[ViewVC] Diff of: cvs/libeio/eio.c

Comparing libeio/eio.c (file contents):
Revision 1.21 by root, Sat Jun 21 00:13:13 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 */
158	static int tvdiff (struct timeval tv1, struct timeval tv2)	202	static 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
177	static mutex_t wrklock = X_MUTEX_INIT;	221	static mutex_t wrklock = X_MUTEX_INIT;
178	static mutex_t reslock = X_MUTEX_INIT;	222	static mutex_t reslock = X_MUTEX_INIT;
179	static mutex_t reqlock = X_MUTEX_INIT;	223	static mutex_t reqlock = X_MUTEX_INIT;
180	static cond_t reqwait = X_COND_INIT;	224	static 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	*/
		232	static mutex_t preadwritelock = X_MUTEX_INIT;
		233	#endif
181		234
182	typedef struct etp_worker	235	typedef 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;
…		…
402	X_UNLOCK (wrklock);	455	X_UNLOCK (wrklock);
403	}	456	}
404		457
405	static void etp_maybe_start_thread (void)	458	static void etp_maybe_start_thread (void)
406	{	459	{
407	if (etp_nthreads () >= wanted)	460	if (expect_true (etp_nthreads () >= wanted))
408	return;	461	return;
409		462
410	/* todo: maybe use idle here, but might be less exact */	463	/* todo: maybe use idle here, but might be less exact */
411	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 ()))
412	return;	465	return;
413		466
414	etp_start_thread ();	467	etp_start_thread ();
415	}	468	}
416		469
…		…
469		522
470	X_LOCK (reqlock);	523	X_LOCK (reqlock);
471	--nreqs;	524	--nreqs;
472	X_UNLOCK (reqlock);	525	X_UNLOCK (reqlock);
473		526
474	if (req->type == EIO_GROUP && req->size)	527	if (expect_false (req->type == EIO_GROUP && req->size))
475	{	528	{
476	req->int1 = 1; /* mark request as delayed */	529	req->int1 = 1; /* mark request as delayed */
477	continue;	530	continue;
478	}	531	}
479	else	532	else
480	{	533	{
481	int res = ETP_FINISH (req);	534	int res = ETP_FINISH (req);
482	if (res)	535	if (expect_false (res))
483	return res;	536	return res;
484	}	537	}
485		538
486	if (maxreqs && !--maxreqs)	539	if (expect_false (maxreqs && !--maxreqs))
487	break;	540	break;
488		541
489	if (maxtime)	542	if (maxtime)
490	{	543	{
491	gettimeofday (&tv_now, 0);	544	gettimeofday (&tv_now, 0);
…		…
510		563
511	static void etp_submit (ETP_REQ *req)	564	static void etp_submit (ETP_REQ *req)
512	{	565	{
513	req->pri -= ETP_PRI_MIN;	566	req->pri -= ETP_PRI_MIN;
514		567
515	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;
516	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;
517		570
		571	if (expect_false (req->type == EIO_GROUP))
		572	{
		573	/* I hope this is worth it :/ */
518	X_LOCK (reqlock);	574	X_LOCK (reqlock);
519	++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;
520	++nready;	591	++nready;
521	reqq_push (&req_queue, req);	592	reqq_push (&req_queue, req);
522	X_COND_SIGNAL (reqwait);	593	X_COND_SIGNAL (reqwait);
523	X_UNLOCK (reqlock);	594	X_UNLOCK (reqlock);
524		595
525	etp_maybe_start_thread ();	596	etp_maybe_start_thread ();
		597	}
526	}	598	}
527		599
528	static void etp_set_max_poll_time (double nseconds)	600	static void etp_set_max_poll_time (double nseconds)
529	{	601	{
530	if (WORDACCESS_UNSAFE) X_LOCK (reslock);	602	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
531	max_poll_time = nseconds;	603	max_poll_time = nseconds * EIO_TICKS;
532	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);	604	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
533	}	605	}
534		606
535	static void etp_set_max_poll_reqs (unsigned int maxreqs)	607	static void etp_set_max_poll_reqs (unsigned int maxreqs)
536	{	608	{
…		…
565		637
566	static void grp_try_feed (eio_req *grp)	638	static void grp_try_feed (eio_req *grp)
567	{	639	{
568	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	640	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
569	{	641	{
570	int old_len = grp->size;	642	grp->flags &= ~EIO_FLAG_GROUPADD;
571		643
572	EIO_FEED (grp);	644	EIO_FEED (grp);
573		645
574	/* stop if no progress has been made */	646	/* stop if no progress has been made */
575	if (old_len == grp->size)	647	if (!(grp->flags & EIO_FLAG_GROUPADD))
576	{	648	{
577	grp->feed = 0;	649	grp->feed = 0;
578	break;	650	break;
579	}	651	}
580	}	652	}
…		…
697		769
698	/*****************************************************************************/	770	/*****************************************************************************/
699	/* work around various missing functions */	771	/* work around various missing functions */
700		772
701	#if !HAVE_PREADWRITE	773	#if !HAVE_PREADWRITE
		774	# undef pread
		775	# undef pwrite
702	# define pread eio__pread	776	# define pread eio__pread
703	# define pwrite eio__pwrite	777	# define pwrite eio__pwrite
704
705	/*
706	* make our pread/pwrite safe against themselves, but not against
707	* normal read/write by using a mutex. slows down execution a lot,
708	* but that's your problem, not mine.
