ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libeio/eio.c
Revision: 1.76
Committed: Tue Jul 5 09:24:12 2011 UTC (12 years, 11 months ago) by root
Content type: text/plain
Branch: MAIN
Changes since 1.75: +14 -0 lines
Log Message:
realpath

File Contents

# Content
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
44 #include "eio.h"
45 #include "ecb.h"
46
47 #ifdef EIO_STACKSIZE
48 # define XTHREAD_STACKSIZE EIO_STACKSIZE
49 #endif
50 #include "xthread.h"
51
52 #include <errno.h>
53 #include <stddef.h>
54 #include <stdlib.h>
55 #include <string.h>
56 #include <errno.h>
57 #include <sys/types.h>
58 #include <sys/stat.h>
59 #include <sys/statvfs.h>
60 #include <limits.h>
61 #include <fcntl.h>
62 #include <assert.h>
63
64 /* intptr_t comes from unistd.h, says POSIX/UNIX/tradition */
65 /* intptr_t only comes form stdint.h, says idiot openbsd coder */
66 #if HAVE_STDINT_H
67 # include <stdint.h>
68 #endif
69
70 #ifndef EIO_FINISH
71 # define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0
72 #endif
73
74 #ifndef EIO_DESTROY
75 # define EIO_DESTROY(req) do { if ((req)->destroy) (req)->destroy (req); } while (0)
76 #endif
77
78 #ifndef EIO_FEED
79 # define EIO_FEED(req) do { if ((req)->feed ) (req)->feed (req); } while (0)
80 #endif
81
82 #ifdef _WIN32
83
84 /*doh*/
85 #else
86
87 # include <sys/time.h>
88 # include <sys/select.h>
89 # include <unistd.h>
90 # include <utime.h>
91 # include <signal.h>
92 # include <dirent.h>
93
94 #if _POSIX_MEMLOCK || _POSIX_MEMLOCK_RANGE || _POSIX_MAPPED_FILES
95 # include <sys/mman.h>
96 #endif
97
98 /* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
99 # if __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__
100 # define _DIRENT_HAVE_D_TYPE /* sigh */
101 # define D_INO(de) (de)->d_fileno
102 # define D_NAMLEN(de) (de)->d_namlen
103 # elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
104 # define D_INO(de) (de)->d_ino
105 # endif
106
107 #ifdef _D_EXACT_NAMLEN
108 # undef D_NAMLEN
109 # define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
110 #endif
111
112 # ifdef _DIRENT_HAVE_D_TYPE
113 # define D_TYPE(de) (de)->d_type
114 # endif
115
116 # ifndef EIO_STRUCT_DIRENT
117 # define EIO_STRUCT_DIRENT struct dirent
118 # endif
119
120 #endif
121
122 #if HAVE_SENDFILE
123 # if __linux
124 # include <sys/sendfile.h>
125 # elif __FreeBSD__ || defined __APPLE__
126 # include <sys/socket.h>
127 # include <sys/uio.h>
128 # elif __hpux
129 # include <sys/socket.h>
130 # elif __solaris
131 # include <sys/sendfile.h>
132 # else
133 # error sendfile support requested but not available
134 # endif
135 #endif
136
137 #ifndef D_TYPE
138 # define D_TYPE(de) 0
139 #endif
140 #ifndef D_INO
141 # define D_INO(de) 0
142 #endif
143 #ifndef D_NAMLEN
144 # define D_NAMLEN(de) strlen ((de)->d_name)
145 #endif
146
147 /* used for struct dirent, AIX doesn't provide it */
148 #ifndef NAME_MAX
149 # define NAME_MAX 4096
150 #endif
151
152 /* used for readlink etc. */
153 #ifndef PATH_MAX
154 # define PATH_MAX 4096
155 #endif
156
157 /* buffer size for various temporary buffers */
158 #define EIO_BUFSIZE 65536
159
160 #define dBUF \
161 char *eio_buf; \
162 ETP_WORKER_LOCK (self); \
163 self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
164 ETP_WORKER_UNLOCK (self); \
165 errno = ENOMEM; \
166 if (!eio_buf) \
167 return -1;
168
169 #define EIO_TICKS ((1000000 + 1023) >> 10)
170
171 #define ETP_PRI_MIN EIO_PRI_MIN
172 #define ETP_PRI_MAX EIO_PRI_MAX
173
174 struct etp_worker;
175
176 #define ETP_REQ eio_req
177 #define ETP_DESTROY(req) eio_destroy (req)
178 static int eio_finish (eio_req *req);
179 #define ETP_FINISH(req) eio_finish (req)
180 static void eio_execute (struct etp_worker *self, eio_req *req);
181 #define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
182
183 #define ETP_WORKER_CLEAR(req) \
184 if (wrk->dbuf) \
185 { \
186 free (wrk->dbuf); \
187 wrk->dbuf = 0; \
188 } \
189 \
190 if (wrk->dirp) \
191 { \
192 closedir (wrk->dirp); \
193 wrk->dirp = 0; \
194 }
195
196 #define ETP_WORKER_COMMON \
197 void *dbuf; \
198 DIR *dirp;
199
200 /*****************************************************************************/
201
202 #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
203
204 /* calculate time difference in ~1/EIO_TICKS of a second */
205 ecb_inline int
206 tvdiff (struct timeval *tv1, struct timeval *tv2)
207 {
208 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
209 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
210 }
211
212 static unsigned int started, idle, wanted = 4;
213
214 static void (*want_poll_cb) (void);
215 static void (*done_poll_cb) (void);
216
217 static unsigned int max_poll_time; /* reslock */
218 static unsigned int max_poll_reqs; /* reslock */
219
220 static volatile unsigned int nreqs; /* reqlock */
221 static volatile unsigned int nready; /* reqlock */
222 static volatile unsigned int npending; /* reqlock */
223 static volatile unsigned int max_idle = 4; /* maximum number of threads that can idle indefinitely */
224 static volatile unsigned int idle_timeout = 10; /* number of seconds after which an idle threads exit */
225
226 static xmutex_t wrklock;
227 static xmutex_t reslock;
228 static xmutex_t reqlock;
229 static xcond_t reqwait;
230
231 #if !HAVE_PREADWRITE
232 /*
233 * make our pread/pwrite emulation safe against themselves, but not against
234 * normal read/write by using a mutex. slows down execution a lot,
235 * but that's your problem, not mine.
236 */
237 static xmutex_t preadwritelock = X_MUTEX_INIT;
238 #endif
239
240 typedef struct etp_worker
241 {
242 /* locked by wrklock */
243 struct etp_worker *prev, *next;
244
245 xthread_t tid;
246
247 /* locked by reslock, reqlock or wrklock */
248 ETP_REQ *req; /* currently processed request */
249
250 ETP_WORKER_COMMON
251 } etp_worker;
252
253 static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
254
255 #define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
256 #define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
257
258 /* worker threads management */
259
260 static void ecb_cold
261 etp_worker_clear (etp_worker *wrk)
262 {
263 ETP_WORKER_CLEAR (wrk);
264 }
265
266 static void ecb_cold
267 etp_worker_free (etp_worker *wrk)
268 {
269 wrk->next->prev = wrk->prev;
270 wrk->prev->next = wrk->next;
271
272 free (wrk);
273 }
274
275 static unsigned int
276 etp_nreqs (void)
277 {
278 int retval;
279 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
280 retval = nreqs;
281 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
282 return retval;
283 }
284
285 static unsigned int
286 etp_nready (void)
287 {
288 unsigned int retval;
289
290 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
291 retval = nready;
292 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
293
294 return retval;
295 }
296
297 static unsigned int
298 etp_npending (void)
299 {
300 unsigned int retval;
301
302 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
303 retval = npending;
304 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
305
306 return retval;
307 }
308
309 static unsigned int
310 etp_nthreads (void)
311 {
312 unsigned int retval;
313
314 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
315 retval = started;
316 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
317
318 return retval;
319 }
320
321 /*
322 * a somewhat faster data structure might be nice, but
323 * with 8 priorities this actually needs <20 insns
324 * per shift, the most expensive operation.
