ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libeio/eio.c
Revision: 1.49
Committed: Sat Jan 2 14:24:32 2010 UTC (14 years, 11 months ago) by root
Content type: text/plain
Branch: MAIN
CVS Tags: rel-3_4
Changes since 1.48: +5 -1 lines
Log Message:
3.4

File Contents

# Content
1 /*
2 * libeio implementation
3 *
4 * Copyright (c) 2007,2008,2009,2010 Marc Alexander Lehmann <libeio@schmorp.de>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without modifica-
8 * tion, are permitted provided that the following conditions are met:
9 *
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
12 *
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
19 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
20 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
21 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
23 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
24 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
25 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
26 * OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Alternatively, the contents of this file may be used under the terms of
29 * the GNU General Public License ("GPL") version 2 or any later version,
30 * in which case the provisions of the GPL are applicable instead of
31 * the above. If you wish to allow the use of your version of this file
32 * only under the terms of the GPL and not to allow others to use your
33 * version of this file under the BSD license, indicate your decision
34 * by deleting the provisions above and replace them with the notice
35 * and other provisions required by the GPL. If you do not delete the
36 * provisions above, a recipient may use your version of this file under
37 * either the BSD or the GPL.
38 */
39
40 #include "eio.h"
41
42 #ifdef EIO_STACKSIZE
43 # define XTHREAD_STACKSIZE EIO_STACKSIZE
44 #endif
45 #include "xthread.h"
46
47 #include <errno.h>
48 #include <stddef.h>
49 #include <stdlib.h>
50 #include <string.h>
51 #include <errno.h>
52 #include <sys/types.h>
53 #include <sys/stat.h>
54 #include <limits.h>
55 #include <fcntl.h>
56 #include <assert.h>
57
58 #ifndef EIO_FINISH
59 # define EIO_FINISH(req) ((req)->finish) && !EIO_CANCELLED (req) ? (req)->finish (req) : 0
60 #endif
61
62 #ifndef EIO_DESTROY
63 # define EIO_DESTROY(req) do { if ((req)->destroy) (req)->destroy (req); } while (0)
64 #endif
65
66 #ifndef EIO_FEED
67 # define EIO_FEED(req) do { if ((req)->feed ) (req)->feed (req); } while (0)
68 #endif
69
70 #ifdef _WIN32
71
72 /*doh*/
73 #else
74
75 # include "config.h"
76 # include <sys/time.h>
77 # include <sys/select.h>
78 # include <sys/mman.h>
79 # include <unistd.h>
80 # include <utime.h>
81 # include <signal.h>
82 # include <dirent.h>
83
84 /* POSIX_SOURCE is useless on bsd's, and XOPEN_SOURCE is unreliable there, too */
85 # if __freebsd || defined __NetBSD__ || defined __OpenBSD__
86 # define _DIRENT_HAVE_D_TYPE /* sigh */
87 # define D_INO(de) (de)->d_fileno
88 # define D_NAMLEN(de) (de)->d_namlen
89 # elif __linux || defined d_ino || _XOPEN_SOURCE >= 600
90 # define D_INO(de) (de)->d_ino
91 # endif
92
93 #ifdef _D_EXACT_NAMLEN
94 # undef D_NAMLEN
95 # define D_NAMLEN(de) _D_EXACT_NAMLEN (de)
96 #endif
97
98 # ifdef _DIRENT_HAVE_D_TYPE
99 # define D_TYPE(de) (de)->d_type
100 # endif
101
102 # ifndef EIO_STRUCT_DIRENT
103 # define EIO_STRUCT_DIRENT struct dirent
104 # endif
105
106 #endif
107
108 #if HAVE_SENDFILE
109 # if __linux
110 # include <sys/sendfile.h>
111 # elif __freebsd || defined __APPLE__
112 # include <sys/socket.h>
113 # include <sys/uio.h>
114 # elif __hpux
115 # include <sys/socket.h>
116 # elif __solaris
117 # include <sys/sendfile.h>
118 # else
119 # error sendfile support requested but not available
120 # endif
121 #endif
122
123 #ifndef D_TYPE
124 # define D_TYPE(de) 0
125 #endif
126 #ifndef D_INO
127 # define D_INO(de) 0
128 #endif
129 #ifndef D_NAMLEN
130 # define D_NAMLEN(de) strlen ((de)->d_name)
131 #endif
132
133 /* number of seconds after which an idle threads exit */
134 #define IDLE_TIMEOUT 10
135
136 /* used for struct dirent, AIX doesn't provide it */
137 #ifndef NAME_MAX
138 # define NAME_MAX 4096
139 #endif
140
141 /* buffer size for various temporary buffers */
142 #define EIO_BUFSIZE 65536
143
144 #define dBUF \
145 char *eio_buf; \
146 ETP_WORKER_LOCK (self); \
147 self->dbuf = eio_buf = malloc (EIO_BUFSIZE); \
148 ETP_WORKER_UNLOCK (self); \
149 errno = ENOMEM; \
150 if (!eio_buf) \
151 return -1;
152
153 #define EIO_TICKS ((1000000 + 1023) >> 10)
154
155 /*****************************************************************************/
156
157 #if __GNUC__ >= 3
158 # define expect(expr,value) __builtin_expect ((expr),(value))
159 #else
160 # define expect(expr,value) (expr)
161 #endif
162
163 #define expect_false(expr) expect ((expr) != 0, 0)
164 #define expect_true(expr) expect ((expr) != 0, 1)
165
166 /*****************************************************************************/
167
168 #define ETP_PRI_MIN EIO_PRI_MIN
169 #define ETP_PRI_MAX EIO_PRI_MAX
170
171 struct etp_worker;
172
173 #define ETP_REQ eio_req
174 #define ETP_DESTROY(req) eio_destroy (req)
175 static int eio_finish (eio_req *req);
176 #define ETP_FINISH(req) eio_finish (req)
177 static void eio_execute (struct etp_worker *self, eio_req *req);
178 #define ETP_EXECUTE(wrk,req) eio_execute (wrk,req)
179
180 #define ETP_WORKER_CLEAR(req) \
181 if (wrk->dbuf) \
182 { \
183 free (wrk->dbuf); \
184 wrk->dbuf = 0; \
185 } \
186 \
187 if (wrk->dirp) \
188 { \
189 closedir (wrk->dirp); \
190 wrk->dirp = 0; \
191 }
192
193 #define ETP_WORKER_COMMON \
194 void *dbuf; \
195 DIR *dirp;
196
197 /*****************************************************************************/
198
199 #define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
200
201 /* calculate time difference in ~1/EIO_TICKS of a second */
202 static int tvdiff (struct timeval *tv1, struct timeval *tv2)
203 {
204 return (tv2->tv_sec - tv1->tv_sec ) * EIO_TICKS
205 + ((tv2->tv_usec - tv1->tv_usec) >> 10);
206 }
207
208 static unsigned int started, idle, wanted = 4;
209
210 static void (*want_poll_cb) (void);
211 static void (*done_poll_cb) (void);
212
213 static unsigned int max_poll_time; /* reslock */
214 static unsigned int max_poll_reqs; /* reslock */
215
216 static volatile unsigned int nreqs; /* reqlock */
217 static volatile unsigned int nready; /* reqlock */
218 static volatile unsigned int npending; /* reqlock */
219 static volatile unsigned int max_idle = 4;
220
221 static mutex_t wrklock = X_MUTEX_INIT;
222 static mutex_t reslock = X_MUTEX_INIT;
223 static mutex_t reqlock = X_MUTEX_INIT;
224 static cond_t reqwait = X_COND_INIT;
225
226 #if !HAVE_PREADWRITE
227 /*
228 * make our pread/pwrite emulation safe against themselves, but not against
229 * normal read/write by using a mutex. slows down execution a lot,
230 * but that's your problem, not mine.
231 */
232 static mutex_t preadwritelock = X_MUTEX_INIT;
233 #endif
234
235 typedef struct etp_worker
236 {
237 /* locked by wrklock */
238 struct etp_worker *prev, *next;
239
240 thread_t tid;
241
242 /* locked by reslock, reqlock or wrklock */
243 ETP_REQ *req; /* currently processed request */
244
245 ETP_WORKER_COMMON
246 } etp_worker;
247
248 static etp_worker wrk_first = { &wrk_first, &wrk_first, 0 }; /* NOT etp */
249
250 #define ETP_WORKER_LOCK(wrk) X_LOCK (wrklock)
251 #define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (wrklock)
252
253 /* worker threads management */
254
255 static void etp_worker_clear (etp_worker *wrk)
256 {
257 ETP_WORKER_CLEAR (wrk);
258 }
259
260 static void etp_worker_free (etp_worker *wrk)
261 {
262 wrk->next->prev = wrk->prev;
263 wrk->prev->next = wrk->next;
264
265 free (wrk);
266 }
267
268 static unsigned int etp_nreqs (void)
269 {
270 int retval;
271 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
272 retval = nreqs;
273 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
274 return retval;
275 }
276
277 static unsigned int etp_nready (void)
278 {
279 unsigned int retval;
280
281 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
282 retval = nready;
283 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
284
285 return retval;
286 }
287
288 static unsigned int etp_npending (void)
289 {
290 unsigned int retval;
291
292 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
293 retval = npending;
294 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
295
296 return retval;
297 }
298
299 static unsigned int etp_nthreads (void)
300 {
301 unsigned int retval;
302
303 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
304 retval = started;
305 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
306
307 return retval;
308 }
309
310 /*
311 * a somewhat faster data structure might be nice, but
312 * with 8 priorities this actually needs <20 insns
313 * per shift, the most expensive operation.
314 */
315 typedef struct {
316 ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
317 int size;
318 } etp_reqq;
319
320 static etp_reqq req_queue;
321 static etp_reqq res_queue;
322
323 static int reqq_push (etp_reqq *q, ETP_REQ *req)
324 {
325 int pri = req->pri;
326 req->next = 0;
327
328 if (q->qe[pri])
329 {
330 q->qe[pri]->next = req;
331 q->qe[pri] = req;
332 }
333 else
334 q->qe[pri] = q->qs[pri] = req;
335
336 return q->size++;
337 }
338
339 static ETP_REQ *reqq_shift (etp_reqq *q)
340 {
341 int pri;
342
343 if (!q->size)
344 return 0;
345
346 --q->size;
347
348 for (pri = ETP_NUM_PRI; pri--; )
349 {
350 eio_req *req = q->qs[pri];
351
352 if (req)
353 {
354 if (!(q->qs[pri] = (eio_req *)req->next))
355 q->qe[pri] = 0;
356
357 return req;
358 }
359 }
360
361 abort ();
362 }
363
364 static void etp_atfork_prepare (void)
365 {
366 X_LOCK (wrklock);
367 X_LOCK (reqlock);
368 X_LOCK (reslock);
369 #if !HAVE_PREADWRITE
370 X_LOCK (preadwritelock);
371 #endif
372 }
373
374 static void etp_atfork_parent (void)
375 {
376 #if !HAVE_PREADWRITE
377 X_UNLOCK (preadwritelock);
378 #endif
379 X_UNLOCK (reslock);
380 X_UNLOCK (reqlock);
381 X_UNLOCK (wrklock);
382 }
383
384 static void etp_atfork_child (void)
385 {
386 ETP_REQ *prv;
387
388 while ((prv = reqq_shift (&req_queue)))
389 ETP_DESTROY (prv);
390
391 while ((prv = reqq_shift (&res_queue)))
392 ETP_DESTROY (prv);
393
394 while (wrk_first.next != &wrk_first)
395 {
396 etp_worker *wrk = wrk_first.next;
397
398 if (wrk->req)
399 ETP_DESTROY (wrk->req);
400
401 etp_worker_clear (wrk);
402 etp_worker_free (wrk);
403 }
404
405 started = 0;
406 idle = 0;
407 nreqs = 0;
408 nready = 0;
409 npending = 0;
410
411 etp_atfork_parent ();
412 }
413
414 static void
415 etp_once_init (void)
416 {
417 X_THREAD_ATFORK (etp_atfork_prepare, etp_atfork_parent, etp_atfork_child);
418 }
419
420 static int
421 etp_init (void (*want_poll)(void), void (*done_poll)(void))
422 {
423 static pthread_once_t doinit = PTHREAD_ONCE_INIT;
424
425 pthread_once (&doinit, etp_once_init);
426
427 want_poll_cb = want_poll;
428 done_poll_cb = done_poll;
429
430 return 0;
431 }
432
433 X_THREAD_PROC (etp_proc);
434
435 static void etp_start_thread (void)
436 {
437 etp_worker *wrk = calloc (1, sizeof (etp_worker));
438
439 /*TODO*/
440 assert (("unable to allocate worker thread data", wrk));
441
442 X_LOCK (wrklock);
443
444 if (thread_create (&wrk->tid, etp_proc, (void *)wrk))
445 {
446 wrk->prev = &wrk_first;
447 wrk->next = wrk_first.next;
448 wrk_first.next->prev = wrk;
449 wrk_first.next = wrk;
450 ++started;
451 }
452 else
453 free (wrk);
454
455 X_UNLOCK (wrklock);
456 }
457
458 static void etp_maybe_start_thread (void)
459 {
460 if (expect_true (etp_nthreads () >= wanted))
461 return;
462
463 /* todo: maybe use idle here, but might be less exact */
464 if (expect_true (0 <= (int)etp_nthreads () + (int)etp_npending () - (int)etp_nreqs ()))
465 return;
466
467 etp_start_thread ();
468 }
469
470 static void etp_end_thread (void)
471 {
472 eio_req *req = calloc (1, sizeof (eio_req));
473
474 req->type = -1;
475 req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
476
477 X_LOCK (reqlock);
478 reqq_push (&req_queue, req);
479 X_COND_SIGNAL (reqwait);
480 X_UNLOCK (reqlock);
481
482 X_LOCK (wrklock);
483 --started;
484 X_UNLOCK (wrklock);
485 }
486
487 static int etp_poll (void)
488 {
489 unsigned int maxreqs;
490 unsigned int maxtime;
491 struct timeval tv_start, tv_now;
492
493 X_LOCK (reslock);
494 maxreqs = max_poll_reqs;
495 maxtime = max_poll_time;
496 X_UNLOCK (reslock);
497
498 if (maxtime)
499 gettimeofday (&tv_start, 0);
500
501 for (;;)
502 {
503 ETP_REQ *req;
504
505 etp_maybe_start_thread ();
506
507 X_LOCK (reslock);
508 req = reqq_shift (&res_queue);
509
510 if (req)
511 {
512 --npending;
513
514 if (!res_queue.size && done_poll_cb)
515 done_poll_cb ();
516 }
517
518 X_UNLOCK (reslock);
519
520 if (!req)
521 return 0;
522
523 X_LOCK (reqlock);
524 --nreqs;
525 X_UNLOCK (reqlock);
526
527 if (expect_false (req->type == EIO_GROUP && req->size))
528 {
529 req->int1 = 1; /* mark request as delayed */
530 continue;
531 }
532 else
533 {
534 int res = ETP_FINISH (req);
535 if (expect_false (res))
536 return res;
537 }
538
539 if (expect_false (maxreqs && !--maxreqs))
540 break;
541
542 if (maxtime)
543 {
544 gettimeofday (&tv_now, 0);
545
546 if (tvdiff (&tv_start, &tv_now) >= maxtime)
547 break;
548 }
549 }
550
551 errno = EAGAIN;
552 return -1;
553 }
554
555 static void etp_cancel (ETP_REQ *req)
556 {
557 X_LOCK (wrklock);
558 req->flags |= EIO_FLAG_CANCELLED;
559 X_UNLOCK (wrklock);
560
561 eio_grp_cancel (req);
562 }
563
564 static void etp_submit (ETP_REQ *req)
565 {
566 req->pri -= ETP_PRI_MIN;
567
568 if (expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
569 if (expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
570
571 if (expect_false (req->type == EIO_GROUP))
572 {
573 /* I hope this is worth it :/ */
574 X_LOCK (reqlock);
575 ++nreqs;
576 X_UNLOCK (reqlock);
577
578 X_LOCK (reslock);
579
580 ++npending;
581
582 if (!reqq_push (&res_queue, req) && want_poll_cb)
583 want_poll_cb ();
584
585 X_UNLOCK (reslock);
586 }
587 else
588 {
589 X_LOCK (reqlock);
590 ++nreqs;
591 ++nready;
592 reqq_push (&req_queue, req);