709	*/
710	static mutex_t preadwritelock = X_MUTEX_INIT;
711		778
712	static ssize_t	779	static ssize_t
713	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)
714	{	781	{
715	ssize_t res;	782	ssize_t res;
…		…
733		800
734	X_LOCK (preadwritelock);	801	X_LOCK (preadwritelock);
735	ooffset = lseek (fd, 0, SEEK_CUR);	802	ooffset = lseek (fd, 0, SEEK_CUR);
736	lseek (fd, offset, SEEK_SET);	803	lseek (fd, offset, SEEK_SET);
737	res = write (fd, buf, count);	804	res = write (fd, buf, count);
738	lseek (fd, offset, SEEK_SET);	805	lseek (fd, ooffset, SEEK_SET);
739	X_UNLOCK (preadwritelock);	806	X_UNLOCK (preadwritelock);
740		807
741	return res;	808	return res;
742	}	809	}
743	#endif	810	#endif
744		811
745	#ifndef HAVE_FUTIMES	812	#ifndef HAVE_FUTIMES
746		813
		814	# undef utimes
		815	# undef futimes
747	# define utimes(path,times) eio__utimes (path, times)	816	# define utimes(path,times) eio__utimes (path, times)
748	# define futimes(fd,times) eio__futimes (fd, times)	817	# define futimes(fd,times) eio__futimes (fd, times)
749		818
750	static int	819	static int
751	eio__utimes (const char *filename, const struct timeval times[2])	820	eio__utimes (const char *filename, const struct timeval times[2])
…		…
770	}	839	}
771		840
772	#endif	841	#endif
773		842
774	#if !HAVE_FDATASYNC	843	#if !HAVE_FDATASYNC
		844	# undef fdatasync
775	# define fdatasync fsync	845	# define fdatasync(fd) fsync (fd)
776	#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	}
777		875
778	#if !HAVE_READAHEAD	876	#if !HAVE_READAHEAD
		877	# undef readahead
779	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	878	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
780		879
781	static ssize_t	880	static ssize_t
782	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)
783	{	882	{
…		…
824	if (res < 0 && sbytes)	923	if (res < 0 && sbytes)
825	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	924	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
826	res = sbytes;	925	res = sbytes;
827	}	926	}
828		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
829	# elif __hpux	938	# elif __hpux
830	res = sendfile (ofd, ifd, offset, count, 0, 0);	939	res = sendfile (ofd, ifd, offset, count, 0, 0);
831		940
832	# elif __solaris	941	# elif __solaris
833	{	942	{
…		…
844	if (res < 0 && sbytes)	953	if (res < 0 && sbytes)
845	res = sbytes;	954	res = sbytes;
846	}	955	}
847		956
848	# 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
849	#else	968	#else
850	res = -1;	969	res = -1;
851	errno = ENOSYS;	970	errno = ENOSYS;
852	#endif	971	#endif
853		972
854	if (res < 0	973	if (res < 0
855	&& (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 */
856	#if __solaris	980	#if __solaris
857	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE	981	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE
858	#endif	982	#endif
859	)	983	)
860	)	984	)
…		…
891	}	1015	}
892		1016
893	return res;	1017	return res;
894	}	1018	}
895		1019
		1020	static signed char
		1021	eio_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
		1032	static void
		1033	eio_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
		1133	static void
		1134	eio_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
		1168	static void
		1169	eio_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
896	/* read a full directory */	1183	/* read a full directory */
897	static void	1184	static void
898	eio__scandir (eio_req req, etp_worker self)	1185	eio__scandir (eio_req req, etp_worker self)
899	{	1186	{
900	DIR *dirp;	1187	DIR *dirp;
901	EIO_STRUCT_DIRENT *entp;	1188	EIO_STRUCT_DIRENT *entp;
902	char name, names;	1189	char name, names;
903	int memlen = 4096;	1190	int namesalloc = 4096;
904	int memofs = 0;	1191	int namesoffs = 0;
