… | |
… | |
53 | * 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 |
54 | */ |
54 | */ |
55 | #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ |
55 | #define EV_LINUXAIO_DEPTH (128 / 2 - 2 - 1) /* max. number of io events per batch */ |
56 | |
56 | |
57 | /*****************************************************************************/ |
57 | /*****************************************************************************/ |
58 | /* syscall wrapdadoop */ |
58 | /* syscall wrapdadoop - this section has the raw syscall definitions */ |
59 | |
59 | |
60 | #include <sys/syscall.h> /* no glibc wrappers */ |
60 | #include <sys/syscall.h> /* no glibc wrappers */ |
61 | |
61 | |
62 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
62 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
63 | #define IOCB_CMD_POLL 5 |
63 | #define IOCB_CMD_POLL 5 |
… | |
… | |
80 | struct io_event io_events[0]; |
80 | struct io_event io_events[0]; |
81 | }; |
81 | }; |
82 | |
82 | |
83 | inline_size |
83 | inline_size |
84 | int |
84 | int |
85 | ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
85 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
86 | { |
86 | { |
87 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
87 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
88 | } |
88 | } |
89 | |
89 | |
90 | inline_size |
90 | inline_size |
91 | int |
91 | int |
92 | ev_io_destroy (aio_context_t ctx_id) |
92 | evsys_io_destroy (aio_context_t ctx_id) |
93 | { |
93 | { |
94 | return syscall (SYS_io_destroy, ctx_id); |
94 | return syscall (SYS_io_destroy, ctx_id); |
95 | } |
95 | } |
96 | |
96 | |
97 | inline_size |
97 | inline_size |
98 | int |
98 | int |
99 | ev_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
99 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
100 | { |
100 | { |
101 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
101 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
102 | } |
102 | } |
103 | |
103 | |
104 | inline_size |
104 | inline_size |
105 | int |
105 | int |
106 | ev_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
106 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
107 | { |
107 | { |
108 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
108 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
109 | } |
109 | } |
110 | |
110 | |
111 | inline_size |
111 | inline_size |
112 | int |
112 | int |
113 | ev_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
113 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
114 | { |
114 | { |
115 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
115 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
116 | } |
116 | } |
117 | |
117 | |
118 | /*****************************************************************************/ |
118 | /*****************************************************************************/ |
… | |
… | |
164 | iocb->io.aio_reqprio = 0; |
164 | iocb->io.aio_reqprio = 0; |
165 | } |
165 | } |
166 | #endif |
166 | #endif |
167 | |
167 | |
168 | if (iocb->io.aio_buf) |
168 | if (iocb->io.aio_buf) |
169 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
169 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
170 | |
170 | |
171 | if (nev) |
171 | if (nev) |
172 | { |
172 | { |
173 | iocb->io.aio_data = fd; |
173 | iocb->io.aio_data = fd; |
174 | iocb->io.aio_fildes = fd; |
174 | iocb->io.aio_fildes = fd; |
… | |
… | |
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)) |
|
|
204 | fd_kill (EV_A_ fd); |
|
|
205 | else |
|
|
206 | fd_event ( |
255 | fd_event ( |
207 | EV_A_ |
256 | EV_A_ |
208 | fd, |
257 | fd, |
209 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
258 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
210 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
259 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
211 | ); |
260 | ); |
212 | |
261 | |
213 | --nr; |
262 | --nr; |
214 | ++ev; |
263 | ++ev; |
215 | } |
264 | } |
216 | } |
265 | } |
… | |
… | |
219 | static int |
268 | static int |
220 | linuxaio_get_events_from_ring (EV_P) |
269 | linuxaio_get_events_from_ring (EV_P) |
221 | { |
270 | { |
222 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
271 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
223 | |
272 | |
224 | unsigned head = ring->head; |
273 | /* the kernel reads and writes both of these variables, */ |
|
|
274 | /* as a C extension, we assume that volatile use here */ |
|
|
275 | /* both makes reads atomic and once-only */ |
|
|
276 | unsigned head = *(volatile unsigned *)&ring->head; |
225 | unsigned tail = *(volatile unsigned *)&ring->tail; |
277 | unsigned tail = *(volatile unsigned *)&ring->tail; |
226 | |
278 | |
227 | if (head == tail) |
279 | if (head == tail) |
228 | return 0; |
280 | return 0; |
229 | |
281 | |
230 | /* bail out if the ring buffer doesn't match the expected layout */ |
282 | /* bail out if the ring buffer doesn't match the expected layout */ |
231 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
283 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
232 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
284 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
233 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
285 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
234 | return 0; |
286 | return 0; |
235 | |
287 | |
|
|
288 | /* make sure the events up to tail are visible */ |
236 | ECB_MEMORY_FENCE_ACQUIRE; |
289 | ECB_MEMORY_FENCE_ACQUIRE; |
237 | |
290 | |
238 | /* parse all available events, but only once, to avoid starvation */ |
291 | /* parse all available events, but only once, to avoid starvation */ |
239 | if (tail > head) /* normal case around */ |
292 | if (tail > head) /* normal case around */ |
240 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
293 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
242 | { |
295 | { |
243 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
296 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
244 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
297 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
245 | } |
298 | } |
246 | |
299 | |
|
|
300 | /* TODO: we only need a compiler barrier here, not a read fence */ |
|
|
301 | ECB_MEMORY_FENCE_RELEASE; |
|
|
302 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
247 | *(volatile unsigned *)&ring->head = tail; |
303 | *(volatile unsigned *)&ring->head = tail; |
248 | |
304 | /* 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; |
305 | ECB_MEMORY_FENCE_RELEASE; |
251 | |
306 | |
252 | return 1; |
307 | return 1; |
253 | } |
308 | } |
254 | |
309 | |
… | |
… | |
256 | inline_size |
311 | inline_size |
257 | void |
312 | void |
258 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
313 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
259 | { |
314 | { |
260 | struct timespec ts; |
315 | struct timespec ts; |
261 | struct io_event ioev; |
316 | struct io_event ioev[1]; |
262 | int res; |
317 | int res; |
263 | |
318 | |
264 | if (linuxaio_get_events_from_ring (EV_A)) |
319 | if (linuxaio_get_events_from_ring (EV_A)) |
265 | return; |
320 | return; |
266 | |
321 | |
267 | /* no events, so wait for at least one, then poll ring buffer again */ |
322 | /* no events, so wait for at least one, then poll ring buffer again */ |
268 | /* this degrades to one event per loop iteration */ |
323 | /* this degrades to one event per loop iteration */ |
269 | /* if the ring buffer changes layout, but so be it */ |
324 | /* if the ring buffer changes layout, but so be it */ |
270 | |
325 | |
|
|
326 | EV_RELEASE_CB; |
|
|
327 | |
271 | ts.tv_sec = (long)timeout; |
328 | ts.tv_sec = (long)timeout; |
272 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
329 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
273 | |
330 | |
274 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
331 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
332 | |
|
|
333 | EV_ACQUIRE_CB; |
275 | |
334 | |
276 | if (res < 0) |
335 | if (res < 0) |
277 | if (errno == EINTR) |
336 | if (errno == EINTR) |
278 | /* ignored */; |
337 | /* ignored */; |
279 | else |
338 | else |
280 | ev_syserr ("(libev) linuxaio io_getevents"); |
339 | ev_syserr ("(libev) linuxaio io_getevents"); |
281 | else if (res) |
340 | else if (res) |
282 | { |
341 | { |
283 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
342 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
284 | linuxaio_parse_events (EV_A_ &ioev, 1); |
343 | linuxaio_parse_events (EV_A_ ioev, res); |
285 | linuxaio_get_events_from_ring (EV_A); |
344 | linuxaio_get_events_from_ring (EV_A); |
286 | } |
345 | } |
287 | } |
346 | } |
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 | |
347 | |
305 | static void |
348 | static void |
306 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
349 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
307 | { |
350 | { |
308 | int submitted; |
351 | int submitted; |
… | |
… | |
312 | /* io_submit might return less than the requested number of iocbs */ |
355 | /* 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, */ |
356 | /* 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 */ |
357 | /* which allows us to pinpoint the errornous iocb */ |
315 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
358 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
316 | { |
359 | { |
|
|
360 | #if 0 |
|
|
361 | int res; |
|
|
362 | if (linuxaio_submits[submitted]->aio_fildes == backend_fd) |
|
|
363 | res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); |
|
|
364 | else |
|
|
365 | { res = -1; errno = EINVAL; }; |
|
|
366 | #else |
317 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
367 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
|
|
368 | #endif |
318 | |
369 | |
319 | if (ecb_expect_false (res < 0)) |
370 | if (expect_false (res < 0)) |
320 | if (errno == EAGAIN) |
371 | if (errno == EAGAIN) |
321 | { |
372 | { |
322 | /* This happens when the ring buffer is full, at least. I assume this means |
373 | /* 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 |
374 | * that the event was queued synchronously during io_submit, and thus |
324 | * the buffer overflowed. |
375 | * the buffer overflowed. |
325 | * In this case, we just try in next loop iteration. |
376 | * In this case, we just try in next loop iteration. |
326 | * This should not result in a few fds taking priority, as the interface |
377 | * 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. |
378 | * is one-shot, and we submit iocb's in a round-robin fashion. |
|
|
379 | * TODO: maybe make "submitted" persistent, so we don't have to memmove? |
328 | */ |
380 | */ |
|
|
381 | if (ecb_expect_false (submitted)) |
|
|
382 | { |
329 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
383 | memmove (linuxaio_submits, linuxaio_submits + submitted, (linuxaio_submitcnt - submitted) * sizeof (*linuxaio_submits)); |
330 | linuxaio_submitcnt -= submitted; |
384 | linuxaio_submitcnt -= submitted; |
|
|
385 | } |
|
|
386 | |
331 | timeout = 0; |
387 | timeout = 0; |
332 | break; |
388 | break; |
333 | } |
389 | } |
334 | #if EPOLL_FALLBACK |
390 | #if EPOLL_FALLBACK |
335 | else if (errno == EINVAL) |
391 | else if (errno == EINVAL) |
336 | { |
392 | { |
337 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
393 | /* This happens for unsupported fds, officially, but in my testing, |
338 | * also randomly happens for supported fds. We fall back to good old |
394 | * 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. |
395 | * poll() here, under the assumption that this is a very rare case. |
|
|
396 | * See https://lore.kernel.org/patchwork/patch/1047453/ to see |
|
|
397 | * discussion about such a case (ttys) where polling for POLLIN |
|
|
398 | * fails but POLLIN|POLLOUT works. |
340 | */ |
399 | */ |
341 | struct iocb *iocb = linuxaio_submits [submitted]; |
400 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
401 | |
|
|
402 | linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD); |
|
|
403 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
404 | |
342 | res = 1; /* skip this iocb */ |
405 | res = 1; /* skip this iocb */ |
343 | |
|
|
344 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
345 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
346 | } |
406 | } |
347 | #endif |
407 | #endif |
|
|
408 | else if (errno == EBADF) |
|
|
409 | { |
|
|
410 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
|
|
411 | |
|
|
412 | res = 1; /* skip this iocb */ |
|
|
413 | } |
348 | else |
414 | else |
349 | ev_syserr ("(libev) linuxaio io_submit"); |
415 | ev_syserr ("(libev) linuxaio io_submit"); |
350 | |
416 | |
351 | submitted += res; |
417 | submitted += res; |
352 | } |
418 | } |
… | |
… | |
356 | /* second phase: fetch and parse events */ |
422 | /* second phase: fetch and parse events */ |
357 | |
423 | |
358 | linuxaio_get_events (EV_A_ timeout); |
424 | linuxaio_get_events (EV_A_ timeout); |
359 | } |
425 | } |
360 | |
426 | |
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 |
427 | inline_size |
400 | int |
428 | int |
401 | linuxaio_init (EV_P_ int flags) |
429 | linuxaio_init (EV_P_ int flags) |
402 | { |
430 | { |
403 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
431 | /* 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 */ |
432 | /* 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 */ |
433 | #if EPOLL_FALLBACK |
|
|
434 | /* 4.19 made epoll work */ |
|
|
435 | if (ev_linux_version () < 0x041300) |
406 | return 0; |
436 | return 0; |
|
|
437 | #else |
|
|
438 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
439 | if (ev_linux_version () < 0x041200) |
|
|
440 | return 0; |
|
|
441 | #endif |
407 | |
442 | |
408 | linuxaio_ctx = 0; |
443 | linuxaio_ctx = 0; |
409 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
444 | if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
410 | return 0; |
445 | return 0; |
411 | |
446 | |
412 | #if EPOLL_FALLBACK |
447 | #if EPOLL_FALLBACK |
413 | backend_fd = ev_epoll_create (); |
448 | backend_fd = ev_epoll_create (); |
414 | if (backend_fd < 0) |
449 | if (backend_fd < 0) |
415 | { |
450 | { |
416 | ev_io_destroy (linuxaio_ctx); |
451 | evsys_io_destroy (linuxaio_ctx); |
417 | return 0; |
452 | return 0; |
418 | } |
453 | } |
419 | |
454 | |
420 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
455 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
456 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
421 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
457 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
458 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
422 | #endif |
459 | #endif |
423 | |
460 | |
424 | backend_modify = linuxaio_modify; |
461 | backend_modify = linuxaio_modify; |
425 | backend_poll = linuxaio_poll; |
462 | backend_poll = linuxaio_poll; |
426 | |
463 | |
… | |
… | |
440 | { |
477 | { |
441 | #if EPOLL_FALLBACK |
478 | #if EPOLL_FALLBACK |
442 | close (backend_fd); |
479 | close (backend_fd); |
443 | #endif |
480 | #endif |
444 | linuxaio_free_iocbp (EV_A); |
481 | linuxaio_free_iocbp (EV_A); |
445 | ev_io_destroy (linuxaio_ctx); |
482 | evsys_io_destroy (linuxaio_ctx); |
446 | } |
483 | } |
447 | |
484 | |
448 | inline_size |
485 | inline_size |
449 | void |
486 | void |
450 | linuxaio_fork (EV_P) |
487 | linuxaio_fork (EV_P) |
… | |
… | |
452 | /* this frees all iocbs, which is very heavy-handed */ |
489 | /* this frees all iocbs, which is very heavy-handed */ |
453 | linuxaio_destroy (EV_A); |
490 | linuxaio_destroy (EV_A); |
454 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
491 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
455 | |
492 | |
456 | linuxaio_ctx = 0; |
493 | linuxaio_ctx = 0; |
457 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
494 | while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
458 | ev_syserr ("(libev) linuxaio io_setup"); |
495 | ev_syserr ("(libev) linuxaio io_setup"); |
459 | |
496 | |
460 | #if EPOLL_FALLBACK |
497 | #if EPOLL_FALLBACK |
461 | while ((backend_fd = ev_epoll_create ()) < 0) |
498 | while ((backend_fd = ev_epoll_create ()) < 0) |
462 | ev_syserr ("(libev) linuxaio epoll_create"); |
499 | ev_syserr ("(libev) linuxaio epoll_create"); |
463 | |
500 | |
464 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
501 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
465 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
502 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
466 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
503 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
467 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
468 | #endif |
504 | #endif |
469 | |
505 | |
470 | fd_rearm_all (EV_A); |
506 | fd_rearm_all (EV_A); |
471 | } |
507 | } |
472 | |
508 | |