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 1.62 by root, Tue Feb 15 03:15:16 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
1	#include "eio.h"	40	#include "eio.h"
		41
		42	#ifdef EIO_STACKSIZE
		43	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		44	#endif
2	#include "xthread.h"	45	#include "xthread.h"
3		46
4	#include <errno.h>	47	#include <errno.h>
5	#include <stddef.h>	48	#include <stddef.h>
6	#include <stdlib.h>	49	#include <stdlib.h>
7	#include <string.h>	50	#include <string.h>
8	#include <errno.h>	51	#include <errno.h>
9	#include <sys/types.h>	52	#include <sys/types.h>
10	#include <sys/stat.h>	53	#include <sys/stat.h>
		54	#include <sys/statvfs.h>
11	#include <limits.h>	55	#include <limits.h>
12	#include <fcntl.h>	56	#include <fcntl.h>
13	#include <assert.h>	57	#include <assert.h>
14		58
15	#ifndef EIO_FINISH	59	#ifndef EIO_FINISH
…		…
25	#endif	69	#endif
26		70
27	#ifdef _WIN32	71	#ifdef _WIN32
28		72
29	/*doh*/	73	/*doh*/
30
31	#else	74	#else
32		75
33	# include "config.h"	76	# include "config.h"
34	# include <sys/time.h>	77	# include <sys/time.h>
35	# include <sys/select.h>	78	# include <sys/select.h>
36	# include <unistd.h>	79	# include <unistd.h>
37	# include <utime.h>	80	# include <utime.h>
38	# include <signal.h>	81	# include <signal.h>
39	# include <dirent.h>	82	# include <dirent.h>
40		83
		84	#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
		85	# include <sys/mman.h>
		86	#endif
		87
		88	/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
		89	# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
		90	# define _DIRENT_HAVE_D_TYPE /* sigh */
		91	# define D_INO(de) (de)->d_fileno
		92	# define D_NAMLEN(de) (de)->d_namlen
		93	# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
		94	# define D_INO(de) (de)->d_ino
		95	# endif
		96
		97	#ifdef _D_EXACT_NAMLEN
		98	# undef D_NAMLEN
		99	# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
		100	#endif
		101
		102	# ifdef _DIRENT_HAVE_D_TYPE
		103	# define D_TYPE(de) (de)->d_type
		104	# endif
		105
41	# ifndef EIO_STRUCT_DIRENT	106	# ifndef EIO_STRUCT_DIRENT
42	# define EIO_STRUCT_DIRENT struct dirent	107	# define EIO_STRUCT_DIRENT struct dirent
43	# endif	108	# endif
44		109
45	#endif	110	#endif
46		111
47	#if HAVE_SENDFILE	112	#if HAVE_SENDFILE
48	# if __linux	113	# if __linux
49	# include <sys/sendfile.h>	114	# include <sys/sendfile.h>
50	# elif __freebsd	115	# elif __FreeBSD__ || defined __APPLE__
51	# include <sys/socket.h>	116	# include <sys/socket.h>
52	# include <sys/uio.h>	117	# include <sys/uio.h>
53	# elif __hpux	118	# elif __hpux
54	# include <sys/socket.h>	119	# include <sys/socket.h>
55	# elif __solaris /* not yet */	120	# elif __solaris
56	# include <sys/sendfile.h>	121	# include <sys/sendfile.h>
57	# else	122	# else
58	# error sendfile support requested but not available	123	# error sendfile support requested but not available
59	# endif	124	# endif
60	#endif	125	#endif
61		126
62	/* number of seconds after which an idle threads exit */	127	#ifndef D_TYPE
63	#define IDLE_TIMEOUT 10	128	# define D_TYPE(de) 0
		129	#endif
		130	#ifndef D_INO
		131	# define D_INO(de) 0
		132	#endif
		133	#ifndef D_NAMLEN
		134	# define D_NAMLEN(de) strlen ((de)->d_name)
		135	#endif
64		136
65	/* used for struct dirent, AIX doesn't provide it */	137	/* used for struct dirent, AIX doesn't provide it */
66	#ifndef NAME_MAX	138	#ifndef NAME_MAX
67	# define NAME_MAX 4096	139	# define NAME_MAX 4096
68	#endif	140	#endif
69		141
		142	/* used for readlink etc. */
		143	#ifndef PATH_MAX
		144	# define PATH_MAX 4096
		145	#endif
		146
70	/* buffer size for various temporary buffers */	147	/* buffer size for various temporary buffers */
71	#define EIO_BUFSIZE 65536	148	#define EIO_BUFSIZE 65536
72		149
73	#define dBUF \	150	#define dBUF \
74	char *eio_buf; \	151	char *eio_buf; \
75	X_LOCK (wrklock); \	152	ETP_WORKER_LOCK (self); \
76	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \	153	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
77	X_UNLOCK (wrklock); \	154	ETP_WORKER_UNLOCK (self); \
78	errno = ENOMEM; \	155	errno = ENOMEM; \
79	if (!eio_buf) \	156	if (!eio_buf) \
80	return -1;	157	return -1;
81		158
82	#define EIO_TICKS ((1000000 + 1023) >> 10)	159	#define EIO_TICKS ((1000000 + 1023) >> 10)
83		160
		161	/*****************************************************************************/
		162
		163	#if __GNUC__ >= 3
		164	# define expect(expr,value) __builtin_expect ((expr),(value))
		165	#else
		166	# define expect(expr,value) (expr)
		167	#endif
		168
		169	#define expect_false(expr) expect ((expr) != 0, 0)
		170	#define expect_true(expr) expect ((expr) != 0, 1)
		171
		172	/*****************************************************************************/
		173
		174	#define ETP_PRI_MIN EIO_PRI_MIN
		175	#define ETP_PRI_MAX EIO_PRI_MAX
		176
		177	struct etp_worker;
		178
		179	#define ETP_REQ eio_req
		180	#define ETP_DESTROY(req) eio_destroy (req)
		181	static int eio_finish (eio_req *req);
		182	#define ETP_FINISH(req) eio_finish (req)
		183	static void eio_execute (struct etp_worker *self, eio_req *req);
		184	#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
		185
		186	#define ETP_WORKER_CLEAR(req) \
		187	if (wrk->dbuf) \
		188	{ \
		189	free (wrk->dbuf); \
		190	wrk->dbuf = 0; \
		191	} \
		192	\
		193	if (wrk->dirp) \
		194	{ \
		195	closedir (wrk->dirp); \
		196	wrk->dirp = 0; \
		197	}
		198
		199	#define ETP_WORKER_COMMON \
		200	void *dbuf; \
		201	DIR *dirp;
		202
		203	/*****************************************************************************/
		204
		205	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
		206
		207	/* calculate time difference in ~1/EIO_TICKS of a second */
		208	static int tvdiff (struct timeval *tv1, struct timeval *tv2)
		209	{
		210	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
		211	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
		212	}
		213
		214	static unsigned int started, idle, wanted = 4;
		215
84	static void (*want_poll_cb) (void);	216	static void (*want_poll_cb) (void);
85	static void (*done_poll_cb) (void);	217	static void (*done_poll_cb) (void);
86		218
87	static unsigned int max_poll_time = 0;	219	static unsigned int max_poll_time; /* reslock */
88	static unsigned int max_poll_reqs = 0;	220	static unsigned int max_poll_reqs; /* reslock */
89		221
90	/* calculcate time difference in ~1/EIO_TICKS of a second */	222	static volatile unsigned int nreqs; /* reqlock */
91	static int tvdiff (struct timeval *tv1, struct timeval *tv2)	223	static volatile unsigned int nready; /* reqlock */
92	{	224	static volatile unsigned int npending; /* reqlock */
93	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS	225	static volatile unsigned int max_idle = 4; /* maximum number of threads that can idle indefinitely */
94	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);	226	static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */
95	}
96		227
97	static unsigned int started, idle, wanted = 4;	228	static xmutex_t wrklock;
		229	static xmutex_t reslock;
		230	static xmutex_t reqlock;
		231	static xcond_t reqwait;
		232
		233	#if !HAVE_PREADWRITE
		234	/*
		235	* make our pread/pwrite emulation safe against themselves, but not against
		236	* normal read/write by using a mutex. slows down execution a lot,
		237	* but that's your problem, not mine.
		238	*/
		239	static xmutex_t preadwritelock = X_MUTEX_INIT;
		240	#endif
		241
		242	typedef struct etp_worker
		243	{
		244	/* locked by wrklock */
		245	struct etp_worker *prev, *next;
		246
		247	xthread_t tid;
		248
		249	/* locked by reslock, reqlock or wrklock */
		250	ETP_REQ *req; /* currently processed request */
		251
		252	ETP_WORKER_COMMON
		253	} etp_worker;
		254
		255	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		256
		257	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		258	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
98		259
99	/* worker threads management */	260	/* worker threads management */
100	static mutex_t wrklock = X_MUTEX_INIT;
101		261
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)	262	static void etp_worker_clear (etp_worker *wrk)
118	{	263	{
119	if (wrk->dirp)	264	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	}	265	}
131		266
132	static void worker_free (worker *wrk)	267	static void 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;
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)	275	static unsigned int etp_nreqs (void)
148	{	276	{
		277	int retval;
		278	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		279	retval = nreqs;
		280	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
149	return nreqs;	281	return retval;
150	}	282	}
151		283
152	unsigned int eio_nready (void)	284	static unsigned int etp_nready (void)
153	{	285	{
154	unsigned int retval;	286	unsigned int retval;
155		287
156	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	288	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
157	retval = nready;	289	retval = nready;
158	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	290	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
159		291
160	return retval;	292	return retval;
161	}	293	}
162		294
163	unsigned int eio_npending (void)	295	static unsigned int etp_npending (void)
164	{	296	{
165	unsigned int retval;	297	unsigned int retval;
166		298
167	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	299	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
168	retval = npending;	300	retval = npending;
169	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	301	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
170		302
171	return retval;	303	return retval;
172	}	304	}
173		305
174	unsigned int eio_nthreads (void)	306	static unsigned int etp_nthreads (void)
175	{	307	{
176	unsigned int retval;	308	unsigned int retval;
177		309
178	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	310	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
179	retval = started;	311	retval = started;
…		…
186	* a somewhat faster data structure might be nice, but	318	* a somewhat faster data structure might be nice, but
187	* with 8 priorities this actually needs <20 insns	319	* with 8 priorities this actually needs <20 insns
188	* per shift, the most expensive operation.	320	* per shift, the most expensive operation.
189	*/	321	*/
190	typedef struct {	322	typedef struct {
191	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	323	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
192	int size;	324	int size;
193	} reqq;	325	} etp_reqq;
194		326
195	static reqq req_queue;	327	static etp_reqq req_queue;
196	static reqq res_queue;	328	static etp_reqq res_queue;
197		329
198	static int reqq_push (reqq *q, eio_req *req)	330	static int reqq_push (etp_reqq *q, ETP_REQ *req)
199	{	331	{
200	int pri = req->pri;	332	int pri = req->pri;
201	req->next = 0;	333	req->next = 0;
202		334
203	if (q->qe[pri])	335	if (q->qe[pri])
…		…
209	q->qe[pri] = q->qs[pri] = req;	341	q->qe[pri] = q->qs[pri] = req;
210		342
211	return q->size++;	343	return q->size++;
212	}	344	}
213		345
214	static eio_req *reqq_shift (reqq *q)	346	static ETP_REQ *reqq_shift (etp_reqq *q)
215	{	347	{
216	int pri;	348	int pri;
217		349
218	if (!q->size)	350	if (!q->size)
219	return 0;	351	return 0;
220		352
221	--q->size;	353	--q->size;
222		354
223	for (pri = EIO_NUM_PRI; pri--; )	355	for (pri = ETP_NUM_PRI; pri--; )
224	{	356	{
225	eio_req *req = q->qs[pri];	357	eio_req *req = q->qs[pri];
226		358
227	if (req)	359	if (req)
228	{	360	{
…		…
234	}	366	}
235		367
236	abort ();	368	abort ();
237	}	369	}
238		370
239	static void grp_try_feed (eio_req *grp)	371	static void etp_thread_init (void)
240	{	372	{
241	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	373	X_MUTEX_CREATE (wrklock);
		374	X_MUTEX_CREATE (reslock);
		375	X_MUTEX_CREATE (reqlock);
		376	X_COND_CREATE (reqwait);
		377	}
		378
		379	static void etp_atfork_prepare (void)
		380	{
		381	X_LOCK (wrklock);
		382	X_LOCK (reqlock);
		383	X_LOCK (reslock);
		384	#if !HAVE_PREADWRITE
		385	X_LOCK (preadwritelock);
		386	#endif
		387	}
		388
		389	static void etp_atfork_parent (void)
		390	{
		391	#if !HAVE_PREADWRITE
		392	X_UNLOCK (preadwritelock);
		393	#endif
		394	X_UNLOCK (reslock);
		395	X_UNLOCK (reqlock);
		396	X_UNLOCK (wrklock);
		397	}
		398
		399	static void etp_atfork_child (void)
		400	{
		401	ETP_REQ *prv;
		402
		403	while ((prv = reqq_shift (&req_queue)))
		404	ETP_DESTROY (prv);
		405
		406	while ((prv = reqq_shift (&res_queue)))
		407	ETP_DESTROY (prv);
		408
		409	while (wrk_first.next != &wrk_first)
242	{	410	{
243	int old_len = grp->size;	411	etp_worker *wrk = wrk_first.next;
244		412
245	EIO_FEED (grp);	413	if (wrk->req)
		414	ETP_DESTROY (wrk->req);
246		415
247	/* stop if no progress has been made */	416	etp_worker_clear (wrk);
248	if (old_len == grp->size)	417	etp_worker_free (wrk);
249	{
250	grp->feed = 0;
251	break;
252	}
253	}	418	}
254	}
255		419
256	static int eio_finish (eio_req *req);	420	started = 0;
		421	idle = 0;
		422	nreqs = 0;
		423	nready = 0;
		424	npending = 0;
257		425
258	static int grp_dec (eio_req *grp)	426	etp_thread_init ();
259	{	427	}
260	--grp->size;
261		428
262	/* call feeder, if applicable */	429	static void
263	grp_try_feed (grp);	430	etp_once_init (void)
		431	{
		432	etp_thread_init ();
		433	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
		434	}
264		435
265	/* finish, if done */	436	static int
266	if (!grp->size && grp->int1)	437	etp_init (void (*want_poll)(void), void (*done_poll)(void))
267	return eio_finish (grp);	438	{
268	else	439	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
		440
		441	pthread_once (&doinit, etp_once_init);
		442
		443	want_poll_cb = want_poll;
		444	done_poll_cb = done_poll;
		445
269	return 0;	446	return 0;
270	}	447	}
271		448
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);	449	X_THREAD_PROC (etp_proc);
323		450
324	static void start_thread (void)	451	static void etp_start_thread (void)
325	{	452	{
326	worker *wrk = calloc (1, sizeof (worker));	453	etp_worker *wrk = calloc (1, sizeof (etp_worker));
327		454
328	/*TODO*/	455	/*TODO*/
329	assert (("unable to allocate worker thread data", wrk));	456	assert (("unable to allocate worker thread data", wrk));
330		457
331	X_LOCK (wrklock);	458	X_LOCK (wrklock);
332		459
333	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	460	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
334	{	461	{
335	wrk->prev = &wrk_first;	462	wrk->prev = &wrk_first;
336	wrk->next = wrk_first.next;	463	wrk->next = wrk_first.next;
337	wrk_first.next->prev = wrk;	464	wrk_first.next->prev = wrk;
338	wrk_first.next = wrk;	465	wrk_first.next = wrk;
…		…
342	free (wrk);	469	free (wrk);
343		470
344	X_UNLOCK (wrklock);	471	X_UNLOCK (wrklock);
345	}	472	}
346		473
347	static void maybe_start_thread (void)	474	static void etp_maybe_start_thread (void)
348	{	475	{
349	if (eio_nthreads () >= wanted)	476	if (expect_true (etp_nthreads () >= wanted))
350	return;	477	return;
351		478
352	/* todo: maybe use idle here, but might be less exact */	479	/* todo: maybe use idle here, but might be less exact */
353	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	480	if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
354	return;	481	return;
355		482
356	start_thread ();	483	etp_start_thread ();
357	}	484	}
358		485
359	void eio_submit (eio_req *req)
360	{
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)	486	static void etp_end_thread (void)
378	{	487	{
379	eio_req *req = calloc (1, sizeof (eio_req));	488	eio_req *req = calloc (1, sizeof (eio_req));
380		489
381	req->type = EIO_QUIT;	490	req->type = -1;
382	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	491	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
383		492
384	X_LOCK (reqlock);	493	X_LOCK (reqlock);
385	reqq_push (&req_queue, req);	494	reqq_push (&req_queue, req);
386	X_COND_SIGNAL (reqwait);	495	X_COND_SIGNAL (reqwait);
387	X_UNLOCK (reqlock);	496	X_UNLOCK (reqlock);
…		…
389	X_LOCK (wrklock);	498	X_LOCK (wrklock);
390	--started;	499	--started;
391	X_UNLOCK (wrklock);	500	X_UNLOCK (wrklock);
392	}	501	}
393		502
394	void eio_set_max_poll_time (double nseconds)
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)	503	static int etp_poll (void)
431	{	504	{
432	int maxreqs = max_poll_reqs;	505	unsigned int maxreqs;
		506	unsigned int maxtime;
433	struct timeval tv_start, tv_now;	507	struct timeval tv_start, tv_now;
434	eio_req *req;
435		508
		509	X_LOCK (reslock);
		510	maxreqs = max_poll_reqs;
		511	maxtime = max_poll_time;
		512	X_UNLOCK (reslock);
		513
436	if (max_poll_time)	514	if (maxtime)
437	gettimeofday (&tv_start, 0);	515	gettimeofday (&tv_start, 0);
438		516
439	for (;;)	517	for (;;)
440	{	518	{
		519	ETP_REQ *req;
		520
441	maybe_start_thread ();	521	etp_maybe_start_thread ();
442		522
443	X_LOCK (reslock);	523	X_LOCK (reslock);
444	req = reqq_shift (&res_queue);	524	req = reqq_shift (&res_queue);
445		525
446	if (req)	526	if (req)
…		…
454	X_UNLOCK (reslock);	534	X_UNLOCK (reslock);
455		535
456	if (!req)	536	if (!req)
457	return 0;	537	return 0;
458		538
		539	X_LOCK (reqlock);
459	--nreqs;	540	--nreqs;
		541	X_UNLOCK (reqlock);
460		542
461	if (req->type == EIO_GROUP && req->size)	543	if (expect_false (req->type == EIO_GROUP && req->size))
462	{	544	{
463	req->int1 = 1; /* mark request as delayed */	545	req->int1 = 1; /* mark request as delayed */
464	continue;	546	continue;
465	}	547	}
466	else	548	else
467	{	549	{
468	int res = eio_finish (req);	550	int res = ETP_FINISH (req);
469	if (res)	551	if (expect_false (res))
470	return res;	552	return res;
471	}	553	}
472		554
473	if (maxreqs && !--maxreqs)	555	if (expect_false (maxreqs && !--maxreqs))
474	break;	556	break;
475		557
476	if (max_poll_time)	558	if (maxtime)
477	{	559	{
478	gettimeofday (&tv_now, 0);	560	gettimeofday (&tv_now, 0);
479		561
480	if (tvdiff (&tv_start, &tv_now) >= max_poll_time)	562	if (tvdiff (&tv_start, &tv_now) >= maxtime)
481	break;	563	break;
482	}	564	}
483	}	565	}
484		566
485	errno = EAGAIN;	567	errno = EAGAIN;
486	return -1;	568	return -1;
487	}	569	}
488		570
		571	static void etp_cancel (ETP_REQ *req)
		572	{
		573	X_LOCK (wrklock);
		574	req->flags |= EIO_FLAG_CANCELLED;
		575	X_UNLOCK (wrklock);
		576
		577	eio_grp_cancel (req);
		578	}
		579
		580	static void etp_submit (ETP_REQ *req)
		581	{
		582	req->pri -= ETP_PRI_MIN;
		583
		584	if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		585	if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		586
		587	if (expect_false (req->type == EIO_GROUP))
		588	{
		589	/* I hope this is worth it :/ */
		590	X_LOCK (reqlock);
		591	++nreqs;
		592	X_UNLOCK (reqlock);
		593
		594	X_LOCK (reslock);
		595
		596	++npending;
		597
		598	if (!reqq_push (&res_queue, req) && want_poll_cb)
		599	want_poll_cb ();
		600
		601	X_UNLOCK (reslock);
		602	}
		603	else
		604	{
		605	X_LOCK (reqlock);
		606	++nreqs;
		607	++nready;
		608	reqq_push (&req_queue, req);
		609	X_COND_SIGNAL (reqwait);
		610	X_UNLOCK (reqlock);
		611
		612	etp_maybe_start_thread ();
		613	}
		614	}
		615
		616	static void etp_set_max_poll_time (double nseconds)
		617	{
		618	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		619	max_poll_time = nseconds * EIO_TICKS;
		620	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		621	}
		622
		623	static void etp_set_max_poll_reqs (unsigned int maxreqs)
		624	{
		625	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		626	max_poll_reqs = maxreqs;
		627	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		628	}
		629
		630	static void etp_set_max_idle (unsigned int nthreads)
		631	{
		632	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		633	max_idle = nthreads;
		634	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		635	}
		636
		637	static void etp_set_idle_timeout (unsigned int seconds)
		638	{
		639	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		640	idle_timeout = seconds;
		641	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		642	}
		643
		644	static void etp_set_min_parallel (unsigned int nthreads)
		645	{
		646	if (wanted < nthreads)
		647	wanted = nthreads;
		648	}
		649
		650	static void etp_set_max_parallel (unsigned int nthreads)
		651	{
		652	if (wanted > nthreads)
		653	wanted = nthreads;
		654
		655	while (started > wanted)
		656	etp_end_thread ();
		657	}
		658
		659	/*****************************************************************************/
		660
		661	static void grp_try_feed (eio_req *grp)
		662	{
		663	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		664	{
		665	grp->flags &= ~EIO_FLAG_GROUPADD;
		666
		667	EIO_FEED (grp);
		668
		669	/* stop if no progress has been made */
		670	if (!(grp->flags & EIO_FLAG_GROUPADD))
		671	{
		672	grp->feed = 0;
		673	break;
		674	}
		675	}
		676	}
		677
		678	static int grp_dec (eio_req *grp)
		679	{
		680	--grp->size;
		681
		682	/* call feeder, if applicable */
		683	grp_try_feed (grp);
		684
		685	/* finish, if done */
		686	if (!grp->size && grp->int1)
		687	return eio_finish (grp);
		688	else
		689	return 0;
		690	}
		691
		692	void eio_destroy (eio_req *req)
		693	{
		694	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		695	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		696
		697	EIO_DESTROY (req);
		698	}
		699
		700	static int eio_finish (eio_req *req)
		701	{
		702	int res = EIO_FINISH (req);
		703
		704	if (req->grp)
		705	{
		706	int res2;
		707	eio_req *grp = req->grp;
		708
		709	/* unlink request */
		710	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		711	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		712
		713	if (grp->grp_first == req)
		714	grp->grp_first = req->grp_next;
		715
		716	res2 = grp_dec (grp);
		717
		718	if (!res && res2)
		719	res = res2;
		720	}
		721
		722	eio_destroy (req);
		723
		724	return res;
		725	}
		726
		727	void eio_grp_cancel (eio_req *grp)
		728	{
		729	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		730	eio_cancel (grp);
		731	}
		732
		733	void eio_cancel (eio_req *req)
		734	{
		735	etp_cancel (req);
		736	}
		737
		738	void eio_submit (eio_req *req)
		739	{
		740	etp_submit (req);
		741	}
		742
		743	unsigned int eio_nreqs (void)
		744	{
		745	return etp_nreqs ();
		746	}
		747
		748	unsigned int eio_nready (void)
		749	{
		750	return etp_nready ();
		751	}
		752
		753	unsigned int eio_npending (void)
		754	{
		755	return etp_npending ();
		756	}
		757
		758	unsigned int eio_nthreads (void)
		759	{
		760	return etp_nthreads ();
		761	}
		762
		763	void eio_set_max_poll_time (double nseconds)
		764	{
		765	etp_set_max_poll_time (nseconds);
		766	}
		767
		768	void eio_set_max_poll_reqs (unsigned int maxreqs)
		769	{
		770	etp_set_max_poll_reqs (maxreqs);
		771	}
		772
		773	void eio_set_max_idle (unsigned int nthreads)
		774	{
		775	etp_set_max_idle (nthreads);
		776	}
		777
		778	void eio_set_idle_timeout (unsigned int seconds)
		779	{
		780	etp_set_idle_timeout (seconds);
		781	}
		782
		783	void eio_set_min_parallel (unsigned int nthreads)
		784	{
		785	etp_set_min_parallel (nthreads);
		786	}
		787
		788	void eio_set_max_parallel (unsigned int nthreads)
		789	{
		790	etp_set_max_parallel (nthreads);
		791	}
		792
		793	int eio_poll (void)
		794	{
		795	return etp_poll ();
		796	}
		797
489	/*****************************************************************************/	798	/*****************************************************************************/
490	/* work around various missing functions */	799	/* work around various missing functions */
491		800
492	#if !HAVE_PREADWRITE	801	#if !HAVE_PREADWRITE
		802	# undef pread
		803	# undef pwrite
493	# define pread eio__pread	804	# define pread eio__pread
494	# define pwrite eio__pwrite	805	# 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		806
503	static ssize_t	807	static ssize_t
504	eio__pread (int fd, void *buf, size_t count, off_t offset)	808	eio__pread (int fd, void *buf, size_t count, off_t offset)
505	{	809	{
506	ssize_t res;	810	ssize_t res;
…		…
524		828
525	X_LOCK (preadwritelock);	829	X_LOCK (preadwritelock);
526	ooffset = lseek (fd, 0, SEEK_CUR);	830	ooffset = lseek (fd, 0, SEEK_CUR);
527	lseek (fd, offset, SEEK_SET);	831	lseek (fd, offset, SEEK_SET);
528	res = write (fd, buf, count);	832	res = write (fd, buf, count);
529	lseek (fd, offset, SEEK_SET);	833	lseek (fd, ooffset, SEEK_SET);
530	X_UNLOCK (preadwritelock);	834	X_UNLOCK (preadwritelock);
531		835
532	return res;	836	return res;
533	}	837	}
534	#endif	838	#endif
535		839
536	#ifndef HAVE_FUTIMES	840	#ifndef HAVE_UTIMES
537		841
		842	# undef utimes
538	# define utimes(path,times) eio__utimes (path, times)	843	# define utimes(path,times) eio__utimes (path, times)
539	# define futimes(fd,times) eio__futimes (fd, times)
540		844
541	static int	845	static int
542	eio__utimes (const char *filename, const struct timeval times[2])	846	eio__utimes (const char *filename, const struct timeval times[2])
543	{	847	{
544	if (times)	848	if (times)
…		…
552	}	856	}
553	else	857	else
554	return utime (filename, 0);	858	return utime (filename, 0);
555	}	859	}
556		860
		861	#endif
		862
		863	#ifndef HAVE_FUTIMES
		864
		865	# undef futimes
		866	# define futimes(fd,times) eio__futimes (fd, times)
		867
557	static int eio__futimes (int fd, const struct timeval tv[2])	868	static int eio__futimes (int fd, const struct timeval tv[2])
558	{	869	{
559	errno = ENOSYS;	870	errno = ENOSYS;
560	return -1;	871	return -1;
561	}	872	}
562		873
563	#endif	874	#endif
564		875
565	#if !HAVE_FDATASYNC	876	#if !HAVE_FDATASYNC
		877	# undef fdatasync
566	# define fdatasync fsync	878	# define fdatasync(fd) fsync (fd)
567	#endif	879	#endif
		880
		881	/* sync_file_range always needs emulation */
		882	int
		883	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		884	{
		885	#if HAVE_SYNC_FILE_RANGE
		886	int res;
		887
		888	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		889	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		890	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		891	{
		892	flags = 0
		893	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		894	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		895	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		896	}
		897
		898	res = sync_file_range (fd, offset, nbytes, flags);
		899
		900	if (!res || errno != ENOSYS)
		901	return res;
		902	#endif
		903
		904	/* even though we could play tricks with the flags, it's better to always
		905	* call fdatasync, as that matches the expectation of its users best */
		906	return fdatasync (fd);
		907	}
568		908
569	#if !HAVE_READAHEAD	909	#if !HAVE_READAHEAD
		910	# undef readahead
570	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	911	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
571		912
572	static ssize_t	913	static ssize_t
573	eio__readahead (int fd, off_t offset, size_t count, worker *self)	914	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
574	{	915	{
575	size_t todo = count;	916	size_t todo = count;
576	dBUF;	917	dBUF;
577		918
578	while (todo > 0)	919	while (todo > 0)
…		…
588	return count;	929	return count;
589	}	930	}
590		931
591	#endif	932	#endif
592		933
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 */	934	/* sendfile always needs emulation */
625	static ssize_t	935	static ssize_t
626	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	936	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
627	{	937	{
628	ssize_t res;	938	ssize_t res;
629		939
630	if (!count)	940	if (!count)
631	return 0;	941	return 0;
632		942
633	#if HAVE_SENDFILE	943	#if HAVE_SENDFILE
634	# if __linux	944	# if __linux
635	res = sendfile (ofd, ifd, &offset, count);	945	res = sendfile (ofd, ifd, &offset, count);
636		946
637	# elif __freebsd	947	# elif __FreeBSD__
638	/*	948	/*
639	* Of course, the freebsd sendfile is a dire hack with no thoughts	949	* Of course, the freebsd sendfile is a dire hack with no thoughts
640	* wasted on making it similar to other I/O functions.	950	* wasted on making it similar to other I/O functions.
641	*/	951	*/
642	{	952	{
643	off_t sbytes;	953	off_t sbytes;
644	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	954	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
645		955
646	if (res < 0 && sbytes)	956	#if 0 /* according to the manpage, this is correct, but broken behaviour */
647	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	957	/* freebsd' sendfile will return 0 on success */
		958	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		959	/* not on e.g. EIO or EPIPE - sounds broken */
		960	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
		961	res = sbytes;
		962	#endif
		963
		964	/* according to source inspection, this is correct, and useful behaviour */
		965	if (sbytes)
		966	res = sbytes;
		967	}
		968
		969	# elif defined (__APPLE__)
		970
		971	{
		972	off_t sbytes = count;
		973	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		974
		975	/* according to the manpage, sbytes is always valid */
		976	if (sbytes)
648	res = sbytes;	977	res = sbytes;
649	}	978	}
650		979
651	# elif __hpux	980	# elif __hpux
652	res = sendfile (ofd, ifd, offset, count, 0, 0);	981	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
666	if (res < 0 && sbytes)	995	if (res < 0 && sbytes)
667	res = sbytes;	996	res = sbytes;
668	}	997	}
669		998
670	# endif	999	# endif
		1000
		1001	#elif defined (_WIN32)
		1002
		1003	/* does not work, just for documentation of what would need to be done */
		1004	{
		1005	HANDLE h = TO_SOCKET (ifd);
		1006	SetFilePointer (h, offset, 0, FILE_BEGIN);
		1007	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		1008	}
		1009
671	#else	1010	#else
672	res = -1;	1011	res = -1;
673	errno = ENOSYS;	1012	errno = ENOSYS;
674	#endif	1013	#endif
675		1014
676	if (res < 0	1015	if (res < 0
677	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	1016	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		1017	/* BSDs */
		1018	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		1019	|| errno == ENOTSUP
		1020	#endif
		1021	|| errno == EOPNOTSUPP /* BSDs */
678	#if __solaris	1022	#if __solaris
679	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	1023	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
680	#endif	1024	#endif
681	)	1025	)
682	)	1026	)
…		…
713	}	1057	}
714		1058
715	return res;	1059	return res;
716	}	1060	}
717		1061
		1062	static signed char
		1063	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1064	{
		1065	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1066	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1067	}
		1068
		1069	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1070
		1071	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1072	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1073
		1074	static void
		1075	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1076	{
		1077	unsigned char bits [9 + sizeof (ino_t) * 8];
		1078	unsigned char *bit = bits;
		1079
		1080	assert (CHAR_BIT == 8);
		1081	assert (sizeof (eio_dirent) * 8 < 256);
		1082	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1083	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1084
		1085	if (size <= EIO_SORT_FAST)
		1086	return;
		1087
		1088	/* first prepare an array of bits to test in our radix sort */
		1089	/* try to take endianness into account, as well as differences in ino_t sizes */
		1090	/* inode_bits must contain all inodes ORed together */
		1091	/* which is used to skip bits that are 0 everywhere, which is very common */
		1092	{
		1093	ino_t endianness;
		1094	int i, j;
		1095
		1096	/* we store the byte offset of byte n into byte n of "endianness" */
		1097	for (i = 0; i < sizeof (ino_t); ++i)
		1098	((unsigned char *)&endianness)[i] = i;
		1099
		1100	*bit++ = 0;
		1101
		1102	for (i = 0; i < sizeof (ino_t); ++i)
		1103	{
		1104	/* shifting off the byte offsets out of "endianness" */
		1105	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1106	endianness >>= 8;
		1107
		1108	for (j = 0; j < 8; ++j)
		1109	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1110	*bit++ = offs + j;
		1111	}
		1112
		1113	for (j = 0; j < 8; ++j)
		1114	if (score_bits & (1 << j))
		1115	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1116	}
		1117
		1118	/* now actually do the sorting (a variant of MSD radix sort) */
		1119	{
		1120	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1121	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1122	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1123	int stk_idx = 0;
		1124
		1125	base_stk [stk_idx] = dents;
		1126	end_stk [stk_idx] = dents + size;
		1127	bit_stk [stk_idx] = bit - 1;
		1128
		1129	do
		1130	{
		1131	base = base_stk [stk_idx];
		1132	end = end_stk [stk_idx];
		1133	bit = bit_stk [stk_idx];
		1134
		1135	for (;;)
		1136	{
		1137	unsigned char O = *bit >> 3;
		1138	unsigned char M = 1 << (*bit & 7);
		1139
		1140	eio_dirent *a = base;
		1141	eio_dirent *b = end;
		1142
		1143	if (b - a < EIO_SORT_CUTOFF)
		1144	break;
		1145
		1146	/* now bit-partition the array on the bit */
		1147	/* this ugly asymmetric loop seems to perform much better than typical */
		1148	/* partition algos found in the literature */
		1149	do
		1150	if (!(((unsigned char *)a)[O] & M))
		1151	++a;
		1152	else if (!(((unsigned char *)--b)[O] & M))
		1153	{
		1154	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1155	++a;
		1156	}
		1157	while (b > a);
		1158
		1159	/* next bit, or stop, if no bits left in this path */
		1160	if (!*--bit)
		1161	break;
		1162
		1163	base_stk [stk_idx] = a;
		1164	end_stk [stk_idx] = end;
		1165	bit_stk [stk_idx] = bit;
		1166	++stk_idx;
		1167
		1168	end = a;
		1169	}
		1170	}
		1171	while (stk_idx--);
		1172	}
		1173	}
		1174
		1175	static void
		1176	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1177	{
		1178	/* first move the smallest element to the front, to act as a sentinel */
		1179	{
		1180	int i;
		1181	eio_dirent *min = dents;
		1182
		1183	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1184	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1185	if (EIO_DENT_CMP (dents [i], <, *min))
		1186	min = &dents [i];
		1187
		1188	/* swap elements 0 and j (minimum) */
		1189	{
		1190	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1191	}
		1192	}
		1193
		1194	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1195	{
		1196	eio_dirent *i, *j;
		1197
		1198	for (i = dents + 1; i < dents + size; ++i)
		1199	{
		1200	eio_dirent value = *i;
		1201
		1202	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1203	j [1] = j [0];
		1204
		1205	j [1] = value;
		1206	}
		1207	}
		1208	}
		1209
		1210	static void
		1211	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1212	{
		1213	if (size <= 1)
		1214	return; /* our insertion sort relies on size > 0 */
		1215
		1216	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1217	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1218	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1219
		1220	/* use an insertion sort at the end, or for small arrays, */
		1221	/* as insertion sort is more efficient for small partitions */
		1222	eio_dent_insertion_sort (dents, size);
		1223	}
		1224
718	/* read a full directory */	1225	/* read a full directory */
719	static void	1226	static void
720	eio__scandir (eio_req *req, worker *self)	1227	eio__scandir (eio_req *req, etp_worker *self)
721	{	1228	{
722	DIR *dirp;	1229	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;	1230	EIO_STRUCT_DIRENT *entp;
729	char *name, *names;	1231	char *name, *names;
730	int memlen = 4096;	1232	int namesalloc = 4096;
731	int memofs = 0;	1233	int namesoffs = 0;
		1234	int flags = req->int1;
		1235	eio_dirent *dents = 0;
		1236	int dentalloc = 128;
732	int res = 0;	1237	int dentoffs = 0;
		1238	ino_t inode_bits = 0;
		1239
		1240	req->result = -1;
		1241
		1242	if (!(flags & EIO_READDIR_DENTS))
		1243	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
733		1244
734	X_LOCK (wrklock);	1245	X_LOCK (wrklock);
		1246	/* the corresponding closedir is in ETP_WORKER_CLEAR */
735	self->dirp = dirp = opendir (req->ptr1);	1247	self->dirp = dirp = opendir (req->ptr1);
736	self->dbuf = u = malloc (sizeof (*u));
737	req->flags |= EIO_FLAG_PTR2_FREE;	1248	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1249	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
738	req->ptr2 = names = malloc (memlen);	1250	req->ptr2 = names = malloc (namesalloc);
739	X_UNLOCK (wrklock);	1251	X_UNLOCK (wrklock);
740		1252
741	if (dirp && u && names)	1253	if (dirp && names && (!flags || dents))
742	for (;;)	1254	for (;;)
743	{	1255	{
744	errno = 0;	1256	errno = 0;
745	readdir_r (dirp, &u->d, &entp);	1257	entp = readdir (dirp);
746		1258
747	if (!entp)	1259	if (!entp)
		1260	{
		1261	if (errno)
		1262	break;
		1263
		1264	/* sort etc. */
		1265	req->int1 = flags;
		1266	req->result = dentoffs;
		1267
		1268	if (flags & EIO_READDIR_STAT_ORDER)
		1269	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1270	else if (flags & EIO_READDIR_DIRS_FIRST)
		1271	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1272	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1273	else
		1274	{
		1275	/* in this case, all is known, and we just put dirs first and sort them */
		1276	eio_dirent *oth = dents + dentoffs;
		1277	eio_dirent *dir = dents;
		1278
		1279	/* now partition dirs to the front, and non-dirs to the back */
		1280	/* by walking from both sides and swapping if necessary */
		1281	/* also clear score, so it doesn't influence sorting */
		1282	while (oth > dir)
		1283	{
		1284	if (dir->type == EIO_DT_DIR)
		1285	++dir;
		1286	else if ((--oth)->type == EIO_DT_DIR)
		1287	{
		1288	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1289
		1290	++dir;
		1291	}
		1292	}
		1293
		1294	/* now sort the dirs only */
		1295	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1296	}
		1297
748	break;	1298	break;
		1299	}
749		1300
		1301	/* now add the entry to our list(s) */
750	name = entp->d_name;	1302	name = entp->d_name;
751		1303
		1304	/* skip . and .. entries */
752	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1305	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
753	{	1306	{
754	int len = strlen (name) + 1;	1307	int len = D_NAMLEN (entp) + 1;
755		1308
756	res++;	1309	while (expect_false (namesoffs + len > namesalloc))
757
758	while (memofs + len > memlen)
759	{	1310	{
760	memlen *= 2;	1311	namesalloc *= 2;
761	X_LOCK (wrklock);	1312	X_LOCK (wrklock);
762	req->ptr2 = names = realloc (names, memlen);	1313	req->ptr2 = names = realloc (names, namesalloc);
763	X_UNLOCK (wrklock);	1314	X_UNLOCK (wrklock);
764		1315
765	if (!names)	1316	if (!names)
766	break;	1317	break;
767	}	1318	}
768		1319
769	memcpy (names + memofs, name, len);	1320	memcpy (names + namesoffs, name, len);
		1321
		1322	if (dents)
		1323	{
		1324	struct eio_dirent *ent;
		1325
		1326	if (expect_false (dentoffs == dentalloc))
		1327	{
		1328	dentalloc *= 2;
		1329	X_LOCK (wrklock);
		1330	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1331	X_UNLOCK (wrklock);
		1332
		1333	if (!dents)
		1334	break;
		1335	}
		1336
		1337	ent = dents + dentoffs;
		1338
		1339	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1340	ent->namelen = len - 1;
		1341	ent->inode = D_INO (entp);
		1342
		1343	inode_bits |= ent->inode;
		1344
		1345	switch (D_TYPE (entp))
		1346	{
		1347	default:
		1348	ent->type = EIO_DT_UNKNOWN;
		1349	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1350	break;
		1351
		1352	#ifdef DT_FIFO
		1353	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1354	#endif
		1355	#ifdef DT_CHR
		1356	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1357	#endif
		1358	#ifdef DT_MPC
		1359	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1360	#endif
		1361	#ifdef DT_DIR
		1362	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1363	#endif
		1364	#ifdef DT_NAM
		1365	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1366	#endif
		1367	#ifdef DT_BLK
		1368	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1369	#endif
		1370	#ifdef DT_MPB
		1371	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1372	#endif
		1373	#ifdef DT_REG
		1374	case DT_REG: ent->type = EIO_DT_REG; break;
		1375	#endif
		1376	#ifdef DT_NWK
		1377	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1378	#endif
		1379	#ifdef DT_CMP
		1380	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1381	#endif
		1382	#ifdef DT_LNK
		1383	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1384	#endif
		1385	#ifdef DT_SOCK
		1386	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1387	#endif
		1388	#ifdef DT_DOOR
		1389	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1390	#endif
		1391	#ifdef DT_WHT
		1392	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1393	#endif
		1394	}
		1395
		1396	ent->score = 7;
		1397
		1398	if (flags & EIO_READDIR_DIRS_FIRST)
		1399	{
		1400	if (ent->type == EIO_DT_UNKNOWN)
		1401	{
		1402	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1403	ent->score = 1;
		1404	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1405	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1406	}
		1407	else if (ent->type == EIO_DT_DIR)
		1408	ent->score = 0;
		1409	}
		1410	}
		1411
770	memofs += len;	1412	namesoffs += len;
		1413	++dentoffs;
		1414	}
		1415
		1416	if (EIO_CANCELLED (req))
		1417	{
		1418	errno = ECANCELED;
		1419	break;
771	}	1420	}
772	}	1421	}
		1422	}
773		1423
774	if (errno)	1424	#ifdef PAGESIZE
775	res = -1;	1425	# define eio_pagesize() PAGESIZE
		1426	#else
		1427	static intptr_t
		1428	eio_pagesize (void)
		1429	{
		1430	static intptr_t page;
		1431
		1432	if (!page)
		1433	page = sysconf (_SC_PAGESIZE);
		1434
		1435	return page;
		1436	}
		1437	#endif
		1438
		1439	static void
		1440	eio_page_align (void **addr, size_t *length)
		1441	{
		1442	intptr_t mask = eio_pagesize () - 1;
		1443
		1444	/* round down addr */
		1445	intptr_t adj = mask & (intptr_t)*addr;
		1446
		1447	*addr = (void *)((intptr_t)*addr - adj);
		1448	*length += adj;
		1449
		1450	/* round up length */
		1451	*length = (*length + mask) & ~mask;
		1452	}
		1453
		1454	#if !_POSIX_MEMLOCK
		1455	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1456	#else
		1457
		1458	static int
		1459	eio__mlockall (int flags)
		1460	{
		1461	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1462	extern int mallopt (int, int);
		1463	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1464	#endif
		1465
		1466	if (EIO_MCL_CURRENT != MCL_CURRENT
		1467	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1468	{
		1469	flags = 0
		1470	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1471	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1472	}
		1473
		1474	return mlockall (flags);
		1475	}
		1476	#endif
		1477
		1478	#if !_POSIX_MEMLOCK_RANGE
		1479	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1480	#else
		1481
		1482	static int
		1483	eio__mlock (void *addr, size_t length)
		1484	{
		1485	eio_page_align (&addr, &length);
		1486
		1487	return mlock (addr, length);
		1488	}
		1489
		1490	#endif
		1491
		1492	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1493	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1494	#else
		1495
		1496	int
		1497	eio__msync (void *mem, size_t len, int flags)
		1498	{
		1499	eio_page_align (&mem, &len);
		1500
		1501	if (EIO_MS_ASYNC != MS_SYNC
		1502	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1503	|| EIO_MS_SYNC != MS_SYNC)
		1504	{
		1505	flags = 0
		1506	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1507	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1508	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1509	}
		1510
		1511	return msync (mem, len, flags);
		1512	}
		1513
		1514	#endif
		1515
		1516	int
		1517	eio__mtouch (void *mem, size_t len, int flags)
		1518	{
		1519	eio_page_align (&mem, &len);
		1520
776		1521	{
777	req->result = res;	1522	intptr_t addr = (intptr_t)mem;
		1523	intptr_t end = addr + len;
		1524	intptr_t page = eio_pagesize ();
		1525
		1526	if (addr < end)
		1527	if (flags & EIO_MT_MODIFY) /* modify */
		1528	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
		1529	else
		1530	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
		1531	}
		1532
		1533	return 0;
778	}	1534	}
779		1535
780	/*****************************************************************************/	1536	/*****************************************************************************/
781		1537
782	#define ALLOC(len) \	1538	#define ALLOC(len) \
…		…
792	req->result = -1; \	1548	req->result = -1; \
793	break; \	1549	break; \
794	} \	1550	} \
795	}	1551	}
796		1552
797	X_THREAD_PROC (eio_proc)	1553	X_THREAD_PROC (etp_proc)
798	{	1554	{
799	eio_req *req;	1555	ETP_REQ *req;
800	struct timespec ts;	1556	struct timespec ts;
801	worker *self = (worker *)thr_arg;	1557	etp_worker *self = (etp_worker *)thr_arg;
802		1558
803	/* try to distribute timeouts somewhat randomly */	1559	/* try to distribute timeouts somewhat randomly */
804	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1560	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
805		1561
806	for (;;)	1562	for (;;)
807	{	1563	{
808	ts.tv_sec = time (0) + IDLE_TIMEOUT;
809
810	X_LOCK (reqlock);	1564	X_LOCK (reqlock);
811		1565
812	for (;;)	1566	for (;;)
813	{	1567	{
814	self->req = req = reqq_shift (&req_queue);	1568	self->req = req = reqq_shift (&req_queue);
…		…
816	if (req)	1570	if (req)
817	break;	1571	break;
818		1572
819	++idle;	1573	++idle;
820		1574
		1575	ts.tv_sec = time (0) + idle_timeout;
821	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1576	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
822	{	1577	{
823	if (idle > max_idle)	1578	if (idle > max_idle)
824	{	1579	{
825	--idle;	1580	--idle;
…		…
830	goto quit;	1585	goto quit;
831	}	1586	}
832		1587
833	/* we are allowed to idle, so do so without any timeout */	1588	/* we are allowed to idle, so do so without any timeout */
834	X_COND_WAIT (reqwait, reqlock);	1589	X_COND_WAIT (reqwait, reqlock);
835	ts.tv_sec = time (0) + IDLE_TIMEOUT;
836	}	1590	}
837		1591
838	--idle;	1592	--idle;
839	}	1593	}
840		1594
841	--nready;	1595	--nready;
842		1596
843	X_UNLOCK (reqlock);	1597	X_UNLOCK (reqlock);
844		1598
845	errno = 0; /* strictly unnecessary */	1599	if (req->type < 0)
		1600	goto quit;
846		1601
847	if (!EIO_CANCELLED (req))	1602	if (!EIO_CANCELLED (req))
848	switch (req->type)	1603	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		1604
949	X_LOCK (reslock);	1605	X_LOCK (reslock);
950		1606
951	++npending;	1607	++npending;
952		1608
953	if (!reqq_push (&res_queue, req) && want_poll_cb)	1609	if (!reqq_push (&res_queue, req) && want_poll_cb)
954	want_poll_cb ();	1610	want_poll_cb ();
955		1611
956	self->req = 0;	1612	self->req = 0;
957	worker_clear (self);	1613	etp_worker_clear (self);
958		1614
959	X_UNLOCK (reslock);	1615	X_UNLOCK (reslock);
960	}	1616	}
961		1617
962	quit:	1618	quit:
963	X_LOCK (wrklock);	1619	X_LOCK (wrklock);
964	worker_free (self);	1620	etp_worker_free (self);
965	X_UNLOCK (wrklock);	1621	X_UNLOCK (wrklock);
966		1622
967	return 0;	1623	return 0;
968	}	1624	}
969		1625
970	/*****************************************************************************/	1626	/*****************************************************************************/
971		1627
972	static void eio_atfork_prepare (void)
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))	1628	int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1029	{	1629	{
1030	want_poll_cb = want_poll;	1630	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	}	1631	}
1046		1632
1047	static void eio_api_destroy (eio_req *req)	1633	static void eio_api_destroy (eio_req *req)
1048	{	1634	{
1049	free (req);	1635	free (req);
…		…
1071	{ \	1657	{ \
1072	eio_api_destroy (req); \	1658	eio_api_destroy (req); \
1073	return 0; \	1659	return 0; \
1074	}	1660	}
1075		1661
		1662	static void eio_execute (etp_worker *self, eio_req *req)
		1663	{
		1664	switch (req->type)
		1665	{
		1666	case EIO_READ: ALLOC (req->size);
		1667	req->result = req->offs >= 0
		1668	? pread (req->int1, req->ptr2, req->size, req->offs)
		1669	: read (req->int1, req->ptr2, req->size); break;
		1670	case EIO_WRITE: req->result = req->offs >= 0
		1671	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1672	: write (req->int1, req->ptr2, req->size); break;
		1673
		1674	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1675	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1676
		1677	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1678	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1679	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1680	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1681	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1682	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1683
		1684	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1685	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1686	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1687	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1688
		1689	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1690	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1691	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1692	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1693	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1694	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1695
		1696	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1697	case EIO_CLOSE: req->result = close (req->int1); break;
		1698	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1699	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1700	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1701	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1702	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1703	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1704	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1705	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
		1706
		1707	case EIO_READLINK: ALLOC (PATH_MAX);
		1708	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1709
		1710	case EIO_SYNC: req->result = 0; sync (); break;
		1711	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1712	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1713	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1714	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1715	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1716	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1717	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1718
		1719	case EIO_READDIR: eio__scandir (req, self); break;
		1720
		1721	case EIO_BUSY:
		1722	#ifdef _WIN32
		1723	Sleep (req->nv1 * 1e3);
		1724	#else
		1725	{
		1726	struct timeval tv;
		1727
		1728	tv.tv_sec = req->nv1;
		1729	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
		1730
		1731	req->result = select (0, 0, 0, 0, &tv);
		1732	}
		1733	#endif
		1734	break;
		1735
		1736	case EIO_UTIME:
		1737	case EIO_FUTIME:
		1738	{
		1739	struct timeval tv[2];
		1740	struct timeval *times;
		1741
		1742	if (req->nv1 != -1. || req->nv2 != -1.)
		1743	{
		1744	tv[0].tv_sec = req->nv1;
		1745	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1746	tv[1].tv_sec = req->nv2;
		1747	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1748
		1749	times = tv;
		1750	}
		1751	else
		1752	times = 0;
		1753
		1754	req->result = req->type == EIO_FUTIME
		1755	? futimes (req->int1, times)
		1756	: utimes (req->ptr1, times);
		1757	}
		1758	break;
		1759
		1760	case EIO_GROUP:
		1761	abort (); /* handled in eio_request */
		1762
		1763	case EIO_NOP:
		1764	req->result = 0;
		1765	break;
		1766
		1767	case EIO_CUSTOM:
		1768	((void (*)(eio_req *))req->feed) (req);
		1769	break;
		1770
		1771	default:
		1772	errno = ENOSYS;
		1773	req->result = -1;
		1774	break;
		1775	}
		1776
		1777	req->errorno = errno;
		1778	}
		1779
		1780	#ifndef EIO_NO_WRAPPERS
		1781
1076	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1782	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1077	{	1783	{
1078	REQ (EIO_NOP); SEND;	1784	REQ (EIO_NOP); SEND;
1079	}	1785	}
1080		1786
…		…
1091	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1797	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1092	{	1798	{
1093	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1799	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1094	}	1800	}
1095		1801
		1802	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1803	{
		1804	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1805	}
		1806
		1807	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1808	{
		1809	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1810	}
		1811
		1812	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1813	{
		1814	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1815	}
		1816
		1817	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1818	{
		1819	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1820	}
		1821
		1822	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1823	{
		1824	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1825	}
		1826
1096	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1827	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1097	{	1828	{
1098	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1829	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1099	}	1830	}
1100		1831
…		…
1119	}	1850	}
1120		1851
1121	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1852	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1122	{	1853	{
1123	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1854	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1855	}
		1856
		1857	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1858	{
		1859	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1124	}	1860	}
1125		1861
1126	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1862	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1127	{	1863	{
1128	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1864	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)	1938	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1203	{	1939	{
1204	return eio__1path (EIO_LSTAT, path, pri, cb, data);	1940	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1205	}	1941	}
1206		1942
		1943	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		1944	{
		1945	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		1946	}
		1947
1207	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	1948	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1208	{	1949	{
1209	return eio__1path (EIO_UNLINK, path, pri, cb, data);	1950	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1210	}	1951	}
1211		1952
1212	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	1953	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1213	{	1954	{
1214	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1955	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1215	}	1956	}
1216		1957
1217	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	1958	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1218	{	1959	{
1219	return eio__1path (EIO_READDIR, path, pri, cb, data);	1960	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1220	}	1961	}
1221		1962
1222	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1963	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1223	{	1964	{
1224	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	1965	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND;
1225	}	1966	}
1226		1967
1227	static eio_req *	1968	static eio_req *
1228	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1969	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1229	{	1970	{
…		…
1253	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1994	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1254	{	1995	{
1255	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1996	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1256	}	1997	}
1257		1998
		1999	eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
		2000	{
		2001	REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
		2002	}
		2003
		2004	#endif
		2005
1258	eio_req *eio_grp (eio_cb cb, void *data)	2006	eio_req *eio_grp (eio_cb cb, void *data)
1259	{	2007	{
1260	const int pri = EIO_PRI_MAX;	2008	const int pri = EIO_PRI_MAX;
1261		2009
1262	REQ (EIO_GROUP); SEND;	2010	REQ (EIO_GROUP); SEND;
…		…
1286		2034
1287	void eio_grp_add (eio_req *grp, eio_req *req)	2035	void eio_grp_add (eio_req *grp, eio_req *req)
1288	{	2036	{
1289	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2037	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1290		2038
		2039	grp->flags |= EIO_FLAG_GROUPADD;
		2040
1291	++grp->size;	2041	++grp->size;
1292	req->grp = grp;	2042	req->grp = grp;
1293		2043
1294	req->grp_prev = 0;	2044	req->grp_prev = 0;
1295	req->grp_next = grp->grp_first;	2045	req->grp_next = grp->grp_first;
…		…
1303	/*****************************************************************************/	2053	/*****************************************************************************/
1304	/* misc garbage */	2054	/* misc garbage */
1305		2055
1306	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2056	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1307	{	2057	{
1308	worker wrk;	2058	etp_worker wrk;
		2059	ssize_t ret;
1309		2060
1310	wrk.dbuf = 0;	2061	wrk.dbuf = 0;
1311		2062
1312	eio__sendfile (ofd, ifd, offset, count, &wrk);	2063	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1313		2064
1314	if (wrk.dbuf)	2065	if (wrk.dbuf)
1315	free (wrk.dbuf);	2066	free (wrk.dbuf);
1316	}
1317		2067
		2068	return ret;
		2069	}
		2070

Diff Legend

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