Re: [PATCH] [RFC] vpci: allow BAR write while mapped

["Alejandro Vallejo" <alejandro.vallejo@cloud.com>]

On Mon Mar 17, 2025 at 12:32 PM GMT, Roger Pau Monné wrote:
> On Fri, Mar 14, 2025 at 10:39:04AM +0000, Alejandro Vallejo wrote:
> > On Thu Mar 13, 2025 at 5:43 PM GMT, Stewart Hildebrand wrote:
> > > I think this just reinforces the need to rework the BAR mapping
> > > machinery.
> > 
> > Right. The most delicate part is dealing with races with another vCPU when the
> > unmap-then-map operation does not complete in a single taking of the vpci lock
> > I'd say. And that much is unavoidable, I think, because either unmapping or
> > mapping might take a while.
>
> The original code was put together for dom0, I bet there are some
> races that can lead to incomplete p2m mappings if the domain attempts
> parallel manipulation of the BARs and memory decoding bits.  However
> there should be no case where an unexpected mfn gets mapped into the
> p2m as a result of such races.
>
> It's fine for a domain to shot it's own foot if it attempts to do
> concurrent PCI accesses to explicitly trigger races in the MMIO
> mapping handling, as long as this doesn't cause issues to other
> guests, and doesn't leak memory.
>
> Regards, Roger.

It's refcounting bugs that would scare me the most. Or something similar, like
having an expectation that a BAR is absent from a certain p2m when it's still
present.

Cheers,
Alejandro