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

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.19 by root, Tue Jun 17 23:34:20 2008 UTC vs.
Revision 1.47 by root, Sat Jan 2 13:02:20 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>
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;
…		…
330		383
331	static void etp_atfork_child (void)	384	static 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;
…		…
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
856	#if __solaris	975	#if __solaris
857	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE	976	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE
858	#endif	977	#endif
		978	\|\| errno == ENOTSUP \|\| errno == EOPNOTSUPP
859	)	979	)
860	)	980	)
861	{	981	{
862	/* emulate sendfile. this is a major pain in the ass */	982	/* emulate sendfile. this is a major pain in the ass */
863	dBUF;	983	dBUF;
…		…
891	}	1011	}
892		1012
893	return res;	1013	return res;
894	}	1014	}
895		1015
		1016	static signed char
		1017	eio_dent_cmp (const eio_dirent a, const eio_dirent b)
		1018	{
		1019	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1020	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1021	}
		1022
		1023	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1024
		1025	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1026	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1027
		1028	static void
		1029	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1030	{
		1031	unsigned char bits [9 + sizeof (ino_t) * 8];
		1032	unsigned char *bit = bits;
		1033
		1034	assert (CHAR_BIT == 8);
		1035	assert (sizeof (eio_dirent) * 8 < 256);
		1036	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1037	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1038
		1039	if (size <= EIO_SORT_FAST)
		1040	return;
		1041
		1042	/* first prepare an array of bits to test in our radix sort */
		1043	/* try to take endianness into account, as well as differences in ino_t sizes */
		1044	/* inode_bits must contain all inodes ORed together */
		1045	/* which is used to skip bits that are 0 everywhere, which is very common */
		1046	{
		1047	ino_t endianness;
		1048	int i, j;
		1049
		1050	/* we store the byte offset of byte n into byte n of "endianness" */
		1051	for (i = 0; i < sizeof (ino_t); ++i)
		1052	((unsigned char *)&endianness)[i] = i;
		1053
		1054	*bit++ = 0;
		1055
		1056	for (i = 0; i < sizeof (ino_t); ++i)
		1057	{
		1058	/* shifting off the byte offsets out of "endianness" */
		1059	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1060	endianness >>= 8;
		1061
		1062	for (j = 0; j < 8; ++j)
		1063	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1064	*bit++ = offs + j;
		1065	}
		1066
		1067	for (j = 0; j < 8; ++j)
		1068	if (score_bits & (1 << j))
		1069	bit++ = offsetof (eio_dirent, score) 8 + j;
		1070	}
		1071
		1072	/* now actually do the sorting (a variant of MSD radix sort) */
		1073	{
		1074	eio_dirent base_stk [9 + sizeof (ino_t) 8], *base;
		1075	eio_dirent end_stk [9 + sizeof (ino_t) 8], *end;
		1076	unsigned char bit_stk [9 + sizeof (ino_t) 8];
		1077	int stk_idx = 0;
		1078
		1079	base_stk [stk_idx] = dents;
		1080	end_stk [stk_idx] = dents + size;
		1081	bit_stk [stk_idx] = bit - 1;
		1082
		1083	do
		1084	{
		1085	base = base_stk [stk_idx];
		1086	end = end_stk [stk_idx];
		1087	bit = bit_stk [stk_idx];
		1088
		1089	for (;;)
		1090	{
		1091	unsigned char O = *bit >> 3;
		1092	unsigned char M = 1 << (*bit & 7);
		1093
		1094	eio_dirent *a = base;
		1095	eio_dirent *b = end;
		1096
		1097	if (b - a < EIO_SORT_CUTOFF)
		1098	break;
		1099
		1100	/* now bit-partition the array on the bit */
		1101	/* this ugly asymmetric loop seems to perform much better than typical */
		1102	/* partition algos found in the literature */
		1103	do
		1104	if (!(((unsigned char *)a)[O] & M))
		1105	++a;
		1106	else if (!(((unsigned char *)--b)[O] & M))
		1107	{
		1108	eio_dirent tmp = a; a = b; b = tmp;
		1109	++a;
		1110	}
		1111	while (b > a);
		1112
		1113	/* next bit, or stop, if no bits left in this path */
		1114	if (!*--bit)
		1115	break;
		1116
		1117	base_stk [stk_idx] = a;
		1118	end_stk [stk_idx] = end;
		1119	bit_stk [stk_idx] = bit;
		1120	++stk_idx;
		1121
		1122	end = a;
		1123	}
		1124	}
		1125	while (stk_idx--);
		1126	}
		1127	}
		1128
		1129	static void
		1130	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1131	{
		1132	/* first move the smallest element to the front, to act as a sentinel */
		1133	{
		1134	int i;
		1135	eio_dirent *min = dents;
		1136
		1137	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1138	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1139	if (EIO_DENT_CMP (dents [i], <, *min))
