On 02/07/2018 10:16 PM, Paolo Bonzini wrote:
On 07/02/2018 07:25, Wanpeng Li wrote:diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index 786cd00..445e702 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -7458,6 +7458,11 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run) goto out; }+ if (unlikely(vcpu->run->exit_reason == KVM_EXIT_SHUTDOWN)) {+ r = -EINVAL; + goto out; + } + if (vcpu->run->kvm_dirty_regs) { r = sync_regs(vcpu); if (r != 0)This most likely breaks triple faults in the usual case where they should result in resetting the system; the KVM API doesn't say that you should clear vcpu->run->exit_reason before entering. What exactly causes the EPT misconfig to reach the WARN? That is, how does kvm_mmu_page_fault end up returning a negative errno value? If I read the code correctly only tdp_page_fault can do so, so my guess would be kvm_handle_bad_page: if (pfn == KVM_PFN_ERR_RO_FAULT) return RET_PF_EMULATE; if (pfn == KVM_PFN_ERR_HWPOISON) { kvm_send_hwpoison_signal(kvm_vcpu_gfn_to_hva(vcpu, gfn), current); return RET_PF_RETRY; } /* KVM_PFN_ERR_FAULT */ return -EFAULT; Maybe it should return RET_PF_EMULATE, which would cause an emulation failure and then an exit with KVM_EXIT_INTERNAL_ERROR.
So the root cause is that a running vCPU accessing the memory whose memslot is being updated (met the condition KVM_MEMSLOT_INVALID is set on the its memslot). The normal #PF handler breaks KVM_RUN and returns -EFAULT to userspace, we'd better to make ept-misconfig's handler follow this style as well. Actually, the WARN_ON in ept-misconfig's handler is unnecessary as kvm_mmu_page_fault() will warn us if it is the real ept misconfig, so we can simply return kvm_mmu_page_fault().
