… | |
… | |
182 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
182 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
184 | } |
184 | } |
185 | } |
185 | } |
186 | |
186 | |
|
|
187 | #if EPOLL_FALLBACK |
|
|
188 | |
|
|
189 | static void |
|
|
190 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
191 | { |
|
|
192 | struct epoll_event eev; |
|
|
193 | |
|
|
194 | eev.events = EPOLLONESHOT; |
|
|
195 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
196 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
197 | eev.data.fd = iocb->aio_fildes; |
|
|
198 | |
|
|
199 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
200 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
201 | } |
|
|
202 | |
|
|
203 | static void |
|
|
204 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
205 | { |
|
|
206 | struct epoll_event events[16]; |
|
|
207 | |
|
|
208 | for (;;) |
|
|
209 | { |
|
|
210 | int idx; |
|
|
211 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
212 | |
|
|
213 | if (expect_false (res < 0)) |
|
|
214 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
215 | else if (!res) |
|
|
216 | break; |
|
|
217 | |
|
|
218 | for (idx = res; idx--; ) |
|
|
219 | { |
|
|
220 | int fd = events [idx].data.fd; |
|
|
221 | uint32_t ev = events [idx].events; |
|
|
222 | |
|
|
223 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
224 | |
|
|
225 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
226 | |
|
|
227 | fd_event (EV_A_ fd, |
|
|
228 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
229 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
230 | } |
|
|
231 | |
|
|
232 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
233 | break; |
|
|
234 | } |
|
|
235 | } |
|
|
236 | |
|
|
237 | #endif |
|
|
238 | |
187 | static void |
239 | static void |
188 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
240 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
189 | { |
241 | { |
190 | while (nr) |
242 | while (nr) |
191 | { |
243 | { |
192 | int fd = ev->data; |
244 | int fd = ev->data; |
193 | int res = ev->res; |
245 | int res = ev->res; |
194 | |
246 | |
195 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
247 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
196 | |
248 | |
197 | /* linux aio is oneshot: rearm fd */ |
249 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
250 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
199 | anfds [fd].events = 0; |
251 | anfds [fd].events = 0; |
200 | fd_change (EV_A_ fd, 0); |
252 | fd_change (EV_A_ fd, 0); |
201 | |
253 | |
202 | /* feed events, we do not expect or handle POLLNVAL */ |
254 | /* feed events, we do not expect or handle POLLNVAL */ |
203 | if (ecb_expect_false (res & POLLNVAL)) |
255 | if (expect_false (res & POLLNVAL)) |
204 | fd_kill (EV_A_ fd); |
256 | fd_kill (EV_A_ fd); |
205 | else |
257 | else |
206 | fd_event ( |
258 | fd_event ( |
207 | EV_A_ |
259 | EV_A_ |
208 | fd, |
260 | fd, |
… | |
… | |
219 | static int |
271 | static int |
220 | linuxaio_get_events_from_ring (EV_P) |
272 | linuxaio_get_events_from_ring (EV_P) |
221 | { |
273 | { |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
274 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
223 | |
275 | |
224 | unsigned head = ring->head; |
276 | /* the kernel reads and writes both of these variables, */ |
|
|
277 | /* as a C extension, we assume that volatile use here */ |
|
|
278 | /* both makes reads atomic and once-only */ |
|
|
279 | unsigned head = *(volatile unsigned *)&ring->head; |
225 | unsigned tail = *(volatile unsigned *)&ring->tail; |
280 | unsigned tail = *(volatile unsigned *)&ring->tail; |
226 | |
281 | |
227 | if (head == tail) |
282 | if (head == tail) |
228 | return 0; |
283 | return 0; |
229 | |
284 | |
230 | /* bail out if the ring buffer doesn't match the expected layout */ |
285 | /* bail out if the ring buffer doesn't match the expected layout */ |
231 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
286 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
232 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
287 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
233 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
288 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
234 | return 0; |
289 | return 0; |
235 | |
290 | |
|
|
291 | /* make sure the events up to tail are visible */ |
236 | ECB_MEMORY_FENCE_ACQUIRE; |
292 | ECB_MEMORY_FENCE_ACQUIRE; |
237 | |
293 | |
238 | /* parse all available events, but only once, to avoid starvation */ |
294 | /* parse all available events, but only once, to avoid starvation */ |
239 | if (tail > head) /* normal case around */ |
295 | if (tail > head) /* normal case around */ |
240 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
296 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
242 | { |
298 | { |
243 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
299 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
300 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
245 | } |
301 | } |
246 | |
302 | |
|
|
303 | /* TODO: we only need a compiler barrier here, not a read fence */ |
