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

Comparing libeio/eio.c (file contents):
Revision 1.17 by root, Tue Jun 3 05:12:51 2008 UTC vs.
Revision 1.49 by root, Sat Jan 2 14:24:32 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>
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
117 errno = ENOMEM; \ 149 errno = ENOMEM; \
118 if (!eio_buf) \ 150 if (!eio_buf) \
119 return -1; 151 return -1;
120 152
121#define EIO_TICKS ((1000000 + 1023) >> 10) 153#define EIO_TICKS ((1000000 + 1023) >> 10)
154
155/*****************************************************************************/
156
157#if __GNUC__ >= 3
158# define expect(expr,value) __builtin_expect ((expr),(value))
159#else
160# define expect(expr,value) (expr)
161#endif
162
163#define expect_false(expr) expect ((expr) != 0, 0)
164#define expect_true(expr) expect ((expr) != 0, 1)
122 165
123/*****************************************************************************/ 166/*****************************************************************************/
124 167
125#define ETP_PRI_MIN EIO_PRI_MIN 168#define ETP_PRI_MIN EIO_PRI_MIN
126#define ETP_PRI_MAX EIO_PRI_MAX 169#define ETP_PRI_MAX EIO_PRI_MAX
144 if (wrk->dirp) \ 187 if (wrk->dirp) \
145 { \ 188 { \
146 closedir (wrk->dirp); \ 189 closedir (wrk->dirp); \
147 wrk->dirp = 0; \ 190 wrk->dirp = 0; \
148 } 191 }
192
149#define ETP_WORKER_COMMON \ 193#define ETP_WORKER_COMMON \
150 void *dbuf; \ 194 void *dbuf; \
151 DIR *dirp; 195 DIR *dirp;
152 196
153/*****************************************************************************/ 197/*****************************************************************************/
154 198
155#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1) 199#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
156 200
157/* calculcate time difference in ~1/EIO_TICKS of a second */ 201/* calculate time difference in ~1/EIO_TICKS of a second */
158static int tvdiff (struct timeval *tv1, struct timeval *tv2) 202static int tvdiff (struct timeval *tv1, struct timeval *tv2)
159{ 203{
160 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS 204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
161 + ((tv2->tv_usec - tv1->tv_usec) >> 10); 205 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
162} 206}
176 220
177static mutex_t wrklock = X_MUTEX_INIT; 221static mutex_t wrklock = X_MUTEX_INIT;
178static mutex_t reslock = X_MUTEX_INIT; 222static mutex_t reslock = X_MUTEX_INIT;
179static mutex_t reqlock = X_MUTEX_INIT; 223static mutex_t reqlock = X_MUTEX_INIT;
180static cond_t reqwait = X_COND_INIT; 224static cond_t reqwait = X_COND_INIT;
225
226#if !HAVE_PREADWRITE
227/*
228 * make our pread/pwrite emulation safe against themselves, but not against
229 * normal read/write by using a mutex. slows down execution a lot,
230 * but that's your problem, not mine.
