On 21/09/2018 04:08, peng.hao2@xxxxxxxxxx wrote: >>> Unqueued, sorry. The hypercall test from kvm-unit-tests fails. A >>> VMCALL on the "edge" of canonical address space, i.e. at 0x7ffffffffffd, >>> raises a general protection fault before this patch and a double fault >>> afterwards. >> Peng Hao submitted a very similar patch[1], the difference being that >> it also modified x86_emulate_instruction() to ignore the result of >> inject_emulated_exception(). I have no idea if that change is correct >> but it may be related to the aforementioned unit test failing. > The return value of inject_emulated_exception for #PF represents whether in nested vcpu, > not about the success or failure. Because we are not in nested vcpu, inject_emulated_exception > must return false. And for other exceptions, inject_emulated_exception just return false. > After propagating exceptions to guest, we just should return to guest mode and let guest handle. Yes, your patch should work, and the x86_emulate_instruction code that is there now is clearly wrong (though I'm not sure how the double-fault comes up, #GP + #UD shouldn't generate one because #UD is a benign exception). Peng Hao, do you have a testcase? Eduardo's kvm-unit-tests patch from Nov 2017 doesn't reproduce the bug anymore for me. Thanks, Paolo
