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

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.22 by root, Fri Jul 11 10:54:50 2008 UTC vs.
Revision 1.50 by root, Thu Jan 7 17:18:08 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	}
…		…
411	X_UNLOCK (wrklock);	455	X_UNLOCK (wrklock);
412	}	456	}
413		457
414	static void etp_maybe_start_thread (void)	458	static void etp_maybe_start_thread (void)
415	{	459	{
416	if (etp_nthreads () >= wanted)	460	if (expect_true (etp_nthreads () >= wanted))
417	return;	461	return;
418		462
419	/* todo: maybe use idle here, but might be less exact */	463	/* todo: maybe use idle here, but might be less exact */
420	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 ()))
421	return;	465	return;
422		466
423	etp_start_thread ();	467	etp_start_thread ();
424	}	468	}
425		469
…		…
478		522
479	X_LOCK (reqlock);	523	X_LOCK (reqlock);
480	--nreqs;	524	--nreqs;
481	X_UNLOCK (reqlock);	525	X_UNLOCK (reqlock);
482		526
483	if (req->type == EIO_GROUP && req->size)	527	if (expect_false (req->type == EIO_GROUP && req->size))
484	{	528	{
485	req->int1 = 1; /* mark request as delayed */	529	req->int1 = 1; /* mark request as delayed */
486	continue;	530	continue;
487	}	531	}
488	else	532	else
489	{	533	{
490	int res = ETP_FINISH (req);	534	int res = ETP_FINISH (req);
491	if (res)	535	if (expect_false (res))
492	return res;	536	return res;
493	}	537	}
494		538
495	if (maxreqs && !--maxreqs)	539	if (expect_false (maxreqs && !--maxreqs))
496	break;	540	break;
497		541
498	if (maxtime)	542	if (maxtime)
499	{	543	{
500	gettimeofday (&tv_now, 0);	544	gettimeofday (&tv_now, 0);
…		…
519		563
520	static void etp_submit (ETP_REQ *req)	564	static void etp_submit (ETP_REQ *req)
521	{	565	{
522	req->pri -= ETP_PRI_MIN;	566	req->pri -= ETP_PRI_MIN;
523		567
524	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;
525	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;
526		570
		571	if (expect_false (req->type == EIO_GROUP))
		572	{
		573	/* I hope this is worth it :/ */
527	X_LOCK (reqlock);	574	X_LOCK (reqlock);
528	++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;
529	++nready;	591	++nready;
530	reqq_push (&req_queue, req);	592	reqq_push (&req_queue, req);
531	X_COND_SIGNAL (reqwait);	593	X_COND_SIGNAL (reqwait);
532	X_UNLOCK (reqlock);	594	X_UNLOCK (reqlock);
533		595
534	etp_maybe_start_thread ();	596	etp_maybe_start_thread ();
		597	}
535	}	598	}
536		599
537	static void etp_set_max_poll_time (double nseconds)	600	static void etp_set_max_poll_time (double nseconds)
538	{	601	{
539	if (WORDACCESS_UNSAFE) X_LOCK (reslock);	602	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
540	max_poll_time = nseconds;	603	max_poll_time = nseconds * EIO_TICKS;
541	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);	604	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
542	}	605	}
543		606
544	static void etp_set_max_poll_reqs (unsigned int maxreqs)	607	static void etp_set_max_poll_reqs (unsigned int maxreqs)
545	{	608	{
…		…
574		637
575	static void grp_try_feed (eio_req *grp)	638	static void grp_try_feed (eio_req *grp)
576	{	639	{
577	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	640	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
578	{	641	{
579	int old_len = grp->size;	642	grp->flags &= ~EIO_FLAG_GROUPADD;
580		643
581	EIO_FEED (grp);	644	EIO_FEED (grp);
582		645
583	/* stop if no progress has been made */	646	/* stop if no progress has been made */
584	if (old_len == grp->size)	647	if (!(grp->flags & EIO_FLAG_GROUPADD))
585	{	648	{
586	grp->feed = 0;	649	grp->feed = 0;
587	break;	650	break;
588	}	651	}
589	}	652	}
…		…
706		769
707	/*****************************************************************************/	770	/*****************************************************************************/
708	/* work around various missing functions */	771	/* work around various missing functions */
709		772
710	#if !HAVE_PREADWRITE	773	#if !HAVE_PREADWRITE
		774	# undef pread
		775	# undef pwrite
711	# define pread eio__pread	776	# define pread eio__pread
712	# define pwrite eio__pwrite	777	# define pwrite eio__pwrite
713		778
714	static ssize_t	779	static ssize_t
715	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)
…		…
735		800
736	X_LOCK (preadwritelock);	801	X_LOCK (preadwritelock);
737	ooffset = lseek (fd, 0, SEEK_CUR);	802	ooffset = lseek (fd, 0, SEEK_CUR);
738	lseek (fd, offset, SEEK_SET);	803	lseek (fd, offset, SEEK_SET);
739	res = write (fd, buf, count);	804	res = write (fd, buf, count);
740	lseek (fd, offset, SEEK_SET);	805	lseek (fd, ooffset, SEEK_SET);
741	X_UNLOCK (preadwritelock);	806	X_UNLOCK (preadwritelock);
742		807
743	return res;	808	return res;
744	}	809	}
745	#endif	810	#endif
746		811
747	#ifndef HAVE_FUTIMES	812	#ifndef HAVE_FUTIMES
748		813
		814	# undef utimes
		815	# undef futimes
749	# define utimes(path,times) eio__utimes (path, times)	816	# define utimes(path,times) eio__utimes (path, times)
750	# define futimes(fd,times) eio__futimes (fd, times)	817	# define futimes(fd,times) eio__futimes (fd, times)
751		818
752	static int	819	static int
753	eio__utimes (const char *filename, const struct timeval times[2])	820	eio__utimes (const char *filename, const struct timeval times[2])
…		…
772	}	839	}
773		840
774	#endif	841	#endif
775		842
776	#if !HAVE_FDATASYNC	843	#if !HAVE_FDATASYNC
		844	# undef fdatasync
777	# define fdatasync fsync	845	# define fdatasync(fd) fsync (fd)
778	#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	}
779		875
780	#if !HAVE_READAHEAD	876	#if !HAVE_READAHEAD
		877	# undef readahead
781	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	878	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
782		879
783	static ssize_t	880	static ssize_t
784	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)
785	{	882	{
…		…
812		909
813	#if HAVE_SENDFILE	910	#if HAVE_SENDFILE
814	# if __linux	911	# if __linux
815	res = sendfile (ofd, ifd, &offset, count);	912	res = sendfile (ofd, ifd, &offset, count);
816		913
817	# elif __freebsd	914	# elif __FreeBSD__
818	/*	915	/*
819	* Of course, the freebsd sendfile is a dire hack with no thoughts	916	* Of course, the freebsd sendfile is a dire hack with no thoughts
820	* wasted on making it similar to other I/O functions.	917	* wasted on making it similar to other I/O functions.
