ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_linuxaio.c
Revision: 1.37
Committed: Wed Jun 26 07:20:09 2019 UTC (5 years ago) by root
Content type: text/plain
Branch: MAIN
CVS Tags: EV-rel-4_27, rel-4_27
Changes since 1.36: +65 -29 lines
Log Message:
*** empty log message ***

File Contents

# Content
1 /*
2 * libev linux aio fd activity backend
3 *
4 * Copyright (c) 2019 Marc Alexander Lehmann <libev@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 /*
41 * general notes about linux aio:
42 *
43 * a) at first, the linux aio IOCB_CMD_POLL functionality introduced in
44 * 4.18 looks too good to be true: both watchers and events can be
45 * batched, and events can even be handled in userspace using
46 * a ring buffer shared with the kernel. watchers can be canceled
47 * regardless of whether the fd has been closed. no problems with fork.
48 * ok, the ring buffer is 200% undocumented (there isn't even a
49 * header file), but otherwise, it's pure bliss!
50 * b) ok, watchers are one-shot, so you have to re-arm active ones
51 * on every iteration. so much for syscall-less event handling,
52 * but at least these re-arms can be batched, no big deal, right?
53 * c) well, linux as usual: the documentation lies to you: io_submit
54 * sometimes returns EINVAL because the kernel doesn't feel like
55 * handling your poll mask - ttys can be polled for POLLOUT,
56 * POLLOUT|POLLIN, but polling for POLLIN fails. just great,
57 * so we have to fall back to something else (hello, epoll),
58 * but at least the fallback can be slow, because these are
59 * exceptional cases, right?
60 * d) hmm, you have to tell the kernel the maximum number of watchers
61 * you want to queue when initialising the aio context. but of
62 * course the real limit is magically calculated in the kernel, and
63 * is often higher then we asked for. so we just have to destroy
64 * the aio context and re-create it a bit larger if we hit the limit.
65 * (starts to remind you of epoll? well, it's a bit more deterministic
66 * and less gambling, but still ugly as hell).
67 * e) that's when you find out you can also hit an arbitrary system-wide
68 * limit. or the kernel simply doesn't want to handle your watchers.
69 * what the fuck do we do then? you guessed it, in the middle
70 * of event handling we have to switch to 100% epoll polling. and
71 * that better is as fast as normal epoll polling, so you practically
72 * have to use the normal epoll backend with all its quirks.
73 * f) end result of this train wreck: it inherits all the disadvantages
74 * from epoll, while adding a number on its own. why even bother to use
75 * it? because if conditions are right and your fds are supported and you
76 * don't hit a limit, this backend is actually faster, doesn't gamble with
77 * your fds, batches watchers and events and doesn't require costly state
78 * recreates. well, until it does.
79 * g) all of this makes this backend use almost twice as much code as epoll.
80 * which in turn uses twice as much code as poll. and that#s not counting
81 * the fact that this backend also depends on the epoll backend, making
82 * it three times as much code as poll, or kqueue.
83 * h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now
84 * it's clear that whatever linux comes up with is far, far, far worse.
85 */
86
87 #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */
88 #include <poll.h>
89 #include <linux/aio_abi.h>
90
91 /*****************************************************************************/
92 /* syscall wrapdadoop - this section has the raw api/abi definitions */
93
94 #include <sys/syscall.h> /* no glibc wrappers */
95
96 /* aio_abi.h is not versioned in any way, so we cannot test for its existance */
97 #define IOCB_CMD_POLL 5
98
99 /* taken from linux/fs/aio.c. yup, that's a .c file.
100 * not only is this totally undocumented, not even the source code
101 * can tell you what the future semantics of compat_features and
102 * incompat_features are, or what header_length actually is for.
103 */
104 #define AIO_RING_MAGIC 0xa10a10a1
105 #define EV_AIO_RING_INCOMPAT_FEATURES 0
106 struct aio_ring
107 {
108 unsigned id; /* kernel internal index number */
109 unsigned nr; /* number of io_events */
110 unsigned head; /* Written to by userland or by kernel. */
111 unsigned tail;
112
113 unsigned magic;
114 unsigned compat_features;
115 unsigned incompat_features;
116 unsigned header_length; /* size of aio_ring */
117
118 struct io_event io_events[0];
119 };
120
121 /*
122 * define some syscall wrappers for common architectures
123 * this is mostly for nice looks during debugging, not performance.
124 * our syscalls return < 0, not == -1, on error. which is good
125 * enough for linux aio.
126 * TODO: arm is also common nowadays, maybe even mips and x86
127 * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove...
128 */
129 #if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__
130 /* the costly errno access probably kills this for size optimisation */
131
132 #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \
133 ({ \
134 long res; \
135 register unsigned long r5 __asm__ ("r8" ); \
136 register unsigned long r4 __asm__ ("r10"); \
137 register unsigned long r3 __asm__ ("rdx"); \
138 register unsigned long r2 __asm__ ("rsi"); \
139 register unsigned long r1 __asm__ ("rdi"); \
140 if (narg >= 5) r5 = (unsigned long)(arg5); \
141 if (narg >= 4) r4 = (unsigned long)(arg4); \
142 if (narg >= 3) r3 = (unsigned long)(arg3); \
143 if (narg >= 2) r2 = (unsigned long)(arg2); \
144 if (narg >= 1) r1 = (unsigned long)(arg1); \
145 __asm__ __volatile__ ( \
146 "syscall\n\t" \
147 : "=a" (res) \
148 : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \
149 : "cc", "r11", "cx", "memory"); \
150 errno = -res; \
151 res; \
152 })
153
154 #endif
155
156 #ifdef ev_syscall
157 #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0
158 #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0)
159 #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0)
160 #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0)
161 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0)
162 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5)
163 #else
164 #define ev_syscall0(nr) syscall (nr)
165 #define ev_syscall1(nr,arg1) syscall (nr, arg1)
166 #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2)
167 #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3)
168 #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4)
169 #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5)
170 #endif
171
172 inline_size
173 int
174 evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp)
175 {
176 return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp);
177 }
178
179 inline_size
180 int
181 evsys_io_destroy (aio_context_t ctx_id)
182 {
183 return ev_syscall1 (SYS_io_destroy, ctx_id);
184 }
185
186 inline_size
187 int
188 evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[])
189 {
190 return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp);
191 }
192
193 inline_size
194 int
195 evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result)
196 {
197 return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result);
198 }
199
200 inline_size
201 int
202 evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout)
203 {
204 return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout);
205 }
206
207 /*****************************************************************************/
208 /* actual backed implementation */
209
210 ecb_cold
211 static int
212 linuxaio_nr_events (EV_P)
213 {
214 /* we start with 16 iocbs and incraese from there
215 * that's tiny, but the kernel has a rather low system-wide
216 * limit that can be reached quickly, so let's be parsimonious
217 * with this resource.
218 * Rest assured, the kernel generously rounds up small and big numbers
219 * in different ways (but doesn't seem to charge you for it).
220 * The 15 here is because the kernel usually has a power of two as aio-max-nr,
221 * and this helps to take advantage of that limit.
222 */
223
224 /* we try to fill 4kB pages exactly.
225 * the ring buffer header is 32 bytes, every io event is 32 bytes.
226 * the kernel takes the io requests number, doubles it, adds 2
227 * and adds the ring buffer.
228 * the way we use this is by starting low, and then roughly doubling the
229 * size each time we hit a limit.
230 */
231
232 int requests = 15 << linuxaio_iteration;
233 int one_page = (4096
234 / sizeof (struct io_event) ) / 2; /* how many fit into one page */
235 int first_page = ((4096 - sizeof (struct aio_ring))
236 / sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */
237
238 /* if everything fits into one page, use count exactly */
239 if (requests > first_page)
240 /* otherwise, round down to full pages and add the first page */
241 requests = requests / one_page * one_page + first_page;
242
243 return requests;
244 }
245
246 /* we use out own wrapper structure in case we ever want to do something "clever" */
247 typedef struct aniocb
248 {
249 struct iocb io;
250 /*int inuse;*/
251 } *ANIOCBP;
252
253 inline_size
254 void
255 linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count)
256 {
257 while (count--)
258 {
259 /* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */
260 ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb));
261
262 /* full zero initialise is probably not required at the moment, but
263 * this is not well documented, so we better do it.
264 */
265 memset (iocb, 0, sizeof (*iocb));
266
267 iocb->io.aio_lio_opcode = IOCB_CMD_POLL;
268 iocb->io.aio_data = offset;
269 iocb->io.aio_fildes = offset;
270
271 base [offset++] = iocb;
272 }
273 }
274
275 ecb_cold
276 static void
277 linuxaio_free_iocbp (EV_P)
278 {
279 while (linuxaio_iocbpmax--)
280 ev_free (linuxaio_iocbps [linuxaio_iocbpmax]);
281
282 linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */
283 }
284
285 static void
286 linuxaio_modify (EV_P_ int fd, int oev, int nev)
287 {
288 array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp);
289 ANIOCBP iocb = linuxaio_iocbps [fd];
290
291 if (iocb->io.aio_reqprio < 0)
292 {
293 /* we handed this fd over to epoll, so undo this first */
294 /* we do it manually because the optimisations on epoll_modify won't do us any good */
295 epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0);
296 anfds [fd].emask = 0;
297 iocb->io.aio_reqprio = 0;
298 }
299
300 if (iocb->io.aio_buf)
301 {
302 evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0);
303 /* on relevant kernels, io_cancel fails with EINPROGRES if everything is fine */
304 assert (("libev: linuxaio unexpected io_cancel failed", errno == EINPROGRESS));
305 }
306
307 if (nev)
308 {
309 iocb->io.aio_buf =
310 (nev & EV_READ ? POLLIN : 0)
311 | (nev & EV_WRITE ? POLLOUT : 0);
312
313 /* queue iocb up for io_submit */
314 /* this assumes we only ever get one call per fd per loop iteration */
315 ++linuxaio_submitcnt;
316 array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit);
317 linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io;
318 }
319 }
320
321 static void
322 linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents)
323 {
324 epoll_poll (EV_A_ 0);
325 }
326
327 inline_speed
328 void
329 linuxaio_fd_rearm (EV_P_ int fd)
330 {
331 anfds [fd].events = 0;
332 linuxaio_iocbps [fd]->io.aio_buf = 0;
333 fd_change (EV_A_ fd, EV_ANFD_REIFY);
334 }
335
336 static void
337 linuxaio_parse_events (EV_P_ struct io_event *ev, int nr)
338 {
339 while (nr)
340 {
341 int fd = ev->data;
342 int res = ev->res;
343
344 assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax));
345
346 /* feed events, we do not expect or handle POLLNVAL */
347 fd_event (
348 EV_A_
349 fd,
350 (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
351 | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
352 );
353
354 /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */
355 linuxaio_fd_rearm (EV_A_ fd);
356
357 --nr;
358 ++ev;
359 }
360 }
361
362 /* get any events from ring buffer, return true if any were handled */
363 static int
364 linuxaio_get_events_from_ring (EV_P)
365 {
366 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
367
368 /* the kernel reads and writes both of these variables, */
369 /* as a C extension, we assume that volatile use here */
370 /* both makes reads atomic and once-only */
371 unsigned head = *(volatile unsigned *)&ring->head;
372 unsigned tail = *(volatile unsigned *)&ring->tail;
373
374 if (head == tail)
375 return 0;
376
377 /* make sure the events up to tail are visible */
378 ECB_MEMORY_FENCE_ACQUIRE;
379
380 /* parse all available events, but only once, to avoid starvation */
381 if (tail > head) /* normal case around */
382 linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head);
383 else /* wrapped around */
384 {
385 linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head);
386 linuxaio_parse_events (EV_A_ ring->io_events, tail);
387 }
388
389 ECB_MEMORY_FENCE_RELEASE;
390 /* as an extension to C, we hope that the volatile will make this atomic and once-only */
391 *(volatile unsigned *)&ring->head = tail;
392
393 return 1;
394 }
395
396 inline_size
397 int
398 linuxaio_ringbuf_valid (EV_P)
399 {
400 struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx;
401
402 return expect_true (ring->magic == AIO_RING_MAGIC)
403 && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES
404 && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */
405 }
406
407 /* read at least one event from kernel, or timeout */
408 inline_size
409 void
410 linuxaio_get_events (EV_P_ ev_tstamp timeout)
411 {
412 struct timespec ts;
413 struct io_event ioev[8]; /* 256 octet stack space */
414 int want = 1; /* how many events to request */
415 int ringbuf_valid = linuxaio_ringbuf_valid (EV_A);
416
417 if (expect_true (ringbuf_valid))
418 {
419 /* if the ring buffer has any events, we don't wait or call the kernel at all */
420 if (linuxaio_get_events_from_ring (EV_A))
421 return;
422
423 /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */
424 if (!timeout)
425 return;
426 }
427 else
428 /* no ringbuffer, request slightly larger batch */
429 want = sizeof (ioev) / sizeof (ioev [0]);
430
431 /* no events, so wait for some
432 * for fairness reasons, we do this in a loop, to fetch all events
433 */
434 for (;;)
435 {
436 int res;
437
438 EV_RELEASE_CB;
439
440 ts.tv_sec = (long)timeout;
441 ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9);
442
443 res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts);
444
445 EV_ACQUIRE_CB;
446
447 if (res < 0)
448 if (errno == EINTR)
449 /* ignored, retry */;
450 else
451 ev_syserr ("(libev) linuxaio io_getevents");
452 else if (res)
453 {
454 /* at least one event available, handle them */
455 linuxaio_parse_events (EV_A_ ioev, res);
456
457 if (expect_true (ringbuf_valid))
458 {
459 /* if we have a ring buffer, handle any remaining events in it */
460 linuxaio_get_events_from_ring (EV_A);
461
462 /* at this point, we should have handled all outstanding events */
463 break;
464 }
465 else if (res < want)
466 /* otherwise, if there were fewere events than we wanted, we assume there are no more */
467 break;
468 }
469 else
470 break; /* no events from the kernel, we are done */
471
472 timeout = 0; /* only wait in the first iteration */
473 }
474 }
475
476 inline_size
477 int
478 linuxaio_io_setup (EV_P)
479 {
480 linuxaio_ctx = 0;
481 return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx);
482 }
483
484 static void
485 linuxaio_poll (EV_P_ ev_tstamp timeout)
486 {
487 int submitted;
488
489 /* first phase: submit new iocbs */
490
491 /* io_submit might return less than the requested number of iocbs */
492 /* this is, afaics, only because of errors, but we go by the book and use a loop, */
493 /* which allows us to pinpoint the erroneous iocb */
494 for (submitted = 0; submitted < linuxaio_submitcnt; )
495 {
496 int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted);
497
498 if (expect_false (res < 0))
499 if (errno == EINVAL)
500 {
501 /* This happens for unsupported fds, officially, but in my testing,
502 * also randomly happens for supported fds. We fall back to good old
503 * poll() here, under the assumption that this is a very rare case.
504 * See https://lore.kernel.org/patchwork/patch/1047453/ to see
505 * discussion about such a case (ttys) where polling for POLLIN
506 * fails but POLLIN|POLLOUT works.
507 */
508 struct iocb *iocb = linuxaio_submits [submitted];
509 epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events);
510 iocb->aio_reqprio = -1; /* mark iocb as epoll */
511
512 res = 1; /* skip this iocb - another iocb, another chance */
513 }
514 else if (errno == EAGAIN)
515 {
516 /* This happens when the ring buffer is full, or some other shit we
517 * don't know and isn't documented. Most likely because we have too
518 * many requests and linux aio can't be assed to handle them.
