ViewVC Help

View File | Revision Log | Show Annotations | Download File

/cvs/cvsroot/libeio/eio.c

Comparing cvsroot/libeio/eio.c (file contents):
Revision 1.12 by root, Tue May 13 18:54:52 2008 UTC vs.
Revision 1.65 by root, Mon May 30 12:56:50 2011 UTC

		1	/*
		2	* libeio implementation
		3	*
		4	* Copyright (c) 2007,2008,2009,2010,2011 Marc Alexander Lehmann <libeio@schmorp.de>
		5	* All rights reserved.
		6	*
		7	* Redistribution and use in source and binary forms, with or without modifica-
		8	* tion, are permitted provided that the following conditions are met:
		9	*
		10	* 1. Redistributions of source code must retain the above copyright notice,
		11	* this list of conditions and the following disclaimer.
		12	*
		13	* 2. Redistributions in binary form must reproduce the above copyright
		14	* notice, this list of conditions and the following disclaimer in the
		15	* documentation and/or other materials provided with the distribution.
		16	*
		17	* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
		18	* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
		19	* CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
		20	* EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
		21	* CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
		22	* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
		23	* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
		24	* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
		25	* ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
		26	* OF THE POSSIBILITY OF SUCH DAMAGE.
		27	*
		28	* Alternatively, the contents of this file may be used under the terms of
		29	* the GNU General Public License ("GPL") version 2 or any later version,
		30	* in which case the provisions of the GPL are applicable instead of
		31	* the above. If you wish to allow the use of your version of this file
		32	* only under the terms of the GPL and not to allow others to use your
		33	* version of this file under the BSD license, indicate your decision
		34	* by deleting the provisions above and replace them with the notice
		35	* and other provisions required by the GPL. If you do not delete the
		36	* provisions above, a recipient may use your version of this file under
		37	* either the BSD or the GPL.
		38	*/
		39
		40	#ifndef _WIN32
		41	# include "config.h"
		42	#endif
		43
1	#include "eio.h"	44	#include "eio.h"
		45
		46	#ifdef EIO_STACKSIZE
		47	# define XTHREAD_STACKSIZE EIO_STACKSIZE
		48	#endif
2	#include "xthread.h"	49	#include "xthread.h"
3		50
4	#include <errno.h>	51	#include <errno.h>
5	#include <stddef.h>	52	#include <stddef.h>
6	#include <stdlib.h>	53	#include <stdlib.h>
7	#include <string.h>	54	#include <string.h>
8	#include <errno.h>	55	#include <errno.h>
9	#include <sys/types.h>	56	#include <sys/types.h>
10	#include <sys/stat.h>	57	#include <sys/stat.h>
		58	#include <sys/statvfs.h>
11	#include <limits.h>	59	#include <limits.h>
12	#include <fcntl.h>	60	#include <fcntl.h>
13	#include <assert.h>	61	#include <assert.h>
14		62
15	#ifndef EIO_FINISH	63	#ifndef EIO_FINISH
…		…
25	#endif	73	#endif
26		74
27	#ifdef _WIN32	75	#ifdef _WIN32
28		76
29	/*doh*/	77	/*doh*/
30
31	#else	78	#else
32		79
33	# include "config.h"
34	# include <sys/time.h>	80	# include <sys/time.h>
35	# include <sys/select.h>	81	# include <sys/select.h>
36	# include <unistd.h>	82	# include <unistd.h>
37	# include <utime.h>	83	# include <utime.h>
38	# include <signal.h>	84	# include <signal.h>
39	# include <dirent.h>	85	# include <dirent.h>
40		86
		87	#if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
		88	# include <sys/mman.h>
		89	#endif
		90
		91	/* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
		92	# if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
		93	# define _DIRENT_HAVE_D_TYPE /* sigh */
		94	# define D_INO(de) (de)->d_fileno
		95	# define D_NAMLEN(de) (de)->d_namlen
		96	# elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
		97	# define D_INO(de) (de)->d_ino
		98	# endif
		99
		100	#ifdef _D_EXACT_NAMLEN
		101	# undef D_NAMLEN
		102	# define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
		103	#endif
		104
		105	# ifdef _DIRENT_HAVE_D_TYPE
		106	# define D_TYPE(de) (de)->d_type
		107	# endif
		108
41	# ifndef EIO_STRUCT_DIRENT	109	# ifndef EIO_STRUCT_DIRENT
42	# define EIO_STRUCT_DIRENT struct dirent	110	# define EIO_STRUCT_DIRENT struct dirent
43	# endif	111	# endif
44		112
45	#endif	113	#endif
46		114
47	#if HAVE_SENDFILE	115	#if HAVE_SENDFILE
48	# if __linux	116	# if __linux
49	# include <sys/sendfile.h>	117	# include <sys/sendfile.h>
50	# elif __freebsd	118	# elif __FreeBSD__ || defined __APPLE__
51	# include <sys/socket.h>	119	# include <sys/socket.h>
52	# include <sys/uio.h>	120	# include <sys/uio.h>
53	# elif __hpux	121	# elif __hpux
54	# include <sys/socket.h>	122	# include <sys/socket.h>
55	# elif __solaris /* not yet */	123	# elif __solaris
56	# include <sys/sendfile.h>	124	# include <sys/sendfile.h>
57	# else	125	# else
58	# error sendfile support requested but not available	126	# error sendfile support requested but not available
59	# endif	127	# endif
60	#endif	128	#endif
61		129
62	/* number of seconds after which an idle threads exit */	130	#ifndef D_TYPE
63	#define IDLE_TIMEOUT 10	131	# define D_TYPE(de) 0
		132	#endif
		133	#ifndef D_INO
		134	# define D_INO(de) 0
		135	#endif
		136	#ifndef D_NAMLEN
		137	# define D_NAMLEN(de) strlen ((de)->d_name)
		138	#endif
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; /* maximum number of threads that can idle indefinitely */
94	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);	229	static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */
95	}
96		230
97	static unsigned int started, idle, wanted = 4;	231	static xmutex_t wrklock;
		232	static xmutex_t reslock;
		233	static xmutex_t reqlock;
		234	static xcond_t reqwait;
		235
		236	#if !HAVE_PREADWRITE
		237	/*
		238	* make our pread/pwrite emulation safe against themselves, but not against
		239	* normal read/write by using a mutex. slows down execution a lot,
		240	* but that's your problem, not mine.