325 */
326 typedef struct {
327 ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
328 int size;
329 } etp_reqq;
330
331 static etp_reqq req_queue;
332 static etp_reqq res_queue;
333
334 static int ecb_noinline
335 reqq_push (etp_reqq *q, ETP_REQ *req)
336 {
337 int pri = req->pri;
338 req->next = 0;
339
340 if (q->qe[pri])
341 {
342 q->qe[pri]->next = req;
343 q->qe[pri] = req;
344 }
345 else
346 q->qe[pri] = q->qs[pri] = req;
347
348 return q->size++;
349 }
350
351 static ETP_REQ * ecb_noinline
352 reqq_shift (etp_reqq *q)
353 {
354 int pri;
355
356 if (!q->size)
357 return 0;
358
359 --q->size;
360
361 for (pri = ETP_NUM_PRI; pri--; )
362 {
363 eio_req *req = q->qs[pri];
364
365 if (req)
366 {
367 if (!(q->qs[pri] = (eio_req *)req->next))
368 q->qe[pri] = 0;
369
370 return req;
371 }
372 }
373
374 abort ();
375 }
376
377 static void ecb_cold
378 etp_thread_init (void)
379 {
380 X_MUTEX_CREATE (wrklock);
381 X_MUTEX_CREATE (reslock);
382 X_MUTEX_CREATE (reqlock);
383 X_COND_CREATE (reqwait);
384 }
385
386 static void ecb_cold
387 etp_atfork_prepare (void)
388 {
389 X_LOCK (wrklock);
390 X_LOCK (reqlock);
391 X_LOCK (reslock);
392 #if !HAVE_PREADWRITE
393 X_LOCK (preadwritelock);
394 #endif
395 }
396
397 static void ecb_cold
398 etp_atfork_parent (void)
399 {
400 #if !HAVE_PREADWRITE
401 X_UNLOCK (preadwritelock);
402 #endif
403 X_UNLOCK (reslock);
404 X_UNLOCK (reqlock);
405 X_UNLOCK (wrklock);
406 }
407
408 static void ecb_cold
409 etp_atfork_child (void)
410 {
411 ETP_REQ *prv;
412
413 while ((prv = reqq_shift (&req_queue)))
414 ETP_DESTROY (prv);
415
416 while ((prv = reqq_shift (&res_queue)))
417 ETP_DESTROY (prv);
418
419 while (wrk_first.next != &wrk_first)
420 {
421 etp_worker *wrk = wrk_first.next;
422
423 if (wrk->req)
424 ETP_DESTROY (wrk->req);
425
426 etp_worker_clear (wrk);
427 etp_worker_free (wrk);
428 }
429
430 started = 0;
431 idle = 0;
432 nreqs = 0;
433 nready = 0;
434 npending = 0;
435
436 etp_thread_init ();
437 }
438
439 static void ecb_cold
440 etp_once_init (void)
441 {
442 etp_thread_init ();
443 X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
444 }
445
446 static int ecb_cold
447 etp_init (void (*want_poll)(void), void (*done_poll)(void))
448 {
449 static pthread_once_t doinit = PTHREAD_ONCE_INIT;
450
451 pthread_once (&doinit, etp_once_init);
452
453 want_poll_cb = want_poll;
454 done_poll_cb = done_poll;
455
456 return 0;
457 }
458
459 X_THREAD_PROC (etp_proc);
460
461 static void ecb_cold
462 etp_start_thread (void)
463 {
464 etp_worker *wrk = calloc (1, sizeof (etp_worker));
465
466 /*TODO*/
467 assert (("unable to allocate worker thread data", wrk));
468
469 X_LOCK (wrklock);
470
471 if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
472 {
473 wrk->prev = &wrk_first;
474 wrk->next = wrk_first.next;
475 wrk_first.next->prev = wrk;
476 wrk_first.next = wrk;
477 ++started;
478 }
479 else
480 free (wrk);
481
482 X_UNLOCK (wrklock);
483 }
484
485 static void
486 etp_maybe_start_thread (void)
487 {
488 if (ecb_expect_true (etp_nthreads () >= wanted))
489 return;
490
491 /* todo: maybe use idle here, but might be less exact */
492 if (ecb_expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
493 return;
494
495 etp_start_thread ();
496 }
497
498 static void ecb_cold
499 etp_end_thread (void)
500 {
501 eio_req *req = calloc (1, sizeof (eio_req));
502
503 req->type = -1;
504 req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
505
506 X_LOCK (reqlock);
507 reqq_push (&req_queue, req);
508 X_COND_SIGNAL (reqwait);
509 X_UNLOCK (reqlock);
510
511 X_LOCK (wrklock);
512 --started;
513 X_UNLOCK (wrklock);
514 }
515
516 static int
517 etp_poll (void)
518 {
519 unsigned int maxreqs;
520 unsigned int maxtime;
521 struct timeval tv_start, tv_now;
522
523 X_LOCK (reslock);
524 maxreqs = max_poll_reqs;
525 maxtime = max_poll_time;
526 X_UNLOCK (reslock);
527
528 if (maxtime)
529 gettimeofday (&tv_start, 0);
530
531 for (;;)
532 {
533 ETP_REQ *req;
534
535 etp_maybe_start_thread ();
536
537 X_LOCK (reslock);
538 req = reqq_shift (&res_queue);
539
540 if (req)
541 {
542 --npending;
543
544 if (!res_queue.size && done_poll_cb)
545 done_poll_cb ();
546 }
547
548 X_UNLOCK (reslock);
549
550 if (!req)
551 return 0;
552
553 X_LOCK (reqlock);
554 --nreqs;
555 X_UNLOCK (reqlock);
556
557 if (ecb_expect_false (req->type == EIO_GROUP && req->size))
558 {
559 req->int1 = 1; /* mark request as delayed */
560 continue;
561 }
562 else
563 {
564 int res = ETP_FINISH (req);
565 if (ecb_expect_false (res))
566 return res;
567 }
568
569 if (ecb_expect_false (maxreqs && !--maxreqs))
570 break;
571
572 if (maxtime)
573 {
574 gettimeofday (&tv_now, 0);
575
576 if (tvdiff (&tv_start, &tv_now) >= maxtime)
577 break;
578 }
579 }
580
581 errno = EAGAIN;
582 return -1;
583 }
584
585 static void
586 etp_cancel (ETP_REQ *req)
587 {
588 X_LOCK (wrklock);
589 req->flags |= EIO_FLAG_CANCELLED;
590 X_UNLOCK (wrklock);
591
592 eio_grp_cancel (req);
593 }
594
595 static void
596 etp_submit (ETP_REQ *req)
597 {
598 req->pri -= ETP_PRI_MIN;
599
600 if (ecb_expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
601 if (ecb_expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
602
603 if (ecb_expect_false (req->type == EIO_GROUP))
604 {
605 /* I hope this is worth it :/ */
606 X_LOCK (reqlock);
607 ++nreqs;
608 X_UNLOCK (reqlock);
609
610 X_LOCK (reslock);
611
612 ++npending;
613
614 if (!reqq_push (&res_queue, req) && want_poll_cb)
615 want_poll_cb ();
616
617 X_UNLOCK (reslock);
618 }
619 else
620 {
621 X_LOCK (reqlock);
622 ++nreqs;
623 ++nready;
624 reqq_push (&req_queue, req);
625 X_COND_SIGNAL (reqwait);
626 X_UNLOCK (reqlock);
627
628 etp_maybe_start_thread ();
629 }
630 }
631
632 static void ecb_cold
633 etp_set_max_poll_time (double nseconds)
634 {
635 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
636 max_poll_time = nseconds * EIO_TICKS;
637 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
638 }
639
640 static void ecb_cold
641 etp_set_max_poll_reqs (unsigned int maxreqs)
642 {
643 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
644 max_poll_reqs = maxreqs;
645 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
646 }
647
648 static void ecb_cold
649 etp_set_max_idle (unsigned int nthreads)
650 {
651 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
652 max_idle = nthreads;
653 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
654 }
655
656 static void ecb_cold
657 etp_set_idle_timeout (unsigned int seconds)
658 {
659 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
660 idle_timeout = seconds;
661 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
662 }
663
664 static void ecb_cold
665 etp_set_min_parallel (unsigned int nthreads)
666 {
667 if (wanted < nthreads)
