ViewVC Help

View File | Revision Log | Show Annotations | Download File

/cvs/libeio/eio.c

Comparing libeio/eio.c (file contents):
Revision 1.11 by root, Tue May 13 18:50:27 2008 UTC vs.
Revision 1.72 by root, Fri Jun 10 12:45:20 2011 UTC

		1	/*
		2	* libeio implementation
		3	*
		4	* Copyright (c) 2007,2008,2009,2010,2011 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
		40	#ifndef _WIN32
		41	# include "config.h"
		42	#endif
		43
1	#include "eio.h"	44	#include "eio.h"
		45	#include "ecb.h"
		46
		47	#ifdef EIO_STACKSIZE
		48	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		49	#endif
2	#include "xthread.h"	50	#include "xthread.h"
3		51
4	#include <errno.h>	52	#include <errno.h>
5	#include <stddef.h>	53	#include <stddef.h>
6	#include <stdlib.h>	54	#include <stdlib.h>
7	#include <string.h>	55	#include <string.h>
8	#include <errno.h>	56	#include <errno.h>
9	#include <sys/types.h>	57	#include <sys/types.h>
10	#include <sys/stat.h>	58	#include <sys/stat.h>
		59	#include <sys/statvfs.h>
11	#include <limits.h>	60	#include <limits.h>
12	#include <fcntl.h>	61	#include <fcntl.h>
13	#include <assert.h>	62	#include <assert.h>
14		63
		64	/* intptr_t comes from unistd.h, says POSIX/UNIX/tradition */
		65	/* intptr_t only comes form stdint.h, says idiot openbsd coder */
		66	#if HAVE_STDINT_H
		67	# include <stdint.h>
		68	#endif
		69
15	#ifndef EIO_FINISH	70	#ifndef EIO_FINISH
16	# define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0	71	# define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0
17	#endif	72	#endif
18		73
19	#ifndef EIO_DESTROY	74	#ifndef EIO_DESTROY
…		…
25	#endif	80	#endif
26		81
27	#ifdef _WIN32	82	#ifdef _WIN32
28		83
29	/*doh*/	84	/*doh*/
30
31	#else	85	#else
32		86
33	# include "config.h"
34	# include <sys/time.h>	87	# include <sys/time.h>
35	# include <sys/select.h>	88	# include <sys/select.h>
36	# include <unistd.h>	89	# include <unistd.h>
37	# include <utime.h>	90	# include <utime.h>
38	# include <signal.h>	91	# include <signal.h>
39	# include <dirent.h>	92	# include <dirent.h>
40		93
		94	#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
		95	# include <sys/mman.h>
		96	#endif
		97
		98	/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
		99	# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
		100	# define _DIRENT_HAVE_D_TYPE /* sigh */
		101	# define D_INO(de) (de)->d_fileno
		102	# define D_NAMLEN(de) (de)->d_namlen
		103	# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
		104	# define D_INO(de) (de)->d_ino
		105	# endif
		106
		107	#ifdef _D_EXACT_NAMLEN
		108	# undef D_NAMLEN
		109	# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
		110	#endif
		111
		112	# ifdef _DIRENT_HAVE_D_TYPE
		113	# define D_TYPE(de) (de)->d_type
		114	# endif
		115
41	# ifndef EIO_STRUCT_DIRENT	116	# ifndef EIO_STRUCT_DIRENT
42	# define EIO_STRUCT_DIRENT struct dirent	117	# define EIO_STRUCT_DIRENT struct dirent
43	# endif	118	# endif
44		119
45	#endif	120	#endif
46		121
47	#if HAVE_SENDFILE	122	#if HAVE_SENDFILE
48	# if __linux	123	# if __linux
49	# include <sys/sendfile.h>	124	# include <sys/sendfile.h>
50	# elif __freebsd	125	# elif __FreeBSD__ || defined __APPLE__
51	# include <sys/socket.h>	126	# include <sys/socket.h>
52	# include <sys/uio.h>	127	# include <sys/uio.h>
53	# elif __hpux	128	# elif __hpux
54	# include <sys/socket.h>	129	# include <sys/socket.h>
55	# elif __solaris /* not yet */	130	# elif __solaris
56	# include <sys/sendfile.h>	131	# include <sys/sendfile.h>
57	# else	132	# else
58	# error sendfile support requested but not available	133	# error sendfile support requested but not available
59	# endif	134	# endif
60	#endif	135	#endif
61		136
62	/* number of seconds after which an idle threads exit */	137	#ifndef D_TYPE
63	#define IDLE_TIMEOUT 10	138	# define D_TYPE(de) 0
		139	#endif
		140	#ifndef D_INO
		141	# define D_INO(de) 0
		142	#endif
		143	#ifndef D_NAMLEN
		144	# define D_NAMLEN(de) strlen ((de)->d_name)
		145	#endif
64		146
65	/* used for struct dirent, AIX doesn't provide it */	147	/* used for struct dirent, AIX doesn't provide it */
66	#ifndef NAME_MAX	148	#ifndef NAME_MAX
67	# define NAME_MAX 4096	149	# define NAME_MAX 4096
68	#endif	150	#endif
69		151
		152	/* used for readlink etc. */
		153	#ifndef PATH_MAX
		154	# define PATH_MAX 4096
		155	#endif
		156
70	/* buffer size for various temporary buffers */	157	/* buffer size for various temporary buffers */
71	#define EIO_BUFSIZE 65536	158	#define EIO_BUFSIZE 65536
72		159
73	#define dBUF \	160	#define dBUF \
74	char *eio_buf; \	161	char *eio_buf; \
75	X_LOCK (wrklock); \	162	ETP_WORKER_LOCK (self); \
76	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \	163	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
77	X_UNLOCK (wrklock); \	164	ETP_WORKER_UNLOCK (self); \
78	errno = ENOMEM; \	165	errno = ENOMEM; \
79	if (!eio_buf) \	166	if (!eio_buf) \
80	return -1;	167	return -1;
81		168
82	#define EIO_TICKS ((1000000 + 1023) >> 10)	169	#define EIO_TICKS ((1000000 + 1023) >> 10)
83		170
		171	#define ETP_PRI_MIN EIO_PRI_MIN
		172	#define ETP_PRI_MAX EIO_PRI_MAX
		173
		174	struct etp_worker;
		175
		176	#define ETP_REQ eio_req
		177	#define ETP_DESTROY(req) eio_destroy (req)
		178	static int eio_finish (eio_req *req);
		179	#define ETP_FINISH(req) eio_finish (req)
		180	static void eio_execute (struct etp_worker *self, eio_req *req);
		181	#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
		182
		183	#define ETP_WORKER_CLEAR(req) \
		184	if (wrk->dbuf) \
		185	{ \
		186	free (wrk->dbuf); \
		187	wrk->dbuf = 0; \
		188	} \
		189	\
		190	if (wrk->dirp) \
		191	{ \
		192	closedir (wrk->dirp); \
		193	wrk->dirp = 0; \
		194	}
		195
		196	#define ETP_WORKER_COMMON \
		197	void *dbuf; \
		198	DIR *dirp;
		199
		200	/*****************************************************************************/
		201
		202	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
		203
		204	/* calculate time difference in ~1/EIO_TICKS of a second */
		205	ECB_INLINE int
		206	tvdiff (struct timeval *tv1, struct timeval *tv2)
		207	{
		208	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
		209	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
		210	}
		211
		212	static unsigned int started, idle, wanted = 4;
		213
84	static void (*want_poll_cb) (void);	214	static void (*want_poll_cb) (void);
85	static void (*done_poll_cb) (void);	215	static void (*done_poll_cb) (void);
86		216
87	static unsigned int max_poll_time = 0;	217	static unsigned int max_poll_time; /* reslock */
88	static unsigned int max_poll_reqs = 0;	218	static unsigned int max_poll_reqs; /* reslock */
89		219
90	/* calculcate time difference in ~1/EIO_TICKS of a second */	220	static volatile unsigned int nreqs; /* reqlock */
91	static int tvdiff (struct timeval *tv1, struct timeval *tv2)	221	static volatile unsigned int nready; /* reqlock */
92	{	222	static volatile unsigned int npending; /* reqlock */
93	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS	223	static volatile unsigned int max_idle = 4; /* maximum number of threads that can idle indefinitely */
94	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);	224	static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */
95	}
96		225
97	static unsigned int started, idle, wanted = 4;	226	static xmutex_t wrklock;
		227	static xmutex_t reslock;
		228	static xmutex_t reqlock;
		229	static xcond_t reqwait;
		230
		231	#if !HAVE_PREADWRITE
		232	/*
		233	* make our pread/pwrite emulation safe against themselves, but not against
		234	* normal read/write by using a mutex. slows down execution a lot,
		235	* but that's your problem, not mine.