		241	*/
		242	static xmutex_t preadwritelock = X_MUTEX_INIT;
		243	#endif
		244
		245	typedef struct etp_worker
		246	{
		247	/* locked by wrklock */
		248	struct etp_worker *prev, *next;
		249
		250	xthread_t tid;
		251
		252	/* locked by reslock, reqlock or wrklock */
		253	ETP_REQ *req; /* currently processed request */
		254
		255	ETP_WORKER_COMMON
		256	} etp_worker;
		257
		258	static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
		259
		260	#define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
		261	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
98		262
99	/* worker threads management */	263	/* worker threads management */
100	static mutex_t wrklock = X_MUTEX_INIT;
101		264
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)	265	static void etp_worker_clear (etp_worker *wrk)
118	{	266	{
119	if (wrk->dirp)	267	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	}	268	}
131		269
132	static void worker_free (worker *wrk)	270	static void etp_worker_free (etp_worker *wrk)
133	{	271	{
134	wrk->next->prev = wrk->prev;	272	wrk->next->prev = wrk->prev;
135	wrk->prev->next = wrk->next;	273	wrk->prev->next = wrk->next;
136		274
137	free (wrk);	275	free (wrk);
138	}	276	}
139		277
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)	278	static unsigned int etp_nreqs (void)
148	{	279	{
		280	int retval;
		281	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		282	retval = nreqs;
		283	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
149	return nreqs;	284	return retval;
150	}	285	}
151		286
152	unsigned int eio_nready (void)	287	static unsigned int etp_nready (void)
153	{	288	{
154	unsigned int retval;	289	unsigned int retval;
155		290
156	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	291	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
157	retval = nready;	292	retval = nready;
158	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	293	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
159		294
160	return retval;	295	return retval;
161	}	296	}
162		297
163	unsigned int eio_npending (void)	298	static unsigned int etp_npending (void)
164	{	299	{
165	unsigned int retval;	300	unsigned int retval;
166		301
167	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	302	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
168	retval = npending;	303	retval = npending;
169	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);	304	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
170		305
171	return retval;	306	return retval;
172	}	307	}
173		308
174	unsigned int eio_nthreads (void)	309	static unsigned int etp_nthreads (void)
175	{	310	{
176	unsigned int retval;	311	unsigned int retval;
177		312
178	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);	313	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
179	retval = started;	314	retval = started;
…		…
186	* a somewhat faster data structure might be nice, but	321	* a somewhat faster data structure might be nice, but
187	* with 8 priorities this actually needs <20 insns	322	* with 8 priorities this actually needs <20 insns
188	* per shift, the most expensive operation.	323	* per shift, the most expensive operation.
189	*/	324	*/
190	typedef struct {	325	typedef struct {
191	eio_req *qs[EIO_NUM_PRI], *qe[EIO_NUM_PRI]; /* qstart, qend */	326	ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
192	int size;	327	int size;
193	} reqq;	328	} etp_reqq;
194		329
195	static reqq req_queue;	330	static etp_reqq req_queue;
196	static reqq res_queue;	331	static etp_reqq res_queue;
197		332
198	static int reqq_push (reqq *q, eio_req *req)	333	static int reqq_push (etp_reqq *q, ETP_REQ *req)
199	{	334	{
200	int pri = req->pri;	335	int pri = req->pri;
201	req->next = 0;	336	req->next = 0;
202		337
203	if (q->qe[pri])	338	if (q->qe[pri])
…		…
209	q->qe[pri] = q->qs[pri] = req;	344	q->qe[pri] = q->qs[pri] = req;
210		345
211	return q->size++;	346	return q->size++;
212	}	347	}
213		348
214	static eio_req *reqq_shift (reqq *q)	349	static ETP_REQ *reqq_shift (etp_reqq *q)
215	{	350	{
216	int pri;	351	int pri;
217		352
218	if (!q->size)	353	if (!q->size)
219	return 0;	354	return 0;
220		355
221	--q->size;	356	--q->size;
222		357
223	for (pri = EIO_NUM_PRI; pri--; )	358	for (pri = ETP_NUM_PRI; pri--; )
224	{	359	{
225	eio_req *req = q->qs[pri];	360	eio_req *req = q->qs[pri];
226		361
227	if (req)	362	if (req)
228	{	363	{
…		…
234	}	369	}
235		370
236	abort ();	371	abort ();
237	}	372	}
238		373
239	static void grp_try_feed (eio_req *grp)	374	static void etp_thread_init (void)
240	{	375	{
241	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))	376	X_MUTEX_CREATE (wrklock);
		377	X_MUTEX_CREATE (reslock);
		378	X_MUTEX_CREATE (reqlock);
		379	X_COND_CREATE (reqwait);
		380	}
		381
		382	static void etp_atfork_prepare (void)
		383	{
		384	X_LOCK (wrklock);
		385	X_LOCK (reqlock);
		386	X_LOCK (reslock);
		387	#if !HAVE_PREADWRITE
		388	X_LOCK (preadwritelock);
		389	#endif
		390	}
		391
		392	static void etp_atfork_parent (void)
		393	{
		394	#if !HAVE_PREADWRITE
		395	X_UNLOCK (preadwritelock);
		396	#endif
		397	X_UNLOCK (reslock);
		398	X_UNLOCK (reqlock);
		399	X_UNLOCK (wrklock);
		400	}
		401
		402	static void etp_atfork_child (void)
		403	{
		404	ETP_REQ *prv;
		405
		406	while ((prv = reqq_shift (&req_queue)))
		407	ETP_DESTROY (prv);
		408
		409	while ((prv = reqq_shift (&res_queue)))
		410	ETP_DESTROY (prv);
		411
		412	while (wrk_first.next != &wrk_first)
242	{	413	{
243	int old_len = grp->size;	414	etp_worker *wrk = wrk_first.next;
244		415
245	EIO_FEED (grp);	416	if (wrk->req)
		417	ETP_DESTROY (wrk->req);
246		418
247	/* stop if no progress has been made */	419	etp_worker_clear (wrk);
248	if (old_len == grp->size)	420	etp_worker_free (wrk);
249	{
250	grp->feed = 0;
251	break;
252	}
253	}	421	}
254	}
255		422
256	static int eio_finish (eio_req *req);	423	started = 0;
		424	idle = 0;
		425	nreqs = 0;
		426	nready = 0;
		427	npending = 0;
257		428
258	static int grp_dec (eio_req *grp)	429	etp_thread_init ();
259	{	430	}
260	--grp->size;
261		431
262	/* call feeder, if applicable */	432	static void
263	grp_try_feed (grp);	433	etp_once_init (void)
		434	{
		435	etp_thread_init ();
		436	X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
		437	}
264		438
265	/* finish, if done */	439	static int
266	if (!grp->size && grp->int1)	440	etp_init (void (*want_poll)(void), void (*done_poll)(void))
267	return eio_finish (grp);	441	{
268	else	442	static pthread_once_t doinit = PTHREAD_ONCE_INIT;
		443
		444	pthread_once (&doinit, etp_once_init);
		445
		446	want_poll_cb = want_poll;
		447	done_poll_cb = done_poll;
		448
269	return 0;	449	return 0;
270	}	450	}
271		451
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);	452	X_THREAD_PROC (etp_proc);
323		453
324	static void start_thread (void)	454	static void etp_start_thread (void)
325	{	455	{
326	worker *wrk = calloc (1, sizeof (worker));	456	etp_worker *wrk = calloc (1, sizeof (etp_worker));
327		457