668 wanted = nthreads;
669 }
670
671 static void ecb_cold
672 etp_set_max_parallel (unsigned int nthreads)
673 {
674 if (wanted > nthreads)
675 wanted = nthreads;
676
677 while (started > wanted)
678 etp_end_thread ();
679 }
680
681 /*****************************************************************************/
682
683 static void
684 grp_try_feed (eio_req *grp)
685 {
686 while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
687 {
688 grp->flags &= ~EIO_FLAG_GROUPADD;
689
690 EIO_FEED (grp);
691
692 /* stop if no progress has been made */
693 if (!(grp->flags & EIO_FLAG_GROUPADD))
694 {
695 grp->feed = 0;
696 break;
697 }
698 }
699 }
700
701 static int
702 grp_dec (eio_req *grp)
703 {
704 --grp->size;
705
706 /* call feeder, if applicable */
707 grp_try_feed (grp);
708
709 /* finish, if done */
710 if (!grp->size && grp->int1)
711 return eio_finish (grp);
712 else
713 return 0;
714 }
715
716 void
717 eio_destroy (eio_req *req)
718 {
719 if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
720 if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
721
722 EIO_DESTROY (req);
723 }
724
725 static int
726 eio_finish (eio_req *req)
727 {
728 int res = EIO_FINISH (req);
729
730 if (req->grp)
731 {
732 int res2;
733 eio_req *grp = req->grp;
734
735 /* unlink request */
736 if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
737 if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
738
739 if (grp->grp_first == req)
740 grp->grp_first = req->grp_next;
741
742 res2 = grp_dec (grp);
743
744 if (!res && res2)
745 res = res2;
746 }
747
748 eio_destroy (req);
749
750 return res;
751 }
752
753 void
754 eio_grp_cancel (eio_req *grp)
755 {
756 for (grp = grp->grp_first; grp; grp = grp->grp_next)
757 eio_cancel (grp);
758 }
759
760 void
761 eio_cancel (eio_req *req)
762 {
763 etp_cancel (req);
764 }
765
766 void
767 eio_submit (eio_req *req)
768 {
769 etp_submit (req);
770 }
771
772 unsigned int
773 eio_nreqs (void)
774 {
775 return etp_nreqs ();
776 }
777
778 unsigned int
779 eio_nready (void)
780 {
781 return etp_nready ();
782 }
783
784 unsigned int
785 eio_npending (void)
786 {
787 return etp_npending ();
788 }
789
790 unsigned int ecb_cold
791 eio_nthreads (void)
792 {
793 return etp_nthreads ();
794 }
795
796 void ecb_cold
797 eio_set_max_poll_time (double nseconds)
798 {
799 etp_set_max_poll_time (nseconds);
800 }
801
802 void ecb_cold
803 eio_set_max_poll_reqs (unsigned int maxreqs)
804 {
805 etp_set_max_poll_reqs (maxreqs);
806 }
807
808 void ecb_cold
809 eio_set_max_idle (unsigned int nthreads)
810 {
811 etp_set_max_idle (nthreads);
812 }
813
814 void ecb_cold
815 eio_set_idle_timeout (unsigned int seconds)
816 {
817 etp_set_idle_timeout (seconds);
818 }
819
820 void ecb_cold
821 eio_set_min_parallel (unsigned int nthreads)
822 {
823 etp_set_min_parallel (nthreads);
824 }
825
826 void ecb_cold
827 eio_set_max_parallel (unsigned int nthreads)
828 {
829 etp_set_max_parallel (nthreads);
830 }
831
832 int eio_poll (void)
833 {
834 return etp_poll ();
835 }
836
837 /*****************************************************************************/
838 /* work around various missing functions */
839
840 #if _POSIX_VERSION < 200809L
841 # define realpath(path,resolved_path) (errno = ENOSYS, 0)
842 #endif
843
844 #if !HAVE_PREADWRITE
845 # undef pread
846 # undef pwrite
847 # define pread eio__pread
848 # define pwrite eio__pwrite
849
850 static ssize_t
851 eio__pread (int fd, void *buf, size_t count, off_t offset)
852 {
853 ssize_t res;
854 off_t ooffset;
855
856 X_LOCK (preadwritelock);
857 ooffset = lseek (fd, 0, SEEK_CUR);
858 lseek (fd, offset, SEEK_SET);
859 res = read (fd, buf, count);
860 lseek (fd, ooffset, SEEK_SET);
861 X_UNLOCK (preadwritelock);
862
863 return res;
864 }
865
866 static ssize_t
867 eio__pwrite (int fd, void *buf, size_t count, off_t offset)
868 {
869 ssize_t res;
870 off_t ooffset;
871
872 X_LOCK (preadwritelock);
873 ooffset = lseek (fd, 0, SEEK_CUR);
874 lseek (fd, offset, SEEK_SET);
875 res = write (fd, buf, count);
876 lseek (fd, ooffset, SEEK_SET);
877 X_UNLOCK (preadwritelock);
878
879 return res;
880 }
881 #endif
882
883 #ifndef HAVE_UTIMES
884
885 # undef utimes
886 # define utimes(path,times) eio__utimes (path, times)
887
888 static int
889 eio__utimes (const char *filename, const struct timeval times[2])
890 {
891 if (times)
892 {
893 struct utimbuf buf;
894
895 buf.actime = times[0].tv_sec;
896 buf.modtime = times[1].tv_sec;
897
898 return utime (filename, &buf);
899 }
900 else
901 return utime (filename, 0);
902 }
903
904 #endif
905
906 #ifndef HAVE_FUTIMES
907
908 # undef futimes
909 # define futimes(fd,times) eio__futimes (fd, times)
910
911 static int
912 eio__futimes (int fd, const struct timeval tv[2])
913 {
914 errno = ENOSYS;
915 return -1;
916 }
917
918 #endif
919
920 #if !HAVE_FDATASYNC
921 # undef fdatasync
922 # define fdatasync(fd) fsync (fd)
923 #endif
924
925 /* sync_file_range always needs emulation */
926 static int
927 eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
928 {
929 #if HAVE_SYNC_FILE_RANGE
930 int res;
931
932 if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
933 || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
934 || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
935 {
936 flags = 0
937 | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
938 | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
939 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
940 }
941
942 res = sync_file_range (fd, offset, nbytes, flags);
943
944 if (!res || errno != ENOSYS)
945 return res;
946 #endif
947
948 /* even though we could play tricks with the flags, it's better to always
949 * call fdatasync, as that matches the expectation of its users best */
950 return fdatasync (fd);
951 }
952
953 #if !HAVE_READAHEAD
954 # undef readahead
955 # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
956
957 static ssize_t
958 eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
959 {
960 size_t todo = count;
961 dBUF;
962
963 while (todo > 0)
964 {
965 size_t len = todo < EIO_BUFSIZE ? todo : EIO_BUFSIZE;
966
967 pread (fd, eio_buf, len, offset);
968 offset += len;
969 todo -= len;
970 }
971
972 errno = 0;
973 return count;
974 }
975
976 #endif
977
978 /* sendfile always needs emulation */
979 static ssize_t
980 eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
981 {
982 ssize_t written = 0;
983 ssize_t res;
984
985 if (!count)
986 return 0;
987
988 for (;;)
989 {
990 #if HAVE_SENDFILE
991 # if __linux
992 off_t soffset = offset;
993 res = sendfile (ofd, ifd, &soffset, count);
994
995 # elif __FreeBSD__
996 /*
997 * Of course, the freebsd sendfile is a dire hack with no thoughts
998 * wasted on making it similar to other I/O functions.
