… | |
… | |
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 | /*****************************************************************************/ |
… | |
… | |
125 | /*int inuse;*/ |
125 | /*int inuse;*/ |
126 | } *ANIOCBP; |
126 | } *ANIOCBP; |
127 | |
127 | |
128 | inline_size |
128 | inline_size |
129 | void |
129 | void |
130 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int count) |
130 | linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
131 | { |
131 | { |
132 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
|
|
133 | while (count--) |
132 | while (count--) |
134 | { |
133 | { |
|
|
134 | /* TODO: quite the overhead to allocate every iocb separately, maybe use our own alocator? */ |
135 | *base = (ANIOCBP)ev_malloc (sizeof (**base)); |
135 | ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
136 | /* TODO: full zero initialize required? */ |
136 | |
|
|
137 | /* full zero initialise is probably not required at the moment, but |
|
|
138 | * this is not well documented, so we better do it. |
|
|
139 | */ |
137 | memset (*base, 0, sizeof (**base)); |
140 | memset (iocb, 0, sizeof (*iocb)); |
138 | /* would be nice to initialize fd/data as well, but array_needsize API doesn't support that */ |
141 | |
139 | (*base)->io.aio_lio_opcode = IOCB_CMD_POLL; |
142 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
140 | ++base; |
143 | iocb->io.aio_data = offset; |
|
|
144 | iocb->io.aio_fildes = offset; |
|
|
145 | |
|
|
146 | base [offset++] = iocb; |
141 | } |
147 | } |
142 | } |
148 | } |
143 | |
149 | |
144 | ecb_cold |
150 | ecb_cold |
145 | static void |
151 | static void |
… | |
… | |
153 | |
159 | |
154 | static void |
160 | static void |
155 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
161 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
156 | { |
162 | { |
157 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
163 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
158 | struct aniocb *iocb = linuxaio_iocbps [fd]; |
164 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
159 | |
165 | |
160 | #if EPOLL_FALLBACK |
166 | #if EPOLL_FALLBACK |
161 | if (iocb->io.aio_reqprio < 0) |
167 | if (iocb->io.aio_reqprio < 0) |
162 | { |
168 | { |
163 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
169 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
164 | iocb->io.aio_reqprio = 0; |
170 | iocb->io.aio_reqprio = 0; |
165 | } |
171 | } |
166 | #endif |
172 | #endif |
167 | |
173 | |
168 | if (iocb->io.aio_buf) |
174 | if (iocb->io.aio_buf) |
169 | ev_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
175 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); /* always returns an error relevant kernels */ |
170 | |
176 | |
171 | if (nev) |
177 | if (nev) |
172 | { |
178 | { |
173 | iocb->io.aio_data = fd; |
|
|
174 | iocb->io.aio_fildes = fd; |
|
|
175 | iocb->io.aio_buf = |
179 | iocb->io.aio_buf = |
176 | (nev & EV_READ ? POLLIN : 0) |
180 | (nev & EV_READ ? POLLIN : 0) |
177 | | (nev & EV_WRITE ? POLLOUT : 0); |
181 | | (nev & EV_WRITE ? POLLOUT : 0); |
178 | |
182 | |
179 | /* queue iocb up for io_submit */ |
183 | /* queue iocb up for io_submit */ |
180 | /* this assumes we only ever get one call per fd per loop iteration */ |
184 | /* this assumes we only ever get one call per fd per loop iteration */ |
… | |
… | |
182 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
186 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
187 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
184 | } |
188 | } |
185 | } |
189 | } |
186 | |
190 | |
|
|
191 | #if EPOLL_FALLBACK |
|
|
192 | |
|
|
193 | static void |
|
|
194 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
195 | { |
|
|
196 | struct epoll_event eev; |
|
|
197 | |
|
|
198 | eev.events = EPOLLONESHOT; |
|
|
199 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
200 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
201 | eev.data.fd = iocb->aio_fildes; |
|
|
