ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/IO-AIO/AIO.xs
Revision: 1.78
Committed: Thu Oct 26 14:35:34 2006 UTC (17 years, 6 months ago) by root
Branch: MAIN
Changes since 1.77: +62 -44 lines
Log Message:
*** empty log message ***

File Contents

# Content
1 /* solaris */
2 #define _POSIX_PTHREAD_SEMANTICS 1
3
4 #if __linux && !defined(_GNU_SOURCE)
5 # define _GNU_SOURCE
6 #endif
7
8 /* just in case */
9 #define _REENTRANT 1
10
11 #include <errno.h>
12
13 #include "EXTERN.h"
14 #include "perl.h"
15 #include "XSUB.h"
16
17 #include "autoconf/config.h"
18
19 #include <pthread.h>
20
21 #include <stddef.h>
22 #include <errno.h>
23 #include <sys/time.h>
24 #include <sys/select.h>
25 #include <sys/types.h>
26 #include <sys/stat.h>
27 #include <limits.h>
28 #include <unistd.h>
29 #include <fcntl.h>
30 #include <signal.h>
31 #include <sched.h>
32
33 #if HAVE_SENDFILE
34 # if __linux
35 # include <sys/sendfile.h>
36 # elif __freebsd
37 # include <sys/socket.h>
38 # include <sys/uio.h>
39 # elif __hpux
40 # include <sys/socket.h>
41 # elif __solaris /* not yet */
42 # include <sys/sendfile.h>
43 # else
44 # error sendfile support requested but not available
45 # endif
46 #endif
47
48 /* used for struct dirent, AIX doesn't provide it */
49 #ifndef NAME_MAX
50 # define NAME_MAX 4096
51 #endif
52
53 #ifndef PTHREAD_STACK_MIN
54 /* care for broken platforms, e.g. windows */
55 # define PTHREAD_STACK_MIN 16384
56 #endif
57
58 #if __ia64
59 # define STACKSIZE 65536
60 #elif __i386 || __x86_64 /* 16k is unreasonably high :( */
61 # define STACKSIZE PTHREAD_STACK_MIN
62 #else
63 # define STACKSIZE 16384
64 #endif
65
66 /* buffer size for various temporary buffers */
67 #define AIO_BUFSIZE 65536
68
69 #define dBUF \
70 char *aio_buf; \
71 LOCK (wrklock); \
72 self->dbuf = aio_buf = malloc (AIO_BUFSIZE); \
73 UNLOCK (wrklock); \
74 if (!aio_buf) \
75 return -1;
76
77 enum {
78 REQ_QUIT,
79 REQ_OPEN, REQ_CLOSE,
80 REQ_READ, REQ_WRITE, REQ_READAHEAD,
81 REQ_SENDFILE,
82 REQ_STAT, REQ_LSTAT, REQ_FSTAT,
83 REQ_FSYNC, REQ_FDATASYNC,
84 REQ_UNLINK, REQ_RMDIR, REQ_RENAME,
85 REQ_READDIR,
86 REQ_LINK, REQ_SYMLINK,
87 REQ_GROUP, REQ_NOP,
88 REQ_BUSY,
89 };
90
91 #define AIO_REQ_KLASS "IO::AIO::REQ"
92 #define AIO_GRP_KLASS "IO::AIO::GRP"
93
94 typedef struct aio_cb
95 {
96 struct aio_cb *volatile next;
97
98 SV *data, *callback;
99 SV *fh, *fh2;
100 void *dataptr, *data2ptr;
101 Stat_t *statdata;
102 off_t offset;
103 size_t length;
104 ssize_t result;
105
106 STRLEN dataoffset;
107 int type;
108 int fd, fd2;
109 int errorno;
110 mode_t mode; /* open */
111
112 unsigned char flags;
113 unsigned char pri;
114
115 SV *self; /* the perl counterpart of this request, if any */
116 struct aio_cb *grp, *grp_prev, *grp_next, *grp_first;
117 } aio_cb;
118
119 enum {
120 FLAG_CANCELLED = 0x01,
121 };
122
123 typedef aio_cb *aio_req;
124 typedef aio_cb *aio_req_ornot;
125
126 enum {
127 PRI_MIN = -4,
128 PRI_MAX = 4,
129
130 DEFAULT_PRI = 0,
131 PRI_BIAS = -PRI_MIN,
132 NUM_PRI = PRI_MAX + PRI_BIAS + 1,
133 };
134
135 static int next_pri = DEFAULT_PRI + PRI_BIAS;
136
137 static unsigned int started, wanted;
138 static volatile unsigned int nreqs;
139 static volatile unsigned int max_outstanding = 0xffffffff;
140 static int respipe [2];
141
142 #if __linux && defined (PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP)
143 # define AIO_MUTEX_INIT PTHREAD_ADAPTIVE_MUTEX_INITIALIZER_NP
144 #else
145 # define AIO_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER
146 #endif
147
148 #define LOCK(mutex) pthread_mutex_lock (&(mutex))
149 #define UNLOCK(mutex) pthread_mutex_unlock (&(mutex))
150
151 /* worker threads management */
152 static pthread_mutex_t wrklock = AIO_MUTEX_INIT;
153
154 typedef struct worker {
155 /* locked by wrklock */
156 struct worker *prev, *next;
157
158 pthread_t tid;
159
160 /* locked by reslock, reqlock or wrklock */
161 aio_req req; /* currently processed request */
162 void *dbuf;
163 DIR *dirp;
164 } worker;
165
166 static worker wrk_first = { &wrk_first, &wrk_first, 0 };
167
168 static void worker_clear (worker *wrk)
169 {
170 if (wrk->dirp)
171 {
172 closedir (wrk->dirp);
173 wrk->dirp = 0;
174 }
175
176 if (wrk->dbuf)
177 {
178 free (wrk->dbuf);
179 wrk->dbuf = 0;
180 }
181 }
182
183 static void worker_free (worker *wrk)
184 {
185 wrk->next->prev = wrk->prev;
186 wrk->prev->next = wrk->next;
187
188 free (wrk);
189 }
190
191 static pthread_mutex_t reslock = AIO_MUTEX_INIT;
192 static pthread_mutex_t reqlock = AIO_MUTEX_INIT;
193 static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER;
194
195 /*
196 * a somewhat faster data structure might be nice, but
197 * with 8 priorities this actually needs <20 insns
198 * per shift, the most expensive operation.
199 */
200 typedef struct {
201 aio_req qs[NUM_PRI], qe[NUM_PRI]; /* qstart, qend */
202 int size;
203 } reqq;
204
205 static reqq req_queue;
206 static reqq res_queue;
207
208 int reqq_push (reqq *q, aio_req req)
209 {
210 int pri = req->pri;
211 req->next = 0;
212
213 if (q->qe[pri])
214 {
215 q->qe[pri]->next = req;
216 q->qe[pri] = req;
217 }
218 else
219 q->qe[pri] = q->qs[pri] = req;
220
221 return q->size++;
222 }
223
224 aio_req reqq_shift (reqq *q)
225 {
226 int pri;
227
228 if (!q->size)
229 return 0;
230
231 --q->size;
232
233 for (pri = NUM_PRI; pri--; )
234 {
235 aio_req req = q->qs[pri];
236
237 if (req)
238 {
239 if (!(q->qs[pri] = req->next))
240 q->qe[pri] = 0;
241
242 return req;
243 }
244 }
245
246 abort ();
247 }
248
249 static int poll_cb (int max);
250 static void req_invoke (aio_req req);
251 static void req_free (aio_req req);
252 static void req_cancel (aio_req req);
253
254 /* must be called at most once */
255 static SV *req_sv (aio_req req, const char *klass)
256 {
257 if (!req->self)
258 {
259 req->self = (SV *)newHV ();
260 sv_magic (req->self, 0, PERL_MAGIC_ext, (char *)req, 0);
261 }
262
263 return sv_2mortal (sv_bless (newRV_inc (req->self), gv_stashpv (klass, 1)));
264 }
265
266 static aio_req SvAIO_REQ (SV *sv)
267 {
268 MAGIC *mg;
269
270 if (!sv_derived_from (sv, AIO_REQ_KLASS) || !SvROK (sv))
271 croak ("object of class " AIO_REQ_KLASS " expected");
272
273 mg = mg_find (SvRV (sv), PERL_MAGIC_ext);
274
275 return mg ? (aio_req)mg->mg_ptr : 0;
276 }
277
278 static void aio_grp_feed (aio_req grp)
279 {
280 while (grp->length < grp->fd2 && !(grp->flags & FLAG_CANCELLED))
281 {
282 int old_len = grp->length;
283
284 if (grp->fh2 && SvOK (grp->fh2))
285 {
286 dSP;
287
288 ENTER;
289 SAVETMPS;
290 PUSHMARK (SP);
291 XPUSHs (req_sv (grp, AIO_GRP_KLASS));
292 PUTBACK;
293 call_sv (grp->fh2, G_VOID | G_EVAL | G_KEEPERR);
294 SPAGAIN;
295 FREETMPS;
296 LEAVE;
297 }
298
299 /* stop if no progress has been made */
300 if (old_len == grp->length)
301 {
302 SvREFCNT_dec (grp->fh2);
303 grp->fh2 = 0;
304 break;
305 }
306 }
307 }
308
309 static void aio_grp_dec (aio_req grp)
310 {
311 --grp->length;
312
313 /* call feeder, if applicable */
314 aio_grp_feed (grp);
315
316 /* finish, if done */
317 if (!grp->length && grp->fd)
318 {
319 req_invoke (grp);
320 req_free (grp);
321 }
322 }
323
324 static void poll_wait ()
325 {
326 fd_set rfd;
327
328 while (nreqs)
329 {
330 int size;
331 #if !(__i386 || __x86_64) /* safe without sempahore on these archs */
332 LOCK (reslock);
333 #endif
334 size = res_queue.size;
335 #if !(__i386 || __x86_64) /* safe without sempahore on these archs */
336 UNLOCK (reslock);
337 #endif
338
339 if (size)
340 return;
341
342 FD_ZERO(&rfd);
343 FD_SET(respipe [0], &rfd);
344
345 select (respipe [0] + 1, &rfd, 0, 0, 0);
346 }
347 }
348
349 static void req_invoke (aio_req req)
350 {
351 dSP;
352
353 if (!(req->flags & FLAG_CANCELLED) && SvOK (req->callback))
354 {
355 errno = req->errorno;
356
357 ENTER;
358 SAVETMPS;
359 PUSHMARK (SP);
360 EXTEND (SP, 1);
361
362 switch (req->type)
363 {
364 case REQ_READDIR:
365 {
366 SV *rv = &PL_sv_undef;
367
368 if (req->result >= 0)
369 {
370 int i;
371 char *buf = req->data2ptr;
372 AV *av = newAV ();
373
374 av_extend (av, req->result - 1);
375
376 for (i = 0; i < req->result; ++i)
377 {
378 SV *sv = newSVpv (buf, 0);
379
380 av_store (av, i, sv);
381 buf += SvCUR (sv) + 1;
382 }
383
384 rv = sv_2mortal (newRV_noinc ((SV *)av));
385 }
386
387 PUSHs (rv);
388 }
389 break;
390
391 case REQ_OPEN:
392 {
393 /* convert fd to fh */
394 SV *fh;
395
396 PUSHs (sv_2mortal (newSViv (req->result)));
397 PUTBACK;
398 call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL);
399 SPAGAIN;
400
401 fh = SvREFCNT_inc (POPs);
402
403 PUSHMARK (SP);
404 XPUSHs (sv_2mortal (fh));
405 }
406 break;
407
408 case REQ_GROUP:
409 req->fd = 2; /* mark group as finished */
410
411 if (req->data)
412 {
413 int i;
414 AV *av = (AV *)req->data;
415
416 EXTEND (SP, AvFILL (av) + 1);
417 for (i = 0; i <= AvFILL (av); ++i)
418 PUSHs (*av_fetch (av, i, 0));
419 }
420 break;
421
422 case REQ_NOP:
423 case REQ_BUSY:
424 break;
425
426 default:
427 PUSHs (sv_2mortal (newSViv (req->result)));
428 break;
429 }
430
431
432 PUTBACK;
433 call_sv (req->callback, G_VOID | G_EVAL);
434 SPAGAIN;
435
436 FREETMPS;
437 LEAVE;
438 }
439
440 if (req->grp)
441 {
442 aio_req grp = req->grp;
443
444 /* unlink request */
445 if (req->grp_next) req->grp_next->grp_prev = req->grp_prev;
446 if (req->grp_prev) req->grp_prev->grp_next = req->grp_next;
447
448 if (grp->grp_first == req)
449 grp->grp_first = req->grp_next;
450
451 aio_grp_dec (grp);
452 }
453
454 if (SvTRUE (ERRSV))
455 {
456 req_free (req);
457 croak (0);
458 }
459 }
460
461 static void req_free (aio_req req)
462 {
463 if (req->self)
464 {
465 sv_unmagic (req->self, PERL_MAGIC_ext);
466 SvREFCNT_dec (req->self);
467 }
468
469 SvREFCNT_dec (req->data);
470 SvREFCNT_dec (req->fh);
471 SvREFCNT_dec (req->fh2);
472 SvREFCNT_dec (req->callback);
473 Safefree (req->statdata);
474
475 if (req->type == REQ_READDIR)
476 free (req->data2ptr);
477
478 Safefree (req);
479 }
480
481 static void req_cancel_subs (aio_req grp)
482 {
483 aio_req sub;
484
485 if (grp->type != REQ_GROUP)
486 return;
487
488 SvREFCNT_dec (grp->fh2);
489 grp->fh2 = 0;
490
491 for (sub = grp->grp_first; sub; sub = sub->grp_next)
492 req_cancel (sub);
493 }
494
495 static void req_cancel (aio_req req)
496 {
497 req->flags |= FLAG_CANCELLED;
498
499 req_cancel_subs (req);
500 }
501
502 static int poll_cb (int max)
503 {
504 dSP;
505 int count = 0;
506 int do_croak = 0;
507 aio_req req;
508
509 for (;;)
510 {
511 while (max <= 0 || count < max)
512 {
513 LOCK (reslock);
514 req = reqq_shift (&res_queue);
515
516 if (req)
517 {
518 if (!res_queue.size)
519 {
520 /* read any signals sent by the worker threads */
521 char buf [32];
522 while (read (respipe [0], buf, 32) == 32)
523 ;
524 }
525 }
526
527 UNLOCK (reslock);
528
529 if (!req)
530 break;
531
532 --nreqs;
533
534 if (req->type == REQ_QUIT)
535 --started;
536 else if (req->type == REQ_GROUP && req->length)
537 {
538 req->fd = 1; /* mark request as delayed */
539 continue;
540 }
541 else
542 {
543 if (req->type == REQ_READ)
544 SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0));
545
546 if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE))
547 SvREADONLY_off (req->data);
548
549 if (req->statdata)
550 {
551 PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT;
552 PL_laststatval = req->result;
553 PL_statcache = *(req->statdata);
554 }
555
556 req_invoke (req);
557
558 count++;
559 }
560
561 req_free (req);
562 }
563
564 if (nreqs <= max_outstanding)
565 break;
566
567 poll_wait ();
568
569 max = 0;
570 }
571
572 return count;
573 }
574
575 static void *aio_proc(void *arg);
576
577 static void start_thread (void)
578 {
579 sigset_t fullsigset, oldsigset;
580 pthread_attr_t attr;
581
582 worker *wrk = calloc (1, sizeof (worker));
583
584 if (!wrk)
585 croak ("unable to allocate worker thread data");
586
587 pthread_attr_init (&attr);
588 pthread_attr_setstacksize (&attr, STACKSIZE);
589 pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
590
591 sigfillset (&fullsigset);
592
593 LOCK (wrklock);
594 sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);
595
596 if (pthread_create (&wrk->tid, &attr, aio_proc, (void *)wrk) == 0)
597 {
598 wrk->prev = &wrk_first;
599 wrk->next = wrk_first.next;
600 wrk_first.next->prev = wrk;
601 wrk_first.next = wrk;
602 ++started;
603 }
604 else
605 free (wrk);
606
607 sigprocmask (SIG_SETMASK, &oldsigset, 0);
608 UNLOCK (wrklock);
609 }
610
611 static void req_send (aio_req req)