231 */
232static mutex_t preadwritelock = X_MUTEX_INIT;
233#endif
181 234
182typedef struct etp_worker 235typedef struct etp_worker
183{ 236{
184 /* locked by wrklock */ 237 /* locked by wrklock */
185 struct etp_worker *prev, *next; 238 struct etp_worker *prev, *next;
330 383
331static void etp_atfork_child (void) 384static void etp_atfork_child (void)
332{ 385{
333 ETP_REQ *prv; 386 ETP_REQ *prv;
334 387
335 while (prv = reqq_shift (&req_queue)) 388 while ((prv = reqq_shift (&req_queue)))
336 ETP_DESTROY (prv); 389 ETP_DESTROY (prv);
337 390
338 while (prv = reqq_shift (&res_queue)) 391 while ((prv = reqq_shift (&res_queue)))
339 ETP_DESTROY (prv); 392 ETP_DESTROY (prv);
340 393
341 while (wrk_first.next != &wrk_first) 394 while (wrk_first.next != &wrk_first)
342 { 395 {
343 etp_worker *wrk = wrk_first.next; 396 etp_worker *wrk = wrk_first.next;
371 424
372 pthread_once (&doinit, etp_once_init); 425 pthread_once (&doinit, etp_once_init);
373 426
374 want_poll_cb = want_poll; 427 want_poll_cb = want_poll;
375 done_poll_cb = done_poll; 428 done_poll_cb = done_poll;
429
430 return 0;
376} 431}
377 432
378X_THREAD_PROC (etp_proc); 433X_THREAD_PROC (etp_proc);
379 434
380static void etp_start_thread (void) 435static void etp_start_thread (void)
400 X_UNLOCK (wrklock); 455 X_UNLOCK (wrklock);
401} 456}
402 457
403static void etp_maybe_start_thread (void) 458static void etp_maybe_start_thread (void)
404{ 459{
405 if (etp_nthreads () >= wanted) 460 if (expect_true (etp_nthreads () >= wanted))
406 return; 461 return;
407 462
408 /* todo: maybe use idle here, but might be less exact */ 463 /* todo: maybe use idle here, but might be less exact */
409 if (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()) 464 if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
410 return; 465 return;
411 466
412 etp_start_thread (); 467 etp_start_thread ();
413} 468}
414 469
467 522
468 X_LOCK (reqlock); 523 X_LOCK (reqlock);
469 --nreqs; 524 --nreqs;
470 X_UNLOCK (reqlock); 525 X_UNLOCK (reqlock);
471 526
472 if (req->type == EIO_GROUP && req->size) 527 if (expect_false (req->type == EIO_GROUP && req->size))
473 { 528 {
474 req->int1 = 1; /* mark request as delayed */ 529 req->int1 = 1; /* mark request as delayed */
475 continue; 530 continue;
476 } 531 }
477 else 532 else
478 { 533 {
479 int res = ETP_FINISH (req); 534 int res = ETP_FINISH (req);
480 if (res) 535 if (expect_false (res))
481 return res; 536 return res;
482 } 537 }
483 538
484 if (maxreqs && !--maxreqs) 539 if (expect_false (maxreqs && !--maxreqs))
485 break; 540 break;
486 541
487 if (maxtime) 542 if (maxtime)
488 { 543 {
489 gettimeofday (&tv_now, 0); 544 gettimeofday (&tv_now, 0);
508 563
509static void etp_submit (ETP_REQ *req) 564static void etp_submit (ETP_REQ *req)
510{ 565{
511 req->pri -= ETP_PRI_MIN; 566 req->pri -= ETP_PRI_MIN;
512 567
513 if (req->pri < ETP_PRI_MIN - ETP_PRI_MIN) req->pri = ETP_PRI_MIN - ETP_PRI_MIN; 568 if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
514 if (req->pri > ETP_PRI_MAX - ETP_PRI_MIN) req->pri = ETP_PRI_MAX - ETP_PRI_MIN; 569 if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
515 570
571 if (expect_false (req->type == EIO_GROUP))
572 {
573 /* I hope this is worth it :/ */
516 X_LOCK (reqlock); 574 X_LOCK (reqlock);
517 ++nreqs; 575 ++nreqs;
576 X_UNLOCK (reqlock);
577
578 X_LOCK (reslock);
579
580 ++npending;
581
582 if (!reqq_push (&res_queue, req) && want_poll_cb)
583 want_poll_cb ();
584
585 X_UNLOCK (reslock);
586 }
587 else
588 {
589 X_LOCK (reqlock);
590 ++nreqs;
518 ++nready; 591 ++nready;
519 reqq_push (&req_queue, req); 592 reqq_push (&req_queue, req);
520 X_COND_SIGNAL (reqwait); 593 X_COND_SIGNAL (reqwait);
521 X_UNLOCK (reqlock); 594 X_UNLOCK (reqlock);
522 595
523 etp_maybe_start_thread (); 596 etp_maybe_start_thread ();
597 }
524} 598}
525 599
526static void etp_set_max_poll_time (double nseconds) 600static void etp_set_max_poll_time (double nseconds)
