[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 1.45 by root, Sat Jan 2 12:50:22 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 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>
		118	# elif defined _WIN32
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 */
158	static int tvdiff (struct timeval tv1, struct timeval tv2)	203	static 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	}
…		…
176		221
177	static mutex_t wrklock = X_MUTEX_INIT;	222	static mutex_t wrklock = X_MUTEX_INIT;
178	static mutex_t reslock = X_MUTEX_INIT;	223	static mutex_t reslock = X_MUTEX_INIT;
179	static mutex_t reqlock = X_MUTEX_INIT;	224	static mutex_t reqlock = X_MUTEX_INIT;
180	static cond_t reqwait = X_COND_INIT;	225	static cond_t reqwait = X_COND_INIT;
		226
		227	#if !HAVE_PREADWRITE
		228	/*
		229	* make our pread/pwrite emulation safe against themselves, but not against
		230	* normal read/write by using a mutex. slows down execution a lot,
		231	* but that's your problem, not mine.
		232	*/
		233	static mutex_t preadwritelock = X_MUTEX_INIT;
		234	#endif
181		235
182	typedef struct etp_worker	236	typedef struct etp_worker
183	{	237	{
184	/* locked by wrklock */	238	/* locked by wrklock */
185	struct etp_worker prev, next;	239	struct etp_worker prev, next;
…		…
402	X_UNLOCK (wrklock);	456	X_UNLOCK (wrklock);
403	}	457	}
404		458
405	static void etp_maybe_start_thread (void)	459	static void etp_maybe_start_thread (void)
406	{	460	{
407	if (etp_nthreads () >= wanted)	461	if (expect_true (etp_nthreads () >= wanted))
408	return;	462	return;
409		463
410	/* todo: maybe use idle here, but might be less exact */	464	/* todo: maybe use idle here, but might be less exact */
411	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 ()))
412	return;	466	return;
413		467
414	etp_start_thread ();	468	etp_start_thread ();
415	}	469	}
416		470
…		…
469		523
470	X_LOCK (reqlock);	524	X_LOCK (reqlock);
471	--nreqs;	525	--nreqs;
472	X_UNLOCK (reqlock);	526	X_UNLOCK (reqlock);
473		527
474	if (req->type == EIO_GROUP && req->size)	528	if (expect_false (req->type == EIO_GROUP && req->size))
475	{	529	{
476	req->int1 = 1; /* mark request as delayed */	530	req->int1 = 1; /* mark request as delayed */
477	continue;	531	continue;
478	}	532	}
479	else	533	else
480	{	534	{
481	int res = ETP_FINISH (req);	535	int res = ETP_FINISH (req);
482	if (res)	536	if (expect_false (res))
483	return res;	537	return res;
484	}	538	}
485		539
486	if (maxreqs && !--maxreqs)	540	if (expect_false (maxreqs && !--maxreqs))
487	break;	541	break;
488		542
489	if (maxtime)	543	if (maxtime)
490	{	544	{
491	gettimeofday (&tv_now, 0);	545	gettimeofday (&tv_now, 0);
…		…
510		564
511	static void etp_submit (ETP_REQ *req)	565	static void etp_submit (ETP_REQ *req)
512	{	566	{
513	req->pri -= ETP_PRI_MIN;	567	req->pri -= ETP_PRI_MIN;
514		568
515	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;
516	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;
517		571
		572	if (expect_false (req->type == EIO_GROUP))
		573	{
		574	/* I hope this is worth it :/ */
518	X_LOCK (reqlock);	575	X_LOCK (reqlock);
519	++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;
520	++nready;	592	++nready;
521	reqq_push (&req_queue, req);	593	reqq_push (&req_queue, req);
522	X_COND_SIGNAL (reqwait);	594	X_COND_SIGNAL (reqwait);
523	X_UNLOCK (reqlock);	595	X_UNLOCK (reqlock);
524		596
525	etp_maybe_start_thread ();	597	etp_maybe_start_thread ();
		598	}
526	}	599	}
527		600
528	static void etp_set_max_poll_time (double nseconds)	601	static void etp_set_max_poll_time (double nseconds)
529	{	602	{
530	if (WORDACCESS_UNSAFE) X_LOCK (reslock);	603	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
