> On Wed, Feb 07, 2024, Xin Li wrote: > > @@ -7382,6 +7419,24 @@ static noinstr void vmx_vcpu_enter_exit(struct > kvm_vcpu *vcpu, > > > > vmx_disable_fb_clear(vmx); > > > > + /* > > + * %cr2 needs to be saved after a VM exit and restored before a VM > > + * entry in case a VM exit happens immediately after delivery of a > > + * guest #PF but before guest reads %cr2. > > + * > > + * A FRED guest should read its #PF faulting linear address from > > + * the event data field in its FRED stack frame instead of %cr2. > > + * But the FRED 5.0 spec still requires a FRED CPU to update %cr2 > > + * in the normal way, thus %cr2 is still updated even for a FRED > > + * guest. > > + * > > + * Note, an NMI could interrupt KVM: > > + * 1) after VM exit but before CR2 is saved. > > + * 2) after CR2 is restored but before VM entry. > > + * And a #PF could happen durng NMI handlng, which overwrites %cr2. > > + * Thus exc_nmi() should save and restore %cr2 upon entering and > > + * before leaving to make sure %cr2 not corrupted. > > + */ > > This is 99.9% noise. What software does or does not do with respect to CR2 is > completely irrelevant. The *only* thing that matters is the architectural > behavior, and architecturally guest CR2 _must_ be up-to-date at all times because > CR2 accesses cannot be intercepted. So, just say: > > /* > * Note, even though FRED delivers the faulting linear address via the > * event data field on the stack, CR2 is still updated. > */ Will do! There is a reason for this comment because it won't be architectural: https://lore.kernel.org/lkml/c0c7c605-d487-483e-a034-983b76ee1dfa@xxxxxxxxx/ FRED is designed to atomically save and restore _full_ supervisor/user context upon event delivery and return. But unfortunately, KVM still has to save/restore guest CR2 explicitly due to the issue mentioned above. Thanks! Xin
