ViewVC Help

View File | Revision Log | Show Annotations | Download File

/cvs/libeio/eio.c

Comparing libeio/eio.c (file contents):
Revision 1.13 by root, Tue May 13 19:34:11 2008 UTC vs.
Revision 1.56 by root, Sun Sep 12 03:36:28 2010 UTC

		1	/*
		2	* libeio implementation
		3	*
		4	* Copyright (c) 2007,2008,2009,2010 Marc Alexander Lehmann <libeio@schmorp.de>
		5	* All rights reserved.
		6	*
		7	* Redistribution and use in source and binary forms, with or without modifica-
		8	* tion, are permitted provided that the following conditions are met:
		9	*
		10	* 1. Redistributions of source code must retain the above copyright notice,
		11	* this list of conditions and the following disclaimer.
		12	*
		13	* 2. Redistributions in binary form must reproduce the above copyright
		14	* notice, this list of conditions and the following disclaimer in the
		15	* documentation and/or other materials provided with the distribution.
		16	*
		17	* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
		18	* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
		19	* CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
		20	* EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
		21	* CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
		22	* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
		23	* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
		24	* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
		25	* ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
		26	* OF THE POSSIBILITY OF SUCH DAMAGE.
		27	*
		28	* Alternatively, the contents of this file may be used under the terms of
		29	* the GNU General Public License ("GPL") version 2 or any later version,
		30	* in which case the provisions of the GPL are applicable instead of
		31	* the above. If you wish to allow the use of your version of this file
		32	* only under the terms of the GPL and not to allow others to use your
		33	* version of this file under the BSD license, indicate your decision
		34	* by deleting the provisions above and replace them with the notice
		35	* and other provisions required by the GPL. If you do not delete the
		36	* provisions above, a recipient may use your version of this file under
		37	* either the BSD or the GPL.
		38	*/
		39
1	#include "eio.h"	40	#include "eio.h"
		41
		42	#ifdef EIO_STACKSIZE
		43	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		44	#endif
2	#include "xthread.h"	45	#include "xthread.h"
3		46
4	#include <errno.h>	47	#include <errno.h>
5	#include <stddef.h>	48	#include <stddef.h>
6	#include <stdlib.h>	49	#include <stdlib.h>
7	#include <string.h>	50	#include <string.h>
8	#include <errno.h>	51	#include <errno.h>
9	#include <sys/types.h>	52	#include <sys/types.h>
10	#include <sys/stat.h>	53	#include <sys/stat.h>
		54	#include <sys/statvfs.h>
11	#include <limits.h>	55	#include <limits.h>
12	#include <fcntl.h>	56	#include <fcntl.h>
13	#include <assert.h>	57	#include <assert.h>
14		58
15	#ifndef EIO_FINISH	59	#ifndef EIO_FINISH
…		…
25	#endif	69	#endif
26		70
27	#ifdef _WIN32	71	#ifdef _WIN32
28		72
29	/*doh*/	73	/*doh*/
30
31	#else	74	#else
32		75
33	# include "config.h"	76	# include "config.h"
34	# include <sys/time.h>	77	# include <sys/time.h>
35	# include <sys/select.h>	78	# include <sys/select.h>
36	# include <unistd.h>	79	# include <unistd.h>
37	# include <utime.h>	80	# include <utime.h>
38	# include <signal.h>	81	# include <signal.h>
39	# include <dirent.h>	82	# include <dirent.h>
40		83
		84	#if _POSIX_MEMLOCK || _POSIX_MAPPED_FILES
		85	# include <sys/mman.h>
		86	#endif
		87
		88	/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
		89	# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
		90	# define _DIRENT_HAVE_D_TYPE /* sigh */
		91	# define D_INO(de) (de)->d_fileno
		92	# define D_NAMLEN(de) (de)->d_namlen
		93	# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
		94	# define D_INO(de) (de)->d_ino
		95	# endif
		96
		97	#ifdef _D_EXACT_NAMLEN
		98	# undef D_NAMLEN
		99	# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
		100	#endif
		101
		102	# ifdef _DIRENT_HAVE_D_TYPE
		103	# define D_TYPE(de) (de)->d_type
		104	# endif
		105
41	# ifndef EIO_STRUCT_DIRENT	106	# ifndef EIO_STRUCT_DIRENT
42	# define EIO_STRUCT_DIRENT struct dirent	107	# define EIO_STRUCT_DIRENT struct dirent
43	# endif	108	# endif
44		109
45	#endif	110	#endif
46		111
47	#if HAVE_SENDFILE	112	#if HAVE_SENDFILE
48	# if __linux	113	# if __linux
49	# include <sys/sendfile.h>	114	# include <sys/sendfile.h>
50	# elif __freebsd	115	# elif __FreeBSD__ || defined __APPLE__
51	# include <sys/socket.h>	116	# include <sys/socket.h>
52	# include <sys/uio.h>	117	# include <sys/uio.h>
53	# elif __hpux	118	# elif __hpux
54	# include <sys/socket.h>	119	# include <sys/socket.h>
55	# elif __solaris /* not yet */	120	# elif __solaris
56	# include <sys/sendfile.h>	121	# include <sys/sendfile.h>
57	# else	122	# else
58	# error sendfile support requested but not available	123	# error sendfile support requested but not available
59	# endif	124	# endif
60	#endif	125	#endif
61		126
		127	#ifndef D_TYPE
		128	# define D_TYPE(de) 0
		129	#endif
		130	#ifndef D_INO
		131	# define D_INO(de) 0
		132	#endif
		133	#ifndef D_NAMLEN
		134	# define D_NAMLEN(de) strlen ((de)->d_name)
		135	#endif
		136
62	/* number of seconds after which an idle threads exit */	137	/* number of seconds after which an idle threads exit */
63	#define IDLE_TIMEOUT 10	138	#define IDLE_TIMEOUT 10
64		139
65	/* used for struct dirent, AIX doesn't provide it */	140	/* used for struct dirent, AIX doesn't provide it */
66	#ifndef NAME_MAX	141	#ifndef NAME_MAX
67	# define NAME_MAX 4096	142	# define NAME_MAX 4096
68	#endif	143	#endif
69		144
		145	/* used for readlink etc. */
		146	#ifndef PATH_MAX
		147	# define PATH_MAX 4096
		148	#endif
		149
70	/* buffer size for various temporary buffers */	150	/* buffer size for various temporary buffers */
71	#define EIO_BUFSIZE 65536	151	#define EIO_BUFSIZE 65536
72		152
73	#define dBUF \	153	#define dBUF \
74	char *eio_buf; \	154	char *eio_buf; \
75	X_LOCK (etplock); \	155	ETP_WORKER_LOCK (self); \
76	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \	156	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
77	X_UNLOCK (etplock); \	157	ETP_WORKER_UNLOCK (self); \
78	errno = ENOMEM; \	158	errno = ENOMEM; \
79	if (!eio_buf) \	159	if (!eio_buf) \
80	return -1;	160	return -1;
81		161
82	#define EIO_TICKS ((1000000 + 1023) >> 10)	162	#define EIO_TICKS ((1000000 + 1023) >> 10)
83		163
84	/*****************************************************************************/	164	/*****************************************************************************/
85		165
		166	#if __GNUC__ >= 3
		167	# define expect(expr,value) __builtin_expect ((expr),(value))
		168	#else
		169	# define expect(expr,value) (expr)
		170	#endif
		171
		172	#define expect_false(expr) expect ((expr) != 0, 0)
		173	#define expect_true(expr) expect ((expr) != 0, 1)
		174
		175	/*****************************************************************************/
		176
		177	#define ETP_PRI_MIN EIO_PRI_MIN
		178	#define ETP_PRI_MAX EIO_PRI_MAX
		179
		180	struct etp_worker;
		181
		182	#define ETP_REQ eio_req
		183	#define ETP_DESTROY(req) eio_destroy (req)
		184	static int eio_finish (eio_req *req);
		185	#define ETP_FINISH(req) eio_finish (req)
		186	static void eio_execute (struct etp_worker *self, eio_req *req);
		187	#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
		188
		189	#define ETP_WORKER_CLEAR(req) \
		190	if (wrk->dbuf) \
		191	{ \
		192	free (wrk->dbuf); \
		193	wrk->dbuf = 0; \
		194	} \
		195	\
		196	if (wrk->dirp) \
		197	{ \
		198	closedir (wrk->dirp); \
		199	wrk->dirp = 0; \
		200	}
		201
		202	#define ETP_WORKER_COMMON \
		203	void *dbuf; \
		204	DIR *dirp;
		205
		206	/*****************************************************************************/
		207
		208	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
		209
86	/* calculcate time difference in ~1/EIO_TICKS of a second */	210	/* calculate time difference in ~1/EIO_TICKS of a second */
87	static int tvdiff (struct timeval *tv1, struct timeval *tv2)	211	static int tvdiff (struct timeval *tv1, struct timeval *tv2)
88	{	212	{
89	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS	213	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
90	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);	214	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
91	}	215	}
92		216
93	static unsigned int started, idle, wanted = 4;	217	static unsigned int started, idle, wanted = 4;
94		218
95	typedef struct etp_pool
96	{
97	void (*want_poll_cb) (void);	219	static void (*want_poll_cb) (void);
98	void (*done_poll_cb) (void);	220	static void (*done_poll_cb) (void);
99		221
100	unsigned int max_poll_time;	222	static unsigned int max_poll_time; /* reslock */
101	unsigned int max_poll_reqs;	223	static unsigned int max_poll_reqs; /* reslock */
102	} etp_pool;
103		224
104	static volatile unsigned int nreqs, nready, npending;	225	static volatile unsigned int nreqs; /* reqlock */
		226	static volatile unsigned int nready; /* reqlock */
		227	static volatile unsigned int npending; /* reqlock */
105	static volatile unsigned int max_idle = 4;	228	static volatile unsigned int max_idle = 4;
106		229
107	static mutex_t etplock = X_MUTEX_INIT;	230	static xmutex_t wrklock = X_MUTEX_INIT;
108	static mutex_t reslock = X_MUTEX_INIT;	231	static xmutex_t reslock = X_MUTEX_INIT;
109	static mutex_t reqlock = X_MUTEX_INIT;	232	static xmutex_t reqlock = X_MUTEX_INIT;
110	static cond_t reqwait = X_COND_INIT;	233	static xcond_t reqwait = X_COND_INIT;
111		234
		235	#if !HAVE_PREADWRITE
		236	/*
		237	* make our pread/pwrite emulation safe against themselves, but not against
		238	* normal read/write by using a mutex. slows down execution a lot,
		239	* but that's your problem, not mine.
		240	*/
		241	static xmutex_t preadwritelock = X_MUTEX_INIT;
		242	#endif
		243
112	typedef struct worker	244	typedef struct etp_worker
113	{	245	{
114	/* locked by etplock */	246	/* locked by wrklock */
115	struct worker *prev, *next;	247	struct etp_worker *prev, *next;
116		248
117	thread_t tid;	249	xthread_t tid;
118		250
119	/* locked by reslock, reqlock or etplock */	251	/* locked by reslock, reqlock or wrklock */
120	eio_req *req; /* currently processed request */	252	ETP_REQ *req; /* currently processed request */
121	void *dbuf;	253
122	DIR *dirp;	254	ETP_WORKER_COMMON
123	} worker;	255	} etp_worker;
124		256
125	static worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */	257	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		258
		259	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		260	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
126		261
127	/* worker threads management */	262	/* worker threads management */
128		263
129	static void worker_clear (worker *wrk)	264	static void etp_worker_clear (etp_worker *wrk)
130	{	265	{
131	if (wrk->dirp)	266	ETP_WORKER_CLEAR (wrk);
132	{
133	closedir (wrk->dirp);
134	wrk->dirp = 0;
135	}
136
137	if (wrk->dbuf)
138	{
139	free (wrk->dbuf);
140	wrk->dbuf = 0;
141	}
142	}	267	}
143		268
144	static void worker_free (worker *wrk)	269	static void etp_worker_free (etp_worker *wrk)
145	{	270	{
146	wrk->next->prev = wrk->prev;	271	wrk->next->prev = wrk->prev;
147	wrk->prev->next = wrk->next;	272	wrk->prev->next = wrk->next;
148		273
149	free (wrk);	274	free (wrk);
150	}	275	}
151		276
152	unsigned int eio_nreqs (void)	277	static unsigned int etp_nreqs (void)
153	{	278	{
		279	int retval;
		280	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		281	retval = nreqs;
		282	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
154	return nreqs;	283	return retval;
155	}	284	}
156		285
157	unsigned int eio_nready (void)	286	static unsigned int etp_nready (void)
158	{	287	{
159	unsigned int retval;	288	unsigned int retval;
160		289
161	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	290	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
162	retval = nready;	291	retval = nready;
163	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	292	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
164		293
165	return retval;	294	return retval;
166	}	295	}
167		296
168	unsigned int eio_npending (void)	297	static unsigned int etp_npending (void)
169	{	298	{
170	unsigned int retval;	299	unsigned int retval;
171		300
172	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	301	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
173	retval = npending;	302	retval = npending;
174	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	303	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
175		304
176	return retval;	305	return retval;
177	}	306	}
178		307
179	unsigned int eio_nthreads (void)	308	static unsigned int etp_nthreads (void)
180	{	309	{
181	unsigned int retval;	310	unsigned int retval;
182		311
183	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	312	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
184	retval = started;	313	retval = started;
…		…
191	* a somewhat faster data structure might be nice, but	320	* a somewhat faster data structure might be nice, but
192	* with 8 priorities this actually needs <20 insns	321	* with 8 priorities this actually needs <20 insns
193	* per shift, the most expensive operation.	322	* per shift, the most expensive operation.
194	*/	323	*/
195	typedef struct {	324	typedef struct {
196	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	325	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
197	int size;	326	int size;
198	} reqq;	327	} etp_reqq;
199		328
200	static reqq req_queue;	329	static etp_reqq req_queue;
201	static reqq res_queue;	330	static etp_reqq res_queue;
202		331
203	static int reqq_push (reqq *q, eio_req *req)	332	static int reqq_push (etp_reqq *q, ETP_REQ *req)
204	{	333	{
205	int pri = req->pri;	334	int pri = req->pri;
206	req->next = 0;	335	req->next = 0;
207		336
208	if (q->qe[pri])	337	if (q->qe[pri])
…		…
214	q->qe[pri] = q->qs[pri] = req;	343	q->qe[pri] = q->qs[pri] = req;
215		344
216	return q->size++;	345	return q->size++;
217	}	346	}
218		347
219	static eio_req *reqq_shift (reqq *q)	348	static ETP_REQ *reqq_shift (etp_reqq *q)
220	{	349	{
221	int pri;	350	int pri;
222		351
223	if (!q->size)	352	if (!q->size)
224	return 0;	353	return 0;
225		354
226	--q->size;	355	--q->size;
227		356
228	for (pri = EIO_NUM_PRI; pri--; )	357	for (pri = ETP_NUM_PRI; pri--; )
229	{	358	{
230	eio_req *req = q->qs[pri];	359	eio_req *req = q->qs[pri];
231		360
232	if (req)	361	if (req)
233	{	362	{
…		…
241	abort ();	370	abort ();
242	}	371	}
243		372
244	static void etp_atfork_prepare (void)	373	static void etp_atfork_prepare (void)
245	{	374	{
246	X_LOCK (etplock);	375	X_LOCK (wrklock);
247	X_LOCK (reqlock);	376	X_LOCK (reqlock);
248	X_LOCK (reslock);	377	X_LOCK (reslock);
249	#if !HAVE_PREADWRITE	378	#if !HAVE_PREADWRITE
250	X_LOCK (preadwritelock);	379	X_LOCK (preadwritelock);
251	#endif	380	#endif
252	#if !HAVE_READDIR_R
253	X_LOCK (readdirlock);
254	#endif
255	}	381	}
256		382
257	static void etp_atfork_parent (void)	383	static void etp_atfork_parent (void)
258	{	384	{
259	#if !HAVE_READDIR_R
260	X_UNLOCK (readdirlock);
261	#endif
262	#if !HAVE_PREADWRITE	385	#if !HAVE_PREADWRITE
263	X_UNLOCK (preadwritelock);	386	X_UNLOCK (preadwritelock);
264	#endif	387	#endif
265	X_UNLOCK (reslock);	388	X_UNLOCK (reslock);
266	X_UNLOCK (reqlock);	389	X_UNLOCK (reqlock);
267	X_UNLOCK (etplock);	390	X_UNLOCK (wrklock);
268	}	391	}
269		392
270	static void etp_atfork_child (void)	393	static void etp_atfork_child (void)
271	{	394	{
272	eio_req *prv;	395	ETP_REQ *prv;
273		396
274	while (prv = reqq_shift (&req_queue))	397	while ((prv = reqq_shift (&req_queue)))
275	eio_destroy (prv);	398	ETP_DESTROY (prv);
276		399
277	while (prv = reqq_shift (&res_queue))	400	while ((prv = reqq_shift (&res_queue)))
278	eio_destroy (prv);	401	ETP_DESTROY (prv);
279		402
280	while (wrk_first.next != &wrk_first)	403	while (wrk_first.next != &wrk_first)
281	{	404	{
282	worker *wrk = wrk_first.next;	405	etp_worker *wrk = wrk_first.next;
283		406
284	if (wrk->req)	407	if (wrk->req)
285	eio_destroy (wrk->req);	408	ETP_DESTROY (wrk->req);
286		409
287	worker_clear (wrk);	410	etp_worker_clear (wrk);
288	worker_free (wrk);	411	etp_worker_free (wrk);
289	}	412	}
290		413
291	started = 0;	414	started = 0;
292	idle = 0;	415	idle = 0;
293	nreqs = 0;	416	nreqs = 0;
…		…
302	{	425	{
303	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);	426	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
304	}	427	}
305		428
306	static int	429	static int
307	etp_init (etp_pool *etp, void (*want_poll)(void), void (*done_poll)(void))	430	etp_init (void (*want_poll)(void), void (*done_poll)(void))
308	{	431	{
309	static pthread_once_t doinit = PTHREAD_ONCE_INIT;	432	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
310		433
311	pthread_once (&doinit, etp_once_init);	434	pthread_once (&doinit, etp_once_init);
312		435
313	memset (etp, 0, sizeof *etp);
314
315	etp->want_poll_cb = want_poll;	436	want_poll_cb = want_poll;
316	etp->done_poll_cb = done_poll;	437	done_poll_cb = done_poll;
317	}
318		438
319	static etp_pool etp;
320
321	/*****************************************************************************/
322
323	static void grp_try_feed (eio_req *grp)
324	{
325	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
326	{
327	int old_len = grp->size;
328
329	EIO_FEED (grp);
330
331	/* stop if no progress has been made */
332	if (old_len == grp->size)
333	{
334	grp->feed = 0;
335	break;
336	}
337	}
338	}
339
340	static int eio_finish (eio_req *req);
341
342	static int grp_dec (eio_req *grp)
343	{
344	--grp->size;
345
346	/* call feeder, if applicable */
347	grp_try_feed (grp);
348
349	/* finish, if done */
350	if (!grp->size && grp->int1)
351	return eio_finish (grp);
352	else
353	return 0;	439	return 0;
354	}	440	}
355		441
356	void eio_destroy (eio_req *req)
357	{
358	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
359	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
360
361	EIO_DESTROY (req);
362	}
363
364	static int eio_finish (eio_req *req)
365	{
366	int res = EIO_FINISH (req);
367
368	if (req->grp)
369	{
370	int res2;
371	eio_req *grp = req->grp;
372
373	/* unlink request */
374	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
375	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
376
377	if (grp->grp_first == req)
378	grp->grp_first = req->grp_next;
379
380	res2 = grp_dec (grp);
381
382	if (!res && res2)
383	res = res2;
384	}
385
386	eio_destroy (req);
387
388	return res;
389	}
390
391	void eio_grp_cancel (eio_req *grp)
392	{
393	for (grp = grp->grp_first; grp; grp = grp->grp_next)
394	eio_cancel (grp);
395	}
396
397	void eio_cancel (eio_req *req)
398	{
399	X_LOCK (etplock);
400	req->flags |= EIO_FLAG_CANCELLED;
401	X_UNLOCK (etplock);
402
403	eio_grp_cancel (req);
404	}
405
406	X_THREAD_PROC (eio_proc);	442	X_THREAD_PROC (etp_proc);
407		443
408	static void start_thread (void)	444	static void etp_start_thread (void)
409	{	445	{
410	worker *wrk = calloc (1, sizeof (worker));	446	etp_worker *wrk = calloc (1, sizeof (etp_worker));
411		447
412	/*TODO*/	448	/*TODO*/
413	assert (("unable to allocate worker thread data", wrk));	449	assert (("unable to allocate worker thread data", wrk));
414		450
415	X_LOCK (etplock);	451	X_LOCK (wrklock);
416		452
417	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	453	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
418	{	454	{
419	wrk->prev = &wrk_first;	455	wrk->prev = &wrk_first;
420	wrk->next = wrk_first.next;	456	wrk->next = wrk_first.next;
421	wrk_first.next->prev = wrk;	457	wrk_first.next->prev = wrk;
422	wrk_first.next = wrk;	458	wrk_first.next = wrk;
423	++started;	459	++started;
424	}	460	}
425	else	461	else
426	free (wrk);	462	free (wrk);
427		463
428	X_UNLOCK (etplock);	464	X_UNLOCK (wrklock);
429	}	465	}
430		466
431	static void maybe_start_thread (void)	467	static void etp_maybe_start_thread (void)
432	{	468	{
433	if (eio_nthreads () >= wanted)	469	if (expect_true (etp_nthreads () >= wanted))
434	return;	470	return;
435		471
436	/* todo: maybe use idle here, but might be less exact */	472	/* todo: maybe use idle here, but might be less exact */
437	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	473	if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
438	return;	474	return;
439		475
440	start_thread ();	476	etp_start_thread ();
441	}	477	}
442		478
443	void eio_submit (eio_req *req)
444	{
445	req->pri += EIO_PRI_BIAS;
446
447	if (req->pri < EIO_PRI_MIN + EIO_PRI_BIAS) req->pri = EIO_PRI_MIN + EIO_PRI_BIAS;
448	if (req->pri > EIO_PRI_MAX + EIO_PRI_BIAS) req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;
449
450	++nreqs;
451
452	X_LOCK (reqlock);
453	++nready;
454	reqq_push (&req_queue, req);
455	X_COND_SIGNAL (reqwait);
456	X_UNLOCK (reqlock);
457
458	maybe_start_thread ();
459	}
460
461	static void end_thread (void)	479	static void etp_end_thread (void)
462	{	480	{
463	eio_req *req = calloc (1, sizeof (eio_req));	481	eio_req *req = calloc (1, sizeof (eio_req));
464		482
465	req->type = EIO_QUIT;	483	req->type = -1;
466	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	484	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
467		485
468	X_LOCK (reqlock);	486	X_LOCK (reqlock);
469	reqq_push (&req_queue, req);	487	reqq_push (&req_queue, req);
470	X_COND_SIGNAL (reqwait);	488	X_COND_SIGNAL (reqwait);
471	X_UNLOCK (reqlock);	489	X_UNLOCK (reqlock);
472		490
473	X_LOCK (etplock);	491	X_LOCK (wrklock);
474	--started;	492	--started;
475	X_UNLOCK (etplock);	493	X_UNLOCK (wrklock);
476	}	494	}
477		495
478	void eio_set_max_poll_time (double nseconds)
479	{
480	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
481	etp.max_poll_time = nseconds;
482	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
483	}
484
485	void eio_set_max_poll_reqs (unsigned int maxreqs)
486	{
487	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
488	etp.max_poll_reqs = maxreqs;
489	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
490	}
491
492	void eio_set_max_idle (unsigned int nthreads)
493	{
494	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
495	max_idle = nthreads <= 0 ? 1 : nthreads;
496	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
497	}
498
499	void eio_set_min_parallel (unsigned int nthreads)
500	{
501	if (wanted < nthreads)
502	wanted = nthreads;
503	}
504
505	void eio_set_max_parallel (unsigned int nthreads)
506	{
507	if (wanted > nthreads)
508	wanted = nthreads;
509
510	while (started > wanted)
511	end_thread ();
512	}
513
514	int eio_poll (void)	496	static int etp_poll (void)
515	{	497	{
516	int maxreqs = etp.max_poll_reqs;	498	unsigned int maxreqs;
		499	unsigned int maxtime;
517	struct timeval tv_start, tv_now;	500	struct timeval tv_start, tv_now;
518	eio_req *req;
519		501
520	if (etp.max_poll_time)	502	X_LOCK (reslock);
		503	maxreqs = max_poll_reqs;
		504	maxtime = max_poll_time;
		505	X_UNLOCK (reslock);
		506
		507	if (maxtime)
521	gettimeofday (&tv_start, 0);	508	gettimeofday (&tv_start, 0);
522		509
523	for (;;)	510	for (;;)
524	{	511	{
		512	ETP_REQ *req;
		513
525	maybe_start_thread ();	514	etp_maybe_start_thread ();
526		515
527	X_LOCK (reslock);	516	X_LOCK (reslock);
528	req = reqq_shift (&res_queue);	517	req = reqq_shift (&res_queue);
529		518
530	if (req)	519	if (req)
531	{	520	{
532	--npending;	521	--npending;
533		522