		1140	min = &dents [i];
		1141
		1142	/* swap elements 0 and j (minimum) */
		1143	{
		1144	eio_dirent tmp = dents; dents = min; min = tmp;
		1145	}
		1146	}
		1147
		1148	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1149	{
		1150	eio_dirent i, j;
		1151
		1152	for (i = dents + 1; i < dents + size; ++i)
		1153	{
		1154	eio_dirent value = *i;
		1155
		1156	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1157	j [1] = j [0];
		1158
		1159	j [1] = value;
		1160	}
		1161	}
		1162	}
		1163
		1164	static void
		1165	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1166	{
		1167	if (size <= 1)
		1168	return; /* our insertion sort relies on size > 0 */
		1169
		1170	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1171	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1172	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1173
		1174	/* use an insertion sort at the end, or for small arrays, */
		1175	/* as insertion sort is more efficient for small partitions */
		1176	eio_dent_insertion_sort (dents, size);
		1177	}
		1178
896	/* read a full directory */	1179	/* read a full directory */
897	static void	1180	static void
898	eio__scandir (eio_req req, etp_worker self)	1181	eio__scandir (eio_req req, etp_worker self)
899	{	1182	{
900	DIR *dirp;	1183	DIR *dirp;
901	EIO_STRUCT_DIRENT *entp;	1184	EIO_STRUCT_DIRENT *entp;
902	char name, names;	1185	char name, names;
903	int memlen = 4096;	1186	int namesalloc = 4096;
904	int memofs = 0;	1187	int namesoffs = 0;
		1188	int flags = req->int1;
		1189	eio_dirent *dents = 0;
		1190	int dentalloc = 128;
905	int res = 0;	1191	int dentoffs = 0;
		1192	ino_t inode_bits = 0;
		1193
		1194	req->result = -1;
		1195
		1196	if (!(flags & EIO_READDIR_DENTS))
		1197	flags &= ~(EIO_READDIR_DIRS_FIRST \| EIO_READDIR_STAT_ORDER);
906		1198
907	X_LOCK (wrklock);	1199	X_LOCK (wrklock);
		1200	/* the corresponding closedir is in ETP_WORKER_CLEAR */
908	self->dirp = dirp = opendir (req->ptr1);	1201	self->dirp = dirp = opendir (req->ptr1);
909	req->flags \|= EIO_FLAG_PTR2_FREE;	1202	req->flags \|= EIO_FLAG_PTR1_FREE \| EIO_FLAG_PTR2_FREE;
		1203	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
910	req->ptr2 = names = malloc (memlen);	1204	req->ptr2 = names = malloc (namesalloc);
911	X_UNLOCK (wrklock);	1205	X_UNLOCK (wrklock);
912		1206
913	if (dirp && names)	1207	if (dirp && names && (!flags \|\| dents))
914	for (;;)	1208	for (;;)
915	{	1209	{
916	errno = 0;	1210	errno = 0;
917	entp = readdir (dirp);	1211	entp = readdir (dirp);
918		1212
919	if (!entp)	1213	if (!entp)
		1214	{
		1215	if (errno)
		1216	break;
		1217
		1218	/* sort etc. */
		1219	req->int1 = flags;
		1220	req->result = dentoffs;
		1221
		1222	if (flags & EIO_READDIR_STAT_ORDER)
		1223	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1224	else if (flags & EIO_READDIR_DIRS_FIRST)
		1225	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1226	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1227	else
		1228	{
		1229	/* in this case, all is known, and we just put dirs first and sort them */
		1230	eio_dirent *oth = dents + dentoffs;
		1231	eio_dirent *dir = dents;
		1232
		1233	/* now partition dirs to the front, and non-dirs to the back */
		1234	/* by walking from both sides and swapping if necessary */
		1235	/* also clear score, so it doesn't influence sorting */
		1236	while (oth > dir)
		1237	{
		1238	if (dir->type == EIO_DT_DIR)
		1239	++dir;
		1240	else if ((--oth)->type == EIO_DT_DIR)
		1241	{
		1242	eio_dirent tmp = dir; dir = oth; oth = tmp;
		1243
		1244	++dir;
		1245	}
		1246	}
		1247
		1248	/* now sort the dirs only */
		1249	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1250	}
		1251
920	break;	1252	break;
		1253	}
921		1254
		1255	/* now add the entry to our list(s) */
922	name = entp->d_name;	1256	name = entp->d_name;
923		1257
		1258	/* skip . and .. entries */
924	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))	1259	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))
925	{	1260	{
926	int len = strlen (name) + 1;	1261	int len = D_NAMLEN (entp) + 1;
927		1262
928	res++;	1263	while (expect_false (namesoffs + len > namesalloc))
929
930	while (memofs + len > memlen)
931	{	1264	{
932	memlen *= 2;	1265	namesalloc *= 2;
933	X_LOCK (wrklock);	1266	X_LOCK (wrklock);
934	req->ptr2 = names = realloc (names, memlen);	1267	req->ptr2 = names = realloc (names, namesalloc);
935	X_UNLOCK (wrklock);	1268	X_UNLOCK (wrklock);
936		1269
937	if (!names)	1270	if (!names)
938	break;	1271	break;
939	}	1272	}
940		1273
941	memcpy (names + memofs, name, len);	1274	memcpy (names + namesoffs, name, len);
		1275
		1276	if (dents)
		1277	{
