On 07/07/21 01:06, stsp wrote:
What I ask is how SHOULD the KVM_SET_REGS and KVM_SET_SREGS behave when someone (mistakenly) calls them with the exception pending. Should they return an error instead of canceling exception?
In theory, KVM_SET_REGS and KVM_SET_SREGS should do nothing but set the value of the registers. They not should clear either vcpu->arch.exception.pending or vcpu->arch.exception.injected. I'm wary of changing that and breaking users of KVM, though.
In this case the problem is that, with a pending exception, you should not inject the interrupt (doesn't matter if it's with KVM_SET_REGS or KVM_INTERRUPT). Raising a page fault is part of executing the previous instruction, and interrupts are only recognized at instruction boundaries. Therefore, you need to test ready_for_interrupt_injection, and possibly use request_interrupt_window, before calling KVM_SET_REGS.
The patch you identified as the culprit does have a bug, but that's fixed in kvm_cpu_accept_dm_intr as I suggested in the other thread.
Paolo