534	if (!res_queue.size && etp.done_poll_cb)	523	if (!res_queue.size && done_poll_cb)
535	etp.done_poll_cb ();	524	done_poll_cb ();
536	}	525	}
537		526
538	X_UNLOCK (reslock);	527	X_UNLOCK (reslock);
539		528
540	if (!req)	529	if (!req)
541	return 0;	530	return 0;
542		531
		532	X_LOCK (reqlock);
543	--nreqs;	533	--nreqs;
		534	X_UNLOCK (reqlock);
544		535
545	if (req->type == EIO_GROUP && req->size)	536	if (expect_false (req->type == EIO_GROUP && req->size))
546	{	537	{
547	req->int1 = 1; /* mark request as delayed */	538	req->int1 = 1; /* mark request as delayed */
548	continue;	539	continue;
549	}	540	}
550	else	541	else
551	{	542	{
552	int res = eio_finish (req);	543	int res = ETP_FINISH (req);
553	if (res)	544	if (expect_false (res))
554	return res;	545	return res;
555	}	546	}
556		547
557	if (maxreqs && !--maxreqs)	548	if (expect_false (maxreqs && !--maxreqs))
558	break;	549	break;
559		550
560	if (etp.max_poll_time)	551	if (maxtime)
561	{	552	{
562	gettimeofday (&tv_now, 0);	553	gettimeofday (&tv_now, 0);
563		554
564	if (tvdiff (&tv_start, &tv_now) >= etp.max_poll_time)	555	if (tvdiff (&tv_start, &tv_now) >= maxtime)
565	break;	556	break;
566	}	557	}
567	}	558	}
568		559
569	errno = EAGAIN;	560	errno = EAGAIN;
570	return -1;	561	return -1;
571	}	562	}
572		563
		564	static void etp_cancel (ETP_REQ *req)
		565	{
		566	X_LOCK (wrklock);
		567	req->flags |= EIO_FLAG_CANCELLED;
		568	X_UNLOCK (wrklock);
		569
		570	eio_grp_cancel (req);
		571	}
		572
		573	static void etp_submit (ETP_REQ *req)
		574	{
		575	req->pri -= ETP_PRI_MIN;
		576
		577	if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		578	if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		579
		580	if (expect_false (req->type == EIO_GROUP))
		581	{
		582	/* I hope this is worth it :/ */
		583	X_LOCK (reqlock);
		584	++nreqs;
		585	X_UNLOCK (reqlock);
		586
		587	X_LOCK (reslock);
		588
		589	++npending;
		590
		591	if (!reqq_push (&res_queue, req) && want_poll_cb)
		592	want_poll_cb ();
		593
		594	X_UNLOCK (reslock);
		595	}
		596	else
		597	{
		598	X_LOCK (reqlock);
		599	++nreqs;
		600	++nready;
		601	reqq_push (&req_queue, req);
		602	X_COND_SIGNAL (reqwait);
		603	X_UNLOCK (reqlock);
		604
		605	etp_maybe_start_thread ();
		606	}
		607	}
		608
		609	static void etp_set_max_poll_time (double nseconds)
		610	{
		611	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		612	max_poll_time = nseconds * EIO_TICKS;
		613	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		614	}
		615
		616	static void etp_set_max_poll_reqs (unsigned int maxreqs)
		617	{
		618	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		619	max_poll_reqs = maxreqs;
		620	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		621	}
		622
		623	static void etp_set_max_idle (unsigned int nthreads)
		624	{
		625	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		626	max_idle = nthreads <= 0 ? 1 : nthreads;
		627	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		628	}
		629
		630	static void etp_set_min_parallel (unsigned int nthreads)
		631	{
		632	if (wanted < nthreads)
		633	wanted = nthreads;
		634	}
		635
		636	static void etp_set_max_parallel (unsigned int nthreads)
		637	{
		638	if (wanted > nthreads)
		639	wanted = nthreads;
		640
		641	while (started > wanted)
		642	etp_end_thread ();
		643	}
		644
		645	/*****************************************************************************/
		646
		647	static void grp_try_feed (eio_req *grp)
		648	{
		649	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		650	{
		651	grp->flags &= ~EIO_FLAG_GROUPADD;
		652
		653	EIO_FEED (grp);
		654
		655	/* stop if no progress has been made */
		656	if (!(grp->flags & EIO_FLAG_GROUPADD))
		657	{
		658	grp->feed = 0;
		659	break;
		660	}
		661	}
		662	}
		663
		664	static int grp_dec (eio_req *grp)
		665	{
		666	--grp->size;
		667
		668	/* call feeder, if applicable */
		669	grp_try_feed (grp);
		670
		671	/* finish, if done */
		672	if (!grp->size && grp->int1)
		673	return eio_finish (grp);
		674	else
		675	return 0;
		676	}
		677
		678	void eio_destroy (eio_req *req)
		679	{
		680	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		681	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		682
		683	EIO_DESTROY (req);
		684	}
		685
		686	static int eio_finish (eio_req *req)
		687	{
		688	int res = EIO_FINISH (req);
		689
		690	if (req->grp)
		691	{
		692	int res2;
		693	eio_req *grp = req->grp;
		694
		695	/* unlink request */
		696	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		697	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		698
		699	if (grp->grp_first == req)
		700	grp->grp_first = req->grp_next;
		701
		702	res2 = grp_dec (grp);
		703
		704	if (!res && res2)
		705	res = res2;
		706	}
		707
		708	eio_destroy (req);
		709
		710	return res;
		711	}
		712
		713	void eio_grp_cancel (eio_req *grp)
		714	{
		715	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		716	eio_cancel (grp);
		717	}
		718
		719	void eio_cancel (eio_req *req)
		720	{
		721	etp_cancel (req);
		722	}
		723
		724	void eio_submit (eio_req *req)
		725	{
		726	etp_submit (req);
		727	}
		728
		729	unsigned int eio_nreqs (void)
		730	{
		731	return etp_nreqs ();
		732	}
		733
		734	unsigned int eio_nready (void)
		735	{
		736	return etp_nready ();
		737	}
		738
		739	unsigned int eio_npending (void)
		740	{
		741	return etp_npending ();
		742	}
		743
		744	unsigned int eio_nthreads (void)
		745	{
		746	return etp_nthreads ();
		747	}
		748
		749	void eio_set_max_poll_time (double nseconds)
		750	{
		751	etp_set_max_poll_time (nseconds);
		752	}
		753
		754	void eio_set_max_poll_reqs (unsigned int maxreqs)
		755	{
		756	etp_set_max_poll_reqs (maxreqs);
		757	}
		758
		759	void eio_set_max_idle (unsigned int nthreads)
		760	{
		761	etp_set_max_idle (nthreads);
		762	}
		763
		764	void eio_set_min_parallel (unsigned int nthreads)
		765	{
		766	etp_set_min_parallel (nthreads);
		767	}
		768
		769	void eio_set_max_parallel (unsigned int nthreads)
		770	{
		771	etp_set_max_parallel (nthreads);
		772	}
		773
		774	int eio_poll (void)
		775	{
		776	return etp_poll ();
		777	}
		778
573	/*****************************************************************************/	779	/*****************************************************************************/
574	/* work around various missing functions */	780	/* work around various missing functions */
575		781
576	#if !HAVE_PREADWRITE	782	#if !HAVE_PREADWRITE
		783	# undef pread
		784	# undef pwrite
577	# define pread eio__pread	785	# define pread eio__pread
578	# define pwrite eio__pwrite	786	# define pwrite eio__pwrite
579
580	/*
581	* make our pread/pwrite safe against themselves, but not against
582	* normal read/write by using a mutex. slows down execution a lot,
583	* but that's your problem, not mine.
