Re: [PATCH] pc-bios/s390-ccw: Use memory barriers in virtio code

[Cornelia Huck <cohuck@redhat.com>]

On Tue, 16 Feb 2021 14:37:22 +0000
Peter Maydell <peter.maydell@linaro.org> wrote:

> On Tue, 16 Feb 2021 at 14:35, Thomas Huth <thuth@redhat.com> wrote:
> >
> > On 16/02/2021 15.32, Peter Maydell wrote:  
> > > On Tue, 16 Feb 2021 at 14:30, Cornelia Huck <cohuck@redhat.com> wrote:  
> > >> Step 4 in "2.7.13 Supplying Buffers to The Device":
> > >>
> > >> "The driver performs a suitable memory barrier to ensure the device
> > >> sees the updated descriptor table and available ring before the next
> > >> step."  
> > >
> > > I thought that my first time through the spec as well, but
> > > I think the whole of section 2.7 is dealing with "packed virtqueues",
> > > which have to be explicitly negotiated and which I don't think
> > > the s390-ccw code is doing.  

2.7 is split, 2.8 packed (at least in my built-from-git version; maybe
I should have mentioned that :)

> >
> > Right. I think the s390-ccw code is still based on virtio v1.0, that's why I
> > also only looked at that version of the spec.  

Yes, the bios code is using split.

> I think the ideal would be to find somebody who's really well
> acquainted with the virtio spec (MST?) and ask them to have
> a quick look-over the s390-ccw code to say where it needs
> changes... The fact that this patch doesn't completely fix
> the bug leaves me suspicious that we're missing something in
> our readings of the spec.

Unfortunately, the bios virtio code is not that easy to follow :/