999 */
1000 off_t sbytes;
1001 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
1002
1003 #if 0 /* according to the manpage, this is correct, but broken behaviour */
1004 /* freebsd' sendfile will return 0 on success */
1005 /* freebsd 8 documents it as only setting *sbytes on EINTR and EAGAIN, but */
1006 /* not on e.g. EIO or EPIPE - sounds broken */
1007 if ((res < 0 && (errno == EAGAIN || errno == EINTR) && sbytes) || res == 0)
1008 res = sbytes;
1009 #endif
1010
1011 /* according to source inspection, this is correct, and useful behaviour */
1012 if (sbytes)
1013 res = sbytes;
1014
1015 # elif defined (__APPLE__) && 0 /* broken, as everything on os x */
1016 off_t sbytes = count;
1017 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
1018
1019 /* according to the manpage, sbytes is always valid */
1020 if (sbytes)
1021 res = sbytes;
1022
1023 # elif __hpux
1024 res = sendfile (ofd, ifd, offset, count, 0, 0);
1025
1026 # elif __solaris
1027 struct sendfilevec vec;
1028 size_t sbytes;
1029
1030 vec.sfv_fd = ifd;
1031 vec.sfv_flag = 0;
1032 vec.sfv_off = offset;
1033 vec.sfv_len = count;
1034
1035 res = sendfilev (ofd, &vec, 1, &sbytes);
1036
1037 if (res < 0 && sbytes)
1038 res = sbytes;
1039
1040 # endif
1041
1042 #elif defined (_WIN32)
1043 /* does not work, just for documentation of what would need to be done */
1044 /* actually, cannot be done like this, as TransmitFile changes the file offset, */
1045 /* libeio guarantees that the file offset does not change, and windows */
1046 /* has no way to get an independent handle to the same file description */
1047 HANDLE h = TO_SOCKET (ifd);
1048 SetFilePointer (h, offset, 0, FILE_BEGIN);
1049 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
1050
1051 #else
1052 res = -1;
1053 errno = ENOSYS;
1054 #endif
1055
1056 /* we assume sendfile can copy at least 128mb in one go */
1057 if (res <= 128 * 1024 * 1024)
1058 {
1059 if (res > 0)
1060 written += res;
1061
1062 if (written)
1063 return written;
1064
1065 break;
1066 }
1067 else
1068 {
1069 /* if we requested more, then probably the kernel was lazy */
1070 written += res;
1071 offset += res;
1072 count -= res;
1073
1074 if (!count)
1075 return written;
1076 }
1077 }
1078
1079 if (res < 0
1080 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
1081 /* BSDs */
1082 #ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
1083 || errno == ENOTSUP
1084 #endif
1085 || errno == EOPNOTSUPP /* BSDs */
1086 #if __solaris
1087 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
1088 #endif
1089 )
1090 )
1091 {
1092 /* emulate sendfile. this is a major pain in the ass */
1093 dBUF;
1094
1095 res = 0;
1096
1097 while (count)
1098 {
1099 ssize_t cnt;
1100
1101 cnt = pread (ifd, eio_buf, count > EIO_BUFSIZE ? EIO_BUFSIZE : count, offset);
1102
1103 if (cnt <= 0)
1104 {
1105 if (cnt && !res) res = -1;
1106 break;
1107 }
1108
1109 cnt = write (ofd, eio_buf, cnt);
1110
1111 if (cnt <= 0)
1112 {
1113 if (cnt && !res) res = -1;
1114 break;
1115 }
1116
1117 offset += cnt;
1118 res += cnt;
1119 count -= cnt;
1120 }
1121 }
1122
1123 return res;
1124 }
1125
1126 static signed char
1127 eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
1128 {
1129 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
1130 : a->inode < b->inode ? -1
1131 : a->inode > b->inode ? 1
1132 : 0;
1133 }
1134
1135 #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1136
1137 #define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1138 #define EIO_SORT_FAST 60 /* when to only use insertion sort */
1139
1140 static void
1141 eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1142 {
1143 unsigned char bits [9 + sizeof (ino_t) * 8];
1144 unsigned char *bit = bits;
1145
1146 assert (CHAR_BIT == 8);
1147 assert (sizeof (eio_dirent) * 8 < 256);
1148 assert (offsetof (eio_dirent, inode)); /* we use bit #0 as sentinel */
1149 assert (offsetof (eio_dirent, score)); /* we use bit #0 as sentinel */
1150
1151 if (size <= EIO_SORT_FAST)
1152 return;
1153
1154 /* first prepare an array of bits to test in our radix sort */
1155 /* try to take endianness into account, as well as differences in ino_t sizes */
1156 /* inode_bits must contain all inodes ORed together */
1157 /* which is used to skip bits that are 0 everywhere, which is very common */
1158 {
1159 ino_t endianness;
1160 int i, j;
1161
1162 /* we store the byte offset of byte n into byte n of "endianness" */
1163 for (i = 0; i < sizeof (ino_t); ++i)
1164 ((unsigned char *)&endianness)[i] = i;
1165
1166 *bit++ = 0;
1167
1168 for (i = 0; i < sizeof (ino_t); ++i)
1169 {
1170 /* shifting off the byte offsets out of "endianness" */
1171 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1172 endianness >>= 8;
1173
1174 for (j = 0; j < 8; ++j)
1175 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1176 *bit++ = offs + j;
1177 }
1178
1179 for (j = 0; j < 8; ++j)
1180 if (score_bits & (1 << j))
1181 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1182 }
1183
1184 /* now actually do the sorting (a variant of MSD radix sort) */
1185 {
1186 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1187 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1188 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1189 int stk_idx = 0;
1190
1191 base_stk [stk_idx] = dents;
1192 end_stk [stk_idx] = dents + size;
1193 bit_stk [stk_idx] = bit - 1;
1194
1195 do
1196 {
1197 base = base_stk [stk_idx];
1198 end = end_stk [stk_idx];
1199 bit = bit_stk [stk_idx];
1200
1201 for (;;)
1202 {
1203 unsigned char O = *bit >> 3;
1204 unsigned char M = 1 << (*bit & 7);
1205
1206 eio_dirent *a = base;
1207 eio_dirent *b = end;
1208
1209 if (b - a < EIO_SORT_CUTOFF)
1210 break;
1211
1212 /* now bit-partition the array on the bit */
1213 /* this ugly asymmetric loop seems to perform much better than typical */
1214 /* partition algos found in the literature */
1215 do
1216 if (!(((unsigned char *)a)[O] & M))
1217 ++a;
1218 else if (!(((unsigned char *)--b)[O] & M))
1219 {
1220 eio_dirent tmp = *a; *a = *b; *b = tmp;
1221 ++a;
1222 }
1223 while (b > a);
1224
1225 /* next bit, or stop, if no bits left in this path */