584	*/
585	static mutex_t preadwritelock = X_MUTEX_INIT;
586		787
587	static ssize_t	788	static ssize_t
588	eio__pread (int fd, void *buf, size_t count, off_t offset)	789	eio__pread (int fd, void *buf, size_t count, off_t offset)
589	{	790	{
590	ssize_t res;	791	ssize_t res;
…		…
608		809
609	X_LOCK (preadwritelock);	810	X_LOCK (preadwritelock);
610	ooffset = lseek (fd, 0, SEEK_CUR);	811	ooffset = lseek (fd, 0, SEEK_CUR);
611	lseek (fd, offset, SEEK_SET);	812	lseek (fd, offset, SEEK_SET);
612	res = write (fd, buf, count);	813	res = write (fd, buf, count);
613	lseek (fd, offset, SEEK_SET);	814	lseek (fd, ooffset, SEEK_SET);
614	X_UNLOCK (preadwritelock);	815	X_UNLOCK (preadwritelock);
615		816
616	return res;	817	return res;
617	}	818	}
618	#endif	819	#endif
619		820
620	#ifndef HAVE_FUTIMES	821	#ifndef HAVE_FUTIMES
621		822
		823	# undef utimes
		824	# undef futimes
622	# define utimes(path,times) eio__utimes (path, times)	825	# define utimes(path,times) eio__utimes (path, times)
623	# define futimes(fd,times) eio__futimes (fd, times)	826	# define futimes(fd,times) eio__futimes (fd, times)
624		827
625	static int	828	static int
626	eio__utimes (const char *filename, const struct timeval times[2])	829	eio__utimes (const char *filename, const struct timeval times[2])
…		…
645	}	848	}
646		849
647	#endif	850	#endif
648		851
649	#if !HAVE_FDATASYNC	852	#if !HAVE_FDATASYNC
		853	# undef fdatasync
650	# define fdatasync fsync	854	# define fdatasync(fd) fsync (fd)
651	#endif	855	#endif
		856
		857	/* sync_file_range always needs emulation */
		858	int
		859	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		860	{
		861	#if HAVE_SYNC_FILE_RANGE
		862	int res;
		863
		864	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		865	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		866	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		867	{
		868	flags = 0
		869	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		870	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		871	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		872	}
		873
		874	res = sync_file_range (fd, offset, nbytes, flags);
		875
		876	if (!res || errno != ENOSYS)
		877	return res;
		878	#endif
		879
		880	/* even though we could play tricks with the flags, it's better to always
		881	* call fdatasync, as that matches the expectation of its users best */
		882	return fdatasync (fd);
		883	}
652		884
653	#if !HAVE_READAHEAD	885	#if !HAVE_READAHEAD
		886	# undef readahead
654	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	887	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
655		888
656	static ssize_t	889	static ssize_t
657	eio__readahead (int fd, off_t offset, size_t count, worker *self)	890	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
658	{	891	{
659	size_t todo = count;	892	size_t todo = count;
660	dBUF;	893	dBUF;
661		894
662	while (todo > 0)	895	while (todo > 0)
…		…
672	return count;	905	return count;
673	}	906	}
674		907
675	#endif	908	#endif
676		909
677	#if !HAVE_READDIR_R
678	# define readdir_r eio__readdir_r
679
680	static mutex_t readdirlock = X_MUTEX_INIT;
681
682	static int
683	eio__readdir_r (DIR *dirp, EIO_STRUCT_DIRENT *ent, EIO_STRUCT_DIRENT **res)
684	{
685	EIO_STRUCT_DIRENT *e;
686	int errorno;
687
688	X_LOCK (readdirlock);
689
690	e = readdir (dirp);
691	errorno = errno;
692
693	if (e)
694	{
695	*res = ent;
696	strcpy (ent->d_name, e->d_name);
697	}
698	else
699	*res = 0;
700
701	X_UNLOCK (readdirlock);
702
703	errno = errorno;
704	return e ? 0 : -1;
705	}
706	#endif
707
708	/* sendfile always needs emulation */	910	/* sendfile always needs emulation */
709	static ssize_t	911	static ssize_t
710	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	912	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
711	{	913	{
712	ssize_t res;	914	ssize_t res;
713		915
714	if (!count)	916	if (!count)
715	return 0;	917	return 0;
716		918
717	#if HAVE_SENDFILE	919	#if HAVE_SENDFILE
718	# if __linux	920	# if __linux
719	res = sendfile (ofd, ifd, &offset, count);	921	res = sendfile (ofd, ifd, &offset, count);
720		922
721	# elif __freebsd	923	# elif __FreeBSD__
722	/*	924	/*
723	* Of course, the freebsd sendfile is a dire hack with no thoughts	925	* Of course, the freebsd sendfile is a dire hack with no thoughts
724	* wasted on making it similar to other I/O functions.	926	* wasted on making it similar to other I/O functions.
725	*/	927	*/
726	{	928	{
727	off_t sbytes;	929	off_t sbytes;
728	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	930	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
729		931
730	if (res < 0 && sbytes)	932	#if 0 /* according to the manpage, this is correct, but broken behaviour */
731	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	933	/* freebsd' sendfile will return 0 on success */
		934	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		935	/* not on e.g. EIO or EPIPE - sounds broken */
		936	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
		937	res = sbytes;
		938	#endif
		939
		940	/* according to source inspection, this is correct, and useful behaviour */
		941	if (sbytes)
		942	res = sbytes;
		943	}
		944
		945	# elif defined (__APPLE__)
		946
		947	{
		948	off_t sbytes = count;
		949	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		950
		951	/* according to the manpage, sbytes is always valid */
		952	if (sbytes)
732	res = sbytes;	953	res = sbytes;
733	}	954	}
734		955
735	# elif __hpux	956	# elif __hpux
736	res = sendfile (ofd, ifd, offset, count, 0, 0);	957	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
750	if (res < 0 && sbytes)	971	if (res < 0 && sbytes)
751	res = sbytes;	972	res = sbytes;
752	}	973	}
753		974
754	# endif	975	# endif
		976
		977	#elif defined (_WIN32)
		978
		979	/* does not work, just for documentation of what would need to be done */
		980	{
		981	HANDLE h = TO_SOCKET (ifd);
		982	SetFilePointer (h, offset, 0, FILE_BEGIN);
		983	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		984	}
		985
755	#else	986	#else
756	res = -1;	987	res = -1;
757	errno = ENOSYS;	988	errno = ENOSYS;
758	#endif	989	#endif
759		990
760	if (res < 0	991	if (res < 0
761	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	992	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		993	/* BSDs */
		994	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		995	|| errno == ENOTSUP
		996	#endif
		997	|| errno == EOPNOTSUPP /* BSDs */
762	#if __solaris	998	#if __solaris
763	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	999	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
764	#endif	1000	#endif
765	)	1001	)
766	)	1002	)
…		…
797	}	1033	}
798		1034
799	return res;	1035	return res;
800	}	1036	}
801		1037
		1038	static signed char
		1039	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1040	{
		1041	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1042	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1043	}
		1044
		1045	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1046
		1047	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1048	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1049
		1050	static void
		1051	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1052	{
		1053	unsigned char bits [9 + sizeof (ino_t) * 8];
		1054	unsigned char *bit = bits;
		1055
		1056	assert (CHAR_BIT == 8);
		1057	assert (sizeof (eio_dirent) * 8 < 256);
		1058	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1059	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1060
		1061	if (size <= EIO_SORT_FAST)
		1062	return;
		1063
		1064	/* first prepare an array of bits to test in our radix sort */
		1065	/* try to take endianness into account, as well as differences in ino_t sizes */
		1066	/* inode_bits must contain all inodes ORed together */
		1067	/* which is used to skip bits that are 0 everywhere, which is very common */
		1068	{
		1069	ino_t endianness;
		1070	int i, j;
		1071
		1072	/* we store the byte offset of byte n into byte n of "endianness" */
		1073	for (i = 0; i < sizeof (ino_t); ++i)
		1074	((unsigned char *)&endianness)[i] = i;
		1075
		1076	*bit++ = 0;
		1077
		1078	for (i = 0; i < sizeof (ino_t); ++i)
		1079	{
		1080	/* shifting off the byte offsets out of "endianness" */
		1081	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1082	endianness >>= 8;
		1083
		1084	for (j = 0; j < 8; ++j)
		1085	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1086	*bit++ = offs + j;
		1087	}
		1088
		1089	for (j = 0; j < 8; ++j)
		1090	if (score_bits & (1 << j))
		1091	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1092	}
		1093
		1094	/* now actually do the sorting (a variant of MSD radix sort) */