328	/*TODO*/	458	/*TODO*/
329	assert (("unable to allocate worker thread data", wrk));	459	assert (("unable to allocate worker thread data", wrk));
330		460
331	X_LOCK (wrklock);	461	X_LOCK (wrklock);
332		462
333	if (thread_create (&wrk->tid, eio_proc, (void *)wrk))	463	if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
334	{	464	{
335	wrk->prev = &wrk_first;	465	wrk->prev = &wrk_first;
336	wrk->next = wrk_first.next;	466	wrk->next = wrk_first.next;
337	wrk_first.next->prev = wrk;	467	wrk_first.next->prev = wrk;
338	wrk_first.next = wrk;	468	wrk_first.next = wrk;
…		…
342	free (wrk);	472	free (wrk);
343		473
344	X_UNLOCK (wrklock);	474	X_UNLOCK (wrklock);
345	}	475	}
346		476
347	static void maybe_start_thread (void)	477	static void etp_maybe_start_thread (void)
348	{	478	{
349	if (eio_nthreads () >= wanted)	479	if (expect_true (etp_nthreads () >= wanted))
350	return;	480	return;
351		481
352	/* todo: maybe use idle here, but might be less exact */	482	/* todo: maybe use idle here, but might be less exact */
353	if (0 <= (int)eio_nthreads () + (int)eio_npending () - (int)eio_nreqs ())	483	if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
354	return;	484	return;
355		485
356	start_thread ();	486	etp_start_thread ();
357	}	487	}
358		488
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)	489	static void etp_end_thread (void)
378	{	490	{
379	eio_req *req = calloc (1, sizeof (eio_req));	491	eio_req *req = calloc (1, sizeof (eio_req));
380		492
381	req->type = EIO_QUIT;	493	req->type = -1;
382	req->pri = EIO_PRI_MAX + EIO_PRI_BIAS;	494	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
383		495
384	X_LOCK (reqlock);	496	X_LOCK (reqlock);
385	reqq_push (&req_queue, req);	497	reqq_push (&req_queue, req);
386	X_COND_SIGNAL (reqwait);	498	X_COND_SIGNAL (reqwait);
387	X_UNLOCK (reqlock);	499	X_UNLOCK (reqlock);
…		…
389	X_LOCK (wrklock);	501	X_LOCK (wrklock);
390	--started;	502	--started;
391	X_UNLOCK (wrklock);	503	X_UNLOCK (wrklock);
392	}	504	}
393		505
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)	506	static int etp_poll (void)
431	{	507	{
432	int maxreqs = max_poll_reqs;	508	unsigned int maxreqs;
		509	unsigned int maxtime;
433	struct timeval tv_start, tv_now;	510	struct timeval tv_start, tv_now;
434	eio_req *req;
435		511
		512	X_LOCK (reslock);
		513	maxreqs = max_poll_reqs;
		514	maxtime = max_poll_time;
		515	X_UNLOCK (reslock);
		516
436	if (max_poll_time)	517	if (maxtime)
437	gettimeofday (&tv_start, 0);	518	gettimeofday (&tv_start, 0);
438		519
439	for (;;)	520	for (;;)
440	{	521	{
		522	ETP_REQ *req;
		523
441	maybe_start_thread ();	524	etp_maybe_start_thread ();
442		525
443	X_LOCK (reslock);	526	X_LOCK (reslock);
444	req = reqq_shift (&res_queue);	527	req = reqq_shift (&res_queue);
445		528
446	if (req)	529	if (req)
…		…
454	X_UNLOCK (reslock);	537	X_UNLOCK (reslock);
455		538
456	if (!req)	539	if (!req)
457	return 0;	540	return 0;
458		541
		542	X_LOCK (reqlock);
459	--nreqs;	543	--nreqs;
		544	X_UNLOCK (reqlock);
460		545
461	if (req->type == EIO_GROUP && req->size)	546	if (expect_false (req->type == EIO_GROUP && req->size))
462	{	547	{
463	req->int1 = 1; /* mark request as delayed */	548	req->int1 = 1; /* mark request as delayed */
464	continue;	549	continue;
465	}	550	}
466	else	551	else
467	{	552	{
468	int res = eio_finish (req);	553	int res = ETP_FINISH (req);
469	if (res)	554	if (expect_false (res))
470	return res;	555	return res;
471	}	556	}
472		557
473	if (maxreqs && !--maxreqs)	558	if (expect_false (maxreqs && !--maxreqs))
474	break;	559	break;
475		560
476	if (max_poll_time)	561	if (maxtime)
477	{	562	{
478	gettimeofday (&tv_now, 0);	563	gettimeofday (&tv_now, 0);
479		564
480	if (tvdiff (&tv_start, &tv_now) >= max_poll_time)	565	if (tvdiff (&tv_start, &tv_now) >= maxtime)
481	break;	566	break;
482	}	567	}
483	}	568	}
484		569
485	errno = EAGAIN;	570	errno = EAGAIN;
486	return -1;	571	return -1;
487	}	572	}
488		573
		574	static void etp_cancel (ETP_REQ *req)
		575	{
		576	X_LOCK (wrklock);
		577	req->flags |= EIO_FLAG_CANCELLED;
		578	X_UNLOCK (wrklock);
		579
		580	eio_grp_cancel (req);
		581	}
		582
		583	static void etp_submit (ETP_REQ *req)
		584	{
		585	req->pri -= ETP_PRI_MIN;
		586
		587	if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
		588	if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
		589
		590	if (expect_false (req->type == EIO_GROUP))
		591	{
		592	/* I hope this is worth it :/ */
		593	X_LOCK (reqlock);
		594	++nreqs;
		595	X_UNLOCK (reqlock);
		596
		597	X_LOCK (reslock);
		598
		599	++npending;
		600
		601	if (!reqq_push (&res_queue, req) && want_poll_cb)
		602	want_poll_cb ();
		603
		604	X_UNLOCK (reslock);
		605	}
		606	else
		607	{
		608	X_LOCK (reqlock);
		609	++nreqs;
		610	++nready;
		611	reqq_push (&req_queue, req);
		612	X_COND_SIGNAL (reqwait);
		613	X_UNLOCK (reqlock);
		614
		615	etp_maybe_start_thread ();
		616	}
		617	}
		618
		619	static void etp_set_max_poll_time (double nseconds)
		620	{
		621	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		622	max_poll_time = nseconds * EIO_TICKS;
		623	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		624	}
		625
		626	static void etp_set_max_poll_reqs (unsigned int maxreqs)
		627	{
		628	if (WORDACCESS_UNSAFE) X_LOCK (reslock);
		629	max_poll_reqs = maxreqs;
		630	if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
		631	}
		632
		633	static void etp_set_max_idle (unsigned int nthreads)
		634	{
		635	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		636	max_idle = nthreads;
		637	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		638	}
		639
		640	static void etp_set_idle_timeout (unsigned int seconds)
		641	{
		642	if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
		643	idle_timeout = seconds;
		644	if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
		645	}
		646
		647	static void etp_set_min_parallel (unsigned int nthreads)
		648	{
		649	if (wanted < nthreads)
		650	wanted = nthreads;
		651	}
		652
		653	static void etp_set_max_parallel (unsigned int nthreads)
		654	{
		655	if (wanted > nthreads)
		656	wanted = nthreads;
		657
		658	while (started > wanted)
		659	etp_end_thread ();
		660	}
		661
		662	/*****************************************************************************/
		663
		664	static void grp_try_feed (eio_req *grp)
		665	{
		666	while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
		667	{
		668	grp->flags &= ~EIO_FLAG_GROUPADD;
		669
		670	EIO_FEED (grp);
		671
		672	/* stop if no progress has been made */
		673	if (!(grp->flags & EIO_FLAG_GROUPADD))
		674	{
		675	grp->feed = 0;
		676	break;
		677	}
		678	}
		679	}
		680
		681	static int grp_dec (eio_req *grp)
		682	{
		683	--grp->size;
		684
		685	/* call feeder, if applicable */
		686	grp_try_feed (grp);
		687
		688	/* finish, if done */
		689	if (!grp->size && grp->int1)
		690	return eio_finish (grp);
		691	else
		692	return 0;
		693	}
		694
		695	void eio_destroy (eio_req *req)
		696	{