527{ 601{
528 if (WORDACCESS_UNSAFE) X_LOCK (reslock); 602 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
529 max_poll_time = nseconds; 603 max_poll_time = nseconds * EIO_TICKS;
530 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock); 604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
531} 605}
532 606
533static void etp_set_max_poll_reqs (unsigned int maxreqs) 607static void etp_set_max_poll_reqs (unsigned int maxreqs)
534{ 608{
563 637
564static void grp_try_feed (eio_req *grp) 638static void grp_try_feed (eio_req *grp)
565{ 639{
566 while (grp->size < grp->int2 && !EIO_CANCELLED (grp)) 640 while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
567 { 641 {
568 int old_len = grp->size; 642 grp->flags &= ~EIO_FLAG_GROUPADD;
569 643
570 EIO_FEED (grp); 644 EIO_FEED (grp);
571 645
572 /* stop if no progress has been made */ 646 /* stop if no progress has been made */
573 if (old_len == grp->size) 647 if (!(grp->flags & EIO_FLAG_GROUPADD))
574 { 648 {
575 grp->feed = 0; 649 grp->feed = 0;
576 break; 650 break;
577 } 651 }
578 } 652 }
695 769
696/*****************************************************************************/ 770/*****************************************************************************/
697/* work around various missing functions */ 771/* work around various missing functions */
698 772
699#if !HAVE_PREADWRITE 773#if !HAVE_PREADWRITE
774# undef pread
775# undef pwrite
700# define pread eio__pread 776# define pread eio__pread
701# define pwrite eio__pwrite 777# define pwrite eio__pwrite
702
703/*
704 * make our pread/pwrite safe against themselves, but not against
705 * normal read/write by using a mutex. slows down execution a lot,
706 * but that's your problem, not mine.
707 */
708static mutex_t preadwritelock = X_MUTEX_INIT;
709 778
710static ssize_t 779static ssize_t
711eio__pread (int fd, void *buf, size_t count, off_t offset) 780eio__pread (int fd, void *buf, size_t count, off_t offset)
712{ 781{
713 ssize_t res; 782 ssize_t res;
731 800
732 X_LOCK (preadwritelock); 801 X_LOCK (preadwritelock);
733 ooffset = lseek (fd, 0, SEEK_CUR); 802 ooffset = lseek (fd, 0, SEEK_CUR);
734 lseek (fd, offset, SEEK_SET); 803 lseek (fd, offset, SEEK_SET);
735 res = write (fd, buf, count); 804 res = write (fd, buf, count);
736 lseek (fd, offset, SEEK_SET); 805 lseek (fd, ooffset, SEEK_SET);
737 X_UNLOCK (preadwritelock); 806 X_UNLOCK (preadwritelock);
738 807
739 return res; 808 return res;
740} 809}
741#endif 810#endif
742 811
743#ifndef HAVE_FUTIMES 812#ifndef HAVE_FUTIMES
744 813
814# undef utimes
815# undef futimes
745# define utimes(path,times) eio__utimes (path, times) 816# define utimes(path,times) eio__utimes (path, times)
746# define futimes(fd,times) eio__futimes (fd, times) 817# define futimes(fd,times) eio__futimes (fd, times)
747 818
748static int 819static int
749eio__utimes (const char *filename, const struct timeval times[2]) 820eio__utimes (const char *filename, const struct timeval times[2])
768} 839}
769 840
770#endif 841#endif
771 842
772#if !HAVE_FDATASYNC 843#if !HAVE_FDATASYNC
844# undef fdatasync
773# define fdatasync fsync 845# define fdatasync(fd) fsync (fd)
774#endif 846#endif
847
848/* sync_file_range always needs emulation */
849int
850eio__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}
775 875
776#if !HAVE_READAHEAD 876#if !HAVE_READAHEAD
877# undef readahead
777# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self) 878# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
778 879
779static ssize_t 880static ssize_t
780eio__readahead (int fd, off_t offset, size_t count, etp_worker *self) 881eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
781{ 882{
822 if (res < 0 && sbytes) 923 if (res < 0 && sbytes)
823 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */ 924 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
824 res = sbytes; 925 res = sbytes;