1226 if (!*--bit)
1227 break;
1228
1229 base_stk [stk_idx] = a;
1230 end_stk [stk_idx] = end;
1231 bit_stk [stk_idx] = bit;
1232 ++stk_idx;
1233
1234 end = a;
1235 }
1236 }
1237 while (stk_idx--);
1238 }
1239 }
1240
1241 static void
1242 eio_dent_insertion_sort (eio_dirent *dents, int size)
1243 {
1244 /* first move the smallest element to the front, to act as a sentinel */
1245 {
1246 int i;
1247 eio_dirent *min = dents;
1248
1249 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1250 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1251 if (EIO_DENT_CMP (dents [i], <, *min))
1252 min = &dents [i];
1253
1254 /* swap elements 0 and j (minimum) */
1255 {
1256 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1257 }
1258 }
1259
1260 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1261 {
1262 eio_dirent *i, *j;
1263
1264 for (i = dents + 1; i < dents + size; ++i)
1265 {
1266 eio_dirent value = *i;
1267
1268 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1269 j [1] = j [0];
1270
1271 j [1] = value;
1272 }
1273 }
1274 }
1275
1276 static void
1277 eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1278 {
1279 if (size <= 1)
1280 return; /* our insertion sort relies on size > 0 */
1281
1282 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1283 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1284 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1285
1286 /* use an insertion sort at the end, or for small arrays, */
1287 /* as insertion sort is more efficient for small partitions */
1288 eio_dent_insertion_sort (dents, size);
1289 }
1290
1291 /* read a full directory */
1292 static void
1293 eio__scandir (eio_req *req, etp_worker *self)
1294 {
1295 DIR *dirp;
1296 EIO_STRUCT_DIRENT *entp;
1297 char *name, *names;
1298 int namesalloc = 4096;
1299 int namesoffs = 0;
1300 int flags = req->int1;
1301 eio_dirent *dents = 0;
1302 int dentalloc = 128;
1303 int dentoffs = 0;
1304 ino_t inode_bits = 0;
1305
1306 req->result = -1;
1307
1308 if (!(flags & EIO_READDIR_DENTS))
1309 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1310
1311 X_LOCK (wrklock);
1312 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1313 self->dirp = dirp = opendir (req->ptr1);
1314
1315 if (req->flags & EIO_FLAG_PTR1_FREE)
1316 free (req->ptr1);
1317
1318 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1319 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1320 req->ptr2 = names = malloc (namesalloc);
1321 X_UNLOCK (wrklock);
1322
1323 if (dirp && names && (!flags || dents))
1324 for (;;)
1325 {
1326 errno = 0;
1327 entp = readdir (dirp);
1328
1329 if (!entp)
1330 {
1331 if (errno)
1332 break;
1333
1334 /* sort etc. */
1335 req->int1 = flags;
1336 req->result = dentoffs;
1337
1338 if (flags & EIO_READDIR_STAT_ORDER)
1339 eio_dent_sort (dents, dentoffs, flags & EIO_READDIR_DIRS_FIRST ? 7 : 0, inode_bits);
1340 else if (flags & EIO_READDIR_DIRS_FIRST)
1341 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1342 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1343 else
1344 {
1345 /* in this case, all is known, and we just put dirs first and sort them */
1346 eio_dirent *oth = dents + dentoffs;
1347 eio_dirent *dir = dents;
1348
1349 /* now partition dirs to the front, and non-dirs to the back */
1350 /* by walking from both sides and swapping if necessary */
1351 while (oth > dir)
1352 {
1353 if (dir->type == EIO_DT_DIR)
1354 ++dir;
1355 else if ((--oth)->type == EIO_DT_DIR)
1356 {
1357 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1358
1359 ++dir;
1360 }
1361 }
1362
1363 /* now sort the dirs only (dirs all have the same score) */
1364 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1365 }
1366
1367 break;
1368 }
1369
1370 /* now add the entry to our list(s) */
1371 name = entp->d_name;
1372
1373 /* skip . and .. entries */
1374 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
1375 {
1376 int len = D_NAMLEN (entp) + 1;
1377
1378 while (ecb_expect_false (namesoffs + len > namesalloc))
1379 {
1380 namesalloc *= 2;
1381 X_LOCK (wrklock);
1382 req->ptr2 = names = realloc (names, namesalloc);
1383 X_UNLOCK (wrklock);
1384
1385 if (!names)
1386 break;
1387 }
1388
1389 memcpy (names + namesoffs, name, len);
1390
1391 if (dents)
1392 {
1393 struct eio_dirent *ent;
1394
1395 if (ecb_expect_false (dentoffs == dentalloc))
1396 {
1397 dentalloc *= 2;
1398 X_LOCK (wrklock);
1399 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1400 X_UNLOCK (wrklock);
1401
1402 if (!dents)
1403 break;
1404 }
1405
1406 ent = dents + dentoffs;
1407
1408 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1409 ent->namelen = len - 1;
1410 ent->inode = D_INO (entp);
1411
1412 inode_bits |= ent->inode;
1413
1414 switch (D_TYPE (entp))
1415 {
1416 default:
1417 ent->type = EIO_DT_UNKNOWN;
1418 flags |= EIO_READDIR_FOUND_UNKNOWN;
1419 break;
1420
1421 #ifdef DT_FIFO
1422 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1423 #endif
1424 #ifdef DT_CHR
1425 case DT_CHR: ent->type = EIO_DT_CHR; break;
1426 #endif
1427 #ifdef DT_MPC
1428 case DT_MPC: ent->type = EIO_DT_MPC; break;
1429 #endif
1430 #ifdef DT_DIR
1431 case DT_DIR: ent->type = EIO_DT_DIR; break;
1432 #endif
1433 #ifdef DT_NAM
1434 case DT_NAM: ent->type = EIO_DT_NAM; break;
1435 #endif
1436 #ifdef DT_BLK
1437 case DT_BLK: ent->type = EIO_DT_BLK; break;
1438 #endif
1439 #ifdef DT_MPB
1440 case DT_MPB: ent->type = EIO_DT_MPB; break;
1441 #endif
1442 #ifdef DT_REG
1443 case DT_REG: ent->type = EIO_DT_REG; break;
1444 #endif
1445 #ifdef DT_NWK
1446 case DT_NWK: ent->type = EIO_DT_NWK; break;
1447 #endif
1448 #ifdef DT_CMP
1449 case DT_CMP: ent->type = EIO_DT_CMP; break;
1450 #endif
1451 #ifdef DT_LNK
1452 case DT_LNK: ent->type = EIO_DT_LNK; break;
1453 #endif
1454 #ifdef DT_SOCK
1455 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1456 #endif
1457 #ifdef DT_DOOR
1458 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1459 #endif
1460 #ifdef DT_WHT
1461 case DT_WHT: ent->type = EIO_DT_WHT; break;
1462 #endif
1463 }
1464
1465 ent->score = 7;
1466
1467 if (flags & EIO_READDIR_DIRS_FIRST)
1468 {
1469 if (ent->type == EIO_DT_UNKNOWN)
