Re: getting the console output for s390 cdrom-test?

[Peter Maydell]

On Tue, 9 Feb 2021 at 17:10, Cornelia Huck <cohuck@redhat.com> wrote:
>
> On Tue, 9 Feb 2021 14:58:53 +0000
> Peter Maydell <peter.maydell@linaro.org> wrote:
> > This change significantly reduces the frequency with which I see
> > the hang; but it doesn't get rid of it altogether. Also I couldn't
> > really figure out from the virtio spec exactly where barriers
> > were required: I think I would need to read the whole thing in
> > more detail rather than trying to fish out the information by
> > just reading small pieces of it.
>
> The Linux virtio-ccw code uses 'weak barriers', i.e. the heavy bcr15
> should not be needed. We might well miss other (lightweight) barriers
> in other parts of the code part, though.

Is that the version the Linux kernel has as
  /* Fast-BCR without checkpoint synchronization */
  #define __ASM_BARRIER "bcr 14,0\n"

?

> > But some of the ordering of
> > operations the spec describes doesn't match how the s390-ccw
> > BIOS code is doing it at all (eg the spec says that when feeding
> > a batch of descriptors to the device the driver isn't supposed to
> > update the flags on the first descriptor until it's finished
> > writing all of the descriptors, but the code doesn't seem to
> > try to do that). So I think the code could use an overhaul from
> > somebody with a better understanding of virtio than me...
>
> Yeah, the bios virtio code could probably use some love.
>
> I'm wondering how much memory ordering on the host platform influences
> things. I doubt many people try to run an s390x guest on an aarch64
> host...

Yes, you won't see this bug unless you're running QEMU on a
host that's pretty enthusiastic about reordering memory
transactions (and you'd never have seen it at all back when
we ran the iothread actions synchronously with the emulated
CPU, which we probably did back in 2013 when the s390-ccw
virtio code was written...) I haven't tested other aarch64
hosts but I would be unsurprised to find that whether you
could repro it and how frequently depended on the particular
h/w implementation.

thanks
-- PMM