821	*/	918	*/
822	{	919	{
823	off_t sbytes;	920	off_t sbytes;
824	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	921	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
825		922
		923	/* freebsd' sendfile will return 0 when success */
826	if (res < 0 && sbytes)	924	if (res == 0 && sbytes)
827	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	925	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
		926	res = sbytes;
		927	}
		928
		929	# elif defined (__APPLE__)
		930
		931	{
		932	off_t sbytes = count;
		933	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		934
		935	if (res == 0 && errno == EAGAIN && sbytes)
828	res = sbytes;	936	res = sbytes;
829	}	937	}
830		938
831	# elif __hpux	939	# elif __hpux
832	res = sendfile (ofd, ifd, offset, count, 0, 0);	940	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
846	if (res < 0 && sbytes)	954	if (res < 0 && sbytes)
847	res = sbytes;	955	res = sbytes;
848	}	956	}
849		957
850	# endif	958	# endif
		959
		960	#elif defined (_WIN32)
		961
		962	/* does not work, just for documentation of what would need to be done */
		963	{
		964	HANDLE h = TO_SOCKET (ifd);
		965	SetFilePointer (h, offset, 0, FILE_BEGIN);
		966	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		967	}
		968
851	#else	969	#else
852	res = -1;	970	res = -1;
853	errno = ENOSYS;	971	errno = ENOSYS;
854	#endif	972	#endif
855		973
856	if (res < 0	974	if (res < 0
857	&& (errno == ENOSYS \|\| errno == EINVAL \|\| errno == ENOTSOCK	975	&& (errno == ENOSYS \|\| errno == EINVAL \|\| errno == ENOTSOCK
		976	/* BSDs */
		977	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		978	\|\| errno == ENOTSUP
		979	#endif
		980	\|\| errno == EOPNOTSUPP /* BSDs */
858	#if __solaris	981	#if __solaris
859	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE	982	\|\| errno == EAFNOSUPPORT \|\| errno == EPROTOTYPE
860	#endif	983	#endif
861	)	984	)
862	)	985	)
…		…
893	}	1016	}
894		1017
895	return res;	1018	return res;
896	}	1019	}
897		1020
		1021	static signed char
		1022	eio_dent_cmp (const eio_dirent a, const eio_dirent b)
		1023	{
		1024	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1025	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1026	}
		1027
		1028	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1029
		1030	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1031	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1032
		1033	static void
		1034	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1035	{
		1036	unsigned char bits [9 + sizeof (ino_t) * 8];
		1037	unsigned char *bit = bits;
		1038
		1039	assert (CHAR_BIT == 8);
		1040	assert (sizeof (eio_dirent) * 8 < 256);
		1041	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1042	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1043
		1044	if (size <= EIO_SORT_FAST)
		1045	return;
		1046
		1047	/* first prepare an array of bits to test in our radix sort */
		1048	/* try to take endianness into account, as well as differences in ino_t sizes */
		1049	/* inode_bits must contain all inodes ORed together */
		1050	/* which is used to skip bits that are 0 everywhere, which is very common */
		1051	{
		1052	ino_t endianness;
		1053	int i, j;
		1054
		1055	/* we store the byte offset of byte n into byte n of "endianness" */
		1056	for (i = 0; i < sizeof (ino_t); ++i)
		1057	((unsigned char *)&endianness)[i] = i;
		1058
		1059	*bit++ = 0;
		1060
		1061	for (i = 0; i < sizeof (ino_t); ++i)
		1062	{
		1063	/* shifting off the byte offsets out of "endianness" */
		1064	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1065	endianness >>= 8;
		1066
		1067	for (j = 0; j < 8; ++j)
		1068	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1069	*bit++ = offs + j;
		1070	}
		1071
		1072	for (j = 0; j < 8; ++j)
		1073	if (score_bits & (1 << j))
		1074	bit++ = offsetof (eio_dirent, score) 8 + j;
		1075	}
		1076
		1077	/* now actually do the sorting (a variant of MSD radix sort) */
		1078	{
		1079	eio_dirent base_stk [9 + sizeof (ino_t) 8], *base;
		1080	eio_dirent end_stk [9 + sizeof (ino_t) 8], *end;
		1081	unsigned char bit_stk [9 + sizeof (ino_t) 8];
		1082	int stk_idx = 0;
		1083
		1084	base_stk [stk_idx] = dents;
		1085	end_stk [stk_idx] = dents + size;
		1086	bit_stk [stk_idx] = bit - 1;
		1087
		1088	do
		1089	{
		1090	base = base_stk [stk_idx];
		1091	end = end_stk [stk_idx];
		1092	bit = bit_stk [stk_idx];
		1093
		1094	for (;;)
		1095	{
		1096	unsigned char O = *bit >> 3;
		1097	unsigned char M = 1 << (*bit & 7);
		1098
		1099	eio_dirent *a = base;
		1100	eio_dirent *b = end;
		1101
		1102	if (b - a < EIO_SORT_CUTOFF)
		1103	break;
		1104
		1105	/* now bit-partition the array on the bit */
		1106	/* this ugly asymmetric loop seems to perform much better than typical */
		1107	/* partition algos found in the literature */
		1108	do
		1109	if (!(((unsigned char *)a)[O] & M))
		1110	++a;
		1111	else if (!(((unsigned char *)--b)[O] & M))
		1112	{
		1113	eio_dirent tmp = a; a = b; b = tmp;
		1114	++a;
		1115	}
		1116	while (b > a);
		1117
		1118	/* next bit, or stop, if no bits left in this path */
		1119	if (!*--bit)
		1120	break;
		1121
		1122	base_stk [stk_idx] = a;
		1123	end_stk [stk_idx] = end;
		1124	bit_stk [stk_idx] = bit;
		1125	++stk_idx;
		1126
		1127	end = a;
		1128	}
		1129	}
		1130	while (stk_idx--);
		1131	}
		1132	}
		1133
		1134	static void
		1135	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1136	{
		1137	/* first move the smallest element to the front, to act as a sentinel */
		1138	{
		1139	int i;
		1140	eio_dirent *min = dents;
		1141
		1142	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1143	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1144	if (EIO_DENT_CMP (dents [i], <, *min))
		1145	min = &dents [i];
		1146
		1147	/* swap elements 0 and j (minimum) */
		1148	{
		1149	eio_dirent tmp = dents; dents = min; min = tmp;
		1150	}
		1151	}
		1152
		1153	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1154	{
		1155	eio_dirent i, j;
		1156
		1157	for (i = dents + 1; i < dents + size; ++i)
		1158	{
		1159	eio_dirent value = *i;
		1160
		1161	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1162	j [1] = j [0];
		1163
		1164	j [1] = value;
		1165	}
		1166	}
		1167	}
		1168
		1169	static void
		1170	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1171	{
		1172	if (size <= 1)
		1173	return; /* our insertion sort relies on size > 0 */
		1174
		1175	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1176	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1177	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1178
		1179	/* use an insertion sort at the end, or for small arrays, */
		1180	/* as insertion sort is more efficient for small partitions */
		1181	eio_dent_insertion_sort (dents, size);
		1182	}
		1183
898	/* read a full directory */	1184	/* read a full directory */
899	static void	1185	static void
900	eio__scandir (eio_req req, etp_worker self)	1186	eio__scandir (eio_req req, etp_worker self)
901	{	1187	{
902	DIR *dirp;	1188	DIR *dirp;
903	EIO_STRUCT_DIRENT *entp;	1189	EIO_STRUCT_DIRENT *entp;
904	char name, names;	1190	char name, names;
905	int memlen = 4096;	1191	int namesalloc = 4096;
906	int memofs = 0;	1192	int namesoffs = 0;
		1193	int flags = req->int1;
		1194	eio_dirent *dents = 0;
		1195	int dentalloc = 128;
907	int res = 0;	1196	int dentoffs = 0;
		1197	ino_t inode_bits = 0;
		1198
		1199	req->result = -1;
		1200
		1201	if (!(flags & EIO_READDIR_DENTS))
		1202	flags &= ~(EIO_READDIR_DIRS_FIRST \| EIO_READDIR_STAT_ORDER);
908		1203
909	X_LOCK (wrklock);	1204	X_LOCK (wrklock);
910	/* the corresponding closedir is in ETP_WORKER_CLEAR */	1205	/* the corresponding closedir is in ETP_WORKER_CLEAR */
911	self->dirp = dirp = opendir (req->ptr1);	1206	self->dirp = dirp = opendir (req->ptr1);
912	req->flags \|= EIO_FLAG_PTR2_FREE;	1207	req->flags \|= EIO_FLAG_PTR1_FREE \| EIO_FLAG_PTR2_FREE;
		1208	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
913	req->ptr2 = names = malloc (memlen);	1209	req->ptr2 = names = malloc (namesalloc);
914	X_UNLOCK (wrklock);	1210	X_UNLOCK (wrklock);
915		1211
916	if (dirp && names)	1212	if (dirp && names && (!flags \|\| dents))
