On Fri, 2024-07-26 at 08:47 -0400, Michael S. Tsirkin wrote:

On Fri, Jul 26, 2024 at 09:06:29AM +0100, David Woodhouse wrote:

quoted

That's great. You don't even need it to be per-vCPU if you let the
hypervisor write directly to the single physical location that's mapped
to userspace. It can do that before it even starts *running* the vCPUs
after migration. It's a whole lot simpler. 

It *seems* simpler, until you realize that there is no way
to change anything in the interface, there is no negotiation
between hypervisor and userspace. If I learned anything at all
in tens of years working on software, it's that it is
never done. So let's have userspace talk to the kernel
and have kernel talk to the devices, please. There's
no compelling reason to have this bypass here.

Thanks for the useful feedback. As you see, I've incorporated most of
it into the v2 post a few minutes ago.

On this particular topic we disagree. I absolutely don't want to take
dependencies on kernel code, on virtio, or on the cross-platform
assumption (even if it's true) that a device can raise an interrupt and
guarantee that no userspace code will run before that interrupt is
handled.

Using virtio does allow for some negotiation — for handling differences
in page sizes, and enabling the timekeeping only on demand. That's
great, but the structure is still an ABI in its own right, and we know
how to do those.

We'll ship the ACPI version, and I look forward to incorporating it
into virtio-rtc too.