825 } 926 }
826 927
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
827# elif __hpux 938# elif __hpux
828 res = sendfile (ofd, ifd, offset, count, 0, 0); 939 res = sendfile (ofd, ifd, offset, count, 0, 0);
829 940
830# elif __solaris 941# elif __solaris
831 { 942 {
842 if (res < 0 && sbytes) 953 if (res < 0 && sbytes)
843 res = sbytes; 954 res = sbytes;
844 } 955 }
845 956
846# 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
847#else 968#else
848 res = -1; 969 res = -1;
849 errno = ENOSYS; 970 errno = ENOSYS;
850#endif 971#endif
851 972
852 if (res < 0 973 if (res < 0
853 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK 974 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
975 /* BSDs */
976#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
977 || errno == ENOTSUP
978#endif
979 || errno == EOPNOTSUPP /* BSDs */
854#if __solaris 980#if __solaris
855 || errno == EAFNOSUPPORT || errno == EPROTOTYPE 981 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
856#endif 982#endif
857 ) 983 )
858 ) 984 )
889 } 1015 }
890 1016
891 return res; 1017 return res;
892} 1018}
893 1019
1020static signed char
1021eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
1022{
1023 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
1024 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1025}
1026
1027#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1028
1029#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1030#define EIO_SORT_FAST 60 /* when to only use insertion sort */
1031
1032static void
1033eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1034{
1035 unsigned char bits [9 + sizeof (ino_t) * 8];
1036 unsigned char *bit = bits;
1037
1038 assert (CHAR_BIT == 8);
1039 assert (sizeof (eio_dirent) * 8 < 256);
1040 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1041 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1042
1043 if (size <= EIO_SORT_FAST)
1044 return;
1045
1046 /* first prepare an array of bits to test in our radix sort */
1047 /* try to take endianness into account, as well as differences in ino_t sizes */
1048 /* inode_bits must contain all inodes ORed together */
1049 /* which is used to skip bits that are 0 everywhere, which is very common */
1050 {
1051 ino_t endianness;
1052 int i, j;
1053
1054 /* we store the byte offset of byte n into byte n of "endianness" */
1055 for (i = 0; i < sizeof (ino_t); ++i)
1056 ((unsigned char *)&endianness)[i] = i;
1057
1058 *bit++ = 0;
1059
1060 for (i = 0; i < sizeof (ino_t); ++i)
1061 {
1062 /* shifting off the byte offsets out of "endianness" */
1063 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1064 endianness >>= 8;
1065
1066 for (j = 0; j < 8; ++j)
1067 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1068 *bit++ = offs + j;
1069 }
1070
1071 for (j = 0; j < 8; ++j)
1072 if (score_bits & (1 << j))
1073 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1074 }
1075
1076 /* now actually do the sorting (a variant of MSD radix sort) */
1077 {
1078 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1079 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1080 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1081 int stk_idx = 0;
1082
1083 base_stk [stk_idx] = dents;
1084 end_stk [stk_idx] = dents + size;
1085 bit_stk [stk_idx] = bit - 1;
1086
1087 do
1088 {
1089 base = base_stk [stk_idx];
1090 end = end_stk [stk_idx];
1091 bit = bit_stk [stk_idx];
1092
1093 for (;;)
1094 {
1095 unsigned char O = *bit >> 3;
1096 unsigned char M = 1 << (*bit & 7);
1097
1098 eio_dirent *a = base;
1099 eio_dirent *b = end;
1100
1101 if (b - a < EIO_SORT_CUTOFF)
1102 break;
1103
1104 /* now bit-partition the array on the bit */
1105 /* this ugly asymmetric loop seems to perform much better than typical */
1106 /* partition algos found in the literature */
1107 do
1108 if (!(((unsigned char *)a)[O] & M))
1109 ++a;
1110 else if (!(((unsigned char *)--b)[O] & M))
