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 1.60 by root, Fri Jan 7 21:45:42 2011 UTC

		1	/*
		2	* libeio implementation
		3	*
		4	* Copyright (c) 2007,2008,2009,2010 Marc Alexander Lehmann <libeio@schmorp.de>
		5	* All rights reserved.
		6	*
		7	* Redistribution and use in source and binary forms, with or without modifica-
		8	* tion, are permitted provided that the following conditions are met:
		9	*
		10	* 1. Redistributions of source code must retain the above copyright notice,
		11	* this list of conditions and the following disclaimer.
		12	*
		13	* 2. Redistributions in binary form must reproduce the above copyright
		14	* notice, this list of conditions and the following disclaimer in the
		15	* documentation and/or other materials provided with the distribution.
		16	*
		17	* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
		18	* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
		19	* CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
		20	* EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
		21	* CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
		22	* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
		23	* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
		24	* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
		25	* ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
		26	* OF THE POSSIBILITY OF SUCH DAMAGE.
		27	*
		28	* Alternatively, the contents of this file may be used under the terms of
		29	* the GNU General Public License ("GPL") version 2 or any later version,
		30	* in which case the provisions of the GPL are applicable instead of
		31	* the above. If you wish to allow the use of your version of this file
		32	* only under the terms of the GPL and not to allow others to use your
		33	* version of this file under the BSD license, indicate your decision
		34	* by deleting the provisions above and replace them with the notice
		35	* and other provisions required by the GPL. If you do not delete the
		36	* provisions above, a recipient may use your version of this file under
		37	* either the BSD or the GPL.
		38	*/
		39
1	#include "eio.h"	40	#include "eio.h"
		41
		42	#ifdef EIO_STACKSIZE
		43	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		44	#endif
2	#include "xthread.h"	45	#include "xthread.h"
3		46
4	#include <errno.h>	47	#include <errno.h>
5	#include <stddef.h>	48	#include <stddef.h>
6	#include <stdlib.h>	49	#include <stdlib.h>
7	#include <string.h>	50	#include <string.h>
8	#include <errno.h>	51	#include <errno.h>
9	#include <sys/types.h>	52	#include <sys/types.h>
10	#include <sys/stat.h>	53	#include <sys/stat.h>
		54	#include <sys/statvfs.h>
11	#include <limits.h>	55	#include <limits.h>
12	#include <fcntl.h>	56	#include <fcntl.h>
13	#include <assert.h>	57	#include <assert.h>
14		58
15	#ifndef EIO_FINISH	59	#ifndef EIO_FINISH
…		…
25	#endif	69	#endif
26		70
27	#ifdef _WIN32	71	#ifdef _WIN32
28		72
29	/*doh*/	73	/*doh*/
30
31	#else	74	#else
32		75
33	# include "config.h"	76	# include "config.h"
34	# include <sys/time.h>	77	# include <sys/time.h>
35	# include <sys/select.h>	78	# include <sys/select.h>
36	# include <unistd.h>	79	# include <unistd.h>
37	# include <utime.h>	80	# include <utime.h>
38	# include <signal.h>	81	# include <signal.h>
39	# include <dirent.h>	82	# include <dirent.h>
40		83
		84	#if _POSIX_MEMLOCK || _POSIX_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
		127	#ifndef D_TYPE
		128	# define D_TYPE(de) 0
		129	#endif
		130	#ifndef D_INO
		131	# define D_INO(de) 0
		132	#endif
		133	#ifndef D_NAMLEN
		134	# define D_NAMLEN(de) strlen ((de)->d_name)
		135	#endif
		136
62	/* number of seconds after which an idle threads exit */	137	/* number of seconds after which an idle threads exit */
63	#define IDLE_TIMEOUT 10	138	#define IDLE_TIMEOUT 10
64		139
65	/* used for struct dirent, AIX doesn't provide it */	140	/* used for struct dirent, AIX doesn't provide it */
66	#ifndef NAME_MAX	141	#ifndef NAME_MAX
67	# define NAME_MAX 4096	142	# define NAME_MAX 4096
68	#endif	143	#endif
69		144
		145	/* used for readlink etc. */
		146	#ifndef PATH_MAX
		147	# define PATH_MAX 4096
		148	#endif
		149
70	/* buffer size for various temporary buffers */	150	/* buffer size for various temporary buffers */
71	#define EIO_BUFSIZE 65536	151	#define EIO_BUFSIZE 65536
72		152
73	#define dBUF \	153	#define dBUF \
74	char *eio_buf; \	154	char *eio_buf; \
75	X_LOCK (wrklock); \	155	ETP_WORKER_LOCK (self); \
76	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \	156	self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
77	X_UNLOCK (wrklock); \	157	ETP_WORKER_UNLOCK (self); \
78	errno = ENOMEM; \	158	errno = ENOMEM; \
79	if (!eio_buf) \	159	if (!eio_buf) \
80	return -1;	160	return -1;
81		161
82	#define EIO_TICKS ((1000000 + 1023) >> 10)	162	#define EIO_TICKS ((1000000 + 1023) >> 10)
83		163
		164	/*****************************************************************************/
		165
		166	#if __GNUC__ >= 3
		167	# define expect(expr,value) __builtin_expect ((expr),(value))
		168	#else
		169	# define expect(expr,value) (expr)
		170	#endif
		171
		172	#define expect_false(expr) expect ((expr) != 0, 0)
		173	#define expect_true(expr) expect ((expr) != 0, 1)
		174
		175	/*****************************************************************************/
		176
		177	#define ETP_PRI_MIN EIO_PRI_MIN
		178	#define ETP_PRI_MAX EIO_PRI_MAX
		179
		180	struct etp_worker;
		181
		182	#define ETP_REQ eio_req
		183	#define ETP_DESTROY(req) eio_destroy (req)
		184	static int eio_finish (eio_req *req);
		185	#define ETP_FINISH(req) eio_finish (req)
		186	static void eio_execute (struct etp_worker *self, eio_req *req);
		187	#define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
		188
		189	#define ETP_WORKER_CLEAR(req) \
		190	if (wrk->dbuf) \
		191	{ \
		192	free (wrk->dbuf); \
		193	wrk->dbuf = 0; \
		194	} \
		195	\
		196	if (wrk->dirp) \
		197	{ \
		198	closedir (wrk->dirp); \
		199	wrk->dirp = 0; \
		200	}
		201
		202	#define ETP_WORKER_COMMON \
		203	void *dbuf; \
		204	DIR *dirp;
		205
		206	/*****************************************************************************/
		207
		208	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
		209
		210	/* calculate time difference in ~1/EIO_TICKS of a second */
		211	static int tvdiff (struct timeval *tv1, struct timeval *tv2)
		212	{
		213	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
		214	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
		215	}
		216
		217	static unsigned int started, idle, wanted = 4;
		218
84	static void (*want_poll_cb) (void);	219	static void (*want_poll_cb) (void);
85	static void (*done_poll_cb) (void);	220	static void (*done_poll_cb) (void);
86		221
87	static unsigned int max_poll_time = 0;	222	static unsigned int max_poll_time; /* reslock */
88	static unsigned int max_poll_reqs = 0;	223	static unsigned int max_poll_reqs; /* reslock */
89		224
90	/* calculcate time difference in ~1/EIO_TICKS of a second */	225	static volatile unsigned int nreqs; /* reqlock */
91	static int tvdiff (struct timeval *tv1, struct timeval *tv2)	226	static volatile unsigned int nready; /* reqlock */
92	{	227	static volatile unsigned int npending; /* reqlock */
93	return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS	228	static volatile unsigned int max_idle = 4;
94	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
95	}
96		229
97	static unsigned int started, idle, wanted = 4;	230	static xmutex_t wrklock = X_MUTEX_INIT;
		231	static xmutex_t reslock = X_MUTEX_INIT;
		232	static xmutex_t reqlock = X_MUTEX_INIT;
		233	static xcond_t reqwait = X_COND_INIT;
		234
		235	#if !HAVE_PREADWRITE
		236	/*
		237	* make our pread/pwrite emulation safe against themselves, but not against
		238	* normal read/write by using a mutex. slows down execution a lot,
		239	* but that's your problem, not mine.
