ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev_iouring.c
(Generate patch)

Comparing libev/ev_iouring.c (file contents):
Revision 1.13 by root, Sat Dec 28 03:29:50 2019 UTC vs.
Revision 1.15 by root, Sat Dec 28 05:53:48 2019 UTC

… …
228 228
229/* the submit/completion queue entries */ 229/* the submit/completion queue entries */
230#define EV_SQES ((struct io_uring_sqe *) iouring_sqes) 230#define EV_SQES ((struct io_uring_sqe *) iouring_sqes)
231#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes)) 231#define EV_CQES ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
232 232
233/* TODO: this is not enough, we might have to reap events */
234/* TODO: but we can't, as that will re-arm events, causing */
235/* TODO: an endless loop in fd_reify */
236static int
237iouring_enter (EV_P_ ev_tstamp timeout)
238{
239 int res;
240
241 EV_RELEASE_CB;
242
243 res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
244 timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
245
246 assert (("libev: io_uring_enter did not consume all sqes", (res < 0 || res == iouring_to_submit)));
247
248 iouring_to_submit = 0;
249
250 EV_ACQUIRE_CB;
251
252 return res;
253}
254
233static 255static
234struct io_uring_sqe * 256struct io_uring_sqe *
235iouring_sqe_get (EV_P) 257iouring_sqe_get (EV_P)
236{ 258{
237 unsigned tail = EV_SQ_VAR (tail); 259 unsigned tail = EV_SQ_VAR (tail);
238 260
239 if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries)) 261 while (ecb_expect_false (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries)))
240 { 262 {
241 /* queue full, flush */ 263 /* queue full, need to flush */
242 evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
243 iouring_to_submit = 0;
244 }
245 264
265 int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
266
267 /* io_uring_enter might fail with EBUSY and won't submit anything */
268 /* unfortunately, we can't handle this at the moment */
269
270 if (res < 0 && errno == EBUSY)
271 /* the sane thing might be to resize, but we can't */
272 //TODO
273 ev_syserr ("(libev) io_uring_enter could not clear sq");
274 else
275 break;
276
277 /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE */
278 }
279
246 assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries))); 280 /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
247 281
248 return EV_SQES + (tail & EV_SQ_VAR (ring_mask)); 282 return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
249} 283}
250 284
251inline_size 285inline_size
… …
405 /* Jens Axboe notified me that user_data is not what is documented, but is 439 /* Jens Axboe notified me that user_data is not what is documented, but is
406 * some kind of unique ID that has to match, otherwise the request cannot 440 * some kind of unique ID that has to match, otherwise the request cannot
407 * be removed. Since we don't *really* have that, we pass in the old 441 * be removed. Since we don't *really* have that, we pass in the old
408 * generation counter - if that fails, too bad, it will hopefully be removed 442 * generation counter - if that fails, too bad, it will hopefully be removed
409 * at close time and then be ignored. */ 443 * at close time and then be ignored. */
410 sqe->user_data = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32); 444 sqe->addr = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
445 sqe->user_data = (uint64_t)-1;
411 iouring_sqe_submit (EV_A_ sqe); 446 iouring_sqe_submit (EV_A_ sqe);
412 447
413 /* increment generation counter to avoid handling old events */ 448 /* increment generation counter to avoid handling old events */
414 ++anfds [fd].egen; 449 ++anfds [fd].egen;
415 } 450 }
… …
455iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe) 490iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
456{ 491{
457 int fd = cqe->user_data & 0xffffffffU; 492 int fd = cqe->user_data & 0xffffffffU;
458 uint32_t gen = cqe->user_data >> 32; 493 uint32_t gen = cqe->user_data >> 32;
459 int res = cqe->res; 494 int res = cqe->res;
495
496 /* user_data -1 is a remove that we are not atm. interested in */
497 if (cqe->user_data == (uint64_t)-1)
498 return;
460 499
461 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax)); 500 assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
462 501
463 /* documentation lies, of course. the result value is NOT like 502 /* documentation lies, of course. the result value is NOT like
464 * normal syscalls, but like linux raw syscalls, i.e. negative 503 * normal syscalls, but like linux raw syscalls, i.e. negative
… …
598 iouring_tfd_update (EV_A_ timeout); 637 iouring_tfd_update (EV_A_ timeout);
599 638
600 /* only enter the kernel if we have something to submit, or we need to wait */ 639 /* only enter the kernel if we have something to submit, or we need to wait */
601 if (timeout || iouring_to_submit) 640 if (timeout || iouring_to_submit)
602 { 641 {
603 int res; 642 int res = iouring_enter (EV_A_ timeout);
604
605 EV_RELEASE_CB;
606
607 res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
608 timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
609 iouring_to_submit = 0;
610
611 EV_ACQUIRE_CB;
612 643
613 if (ecb_expect_false (res < 0)) 644 if (ecb_expect_false (res < 0))
614 if (errno == EINTR) 645 if (errno == EINTR)
615 /* ignore */; 646 /* ignore */;
647 else if (errno == EBUSY)
648 /* cq full, cannot submit - should be rare because we flush the cq first, so simply ignore */;
616 else 649 else
617 ev_syserr ("(libev) iouring setup"); 650 ev_syserr ("(libev) iouring setup");
618 else 651 else
619 iouring_handle_cq (EV_A); 652 iouring_handle_cq (EV_A);
620 } 653 }

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines