On Thu, Apr 18, 2024, Thomas Prescher wrote: > You are right. After your pointers and looking at the nesting code > again, I think I know what to do. Just to make sure I understand this > correctly: > > If L0 exits with L2 state, KVM_GET_NESTED_STATE will have > KVM_STATE_NESTED_RUN_PENDING set in the flags field. Not necessarily. KVM_STATE_NESTED_GUEST_MODE is the flag that says "L2 state is loaded", the NESTED_RUN_PENDING flag is effectively a modifier on top of that. KVM_STATE_NESTED_RUN_PENDING is set when userspace interrupts KVM in the middle of nested VM-Enter emulation. In that case, KVM needs to complete emulation of the VM-Enter instruction (VMLAUNCH, VMRESUME, or VMRUN) before doing anything. I.e. KVM has loaded L2 state and is committed to completing VM-Enter, but hasn't actually done so yet. In retrospect, KVM probably should have forced userspace to call back into KVM to complete emulation before allowing KVM_GET_NESTED_STATE to succeed, but it's a minor blip. > So when we restore the vCPU state after a vmsave/vmload cycle, we don't need > to update anything in kvm_run.s.regs because KVM will enter the L2 > immediately. Is that correct? No? Presumably your touching vCPU state, otherwise you wouldn't be doing vmsave/vmload. And if you touch vCPU state, then you need to restore the old state for things to work. Again, what are you trying to do, at a higher level? I.e. _why_ are you doing a save/restore cycle? If it's for something akin to live migration, where you need to save and restore *everything*, then stating the obvious, you need to save and restore everything in KVM too, which includes nested state.