		236	*/
		237	static xmutex_t preadwritelock = X_MUTEX_INIT;
		238	#endif
		239
		240	typedef struct etp_worker
		241	{
		242	/* locked by wrklock */
		243	struct etp_worker *prev, *next;
		244
		245	xthread_t tid;
		246
		247	/* locked by reslock, reqlock or wrklock */
		248	ETP_REQ *req; /* currently processed request */
		249
		250	ETP_WORKER_COMMON
		251	} etp_worker;
		252
		253	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		254
		255	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		256	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
98		257
99	/* worker threads management */	258	/* worker threads management */
100	static mutex_t wrklock = X_MUTEX_INIT;
101		259
102	typedef struct worker	260	static void ecb_cold
103	{
104	/* locked by wrklock */
105	struct worker *prev, *next;
106
107	thread_t tid;
108
109	/* locked by reslock, reqlock or wrklock */
110	eio_req *req; /* currently processed request */
111	void *dbuf;
112	DIR *dirp;
113	} worker;
114
115	static worker wrk_first = { &wrk_first, &wrk_first, 0 };
116
117	static void worker_clear (worker *wrk)	261	etp_worker_clear (etp_worker *wrk)
118	{	262	{
119	if (wrk->dirp)	263	ETP_WORKER_CLEAR (wrk);
120	{
121	closedir (wrk->dirp);
122	wrk->dirp = 0;
123	}
124
125	if (wrk->dbuf)
126	{
127	free (wrk->dbuf);
128	wrk->dbuf = 0;
129	}
130	}	264	}
131		265
132	static void worker_free (worker *wrk)	266	static void ecb_cold
		267	etp_worker_free (etp_worker *wrk)
133	{	268	{
134	wrk->next->prev = wrk->prev;	269	wrk->next->prev = wrk->prev;
135	wrk->prev->next = wrk->next;	270	wrk->prev->next = wrk->next;
136		271
137	free (wrk);	272	free (wrk);
138	}	273	}
139		274
140	static volatile unsigned int nreqs, nready, npending;	275	static unsigned int
141	static volatile unsigned int max_idle = 4;	276	etp_nreqs (void)
142
143	static mutex_t reslock = X_MUTEX_INIT;
144	static mutex_t reqlock = X_MUTEX_INIT;
145	static cond_t reqwait = X_COND_INIT;
146
147	unsigned int eio_nreqs (void)
148	{	277	{
		278	int retval;
		279	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		280	retval = nreqs;
		281	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
149	return nreqs;	282	return retval;
150	}	283	}
151		284
152	unsigned int eio_nready (void)	285	static unsigned int
		286	etp_nready (void)
153	{	287	{
154	unsigned int retval;	288	unsigned int retval;
155		289
156	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	290	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
157	retval = nready;	291	retval = nready;
158	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	292	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
159		293
160	return retval;	294	return retval;
161	}	295	}
162		296
163	unsigned int eio_npending (void)	297	static unsigned int
		298	etp_npending (void)
164	{	299	{
165	unsigned int retval;	300	unsigned int retval;
166		301
167	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	302	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
168	retval = npending;	303	retval = npending;
169	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	304	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
170		305
171	return retval;	306	return retval;
172	}	307	}
173		308
174	unsigned int eio_nthreads (void)	309	static unsigned int
		310	etp_nthreads (void)
175	{	311	{
176	unsigned int retval;	312	unsigned int retval;
177		313
178	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	314	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
179	retval = started;	315	retval = started;
…		…
186	* a somewhat faster data structure might be nice, but	322	* a somewhat faster data structure might be nice, but
187	* with 8 priorities this actually needs <20 insns	323	* with 8 priorities this actually needs <20 insns
188	* per shift, the most expensive operation.	324	* per shift, the most expensive operation.
189	*/	325	*/
190	typedef struct {	326	typedef struct {
191	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	327	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
192	int size;	328	int size;
193	} reqq;	329	} etp_reqq;
194		330
195	static reqq req_queue;	331	static etp_reqq req_queue;
196	static reqq res_queue;	332	static etp_reqq res_queue;
197		333
198	static int reqq_push (reqq *q, eio_req *req)	334	static int ecb_noinline
		335	reqq_push (etp_reqq *q, ETP_REQ *req)
199	{	336	{
200	int pri = req->pri;	337	int pri = req->pri;
201	req->next = 0;	338	req->next = 0;
202		339
203	if (q->qe[pri])	340	if (q->qe[pri])
…		…
209	q->qe[pri] = q->qs[pri] = req;	346	q->qe[pri] = q->qs[pri] = req;
210		347
211	return q->size++;	348	return q->size++;
212	}	349	}
213		350
214	static eio_req *reqq_shift (reqq *q)	351	static ETP_REQ * ecb_noinline
		352	reqq_shift (etp_reqq *q)
215	{	353	{
216	int pri;	354	int pri;
217		355
218	if (!q->size)	356	if (!q->size)
219	return 0;	357	return 0;
220		358
221	--q->size;	359	--q->size;
222		360
223	for (pri = EIO_NUM_PRI; pri--; )	361	for (pri = ETP_NUM_PRI; pri--; )
224	{	362	{
225	eio_req *req = q->qs[pri];	363	eio_req *req = q->qs[pri];
226		364
227	if (req)	365	if (req)
228	{	366	{
…		…
234	}	372	}
235		373
236	abort ();	374	abort ();
237	}	375	}
238		376
239	static void grp_try_feed (eio_req *grp)	377	static void ecb_cold
		378	etp_thread_init (void)
240	{	379	{
241	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	380	X_MUTEX_CREATE (wrklock);
		381	X_MUTEX_CREATE (reslock);
		382	X_MUTEX_CREATE (reqlock);
		383	X_COND_CREATE (reqwait);
		384	}
		385
		386	static void ecb_cold
		387	etp_atfork_prepare (void)
		388	{
		389	X_LOCK (wrklock);
		390	X_LOCK (reqlock);
		391	X_LOCK (reslock);
		392	#if !HAVE_PREADWRITE
		393	X_LOCK (preadwritelock);
		394	#endif
		395	}
		396
		397	static void ecb_cold
		398	etp_atfork_parent (void)
		399	{
		400	#if !HAVE_PREADWRITE
		401	X_UNLOCK (preadwritelock);
		402	#endif
		403	X_UNLOCK (reslock);
		404	X_UNLOCK (reqlock);
		405	X_UNLOCK (wrklock);
		406	}
		407
		408	static void ecb_cold
		409	etp_atfork_child (void)
		410	{
		411	ETP_REQ *prv;
		412
		413	while ((prv = reqq_shift (&req_queue)))
		414	ETP_DESTROY (prv);
		415
		416	while ((prv = reqq_shift (&res_queue)))
		417	ETP_DESTROY (prv);
		418
		419	while (wrk_first.next != &wrk_first)
242	{	420	{
243	int old_len = grp->size;	421	etp_worker *wrk = wrk_first.next;
244		422
245	EIO_FEED (grp);	423	if (wrk->req)
		424	ETP_DESTROY (wrk->req);
246		425
247	/* stop if no progress has been made */	426	etp_worker_clear (wrk);
248	if (old_len == grp->size)	427	etp_worker_free (wrk);
249	{
250	grp->feed = 0;
251	break;
252	}
253	}	428	}
254	}
255		429
256	static int eio_finish (eio_req *req);	430	started = 0;
		431	idle = 0;
		432	nreqs = 0;
		433	nready = 0;
		434	npending = 0;
257		435
258	static int grp_dec (eio_req *grp)	436	etp_thread_init ();
259	{	437	}
260	--grp->size;
261		438
262	/* call feeder, if applicable */	439	static void ecb_cold
263	grp_try_feed (grp);	440	etp_once_init (void)
		441	{
		442	etp_thread_init ();
		443	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
		444	}
264		445
265	/* finish, if done */	446	static int ecb_cold
266	if (!grp->size && grp->int1)	447	etp_init (void (*want_poll)(void), void (*done_poll)(void))
267	return eio_finish (grp);	448	{
268	else	449	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
		450
		451	pthread_once (&doinit, etp_once_init);
		452
		453	want_poll_cb = want_poll;
		454	done_poll_cb = done_poll;
		455
269	return 0;	456	return 0;
270	}	457	}
271		458
272	void eio_destroy (eio_req *req)
273	{
274	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
275	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
276
277	EIO_DESTROY (req);
278	}
279
280	static int eio_finish (eio_req *req)
281	{
282	int res = EIO_FINISH (req);
283
284	if (req->grp)
285	{
286	int res2;
287	eio_req *grp = req->grp;
288
289	/* unlink request */
290	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
291	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
292
293	if (grp->grp_first == req)
294	grp->grp_first = req->grp_next;
295
296	res2 = grp_dec (grp);
297
298	if (!res && res2)
299	res = res2;
300	}
301
302	eio_destroy (req);
303
304	return res;
305	}
306
307	void eio_grp_cancel (eio_req *grp)
308	{
309	for (grp = grp->grp_first; grp; grp = grp->grp_next)
310	eio_cancel (grp);
311	}
312
313	void eio_cancel (eio_req *req)
314	{
315	X_LOCK (wrklock);
316	req->flags |= EIO_FLAG_CANCELLED;
317	X_UNLOCK (wrklock);
318
319	eio_grp_cancel (req);
320	}
321
322	X_THREAD_PROC (eio_proc);	459	X_THREAD_PROC (etp_proc);
323		460
324	static void start_thread (void)	461	static void ecb_cold
		462	etp_start_thread (void)
325	{	463	{
326	worker *wrk = calloc (1, sizeof (worker));	464	etp_worker *wrk = calloc (1, sizeof (etp_worker));
327		465
328	/*TODO*/	466	/*TODO*/
329	assert (("unable to allocate worker thread data", wrk));	467	assert (("unable to allocate worker thread data", wrk));
330		468
331	X_LOCK (wrklock);	469	X_LOCK (wrklock);
332		470
333	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	471	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
334	{	472	{
335	wrk->prev = &wrk_first;	473	wrk->prev = &wrk_first;
336	wrk->next = wrk_first.next;	474	wrk->next = wrk_first.next;
337	wrk_first.next->prev = wrk;	475	wrk_first.next->prev = wrk;
338	wrk_first.next = wrk;	476	wrk_first.next = wrk;
…		…
342	free (wrk);	480	free (wrk);
343		481
344	X_UNLOCK (wrklock);	482	X_UNLOCK (wrklock);
345	}	483	}
346		484
		485	static void
347	static void maybe_start_thread (void)	486	etp_maybe_start_thread (void)
348	{	487	{
349	if (eio_nthreads () >= wanted)	488	if (ecb_expect_true (etp_nthreads () >= wanted))
350	return;	489	return;
351		490
352	/* todo: maybe use idle here, but might be less exact */	491	/* todo: maybe use idle here, but might be less exact */
353	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	492	if (ecb_expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
354	return;	493	return;
355		494
356	start_thread ();	495	etp_start_thread ();
357	}	496	}
358		497
359	void eio_submit (eio_req *req)	498	static void ecb_cold
360	{	499	etp_end_thread (void)
361	req->pri += EIO_PRI_BIAS;
362
363	if (req->pri < EIO_PRI_MIN + EIO_PRI_BIAS) req->pri = EIO_PRI_MIN + EIO_PRI_BIAS;
364	if (req->pri > EIO_PRI_MAX + EIO_PRI_BIAS) req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;
365
366	++nreqs;
367
368	X_LOCK (reqlock);
369	++nready;
370	reqq_push (&req_queue, req);
371	X_COND_SIGNAL (reqwait);
372	X_UNLOCK (reqlock);
373
374	maybe_start_thread ();
375	}
376
377	static void end_thread (void)
378	{	500	{
379	eio_req *req = calloc (1, sizeof (eio_req));	501	eio_req *req = calloc (1, sizeof (eio_req));
380		502
381	req->type = EIO_QUIT;	503	req->type = -1;
382	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	504	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
383		505
384	X_LOCK (reqlock);	506	X_LOCK (reqlock);
385	reqq_push (&req_queue, req);	507	reqq_push (&req_queue, req);
386	X_COND_SIGNAL (reqwait);	508	X_COND_SIGNAL (reqwait);
387	X_UNLOCK (reqlock);	509	X_UNLOCK (reqlock);
…		…
389	X_LOCK (wrklock);	511	X_LOCK (wrklock);
390	--started;	512	--started;
391	X_UNLOCK (wrklock);	513	X_UNLOCK (wrklock);
392	}	514	}
393		515
394	void eio_set_max_poll_time (double nseconds)	516	static int
395	{
396	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
397	max_poll_time = nseconds;
398	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
399	}
400
401	void eio_set_max_poll_reqs (unsigned int maxreqs)
402	{
403	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
404	max_poll_reqs = maxreqs;
405	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
406	}
407
408	void eio_set_max_idle (unsigned int nthreads)
409	{
410	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
411	max_idle = nthreads <= 0 ? 1 : nthreads;
412	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
413	}
414
415	void eio_set_min_parallel (unsigned int nthreads)
416	{
417	if (wanted < nthreads)
418	wanted = nthreads;
419	}
420
421	void eio_set_max_parallel (unsigned int nthreads)
422	{
423	if (wanted > nthreads)
424	wanted = nthreads;
425
426	while (started > wanted)
427	end_thread ();
428	}
429
430	int eio_poll (void)	517	etp_poll (void)
431	{	518	{
432	int maxreqs = max_poll_reqs;	519	unsigned int maxreqs;
		520	unsigned int maxtime;
433	struct timeval tv_start, tv_now;	521	struct timeval tv_start, tv_now;
434	eio_req *req;
435		522
		523	X_LOCK (reslock);
		524	maxreqs = max_poll_reqs;
		525	maxtime = max_poll_time;
		526	X_UNLOCK (reslock);
		527
436	if (max_poll_time)	528	if (maxtime)
437	gettimeofday (&tv_start, 0);	529	gettimeofday (&tv_start, 0);
438		530
439	for (;;)	531	for (;;)
440	{	532	{
		533	ETP_REQ *req;
		534
441	maybe_start_thread ();	535	etp_maybe_start_thread ();
442		536
443	X_LOCK (reslock);	537	X_LOCK (reslock);
444	req = reqq_shift (&res_queue);	538	req = reqq_shift (&res_queue);
445		539
446	if (req)	540	if (req)
…		…
454	X_UNLOCK (reslock);	548	X_UNLOCK (reslock);
455		549
456	if (!req)	550	if (!req)
457	return 0;	551	return 0;
458		552
		553	X_LOCK (reqlock);
459	--nreqs;	554	--nreqs;
		555	X_UNLOCK (reqlock);
460		556
461	if (req->type == EIO_GROUP && req->size)	557	if (ecb_expect_false (req->type == EIO_GROUP && req->size))
462	{	558	{
463	req->int1 = 1; /* mark request as delayed */	559	req->int1 = 1; /* mark request as delayed */
464	continue;	560	continue;
465	}	561	}
466	else	562	else
467	{	563	{
468	int res = eio_finish (req);	564	int res = ETP_FINISH (req);
469	if (res)	565	if (ecb_expect_false (res))
470	return res;	566	return res;
471	}	567	}
472		568
473	if (maxreqs && !--maxreqs)	569	if (ecb_expect_false (maxreqs && !--maxreqs))
474	break;	570	break;
475		571
476	if (max_poll_time)	572	if (maxtime)
477	{	573	{
478	gettimeofday (&tv_now, 0);	574	gettimeofday (&tv_now, 0);
479		575
480	if (tvdiff (&tv_start, &tv_now) >= max_poll_time)	576	if (tvdiff (&tv_start, &tv_now) >= maxtime)
481	break;	577	break;
482	}	578	}
483	}	579	}
484		580
485	errno = EAGAIN;	581	errno = EAGAIN;
486	return -1;	582	return -1;
487	}	583	}
488		584
		585	static void
		586	etp_cancel (ETP_REQ *req)
		587	{
		588	X_LOCK (wrklock);
		589	req->flags |= EIO_FLAG_CANCELLED;
		590	X_UNLOCK (wrklock);
		591
		592	eio_grp_cancel (req);
		593	}
		594
		595	static void
		596	etp_submit (ETP_REQ *req)
		597	{
		598	req->pri -= ETP_PRI_MIN;
		599
		600	if (ecb_expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		601	if (ecb_expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		602
		603	if (ecb_expect_false (req->type == EIO_GROUP))
		604	{
		605	/* I hope this is worth it :/ */
		606	X_LOCK (reqlock);
		607	++nreqs;
		608	X_UNLOCK (reqlock);
		609
		610	X_LOCK (reslock);
		611
		612	++npending;
		613
		614	if (!reqq_push (&res_queue, req) && want_poll_cb)
		615	want_poll_cb ();
		616
		617	X_UNLOCK (reslock);
		618	}
		619	else
		620	{
		621	X_LOCK (reqlock);
		622	++nreqs;
		623	++nready;
		624	reqq_push (&req_queue, req);
		625	X_COND_SIGNAL (reqwait);
		626	X_UNLOCK (reqlock);
		627
		628	etp_maybe_start_thread ();
		629	}
		630	}
		631
		632	static void ecb_cold
		633	etp_set_max_poll_time (double nseconds)
		634	{
		635	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		636	max_poll_time = nseconds * EIO_TICKS;
		637	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		638	}
		639
		640	static void ecb_cold
		641	etp_set_max_poll_reqs (unsigned int maxreqs)
		642	{
		643	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		644	max_poll_reqs = maxreqs;
		645	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		646	}
		647
		648	static void ecb_cold
		649	etp_set_max_idle (unsigned int nthreads)
		650	{
		651	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		652	max_idle = nthreads;
		653	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		654	}
		655
		656	static void ecb_cold
		657	etp_set_idle_timeout (unsigned int seconds)
		658	{
		659	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		660	idle_timeout = seconds;
		661	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		662	}
		663
		664	static void ecb_cold
		665	etp_set_min_parallel (unsigned int nthreads)
		666	{
		667	if (wanted < nthreads)
		668	wanted = nthreads;
		669	}
		670
		671	static void ecb_cold
		672	etp_set_max_parallel (unsigned int nthreads)
		673	{
		674	if (wanted > nthreads)
		675	wanted = nthreads;
		676
		677	while (started > wanted)
		678	etp_end_thread ();
		679	}
		680
		681	/*****************************************************************************/
		682
		683	static void
		684	grp_try_feed (eio_req *grp)
		685	{
		686	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		687	{
		688	grp->flags &= ~EIO_FLAG_GROUPADD;
		689
		690	EIO_FEED (grp);
		691
		692	/* stop if no progress has been made */
		693	if (!(grp->flags & EIO_FLAG_GROUPADD))
		694	{
		695	grp->feed = 0;
		696	break;
		697	}
		698	}
		699	}
		700
		701	static int
		702	grp_dec (eio_req *grp)
		703	{
		704	--grp->size;
		705
		706	/* call feeder, if applicable */
		707	grp_try_feed (grp);
		708
		709	/* finish, if done */
		710	if (!grp->size && grp->int1)
		711	return eio_finish (grp);
		712	else
		713	return 0;
		714	}
		715
		716	void
		717	eio_destroy (eio_req *req)
		718	{
		719	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		720	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		721
		722	EIO_DESTROY (req);
		723	}
		724
		725	static int
		726	eio_finish (eio_req *req)
		727	{
		728	int res = EIO_FINISH (req);
		729
		730	if (req->grp)
		731	{
		732	int res2;
		733	eio_req *grp = req->grp;
		734
		735	/* unlink request */
		736	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		737	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		738
		739	if (grp->grp_first == req)
		740	grp->grp_first = req->grp_next;
		741
		742	res2 = grp_dec (grp);
		743
		744	if (!res && res2)
		745	res = res2;
		746	}
		747
		748	eio_destroy (req);
		749
		750	return res;
		751	}
		752
		753	void
		754	eio_grp_cancel (eio_req *grp)
		755	{
		756	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		757	eio_cancel (grp);
		758	}
		759
		760	void
		761	eio_cancel (eio_req *req)
		762	{
		763	etp_cancel (req);
		764	}
		765
		766	void
		767	eio_submit (eio_req *req)
		768	{
		769	etp_submit (req);
		770	}
		771
		772	unsigned int
		773	eio_nreqs (void)
		774	{
		775	return etp_nreqs ();
		776	}
		777
		778	unsigned int
		779	eio_nready (void)
		780	{
		781	return etp_nready ();
		782	}
		783
		784	unsigned int
		785	eio_npending (void)
		786	{
		787	return etp_npending ();
		788	}
		789
		790	unsigned int ecb_cold
		791	eio_nthreads (void)
		792	{
		793	return etp_nthreads ();
		794	}
		795
		796	void ecb_cold
		797	eio_set_max_poll_time (double nseconds)
		798	{
		799	etp_set_max_poll_time (nseconds);
		800	}
		801
		802	void ecb_cold
		803	eio_set_max_poll_reqs (unsigned int maxreqs)
		804	{
		805	etp_set_max_poll_reqs (maxreqs);
		806	}
		807
		808	void ecb_cold
		809	eio_set_max_idle (unsigned int nthreads)
		810	{
		811	etp_set_max_idle (nthreads);
		812	}
		813
		814	void ecb_cold
		815	eio_set_idle_timeout (unsigned int seconds)
		816	{
		817	etp_set_idle_timeout (seconds);
		818	}
		819
		820	void ecb_cold
		821	eio_set_min_parallel (unsigned int nthreads)
		822	{
		823	etp_set_min_parallel (nthreads);
		824	}
		825
		826	void ecb_cold
		827	eio_set_max_parallel (unsigned int nthreads)
		828	{
		829	etp_set_max_parallel (nthreads);
		830	}
		831
		832	int eio_poll (void)
		833	{
		834	return etp_poll ();
		835	}
		836
489	/*****************************************************************************/	837	/*****************************************************************************/
490	/* work around various missing functions */	838	/* work around various missing functions */
491		839
492	#if !HAVE_PREADWRITE	840	#if !HAVE_PREADWRITE
		841	# undef pread
		842	# undef pwrite
493	# define pread eio__pread	843	# define pread eio__pread
494	# define pwrite eio__pwrite	844	# define pwrite eio__pwrite
495
496	/*
497	* make our pread/pwrite safe against themselves, but not against
498	* normal read/write by using a mutex. slows down execution a lot,
499	* but that's your problem, not mine.
500	*/
501	static mutex_t preadwritelock = X_MUTEX_INIT;
502		845
503	static ssize_t	846	static ssize_t
504	eio__pread (int fd, void *buf, size_t count, off_t offset)	847	eio__pread (int fd, void *buf, size_t count, off_t offset)
505	{	848	{
506	ssize_t res;	849	ssize_t res;
…		…
524		867
525	X_LOCK (preadwritelock);	868	X_LOCK (preadwritelock);
526	ooffset = lseek (fd, 0, SEEK_CUR);	869	ooffset = lseek (fd, 0, SEEK_CUR);
527	lseek (fd, offset, SEEK_SET);	870	lseek (fd, offset, SEEK_SET);
528	res = write (fd, buf, count);	871	res = write (fd, buf, count);
529	lseek (fd, offset, SEEK_SET);	872	lseek (fd, ooffset, SEEK_SET);
530	X_UNLOCK (preadwritelock);	873	X_UNLOCK (preadwritelock);
531		874
532	return res;	875	return res;
533	}	876	}
534	#endif	877	#endif
535		878
536	#ifndef HAVE_FUTIMES	879	#ifndef HAVE_UTIMES
537		880
		881	# undef utimes
538	# define utimes(path,times) eio__utimes (path, times)	882	# define utimes(path,times) eio__utimes (path, times)
539	# define futimes(fd,times) eio__futimes (fd, times)
540		883
541	static int	884	static int
542	eio__utimes (const char *filename, const struct timeval times[2])	885	eio__utimes (const char *filename, const struct timeval times[2])
543	{	886	{
544	if (times)	887	if (times)
…		…
552	}	895	}
553	else	896	else
554	return utime (filename, 0);	897	return utime (filename, 0);
555	}	898	}
556		899
		900	#endif
		901
		902	#ifndef HAVE_FUTIMES
		903
		904	# undef futimes
		905	# define futimes(fd,times) eio__futimes (fd, times)
		906
		907	static int
557	static int eio__futimes (int fd, const struct timeval tv[2])	908	eio__futimes (int fd, const struct timeval tv[2])
558	{	909	{
559	errno = ENOSYS;	910	errno = ENOSYS;
560	return -1;	911	return -1;
561	}	912	}
562		913
563	#endif	914	#endif
564		915
565	#if !HAVE_FDATASYNC	916	#if !HAVE_FDATASYNC
		917	# undef fdatasync
566	# define fdatasync fsync	918	# define fdatasync(fd) fsync (fd)
567	#endif	919	#endif
		920
		921	/* sync_file_range always needs emulation */
		922	static int
		923	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		924	{
		925	#if HAVE_SYNC_FILE_RANGE
		926	int res;
		927
		928	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		929	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		930	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		931	{
		932	flags = 0
		933	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		934	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		935	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		936	}
		937
		938	res = sync_file_range (fd, offset, nbytes, flags);
		939
		940	if (!res || errno != ENOSYS)
		941	return res;
		942	#endif
		943
		944	/* even though we could play tricks with the flags, it's better to always
		945	* call fdatasync, as that matches the expectation of its users best */
		946	return fdatasync (fd);
		947	}
568		948
569	#if !HAVE_READAHEAD	949	#if !HAVE_READAHEAD
		950	# undef readahead
570	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	951	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
571		952
572	static ssize_t	953	static ssize_t
573	eio__readahead (int fd, off_t offset, size_t count, worker *self)	954	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
574	{	955	{
575	size_t todo = count;	956	size_t todo = count;
576	dBUF;	957	dBUF;
577		958
578	while (todo > 0)	959	while (todo > 0)
…		…
588	return count;	969	return count;
589	}	970	}
590		971
591	#endif	972	#endif
592		973
593	#if !HAVE_READDIR_R
594	# define readdir_r eio__readdir_r
595
596	static mutex_t readdirlock = X_MUTEX_INIT;
597
598	static int
599	eio__readdir_r (DIR *dirp, EIO_STRUCT_DIRENT *ent, EIO_STRUCT_DIRENT **res)
600	{
601	EIO_STRUCT_DIRENT *e;
602	int errorno;
603
604	X_LOCK (readdirlock);
605
606	e = readdir (dirp);
607	errorno = errno;
608
609	if (e)
610	{
611	*res = ent;
612	strcpy (ent->d_name, e->d_name);
613	}
614	else
615	*res = 0;
616
617	X_UNLOCK (readdirlock);
618
619	errno = errorno;
620	return e ? 0 : -1;
621	}
622	#endif
623
624	/* sendfile always needs emulation */	974	/* sendfile always needs emulation */
625	static ssize_t	975	static ssize_t
626	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	976	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
627	{	977	{
		978	ssize_t written = 0;
628	ssize_t res;	979	ssize_t res;
629		980
630	if (!count)	981	if (!count)
631	return 0;	982	return 0;
632		983
		984	for (;;)
		985	{
633	#if HAVE_SENDFILE	986	#if HAVE_SENDFILE
634	# if __linux	987	# if __linux
		988	off_t soffset = offset;
635	res = sendfile (ofd, ifd, &offset, count);	989	res = sendfile (ofd, ifd, &soffset, count);
636		990
637	# elif __freebsd	991	# elif __FreeBSD__
638	/*	992	/*
639	* Of course, the freebsd sendfile is a dire hack with no thoughts	993	* Of course, the freebsd sendfile is a dire hack with no thoughts
640	* wasted on making it similar to other I/O functions.	994	* wasted on making it similar to other I/O functions.
