On Mon, Dec 14, 2015 at 07:17:27PM -0500, Raj, Ashok wrote:

On Mon, Dec 14, 2015 at 11:37:16PM +0100, Borislav Petkov wrote:

...

On Mon, Dec 14, 2015 at 02:11:46PM -0500, Raj, Ashok wrote:

...

This is mostly harmless.. since the MCG_CAP space is shared and has no conflict between vendors. Also just the CAP being set has no effect.

Of course it does - we check SER_P in machine_check_poll() and when I emulate an AMD guest and inject errors into it, error handling is obviously wrong, see:

https://lkml.kernel.org/r/20151123150355.GE5134@pd.tnic

I can see how this hurts.. since the poller isn't doing cpu model specific stuff..?

in the LMCE case, even if you advertise MCG_LMCE_P in MCG_CAP, the guest kernel wont call intel_init_lmce() only from mce_intel.c.. so the same problem won't happen.

but the issue Eduardo mentioned seems like the following.

New QEMU_LMCE + New KVM_LMCE + New_GUEST_LMCE - No problem

but if you were to migrage the Guest_LMCE to a non-LMCE supported KVM host we could run into an issue..

is this the compatibility issue that you were looking to fix Eduardo?

If I understood you correctly, yes. Also, note that currently kvm_arch_init_vcpu() simply warns about missing capabilities, instead of preventing the VM from running/migrating (as it should). We need to change that, and figure out a good way to report "feature FOO can't be enabled in this host" errors to management software[1]. The main problem is that we don't even have a QMP console available anymore if machine initialization is aborted. CCing libvir-list so they get in the loop. [1] This is similar to what we need for CPUID checks, but the new MCE feature means we need something more generic (that just reports QOM property names, probably?) -- Eduardo