202 | |
|
|
203 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
204 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
205 | } |
|
|
206 | |
|
|
207 | static void |
|
|
208 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
209 | { |
|
|
210 | struct epoll_event events[16]; |
|
|
211 | |
|
|
212 | for (;;) |
|
|
213 | { |
|
|
214 | int idx; |
|
|
215 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
216 | |
|
|
217 | if (expect_false (res < 0)) |
|
|
218 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
219 | else if (!res) |
|
|
220 | break; |
|
|
221 | |
|
|
222 | for (idx = res; idx--; ) |
|
|
223 | { |
|
|
224 | int fd = events [idx].data.fd; |
|
|
225 | uint32_t ev = events [idx].events; |
|
|
226 | |
|
|
227 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
228 | |
|
|
229 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
230 | |
|
|
231 | fd_event (EV_A_ fd, |
|
|
232 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
233 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
234 | } |
|
|
235 | |
|
|
236 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
237 | break; |
|
|
238 | } |
|
|
239 | } |
|
|
240 | |
|
|
241 | #endif |
|
|
242 | |
187 | static void |
243 | static void |
188 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
244 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
189 | { |
245 | { |
190 | while (nr) |
246 | while (nr) |
191 | { |
247 | { |
192 | int fd = ev->data; |
248 | int fd = ev->data; |
193 | int res = ev->res; |
249 | int res = ev->res; |
194 | |
250 | |
195 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
251 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
196 | |
252 | |
197 | /* linux aio is oneshot: rearm fd */ |
253 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
198 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
254 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
199 | anfds [fd].events = 0; |
255 | anfds [fd].events = 0; |
200 | fd_change (EV_A_ fd, 0); |
256 | fd_change (EV_A_ fd, 0); |
201 | |
257 | |
202 | /* feed events, we do not expect or handle POLLNVAL */ |
258 | /* 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 ( |
259 | fd_event ( |
207 | EV_A_ |
260 | EV_A_ |
208 | fd, |
261 | fd, |
209 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
262 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
210 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
263 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
211 | ); |
264 | ); |
212 | |
265 | |
213 | --nr; |
266 | --nr; |
214 | ++ev; |
267 | ++ev; |
215 | } |
268 | } |
216 | } |
269 | } |
… | |
… | |
229 | |
282 | |
230 | if (head == tail) |
283 | if (head == tail) |
231 | return 0; |
284 | return 0; |
232 | |
285 | |
233 | /* bail out if the ring buffer doesn't match the expected layout */ |
286 | /* bail out if the ring buffer doesn't match the expected layout */ |
234 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
287 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
235 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
288 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
236 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
289 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
237 | return 0; |
290 | return 0; |
238 | |
291 | |
239 | /* make sure the events up to tail are visible */ |
292 | /* make sure the events up to tail are visible */ |
240 | ECB_MEMORY_FENCE_ACQUIRE; |
293 | ECB_MEMORY_FENCE_ACQUIRE; |
241 | |
294 | |
… | |
… | |
246 | { |
299 | { |
247 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
300 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
248 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
301 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
249 | } |
302 | } |
250 | |
303 | |
|
|
304 | ECB_MEMORY_FENCE_RELAXED; |
251 | /* as an extension to C, we hope that the volatile will makethis atomic and once-only */ |