641	*/	995	*/
642	{
643	off_t sbytes;	996	off_t sbytes;
644	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	997	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
645		998
646	if (res < 0 && sbytes)	999	#if 0 /* according to the manpage, this is correct, but broken behaviour */
647	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	1000	/* freebsd' sendfile will return 0 on success */
		1001	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		1002	/* not on e.g. EIO or EPIPE - sounds broken */
		1003	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
648	res = sbytes;	1004	res = sbytes;
649	}	1005	#endif
		1006
		1007	/* according to source inspection, this is correct, and useful behaviour */
		1008	if (sbytes)
		1009	res = sbytes;
		1010
		1011	# elif defined (__APPLE__)
		1012	off_t sbytes = count;
		1013	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		1014
		1015	/* according to the manpage, sbytes is always valid */
		1016	if (sbytes)
		1017	res = sbytes;
650		1018
651	# elif __hpux	1019	# elif __hpux
652	res = sendfile (ofd, ifd, offset, count, 0, 0);	1020	res = sendfile (ofd, ifd, offset, count, 0, 0);
653		1021
654	# elif __solaris	1022	# elif __solaris
655	{
656	struct sendfilevec vec;	1023	struct sendfilevec vec;
657	size_t sbytes;	1024	size_t sbytes;
658		1025
659	vec.sfv_fd = ifd;	1026	vec.sfv_fd = ifd;
660	vec.sfv_flag = 0;	1027	vec.sfv_flag = 0;
661	vec.sfv_off = offset;	1028	vec.sfv_off = offset;
662	vec.sfv_len = count;	1029	vec.sfv_len = count;
663		1030
664	res = sendfilev (ofd, &vec, 1, &sbytes);	1031	res = sendfilev (ofd, &vec, 1, &sbytes);
665		1032
666	if (res < 0 && sbytes)	1033	if (res < 0 && sbytes)
667	res = sbytes;	1034	res = sbytes;
668	}
669		1035
670	# endif	1036	# endif
		1037
		1038	#elif defined (_WIN32)
		1039	/* does not work, just for documentation of what would need to be done */
		1040	/* actually, cannot be done like this, as TransmitFile changes the file offset, */
		1041	/* libeio guarantees that the file offset does not change, and windows */
		1042	/* has no way to get an independent handle to the same file description */
		1043	HANDLE h = TO_SOCKET (ifd);
		1044	SetFilePointer (h, offset, 0, FILE_BEGIN);
		1045	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		1046
671	#else	1047	#else
672	res = -1;	1048	res = -1;
673	errno = ENOSYS;	1049	errno = ENOSYS;
674	#endif	1050	#endif
675		1051
		1052	/* we assume sendfile can copy at least 128mb in one go */
		1053	if (res <= 128 * 1024 * 1024)
		1054	{
		1055	if (res > 0)
		1056	written += res;
		1057
		1058	if (written)
		1059	return written;
		1060
		1061	break;
		1062	}
		1063	else
		1064	{
		1065	/* if we requested more, then probably the kernel was lazy */
		1066	written += res;
		1067	offset += res;
		1068	count -= res;
		1069
		1070	if (!count)
		1071	return written;
		1072	}
		1073	}
		1074
676	if (res < 0	1075	if (res < 0
677	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	1076	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		1077	/* BSDs */
		1078	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		1079	|| errno == ENOTSUP
		1080	#endif
		1081	|| errno == EOPNOTSUPP /* BSDs */
678	#if __solaris	1082	#if __solaris
679	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	1083	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
680	#endif	1084	#endif
681	)	1085	)
682	)	1086	)
…		…
713	}	1117	}
714		1118
715	return res;	1119	return res;
716	}	1120	}
717		1121
		1122	static signed char
		1123	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1124	{
		1125	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1126	: a->inode < b->inode ? -1
		1127	: a->inode > b->inode ? 1
		1128	: 0;
		1129	}
		1130
		1131	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1132
		1133	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1134	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1135
		1136	static void
		1137	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1138	{
		1139	unsigned char bits [9 + sizeof (ino_t) * 8];
		1140	unsigned char *bit = bits;
		1141
		1142	assert (CHAR_BIT == 8);
		1143	assert (sizeof (eio_dirent) * 8 < 256);
		1144	assert (offsetof (eio_dirent, inode)); /* we use bit #0 as sentinel */
		1145	assert (offsetof (eio_dirent, score)); /* we use bit #0 as sentinel */
		1146
		1147	if (size <= EIO_SORT_FAST)
		1148	return;
		1149
		1150	/* first prepare an array of bits to test in our radix sort */
		1151	/* try to take endianness into account, as well as differences in ino_t sizes */
		1152	/* inode_bits must contain all inodes ORed together */
		1153	/* which is used to skip bits that are 0 everywhere, which is very common */
		1154	{
		1155	ino_t endianness;
		1156	int i, j;
		1157
		1158	/* we store the byte offset of byte n into byte n of "endianness" */
		1159	for (i = 0; i < sizeof (ino_t); ++i)
		1160	((unsigned char *)&endianness)[i] = i;
		1161
		1162	*bit++ = 0;
		1163
		1164	for (i = 0; i < sizeof (ino_t); ++i)
		1165	{
		1166	/* shifting off the byte offsets out of "endianness" */
		1167	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1168	endianness >>= 8;
		1169
		1170	for (j = 0; j < 8; ++j)
		1171	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1172	*bit++ = offs + j;
		1173	}
		1174
		1175	for (j = 0; j < 8; ++j)
		1176	if (score_bits & (1 << j))
		1177	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1178	}
		1179
		1180	/* now actually do the sorting (a variant of MSD radix sort) */
		1181	{
		1182	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1183	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1184	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1185	int stk_idx = 0;
		1186
		1187	base_stk [stk_idx] = dents;
		1188	end_stk [stk_idx] = dents + size;
		1189	bit_stk [stk_idx] = bit - 1;
		1190
		1191	do
		1192	{
		1193	base = base_stk [stk_idx];
		1194	end = end_stk [stk_idx];
		1195	bit = bit_stk [stk_idx];
		1196
		1197	for (;;)
		1198	{
		1199	unsigned char O = *bit >> 3;
		1200	unsigned char M = 1 << (*bit & 7);
		1201
		1202	eio_dirent *a = base;
		1203	eio_dirent *b = end;
		1204
		1205	if (b - a < EIO_SORT_CUTOFF)
		1206	break;
		1207
		1208	/* now bit-partition the array on the bit */
		1209	/* this ugly asymmetric loop seems to perform much better than typical */
		1210	/* partition algos found in the literature */
		1211	do
		1212	if (!(((unsigned char *)a)[O] & M))
		1213	++a;
		1214	else if (!(((unsigned char *)--b)[O] & M))
		1215	{
		1216	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1217	++a;
		1218	}
		1219	while (b > a);
		1220
		1221	/* next bit, or stop, if no bits left in this path */
		1222	if (!*--bit)
		1223	break;
		1224
		1225	base_stk [stk_idx] = a;
		1226	end_stk [stk_idx] = end;
		1227	bit_stk [stk_idx] = bit;
		1228	++stk_idx;
		1229
		1230	end = a;
		1231	}
		1232	}
		1233	while (stk_idx--);
		1234	}
		1235	}
		1236
		1237	static void
		1238	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1239	{
		1240	/* first move the smallest element to the front, to act as a sentinel */
		1241	{
		1242	int i;
		1243	eio_dirent *min = dents;
		1244
		1245	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1246	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1247	if (EIO_DENT_CMP (dents [i], <, *min))
		1248	min = &dents [i];
		1249
		1250	/* swap elements 0 and j (minimum) */
		1251	{
		1252	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1253	}
		1254	}
		1255
		1256	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1257	{
		1258	eio_dirent *i, *j;
		1259
		1260	for (i = dents + 1; i < dents + size; ++i)
		1261	{
		1262	eio_dirent value = *i;
		1263
		1264	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1265	j [1] = j [0];
		1266
		1267	j [1] = value;
		1268	}
		1269	}
		1270	}
		1271
		1272	static void
		1273	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1274	{
		1275	if (size <= 1)
		1276	return; /* our insertion sort relies on size > 0 */
		1277
		1278	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1279	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1280	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1281
		1282	/* use an insertion sort at the end, or for small arrays, */
		1283	/* as insertion sort is more efficient for small partitions */
		1284	eio_dent_insertion_sort (dents, size);
		1285	}
		1286
718	/* read a full directory */	1287	/* read a full directory */
719	static void	1288	static void
720	eio__scandir (eio_req *req, worker *self)	1289	eio__scandir (eio_req *req, etp_worker *self)
721	{	1290	{
722	DIR *dirp;	1291	DIR *dirp;
723	union
724	{
725	EIO_STRUCT_DIRENT d;
726	char b [offsetof (EIO_STRUCT_DIRENT, d_name) + NAME_MAX + 1];
727	} *u;
728	EIO_STRUCT_DIRENT *entp;	1292	EIO_STRUCT_DIRENT *entp;
729	char *name, *names;	1293	char *name, *names;
730	int memlen = 4096;	1294	int namesalloc = 4096;
731	int memofs = 0;	1295	int namesoffs = 0;
		1296	int flags = req->int1;
		1297	eio_dirent *dents = 0;
		1298	int dentalloc = 128;
732	int res = 0;	1299	int dentoffs = 0;
		1300	ino_t inode_bits = 0;
		1301
		1302	req->result = -1;
		1303
		1304	if (!(flags & EIO_READDIR_DENTS))
