… | |
… | |
35 | * and other provisions required by the GPL. If you do not delete the |
35 | * and other provisions required by the GPL. If you do not delete the |
36 | * provisions above, a recipient may use your version of this file under |
36 | * provisions above, a recipient may use your version of this file under |
37 | * either the BSD or the GPL. |
37 | * either the BSD or the GPL. |
38 | */ |
38 | */ |
39 | |
39 | |
|
|
40 | #define EPOLL_FALLBACK 1 |
|
|
41 | |
40 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
42 | #include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
41 | #include <poll.h> |
43 | #include <poll.h> |
42 | #include <linux/aio_abi.h> |
44 | #include <linux/aio_abi.h> |
|
|
45 | |
|
|
46 | #if EPOLL_FALLBACK |
|
|
47 | # include <sys/epoll.h> |
|
|
48 | #endif |
43 | |
49 | |
44 | /* we try to fill 4kB pages exactly. |
50 | /* we try to fill 4kB pages exactly. |
45 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
51 | * the ring buffer header is 32 bytes, every io event is 32 bytes. |
46 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. |
52 | * the kernel takes the io event number, doubles it, adds 2, adds the ring buffer. |
47 | * therefore the calculation below will use "exactly" 4kB for the ring buffer |
53 | * therefore the calculation below will use "exactly" 4kB for the ring buffer |
… | |
… | |
149 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
155 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
150 | { |
156 | { |
151 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
157 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
152 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
158 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
153 | |
159 | |
|
|
160 | #if EPOLL_FALLBACK |
|
|
161 | if (iocb->io.aio_reqprio < 0) |
|
|
162 | { |
|
|
163 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
164 | iocb->io.aio_reqprio = 0; |
|
|
165 | } |
|
|
166 | #endif |
|
|
167 | |
154 | if (iocb->io.aio_buf) |
168 | if (iocb->io.aio_buf) |
155 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
169 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
156 | |
170 | |
157 | if (nev) |
171 | if (nev) |
158 | { |
172 | { |
… | |
… | |
184 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
185 | anfds [fd].events = 0; |
199 | anfds [fd].events = 0; |
186 | fd_change (EV_A_ fd, 0); |
200 | fd_change (EV_A_ fd, 0); |
187 | |
201 | |
188 | /* feed events, we do not expect or handle POLLNVAL */ |
202 | /* feed events, we do not expect or handle POLLNVAL */ |
189 | if (ecb_expect_false (res & POLLNVAL)) |
203 | if (expect_false (res & POLLNVAL)) |
190 | fd_kill (EV_A_ fd); |
204 | fd_kill (EV_A_ fd); |
191 | else |
205 | else |
192 | fd_event ( |
206 | fd_event ( |
193 | EV_A_ |
207 | EV_A_ |
194 | fd, |
208 | fd, |
… | |
… | |
205 | static int |
219 | static int |
206 | linuxaio_get_events_from_ring (EV_P) |
220 | linuxaio_get_events_from_ring (EV_P) |
207 | { |
221 | { |
208 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
209 | |
223 | |
210 | 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; |
211 | unsigned tail = *(volatile unsigned *)&ring->tail; |
228 | unsigned tail = *(volatile unsigned *)&ring->tail; |
212 | |
229 | |
213 | if (head == tail) |
230 | if (head == tail) |
214 | return 0; |
231 | return 0; |
215 | |
232 | |
216 | /* 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 */ |
217 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
234 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
218 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
235 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
219 | || 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]? */ |
220 | return 0; |
237 | return 0; |
221 | |
238 | |
|
|
239 | /* make sure the events up to tail are visible */ |
222 | ECB_MEMORY_FENCE_ACQUIRE; |
240 | ECB_MEMORY_FENCE_ACQUIRE; |
223 | |
241 | |
224 | /* parse all available events, but only once, to avoid starvation */ |
242 | /* parse all available events, but only once, to avoid starvation */ |
225 | if (tail > head) /* normal case around */ |
243 | if (tail > head) /* normal case around */ |
226 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
