On Sat, Mar 7, 2020 at 7:47 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > > Andy Lutomirski <luto@xxxxxxxxxx> writes: > > On Fri, Mar 6, 2020 at 6:26 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote: > >> + /* > >> + * 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. > > I think even with the current mode this is fixable on the host side when > it keeps track of the state. > > The host knows exactly when it injects a async PF and it can store CR2 > and reason of that async PF in flight. > > On the next VMEXIT it checks whether apf_reason is 0. If apf_reason is 0 > then it knows that the guest has read CR2 and apf_reason. All good > nothing to worry about. > > If not it needs to be careful. > > As long as the apf_reason of the last async #PF is not cleared by the > guest no new async #PF can be injected. That's already correct because > in that case IF==0 which prevents a nested async #PF. > > If MCE, NMI trigger a real pagefault then the #PF injection needs to > clear apf_reason and set the correct CR2. When that #PF returns then the > old CR2 and apf_reason need to be restored. How is the host supposed to know when the #PF returns? Intercepting IRET sounds like a bad idea and, in any case, is not actually a reliable indication that #PF returned.