		697	if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
		698	if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
		699
		700	EIO_DESTROY (req);
		701	}
		702
		703	static int eio_finish (eio_req *req)
		704	{
		705	int res = EIO_FINISH (req);
		706
		707	if (req->grp)
		708	{
		709	int res2;
		710	eio_req *grp = req->grp;
		711
		712	/* unlink request */
		713	if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
		714	if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
		715
		716	if (grp->grp_first == req)
		717	grp->grp_first = req->grp_next;
		718
		719	res2 = grp_dec (grp);
		720
		721	if (!res && res2)
		722	res = res2;
		723	}
		724
		725	eio_destroy (req);
		726
		727	return res;
		728	}
		729
		730	void eio_grp_cancel (eio_req *grp)
		731	{
		732	for (grp = grp->grp_first; grp; grp = grp->grp_next)
		733	eio_cancel (grp);
		734	}
		735
		736	void eio_cancel (eio_req *req)
		737	{
		738	etp_cancel (req);
		739	}
		740
		741	void eio_submit (eio_req *req)
		742	{
		743	etp_submit (req);
		744	}
		745
		746	unsigned int eio_nreqs (void)
		747	{
		748	return etp_nreqs ();
		749	}
		750
		751	unsigned int eio_nready (void)
		752	{
		753	return etp_nready ();
		754	}
		755
		756	unsigned int eio_npending (void)
		757	{
		758	return etp_npending ();
		759	}
		760
		761	unsigned int eio_nthreads (void)
		762	{
		763	return etp_nthreads ();
		764	}
		765
		766	void eio_set_max_poll_time (double nseconds)
		767	{
		768	etp_set_max_poll_time (nseconds);
		769	}
		770
		771	void eio_set_max_poll_reqs (unsigned int maxreqs)
		772	{
		773	etp_set_max_poll_reqs (maxreqs);
		774	}
		775
		776	void eio_set_max_idle (unsigned int nthreads)
		777	{
		778	etp_set_max_idle (nthreads);
		779	}
		780
		781	void eio_set_idle_timeout (unsigned int seconds)
		782	{
		783	etp_set_idle_timeout (seconds);
		784	}
		785
		786	void eio_set_min_parallel (unsigned int nthreads)
		787	{
		788	etp_set_min_parallel (nthreads);
		789	}
		790
		791	void eio_set_max_parallel (unsigned int nthreads)
		792	{
		793	etp_set_max_parallel (nthreads);
		794	}
		795
		796	int eio_poll (void)
		797	{
		798	return etp_poll ();
		799	}
		800
489	/*****************************************************************************/	801	/*****************************************************************************/
490	/* work around various missing functions */	802	/* work around various missing functions */
491		803
492	#if !HAVE_PREADWRITE	804	#if !HAVE_PREADWRITE
		805	# undef pread
		806	# undef pwrite
493	# define pread eio__pread	807	# define pread eio__pread
494	# define pwrite eio__pwrite	808	# 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		809
503	static ssize_t	810	static ssize_t
504	eio__pread (int fd, void *buf, size_t count, off_t offset)	811	eio__pread (int fd, void *buf, size_t count, off_t offset)
505	{	812	{
506	ssize_t res;	813	ssize_t res;
…		…
524		831
525	X_LOCK (preadwritelock);	832	X_LOCK (preadwritelock);
526	ooffset = lseek (fd, 0, SEEK_CUR);	833	ooffset = lseek (fd, 0, SEEK_CUR);
527	lseek (fd, offset, SEEK_SET);	834	lseek (fd, offset, SEEK_SET);
528	res = write (fd, buf, count);	835	res = write (fd, buf, count);
529	lseek (fd, offset, SEEK_SET);	836	lseek (fd, ooffset, SEEK_SET);
530	X_UNLOCK (preadwritelock);	837	X_UNLOCK (preadwritelock);
531		838
532	return res;	839	return res;
533	}	840	}
534	#endif	841	#endif
535		842
536	#ifndef HAVE_FUTIMES	843	#ifndef HAVE_UTIMES
537		844
		845	# undef utimes
538	# define utimes(path,times) eio__utimes (path, times)	846	# define utimes(path,times) eio__utimes (path, times)
539	# define futimes(fd,times) eio__futimes (fd, times)
540		847
541	static int	848	static int
542	eio__utimes (const char *filename, const struct timeval times[2])	849	eio__utimes (const char *filename, const struct timeval times[2])
543	{	850	{
544	if (times)	851	if (times)
…		…
552	}	859	}
553	else	860	else
554	return utime (filename, 0);	861	return utime (filename, 0);
555	}	862	}
556		863
		864	#endif
		865
		866	#ifndef HAVE_FUTIMES
		867
		868	# undef futimes
		869	# define futimes(fd,times) eio__futimes (fd, times)
		870
557	static int eio__futimes (int fd, const struct timeval tv[2])	871	static int eio__futimes (int fd, const struct timeval tv[2])
558	{	872	{
559	errno = ENOSYS;	873	errno = ENOSYS;
560	return -1;	874	return -1;
561	}	875	}
562		876
563	#endif	877	#endif
564		878
565	#if !HAVE_FDATASYNC	879	#if !HAVE_FDATASYNC
		880	# undef fdatasync
566	# define fdatasync fsync	881	# define fdatasync(fd) fsync (fd)
567	#endif	882	#endif
		883
		884	/* sync_file_range always needs emulation */
		885	int
		886	eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
		887	{
		888	#if HAVE_SYNC_FILE_RANGE
		889	int res;
		890
		891	if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
		892	|| EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
		893	|| EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
		894	{
		895	flags = 0
		896	| (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
		897	| (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
		898	| (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
		899	}
		900
		901	res = sync_file_range (fd, offset, nbytes, flags);
		902
		903	if (!res || errno != ENOSYS)
		904	return res;
		905	#endif
		906
		907	/* even though we could play tricks with the flags, it's better to always
		908	* call fdatasync, as that matches the expectation of its users best */
		909	return fdatasync (fd);
		910	}
568		911
569	#if !HAVE_READAHEAD	912	#if !HAVE_READAHEAD
		913	# undef readahead
570	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)	914	# define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
571		915
572	static ssize_t	916	static ssize_t
573	eio__readahead (int fd, off_t offset, size_t count, worker *self)	917	eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
574	{	918	{
575	size_t todo = count;	919	size_t todo = count;
576	dBUF;	920	dBUF;
577		921
578	while (todo > 0)	922	while (todo > 0)
…		…
588	return count;	932	return count;
589	}	933	}
590		934
591	#endif	935	#endif
592		936
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 */	937	/* sendfile always needs emulation */
625	static ssize_t	938	static ssize_t
626	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, worker *self)	939	eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
627	{	940	{
628	ssize_t res;	941	ssize_t res;
629		942
630	if (!count)	943	if (!count)
631	return 0;	944	return 0;
632		945
633	#if HAVE_SENDFILE	946	#if HAVE_SENDFILE
634	# if __linux	947	# if __linux
635	res = sendfile (ofd, ifd, &offset, count);	948	res = sendfile (ofd, ifd, &offset, count);
636		949
637	# elif __freebsd	950	# elif __FreeBSD__
638	/*	951	/*
639	* Of course, the freebsd sendfile is a dire hack with no thoughts	952	* Of course, the freebsd sendfile is a dire hack with no thoughts
640	* wasted on making it similar to other I/O functions.	953	* wasted on making it similar to other I/O functions.
