… | |
… | |
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 |
102 | * incompat_features are, or what header_length actually is for. |
102 | * incompat_features are, or what header_length actually is for. |
103 | */ |
103 | */ |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
104 | #define AIO_RING_MAGIC 0xa10a10a1 |
105 | #define AIO_RING_INCOMPAT_FEATURES 0 |
105 | #define EV_AIO_RING_INCOMPAT_FEATURES 0 |
106 | struct aio_ring |
106 | struct aio_ring |
107 | { |
107 | { |
108 | unsigned id; /* kernel internal index number */ |
108 | unsigned id; /* kernel internal index number */ |
109 | unsigned nr; /* number of io_events */ |
109 | unsigned nr; /* number of io_events */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
110 | unsigned head; /* Written to by userland or by kernel. */ |
… | |
… | |
120 | |
120 | |
121 | inline_size |
121 | inline_size |
122 | int |
122 | int |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
123 | evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
124 | { |
124 | { |
125 | return syscall (SYS_io_setup, nr_events, ctx_idp); |
125 | return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
126 | } |
126 | } |
127 | |
127 | |
128 | inline_size |
128 | inline_size |
129 | int |
129 | int |
130 | evsys_io_destroy (aio_context_t ctx_id) |
130 | evsys_io_destroy (aio_context_t ctx_id) |
131 | { |
131 | { |
132 | return syscall (SYS_io_destroy, ctx_id); |
132 | return ev_syscall1 (SYS_io_destroy, ctx_id); |
133 | } |
133 | } |
134 | |
134 | |
135 | inline_size |
135 | inline_size |
136 | int |
136 | int |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
137 | evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
138 | { |
138 | { |
139 | return syscall (SYS_io_submit, ctx_id, nr, cbp); |
139 | return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
140 | } |
140 | } |
141 | |
141 | |
142 | inline_size |
142 | inline_size |
143 | int |
143 | int |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
144 | evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
145 | { |
145 | { |
146 | return syscall (SYS_io_cancel, ctx_id, cbp, result); |
146 | return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
147 | } |
147 | } |
148 | |
148 | |
149 | inline_size |
149 | inline_size |
150 | int |
150 | int |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
151 | evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
152 | { |
152 | { |
153 | return syscall (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
153 | return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
154 | } |
154 | } |
155 | |
155 | |
156 | /*****************************************************************************/ |
156 | /*****************************************************************************/ |
157 | /* actual backed implementation */ |
157 | /* actual backed implementation */ |
158 | |
158 | |
… | |
… | |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
235 | linuxaio_modify (EV_P_ int fd, int oev, int nev) |
236 | { |
236 | { |
237 | 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); |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
238 | ANIOCBP iocb = linuxaio_iocbps [fd]; |
239 | |
239 | |
240 | if (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_modfy 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 | iocb->io.aio_reqprio = 0; |
246 | iocb->io.aio_reqprio = 0; |
246 | } |
247 | } |
247 | |
248 | |
|
|
249 | if (ecb_expect_false (iocb->io.aio_buf)) |
|
|
250 | { |
|
|
251 | /* iocb active, so cancel it first before resubmit */ |
|
|
252 | for (;;) |
|
|
253 | { |
|
|
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)) |
|
|
256 | break; |
|
|
257 | |
|
|
258 | if (ecb_expect_true (errno == EINPROGRESS)) |
|
|
259 | break; |
|
|
260 | |
|
|
261 | /* the EINPROGRESS test is for nicer error message. clumsy. */ |
|
|
262 | assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
|
|
263 | } |
|
|
264 | } |
|
|
265 | |
248 | if (iocb->io.aio_buf) |
266 | iocb->io.aio_buf = |
249 | /* io_cancel always returns some error on relevant kernels, but works */ |
267 | (nev & EV_READ ? POLLIN : 0) |
250 | evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0); |
268 | | (nev & EV_WRITE ? POLLOUT : 0); |
251 | |
269 | |
252 | if (nev) |
270 | if (nev) |
253 | { |
271 | { |
254 | iocb->io.aio_buf = |
|
|
255 | (nev & EV_READ ? POLLIN : 0) |
|
|
256 | | (nev & EV_WRITE ? POLLOUT : 0); |
|
|
257 | |
|
|
258 | /* queue iocb up for io_submit */ |
272 | /* queue iocb up for io_submit */ |
259 | /* this assumes we only ever get one call per fd per loop iteration */ |
273 | /* this assumes we only ever get one call per fd per loop iteration */ |
260 | ++linuxaio_submitcnt; |
274 | ++linuxaio_submitcnt; |
261 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
275 | array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
262 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
276 | linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
… | |
… | |
267 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
281 | linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
268 | { |
282 | { |
269 | epoll_poll (EV_A_ 0); |
283 | epoll_poll (EV_A_ 0); |
270 | } |
284 | } |
271 | |
285 | |
272 | static void |
286 | inline_speed |
|
|
287 | void |
273 | linuxaio_fd_rearm (EV_P_ int fd) |
288 | linuxaio_fd_rearm (EV_P_ int fd) |
274 | { |
289 | { |
275 | anfds [fd].events = 0; |
290 | anfds [fd].events = 0; |
276 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
291 | linuxaio_iocbps [fd]->io.aio_buf = 0; |
277 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
292 | fd_change (EV_A_ fd, EV_ANFD_REIFY); |
… | |
… | |
293 | fd, |
308 | fd, |
294 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
309 | (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
295 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
310 | | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
296 | ); |
311 | ); |
297 | |
312 | |
298 | /* linux aio is oneshot: rearm fd. TODO: this does more work than needed */ |
313 | /* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
299 | linuxaio_fd_rearm (EV_A_ fd); |
314 | linuxaio_fd_rearm (EV_A_ fd); |
300 | |
315 | |
301 | --nr; |
316 | --nr; |
302 | ++ev; |
317 | ++ev; |
303 | } |
318 | } |
… | |
… | |
316 | unsigned tail = *(volatile unsigned *)&ring->tail; |
331 | unsigned tail = *(volatile unsigned *)&ring->tail; |
317 | |
332 | |
318 | if (head == tail) |
333 | if (head == tail) |
319 | return 0; |
334 | return 0; |
320 | |
335 | |
321 | /* bail out if the ring buffer doesn't match the expected layout */ |
|
|
322 | if (expect_false (ring->magic != AIO_RING_MAGIC) |
|
|
323 | || ring->incompat_features != AIO_RING_INCOMPAT_FEATURES |
|
|
324 | || ring->header_length != sizeof (struct aio_ring)) /* TODO: or use it to find io_event[0]? */ |
|
|
325 | return 0; |
|
|
326 | |
|
|
327 | /* make sure the events up to tail are visible */ |
336 | /* make sure the events up to tail are visible */ |
328 | ECB_MEMORY_FENCE_ACQUIRE; |
337 | ECB_MEMORY_FENCE_ACQUIRE; |
329 | |
338 | |
330 | /* parse all available events, but only once, to avoid starvation */ |
339 | /* parse all available events, but only once, to avoid starvation */ |
331 | if (tail > head) /* normal case around */ |
340 | if (tail > head) /* normal case around */ |
… | |
… | |
341 | *(volatile unsigned *)&ring->head = tail; |
350 | *(volatile unsigned *)&ring->head = tail; |
342 | |
351 | |
343 | return 1; |
352 | return 1; |
344 | } |
353 | } |
345 | |
354 | |
|
|
355 | inline_size |
|
|
356 | int |
|
|
357 | linuxaio_ringbuf_valid (EV_P) |
|
|
358 | { |
|
|
359 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
360 | |
|
|
361 | return ecb_expect_true (ring->magic == AIO_RING_MAGIC) |
|
|
362 | && ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
|
|
363 | && ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
|
|
364 | } |
|
|
365 | |
346 | /* read at least one event from kernel, or timeout */ |
366 | /* read at least one event from kernel, or timeout */ |
347 | inline_size |
367 | inline_size |
348 | void |
368 | void |
349 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
369 | linuxaio_get_events (EV_P_ ev_tstamp timeout) |
350 | { |
370 | { |
351 | struct timespec ts; |
371 | struct timespec ts; |
352 | struct io_event ioev[1]; |
372 | struct io_event ioev[8]; /* 256 octet stack space */ |
353 | int res; |
373 | int want = 1; /* how many events to request */ |
|
|
374 | int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
354 | |
375 | |
|
|
376 | if (ecb_expect_true (ringbuf_valid)) |
|
|
377 | { |
|
|
378 | /* if the ring buffer has any events, we don't wait or call the kernel at all */ |
355 | if (linuxaio_get_events_from_ring (EV_A)) |
379 | if (linuxaio_get_events_from_ring (EV_A)) |
356 | return; |
380 | return; |
357 | |
381 | |
358 | /* no events, so wait for at least one, then poll ring buffer again */ |
382 | /* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
359 | /* this degrades to one event per loop iteration */ |
383 | if (!timeout) |
360 | /* if the ring buffer changes layout, but so be it */ |
384 | return; |
|
|
385 | } |
|
|
386 | else |
|
|
387 | /* no ringbuffer, request slightly larger batch */ |
|
|
388 | want = sizeof (ioev) / sizeof (ioev [0]); |
361 | |
389 | |
|
|
390 | /* no events, so wait for some |
|
|
391 | * for fairness reasons, we do this in a loop, to fetch all events |
|
|
392 | */ |
|
|
393 | for (;;) |
|
|
394 | { |
|
|
395 | int res; |
|
|
396 | |
362 | EV_RELEASE_CB; |
397 | EV_RELEASE_CB; |
363 | |
398 | |
364 | ts.tv_sec = (long)timeout; |
399 | ts.tv_sec = (long)timeout; |
365 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
400 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
366 | |
401 | |
367 | res = evsys_io_getevents (linuxaio_ctx, 1, sizeof (ioev) / sizeof (ioev [0]), ioev, &ts); |
402 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
368 | |
403 | |
369 | EV_ACQUIRE_CB; |
404 | EV_ACQUIRE_CB; |
370 | |
405 | |
371 | if (res < 0) |
406 | if (res < 0) |
372 | if (errno == EINTR) |
407 | if (errno == EINTR) |
373 | /* ignored */; |
408 | /* ignored, retry */; |
374 | else |
409 | else |
375 | ev_syserr ("(libev) linuxaio io_getevents"); |
410 | ev_syserr ("(libev) linuxaio io_getevents"); |
376 | else if (res) |
411 | else if (res) |
377 | { |
412 | { |
378 | /* at least one event received, handle it and any remaining ones in the ring buffer */ |
413 | /* at least one event available, handle them */ |
379 | linuxaio_parse_events (EV_A_ ioev, res); |
414 | linuxaio_parse_events (EV_A_ ioev, res); |
|
|
415 | |
|
|
416 | if (ecb_expect_true (ringbuf_valid)) |
|
|
417 | { |
|
|
418 | /* if we have a ring buffer, handle any remaining events in it */ |
380 | linuxaio_get_events_from_ring (EV_A); |
419 | linuxaio_get_events_from_ring (EV_A); |
381 | } |
|
|
382 | } |
|
|
383 | |
420 | |
384 | static int |
421 | /* at this point, we should have handled all outstanding events */ |
|
|
422 | break; |
|
|
423 | } |
|
|
424 | else if (res < want) |
|
|
425 | /* otherwise, if there were fewere events than we wanted, we assume there are no more */ |
|
|
426 | break; |
|
|
427 | } |
|
|
428 | else |
|
|
429 | break; /* no events from the kernel, we are done */ |
|
|
430 | |
|
|
431 | timeout = 0; /* only wait in the first iteration */ |
|
|
432 | } |
|
|
433 | } |
|
|
434 | |
|
|
435 | inline_size |
|
|
436 | int |
385 | linuxaio_io_setup (EV_P) |
437 | linuxaio_io_setup (EV_P) |
386 | { |
438 | { |
387 | linuxaio_ctx = 0; |
439 | linuxaio_ctx = 0; |
388 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
440 | return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
389 | } |
441 | } |
… | |
… | |
400 | /* which allows us to pinpoint the erroneous iocb */ |
452 | /* which allows us to pinpoint the erroneous iocb */ |
401 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
453 | for (submitted = 0; submitted < linuxaio_submitcnt; ) |
402 | { |
454 | { |
403 | 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); |
404 | |
456 | |
405 | if (expect_false (res < 0)) |
457 | if (ecb_expect_false (res < 0)) |
406 | if (errno == EINVAL) |
458 | if (errno == EINVAL) |
407 | { |
459 | { |
408 | /* This happens for unsupported fds, officially, but in my testing, |
460 | /* This happens for unsupported fds, officially, but in my testing, |
409 | * 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 |
410 | * 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. |
… | |
… | |
455 | /* it's easiest to handle this mess in another iteration */ |
507 | /* it's easiest to handle this mess in another iteration */ |
456 | return; |
508 | return; |
457 | } |
509 | } |
458 | else if (errno == EBADF) |
510 | else if (errno == EBADF) |
459 | { |
511 | { |
|
|
512 | assert (("libev: event loop rejected bad fd", errno != EBADF)); |
460 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
513 | fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
461 | |
514 | |
462 | res = 1; /* skip this iocb */ |
515 | res = 1; /* skip this iocb */ |
463 | } |
516 | } |
|
|
517 | else if (errno == EINTR) /* not seen in reality, not documented */ |
|
|
518 | res = 0; /* silently ignore and retry */ |
464 | else |
519 | else |
465 | ev_syserr ("(libev) linuxaio io_submit"); |
520 | ev_syserr ("(libev) linuxaio io_submit"); |
466 | |
521 | |
467 | submitted += res; |
522 | submitted += res; |
468 | } |
523 | } |
… | |
… | |
517 | void |
572 | void |
518 | linuxaio_destroy (EV_P) |
573 | linuxaio_destroy (EV_P) |
519 | { |
574 | { |
520 | epoll_destroy (EV_A); |
575 | epoll_destroy (EV_A); |
521 | linuxaio_free_iocbp (EV_A); |
576 | linuxaio_free_iocbp (EV_A); |
522 | evsys_io_destroy (linuxaio_ctx); |
577 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
523 | } |
578 | } |
524 | |
579 | |
525 | inline_size |
580 | inline_size |
526 | void |
581 | void |
527 | linuxaio_fork (EV_P) |
582 | linuxaio_fork (EV_P) |
… | |
… | |
533 | linuxaio_iteration = 0; /* we start over in the child */ |
588 | linuxaio_iteration = 0; /* we start over in the child */ |
534 | |
589 | |
535 | while (linuxaio_io_setup (EV_A) < 0) |
590 | while (linuxaio_io_setup (EV_A) < 0) |
536 | ev_syserr ("(libev) linuxaio io_setup"); |
591 | ev_syserr ("(libev) linuxaio io_setup"); |
537 | |
592 | |
|
|
593 | /* forking epoll should also effectively unregister all fds from the backend */ |
538 | epoll_fork (EV_A); |
594 | epoll_fork (EV_A); |
539 | |
595 | |
540 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
596 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
541 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
597 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
542 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
598 | ev_io_start (EV_A_ &linuxaio_epoll_w); |