… | |
… | |
168 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
168 | int fd = (uint32_t)ev->data.u64; /* mask out the lower 32 bits */ |
169 | int want = anfds [fd].events; |
169 | int want = anfds [fd].events; |
170 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
170 | int got = (ev->events & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
171 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
171 | | (ev->events & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0); |
172 | |
172 | |
|
|
173 | /* |
173 | /* check for spurious notification */ |
174 | * check for spurious notification. |
|
|
175 | * this only finds spurious notifications on egen updates |
|
|
176 | * other spurious notifications will be found by epoll_ctl, below |
174 | /* we assume that fd is always in range, as we never shrink the anfds array */ |
177 | * we assume that fd is always in range, as we never shrink the anfds array |
|
|
178 | */ |
175 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
179 | if (expect_false ((uint32_t)anfds [fd].egen != (uint32_t)(ev->data.u64 >> 32))) |
176 | { |
180 | { |
177 | /* recreate kernel state */ |
181 | /* recreate kernel state */ |
178 | postfork = 1; |
182 | postfork |= 2; |
179 | continue; |
183 | continue; |
180 | } |
184 | } |
181 | |
185 | |
182 | if (expect_false (got & ~want)) |
186 | if (expect_false (got & ~want)) |
183 | { |
187 | { |
184 | anfds [fd].emask = want; |
188 | anfds [fd].emask = want; |
185 | |
189 | |
|
|
190 | /* |
186 | /* we received an event but are not interested in it, try mod or del */ |
191 | * we received an event but are not interested in it, try mod or del |
187 | /* I don't think we ever need MOD, but let's handle it anyways */ |
192 | * this often happens because we optimistically do not unregister fds |
|
|
193 | * when we are no longer interested in them, but also when we get spurious |
|
|
194 | * notifications for fds from another process. this is partially handled |
|
|
195 | * above with the gencounter check (== our fd is not the event fd), and |
|
|
196 | * partially here, when epoll_ctl returns an error (== a child has the fd |
|
|
197 | * but we closed it). |
|
|
198 | */ |
188 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
199 | ev->events = (want & EV_READ ? EPOLLIN : 0) |
189 | | (want & EV_WRITE ? EPOLLOUT : 0); |
200 | | (want & EV_WRITE ? EPOLLOUT : 0); |
190 | |
201 | |
191 | /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ |
202 | /* pre-2.6.9 kernels require a non-null pointer with EPOLL_CTL_DEL, */ |
192 | /* which is fortunately easy to do for us. */ |
203 | /* which is fortunately easy to do for us. */ |
193 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
204 | if (epoll_ctl (backend_fd, want ? EPOLL_CTL_MOD : EPOLL_CTL_DEL, fd, ev)) |
194 | { |
205 | { |
195 | postfork = 1; /* an error occurred, recreate kernel state */ |
206 | postfork |= 2; /* an error occurred, recreate kernel state */ |
196 | continue; |
207 | continue; |
197 | } |
208 | } |
198 | } |
209 | } |
199 | |
210 | |
200 | fd_event (EV_A_ fd, got); |
211 | fd_event (EV_A_ fd, got); |
… | |
… | |
215 | unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE); |
226 | unsigned char events = anfds [fd].events & (EV_READ | EV_WRITE); |
216 | |
227 | |
217 | if (anfds [fd].emask & EV_EMASK_EPERM && events) |
228 | if (anfds [fd].emask & EV_EMASK_EPERM && events) |
218 | fd_event (EV_A_ fd, events); |
229 | fd_event (EV_A_ fd, events); |
219 | else |
230 | else |
|
|
231 | { |
220 | epoll_eperms [i] = epoll_eperms [--epoll_epermcnt]; |
232 | epoll_eperms [i] = epoll_eperms [--epoll_epermcnt]; |
|
|
233 | anfds [fd].emask = 0; |
|
|
234 | } |
221 | } |
235 | } |
222 | } |
236 | } |
223 | |
237 | |
224 | int inline_size |
238 | inline_size |
|
|
239 | int |
225 | epoll_init (EV_P_ int flags) |
240 | epoll_init (EV_P_ int flags) |
226 | { |
241 | { |
227 | #ifdef EPOLL_CLOEXEC |
242 | #ifdef EPOLL_CLOEXEC |
228 | backend_fd = epoll_create1 (EPOLL_CLOEXEC); |
243 | backend_fd = epoll_create1 (EPOLL_CLOEXEC); |
229 | |
244 | |
230 | if (backend_fd <= 0) |
245 | if (backend_fd < 0 && (errno == EINVAL || errno == ENOSYS)) |
231 | #endif |
246 | #endif |
232 | backend_fd = epoll_create (256); |
247 | backend_fd = epoll_create (256); |
233 | |
248 | |
234 | if (backend_fd < 0) |
249 | if (backend_fd < 0) |
235 | return 0; |
250 | return 0; |
236 | |
251 | |
237 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
252 | fcntl (backend_fd, F_SETFD, FD_CLOEXEC); |
238 | |
253 | |
239 | backend_mintime = 1./1024.; /* epoll does sometimes return early, this is just to avoid the worst */ |
254 | backend_mintime = 1e-3; /* epoll does sometimes return early, this is just to avoid the worst */ |
240 | backend_modify = epoll_modify; |
255 | backend_modify = epoll_modify; |
241 | backend_poll = epoll_poll; |
256 | backend_poll = epoll_poll; |
242 | |
257 | |
243 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
258 | epoll_eventmax = 64; /* initial number of events receivable per poll */ |
244 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
259 | epoll_events = (struct epoll_event *)ev_malloc (sizeof (struct epoll_event) * epoll_eventmax); |
245 | |
260 | |
246 | return EVBACKEND_EPOLL; |
261 | return EVBACKEND_EPOLL; |
247 | } |
262 | } |
248 | |
263 | |
249 | void inline_size |
264 | inline_size |
|
|
265 | void |
250 | epoll_destroy (EV_P) |
266 | epoll_destroy (EV_P) |
251 | { |
267 | { |
252 | ev_free (epoll_events); |
268 | ev_free (epoll_events); |
253 | array_free (epoll_eperm, EMPTY); |
269 | array_free (epoll_eperm, EMPTY); |
254 | } |
270 | } |
255 | |
271 | |
256 | void inline_size |
272 | inline_size |
|
|
273 | void |
257 | epoll_fork (EV_P) |
274 | epoll_fork (EV_P) |
258 | { |
275 | { |
259 | close (backend_fd); |
276 | close (backend_fd); |
260 | |
277 | |
261 | while ((backend_fd = epoll_create (256)) < 0) |
278 | while ((backend_fd = epoll_create (256)) < 0) |