641	*/	954	*/
642	{	955	{
643	off_t sbytes;	956	off_t sbytes;
644	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);	957	res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
645		958
646	if (res < 0 && sbytes)	959	#if 0 /* according to the manpage, this is correct, but broken behaviour */
647	/* maybe only on EAGAIN: as usual, the manpage leaves you guessing */	960	/* freebsd' sendfile will return 0 on success */
		961	/* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
		962	/* not on e.g. EIO or EPIPE - sounds broken */
		963	if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
		964	res = sbytes;
		965	#endif
		966
		967	/* according to source inspection, this is correct, and useful behaviour */
		968	if (sbytes)
		969	res = sbytes;
		970	}
		971
		972	# elif defined (__APPLE__)
		973
		974	{
		975	off_t sbytes = count;
		976	res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
		977
		978	/* according to the manpage, sbytes is always valid */
		979	if (sbytes)
648	res = sbytes;	980	res = sbytes;
649	}	981	}
650		982
651	# elif __hpux	983	# elif __hpux
652	res = sendfile (ofd, ifd, offset, count, 0, 0);	984	res = sendfile (ofd, ifd, offset, count, 0, 0);
…		…
666	if (res < 0 && sbytes)	998	if (res < 0 && sbytes)
667	res = sbytes;	999	res = sbytes;
668	}	1000	}
669		1001
670	# endif	1002	# endif
		1003
		1004	#elif defined (_WIN32)
		1005
		1006	/* does not work, just for documentation of what would need to be done */
		1007	{
		1008	HANDLE h = TO_SOCKET (ifd);
		1009	SetFilePointer (h, offset, 0, FILE_BEGIN);
		1010	res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
		1011	}
		1012
671	#else	1013	#else
672	res = -1;	1014	res = -1;
673	errno = ENOSYS;	1015	errno = ENOSYS;
674	#endif	1016	#endif
675		1017
676	if (res < 0	1018	if (res < 0
677	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK	1019	&& (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
		1020	/* BSDs */
		1021	#ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
		1022	|| errno == ENOTSUP
		1023	#endif
		1024	|| errno == EOPNOTSUPP /* BSDs */
678	#if __solaris	1025	#if __solaris
679	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE	1026	|| errno == EAFNOSUPPORT || errno == EPROTOTYPE
680	#endif	1027	#endif
681	)	1028	)
682	)	1029	)
…		…
713	}	1060	}
714		1061
715	return res;	1062	return res;
716	}	1063	}
717		1064
		1065	static signed char
		1066	eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
		1067	{
		1068	return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
		1069	: a->inode < b->inode ? -1
		1070	: a->inode > b->inode ? 1
		1071	: 0;
		1072	}
		1073
		1074	#define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
		1075
		1076	#define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
		1077	#define EIO_SORT_FAST 60 /* when to only use insertion sort */
		1078
		1079	static void
		1080	eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1081	{
		1082	unsigned char bits [9 + sizeof (ino_t) * 8];
		1083	unsigned char *bit = bits;
		1084
		1085	assert (CHAR_BIT == 8);
		1086	assert (sizeof (eio_dirent) * 8 < 256);
		1087	assert (offsetof (eio_dirent, inode)); /* we use bit #0 as sentinel */
		1088	assert (offsetof (eio_dirent, score)); /* we use bit #0 as sentinel */
		1089
		1090	if (size <= EIO_SORT_FAST)
		1091	return;
		1092
		1093	/* first prepare an array of bits to test in our radix sort */
		1094	/* try to take endianness into account, as well as differences in ino_t sizes */
		1095	/* inode_bits must contain all inodes ORed together */
		1096	/* which is used to skip bits that are 0 everywhere, which is very common */
		1097	{
		1098	ino_t endianness;
		1099	int i, j;
		1100
		1101	/* we store the byte offset of byte n into byte n of "endianness" */
		1102	for (i = 0; i < sizeof (ino_t); ++i)
		1103	((unsigned char *)&endianness)[i] = i;
		1104
		1105	*bit++ = 0;
		1106
		1107	for (i = 0; i < sizeof (ino_t); ++i)
		1108	{
		1109	/* shifting off the byte offsets out of "endianness" */
		1110	int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
		1111	endianness >>= 8;
		1112
		1113	for (j = 0; j < 8; ++j)
		1114	if (inode_bits & (((ino_t)1) << (i * 8 + j)))
		1115	*bit++ = offs + j;
		1116	}
		1117
		1118	for (j = 0; j < 8; ++j)
		1119	if (score_bits & (1 << j))
		1120	*bit++ = offsetof (eio_dirent, score) * 8 + j;
		1121	}
		1122
		1123	/* now actually do the sorting (a variant of MSD radix sort) */
		1124	{
		1125	eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
		1126	eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
		1127	unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
		1128	int stk_idx = 0;
		1129
		1130	base_stk [stk_idx] = dents;
		1131	end_stk [stk_idx] = dents + size;
		1132	bit_stk [stk_idx] = bit - 1;
		1133
		1134	do
		1135	{
		1136	base = base_stk [stk_idx];
		1137	end = end_stk [stk_idx];
		1138	bit = bit_stk [stk_idx];
		1139
		1140	for (;;)
		1141	{
		1142	unsigned char O = *bit >> 3;
		1143	unsigned char M = 1 << (*bit & 7);
		1144
		1145	eio_dirent *a = base;
		1146	eio_dirent *b = end;
		1147
		1148	if (b - a < EIO_SORT_CUTOFF)
		1149	break;
		1150
		1151	/* now bit-partition the array on the bit */
		1152	/* this ugly asymmetric loop seems to perform much better than typical */
		1153	/* partition algos found in the literature */
		1154	do
		1155	if (!(((unsigned char *)a)[O] & M))
		1156	++a;
		1157	else if (!(((unsigned char *)--b)[O] & M))
		1158	{
		1159	eio_dirent tmp = *a; *a = *b; *b = tmp;
		1160	++a;
		1161	}
		1162	while (b > a);
		1163
		1164	/* next bit, or stop, if no bits left in this path */
		1165	if (!*--bit)
		1166	break;
		1167
		1168	base_stk [stk_idx] = a;
		1169	end_stk [stk_idx] = end;
		1170	bit_stk [stk_idx] = bit;
		1171	++stk_idx;
		1172
		1173	end = a;
		1174	}
		1175	}
		1176	while (stk_idx--);
		1177	}
		1178	}
		1179
		1180	static void
		1181	eio_dent_insertion_sort (eio_dirent *dents, int size)
		1182	{
		1183	/* first move the smallest element to the front, to act as a sentinel */
		1184	{
		1185	int i;
		1186	eio_dirent *min = dents;
		1187
		1188	/* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
		1189	for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
		1190	if (EIO_DENT_CMP (dents [i], <, *min))
		1191	min = &dents [i];
		1192
		1193	/* swap elements 0 and j (minimum) */
		1194	{
		1195	eio_dirent tmp = *dents; *dents = *min; *min = tmp;
		1196	}
		1197	}
		1198
		1199	/* then do standard insertion sort, assuming that all elements are >= dents [0] */
		1200	{
		1201	eio_dirent *i, *j;
		1202
		1203	for (i = dents + 1; i < dents + size; ++i)
		1204	{
		1205	eio_dirent value = *i;
		1206
		1207	for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
		1208	j [1] = j [0];
		1209
		1210	j [1] = value;
		1211	}
		1212	}
		1213	}
		1214
		1215	static void
		1216	eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
		1217	{
		1218	if (size <= 1)
		1219	return; /* our insertion sort relies on size > 0 */
		1220
		1221	/* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
		1222	/* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
		1223	eio_dent_radix_sort (dents, size, score_bits, inode_bits);
		1224
		1225	/* use an insertion sort at the end, or for small arrays, */
		1226	/* as insertion sort is more efficient for small partitions */
		1227	eio_dent_insertion_sort (dents, size);
		1228	}
		1229
718	/* read a full directory */	1230	/* read a full directory */
719	static void	1231	static void
720	eio__scandir (eio_req *req, worker *self)	1232	eio__scandir (eio_req *req, etp_worker *self)
721	{	1233	{
722	DIR *dirp;	1234	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;	1235	EIO_STRUCT_DIRENT *entp;
729	char *name, *names;	1236	char *name, *names;
730	int memlen = 4096;	1237	int namesalloc = 4096;
731	int memofs = 0;	1238	int namesoffs = 0;
		1239	int flags = req->int1;
		1240	eio_dirent *dents = 0;
		1241	int dentalloc = 128;
732	int res = 0;	1242	int dentoffs = 0;
		1243	ino_t inode_bits = 0;
		1244
		1245	req->result = -1;
		1246
		1247	if (!(flags & EIO_READDIR_DENTS))
		1248	flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
733		1249
734	X_LOCK (wrklock);	1250	X_LOCK (wrklock);
		1251	/* the corresponding closedir is in ETP_WORKER_CLEAR */
735	self->dirp = dirp = opendir (req->ptr1);	1252	self->dirp = dirp = opendir (req->ptr1);
736	self->dbuf = u = malloc (sizeof (*u));	1253
737	req->flags |= EIO_FLAG_PTR2_FREE;	1254	req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
		1255	req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
738	req->ptr2 = names = malloc (memlen);	1256	req->ptr2 = names = malloc (namesalloc);
739	X_UNLOCK (wrklock);	1257	X_UNLOCK (wrklock);
740		1258
741	if (dirp && u && names)	1259	if (dirp && names && (!flags || dents))
742	for (;;)	1260	for (;;)
743	{	1261	{
744	errno = 0;	1262	errno = 0;
745	readdir_r (dirp, &u->d, &entp);	1263	entp = readdir (dirp);
746		1264
747	if (!entp)	1265	if (!entp)
		1266	{
		1267	if (errno)
		1268	break;
		1269
		1270	/* sort etc. */
		1271	req->int1 = flags;
		1272	req->result = dentoffs;
		1273
		1274	if (flags & EIO_READDIR_STAT_ORDER)
		1275	eio_dent_sort (dents, dentoffs, flags & EIO_READDIR_DIRS_FIRST ? 7 : 0, inode_bits);
		1276	else if (flags & EIO_READDIR_DIRS_FIRST)
		1277	if (flags & EIO_READDIR_FOUND_UNKNOWN)
		1278	eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
		1279	else
		1280	{
		1281	/* in this case, all is known, and we just put dirs first and sort them */
		1282	eio_dirent *oth = dents + dentoffs;
		1283	eio_dirent *dir = dents;
		1284
		1285	/* now partition dirs to the front, and non-dirs to the back */
		1286	/* by walking from both sides and swapping if necessary */
		1287	while (oth > dir)
		1288	{
		1289	if (dir->type == EIO_DT_DIR)
		1290	++dir;
		1291	else if ((--oth)->type == EIO_DT_DIR)
		1292	{
		1293	eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
		1294
		1295	++dir;
		1296	}
		1297	}
		1298
		1299	/* now sort the dirs only (dirs all have the same score) */
		1300	eio_dent_sort (dents, dir - dents, 0, inode_bits);
		1301	}
		1302
748	break;	1303	break;
		1304	}
749		1305
		1306	/* now add the entry to our list(s) */
750	name = entp->d_name;	1307	name = entp->d_name;
751		1308
		1309	/* skip . and .. entries */
752	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))	1310	if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
753	{	1311	{
754	int len = strlen (name) + 1;	1312	int len = D_NAMLEN (entp) + 1;
755		1313
756	res++;	1314	while (expect_false (namesoffs + len > namesalloc))
757
758	while (memofs + len > memlen)
759	{	1315	{
760	memlen *= 2;	1316	namesalloc *= 2;
761	X_LOCK (wrklock);	1317	X_LOCK (wrklock);
762	req->ptr2 = names = realloc (names, memlen);	1318	req->ptr2 = names = realloc (names, namesalloc);
763	X_UNLOCK (wrklock);	1319	X_UNLOCK (wrklock);
764		1320
765	if (!names)	1321	if (!names)
766	break;	1322	break;
767	}	1323	}
768		1324
769	memcpy (names + memofs, name, len);	1325	memcpy (names + namesoffs, name, len);
		1326
		1327	if (dents)
		1328	{
		1329	struct eio_dirent *ent;
		1330
		1331	if (expect_false (dentoffs == dentalloc))
		1332	{
		1333	dentalloc *= 2;
		1334	X_LOCK (wrklock);
		1335	req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
		1336	X_UNLOCK (wrklock);
		1337
		1338	if (!dents)
		1339	break;
		1340	}
		1341
		1342	ent = dents + dentoffs;
		1343
		1344	ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
		1345	ent->namelen = len - 1;
		1346	ent->inode = D_INO (entp);
		1347
		1348	inode_bits |= ent->inode;
		1349
		1350	switch (D_TYPE (entp))
		1351	{
		1352	default:
		1353	ent->type = EIO_DT_UNKNOWN;
		1354	flags |= EIO_READDIR_FOUND_UNKNOWN;
		1355	break;
		1356
		1357	#ifdef DT_FIFO
		1358	case DT_FIFO: ent->type = EIO_DT_FIFO; break;
		1359	#endif
		1360	#ifdef DT_CHR
		1361	case DT_CHR: ent->type = EIO_DT_CHR; break;
		1362	#endif
		1363	#ifdef DT_MPC
		1364	case DT_MPC: ent->type = EIO_DT_MPC; break;
		1365	#endif
		1366	#ifdef DT_DIR
		1367	case DT_DIR: ent->type = EIO_DT_DIR; break;
		1368	#endif
		1369	#ifdef DT_NAM
		1370	case DT_NAM: ent->type = EIO_DT_NAM; break;
		1371	#endif
		1372	#ifdef DT_BLK
		1373	case DT_BLK: ent->type = EIO_DT_BLK; break;
		1374	#endif
		1375	#ifdef DT_MPB
		1376	case DT_MPB: ent->type = EIO_DT_MPB; break;
		1377	#endif
		1378	#ifdef DT_REG
		1379	case DT_REG: ent->type = EIO_DT_REG; break;
		1380	#endif
		1381	#ifdef DT_NWK
		1382	case DT_NWK: ent->type = EIO_DT_NWK; break;
		1383	#endif
		1384	#ifdef DT_CMP
		1385	case DT_CMP: ent->type = EIO_DT_CMP; break;
		1386	#endif
		1387	#ifdef DT_LNK
		1388	case DT_LNK: ent->type = EIO_DT_LNK; break;
		1389	#endif
		1390	#ifdef DT_SOCK
		1391	case DT_SOCK: ent->type = EIO_DT_SOCK; break;
		1392	#endif
		1393	#ifdef DT_DOOR
		1394	case DT_DOOR: ent->type = EIO_DT_DOOR; break;
		1395	#endif
		1396	#ifdef DT_WHT
		1397	case DT_WHT: ent->type = EIO_DT_WHT; break;
		1398	#endif
		1399	}
		1400
		1401	ent->score = 7;
		1402
		1403	if (flags & EIO_READDIR_DIRS_FIRST)
		1404	{
		1405	if (ent->type == EIO_DT_UNKNOWN)
		1406	{
		1407	if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
		1408	ent->score = 1;
		1409	else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
		1410	ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
		1411	}
		1412	else if (ent->type == EIO_DT_DIR)
		1413	ent->score = 0;
		1414	}
		1415	}
		1416
770	memofs += len;	1417	namesoffs += len;
		1418	++dentoffs;
		1419	}
		1420
		1421	if (EIO_CANCELLED (req))
		1422	{
		1423	errno = ECANCELED;
		1424	break;
771	}	1425	}
772	}	1426	}
		1427	}
773		1428
774	if (errno)	1429	#ifdef PAGESIZE
775	res = -1;	1430	# define eio_pagesize() PAGESIZE
		1431	#else
		1432	static intptr_t
		1433	eio_pagesize (void)
		1434	{
		1435	static intptr_t page;
		1436
		1437	if (!page)
		1438	page = sysconf (_SC_PAGESIZE);
		1439
		1440	return page;
		1441	}
		1442	#endif
		1443
		1444	static void
		1445	eio_page_align (void **addr, size_t *length)
		1446	{
		1447	intptr_t mask = eio_pagesize () - 1;
		1448
		1449	/* round down addr */
		1450	intptr_t adj = mask & (intptr_t)*addr;
		1451
		1452	*addr = (void *)((intptr_t)*addr - adj);
		1453	*length += adj;
		1454
		1455	/* round up length */
		1456	*length = (*length + mask) & ~mask;
		1457	}
		1458
		1459	#if !_POSIX_MEMLOCK
		1460	# define eio__mlockall(a) ((errno = ENOSYS), -1)
		1461	#else
		1462
		1463	static int
		1464	eio__mlockall (int flags)
		1465	{
		1466	#if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
		1467	extern int mallopt (int, int);
		1468	mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
		1469	#endif
		1470
		1471	if (EIO_MCL_CURRENT != MCL_CURRENT
		1472	|| EIO_MCL_FUTURE != MCL_FUTURE)
		1473	{
		1474	flags = 0
		1475	| (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
		1476	| (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
		1477	}
		1478
		1479	return mlockall (flags);
		1480	}
		1481	#endif
		1482
		1483	#if !_POSIX_MEMLOCK_RANGE
		1484	# define eio__mlock(a,b) ((errno = ENOSYS), -1)
		1485	#else
		1486
		1487	static int
		1488	eio__mlock (void *addr, size_t length)
		1489	{
		1490	eio_page_align (&addr, &length);
		1491
		1492	return mlock (addr, length);
		1493	}
		1494
		1495	#endif
		1496
		1497	#if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
		1498	# define eio__msync(a,b,c) ((errno = ENOSYS), -1)
		1499	#else
		1500
		1501	int
		1502	eio__msync (void *mem, size_t len, int flags)
		1503	{
		1504	eio_page_align (&mem, &len);
		1505
		1506	if (EIO_MS_ASYNC != MS_SYNC
		1507	|| EIO_MS_INVALIDATE != MS_INVALIDATE
		1508	|| EIO_MS_SYNC != MS_SYNC)
		1509	{
		1510	flags = 0
		1511	| (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
		1512	| (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
		1513	| (flags & EIO_MS_SYNC ? MS_SYNC : 0);
		1514	}
		1515
		1516	return msync (mem, len, flags);
		1517	}
		1518
		1519	#endif
		1520
		1521	int
		1522	eio__mtouch (void *mem, size_t len, int flags)
		1523	{
		1524	eio_page_align (&mem, &len);
		1525
776		1526	{
777	req->result = res;	1527	intptr_t addr = (intptr_t)mem;
		1528	intptr_t end = addr + len;
		1529	intptr_t page = eio_pagesize ();
		1530
		1531	if (addr < end)
		1532	if (flags & EIO_MT_MODIFY) /* modify */
		1533	do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
		1534	else
		1535	do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
		1536	}
		1537
		1538	return 0;
778	}	1539	}
779		1540
780	/*****************************************************************************/	1541	/*****************************************************************************/
781		1542
782	#define ALLOC(len) \	1543	#define ALLOC(len) \
…		…
792	req->result = -1; \	1553	req->result = -1; \
793	break; \	1554	break; \
794	} \	1555	} \
795	}	1556	}
796		1557
797	X_THREAD_PROC (eio_proc)	1558	X_THREAD_PROC (etp_proc)
798	{	1559	{
799	eio_req *req;	1560	ETP_REQ *req;
800	struct timespec ts;	1561	struct timespec ts;
801	worker *self = (worker *)thr_arg;	1562	etp_worker *self = (etp_worker *)thr_arg;
802		1563
803	/* try to distribute timeouts somewhat randomly */	1564	/* try to distribute timeouts somewhat randomly */
804	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);	1565	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
805		1566
806	for (;;)	1567	for (;;)
807	{	1568	{
808	ts.tv_sec = time (0) + IDLE_TIMEOUT;
809
810	X_LOCK (reqlock);	1569	X_LOCK (reqlock);
811		1570
812	for (;;)	1571	for (;;)
813	{	1572	{
814	self->req = req = reqq_shift (&req_queue);	1573	self->req = req = reqq_shift (&req_queue);
…		…
816	if (req)	1575	if (req)
817	break;	1576	break;
818		1577
819	++idle;	1578	++idle;
820		1579
		1580	ts.tv_sec = time (0) + idle_timeout;
821	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)	1581	if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
822	{	1582	{
823	if (idle > max_idle)	1583	if (idle > max_idle)
824	{	1584	{
825	--idle;	1585	--idle;
…		…
830	goto quit;	1590	goto quit;
831	}	1591	}
832		1592
833	/* we are allowed to idle, so do so without any timeout */	1593	/* we are allowed to idle, so do so without any timeout */
834	X_COND_WAIT (reqwait, reqlock);	1594	X_COND_WAIT (reqwait, reqlock);
835	ts.tv_sec = time (0) + IDLE_TIMEOUT;
836	}	1595	}
837		1596
838	--idle;	1597	--idle;
839	}	1598	}
840		1599
841	--nready;	1600	--nready;
842		1601
843	X_UNLOCK (reqlock);	1602	X_UNLOCK (reqlock);
844		1603
845	errno = 0; /* strictly unnecessary */	1604	if (req->type < 0)
		1605	goto quit;
846		1606
847	if (!EIO_CANCELLED (req))	1607	if (!EIO_CANCELLED (req))
848	switch (req->type)	1608	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		1609
949	X_LOCK (reslock);	1610	X_LOCK (reslock);
950		1611
951	++npending;	1612	++npending;
952		1613
953	if (!reqq_push (&res_queue, req) && want_poll_cb)	1614	if (!reqq_push (&res_queue, req) && want_poll_cb)
954	want_poll_cb ();	1615	want_poll_cb ();
955		1616
956	self->req = 0;	1617	self->req = 0;
957	worker_clear (self);	1618	etp_worker_clear (self);
958		1619
959	X_UNLOCK (reslock);	1620	X_UNLOCK (reslock);
960	}	1621	}
961		1622
962	quit:	1623	quit:
963	X_LOCK (wrklock);	1624	X_LOCK (wrklock);
964	worker_free (self);	1625	etp_worker_free (self);
965	X_UNLOCK (wrklock);	1626	X_UNLOCK (wrklock);
966		1627
967	return 0;	1628	return 0;
968	}	1629	}
969		1630
970	/*****************************************************************************/	1631	/*****************************************************************************/
971		1632
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))	1633	int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1029	{	1634	{
1030	want_poll_cb = want_poll;	1635	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	}	1636	}
1046		1637
1047	static void eio_api_destroy (eio_req *req)	1638	static void eio_api_destroy (eio_req *req)
1048	{	1639	{
1049	free (req);	1640	free (req);
…		…
1071	{ \	1662	{ \
1072	eio_api_destroy (req); \	1663	eio_api_destroy (req); \
1073	return 0; \	1664	return 0; \
1074	}	1665	}
1075		1666
		1667	static void eio_execute (etp_worker *self, eio_req *req)
		1668	{
		1669	switch (req->type)
		1670	{
		1671	case EIO_READ: ALLOC (req->size);
		1672	req->result = req->offs >= 0
		1673	? pread (req->int1, req->ptr2, req->size, req->offs)
		1674	: read (req->int1, req->ptr2, req->size); break;
		1675	case EIO_WRITE: req->result = req->offs >= 0
		1676	? pwrite (req->int1, req->ptr2, req->size, req->offs)
		1677	: write (req->int1, req->ptr2, req->size); break;
		1678
		1679	case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
		1680	case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
		1681
		1682	case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1683	req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1684	case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1685	req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1686	case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
		1687	req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
		1688
		1689	case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1690	req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1691	case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
		1692	req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
		1693
		1694	case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
		1695	case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
		1696	case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
		1697	case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
		1698	case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
		1699	case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
		1700
		1701	case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
		1702	case EIO_CLOSE: req->result = close (req->int1); break;
		1703	case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
		1704	case EIO_UNLINK: req->result = unlink (req->ptr1); break;
		1705	case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
		1706	case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
		1707	case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
		1708	case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
		1709	case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
		1710	case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
		1711
		1712	case EIO_READLINK: ALLOC (PATH_MAX);
		1713	req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
		1714
		1715	case EIO_SYNC: req->result = 0; sync (); break;
		1716	case EIO_FSYNC: req->result = fsync (req->int1); break;
		1717	case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
		1718	case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
		1719	case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
		1720	case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
		1721	case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
		1722	case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
		1723
		1724	case EIO_READDIR: eio__scandir (req, self); break;
		1725
		1726	case EIO_BUSY:
		1727	#ifdef _WIN32
		1728	Sleep (req->nv1 * 1e3);
		1729	#else
		1730	{
		1731	struct timeval tv;
		1732
		1733	tv.tv_sec = req->nv1;
		1734	tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
		1735
		1736	req->result = select (0, 0, 0, 0, &tv);
		1737	}
		1738	#endif
		1739	break;
		1740
		1741	case EIO_UTIME:
		1742	case EIO_FUTIME:
		1743	{
		1744	struct timeval tv[2];
		1745	struct timeval *times;
		1746
		1747	if (req->nv1 != -1. || req->nv2 != -1.)