		240	*/
		241	static xmutex_t preadwritelock = X_MUTEX_INIT;
		242	#endif
		243
		244	typedef struct etp_worker
		245	{
		246	/* locked by wrklock */
		247	struct etp_worker *prev, *next;
		248
		249	xthread_t tid;
		250
		251	/* locked by reslock, reqlock or wrklock */
		252	ETP_REQ *req; /* currently processed request */
		253
		254	ETP_WORKER_COMMON
		255	} etp_worker;
		256
		257	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		258
		259	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		260	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
98		261
99	/* worker threads management */	262	/* worker threads management */
100	static mutex_t wrklock = X_MUTEX_INIT;
101		263
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)	264	static void etp_worker_clear (etp_worker *wrk)
118	{	265	{
119	if (wrk->dirp)	266	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	}	267	}
131		268
132	static void worker_free (worker *wrk)	269	static void etp_worker_free (etp_worker *wrk)
133	{	270	{
134	wrk->next->prev = wrk->prev;	271	wrk->next->prev = wrk->prev;
135	wrk->prev->next = wrk->next;	272	wrk->prev->next = wrk->next;
136		273
137	free (wrk);	274	free (wrk);
138	}	275	}
139		276
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)	277	static unsigned int etp_nreqs (void)
148	{	278	{
		279	int retval;
		280	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		281	retval = nreqs;
		282	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
149	return nreqs;	283	return retval;
150	}	284	}
151		285
152	unsigned int eio_nready (void)	286	static unsigned int etp_nready (void)
153	{	287	{
154	unsigned int retval;	288	unsigned int retval;
155		289
156	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	290	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
157	retval = nready;	291	retval = nready;
158	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	292	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
159		293
160	return retval;	294	return retval;
161	}	295	}
162		296
163	unsigned int eio_npending (void)	297	static unsigned int etp_npending (void)
164	{	298	{
165	unsigned int retval;	299	unsigned int retval;
166		300
167	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	301	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
168	retval = npending;	302	retval = npending;
169	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	303	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
170		304
171	return retval;	305	return retval;
172	}	306	}
173		307
174	unsigned int eio_nthreads (void)	308	static unsigned int etp_nthreads (void)
175	{	309	{
176	unsigned int retval;	310	unsigned int retval;
177		311
178	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	312	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
179	retval = started;	313	retval = started;
…		…
186	* a somewhat faster data structure might be nice, but	320	* a somewhat faster data structure might be nice, but
187	* with 8 priorities this actually needs <20 insns	321	* with 8 priorities this actually needs <20 insns
188	* per shift, the most expensive operation.	322	* per shift, the most expensive operation.
189	*/	323	*/
190	typedef struct {	324	typedef struct {
191	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	325	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
192	int size;	326	int size;
193	} reqq;	327	} etp_reqq;
194		328
195	static reqq req_queue;	329	static etp_reqq req_queue;
196	static reqq res_queue;	330	static etp_reqq res_queue;
197		331
198	static int reqq_push (reqq *q, eio_req *req)	332	static int reqq_push (etp_reqq *q, ETP_REQ *req)
199	{	333	{
200	int pri = req->pri;	334	int pri = req->pri;
201	req->next = 0;	335	req->next = 0;
202		336
203	if (q->qe[pri])	337	if (q->qe[pri])
…		…
209	q->qe[pri] = q->qs[pri] = req;	343	q->qe[pri] = q->qs[pri] = req;
210		344
211	return q->size++;	345	return q->size++;
212	}	346	}
213		347
214	static eio_req *reqq_shift (reqq *q)	348	static ETP_REQ *reqq_shift (etp_reqq *q)
215	{	349	{
216	int pri;	350	int pri;
217		351
218	if (!q->size)	352	if (!q->size)
219	return 0;	353	return 0;
220		354
221	--q->size;	355	--q->size;
222		356
223	for (pri = EIO_NUM_PRI; pri--; )	357	for (pri = ETP_NUM_PRI; pri--; )
224	{	358	{
225	eio_req *req = q->qs[pri];	359	eio_req *req = q->qs[pri];
226		360
227	if (req)	361	if (req)
228	{	362	{
…		…
234	}	368	}
235		369
236	abort ();	370	abort ();
237	}	371	}
238		372
239	static void grp_try_feed (eio_req *grp)	373	static void etp_atfork_prepare (void)
240	{	374	{
241	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	375	X_LOCK (wrklock);
		376	X_LOCK (reqlock);
		377	X_LOCK (reslock);
		378	#if !HAVE_PREADWRITE
		379	X_LOCK (preadwritelock);
		380	#endif
		381	}
		382
		383	static void etp_atfork_parent (void)
		384	{
		385	#if !HAVE_PREADWRITE
		386	X_UNLOCK (preadwritelock);
		387	#endif
		388	X_UNLOCK (reslock);
		389	X_UNLOCK (reqlock);
		390	X_UNLOCK (wrklock);
		391	}
		392
		393	static void etp_atfork_child (void)
		394	{
		395	ETP_REQ *prv;
		396
		397	while ((prv = reqq_shift (&req_queue)))
		398	ETP_DESTROY (prv);
		399
		400	while ((prv = reqq_shift (&res_queue)))
		401	ETP_DESTROY (prv);
		402
		403	while (wrk_first.next != &wrk_first)
242	{	404	{
243	int old_len = grp->size;	405	etp_worker *wrk = wrk_first.next;
244		406
245	EIO_FEED (grp);	407	if (wrk->req)
		408	ETP_DESTROY (wrk->req);
246		409
247	/* stop if no progress has been made */	410	etp_worker_clear (wrk);
248	if (old_len == grp->size)	411	etp_worker_free (wrk);
249	{
250	grp->feed = 0;
251	break;
252	}
253	}	412	}
254	}
255		413
256	static int eio_finish (eio_req *req);	414	started = 0;
		415	idle = 0;
		416	nreqs = 0;
		417	nready = 0;
		418	npending = 0;
257		419
258	static int grp_dec (eio_req *grp)	420	etp_atfork_parent ();
259	{	421	}
260	--grp->size;
261		422
262	/* call feeder, if applicable */	423	static void
263	grp_try_feed (grp);	424	etp_once_init (void)
		425	{
		426	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
		427	}
264		428
265	/* finish, if done */	429	static int
266	if (!grp->size && grp->int1)	430	etp_init (void (*want_poll)(void), void (*done_poll)(void))
267	return eio_finish (grp);	431	{
268	else	432	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
		433
		434	pthread_once (&doinit, etp_once_init);
		435
		436	want_poll_cb = want_poll;
		437	done_poll_cb = done_poll;
		438
269	return 0;	439	return 0;
270	}	440	}
271		441
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);	442	X_THREAD_PROC (etp_proc);
323		443
324	static void start_thread (void)	444	static void etp_start_thread (void)
325	{	445	{
326	worker *wrk = calloc (1, sizeof (worker));	446	etp_worker *wrk = calloc (1, sizeof (etp_worker));
327		447
328	/*TODO*/	448	/*TODO*/
