On 2/6/20 12:15 PM, Jens Axboe wrote: > On 2/6/20 10:33 AM, Stefano Garzarella wrote: >> >> >> On Fri, Jan 31, 2020 at 4:39 PM Jens Axboe <axboe@xxxxxxxxx> wrote: >>> >>> On 1/31/20 7:29 AM, Stefano Garzarella wrote: >>>> Hi Jens, >>>> this is a v2 of the epoll test. >>>> >>>> v1 -> v2: >>>> - if IORING_FEAT_NODROP is not available, avoid to overflow the CQ >>>> - add 2 new tests to test epoll with IORING_FEAT_NODROP >>>> - cleanups >>>> >>>> There are 4 sub-tests: >>>> 1. test_epoll >>>> 2. test_epoll_sqpoll >>>> 3. test_epoll_nodrop >>>> 4. test_epoll_sqpoll_nodrop >>>> >>>> In the first 2 tests, I try to avoid to queue more requests than we have room >>>> for in the CQ ring. These work fine, I have no faults. >>> >>> Thanks! >>> >>>> In the tests 3 and 4, if IORING_FEAT_NODROP is supported, I try to submit as >>>> much as I can until I get a -EBUSY, but they often fail in this way: >>>> the submitter manages to submit everything, the receiver receives all the >>>> submitted bytes, but the cleaner loses completion events (I also tried to put a >>>> timeout to epoll_wait() in the cleaner to be sure that it is not related to the >>>> patch that I send some weeks ago, but the situation doesn't change, it's like >>>> there is still overflow in the CQ). >>>> >>>> Next week I'll try to investigate better which is the problem. >>> >>> Does it change if you have an io_uring_enter() with GETEVENTS set? I wonder if >>> you just pruned the CQ ring but didn't flush the internal side. >> >> If I do io_uring_enter() with GETEVENTS set and wait_nr = 0 it solves >> the issue, I think because we call io_cqring_events() that flushes the >> overflow list. >> >> At this point, should we call io_cqring_events() (that flushes the >> overflow list) in io_uring_poll()? >> I mean something like this: >> >> diff --git a/fs/io_uring.c b/fs/io_uring.c >> index 77f22c3da30f..2769451af89a 100644 >> --- a/fs/io_uring.c >> +++ b/fs/io_uring.c >> @@ -6301,7 +6301,7 @@ static __poll_t io_uring_poll(struct file *file, poll_table *wait) >> if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != >> ctx->rings->sq_ring_entries) >> mask |= EPOLLOUT | EPOLLWRNORM; >> - if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) >> + if (!io_cqring_events(ctx, false)) >> mask |= EPOLLIN | EPOLLRDNORM; >> >> return mask; > > That's not a bad idea, would just have to verify that it is indeed safe > to always call the flushing variant from there. Double checked, and it should be fine. We may be invoked with ctx->uring_lock held, but that's fine. -- Jens Axboe