		1095	{
		1096	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1097	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1098	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1099	int stk_idx = 0;
		1100
		1101	base_stk [stk_idx] = dents;
		1102	end_stk [stk_idx] = dents + size;
		1103	bit_stk [stk_idx] = bit - 1;
		1104
		1105	do
		1106	{
		1107	base = base_stk [stk_idx];
		1108	end = end_stk [stk_idx];
		1109	bit = bit_stk [stk_idx];
		1110
		1111	for (;;)
		1112	{
		1113	unsigned char O = *bit >> 3;
		1114	unsigned char M = 1 << (*bit & 7);
		1115
		1116	eio_dirent *a = base;
		1117	eio_dirent *b = end;
		1118
		1119	if (b - a < EIO_SORT_CUTOFF)
		1120	break;
		1121
		1122	/* now bit-partition the array on the bit */
		1123	/* this ugly asymmetric loop seems to perform much better than typical */
		1124	/* partition algos found in the literature */
		1125	do
		1126	if (!(((unsigned char *)a)[O] & M))
		1127	++a;
		1128	else if (!(((unsigned char *)--b)[O] & M))
		1129	{
		1130	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1131	++a;
		1132	}
		1133	while (b > a);
		1134
		1135	/* next bit, or stop, if no bits left in this path */
		1136	if (!*--bit)
		1137	break;
		1138
		1139	base_stk [stk_idx] = a;
		1140	end_stk [stk_idx] = end;
		1141	bit_stk [stk_idx] = bit;
		1142	++stk_idx;
		1143
		1144	end = a;
		1145	}
		1146	}
		1147	while (stk_idx--);
		1148	}
		1149	}
		1150
		1151	static void
		1152	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1153	{
		1154	/* first move the smallest element to the front, to act as a sentinel */
		1155	{
		1156	int i;
		1157	eio_dirent *min = dents;
		1158
		1159	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1160	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1161	if (EIO_DENT_CMP (dents [i], <, *min))
		1162	min = &dents [i];
		1163
		1164	/* swap elements 0 and j (minimum) */
		1165	{
		1166	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1167	}
		1168	}
		1169
		1170	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1171	{
		1172	eio_dirent *i, *j;
		1173
		1174	for (i = dents + 1; i < dents + size; ++i)
		1175	{
		1176	eio_dirent value = *i;
		1177
		1178	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1179	j [1] = j [0];
		1180
		1181	j [1] = value;
		1182	}
		1183	}
		1184	}
		1185
		1186	static void
		1187	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1188	{
		1189	if (size <= 1)
		1190	return; /* our insertion sort relies on size > 0 */
		1191
		1192	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1193	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1194	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1195
		1196	/* use an insertion sort at the end, or for small arrays, */
		1197	/* as insertion sort is more efficient for small partitions */
		1198	eio_dent_insertion_sort (dents, size);
		1199	}
		1200
802	/* read a full directory */	1201	/* read a full directory */
803	static void	1202	static void
804	eio__scandir (eio_req *req, worker *self)	1203	eio__scandir (eio_req *req, etp_worker *self)
805	{	1204	{
806	DIR *dirp;	1205	DIR *dirp;
807	union
808	{
809	EIO_STRUCT_DIRENT d;
810	char b [offsetof (EIO_STRUCT_DIRENT, d_name) + NAME_MAX + 1];
811	} *u;
812	EIO_STRUCT_DIRENT *entp;	1206	EIO_STRUCT_DIRENT *entp;
813	char *name, *names;	1207	char *name, *names;
814	int memlen = 4096;	1208	int namesalloc = 4096;
815	int memofs = 0;	1209	int namesoffs = 0;
		1210	int flags = req->int1;
		1211	eio_dirent *dents = 0;
		1212	int dentalloc = 128;
816	int res = 0;	1213	int dentoffs = 0;
		1214	ino_t inode_bits = 0;
817		1215
		1216	req->result = -1;
		1217
		1218	if (!(flags & EIO_READDIR_DENTS))
		1219	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
		1220
818	X_LOCK (etplock);	1221	X_LOCK (wrklock);
		1222	/* the corresponding closedir is in ETP_WORKER_CLEAR */
819	self->dirp = dirp = opendir (req->ptr1);	1223	self->dirp = dirp = opendir (req->ptr1);
820	self->dbuf = u = malloc (sizeof (*u));
821	req->flags |= EIO_FLAG_PTR2_FREE;	1224	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1225	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
822	req->ptr2 = names = malloc (memlen);	1226	req->ptr2 = names = malloc (namesalloc);
823	X_UNLOCK (etplock);	1227	X_UNLOCK (wrklock);
824		1228
825	if (dirp && u && names)	1229	if (dirp && names && (!flags || dents))
826	for (;;)	1230	for (;;)
827	{	1231	{
828	errno = 0;	1232	errno = 0;
829	readdir_r (dirp, &u->d, &entp);	1233	entp = readdir (dirp);
830		1234
831	if (!entp)	1235	if (!entp)
		1236	{
		1237	if (errno)
		1238	break;
		1239
		1240	/* sort etc. */
		1241	req->int1 = flags;
		1242	req->result = dentoffs;
		1243
		1244	if (flags & EIO_READDIR_STAT_ORDER)
		1245	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1246	else if (flags & EIO_READDIR_DIRS_FIRST)
		1247	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1248	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1249	else
		1250	{
		1251	/* in this case, all is known, and we just put dirs first and sort them */
		1252	eio_dirent *oth = dents + dentoffs;
		1253	eio_dirent *dir = dents;
		1254
		1255	/* now partition dirs to the front, and non-dirs to the back */
		1256	/* by walking from both sides and swapping if necessary */
		1257	/* also clear score, so it doesn't influence sorting */
		1258	while (oth > dir)
		1259	{
		1260	if (dir->type == EIO_DT_DIR)
		1261	++dir;
		1262	else if ((--oth)->type == EIO_DT_DIR)
		1263	{
		1264	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1265
		1266	++dir;
		1267	}
		1268	}
		1269
		1270	/* now sort the dirs only */
		1271	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1272	}
		1273
832	break;	1274	break;
		1275	}
833		1276
		1277	/* now add the entry to our list(s) */
834	name = entp->d_name;	1278	name = entp->d_name;
835		1279
		1280	/* skip . and .. entries */
836	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1281	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
837	{	1282	{
838	int len = strlen (name) + 1;	1283	int len = D_NAMLEN (entp) + 1;
839		1284
840	res++;	1285	while (expect_false (namesoffs + len > namesalloc))
841
842	while (memofs + len > memlen)
843	{	1286	{
844	memlen *= 2;	1287	namesalloc *= 2;
845	X_LOCK (etplock);	1288	X_LOCK (wrklock);
846	req->ptr2 = names = realloc (names, memlen);	1289	req->ptr2 = names = realloc (names, namesalloc);
847	X_UNLOCK (etplock);	1290	X_UNLOCK (wrklock);
848		1291
849	if (!names)	1292	if (!names)
850	break;	1293	break;
851	}	1294	}
852		1295
853	memcpy (names + memofs, name, len);	1296	memcpy (names + namesoffs, name, len);
		1297
		1298	if (dents)
		1299	{
		1300	struct eio_dirent *ent;
		1301
		1302	if (expect_false (dentoffs == dentalloc))
		1303	{
		1304	dentalloc *= 2;
		1305	X_LOCK (wrklock);
		1306	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1307	X_UNLOCK (wrklock);
		1308
		1309	if (!dents)
		1310	break;
		1311	}
		1312
		1313	ent = dents + dentoffs;
		1314
		1315	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1316	ent->namelen = len - 1;
		1317	ent->inode = D_INO (entp);
		1318
		1319	inode_bits |= ent->inode;
		1320
		1321	switch (D_TYPE (entp))
		1322	{
		1323	default:
		1324	ent->type = EIO_DT_UNKNOWN;
		1325	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1326	break;
		1327
		1328	#ifdef DT_FIFO
		1329	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1330	#endif
		1331	#ifdef DT_CHR
		1332	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1333	#endif
		1334	#ifdef DT_MPC
		1335	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1336	#endif
		1337	#ifdef DT_DIR
		1338	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1339	#endif
		1340	#ifdef DT_NAM
		1341	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1342	#endif
		1343	#ifdef DT_BLK
		1344	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1345	#endif
		1346	#ifdef DT_MPB
		1347	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1348	#endif
		1349	#ifdef DT_REG
		1350	case DT_REG: ent->type = EIO_DT_REG; break;
		1351	#endif
		1352	#ifdef DT_NWK
		1353	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1354	#endif
		1355	#ifdef DT_CMP
		1356	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1357	#endif
		1358	#ifdef DT_LNK
		1359	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1360	#endif
		1361	#ifdef DT_SOCK
		1362	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1363	#endif
		1364	#ifdef DT_DOOR
		1365	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1366	#endif
		1367	#ifdef DT_WHT
		1368	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1369	#endif
		1370	}
		1371
		1372	ent->score = 7;