593 X_COND_SIGNAL (reqwait);
594 X_UNLOCK (reqlock);
595
596 etp_maybe_start_thread ();
597 }
598 }
599
600 static void etp_set_max_poll_time (double nseconds)
601 {
602 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
603 max_poll_time = nseconds * EIO_TICKS;
604 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
605 }
606
607 static void etp_set_max_poll_reqs (unsigned int maxreqs)
608 {
609 if (WORDACCESS_UNSAFE) X_LOCK (reslock);
610 max_poll_reqs = maxreqs;
611 if (WORDACCESS_UNSAFE) X_UNLOCK (reslock);
612 }
613
614 static void etp_set_max_idle (unsigned int nthreads)
615 {
616 if (WORDACCESS_UNSAFE) X_LOCK (reqlock);
617 max_idle = nthreads <= 0 ? 1 : nthreads;
618 if (WORDACCESS_UNSAFE) X_UNLOCK (reqlock);
619 }
620
621 static void etp_set_min_parallel (unsigned int nthreads)
622 {
623 if (wanted < nthreads)
624 wanted = nthreads;
625 }
626
627 static void etp_set_max_parallel (unsigned int nthreads)
628 {
629 if (wanted > nthreads)
630 wanted = nthreads;
631
632 while (started > wanted)
633 etp_end_thread ();
634 }
635
636 /*****************************************************************************/
637
638 static void grp_try_feed (eio_req *grp)
639 {
640 while (grp->size < grp->int2 && !EIO_CANCELLED (grp))
641 {
642 grp->flags &= ~EIO_FLAG_GROUPADD;
643
644 EIO_FEED (grp);
645
646 /* stop if no progress has been made */
647 if (!(grp->flags & EIO_FLAG_GROUPADD))
648 {
649 grp->feed = 0;
650 break;
651 }
652 }
653 }
654
655 static int grp_dec (eio_req *grp)
656 {
657 --grp->size;
658
659 /* call feeder, if applicable */
660 grp_try_feed (grp);
661
662 /* finish, if done */
663 if (!grp->size && grp->int1)
664 return eio_finish (grp);
665 else
666 return 0;
667 }
668
669 void eio_destroy (eio_req *req)
670 {
671 if ((req)->flags & EIO_FLAG_PTR1_FREE) free (req->ptr1);
672 if ((req)->flags & EIO_FLAG_PTR2_FREE) free (req->ptr2);
673
674 EIO_DESTROY (req);
675 }
676
677 static int eio_finish (eio_req *req)
678 {
679 int res = EIO_FINISH (req);
680
681 if (req->grp)
682 {
683 int res2;
684 eio_req *grp = req->grp;
685
686 /* unlink request */
687 if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
688 if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
689
690 if (grp->grp_first == req)
691 grp->grp_first = req->grp_next;
692
693 res2 = grp_dec (grp);
694
695 if (!res && res2)
696 res = res2;
697 }
698
699 eio_destroy (req);
700
701 return res;
702 }
703
704 void eio_grp_cancel (eio_req *grp)
705 {
706 for (grp = grp->grp_first; grp; grp = grp->grp_next)
707 eio_cancel (grp);
708 }
709
710 void eio_cancel (eio_req *req)
711 {
712 etp_cancel (req);
713 }
714
715 void eio_submit (eio_req *req)
716 {
717 etp_submit (req);
718 }
719
720 unsigned int eio_nreqs (void)
721 {
722 return etp_nreqs ();
723 }
724
725 unsigned int eio_nready (void)
726 {
727 return etp_nready ();
728 }
729
730 unsigned int eio_npending (void)
731 {
732 return etp_npending ();
733 }
734
735 unsigned int eio_nthreads (void)
736 {
737 return etp_nthreads ();
738 }
739
740 void eio_set_max_poll_time (double nseconds)
741 {
742 etp_set_max_poll_time (nseconds);
743 }
744
745 void eio_set_max_poll_reqs (unsigned int maxreqs)
746 {
747 etp_set_max_poll_reqs (maxreqs);
748 }
749
750 void eio_set_max_idle (unsigned int nthreads)
751 {
752 etp_set_max_idle (nthreads);
753 }
754
755 void eio_set_min_parallel (unsigned int nthreads)
756 {
757 etp_set_min_parallel (nthreads);
758 }
759
760 void eio_set_max_parallel (unsigned int nthreads)
761 {
762 etp_set_max_parallel (nthreads);
763 }
764
765 int eio_poll (void)
766 {
767 return etp_poll ();
768 }
769
770 /*****************************************************************************/
771 /* work around various missing functions */
772
773 #if !HAVE_PREADWRITE
774 # undef pread
775 # undef pwrite
776 # define pread eio__pread
777 # define pwrite eio__pwrite
778
779 static ssize_t
780 eio__pread (int fd, void *buf, size_t count, off_t offset)
781 {
782 ssize_t res;
783 off_t ooffset;
784
785 X_LOCK (preadwritelock);
786 ooffset = lseek (fd, 0, SEEK_CUR);
787 lseek (fd, offset, SEEK_SET);
788 res = read (fd, buf, count);
789 lseek (fd, ooffset, SEEK_SET);
790 X_UNLOCK (preadwritelock);
791
792 return res;
793 }
794
795 static ssize_t
796 eio__pwrite (int fd, void *buf, size_t count, off_t offset)
797 {
798 ssize_t res;
799 off_t ooffset;
800
801 X_LOCK (preadwritelock);
802 ooffset = lseek (fd, 0, SEEK_CUR);
803 lseek (fd, offset, SEEK_SET);
804 res = write (fd, buf, count);
805 lseek (fd, ooffset, SEEK_SET);
806 X_UNLOCK (preadwritelock);
807
808 return res;
809 }
810 #endif
811
812 #ifndef HAVE_FUTIMES
813
814 # undef utimes
815 # undef futimes
816 # define utimes(path,times) eio__utimes (path, times)
817 # define futimes(fd,times) eio__futimes (fd, times)
818
819 static int
820 eio__utimes (const char *filename, const struct timeval times[2])
821 {
822 if (times)
823 {
824 struct utimbuf buf;
825
826 buf.actime = times[0].tv_sec;
827 buf.modtime = times[1].tv_sec;
828
829 return utime (filename, &buf);
830 }
831 else
832 return utime (filename, 0);
833 }
834
835 static int eio__futimes (int fd, const struct timeval tv[2])
836 {
837 errno = ENOSYS;
838 return -1;
839 }
840
841 #endif
842
843 #if !HAVE_FDATASYNC
844 # undef fdatasync
845 # define fdatasync(fd) fsync (fd)
846 #endif
847
848 /* sync_file_range always needs emulation */
849 int
850 eio__sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags)
851 {
852 #if HAVE_SYNC_FILE_RANGE
853 int res;
854
855 if (EIO_SYNC_FILE_RANGE_WAIT_BEFORE != SYNC_FILE_RANGE_WAIT_BEFORE
856 || EIO_SYNC_FILE_RANGE_WRITE != SYNC_FILE_RANGE_WRITE
857 || EIO_SYNC_FILE_RANGE_WAIT_AFTER != SYNC_FILE_RANGE_WAIT_AFTER)
858 {
859 flags = 0
860 | (flags & EIO_SYNC_FILE_RANGE_WAIT_BEFORE ? SYNC_FILE_RANGE_WAIT_BEFORE : 0)
861 | (flags & EIO_SYNC_FILE_RANGE_WRITE ? SYNC_FILE_RANGE_WRITE : 0)
862 | (flags & EIO_SYNC_FILE_RANGE_WAIT_AFTER ? SYNC_FILE_RANGE_WAIT_AFTER : 0);
863 }
864
865 res = sync_file_range (fd, offset, nbytes, flags);
866
867 if (!res || errno != ENOSYS)
868 return res;
869 #endif
870
871 /* even though we could play tricks with the flags, it's better to always
872 * call fdatasync, as that matches the expectation of its users best */
873 return fdatasync (fd);
874 }
875
876 #if !HAVE_READAHEAD
877 # undef readahead
878 # define readahead(fd,offset,count) eio__readahead (fd, offset, count, self)
879
880 static ssize_t
881 eio__readahead (int fd, off_t offset, size_t count, etp_worker *self)
882 {
883 size_t todo = count;
884 dBUF;
885
886 while (todo > 0)
887 {
888 size_t len = todo < EIO_BUFSIZE ? todo : EIO_BUFSIZE;
889
890 pread (fd, eio_buf, len, offset);
891 offset += len;
892 todo -= len;
893 }
894
895 errno = 0;
896 return count;
897 }
898
899 #endif
900
901 /* sendfile always needs emulation */
902 static ssize_t
903 eio__sendfile (int ofd, int ifd, off_t offset, size_t count, etp_worker *self)
904 {
905 ssize_t res;
906
907 if (!count)
908 return 0;
909
910 #if HAVE_SENDFILE
911 # if __linux
912 res = sendfile (ofd, ifd, &offset, count);
913
914 # elif __freebsd
915 /*
916 * Of course, the freebsd sendfile is a dire hack with no thoughts
917 * wasted on making it similar to other I/O functions.