305 | /* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
252 | *(volatile unsigned *)&ring->head = tail; |
306 | *(volatile unsigned *)&ring->head = tail; |
253 | /* make sure kernel can see our new head value - probably not required */ |
307 | /* make sure kernel can see our new head value - probably not required */ |
254 | ECB_MEMORY_FENCE_RELEASE; |
308 | ECB_MEMORY_FENCE_RELEASE; |
255 | |
309 | |
256 | return 1; |
310 | return 1; |
… | |
… | |
260 | inline_size |
314 | inline_size |
261 | void |
315 | void |
262 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
316 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
263 | { |
317 | { |
264 | struct timespec ts; |
318 | struct timespec ts; |
265 | struct io_event ioev; |
319 | struct io_event ioev[1]; |
266 | int res; |
320 | int res; |
267 | |
321 | |
268 | if (linuxaio_get_events_from_ring (EV_A)) |
322 | if (linuxaio_get_events_from_ring (EV_A)) |
269 | return; |
323 | return; |
270 | |
324 | |
271 | /* 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 */ |
272 | /* this degrades to one event per loop iteration */ |
326 | /* this degrades to one event per loop iteration */ |
273 | /* if the ring buffer changes layout, but so be it */ |
327 | /* if the ring buffer changes layout, but so be it */ |
274 | |
328 | |
|
|
329 | EV_RELEASE_CB; |
|
|
330 | |
275 | ts.tv_sec = (long)timeout; |
331 | ts.tv_sec = (long)timeout; |
276 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
332 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
277 | |
333 | |
278 | res = ev_io_getevents (linuxaio_ctx, 1, 1, &ioev, &ts); |
334 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
|
|
335 | |
|
|
336 | EV_ACQUIRE_CB; |
279 | |
337 | |
280 | if (res < 0) |
338 | if (res < 0) |
281 | if (errno == EINTR) |
339 | if (errno == EINTR) |
282 | /* ignored */; |
340 | /* ignored */; |
283 | else |
341 | else |
284 | ev_syserr ("(libev) linuxaio io_getevents"); |
342 | ev_syserr ("(libev) linuxaio io_getevents"); |
285 | else if (res) |
343 | else if (res) |
286 | { |
344 | { |
287 | /* 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 */ |
288 | linuxaio_parse_events (EV_A_ &ioev, 1); |
346 | linuxaio_parse_events (EV_A_ ioev, res); |
289 | linuxaio_get_events_from_ring (EV_A); |
347 | linuxaio_get_events_from_ring (EV_A); |
290 | } |
348 | } |
291 | } |
349 | } |
292 | |
|
|
293 | #if EPOLL_FALLBACK |
|
|
294 | static void |
|
|
295 | linuxaio_rearm_epoll (EV_P_ struct iocb *iocb, int op) |
|
|
296 | { |
|
|
297 | struct epoll_event eev; |
|
|
298 | |
|
|
299 | eev.events = EPOLLONESHOT; |
|
|
300 | if (iocb->aio_buf & POLLIN ) eev.events |= EPOLLIN ; |
|
|
301 | if (iocb->aio_buf & POLLOUT) eev.events |= EPOLLOUT; |
|
|
302 | eev.data.fd = iocb->aio_fildes; |
|
|
303 | |
|
|
304 | if (epoll_ctl (backend_fd, op, iocb->aio_fildes, &eev) < 0) |
|
|
305 | ev_syserr ("(libeio) linuxaio epoll_ctl"); |
|
|
306 | } |
|
|
307 | #endif |
|
|
308 | |
350 | |
309 | static void |
351 | static void |
310 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
352 | linuxaio_poll (EV_P_ ev_tstamp timeout) |
311 | { |
353 | { |
312 | int submitted; |
354 | int submitted; |
… | |
… | |
316 | /* io_submit might return less than the requested number of iocbs */ |
358 | /* io_submit might return less than the requested number of iocbs */ |
317 | /* 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, */ |
318 | /* which allows us to pinpoint the errornous iocb */ |
360 | /* which allows us to pinpoint the errornous iocb */ |
319 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
361 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
320 | { |
362 | { |
|
|
363 | #if 0 |
|
|
364 | int res; |
|
|
365 | if (linuxaio_submits[submitted]->aio_fildes == backend_fd) |
|
|
366 | res = evsys_io_submit (linuxaio_ctx, 1, linuxaio_submits + submitted); |
|
|
367 | else |
|
|
368 | { res = -1; errno = EINVAL; }; |
|
|
369 | #else |
321 | int res = ev_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
