Re: [PATCH 0/8] Add support for pvpanic mmio device

[Peter Maydell]

On Mon, 26 Oct 2020 at 13:51, Mihai Carabas <mihai.carabas@oracle.com> wrote:
> On 10/22/2020 1:17 PM, Peter Maydell wrote:
> > I'm afraid the answer is still the same. You need to provide
> > a convincing argument for why this needs to be an MMIO
> > device rather than a PCI device. I really don't want to
> > add MMIO devices to the virt board if I can avoid it,
> > because they're all extra code and potential extra
> > security boundary attack surface. PCI devices are guest
> > probeable and user-pluggable so they're almost always
> > nicer to use than MMIO.
> >
>
> Thank you for your input.
>
> The reason why pvpanic should be MMIO is that is a special device which
> is not used commonly by the user (aka VM), it is not need to be
> hot-plugable and it does not have a hardware correspondent to be a PCI
> device.

"Not used commonly by the user" is a good argument for
"just use PCI" -- users who want the functionality can add
it to their QEMU command line or VM config, and the bulk of
users who don't don't have to worry about it.

PCI devices don't have to support hotplug; I agree you don't
need hotplug for this use case.

Using a PCI device would match up with the way the x86 pvpanic
device is an ISA bus device.

> Another reason is that MMIO support was accepted in the kernel
> driver and it is pretty useless there without a device.

> I know it seems that I want to get this on the short-path, but at this
> point having a kernel driver in the upstream and no device to test it
> against it is pretty weird.

That's the kernel folks' problem to deal with, because they accepted
the kernel driver without having confirmed that QEMU was going to
implement it. If I were them I'd deprecate and delete that driver code.
(They should certainly fix the bit of the documentation that claims
that QEMU implements it.)

thanks
-- PMM