917	for (;;)	1213	for (;;)
918	{	1214	{
919	errno = 0;	1215	errno = 0;
920	entp = readdir (dirp);	1216	entp = readdir (dirp);
921		1217
922	if (!entp)	1218	if (!entp)
		1219	{
		1220	if (errno)
		1221	break;
		1222
		1223	/* sort etc. */
		1224	req->int1 = flags;
		1225	req->result = dentoffs;
		1226
		1227	if (flags & EIO_READDIR_STAT_ORDER)
		1228	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1229	else if (flags & EIO_READDIR_DIRS_FIRST)
		1230	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1231	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1232	else
		1233	{
		1234	/* in this case, all is known, and we just put dirs first and sort them */
		1235	eio_dirent *oth = dents + dentoffs;
		1236	eio_dirent *dir = dents;
		1237
		1238	/* now partition dirs to the front, and non-dirs to the back */
		1239	/* by walking from both sides and swapping if necessary */
		1240	/* also clear score, so it doesn't influence sorting */
		1241	while (oth > dir)
		1242	{
		1243	if (dir->type == EIO_DT_DIR)
		1244	++dir;
		1245	else if ((--oth)->type == EIO_DT_DIR)
		1246	{
		1247	eio_dirent tmp = dir; dir = oth; oth = tmp;
		1248
		1249	++dir;
		1250	}
		1251	}
		1252
		1253	/* now sort the dirs only */
		1254	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1255	}
		1256
923	break;	1257	break;
		1258	}
924		1259
		1260	/* now add the entry to our list(s) */
925	name = entp->d_name;	1261	name = entp->d_name;
926		1262
		1263	/* skip . and .. entries */
927	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))	1264	if (name [0] != '.' \|\| (name [1] && (name [1] != '.' \|\| name [2])))
928	{	1265	{
929	int len = strlen (name) + 1;	1266	int len = D_NAMLEN (entp) + 1;
930		1267
931	res++;	1268	while (expect_false (namesoffs + len > namesalloc))
932
933	while (memofs + len > memlen)
934	{	1269	{
935	memlen *= 2;	1270	namesalloc *= 2;
936	X_LOCK (wrklock);	1271	X_LOCK (wrklock);
937	req->ptr2 = names = realloc (names, memlen);	1272	req->ptr2 = names = realloc (names, namesalloc);
938	X_UNLOCK (wrklock);	1273	X_UNLOCK (wrklock);
939		1274
940	if (!names)	1275	if (!names)
941	break;	1276	break;
942	}	1277	}
943		1278
944	memcpy (names + memofs, name, len);	1279	memcpy (names + namesoffs, name, len);
		1280
		1281	if (dents)
		1282	{
		1283	struct eio_dirent *ent;
		1284
		1285	if (expect_false (dentoffs == dentalloc))
		1286	{
		1287	dentalloc *= 2;
		1288	X_LOCK (wrklock);
		1289	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1290	X_UNLOCK (wrklock);
		1291
		1292	if (!dents)
		1293	break;
		1294	}
		1295
		1296	ent = dents + dentoffs;
		1297
		1298	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1299	ent->namelen = len - 1;
		1300	ent->inode = D_INO (entp);
		1301
		1302	inode_bits \|= ent->inode;
		1303
		1304	switch (D_TYPE (entp))
		1305	{
		1306	default:
		1307	ent->type = EIO_DT_UNKNOWN;
		1308	flags \|= EIO_READDIR_FOUND_UNKNOWN;
		1309	break;
		1310
		1311	#ifdef DT_FIFO
		1312	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1313	#endif
		1314	#ifdef DT_CHR
		1315	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1316	#endif
		1317	#ifdef DT_MPC
		1318	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1319	#endif
		1320	#ifdef DT_DIR
		1321	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1322	#endif
		1323	#ifdef DT_NAM
		1324	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1325	#endif
		1326	#ifdef DT_BLK
		1327	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1328	#endif
		1329	#ifdef DT_MPB
		1330	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1331	#endif
		1332	#ifdef DT_REG
		1333	case DT_REG: ent->type = EIO_DT_REG; break;
		1334	#endif
		1335	#ifdef DT_NWK
		1336	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1337	#endif
		1338	#ifdef DT_CMP
		1339	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1340	#endif
		1341	#ifdef DT_LNK
		1342	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1343	#endif
		1344	#ifdef DT_SOCK
		1345	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1346	#endif