1470 {
1471 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1472 ent->score = 1;
1473 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1474 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1475 }
1476 else if (ent->type == EIO_DT_DIR)
1477 ent->score = 0;
1478 }
1479 }
1480
1481 namesoffs += len;
1482 ++dentoffs;
1483 }
1484
1485 if (EIO_CANCELLED (req))
1486 {
1487 errno = ECANCELED;
1488 break;
1489 }
1490 }
1491 }
1492
1493 #ifdef PAGESIZE
1494 # define eio_pagesize() PAGESIZE
1495 #else
1496 static intptr_t
1497 eio_pagesize (void)
1498 {
1499 static intptr_t page;
1500
1501 if (!page)
1502 page = sysconf (_SC_PAGESIZE);
1503
1504 return page;
1505 }
1506 #endif
1507
1508 static void
1509 eio_page_align (void **addr, size_t *length)
1510 {
1511 intptr_t mask = eio_pagesize () - 1;
1512
1513 /* round down addr */
1514 intptr_t adj = mask & (intptr_t)*addr;
1515
1516 *addr = (void *)((intptr_t)*addr - adj);
1517 *length += adj;
1518
1519 /* round up length */
1520 *length = (*length + mask) & ~mask;
1521 }
1522
1523 #if !_POSIX_MEMLOCK
1524 # define eio__mlockall(a) ((errno = ENOSYS), -1)
1525 #else
1526
1527 static int
1528 eio__mlockall (int flags)
1529 {
1530 #if __GLIBC__ == 2 && __GLIBC_MINOR__ <= 7
1531 extern int mallopt (int, int);
1532 mallopt (-6, 238); /* http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=473812 */
1533 #endif
1534
1535 if (EIO_MCL_CURRENT != MCL_CURRENT
1536 || EIO_MCL_FUTURE != MCL_FUTURE)
1537 {
1538 flags = 0
1539 | (flags & EIO_MCL_CURRENT ? MCL_CURRENT : 0)
1540 | (flags & EIO_MCL_FUTURE ? MCL_FUTURE : 0);
1541 }
1542
1543 return mlockall (flags);
1544 }
1545 #endif
1546
1547 #if !_POSIX_MEMLOCK_RANGE
1548 # define eio__mlock(a,b) ((errno = ENOSYS), -1)
1549 #else
1550
1551 static int
1552 eio__mlock (void *addr, size_t length)
1553 {
1554 eio_page_align (&addr, &length);
1555
1556 return mlock (addr, length);
1557 }
1558
1559 #endif
1560
1561 #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1562 # define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1563 #else
1564
1565 static int
1566 eio__msync (void *mem, size_t len, int flags)
1567 {
1568 eio_page_align (&mem, &len);
1569
1570 if (EIO_MS_ASYNC != MS_SYNC
1571 || EIO_MS_INVALIDATE != MS_INVALIDATE
1572 || EIO_MS_SYNC != MS_SYNC)
1573 {
1574 flags = 0
1575 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1576 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1577 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1578 }
1579
1580 return msync (mem, len, flags);
1581 }
1582
1583 #endif
1584
1585 static int
1586 eio__mtouch (eio_req *req)
1587 {
1588 void *mem = req->ptr2;
1589 size_t len = req->size;
1590 int flags = req->int1;
1591
1592 eio_page_align (&mem, &len);
1593
1594 {
1595 intptr_t addr = (intptr_t)mem;
1596 intptr_t end = addr + len;
1597 intptr_t page = eio_pagesize ();
1598
1599 if (addr < end)
1600 if (flags & EIO_MT_MODIFY) /* modify */
1601 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len && !EIO_CANCELLED (req));
1602 else
1603 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len && !EIO_CANCELLED (req));
1604 }
1605
1606 return 0;
1607 }
1608
1609 /*****************************************************************************/
1610
1611 #define ALLOC(len) \
1612 if (!req->ptr2) \
1613 { \
1614 X_LOCK (wrklock); \
1615 req->flags |= EIO_FLAG_PTR2_FREE; \
1616 X_UNLOCK (wrklock); \
1617 req->ptr2 = malloc (len); \
1618 if (!req->ptr2) \
1619 { \
1620 errno = ENOMEM; \
1621 req->result = -1; \
1622 break; \
1623 } \
1624 }
1625
1626 X_THREAD_PROC (etp_proc)
1627 {
1628 ETP_REQ *req;
1629 struct timespec ts;
1630 etp_worker *self = (etp_worker *)thr_arg;
1631
1632 /* try to distribute timeouts somewhat randomly */
1633 ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
1634
1635 for (;;)
1636 {
1637 X_LOCK (reqlock);
1638
1639 for (;;)
1640 {
1641 self->req = req = reqq_shift (&req_queue);
1642
1643 if (req)
1644 break;
1645
1646 ++idle;
1647
1648 ts.tv_sec = time (0) + idle_timeout;
1649 if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
1650 {
1651 if (idle > max_idle)
1652 {
1653 --idle;
1654 X_UNLOCK (reqlock);
1655 X_LOCK (wrklock);
1656 --started;
1657 X_UNLOCK (wrklock);
1658 goto quit;
1659 }
1660
1661 /* we are allowed to idle, so do so without any timeout */
1662 X_COND_WAIT (reqwait, reqlock);
1663 }
1664
1665 --idle;
1666 }
1667
1668 --nready;
1669
1670 X_UNLOCK (reqlock);
1671
1672 if (req->type < 0)
1673 goto quit;
1674
1675 if (!EIO_CANCELLED (req))
1676 ETP_EXECUTE (self, req);
1677
1678 X_LOCK (reslock);
1679
1680 ++npending;
1681
1682 if (!reqq_push (&res_queue, req) && want_poll_cb)
1683 want_poll_cb ();
1684
1685 self->req = 0;
1686 etp_worker_clear (self);
1687
1688 X_UNLOCK (reslock);
1689 }
1690
1691 quit:
1692 X_LOCK (wrklock);
1693 etp_worker_free (self);
1694 X_UNLOCK (wrklock);
1695
1696 return 0;
1697 }
1698
1699 /*****************************************************************************/
1700
1701 int ecb_cold
1702 eio_init (void (*want_poll)(void), void (*done_poll)(void))
1703 {
1704 return etp_init (want_poll, done_poll);
1705 }
1706
1707 ecb_inline void
1708 eio_api_destroy (eio_req *req)
1709 {
1710 free (req);
1711 }
1712
1713 #define REQ(rtype) \
1714 eio_req *req; \
1715 \
1716 req = (eio_req *)calloc (1, sizeof *req); \
1717 if (!req) \
1718 return 0; \
1719 \
1720 req->type = rtype; \
1721 req->pri = pri; \
1722 req->finish = cb; \
1723 req->data = data; \
1724 req->destroy = eio_api_destroy;
1725
1726 #define SEND eio_submit (req); return req
1727
1728 #define PATH \
1729 req->flags |= EIO_FLAG_PTR1_FREE; \
1730 req->ptr1 = strdup (path); \
1731 if (!req->ptr1) \
1732 { \
1733 eio_api_destroy (req); \
1734 return 0; \
1735 }
1736
1737 static void
1738 eio_execute (etp_worker *self, eio_req *req)
1739 {
1740 switch (req->type)
1741 {
1742 case EIO_READ: ALLOC (req->size);
1743 req->result = req->offs >= 0
1744 ? pread (req->int1, req->ptr2, req->size, req->offs)
1745 : read (req->int1, req->ptr2, req->size); break;
1746 case EIO_WRITE: req->result = req->offs >= 0
