ViewVC Help

View File | Revision Log | Show Annotations | Download File

/cvs/libeio/eio.c

Comparing libeio/eio.c (file contents):
Revision 1.10 by root, Tue May 13 17:08:15 2008 UTC vs.
Revision 1.68 by root, Sun Jun 5 22:44:30 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
		46	#ifdef EIO_STACKSIZE
		47	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		48	#endif
2	#include "xthread.h"	49	#include "xthread.h"
3		50
4	#include <errno.h>	51	#include <errno.h>
5	#include <stddef.h>	52	#include <stddef.h>
6	#include <stdlib.h>	53	#include <stdlib.h>
7	#include <string.h>	54	#include <string.h>
8	#include <errno.h>	55	#include <errno.h>
9	#include <sys/types.h>	56	#include <sys/types.h>
10	#include <sys/stat.h>	57	#include <sys/stat.h>
		58	#include <sys/statvfs.h>
11	#include <limits.h>	59	#include <limits.h>
12	#include <fcntl.h>	60	#include <fcntl.h>
13	#include <assert.h>	61	#include <assert.h>
14		62
		63	/* intptr_t comes from unistd.h, says POSIX/UNIX/tradition */
		64	/* intptr_t only comes form stdint.h, says idiot openbsd coder */
		65	#if HAVE_STDINT_H
		66	# include <stdint.h>
		67	#endif
		68
15	#ifndef EIO_FINISH	69	#ifndef EIO_FINISH
16	# define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0	70	# define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0
17	#endif	71	#endif
18		72
19	#ifndef EIO_DESTROY	73	#ifndef EIO_DESTROY
…		…
25	#endif	79	#endif
26		80
27	#ifdef _WIN32	81	#ifdef _WIN32
28		82
29	/*doh*/	83	/*doh*/
30
31	#else	84	#else
32		85
33	# include "config.h"
34	# include <sys/time.h>	86	# include <sys/time.h>
35	# include <sys/select.h>	87	# include <sys/select.h>
36	# include <unistd.h>	88	# include <unistd.h>
37	# include <utime.h>	89	# include <utime.h>
38	# include <signal.h>	90	# include <signal.h>
39	# include <dirent.h>	91	# include <dirent.h>
40		92
		93	#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
		94	# include <sys/mman.h>
		95	#endif
		96
		97	/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
		98	# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
		99	# define _DIRENT_HAVE_D_TYPE /* sigh */
		100	# define D_INO(de) (de)->d_fileno
		101	# define D_NAMLEN(de) (de)->d_namlen
		102	# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
		103	# define D_INO(de) (de)->d_ino
		104	# endif
		105
		106	#ifdef _D_EXACT_NAMLEN
		107	# undef D_NAMLEN
		108	# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
		109	#endif
		110
		111	# ifdef _DIRENT_HAVE_D_TYPE
		112	# define D_TYPE(de) (de)->d_type
		113	# endif
		114
41	# ifndef EIO_STRUCT_DIRENT	115	# ifndef EIO_STRUCT_DIRENT
42	# define EIO_STRUCT_DIRENT struct dirent	116	# define EIO_STRUCT_DIRENT struct dirent
43	# endif	117	# endif
44		118
45	#endif	119	#endif
46		120
47	#if HAVE_SENDFILE	121	#if HAVE_SENDFILE
48	# if __linux	122	# if __linux
49	# include <sys/sendfile.h>	123	# include <sys/sendfile.h>
50	# elif __freebsd	124	# elif __FreeBSD__ || defined __APPLE__
51	# include <sys/socket.h>	125	# include <sys/socket.h>
52	# include <sys/uio.h>	126	# include <sys/uio.h>
53	# elif __hpux	127	# elif __hpux
54	# include <sys/socket.h>	128	# include <sys/socket.h>
55	# elif __solaris /* not yet */	129	# elif __solaris
56	# include <sys/sendfile.h>	130	# include <sys/sendfile.h>
57	# else	131	# else
58	# error sendfile support requested but not available	132	# error sendfile support requested but not available
59	# endif	133	# endif
60	#endif	134	#endif
61		135
62	/* number of seconds after which an idle threads exit */	136	#ifndef D_TYPE
63	#define IDLE_TIMEOUT 10	137	# define D_TYPE(de) 0
		138	#endif
		139	#ifndef D_INO
		140	# define D_INO(de) 0
		141	#endif
		142	#ifndef D_NAMLEN
		143	# define D_NAMLEN(de) strlen ((de)->d_name)
		144	#endif
64		145
65	/* used for struct dirent, AIX doesn't provide it */	146	/* used for struct dirent, AIX doesn't provide it */
66	#ifndef NAME_MAX	147	#ifndef NAME_MAX
67	# define NAME_MAX 4096	148	# define NAME_MAX 4096
68	#endif	149	#endif
69		150
		151	/* used for readlink etc. */
		152	#ifndef PATH_MAX
		153	# define PATH_MAX 4096
		154	#endif
		155
70	/* buffer size for various temporary buffers */	156	/* buffer size for various temporary buffers */
71	#define EIO_BUFSIZE 65536	157	#define EIO_BUFSIZE 65536
72		158
73	#define dBUF \	159	#define dBUF \
74	char *eio_buf; \	160	char *eio_buf; \
75	X_LOCK (wrklock); \	161	ETP_WORKER_LOCK (self); \
76	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \	162	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
77	X_UNLOCK (wrklock); \	163	ETP_WORKER_UNLOCK (self); \
78	errno = ENOMEM; \	164	errno = ENOMEM; \
79	if (!eio_buf) \	165	if (!eio_buf) \
80	return -1;	166	return -1;
81		167
82	#define EIO_TICKS ((1000000 + 1023) >> 10)	168	#define EIO_TICKS ((1000000 + 1023) >> 10)
83		169
		170	/*****************************************************************************/
		171
		172	#if __GNUC__ >= 3
		173	# define expect(expr,value) __builtin_expect ((expr),(value))
		174	#else
		175	# define expect(expr,value) (expr)
		176	#endif
		177
		178	#define expect_false(expr) expect ((expr) != 0, 0)
		179	#define expect_true(expr) expect ((expr) != 0, 1)
		180
		181	/*****************************************************************************/
		182
		183	#define ETP_PRI_MIN EIO_PRI_MIN
		184	#define ETP_PRI_MAX EIO_PRI_MAX
		185
		186	struct etp_worker;
		187
		188	#define ETP_REQ eio_req
		189	#define ETP_DESTROY(req) eio_destroy (req)
		190	static int eio_finish (eio_req *req);
		191	#define ETP_FINISH(req) eio_finish (req)
		192	static void eio_execute (struct etp_worker *self, eio_req *req);
		193	#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
		194
		195	#define ETP_WORKER_CLEAR(req) \
		196	if (wrk->dbuf) \
		197	{ \
		198	free (wrk->dbuf); \
		199	wrk->dbuf = 0; \
		200	} \
		201	\
		202	if (wrk->dirp) \
		203	{ \
		204	closedir (wrk->dirp); \
		205	wrk->dirp = 0; \
		206	}
		207
		208	#define ETP_WORKER_COMMON \
		209	void *dbuf; \
		210	DIR *dirp;
		211
		212	/*****************************************************************************/
		213
		214	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
		215
		216	/* calculate time difference in ~1/EIO_TICKS of a second */
		217	static int tvdiff (struct timeval *tv1, struct timeval *tv2)
		218	{
		219	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
		220	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
		221	}
		222
		223	static unsigned int started, idle, wanted = 4;
		224
84	static void (*want_poll_cb) (void);	225	static void (*want_poll_cb) (void);
85	static void (*done_poll_cb) (void);	226	static void (*done_poll_cb) (void);
86		227
87	static unsigned int max_poll_time = 0;	228	static unsigned int max_poll_time; /* reslock */
88	static unsigned int max_poll_reqs = 0;	229	static unsigned int max_poll_reqs; /* reslock */
89		230
90	/* calculcate time difference in ~1/EIO_TICKS of a second */	231	static volatile unsigned int nreqs; /* reqlock */
91	static int tvdiff (struct timeval *tv1, struct timeval *tv2)	232	static volatile unsigned int nready; /* reqlock */
92	{	233	static volatile unsigned int npending; /* reqlock */
93	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS	234	static volatile unsigned int max_idle = 4; /* maximum number of threads that can idle indefinitely */
94	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);	235	static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */
95	}
96		236
97	static unsigned int started, idle, wanted = 4;	237	static xmutex_t wrklock;
		238	static xmutex_t reslock;
		239	static xmutex_t reqlock;
		240	static xcond_t reqwait;
		241
		242	#if !HAVE_PREADWRITE
		243	/*
		244	* make our pread/pwrite emulation safe against themselves, but not against
		245	* normal read/write by using a mutex. slows down execution a lot,
		246	* but that's your problem, not mine.
		247	*/
		248	static xmutex_t preadwritelock = X_MUTEX_INIT;
		249	#endif
		250
		251	typedef struct etp_worker
		252	{
		253	/* locked by wrklock */
		254	struct etp_worker *prev, *next;
		255
		256	xthread_t tid;
		257
		258	/* locked by reslock, reqlock or wrklock */
		259	ETP_REQ *req; /* currently processed request */
		260
		261	ETP_WORKER_COMMON
		262	} etp_worker;
		263
		264	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		265
		266	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		267	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
98		268
99	/* worker threads management */	269	/* worker threads management */
100	static mutex_t wrklock = X_MUTEX_INIT;
101		270
102	typedef struct worker
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)	271	static void etp_worker_clear (etp_worker *wrk)
118	{	272	{
119	if (wrk->dirp)	273	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	}	274	}
131		275
132	static void worker_free (worker *wrk)	276	static void etp_worker_free (etp_worker *wrk)
133	{	277	{
134	wrk->next->prev = wrk->prev;	278	wrk->next->prev = wrk->prev;
135	wrk->prev->next = wrk->next;	279	wrk->prev->next = wrk->next;
136		280
137	free (wrk);	281	free (wrk);
138	}	282	}
139		283
140	static volatile unsigned int nreqs, nready, npending;
141	static volatile unsigned int max_idle = 4;
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)	284	static unsigned int etp_nreqs (void)
148	{	285	{
		286	int retval;
		287	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		288	retval = nreqs;
		289	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
149	return nreqs;	290	return retval;
150	}	291	}
151		292
152	unsigned int eio_nready (void)	293	static unsigned int etp_nready (void)
153	{	294	{
154	unsigned int retval;	295	unsigned int retval;
155		296
156	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	297	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
157	retval = nready;	298	retval = nready;
158	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	299	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
159		300
160	return retval;	301	return retval;
161	}	302	}
162		303
163	unsigned int eio_npending (void)	304	static unsigned int etp_npending (void)
164	{	305	{
165	unsigned int retval;	306	unsigned int retval;
166		307
167	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	308	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
168	retval = npending;	309	retval = npending;
169	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	310	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
170		311
171	return retval;	312	return retval;
172	}	313	}
173		314
174	unsigned int eio_nthreads (void)	315	static unsigned int etp_nthreads (void)
175	{	316	{
176	unsigned int retval;	317	unsigned int retval;
177		318
178	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	319	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
179	retval = started;	320	retval = started;
…		…
186	* a somewhat faster data structure might be nice, but	327	* a somewhat faster data structure might be nice, but
187	* with 8 priorities this actually needs <20 insns	328	* with 8 priorities this actually needs <20 insns
188	* per shift, the most expensive operation.	329	* per shift, the most expensive operation.
189	*/	330	*/
190	typedef struct {	331	typedef struct {
191	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	332	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
192	int size;	333	int size;
193	} reqq;	334	} etp_reqq;
194		335
195	static reqq req_queue;	336	static etp_reqq req_queue;
196	static reqq res_queue;	337	static etp_reqq res_queue;
197		338
198	static int reqq_push (reqq *q, eio_req *req)	339	static int reqq_push (etp_reqq *q, ETP_REQ *req)
199	{	340	{
200	int pri = req->pri;	341	int pri = req->pri;
201	req->next = 0;	342	req->next = 0;
202		343
203	if (q->qe[pri])	344	if (q->qe[pri])
…		…
209	q->qe[pri] = q->qs[pri] = req;	350	q->qe[pri] = q->qs[pri] = req;
210		351
211	return q->size++;	352	return q->size++;
212	}	353	}
213		354
214	static eio_req *reqq_shift (reqq *q)	355	static ETP_REQ *reqq_shift (etp_reqq *q)
215	{	356	{
216	int pri;	357	int pri;
217		358
218	if (!q->size)	359	if (!q->size)
219	return 0;	360	return 0;
220		361
221	--q->size;	362	--q->size;
222		363
223	for (pri = EIO_NUM_PRI; pri--; )	364	for (pri = ETP_NUM_PRI; pri--; )
224	{	365	{
225	eio_req *req = q->qs[pri];	366	eio_req *req = q->qs[pri];
226		367
227	if (req)	368	if (req)
228	{	369	{
…		…
234	}	375	}
235		376
236	abort ();	377	abort ();
237	}	378	}
238		379
239	static void grp_try_feed (eio_req *grp)	380	static void etp_thread_init (void)
240	{	381	{
241	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	382	X_MUTEX_CREATE (wrklock);
		383	X_MUTEX_CREATE (reslock);
		384	X_MUTEX_CREATE (reqlock);
		385	X_COND_CREATE (reqwait);
		386	}
		387
		388	static void etp_atfork_prepare (void)
		389	{
		390	X_LOCK (wrklock);
		391	X_LOCK (reqlock);
		392	X_LOCK (reslock);
		393	#if !HAVE_PREADWRITE
		394	X_LOCK (preadwritelock);
		395	#endif
		396	}
		397
		398	static void 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 etp_atfork_child (void)
		409	{
		410	ETP_REQ *prv;
		411
		412	while ((prv = reqq_shift (&req_queue)))
		413	ETP_DESTROY (prv);
		414
		415	while ((prv = reqq_shift (&res_queue)))
		416	ETP_DESTROY (prv);
		417
		418	while (wrk_first.next != &wrk_first)
242	{	419	{
243	int old_len = grp->size;	420	etp_worker *wrk = wrk_first.next;
244		421
245	EIO_FEED (grp);	422	if (wrk->req)
		423	ETP_DESTROY (wrk->req);
246		424
247	/* stop if no progress has been made */	425	etp_worker_clear (wrk);
248	if (old_len == grp->size)	426	etp_worker_free (wrk);
249	{
250	grp->feed = 0;
251	break;
252	}
253	}	427	}
254	}
255		428
256	static int eio_finish (eio_req *req);	429	started = 0;
		430	idle = 0;
		431	nreqs = 0;
		432	nready = 0;
		433	npending = 0;
257		434
258	static int grp_dec (eio_req *grp)	435	etp_thread_init ();
259	{	436	}
260	--grp->size;
261		437
262	/* call feeder, if applicable */	438	static void
263	grp_try_feed (grp);	439	etp_once_init (void)
		440	{
		441	etp_thread_init ();
		442	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
		443	}
264		444
265	/* finish, if done */	445	static int
266	if (!grp->size && grp->int1)	446	etp_init (void (*want_poll)(void), void (*done_poll)(void))
267	return eio_finish (grp);	447	{
268	else	448	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
		449
		450	pthread_once (&doinit, etp_once_init);
		451
		452	want_poll_cb = want_poll;
		453	done_poll_cb = done_poll;
		454
269	return 0;	455	return 0;
270	}	456	}
271		457
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);	458	X_THREAD_PROC (etp_proc);
323		459
324	static void start_thread (void)	460	static void etp_start_thread (void)
325	{	461	{
326	worker *wrk = calloc (1, sizeof (worker));	462	etp_worker *wrk = calloc (1, sizeof (etp_worker));
327		463
328	/*TODO*/	464	/*TODO*/
329	assert (("unable to allocate worker thread data", wrk));	465	assert (("unable to allocate worker thread data", wrk));
330		466
331	X_LOCK (wrklock);	467	X_LOCK (wrklock);
332		468
333	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	469	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
334	{	470	{
335	wrk->prev = &wrk_first;	471	wrk->prev = &wrk_first;
336	wrk->next = wrk_first.next;	472	wrk->next = wrk_first.next;
337	wrk_first.next->prev = wrk;	473	wrk_first.next->prev = wrk;
338	wrk_first.next = wrk;	474	wrk_first.next = wrk;
…		…
342	free (wrk);	478	free (wrk);
343		479
344	X_UNLOCK (wrklock);	480	X_UNLOCK (wrklock);
345	}	481	}
346		482
347	static void maybe_start_thread (void)	483	static void etp_maybe_start_thread (void)
348	{	484	{
349	if (eio_nthreads () >= wanted)	485	if (expect_true (etp_nthreads () >= wanted))
350	return;	486	return;
351		487
352	/* todo: maybe use idle here, but might be less exact */	488	/* todo: maybe use idle here, but might be less exact */
353	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	489	if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
354	return;	490	return;
355		491
356	start_thread ();	492	etp_start_thread ();
357	}	493	}
358		494
359	void eio_submit (eio_req *req)
360	{
361	++nreqs;
362
363	X_LOCK (reqlock);
364	++nready;
365	reqq_push (&req_queue, req);
366	X_COND_SIGNAL (reqwait);
367	X_UNLOCK (reqlock);
368
369	maybe_start_thread ();
370	}
371
372	static void end_thread (void)	495	static void etp_end_thread (void)
373	{	496	{
374	eio_req *req = calloc (1, sizeof (eio_req));	497	eio_req *req = calloc (1, sizeof (eio_req));
375		498
376	req->type = EIO_QUIT;	499	req->type = -1;
377	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	500	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
378		501
379	X_LOCK (reqlock);	502	X_LOCK (reqlock);
380	reqq_push (&req_queue, req);	503	reqq_push (&req_queue, req);
381	X_COND_SIGNAL (reqwait);	504	X_COND_SIGNAL (reqwait);
382	X_UNLOCK (reqlock);	505	X_UNLOCK (reqlock);
…		…
384	X_LOCK (wrklock);	507	X_LOCK (wrklock);
385	--started;	508	--started;
386	X_UNLOCK (wrklock);	509	X_UNLOCK (wrklock);
387	}	510	}
388		511
389	void eio_set_max_poll_time (double nseconds)
390	{
391	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
392	max_poll_time = nseconds;
393	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
394	}
395
396	void eio_set_max_poll_reqs (unsigned int maxreqs)
397	{
398	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
399	max_poll_reqs = maxreqs;
400	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
401	}
402
403	void eio_set_max_idle (unsigned int nthreads)
404	{
405	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
406	max_idle = nthreads <= 0 ? 1 : nthreads;
407	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
408	}
409
410	void eio_set_min_parallel (unsigned int nthreads)
411	{
412	if (wanted < nthreads)
413	wanted = nthreads;
414	}
415
416	void eio_set_max_parallel (unsigned int nthreads)
417	{
418	if (wanted > nthreads)
419	wanted = nthreads;
420
421	while (started > wanted)
422	end_thread ();
423	}
424
425	int eio_poll (void)	512	static int etp_poll (void)
426	{	513	{
427	int maxreqs = max_poll_reqs;	514	unsigned int maxreqs;
		515	unsigned int maxtime;
428	struct timeval tv_start, tv_now;	516	struct timeval tv_start, tv_now;
429	eio_req *req;
430		517
		518	X_LOCK (reslock);
		519	maxreqs = max_poll_reqs;
		520	maxtime = max_poll_time;
		521	X_UNLOCK (reslock);
		522
431	if (max_poll_time)	523	if (maxtime)
432	gettimeofday (&tv_start, 0);	524	gettimeofday (&tv_start, 0);
433		525
434	for (;;)	526	for (;;)
435	{	527	{
		528	ETP_REQ *req;
		529
436	maybe_start_thread ();	530	etp_maybe_start_thread ();
437		531
438	X_LOCK (reslock);	532	X_LOCK (reslock);
439	req = reqq_shift (&res_queue);	533	req = reqq_shift (&res_queue);
440		534
441	if (req)	535	if (req)
…		…
449	X_UNLOCK (reslock);	543	X_UNLOCK (reslock);
450		544
451	if (!req)	545	if (!req)
452	return 0;	546	return 0;
453		547
		548	X_LOCK (reqlock);
454	--nreqs;	549	--nreqs;
		550	X_UNLOCK (reqlock);
455		551
456	if (req->type == EIO_GROUP && req->size)	552	if (expect_false (req->type == EIO_GROUP && req->size))
457	{	553	{
458	req->int1 = 1; /* mark request as delayed */	554	req->int1 = 1; /* mark request as delayed */
459	continue;	555	continue;
460	}	556	}
461	else	557	else
462	{	558	{
463	int res = eio_finish (req);	559	int res = ETP_FINISH (req);
464	if (res)	560	if (expect_false (res))
465	return res;	561	return res;
466	}	562	}
467		563
468	if (maxreqs && !--maxreqs)	564	if (expect_false (maxreqs && !--maxreqs))
469	break;	565	break;
470		566
471	if (max_poll_time)	567	if (maxtime)
472	{	568	{
473	gettimeofday (&tv_now, 0);	569	gettimeofday (&tv_now, 0);
474		570
475	if (tvdiff (&tv_start, &tv_now) >= max_poll_time)	571	if (tvdiff (&tv_start, &tv_now) >= maxtime)
476	break;	572	break;
477	}	573	}
478	}	574	}
479		575
480	errno = EAGAIN;	576	errno = EAGAIN;
481	return -1;	577	return -1;
482	}	578	}
483		579
		580	static void etp_cancel (ETP_REQ *req)
		581	{
		582	X_LOCK (wrklock);
		583	req->flags |= EIO_FLAG_CANCELLED;
		584	X_UNLOCK (wrklock);
		585
		586	eio_grp_cancel (req);
		587	}
		588
		589	static void etp_submit (ETP_REQ *req)
		590	{
		591	req->pri -= ETP_PRI_MIN;
		592
		593	if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		594	if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		595
		596	if (expect_false (req->type == EIO_GROUP))
		597	{
		598	/* I hope this is worth it :/ */
		599	X_LOCK (reqlock);
		600	++nreqs;
		601	X_UNLOCK (reqlock);
		602
		603	X_LOCK (reslock);
		604
		605	++npending;
		606
		607	if (!reqq_push (&res_queue, req) && want_poll_cb)
		608	want_poll_cb ();
		609
		610	X_UNLOCK (reslock);
		611	}
		612	else
		613	{
		614	X_LOCK (reqlock);
		615	++nreqs;
		616	++nready;
		617	reqq_push (&req_queue, req);
		618	X_COND_SIGNAL (reqwait);
		619	X_UNLOCK (reqlock);
		620
		621	etp_maybe_start_thread ();
		622	}
		623	}
		624
		625	static void etp_set_max_poll_time (double nseconds)
		626	{
		627	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		628	max_poll_time = nseconds * EIO_TICKS;
		629	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		630	}
		631
		632	static void etp_set_max_poll_reqs (unsigned int maxreqs)
		633	{
		634	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		635	max_poll_reqs = maxreqs;
		636	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		637	}
		638
		639	static void etp_set_max_idle (unsigned int nthreads)
		640	{
		641	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		642	max_idle = nthreads;
		643	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		644	}
		645
		646	static void etp_set_idle_timeout (unsigned int seconds)
		647	{
		648	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		649	idle_timeout = seconds;
		650	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		651	}
		652
		653	static void etp_set_min_parallel (unsigned int nthreads)
		654	{
		655	if (wanted < nthreads)
		656	wanted = nthreads;
		657	}
		658
		659	static void etp_set_max_parallel (unsigned int nthreads)
		660	{
		661	if (wanted > nthreads)
		662	wanted = nthreads;
		663
		664	while (started > wanted)
		665	etp_end_thread ();
		666	}
		667
		668	/*****************************************************************************/
		669
		670	static void grp_try_feed (eio_req *grp)
		671	{
		672	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		673	{
		674	grp->flags &= ~EIO_FLAG_GROUPADD;
		675
		676	EIO_FEED (grp);
		677
		678	/* stop if no progress has been made */
		679	if (!(grp->flags & EIO_FLAG_GROUPADD))
		680	{
		681	grp->feed = 0;
		682	break;
		683	}
		684	}
		685	}
		686
		687	static int grp_dec (eio_req *grp)
		688	{
		689	--grp->size;
		690
		691	/* call feeder, if applicable */
		692	grp_try_feed (grp);
		693
		694	/* finish, if done */
		695	if (!grp->size && grp->int1)
		696	return eio_finish (grp);
		697	else
		698	return 0;
		699	}
		700
		701	void eio_destroy (eio_req *req)
		702	{
		703	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		704	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		705
		706	EIO_DESTROY (req);
		707	}
		708
		709	static int eio_finish (eio_req *req)
		710	{
		711	int res = EIO_FINISH (req);
		712
		713	if (req->grp)
		714	{
		715	int res2;
		716	eio_req *grp = req->grp;
		717
		718	/* unlink request */
		719	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		720	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		721
		722	if (grp->grp_first == req)
		723	grp->grp_first = req->grp_next;
		724
		725	res2 = grp_dec (grp);
		726
		727	if (!res && res2)
		728	res = res2;
		729	}
		730
		731	eio_destroy (req);
		732
		733	return res;
		734	}
		735
		736	void eio_grp_cancel (eio_req *grp)
		737	{
		738	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		739	eio_cancel (grp);
		740	}
		741
		742	void eio_cancel (eio_req *req)
		743	{
		744	etp_cancel (req);
		745	}
		746
		747	void eio_submit (eio_req *req)
		748	{
		749	etp_submit (req);
		750	}
		751
		752	unsigned int eio_nreqs (void)
		753	{
		754	return etp_nreqs ();
		755	}
		756
		757	unsigned int eio_nready (void)
		758	{
		759	return etp_nready ();
		760	}
		761
		762	unsigned int eio_npending (void)
		763	{
		764	return etp_npending ();
		765	}
		766
		767	unsigned int eio_nthreads (void)
		768	{
		769	return etp_nthreads ();
		770	}
		771
		772	void eio_set_max_poll_time (double nseconds)
		773	{
		774	etp_set_max_poll_time (nseconds);
		775	}
		776
		777	void eio_set_max_poll_reqs (unsigned int maxreqs)
		778	{
		779	etp_set_max_poll_reqs (maxreqs);
		780	}
		781
		782	void eio_set_max_idle (unsigned int nthreads)
		783	{
		784	etp_set_max_idle (nthreads);
		785	}
		786
		787	void eio_set_idle_timeout (unsigned int seconds)
		788	{
		789	etp_set_idle_timeout (seconds);
		790	}
		791
		792	void eio_set_min_parallel (unsigned int nthreads)
		793	{
		794	etp_set_min_parallel (nthreads);
		795	}
		796
		797	void eio_set_max_parallel (unsigned int nthreads)
		798	{
		799	etp_set_max_parallel (nthreads);
		800	}
		801
		802	int eio_poll (void)
		803	{
		804	return etp_poll ();
		805	}
		806
484	/*****************************************************************************/	807	/*****************************************************************************/
485	/* work around various missing functions */	808	/* work around various missing functions */
486		809
487	#if !HAVE_PREADWRITE	810	#if !HAVE_PREADWRITE
		811	# undef pread
		812	# undef pwrite
488	# define pread eio__pread	813	# define pread eio__pread
489	# define pwrite eio__pwrite	814	# define pwrite eio__pwrite
490
491	/*
492	* make our pread/pwrite safe against themselves, but not against
493	* normal read/write by using a mutex. slows down execution a lot,
494	* but that's your problem, not mine.
495	*/
496	static mutex_t preadwritelock = X_MUTEX_INIT;
497		815
498	static ssize_t	816	static ssize_t
499	eio__pread (int fd, void *buf, size_t count, off_t offset)	817	eio__pread (int fd, void *buf, size_t count, off_t offset)
500	{	818	{
501	ssize_t res;	819	ssize_t res;
…		…
519		837
520	X_LOCK (preadwritelock);	838	X_LOCK (preadwritelock);
521	ooffset = lseek (fd, 0, SEEK_CUR);	839	ooffset = lseek (fd, 0, SEEK_CUR);
522	lseek (fd, offset, SEEK_SET);	840	lseek (fd, offset, SEEK_SET);
523	res = write (fd, buf, count);	841	res = write (fd, buf, count);
524	lseek (fd, offset, SEEK_SET);	842	lseek (fd, ooffset, SEEK_SET);
525	X_UNLOCK (preadwritelock);	843	X_UNLOCK (preadwritelock);
526		844
527	return res;	845	return res;
528	}	846	}
529	#endif	847	#endif
530		848
531	#ifndef HAVE_FUTIMES	849	#ifndef HAVE_UTIMES
532		850
		851	# undef utimes
533	# define utimes(path,times) eio__utimes (path, times)	852	# define utimes(path,times) eio__utimes (path, times)
534	# define futimes(fd,times) eio__futimes (fd, times)
535		853
536	static int	854	static int
537	eio__utimes (const char *filename, const struct timeval times[2])	855	eio__utimes (const char *filename, const struct timeval times[2])
538	{	856	{
539	if (times)	857	if (times)
…		…
547	}	865	}
548	else	866	else
549	return utime (filename, 0);	867	return utime (filename, 0);
550	}	868	}
551		869
		870	#endif
		871
		872	#ifndef HAVE_FUTIMES
		873
		874	# undef futimes
		875	# define futimes(fd,times) eio__futimes (fd, times)
		876
552	static int eio__futimes (int fd, const struct timeval tv[2])	877	static int eio__futimes (int fd, const struct timeval tv[2])
553	{	878	{
554	errno = ENOSYS;	879	errno = ENOSYS;
555	return -1;	880	return -1;
556	}	881	}
557		882
558	#endif	883	#endif
559		884
560	#if !HAVE_FDATASYNC	885	#if !HAVE_FDATASYNC
		886	# undef fdatasync
561	# define fdatasync fsync	887	# define fdatasync(fd) fsync (fd)
562	#endif	888	#endif
		889
		890	/* sync_file_range always needs emulation */
		891	int
		892	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		893	{
		894	#if HAVE_SYNC_FILE_RANGE
		895	int res;
		896
		897	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		898	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		899	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		900	{
		901	flags = 0
		902	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		903	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		904	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		905	}
		906
		907	res = sync_file_range (fd, offset, nbytes, flags);
		908
		909	if (!res || errno != ENOSYS)
		910	return res;
		911	#endif
		912
		913	/* even though we could play tricks with the flags, it's better to always
		914	* call fdatasync, as that matches the expectation of its users best */
		915	return fdatasync (fd);
		916	}
563		917
564	#if !HAVE_READAHEAD	918	#if !HAVE_READAHEAD
		919	# undef readahead
565	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	920	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
566		921
567	static ssize_t	922	static ssize_t
568	eio__readahead (int fd, off_t offset, size_t count, worker *self)	923	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
569	{	924	{
570	size_t todo = count;	925	size_t todo = count;
571	dBUF;	926	dBUF;
572		927
573	while (todo > 0)	928	while (todo > 0)
…		…
583	return count;	938	return count;
584	}	939	}
585		940
586	#endif	941	#endif
587		942
588	#if !HAVE_READDIR_R
589	# define readdir_r eio__readdir_r
590
591	static mutex_t readdirlock = X_MUTEX_INIT;
592
593	static int
594	eio__readdir_r (DIR *dirp, EIO_STRUCT_DIRENT *ent, EIO_STRUCT_DIRENT **res)
595	{
596	EIO_STRUCT_DIRENT *e;
597	int errorno;
598
599	X_LOCK (readdirlock);
600
601	e = readdir (dirp);
602	errorno = errno;
603
604	if (e)
605	{
606	*res = ent;
607	strcpy (ent->d_name, e->d_name);
608	}
609	else
610	*res = 0;
611
612	X_UNLOCK (readdirlock);
613
614	errno = errorno;
615	return e ? 0 : -1;
616	}
617	#endif
618
619	/* sendfile always needs emulation */	943	/* sendfile always needs emulation */
620	static ssize_t	944	static ssize_t
621	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	945	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
622	{	946	{
623	ssize_t res;	947	ssize_t res;
624		948
625	if (!count)	949	if (!count)
626	return 0;	950	return 0;
627		951
628	#if HAVE_SENDFILE	952	#if HAVE_SENDFILE
629	# if __linux	953	# if __linux
630	res = sendfile (ofd, ifd, &offset, count);	954	res = sendfile (ofd, ifd, &offset, count);
631		955
632	# elif __freebsd	956	# elif __FreeBSD__
633	/*	957	/*
634	* Of course, the freebsd sendfile is a dire hack with no thoughts	958	* Of course, the freebsd sendfile is a dire hack with no thoughts
635	* wasted on making it similar to other I/O functions.	959	* wasted on making it similar to other I/O functions.
636	*/	960	*/
637	{	961	{
638	off_t sbytes;	962	off_t sbytes;
639	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	963	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
640		964
641	if (res < 0 && sbytes)	965	#if 0 /* according to the manpage, this is correct, but broken behaviour */
642	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	966	/* freebsd' sendfile will return 0 on success */
		967	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		968	/* not on e.g. EIO or EPIPE - sounds broken */
		969	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
		970	res = sbytes;
		971	#endif
		972
		973	/* according to source inspection, this is correct, and useful behaviour */
		974	if (sbytes)
		975	res = sbytes;
		976	}
		977
		978	# elif defined (__APPLE__)
		979
		980	{
		981	off_t sbytes = count;
		982	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		983
		984	/* according to the manpage, sbytes is always valid */
		985	if (sbytes)
643	res = sbytes;	986	res = sbytes;
644	}	987	}
645		988
646	# elif __hpux	989	# elif __hpux
647	res = sendfile (ofd, ifd, offset, count, 0, 0);	990	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
661	if (res < 0 && sbytes)	1004	if (res < 0 && sbytes)
662	res = sbytes;	1005	res = sbytes;
663	}	1006	}
664		1007
665	# endif	1008	# endif
		1009
		1010	#elif defined (_WIN32)
		1011
		1012	/* does not work, just for documentation of what would need to be done */
		1013	{
		1014	HANDLE h = TO_SOCKET (ifd);
		1015	SetFilePointer (h, offset, 0, FILE_BEGIN);
		1016	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		1017	}
		1018
666	#else	1019	#else
667	res = -1;	1020	res = -1;
668	errno = ENOSYS;	1021	errno = ENOSYS;
669	#endif	1022	#endif
670		1023
671	if (res < 0	1024	if (res < 0
672	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	1025	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		1026	/* BSDs */
		1027	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		1028	|| errno == ENOTSUP
		1029	#endif
		1030	|| errno == EOPNOTSUPP /* BSDs */
673	#if __solaris	1031	#if __solaris
674	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	1032	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
675	#endif	1033	#endif
676	)	1034	)
677	)	1035	)
…		…
708	}	1066	}
709		1067
710	return res;	1068	return res;
711	}	1069	}
712		1070
		1071	static signed char
		1072	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1073	{
		1074	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1075	: a->inode < b->inode ? -1
		1076	: a->inode > b->inode ? 1
		1077	: 0;
		1078	}
		1079
		1080	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1081
		1082	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1083	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1084
		1085	static void
		1086	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1087	{
		1088	unsigned char bits [9 + sizeof (ino_t) * 8];
		1089	unsigned char *bit = bits;
		1090
		1091	assert (CHAR_BIT == 8);
		1092	assert (sizeof (eio_dirent) * 8 < 256);
		1093	assert (offsetof (eio_dirent, inode)); /* we use bit #0 as sentinel */
		1094	assert (offsetof (eio_dirent, score)); /* we use bit #0 as sentinel */
		1095
		1096	if (size <= EIO_SORT_FAST)
		1097	return;
		1098
		1099	/* first prepare an array of bits to test in our radix sort */
		1100	/* try to take endianness into account, as well as differences in ino_t sizes */
		1101	/* inode_bits must contain all inodes ORed together */
		1102	/* which is used to skip bits that are 0 everywhere, which is very common */
		1103	{
		1104	ino_t endianness;
		1105	int i, j;
		1106
		1107	/* we store the byte offset of byte n into byte n of "endianness" */
		1108	for (i = 0; i < sizeof (ino_t); ++i)
		1109	((unsigned char *)&endianness)[i] = i;
		1110
		1111	*bit++ = 0;
		1112
		1113	for (i = 0; i < sizeof (ino_t); ++i)
		1114	{
		1115	/* shifting off the byte offsets out of "endianness" */
		1116	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1117	endianness >>= 8;
		1118
		1119	for (j = 0; j < 8; ++j)
		1120	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1121	*bit++ = offs + j;
		1122	}
		1123
		1124	for (j = 0; j < 8; ++j)
		1125	if (score_bits & (1 << j))
		1126	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1127	}
		1128
		1129	/* now actually do the sorting (a variant of MSD radix sort) */
		1130	{
		1131	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1132	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1133	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1134	int stk_idx = 0;
		1135
		1136	base_stk [stk_idx] = dents;
		1137	end_stk [stk_idx] = dents + size;
		1138	bit_stk [stk_idx] = bit - 1;
		1139
		1140	do
		1141	{
		1142	base = base_stk [stk_idx];
		1143	end = end_stk [stk_idx];
		1144	bit = bit_stk [stk_idx];
		1145
		1146	for (;;)
		1147	{
		1148	unsigned char O = *bit >> 3;
		1149	unsigned char M = 1 << (*bit & 7);
		1150
		1151	eio_dirent *a = base;
		1152	eio_dirent *b = end;
		1153
		1154	if (b - a < EIO_SORT_CUTOFF)
		1155	break;
		1156
		1157	/* now bit-partition the array on the bit */
		1158	/* this ugly asymmetric loop seems to perform much better than typical */
		1159	/* partition algos found in the literature */
		1160	do
		1161	if (!(((unsigned char *)a)[O] & M))
		1162	++a;
		1163	else if (!(((unsigned char *)--b)[O] & M))
		1164	{
		1165	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1166	++a;
		1167	}
		1168	while (b > a);
		1169
		1170	/* next bit, or stop, if no bits left in this path */
		1171	if (!*--bit)
		1172	break;
		1173
		1174	base_stk [stk_idx] = a;
		1175	end_stk [stk_idx] = end;
		1176	bit_stk [stk_idx] = bit;
		1177	++stk_idx;
		1178
		1179	end = a;
		1180	}
		1181	}
		1182	while (stk_idx--);
		1183	}
		1184	}
		1185
		1186	static void
		1187	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1188	{
		1189	/* first move the smallest element to the front, to act as a sentinel */
		1190	{
		1191	int i;
		1192	eio_dirent *min = dents;
		1193
		1194	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1195	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1196	if (EIO_DENT_CMP (dents [i], <, *min))
		1197	min = &dents [i];
		1198
		1199	/* swap elements 0 and j (minimum) */
		1200	{
		1201	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1202	}
		1203	}
		1204
		1205	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1206	{
		1207	eio_dirent *i, *j;
		1208
		1209	for (i = dents + 1; i < dents + size; ++i)
		1210	{
		1211	eio_dirent value = *i;
		1212
		1213	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1214	j [1] = j [0];
		1215
		1216	j [1] = value;
		1217	}
		1218	}
		1219	}
		1220
		1221	static void
		1222	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1223	{
		1224	if (size <= 1)
		1225	return; /* our insertion sort relies on size > 0 */
		1226
		1227	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1228	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1229	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1230
		1231	/* use an insertion sort at the end, or for small arrays, */
		1232	/* as insertion sort is more efficient for small partitions */
		1233	eio_dent_insertion_sort (dents, size);
		1234	}
		1235
713	/* read a full directory */	1236	/* read a full directory */
714	static void	1237	static void
715	eio__scandir (eio_req *req, worker *self)	1238	eio__scandir (eio_req *req, etp_worker *self)
716	{	1239	{
717	DIR *dirp;	1240	DIR *dirp;
718	union
719	{
720	EIO_STRUCT_DIRENT d;
721	char b [offsetof (EIO_STRUCT_DIRENT, d_name) + NAME_MAX + 1];
722	} *u;
723	EIO_STRUCT_DIRENT *entp;	1241	EIO_STRUCT_DIRENT *entp;
724	char *name, *names;	1242	char *name, *names;
725	int memlen = 4096;	1243	int namesalloc = 4096;
726	int memofs = 0;	1244	int namesoffs = 0;
		1245	int flags = req->int1;
		1246	eio_dirent *dents = 0;
		1247	int dentalloc = 128;
727	int res = 0;	1248	int dentoffs = 0;
		1249	ino_t inode_bits = 0;
		1250
		1251	req->result = -1;
		1252
		1253	if (!(flags & EIO_READDIR_DENTS))
		1254	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
728		1255
729	X_LOCK (wrklock);	1256	X_LOCK (wrklock);
		1257	/* the corresponding closedir is in ETP_WORKER_CLEAR */
730	self->dirp = dirp = opendir (req->ptr1);	1258	self->dirp = dirp = opendir (req->ptr1);
731	self->dbuf = u = malloc (sizeof (*u));	1259
732	req->flags |= EIO_FLAG_PTR2_FREE;	1260	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1261	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
733	req->ptr2 = names = malloc (memlen);	1262	req->ptr2 = names = malloc (namesalloc);
734	X_UNLOCK (wrklock);	1263	X_UNLOCK (wrklock);
735		1264
736	if (dirp && u && names)	1265	if (dirp && names && (!flags || dents))
737	for (;;)	1266	for (;;)
738	{	1267	{
739	errno = 0;	1268	errno = 0;
740	readdir_r (dirp, &u->d, &entp);	1269	entp = readdir (dirp);
741		1270
742	if (!entp)	1271	if (!entp)
		1272	{
		1273	if (errno)
		1274	break;
		1275
		1276	/* sort etc. */
		1277	req->int1 = flags;
		1278	req->result = dentoffs;
		1279
		1280	if (flags & EIO_READDIR_STAT_ORDER)
		1281	eio_dent_sort (dents, dentoffs, flags & EIO_READDIR_DIRS_FIRST ? 7 : 0, inode_bits);
		1282	else if (flags & EIO_READDIR_DIRS_FIRST)
		1283	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1284	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1285	else
		1286	{
		1287	/* in this case, all is known, and we just put dirs first and sort them */
		1288	eio_dirent *oth = dents + dentoffs;
		1289	eio_dirent *dir = dents;
		1290
		1291	/* now partition dirs to the front, and non-dirs to the back */
		1292	/* by walking from both sides and swapping if necessary */
		1293	while (oth > dir)
		1294	{
		1295	if (dir->type == EIO_DT_DIR)
		1296	++dir;
		1297	else if ((--oth)->type == EIO_DT_DIR)
		1298	{
		1299	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1300
		1301	++dir;
		1302	}
		1303	}
		1304
		1305	/* now sort the dirs only (dirs all have the same score) */
		1306	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1307	}
		1308
743	break;	1309	break;
		1310	}
744		1311
		1312	/* now add the entry to our list(s) */
745	name = entp->d_name;	1313	name = entp->d_name;
746		1314
		1315	/* skip . and .. entries */
747	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1316	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
748	{	1317	{
749	int len = strlen (name) + 1;	1318	int len = D_NAMLEN (entp) + 1;
750		1319
751	res++;	1320	while (expect_false (namesoffs + len > namesalloc))
752
753	while (memofs + len > memlen)
754	{	1321	{
755	memlen *= 2;	1322	namesalloc *= 2;
756	X_LOCK (wrklock);	1323	X_LOCK (wrklock);
757	req->ptr2 = names = realloc (names, memlen);	1324	req->ptr2 = names = realloc (names, namesalloc);
758	X_UNLOCK (wrklock);	1325	X_UNLOCK (wrklock);
759		1326
760	if (!names)	1327	if (!names)
761	break;	1328	break;
762	}	1329	}
763		1330
764	memcpy (names + memofs, name, len);	1331	memcpy (names + namesoffs, name, len);
		1332
		1333	if (dents)
		1334	{
		1335	struct eio_dirent *ent;
		1336
		1337	if (expect_false (dentoffs == dentalloc))
		1338	{
		1339	dentalloc *= 2;
		1340	X_LOCK (wrklock);
		1341	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1342	X_UNLOCK (wrklock);
		1343
		1344	if (!dents)
		1345	break;
		1346	}
		1347
		1348	ent = dents + dentoffs;
		1349
		1350	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1351	ent->namelen = len - 1;
		1352	ent->inode = D_INO (entp);
		1353
		1354	inode_bits |= ent->inode;
		1355
		1356	switch (D_TYPE (entp))
		1357	{
		1358	default:
		1359	ent->type = EIO_DT_UNKNOWN;
		1360	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1361	break;
		1362
		1363	#ifdef DT_FIFO
		1364	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1365	#endif
		1366	#ifdef DT_CHR
		1367	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1368	#endif
		1369	#ifdef DT_MPC
		1370	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1371	#endif
		1372	#ifdef DT_DIR
		1373	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1374	#endif
		1375	#ifdef DT_NAM
		1376	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1377	#endif
		1378	#ifdef DT_BLK
		1379	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1380	#endif
		1381	#ifdef DT_MPB
		1382	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1383	#endif
		1384	#ifdef DT_REG
		1385	case DT_REG: ent->type = EIO_DT_REG; break;
		1386	#endif
		1387	#ifdef DT_NWK
		1388	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1389	#endif
		1390	#ifdef DT_CMP
		1391	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1392	#endif
		1393	#ifdef DT_LNK
		1394	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1395	#endif
		1396	#ifdef DT_SOCK
		1397	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1398	#endif
		1399	#ifdef DT_DOOR
		1400	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1401	#endif
		1402	#ifdef DT_WHT
		1403	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1404	#endif
		1405	}
		1406
		1407	ent->score = 7;
		1408
		1409	if (flags & EIO_READDIR_DIRS_FIRST)
		1410	{
		1411	if (ent->type == EIO_DT_UNKNOWN)
		1412	{
		1413	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1414	ent->score = 1;
		1415	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1416	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1417	}
		1418	else if (ent->type == EIO_DT_DIR)
		1419	ent->score = 0;
		1420	}
		1421	}
		1422
765	memofs += len;	1423	namesoffs += len;
		1424	++dentoffs;
		1425	}
		1426
		1427	if (EIO_CANCELLED (req))
		1428	{
		1429	errno = ECANCELED;
		1430	break;
766	}	1431	}
767	}	1432	}
		1433	}
768		1434
769	if (errno)	1435	#ifdef PAGESIZE
770	res = -1;	1436	# define eio_pagesize() PAGESIZE
		1437	#else
		1438	static intptr_t
		1439	eio_pagesize (void)
		1440	{
		1441	static intptr_t page;
		1442
		1443	if (!page)
		1444	page = sysconf (_SC_PAGESIZE);
		1445
		1446	return page;
		1447	}
		1448	#endif
		1449
		1450	static void
		1451	eio_page_align (void **addr, size_t *length)
		1452	{
		1453	intptr_t mask = eio_pagesize () - 1;
		1454
		1455	/* round down addr */
		1456	intptr_t adj = mask & (intptr_t)*addr;
		1457
		1458	*addr = (void *)((intptr_t)*addr - adj);
		1459	*length += adj;
		1460
		1461	/* round up length */
		1462	*length = (*length + mask) & ~mask;
		1463	}
		1464
		1465	#if !_POSIX_MEMLOCK
		1466	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1467	#else
		1468
		1469	static int
		1470	eio__mlockall (int flags)
		1471	{
		1472	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1473	extern int mallopt (int, int);
		1474	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1475	#endif
		1476
		1477	if (EIO_MCL_CURRENT != MCL_CURRENT
		1478	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1479	{
		1480	flags = 0
		1481	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1482	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1483	}
		1484
		1485	return mlockall (flags);
		1486	}
		1487	#endif
		1488
		1489	#if !_POSIX_MEMLOCK_RANGE
		1490	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1491	#else
		1492
		1493	static int
		1494	eio__mlock (void *addr, size_t length)
		1495	{
		1496	eio_page_align (&addr, &length);
		1497
		1498	return mlock (addr, length);
		1499	}
		1500
		1501	#endif
		1502
		1503	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1504	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1505	#else
		1506
		1507	int
		1508	eio__msync (void *mem, size_t len, int flags)
		1509	{
		1510	eio_page_align (&mem, &len);
		1511
		1512	if (EIO_MS_ASYNC != MS_SYNC
		1513	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1514	|| EIO_MS_SYNC != MS_SYNC)
		1515	{
		1516	flags = 0
		1517	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1518	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1519	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1520	}
		1521
		1522	return msync (mem, len, flags);
		1523	}
		1524
		1525	#endif
		1526
		1527	int
		1528	eio__mtouch (eio_req *req)
		1529	{
		1530	void *mem = req->ptr2;
		1531	size_t len = req->size;
		1532	int flags = req->int1;
		1533
		1534	eio_page_align (&mem, &len);
		1535
771		1536	{
772	req->result = res;	1537	intptr_t addr = (intptr_t)mem;
		1538	intptr_t end = addr + len;
		1539	intptr_t page = eio_pagesize ();
		1540
		1541	if (addr < end)
		1542	if (flags & EIO_MT_MODIFY) /* modify */
		1543	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len && !EIO_CANCELLED (req));
		1544	else
		1545	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len && !EIO_CANCELLED (req));
		1546	}
		1547
		1548	return 0;
773	}	1549	}
774		1550
775	/*****************************************************************************/	1551	/*****************************************************************************/
776		1552
777	#define ALLOC(len) \	1553	#define ALLOC(len) \
…		…
787	req->result = -1; \	1563	req->result = -1; \
788	break; \	1564	break; \
789	} \	1565	} \
790	}	1566	}
791		1567
792	X_THREAD_PROC (eio_proc)	1568	X_THREAD_PROC (etp_proc)
793	{	1569	{
794	eio_req *req;	1570	ETP_REQ *req;
795	struct timespec ts;	1571	struct timespec ts;
796	worker *self = (worker *)thr_arg;	1572	etp_worker *self = (etp_worker *)thr_arg;
797		1573
798	/* try to distribute timeouts somewhat randomly */	1574	/* try to distribute timeouts somewhat randomly */
799	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1575	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
800		1576
801	for (;;)	1577	for (;;)
802	{	1578	{
803	ts.tv_sec = time (0) + IDLE_TIMEOUT;
804
805	X_LOCK (reqlock);	1579	X_LOCK (reqlock);
806		1580
807	for (;;)	1581	for (;;)
808	{	1582	{