|
|
304 | ECB_MEMORY_FENCE_RELEASE; |
|
|
305 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
247 | *(volatile unsigned *)&ring->head = tail; |
306 | *(volatile unsigned *)&ring->head = tail; |
248 | |
307 | /* 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; |
308 | ECB_MEMORY_FENCE_RELEASE; |
251 | |
309 | |
252 | return 1; |
310 | return 1; |
253 | } |
311 | } |
254 | |
312 | |
… | |
… | |
256 | inline_size |
314 | inline_size |
257 | void |
315 | void |
258 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
316 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
259 | { |
317 | { |
260 | struct timespec ts; |
318 | struct timespec ts; |
261 | struct io_event ioev; |
319 | struct io_event ioev[1]; |
262 | int res; |
320 | int res; |
263 | |
321 | |
264 | if (linuxaio_get_events_from_ring (EV_A)) |
322 | if (linuxaio_get_events_from_ring (EV_A)) |
265 | return; |
323 | return; |
266 | |
324 | |
267 | /* no events, so wait for at least one, then poll ring buffer again */ |
325 | /* no events, so wait for at least one, then poll ring buffer again */ |
268 | /* this degrades to one event per loop iteration */ |
326 | /* this degrades to one event per loop iteration */ |
269 | /* if the ring buffer changes layout, but so be it */ |
327 | /* if the ring buffer changes layout, but so be it */ |
270 | |
328 | |
|
|
329 | EV_RELEASE_CB; |
|
|
330 | |
271 | ts.tv_sec = (long)timeout; |
331 | ts.tv_sec = (long)timeout; |
272 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
332 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
273 | |
333 | |
274 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
334 | res = ev_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
335 | |
|
|
336 | EV_ACQUIRE_CB; |
275 | |
337 | |
276 | if (res < 0) |
338 | if (res < 0) |
277 | if (errno == EINTR) |
339 | if (errno == EINTR) |
278 | /* ignored */; |
340 | /* ignored */; |
279 | else |
341 | else |
280 | ev_syserr ("(libev) linuxaio io_getevents"); |
342 | ev_syserr ("(libev) linuxaio io_getevents"); |
281 | else if (res) |
343 | else if (res) |
282 | { |
344 | { |
283 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
345 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
284 | linuxaio_parse_events (EV_A_ &ioev, 1); |
346 | linuxaio_parse_events (EV_A_ ioev, res); |
285 | linuxaio_get_events_from_ring (EV_A); |
347 | linuxaio_get_events_from_ring (EV_A); |
286 | } |
348 | } |
287 | } |
349 | } |
288 | |
|
|
289 | #if EPOLL_FALLBACK |
|
|
290 | static void |
|
|
291 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
292 | { |
|
|
293 | struct epoll_event eev; |
|
|
294 | |
|
|
295 | eev.events = EPOLLONESHOT; |
|
|
296 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
297 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
298 | eev.data.fd = iocb->aio_fildes; |
|
|
299 | |
|
|
300 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
301 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
302 | } |
|
|
303 | #endif |
|
|
304 | |
350 | |
305 | static void |
351 | static void |
306 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
352 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
307 | { |
353 | { |
308 | int submitted; |
354 | int submitted; |
… | |
… | |
312 | /* io_submit might return less than the requested number of iocbs */ |
358 | /* io_submit might return less than the requested number of iocbs */ |
313 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
359 | /* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
314 | /* which allows us to pinpoint the errornous iocb */ |
360 | /* which allows us to pinpoint the errornous iocb */ |
315 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
361 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
316 | { |
362 | { |
|
|
363 | #if 0 |
|
|
364 | int res; |
|
|
365 | if (linuxaio_submits[submitted]->aio_fildes == backend_fd) |
|
|
366 | res = ev_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); |
|
|
367 | else |
|
|
368 | { res = -1; errno = EINVAL; }; |
|
|
369 | #else |
317 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
370 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
|
|
371 | #endif |
318 | |
372 | |
319 | if (ecb_expect_false (res < 0)) |
373 | if (expect_false (res < 0)) |
320 | if (errno == EAGAIN) |
374 | if (errno == EAGAIN) |
321 | { |
375 | { |
322 | /* This happens when the ring buffer is full, at least. I assume this means |
376 | /* 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 |
377 | * that the event was queued synchronously during io_submit, and thus |
324 | * the buffer overflowed. |
378 | * the buffer overflowed. |
325 | * In this case, we just try in next loop iteration. |
379 | * In this case, we just try in next loop iteration. |
326 | * This should not result in a few fds taking priority, as the interface |
380 | * This should not result in a few fds taking priority, as the interface |
327 | * is one-shot, and we submit iocb's in a round-robin fashion. |
381 | * is one-shot, and we submit iocb's in a round-robin fashion. |
|
|
382 | * TODO: maybe make "submitted" persistent, so we don't have to memmove? |
