… | |
… | |
219 | static int |
219 | static int |
220 | linuxaio_get_events_from_ring (EV_P) |
220 | linuxaio_get_events_from_ring (EV_P) |
221 | { |
221 | { |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
223 | |
223 | |
224 | unsigned head = ring->head; |
224 | /* the kernel reads and writes both of these variables, */ |
|
|
225 | /* as a C extension, we assume that volatile use here */ |
|
|
226 | /* both makes reads atomic and once-only */ |
|
|
227 | unsigned head = *(volatile unsigned *)&ring->head; |
225 | unsigned tail = *(volatile unsigned *)&ring->tail; |
228 | unsigned tail = *(volatile unsigned *)&ring->tail; |
226 | |
229 | |
227 | if (head == tail) |
230 | if (head == tail) |
228 | return 0; |
231 | return 0; |
229 | |
232 | |
… | |
… | |
231 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
234 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
232 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
235 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
233 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
236 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
234 | return 0; |
237 | return 0; |
235 | |
238 | |
|
|
239 | /* make sure the events up to tail are visible */ |
236 | ECB_MEMORY_FENCE_ACQUIRE; |
240 | ECB_MEMORY_FENCE_ACQUIRE; |
237 | |
241 | |
238 | /* parse all available events, but only once, to avoid starvation */ |
242 | /* parse all available events, but only once, to avoid starvation */ |
239 | if (tail > head) /* normal case around */ |
243 | if (tail > head) /* normal case around */ |
240 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
242 | { |
246 | { |
243 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
247 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
248 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
245 | } |
249 | } |
246 | |
250 | |
|
|
251 | /* as an extension to C, we hope that the volatile will makethis atomic and once-only */ |
247 | *(volatile unsigned *)&ring->head = tail; |
252 | *(volatile unsigned *)&ring->head = tail; |
248 | |
253 | /* make sure kernel can see our new head value - probably not required */ |
249 | /* again, other implementations don't do this, but I think it's required for the kernel to see free slots */ |
|
|
250 | ECB_MEMORY_FENCE_RELEASE; |
254 | ECB_MEMORY_FENCE_RELEASE; |
251 | |
255 | |
252 | return 1; |
256 | return 1; |
253 | } |
257 | } |
254 | |
258 | |
… | |
… | |
332 | break; |
336 | break; |
333 | } |
337 | } |
334 | #if EPOLL_FALLBACK |
338 | #if EPOLL_FALLBACK |
335 | else if (errno == EINVAL) |
339 | else if (errno == EINVAL) |
336 | { |
340 | { |
337 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
341 | /* This happens for unsupported fds, officially, but in my testing, |
338 | * also randomly happens for supported fds. We fall back to good old |
342 | * also randomly happens for supported fds. We fall back to good old |
339 | * poll() here, under the assumption that this is a very rare case. |
343 | * poll() here, under the assumption that this is a very rare case. |
|
|
344 | * See https://lore.kernel.org/patchwork/patch/1047453/ for evidence |
|
|
345 | * that the problem is known, but ignored. |
340 | */ |
346 | */ |
341 | struct iocb *iocb = linuxaio_submits [submitted]; |
347 | struct iocb *iocb = linuxaio_submits [submitted]; |
342 | res = 1; /* skip this iocb */ |
348 | res = 1; /* skip this iocb */ |
343 | |
349 | |
344 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
350 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
… | |
… | |
400 | int |
406 | int |
401 | linuxaio_init (EV_P_ int flags) |
407 | linuxaio_init (EV_P_ int flags) |
402 | { |
408 | { |
403 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
409 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
404 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
410 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
405 | if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ |
411 | #if EPOLL_FALLBACK |
|
|
412 | /* 4.19 made epoll work */ |
|
|
413 | if (ev_linux_version () < 0x041300) |
406 | return 0; |
414 | return 0; |
|
|
415 | #else |
|
|
416 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
417 | if (ev_linux_version () < 0x041200) |
|
|
418 | return 0; |
|
|
419 | #endif |
407 | |
420 | |
408 | linuxaio_ctx = 0; |
421 | linuxaio_ctx = 0; |
409 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
422 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
410 | return 0; |
423 | return 0; |
411 | |
424 | |
… | |
… | |
417 | return 0; |
430 | return 0; |
418 | } |
431 | } |
419 | |
432 | |
420 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
433 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
421 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
434 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
435 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
422 | #endif |
436 | #endif |
423 | |
437 | |
424 | backend_modify = linuxaio_modify; |
438 | backend_modify = linuxaio_modify; |
425 | backend_poll = linuxaio_poll; |
439 | backend_poll = linuxaio_poll; |
426 | |
440 | |
… | |
… | |
462 | ev_syserr ("(libev) linuxaio epoll_create"); |
476 | ev_syserr ("(libev) linuxaio epoll_create"); |
463 | |
477 | |
464 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
478 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
465 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
479 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
466 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
480 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
467 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
468 | #endif |
481 | #endif |
469 | |
482 | |
470 | fd_rearm_all (EV_A); |
483 | fd_rearm_all (EV_A); |
471 | } |
484 | } |
472 | |
485 | |