> On 15 Oct 2018, at 20:07, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > On 08/10/2018 20:29, Jim Mattson wrote: >> Under nested virtualization, the L1 hypervisor may intercept an >> exception raised during the execution of L2 before the exception >> is delivered. When the intercepted exception is #PF, the VM-exit >> to the L1 hypervisor precedes the modification of CR2. When the >> intercepted exception is #DB, the VM-exit to the L1 hypervisor >> precedes the modifications of DR6 and DR7 under VMX, but the >> VM-exit to the L1 hypervisor follows the modifications of DR6 and >> DR7 under SVM. >> >> At present, CR2 is modified too early under both VMX and SVM. DR6 is >> modified too early under VMX. DR7 is modified at the appropriate time. >> Unfortunately, it is possible to exit to userspace with one of these >> exceptions pending, and userspace may rely on the premature >> side-effects. It is also possible for userspace to inject one of these >> exceptions, in which case, userspace will presumably have already >> processed the side-effects. >> >> To address this problem, a new per-VM capability >> (KVM_CAP_EXCEPTION_PAYLOAD) will be introduced. When this capability >> is enabled by userspace, the faulting linear address will be included >> with the information about a pending #PF in L2, and the "new DR6 bits" >> will be included with the information about a pending #DB in L2. This >> ancillary exception information is carried in a new "payload" field. >> >> Reported-by: Jim Mattson <jmattson@xxxxxxxxxx> >> Suggested-by: Paolo Bonzini <pbonzini@xxxxxxxxxx> >> Signed-off-by: Jim Mattson <jmattson@xxxxxxxxxx> >> Reviewed-by: Peter Shier <pshier@xxxxxxxxxx> >> > > I'm not applying any of the patches yet, but I'd be happy to queue it > for 4.20 even (slightly) after -rc1. Once this is done, I think we can > flip nested=1. > > Paolo Just out of curiosity, why is this specifically a decision-point for flipping nested=1? Because this is a break of userspace APIs in case of nVMX workloads? What are the conditions for flipping it? If that’s the case, this is also true for the get/set vCPU events IOCTLs. So they need to also be applied before the flip. Don’t get me wrong, I will be happy to see nested=1 by default. Just wondering… -Liran
