… | |
… | |
212 | * this is not well documented, so we better do it. |
212 | * this is not well documented, so we better do it. |
213 | */ |
213 | */ |
214 | memset (iocb, 0, sizeof (*iocb)); |
214 | memset (iocb, 0, sizeof (*iocb)); |
215 | |
215 | |
216 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
216 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
217 | iocb->io.aio_data = offset; |
|
|
218 | iocb->io.aio_fildes = offset; |
217 | iocb->io.aio_fildes = offset; |
219 | |
218 | |
220 | base [offset++] = iocb; |
219 | base [offset++] = iocb; |
221 | } |
220 | } |
222 | } |
221 | } |
… | |
… | |
234 | static void |
233 | static void |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
234 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
236 | { |
235 | { |
237 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
236 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
237 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
|
|
238 | ANFD *anfd = &anfds [fd]; |
239 | |
239 | |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
241 | { |
241 | { |
242 | /* we handed this fd over to epoll, so undo this first */ |
242 | /* we handed this fd over to epoll, so undo this first */ |
243 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
243 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
245 | anfds [fd].emask = 0; |
245 | anfd->emask = 0; |
246 | iocb->io.aio_reqprio = 0; |
246 | iocb->io.aio_reqprio = 0; |
247 | } |
247 | } |
248 | |
|
|
249 | if (ecb_expect_false (iocb->io.aio_buf)) |
248 | else if (ecb_expect_false (iocb->io.aio_buf)) |
250 | { |
249 | { |
251 | /* iocb active, so cancel it first before resubmit */ |
250 | /* iocb active, so cancel it first before resubmit */ |
|
|
251 | /* this assumes we only ever get one call per fd per loop iteration */ |
252 | for (;;) |
252 | for (;;) |
253 | { |
253 | { |
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
256 | break; |
256 | break; |
257 | |
257 | |
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
259 | break; |
259 | break; |
260 | |
260 | |
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
262 | if (errno != EINTR) |
|
|
263 | { |
262 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
264 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS)); |
|
|
265 | break; |
|
|
266 | } |
263 | } |
267 | } |
|
|
268 | |
|
|
269 | /* increment generation counter to avoid handling old events */ |
|
|
270 | ++anfd->egen; |
264 | } |
271 | } |
265 | |
272 | |
266 | iocb->io.aio_buf = |
273 | iocb->io.aio_buf = |
267 | (nev & EV_READ ? POLLIN : 0) |
274 | (nev & EV_READ ? POLLIN : 0) |
268 | | (nev & EV_WRITE ? POLLOUT : 0); |
275 | | (nev & EV_WRITE ? POLLOUT : 0); |
269 | |
276 | |
270 | if (nev) |
277 | if (nev) |
271 | { |
278 | { |
|
|
279 | iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32); |
|
|
280 | |
272 | /* queue iocb up for io_submit */ |
281 | /* queue iocb up for io_submit */ |
273 | /* this assumes we only ever get one call per fd per loop iteration */ |
282 | /* this assumes we only ever get one call per fd per loop iteration */ |
274 | ++linuxaio_submitcnt; |
283 | ++linuxaio_submitcnt; |
275 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
284 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
276 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
285 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
… | |
… | |
295 | static void |
304 | static void |
296 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
305 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
297 | { |
306 | { |
298 | while (nr) |
307 | while (nr) |
299 | { |
308 | { |
300 | int fd = ev->data; |
309 | int fd = ev->data & 0xffffffff; |
|
|
310 | uint32_t gen = ev->data >> 32; |
301 | int res = ev->res; |
311 | int res = ev->res; |
302 | |
312 | |
303 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
313 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
314 | |
|
|
315 | /* ignore event if generation doesn't match */ |
|
|
316 | if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen)) |
|
|
317 | continue; |
304 | |
318 | |
305 | /* feed events, we do not expect or handle POLLNVAL */ |
319 | /* feed events, we do not expect or handle POLLNVAL */ |
306 | fd_event ( |
320 | fd_event ( |
307 | EV_A_ |
321 | EV_A_ |
308 | fd, |
322 | fd, |
… | |
… | |
321 | /* get any events from ring buffer, return true if any were handled */ |
335 | /* get any events from ring buffer, return true if any were handled */ |
322 | static int |
336 | static int |
323 | linuxaio_get_events_from_ring (EV_P) |
337 | linuxaio_get_events_from_ring (EV_P) |
324 | { |
338 | { |
325 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
339 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
340 | unsigned head, tail; |
326 | |
341 | |
327 | /* the kernel reads and writes both of these variables, */ |
342 | /* the kernel reads and writes both of these variables, */ |
328 | /* as a C extension, we assume that volatile use here */ |
343 | /* as a C extension, we assume that volatile use here */ |
329 | /* both makes reads atomic and once-only */ |
344 | /* both makes reads atomic and once-only */ |
330 | unsigned head = *(volatile unsigned *)&ring->head; |
345 | head = *(volatile unsigned *)&ring->head; |
|
|
346 | ECB_MEMORY_FENCE_ACQUIRE; |
331 | unsigned tail = *(volatile unsigned *)&ring->tail; |
347 | tail = *(volatile unsigned *)&ring->tail; |
332 | |
348 | |
333 | if (head == tail) |
349 | if (head == tail) |
334 | return 0; |
350 | return 0; |
335 | |
351 | |
336 | /* make sure the events up to tail are visible */ |
|
|
337 | ECB_MEMORY_FENCE_ACQUIRE; |
|
|
338 | |
|
|
339 | /* parse all available events, but only once, to avoid starvation */ |
352 | /* parse all available events, but only once, to avoid starvation */ |
340 | if (tail > head) /* normal case around */ |
353 | if (ecb_expect_true (tail > head)) /* normal case around */ |
341 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
354 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
342 | else /* wrapped around */ |
355 | else /* wrapped around */ |
343 | { |
356 | { |
344 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
357 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
345 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
358 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
… | |
… | |
394 | { |
407 | { |
395 | int res; |
408 | int res; |
396 | |
409 | |
397 | EV_RELEASE_CB; |
410 | EV_RELEASE_CB; |
398 | |
411 | |
399 | ts.tv_sec = (long)timeout; |
412 | EV_TS_SET (ts, timeout); |
400 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
|
|
401 | |
|
|
402 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
413 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
403 | |
414 | |
404 | EV_ACQUIRE_CB; |
415 | EV_ACQUIRE_CB; |
405 | |
416 | |
406 | if (res < 0) |
417 | if (res < 0) |
… | |
… | |
492 | } |
503 | } |
493 | |
504 | |
494 | ++linuxaio_iteration; |
505 | ++linuxaio_iteration; |
495 | if (linuxaio_io_setup (EV_A) < 0) |
506 | if (linuxaio_io_setup (EV_A) < 0) |
496 | { |
507 | { |
|
|
508 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
509 | /* TODO: might be more prudent? */ |
|
|
510 | |
497 | /* to bad, we can't get a new aio context, go 100% epoll */ |
511 | /* to bad, we can't get a new aio context, go 100% epoll */ |
498 | linuxaio_free_iocbp (EV_A); |
512 | linuxaio_free_iocbp (EV_A); |
499 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
513 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
500 | ev_ref (EV_A); |
514 | ev_ref (EV_A); |
501 | linuxaio_ctx = 0; |
515 | linuxaio_ctx = 0; |
|
|
516 | |
|
|
517 | backend = EVBACKEND_EPOLL; |
502 | backend_modify = epoll_modify; |
518 | backend_modify = epoll_modify; |
503 | backend_poll = epoll_poll; |
519 | backend_poll = epoll_poll; |
504 | } |
520 | } |
505 | |
521 | |
506 | timeout = 0; |
522 | timeout = 0; |
… | |
… | |
515 | res = 1; /* skip this iocb */ |
531 | res = 1; /* skip this iocb */ |
516 | } |
532 | } |
517 | else if (errno == EINTR) /* not seen in reality, not documented */ |
533 | else if (errno == EINTR) /* not seen in reality, not documented */ |
518 | res = 0; /* silently ignore and retry */ |
534 | res = 0; /* silently ignore and retry */ |
519 | else |
535 | else |
|
|
536 | { |
520 | ev_syserr ("(libev) linuxaio io_submit"); |
537 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
538 | res = 0; |
|
|
539 | } |
521 | |
540 | |
522 | submitted += res; |
541 | submitted += res; |
523 | } |
542 | } |
524 | |
543 | |
525 | linuxaio_submitcnt = 0; |
544 | linuxaio_submitcnt = 0; |
… | |
… | |
553 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
572 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
554 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
573 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
555 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
574 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
556 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
575 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
557 | |
576 | |
558 | backend_modify = linuxaio_modify; |
577 | backend_modify = linuxaio_modify; |
559 | backend_poll = linuxaio_poll; |
578 | backend_poll = linuxaio_poll; |
560 | |
579 | |
561 | linuxaio_iocbpmax = 0; |
580 | linuxaio_iocbpmax = 0; |
562 | linuxaio_iocbps = 0; |
581 | linuxaio_iocbps = 0; |
563 | |
582 | |
564 | linuxaio_submits = 0; |
583 | linuxaio_submits = 0; |
… | |
… | |
575 | epoll_destroy (EV_A); |
594 | epoll_destroy (EV_A); |
576 | linuxaio_free_iocbp (EV_A); |
595 | linuxaio_free_iocbp (EV_A); |
577 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
596 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
578 | } |
597 | } |
579 | |
598 | |
580 | inline_size |
599 | ecb_cold |
581 | void |
600 | static void |
582 | linuxaio_fork (EV_P) |
601 | linuxaio_fork (EV_P) |
583 | { |
602 | { |
584 | /* this frees all iocbs, which is very heavy-handed */ |
|
|
585 | linuxaio_destroy (EV_A); |
|
|
586 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
603 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
|
|
604 | linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */ |
|
|
605 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
587 | |
606 | |
588 | linuxaio_iteration = 0; /* we start over in the child */ |
607 | linuxaio_iteration = 0; /* we start over in the child */ |
589 | |
608 | |
590 | while (linuxaio_io_setup (EV_A) < 0) |
609 | while (linuxaio_io_setup (EV_A) < 0) |
591 | ev_syserr ("(libev) linuxaio io_setup"); |
610 | ev_syserr ("(libev) linuxaio io_setup"); |
592 | |
611 | |
593 | /* forking epoll should also effectively unregister all fds from the backend */ |
612 | /* forking epoll should also effectively unregister all fds from the backend */ |
594 | epoll_fork (EV_A); |
613 | epoll_fork (EV_A); |
|
|
614 | /* epoll_fork already did this. hopefully */ |
|
|
615 | /*fd_rearm_all (EV_A);*/ |
595 | |
616 | |
596 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
617 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
597 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
618 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
598 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
619 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
599 | |
|
|
600 | /* epoll_fork already did this. hopefully */ |
|
|
601 | /*fd_rearm_all (EV_A);*/ |
|
|
602 | } |
620 | } |
603 | |
621 | |