Re: [PATCH] qrap: Pass PCI device numbers to qemu in base 10, not in base 16

[Andrea Bolognani <abologna@redhat.com>]

On Fri, Feb 24, 2023 at 08:14:16AM +0100, Stefano Brivio wrote:
> On Thu, 23 Feb 2023 06:06:17 -0800 Andrea Bolognani <abologna@redhat.com> wrote:
> > On Thu, Feb 23, 2023 at 09:27:14AM +1100, David Gibson wrote:
> > > On Wed, Feb 22, 2023 at 02:40:32AM -0800, Andrea Bolognani wrote:
> > > > I don't think this is going to work.
> > > >
> > > > The problem is that, while PCI buses are indeed named with increasing
> > > > numbers in integer format (pci.9, pci.10 and so on), PCI slots are
> > > > addressed using hexadecimal format (0x9, 0xa and so on). libvirt uses
> > > > this naming convention because it matches QEMU's.
> > >
> > > Actually, I think we're ok.  PCI slots are addressed in hex by
> > > convention, but AFAICT if you *just* give a slot number, it will
> > > accept either decimal or hex (so addr=10 and addr=0xa are equivalent).
> > > That's *not* true if you use SS.F format to include the function
> > > number - then it expects hex only.  But we're not doing that, so so
> > > always using decimal should be ok here.
> > >
> > > Source: set_pci_devfn() in the qemu source
> > >
> > > Obviously that's a pretty fragile hack, but that's 'qrap' for you.
> >
> > Yeah, even if that happens to work I'd rather not rely on it,
> > especially since a proper solution doesn't look like it would be a
> > lot of additional effort.
> >
> > I've managed to reproduce the original issue in the context of
> > KubeVirt. I'll hopefully have a patch ready soon.
>
> Andrea, allow me to do this: I would push this patch meanwhile, along
> with the changes for the DNS issue you reported, because that one might
> impact many users, and I think it makes sense to have a fix out soon.
>
> I start thinking it's also part of the issue Paul reported for Podman
> with pasta here:
>   https://github.com/containers/podman/issues/17074
>
> This patch itself can't hurt, and it changes exactly two letters.

I strongly disagree with this assessment. This patch merely trades
one set of issues for another one.

In particular, for pc machine types we'd end up producing
bus=pci.0,addr=0x10 for slot 10 instead of bus=pci.0,addr=0xa,
because the addr=0x part is baked into the template. So the QEMU
logic David mentioned above wouldn't kick in at all.

More importantly, for q35 machines we'd start producing decimal bus
numbers while still parsing the ones present in the original command
line as hexadecimal, so things would stop lining up as soon as enough
devices are present, meaning that the issue reported by Alona would
still exist.

> As soon as you have something less qrappy we'll go with that (you don't
> even need to rebase, I'll revert this one on the tree first).

Patches fixing this issue, as well as a few additional ones, are now
on the list. I'll follow up on that thread with some considerations
related to testing the changes.

-- 
Andrea Bolognani / Red Hat / Virtualization