809	self->req = req = reqq_shift (&req_queue);	1583	self->req = req = reqq_shift (&req_queue);
…		…
811	if (req)	1585	if (req)
812	break;	1586	break;
813		1587
814	++idle;	1588	++idle;
815		1589
		1590	ts.tv_sec = time (0) + idle_timeout;
816	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1591	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
817	{	1592	{
818	if (idle > max_idle)	1593	if (idle > max_idle)
819	{	1594	{
820	--idle;	1595	--idle;
…		…
825	goto quit;	1600	goto quit;
826	}	1601	}
827		1602
828	/* we are allowed to idle, so do so without any timeout */	1603	/* we are allowed to idle, so do so without any timeout */
829	X_COND_WAIT (reqwait, reqlock);	1604	X_COND_WAIT (reqwait, reqlock);
830	ts.tv_sec = time (0) + IDLE_TIMEOUT;
831	}	1605	}
832		1606
833	--idle;	1607	--idle;
834	}	1608	}
835		1609
836	--nready;	1610	--nready;
837		1611
838	X_UNLOCK (reqlock);	1612	X_UNLOCK (reqlock);
839		1613
840	errno = 0; /* strictly unnecessary */	1614	if (req->type < 0)
		1615	goto quit;
841		1616
842	if (!EIO_CANCELLED (req))	1617	if (!EIO_CANCELLED (req))
843	switch (req->type)	1618	ETP_EXECUTE (self, req);
844	{
845	case EIO_READ: ALLOC (req->size);
846	req->result = req->offs >= 0
847	? pread (req->int1, req->ptr2, req->size, req->offs)
848	: read (req->int1, req->ptr2, req->size); break;
849	case EIO_WRITE: req->result = req->offs >= 0
850	? pwrite (req->int1, req->ptr2, req->size, req->offs)
851	: write (req->int1, req->ptr2, req->size); break;
852
853	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
854	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
855
856	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
857	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
858	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
859	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
860	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
861	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
862
863	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
864	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
865	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
866	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
867	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
868	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
869
870	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
871	case EIO_CLOSE: req->result = close (req->int1); break;
872	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
873	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
874	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
875	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
876	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
877	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
878	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
879	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
880
881	case EIO_READLINK: ALLOC (NAME_MAX);
882	req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
883
884	case EIO_SYNC: req->result = 0; sync (); break;
885	case EIO_FSYNC: req->result = fsync (req->int1); break;
886	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
887
888	case EIO_READDIR: eio__scandir (req, self); break;
889
890	case EIO_BUSY:
891	#ifdef _WIN32
892	Sleep (req->nv1 * 1000.);
893	#else
894	{
895	struct timeval tv;
896
897	tv.tv_sec = req->nv1;
898	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.;
899
900	req->result = select (0, 0, 0, 0, &tv);
901	}
902	#endif
903	break;
904
905	case EIO_UTIME:
906	case EIO_FUTIME:
907	{
908	struct timeval tv[2];
909	struct timeval *times;
910
911	if (req->nv1 != -1. || req->nv2 != -1.)
912	{
913	tv[0].tv_sec = req->nv1;
914	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
915	tv[1].tv_sec = req->nv2;
916	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
917
918	times = tv;
919	}
920	else
921	times = 0;
922
923
924	req->result = req->type == EIO_FUTIME
925	? futimes (req->int1, times)
926	: utimes (req->ptr1, times);
927	}
928
929	case EIO_GROUP:
930	case EIO_NOP:
931	req->result = 0;
932	break;
933
934	case EIO_QUIT:
935	goto quit;
936
937	default:
938	req->result = -1;
939	break;
940	}
941
942	req->errorno = errno;
943		1619
944	X_LOCK (reslock);	1620	X_LOCK (reslock);
945		1621
946	++npending;	1622	++npending;
947		1623
948	if (!reqq_push (&res_queue, req) && want_poll_cb)	1624	if (!reqq_push (&res_queue, req) && want_poll_cb)
949	want_poll_cb ();	1625	want_poll_cb ();
950		1626
951	self->req = 0;	1627	self->req = 0;
952	worker_clear (self);	1628	etp_worker_clear (self);
953		1629
954	X_UNLOCK (reslock);	1630	X_UNLOCK (reslock);
955	}	1631	}
956		1632
957	quit:	1633	quit:
958	X_LOCK (wrklock);	1634	X_LOCK (wrklock);
959	worker_free (self);	1635	etp_worker_free (self);
960	X_UNLOCK (wrklock);	1636	X_UNLOCK (wrklock);
961		1637
962	return 0;	1638	return 0;
963	}	1639	}
964		1640
965	/*****************************************************************************/	1641	/*****************************************************************************/
966		1642
967	static void eio_atfork_prepare (void)
968	{
969	X_LOCK (wrklock);
970	X_LOCK (reqlock);
971	X_LOCK (reslock);
972	#if !HAVE_PREADWRITE
973	X_LOCK (preadwritelock);
974	#endif
975	#if !HAVE_READDIR_R
976	X_LOCK (readdirlock);
977	#endif
978	}
979
980	static void eio_atfork_parent (void)
981	{
982	#if !HAVE_READDIR_R
983	X_UNLOCK (readdirlock);
984	#endif
985	#if !HAVE_PREADWRITE
986	X_UNLOCK (preadwritelock);
987	#endif
988	X_UNLOCK (reslock);
989	X_UNLOCK (reqlock);
990	X_UNLOCK (wrklock);
991	}
992
993	static void eio_atfork_child (void)
994	{
995	eio_req *prv;
996
997	while (prv = reqq_shift (&req_queue))
998	eio_destroy (prv);
999
1000	while (prv = reqq_shift (&res_queue))
1001	eio_destroy (prv);
1002
1003	while (wrk_first.next != &wrk_first)
1004	{
1005	worker *wrk = wrk_first.next;
1006
1007	if (wrk->req)
1008	eio_destroy (wrk->req);
1009
1010	worker_clear (wrk);
1011	worker_free (wrk);
1012	}
1013
1014	started = 0;
1015	idle = 0;
1016	nreqs = 0;
1017	nready = 0;
1018	npending = 0;
1019
1020	eio_atfork_parent ();
1021	}
1022
1023	int eio_init (void (*want_poll)(void), void (*done_poll)(void))	1643	int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1024	{	1644	{
1025	want_poll_cb = want_poll;	1645	return etp_init (want_poll, done_poll);
1026	done_poll_cb = done_poll;
1027
1028	#ifdef _WIN32
1029	X_MUTEX_CHECK (wrklock);
1030	X_MUTEX_CHECK (reslock);
1031	X_MUTEX_CHECK (reqlock);
1032	X_MUTEX_CHECK (reqwait);
1033	X_MUTEX_CHECK (preadwritelock);
1034	X_MUTEX_CHECK (readdirlock);
1035
1036	X_COND_CHECK (reqwait);
1037	#endif
1038
1039	X_THREAD_ATFORK (eio_atfork_prepare, eio_atfork_parent, eio_atfork_child);
1040	}	1646	}
1041		1647
1042	static void eio_api_destroy (eio_req *req)	1648	static void eio_api_destroy (eio_req *req)
1043	{	1649	{
1044	free (req);	1650	free (req);
…		…
1049	\	1655	\
1050	req = (eio_req *)calloc (1, sizeof *req); \	1656	req = (eio_req *)calloc (1, sizeof *req); \
1051	if (!req) \	1657	if (!req) \
1052	return 0; \	1658	return 0; \
1053	\	1659	\
1054	req->type = rtype; \	1660	req->type = rtype; \
1055	req->pri = EIO_DEFAULT_PRI + EIO_PRI_BIAS; \	1661	req->pri = pri; \
1056	req->finish = cb; \	1662	req->finish = cb; \
		1663	req->data = data; \
1057	req->destroy = eio_api_destroy;	1664	req->destroy = eio_api_destroy;
1058		1665
1059	#define SEND eio_submit (req); return req	1666	#define SEND eio_submit (req); return req
1060		1667
1061	#define PATH \	1668	#define PATH \
…		…
1065	{ \	1672	{ \
1066	eio_api_destroy (req); \	1673	eio_api_destroy (req); \
1067	return 0; \	1674	return 0; \
1068	}	1675	}
1069		1676
		1677	static void eio_execute (etp_worker *self, eio_req *req)
		1678	{
		1679	switch (req->type)
		1680	{
		1681	case EIO_READ: ALLOC (req->size);
		1682	req->result = req->offs >= 0
		1683	? pread (req->int1, req->ptr2, req->size, req->offs)
		1684	: read (req->int1, req->ptr2, req->size); break;
		1685	case EIO_WRITE: req->result = req->offs >= 0
		1686	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1687	: write (req->int1, req->ptr2, req->size); break;
		1688
		1689	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1690	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1691
		1692	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1693	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1694	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1695	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1696	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1697	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1698
		1699	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1700	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1701	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1702	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1703