918 */
919 {
920 off_t sbytes;
921 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
922
923 if (res < 0 && sbytes)
924 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
925 res = sbytes;
926 }
927
928 # elif defined (__APPLE__)
929
930 {
931 off_t sbytes = count;
932 res = sendfile (ifd, ofd, offset, &sbytes, 0, 0);
933
934 if (res < 0 && errno == EAGAIN && sbytes)
935 res = sbytes;
936 }
937
938 # elif __hpux
939 res = sendfile (ofd, ifd, offset, count, 0, 0);
940
941 # elif __solaris
942 {
943 struct sendfilevec vec;
944 size_t sbytes;
945
946 vec.sfv_fd = ifd;
947 vec.sfv_flag = 0;
948 vec.sfv_off = offset;
949 vec.sfv_len = count;
950
951 res = sendfilev (ofd, &vec, 1, &sbytes);
952
953 if (res < 0 && sbytes)
954 res = sbytes;
955 }
956
957 # endif
958
959 #elif defined (_WIN32)
960
961 /* does not work, just for documentation of what would need to be done */
962 {
963 HANDLE h = TO_SOCKET (ifd);
964 SetFilePointer (h, offset, 0, FILE_BEGIN);
965 res = TransmitFile (TO_SOCKET (ofd), h, count, 0, 0, 0, 0);
966 }
967
968 #else
969 res = -1;
970 errno = ENOSYS;
971 #endif
972
973 if (res < 0
974 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
975 /* BSDs */
976 #ifdef ENOTSUP /* sigh, if the steenking pile called openbsd would only try to at least compile posix code... */
977 || errno == ENOTSUP
978 #endif
979 || errno == EOPNOTSUPP /* BSDs */
980 #if __solaris
981 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
982 #endif
983 )
984 )
985 {
986 /* emulate sendfile. this is a major pain in the ass */
987 dBUF;
988
989 res = 0;
990
991 while (count)
992 {
993 ssize_t cnt;
994
995 cnt = pread (ifd, eio_buf, count > EIO_BUFSIZE ? EIO_BUFSIZE : count, offset);
996
997 if (cnt <= 0)
998 {
999 if (cnt && !res) res = -1;
1000 break;
1001 }
1002
1003 cnt = write (ofd, eio_buf, cnt);
1004
1005 if (cnt <= 0)
1006 {
1007 if (cnt && !res) res = -1;
1008 break;
1009 }
1010
1011 offset += cnt;
1012 res += cnt;
1013 count -= cnt;
1014 }
1015 }
1016
1017 return res;
1018 }
1019
1020 static signed char
1021 eio_dent_cmp (const eio_dirent *a, const eio_dirent *b)
1022 {
1023 return a->score - b->score ? a->score - b->score /* works because our signed char is always 0..100 */
1024 : a->inode < b->inode ? -1 : a->inode > b->inode ? 1 : 0;
1025 }
1026
1027 #define EIO_DENT_CMP(i,op,j) eio_dent_cmp (&i, &j) op 0
1028
1029 #define EIO_SORT_CUTOFF 30 /* quite high, but performs well on many filesystems */
1030 #define EIO_SORT_FAST 60 /* when to only use insertion sort */
1031
1032 static void
1033 eio_dent_radix_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1034 {
1035 unsigned char bits [9 + sizeof (ino_t) * 8];
1036 unsigned char *bit = bits;
1037
1038 assert (CHAR_BIT == 8);
1039 assert (sizeof (eio_dirent) * 8 < 256);
1040 assert (offsetof (eio_dirent, inode)); /* we use 0 as sentinel */
1041 assert (offsetof (eio_dirent, score)); /* we use 0 as sentinel */
1042
1043 if (size <= EIO_SORT_FAST)
1044 return;
1045
1046 /* first prepare an array of bits to test in our radix sort */
1047 /* try to take endianness into account, as well as differences in ino_t sizes */
1048 /* inode_bits must contain all inodes ORed together */
1049 /* which is used to skip bits that are 0 everywhere, which is very common */
1050 {
1051 ino_t endianness;
1052 int i, j;
1053
1054 /* we store the byte offset of byte n into byte n of "endianness" */
1055 for (i = 0; i < sizeof (ino_t); ++i)
1056 ((unsigned char *)&endianness)[i] = i;
1057
1058 *bit++ = 0;
1059
1060 for (i = 0; i < sizeof (ino_t); ++i)
1061 {
1062 /* shifting off the byte offsets out of "endianness" */
1063 int offs = (offsetof (eio_dirent, inode) + (endianness & 0xff)) * 8;
1064 endianness >>= 8;
1065
1066 for (j = 0; j < 8; ++j)
1067 if (inode_bits & (((ino_t)1) << (i * 8 + j)))
1068 *bit++ = offs + j;
1069 }
1070
1071 for (j = 0; j < 8; ++j)
1072 if (score_bits & (1 << j))
1073 *bit++ = offsetof (eio_dirent, score) * 8 + j;
1074 }
1075
1076 /* now actually do the sorting (a variant of MSD radix sort) */
1077 {
1078 eio_dirent *base_stk [9 + sizeof (ino_t) * 8], *base;
1079 eio_dirent *end_stk [9 + sizeof (ino_t) * 8], *end;
1080 unsigned char *bit_stk [9 + sizeof (ino_t) * 8];
1081 int stk_idx = 0;
1082
1083 base_stk [stk_idx] = dents;
1084 end_stk [stk_idx] = dents + size;
1085 bit_stk [stk_idx] = bit - 1;
1086
1087 do
1088 {
1089 base = base_stk [stk_idx];
1090 end = end_stk [stk_idx];
1091 bit = bit_stk [stk_idx];
1092
1093 for (;;)
1094 {
1095 unsigned char O = *bit >> 3;
1096 unsigned char M = 1 << (*bit & 7);
1097
1098 eio_dirent *a = base;
1099 eio_dirent *b = end;
1100
1101 if (b - a < EIO_SORT_CUTOFF)
1102 break;
1103
1104 /* now bit-partition the array on the bit */
1105 /* this ugly asymmetric loop seems to perform much better than typical */
1106 /* partition algos found in the literature */
1107 do
1108 if (!(((unsigned char *)a)[O] & M))
1109 ++a;
1110 else if (!(((unsigned char *)--b)[O] & M))
1111 {
1112 eio_dirent tmp = *a; *a = *b; *b = tmp;
1113 ++a;
1114 }
1115 while (b > a);
1116
1117 /* next bit, or stop, if no bits left in this path */
1118 if (!*--bit)
1119 break;
1120
1121 base_stk [stk_idx] = a;
1122 end_stk [stk_idx] = end;
1123 bit_stk [stk_idx] = bit;
1124 ++stk_idx;
1125
1126 end = a;
1127 }
1128 }
1129 while (stk_idx--);
1130 }
1131 }
1132
1133 static void
1134 eio_dent_insertion_sort (eio_dirent *dents, int size)
1135 {
1136 /* first move the smallest element to the front, to act as a sentinel */
1137 {
1138 int i;
1139 eio_dirent *min = dents;
1140
1141 /* the radix pre-pass ensures that the minimum element is in the first EIO_SORT_CUTOFF + 1 elements */