1747 ? pwrite (req->int1, req->ptr2, req->size, req->offs)
1748 : write (req->int1, req->ptr2, req->size); break;
1749
1750 case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
1751 case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
1752
1753 case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1754 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1755 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1756 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1757 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1758 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1759
1760 case EIO_STATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1761 req->result = statvfs (req->ptr1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1762 case EIO_FSTATVFS: ALLOC (sizeof (EIO_STRUCT_STATVFS));
1763 req->result = fstatvfs (req->int1, (EIO_STRUCT_STATVFS *)req->ptr2); break;
1764
1765 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1766 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1767 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1768 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
1769 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
1770 case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
1771
1772 case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
1773 case EIO_CLOSE: req->result = close (req->int1); break;
1774 case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
1775 case EIO_UNLINK: req->result = unlink (req->ptr1); break;
1776 case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
1777 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
1778 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1779 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1780 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1781 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->offs); break;
1782
1783 case EIO_REALPATH: req->flags |= EIO_FLAG_PTR2_FREE;
1784 req->ptr2 = realpath (req->ptr1, 0);
1785 req->result = req->ptr2 ? strlen (req->ptr2) : -1;
1786 break;
1787
1788 case EIO_READLINK: ALLOC (PATH_MAX);
1789 req->result = readlink (req->ptr1, req->ptr2, PATH_MAX); break;
1790
1791 case EIO_SYNC: req->result = 0; sync (); break;
1792 case EIO_FSYNC: req->result = fsync (req->int1); break;
1793 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1794 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1795 case EIO_MTOUCH: req->result = eio__mtouch (req); break;
1796 case EIO_MLOCK: req->result = eio__mlock (req->ptr2, req->size); break;
1797 case EIO_MLOCKALL: req->result = eio__mlockall (req->int1); break;
1798 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1799
1800 case EIO_READDIR: eio__scandir (req, self); break;
1801
1802 case EIO_BUSY:
1803 #ifdef _WIN32
1804 Sleep (req->nv1 * 1e3);
1805 #else
1806 {
1807 struct timeval tv;
1808
1809 tv.tv_sec = req->nv1;
1810 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1e6;
1811
1812 req->result = select (0, 0, 0, 0, &tv);
1813 }
1814 #endif
1815 break;
1816
1817 case EIO_UTIME:
1818 case EIO_FUTIME:
1819 {
1820 struct timeval tv[2];
1821 struct timeval *times;
1822
1823 if (req->nv1 != -1. || req->nv2 != -1.)
1824 {
1825 tv[0].tv_sec = req->nv1;
1826 tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
1827 tv[1].tv_sec = req->nv2;
1828 tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
1829
1830 times = tv;
1831 }
1832 else
1833 times = 0;
1834
1835 req->result = req->type == EIO_FUTIME
1836 ? futimes (req->int1, times)
1837 : utimes (req->ptr1, times);
1838 }
1839 break;
1840
1841 case EIO_GROUP:
1842 abort (); /* handled in eio_request */
1843
1844 case EIO_NOP:
1845 req->result = 0;
1846 break;
1847
1848 case EIO_CUSTOM:
1849 req->feed (req);
1850 break;
1851
1852 default:
1853 errno = ENOSYS;
1854 req->result = -1;
1855 break;
1856 }
1857
1858 req->errorno = errno;
1859 }
1860
1861 #ifndef EIO_NO_WRAPPERS
1862
1863 eio_req *eio_nop (int pri, eio_cb cb, void *data)
1864 {
1865 REQ (EIO_NOP); SEND;
1866 }
1867
1868 eio_req *eio_busy (double delay, int pri, eio_cb cb, void *data)
1869 {
1870 REQ (EIO_BUSY); req->nv1 = delay; SEND;
1871 }
1872
1873 eio_req *eio_sync (int pri, eio_cb cb, void *data)
1874 {
1875 REQ (EIO_SYNC); SEND;
1876 }
1877
1878 eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1879 {
1880 REQ (EIO_FSYNC); req->int1 = fd; SEND;
1881 }
1882
1883 eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1884 {
1885 REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1886 }
1887
1888 eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1889 {
1890 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1891 }
1892
1893 eio_req *eio_mlock (void *addr, size_t length, int pri, eio_cb cb, void *data)
1894 {
1895 REQ (EIO_MLOCK); req->ptr2 = addr; req->size = length; SEND;
1896 }
1897
1898 eio_req *eio_mlockall (int flags, int pri, eio_cb cb, void *data)
1899 {
1900 REQ (EIO_MLOCKALL); req->int1 = flags; SEND;
1901 }
1902
1903 eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1904 {
1905 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1906 }
1907
1908 eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1909 {
1910 REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1911 }
1912
1913 eio_req *eio_close (int fd, int pri, eio_cb cb, void *data)
1914 {
1915 REQ (EIO_CLOSE); req->int1 = fd; SEND;
1916 }
1917
1918 eio_req *eio_readahead (int fd, off_t offset, size_t length, int pri, eio_cb cb, void *data)
1919 {
1920 REQ (EIO_READAHEAD); req->int1 = fd; req->offs = offset; req->size = length; SEND;
1921 }
1922
1923 eio_req *eio_read (int fd, void *buf, size_t length, off_t offset, int pri, eio_cb cb, void *data)
1924 {
1925 REQ (EIO_READ); req->int1 = fd; req->offs = offset; req->size = length; req->ptr2 = buf; SEND;
1926 }
1927
1928 eio_req *eio_write (int fd, void *buf, size_t length, off_t offset, int pri, eio_cb cb, void *data)
1929 {
1930 REQ (EIO_WRITE); req->int1 = fd; req->offs = offset; req->size = length; req->ptr2 = buf; SEND;
1931 }
1932
1933 eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1934 {
1935 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1936 }