612 {
613 while (started < wanted && nreqs >= started)
614 start_thread ();
615
616 ++nreqs;
617
618 LOCK (reqlock);
619 reqq_push (&req_queue, req);
620 pthread_cond_signal (&reqwait);
621 UNLOCK (reqlock);
622 }
623
624 static void end_thread (void)
625 {
626 aio_req req;
627
628 Newz (0, req, 1, aio_cb);
629
630 req->type = REQ_QUIT;
631 req->pri = PRI_MAX + PRI_BIAS;
632
633 req_send (req);
634 }
635
636 static void min_parallel (int nthreads)
637 {
638 if (wanted < nthreads)
639 wanted = nthreads;
640 }
641
642 static void max_parallel (int nthreads)
643 {
644 int cur = started;
645
646 if (wanted > nthreads)
647 wanted = nthreads;
648
649 while (cur > wanted)
650 {
651 end_thread ();
652 cur--;
653 }
654
655 while (started > wanted)
656 {
657 poll_wait ();
658 poll_cb (0);
659 }
660 }
661
662 static void create_pipe ()
663 {
664 if (pipe (respipe))
665 croak ("unable to initialize result pipe");
666
667 if (fcntl (respipe [0], F_SETFL, O_NONBLOCK))
668 croak ("cannot set result pipe to nonblocking mode");
669
670 if (fcntl (respipe [1], F_SETFL, O_NONBLOCK))
671 croak ("cannot set result pipe to nonblocking mode");
672 }
673
674 /*****************************************************************************/
675 /* work around various missing functions */
676
677 #if !HAVE_PREADWRITE
678 # define pread aio_pread
679 # define pwrite aio_pwrite
680
681 /*
682 * make our pread/pwrite safe against themselves, but not against
683 * normal read/write by using a mutex. slows down execution a lot,
684 * but that's your problem, not mine.
685 */
686 static pthread_mutex_t preadwritelock = PTHREAD_MUTEX_INITIALIZER;
687
688 static ssize_t pread (int fd, void *buf, size_t count, off_t offset)
689 {
690 ssize_t res;
691 off_t ooffset;
692
693 LOCK (preadwritelock);
694 ooffset = lseek (fd, 0, SEEK_CUR);
695 lseek (fd, offset, SEEK_SET);
696 res = read (fd, buf, count);
697 lseek (fd, ooffset, SEEK_SET);
698 UNLOCK (preadwritelock);
699
700 return res;
701 }
702
703 static ssize_t pwrite (int fd, void *buf, size_t count, off_t offset)
704 {
705 ssize_t res;
706 off_t ooffset;
707
708 LOCK (preadwritelock);
709 ooffset = lseek (fd, 0, SEEK_CUR);
710 lseek (fd, offset, SEEK_SET);
711 res = write (fd, buf, count);
712 lseek (fd, offset, SEEK_SET);
713 UNLOCK (preadwritelock);
714
715 return res;
716 }
717 #endif
718
719 #if !HAVE_FDATASYNC
720 # define fdatasync fsync
721 #endif
722
723 #if !HAVE_READAHEAD
724 # define readahead(fd,offset,count) aio_readahead (fd, offset, count, self)
725
726 static ssize_t aio_readahead (int fd, off_t offset, size_t count, worker *self)
727 {
728 dBUF;
729
730 while (count > 0)
731 {
732 size_t len = count < AIO_BUFSIZE ? count : AIO_BUFSIZE;
733
734 pread (fd, aio_buf, len, offset);
735 offset += len;
736 count -= len;
737 }
738
739 errno = 0;
740 }
741
742 #endif
743
744 #if !HAVE_READDIR_R
745 # define readdir_r aio_readdir_r
746
747 static pthread_mutex_t readdirlock = PTHREAD_MUTEX_INITIALIZER;
748
749 static int readdir_r (DIR *dirp, struct dirent *ent, struct dirent **res)
750 {
751 struct dirent *e;
752 int errorno;
753
754 LOCK (readdirlock);
755
756 e = readdir (dirp);
757 errorno = errno;
758
759 if (e)
760 {
761 *res = ent;
762 strcpy (ent->d_name, e->d_name);
763 }
764 else
765 *res = 0;
766
767 UNLOCK (readdirlock);
768
769 errno = errorno;
770 return e ? 0 : -1;
771 }
772 #endif
773
774 /* sendfile always needs emulation */
775 static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count, worker *self)
776 {
777 ssize_t res;
778
779 if (!count)
780 return 0;
781
782 #if HAVE_SENDFILE
783 # if __linux
784 res = sendfile (ofd, ifd, &offset, count);
785
786 # elif __freebsd
787 /*
788 * Of course, the freebsd sendfile is a dire hack with no thoughts
789 * wasted on making it similar to other I/O functions.
790 */
791 {
792 off_t sbytes;
793 res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
794
795 if (res < 0 && sbytes)
796 /* maybe only on EAGAIN: as usual, the manpage leaves you guessing */
797 res = sbytes;
798 }
799
800 # elif __hpux
801 res = sendfile (ofd, ifd, offset, count, 0, 0);
802
803 # elif __solaris
804 {
805 struct sendfilevec vec;
806 size_t sbytes;
807
808 vec.sfv_fd = ifd;
809 vec.sfv_flag = 0;
810 vec.sfv_off = offset;
811 vec.sfv_len = count;
812
813 res = sendfilev (ofd, &vec, 1, &sbytes);
814
815 if (res < 0 && sbytes)
816 res = sbytes;
817 }
818
819 # endif
820 #else
821 res = -1;
822 errno = ENOSYS;
823 #endif
824
825 if (res < 0
826 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK
827 #if __solaris
828 || errno == EAFNOSUPPORT || errno == EPROTOTYPE
829 #endif
830 )
831 )
832 {
833 /* emulate sendfile. this is a major pain in the ass */
834 dBUF;
835
836 res = 0;
837
838 while (count)
839 {
840 ssize_t cnt;
841
842 cnt = pread (ifd, aio_buf, count > AIO_BUFSIZE ? AIO_BUFSIZE : count, offset);
843
844 if (cnt <= 0)
845 {
846 if (cnt && !res) res = -1;
847 break;
848 }
849
850 cnt = write (ofd, aio_buf, cnt);
851
852 if (cnt <= 0)
853 {
854 if (cnt && !res) res = -1;
855 break;
856 }
857
858 offset += cnt;
859 res += cnt;
860 count -= cnt;
861 }
862 }
863
864 return res;
865 }
866
867 /* read a full directory */
868 static void scandir_ (aio_req req, worker *self)
869 {
870 DIR *dirp;
871 union
872 {
873 struct dirent d;
874 char b [offsetof (struct dirent, d_name) + NAME_MAX + 1];
875 } *u;
876 struct dirent *entp;
877 char *name, *names;
878 int memlen = 4096;
879 int memofs = 0;
880 int res = 0;
881 int errorno;
882
883 LOCK (wrklock);
884 self->dirp = dirp = opendir (req->dataptr);
885 self->dbuf = u = malloc (sizeof (*u));
886 req->data2ptr = names = malloc (memlen);
887 UNLOCK (wrklock);
888
889 if (dirp && u && names)
890 for (;;)
891 {
892 errno = 0;
893 readdir_r (dirp, &u->d, &entp);
894
895 if (!entp)
896 break;
897
898 name = entp->d_name;
899
900 if (name [0] != '.' || (name [1] && (name [1] != '.' || name [2])))
901 {
902 int len = strlen (name) + 1;
903
904 res++;
905
906 while (memofs + len > memlen)
907 {
908 memlen *= 2;
909 LOCK (wrklock);
910 req->data2ptr = names = realloc (names, memlen);
911 UNLOCK (wrklock);
912
913 if (!names)
914 break;
915 }
916
917 memcpy (names + memofs, name, len);
918 memofs += len;
919 }
920 }
921
922 if (errno)
923 res = -1;
924
925 req->result = res;
926 }
927
928 /*****************************************************************************/
929
930 static void *aio_proc (void *thr_arg)
931 {
932 aio_req req;
933 int type;
934 worker *self = (worker *)thr_arg;
935
936 do
937 {
938 LOCK (reqlock);
939
940 for (;;)
941 {
942 self->req = req = reqq_shift (&req_queue);
943
944 if (req)
945 break;
946
947 pthread_cond_wait (&reqwait, &reqlock);
948 }
949
950 UNLOCK (reqlock);
951
952 errno = 0; /* strictly unnecessary */
953 type = req->type; /* remember type for QUIT check */
954
955 if (!(req->flags & FLAG_CANCELLED))
956 switch (type)
957 {
958 case REQ_READ: req->result = pread (req->fd, req->dataptr, req->length, req->offset); break;
959 case REQ_WRITE: req->result = pwrite (req->fd, req->dataptr, req->length, req->offset); break;
960
961 case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break;
962 case REQ_SENDFILE: req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length, self); break;
963
964 case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break;
965 case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break;
966 case REQ_FSTAT: req->result = fstat (req->fd , req->statdata); break;
967
968 case REQ_OPEN: req->result = open (req->dataptr, req->fd, req->mode); break;
969 case REQ_CLOSE: req->result = close (req->fd); break;
970 case REQ_UNLINK: req->result = unlink (req->dataptr); break;
971 case REQ_RMDIR: req->result = rmdir (req->dataptr); break;
972 case REQ_RENAME: req->result = rename (req->data2ptr, req->dataptr); break;
973 case REQ_LINK: req->result = link (req->data2ptr, req->dataptr); break;
974 case REQ_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break;
975
976 case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
977 case REQ_FSYNC: req->result = fsync (req->fd); break;
978 case REQ_READDIR: scandir_ (req, self); break;
979
980 case REQ_BUSY:
981 {
982 struct timeval tv;
983
984 tv.tv_sec = req->fd;
985 tv.tv_usec = req->fd2;
986
987 req->result = select (0, 0, 0, 0, &tv);
988 }
989
990 case REQ_GROUP:
991 case REQ_NOP:
992 case REQ_QUIT:
993 break;
994
995 default:
996 req->result = ENOSYS;
997 break;
998 }
999
1000 req->errorno = errno;
1001
1002 LOCK (reslock);
1003
1004 if (!reqq_push (&res_queue, req))
1005 /* write a dummy byte to the pipe so fh becomes ready */
1006 write (respipe [1], &respipe, 1);
1007
1008 self->req = 0;
1009 worker_clear (self);
1010
1011 UNLOCK (reslock);
1012 }
1013 while (type != REQ_QUIT);
1014
1015 LOCK (wrklock);
1016 worker_free (self);
1017 UNLOCK (wrklock);
1018
1019 return 0;
1020 }
1021
1022 /*****************************************************************************/
1023
1024 static void atfork_prepare (void)
1025 {
1026 LOCK (wrklock);
1027 LOCK (reqlock);
1028 LOCK (reslock);
1029 #if !HAVE_PREADWRITE
1030 LOCK (preadwritelock);
1031 #endif
1032 #if !HAVE_READDIR_R
1033 LOCK (readdirlock);
1034 #endif
1035 }
1036
1037 static void atfork_parent (void)
1038 {
1039 #if !HAVE_READDIR_R
1040 UNLOCK (readdirlock);
1041 #endif
1042 #if !HAVE_PREADWRITE
1043 UNLOCK (preadwritelock);
1044 #endif
1045 UNLOCK (reslock);
1046 UNLOCK (reqlock);
1047 UNLOCK (wrklock);
1048 }
1049
1050 static void atfork_child (void)
1051 {
1052 aio_req prv;
1053
1054 while (prv = reqq_shift (&req_queue))
1055 req_free (prv);
1056
1057 while (prv = reqq_shift (&res_queue))
1058 req_free (prv);
1059
1060 while (wrk_first.next != &wrk_first)
1061 {
1062 worker *wrk = wrk_first.next;
1063
1064 if (wrk->req)
1065 req_free (wrk->req);
1066
1067 worker_clear (wrk);
1068 worker_free (wrk);
1069 }
1070
1071 started = 0;
1072 nreqs = 0;
1073
1074 close (respipe [0]);
1075 close (respipe [1]);
1076 create_pipe ();
1077
1078 atfork_parent ();
1079 }
1080
1081 #define dREQ \
1082 aio_req req; \
1083 int req_pri = next_pri; \
1084 next_pri = DEFAULT_PRI + PRI_BIAS; \
1085 \
1086 if (SvOK (callback) && !SvROK (callback)) \
1087 croak ("callback must be undef or of reference type"); \
1088 \
1089 Newz (0, req, 1, aio_cb); \
1090 if (!req) \
1091 croak ("out of memory during aio_req allocation"); \
1092 \
1093 req->callback = newSVsv (callback); \
1094 req->pri = req_pri
1095
1096 #define REQ_SEND \
1097 req_send (req); \
1098 \
1099 if (GIMME_V != G_VOID) \
1100 XPUSHs (req_sv (req, AIO_REQ_KLASS));
1101
1102 MODULE = IO::AIO PACKAGE = IO::AIO
1103
1104 PROTOTYPES: ENABLE
1105
1106 BOOT:
1107 {
1108 HV *stash = gv_stashpv ("IO::AIO", 1);
1109 newCONSTSUB (stash, "EXDEV", newSViv (EXDEV));
1110 newCONSTSUB (stash, "O_RDONLY", newSViv (O_RDONLY));
1111 newCONSTSUB (stash, "O_WRONLY", newSViv (O_WRONLY));
1112
1113 create_pipe ();
1114 pthread_atfork (atfork_prepare, atfork_parent, atfork_child);
1115 }
1116
1117 void
1118 min_parallel (int nthreads)
1119 PROTOTYPE: $
1120
1121 void
1122 max_parallel (int nthreads)
1123 PROTOTYPE: $
1124
1125 int
1126 max_outstanding (int maxreqs)
1127 PROTOTYPE: $
1128 CODE:
1129 RETVAL = max_outstanding;
1130 max_outstanding = maxreqs;
1131 OUTPUT:
1132 RETVAL
1133
1134 void
1135 aio_open (pathname,flags,mode,callback=&PL_sv_undef)
1136 SV * pathname
1137 int flags
1138 int mode
1139 SV * callback
1140 PROTOTYPE: $$$;$
1141 PPCODE:
1142 {
1143 dREQ;
1144
1145 req->type = REQ_OPEN;
1146 req->data = newSVsv (pathname);
1147 req->dataptr = SvPVbyte_nolen (req->data);
1148 req->fd = flags;
1149 req->mode = mode;
1150
1151 REQ_SEND;
1152 }
1153
1154 void
1155 aio_close (fh,callback=&PL_sv_undef)
1156 SV * fh
1157 SV * callback
1158 PROTOTYPE: $;$
1159 ALIAS:
1160 aio_close = REQ_CLOSE
1161 aio_fsync = REQ_FSYNC
1162 aio_fdatasync = REQ_FDATASYNC
1163 PPCODE:
1164 {
1165 dREQ;
1166
1167 req->type = ix;
1168 req->fh = newSVsv (fh);
1169 req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
1170
1171 REQ_SEND (req);
1172 }
1173
1174 void
1175 aio_read (fh,offset,length,data,dataoffset,callback=&PL_sv_undef)
1176 SV * fh
1177 UV offset
1178 UV length
1179 SV * data
1180 UV dataoffset
1181 SV * callback
1182 ALIAS:
1183 aio_read = REQ_READ
1184 aio_write = REQ_WRITE
1185 PROTOTYPE: $$$$$;$
1186 PPCODE:
1187 {
1188 aio_req req;
1189 STRLEN svlen;
1190 char *svptr = SvPVbyte (data, svlen);
1191
1192 SvUPGRADE (data, SVt_PV);
1193 SvPOK_on (data);
1194
1195 if (dataoffset < 0)
1196 dataoffset += svlen;
1197
1198 if (dataoffset < 0 || dataoffset > svlen)
1199 croak ("data offset outside of string");
1200
1201 if (ix == REQ_WRITE)
1202 {
1203 /* write: check length and adjust. */
1204 if (length < 0 || length + dataoffset > svlen)
1205 length = svlen - dataoffset;
1206 }
1207 else
1208 {
1209 /* read: grow scalar as necessary */
1210 svptr = SvGROW (data, length + dataoffset);
1211 }
1212
1213 if (length < 0)
1214 croak ("length must not be negative");
1215
1216 {
1217 dREQ;
1218
1219 req->type = ix;
1220 req->fh = newSVsv (fh);
1221 req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh))
1222 : IoOFP (sv_2io (fh)));
1223 req->offset = offset;
1224 req->length = length;
1225 req->data = SvREFCNT_inc (data);
1226 req->dataptr = (char *)svptr + dataoffset;
1227
1228 if (!SvREADONLY (data))
1229 {
1230 SvREADONLY_on (data);
1231 req->data2ptr = (void *)data;
1232 }
1233
1234 REQ_SEND;
1235 }
1236 }
1237
1238 void
1239 aio_sendfile (out_fh,in_fh,in_offset,length,callback=&PL_sv_undef)
1240 SV * out_fh
1241 SV * in_fh
1242 UV in_offset
1243 UV length
1244 SV * callback
1245 PROTOTYPE: $$$$;$
1246 PPCODE:
1247 {
1248 dREQ;
1249
1250 req->type = REQ_SENDFILE;
1251 req->fh = newSVsv (out_fh);
1252 req->fd = PerlIO_fileno (IoIFP (sv_2io (out_fh)));
1253 req->fh2 = newSVsv (in_fh);
1254 req->fd2 = PerlIO_fileno (IoIFP (sv_2io (in_fh)));
1255 req->offset = in_offset;
1256 req->length = length;
1257
1258 REQ_SEND;
1259 }
1260
1261 void
1262 aio_readahead (fh,offset,length,callback=&PL_sv_undef)
1263 SV * fh
1264 UV offset
1265 IV length
1266 SV * callback
1267 PROTOTYPE: $$$;$
1268 PPCODE:
1269 {
1270 dREQ;
1271
1272 req->type = REQ_READAHEAD;
1273 req->fh = newSVsv (fh);
1274 req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
1275 req->offset = offset;
1276 req->length = length;
1277
1278 REQ_SEND;
1279 }
1280
1281 void
1282 aio_stat (fh_or_path,callback=&PL_sv_undef)
1283 SV * fh_or_path
1284 SV * callback
1285 ALIAS:
1286 aio_stat = REQ_STAT
1287 aio_lstat = REQ_LSTAT
1288 PPCODE:
1289 {
1290 dREQ;
1291
1292 New (0, req->statdata, 1, Stat_t);
1293 if (!req->statdata)
1294 {
1295 req_free (req);
1296 croak ("out of memory during aio_req->statdata allocation");
1297 }
1298
1299 if (SvPOK (fh_or_path))
1300 {
1301 req->type = ix;
1302 req->data = newSVsv (fh_or_path);
1303 req->dataptr = SvPVbyte_nolen (req->data);
1304 }
1305 else
1306 {
1307 req->type = REQ_FSTAT;
1308 req->fh = newSVsv (fh_or_path);
1309 req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path)));
1310 }
1311
1312 REQ_SEND;
1313 }
1314
1315 void
1316 aio_unlink (pathname,callback=&PL_sv_undef)
1317 SV * pathname
1318 SV * callback
1319 ALIAS:
1320 aio_unlink = REQ_UNLINK
1321 aio_rmdir = REQ_RMDIR
1322 aio_readdir = REQ_READDIR
1323 PPCODE:
1324 {
1325 dREQ;
1326
1327 req->type = ix;
1328 req->data = newSVsv (pathname);
1329 req->dataptr = SvPVbyte_nolen (req->data);
1330
1331 REQ_SEND;
1332 }
1333
1334 void
1335 aio_link (oldpath,newpath,callback=&PL_sv_undef)
1336 SV * oldpath
1337 SV * newpath
1338 SV * callback
1339 ALIAS:
1340 aio_link = REQ_LINK
1341 aio_symlink = REQ_SYMLINK
1342 aio_rename = REQ_RENAME
1343 PPCODE:
1344 {
1345 dREQ;
1346
1347 req->type = ix;
1348 req->fh = newSVsv (oldpath);
1349 req->data2ptr = SvPVbyte_nolen (req->fh);
1350 req->data = newSVsv (newpath);
1351 req->dataptr = SvPVbyte_nolen (req->data);
1352
1353 REQ_SEND;
1354 }
1355
1356 void
1357 aio_busy (delay,callback=&PL_sv_undef)
1358 double delay
1359 SV * callback
1360 PPCODE:
1361 {
1362 dREQ;
1363
1364 req->type = REQ_BUSY;
1365 req->fd = delay < 0. ? 0 : delay;
1366 req->fd2 = delay < 0. ? 0 : 1000. * (delay - req->fd);
1367
1368 REQ_SEND;
1369 }
1370
1371 void
1372 aio_group (callback=&PL_sv_undef)
1373 SV * callback
1374 PROTOTYPE: ;$
1375 PPCODE:
1376 {
1377 dREQ;
1378
1379 req->type = REQ_GROUP;
1380 req_send (req);
1381
1382 XPUSHs (req_sv (req, AIO_GRP_KLASS));
1383 }
1384
1385 void
1386 aio_nop (callback=&PL_sv_undef)
1387 SV * callback
1388 PPCODE:
1389 {
1390 dREQ;
1391
1392 req->type = REQ_NOP;
1393
1394 REQ_SEND;
1395 }
1396
1397 void
1398 aioreq_pri (int pri = DEFAULT_PRI)
1399 CODE:
1400 if (pri < PRI_MIN) pri = PRI_MIN;
1401 if (pri > PRI_MAX) pri = PRI_MAX;
1402 next_pri = pri + PRI_BIAS;
1403
1404 void
1405 aioreq_nice (int nice = 0)
1406 CODE:
1407 nice = next_pri - nice;
1408 if (nice < PRI_MIN) nice = PRI_MIN;
1409 if (nice > PRI_MAX) nice = PRI_MAX;
1410 next_pri = nice + PRI_BIAS;
1411
1412 void
1413 flush ()
1414 PROTOTYPE:
1415 CODE:
1416 while (nreqs)
1417 {
1418 poll_wait ();
1419 poll_cb (0);
1420 }
1421
1422 void
1423 poll()
1424 PROTOTYPE:
1425 CODE:
1426 if (nreqs)
1427 {
1428 poll_wait ();
1429 poll_cb (0);
1430 }
1431
1432 int
1433 poll_fileno()
1434 PROTOTYPE:
1435 CODE:
1436 RETVAL = respipe [0];
1437 OUTPUT:
1438 RETVAL
1439
1440 int
1441 poll_cb(...)
1442 PROTOTYPE:
1443 CODE:
1444 RETVAL = poll_cb (0);
1445 OUTPUT:
1446 RETVAL
1447
1448 int
1449 poll_some(int max = 0)
1450 PROTOTYPE: $
1451 CODE:
1452 RETVAL = poll_cb (max);
1453 OUTPUT:
1454 RETVAL
1455
1456 void
1457 poll_wait()
1458 PROTOTYPE:
1459 CODE:
1460 if (nreqs)
1461 poll_wait ();
1462
1463 int
1464 nreqs()
1465 PROTOTYPE:
1466 CODE:
1467 RETVAL = nreqs;
1468 OUTPUT:
1469 RETVAL
1470
1471 PROTOTYPES: DISABLE
1472
1473 MODULE = IO::AIO PACKAGE = IO::AIO::REQ
1474
1475 void
1476 cancel (aio_req_ornot req)
1477 CODE:
1478 req_cancel (req);
1479
1480 void
1481 cb (aio_req_ornot req, SV *callback=&PL_sv_undef)
1482 CODE:
1483 SvREFCNT_dec (req->callback);
1484 req->callback = newSVsv (callback);
1485
1486 MODULE = IO::AIO PACKAGE = IO::AIO::GRP
1487
1488 void
1489 add (aio_req grp, ...)
1490 PPCODE:
1491 {
1492 int i;
1493 aio_req req;
1494
1495 if (grp->fd == 2)
1496 croak ("cannot add requests to IO::AIO::GRP after the group finished");
1497
1498 for (i = 1; i < items; ++i )
1499 {
1500 if (GIMME_V != G_VOID)
1501 XPUSHs (sv_2mortal (newSVsv (ST (i))));
1502
1503 req = SvAIO_REQ (ST (i));
1504
1505 if (req)
1506 {
1507 ++grp->length;
1508 req->grp = grp;
1509
1510 req->grp_prev = 0;
1511 req->grp_next = grp->grp_first;
1512
1513 if (grp->grp_first)
1514 grp->grp_first->grp_prev = req;
1515
1516 grp->grp_first = req;
1517 }
1518 }
1519 }
1520
1521 void
1522 cancel_subs (aio_req_ornot req)
1523 CODE:
1524 req_cancel_subs (req);
1525
1526 void
1527 result (aio_req grp, ...)
1528 CODE:
1529 {
1530 int i;
1531 AV *av = newAV ();
1532
1533 for (i = 1; i < items; ++i )
1534 av_push (av, newSVsv (ST (i)));
1535
1536 SvREFCNT_dec (grp->data);
1537 grp->data = (SV *)av;
1538 }
1539
1540 void
1541 limit (aio_req grp, int limit)
1542 CODE:
1543 grp->fd2 = limit;
1544 aio_grp_feed (grp);
1545
1546 void
1547 feed (aio_req grp, SV *callback=&PL_sv_undef)
1548 CODE:
1549 {
1550 SvREFCNT_dec (grp->fh2);
1551 grp->fh2 = newSVsv (callback);
1552
1553 if (grp->fd2 <= 0)
1554 grp->fd2 = 2;
1555
1556 aio_grp_feed (grp);
1557 }
1558