		1192	int flags = req->int1;
		1193	eio_dirent *dents = 0;
		1194	int dentalloc = 128;
905	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);
906		1202
907	X_LOCK (wrklock);	1203	X_LOCK (wrklock);
908	/* the corresponding closedir is in ETP_WORKER_CLEAR */	1204	/* the corresponding closedir is in ETP_WORKER_CLEAR */
909	self->dirp = dirp = opendir (req->ptr1);	1205	self->dirp = dirp = opendir (req->ptr1);
910	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;
911	req->ptr2 = names = malloc (memlen);	1208	req->ptr2 = names = malloc (namesalloc);
912	X_UNLOCK (wrklock);	1209	X_UNLOCK (wrklock);
913		1210
914	if (dirp && names)	1211	if (dirp && names && (!flags \|\| dents))
915	for (;;)	1212	for (;;)
916	{	1213	{
917	errno = 0;	1214	errno = 0;
918	entp = readdir (dirp);	1215	entp = readdir (dirp);
919		1216
920	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
921	break;	1256	break;
		1257	}
922		1258
		1259	/* now add the entry to our list(s) */
923	name = entp->d_name;	1260	name = entp->d_name;
924		1261
		1262	/* skip . and .. entries */
925	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))	1263	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))
926	{	1264	{
927	int len = strlen (name) + 1;	1265	int len = D_NAMLEN (entp) + 1;
928		1266
929	res++;	1267	while (expect_false (namesoffs + len > namesalloc))
930
931	while (memofs + len > memlen)
932	{	1268	{
933	memlen *= 2;	1269	namesalloc *= 2;
934	X_LOCK (wrklock);	1270	X_LOCK (wrklock);
935	req->ptr2 = names = realloc (names, memlen);	1271	req->ptr2 = names = realloc (names, namesalloc);
936	X_UNLOCK (wrklock);	1272	X_UNLOCK (wrklock);
937		1273
938	if (!names)	1274	if (!names)
939	break;	1275	break;
940	}	1276	}
941		1277
942	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
943	memofs += len;	1370	namesoffs += len;
		1371	++dentoffs;
		1372	}
		1373
		1374	if (EIO_CANCELLED (req))
		1375	{
		1376	errno = ECANCELED;
		1377	break;
944	}	1378	}
945	}	1379	}
		1380	}
946		1381
947	if (errno)	1382	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
948	res = -1;	1383	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
949		1384	#else
950	req->result = res;	1385
		1386	int
		1387	eio__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
		1404	int
		1405	eio__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;
951	}	1428	}
952		1429
953	/*****************************************************************************/	1430	/*****************************************************************************/
954		1431
955	#define ALLOC(len) \	1432	#define ALLOC(len) \
…		…
1122	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;	1599	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1123		1600
1124	case EIO_SYNC: req->result = 0; sync (); break;	1601	case EIO_SYNC: req->result = 0; sync (); break;
1125	case EIO_FSYNC: req->result = fsync (req->int1); break;	1602	case EIO_FSYNC: req->result = fsync (req->int1); break;
1126	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;
1127		1607
1128	case EIO_READDIR: eio__scandir (req, self); break;	1608	case EIO_READDIR: eio__scandir (req, self); break;
1129		1609
1130	case EIO_BUSY:	1610	case EIO_BUSY:
1131	#ifdef _WIN32	1611	#ifdef _WIN32
…		…
1163		1643
1164	req->result = req->type == EIO_FUTIME	1644	req->result = req->type == EIO_FUTIME
1165	? futimes (req->int1, times)	1645	? futimes (req->int1, times)
1166	: utimes (req->ptr1, times);	1646	: utimes (req->ptr1, times);
1167	}	1647	}
		1648	break;
1168		1649
1169	case EIO_GROUP:	1650	case EIO_GROUP:
		1651	abort (); /* handled in eio_request */
		1652
1170	case EIO_NOP:	1653	case EIO_NOP:
1171	req->result = 0;	1654	req->result = 0;
1172	break;	1655	break;
1173		1656
1174	case EIO_CUSTOM:	1657	case EIO_CUSTOM:
1175	req->feed (req);	1658	((void ()(eio_req ))req->feed) (req);
1176	break;	1659	break;
1177		1660
1178	default:	1661	default:
1179	req->result = -1;	1662	req->result = -1;
1180	break;	1663	break;
…		…
1203	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)	1686	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)
1204	{	1687	{
1205	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1688	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1206	}	1689	}
1207		1690
		1691	eio_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
		1696	eio_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
		1701	eio_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
1208	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)	1706	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)
1209	{	1707	{
1210	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1708	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1211	}	1709	}
1212		1710
…		…
1324	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)	1822	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)
1325	{	1823	{
1326	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1824	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1327	}	1825	}
1328		1826
1329	eio_req eio_readdir (const char path, int pri, eio_cb cb, void *data)	1827	eio_req eio_readdir (const char path, int flags, int pri, eio_cb cb, void *data)
1330	{	1828	{
1331	return eio__1path (EIO_READDIR, path, pri, cb, data);	1829	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1332	}	1830	}
1333		1831
1334	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1832	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1335	{	1833	{
1336	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;	1834	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
…		…
1367	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1865	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1368	}	1866	}
1369		1867
1370	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)	1868	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)
1371	{	1869	{
1372	REQ (EIO_CUSTOM); req->feed = execute; SEND;	1870	REQ (EIO_CUSTOM); req->feed = (void ()(eio_req ))execute; SEND;
1373	}	1871	}
1374		1872
1375	#endif	1873	#endif
1376		1874
1377	eio_req eio_grp (eio_cb cb, void data)	1875	eio_req eio_grp (eio_cb cb, void data)
…		…
1405		1903
1406	void eio_grp_add (eio_req grp, eio_req req)	1904	void eio_grp_add (eio_req grp, eio_req req)
1407	{	1905	{
1408	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));
1409		1907
		1908	grp->flags \|= EIO_FLAG_GROUPADD;
		1909
1410	++grp->size;	1910	++grp->size;
1411	req->grp = grp;	1911	req->grp = grp;
1412		1912
1413	req->grp_prev = 0;	1913	req->grp_prev = 0;
1414	req->grp_next = grp->grp_first;	1914	req->grp_next = grp->grp_first;
…		…
1423	/* misc garbage */	1923	/* misc garbage */
1424		1924
1425	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	1925	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1426	{	1926	{
1427	etp_worker wrk;	1927	etp_worker wrk;
		1928	ssize_t ret;
1428		1929
1429	wrk.dbuf = 0;	1930	wrk.dbuf = 0;
1430		1931
1431	eio__sendfile (ofd, ifd, offset, count, &wrk);	1932	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1432		1933
1433	if (wrk.dbuf)	1934	if (wrk.dbuf)
1434	free (wrk.dbuf);	1935	free (wrk.dbuf);
1435	}
1436		1936
		1937	return ret;
		1938	}
		1939

Diff Legend

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

Comparing libeio/eio.c (file contents): Revision 1.21 by root, Sat Jun 21 00:13:13 2008 UTC vs. Revision 1.49 by root, Sat Jan 2 14:24:32 2010 UTC

Diff Legend

Comparing libeio/eio.c (file contents):
Revision 1.21 by root, Sat Jun 21 00:13:13 2008 UTC vs.
Revision 1.49 by root, Sat Jan 2 14:24:32 2010 UTC