… | |
… | |
100 | * not only is this totally undocumented, not even the source code |
100 | * not only is this totally undocumented, not even the source code |
101 | * can tell you what the future semantics of compat_features and |
101 | * can tell you what the future semantics of compat_features and |
102 | * incompat_features are, or what header_length actually is for. |
102 | * incompat_features are, or what header_length actually is for. |
103 | */ |
103 | */ |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
105 | #define AIO_RING_INCOMPAT_FEATURES 0 |
105 | #define EV_AIO_RING_INCOMPAT_FEATURES 0 |
106 | struct aio_ring |
106 | struct aio_ring |
107 | { |
107 | { |
108 | unsigned id; /* kernel internal index number */ |
108 | unsigned id; /* kernel internal index number */ |
109 | unsigned nr; /* number of io_events */ |
109 | unsigned nr; /* number of io_events */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
… | |
… | |
289 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
289 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
290 | |
290 | |
291 | if (iocb->io.aio_reqprio < 0) |
291 | if (iocb->io.aio_reqprio < 0) |
292 | { |
292 | { |
293 | /* we handed this fd over to epoll, so undo this first */ |
293 | /* we handed this fd over to epoll, so undo this first */ |
294 | /* we do it manually because the optimisations on epoll_modfy won't do us any good */ |
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); |
295 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
296 | anfds [fd].emask = 0; |
296 | anfds [fd].emask = 0; |
297 | iocb->io.aio_reqprio = 0; |
297 | iocb->io.aio_reqprio = 0; |
298 | } |
298 | } |
299 | |
299 | |
300 | if (iocb->io.aio_buf) |
300 | if (iocb->io.aio_buf) |
301 | /* io_cancel always returns some error on relevant kernels, but works */ |
301 | { |
302 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
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 | } |
303 | |
306 | |
304 | if (nev) |
307 | if (nev) |
305 | { |
308 | { |
306 | iocb->io.aio_buf = |
309 | iocb->io.aio_buf = |
307 | (nev & EV_READ ? POLLIN : 0) |
310 | (nev & EV_READ ? POLLIN : 0) |
… | |
… | |
319 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
322 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
320 | { |
323 | { |
321 | epoll_poll (EV_A_ 0); |
324 | epoll_poll (EV_A_ 0); |
322 | } |
325 | } |
323 | |
326 | |
324 | static void |
327 | inline_speed |
|
|
328 | void |
325 | linuxaio_fd_rearm (EV_P_ int fd) |
329 | linuxaio_fd_rearm (EV_P_ int fd) |
326 | { |
330 | { |
327 | anfds [fd].events = 0; |
331 | anfds [fd].events = 0; |
328 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
332 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
329 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
333 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
… | |
… | |
345 | fd, |
349 | fd, |
346 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
350 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
347 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
351 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
348 | ); |
352 | ); |
349 | |
353 | |
350 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
354 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
351 | linuxaio_fd_rearm (EV_A_ fd); |
355 | linuxaio_fd_rearm (EV_A_ fd); |
352 | |
356 | |
353 | --nr; |
357 | --nr; |
354 | ++ev; |
358 | ++ev; |
355 | } |
359 | } |
… | |
… | |
368 | unsigned tail = *(volatile unsigned *)&ring->tail; |
372 | unsigned tail = *(volatile unsigned *)&ring->tail; |
369 | |
373 | |
370 | if (head == tail) |
374 | if (head == tail) |
371 | return 0; |
375 | return 0; |
372 | |
376 | |
373 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
374 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
375 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
376 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
377 | return 0; |
|
|
378 | |
|
|
379 | /* make sure the events up to tail are visible */ |
377 | /* make sure the events up to tail are visible */ |
380 | ECB_MEMORY_FENCE_ACQUIRE; |
378 | ECB_MEMORY_FENCE_ACQUIRE; |
381 | |
379 | |
382 | /* parse all available events, but only once, to avoid starvation */ |
380 | /* parse all available events, but only once, to avoid starvation */ |
383 | if (tail > head) /* normal case around */ |
381 | if (tail > head) /* normal case around */ |
… | |
… | |
393 | *(volatile unsigned *)&ring->head = tail; |
391 | *(volatile unsigned *)&ring->head = tail; |
394 | |
392 | |
395 | return 1; |
393 | return 1; |
396 | } |
394 | } |
397 | |
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 | |
398 | /* read at least one event from kernel, or timeout */ |
407 | /* read at least one event from kernel, or timeout */ |
399 | inline_size |
408 | inline_size |
400 | void |
409 | void |
401 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
410 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
402 | { |
411 | { |
403 | struct timespec ts; |
412 | struct timespec ts; |
404 | struct io_event ioev[1]; |
413 | struct io_event ioev[8]; /* 256 octet stack space */ |
405 | int res; |
414 | int want = 1; /* how many events to request */ |
|
|
415 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
406 | |
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 */ |
407 | if (linuxaio_get_events_from_ring (EV_A)) |
420 | if (linuxaio_get_events_from_ring (EV_A)) |
408 | return; |
421 | return; |
409 | |
422 | |
410 | /* no events, so wait for at least one, then poll ring buffer again */ |
423 | /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
411 | /* this degrades to one event per loop iteration */ |
424 | if (!timeout) |
412 | /* if the ring buffer changes layout, but so be it */ |
425 | return; |
|
|
426 | } |
|
|
427 | else |
|
|
428 | /* no ringbuffer, request slightly larger batch */ |
|
|
429 | want = sizeof (ioev) / sizeof (ioev [0]); |
413 | |
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 | |
414 | EV_RELEASE_CB; |
438 | EV_RELEASE_CB; |
415 | |
439 | |
416 | ts.tv_sec = (long)timeout; |
440 | ts.tv_sec = (long)timeout; |
417 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
441 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
418 | |
442 | |
419 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
443 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
420 | |
444 | |
421 | EV_ACQUIRE_CB; |
445 | EV_ACQUIRE_CB; |
422 | |
446 | |
423 | if (res < 0) |
447 | if (res < 0) |
424 | if (errno == EINTR) |
448 | if (errno == EINTR) |
425 | /* ignored */; |
449 | /* ignored, retry */; |
426 | else |
450 | else |
427 | ev_syserr ("(libev) linuxaio io_getevents"); |
451 | ev_syserr ("(libev) linuxaio io_getevents"); |
428 | else if (res) |
452 | else if (res) |
429 | { |
453 | { |
430 | /* at least one event available, handle it and any remaining ones in the ring buffer */ |
454 | /* at least one event available, handle them */ |
431 | linuxaio_parse_events (EV_A_ ioev, res); |
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 */ |
432 | linuxaio_get_events_from_ring (EV_A); |
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 */ |
433 | } |
473 | } |
434 | } |
474 | } |
435 | |
475 | |
436 | inline_size |
476 | inline_size |
437 | int |
477 | int |
… | |
… | |
508 | /* it's easiest to handle this mess in another iteration */ |
548 | /* it's easiest to handle this mess in another iteration */ |
509 | return; |
549 | return; |
510 | } |
550 | } |
511 | else if (errno == EBADF) |
551 | else if (errno == EBADF) |
512 | { |
552 | { |
|
|
553 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
513 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
554 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
514 | |
555 | |
515 | res = 1; /* skip this iocb */ |
556 | res = 1; /* skip this iocb */ |
516 | } |
557 | } |
517 | else |
558 | else |
… | |
… | |
570 | void |
611 | void |
571 | linuxaio_destroy (EV_P) |
612 | linuxaio_destroy (EV_P) |
572 | { |
613 | { |
573 | epoll_destroy (EV_A); |
614 | epoll_destroy (EV_A); |
574 | linuxaio_free_iocbp (EV_A); |
615 | linuxaio_free_iocbp (EV_A); |
575 | evsys_io_destroy (linuxaio_ctx); |
616 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
576 | } |
617 | } |
577 | |
618 | |
578 | inline_size |
619 | inline_size |
579 | void |
620 | void |
580 | linuxaio_fork (EV_P) |
621 | linuxaio_fork (EV_P) |
… | |
… | |
586 | linuxaio_iteration = 0; /* we start over in the child */ |
627 | linuxaio_iteration = 0; /* we start over in the child */ |
587 | |
628 | |
588 | while (linuxaio_io_setup (EV_A) < 0) |
629 | while (linuxaio_io_setup (EV_A) < 0) |
589 | ev_syserr ("(libev) linuxaio io_setup"); |
630 | ev_syserr ("(libev) linuxaio io_setup"); |
590 | |
631 | |
|
|
632 | /* forking epoll should also effectively unregister all fds from the backend */ |
591 | epoll_fork (EV_A); |
633 | epoll_fork (EV_A); |
592 | |
634 | |
593 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
635 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
594 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
636 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
595 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
637 | ev_io_start (EV_A_ &linuxaio_epoll_w); |