1937
1938 eio_req *eio_fstatvfs (int fd, int pri, eio_cb cb, void *data)
1939 {
1940 REQ (EIO_FSTATVFS); req->int1 = fd; SEND;
1941 }
1942
1943 eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1944 {
1945 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
1946 }
1947
1948 eio_req *eio_ftruncate (int fd, off_t offset, int pri, eio_cb cb, void *data)
1949 {
1950 REQ (EIO_FTRUNCATE); req->int1 = fd; req->offs = offset; SEND;
1951 }
1952
1953 eio_req *eio_fchmod (int fd, mode_t mode, int pri, eio_cb cb, void *data)
1954 {
1955 REQ (EIO_FCHMOD); req->int1 = fd; req->int2 = (long)mode; SEND;
1956 }
1957
1958 eio_req *eio_fchown (int fd, uid_t uid, gid_t gid, int pri, eio_cb cb, void *data)
1959 {
1960 REQ (EIO_FCHOWN); req->int1 = fd; req->int2 = (long)uid; req->int3 = (long)gid; SEND;
1961 }
1962
1963 eio_req *eio_dup2 (int fd, int fd2, int pri, eio_cb cb, void *data)
1964 {
1965 REQ (EIO_DUP2); req->int1 = fd; req->int2 = fd2; SEND;
1966 }
1967
1968 eio_req *eio_sendfile (int out_fd, int in_fd, off_t in_offset, size_t length, int pri, eio_cb cb, void *data)
1969 {
1970 REQ (EIO_SENDFILE); req->int1 = out_fd; req->int2 = in_fd; req->offs = in_offset; req->size = length; SEND;
1971 }
1972
1973 eio_req *eio_open (const char *path, int flags, mode_t mode, int pri, eio_cb cb, void *data)
1974 {
1975 REQ (EIO_OPEN); PATH; req->int1 = flags; req->int2 = (long)mode; SEND;
1976 }
1977
1978 eio_req *eio_utime (const char *path, double atime, double mtime, int pri, eio_cb cb, void *data)
1979 {
1980 REQ (EIO_UTIME); PATH; req->nv1 = atime; req->nv2 = mtime; SEND;
1981 }
1982
1983 eio_req *eio_truncate (const char *path, off_t offset, int pri, eio_cb cb, void *data)
1984 {
1985 REQ (EIO_TRUNCATE); PATH; req->offs = offset; SEND;
1986 }
1987
1988 eio_req *eio_chown (const char *path, uid_t uid, gid_t gid, int pri, eio_cb cb, void *data)
1989 {
1990 REQ (EIO_CHOWN); PATH; req->int2 = (long)uid; req->int3 = (long)gid; SEND;
1991 }
1992
1993 eio_req *eio_chmod (const char *path, mode_t mode, int pri, eio_cb cb, void *data)
1994 {
1995 REQ (EIO_CHMOD); PATH; req->int2 = (long)mode; SEND;
1996 }
1997
1998 eio_req *eio_mkdir (const char *path, mode_t mode, int pri, eio_cb cb, void *data)
1999 {
2000 REQ (EIO_MKDIR); PATH; req->int2 = (long)mode; SEND;
2001 }
2002
2003 static eio_req *
2004 eio__1path (int type, const char *path, int pri, eio_cb cb, void *data)
2005 {
2006 REQ (type); PATH; SEND;
2007 }
2008
2009 eio_req *eio_readlink (const char *path, int pri, eio_cb cb, void *data)
2010 {
2011 return eio__1path (EIO_READLINK, path, pri, cb, data);
2012 }
2013
2014 eio_req *eio_realpath (const char *path, int pri, eio_cb cb, void *data)
2015 {
2016 return eio__1path (EIO_REALPATH, path, pri, cb, data);
2017 }
2018
2019 eio_req *eio_stat (const char *path, int pri, eio_cb cb, void *data)
2020 {
2021 return eio__1path (EIO_STAT, path, pri, cb, data);
2022 }
2023
2024 eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
2025 {
2026 return eio__1path (EIO_LSTAT, path, pri, cb, data);
2027 }
2028
2029 eio_req *eio_statvfs (const char *path, int pri, eio_cb cb, void *data)
2030 {
2031 return eio__1path (EIO_STATVFS, path, pri, cb, data);
2032 }
2033
2034 eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
2035 {
2036 return eio__1path (EIO_UNLINK, path, pri, cb, data);
2037 }
2038
2039 eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
2040 {
2041 return eio__1path (EIO_RMDIR, path, pri, cb, data);
2042 }
2043
2044 eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
2045 {
2046 REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
2047 }
2048
2049 eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
2050 {
2051 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->offs = (off_t)dev; SEND;
2052 }
2053
2054 static eio_req *
2055 eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
2056 {
2057 REQ (type); PATH;
2058
2059 req->flags |= EIO_FLAG_PTR2_FREE;
2060 req->ptr2 = strdup (new_path);
2061 if (!req->ptr2)
2062 {
2063 eio_api_destroy (req);
2064 return 0;
2065 }
2066
2067 SEND;
2068 }
2069
2070 eio_req *eio_link (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
2071 {
2072 return eio__2path (EIO_LINK, path, new_path, pri, cb, data);
2073 }
2074
2075 eio_req *eio_symlink (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
2076 {
2077 return eio__2path (EIO_SYMLINK, path, new_path, pri, cb, data);
2078 }
2079
2080 eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
2081 {
2082 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
2083 }
2084
2085 eio_req *eio_custom (void (*execute)(eio_req *), int pri, eio_cb cb, void *data)
2086 {
2087 REQ (EIO_CUSTOM); req->feed = execute; SEND;
2088 }
2089
2090 #endif
2091
2092 eio_req *eio_grp (eio_cb cb, void *data)
2093 {
2094 const int pri = EIO_PRI_MAX;
2095
2096 REQ (EIO_GROUP); SEND;
2097 }
2098
2099 #undef REQ
2100 #undef PATH
2101 #undef SEND
2102
2103 /*****************************************************************************/
2104 /* grp functions */
2105
2106 void
2107 eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit)
2108 {
2109 grp->int2 = limit;
2110 grp->feed = feed;
2111
2112 grp_try_feed (grp);
2113 }
2114
2115 void
2116 eio_grp_limit (eio_req *grp, int limit)
2117 {
2118 grp->int2 = limit;
2119
2120 grp_try_feed (grp);
2121 }
2122
2123 void
2124 eio_grp_add (eio_req *grp, eio_req *req)
2125 {
2126 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
2127
2128 grp->flags |= EIO_FLAG_GROUPADD;
2129
2130 ++grp->size;
2131 req->grp = grp;
2132
2133 req->grp_prev = 0;
2134 req->grp_next = grp->grp_first;
2135
2136 if (grp->grp_first)
2137 grp->grp_first->grp_prev = req;
2138
2139 grp->grp_first = req;
2140 }
2141
2142 /*****************************************************************************/
2143 /* misc garbage */
2144
2145 ssize_t
2146 eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
2147 {
2148 etp_worker wrk;
2149 ssize_t ret;
2150
2151 wrk.dbuf = 0;
2152
2153 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
2154
2155 if (wrk.dbuf)
2156 free (wrk.dbuf);
2157
2158 return ret;
2159 }
2160