Re: [Qemu-devel] [RFC] [PATCH 0/3] qemu: arm: Migration between machines with different MIDR values

[Peter Maydell]

On 4 October 2018 at 16:05, Andrew Jones <address@hidden> wrote:
> On Tue, Oct 02, 2018 at 02:07:38PM +0100, Peter Maydell wrote:
>> Hi; thanks for this patch. The issue I see with this patch
>> is that the KVM/ARM QEMU approach to system registers so far
>> has been "the kernel knows about these and it is in control".
>> So we ask the kernel for the list of registers, and just save
>> and restore those. That would suggest that if there are sysregs
>> where it's OK in fact to ignore a difference between two constant
>> register values, it should be the kernel doing the "actually, this
>> mismatch is OK" behaviour...
>
> I don't think the kernel should have to maintain all that logic. If a user
> wants to load guest registers, then the kernel should do it, as long as
> it's safe from the host's integrity/security PoV, and the hardware would
> actually support it. Anything that can only break the guest, but not the
> host, should be allowed. The KVM userspace can certainly ask the kernel
> what it recommends first (i.e. read the invariant registers first, before
> deciding what to write), but the decision of what to write should be left
> up to the user.

I can see the logic of that approach. But right now QEMU
userspace knows basically nothing about the system registers
when we're using KVM: all that knowledge is in the kernel.
So we don't have a place really to put policy info (and
not really anywhere to put policy info that depends on the
host CPU type when mostly we let the kernel care about that).

>> For instance, it's probably OK to ignore a MIDR_EL1 difference
>> that just indicates a minor revision bump; but not to ignore
>> one that indicates you just tried to migrate a Cortex-A53
>> over to a Cavium CPU.
>
> If the user does that, then the guest will break - oh well. That's not the
> host kernel's problem.

Patching QEMU to ignore all attempts to write wrong values
to read-only registers would be easy. It just wouldn't
be very helpful to the user...

thanks
-- PMM