370 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
|
|
371 | #endif |
322 | |
372 | |
323 | if (ecb_expect_false (res < 0)) |
373 | if (expect_false (res < 0)) |
324 | if (errno == EAGAIN) |
374 | if (errno == EAGAIN) |
325 | { |
375 | { |
326 | /* 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 |
327 | * that the event was queued synchronously during io_submit, and thus |
377 | * that the event was queued synchronously during io_submit, and thus |
328 | * the buffer overflowed. |
378 | * the buffer overflowed. |
329 | * In this case, we just try in next loop iteration. |
379 | * In this case, we just try in next loop iteration. |
330 | * 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 |
331 | * 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? |
332 | */ |
383 | */ |
|
|
384 | if (ecb_expect_false (submitted)) |
|
|
385 | { |
333 | 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)); |
334 | linuxaio_submitcnt -= submitted; |
387 | linuxaio_submitcnt -= submitted; |
|
|
388 | } |
|
|
389 | |
335 | timeout = 0; |
390 | timeout = 0; |
336 | break; |
391 | break; |
337 | } |
392 | } |
338 | #if EPOLL_FALLBACK |
393 | #if EPOLL_FALLBACK |
339 | else if (errno == EINVAL) |
394 | else if (errno == EINVAL) |
340 | { |
395 | { |
341 | /* This hapΓΌpens for unsupported fds, officially, but in my testing, |
396 | /* This happens for unsupported fds, officially, but in my testing, |
342 | * 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 |
343 | * 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. |
344 | */ |
402 | */ |
345 | struct iocb *iocb = linuxaio_submits [submitted]; |
403 | struct iocb *iocb = linuxaio_submits [submitted]; |
|
|
404 | |
|
|
405 | linuxaio_rearm_epoll (EV_A_ linuxaio_submits [submitted], EPOLL_CTL_ADD); |
|
|
406 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
407 | |
346 | res = 1; /* skip this iocb */ |
408 | res = 1; /* skip this iocb */ |
347 | |
|
|
348 | linuxaio_rearm_epoll (EV_A_ iocb, EPOLL_CTL_ADD); |
|
|
349 | iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
|
|
350 | } |
409 | } |
351 | #endif |
410 | #endif |
|
|
411 | else if (errno == EBADF) |
|
|
412 | { |
|
|
413 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
|
|
414 | |
|
|
415 | res = 1; /* skip this iocb */ |
|
|
416 | } |
352 | else |
417 | else |
353 | ev_syserr ("(libev) linuxaio io_submit"); |
418 | ev_syserr ("(libev) linuxaio io_submit"); |
354 | |
419 | |
355 | submitted += res; |
420 | submitted += res; |
356 | } |
421 | } |
… | |
… | |
360 | /* second phase: fetch and parse events */ |
425 | /* second phase: fetch and parse events */ |
361 | |
426 | |
362 | linuxaio_get_events (EV_A_ timeout); |
427 | linuxaio_get_events (EV_A_ timeout); |
363 | } |
428 | } |
364 | |
429 | |
365 | #if EPOLL_FALLBACK |
|
|
366 | |
|
|
367 | static void |
|
|
368 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
|
|
369 | { |
|
|
370 | struct epoll_event events[16]; |
|
|
371 | |
|
|
372 | for (;;) |
|
|
373 | { |
|
|
374 | int idx; |
|
|
375 | int res = epoll_wait (backend_fd, events, sizeof (events) / sizeof (events [0]), 0); |
|
|
376 | |
|
|
377 | if (ecb_expect_false (res < 0)) |
|
|
378 | ev_syserr ("(libev) linuxaio epoll_wait"); |
|
|
379 | else if (!res) |
|
|
380 | break; |
|
|
381 | |
|
|
382 | for (idx = res; idx--; ) |
|
|
383 | { |
|
|
384 | int fd = events [idx].data.fd; |
|
|
385 | uint32_t ev = events [idx].events; |
|
|
386 | |
|
|
387 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
|
|
388 | |
|
|
389 | linuxaio_rearm_epoll (EV_A_ &linuxaio_iocbps [fd]->io, EPOLL_CTL_MOD); |
|
|
390 | |
|
|
391 | fd_event (EV_A_ fd, |
|
|
392 | (ev & (EPOLLOUT | EPOLLERR | EPOLLHUP) ? EV_WRITE : 0) |
|
|
393 | | (ev & (EPOLLIN | EPOLLERR | EPOLLHUP) ? EV_READ : 0)); |
|
|
394 | } |
|
|
395 | |
|
|
396 | if (res < sizeof (events) / sizeof (events [0])) |
|
|
397 | break; |
|
|
398 | } |
|
|
399 | } |
|
|
