… | |
… | |
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 | |
|
|
172 | inline_size |
121 | inline_size |
173 | int |
122 | int |
174 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
175 | { |
124 | { |
176 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
125 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
… | |
… | |
286 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
287 | { |
236 | { |
288 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
237 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
289 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
290 | |
239 | |
291 | if (iocb->io.aio_reqprio < 0) |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
292 | { |
241 | { |
293 | /* we handed this fd over to epoll, so undo this first */ |
242 | /* we handed this fd over to epoll, so undo this first */ |
294 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
243 | /* we do it manually because the optimisations on epoll_modify won't do us any good */ |
295 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
296 | anfds [fd].emask = 0; |
245 | anfds [fd].emask = 0; |
297 | iocb->io.aio_reqprio = 0; |
246 | iocb->io.aio_reqprio = 0; |
298 | } |
247 | } |
299 | |
248 | else if (ecb_expect_false (iocb->io.aio_buf)) |
300 | if (iocb->io.aio_buf) |
|
|
301 | { |
249 | { |
|
|
250 | /* iocb active, so cancel it first before resubmit */ |
302 | for (;;) |
251 | for (;;) |
303 | { |
252 | { |
304 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
253 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
305 | if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 09) |
254 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
306 | break; |
255 | break; |
307 | |
256 | |
308 | if (expect_true (errno == EINPROGRESS)) |
257 | if (ecb_expect_true (errno == EINPROGRESS)) |
309 | break; |
258 | break; |
310 | |
259 | |
311 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
260 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
261 | if (errno != EINTR) |
|
|
262 | { |
312 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
263 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS)); |
|
|
264 | break; |
|
|
265 | } |
313 | } |
266 | } |
314 | } |
267 | } |
315 | |
268 | |
|
|
269 | iocb->io.aio_buf = |
|
|
270 | (nev & EV_READ ? POLLIN : 0) |
|
|
271 | | (nev & EV_WRITE ? POLLOUT : 0); |
|
|
272 | |
316 | if (nev) |
273 | if (nev) |
317 | { |
274 | { |
318 | iocb->io.aio_buf = |
|
|
319 | (nev & EV_READ ? POLLIN : 0) |
|
|
320 | | (nev & EV_WRITE ? POLLOUT : 0); |
|
|
321 | |
|
|
322 | /* queue iocb up for io_submit */ |
275 | /* queue iocb up for io_submit */ |
323 | /* this assumes we only ever get one call per fd per loop iteration */ |
276 | /* this assumes we only ever get one call per fd per loop iteration */ |
324 | ++linuxaio_submitcnt; |
277 | ++linuxaio_submitcnt; |
325 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
278 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
326 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
279 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
… | |
… | |
371 | /* get any events from ring buffer, return true if any were handled */ |
324 | /* get any events from ring buffer, return true if any were handled */ |
372 | static int |
325 | static int |
373 | linuxaio_get_events_from_ring (EV_P) |
326 | linuxaio_get_events_from_ring (EV_P) |
374 | { |
327 | { |
375 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
328 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
329 | unsigned head, tail; |
376 | |
330 | |
377 | /* the kernel reads and writes both of these variables, */ |
331 | /* the kernel reads and writes both of these variables, */ |
378 | /* as a C extension, we assume that volatile use here */ |
332 | /* as a C extension, we assume that volatile use here */ |
379 | /* both makes reads atomic and once-only */ |
333 | /* both makes reads atomic and once-only */ |
380 | unsigned head = *(volatile unsigned *)&ring->head; |
334 | head = *(volatile unsigned *)&ring->head; |
|
|
335 | ECB_MEMORY_FENCE_ACQUIRE; |
381 | unsigned tail = *(volatile unsigned *)&ring->tail; |
336 | tail = *(volatile unsigned *)&ring->tail; |
382 | |
337 | |
383 | if (head == tail) |
338 | if (head == tail) |
384 | return 0; |
339 | return 0; |
385 | |
340 | |
386 | /* make sure the events up to tail are visible */ |
|
|
387 | ECB_MEMORY_FENCE_ACQUIRE; |
|
|
388 | |
|
|
389 | /* parse all available events, but only once, to avoid starvation */ |
341 | /* parse all available events, but only once, to avoid starvation */ |
390 | if (tail > head) /* normal case around */ |
342 | if (ecb_expect_true (tail > head)) /* normal case around */ |
391 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
343 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
392 | else /* wrapped around */ |
344 | else /* wrapped around */ |
393 | { |
345 | { |
394 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
346 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
395 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
347 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
… | |
… | |
406 | int |
358 | int |
407 | linuxaio_ringbuf_valid (EV_P) |
359 | linuxaio_ringbuf_valid (EV_P) |
408 | { |
360 | { |
409 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
361 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
410 | |
362 | |
411 | return expect_true (ring->magic == AIO_RING_MAGIC) |
363 | return ecb_expect_true (ring->magic == AIO_RING_MAGIC) |
412 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
364 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
413 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
365 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
414 | } |
366 | } |
415 | |
367 | |
416 | /* read at least one event from kernel, or timeout */ |
368 | /* read at least one event from kernel, or timeout */ |
… | |
… | |
421 | struct timespec ts; |
373 | struct timespec ts; |
422 | struct io_event ioev[8]; /* 256 octet stack space */ |
374 | struct io_event ioev[8]; /* 256 octet stack space */ |
423 | int want = 1; /* how many events to request */ |
375 | int want = 1; /* how many events to request */ |
424 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
376 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
425 | |
377 | |
426 | if (expect_true (ringbuf_valid)) |
378 | if (ecb_expect_true (ringbuf_valid)) |
427 | { |
379 | { |
428 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
380 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
429 | if (linuxaio_get_events_from_ring (EV_A)) |
381 | if (linuxaio_get_events_from_ring (EV_A)) |
430 | return; |
382 | return; |
431 | |
383 | |
… | |
… | |
444 | { |
396 | { |
445 | int res; |
397 | int res; |
446 | |
398 | |
447 | EV_RELEASE_CB; |
399 | EV_RELEASE_CB; |
448 | |
400 | |
449 | ts.tv_sec = (long)timeout; |
401 | EV_TS_SET (ts, timeout); |
450 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
|
|
451 | |
|
|
452 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
402 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
453 | |
403 | |
454 | EV_ACQUIRE_CB; |
404 | EV_ACQUIRE_CB; |
455 | |
405 | |
456 | if (res < 0) |
406 | if (res < 0) |
… | |
… | |
461 | else if (res) |
411 | else if (res) |
462 | { |
412 | { |
463 | /* at least one event available, handle them */ |
413 | /* at least one event available, handle them */ |
464 | linuxaio_parse_events (EV_A_ ioev, res); |
414 | linuxaio_parse_events (EV_A_ ioev, res); |
465 | |
415 | |
466 | if (expect_true (ringbuf_valid)) |
416 | if (ecb_expect_true (ringbuf_valid)) |
467 | { |
417 | { |
468 | /* if we have a ring buffer, handle any remaining events in it */ |
418 | /* if we have a ring buffer, handle any remaining events in it */ |
469 | linuxaio_get_events_from_ring (EV_A); |
419 | linuxaio_get_events_from_ring (EV_A); |
470 | |
420 | |
471 | /* at this point, we should have handled all outstanding events */ |
421 | /* at this point, we should have handled all outstanding events */ |
… | |
… | |
502 | /* which allows us to pinpoint the erroneous iocb */ |
452 | /* which allows us to pinpoint the erroneous iocb */ |
503 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
453 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
504 | { |
454 | { |
505 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
455 | int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
506 | |
456 | |
507 | if (expect_false (res < 0)) |
457 | if (ecb_expect_false (res < 0)) |
508 | if (errno == EINVAL) |
458 | if (errno == EINVAL) |
509 | { |
459 | { |
510 | /* This happens for unsupported fds, officially, but in my testing, |
460 | /* This happens for unsupported fds, officially, but in my testing, |
511 | * also randomly happens for supported fds. We fall back to good old |
461 | * also randomly happens for supported fds. We fall back to good old |
512 | * poll() here, under the assumption that this is a very rare case. |
462 | * poll() here, under the assumption that this is a very rare case. |
… | |
… | |
542 | } |
492 | } |
543 | |
493 | |
544 | ++linuxaio_iteration; |
494 | ++linuxaio_iteration; |
545 | if (linuxaio_io_setup (EV_A) < 0) |
495 | if (linuxaio_io_setup (EV_A) < 0) |
546 | { |
496 | { |
|
|
497 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
498 | /* TODO: might be more prudent? */ |
|
|
499 | |
547 | /* to bad, we can't get a new aio context, go 100% epoll */ |
500 | /* to bad, we can't get a new aio context, go 100% epoll */ |
548 | linuxaio_free_iocbp (EV_A); |
501 | linuxaio_free_iocbp (EV_A); |
549 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
502 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
550 | ev_ref (EV_A); |
503 | ev_ref (EV_A); |
551 | linuxaio_ctx = 0; |
504 | linuxaio_ctx = 0; |
|
|
505 | |
|
|
506 | backend = EVBACKEND_EPOLL; |
552 | backend_modify = epoll_modify; |
507 | backend_modify = epoll_modify; |
553 | backend_poll = epoll_poll; |
508 | backend_poll = epoll_poll; |
554 | } |
509 | } |
555 | |
510 | |
556 | timeout = 0; |
511 | timeout = 0; |
… | |
… | |
565 | res = 1; /* skip this iocb */ |
520 | res = 1; /* skip this iocb */ |
566 | } |
521 | } |
567 | else if (errno == EINTR) /* not seen in reality, not documented */ |
522 | else if (errno == EINTR) /* not seen in reality, not documented */ |
568 | res = 0; /* silently ignore and retry */ |
523 | res = 0; /* silently ignore and retry */ |
569 | else |
524 | else |
|
|
525 | { |
570 | ev_syserr ("(libev) linuxaio io_submit"); |
526 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
527 | res = 0; |
|
|
528 | } |
571 | |
529 | |
572 | submitted += res; |
530 | submitted += res; |
573 | } |
531 | } |
574 | |
532 | |
575 | linuxaio_submitcnt = 0; |
533 | linuxaio_submitcnt = 0; |
… | |
… | |
603 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
561 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
604 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
562 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
605 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
563 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
606 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
564 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
607 | |
565 | |
608 | backend_modify = linuxaio_modify; |
566 | backend_modify = linuxaio_modify; |
609 | backend_poll = linuxaio_poll; |
567 | backend_poll = linuxaio_poll; |
610 | |
568 | |
611 | linuxaio_iocbpmax = 0; |
569 | linuxaio_iocbpmax = 0; |
612 | linuxaio_iocbps = 0; |
570 | linuxaio_iocbps = 0; |
613 | |
571 | |
614 | linuxaio_submits = 0; |
572 | linuxaio_submits = 0; |
… | |
… | |
625 | epoll_destroy (EV_A); |
583 | epoll_destroy (EV_A); |
626 | linuxaio_free_iocbp (EV_A); |
584 | linuxaio_free_iocbp (EV_A); |
627 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
585 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
628 | } |
586 | } |
629 | |
587 | |
630 | inline_size |
588 | ecb_cold |
631 | void |
589 | static void |
632 | linuxaio_fork (EV_P) |
590 | linuxaio_fork (EV_P) |
633 | { |
591 | { |
634 | /* this frees all iocbs, which is very heavy-handed */ |
|
|
635 | linuxaio_destroy (EV_A); |
|
|
636 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
592 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
|
|
593 | linuxaio_free_iocp (EV_A); /* this frees all iocbs, which is very heavy-handed */ |
|
|
594 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
637 | |
595 | |
638 | linuxaio_iteration = 0; /* we start over in the child */ |
596 | linuxaio_iteration = 0; /* we start over in the child */ |
639 | |
597 | |
640 | while (linuxaio_io_setup (EV_A) < 0) |
598 | while (linuxaio_io_setup (EV_A) < 0) |
641 | ev_syserr ("(libev) linuxaio io_setup"); |
599 | ev_syserr ("(libev) linuxaio io_setup"); |
642 | |
600 | |
643 | /* forking epoll should also effectively unregister all fds from the backend */ |
601 | /* forking epoll should also effectively unregister all fds from the backend */ |
644 | epoll_fork (EV_A); |
602 | epoll_fork (EV_A); |
|
|
603 | /* epoll_fork already did this. hopefully */ |
|
|
604 | /*fd_rearm_all (EV_A);*/ |
645 | |
605 | |
646 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
606 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
647 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
607 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
648 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
608 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
649 | |
|
|
650 | /* epoll_fork already did this. hopefully */ |
|
|
651 | /*fd_rearm_all (EV_A);*/ |
|
|
652 | } |
609 | } |
653 | |
610 | |