On 29/11/2017 12:44, Eduardo Habkost wrote: > On Mon, Nov 13, 2017 at 09:32:09AM +0100, Paolo Bonzini wrote: >> On 13/11/2017 08:15, Wanpeng Li wrote: >>> 2017-11-10 17:49 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>: >>>> Sometimes, a processor might execute an instruction while another >>>> processor is updating the page tables for that instruction's code page, >>>> but before the TLB shootdown completes. The interesting case happens >>>> if the page is in the TLB. >>>> >>>> In general, the processor will succeed in executing the instruction and >>>> nothing bad happens. However, what if the instruction is an MMIO access? >>>> If *that* happens, KVM invokes the emulator, and the emulator gets the >>>> updated page tables. If the update side had marked the code page as non >>>> present, the page table walk then will fail and so will x86_decode_insn. >>>> >>>> Unfortunately, even though kvm_fetch_guest_virt is correctly returning >>>> X86EMUL_PROPAGATE_FAULT, x86_decode_insn's caller treats the failure as >>>> a fatal error if the instruction cannot simply be reexecuted (as is the >>>> case for MMIO). And this in fact happened sometimes when rebooting >>>> Windows 2012r2 guests. Just checking ctxt->have_exception and injecting >>>> the exception if true is enough to fix the case. >>> >>> I found the only place which can set ctxt->have_exception is in the >>> function x86_emulate_insn(), and x86_decode_insn() will not set >>> ctxt->have_exception even if kvm_fetch_guest_virt() returns >>> X86_EMUL_PROPAGATE_FAULT. >> >> Hmm, you're right. Looks like Yanan has been (un)lucky when trying out >> this patch! :( >> >> Yanan, can you double check that you can reproduce the issue with an >> unpatched kernel? I will work on a kvm-unit-tests testcsae > > We don't have a kvm-unit-tests reproducer for this yet, right? > > I'm considering trying to write one, but I don't want to > duplicate work. No, I haven't written one yet. Paolo
