On 8/17/22 18:49, David Matlack wrote:
On Wed, Aug 17, 2022 at 3:05 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
On 8/16/22 01:01, David Matlack wrote:
Delete the module parameter tdp_mmu and force KVM to always use the TDP
MMU when TDP hardware support is enabled.
The TDP MMU was introduced in 5.10 and has been enabled by default since
5.15. At this point there are no known functionality gaps between the
TDP MMU and the shadow MMU, and the TDP MMU uses less memory and scales
better with the number of vCPUs. In other words, there is no good reason
to disable the TDP MMU.
Dropping the ability to disable the TDP MMU reduces the number of
possible configurations that need to be tested to validate KVM (i.e. no
need to test with tdp_mmu=N), and simplifies the code.
The snag is that the shadow MMU is only supported on 64-bit systems;
testing tdp_mmu=0 is not a full replacement for booting up a 32-bit
distro, but it's easier (I do 32-bit testing only with nested virt).
Ah, I did forget about 32-bit systems :(. Do Intel or AMD CPUs support
TDP in 32-bit mode?
Both do. Intel theoretically on some 32-bit processors that were
actually sold, too.
Personally I'd have no problem restricting KVM to x86-64 but I know
people would complain.
As a middle-ground we could stop supporting TDP on 32-bit
systems. 32-bit KVM would continue working but just with shadow
paging.
The main problem is, shadow paging is awfully slow due to Meltdown
mitigations these days. I would start with the read-only parameter and
then see whether there's more low-hanging fruit (e.g. make fast page
fault TDP MMU-only).
Paolo
What about making the tdp_mmu module parameter read-only, so that at
least kvm->arch.tdp_mmu_enabled can be replaced by a global variable?
Paolo