		1305	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
733		1306
734	X_LOCK (wrklock);	1307	X_LOCK (wrklock);
		1308	/* the corresponding closedir is in ETP_WORKER_CLEAR */
735	self->dirp = dirp = opendir (req->ptr1);	1309	self->dirp = dirp = opendir (req->ptr1);
736	self->dbuf = u = malloc (sizeof (*u));	1310
737	req->flags |= EIO_FLAG_PTR2_FREE;	1311	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1312	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
738	req->ptr2 = names = malloc (memlen);	1313	req->ptr2 = names = malloc (namesalloc);
739	X_UNLOCK (wrklock);	1314	X_UNLOCK (wrklock);
740		1315
741	if (dirp && u && names)	1316	if (dirp && names && (!flags || dents))
742	for (;;)	1317	for (;;)
743	{	1318	{
744	errno = 0;	1319	errno = 0;
745	readdir_r (dirp, &u->d, &entp);	1320	entp = readdir (dirp);
746		1321
747	if (!entp)	1322	if (!entp)
		1323	{
		1324	if (errno)
		1325	break;
		1326
		1327	/* sort etc. */
		1328	req->int1 = flags;
		1329	req->result = dentoffs;
		1330
		1331	if (flags & EIO_READDIR_STAT_ORDER)
		1332	eio_dent_sort (dents, dentoffs, flags & EIO_READDIR_DIRS_FIRST ? 7 : 0, inode_bits);
		1333	else if (flags & EIO_READDIR_DIRS_FIRST)
		1334	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1335	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1336	else
		1337	{
		1338	/* in this case, all is known, and we just put dirs first and sort them */
		1339	eio_dirent *oth = dents + dentoffs;
		1340	eio_dirent *dir = dents;
		1341
		1342	/* now partition dirs to the front, and non-dirs to the back */
		1343	/* by walking from both sides and swapping if necessary */
		1344	while (oth > dir)
		1345	{
		1346	if (dir->type == EIO_DT_DIR)
		1347	++dir;
		1348	else if ((--oth)->type == EIO_DT_DIR)
		1349	{
		1350	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1351
		1352	++dir;
		1353	}
		1354	}
		1355
		1356	/* now sort the dirs only (dirs all have the same score) */
		1357	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1358	}
		1359
748	break;	1360	break;
		1361	}
749		1362
		1363	/* now add the entry to our list(s) */
750	name = entp->d_name;	1364	name = entp->d_name;
751		1365
		1366	/* skip . and .. entries */
752	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1367	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
753	{	1368	{
754	int len = strlen (name) + 1;	1369	int len = D_NAMLEN (entp) + 1;
755		1370
756	res++;	1371	while (ecb_expect_false (namesoffs + len > namesalloc))
757
758	while (memofs + len > memlen)
759	{	1372	{
760	memlen *= 2;	1373	namesalloc *= 2;
761	X_LOCK (wrklock);	1374	X_LOCK (wrklock);
762	req->ptr2 = names = realloc (names, memlen);	1375	req->ptr2 = names = realloc (names, namesalloc);
763	X_UNLOCK (wrklock);	1376	X_UNLOCK (wrklock);
764		1377
765	if (!names)	1378	if (!names)
766	break;	1379	break;
767	}	1380	}
768		1381
769	memcpy (names + memofs, name, len);	1382	memcpy (names + namesoffs, name, len);
		1383
		1384	if (dents)
		1385	{
		1386	struct eio_dirent *ent;
		1387
		1388	if (ecb_expect_false (dentoffs == dentalloc))
		1389	{
		1390	dentalloc *= 2;
		1391	X_LOCK (wrklock);
		1392	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1393	X_UNLOCK (wrklock);
		1394
		1395	if (!dents)
		1396	break;
		1397	}
		1398
		1399	ent = dents + dentoffs;
		1400
		1401	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1402	ent->namelen = len - 1;
		1403	ent->inode = D_INO (entp);
		1404
		1405	inode_bits |= ent->inode;
		1406
		1407	switch (D_TYPE (entp))
		1408	{
		1409	default:
		1410	ent->type = EIO_DT_UNKNOWN;
		1411	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1412	break;
		1413
		1414	#ifdef DT_FIFO
		1415	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1416	#endif
		1417	#ifdef DT_CHR
		1418	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1419	#endif
		1420	#ifdef DT_MPC
		1421	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1422	#endif
		1423	#ifdef DT_DIR
		1424	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1425	#endif
		1426	#ifdef DT_NAM
		1427	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1428	#endif
		1429	#ifdef DT_BLK
		1430	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1431	#endif
		1432	#ifdef DT_MPB
		1433	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1434	#endif
		1435	#ifdef DT_REG
		1436	case DT_REG: ent->type = EIO_DT_REG; break;
		1437	#endif
		1438	#ifdef DT_NWK
		1439	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1440	#endif
		1441	#ifdef DT_CMP
		1442	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1443	#endif
		1444	#ifdef DT_LNK
		1445	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1446	#endif
		1447	#ifdef DT_SOCK
		1448	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1449	#endif
		1450	#ifdef DT_DOOR
		1451	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1452	#endif
		1453	#ifdef DT_WHT
		1454	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1455	#endif
		1456	}
		1457
		1458	ent->score = 7;
		1459
		1460	if (flags & EIO_READDIR_DIRS_FIRST)
		1461	{
		1462	if (ent->type == EIO_DT_UNKNOWN)
		1463	{
		1464	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1465	ent->score = 1;
		1466	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1467	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1468	}
		1469	else if (ent->type == EIO_DT_DIR)
		1470	ent->score = 0;
		1471	}
		1472	}
		1473
770	memofs += len;	1474	namesoffs += len;
		1475	++dentoffs;
		1476	}
		1477
		1478	if (EIO_CANCELLED (req))
		1479	{
		1480	errno = ECANCELED;
		1481	break;
771	}	1482	}
772	}	1483	}
		1484	}
773		1485
774	if (errno)	1486	#ifdef PAGESIZE
775	res = -1;	1487	# define eio_pagesize() PAGESIZE
		1488	#else
		1489	static intptr_t
		1490	eio_pagesize (void)
		1491	{
		1492	static intptr_t page;
		1493
		1494	if (!page)
		1495	page = sysconf (_SC_PAGESIZE);
		1496
		1497	return page;
		1498	}
		1499	#endif
		1500
		1501	static void
		1502	eio_page_align (void **addr, size_t *length)
		1503	{
		1504	intptr_t mask = eio_pagesize () - 1;
		1505
		1506	/* round down addr */
		1507	intptr_t adj = mask & (intptr_t)*addr;
		1508
		1509	*addr = (void *)((intptr_t)*addr - adj);
		1510	*length += adj;
		1511
		1512	/* round up length */
		1513	*length = (*length + mask) & ~mask;
		1514	}
		1515
		1516	#if !_POSIX_MEMLOCK
		1517	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1518	#else
		1519
		1520	static int
		1521	eio__mlockall (int flags)
		1522	{
		1523	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1524	extern int mallopt (int, int);
		1525	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1526	#endif
		1527
		1528	if (EIO_MCL_CURRENT != MCL_CURRENT
		1529	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1530	{
		1531	flags = 0
		1532	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1533	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1534	}
		1535
		1536	return mlockall (flags);
		1537	}
		1538	#endif
		1539
		1540	#if !_POSIX_MEMLOCK_RANGE
		1541	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1542	#else
		1543
		1544	static int
		1545	eio__mlock (void *addr, size_t length)
		1546	{
		1547	eio_page_align (&addr, &length);
		1548
		1549	return mlock (addr, length);
		1550	}
		1551
		1552	#endif
		1553
		1554	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1555	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1556	#else
		1557
		1558	static int
		1559	eio__msync (void *mem, size_t len, int flags)
		1560	{
		1561	eio_page_align (&mem, &len);
		1562
		1563	if (EIO_MS_ASYNC != MS_SYNC
		1564	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1565	|| EIO_MS_SYNC != MS_SYNC)
		1566	{
		1567	flags = 0
		1568	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1569	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1570	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1571	}
		1572
		1573	return msync (mem, len, flags);
		1574	}
		1575
		1576	#endif
		1577
		1578	static int
		1579	eio__mtouch (eio_req *req)
		1580	{
		1581	void *mem = req->ptr2;
		1582	size_t len = req->size;
		1583	int flags = req->int1;
		1584
		1585	eio_page_align (&mem, &len);
		1586
776		1587	{
777	req->result = res;	1588	intptr_t addr = (intptr_t)mem;
		1589	intptr_t end = addr + len;
		1590	intptr_t page = eio_pagesize ();
		1591
		1592	if (addr < end)
		1593	if (flags & EIO_MT_MODIFY) /* modify */
		1594	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len && !EIO_CANCELLED (req));
		1595	else
		1596	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len && !EIO_CANCELLED (req));
		1597	}
		1598
		1599	return 0;
778	}	1600	}
779		1601
780	/*****************************************************************************/	1602	/*****************************************************************************/
781		1603
782	#define ALLOC(len) \	1604	#define ALLOC(len) \
…		…
792	req->result = -1; \	1614	req->result = -1; \
793	break; \	1615	break; \
794	} \	1616	} \
795	}	1617	}
796		1618
797	X_THREAD_PROC (eio_proc)	1619	X_THREAD_PROC (etp_proc)
798	{	1620	{
799	eio_req *req;	1621	ETP_REQ *req;
800	struct timespec ts;	1622	struct timespec ts;
801	worker *self = (worker *)thr_arg;	1623	etp_worker *self = (etp_worker *)thr_arg;
802		1624