1111 {
1112 eio_dirent tmp = *a; *a = *b; *b = tmp;
1113 ++a;
1114 }
1115 while (b > a);
1116
1117 /* next bit, or stop, if no bits left in this path */
1118 if (!*--bit)
1119 break;
1120
1121 base_stk [stk_idx] = a;
1122 end_stk [stk_idx] = end;
1123 bit_stk [stk_idx] = bit;
1124 ++stk_idx;
1125
1126 end = a;
1127 }
1128 }
1129 while (stk_idx--);
1130 }
1131}
1132
1133static void
1134eio_dent_insertion_sort (eio_dirent *dents, int size)
1135{
1136 /* first move the smallest element to the front, to act as a sentinel */
1137 {
1138 int i;
1139 eio_dirent *min = dents;
1140
1141 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1142 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1143 if (EIO_DENT_CMP (dents [i], <, *min))
1144 min = &dents [i];
1145
1146 /* swap elements 0 and j (minimum) */
1147 {
1148 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1149 }
1150 }
1151
1152 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1153 {
1154 eio_dirent *i, *j;
1155
1156 for (i = dents + 1; i < dents + size; ++i)
1157 {
1158 eio_dirent value = *i;
1159
1160 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1161 j [1] = j [0];
1162
1163 j [1] = value;
1164 }
1165 }
1166}
1167
1168static void
1169eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1170{
1171 if (size <= 1)
1172 return; /* our insertion sort relies on size > 0 */
1173
1174 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1175 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1176 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1177
1178 /* use an insertion sort at the end, or for small arrays, */
1179 /* as insertion sort is more efficient for small partitions */
1180 eio_dent_insertion_sort (dents, size);
1181}
1182
894/* read a full directory */ 1183/* read a full directory */
895static void 1184static void
896eio__scandir (eio_req *req, etp_worker *self) 1185eio__scandir (eio_req *req, etp_worker *self)
897{ 1186{
898 DIR *dirp; 1187 DIR *dirp;
899 EIO_STRUCT_DIRENT *entp; 1188 EIO_STRUCT_DIRENT *entp;
900 char *name, *names; 1189 char *name, *names;
901 int memlen = 4096; 1190 int namesalloc = 4096;
902 int memofs = 0; 1191 int namesoffs = 0;
1192 int flags = req->int1;
1193 eio_dirent *dents = 0;
1194 int dentalloc = 128;
903 int res = 0; 1195 int dentoffs = 0;
1196 ino_t inode_bits = 0;
1197
1198 req->result = -1;
1199
1200 if (!(flags & EIO_READDIR_DENTS))
1201 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
904 1202
905 X_LOCK (wrklock); 1203 X_LOCK (wrklock);
1204 /* the corresponding closedir is in ETP_WORKER_CLEAR */
906 self->dirp = dirp = opendir (req->ptr1); 1205 self->dirp = dirp = opendir (req->ptr1);
907 req->flags |= EIO_FLAG_PTR2_FREE; 1206 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1207 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
908 req->ptr2 = names = malloc (memlen); 1208 req->ptr2 = names = malloc (namesalloc);
909 X_UNLOCK (wrklock); 1209 X_UNLOCK (wrklock);
910 1210
911 if (dirp && names) 1211 if (dirp && names && (!flags || dents))
912 for (;;) 1212 for (;;)
913 { 1213 {
914 errno = 0; 1214 errno = 0;
915 entp = readdir (dirp); 1215 entp = readdir (dirp);
916 1216
917 if (!entp) 1217 if (!entp)
1218 {
1219 if (errno)
1220 break;
1221
1222 /* sort etc. */
1223 req->int1 = flags;
1224 req->result = dentoffs;
1225
1226 if (flags & EIO_READDIR_STAT_ORDER)
1227 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1228 else if (flags & EIO_READDIR_DIRS_FIRST)
1229 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1230 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1231 else
1232 {
1233 /* in this case, all is known, and we just put dirs first and sort them */
1234 eio_dirent *oth = dents + dentoffs;
1235 eio_dirent *dir = dents;
1236
1237 /* now partition dirs to the front, and non-dirs to the back */
1238 /* by walking from both sides and swapping if necessary */
