… | |
… | |
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); |
… | |
… | |
372 | /* get any events from ring buffer, return true if any were handled */ |
321 | /* get any events from ring buffer, return true if any were handled */ |
373 | static int |
322 | static int |
374 | linuxaio_get_events_from_ring (EV_P) |
323 | linuxaio_get_events_from_ring (EV_P) |
375 | { |
324 | { |
376 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
325 | struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
|
|
326 | unsigned head, tail; |
377 | |
327 | |
378 | /* the kernel reads and writes both of these variables, */ |
328 | /* the kernel reads and writes both of these variables, */ |
379 | /* as a C extension, we assume that volatile use here */ |
329 | /* as a C extension, we assume that volatile use here */ |
380 | /* both makes reads atomic and once-only */ |
330 | /* both makes reads atomic and once-only */ |
381 | unsigned head = *(volatile unsigned *)&ring->head; |
331 | head = *(volatile unsigned *)&ring->head; |
|
|
332 | ECB_MEMORY_FENCE_ACQUIRE; |
382 | unsigned tail = *(volatile unsigned *)&ring->tail; |
333 | tail = *(volatile unsigned *)&ring->tail; |
383 | |
334 | |
384 | if (head == tail) |
335 | if (head == tail) |
385 | return 0; |
336 | return 0; |
386 | |
337 | |
387 | /* make sure the events up to tail are visible */ |
|
|
388 | ECB_MEMORY_FENCE_ACQUIRE; |
|
|
389 | |
|
|
390 | /* parse all available events, but only once, to avoid starvation */ |
338 | /* parse all available events, but only once, to avoid starvation */ |
391 | if (tail > head) /* normal case around */ |
339 | if (ecb_expect_true (tail > head)) /* normal case around */ |
392 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
340 | linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
393 | else /* wrapped around */ |
341 | else /* wrapped around */ |
394 | { |
342 | { |
395 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
343 | linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
396 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
344 | linuxaio_parse_events (EV_A_ ring->io_events, tail); |
… | |
… | |
445 | { |
393 | { |
446 | int res; |
394 | int res; |
447 | |
395 | |
448 | EV_RELEASE_CB; |
396 | EV_RELEASE_CB; |
449 | |
397 | |
450 | ts.tv_sec = (long)timeout; |
398 | EV_TS_SET (ts, timeout); |
451 | ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
|
|
452 | |
|
|
453 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
399 | res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
454 | |
400 | |
455 | EV_ACQUIRE_CB; |
401 | EV_ACQUIRE_CB; |
456 | |
402 | |
457 | if (res < 0) |
403 | if (res < 0) |
… | |
… | |
543 | } |
489 | } |
544 | |
490 | |
545 | ++linuxaio_iteration; |
491 | ++linuxaio_iteration; |
546 | if (linuxaio_io_setup (EV_A) < 0) |
492 | if (linuxaio_io_setup (EV_A) < 0) |
547 | { |
493 | { |
|
|
494 | /* TODO: rearm all and recreate epoll backend from scratch */ |
|
|
495 | /* TODO: might be more prudent? */ |
|
|
496 | |
548 | /* to bad, we can't get a new aio context, go 100% epoll */ |
497 | /* to bad, we can't get a new aio context, go 100% epoll */ |
549 | linuxaio_free_iocbp (EV_A); |
498 | linuxaio_free_iocbp (EV_A); |
550 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
499 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
551 | ev_ref (EV_A); |
500 | ev_ref (EV_A); |
552 | linuxaio_ctx = 0; |
501 | linuxaio_ctx = 0; |
|
|
502 | |
|
|
503 | backend = EVBACKEND_EPOLL; |
553 | backend_modify = epoll_modify; |
504 | backend_modify = epoll_modify; |
554 | backend_poll = epoll_poll; |
505 | backend_poll = epoll_poll; |
555 | } |
506 | } |
556 | |
507 | |
557 | timeout = 0; |
508 | timeout = 0; |
… | |
… | |
566 | res = 1; /* skip this iocb */ |
517 | res = 1; /* skip this iocb */ |
567 | } |
518 | } |
568 | else if (errno == EINTR) /* not seen in reality, not documented */ |
519 | else if (errno == EINTR) /* not seen in reality, not documented */ |
569 | res = 0; /* silently ignore and retry */ |
520 | res = 0; /* silently ignore and retry */ |
570 | else |
521 | else |
|
|
522 | { |
571 | ev_syserr ("(libev) linuxaio io_submit"); |
523 | ev_syserr ("(libev) linuxaio io_submit"); |
|
|
524 | res = 0; |
|
|
525 | } |
572 | |
526 | |
573 | submitted += res; |
527 | submitted += res; |
574 | } |
528 | } |
575 | |
529 | |
576 | linuxaio_submitcnt = 0; |
530 | linuxaio_submitcnt = 0; |
… | |
… | |
604 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
558 | ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
605 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
559 | ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
606 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
560 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
607 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
561 | ev_unref (EV_A); /* watcher should not keep loop alive */ |
608 | |
562 | |
609 | backend_modify = linuxaio_modify; |
563 | backend_modify = linuxaio_modify; |
610 | backend_poll = linuxaio_poll; |
564 | backend_poll = linuxaio_poll; |
611 | |
565 | |
612 | linuxaio_iocbpmax = 0; |
566 | linuxaio_iocbpmax = 0; |
613 | linuxaio_iocbps = 0; |
567 | linuxaio_iocbps = 0; |
614 | |
568 | |
615 | linuxaio_submits = 0; |
569 | linuxaio_submits = 0; |
… | |
… | |
626 | epoll_destroy (EV_A); |
580 | epoll_destroy (EV_A); |
627 | linuxaio_free_iocbp (EV_A); |
581 | linuxaio_free_iocbp (EV_A); |
628 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
582 | evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
629 | } |
583 | } |
630 | |
584 | |
631 | inline_size |
585 | ecb_cold |
632 | void |
586 | static void |
633 | linuxaio_fork (EV_P) |
587 | linuxaio_fork (EV_P) |
634 | { |
588 | { |
635 | /* this frees all iocbs, which is very heavy-handed */ |
589 | /* this frees all iocbs, which is very heavy-handed */ |
636 | linuxaio_destroy (EV_A); |
590 | linuxaio_destroy (EV_A); |
637 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
591 | linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
… | |
… | |
641 | while (linuxaio_io_setup (EV_A) < 0) |
595 | while (linuxaio_io_setup (EV_A) < 0) |
642 | ev_syserr ("(libev) linuxaio io_setup"); |
596 | ev_syserr ("(libev) linuxaio io_setup"); |
643 | |
597 | |
644 | /* forking epoll should also effectively unregister all fds from the backend */ |
598 | /* forking epoll should also effectively unregister all fds from the backend */ |
645 | epoll_fork (EV_A); |
599 | epoll_fork (EV_A); |
|
|
600 | /* epoll_fork already did this. hopefully */ |
|
|
601 | /*fd_rearm_all (EV_A);*/ |
646 | |
602 | |
647 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
603 | ev_io_stop (EV_A_ &linuxaio_epoll_w); |
648 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
604 | ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
649 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
605 | ev_io_start (EV_A_ &linuxaio_epoll_w); |
650 | |
|
|
651 | /* epoll_fork already did this. hopefully */ |
|
|
652 | /*fd_rearm_all (EV_A);*/ |
|
|
653 | } |
606 | } |
654 | |
607 | |