		1278	struct eio_dirent *ent;
		1279
		1280	if (expect_false (dentoffs == dentalloc))
		1281	{
		1282	dentalloc *= 2;
		1283	X_LOCK (wrklock);
		1284	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1285	X_UNLOCK (wrklock);
		1286
		1287	if (!dents)
		1288	break;
		1289	}
		1290
		1291	ent = dents + dentoffs;
		1292
		1293	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1294	ent->namelen = len - 1;
		1295	ent->inode = D_INO (entp);
		1296
		1297	inode_bits \|= ent->inode;
		1298
		1299	switch (D_TYPE (entp))
		1300	{
		1301	default:
		1302	ent->type = EIO_DT_UNKNOWN;
		1303	flags \|= EIO_READDIR_FOUND_UNKNOWN;
		1304	break;
		1305
		1306	#ifdef DT_FIFO
		1307	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1308	#endif
		1309	#ifdef DT_CHR
		1310	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1311	#endif
		1312	#ifdef DT_MPC
		1313	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1314	#endif
		1315	#ifdef DT_DIR
		1316	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1317	#endif
		1318	#ifdef DT_NAM
		1319	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1320	#endif
		1321	#ifdef DT_BLK
		1322	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1323	#endif
		1324	#ifdef DT_MPB
		1325	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1326	#endif
		1327	#ifdef DT_REG
		1328	case DT_REG: ent->type = EIO_DT_REG; break;
		1329	#endif
		1330	#ifdef DT_NWK
		1331	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1332	#endif
		1333	#ifdef DT_CMP
		1334	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1335	#endif
		1336	#ifdef DT_LNK
		1337	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1338	#endif
		1339	#ifdef DT_SOCK
		1340	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1341	#endif
		1342	#ifdef DT_DOOR
		1343	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1344	#endif
		1345	#ifdef DT_WHT
		1346	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1347	#endif
		1348	}
		1349
		1350	ent->score = 7;
		1351
		1352	if (flags & EIO_READDIR_DIRS_FIRST)
		1353	{
		1354	if (ent->type == EIO_DT_UNKNOWN)
		1355	{
		1356	if (name == '.') / leading dots are likely directories, and, in any case, rare */
		1357	ent->score = 1;
		1358	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1359	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1360	}
		1361	else if (ent->type == EIO_DT_DIR)
		1362	ent->score = 0;
		1363	}
		1364	}
		1365
942	memofs += len;	1366	namesoffs += len;
		1367	++dentoffs;
		1368	}
		1369
		1370	if (EIO_CANCELLED (req))
		1371	{
		1372	errno = ECANCELED;
		1373	break;
943	}	1374	}
944	}	1375	}
		1376	}
945		1377
946	if (errno)	1378	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
947	res = -1;	1379	# undef msync
948		1380	# define msync(a,b,c) ((errno = ENOSYS), -1)
949	req->result = res;	1381	#endif
		1382
		1383	int
		1384	eio__mtouch (void *mem, size_t len, int flags)
		1385	{
		1386	intptr_t addr = (intptr_t)mem;
		1387	intptr_t end = addr + len;
		1388	#ifdef PAGESIZE
		1389	const intptr_t page = PAGESIZE;
		1390	#else
		1391	static intptr_t page;
		1392
		1393	if (!page)
		1394	page = sysconf (_SC_PAGESIZE);
		1395	#endif
		1396
		1397	/* round down to start of page, although this is probably useless */
		1398	addr &= ~(page - 1); /* assume page size is always a power of two */
		1399
		1400	if (addr < end)
		1401	if (flags) /* modify */
		1402	do { ((volatile sig_atomic_t )addr) \|= 0; } while ((addr += page) < len);
		1403	else
		1404	do { ((volatile sig_atomic_t )addr) ; } while ((addr += page) < len);
		1405
		1406	return 0;
950	}	1407	}
951		1408
952	/*****************************************************************************/	1409	/*****************************************************************************/
953		1410
954	#define ALLOC(len) \	1411	#define ALLOC(len) \
…		…
1121	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;	1578	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1122		1579
1123	case EIO_SYNC: req->result = 0; sync (); break;	1580	case EIO_SYNC: req->result = 0; sync (); break;
1124	case EIO_FSYNC: req->result = fsync (req->int1); break;	1581	case EIO_FSYNC: req->result = fsync (req->int1); break;
1125	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;	1582	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1583	case EIO_MSYNC: req->result = msync (req->ptr2, req->size, req->int1); break;
		1584	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1585	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1126		1586
1127	case EIO_READDIR: eio__scandir (req, self); break;	1587	case EIO_READDIR: eio__scandir (req, self); break;
1128		1588
1129	case EIO_BUSY:	1589	case EIO_BUSY:
