On Fri, 2022-09-23 at 14:53 +0100, Pavel Begunkov wrote:
> Overflowing CQEs may result in reordeing, which is buggy in case of
> links, F_MORE and so.
>
Maybe the commit message got cut off?
I think this is probably ok, the downside being that CQE's with no
ordering constraints will have ordering forced on them. An alternative
would be for each case (eg linked, zerocopy, multishot) to either pause
or force CQE's to be overflow ones. This wouldnt slow down the other
codepaths. I don't have an idea for how difficult this might be.
By the way, if you do go with this approach then I believe you can
revert these patches:
a2da676376fe ("io_uring: fix multishot poll on overflow")
cbd25748545c ("io_uring: fix multishot accept ordering")
and do something similar for multishot receive.
Dylan