1142 for (i = size > EIO_SORT_FAST ? EIO_SORT_CUTOFF + 1 : size; --i; )
1143 if (EIO_DENT_CMP (dents [i], <, *min))
1144 min = &dents [i];
1145
1146 /* swap elements 0 and j (minimum) */
1147 {
1148 eio_dirent tmp = *dents; *dents = *min; *min = tmp;
1149 }
1150 }
1151
1152 /* then do standard insertion sort, assuming that all elements are >= dents [0] */
1153 {
1154 eio_dirent *i, *j;
1155
1156 for (i = dents + 1; i < dents + size; ++i)
1157 {
1158 eio_dirent value = *i;
1159
1160 for (j = i - 1; EIO_DENT_CMP (*j, >, value); --j)
1161 j [1] = j [0];
1162
1163 j [1] = value;
1164 }
1165 }
1166 }
1167
1168 static void
1169 eio_dent_sort (eio_dirent *dents, int size, signed char score_bits, ino_t inode_bits)
1170 {
1171 if (size <= 1)
1172 return; /* our insertion sort relies on size > 0 */
1173
1174 /* first we use a radix sort, but only for dirs >= EIO_SORT_FAST */
1175 /* and stop sorting when the partitions are <= EIO_SORT_CUTOFF */
1176 eio_dent_radix_sort (dents, size, score_bits, inode_bits);
1177
1178 /* use an insertion sort at the end, or for small arrays, */
1179 /* as insertion sort is more efficient for small partitions */
1180 eio_dent_insertion_sort (dents, size);
1181 }
1182
1183 /* read a full directory */
1184 static void
1185 eio__scandir (eio_req *req, etp_worker *self)
1186 {
1187 DIR *dirp;
1188 EIO_STRUCT_DIRENT *entp;
1189 char *name, *names;
1190 int namesalloc = 4096;
1191 int namesoffs = 0;
1192 int flags = req->int1;
1193 eio_dirent *dents = 0;
1194 int dentalloc = 128;
1195 int dentoffs = 0;
1196 ino_t inode_bits = 0;
1197
1198 req->result = -1;
1199
1200 if (!(flags & EIO_READDIR_DENTS))
1201 flags &= ~(EIO_READDIR_DIRS_FIRST | EIO_READDIR_STAT_ORDER);
1202
1203 X_LOCK (wrklock);
1204 /* the corresponding closedir is in ETP_WORKER_CLEAR */
1205 self->dirp = dirp = opendir (req->ptr1);
1206 req->flags |= EIO_FLAG_PTR1_FREE | EIO_FLAG_PTR2_FREE;
1207 req->ptr1 = dents = flags ? malloc (dentalloc * sizeof (eio_dirent)) : 0;
1208 req->ptr2 = names = malloc (namesalloc);
1209 X_UNLOCK (wrklock);
1210
1211 if (dirp && names && (!flags || dents))
1212 for (;;)
1213 {
1214 errno = 0;
1215 entp = readdir (dirp);
1216
1217 if (!entp)
1218 {
1219 if (errno)
1220 break;
1221
1222 /* sort etc. */
1223 req->int1 = flags;
1224 req->result = dentoffs;
1225
1226 if (flags & EIO_READDIR_STAT_ORDER)
1227 eio_dent_sort (dents, dentoffs, 0, inode_bits); /* sort by inode exclusively */
1228 else if (flags & EIO_READDIR_DIRS_FIRST)
1229 if (flags & EIO_READDIR_FOUND_UNKNOWN)
1230 eio_dent_sort (dents, dentoffs, 7, inode_bits); /* sort by score and inode */
1231 else
1232 {
1233 /* in this case, all is known, and we just put dirs first and sort them */
1234 eio_dirent *oth = dents + dentoffs;
1235 eio_dirent *dir = dents;
1236
1237 /* now partition dirs to the front, and non-dirs to the back */
1238 /* by walking from both sides and swapping if necessary */
1239 /* also clear score, so it doesn't influence sorting */
1240 while (oth > dir)
1241 {
1242 if (dir->type == EIO_DT_DIR)
1243 ++dir;
1244 else if ((--oth)->type == EIO_DT_DIR)
1245 {
1246 eio_dirent tmp = *dir; *dir = *oth; *oth = tmp;
1247
1248 ++dir;
1249 }
1250 }
1251
1252 /* now sort the dirs only */
1253 eio_dent_sort (dents, dir - dents, 0, inode_bits);
1254 }
1255
1256 break;
1257 }
1258
1259 /* now add the entry to our list(s) */
1260 name = entp->d_name;
1261
1262 /* skip . and .. entries */
1263 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
1264 {
1265 int len = D_NAMLEN (entp) + 1;
1266
1267 while (expect_false (namesoffs + len > namesalloc))
1268 {
1269 namesalloc *= 2;
1270 X_LOCK (wrklock);
1271 req->ptr2 = names = realloc (names, namesalloc);
1272 X_UNLOCK (wrklock);
1273
1274 if (!names)
1275 break;
1276 }
1277
1278 memcpy (names + namesoffs, name, len);
1279
1280 if (dents)
1281 {
1282 struct eio_dirent *ent;
1283
1284 if (expect_false (dentoffs == dentalloc))
1285 {
1286 dentalloc *= 2;
1287 X_LOCK (wrklock);
1288 req->ptr1 = dents = realloc (dents, dentalloc * sizeof (eio_dirent));
1289 X_UNLOCK (wrklock);
1290
1291 if (!dents)
1292 break;
1293 }
1294
1295 ent = dents + dentoffs;
1296
1297 ent->nameofs = namesoffs; /* rather dirtily we store the offset in the pointer */
1298 ent->namelen = len - 1;
1299 ent->inode = D_INO (entp);
1300
1301 inode_bits |= ent->inode;
1302
1303 switch (D_TYPE (entp))
1304 {
1305 default:
1306 ent->type = EIO_DT_UNKNOWN;
1307 flags |= EIO_READDIR_FOUND_UNKNOWN;
1308 break;
1309
1310 #ifdef DT_FIFO
1311 case DT_FIFO: ent->type = EIO_DT_FIFO; break;
1312 #endif
1313 #ifdef DT_CHR
1314 case DT_CHR: ent->type = EIO_DT_CHR; break;
1315 #endif
1316 #ifdef DT_MPC
1317 case DT_MPC: ent->type = EIO_DT_MPC; break;
1318 #endif
1319 #ifdef DT_DIR
1320 case DT_DIR: ent->type = EIO_DT_DIR; break;
1321 #endif
1322 #ifdef DT_NAM
1323 case DT_NAM: ent->type = EIO_DT_NAM; break;
1324 #endif
1325 #ifdef DT_BLK
1326 case DT_BLK: ent->type = EIO_DT_BLK; break;
1327 #endif
1328 #ifdef DT_MPB
1329 case DT_MPB: ent->type = EIO_DT_MPB; break;
1330 #endif
1331 #ifdef DT_REG
1332 case DT_REG: ent->type = EIO_DT_REG; break;
1333 #endif
1334 #ifdef DT_NWK
1335 case DT_NWK: ent->type = EIO_DT_NWK; break;
1336 #endif
1337 #ifdef DT_CMP
1338 case DT_CMP: ent->type = EIO_DT_CMP; break;
1339 #endif
1340 #ifdef DT_LNK
1341 case DT_LNK: ent->type = EIO_DT_LNK; break;
1342 #endif
1343 #ifdef DT_SOCK
1344 case DT_SOCK: ent->type = EIO_DT_SOCK; break;
1345 #endif
1346 #ifdef DT_DOOR
1347 case DT_DOOR: ent->type = EIO_DT_DOOR; break;
1348 #endif
1349 #ifdef DT_WHT