1130	#ifdef _WIN32	1590	#ifdef _WIN32
…		…
1162		1622
1163	req->result = req->type == EIO_FUTIME	1623	req->result = req->type == EIO_FUTIME
1164	? futimes (req->int1, times)	1624	? futimes (req->int1, times)
1165	: utimes (req->ptr1, times);	1625	: utimes (req->ptr1, times);
1166	}	1626	}
		1627	break;
1167		1628
1168	case EIO_GROUP:	1629	case EIO_GROUP:
		1630	abort (); /* handled in eio_request */
		1631
1169	case EIO_NOP:	1632	case EIO_NOP:
1170	req->result = 0;	1633	req->result = 0;
1171	break;	1634	break;
1172		1635
1173	case EIO_CUSTOM:	1636	case EIO_CUSTOM:
1174	req->feed (req);	1637	((void ()(eio_req ))req->feed) (req);
1175	break;	1638	break;
1176		1639
1177	default:	1640	default:
1178	req->result = -1;	1641	req->result = -1;
1179	break;	1642	break;
…		…
1202	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)	1665	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)
1203	{	1666	{
1204	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1667	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1205	}	1668	}
1206		1669
		1670	eio_req eio_msync (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1671	{
		1672	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1673	}
		1674
		1675	eio_req eio_mtouch (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1676	{
		1677	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1678	}
		1679
		1680	eio_req eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void data)
		1681	{
		1682	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1683	}
		1684
1207	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)	1685	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)
1208	{	1686	{
1209	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1687	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1210	}	1688	}
1211		1689
…		…
1323	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)	1801	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)
1324	{	1802	{
1325	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1803	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1326	}	1804	}
1327		1805
1328	eio_req eio_readdir (const char path, int pri, eio_cb cb, void *data)	1806	eio_req eio_readdir (const char path, int flags, int pri, eio_cb cb, void *data)
1329	{	1807	{
1330	return eio__1path (EIO_READDIR, path, pri, cb, data);	1808	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1331	}	1809	}
1332		1810
1333	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1811	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1334	{	1812	{
1335	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;	1813	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
…		…
1366	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1844	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1367	}	1845	}
1368		1846
1369	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)	1847	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)
1370	{	1848	{
1371	REQ (EIO_CUSTOM); req->feed = execute; SEND;	1849	REQ (EIO_CUSTOM); req->feed = (void ()(eio_req ))execute; SEND;
1372	}	1850	}
1373		1851
1374	#endif	1852	#endif
1375		1853
1376	eio_req eio_grp (eio_cb cb, void data)	1854	eio_req eio_grp (eio_cb cb, void data)
…		…
1404		1882
1405	void eio_grp_add (eio_req grp, eio_req req)	1883	void eio_grp_add (eio_req grp, eio_req req)
1406	{	1884	{
1407	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	1885	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1408		1886
		1887	grp->flags \|= EIO_FLAG_GROUPADD;
		1888
1409	++grp->size;	1889	++grp->size;
1410	req->grp = grp;	1890	req->grp = grp;
1411		1891
1412	req->grp_prev = 0;	1892	req->grp_prev = 0;
1413	req->grp_next = grp->grp_first;	1893	req->grp_next = grp->grp_first;
…		…
1422	/* misc garbage */	1902	/* misc garbage */
1423		1903
1424	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	1904	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1425	{	1905	{
1426	etp_worker wrk;	1906	etp_worker wrk;
		1907	ssize_t ret;
1427		1908
1428	wrk.dbuf = 0;	1909	wrk.dbuf = 0;
1429		1910
1430	eio__sendfile (ofd, ifd, offset, count, &wrk);	1911	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1431		1912
1432	if (wrk.dbuf)	1913	if (wrk.dbuf)
1433	free (wrk.dbuf);	1914	free (wrk.dbuf);
1434	}
1435		1915
		1916	return ret;
		1917	}
		1918

Diff Legend

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

Comparing cvsroot/libeio/eio.c (file contents): Revision 1.19 by root, Tue Jun 17 23:34:20 2008 UTC vs. Revision 1.47 by root, Sat Jan 2 13:02:20 2010 UTC

Diff Legend

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.19 by root, Tue Jun 17 23:34:20 2008 UTC vs.
Revision 1.47 by root, Sat Jan 2 13:02:20 2010 UTC