… | |
… | |
91 | /*****************************************************************************/ |
91 | /*****************************************************************************/ |
92 | /* syscall wrapdadoop - this section has the raw api/abi definitions */ |
92 | /* syscall wrapdadoop - this section has the raw api/abi definitions */ |
93 | |
93 | |
94 | #include <sys/syscall.h> /* no glibc wrappers */ |
94 | #include <sys/syscall.h> /* no glibc wrappers */ |
95 | |
95 | |
96 | /* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
96 | /* aio_abi.h is not versioned in any way, so we cannot test for its existence */ |
97 | #define IOCB_CMD_POLL 5 |
97 | #define IOCB_CMD_POLL 5 |
98 | |
98 | |
99 | /* taken from linux/fs/aio.c. yup, that's a .c file. |
99 | /* taken from linux/fs/aio.c. yup, that's a .c file. |
100 | * not only is this totally undocumented, not even the source code |
100 | * not only is this totally undocumented, not even the source code |
101 | * can tell you what the future semantics of compat_features and |
101 | * can tell you what the future semantics of compat_features and |
… | |
… | |
212 | * this is not well documented, so we better do it. |
212 | * this is not well documented, so we better do it. |
213 | */ |
213 | */ |
214 | memset (iocb, 0, sizeof (*iocb)); |
214 | memset (iocb, 0, sizeof (*iocb)); |
215 | |
215 | |
216 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
216 | iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
217 | iocb->io.aio_data = offset; |
|
|
218 | iocb->io.aio_fildes = offset; |
217 | iocb->io.aio_fildes = offset; |
219 | |
218 | |
220 | base [offset++] = iocb; |
219 | base [offset++] = iocb; |
221 | } |
220 | } |
222 | } |
221 | } |
… | |
… | |
234 | static void |
233 | static void |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
234 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
236 | { |
235 | { |
237 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
236 | array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
237 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
|
|
238 | ANFD *anfd = &anfds [fd]; |
239 | |
239 | |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
240 | if (ecb_expect_false (iocb->io.aio_reqprio < 0)) |
241 | { |
241 | { |
242 | /* we handed this fd over to epoll, so undo this first */ |
242 | /* we handed this fd over to epoll, so undo this first */ |
243 | /* 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 */ |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
244 | epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
245 | anfds [fd].emask = 0; |
245 | anfd->emask = 0; |
246 | iocb->io.aio_reqprio = 0; |
246 | iocb->io.aio_reqprio = 0; |
247 | } |
247 | } |
248 | |
|
|
249 | if (ecb_expect_false (iocb->io.aio_buf)) |
248 | else if (ecb_expect_false (iocb->io.aio_buf)) |
250 | { |
249 | { |
251 | /* iocb active, so cancel it first before resubmit */ |
250 | /* iocb active, so cancel it first before resubmit */ |
|
|
251 | /* this assumes we only ever get one call per fd per loop iteration */ |
252 | for (;;) |
252 | for (;;) |
253 | { |
253 | { |
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
254 | /* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
255 | if (ecb_expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
256 | break; |
256 | break; |
257 | |
257 | |
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
259 | break; |
259 | break; |
260 | |
260 | |
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
262 | if (errno != EINTR) |
|
|
263 | { |
262 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
264 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINTR && errno != EINPROGRESS)); |
|
|
265 | break; |
|
|
266 | } |
263 | } |
267 | } |
264 | } |
|
|
265 | |
268 | |
266 | iocb->io.aio_buf = |
269 | /* increment generation counter to avoid handling old events */ |
267 | (nev & EV_READ ? POLLIN : 0) |
270 | ++anfd->egen; |
|
|
271 | } |
|
|
272 | |
|
|
273 | iocb->io.aio_buf = (nev & EV_READ ? POLLIN : 0) |
268 | | (nev & EV_WRITE ? POLLOUT : 0); |
274 | | (nev & EV_WRITE ? POLLOUT : 0); |
269 | |
275 | |
270 | if (nev) |
276 | if (nev) |
271 | { |
277 | { |
|
|
278 | iocb->io.aio_data = (uint32_t)fd | ((__u64)(uint32_t)anfd->egen << 32); |
|
|
279 | |
272 | /* queue iocb up for io_submit */ |
280 | /* queue iocb up for io_submit */ |
273 | /* this assumes we only ever get one call per fd per loop iteration */ |
281 | /* this assumes we only ever get one call per fd per loop iteration */ |
274 | ++linuxaio_submitcnt; |
282 | ++linuxaio_submitcnt; |
275 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
283 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
276 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
284 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
… | |
… | |
295 | static void |
303 | static void |
296 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
304 | linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
297 | { |
305 | { |
298 | while (nr) |
306 | while (nr) |
299 | { |
307 | { |
300 | int fd = ev->data; |
308 | int fd = ev->data & 0xffffffff; |
|
|
309 | uint32_t gen = ev->data >> 32; |
301 | int res = ev->res; |
310 | int res = ev->res; |
302 | |
311 | |
303 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
312 | assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
304 | |
313 | |
|
|
314 | /* only accept events if generation counter matches */ |
|
|
315 | if (ecb_expect_true (gen == (uint32_t)anfds [fd].egen)) |
|
|
316 | { |
305 | /* feed events, we do not expect or handle POLLNVAL */ |
317 | /* feed events, we do not expect or handle POLLNVAL */ |
306 | fd_event ( |
318 | fd_event ( |
307 | EV_A_ |
319 | EV_A_ |
308 | fd, |
320 | fd, |
309 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
321 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
310 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
322 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
311 | ); |
323 | ); |
312 | |
324 | |
313 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
325 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
314 | linuxaio_fd_rearm (EV_A_ fd); |
326 | linuxaio_fd_rearm (EV_A_ fd); |
|
|
327 | } |
315 | |
328 | |
316 | --nr; |
329 | --nr; |
317 | ++ev; |
330 | ++ev; |
318 | } |
331 | } |
319 | } |
332 | } |
… | |
… | |
321 | /* get any events from ring buffer, return true if any were handled */ |
334 | /* get any events from ring buffer, return true if any were handled */ |
322 | static int |
335 | static int |
323 | linuxaio_get_events_from_ring (EV_P) |
336 | linuxaio_get_events_from_ring (EV_P) |
324 | { |
337 | { |
325 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
338 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
339 | unsigned head, tail; |
326 | |
340 | |
327 | /* the kernel reads and writes both of these variables, */ |
341 | /* the kernel reads and writes both of these variables, */ |
328 | /* as a C extension, we assume that volatile use here */ |
342 | /* as a C extension, we assume that volatile use here */ |
329 | /* both makes reads atomic and once-only */ |
343 | /* both makes reads atomic and once-only */ |
330 | unsigned head = *(volatile unsigned *)&ring->head; |
344 | head = *(volatile unsigned *)&ring->head; |
|
|
345 | ECB_MEMORY_FENCE_ACQUIRE; |
331 | unsigned tail = *(volatile unsigned *)&ring->tail; |
346 | tail = *(volatile unsigned *)&ring->tail; |
332 | |
347 | |
333 | if (head == tail) |
348 | if (head == tail) |
334 | return 0; |
349 | return 0; |
335 | |
350 | |
336 | /* make sure the events up to tail are visible */ |
|
|
337 | ECB_MEMORY_FENCE_ACQUIRE; |
|
|
338 | |
|
|
339 | /* parse all available events, but only once, to avoid starvation */ |
351 | /* parse all available events, but only once, to avoid starvation */ |
340 | if (tail > head) /* normal case around */ |
352 | if (ecb_expect_true (tail > head)) /* normal case around */ |
341 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
353 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
342 | else /* wrapped around */ |
354 | else /* wrapped around */ |
343 | { |
355 | { |
344 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
356 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
345 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
357 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
… | |
… | |
394 | { |
406 | { |
395 | int res; |
407 | int res; |
396 | |
408 | |
397 | EV_RELEASE_CB; |
409 | EV_RELEASE_CB; |
398 | |
410 | |
399 | ts.tv_sec = (long)timeout; |
411 | EV_TS_SET (ts, timeout); |
400 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
|
|
401 | |
|
|
402 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
412 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
403 | |
413 | |
404 | EV_ACQUIRE_CB; |
414 | EV_ACQUIRE_CB; |
405 | |
415 | |
406 | if (res < 0) |
416 | if (res < 0) |
… | |
… | |
426 | break; |
436 | break; |
427 | } |
437 | } |
428 | else |
438 | else |
429 | break; /* no events from the kernel, we are done */ |
439 | break; /* no events from the kernel, we are done */ |
430 | |
440 | |
431 | timeout = 0; /* only wait in the first iteration */ |
441 | timeout = EV_TS_CONST (0.); /* only wait in the first iteration */ |
432 | } |
442 | } |
433 | } |
443 | } |
434 | |
444 | |
435 | inline_size |
445 | inline_size |
436 | int |
446 | int |
… | |
… | |
492 | } |
502 | } |
493 | |
503 | |
494 | ++linuxaio_iteration; |
504 | ++linuxaio_iteration; |
495 | if (linuxaio_io_setup (EV_A) < 0) |
505 | if (linuxaio_io_setup (EV_A) < 0) |
496 | { |
506 | { |
|
|
507 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
508 | /* TODO: might be more prudent? */ |
|
|
509 | |
497 | /* to bad, we can't get a new aio context, go 100% epoll */ |
510 | /* to bad, we can't get a new aio context, go 100% epoll */ |
498 | linuxaio_free_iocbp (EV_A); |
511 | linuxaio_free_iocbp (EV_A); |
499 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
512 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
500 | ev_ref (EV_A); |
513 | ev_ref (EV_A); |
501 | linuxaio_ctx = 0; |
514 | linuxaio_ctx = 0; |
|
|
515 | |
|
|
516 | backend = EVBACKEND_EPOLL; |
502 | backend_modify = epoll_modify; |
517 | backend_modify = epoll_modify; |
503 | backend_poll = epoll_poll; |
518 | backend_poll = epoll_poll; |
504 | } |
519 | } |
505 | |
520 | |
506 | timeout = 0; |
521 | timeout = EV_TS_CONST (0.); |
507 | /* it's easiest to handle this mess in another iteration */ |
522 | /* it's easiest to handle this mess in another iteration */ |
508 | return; |
523 | return; |
509 | } |
524 | } |
510 | else if (errno == EBADF) |
525 | else if (errno == EBADF) |
511 | { |
526 | { |
… | |
… | |
515 | res = 1; /* skip this iocb */ |
530 | res = 1; /* skip this iocb */ |
516 | } |
531 | } |
517 | else if (errno == EINTR) /* not seen in reality, not documented */ |
532 | else if (errno == EINTR) /* not seen in reality, not documented */ |
518 | res = 0; /* silently ignore and retry */ |
533 | res = 0; /* silently ignore and retry */ |
519 | else |
534 | else |
|
|
535 | { |
520 | ev_syserr ("(libev) linuxaio io_submit"); |
536 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
537 | res = 0; |
|
|
538 | } |
521 | |
539 | |
522 | submitted += res; |
540 | submitted += res; |
523 | } |
541 | } |
524 | |
542 | |
525 | linuxaio_submitcnt = 0; |
543 | linuxaio_submitcnt = 0; |
… | |
… | |
548 | { |
566 | { |
549 | epoll_destroy (EV_A); |
567 | epoll_destroy (EV_A); |
550 | return 0; |
568 | return 0; |
551 | } |
569 | } |
552 | |
570 | |
553 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
571 | ev_io_init (&linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
554 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
572 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
555 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
573 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
556 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
574 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
557 | |
575 | |
558 | backend_modify = linuxaio_modify; |
576 | backend_modify = linuxaio_modify; |
559 | backend_poll = linuxaio_poll; |
577 | backend_poll = linuxaio_poll; |
560 | |
578 | |
561 | linuxaio_iocbpmax = 0; |
579 | linuxaio_iocbpmax = 0; |
562 | linuxaio_iocbps = 0; |
580 | linuxaio_iocbps = 0; |
563 | |
581 | |
564 | linuxaio_submits = 0; |
582 | linuxaio_submits = 0; |
… | |
… | |
575 | epoll_destroy (EV_A); |
593 | epoll_destroy (EV_A); |
576 | linuxaio_free_iocbp (EV_A); |
594 | linuxaio_free_iocbp (EV_A); |
577 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
595 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
578 | } |
596 | } |
579 | |
597 | |
580 | inline_size |
598 | ecb_cold |
581 | void |
599 | static void |
582 | linuxaio_fork (EV_P) |
600 | linuxaio_fork (EV_P) |
583 | { |
601 | { |
584 | /* this frees all iocbs, which is very heavy-handed */ |
|
|
585 | linuxaio_destroy (EV_A); |
|
|
586 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
602 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
|
|
603 | linuxaio_free_iocbp (EV_A); /* this frees all iocbs, which is very heavy-handed */ |
|
|
604 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
587 | |
605 | |
588 | linuxaio_iteration = 0; /* we start over in the child */ |
606 | linuxaio_iteration = 0; /* we start over in the child */ |
589 | |
607 | |
590 | while (linuxaio_io_setup (EV_A) < 0) |
608 | while (linuxaio_io_setup (EV_A) < 0) |
591 | ev_syserr ("(libev) linuxaio io_setup"); |
609 | ev_syserr ("(libev) linuxaio io_setup"); |
592 | |
610 | |
593 | /* forking epoll should also effectively unregister all fds from the backend */ |
611 | /* forking epoll should also effectively unregister all fds from the backend */ |
594 | epoll_fork (EV_A); |
612 | epoll_fork (EV_A); |
|
|
613 | /* epoll_fork already did this. hopefully */ |
|
|
614 | /*fd_rearm_all (EV_A);*/ |
595 | |
615 | |
596 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
616 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
597 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
617 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
598 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
618 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
599 | |
|
|
600 | /* epoll_fork already did this. hopefully */ |
|
|
601 | /*fd_rearm_all (EV_A);*/ |
|
|
602 | } |
619 | } |
603 | |
620 | |