1239 /* also clear score, so it doesn't influence sorting */
1240 while (oth > dir)
1241 {
1242 if (dir->type == EIO_DT_DIR)
1243 ++dir;
1244 else if ((--oth)->type == EIO_DT_DIR)
1245 {
1246 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1247
1248 ++dir;
1249 }
1250 }
1251
1252 /* now sort the dirs only */
1253 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1254 }
1255
918 break; 1256 break;
1257 }
919 1258
1259 /* now add the entry to our list(s) */
920 name = entp->d_name; 1260 name = entp->d_name;
921 1261
1262 /* skip . and .. entries */
922 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2]))) 1263 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
923 { 1264 {
924 int len = strlen (name) + 1; 1265 int len = D_NAMLEN (entp) + 1;
925 1266
926 res++; 1267 while (expect_false (namesoffs + len > namesalloc))
927
928 while (memofs + len > memlen)
929 { 1268 {
930 memlen *= 2; 1269 namesalloc *= 2;
931 X_LOCK (wrklock); 1270 X_LOCK (wrklock);
932 req->ptr2 = names = realloc (names, memlen); 1271 req->ptr2 = names = realloc (names, namesalloc);
933 X_UNLOCK (wrklock); 1272 X_UNLOCK (wrklock);
934 1273
935 if (!names) 1274 if (!names)
936 break; 1275 break;
937 } 1276 }
938 1277
939 memcpy (names + memofs, name, len); 1278 memcpy (names + namesoffs, name, len);
1279
1280 if (dents)
1281 {
1282 struct eio_dirent *ent;
1283
1284 if (expect_false (dentoffs == dentalloc))
1285 {
1286 dentalloc *= 2;
1287 X_LOCK (wrklock);
1288 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1289 X_UNLOCK (wrklock);
1290
1291 if (!dents)
1292 break;
1293 }
1294
1295 ent = dents + dentoffs;
1296
1297 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1298 ent->namelen = len - 1;
1299 ent->inode = D_INO (entp);
1300
1301 inode_bits |= ent->inode;
1302
1303 switch (D_TYPE (entp))
1304 {
1305 default:
1306 ent->type = EIO_DT_UNKNOWN;
1307 flags |= EIO_READDIR_FOUND_UNKNOWN;
1308 break;
1309
1310 #ifdef DT_FIFO
1311 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1312 #endif
1313 #ifdef DT_CHR
1314 case DT_CHR: ent->type = EIO_DT_CHR; break;
1315 #endif
1316 #ifdef DT_MPC
1317 case DT_MPC: ent->type = EIO_DT_MPC; break;
1318 #endif
1319 #ifdef DT_DIR
1320 case DT_DIR: ent->type = EIO_DT_DIR; break;
1321 #endif
1322 #ifdef DT_NAM
1323 case DT_NAM: ent->type = EIO_DT_NAM; break;
1324 #endif
1325 #ifdef DT_BLK
1326 case DT_BLK: ent->type = EIO_DT_BLK; break;
1327 #endif
1328 #ifdef DT_MPB
1329 case DT_MPB: ent->type = EIO_DT_MPB; break;
1330 #endif
1331 #ifdef DT_REG
1332 case DT_REG: ent->type = EIO_DT_REG; break;
1333 #endif
1334 #ifdef DT_NWK
1335 case DT_NWK: ent->type = EIO_DT_NWK; break;
1336 #endif
1337 #ifdef DT_CMP
1338 case DT_CMP: ent->type = EIO_DT_CMP; break;
1339 #endif
1340 #ifdef DT_LNK
1341 case DT_LNK: ent->type = EIO_DT_LNK; break;
1342 #endif
1343 #ifdef DT_SOCK
1344 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1345 #endif
1346 #ifdef DT_DOOR
1347 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1348 #endif
1349 #ifdef DT_WHT
1350 case DT_WHT: ent->type = EIO_DT_WHT; break;
1351 #endif
1352 }
1353
1354 ent->score = 7;
1355
1356 if (flags & EIO_READDIR_DIRS_FIRST)
1357 {
1358 if (ent->type == EIO_DT_UNKNOWN)
1359 {
1360 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1361 ent->score = 1;
1362 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1363 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1364 }
1365 else if (ent->type == EIO_DT_DIR)
1366 ent->score = 0;
1367 }
1368 }
1369
940 memofs += len; 1370 namesoffs += len;
1371 ++dentoffs;
1372 }
1373
1374 if (EIO_CANCELLED (req))
1375 {
1376 errno = ECANCELED;
1377 break;
941 } 1378 }
942 } 1379 }
1380}
943 1381
944 if (errno) 1382#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
945 res = -1; 1383# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
946 1384#else
947 req->result = res; 1385
1386int
1387eio__msync (void *mem, size_t len, int flags)
1388{
1389 if (EIO_MS_ASYNC != MS_SYNC
1390 || EIO_MS_INVALIDATE != MS_INVALIDATE
1391 || EIO_MS_SYNC != MS_SYNC)
1392 {
1393 flags = 0
1394 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1395 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1396 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1397 }
1398
1399 return msync (mem, len, flags);
1400}
1401
1402#endif
1403
1404int
1405eio__mtouch (void *mem, size_t len, int flags)
1406{
1407 intptr_t addr = (intptr_t)mem;
1408 intptr_t end = addr + len;
1409#ifdef PAGESIZE
1410 const intptr_t page = PAGESIZE;
1411#else
1412 static intptr_t page;
1413
1414 if (!page)
1415 page = sysconf (_SC_PAGESIZE);
1416#endif
1417
1418 /* round down to start of page, although this is probably useless */
1419 addr &= ~(page - 1); /* assume page size is always a power of two */
1420
1421 if (addr < end)
1422 if (flags & EIO_MT_MODIFY) /* modify */
1423 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1424 else
1425 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1426
1427 return 0;
948} 1428}
949 1429
950/*****************************************************************************/ 1430/*****************************************************************************/
951 1431
952#define ALLOC(len) \ 1432#define ALLOC(len) \
1039 1519
1040/*****************************************************************************/ 1520/*****************************************************************************/
1041 1521
1042int eio_init (void (*want_poll)(void), void (*done_poll)(void)) 1522int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1043{ 1523{
1044 etp_init (want_poll, done_poll); 1524 return etp_init (want_poll, done_poll);
1045} 1525}
1046 1526
1047static void eio_api_destroy (eio_req *req) 1527static void eio_api_destroy (eio_req *req)
1048{ 1528{
1049 free (req); 1529 free (req);
1111 case EIO_RMDIR: req->result = rmdir (req->ptr1); break; 1591 case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
1112 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break; 1592 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
1113 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break; 1593 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1114 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break; 1594 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1115 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break; 1595 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1116 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break; 1596 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1117 1597
1118 case EIO_READLINK: ALLOC (NAME_MAX); 1598 case EIO_READLINK: ALLOC (NAME_MAX);
1119 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break; 1599 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1120 1600
1121 case EIO_SYNC: req->result = 0; sync (); break; 1601 case EIO_SYNC: req->result = 0; sync (); break;
1122 case EIO_FSYNC: req->result = fsync (req->int1); break; 1602 case EIO_FSYNC: req->result = fsync (req->int1); break;
1123 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break; 1603 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1604 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1605 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1606 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1124 1607
1125 case EIO_READDIR: eio__scandir (req, self); break; 1608 case EIO_READDIR: eio__scandir (req, self); break;
1126 1609
1127 case EIO_BUSY: 1610 case EIO_BUSY:
1128#ifdef _WIN32 1611#ifdef _WIN32
1160 1643
1161 req->result = req->type == EIO_FUTIME 1644 req->result = req->type == EIO_FUTIME
1162 ? futimes (req->int1, times) 1645 ? futimes (req->int1, times)
1163 : utimes (req->ptr1, times); 1646 : utimes (req->ptr1, times);
1164 } 1647 }
1648 break;
1165 1649
1166 case EIO_GROUP: 1650 case EIO_GROUP:
1651 abort (); /* handled in eio_request */
1652
