY:
1133#ifdef _WIN32 1618#ifdef _WIN32
1165 1650
1166 req->result = req->type == EIO_FUTIME 1651 req->result = req->type == EIO_FUTIME
1167 ? futimes (req->int1, times) 1652 ? futimes (req->int1, times)
1168 : utimes (req->ptr1, times); 1653 : utimes (req->ptr1, times);
1169 } 1654 }
1655 break;
1170 1656
1171 case EIO_GROUP: 1657 case EIO_GROUP:
1658 abort (); /* handled in eio_request */
1659
1172 case EIO_NOP: 1660 case EIO_NOP:
1173 req->result = 0; 1661 req->result = 0;
1174 break; 1662 break;
1175 1663
1176 case EIO_CUSTOM: 1664 case EIO_CUSTOM:
1177 req->feed (req); 1665 ((void (*)(eio_req *))req->feed) (req);
1178 break; 1666 break;
1179 1667
1180 default: 1668 default:
1181 req->result = -1; 1669 req->result = -1;
1182 break; 1670 break;
1205eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data) 1693eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1206{ 1694{
1207 REQ (EIO_FSYNC); req->int1 = fd; SEND; 1695 REQ (EIO_FSYNC); req->int1 = fd; SEND;
1208} 1696}
1209 1697
1698eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1699{
1700 REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1701}
1702
1703eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1704{
1705 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1706}
1707
1708eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1709{
1710 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1711}
1712
1210eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) 1713eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1211{ 1714{
1212 REQ (EIO_FDATASYNC); req->int1 = fd; SEND; 1715 REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1213} 1716}
1214 1717
1233} 1736}
1234 1737
1235eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data) 1738eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1236{ 1739{
1237 REQ (EIO_FSTAT); req->int1 = fd; SEND; 1740 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1741}
1742
1743eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1744{
1745 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1238} 1746}
1239 1747
1240eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data) 1748eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1241{ 1749{
1242 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND; 1750 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
1316eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data) 1824eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1317{ 1825{
1318 return eio__1path (EIO_LSTAT, path, pri, cb, data); 1826 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1319} 1827}
1320 1828
1829eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
1830{
1831 return eio__1path (EIO_STATVFS, path, pri, cb, data);
1832}
1833
1321eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data) 1834eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1322{ 1835{
1323 return eio__1path (EIO_UNLINK, path, pri, cb, data); 1836 return eio__1path (EIO_UNLINK, path, pri, cb, data);
1324} 1837}
1325 1838
1326eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data) 1839eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1327{ 1840{
1328 return eio__1path (EIO_RMDIR, path, pri, cb, data); 1841 return eio__1path (EIO_RMDIR, path, pri, cb, data);
1329} 1842}
1330 1843
1331eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data) 1844eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1332{ 1845{
1333 return eio__1path (EIO_READDIR, path, pri, cb, data); 1846 REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1334} 1847}
1335 1848
1336eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) 1849eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1337{ 1850{
1338 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND; 1851 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1369 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data); 1882 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1370} 1883}
1371 1884
1372eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data) 1885eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
1373{ 1886{
1374 REQ (EIO_CUSTOM); req->feed = execute; SEND; 1887 REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
1375} 1888}
1376 1889
1377#endif 1890#endif
1378 1891
1379eio_req *eio_grp (eio_cb cb, void *data) 1892eio_req *eio_grp (eio_cb cb, void *data)
1407 1920
1408void eio_grp_add (eio_req *grp, eio_req *req) 1921void eio_grp_add (eio_req *grp, eio_req *req)
1409{ 1922{
1410 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); 1923 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1411 1924
1925 grp->flags |= EIO_FLAG_GROUPADD;
1926
1412 ++grp->size; 1927 ++grp->size;
1413 req->grp = grp; 1928 req->grp = grp;
1414 1929
1415 req->grp_prev = 0; 1930 req->grp_prev = 0;
1416 req->grp_next = grp->grp_first; 1931 req->grp_next = grp->grp_first;
1425/* misc garbage */ 1940/* misc garbage */
1426 1941
1427ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) 1942ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1428{ 1943{
1429 etp_worker wrk; 1944 etp_worker wrk;
1945 ssize_t ret;
1430 1946
1431 wrk.dbuf = 0; 1947 wrk.dbuf = 0;
1432 1948
1433 eio__sendfile (ofd, ifd, offset, count, &wrk); 1949 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1434 1950
1435 if (wrk.dbuf) 1951 if (wrk.dbuf)
1436 free (wrk.dbuf); 1952 free (wrk.dbuf);
1437}
1438 1953
1954 return ret;
1955}
1956

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines