On Tue, 11 May 2021 09:22:37 +0100, Fuad Tabba tabba@google.com wrote:
Hi Marc,
On Tue, May 11, 2021 at 9:14 AM Marc Zyngier maz@kernel.org wrote:
Hi Fuad,
On Tue, 11 May 2021 09:03:40 +0100, Fuad Tabba tabba@google.com wrote:
Hi Marc,
KVM: arm64: Commit pending PC adjustemnts before returning to userspace
s/adjustments/adjustments
Looks like Gmail refuses to let you mimic my spelling mistakes! :D
On Mon, May 10, 2021 at 10:49 AM Marc Zyngier maz@kernel.org wrote:
KVM currently updates PC (and the corresponding exception state) using a two phase approach: first by setting a set of flags, then by converting these flags into a state update when the vcpu is about to enter the guest.
However, this creates a disconnect with userspace if the vcpu thread returns there with any exception/PC flag set. In this case, the exposed context is wrong, as userpsace doesn't have access to these flags (they aren't architectural). It also means that these flags are preserved across a reset, which isn't expected.
To solve this problem, force an explicit synchronisation of the exception state on vcpu exit to userspace. As an optimisation for nVHE systems, only perform this when there is something pending.
I've tested this with a few nvhe and vhe tests that exercise both __kvm_adjust_pc call paths (__kvm_vcpu_run and kvm_arch_vcpu_ioctl_run), and the tests ran as expected. I'll do the same for v2 when you send it out.
Ah, that's interesting. Do you have tests that actually fail when hitting this bug? Given that this is pretty subtle, it'd be good to have a way to make sure it doesn't crop up again.
Nothing that fails, just code that generates exceptions or emulates instructions at various points. That said, I think it should be straightforward to write a selftest for this. I'll give it a go.
PC adjustment is easy-ish: have a vcpu to hit WFI with no interrupt pending, send the thread a signal to make it exit to userspace, update the PC to another address, and check that the instruction at that address is actually executed.
Exception injection is a lot more difficult: you need to force a vcpu exit to userspace right after having caused an exception to be injected by KVM. I can't think of an easy way to do that other than repeatedly executing an instruction that generates an exception while signalling the thread to force the exit. Ugly.
M.