… | |
… | |
116 | unsigned header_length; /* size of aio_ring */ |
116 | unsigned header_length; /* size of aio_ring */ |
117 | |
117 | |
118 | struct io_event io_events[0]; |
118 | struct io_event io_events[0]; |
119 | }; |
119 | }; |
120 | |
120 | |
|
|
121 | /* |
|
|
122 | * define some syscall wrappers for common architectures |
|
|
123 | * this is mostly for nice looks during debugging, not performance. |
|
|
124 | * our syscalls return < 0, not == -1, on error. which is good |
|
|
125 | * enough for linux aio. |
|
|
126 | * TODO: arm is also common nowadays, maybe even mips and x86 |
|
|
127 | * TODO: after implementing this, it suddenly looks like overkill, but its hard to remove... |
|
|
128 | */ |
|
|
129 | #if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__ |
|
|
130 | /* the costly errno access probably kills this for size optimisation */ |
|
|
131 | |
|
|
132 | #define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \ |
|
|
133 | ({ \ |
|
|
134 | long res; \ |
|
|
135 | register unsigned long r5 __asm__ ("r8" ); \ |
|
|
136 | register unsigned long r4 __asm__ ("r10"); \ |
|
|
137 | register unsigned long r3 __asm__ ("rdx"); \ |
|
|
138 | register unsigned long r2 __asm__ ("rsi"); \ |
|
|
139 | register unsigned long r1 __asm__ ("rdi"); \ |
|
|
140 | if (narg >= 5) r5 = (unsigned long)(arg5); \ |
|
|
141 | if (narg >= 4) r4 = (unsigned long)(arg4); \ |
|
|
142 | if (narg >= 3) r3 = (unsigned long)(arg3); \ |
|
|
143 | if (narg >= 2) r2 = (unsigned long)(arg2); \ |
|
|
144 | if (narg >= 1) r1 = (unsigned long)(arg1); \ |
|
|
145 | __asm__ __volatile__ ( \ |
|
|
146 | "syscall\n\t" \ |
|
|
147 | : "=a" (res) \ |
|
|
148 | : "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \ |
|
|
149 | : "cc", "r11", "cx", "memory"); \ |
|
|
150 | errno = -res; \ |
|
|
151 | res; \ |
|
|
152 | }) |
|
|
153 | |
|
|
154 | #endif |
|
|
155 | |
|
|
156 | #ifdef ev_syscall |
|
|
157 | #define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0 |
|
|
158 | #define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0) |
|
|
159 | #define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0) |
|
|
160 | #define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0) |
|
|
161 | #define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0) |
|
|
162 | #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5) |
|
|
163 | #else |
|
|
164 | #define ev_syscall0(nr) syscall (nr) |
|
|
165 | #define ev_syscall1(nr,arg1) syscall (nr, arg1) |
|
|
166 | #define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2) |
|
|
167 | #define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3) |
|
|
168 | #define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4) |
|
|
169 | #define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5) |
|
|
170 | #endif |
|
|
171 | |
121 | inline_size |
172 | inline_size |
122 | int |
173 | int |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
174 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
124 | { |
175 | { |
125 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
176 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
126 | } |
177 | } |
127 | |
178 | |
128 | inline_size |
179 | inline_size |
129 | int |
180 | int |
130 | evsys_io_destroy (aio_context_t ctx_id) |
181 | evsys_io_destroy (aio_context_t ctx_id) |
131 | { |
182 | { |
132 | return syscall (SYS_io_destroy, ctx_id); |
183 | return ev_syscall1 (SYS_io_destroy, ctx_id); |
133 | } |
184 | } |
134 | |
185 | |
135 | inline_size |
186 | inline_size |
136 | int |
187 | int |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
188 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
138 | { |
189 | { |
139 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
190 | return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
140 | } |
191 | } |
141 | |
192 | |
142 | inline_size |
193 | inline_size |
143 | int |
194 | int |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
195 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
145 | { |
196 | { |
146 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
197 | return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
147 | } |
198 | } |
148 | |
199 | |
149 | inline_size |
200 | inline_size |
150 | int |
201 | int |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
202 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
152 | { |
203 | { |
153 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
204 | return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
154 | } |
205 | } |
155 | |
206 | |
156 | /*****************************************************************************/ |
207 | /*****************************************************************************/ |
157 | /* actual backed implementation */ |
208 | /* actual backed implementation */ |
158 | |
209 | |
… | |
… | |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
289 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
239 | |
290 | |
240 | if (iocb->io.aio_reqprio < 0) |
291 | if (iocb->io.aio_reqprio < 0) |
241 | { |
292 | { |
242 | /* we handed this fd over to epoll, so undo this first */ |
293 | /* we handed this fd over to epoll, so undo this first */ |
243 | /* we do it manually because the optimisations on epoll_modfy won't do us any good */ |
294 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
295 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
|
|
296 | anfds [fd].emask = 0; |
245 | iocb->io.aio_reqprio = 0; |
297 | iocb->io.aio_reqprio = 0; |
246 | } |
298 | } |
247 | |
299 | |
248 | if (iocb->io.aio_buf) |
300 | if (iocb->io.aio_buf) |
249 | /* io_cancel always returns some error on relevant kernels, but works */ |
301 | { |
250 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
302 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
|
|
303 | /* on relevant kernels, io_cancel fails with EINPROGRES if everything is fine */ |
|
|
304 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS)); |
|
|
305 | } |
251 | |
306 | |
252 | if (nev) |
307 | if (nev) |
253 | { |
308 | { |
254 | iocb->io.aio_buf = |
309 | iocb->io.aio_buf = |
255 | (nev & EV_READ ? POLLIN : 0) |
310 | (nev & EV_READ ? POLLIN : 0) |
… | |
… | |
267 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
322 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
268 | { |
323 | { |
269 | epoll_poll (EV_A_ 0); |
324 | epoll_poll (EV_A_ 0); |
270 | } |
325 | } |
271 | |
326 | |
272 | static void |
327 | inline_speed |
|
|
328 | void |
273 | linuxaio_fd_rearm (EV_P_ int fd) |
329 | linuxaio_fd_rearm (EV_P_ int fd) |
274 | { |
330 | { |
275 | anfds [fd].events = 0; |
331 | anfds [fd].events = 0; |
276 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
332 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
277 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
333 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
… | |
… | |
293 | fd, |
349 | fd, |
294 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
350 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
295 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
351 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
296 | ); |
352 | ); |
297 | |
353 | |
298 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
354 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
299 | linuxaio_fd_rearm (EV_A_ fd); |
355 | linuxaio_fd_rearm (EV_A_ fd); |
300 | |
356 | |
301 | --nr; |
357 | --nr; |
302 | ++ev; |
358 | ++ev; |
303 | } |
359 | } |
… | |
… | |
373 | /* ignored */; |
429 | /* ignored */; |
374 | else |
430 | else |
375 | ev_syserr ("(libev) linuxaio io_getevents"); |
431 | ev_syserr ("(libev) linuxaio io_getevents"); |
376 | else if (res) |
432 | else if (res) |
377 | { |
433 | { |
378 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
434 | /* at least one event available, handle it and any remaining ones in the ring buffer */ |
379 | linuxaio_parse_events (EV_A_ ioev, res); |
435 | linuxaio_parse_events (EV_A_ ioev, res); |
380 | linuxaio_get_events_from_ring (EV_A); |
436 | linuxaio_get_events_from_ring (EV_A); |
381 | } |
437 | } |
382 | } |
438 | } |
383 | |
439 | |
384 | static int |
440 | inline_size |
|
|
441 | int |
385 | linuxaio_io_setup (EV_P) |
442 | linuxaio_io_setup (EV_P) |
386 | { |
443 | { |
387 | linuxaio_ctx = 0; |
444 | linuxaio_ctx = 0; |
388 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
445 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
389 | } |
446 | } |
… | |
… | |
455 | /* it's easiest to handle this mess in another iteration */ |
512 | /* it's easiest to handle this mess in another iteration */ |
456 | return; |
513 | return; |
457 | } |
514 | } |
458 | else if (errno == EBADF) |
515 | else if (errno == EBADF) |
459 | { |
516 | { |
|
|
517 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
460 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
518 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
461 | |
519 | |
462 | res = 1; /* skip this iocb */ |
520 | res = 1; /* skip this iocb */ |
463 | } |
521 | } |
464 | else |
522 | else |
… | |
… | |
517 | void |
575 | void |
518 | linuxaio_destroy (EV_P) |
576 | linuxaio_destroy (EV_P) |
519 | { |
577 | { |
520 | epoll_destroy (EV_A); |
578 | epoll_destroy (EV_A); |
521 | linuxaio_free_iocbp (EV_A); |
579 | linuxaio_free_iocbp (EV_A); |
522 | evsys_io_destroy (linuxaio_ctx); |
580 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
523 | } |
581 | } |
524 | |
582 | |
525 | inline_size |
583 | inline_size |
526 | void |
584 | void |
527 | linuxaio_fork (EV_P) |
585 | linuxaio_fork (EV_P) |
… | |
… | |
533 | linuxaio_iteration = 0; /* we start over in the child */ |
591 | linuxaio_iteration = 0; /* we start over in the child */ |
534 | |
592 | |
535 | while (linuxaio_io_setup (EV_A) < 0) |
593 | while (linuxaio_io_setup (EV_A) < 0) |
536 | ev_syserr ("(libev) linuxaio io_setup"); |
594 | ev_syserr ("(libev) linuxaio io_setup"); |
537 | |
595 | |
|
|
596 | /* forking epoll should also effectively unregister all fds from the backend */ |
538 | epoll_fork (EV_A); |
597 | epoll_fork (EV_A); |
539 | |
598 | |
540 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
599 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
541 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
600 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
542 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
601 | ev_io_start (EV_A_ &linuxaio_epoll_w); |