1350 case DT_WHT: ent->type = EIO_DT_WHT; break;
1351 #endif
1352 }
1353
1354 ent->score = 7;
1355
1356 if (flags & EIO_READDIR_DIRS_FIRST)
1357 {
1358 if (ent->type == EIO_DT_UNKNOWN)
1359 {
1360 if (*name == '.') /* leading dots are likely directories, and, in any case, rare */
1361 ent->score = 1;
1362 else if (!strchr (name, '.')) /* absense of dots indicate likely dirs */
1363 ent->score = len <= 2 ? 4 - len : len <= 4 ? 4 : len <= 7 ? 5 : 6; /* shorter == more likely dir, but avoid too many classes */
1364 }
1365 else if (ent->type == EIO_DT_DIR)
1366 ent->score = 0;
1367 }
1368 }
1369
1370 namesoffs += len;
1371 ++dentoffs;
1372 }
1373
1374 if (EIO_CANCELLED (req))
1375 {
1376 errno = ECANCELED;
1377 break;
1378 }
1379 }
1380 }
1381
1382 #if !(_POSIX_MAPPED_FILES && _POSIX_SYNCHRONIZED_IO)
1383 # define eio__msync(a,b,c) ((errno = ENOSYS), -1)
1384 #else
1385
1386 int
1387 eio__msync (void *mem, size_t len, int flags)
1388 {
1389 if (EIO_MS_ASYNC != MS_SYNC
1390 || EIO_MS_INVALIDATE != MS_INVALIDATE
1391 || EIO_MS_SYNC != MS_SYNC)
1392 {
1393 flags = 0
1394 | (flags & EIO_MS_ASYNC ? MS_ASYNC : 0)
1395 | (flags & EIO_MS_INVALIDATE ? MS_INVALIDATE : 0)
1396 | (flags & EIO_MS_SYNC ? MS_SYNC : 0);
1397 }
1398
1399 return msync (mem, len, flags);
1400 }
1401
1402 #endif
1403
1404 int
1405 eio__mtouch (void *mem, size_t len, int flags)
1406 {
1407 intptr_t addr = (intptr_t)mem;
1408 intptr_t end = addr + len;
1409 #ifdef PAGESIZE
1410 const intptr_t page = PAGESIZE;
1411 #else
1412 static intptr_t page;
1413
1414 if (!page)
1415 page = sysconf (_SC_PAGESIZE);
1416 #endif
1417
1418 /* round down to start of page, although this is probably useless */
1419 addr &= ~(page - 1); /* assume page size is always a power of two */
1420
1421 if (addr < end)
1422 if (flags & EIO_MT_MODIFY) /* modify */
1423 do { *((volatile sig_atomic_t *)addr) |= 0; } while ((addr += page) < len);
1424 else
1425 do { *((volatile sig_atomic_t *)addr) ; } while ((addr += page) < len);
1426
1427 return 0;
1428 }
1429
1430 /*****************************************************************************/
1431
1432 #define ALLOC(len) \
1433 if (!req->ptr2) \
1434 { \
1435 X_LOCK (wrklock); \
1436 req->flags |= EIO_FLAG_PTR2_FREE; \
1437 X_UNLOCK (wrklock); \
1438 req->ptr2 = malloc (len); \
1439 if (!req->ptr2) \
1440 { \
1441 errno = ENOMEM; \
1442 req->result = -1; \
1443 break; \
1444 } \
1445 }
1446
1447 X_THREAD_PROC (etp_proc)
1448 {
1449 ETP_REQ *req;
1450 struct timespec ts;
1451 etp_worker *self = (etp_worker *)thr_arg;
1452
1453 /* try to distribute timeouts somewhat randomly */
1454 ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
1455
1456 for (;;)
1457 {
1458 X_LOCK (reqlock);
1459
1460 for (;;)
1461 {
1462 self->req = req = reqq_shift (&req_queue);
1463
1464 if (req)
1465 break;
1466
1467 ++idle;
1468
1469 ts.tv_sec = time (0) + IDLE_TIMEOUT;
1470 if (X_COND_TIMEDWAIT (reqwait, reqlock, ts) == ETIMEDOUT)
1471 {
1472 if (idle > max_idle)
1473 {
1474 --idle;
1475 X_UNLOCK (reqlock);
1476 X_LOCK (wrklock);
1477 --started;
1478 X_UNLOCK (wrklock);
1479 goto quit;
1480 }
1481
1482 /* we are allowed to idle, so do so without any timeout */
1483 X_COND_WAIT (reqwait, reqlock);
1484 }
1485
1486 --idle;
1487 }
1488
1489 --nready;
1490
1491 X_UNLOCK (reqlock);
1492
1493 if (req->type < 0)
1494 goto quit;
1495
1496 if (!EIO_CANCELLED (req))
1497 ETP_EXECUTE (self, req);
1498
1499 X_LOCK (reslock);
1500
1501 ++npending;
1502
1503 if (!reqq_push (&res_queue, req) && want_poll_cb)
1504 want_poll_cb ();
1505
1506 self->req = 0;
1507 etp_worker_clear (self);
1508
1509 X_UNLOCK (reslock);
1510 }
1511
1512 quit:
1513 X_LOCK (wrklock);
1514 etp_worker_free (self);
1515 X_UNLOCK (wrklock);
1516
1517 return 0;
1518 }
1519
1520 /*****************************************************************************/
1521
1522 int eio_init (void (*want_poll)(void), void (*done_poll)(void))
1523 {
1524 return etp_init (want_poll, done_poll);
1525 }
1526
1527 static void eio_api_destroy (eio_req *req)
1528 {
1529 free (req);
1530 }
1531
1532 #define REQ(rtype) \
1533 eio_req *req; \
1534 \
1535 req = (eio_req *)calloc (1, sizeof *req); \
1536 if (!req) \
1537 return 0; \
1538 \
1539 req->type = rtype; \
1540 req->pri = pri; \
1541 req->finish = cb; \
1542 req->data = data; \
1543 req->destroy = eio_api_destroy;
1544
1545 #define SEND eio_submit (req); return req
1546
1547 #define PATH \
1548 req->flags |= EIO_FLAG_PTR1_FREE; \
1549 req->ptr1 = strdup (path); \
1550 if (!req->ptr1) \
1551 { \
1552 eio_api_destroy (req); \
1553 return 0; \
1554 }
1555
1556 static void eio_execute (etp_worker *self, eio_req *req)
1557 {
1558 errno = 0;
1559
1560 switch (req->type)
1561 {
1562 case EIO_READ: ALLOC (req->size);
1563 req->result = req->offs >= 0
1564 ? pread (req->int1, req->ptr2, req->size, req->offs)
1565 : read (req->int1, req->ptr2, req->size); break;
1566 case EIO_WRITE: req->result = req->offs >= 0
1567 ? pwrite (req->int1, req->ptr2, req->size, req->offs)
1568 : write (req->int1, req->ptr2, req->size); break;
1569
1570 case EIO_READAHEAD: req->result = readahead (req->int1, req->offs, req->size); break;
1571 case EIO_SENDFILE: req->result = eio__sendfile (req->int1, req->int2, req->offs, req->size, self); break;
1572
1573 case EIO_STAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1574 req->result = stat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1575 case EIO_LSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1576 req->result = lstat (req->ptr1, (EIO_STRUCT_STAT *)req->ptr2); break;
1577 case EIO_FSTAT: ALLOC (sizeof (EIO_STRUCT_STAT));
1578 req->result = fstat (req->int1, (EIO_STRUCT_STAT *)req->ptr2); break;
1579
1580 case EIO_CHOWN: req->result = chown (req->ptr1, req->int2, req->int3); break;
1581 case EIO_FCHOWN: req->result = fchown (req->int1, req->int2, req->int3); break;
1582 case EIO_CHMOD: req->result = chmod (req->ptr1, (mode_t)req->int2); break;
1583 case EIO_FCHMOD: req->result = fchmod (req->int1, (mode_t)req->int2); break;
1584 case EIO_TRUNCATE: req->result = truncate (req->ptr1, req->offs); break;
1585 case EIO_FTRUNCATE: req->result = ftruncate (req->int1, req->offs); break;
1586
1587 case EIO_OPEN: req->result = open (req->ptr1, req->int1, (mode_t)req->int2); break;
1588 case EIO_CLOSE: req->result = close (req->int1); break;
1589 case EIO_DUP2: req->result = dup2 (req->int1, req->int2); break;
1590 case EIO_UNLINK: req->result = unlink (req->ptr1); break;
1591 case EIO_RMDIR: req->result = rmdir (req->ptr1); break;
1592 case EIO_MKDIR: req->result = mkdir (req->ptr1, (mode_t)req->int2); break;
1593 case EIO_RENAME: req->result = rename (req->ptr1, req->ptr2); break;
1594 case EIO_LINK: req->result = link (req->ptr1, req->ptr2); break;
1595 case EIO_SYMLINK: req->result = symlink (req->ptr1, req->ptr2); break;
1596 case EIO_MKNOD: req->result = mknod (req->ptr1, (mode_t)req->int2, (dev_t)req->int3); break;
1597
1598 case EIO_READLINK: ALLOC (NAME_MAX);
1599 req->result = readlink (req->ptr1, req->ptr2, NAME_MAX); break;
1600
1601 case EIO_SYNC: req->result = 0; sync (); break;
1602 case EIO_FSYNC: req->result = fsync (req->int1); break;
1603 case EIO_FDATASYNC: req->result = fdatasync (req->int1); break;
1604 case EIO_MSYNC: req->result = eio__msync (req->ptr2, req->size, req->int1); break;
1605 case EIO_MTOUCH: req->result = eio__mtouch (req->ptr2, req->size, req->int1); break;
1606 case EIO_SYNC_FILE_RANGE: req->result = eio__sync_file_range (req->int1, req->offs, req->size, req->int2); break;
1607
1608 case EIO_READDIR: eio__scandir (req, self); break;
1609
1610 case EIO_BUSY:
1611 #ifdef _WIN32
1612 Sleep (req->nv1 * 1000.);
1613 #else
1614 {
1615 struct timeval tv;
1616
1617 tv.tv_sec = req->nv1;
1618 tv.tv_usec = (req->nv1 - tv.tv_sec) * 1000000.;
1619
1620 req->result = select (0, 0, 0, 0, &tv);
1621 }
1622 #endif
1623 break;
1624
1625 case EIO_UTIME:
1626 case EIO_FUTIME:
1627 {
1628 struct timeval tv[2];
1629 struct timeval *times;
1630
1631 if (req->nv1 != -1. || req->nv2 != -1.)
1632 {
1633 tv[0].tv_sec = req->nv1;
1634 tv[0].tv_usec = (req->nv1 - tv[0].tv_sec) * 1000000.;
1635 tv[1].tv_sec = req->nv2;
1636 tv[1].tv_usec = (req->nv2 - tv[1].tv_sec) * 1000000.;
1637
1638 times = tv;
1639 }
1640 else
1641 times = 0;
1642
1643
1644 req->result = req->type == EIO_FUTIME
1645 ? futimes (req->int1, times)
1646 : utimes (req->ptr1, times);
1647 }
1648 break;
1649
1650 case EIO_GROUP:
1651 abort (); /* handled in eio_request */
1652
1653 case EIO_NOP:
1654 req->result = 0;
1655 break;
1656
1657 case EIO_CUSTOM:
1658 ((void (*)(eio_req *))req->feed) (req);
1659 break;
1660
1661 default:
1662 req->result = -1;
1663 break;
1664 }
1665
1666 req->errorno = errno;
1667 }
1668
1669 #ifndef EIO_NO_WRAPPERS
1670
1671 eio_req *eio_nop (int pri, eio_cb cb, void *data)
1672 {
1673 REQ (EIO_NOP); SEND;
1674 }
1675
1676 eio_req *eio_busy (double delay, int pri, eio_cb cb, void *data)
1677 {
1678 REQ (EIO_BUSY); req->nv1 = delay; SEND;
1679 }
1680
1681 eio_req *eio_sync (int pri, eio_cb cb, void *data)
1682 {
1683 REQ (EIO_SYNC); SEND;
1684 }
1685
1686 eio_req *eio_fsync (int fd, int pri, eio_cb cb, void *data)
1687 {
1688 REQ (EIO_FSYNC); req->int1 = fd; SEND;
1689 }
1690
1691 eio_req *eio_msync (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1692 {
1693 REQ (EIO_MSYNC); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1694 }
1695
1696 eio_req *eio_mtouch (void *addr, size_t length, int flags, int pri, eio_cb cb, void *data)
1697 {
1698 REQ (EIO_MTOUCH); req->ptr2 = addr; req->size = length; req->int1 = flags; SEND;
1699 }
1700
1701 eio_req *eio_sync_file_range (int fd, off_t offset, size_t nbytes, unsigned int flags, int pri, eio_cb cb, void *data)
1702 {
1703 REQ (EIO_SYNC_FILE_RANGE); req->int1 = fd; req->offs = offset; req->size = nbytes; req->int2 = flags; SEND;
1704 }
1705
1706 eio_req *eio_fdatasync (int fd, int pri, eio_cb cb, void *data)
1707 {
1708 REQ (EIO_FDATASYNC); req->int1 = fd; SEND;
1709 }
1710
1711 eio_req *eio_close (int fd, int pri, eio_cb cb, void *data)
1712 {
1713 REQ (EIO_CLOSE); req->int1 = fd; SEND;
1714 }
1715
1716 eio_req *eio_readahead (int fd, off_t offset, size_t length, int pri, eio_cb cb, void *data)
1717 {
1718 REQ (EIO_READAHEAD); req->int1 = fd; req->offs = offset; req->size = length; SEND;
1719 }
1720
1721 eio_req *eio_read (int fd, void *buf, size_t length, off_t offset, int pri, eio_cb cb, void *data)
1722 {
1723 REQ (EIO_READ); req->int1 = fd; req->offs = offset; req->size = length; req->ptr2 = buf; SEND;
1724 }
1725
1726 eio_req *eio_write (int fd, void *buf, size_t length, off_t offset, int pri, eio_cb cb, void *data)
1727 {
1728 REQ (EIO_WRITE); req->int1 = fd; req->offs = offset; req->size = length; req->ptr2 = buf; SEND;
1729 }
1730
1731 eio_req *eio_fstat (int fd, int pri, eio_cb cb, void *data)
1732 {
1733 REQ (EIO_FSTAT); req->int1 = fd; SEND;
1734 }
1735
1736 eio_req *eio_futime (int fd, double atime, double mtime, int pri, eio_cb cb, void *data)
1737 {
1738 REQ (EIO_FUTIME); req->int1 = fd; req->nv1 = atime; req->nv2 = mtime; SEND;
1739 }
1740
1741 eio_req *eio_ftruncate (int fd, off_t offset, int pri, eio_cb cb, void *data)
1742 {