228 | { |
246 | { |
229 | 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); |
230 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
248 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
231 | } |
249 | } |
232 | |
250 | |
233 | 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; |
234 | |
257 | |
235 | return 1; |
258 | return 1; |
236 | } |
259 | } |
237 | |
260 | |
238 | /* read at least one event from kernel, or timeout */ |
261 | /* read at least one event from kernel, or timeout */ |
… | |
… | |
255 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
278 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
256 | |
279 | |
257 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
280 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
258 | |
281 | |
259 | if (res < 0) |
282 | if (res < 0) |
|
|
283 | if (errno == EINTR) |
|
|
284 | /* ignored */; |
|
|
285 | else |
260 | ev_syserr ("(libev) linuxaio io_getevents"); |
286 | ev_syserr ("(libev) linuxaio io_getevents"); |
261 | else if (res) |
287 | else if (res) |
262 | { |
288 | { |
263 | /* 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 */ |
264 | linuxaio_parse_events (EV_A_ &ioev, 1); |
290 | linuxaio_parse_events (EV_A_ &ioev, 1); |
265 | linuxaio_get_events_from_ring (EV_A); |
291 | linuxaio_get_events_from_ring (EV_A); |
266 | } |
292 | } |
267 | } |
293 | } |
|
|
294 | |
|
|
295 | #if EPOLL_FALLBACK |
|
|
296 | static void |
|
|
297 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
298 | { |
|
|
299 | struct epoll_event eev; |
|
|
300 | |
|
|
301 | eev.events = EPOLLONESHOT; |
|
|
302 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
303 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
304 | eev.data.fd = iocb->aio_fildes; |
|
|
305 | |
|
|
306 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
307 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
308 | } |
|
|
309 | #endif |
268 | |
310 | |
269 | static void |
311 | static void |
270 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
312 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
271 | { |
313 | { |
272 | int submitted; |
314 | int submitted; |
… | |
… | |
278 | /* which allows us to pinpoint the errornous iocb */ |
320 | /* which allows us to pinpoint the errornous iocb */ |
279 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
321 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
280 | { |
322 | { |
281 | 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); |
282 | |
324 | |
283 | if (ecb_expect_false (res < 0)) |
325 | if (expect_false (res < 0)) |
284 | if (errno == EAGAIN) |
326 | if (errno == EAGAIN) |
285 | { |
327 | { |
286 | /* 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 |
287 | * that the event was queued synchronously during io_submit, and thus |
329 | * that the event was queued synchronously during io_submit, and thus |
288 | * the buffer overflowd. |
330 | * the buffer overflowed. |
289 | * In this case, we just try next loop iteration. |
331 | * In this case, we just try in next loop iteration. |
290 | * This should not result in a few fds taking priority, as the interface |
332 | * This should not result in a few fds taking priority, as the interface |
291 | * is one-shot, and we submit iocb's in a round-robin fashion. |
333 | * is one-shot, and we submit iocb's in a round-robin fashion. |
292 | */ |
334 | */ |
293 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
335 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
294 | linuxaio_submitcnt -= submitted; |
336 | linuxaio_submitcnt -= submitted; |
295 | timeout = 0; |
337 | timeout = 0; |
296 | break; |
338 | break; |
297 | } |
339 | } |
|
|
340 | #if EPOLL_FALLBACK |
|
|
341 | else if (errno == EINVAL) |
|
|
342 | { |
|
|
343 | /* This happens for unsupported fds, officially, but in my testing, |
|
|
344 | * also randomly happens for supported fds. We fall back to good old |
|
|
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. |
|
|
348 | */ |
|
|
349 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
350 | res = 1; /* skip this iocb */ |
|
|
351 | |
|
|
352 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
353 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