329	assert (("unable to allocate worker thread data", wrk));	449	assert (("unable to allocate worker thread data", wrk));
330		450
331	X_LOCK (wrklock);	451	X_LOCK (wrklock);
332		452
333	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	453	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
334	{	454	{
335	wrk->prev = &wrk_first;	455	wrk->prev = &wrk_first;
336	wrk->next = wrk_first.next;	456	wrk->next = wrk_first.next;
337	wrk_first.next->prev = wrk;	457	wrk_first.next->prev = wrk;
338	wrk_first.next = wrk;	458	wrk_first.next = wrk;
…		…
342	free (wrk);	462	free (wrk);
343		463
344	X_UNLOCK (wrklock);	464	X_UNLOCK (wrklock);
345	}	465	}
346		466
347	static void maybe_start_thread (void)	467	static void etp_maybe_start_thread (void)
348	{	468	{
349	if (eio_nthreads () >= wanted)	469	if (expect_true (etp_nthreads () >= wanted))
350	return;	470	return;
351		471
352	/* todo: maybe use idle here, but might be less exact */	472	/* todo: maybe use idle here, but might be less exact */
353	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	473	if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
354	return;	474	return;
355		475
356	start_thread ();	476	etp_start_thread ();
357	}	477	}
358		478
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)	479	static void etp_end_thread (void)
373	{	480	{
374	eio_req *req = calloc (1, sizeof (eio_req));	481	eio_req *req = calloc (1, sizeof (eio_req));
375		482
376	req->type = EIO_QUIT;	483	req->type = -1;
377	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	484	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
378		485
379	X_LOCK (reqlock);	486	X_LOCK (reqlock);
380	reqq_push (&req_queue, req);	487	reqq_push (&req_queue, req);
381	X_COND_SIGNAL (reqwait);	488	X_COND_SIGNAL (reqwait);
382	X_UNLOCK (reqlock);	489	X_UNLOCK (reqlock);
…		…
384	X_LOCK (wrklock);	491	X_LOCK (wrklock);
385	--started;	492	--started;
386	X_UNLOCK (wrklock);	493	X_UNLOCK (wrklock);
387	}	494	}
388		495
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)	496	static int etp_poll (void)
426	{	497	{
427	int maxreqs = max_poll_reqs;	498	unsigned int maxreqs;
		499	unsigned int maxtime;
428	struct timeval tv_start, tv_now;	500	struct timeval tv_start, tv_now;
429	eio_req *req;
430		501
		502	X_LOCK (reslock);
		503	maxreqs = max_poll_reqs;
		504	maxtime = max_poll_time;
		505	X_UNLOCK (reslock);
		506
431	if (max_poll_time)	507	if (maxtime)
432	gettimeofday (&tv_start, 0);	508	gettimeofday (&tv_start, 0);
433		509
434	for (;;)	510	for (;;)
435	{	511	{
		512	ETP_REQ *req;
		513
436	maybe_start_thread ();	514	etp_maybe_start_thread ();
437		515
438	X_LOCK (reslock);	516	X_LOCK (reslock);
439	req = reqq_shift (&res_queue);	517	req = reqq_shift (&res_queue);
440		518
441	if (req)	519	if (req)
…		…
449	X_UNLOCK (reslock);	527	X_UNLOCK (reslock);
450		528
451	if (!req)	529	if (!req)
452	return 0;	530	return 0;
453		531
		532	X_LOCK (reqlock);
454	--nreqs;	533	--nreqs;
		534	X_UNLOCK (reqlock);
455		535
456	if (req->type == EIO_GROUP && req->size)	536	if (expect_false (req->type == EIO_GROUP && req->size))
457	{	537	{
458	req->int1 = 1; /* mark request as delayed */	538	req->int1 = 1; /* mark request as delayed */
459	continue;	539	continue;
460	}	540	}
461	else	541	else
462	{	542	{
463	int res = eio_finish (req);	543	int res = ETP_FINISH (req);
464	if (res)	544	if (expect_false (res))
465	return res;	545	return res;
466	}	546	}
467		547
468	if (maxreqs && !--maxreqs)	548	if (expect_false (maxreqs && !--maxreqs))
469	break;	549	break;
470		550
471	if (max_poll_time)	551	if (maxtime)
472	{	552	{
473	gettimeofday (&tv_now, 0);	553	gettimeofday (&tv_now, 0);
474		554
475	if (tvdiff (&tv_start, &tv_now) >= max_poll_time)	555	if (tvdiff (&tv_start, &tv_now) >= maxtime)
476	break;	556	break;
477	}	557	}
478	}	558	}
479		559
480	errno = EAGAIN;	560	errno = EAGAIN;
481	return -1;	561	return -1;
482	}	562	}
483		563
		564	static void etp_cancel (ETP_REQ *req)
		565	{
		566	X_LOCK (wrklock);
		567	req->flags |= EIO_FLAG_CANCELLED;
		568	X_UNLOCK (wrklock);
		569
		570	eio_grp_cancel (req);
		571	}
		572
		573	static void etp_submit (ETP_REQ *req)
		574	{
		575	req->pri -= ETP_PRI_MIN;
		576
		577	if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		578	if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		579
		580	if (expect_false (req->type == EIO_GROUP))
		581	{
		582	/* I hope this is worth it :/ */
		583	X_LOCK (reqlock);
		584	++nreqs;
		585	X_UNLOCK (reqlock);
		586
		587	X_LOCK (reslock);
		588
		589	++npending;
		590
		591	if (!reqq_push (&res_queue, req) && want_poll_cb)
		592	want_poll_cb ();
		593
		594	X_UNLOCK (reslock);
		595	}
		596	else
		597	{
		598	X_LOCK (reqlock);
		599	++nreqs;
		600	++nready;
		601	reqq_push (&req_queue, req);
		602	X_COND_SIGNAL (reqwait);
		603	X_UNLOCK (reqlock);
		604
		605	etp_maybe_start_thread ();
		606	}
		607	}
		608
		609	static void etp_set_max_poll_time (double nseconds)
		610	{
		611	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		612	max_poll_time = nseconds * EIO_TICKS;
		613	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		614	}
		615
		616	static void etp_set_max_poll_reqs (unsigned int maxreqs)
		617	{
		618	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		619	max_poll_reqs = maxreqs;
		620	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		621	}
		622
		623	static void etp_set_max_idle (unsigned int nthreads)
		624	{
		625	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		626	max_idle = nthreads <= 0 ? 1 : nthreads;
		627	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		628	}
		629
		630	static void etp_set_min_parallel (unsigned int nthreads)
		631	{
		632	if (wanted < nthreads)
		633	wanted = nthreads;
		634	}
		635
		636	static void etp_set_max_parallel (unsigned int nthreads)
		637	{
		638	if (wanted > nthreads)
		639	wanted = nthreads;
		640
		641	while (started > wanted)
		642	etp_end_thread ();
		643	}
		644
		645	/*****************************************************************************/
		646
		647	static void grp_try_feed (eio_req *grp)
		648	{
		649	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		650	{
		651	grp->flags &= ~EIO_FLAG_GROUPADD;
		652
		653	EIO_FEED (grp);
		654
		655	/* stop if no progress has been made */
		656	if (!(grp->flags & EIO_FLAG_GROUPADD))
		657	{
		658	grp->feed = 0;
		659	break;
		660	}
		661	}
		662	}
		663
		664	static int grp_dec (eio_req *grp)
		665	{
		666	--grp->size;
		667
		668	/* call feeder, if applicable */
		669	grp_try_feed (grp);
		670
		671	/* finish, if done */
		672	if (!grp->size && grp->int1)
		673	return eio_finish (grp);
		674	else
		675	return 0;
		676	}
		677
		678	void eio_destroy (eio_req *req)
		679	{