519 * In this case, we try to allocate a larger ring buffer, freeing
520 * ours first. This might fail, in which case we have to fall back to 100%
521 * epoll.
522 * God, how I hate linux not getting its act together. Ever.
523 */
524 evsys_io_destroy (linuxaio_ctx);
525 linuxaio_submitcnt = 0;
526
527 /* rearm all fds with active iocbs */
528 {
529 int fd;
530 for (fd = 0; fd < linuxaio_iocbpmax; ++fd)
531 if (linuxaio_iocbps [fd]->io.aio_buf)
532 linuxaio_fd_rearm (EV_A_ fd);
533 }
534
535 ++linuxaio_iteration;
536 if (linuxaio_io_setup (EV_A) < 0)
537 {
538 /* to bad, we can't get a new aio context, go 100% epoll */
539 linuxaio_free_iocbp (EV_A);
540 ev_io_stop (EV_A_ &linuxaio_epoll_w);
541 ev_ref (EV_A);
542 linuxaio_ctx = 0;
543 backend_modify = epoll_modify;
544 backend_poll = epoll_poll;
545 }
546
547 timeout = 0;
548 /* it's easiest to handle this mess in another iteration */
549 return;
550 }
551 else if (errno == EBADF)
552 {
553 assert (("libev: event loop rejected bad fd", errno != EBADF));
554 fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes);
555
556 res = 1; /* skip this iocb */
557 }
558 else
559 ev_syserr ("(libev) linuxaio io_submit");
560
561 submitted += res;
562 }
563
564 linuxaio_submitcnt = 0;
565
566 /* second phase: fetch and parse events */
567
568 linuxaio_get_events (EV_A_ timeout);
569 }
570
571 inline_size
572 int
573 linuxaio_init (EV_P_ int flags)
574 {
575 /* would be great to have a nice test for IOCB_CMD_POLL instead */
576 /* also: test some semi-common fd types, such as files and ttys in recommended_backends */
577 /* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */
578 if (ev_linux_version () < 0x041300)
579 return 0;
580
581 if (!epoll_init (EV_A_ 0))
582 return 0;
583
584 linuxaio_iteration = 0;
585
586 if (linuxaio_io_setup (EV_A) < 0)
587 {
588 epoll_destroy (EV_A);
589 return 0;
590 }
591
592 ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ);
593 ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI);
594 ev_io_start (EV_A_ &linuxaio_epoll_w);
595 ev_unref (EV_A); /* watcher should not keep loop alive */
596
597 backend_modify = linuxaio_modify;
598 backend_poll = linuxaio_poll;
599
600 linuxaio_iocbpmax = 0;
601 linuxaio_iocbps = 0;
602
603 linuxaio_submits = 0;
604 linuxaio_submitmax = 0;
605 linuxaio_submitcnt = 0;
606
607 return EVBACKEND_LINUXAIO;
608 }
609
610 inline_size
611 void
612 linuxaio_destroy (EV_P)
613 {
614 epoll_destroy (EV_A);
615 linuxaio_free_iocbp (EV_A);
616 evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */
617 }
618
619 inline_size
620 void
621 linuxaio_fork (EV_P)
622 {
623 /* this frees all iocbs, which is very heavy-handed */
624 linuxaio_destroy (EV_A);
625 linuxaio_submitcnt = 0; /* all pointers were invalidated */
626
627 linuxaio_iteration = 0; /* we start over in the child */
628
629 while (linuxaio_io_setup (EV_A) < 0)
630 ev_syserr ("(libev) linuxaio io_setup");
631
632 /* forking epoll should also effectively unregister all fds from the backend */
633 epoll_fork (EV_A);
634
635 ev_io_stop (EV_A_ &linuxaio_epoll_w);
636 ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ);
637 ev_io_start (EV_A_ &linuxaio_epoll_w);
638
639 /* epoll_fork already did this. hopefully */
640 /*fd_rearm_all (EV_A);*/
641 }
642