400 | |
|
|
401 | #endif |
|
|
402 | |
|
|
403 | inline_size |
430 | inline_size |
404 | int |
431 | int |
405 | linuxaio_init (EV_P_ int flags) |
432 | linuxaio_init (EV_P_ int flags) |
406 | { |
433 | { |
407 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
434 | /* would be great to have a nice test for IOCB_CMD_POLL instead */ |
408 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
435 | /* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
409 | if (ev_linux_version () < 0x041200) /* 4.18 introduced IOCB_CMD_POLL */ |
436 | #if EPOLL_FALLBACK |
|
|
437 | /* 4.19 made epoll work */ |
|
|
438 | if (ev_linux_version () < 0x041300) |
410 | return 0; |
439 | return 0; |
|
|
440 | #else |
|
|
441 | /* 4.18 introduced IOCB_CMD_POLL */ |
|
|
442 | if (ev_linux_version () < 0x041200) |
|
|
443 | return 0; |
|
|
444 | #endif |
411 | |
445 | |
412 | linuxaio_ctx = 0; |
446 | linuxaio_ctx = 0; |
413 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
447 | if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
414 | return 0; |
448 | return 0; |
415 | |
449 | |
416 | #if EPOLL_FALLBACK |
450 | #if EPOLL_FALLBACK |
417 | backend_fd = ev_epoll_create (); |
451 | backend_fd = ev_epoll_create (); |
418 | if (backend_fd < 0) |
452 | if (backend_fd < 0) |
419 | { |
453 | { |
420 | ev_io_destroy (linuxaio_ctx); |
454 | evsys_io_destroy (linuxaio_ctx); |
421 | return 0; |
455 | return 0; |
422 | } |
456 | } |
423 | |
457 | |
424 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
458 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
459 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
425 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
460 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
461 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
426 | #endif |
462 | #endif |
427 | |
463 | |
428 | backend_modify = linuxaio_modify; |
464 | backend_modify = linuxaio_modify; |
429 | backend_poll = linuxaio_poll; |
465 | backend_poll = linuxaio_poll; |
430 | |
466 | |
… | |
… | |
444 | { |
480 | { |
445 | #if EPOLL_FALLBACK |
481 | #if EPOLL_FALLBACK |
446 | close (backend_fd); |
482 | close (backend_fd); |
447 | #endif |
483 | #endif |
448 | linuxaio_free_iocbp (EV_A); |
484 | linuxaio_free_iocbp (EV_A); |
449 | ev_io_destroy (linuxaio_ctx); |
485 | evsys_io_destroy (linuxaio_ctx); |
450 | } |
486 | } |
451 | |
487 | |
452 | inline_size |
488 | inline_size |
453 | void |
489 | void |
454 | linuxaio_fork (EV_P) |
490 | linuxaio_fork (EV_P) |
… | |
… | |
456 | /* this frees all iocbs, which is very heavy-handed */ |
492 | /* this frees all iocbs, which is very heavy-handed */ |
457 | linuxaio_destroy (EV_A); |
493 | linuxaio_destroy (EV_A); |
458 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
494 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
459 | |
495 | |
460 | linuxaio_ctx = 0; |
496 | linuxaio_ctx = 0; |
461 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
497 | while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
462 | ev_syserr ("(libev) linuxaio io_setup"); |
498 | ev_syserr ("(libev) linuxaio io_setup"); |
463 | |
499 | |
464 | #if EPOLL_FALLBACK |
500 | #if EPOLL_FALLBACK |
465 | while ((backend_fd = ev_epoll_create ()) < 0) |
501 | while ((backend_fd = ev_epoll_create ()) < 0) |
466 | ev_syserr ("(libev) linuxaio epoll_create"); |
502 | ev_syserr ("(libev) linuxaio epoll_create"); |
467 | |
503 | |
468 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
504 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
469 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
505 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
470 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
506 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
471 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
472 | #endif |
507 | #endif |
473 | |
508 | |
474 | fd_rearm_all (EV_A); |
509 | fd_rearm_all (EV_A); |
475 | } |
510 | } |
476 | |
511 | |