803	/* try to distribute timeouts somewhat randomly */	1625	/* try to distribute timeouts somewhat randomly */
804	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1626	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
805		1627
806	for (;;)	1628	for (;;)
807	{	1629	{
808	ts.tv_sec = time (0) + IDLE_TIMEOUT;
809
810	X_LOCK (reqlock);	1630	X_LOCK (reqlock);
811		1631
812	for (;;)	1632	for (;;)
813	{	1633	{
814	self->req = req = reqq_shift (&req_queue);	1634	self->req = req = reqq_shift (&req_queue);
…		…
816	if (req)	1636	if (req)
817	break;	1637	break;
818		1638
819	++idle;	1639	++idle;
820		1640
		1641	ts.tv_sec = time (0) + idle_timeout;
821	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1642	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
822	{	1643	{
823	if (idle > max_idle)	1644	if (idle > max_idle)
824	{	1645	{
825	--idle;	1646	--idle;
…		…
830	goto quit;	1651	goto quit;
831	}	1652	}
832		1653
833	/* we are allowed to idle, so do so without any timeout */	1654	/* we are allowed to idle, so do so without any timeout */
834	X_COND_WAIT (reqwait, reqlock);	1655	X_COND_WAIT (reqwait, reqlock);
835	ts.tv_sec = time (0) + IDLE_TIMEOUT;
836	}	1656	}
837		1657
838	--idle;	1658	--idle;
839	}	1659	}
840		1660
841	--nready;	1661	--nready;
842		1662
843	X_UNLOCK (reqlock);	1663	X_UNLOCK (reqlock);
844		1664
845	errno = 0; /* strictly unnecessary */	1665	if (req->type < 0)
		1666	goto quit;
846		1667
847	if (!EIO_CANCELLED (req))	1668	if (!EIO_CANCELLED (req))
848	switch (req->type)	1669	ETP_EXECUTE (self, req);
849	{
850	case EIO_READ: ALLOC (req->size);
851	req->result = req->offs >= 0
852	? pread (req->int1, req->ptr2, req->size, req->offs)
853	: read (req->int1, req->ptr2, req->size); break;
854	case EIO_WRITE: req->result = req->offs >= 0
855	? pwrite (req->int1, req->ptr2, req->size, req->offs)
856	: write (req->int1, req->ptr2, req->size); break;
857
858	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
859	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
860
861	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
862	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
863	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
864	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
865	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
866	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
867
868	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
869	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
870	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
871	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
872	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
873	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
874
875	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
876	case EIO_CLOSE: req->result = close (req->int1); break;
877	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
878	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
879	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
880	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
881	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
882	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
883	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
884	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
885
886	case EIO_READLINK: ALLOC (NAME_MAX);
887	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
888
889	case EIO_SYNC: req->result = 0; sync (); break;
890	case EIO_FSYNC: req->result = fsync (req->int1); break;
891	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
892
893	case EIO_READDIR: eio__scandir (req, self); break;
894
895	case EIO_BUSY:
896	#ifdef _WIN32
897	Sleep (req->nv1 * 1000.);
898	#else
899	{
900	struct timeval tv;
901
902	tv.tv_sec = req->nv1;
903	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.;
904
905	req->result = select (0, 0, 0, 0, &tv);
906	}
907	#endif
908	break;
909
910	case EIO_UTIME:
911	case EIO_FUTIME:
912	{
913	struct timeval tv[2];
914	struct timeval *times;
915
916	if (req->nv1 != -1. || req->nv2 != -1.)
917	{
918	tv[0].tv_sec = req->nv1;
919	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
920	tv[1].tv_sec = req->nv2;
921	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
922
923	times = tv;
924	}
925	else
926	times = 0;
927
928
929	req->result = req->type == EIO_FUTIME
930	? futimes (req->int1, times)
931	: utimes (req->ptr1, times);
932	}
933
934	case EIO_GROUP:
935	case EIO_NOP:
936	req->result = 0;
937	break;
938
939	case EIO_QUIT:
940	goto quit;
941
942	default:
943	req->result = -1;
944	break;
945	}
946
947	req->errorno = errno;
948		1670
949	X_LOCK (reslock);	1671	X_LOCK (reslock);
950		1672
951	++npending;	1673	++npending;
952		1674
953	if (!reqq_push (&res_queue, req) && want_poll_cb)	1675	if (!reqq_push (&res_queue, req) && want_poll_cb)
954	want_poll_cb ();	1676	want_poll_cb ();
955		1677
956	self->req = 0;	1678	self->req = 0;
957	worker_clear (self);	1679	etp_worker_clear (self);
958		1680
959	X_UNLOCK (reslock);	1681	X_UNLOCK (reslock);
960	}	1682	}
961		1683
962	quit:	1684	quit:
963	X_LOCK (wrklock);	1685	X_LOCK (wrklock);
964	worker_free (self);	1686	etp_worker_free (self);
965	X_UNLOCK (wrklock);	1687	X_UNLOCK (wrklock);
966		1688
967	return 0;	1689	return 0;
968	}	1690	}
969		1691
970	/*****************************************************************************/	1692	/*****************************************************************************/
971		1693
972	static void eio_atfork_prepare (void)	1694	int ecb_cold
973	{
974	X_LOCK (wrklock);
975	X_LOCK (reqlock);
976	X_LOCK (reslock);
977	#if !HAVE_PREADWRITE
978	X_LOCK (preadwritelock);
979	#endif
980	#if !HAVE_READDIR_R
981	X_LOCK (readdirlock);
982	#endif
983	}
984
985	static void eio_atfork_parent (void)
986	{
987	#if !HAVE_READDIR_R
988	X_UNLOCK (readdirlock);
989	#endif
990	#if !HAVE_PREADWRITE
991	X_UNLOCK (preadwritelock);
992	#endif
993	X_UNLOCK (reslock);
994	X_UNLOCK (reqlock);
995	X_UNLOCK (wrklock);
996	}
997
998	static void eio_atfork_child (void)
999	{
1000	eio_req *prv;
1001
1002	while (prv = reqq_shift (&req_queue))
1003	eio_destroy (prv);
1004
1005	while (prv = reqq_shift (&res_queue))
1006	eio_destroy (prv);
1007
1008	while (wrk_first.next != &wrk_first)
1009	{
1010	worker *wrk = wrk_first.next;
1011
1012	if (wrk->req)
1013	eio_destroy (wrk->req);
1014
1015	worker_clear (wrk);
1016	worker_free (wrk);
1017	}
1018
1019	started = 0;
1020	idle = 0;
1021	nreqs = 0;
1022	nready = 0;
1023	npending = 0;
1024
1025	eio_atfork_parent ();
1026	}
1027
1028	int eio_init (void (*want_poll)(void), void (*done_poll)(void))	1695	eio_init (void (*want_poll)(void), void (*done_poll)(void))
1029	{	1696	{
1030	want_poll_cb = want_poll;	1697	return etp_init (want_poll, done_poll);
1031	done_poll_cb = done_poll;
1032
1033	#ifdef _WIN32
1034	X_MUTEX_CHECK (wrklock);
1035	X_MUTEX_CHECK (reslock);
1036	X_MUTEX_CHECK (reqlock);
1037	X_MUTEX_CHECK (reqwait);
1038	X_MUTEX_CHECK (preadwritelock);
1039	X_MUTEX_CHECK (readdirlock);
1040
1041	X_COND_CHECK (reqwait);
1042	#endif
1043
1044	X_THREAD_ATFORK (eio_atfork_prepare, eio_atfork_parent, eio_atfork_child);
1045	}	1698	}
1046		1699
		1700	ECB_INLINE void
1047	static void eio_api_destroy (eio_req *req)	1701	eio_api_destroy (eio_req *req)
1048	{	1702	{
1049	free (req);	1703	free (req);
1050	}	1704	}
1051		1705
1052	#define REQ(rtype) \	1706	#define REQ(rtype) \
…		…
1071	{ \	1725	{ \
1072	eio_api_destroy (req); \	1726	eio_api_destroy (req); \
1073	return 0; \	1727	return 0; \
1074	}	1728	}
1075		1729
		1730	static void
		1731	eio_execute (etp_worker *self, eio_req *req)
		1732	{
		1733	switch (req->type)
		1734	{
		1735	case EIO_READ: ALLOC (req->size);
		1736	req->result = req->offs >= 0
		1737	? pread (req->int1, req->ptr2, req->size, req->offs)
		1738	: read (req->int1, req->ptr2, req->size); break;
		1739	case EIO_WRITE: req->result = req->offs >= 0
		1740	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1741	: write (req->int1, req->ptr2, req->size); break;
		1742
		1743	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1744	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1745
		1746	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1747	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1748	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1749	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1750	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1751	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1752
		1753	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1754	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1755	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1756	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1757
		1758	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1759	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1760	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1761	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1762	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1763	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1764
		1765	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1766	case EIO_CLOSE: req->result = close (req->int1); break;
		1767	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1768	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1769	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1770	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1771	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1772	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1773	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1774	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
		1775
		1776	case EIO_READLINK: ALLOC (PATH_MAX);
		1777	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1778
		1779	case EIO_SYNC: req->result = 0; sync (); break;
		1780	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1781	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1782	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1783	case EIO_MTOUCH: req->result = eio__mtouch (req); break;
		1784	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1785	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1786	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1787
		1788	case EIO_READDIR: eio__scandir (req, self); break;
		1789
		1790	case EIO_BUSY:
		1791	#ifdef _WIN32
		1792	Sleep (req->nv1 * 1e3);
		1793	#else
		1794	{
		1795	struct timeval tv;
		1796
		1797	tv.tv_sec = req->nv1;
		1798	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
		1799
		1800	req->result = select (0, 0, 0, 0, &tv);
		1801	}
		1802	#endif
		1803	break;
		1804
		1805	case EIO_UTIME:
		1806	case EIO_FUTIME:
		1807	{
		1808	struct timeval tv[2];
		1809	struct timeval *times;
		1810
		1811	if (req->nv1 != -1. || req->nv2 != -1.)
		1812	{
		1813	tv[0].tv_sec = req->nv1;
		1814	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1815	tv[1].tv_sec = req->nv2;
		1816	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1817
		1818	times = tv;
		1819	}
		1820	else
		1821	times = 0;
		1822
		1823	req->result = req->type == EIO_FUTIME
		1824	? futimes (req->int1, times)
		1825	: utimes (req->ptr1, times);
		1826	}
		1827	break;
		1828
		1829	case EIO_GROUP:
		1830	abort (); /* handled in eio_request */
		1831
		1832	case EIO_NOP:
		1833	req->result = 0;
		1834	break;
		1835
		1836	case EIO_CUSTOM:
		1837	req->feed (req);
		1838	break;
		1839
		1840	default:
		1841	errno = ENOSYS;
		1842	req->result = -1;
		1843	break;
		1844	}
		1845
		1846	req->errorno = errno;
		1847	}
		1848
		1849	#ifndef EIO_NO_WRAPPERS
		1850
1076	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1851	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1077	{	1852	{
1078	REQ (EIO_NOP); SEND;	1853	REQ (EIO_NOP); SEND;
1079	}	1854	}
1080		1855
…		…
1091	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1866	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1092	{	1867	{
1093	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1868	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1094	}	1869	}
1095		1870
		1871	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1872	{
		1873	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1874	}
		1875
		1876	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1877	{
		1878	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1879	}
		1880
		1881	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1882	{
		1883	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1884	}
		1885
		1886	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1887	{
		1888	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1889	}
		1890
		1891	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1892	{
		1893	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1894	}
		1895
1096	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1896	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1097	{	1897	{
1098	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1898	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1099	}	1899	}
1100		1900
…		…
1119	}	1919	}
1120		1920
1121	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1921	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1122	{	1922	{
1123	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1923	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1924	}
		1925
		1926	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1927	{
		1928	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1124	}	1929	}
1125		1930
1126	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1931	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1127	{	1932	{
1128	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1933	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
…		…
1202	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)	2007	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1203	{	2008	{
1204	return eio__1path (EIO_LSTAT, path, pri, cb, data);	2009	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1205	}	2010	}
1206		2011
		2012	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		2013	{
		2014	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		2015	}
		2016
1207	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	2017	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1208	{	2018	{
1209	return eio__1path (EIO_UNLINK, path, pri, cb, data);	2019	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1210	}	2020	}
1211		2021
1212	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	2022	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1213	{	2023	{
1214	return eio__1path (EIO_RMDIR, path, pri, cb, data);	2024	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1215	}	2025	}
1216		2026
1217	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	2027	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1218	{	2028	{
1219	return eio__1path (EIO_READDIR, path, pri, cb, data);	2029	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1220	}	2030	}
1221		2031
1222	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	2032	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1223	{	2033	{
1224	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	2034	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND;
1225	}	2035	}
1226		2036
1227	static eio_req *	2037	static eio_req *
1228	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	2038	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1229	{	2039	{
…		…
1253	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	2063	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1254	{	2064	{
1255	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	2065	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1256	}	2066	}
1257		2067
		2068	eio_req *eio_custom (void (*execute)(eio_req *), int pri, eio_cb cb, void *data)
		2069	{
		2070	REQ (EIO_CUSTOM); req->feed = execute; SEND;
		2071	}
		2072
		2073	#endif
		2074
1258	eio_req *eio_grp (eio_cb cb, void *data)	2075	eio_req *eio_grp (eio_cb cb, void *data)
1259	{	2076	{
1260	const int pri = EIO_PRI_MAX;	2077	const int pri = EIO_PRI_MAX;
1261		2078
1262	REQ (EIO_GROUP); SEND;	2079	REQ (EIO_GROUP); SEND;
…		…
1267	#undef SEND	2084	#undef SEND
1268		2085
1269	/*****************************************************************************/	2086	/*****************************************************************************/
1270	/* grp functions */	2087	/* grp functions */
1271		2088
		2089	void
1272	void eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit)	2090	eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit)
1273	{	2091	{
1274	grp->int2 = limit;	2092	grp->int2 = limit;
1275	grp->feed = feed;	2093	grp->feed = feed;
1276		2094
1277	grp_try_feed (grp);	2095	grp_try_feed (grp);
1278	}	2096	}
1279		2097
		2098	void
1280	void eio_grp_limit (eio_req *grp, int limit)	2099	eio_grp_limit (eio_req *grp, int limit)
1281	{	2100	{
1282	grp->int2 = limit;	2101	grp->int2 = limit;
1283		2102
1284	grp_try_feed (grp);	2103	grp_try_feed (grp);
1285	}	2104	}
1286		2105
		2106	void
1287	void eio_grp_add (eio_req *grp, eio_req *req)	2107	eio_grp_add (eio_req *grp, eio_req *req)
1288	{	2108	{
1289	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2109	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
		2110
		2111	grp->flags |= EIO_FLAG_GROUPADD;
1290		2112
1291	++grp->size;	2113	++grp->size;
1292	req->grp = grp;	2114	req->grp = grp;
1293		2115
1294	req->grp_prev = 0;	2116	req->grp_prev = 0;
…		…
1301	}	2123	}
1302		2124
1303	/*****************************************************************************/	2125	/*****************************************************************************/
1304	/* misc garbage */	2126	/* misc garbage */
1305		2127
		2128	ssize_t
1306	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2129	eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1307	{	2130	{
1308	worker wrk;	2131	etp_worker wrk;
		2132	ssize_t ret;
1309		2133
1310	wrk.dbuf = 0;	2134	wrk.dbuf = 0;
1311		2135
1312	eio__sendfile (ofd, ifd, offset, count, &wrk);	2136	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1313		2137
1314	if (wrk.dbuf)	2138	if (wrk.dbuf)
1315	free (wrk.dbuf);	2139	free (wrk.dbuf);
1316	}
1317		2140
		2141	return ret;
		2142	}
		2143

Diff Legend

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