1167 case EIO_NOP: 1653 case EIO_NOP:
1168 req->result = 0; 1654 req->result = 0;
1169 break; 1655 break;
1170 1656
1171 case EIO_CUSTOM: 1657 case EIO_CUSTOM:
1172 req->feed (req); 1658 ((void (*)(eio_req *))req->feed) (req);
1173 break; 1659 break;
1174 1660
1175 default: 1661 default:
1176 req->result = -1; 1662 req->result = -1;
1177 break; 1663 break;
1200eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data) 1686eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1201{ 1687{
1202 REQ (EIO_FSYNC); req->int1 = fd; SEND; 1688 REQ (EIO_FSYNC); req->int1 = fd; SEND;
1203} 1689}
1204 1690
1691eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1692{
1693 REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1694}
1695
1696eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1697{
1698 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1699}
1700
1701eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1702{
1703 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1704}
1705
1205eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data) 1706eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1206{ 1707{
1207 REQ (EIO_FDATASYNC); req->int1 = fd; SEND; 1708 REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1208} 1709}
1209 1710
1321eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data) 1822eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1322{ 1823{
1323 return eio__1path (EIO_RMDIR, path, pri, cb, data); 1824 return eio__1path (EIO_RMDIR, path, pri, cb, data);
1324} 1825}
1325 1826
1326eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data) 1827eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1327{ 1828{
1328 return eio__1path (EIO_READDIR, path, pri, cb, data); 1829 REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1329} 1830}
1330 1831
1331eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data) 1832eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1332{ 1833{
1333 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND; 1834 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1334} 1835}
1335 1836
1336static eio_req * 1837static eio_req *
1337eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data) 1838eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1338{ 1839{
1364 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data); 1865 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1365} 1866}
1366 1867
1367eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data) 1868eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
1368{ 1869{
1369 REQ (EIO_CUSTOM); req->feed = execute; SEND; 1870 REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
1370} 1871}
1371 1872
1372#endif 1873#endif
1373 1874
1374eio_req *eio_grp (eio_cb cb, void *data) 1875eio_req *eio_grp (eio_cb cb, void *data)
1402 1903
1403void eio_grp_add (eio_req *grp, eio_req *req) 1904void eio_grp_add (eio_req *grp, eio_req *req)
1404{ 1905{
1405 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2)); 1906 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1406 1907
1908 grp->flags |= EIO_FLAG_GROUPADD;
1909
1407 ++grp->size; 1910 ++grp->size;
1408 req->grp = grp; 1911 req->grp = grp;
1409 1912
1410 req->grp_prev = 0; 1913 req->grp_prev = 0;
1411 req->grp_next = grp->grp_first; 1914 req->grp_next = grp->grp_first;
1420/* misc garbage */ 1923/* misc garbage */
1421 1924
1422ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count) 1925ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1423{ 1926{
1424 etp_worker wrk; 1927 etp_worker wrk;
1928 ssize_t ret;
1425 1929
1426 wrk.dbuf = 0; 1930 wrk.dbuf = 0;
1427 1931
1428 eio__sendfile (ofd, ifd, offset, count, &wrk); 1932 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1429 1933
1430 if (wrk.dbuf) 1934 if (wrk.dbuf)
1431 free (wrk.dbuf); 1935 free (wrk.dbuf);
1432}
1433 1936
1937 return ret;
1938}
1939

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines