… | |
… | |
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
199 | anfds [fd].events = 0; |
199 | anfds [fd].events = 0; |
200 | fd_change (EV_A_ fd, 0); |
200 | fd_change (EV_A_ fd, 0); |
201 | |
201 | |
202 | /* feed events, we do not expect or handle POLLNVAL */ |
202 | /* feed events, we do not expect or handle POLLNVAL */ |
203 | if (ecb_expect_false (res & POLLNVAL)) |
203 | if (expect_false (res & POLLNVAL)) |
204 | fd_kill (EV_A_ fd); |
204 | fd_kill (EV_A_ fd); |
205 | else |
205 | else |
206 | fd_event ( |
206 | fd_event ( |
207 | EV_A_ |
207 | EV_A_ |
208 | fd, |
208 | fd, |
… | |
… | |
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 | |
230 | /* bail out if the ring buffer doesn't match the expected layout */ |
233 | /* bail out if the ring buffer doesn't match the expected layout */ |
231 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
234 | if (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 | /* 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 */ |
247 | *(volatile unsigned *)&ring->head = tail; |
254 | *(volatile unsigned *)&ring->head = tail; |
248 | |
255 | /* 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; |
256 | ECB_MEMORY_FENCE_RELEASE; |
251 | |
257 | |
252 | return 1; |
258 | return 1; |
253 | } |
259 | } |
254 | |
260 | |
… | |
… | |
256 | inline_size |
262 | inline_size |
257 | void |
263 | void |
258 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
264 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
259 | { |
265 | { |
260 | struct timespec ts; |
266 | struct timespec ts; |
261 | struct io_event ioev; |
267 | struct io_event ioev[1]; |
262 | int res; |
268 | int res; |
263 | |
269 | |
264 | if (linuxaio_get_events_from_ring (EV_A)) |
270 | if (linuxaio_get_events_from_ring (EV_A)) |
265 | return; |
271 | return; |
266 | |
272 | |
… | |
… | |
269 | /* if the ring buffer changes layout, but so be it */ |
275 | /* if the ring buffer changes layout, but so be it */ |
270 | |
276 | |
271 | ts.tv_sec = (long)timeout; |
277 | ts.tv_sec = (long)timeout; |
272 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
278 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
273 | |
279 | |
274 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
280 | res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
275 | |
281 | |
276 | if (res < 0) |
282 | if (res < 0) |
277 | if (errno == EINTR) |
283 | if (errno == EINTR) |
278 | /* ignored */; |
284 | /* ignored */; |
279 | else |
285 | else |
280 | ev_syserr ("(libev) linuxaio io_getevents"); |
286 | ev_syserr ("(libev) linuxaio io_getevents"); |
281 | else if (res) |
287 | else if (res) |
282 | { |
288 | { |
283 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
289 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
284 | linuxaio_parse_events (EV_A_ &ioev, 1); |
290 | linuxaio_parse_events (EV_A_ ioev, res); |
285 | linuxaio_get_events_from_ring (EV_A); |
291 | linuxaio_get_events_from_ring (EV_A); |
286 | } |
292 | } |
287 | } |
293 | } |
288 | |
294 | |
289 | #if EPOLL_FALLBACK |
295 | #if EPOLL_FALLBACK |
… | |
… | |
314 | /* which allows us to pinpoint the errornous iocb */ |
320 | /* which allows us to pinpoint the errornous iocb */ |
315 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
321 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
316 | { |
322 | { |
317 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
323 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
318 | |
324 | |
319 | if (ecb_expect_false (res < 0)) |
325 | if (expect_false (res < 0)) |
320 | if (errno == EAGAIN) |
326 | if (errno == EAGAIN) |
321 | { |
327 | { |
322 | /* This happens when the ring buffer is full, at least. I assume this means |
328 | /* This happens when the ring buffer is full, at least. I assume this means |
323 | * that the event was queued synchronously during io_submit, and thus |
329 | * that the event was queued synchronously during io_submit, and thus |
324 | * the buffer overflowed. |
330 | * the buffer overflowed. |
… | |
… | |
332 | break; |
338 | break; |
333 | } |
339 | } |
334 | #if EPOLL_FALLBACK |
340 | #if EPOLL_FALLBACK |
335 | else if (errno == EINVAL) |
341 | else if (errno == EINVAL) |
336 | { |
342 | { |
337 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
343 | /* This happens for unsupported fds, officially, but in my testing, |
338 | * 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 |
339 | * 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. |
340 | */ |
348 | */ |
341 | struct iocb *iocb = linuxaio_submits [submitted]; |
349 | struct iocb *iocb = linuxaio_submits [submitted]; |
342 | res = 1; /* skip this iocb */ |
350 | res = 1; /* skip this iocb */ |
343 | |
351 | |
344 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
352 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
… | |
… | |
368 | for (;;) |
376 | for (;;) |
369 | { |
377 | { |
370 | int idx; |
378 | int idx; |
371 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
379 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
372 | |
380 | |
373 | if (ecb_expect_false (res < 0)) |
381 | if (expect_false (res < 0)) |
374 | ev_syserr ("(libev) linuxaio epoll_wait"); |
382 | ev_syserr ("(libev) linuxaio epoll_wait"); |
375 | else if (!res) |
383 | else if (!res) |
376 | break; |
384 | break; |
377 | |
385 | |
378 | for (idx = res; idx--; ) |
386 | for (idx = res; idx--; ) |
… | |
… | |
400 | int |
408 | int |
401 | linuxaio_init (EV_P_ int flags) |
409 | linuxaio_init (EV_P_ int flags) |
402 | { |
410 | { |
403 | /* 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 */ |
404 | /* 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 */ |
405 | 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) |
406 | 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 |
407 | |
422 | |
408 | linuxaio_ctx = 0; |
423 | linuxaio_ctx = 0; |
409 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
424 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
410 | return 0; |
425 | return 0; |
411 | |
426 | |
… | |
… | |
417 | return 0; |
432 | return 0; |
418 | } |
433 | } |
419 | |
434 | |
420 | 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); |
421 | 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 */ |
422 | #endif |
438 | #endif |
423 | |
439 | |
424 | backend_modify = linuxaio_modify; |
440 | backend_modify = linuxaio_modify; |
425 | backend_poll = linuxaio_poll; |
441 | backend_poll = linuxaio_poll; |
426 | |
442 | |
… | |
… | |
462 | ev_syserr ("(libev) linuxaio epoll_create"); |
478 | ev_syserr ("(libev) linuxaio epoll_create"); |
463 | |
479 | |
464 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
480 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
465 | 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); |
466 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
482 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
467 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
468 | #endif |
483 | #endif |
469 | |
484 | |
470 | fd_rearm_all (EV_A); |
485 | fd_rearm_all (EV_A); |
471 | } |
486 | } |
472 | |
487 | |