		1347	#ifdef DT_DOOR
		1348	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1349	#endif
		1350	#ifdef DT_WHT
		1351	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1352	#endif
		1353	}
		1354
		1355	ent->score = 7;
		1356
		1357	if (flags & EIO_READDIR_DIRS_FIRST)
		1358	{
		1359	if (ent->type == EIO_DT_UNKNOWN)
		1360	{
		1361	if (name == '.') / leading dots are likely directories, and, in any case, rare */
		1362	ent->score = 1;
		1363	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1364	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1365	}
		1366	else if (ent->type == EIO_DT_DIR)
		1367	ent->score = 0;
		1368	}
		1369	}
		1370
945	memofs += len;	1371	namesoffs += len;
		1372	++dentoffs;
		1373	}
		1374
		1375	if (EIO_CANCELLED (req))
		1376	{
		1377	errno = ECANCELED;
		1378	break;
946	}	1379	}
947	}	1380	}
		1381	}
948		1382
949	if (errno)	1383	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
950	res = -1;	1384	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
951		1385	#else
952	req->result = res;	1386
		1387	int
		1388	eio__msync (void *mem, size_t len, int flags)
		1389	{
		1390	if (EIO_MS_ASYNC != MS_SYNC
		1391	\|\| EIO_MS_INVALIDATE != MS_INVALIDATE
		1392	\|\| EIO_MS_SYNC != MS_SYNC)
		1393	{
		1394	flags = 0
		1395	\| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1396	\| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1397	\| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1398	}
		1399
		1400	return msync (mem, len, flags);
		1401	}
		1402
		1403	#endif
		1404
		1405	int
		1406	eio__mtouch (void *mem, size_t len, int flags)
		1407	{
		1408	intptr_t addr = (intptr_t)mem;
		1409	intptr_t end = addr + len;
		1410	#ifdef PAGESIZE
		1411	const intptr_t page = PAGESIZE;
		1412	#else
		1413	static intptr_t page;
		1414
		1415	if (!page)
		1416	page = sysconf (_SC_PAGESIZE);
		1417	#endif
		1418
		1419	/* round down to start of page, although this is probably useless */
		1420	addr &= ~(page - 1); /* assume page size is always a power of two */
		1421
		1422	if (addr < end)
		1423	if (flags & EIO_MT_MODIFY) /* modify */
		1424	do { ((volatile sig_atomic_t )addr) \|= 0; } while ((addr += page) < len);
		1425	else
		1426	do { ((volatile sig_atomic_t )addr) ; } while ((addr += page) < len);
		1427
		1428	return 0;
953	}	1429	}
954		1430
955	/*****************************************************************************/	1431	/*****************************************************************************/
956		1432
957	#define ALLOC(len) \	1433	#define ALLOC(len) \
…		…
1124	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;	1600	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1125		1601
1126	case EIO_SYNC: req->result = 0; sync (); break;	1602	case EIO_SYNC: req->result = 0; sync (); break;
1127	case EIO_FSYNC: req->result = fsync (req->int1); break;	1603	case EIO_FSYNC: req->result = fsync (req->int1); break;
1128	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;	1604	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1605	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1606	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1607	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1129		1608
1130	case EIO_READDIR: eio__scandir (req, self); break;	1609	case EIO_READDIR: eio__scandir (req, self); break;
1131		1610
1132	case EIO_BUSY:	1611	case EIO_BUSY:
1133	#ifdef _WIN32	1612	#ifdef _WIN32
…		…
1165		1644
1166	req->result = req->type == EIO_FUTIME	1645	req->result = req->type == EIO_FUTIME
1167	? futimes (req->int1, times)	1646	? futimes (req->int1, times)
1168	: utimes (req->ptr1, times);	1647	: utimes (req->ptr1, times);
1169	}	1648	}
		1649	break;
1170		1650
1171	case EIO_GROUP:	1651	case EIO_GROUP:
		1652	abort (); /* handled in eio_request */
		1653
1172	case EIO_NOP:	1654	case EIO_NOP:
1173	req->result = 0;	1655	req->result = 0;
1174	break;	1656	break;
1175		1657
1176	case EIO_CUSTOM:	1658	case EIO_CUSTOM:
1177	req->feed (req);	1659	((void ()(eio_req ))req->feed) (req);
1178	break;	1660	break;
1179		1661
1180	default:	1662	default:
1181	req->result = -1;	1663	req->result = -1;
1182	break;	1664	break;
…		…
1205	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)	1687	eio_req eio_fsync (int fd, int pri, eio_cb cb, void data)
1206	{	1688	{
1207	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1689	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1208	}	1690	}
1209		1691
		1692	eio_req eio_msync (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1693	{
		1694	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1695	}
		1696
		1697	eio_req eio_mtouch (void addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1698	{
		1699	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1700	}
		1701
		1702	eio_req eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void data)
		1703	{
		1704	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1705	}
		1706
1210	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)	1707	eio_req eio_fdatasync (int fd, int pri, eio_cb cb, void data)
1211	{	1708	{
1212	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1709	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1213	}	1710	}
1214		1711
…		…
1326	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)	1823	eio_req eio_rmdir (const char path, int pri, eio_cb cb, void *data)
1327	{	1824	{
1328	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1825	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1329	}	1826	}
1330		1827
1331	eio_req eio_readdir (const char path, int pri, eio_cb cb, void *data)	1828	eio_req eio_readdir (const char path, int flags, int pri, eio_cb cb, void *data)
1332	{	1829	{
1333	return eio__1path (EIO_READDIR, path, pri, cb, data);	1830	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1334	}	1831	}
1335		1832
1336	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1833	eio_req eio_mknod (const char path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1337	{	1834	{
1338	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;	1835	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
…		…
1369	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1866	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1370	}	1867	}
1371		1868
1372	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)	1869	eio_req eio_custom (eio_cb execute, int pri, eio_cb cb, void data)
1373	{	1870	{
1374	REQ (EIO_CUSTOM); req->feed = execute; SEND;	1871	REQ (EIO_CUSTOM); req->feed = (void ()(eio_req ))execute; SEND;
1375	}	1872	}
1376		1873
1377	#endif	1874	#endif
1378		1875
1379	eio_req eio_grp (eio_cb cb, void data)	1876	eio_req eio_grp (eio_cb cb, void data)
…		…
1407		1904
1408	void eio_grp_add (eio_req grp, eio_req req)	1905	void eio_grp_add (eio_req grp, eio_req req)
1409	{	1906	{
1410	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	1907	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1411		1908
		1909	grp->flags \|= EIO_FLAG_GROUPADD;
		1910
1412	++grp->size;	1911	++grp->size;
1413	req->grp = grp;	1912	req->grp = grp;
1414		1913
1415	req->grp_prev = 0;	1914	req->grp_prev = 0;
1416	req->grp_next = grp->grp_first;	1915	req->grp_next = grp->grp_first;
…		…
1425	/* misc garbage */	1924	/* misc garbage */
1426		1925
1427	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	1926	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1428	{	1927	{
1429	etp_worker wrk;	1928	etp_worker wrk;
		1929	ssize_t ret;
1430		1930
1431	wrk.dbuf = 0;	1931	wrk.dbuf = 0;
1432		1932
1433	eio__sendfile (ofd, ifd, offset, count, &wrk);	1933	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1434		1934
1435	if (wrk.dbuf)	1935	if (wrk.dbuf)
1436	free (wrk.dbuf);	1936	free (wrk.dbuf);
1437	}
1438		1937
		1938	return ret;
		1939	}
		1940

Diff Legend

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

Comparing cvsroot/libeio/eio.c (file contents): Revision 1.22 by root, Fri Jul 11 10:54:50 2008 UTC vs. Revision 1.50 by root, Thu Jan 7 17:18:08 2010 UTC

Diff Legend

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.22 by root, Fri Jul 11 10:54:50 2008 UTC vs.
Revision 1.50 by root, Thu Jan 7 17:18:08 2010 UTC