531	max_poll_time = nseconds;	604	max_poll_time = nseconds * EIO_TICKS;
532	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);	605	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
533	}	606	}
534		607
535	static void etp_set_max_poll_reqs (unsigned int maxreqs)	608	static void etp_set_max_poll_reqs (unsigned int maxreqs)
536	{	609	{
…		…
565		638
566	static void grp_try_feed (eio_req *grp)	639	static void grp_try_feed (eio_req *grp)
567	{	640	{
568	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	641	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
569	{	642	{
570	int old_len = grp->size;	643	grp->flags &= ~EIO_FLAG_GROUPADD;
571		644
572	EIO_FEED (grp);	645	EIO_FEED (grp);
573		646
574	/* stop if no progress has been made */	647	/* stop if no progress has been made */
575	if (old_len == grp->size)	648	if (!(grp->flags & EIO_FLAG_GROUPADD))
576	{	649	{
577	grp->feed = 0;	650	grp->feed = 0;
578	break;	651	break;
579	}	652	}
580	}	653	}
…		…
697		770
698	/*****************************************************************************/	771	/*****************************************************************************/
699	/* work around various missing functions */	772	/* work around various missing functions */
700		773
701	#if !HAVE_PREADWRITE	774	#if !HAVE_PREADWRITE
		775	# undef pread
		776	# undef pwrite
702	# define pread eio__pread	777	# define pread eio__pread
703	# define pwrite eio__pwrite	778	# 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		779
712	static ssize_t	780	static ssize_t
713	eio__pread (int fd, void *buf, size_t count, off_t offset)	781	eio__pread (int fd, void *buf, size_t count, off_t offset)
714	{	782	{
715	ssize_t res;	783	ssize_t res;
…		…
733		801
734	X_LOCK (preadwritelock);	802	X_LOCK (preadwritelock);
735	ooffset = lseek (fd, 0, SEEK_CUR);	803	ooffset = lseek (fd, 0, SEEK_CUR);
736	lseek (fd, offset, SEEK_SET);	804	lseek (fd, offset, SEEK_SET);
737	res = write (fd, buf, count);	805	res = write (fd, buf, count);
738	lseek (fd, offset, SEEK_SET);	806	lseek (fd, ooffset, SEEK_SET);
739	X_UNLOCK (preadwritelock);	807	X_UNLOCK (preadwritelock);
740		808
741	return res;	809	return res;
742	}	810	}
743	#endif	811	#endif
744		812
745	#ifndef HAVE_FUTIMES	813	#ifndef HAVE_FUTIMES
746		814
		815	# undef utimes
		816	# undef futimes
747	# define utimes(path,times) eio__utimes (path, times)	817	# define utimes(path,times) eio__utimes (path, times)
748	# define futimes(fd,times) eio__futimes (fd, times)	818	# define futimes(fd,times) eio__futimes (fd, times)
749		819
750	static int	820	static int
751	eio__utimes (const char *filename, const struct timeval times[2])	821	eio__utimes (const char *filename, const struct timeval times[2])
…		…
770	}	840	}
771		841
772	#endif	842	#endif
773		843
774	#if !HAVE_FDATASYNC	844	#if !HAVE_FDATASYNC
		845	# undef fdatasync
775	# define fdatasync fsync	846	# define fdatasync(fd) fsync (fd)
776	#endif	847	#endif
		848
		849	/* sync_file_range always needs emulation */
		850	int
		851	eio__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	}
777		876
778	#if !HAVE_READAHEAD	877	#if !HAVE_READAHEAD
		878	# undef readahead
779	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	879	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
780		880
781	static ssize_t	881	static ssize_t
782	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)	882	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
783	{	883	{
…		…
824	if (res < 0 && sbytes)	924	if (res < 0 && sbytes)
825	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	925	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