		1704	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1705	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1706	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1707	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1708	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1709	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1710
		1711	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1712	case EIO_CLOSE: req->result = close (req->int1); break;
		1713	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1714	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1715	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1716	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1717	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1718	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1719	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1720	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
		1721
		1722	case EIO_READLINK: ALLOC (PATH_MAX);
		1723	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1724
		1725	case EIO_SYNC: req->result = 0; sync (); break;
		1726	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1727	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1728	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1729	case EIO_MTOUCH: req->result = eio__mtouch (req); break;
		1730	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1731	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1732	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1733
		1734	case EIO_READDIR: eio__scandir (req, self); break;
		1735
		1736	case EIO_BUSY:
		1737	#ifdef _WIN32
		1738	Sleep (req->nv1 * 1e3);
		1739	#else
		1740	{
		1741	struct timeval tv;
		1742
		1743	tv.tv_sec = req->nv1;
		1744	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
		1745
		1746	req->result = select (0, 0, 0, 0, &tv);
		1747	}
		1748	#endif
		1749	break;
		1750
		1751	case EIO_UTIME:
		1752	case EIO_FUTIME:
		1753	{
		1754	struct timeval tv[2];
		1755	struct timeval *times;
		1756
		1757	if (req->nv1 != -1. || req->nv2 != -1.)
		1758	{
		1759	tv[0].tv_sec = req->nv1;
		1760	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1761	tv[1].tv_sec = req->nv2;
		1762	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1763
		1764	times = tv;
		1765	}
		1766	else
		1767	times = 0;
		1768
		1769	req->result = req->type == EIO_FUTIME
		1770	? futimes (req->int1, times)
		1771	: utimes (req->ptr1, times);
		1772	}
		1773	break;
		1774
		1775	case EIO_GROUP:
		1776	abort (); /* handled in eio_request */
		1777
		1778	case EIO_NOP:
		1779	req->result = 0;
		1780	break;
		1781
		1782	case EIO_CUSTOM:
		1783	req->feed (req);
		1784	break;
		1785
		1786	default:
		1787	errno = ENOSYS;
		1788	req->result = -1;
		1789	break;
		1790	}
		1791
		1792	req->errorno = errno;
		1793	}
		1794
		1795	#ifndef EIO_NO_WRAPPERS
		1796
1070	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1797	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1071	{	1798	{
1072	REQ (EIO_NOP); SEND;	1799	REQ (EIO_NOP); SEND;
1073	}	1800	}
1074		1801
…		…
1085	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1812	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1086	{	1813	{
1087	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1814	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1088	}	1815	}
1089		1816
		1817	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1818	{
		1819	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1820	}
		1821
		1822	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1823	{
		1824	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1825	}
		1826
		1827	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1828	{
		1829	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1830	}
		1831
		1832	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1833	{
		1834	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1835	}
		1836
		1837	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1838	{
		1839	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1840	}
		1841
1090	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1842	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1091	{	1843	{
1092	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1844	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1093	}	1845	}
1094		1846
…		…
1113	}	1865	}
1114		1866
1115	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1867	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1116	{	1868	{
1117	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1869	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1870	}
		1871
		1872	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1873	{
		1874	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1118	}	1875	}
1119		1876
1120	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1877	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1121	{	1878	{
1122	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1879	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
…		…
1196	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)	1953	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1197	{	1954	{
1198	return eio__1path (EIO_LSTAT, path, pri, cb, data);	1955	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1199	}	1956	}
1200		1957
		1958	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		1959	{
		1960	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		1961	}
		1962
1201	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	1963	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1202	{	1964	{
1203	return eio__1path (EIO_UNLINK, path, pri, cb, data);	1965	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1204	}	1966	}
1205		1967
1206	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	1968	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1207	{	1969	{
1208	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1970	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1209	}	1971	}
1210		1972
1211	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	1973	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1212	{	1974	{
1213	return eio__1path (EIO_READDIR, path, pri, cb, data);	1975	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1214	}	1976	}
1215		1977
1216	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1978	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1217	{	1979	{
1218	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	1980	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND;
1219	}	1981	}
1220		1982
1221	static eio_req *	1983	static eio_req *
1222	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1984	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1223	{	1985	{
…		…
1247	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	2009	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1248	{	2010	{
1249	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	2011	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1250	}	2012	}
1251		2013
		2014	eio_req *eio_custom (void (*)(eio_req *) execute, int pri, eio_cb cb, void *data);
		2015	{
		2016	REQ (EIO_CUSTOM); req->feed = execute; SEND;
		2017	}
		2018
		2019	#endif
		2020
1252	eio_req *eio_grp (eio_cb cb, void *data)	2021	eio_req *eio_grp (eio_cb cb, void *data)
1253	{	2022	{
1254	const int pri = EIO_PRI_MAX;	2023	const int pri = EIO_PRI_MAX;
1255		2024
1256	REQ (EIO_GROUP); SEND;	2025	REQ (EIO_GROUP); SEND;
…		…
1280		2049
1281	void eio_grp_add (eio_req *grp, eio_req *req)	2050	void eio_grp_add (eio_req *grp, eio_req *req)
1282	{	2051	{
1283	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2052	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1284		2053
		2054	grp->flags |= EIO_FLAG_GROUPADD;
		2055
1285	++grp->size;	2056	++grp->size;
1286	req->grp = grp;	2057	req->grp = grp;
1287		2058
1288	req->grp_prev = 0;	2059	req->grp_prev = 0;
1289	req->grp_next = grp->grp_first;	2060	req->grp_next = grp->grp_first;
…		…
1297	/*****************************************************************************/	2068	/*****************************************************************************/
1298	/* misc garbage */	2069	/* misc garbage */
1299		2070
1300	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2071	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1301	{	2072	{
1302	worker wrk;	2073	etp_worker wrk;
		2074	ssize_t ret;
1303		2075
1304	wrk.dbuf = 0;	2076	wrk.dbuf = 0;
1305		2077
1306	eio__sendfile (ofd, ifd, offset, count, &wrk);	2078	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1307		2079
1308	if (wrk.dbuf)	2080	if (wrk.dbuf)
1309	free (wrk.dbuf);	2081	free (wrk.dbuf);
1310	}
1311		2082
		2083	return ret;
		2084	}
		2085

Diff Legend

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