		1373
		1374	if (flags & EIO_READDIR_DIRS_FIRST)
		1375	{
		1376	if (ent->type == EIO_DT_UNKNOWN)
		1377	{
		1378	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1379	ent->score = 1;
		1380	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1381	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1382	}
		1383	else if (ent->type == EIO_DT_DIR)
		1384	ent->score = 0;
		1385	}
		1386	}
		1387
854	memofs += len;	1388	namesoffs += len;
		1389	++dentoffs;
		1390	}
		1391
		1392	if (EIO_CANCELLED (req))
		1393	{
		1394	errno = ECANCELED;
		1395	break;
855	}	1396	}
856	}	1397	}
		1398	}
857		1399
858	if (errno)	1400	#ifdef PAGESIZE
859	res = -1;	1401	# define eio_pagesize() PAGESIZE
		1402	#else
		1403	static intptr_t
		1404	eio_pagesize (void)
		1405	{
		1406	static intptr_t page;
		1407
		1408	if (!page)
		1409	page = sysconf (_SC_PAGESIZE);
		1410
		1411	return page;
		1412	}
		1413	#endif
		1414
		1415	static void
		1416	eio_page_align (void **addr, size_t *length)
		1417	{
		1418	intptr_t mask = eio_pagesize () - 1;
		1419
		1420	/* round down addr */
		1421	intptr_t adj = mask & (intptr_t)*addr;
		1422
		1423	*addr = (void *)((intptr_t)*addr - adj);
		1424	*length += adj;
		1425
		1426	/* round up length */
		1427	*length = (*length + mask) & ~mask;
		1428	}
		1429
		1430	#if !_POSIX_MEMLOCK
		1431	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1432	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1433	#else
		1434
		1435	static int
		1436	eio__mlock (void *addr, size_t length)
		1437	{
		1438	eio_page_align (&addr, &length);
		1439
		1440	mlock (addr, length);
		1441	}
		1442
		1443	static int
		1444	eio__mlockall (int flags)
		1445	{
		1446	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1447	extern int mallopt (int, int);
		1448	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1449	#endif
		1450
		1451	if (EIO_MCL_CURRENT != MCL_CURRENT
		1452	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1453	{
		1454	flags = 0
		1455	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1456	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1457	}
		1458
		1459	mlockall (flags);
		1460	}
		1461	#endif
		1462
		1463	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1464	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1465	#else
		1466
		1467	int
		1468	eio__msync (void *mem, size_t len, int flags)
		1469	{
		1470	eio_page_align (&mem, &len);
		1471
		1472	if (EIO_MS_ASYNC != MS_SYNC
		1473	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1474	|| EIO_MS_SYNC != MS_SYNC)
		1475	{
		1476	flags = 0
		1477	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1478	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1479	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1480	}
		1481
		1482	return msync (mem, len, flags);
		1483	}
		1484
		1485	#endif
		1486
		1487	int
		1488	eio__mtouch (void *mem, size_t len, int flags)
		1489	{
		1490	eio_page_align (&mem, &len);
		1491
860		1492	{
861	req->result = res;	1493	intptr_t addr = (intptr_t)mem;
		1494	intptr_t end = addr + len;
		1495	intptr_t page = eio_pagesize ();
		1496
		1497	if (addr < end)
		1498	if (flags & EIO_MT_MODIFY) /* modify */
		1499	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
		1500	else
		1501	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
		1502	}
		1503
		1504	return 0;
862	}	1505	}
863		1506
864	/*****************************************************************************/	1507	/*****************************************************************************/
865		1508
866	#define ALLOC(len) \	1509	#define ALLOC(len) \
867	if (!req->ptr2) \	1510	if (!req->ptr2) \
868	{ \	1511	{ \
869	X_LOCK (etplock); \	1512	X_LOCK (wrklock); \
870	req->flags |= EIO_FLAG_PTR2_FREE; \	1513	req->flags |= EIO_FLAG_PTR2_FREE; \
871	X_UNLOCK (etplock); \	1514	X_UNLOCK (wrklock); \
872	req->ptr2 = malloc (len); \	1515	req->ptr2 = malloc (len); \
873	if (!req->ptr2) \	1516	if (!req->ptr2) \
874	{ \	1517	{ \
875	errno = ENOMEM; \	1518	errno = ENOMEM; \
876	req->result = -1; \	1519	req->result = -1; \
877	break; \	1520	break; \
878	} \	1521	} \
879	}	1522	}
880		1523
881	X_THREAD_PROC (eio_proc)	1524	X_THREAD_PROC (etp_proc)
882	{	1525	{
883	eio_req *req;	1526	ETP_REQ *req;
884	struct timespec ts;	1527	struct timespec ts;
885	worker *self = (worker *)thr_arg;	1528	etp_worker *self = (etp_worker *)thr_arg;
886		1529
887	/* try to distribute timeouts somewhat randomly */	1530	/* try to distribute timeouts somewhat randomly */
888	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1531	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
889		1532
890	for (;;)	1533	for (;;)
891	{	1534	{
892	ts.tv_sec = time (0) + IDLE_TIMEOUT;
893
894	X_LOCK (reqlock);	1535	X_LOCK (reqlock);
895		1536
896	for (;;)	1537	for (;;)
897	{	1538	{
898	self->req = req = reqq_shift (&req_queue);	1539	self->req = req = reqq_shift (&req_queue);
…		…
900	if (req)	1541	if (req)
901	break;	1542	break;
902		1543
903	++idle;	1544	++idle;
904		1545
		1546	ts.tv_sec = time (0) + IDLE_TIMEOUT;
905	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1547	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
906	{	1548	{
907	if (idle > max_idle)	1549	if (idle > max_idle)
908	{	1550	{
909	--idle;	1551	--idle;
910	X_UNLOCK (reqlock);	1552	X_UNLOCK (reqlock);
911	X_LOCK (etplock);	1553	X_LOCK (wrklock);
912	--started;	1554	--started;
913	X_UNLOCK (etplock);	1555	X_UNLOCK (wrklock);
914	goto quit;	1556	goto quit;
915	}	1557	}
916		1558
917	/* we are allowed to idle, so do so without any timeout */	1559	/* we are allowed to idle, so do so without any timeout */
918	X_COND_WAIT (reqwait, reqlock);	1560	X_COND_WAIT (reqwait, reqlock);
919	ts.tv_sec = time (0) + IDLE_TIMEOUT;
920	}	1561	}
921		1562
922	--idle;	1563	--idle;
923	}	1564	}
924		1565
925	--nready;	1566	--nready;
926		1567
927	X_UNLOCK (reqlock);	1568	X_UNLOCK (reqlock);
928		1569
929	errno = 0; /* strictly unnecessary */	1570	if (req->type < 0)
		1571	goto quit;
930		1572
931	if (!EIO_CANCELLED (req))	1573	if (!EIO_CANCELLED (req))
932	switch (req->type)	1574	ETP_EXECUTE (self, req);
933	{
934	case EIO_READ: ALLOC (req->size);
935	req->result = req->offs >= 0
936	? pread (req->int1, req->ptr2, req->size, req->offs)
937	: read (req->int1, req->ptr2, req->size); break;
938	case EIO_WRITE: req->result = req->offs >= 0
939	? pwrite (req->int1, req->ptr2, req->size, req->offs)
940	: write (req->int1, req->ptr2, req->size); break;
941
942	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
943	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
944
945	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
946	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
947	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
948	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
949	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
950	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
951
952	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
953	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
954	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
955	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
956	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
957	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
958
959	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
960	case EIO_CLOSE: req->result = close (req->int1); break;
961	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
962	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
963	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
964	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
965	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
966	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
967	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
968	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
969
970	case EIO_READLINK: ALLOC (NAME_MAX);
971	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
972
973	case EIO_SYNC: req->result = 0; sync (); break;
974	case EIO_FSYNC: req->result = fsync (req->int1); break;
975	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
976
977	case EIO_READDIR: eio__scandir (req, self); break;
978
979	case EIO_BUSY:
980	#ifdef _WIN32
981	Sleep (req->nv1 * 1000.);
982	#else
983	{
984	struct timeval tv;
985
986	tv.tv_sec = req->nv1;
987	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.;
988
989	req->result = select (0, 0, 0, 0, &tv);
990	}
991	#endif
992	break;
993
994	case EIO_UTIME:
995	case EIO_FUTIME:
996	{
997	struct timeval tv[2];
998	struct timeval *times;
999
1000	if (req->nv1 != -1. || req->nv2 != -1.)