826	res = sbytes;	926	res = sbytes;
827	}	927	}
828		928
		929	# elif defined __APPLE__
		930
		931	{
		932	off_t bytes = count;
		933	res = sendfile (ifd, ofd, offset, &bytes, 0, 0);
		934
		935	if (res < 0 && errno == EAGAIN && bytes)
		936	res = sbytes;
		937	}
		938
829	# elif __hpux	939	# elif __hpux
830	res = sendfile (ofd, ifd, offset, count, 0, 0);	940	res = sendfile (ofd, ifd, offset, count, 0, 0);
831		941
832	# elif __solaris	942	# elif __solaris
833	{	943	{
…		…
844	if (res < 0 && sbytes)	954	if (res < 0 && sbytes)
845	res = sbytes;	955	res = sbytes;
846	}	956	}
847		957
848	# endif	958	# endif
		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
…		…
891	}	1010	}
892		1011
893	return res;	1012	return res;
894	}	1013	}
895		1014
		1015	static signed char
		1016	eio_dent_cmp (const eio_dirent a, const eio_dirent b)
		1017	{
		1018	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1019	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1020	}
		1021
		1022	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1023
		1024	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1025	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1026
		1027	static void
		1028	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1029	{
		1030	unsigned char bits [9 + sizeof (ino_t) * 8];
		1031	unsigned char *bit = bits;
		1032
		1033	assert (CHAR_BIT == 8);
		1034	assert (sizeof (eio_dirent) * 8 < 256);
		1035	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1036	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1037
		1038	if (size <= EIO_SORT_FAST)
		1039	return;
		1040
		1041	/* first prepare an array of bits to test in our radix sort */
		1042	/* try to take endianness into account, as well as differences in ino_t sizes */
		1043	/* inode_bits must contain all inodes ORed together */
		1044	/* which is used to skip bits that are 0 everywhere, which is very common */
		1045	{
		1046	ino_t endianness;
		1047	int i, j;
		1048
		1049	/* we store the byte offset of byte n into byte n of "endianness" */
		1050	for (i = 0; i < sizeof (ino_t); ++i)
		1051	((unsigned char *)&endianness)[i] = i;
		1052
		1053	*bit++ = 0;
		1054
		1055	for (i = 0; i < sizeof (ino_t); ++i)
		1056	{
		1057	/* shifting off the byte offsets out of "endianness" */
		1058	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1059	endianness >>= 8;
		1060
		1061	for (j = 0; j < 8; ++j)
		1062	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1063	*bit++ = offs + j;
		1064	}
		1065
		1066	for (j = 0; j < 8; ++j)
		1067	if (score_bits & (1 << j))
		1068	bit++ = offsetof (eio_dirent, score) 8 + j;
		1069	}
		1070
		1071	/* now actually do the sorting (a variant of MSD radix sort) */
		1072	{
		1073	eio_dirent base_stk [9 + sizeof (ino_t) 8], *base;
		1074	eio_dirent end_stk [9 + sizeof (ino_t) 8], *end;
		1075	unsigned char bit_stk [9 + sizeof (ino_t) 8];
		1076	int stk_idx = 0;
		1077
		1078	base_stk [stk_idx] = dents;
		1079	end_stk [stk_idx] = dents + size;
		1080	bit_stk [stk_idx] = bit - 1;
		1081
		1082	do
		1083	{
		1084	base = base_stk [stk_idx];
		1085	end = end_stk [stk_idx];
		1086	bit = bit_stk [stk_idx];
		1087
		1088	for (;;)
		1089	{
		1090	unsigned char O = *bit >> 3;
		1091	unsigned char M = 1 << (*bit & 7);
		1092
		1093	eio_dirent *a = base;
		1094	eio_dirent *b = end;
		1095
		1096	if (b - a < EIO_SORT_CUTOFF)
		1097	break;
		1098
		1099	/* now bit-partition the array on the bit */
		1100	/* this ugly asymmetric loop seems to perform much better than typical */
		1101	/* partition algos found in the literature */
		1102	do
		1103	if (!(((unsigned char *)a)[O] & M))
		1104	++a;