354 | } |
|
|
355 | #endif |
298 | else |
356 | else |
299 | ev_syserr ("(libev) linuxaio io_submit"); |
357 | ev_syserr ("(libev) linuxaio io_submit"); |
300 | |
358 | |
301 | submitted += res; |
359 | submitted += res; |
302 | } |
360 | } |
… | |
… | |
306 | /* second phase: fetch and parse events */ |
364 | /* second phase: fetch and parse events */ |
307 | |
365 | |
308 | linuxaio_get_events (EV_A_ timeout); |
366 | linuxaio_get_events (EV_A_ timeout); |
309 | } |
367 | } |
310 | |
368 | |
|
|
369 | #if EPOLL_FALLBACK |
|
|
370 | |
|
|
371 | static void |
|
|
372 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
373 | { |
|
|
374 | struct epoll_event events[16]; |
|
|
375 | |
|
|
376 | for (;;) |
|
|
377 | { |
|
|
378 | int idx; |
|
|
379 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
380 | |
|
|
381 | if (expect_false (res < 0)) |
|
|
382 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
383 | else if (!res) |
|
|
384 | break; |
|
|
385 | |
|
|
386 | for (idx = res; idx--; ) |
|
|
387 | { |
|
|
388 | int fd = events [idx].data.fd; |
|
|
389 | uint32_t ev = events [idx].events; |
|
|
390 | |
|
|
391 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
392 | |
|
|
393 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
394 | |
|
|
395 | fd_event (EV_A_ fd, |
|
|
396 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
397 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
398 | } |
|
|
399 | |
|
|
400 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
401 | break; |
|
|
402 | } |
|
|
403 | } |
|
|
404 | |
|
|
405 | #endif |
|
|
406 | |
311 | inline_size |
407 | inline_size |
312 | int |
408 | int |
313 | linuxaio_init (EV_P_ int flags) |
409 | linuxaio_init (EV_P_ int flags) |
314 | { |
410 | { |
315 | /* 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 */ |
316 | /* 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 */ |
317 | 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) |
318 | 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 |
319 | |
422 | |
320 | linuxaio_ctx = 0; |
423 | linuxaio_ctx = 0; |
321 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
424 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
322 | return 0; |
425 | return 0; |
323 | |
426 | |
|
|
427 | #if EPOLL_FALLBACK |
|
|
428 | backend_fd = ev_epoll_create (); |
|
|
429 | if (backend_fd < 0) |
|
|
430 | { |
|
|
431 | ev_io_destroy (linuxaio_ctx); |
|
|
432 | return 0; |
|
|
433 | } |
|
|
434 | |
|
|
435 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
436 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
437 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
438 | #endif |
|
|
439 | |
324 | backend_modify = linuxaio_modify; |
440 | backend_modify = linuxaio_modify; |
325 | backend_poll = linuxaio_poll; |
441 | backend_poll = linuxaio_poll; |
326 | |
442 | |
327 | linuxaio_iocbpmax = 0; |
443 | linuxaio_iocbpmax = 0; |
328 | linuxaio_iocbps = 0; |
444 | linuxaio_iocbps = 0; |
… | |
… | |
336 | |
452 | |
337 | inline_size |
453 | inline_size |
338 | void |
454 | void |
339 | linuxaio_destroy (EV_P) |
455 | linuxaio_destroy (EV_P) |
340 | { |
456 | { |
|
|
457 | #if EPOLL_FALLBACK |
|
|
458 | close (backend_fd); |
|
|
459 | #endif |
341 | linuxaio_free_iocbp (EV_A); |
460 | linuxaio_free_iocbp (EV_A); |
342 | ev_io_destroy (linuxaio_ctx); |
461 | ev_io_destroy (linuxaio_ctx); |
343 | } |
462 | } |
344 | |
463 | |
345 | inline_size |
464 | inline_size |
… | |
… | |
352 | |
471 | |
353 | linuxaio_ctx = 0; |
472 | linuxaio_ctx = 0; |
354 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
473 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
355 | ev_syserr ("(libev) linuxaio io_setup"); |
474 | ev_syserr ("(libev) linuxaio io_setup"); |
356 | |
475 | |
|
|
476 | #if EPOLL_FALLBACK |
|
|
477 | while ((backend_fd = ev_epoll_create ()) < 0) |
|
|
478 | ev_syserr ("(libev) linuxaio epoll_create"); |
|
|
479 | |
|
|
480 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
481 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
482 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
483 | #endif |
|
|
484 | |
357 | fd_rearm_all (EV_A); |
485 | fd_rearm_all (EV_A); |
358 | } |
486 | } |
359 | |
487 | |