… | |
… | |
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 |
48 | */ |
54 | */ |
49 | #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 */ |
50 | |
56 | |
51 | /*****************************************************************************/ |
57 | /*****************************************************************************/ |
52 | /* syscall wrapdadoop */ |
58 | /* syscall wrapdadoop - this section has the raw syscall definitions */ |
53 | |
59 | |
54 | #include <sys/syscall.h> /* no glibc wrappers */ |
60 | #include <sys/syscall.h> /* no glibc wrappers */ |
55 | |
61 | |
56 | /* 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 */ |
57 | #define IOCB_CMD_POLL 5 |
63 | #define IOCB_CMD_POLL 5 |
… | |
… | |
74 | struct io_event io_events[0]; |
80 | struct io_event io_events[0]; |
75 | }; |
81 | }; |
76 | |
82 | |
77 | inline_size |
83 | inline_size |
78 | int |
84 | int |
79 | ev_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
85 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
80 | { |
86 | { |
81 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
87 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
82 | } |
88 | } |
83 | |
89 | |
84 | inline_size |
90 | inline_size |
85 | int |
91 | int |
86 | ev_io_destroy (aio_context_t ctx_id) |
92 | evsys_io_destroy (aio_context_t ctx_id) |
87 | { |
93 | { |
88 | return syscall (SYS_io_destroy, ctx_id); |
94 | return syscall (SYS_io_destroy, ctx_id); |
89 | } |
95 | } |
90 | |
96 | |
91 | inline_size |
97 | inline_size |
92 | int |
98 | int |
93 | 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[]) |
94 | { |
100 | { |
95 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
101 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
96 | } |
102 | } |
97 | |
103 | |
98 | inline_size |
104 | inline_size |
99 | int |
105 | int |
100 | 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) |
101 | { |
107 | { |
102 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
108 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
103 | } |
109 | } |
104 | |
110 | |
105 | inline_size |
111 | inline_size |
106 | int |
112 | int |
107 | 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) |
108 | { |
114 | { |
109 | 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); |
110 | } |
116 | } |
111 | |
117 | |
112 | /*****************************************************************************/ |
118 | /*****************************************************************************/ |
… | |
… | |
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 | evsys_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 | { |
159 | iocb->io.aio_data = fd; |
173 | iocb->io.aio_data = fd; |
160 | iocb->io.aio_fildes = fd; |
174 | iocb->io.aio_fildes = fd; |
… | |
… | |
168 | 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); |
169 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
183 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
170 | } |
184 | } |
171 | } |
185 | } |
172 | |
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 | |
173 | static void |
239 | static void |
174 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
240 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
175 | { |
241 | { |
176 | while (nr) |
242 | while (nr) |
177 | { |
243 | { |
178 | int fd = ev->data; |
244 | int fd = ev->data; |
179 | int res = ev->res; |
245 | int res = ev->res; |
180 | |
246 | |
181 | 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)); |
182 | |
248 | |
183 | /* linux aio is oneshot: rearm fd */ |
249 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
184 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
250 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
185 | anfds [fd].events = 0; |
251 | anfds [fd].events = 0; |
186 | fd_change (EV_A_ fd, 0); |
252 | fd_change (EV_A_ fd, 0); |
187 | |
253 | |
188 | /* feed events, we do not expect or handle POLLNVAL */ |
254 | /* feed events, we do not expect or handle POLLNVAL */ |
189 | if (ecb_expect_false (res & POLLNVAL)) |
|
|
190 | fd_kill (EV_A_ fd); |
|
|
191 | else |
|
|
192 | fd_event ( |
255 | fd_event ( |
193 | EV_A_ |
256 | EV_A_ |
194 | fd, |
257 | fd, |
195 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
258 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
196 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
259 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
197 | ); |
260 | ); |
198 | |
261 | |
199 | --nr; |
262 | --nr; |
200 | ++ev; |
263 | ++ev; |
201 | } |
264 | } |
202 | } |
265 | } |
… | |
… | |
205 | static int |
268 | static int |
206 | linuxaio_get_events_from_ring (EV_P) |