		1105	else if (!(((unsigned char *)--b)[O] & M))
		1106	{
		1107	eio_dirent tmp = a; a = b; b = tmp;
		1108	++a;
		1109	}
		1110	while (b > a);
		1111
		1112	/* next bit, or stop, if no bits left in this path */
		1113	if (!*--bit)
		1114	break;
		1115
		1116	base_stk [stk_idx] = a;
		1117	end_stk [stk_idx] = end;
		1118	bit_stk [stk_idx] = bit;
		1119	++stk_idx;
		1120
		1121	end = a;
		1122	}
		1123	}
		1124	while (stk_idx--);
		1125	}
		1126	}
		1127
		1128	static void
		1129	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1130	{
		1131	/* first move the smallest element to the front, to act as a sentinel */
		1132	{
		1133	int i;
		1134	eio_dirent *min = dents;
		1135
		1136	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1137	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1138	if (EIO_DENT_CMP (dents [i], <, *min))
		1139	min = &dents [i];
		1140
		1141	/* swap elements 0 and j (minimum) */
		1142	{
		1143	eio_dirent tmp = dents; dents = min; min = tmp;
		1144	}
		1145	}
		1146
		1147	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1148	{
		1149	eio_dirent i, j;
		1150
		1151	for (i = dents + 1; i < dents + size; ++i)
		1152	{
		1153	eio_dirent value = *i;
		1154
		1155	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1156	j [1] = j [0];
		1157
		1158	j [1] = value;
		1159	}
		1160	}
		1161	}
		1162
		1163	static void
		1164	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1165	{
		1166	if (size <= 1)
		1167	return; /* our insertion sort relies on size > 0 */
		1168
		1169	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1170	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1171	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1172
		1173	/* use an insertion sort at the end, or for small arrays, */
		1174	/* as insertion sort is more efficient for small partitions */
		1175	eio_dent_insertion_sort (dents, size);
		1176	}
		1177
896	/* read a full directory */	1178	/* read a full directory */
897	static void	1179	static void
898	eio__scandir (eio_req req, etp_worker self)	1180	eio__scandir (eio_req req, etp_worker self)
899	{	1181	{
900	DIR *dirp;	1182	DIR *dirp;
901	EIO_STRUCT_DIRENT *entp;	1183	EIO_STRUCT_DIRENT *entp;
902	char name, names;	1184	char name, names;
903	int memlen = 4096;	1185	int namesalloc = 4096;
904	int memofs = 0;	1186	int namesoffs = 0;
		1187	int flags = req->int1;
		1188	eio_dirent *dents = 0;
		1189	int dentalloc = 128;
905	int res = 0;	1190	int dentoffs = 0;
		1191	ino_t inode_bits = 0;
		1192
		1193	req->result = -1;
		1194
		1195	if (!(flags & EIO_READDIR_DENTS))
		1196	flags &= ~(EIO_READDIR_DIRS_FIRST \| EIO_READDIR_STAT_ORDER);
906		1197
907	X_LOCK (wrklock);	1198	X_LOCK (wrklock);
908	/* the corresponding closedir is in ETP_WORKER_CLEAR */	1199	/* the corresponding closedir is in ETP_WORKER_CLEAR */
909	self->dirp = dirp = opendir (req->ptr1);	1200	self->dirp = dirp = opendir (req->ptr1);
910	req->flags \|= EIO_FLAG_PTR2_FREE;	1201	req->flags \|= EIO_FLAG_PTR1_FREE \| EIO_FLAG_PTR2_FREE;
		1202	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
911	req->ptr2 = names = malloc (memlen);	1203	req->ptr2 = names = malloc (namesalloc);
912	X_UNLOCK (wrklock);	1204	X_UNLOCK (wrklock);
913		1205
914	if (dirp && names)	1206	if (dirp && names && (!flags \|\| dents))
915	for (;;)	1207	for (;;)
916	{	1208	{
917	errno = 0;	1209	errno = 0;
918	entp = readdir (dirp);	1210	entp = readdir (dirp);
919		1211
920	if (!entp)	1212	if (!entp)
		1213	{
		1214	if (errno)
		1215	break;
		1216
		1217	/* sort etc. */
		1218	req->int1 = flags;
		1219	req->result = dentoffs;
		1220
		1221	if (flags & EIO_READDIR_STAT_ORDER)
