On Mon, 2021-05-10 at 16:56 +0000, Sean Christopherson wrote: > On Mon, May 10, 2021, Maxim Levitsky wrote: > > On Tue, 2021-05-04 at 14:58 -0700, Jim Mattson wrote: > > > On Tue, May 4, 2021 at 2:57 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > > > On 04/05/21 23:53, Sean Christopherson wrote: > > > > > > Does the right thing happen here if the vCPU is in guest mode when > > > > > > userspace decides to toggle the CPUID.80000001H:EDX.RDTSCP bit on or > > > > > > off? > > > > > I hate our terminology. By "guest mode", do you mean running the vCPU, or do > > > > > you specifically mean running in L2? > > > > > > > > > > > > > Guest mode should mean L2. > > > > > > > > (I wonder if we should have a capability that says "KVM_SET_CPUID2 can > > > > only be called prior to KVM_RUN"). > > > > > > It would certainly make it easier to reason about potential security issues. > > > > > I vote too for this. > > Alternatively, what about adding KVM_VCPU_RESET to let userspace explicitly > pull RESET#, and defining that ioctl() to freeze the vCPU model? I.e. after > userspace resets the vCPU, KVM_SET_CPUID (and any other relevant ioctls() is > disallowed. I vote even stronger for this! I recently created what Paulo called an 'evil' test to stress KVM related bugs in nested migration by simulating a nested migration with a vCPU reset, followed by reload of all of its state including nested state. It is ugly since as you say I have to manually load the reset state, and thus using 'KVM_VCPU_RESET' ioctl instead would make this test much more cleaner and even more 'evil'. Best regards, Maxim Levitsky > > Lack of proper RESET emulation is annoying, e.g. userspace has to manually stuff > EDX after vCPU creation to get the right value at RESET. A dedicated ioctl() > would kill two birds with one stone, without having to add yet another "2" > ioctl(). >