1001	{
1002	tv[0].tv_sec = req->nv1;
1003	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
1004	tv[1].tv_sec = req->nv2;
1005	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
1006
1007	times = tv;
1008	}
1009	else
1010	times = 0;
1011
1012
1013	req->result = req->type == EIO_FUTIME
1014	? futimes (req->int1, times)
1015	: utimes (req->ptr1, times);
1016	}
1017
1018	case EIO_GROUP:
1019	case EIO_NOP:
1020	req->result = 0;
1021	break;
1022
1023	case EIO_QUIT:
1024	goto quit;
1025
1026	default:
1027	req->result = -1;
1028	break;
1029	}
1030
1031	req->errorno = errno;
1032		1575
1033	X_LOCK (reslock);	1576	X_LOCK (reslock);
1034		1577
1035	++npending;	1578	++npending;
1036		1579
1037	if (!reqq_push (&res_queue, req) && etp.want_poll_cb)	1580	if (!reqq_push (&res_queue, req) && want_poll_cb)
1038	etp.want_poll_cb ();	1581	want_poll_cb ();
1039		1582
1040	self->req = 0;	1583	self->req = 0;
1041	worker_clear (self);	1584	etp_worker_clear (self);
1042		1585
1043	X_UNLOCK (reslock);	1586	X_UNLOCK (reslock);
1044	}	1587	}
1045		1588
1046	quit:	1589	quit:
1047	X_LOCK (etplock);	1590	X_LOCK (wrklock);
1048	worker_free (self);	1591	etp_worker_free (self);
1049	X_UNLOCK (etplock);	1592	X_UNLOCK (wrklock);
1050		1593
1051	return 0;	1594	return 0;
1052	}	1595	}
1053		1596
1054	/*****************************************************************************/	1597	/*****************************************************************************/
1055		1598
1056	int eio_init (void (*want_poll)(void), void (*done_poll)(void))	1599	int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1057	{	1600	{
1058	etp_init (&etp, want_poll, done_poll);	1601	return etp_init (want_poll, done_poll);
1059	}	1602	}
1060		1603
1061	static void eio_api_destroy (eio_req *req)	1604	static void eio_api_destroy (eio_req *req)
1062	{	1605	{
1063	free (req);	1606	free (req);
…		…
1085	{ \	1628	{ \
1086	eio_api_destroy (req); \	1629	eio_api_destroy (req); \
1087	return 0; \	1630	return 0; \
1088	}	1631	}
1089		1632
		1633	static void eio_execute (etp_worker *self, eio_req *req)
		1634	{
		1635	errno = 0;
		1636
		1637	switch (req->type)
		1638	{
		1639	case EIO_READ: ALLOC (req->size);
		1640	req->result = req->offs >= 0
		1641	? pread (req->int1, req->ptr2, req->size, req->offs)
		1642	: read (req->int1, req->ptr2, req->size); break;
		1643	case EIO_WRITE: req->result = req->offs >= 0
		1644	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1645	: write (req->int1, req->ptr2, req->size); break;
		1646
		1647	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1648	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1649
		1650	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1651	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1652	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1653	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1654	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1655	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1656
		1657	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1658	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1659	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1660	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1661
		1662	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1663	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1664	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1665	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1666	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1667	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1668
		1669	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1670	case EIO_CLOSE: req->result = close (req->int1); break;
		1671	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1672	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1673	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1674	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1675	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1676	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1677	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1678	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
		1679
		1680	case EIO_READLINK: ALLOC (PATH_MAX);
		1681	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1682
		1683	case EIO_SYNC: req->result = 0; sync (); break;
		1684	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1685	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1686	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1687	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1688	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1689	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1690	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1691
		1692	case EIO_READDIR: eio__scandir (req, self); break;
		1693
		1694	case EIO_BUSY:
		1695	#ifdef _WIN32
		1696	Sleep (req->nv1 * 1000.);
		1697	#else
		1698	{
		1699	struct timeval tv;
		1700
		1701	tv.tv_sec = req->nv1;
		1702	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.;
		1703
		1704	req->result = select (0, 0, 0, 0, &tv);
		1705	}
		1706	#endif
		1707	break;
		1708
		1709	case EIO_UTIME:
		1710	case EIO_FUTIME:
		1711	{
		1712	struct timeval tv[2];
		1713	struct timeval *times;
		1714
		1715	if (req->nv1 != -1. || req->nv2 != -1.)
		1716	{
		1717	tv[0].tv_sec = req->nv1;
		1718	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1719	tv[1].tv_sec = req->nv2;
		1720	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1721
		1722	times = tv;
		1723	}
		1724	else
		1725	times = 0;
		1726
		1727
		1728	req->result = req->type == EIO_FUTIME
		1729	? futimes (req->int1, times)
		1730	: utimes (req->ptr1, times);
		1731	}
		1732	break;
		1733
		1734	case EIO_GROUP:
		1735	abort (); /* handled in eio_request */
		1736
		1737	case EIO_NOP:
		1738	req->result = 0;
		1739	break;
		1740
		1741	case EIO_CUSTOM:
		1742	((void (*)(eio_req *))req->feed) (req);
		1743	break;
		1744
		1745	default:
		1746	req->result = -1;
		1747	break;
		1748	}
		1749
		1750	req->errorno = errno;
		1751	}
		1752
1090	#ifndef EIO_NO_WRAPPERS	1753	#ifndef EIO_NO_WRAPPERS
1091		1754
1092	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1755	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1093	{	1756	{
1094	REQ (EIO_NOP); SEND;	1757	REQ (EIO_NOP); SEND;
…		…
1107	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1770	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1108	{	1771	{
1109	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1772	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1110	}	1773	}
1111		1774
		1775	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1776	{
		1777	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1778	}
		1779
		1780	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1781	{
		1782	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1783	}
		1784
		1785	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1786	{
		1787	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1788	}
		1789
		1790	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1791	{
		1792	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1793	}
		1794
		1795	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1796	{
		1797	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1798	}
		1799
1112	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1800	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1113	{	1801	{
1114	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1802	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1115	}	1803	}
1116		1804
…		…
1135	}	1823	}
1136		1824
1137	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1825	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1138	{	1826	{
1139	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1827	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1828	}
		1829
		1830	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1831	{
		1832	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1140	}	1833	}
1141		1834
1142	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1835	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1143	{	1836	{
1144	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1837	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
…		…
1218	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)	1911	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1219	{	1912	{
1220	return eio__1path (EIO_LSTAT, path, pri, cb, data);	1913	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1221	}	1914	}
1222		1915
		1916	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		1917	{
		1918	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		1919	}
		1920
1223	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	1921	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1224	{	1922	{
1225	return eio__1path (EIO_UNLINK, path, pri, cb, data);	1923	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1226	}	1924	}
1227		1925
1228	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	1926	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1229	{	1927	{
1230	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1928	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1231	}	1929	}
1232		1930
1233	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	1931	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1234	{	1932	{
1235	return eio__1path (EIO_READDIR, path, pri, cb, data);	1933	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1236	}	1934	}
1237		1935
1238	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1936	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1239	{	1937	{
1240	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	1938	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1241	}	1939	}
1242		1940
1243	static eio_req *	1941	static eio_req *
1244	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1942	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1245	{	1943	{
…		…
1269	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1967	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1270	{	1968	{
1271	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1969	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1272	}	1970	}
1273		1971
		1972	eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
		1973	{
		1974	REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
		1975	}
		1976
1274	#endif	1977	#endif
1275		1978
1276	eio_req *eio_grp (eio_cb cb, void *data)	1979	eio_req *eio_grp (eio_cb cb, void *data)
1277	{	1980	{
1278	const int pri = EIO_PRI_MAX;	1981	const int pri = EIO_PRI_MAX;
…		…
1304		2007
1305	void eio_grp_add (eio_req *grp, eio_req *req)	2008	void eio_grp_add (eio_req *grp, eio_req *req)
1306	{	2009	{
1307	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2010	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1308		2011
		2012	grp->flags |= EIO_FLAG_GROUPADD;
		2013
1309	++grp->size;	2014	++grp->size;
1310	req->grp = grp;	2015	req->grp = grp;
1311		2016
1312	req->grp_prev = 0;	2017	req->grp_prev = 0;
1313	req->grp_next = grp->grp_first;	2018	req->grp_next = grp->grp_first;
…		…
1321	/*****************************************************************************/	2026	/*****************************************************************************/
1322	/* misc garbage */	2027	/* misc garbage */
1323		2028
1324	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2029	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1325	{	2030	{
1326	worker wrk;	2031	etp_worker wrk;
		2032	ssize_t ret;
1327		2033
1328	wrk.dbuf = 0;	2034	wrk.dbuf = 0;
1329		2035
1330	eio__sendfile (ofd, ifd, offset, count, &wrk);	2036	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1331		2037
1332	if (wrk.dbuf)	2038	if (wrk.dbuf)
1333	free (wrk.dbuf);	2039	free (wrk.dbuf);
1334	}
1335		2040
		2041	return ret;
		2042	}
		2043

Diff Legend

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