		1222	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1223	else if (flags & EIO_READDIR_DIRS_FIRST)
		1224	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1225	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1226	else
		1227	{
		1228	/* in this case, all is known, and we just put dirs first and sort them */
		1229	eio_dirent *oth = dents + dentoffs;
		1230	eio_dirent *dir = dents;
		1231
		1232	/* now partition dirs to the front, and non-dirs to the back */
		1233	/* by walking from both sides and swapping if necessary */
		1234	/* also clear score, so it doesn't influence sorting */
		1235	while (oth > dir)
		1236	{
		1237	if (dir->type == EIO_DT_DIR)
		1238	++dir;
		1239	else if ((--oth)->type == EIO_DT_DIR)
		1240	{
		1241	eio_dirent tmp = dir; dir = oth; oth = tmp;
		1242
		1243	++dir;
		1244	}
		1245	}
		1246
		1247	/* now sort the dirs only */
		1248	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1249	}
		1250
921	break;	1251	break;
		1252	}
922		1253
		1254	/* now add the entry to our list(s) */
923	name = entp->d_name;	1255	name = entp->d_name;
924		1256
		1257	/* skip . and .. entries */
925	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))	1258	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))
926	{	1259	{
927	int len = strlen (name) + 1;	1260	int len = D_NAMLEN (entp) + 1;
928		1261
929	res++;	1262	while (expect_false (namesoffs + len > namesalloc))
930
931	while (memofs + len > memlen)
932	{	1263	{
933	memlen *= 2;	1264	namesalloc *= 2;
934	X_LOCK (wrklock);	1265	X_LOCK (wrklock);
935	req->ptr2 = names = realloc (names, memlen);	1266	req->ptr2 = names = realloc (names, namesalloc);
936	X_UNLOCK (wrklock);	1267	X_UNLOCK (wrklock);
937		1268
938	if (!names)	1269	if (!names)
939	break;	1270	break;
940	}	1271	}
941		1272
942	memcpy (names + memofs, name, len);	1273	memcpy (names + namesoffs, name, len);
		1274
		1275	if (dents)
		1276	{
		1277	struct eio_dirent *ent;
		1278
		1279	if (expect_false (dentoffs == dentalloc))
		1280	{
		1281	dentalloc *= 2;
		1282	X_LOCK (wrklock);
		1283	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1284	X_UNLOCK (wrklock);
		1285
		1286	if (!dents)
		1287	break;
		1288	}
		1289
		1290	ent = dents + dentoffs;
		1291
		1292	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1293	ent->namelen = len - 1;
		1294	ent->inode = D_INO (entp);
		1295
		1296	inode_bits \|= ent->inode;
		1297
		1298	switch (D_TYPE (entp))
		1299	{
		1300	default:
		1301	ent->type = EIO_DT_UNKNOWN;
		1302	flags \|= EIO_READDIR_FOUND_UNKNOWN;
		1303	break;
		1304
		1305	#ifdef DT_FIFO
		1306	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1307	#endif
		1308	#ifdef DT_CHR
		1309	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1310	#endif
		1311	#ifdef DT_MPC
		1312	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1313	#endif
		1314	#ifdef DT_DIR
		1315	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1316	#endif
		1317	#ifdef DT_NAM
		1318	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1319	#endif
		1320	#ifdef DT_BLK
		1321	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1322	#endif
		1323	#ifdef DT_MPB
		1324	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1325	#endif
		1326	#ifdef DT_REG
		1327	case DT_REG: ent->type = EIO_DT_REG; break;
		1328	#endif
		1329	#ifdef DT_NWK
		1330	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1331	#endif
		1332	#ifdef DT_CMP
		1333	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1334	#endif
		1335	#ifdef DT_LNK
		1336	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1337	#endif
		1338	#ifdef DT_SOCK
		1339	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1340	#endif
		1341	#ifdef DT_DOOR
		1342	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1343	#endif
		1344	#ifdef DT_WHT
		1345	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1346	#endif
		1347	}
		1348
		1349	ent->score = 7;
		1350
		1351	if (flags & EIO_READDIR_DIRS_FIRST)
		1352	{
		1353	if (ent->type == EIO_DT_UNKNOWN)
		1354	{
		1355	if (name == '.') / leading dots are likely directories, and, in any case, rare */
		1356	ent->score = 1;
		1357	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1358	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1359	}
		1360	else if (ent->type == EIO_DT_DIR)
		1361	ent->score = 0;
		1362	}
		1363	}
		1364
943	memofs += len;	1365	namesoffs += len;
		1366	++dentoffs;
		1367	}
		1368
		1369	if (EIO_CANCELLED (req))
		1370	{
		1371	errno = ECANCELED;
		1372	break;
944	}	1373	}
945	}	1374	}
		1375	}
946		1376
947	if (errno)	1377	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
948	res = -1;	1378	# undef msync
949		1379	# define msync(a,b,c) ((errno = ENOSYS), -1)
950	req->result = res;	1380	#endif
		1381
		1382	int
		1383	eio__mtouch (void *mem, size_t len, int flags)
		1384	{
		1385	intptr_t addr = (intptr_t)mem;
		1386	intptr_t end = addr + len;
		1387	#ifdef PAGESIZE
		1388	const intptr_t page = PAGESIZE;
		1389	#else
		1390	static intptr_t page;
		1391
		1392	if (!page)
		1393	page = sysconf (_SC_PAGESIZE);
		1394	#endif
		1395
		1396	/* round down to start of page, although this is probably useless */
		1397	addr &= ~(page - 1); /* assume page size is always a power of two */
		1398
		1399	if (addr < end)
		1400	if (flags) /* modify */
		1401	do { ((volatile sig_atomic_t )addr) \|= 0; } while ((addr += page) < len);
		1402	else
		1403	do { ((volatile sig_atomic_t )addr) ; } while ((addr += page) < len);
		1404
		1405	return 0;
951	}	1406	}
952		1407
953	/*****************************************************************************/	1408	/*****************************************************************************/
954		1409
955	#define ALLOC(len) \	1410	#define ALLOC(len) \
…		…
1122	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;	1577	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1123		1578
1124	case EIO_SYNC: req->result = 0; sync (); break;	1579	case EIO_SYNC: req->result = 0; sync (); break;
1125	case EIO_FSYNC: req->result = fsync (req->int1); break;	1580	case EIO_FSYNC: req->result = fsync (req->int1); break;
1126	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;	1581	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1582	case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break;
		1583	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1584	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1127		1585
1128	case EIO_READDIR: eio__scandir (req, self); break;	1586	case EIO_READDIR: eio__scandir (req, self); break;
1129		1587
1130	case EIO_BUSY:	1588	case EIO_BUSY:
1131	#ifdef _WIN32	1589	#ifdef _WIN32
…		…
1163		1621
1164	req->result = req->type == EIO_FUTIME	1622	req->result = req->type == EIO_FUTIME
1165	? futimes (req->int1, times)	1623	? futimes (req->int1, times)
1166	: utimes (req->ptr1, times);	1624	: utimes (req->ptr1, times);
1167	}	1625	}
		1626	break;
1168		1627
1169	case EIO_GROUP:	1628	case EIO_GROUP:
		1629	abort (); /* handled in eio_request */
		1630
1170	case EIO_NOP:	1631	case EIO_NOP:
1171	req->result = 0;	1632	req->result = 0;
1172	break;	1633	break;
1173		1634
1174	case EIO_CUSTOM:	1635	case EIO_CUSTOM:
1175	req->feed (req);	1636	((void ()(eio_req ))req->feed) (req);
1176	break;	1637	break;
1177		1638
1178	default:	1639	default:
1179	req->result = -1;	1640	req->result = -1;
