… | |
… | |
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 | |
247 | ring->head = tail; |
251 | /* TODO: we only need a compiler barrier here, not a read fence */ |
|
|
252 | ECB_MEMORY_FENCE_RELEASE; |
|
|
253 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
|
|
254 | *(volatile unsigned *)&ring->head = tail; |
|
|
255 | /* make sure kernel can see our new head value - probably not required */ |
|
|
256 | ECB_MEMORY_FENCE_RELEASE; |
248 | |
257 | |
249 | return 1; |
258 | return 1; |
250 | } |
259 | } |
251 | |
260 | |
252 | /* read at least one event from kernel, or timeout */ |
261 | /* read at least one event from kernel, or timeout */ |
… | |
… | |
329 | break; |
338 | break; |
330 | } |
339 | } |
331 | #if EPOLL_FALLBACK |
340 | #if EPOLL_FALLBACK |
332 | else if (errno == EINVAL) |
341 | else if (errno == EINVAL) |
333 | { |
342 | { |
334 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
343 | /* This happens for unsupported fds, officially, but in my testing, |
335 | * also randomly happens for supported fds. We fall back to good old |
344 | * also randomly happens for supported fds. We fall back to good old |
336 | * poll() here, under the assumption that this is a very rare case. |
345 | * poll() here, under the assumption that this is a very rare case. |
|
|
346 | * See https://lore.kernel.org/patchwork/patch/1047453/ for evidence |
|
|
347 | * that the problem is known, but ignored. |
337 | */ |
348 | */ |
338 | struct iocb *iocb = linuxaio_submits [submitted]; |
349 | struct iocb *iocb = linuxaio_submits [submitted]; |
339 | res = 1; /* skip this iocb */ |
350 | res = 1; /* skip this iocb */ |
340 | |
351 | |
341 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
352 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
… | |
… | |
397 | int |
408 | int |
398 | linuxaio_init (EV_P_ int flags) |
409 | linuxaio_init (EV_P_ int flags) |
399 | { |
410 | { |
400 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
411 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
401 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
412 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
402 | if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ |
413 | #if EPOLL_FALLBACK |
|
|
414 | /* 4.19 made epoll work */ |
|
|
415 | if (ev_linux_version () < 0x041300) |
403 | return 0; |
416 | return 0; |
|
|
417 | #else |
|
|
418 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
419 | if (ev_linux_version () < 0x041200) |
|
|
420 | return 0; |
|
|
421 | #endif |
404 | |
422 | |
405 | linuxaio_ctx = 0; |
423 | linuxaio_ctx = 0; |
406 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
424 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
407 | return 0; |
425 | return 0; |
408 | |
426 | |
… | |
… | |
414 | return 0; |
432 | return 0; |
415 | } |
433 | } |
416 | |
434 | |
417 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
435 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
418 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
436 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
437 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
419 | #endif |
438 | #endif |
420 | |
439 | |
421 | backend_modify = linuxaio_modify; |
440 | backend_modify = linuxaio_modify; |
422 | backend_poll = linuxaio_poll; |
441 | backend_poll = linuxaio_poll; |
423 | |
442 | |
… | |
… | |
459 | ev_syserr ("(libev) linuxaio epoll_create"); |
478 | ev_syserr ("(libev) linuxaio epoll_create"); |
460 | |
479 | |
461 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
480 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
462 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
481 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
463 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
482 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
464 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
465 | #endif |
483 | #endif |
466 | |
484 | |
467 | fd_rearm_all (EV_A); |
485 | fd_rearm_all (EV_A); |
468 | } |
486 | } |
469 | |
487 | |