		1748	{
		1749	tv[0].tv_sec = req->nv1;
		1750	tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
		1751	tv[1].tv_sec = req->nv2;
		1752	tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
		1753
		1754	times = tv;
		1755	}
		1756	else
		1757	times = 0;
		1758
		1759	req->result = req->type == EIO_FUTIME
		1760	? futimes (req->int1, times)
		1761	: utimes (req->ptr1, times);
		1762	}
		1763	break;
		1764
		1765	case EIO_GROUP:
		1766	abort (); /* handled in eio_request */
		1767
		1768	case EIO_NOP:
		1769	req->result = 0;
		1770	break;
		1771
		1772	case EIO_CUSTOM:
		1773	((void (*)(eio_req *))req->feed) (req);
		1774	break;
		1775
		1776	default:
		1777	errno = ENOSYS;
		1778	req->result = -1;
		1779	break;
		1780	}
		1781
		1782	req->errorno = errno;
		1783	}
		1784
1076	#ifndef EIO_NO_WRAPPERS	1785	#ifndef EIO_NO_WRAPPERS
1077		1786
1078	eio_req *eio_nop (int pri, eio_cb cb, void *data)	1787	eio_req *eio_nop (int pri, eio_cb cb, void *data)
1079	{	1788	{
1080	REQ (EIO_NOP); SEND;	1789	REQ (EIO_NOP); SEND;
…		…
1093	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)	1802	eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1094	{	1803	{
1095	REQ (EIO_FSYNC); req->int1 = fd; SEND;	1804	REQ (EIO_FSYNC); req->int1 = fd; SEND;
1096	}	1805	}
1097		1806
		1807	eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1808	{
		1809	REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1810	}
		1811
		1812	eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
		1813	{
		1814	REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
		1815	}
		1816
		1817	eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
		1818	{
		1819	REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
		1820	}
		1821
		1822	eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
		1823	{
		1824	REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
		1825	}
		1826
		1827	eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
		1828	{
		1829	REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
		1830	}
		1831
1098	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)	1832	eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1099	{	1833	{
1100	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;	1834	REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1101	}	1835	}
1102		1836
…		…
1121	}	1855	}
1122		1856
1123	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)	1857	eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1124	{	1858	{
1125	REQ (EIO_FSTAT); req->int1 = fd; SEND;	1859	REQ (EIO_FSTAT); req->int1 = fd; SEND;
		1860	}
		1861
		1862	eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
		1863	{
		1864	REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1126	}	1865	}
1127		1866
1128	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)	1867	eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1129	{	1868	{
1130	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;	1869	REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
…		…
1204	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)	1943	eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1205	{	1944	{
1206	return eio__1path (EIO_LSTAT, path, pri, cb, data);	1945	return eio__1path (EIO_LSTAT, path, pri, cb, data);
1207	}	1946	}
1208		1947
		1948	eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
		1949	{
		1950	return eio__1path (EIO_STATVFS, path, pri, cb, data);
		1951	}
		1952
1209	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)	1953	eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1210	{	1954	{
1211	return eio__1path (EIO_UNLINK, path, pri, cb, data);	1955	return eio__1path (EIO_UNLINK, path, pri, cb, data);
1212	}	1956	}
1213		1957
1214	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)	1958	eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1215	{	1959	{
1216	return eio__1path (EIO_RMDIR, path, pri, cb, data);	1960	return eio__1path (EIO_RMDIR, path, pri, cb, data);
1217	}	1961	}
1218		1962
1219	eio_req *eio_readdir (const char *path, int pri, eio_cb cb, void *data)	1963	eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1220	{	1964	{
1221	return eio__1path (EIO_READDIR, path, pri, cb, data);	1965	REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1222	}	1966	}
1223		1967
1224	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)	1968	eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1225	{	1969	{
1226	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int2 = (long)dev; SEND;	1970	REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND;
1227	}	1971	}
1228		1972
1229	static eio_req *	1973	static eio_req *
1230	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1974	eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1231	{	1975	{
…		…
1255	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)	1999	eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1256	{	2000	{
1257	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);	2001	return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1258	}	2002	}
1259		2003
		2004	eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
		2005	{
		2006	REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
		2007	}
		2008
1260	#endif	2009	#endif
1261		2010
1262	eio_req *eio_grp (eio_cb cb, void *data)	2011	eio_req *eio_grp (eio_cb cb, void *data)
1263	{	2012	{
1264	const int pri = EIO_PRI_MAX;	2013	const int pri = EIO_PRI_MAX;
…		…
1290		2039
1291	void eio_grp_add (eio_req *grp, eio_req *req)	2040	void eio_grp_add (eio_req *grp, eio_req *req)
1292	{	2041	{
1293	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));	2042	assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1294		2043
		2044	grp->flags |= EIO_FLAG_GROUPADD;
		2045
1295	++grp->size;	2046	++grp->size;
1296	req->grp = grp;	2047	req->grp = grp;
1297		2048
1298	req->grp_prev = 0;	2049	req->grp_prev = 0;
1299	req->grp_next = grp->grp_first;	2050	req->grp_next = grp->grp_first;
…		…
1307	/*****************************************************************************/	2058	/*****************************************************************************/
1308	/* misc garbage */	2059	/* misc garbage */
1309		2060
1310	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)	2061	ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1311	{	2062	{
1312	worker wrk;	2063	etp_worker wrk;
		2064	ssize_t ret;
1313		2065
1314	wrk.dbuf = 0;	2066	wrk.dbuf = 0;
1315		2067
1316	eio__sendfile (ofd, ifd, offset, count, &wrk);	2068	ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1317		2069
1318	if (wrk.dbuf)	2070	if (wrk.dbuf)
1319	free (wrk.dbuf);	2071	free (wrk.dbuf);
1320	}
1321		2072
		2073	return ret;
		2074	}
		2075

Diff Legend

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