1180	break;	1641	break;
…		…
1203	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)	1664	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)
1204	{	1665	{
1205	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1666	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1206	}	1667	}
1207		1668
		1669	eio_req eio_msync (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1670	{
		1671	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1672	}
		1673
		1674	eio_req eio_mtouch (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1675	{
		1676	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1677	}
		1678
		1679	eio_req eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void data)
		1680	{
		1681	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1682	}
		1683
1208	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)	1684	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)
1209	{	1685	{
1210	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1686	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1211	}	1687	}
1212		1688
…		…
1324	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)	1800	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)
1325	{	1801	{
1326	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1802	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1327	}	1803	}
1328		1804
1329	eio_req eio_readdir (const char path, int pri, eio_cb cb, void *data)	1805	eio_req eio_readdir (const char path, int flags, int pri, eio_cb cb, void *data)
1330	{	1806	{
1331	return eio__1path (EIO_READDIR, path, pri, cb, data);	1807	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1332	}	1808	}
1333		1809
1334	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1810	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1335	{	1811	{
1336	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;	1812	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);	1843	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1368	}	1844	}
1369		1845
1370	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)	1846	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)
1371	{	1847	{
1372	REQ (EIO_CUSTOM); req->feed = execute; SEND;	1848	REQ (EIO_CUSTOM); req->feed = (void ()(eio_req ))execute; SEND;
1373	}	1849	}
1374		1850
1375	#endif	1851	#endif
1376		1852
1377	eio_req eio_grp (eio_cb cb, void data)	1853	eio_req eio_grp (eio_cb cb, void data)
…		…
1405		1881
1406	void eio_grp_add (eio_req grp, eio_req req)	1882	void eio_grp_add (eio_req grp, eio_req req)
1407	{	1883	{
1408	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	1884	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1409		1885
		1886	grp->flags \|= EIO_FLAG_GROUPADD;
		1887
1410	++grp->size;	1888	++grp->size;
1411	req->grp = grp;	1889	req->grp = grp;
1412		1890
1413	req->grp_prev = 0;	1891	req->grp_prev = 0;
1414	req->grp_next = grp->grp_first;	1892	req->grp_next = grp->grp_first;
…		…
1423	/* misc garbage */	1901	/* misc garbage */
1424		1902
1425	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	1903	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1426	{	1904	{
1427	etp_worker wrk;	1905	etp_worker wrk;
		1906	ssize_t ret;
1428		1907
1429	wrk.dbuf = 0;	1908	wrk.dbuf = 0;
1430		1909
1431	eio__sendfile (ofd, ifd, offset, count, &wrk);	1910	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1432		1911
1433	if (wrk.dbuf)	1912	if (wrk.dbuf)
1434	free (wrk.dbuf);	1913	free (wrk.dbuf);
1435	}
1436		1914
		1915	return ret;
		1916	}
		1917

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 1.45 by root, Sat Jan 2 12:50:22 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 1.45 by root, Sat Jan 2 12:50:22 2010 UTC