328 | */ |
383 | */ |
|
|
384 | if (ecb_expect_false (submitted)) |
|
|
385 | { |
329 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
386 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
330 | linuxaio_submitcnt -= submitted; |
387 | linuxaio_submitcnt -= submitted; |
|
|
388 | } |
|
|
389 | |
331 | timeout = 0; |
390 | timeout = 0; |
332 | break; |
391 | break; |
333 | } |
392 | } |
334 | #if EPOLL_FALLBACK |
393 | #if EPOLL_FALLBACK |
335 | else if (errno == EINVAL) |
394 | else if (errno == EINVAL) |
336 | { |
395 | { |
337 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
396 | /* This happens for unsupported fds, officially, but in my testing, |
338 | * also randomly happens for supported fds. We fall back to good old |
397 | * 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. |
398 | * poll() here, under the assumption that this is a very rare case. |
|
|
399 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
|
|
400 | * discussion about such a case (ttys) where polling for POLLIN |
|
|
401 | * fails but POLLIN|POLLOUT works. |
340 | */ |
402 | */ |
341 | struct iocb *iocb = linuxaio_submits [submitted]; |
403 | struct iocb *iocb = linuxaio_submits [submitted]; |
342 | res = 1; /* skip this iocb */ |
404 | res = 1; /* skip this iocb */ |
343 | |
405 | |
344 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
406 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
… | |
… | |
356 | /* second phase: fetch and parse events */ |
418 | /* second phase: fetch and parse events */ |
357 | |
419 | |
358 | linuxaio_get_events (EV_A_ timeout); |
420 | linuxaio_get_events (EV_A_ timeout); |
359 | } |
421 | } |
360 | |
422 | |
361 | #if EPOLL_FALLBACK |
|
|
362 | |
|
|
363 | static void |
|
|
364 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
365 | { |
|
|
366 | struct epoll_event events[16]; |
|
|
367 | |
|
|
368 | for (;;) |
|
|
369 | { |
|
|
370 | int idx; |
|
|
371 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
372 | |
|
|
373 | if (ecb_expect_false (res < 0)) |
|
|
374 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
375 | else if (!res) |
|
|
376 | break; |
|
|
377 | |
|
|
378 | for (idx = res; idx--; ) |
|
|
379 | { |
|
|
380 | int fd = events [idx].data.fd; |
|
|
381 | uint32_t ev = events [idx].events; |
|
|
382 | |
|
|
383 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
384 | |
|
|
385 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
386 | |
|
|
387 | fd_event (EV_A_ fd, |
|
|
388 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
389 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
390 | } |
|
|
391 | |
|
|
392 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
393 | break; |
|
|
394 | } |
|
|
395 | } |
|
|
396 | |
|
|
397 | #endif |
|
|
398 | |
|
|
399 | inline_size |
423 | inline_size |
400 | int |
424 | int |
401 | linuxaio_init (EV_P_ int flags) |
425 | linuxaio_init (EV_P_ int flags) |
402 | { |
426 | { |
403 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
427 | /* 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 */ |
428 | /* 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 */ |
429 | #if EPOLL_FALLBACK |
|
|
430 | /* 4.19 made epoll work */ |
|
|
431 | if (ev_linux_version () < 0x041300) |
406 | return 0; |
432 | return 0; |
|
|
433 | #else |
|
|
434 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
435 | if (ev_linux_version () < 0x041200) |
|
|
436 | return 0; |
|
|
437 | #endif |
407 | |
438 | |
408 | linuxaio_ctx = 0; |
439 | linuxaio_ctx = 0; |
409 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
440 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
410 | return 0; |
441 | return 0; |
411 | |
442 | |
… | |
… | |
416 | ev_io_destroy (linuxaio_ctx); |
447 | ev_io_destroy (linuxaio_ctx); |
417 | return 0; |
448 | return 0; |
418 | } |
449 | } |
419 | |
450 | |
420 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
451 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
452 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
421 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
453 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
454 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
422 | #endif |
455 | #endif |
423 | |
456 | |
424 | backend_modify = linuxaio_modify; |
457 | backend_modify = linuxaio_modify; |
425 | backend_poll = linuxaio_poll; |
458 | backend_poll = linuxaio_poll; |
426 | |
459 | |
… | |
… | |
462 | ev_syserr ("(libev) linuxaio epoll_create"); |
495 | ev_syserr ("(libev) linuxaio epoll_create"); |
463 | |
496 | |
464 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
497 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
465 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
498 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
466 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
499 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
467 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
468 | #endif |
500 | #endif |
469 | |
501 | |
470 | fd_rearm_all (EV_A); |
502 | fd_rearm_all (EV_A); |
471 | } |
503 | } |
472 | |
504 | |