		680	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		681	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		682
		683	EIO_DESTROY (req);
		684	}
		685
		686	static int eio_finish (eio_req *req)
		687	{
		688	int res = EIO_FINISH (req);
		689
		690	if (req->grp)
		691	{
		692	int res2;
		693	eio_req *grp = req->grp;
		694
		695	/* unlink request */
		696	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		697	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		698
		699	if (grp->grp_first == req)
		700	grp->grp_first = req->grp_next;
		701
		702	res2 = grp_dec (grp);
		703
		704	if (!res && res2)
		705	res = res2;
		706	}
		707
		708	eio_destroy (req);
		709
		710	return res;
		711	}
		712
		713	void eio_grp_cancel (eio_req *grp)
		714	{
		715	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		716	eio_cancel (grp);
		717	}
		718
		719	void eio_cancel (eio_req *req)
		720	{
		721	etp_cancel (req);
		722	}
		723
		724	void eio_submit (eio_req *req)
		725	{
		726	etp_submit (req);
		727	}
		728
		729	unsigned int eio_nreqs (void)
		730	{
		731	return etp_nreqs ();
		732	}
		733
		734	unsigned int eio_nready (void)
		735	{
		736	return etp_nready ();
		737	}
		738
		739	unsigned int eio_npending (void)
		740	{
		741	return etp_npending ();
		742	}
		743
		744	unsigned int eio_nthreads (void)
		745	{
		746	return etp_nthreads ();
		747	}
		748
		749	void eio_set_max_poll_time (double nseconds)
		750	{
		751	etp_set_max_poll_time (nseconds);
		752	}
		753
		754	void eio_set_max_poll_reqs (unsigned int maxreqs)
		755	{
		756	etp_set_max_poll_reqs (maxreqs);
		757	}
		758
		759	void eio_set_max_idle (unsigned int nthreads)
		760	{
		761	etp_set_max_idle (nthreads);
		762	}
		763
		764	void eio_set_min_parallel (unsigned int nthreads)
		765	{
		766	etp_set_min_parallel (nthreads);
		767	}
		768
		769	void eio_set_max_parallel (unsigned int nthreads)
		770	{
		771	etp_set_max_parallel (nthreads);
		772	}
		773
		774	int eio_poll (void)
		775	{
		776	return etp_poll ();
		777	}
		778
484	/*****************************************************************************/	779	/*****************************************************************************/
485	/* work around various missing functions */	780	/* work around various missing functions */
486		781
487	#if !HAVE_PREADWRITE	782	#if !HAVE_PREADWRITE
		783	# undef pread
		784	# undef pwrite
488	# define pread eio__pread	785	# define pread eio__pread
489	# define pwrite eio__pwrite	786	# 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		787
498	static ssize_t	788	static ssize_t
499	eio__pread (int fd, void *buf, size_t count, off_t offset)	789	eio__pread (int fd, void *buf, size_t count, off_t offset)
500	{	790	{
501	ssize_t res;	791	ssize_t res;
…		…
519		809
520	X_LOCK (preadwritelock);	810	X_LOCK (preadwritelock);
521	ooffset = lseek (fd, 0, SEEK_CUR);	811	ooffset = lseek (fd, 0, SEEK_CUR);
522	lseek (fd, offset, SEEK_SET);	812	lseek (fd, offset, SEEK_SET);
523	res = write (fd, buf, count);	813	res = write (fd, buf, count);
524	lseek (fd, offset, SEEK_SET);	814	lseek (fd, ooffset, SEEK_SET);
525	X_UNLOCK (preadwritelock);	815	X_UNLOCK (preadwritelock);
526		816
527	return res;	817	return res;
528	}	818	}
529	#endif	819	#endif
530		820
531	#ifndef HAVE_FUTIMES	821	#ifndef HAVE_UTIMES
532		822
		823	# undef utimes
533	# define utimes(path,times) eio__utimes (path, times)	824	# define utimes(path,times) eio__utimes (path, times)
534	# define futimes(fd,times) eio__futimes (fd, times)
535		825
536	static int	826	static int
537	eio__utimes (const char *filename, const struct timeval times[2])	827	eio__utimes (const char *filename, const struct timeval times[2])
538	{	828	{
539	if (times)	829	if (times)
…		…
547	}	837	}
548	else	838	else
549	return utime (filename, 0);	839	return utime (filename, 0);
550	}	840	}
551		841
		842	#endif
		843
		844	#ifndef HAVE_FUTIMES
		845
		846	# undef futimes
		847	# define futimes(fd,times) eio__futimes (fd, times)
		848
552	static int eio__futimes (int fd, const struct timeval tv[2])	849	static int eio__futimes (int fd, const struct timeval tv[2])
553	{	850	{
554	errno = ENOSYS;	851	errno = ENOSYS;
555	return -1;	852	return -1;
556	}	853	}
557		854
558	#endif	855	#endif
559		856
560	#if !HAVE_FDATASYNC	857	#if !HAVE_FDATASYNC
		858	# undef fdatasync
561	# define fdatasync fsync	859	# define fdatasync(fd) fsync (fd)
562	#endif	860	#endif
		861
		862	/* sync_file_range always needs emulation */
		863	int
		864	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		865	{
		866	#if HAVE_SYNC_FILE_RANGE
		867	int res;
		868
		869	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		870	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		871	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		872	{
		873	flags = 0
		874	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		875	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		876	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		877	}
		878
		879	res = sync_file_range (fd, offset, nbytes, flags);
		880
		881	if (!res || errno != ENOSYS)
		882	return res;
		883	#endif
		884
		885	/* even though we could play tricks with the flags, it's better to always
		886	* call fdatasync, as that matches the expectation of its users best */
		887	return fdatasync (fd);
		888	}
563		889
564	#if !HAVE_READAHEAD	890	#if !HAVE_READAHEAD
		891	# undef readahead
565	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	892	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
566		893
567	static ssize_t	894	static ssize_t
568	eio__readahead (int fd, off_t offset, size_t count, worker *self)	895	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
569	{	896	{
570	size_t todo = count;	897	size_t todo = count;
571	dBUF;	898	dBUF;
572		899
573	while (todo > 0)	900	while (todo > 0)
…		…
583	return count;	910	return count;
584	}	911	}
585		912
586	#endif	913	#endif
587		914
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 */	915	/* sendfile always needs emulation */
620	static ssize_t	916	static ssize_t
621	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	917	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
622	{	918	{
623	ssize_t res;	919	ssize_t res;
624		920
625	if (!count)	921	if (!count)
626	return 0;	922	return 0;
627		923
628	#if HAVE_SENDFILE	924	#if HAVE_SENDFILE
629	# if __linux	925	# if __linux
630	res = sendfile (ofd, ifd, &offset, count);	926	res = sendfile (ofd, ifd, &offset, count);
631		927
632	# elif __freebsd	928	# elif __FreeBSD__
633	/*	929	/*
634	* Of course, the freebsd sendfile is a dire hack with no thoughts	930	* Of course, the freebsd sendfile is a dire hack with no thoughts
635	* wasted on making it similar to other I/O functions.	931	* wasted on making it similar to other I/O functions.
636	*/	932	*/
637	{	933	{
638	off_t sbytes;	934	off_t sbytes;
639	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	935	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
640		936
641	if (res < 0 && sbytes)	937	#if 0 /* according to the manpage, this is correct, but broken behaviour */
642	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	938	/* freebsd' sendfile will return 0 on success */
		939	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		940	/* not on e.g. EIO or EPIPE - sounds broken */
		941	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
		942	res = sbytes;
		943	#endif
		944
		945	/* according to source inspection, this is correct, and useful behaviour */
		946	if (sbytes)
		947	res = sbytes;
		948	}
		949
		950	# elif defined (__APPLE__)
		951
		952	{
		953	off_t sbytes = count;
		954	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		955
		956	/* according to the manpage, sbytes is always valid */
		957	if (sbytes)
643	res = sbytes;	958	res = sbytes;
644	}	959	}
645		960
646	# elif __hpux	961	# elif __hpux
647	res = sendfile (ofd, ifd, offset, count, 0, 0);	962	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
661	if (res < 0 && sbytes)	976	if (res < 0 && sbytes)
662	res = sbytes;	977	res = sbytes;
663	}	978	}
664		979
665	# endif	980	# endif
		981
		982	#elif defined (_WIN32)
		983
		984	/* does not work, just for documentation of what would need to be done */
		985	{
		986	HANDLE h = TO_SOCKET (ifd);
		987	SetFilePointer (h, offset, 0, FILE_BEGIN);
		988	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		989	}
		990
666	#else	991	#else
667	res = -1;	992	res = -1;
668	errno = ENOSYS;	993	errno = ENOSYS;
669	#endif	994	#endif
670		995
671	if (res < 0	996	if (res < 0
672	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	997	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		998	/* BSDs */
		999	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		1000	|| errno == ENOTSUP
		1001	#endif
		1002	|| errno == EOPNOTSUPP /* BSDs */
673	#if __solaris	1003	#if __solaris
674	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	1004	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
675	#endif	1005	#endif
676	)	1006	)
677	)	1007	)
…		…
708	}	1038	}
709		1039
710	return res;	1040	return res;
711	}	1041	}
712		1042
		1043	static signed char
		1044	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1045	{
		1046	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1047	: a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
		1048	}
		1049
		1050	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1051
		1052	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1053	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1054
		1055	static void
		1056	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1057	{
		1058	unsigned char bits [9 + sizeof (ino_t) * 8];
		1059	unsigned char *bit = bits;
		1060
		1061	assert (CHAR_BIT == 8);
		1062	assert (sizeof (eio_dirent) * 8 < 256);
		1063	assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
		1064	assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
		1065
		1066	if (size <= EIO_SORT_FAST)
		1067	return;
		1068
		1069	/* first prepare an array of bits to test in our radix sort */
		1070	/* try to take endianness into account, as well as differences in ino_t sizes */
		1071	/* inode_bits must contain all inodes ORed together */
		1072	/* which is used to skip bits that are 0 everywhere, which is very common */
		1073	{
		1074	ino_t endianness;
		1075	int i, j;
		1076
		1077	/* we store the byte offset of byte n into byte n of "endianness" */
		1078	for (i = 0; i < sizeof (ino_t); ++i)
		1079	((unsigned char *)&endianness)[i] = i;
		1080
		1081	*bit++ = 0;
		1082
		1083	for (i = 0; i < sizeof (ino_t); ++i)
		1084	{
		1085	/* shifting off the byte offsets out of "endianness" */
		1086	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1087	endianness >>= 8;
		1088
		1089	for (j = 0; j < 8; ++j)
		1090	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1091	*bit++ = offs + j;
		1092	}
		1093
		1094	for (j = 0; j < 8; ++j)
		1095	if (score_bits & (1 << j))
		1096	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1097	}
		1098
		1099	/* now actually do the sorting (a variant of MSD radix sort) */
		1100	{
		1101	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1102	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1103	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1104	int stk_idx = 0;
		1105
		1106	base_stk [stk_idx] = dents;
		1107	end_stk [stk_idx] = dents + size;
		1108	bit_stk [stk_idx] = bit - 1;
		1109
		1110	do
		1111	{
		1112	base = base_stk [stk_idx];
		1113	end = end_stk [stk_idx];
		1114	bit = bit_stk [stk_idx];
		1115
		1116	for (;;)
		1117	{
		1118	unsigned char O = *bit >> 3;
		1119	unsigned char M = 1 << (*bit & 7);
		1120
		1121	eio_dirent *a = base;
		1122	eio_dirent *b = end;
		1123
		1124	if (b - a < EIO_SORT_CUTOFF)
		1125	break;
		1126
		1127	/* now bit-partition the array on the bit */
		1128	/* this ugly asymmetric loop seems to perform much better than typical */
		1129	/* partition algos found in the literature */
		1130	do
		1131	if (!(((unsigned char *)a)[O] & M))
		1132	++a;
		1133	else if (!(((unsigned char *)--b)[O] & M))
		1134	{
		1135	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1136	++a;
		1137	}
		1138	while (b > a);
		1139
		1140	/* next bit, or stop, if no bits left in this path */
		1141	if (!*--bit)
		1142	break;
		1143
		1144	base_stk [stk_idx] = a;
		1145	end_stk [stk_idx] = end;
		1146	bit_stk [stk_idx] = bit;
		1147	++stk_idx;
		1148
		1149	end = a;
		1150	}
		1151	}
		1152	while (stk_idx--);
		1153	}
		1154	}
		1155
		1156	static void
		1157	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1158	{
		1159	/* first move the smallest element to the front, to act as a sentinel */
		1160	{
		1161	int i;
		1162	eio_dirent *min = dents;
		1163
		1164	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1165	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1166	if (EIO_DENT_CMP (dents [i], <, *min))
		1167	min = &dents [i];
		1168
		1169	/* swap elements 0 and j (minimum) */
		1170	{
		1171	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1172	}
		1173	}
		1174
		1175	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1176	{
		1177	eio_dirent *i, *j;
		1178
		1179	for (i = dents + 1; i < dents + size; ++i)
		1180	{
		1181	eio_dirent value = *i;
		1182
		1183	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1184	j [1] = j [0];
		1185
		1186	j [1] = value;
		1187	}
		1188	}
		1189	}
		1190
		1191	static void
		1192	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1193	{
		1194	if (size <= 1)
		1195	return; /* our insertion sort relies on size > 0 */
		1196
		1197	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1198	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1199	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1200
		1201	/* use an insertion sort at the end, or for small arrays, */
		1202	/* as insertion sort is more efficient for small partitions */
		1203	eio_dent_insertion_sort (dents, size);
		1204	}
		1205
713	/* read a full directory */	1206	/* read a full directory */
714	static void	1207	static void
715	eio__scandir (eio_req *req, worker *self)	1208	eio__scandir (eio_req *req, etp_worker *self)
716	{	1209	{
717	DIR *dirp;	1210	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;	1211	EIO_STRUCT_DIRENT *entp;
724	char *name, *names;	1212	char *name, *names;
725	int memlen = 4096;	1213	int namesalloc = 4096;
726	int memofs = 0;	1214	int namesoffs = 0;
		1215	int flags = req->int1;
		1216	eio_dirent *dents = 0;
		1217	int dentalloc = 128;
727	int res = 0;	1218	int dentoffs = 0;
		1219	ino_t inode_bits = 0;
		1220
		1221	req->result = -1;
		1222
		1223	if (!(flags & EIO_READDIR_DENTS))
		1224	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
728		1225
729	X_LOCK (wrklock);	1226	X_LOCK (wrklock);
		1227	/* the corresponding closedir is in ETP_WORKER_CLEAR */
730	self->dirp = dirp = opendir (req->ptr1);	1228	self->dirp = dirp = opendir (req->ptr1);
731	self->dbuf = u = malloc (sizeof (*u));
732	req->flags |= EIO_FLAG_PTR2_FREE;	1229	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1230	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
733	req->ptr2 = names = malloc (memlen);	1231	req->ptr2 = names = malloc (namesalloc);
734	X_UNLOCK (wrklock);	1232	X_UNLOCK (wrklock);
735		1233
736	if (dirp && u && names)	1234	if (dirp && names && (!flags || dents))
737	for (;;)	1235	for (;;)
738	{	1236	{
739	errno = 0;	1237	errno = 0;
740	readdir_r (dirp, &u->d, &entp);	1238	entp = readdir (dirp);
741		1239
742	if (!entp)	1240	if (!entp)
		1241	{
		1242	if (errno)
		1243	break;
		1244
		1245	/* sort etc. */
		1246	req->int1 = flags;
		1247	req->result = dentoffs;
		1248
		1249	if (flags & EIO_READDIR_STAT_ORDER)
		1250	eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
		1251	else if (flags & EIO_READDIR_DIRS_FIRST)
		1252	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1253	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1254	else
		1255	{
		1256	/* in this case, all is known, and we just put dirs first and sort them */
		1257	eio_dirent *oth = dents + dentoffs;
		1258	eio_dirent *dir = dents;
		1259
		1260	/* now partition dirs to the front, and non-dirs to the back */
		1261	/* by walking from both sides and swapping if necessary */
		1262	/* also clear score, so it doesn't influence sorting */
		1263	while (oth > dir)
		1264	{
		1265	if (dir->type == EIO_DT_DIR)
		1266	++dir;
		1267	else if ((--oth)->type == EIO_DT_DIR)
		1268	{
		1269	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1270
		1271	++dir;
		1272	}
		1273	}
		1274
		1275	/* now sort the dirs only */
		1276	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1277	}
		1278
743	break;	1279	break;
		1280	}
744		1281
		1282	/* now add the entry to our list(s) */
745	name = entp->d_name;	1283	name = entp->d_name;
746		1284
		1285	/* skip . and .. entries */
747	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1286	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
748	{	1287	{
749	int len = strlen (name) + 1;	1288	int len = D_NAMLEN (entp) + 1;
750		1289
751	res++;	1290	while (expect_false (namesoffs + len > namesalloc))
752
753	while (memofs + len > memlen)
754	{	1291	{
755	memlen *= 2;	1292	namesalloc *= 2;
756	X_LOCK (wrklock);	1293	X_LOCK (wrklock);
757	req->ptr2 = names = realloc (names, memlen);	1294	req->ptr2 = names = realloc (names, namesalloc);
758	X_UNLOCK (wrklock);	1295	X_UNLOCK (wrklock);
759		1296
760	if (!names)	1297	if (!names)
761	break;	1298	break;
762	}	1299	}
763		1300
764	memcpy (names + memofs, name, len);	1301	memcpy (names + namesoffs, name, len);
		1302
		1303	if (dents)
		1304	{
		1305	struct eio_dirent *ent;
		1306
		1307	if (expect_false (dentoffs == dentalloc))
		1308	{
		1309	dentalloc *= 2;
		1310	X_LOCK (wrklock);
		1311	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1312	X_UNLOCK (wrklock);
		1313
		1314	if (!dents)
		1315	break;
		1316	}
		1317
		1318	ent = dents + dentoffs;
		1319
		1320	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1321	ent->namelen = len - 1;
		1322	ent->inode = D_INO (entp);
		1323
		1324	inode_bits |= ent->inode;
		1325
		1326	switch (D_TYPE (entp))
		1327	{
		1328	default:
		1329	ent->type = EIO_DT_UNKNOWN;
		1330	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1331	break;
		1332
		1333	#ifdef DT_FIFO
		1334	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1335	#endif
		1336	#ifdef DT_CHR
		1337	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1338	#endif
		1339	#ifdef DT_MPC
		1340	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1341	#endif
		1342	#ifdef DT_DIR
		1343	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1344	#endif
		1345	#ifdef DT_NAM
		1346	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1347	#endif
		1348	#ifdef DT_BLK
		1349	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1350	#endif
		1351	#ifdef DT_MPB
		1352	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1353	#endif
		1354	#ifdef DT_REG
		1355	case DT_REG: ent->type = EIO_DT_REG; break;
		1356	#endif
		1357	#ifdef DT_NWK
		1358	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1359	#endif
		1360	#ifdef DT_CMP
		1361	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1362	#endif
		1363	#ifdef DT_LNK
		1364	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1365	#endif
		1366	#ifdef DT_SOCK
		1367	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1368	#endif
		1369	#ifdef DT_DOOR
		1370	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1371	#endif
		1372	#ifdef DT_WHT
		1373	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1374	#endif
		1375	}
		1376
		1377	ent->score = 7;
		1378
		1379	if (flags & EIO_READDIR_DIRS_FIRST)
		1380	{
		1381	if (ent->type == EIO_DT_UNKNOWN)
		1382	{
		1383	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1384	ent->score = 1;
		1385	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1386	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1387	}
		1388	else if (ent->type == EIO_DT_DIR)
		1389	ent->score = 0;
		1390	}
		1391	}
		1392
765	memofs += len;	1393	namesoffs += len;
		1394	++dentoffs;
		1395	}
		1396
		1397	if (EIO_CANCELLED (req))
		1398	{
		1399	errno = ECANCELED;
		1400	break;
766	}	1401	}
767	}	1402	}
		1403	}
768		1404
769	if (errno)	1405	#ifdef PAGESIZE
770	res = -1;	1406	# define eio_pagesize() PAGESIZE
		1407	#else
		1408	static intptr_t
		1409	eio_pagesize (void)
		1410	{
		1411	static intptr_t page;
		1412
		1413	if (!page)
		1414	page = sysconf (_SC_PAGESIZE);
		1415
		1416	return page;
		1417	}
		1418	#endif
		1419
		1420	static void
		1421	eio_page_align (void **addr, size_t *length)
		1422	{
		1423	intptr_t mask = eio_pagesize () - 1;
		1424
		1425	/* round down addr */
		1426	intptr_t adj = mask & (intptr_t)*addr;
		1427
		1428	*addr = (void *)((intptr_t)*addr - adj);
		1429	*length += adj;
		1430
		1431	/* round up length */
		1432	*length = (*length + mask) & ~mask;
		1433	}
		1434
		1435	#if !_POSIX_MEMLOCK
		1436	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1437	#else
		1438
		1439	static int
		1440	eio__mlockall (int flags)
		1441	{
		1442	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1443	extern int mallopt (int, int);
		1444	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1445	#endif
		1446
		1447	if (EIO_MCL_CURRENT != MCL_CURRENT
		1448	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1449	{
		1450	flags = 0
		1451	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1452	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1453	}
		1454
		1455	return mlockall (flags);
		1456	}
		1457	#endif
		1458
		1459	#if !_POSIX_MEMLOCK_RANGE
		1460	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1461	#else
		1462
		1463	static int
		1464	eio__mlock (void *addr, size_t length)
		1465	{
		1466	eio_page_align (&addr, &length);
		1467
		1468	return mlock (addr, length);
		1469	}
		1470
		1471	#endif
		1472
		1473	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1474	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1475	#else
		1476
		1477	int
		1478	eio__msync (void *mem, size_t len, int flags)
		1479	{
		1480	eio_page_align (&mem, &len);
		1481
		1482	if (EIO_MS_ASYNC != MS_SYNC
		1483	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1484	|| EIO_MS_SYNC != MS_SYNC)
		1485	{
		1486	flags = 0
		1487	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1488	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1489	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1490	}
		1491
		1492	return msync (mem, len, flags);
		1493	}
		1494
		1495	#endif
		1496
		1497	int
		1498	eio__mtouch (void *mem, size_t len, int flags)
		1499	{
		1500	eio_page_align (&mem, &len);
		1501
771		1502	{
772	req->result = res;	1503	intptr_t addr = (intptr_t)mem;
		1504	intptr_t end = addr + len;
		1505	intptr_t page = eio_pagesize ();
		1506
		1507	if (addr < end)
		1508	if (flags & EIO_MT_MODIFY) /* modify */
		1509	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
		1510	else
		1511	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
		1512	}
		1513
		1514	return 0;
773	}	1515	}
774		1516
775	/*****************************************************************************/	1517	/*****************************************************************************/
776		1518
777	#define ALLOC(len) \	1519	#define ALLOC(len) \
…		…
787	req->result = -1; \	1529	req->result = -1; \
788	break; \	1530	break; \
789	} \	1531	} \
790	}	1532	}
791		1533
792	X_THREAD_PROC (eio_proc)	1534	X_THREAD_PROC (etp_proc)
793	{	1535	{
794	eio_req *req;	1536	ETP_REQ *req;
795	struct timespec ts;	1537	struct timespec ts;
796	worker *self = (worker *)thr_arg;	1538	etp_worker *self = (etp_worker *)thr_arg;
797		1539
798	/* try to distribute timeouts somewhat randomly */	1540	/* try to distribute timeouts somewhat randomly */
799	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1541	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
800		1542
801	for (;;)	1543	for (;;)
802	{	1544	{
803	ts.tv_sec = time (0) + IDLE_TIMEOUT;
804
805	X_LOCK (reqlock);	1545	X_LOCK (reqlock);
806		1546
807	for (;;)	1547	for (;;)
808	{	1548	{
809	self->req = req = reqq_shift (&req_queue);	1549	self->req = req = reqq_shift (&req_queue);
…		…
811	if (req)	1551	if (req)
812	break;	1552	break;
813		1553
814	++idle;	1554	++idle;
815		1555
		1556	ts.tv_sec = time (0) + IDLE_TIMEOUT;
816	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1557	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
817	{	1558	{
818	if (idle > max_idle)	1559	if (idle > max_idle)
819	{	1560	{
820	--idle;	1561	--idle;
…		…
825	goto quit;	1566	goto quit;
826	}	1567	}
827		1568
828	/* we are allowed to idle, so do so without any timeout */	1569	/* we are allowed to idle, so do so without any timeout */
829	X_COND_WAIT (reqwait, reqlock);	1570	X_COND_WAIT (reqwait, reqlock);
830	ts.tv_sec = time (0) + IDLE_TIMEOUT;
831	}	1571	}
832		1572
833	--idle;	1573	--idle;
834	}	1574	}
835		1575
836	--nready;	1576	--nready;
837		1577
838	X_UNLOCK (reqlock);	1578	X_UNLOCK (reqlock);
839		1579
840	errno = 0; /* strictly unnecessary */	1580	if (req->type < 0)
		1581	goto quit;
841		1582
842	if (!EIO_CANCELLED (req))	1583	if (!EIO_CANCELLED (req))
843	switch (req->type)	1584	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		1585
944	X_LOCK (reslock);	1586	X_LOCK (reslock);
945		1587
946	++npending;	1588	++npending;
947		1589
948	if (!reqq_push (&res_queue, req) && want_poll_cb)	1590	if (!reqq_push (&res_queue, req) && want_poll_cb)
949	want_poll_cb ();	1591	want_poll_cb ();
950		1592
951	self->req = 0;	1593	self->req = 0;
952	worker_clear (self);	1594	etp_worker_clear (self);
953		1595
954	X_UNLOCK (reslock);	1596	X_UNLOCK (reslock);
955	}	1597	}
956		1598
957	quit:	1599	quit:
958	X_LOCK (wrklock);	1600	X_LOCK (wrklock);
959	worker_free (self);	1601	etp_worker_free (self);
960	X_UNLOCK (wrklock);	1602	X_UNLOCK (wrklock);
961		1603
962	return 0;	1604	return 0;
963	}	1605	}
964		1606
965	/*****************************************************************************/	1607	/*****************************************************************************/
966		1608
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))	1609	int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1024	{	1610	{
1025	want_poll_cb = want_poll;	1611	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	}	1612	}
1041		1613
1042	static void eio_api_destroy (eio_req *req)	1614	static void eio_api_destroy (eio_req *req)
1043	{	1615	{
1044	free (req);	1616	free (req);
…		…
1049	\	1621	\
1050	req = (eio_req *)calloc (1, sizeof *req); \	1622	req = (eio_req *)calloc (1, sizeof *req); \
1051	if (!req) \	1623	if (!req) \
1052	return 0; \	1624	return 0; \
1053	\	1625	\
1054	req->type = rtype; \	1626	req->type = rtype; \
1055	req->pri = EIO_DEFAULT_PRI + EIO_PRI_BIAS; \	1627	req->pri = pri; \
1056	req->finish = cb; \	1628	req->finish = cb; \
		1629	req->data = data; \
1057	req->destroy = eio_api_destroy;	1630	req->destroy = eio_api_destroy;
1058		1631
1059	#define SEND eio_submit (req); return req	1632	#define SEND eio_submit (req); return req
1060		1633
1061	#define PATH \	1634	#define PATH \
…		…
1065	{ \	1638	{ \
1066	eio_api_destroy (req); \	1639	eio_api_destroy (req); \
1067	return 0; \	1640	return 0; \
1068	}	1641	}
1069		1642
		1643	static void eio_execute (etp_worker *self, eio_req *req)
		1644	{
		1645	switch (req->type)
		1646	{
		1647	case EIO_READ: ALLOC (req->size);
		1648	req->result = req->offs >= 0
		1649	? pread (req->int1, req->ptr2, req->size, req->offs)
		1650	: read (req->int1, req->ptr2, req->size); break;
		1651	case EIO_WRITE: req->result = req->offs >= 0
		1652	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1653	: write (req->int1, req->ptr2, req->size); break;
		1654
		1655	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1656	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1657
		1658	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1659	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1660	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1661	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1662	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1663	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1664
		1665	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1666	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1667	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1668	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1669
		1670	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1671	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1672	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1673	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1674	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1675	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1676
		1677	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1678	case EIO_CLOSE: req->result = close (req->int1); break;
		1679	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1680	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1681	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1682	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1683	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1684	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1685	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1686	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
		1687
		1688	case EIO_READLINK: ALLOC (PATH_MAX);
		1689	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1690
		1691	case EIO_SYNC: req->result = 0; sync (); break;
		1692	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1693	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1694	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1695	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1696	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1697	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1698	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1699
		1700	case EIO_READDIR: eio__scandir (req, self); break;
		1701
		1702	case EIO_BUSY:
		1703	#ifdef _WIN32
		1704	Sleep (req->nv1 * 1e3);
		1705	#else
		1706	{
		1707	struct timeval tv;
		1708
		1709	tv.tv_sec = req->nv1;
		1710	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
		1711
		1712	req->result = select (0, 0, 0, 0, &tv);
		1713	}
		1714	#endif
		1715	break;
		1716
		1717	case EIO_UTIME:
		1718	case EIO_FUTIME:
		1719	{
		1720	struct timeval tv[2];
		1721	struct timeval *times;
		1722
		1723	if (req->nv1 != -1. || req->nv2 != -1.)
		1724	{
		1725	tv[0].tv_sec = req->nv1;
		1726	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1727	tv[1].tv_sec = req->nv2;
		1728	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1729
		1730	times = tv;
		1731	}
		1732	else
		1733	times = 0;
		1734
		1735	req->result = req->type == EIO_FUTIME
		1736	? futimes (req->int1, times)
		1737	: utimes (req->ptr1, times);
		1738	}
		1739	break;
		1740
		1741	case EIO_GROUP:
		1742	abort (); /* handled in eio_request */
		1743
		1744	case EIO_NOP:
		1745	req->result = 0;
		1746	break;
		1747
		1748	case EIO_CUSTOM:
		1749	((void (*)(eio_req *))req->feed) (req);
		1750	break;
		1751
		1752	default:
		1753	errno = ENOSYS;
		1754	req->result = -1;
		1755	break;
		1756	}
		1757
		1758	req->errorno = errno;
		1759	}
		1760
		1761	#ifndef EIO_NO_WRAPPERS
		1762
1070	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1763	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1071	{	1764	{
1072	REQ (EIO_NOP); SEND;	1765	REQ (EIO_NOP); SEND;
1073	}	1766	}
1074		1767
…		…
1085	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1778	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1086	{	1779	{
1087	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1780	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1088	}	1781	}
1089		1782
		1783	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1784	{
		1785	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1786	}
		1787
		1788	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1789	{
		1790	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1791	}
		1792
		1793	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1794	{
		1795	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1796	}
		1797
		1798	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1799	{
		1800	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1801	}
		1802
		1803	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1804	{
		1805	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1806	}
		1807
1090	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1808	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1091	{	1809	{
1092	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1810	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1093	}	1811	}
1094		1812
…		…
1113	}	1831	}
1114		1832
1115	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1833	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1116	{	1834	{
1117	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1835	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1836	}
		1837
		1838	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1839	{
		1840	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1118	}	1841	}
1119		1842
1120	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1843	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1121	{	1844	{
1122	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1845	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)	1919	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1197	{	1920	{
1198	return eio__1path (EIO_LSTAT, path, pri, cb, data);	1921	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1199	}	1922	}
1200		1923
		1924	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		1925	{
		1926	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		1927	}
		1928
1201	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	1929	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1202	{	1930	{
1203	return eio__1path (EIO_UNLINK, path, pri, cb, data);	1931	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1204	}	1932	}
1205		1933
1206	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	1934	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1207	{	1935	{
1208	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1936	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1209	}	1937	}
1210		1938
1211	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	1939	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1212	{	1940	{
1213	return eio__1path (EIO_READDIR, path, pri, cb, data);	1941	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1214	}	1942	}
1215		1943
1216	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1944	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1217	{	1945	{
1218	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	1946	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1219	}	1947	}
1220		1948
1221	static eio_req *	1949	static eio_req *
1222	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1950	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1223	{	1951	{
…		…
1247	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1975	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1248	{	1976	{
1249	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	1977	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1250	}	1978	}
1251		1979
		1980	eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
		1981	{
		1982	REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
		1983	}
		1984
		1985	#endif
		1986
1252	eio_req *eio_grp (eio_cb cb, void *data)	1987	eio_req *eio_grp (eio_cb cb, void *data)
1253	{	1988	{
1254	const int pri = EIO_PRI_MAX;	1989	const int pri = EIO_PRI_MAX;
1255		1990
1256	REQ (EIO_GROUP); SEND;	1991	REQ (EIO_GROUP); SEND;
…		…
1280		2015
1281	void eio_grp_add (eio_req *grp, eio_req *req)	2016	void eio_grp_add (eio_req *grp, eio_req *req)
1282	{	2017	{
1283	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2018	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1284		2019
		2020	grp->flags |= EIO_FLAG_GROUPADD;
		2021
1285	++grp->size;	2022	++grp->size;
1286	req->grp = grp;	2023	req->grp = grp;
1287		2024
1288	req->grp_prev = 0;	2025	req->grp_prev = 0;
1289	req->grp_next = grp->grp_first;	2026	req->grp_next = grp->grp_first;
…		…
1297	/*****************************************************************************/	2034	/*****************************************************************************/
1298	/* misc garbage */	2035	/* misc garbage */
1299		2036
1300	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2037	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1301	{	2038	{
1302	worker wrk;	2039	etp_worker wrk;
		2040	ssize_t ret;
1303		2041
1304	wrk.dbuf = 0;	2042	wrk.dbuf = 0;
1305		2043
1306	eio__sendfile (ofd, ifd, offset, count, &wrk);	2044	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1307		2045
1308	if (wrk.dbuf)	2046	if (wrk.dbuf)
1309	free (wrk.dbuf);	2047	free (wrk.dbuf);
1310	}
1311		2048
		2049	return ret;
		2050	}
		2051

Diff Legend

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