On Mon, Jan 14, 2013 at 2:00 PM, Will Deacon [off-list ref] wrote:

On Mon, Jan 14, 2013 at 06:53:14PM +0000, Alexander Graf wrote:

quoted

On 01/14/2013 07:50 PM, Will Deacon wrote:

quoted

FWIW, KVM only needs this code for handling complex MMIO instructions, which
aren't even generated by recent guest kernels. I'm inclined to suggest removing
this emulation code from KVM entirely given that it's likely to bitrot as
it is executed less and less often.

That'd mean that you heavily limit what type of guests you're executing,
which I don't think is a good idea.

To be honest, I don't think we know whether that's true or not. How many
guests out there do writeback accesses to MMIO devices? Even on older
Linux guests, it was dependent on how GCC felt.

I don't think bitrot'ing is a valid argument: the code doesn't depend
on any other implementation state that's likely to change and break
this code (the instruction encoding is not exactly going to change).
And we should simply finish the selftest code to test this stuff
(which should be finished if the code is unified or not, and is on my
todo list).

I see where you're coming from, I just don't think we can quantify it either
way outside of Linux.

FWIW, I know of at least a couple of companies wanting to use KVM for
running non-Linux guests as well.

But, however a shame, I can more easily maintain this single patch
out-of-tree, so I'm willing to drop this logic for now if it gets
things moving.

-Christoffer