Re: [PATCH liburing 1/1] test: add epoll test case

[Stefano Garzarella]

On Mon, Jan 27, 2020 at 11:46:34AM -0700, Jens Axboe wrote:
> On 1/27/20 11:25 AM, Stefano Garzarella wrote:
> > On Mon, Jan 27, 2020 at 09:32:43AM -0700, Jens Axboe wrote:
> >> On 1/27/20 9:17 AM, Stefano Garzarella wrote:
> >>> Signed-off-by: Stefano Garzarella <sgarzare@xxxxxxxxxx>
> >>
> >> You're not reaping CQ events, and hence you overflow the ring. Once
> >> overflown, an attempt to submit new IO will returns in a -16/-EBUSY
> >> return value. This is io_uring telling you that it won't submit more
> >> IO until you've emptied the completion ring so io_uring can flush
> >> the overflown entries to the ring.
> > 
> > How can I reaping CQ events? (I was hoping the epoll would help me with that)
> > 
> > What I'm seeing is that the producer (EPOLLOUT) can fill the SQ without issues,
> > the consumer (read()) is receiving all the buffers produced, but the thread
> > that frees the buffers (EPOLLIN) is not woken up.
> > 
> > I tried to set a timeout to the epoll_wait(), but the io_uring_peek_cqe()
> > returns -EAGAIN.
> > 
> > If I'm using a ring with 16 entries, it seems to work better, but
> > sometimes I lose events and the thread that frees the buffer doesn't wake up.
> > 
> > Maybe I'm missing something...
> 
> OK, so that helps in terms of understanding the issue you are seeing with
> it. I'll take a look at this, but it'll probably be a few days. You can
> try and enable tracing, I see events completed just fine. Maybe a race
> with your epoll wait and event reaping?

(discard previous email wrongly sent by my phone, sorry for the noise)

Okay, the issue was that my kernel doesnât support IORING_FEAT_NODROP,
so for this reason I missed the CQ events.

Avoiding the CQ overflow keeping this condition true
(submitted - completed < n_entries), solves the issue.

Iâll try with a mainline kernel, handling also the -EBUSY returned by
io_uring_submit().

Thanks,
Stefano