On Fri, Mar 6, 2020 at 6:26 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote: > > The ABI is broken and we cannot support it properly. Turn it off. > > If this causes a meaningful performance regression for someone, KVM > can introduce an improved ABI that is supportable. > > Cc: stable@xxxxxxxxxxxxxxx > Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx> > --- > arch/x86/kernel/kvm.c | 21 ++++++++++++++++++--- > 1 file changed, 18 insertions(+), 3 deletions(-) > > diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c > index 93ab0cbd304e..e6f2aefa298b 100644 > --- a/arch/x86/kernel/kvm.c > +++ b/arch/x86/kernel/kvm.c > @@ -318,11 +318,26 @@ static void kvm_guest_cpu_init(void) > > pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); > > -#ifdef CONFIG_PREEMPTION > - pa |= KVM_ASYNC_PF_SEND_ALWAYS; > -#endif > pa |= KVM_ASYNC_PF_ENABLED; > > + /* > + * We do not set KVM_ASYNC_PF_SEND_ALWAYS. With the current > + * KVM paravirt ABI, the following scenario is possible: > + * > + * #PF: async page fault (KVM_PV_REASON_PAGE_NOT_PRESENT) > + * NMI before CR2 or KVM_PF_REASON_PAGE_NOT_PRESENT > + * NMI accesses user memory, e.g. due to perf > + * #PF: normal page fault > + * #PF reads CR2 and apf_reason -- apf_reason should be 0 > + * > + * outer #PF reads CR2 and apf_reason -- apf_reason should be > + * KVM_PV_REASON_PAGE_NOT_PRESENT > + * > + * There is no possible way that both reads of CR2 and > + * apf_reason get the correct values. Fixing this would > + * require paravirt ABI changes. > + */ > + Upon re-reading my own comment, I think the problem is real, but I don't think my patch fixes it. The outer #PF could just as easily have come from user mode. We may actually need the NMI code (and perhaps MCE and maybe #DB too) to save, clear, and restore apf_reason. If we do this, then maybe CPL0 async PFs are actually okay, but the semantics are so poorly defined that I'm not very confident about that.