1743 REQ (EIO_FTRUNCATE); req->int1 = fd; req->offs = offset; SEND;
1744 }
1745
1746 eio_req *eio_fchmod (int fd, mode_t mode, int pri, eio_cb cb, void *data)
1747 {
1748 REQ (EIO_FCHMOD); req->int1 = fd; req->int2 = (long)mode; SEND;
1749 }
1750
1751 eio_req *eio_fchown (int fd, uid_t uid, gid_t gid, int pri, eio_cb cb, void *data)
1752 {
1753 REQ (EIO_FCHOWN); req->int1 = fd; req->int2 = (long)uid; req->int3 = (long)gid; SEND;
1754 }
1755
1756 eio_req *eio_dup2 (int fd, int fd2, int pri, eio_cb cb, void *data)
1757 {
1758 REQ (EIO_DUP2); req->int1 = fd; req->int2 = fd2; SEND;
1759 }
1760
1761 eio_req *eio_sendfile (int out_fd, int in_fd, off_t in_offset, size_t length, int pri, eio_cb cb, void *data)
1762 {
1763 REQ (EIO_SENDFILE); req->int1 = out_fd; req->int2 = in_fd; req->offs = in_offset; req->size = length; SEND;
1764 }
1765
1766 eio_req *eio_open (const char *path, int flags, mode_t mode, int pri, eio_cb cb, void *data)
1767 {
1768 REQ (EIO_OPEN); PATH; req->int1 = flags; req->int2 = (long)mode; SEND;
1769 }
1770
1771 eio_req *eio_utime (const char *path, double atime, double mtime, int pri, eio_cb cb, void *data)
1772 {
1773 REQ (EIO_UTIME); PATH; req->nv1 = atime; req->nv2 = mtime; SEND;
1774 }
1775
1776 eio_req *eio_truncate (const char *path, off_t offset, int pri, eio_cb cb, void *data)
1777 {
1778 REQ (EIO_TRUNCATE); PATH; req->offs = offset; SEND;
1779 }
1780
1781 eio_req *eio_chown (const char *path, uid_t uid, gid_t gid, int pri, eio_cb cb, void *data)
1782 {
1783 REQ (EIO_CHOWN); PATH; req->int2 = (long)uid; req->int3 = (long)gid; SEND;
1784 }
1785
1786 eio_req *eio_chmod (const char *path, mode_t mode, int pri, eio_cb cb, void *data)
1787 {
1788 REQ (EIO_CHMOD); PATH; req->int2 = (long)mode; SEND;
1789 }
1790
1791 eio_req *eio_mkdir (const char *path, mode_t mode, int pri, eio_cb cb, void *data)
1792 {
1793 REQ (EIO_MKDIR); PATH; req->int2 = (long)mode; SEND;
1794 }
1795
1796 static eio_req *
1797 eio__1path (int type, const char *path, int pri, eio_cb cb, void *data)
1798 {
1799 REQ (type); PATH; SEND;
1800 }
1801
1802 eio_req *eio_readlink (const char *path, int pri, eio_cb cb, void *data)
1803 {
1804 return eio__1path (EIO_READLINK, path, pri, cb, data);
1805 }
1806
1807 eio_req *eio_stat (const char *path, int pri, eio_cb cb, void *data)
1808 {
1809 return eio__1path (EIO_STAT, path, pri, cb, data);
1810 }
1811
1812 eio_req *eio_lstat (const char *path, int pri, eio_cb cb, void *data)
1813 {
1814 return eio__1path (EIO_LSTAT, path, pri, cb, data);
1815 }
1816
1817 eio_req *eio_unlink (const char *path, int pri, eio_cb cb, void *data)
1818 {
1819 return eio__1path (EIO_UNLINK, path, pri, cb, data);
1820 }
1821
1822 eio_req *eio_rmdir (const char *path, int pri, eio_cb cb, void *data)
1823 {
1824 return eio__1path (EIO_RMDIR, path, pri, cb, data);
1825 }
1826
1827 eio_req *eio_readdir (const char *path, int flags, int pri, eio_cb cb, void *data)
1828 {
1829 REQ (EIO_READDIR); PATH; req->int1 = flags; SEND;
1830 }
1831
1832 eio_req *eio_mknod (const char *path, mode_t mode, dev_t dev, int pri, eio_cb cb, void *data)
1833 {
1834 REQ (EIO_MKNOD); PATH; req->int2 = (long)mode; req->int3 = (long)dev; SEND;
1835 }
1836
1837 static eio_req *
1838 eio__2path (int type, const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1839 {
1840 REQ (type); PATH;
1841
1842 req->flags |= EIO_FLAG_PTR2_FREE;
1843 req->ptr2 = strdup (new_path);
1844 if (!req->ptr2)
1845 {
1846 eio_api_destroy (req);
1847 return 0;
1848 }
1849
1850 SEND;
1851 }
1852
1853 eio_req *eio_link (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1854 {
1855 return eio__2path (EIO_LINK, path, new_path, pri, cb, data);
1856 }
1857
1858 eio_req *eio_symlink (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1859 {
1860 return eio__2path (EIO_SYMLINK, path, new_path, pri, cb, data);
1861 }
1862
1863 eio_req *eio_rename (const char *path, const char *new_path, int pri, eio_cb cb, void *data)
1864 {
1865 return eio__2path (EIO_RENAME, path, new_path, pri, cb, data);
1866 }
1867
1868 eio_req *eio_custom (eio_cb execute, int pri, eio_cb cb, void *data)
1869 {
1870 REQ (EIO_CUSTOM); req->feed = (void (*)(eio_req *))execute; SEND;
1871 }
1872
1873 #endif
1874
1875 eio_req *eio_grp (eio_cb cb, void *data)
1876 {
1877 const int pri = EIO_PRI_MAX;
1878
1879 REQ (EIO_GROUP); SEND;
1880 }
1881
1882 #undef REQ
1883 #undef PATH
1884 #undef SEND
1885
1886 /*****************************************************************************/
1887 /* grp functions */
1888
1889 void eio_grp_feed (eio_req *grp, void (*feed)(eio_req *req), int limit)
1890 {
1891 grp->int2 = limit;
1892 grp->feed = feed;
1893
1894 grp_try_feed (grp);
1895 }
1896
1897 void eio_grp_limit (eio_req *grp, int limit)
1898 {
1899 grp->int2 = limit;
1900
1901 grp_try_feed (grp);
1902 }
1903
1904 void eio_grp_add (eio_req *grp, eio_req *req)
1905 {
1906 assert (("cannot add requests to IO::AIO::GRP after the group finished", grp->int1 != 2));
1907
1908 grp->flags |= EIO_FLAG_GROUPADD;
1909
1910 ++grp->size;
1911 req->grp = grp;
1912
1913 req->grp_prev = 0;
1914 req->grp_next = grp->grp_first;
1915
1916 if (grp->grp_first)
1917 grp->grp_first->grp_prev = req;
1918
1919 grp->grp_first = req;
1920 }
1921
1922 /*****************************************************************************/
1923 /* misc garbage */
1924
1925 ssize_t eio_sendfile_sync (int ofd, int ifd, off_t offset, size_t count)
1926 {
1927 etp_worker wrk;
1928 ssize_t ret;
1929
1930 wrk.dbuf = 0;
1931
1932 ret = eio__sendfile (ofd, ifd, offset, count, &wrk);
1933
1934 if (wrk.dbuf)
1935 free (wrk.dbuf);
1936
1937 return ret;
1938 }
1939