269 | linuxaio_get_events_from_ring (EV_P) |
207 | { |
270 | { |
208 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
271 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
209 | |
272 | |
210 | 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; |
211 | unsigned tail = *(volatile unsigned *)&ring->tail; |
277 | unsigned tail = *(volatile unsigned *)&ring->tail; |
212 | |
278 | |
213 | if (head == tail) |
279 | if (head == tail) |
214 | return 0; |
280 | return 0; |
215 | |
281 | |
216 | /* 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 */ |
217 | if (ecb_expect_false (ring->magic != AIO_RING_MAGIC) |
283 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
218 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
284 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
219 | || 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]? */ |
220 | return 0; |
286 | return 0; |
221 | |
287 | |
|
|
288 | /* make sure the events up to tail are visible */ |
222 | ECB_MEMORY_FENCE_ACQUIRE; |
289 | ECB_MEMORY_FENCE_ACQUIRE; |
223 | |
290 | |
224 | /* parse all available events, but only once, to avoid starvation */ |
291 | /* parse all available events, but only once, to avoid starvation */ |
225 | if (tail > head) /* normal case around */ |
292 | if (tail > head) /* normal case around */ |
226 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
293 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
… | |
… | |
228 | { |
295 | { |
229 | 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); |
230 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
297 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
231 | } |
298 | } |
232 | |
299 | |
233 | ring->head = tail; |
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 */ |
|
|
303 | *(volatile unsigned *)&ring->head = tail; |
|
|
304 | /* make sure kernel can see our new head value - probably not required */ |
|
|
305 | ECB_MEMORY_FENCE_RELEASE; |
234 | |
306 | |
235 | return 1; |
307 | return 1; |
236 | } |
308 | } |
237 | |
309 | |
238 | /* read at least one event from kernel, or timeout */ |
310 | /* read at least one event from kernel, or timeout */ |
239 | inline_size |
311 | inline_size |
240 | void |
312 | void |
241 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
313 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
242 | { |
314 | { |
243 | struct timespec ts; |
315 | struct timespec ts; |
244 | struct io_event ioev; |
316 | struct io_event ioev[1]; |
245 | int res; |
317 | int res; |
246 | |
318 | |
247 | if (linuxaio_get_events_from_ring (EV_A)) |
319 | if (linuxaio_get_events_from_ring (EV_A)) |
248 | return; |
320 | return; |
249 | |
321 | |
250 | /* 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 */ |
251 | /* this degrades to one event per loop iteration */ |
323 | /* this degrades to one event per loop iteration */ |
252 | /* if the ring buffer changes layout, but so be it */ |
324 | /* if the ring buffer changes layout, but so be it */ |
253 | |
325 | |
|
|
326 | EV_RELEASE_CB; |
|
|
327 | |
254 | ts.tv_sec = (long)timeout; |
328 | ts.tv_sec = (long)timeout; |
255 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
329 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
256 | |
330 | |
257 | 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; |
258 | |
334 | |
259 | if (res < 0) |
335 | if (res < 0) |
|
|
336 | if (errno == EINTR) |
|
|
337 | /* ignored */; |
|
|
338 | else |
260 | ev_syserr ("(libev) linuxaio io_getevents"); |
339 | ev_syserr ("(libev) linuxaio io_getevents"); |
261 | else if (res) |
340 | else if (res) |
262 | { |
341 | { |
263 | /* 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 */ |
264 | linuxaio_parse_events (EV_A_ &ioev, 1); |
343 | linuxaio_parse_events (EV_A_ ioev, res); |
265 | linuxaio_get_events_from_ring (EV_A); |
344 | linuxaio_get_events_from_ring (EV_A); |
266 | } |
345 | } |
267 | } |
346 | } |
268 | |
347 | |
269 | static void |
348 | static void |
… | |
… | |
276 | /* io_submit might return less than the requested number of iocbs */ |
355 | /* io_submit might return less than the requested number of iocbs */ |
277 | /* 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, */ |
278 | /* which allows us to pinpoint the errornous iocb */ |
357 | /* which allows us to pinpoint the errornous iocb */ |
279 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
358 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
280 | { |
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 |
281 | 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 |
282 | |
369 | |
283 | if (ecb_expect_false (res < 0)) |
370 | if (expect_false (res < 0)) |
284 | if (errno == EAGAIN) |
371 | if (errno == EAGAIN) |
285 | { |
372 | { |
286 | /* 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 |
287 | * that the event was queued synchronously during io_submit, and thus |
374 | * that the event was queued synchronously during io_submit, and thus |
288 | * the buffer overflowd. |
375 | * the buffer overflowed. |
289 | * In this case, we just try next loop iteration. |
376 | * In this case, we just try in next loop iteration. |
290 | * 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 |
291 | * 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? |
292 | */ |
380 | */ |
|
|
381 | if (ecb_expect_false (submitted)) |
|
|
382 | { |
293 | 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)); |
294 | linuxaio_submitcnt -= submitted; |
384 | linuxaio_submitcnt -= submitted; |
|
|
385 | } |
|
|
386 | |
295 | timeout = 0; |
387 | timeout = 0; |
296 | break; |
388 | break; |
297 | } |
389 | } |
|
|
390 | #if EPOLL_FALLBACK |
|
|
391 | else if (errno == EINVAL) |
|
|
392 | { |
|
|
393 | /* This happens for unsupported fds, officially, but in my testing, |
|
|
394 | * also randomly happens for supported fds. We fall back to good old |
|
|
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. |
|
|
399 | */ |
|
|
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 | |
|
|
405 | res = 1; /* skip this iocb */ |
|
|
406 | } |
|
|
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 | } |
298 | else |
414 | else |
299 | ev_syserr ("(libev) linuxaio io_submit"); |
415 | ev_syserr ("(libev) linuxaio io_submit"); |
300 | |
416 | |
301 | submitted += res; |
417 | submitted += res; |
302 | } |
418 | } |
… | |
… | |
312 | int |
428 | int |
313 | linuxaio_init (EV_P_ int flags) |
429 | linuxaio_init (EV_P_ int flags) |
314 | { |
430 | { |
315 | /* 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 */ |
316 | /* 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 */ |
317 | 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) |
318 | 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 |
319 | |
442 | |
320 | linuxaio_ctx = 0; |
443 | linuxaio_ctx = 0; |
321 | if (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
444 | if (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
322 | return 0; |
445 | return 0; |
|
|
446 | |
|
|
447 | #if EPOLL_FALLBACK |
|
|
448 | backend_fd = ev_epoll_create (); |
|
|
449 | if (backend_fd < 0) |
|
|
450 | { |
|
|
451 | evsys_io_destroy (linuxaio_ctx); |
|
|
452 | return 0; |
|
|
453 | } |
|
|
454 | |
|
|
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); |
|
|
457 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
458 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
|
|
459 | #endif |
323 | |
460 | |
324 | backend_modify = linuxaio_modify; |
461 | backend_modify = linuxaio_modify; |
325 | backend_poll = linuxaio_poll; |
462 | backend_poll = linuxaio_poll; |
326 | |
463 | |
327 | linuxaio_iocbpmax = 0; |
464 | linuxaio_iocbpmax = 0; |
… | |
… | |
336 | |
473 | |
337 | inline_size |
474 | inline_size |
338 | void |
475 | void |
339 | linuxaio_destroy (EV_P) |
476 | linuxaio_destroy (EV_P) |
340 | { |
477 | { |
|
|
478 | #if EPOLL_FALLBACK |
|
|
479 | close (backend_fd); |
|
|
480 | #endif |
341 | linuxaio_free_iocbp (EV_A); |
481 | linuxaio_free_iocbp (EV_A); |
342 | ev_io_destroy (linuxaio_ctx); |
482 | evsys_io_destroy (linuxaio_ctx); |
343 | } |
483 | } |
344 | |
484 | |
345 | inline_size |
485 | inline_size |
346 | void |
486 | void |
347 | linuxaio_fork (EV_P) |
487 | linuxaio_fork (EV_P) |
… | |
… | |
349 | /* this frees all iocbs, which is very heavy-handed */ |
489 | /* this frees all iocbs, which is very heavy-handed */ |
350 | linuxaio_destroy (EV_A); |
490 | linuxaio_destroy (EV_A); |
351 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
491 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
352 | |
492 | |
353 | linuxaio_ctx = 0; |
493 | linuxaio_ctx = 0; |
354 | while (ev_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
494 | while (evsys_io_setup (EV_LINUXAIO_DEPTH, &linuxaio_ctx) < 0) |
355 | ev_syserr ("(libev) linuxaio io_setup"); |
495 | ev_syserr ("(libev) linuxaio io_setup"); |
356 | |
496 | |
|
|
497 | #if EPOLL_FALLBACK |
|
|
498 | while ((backend_fd = ev_epoll_create ()) < 0) |
|
|
499 | ev_syserr ("(libev) linuxaio epoll_create"); |
|
|
500 | |
|
|
501 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
|
|
502 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
|
|
503 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
|
|
504 | #endif |
|
|
505 | |
357 | fd_rearm_all (EV_A); |
506 | fd_rearm_all (EV_A); |
358 | } |
507 | } |
359 | |
508 | |