On Wed, 2022-12-21 at 10:55 +0800, Yuan Yao wrote:
> > +#ifdef CONFIG_X86_64
> > +/* untag addr for guest, according to vCPU CR3 and CR4 settings */
> > +static inline u64 kvm_untagged_addr(u64 addr, struct kvm_vcpu
> > *vcpu)
> > +{
> > + if (addr >> 63 == 0) {
> > + /* User pointers */
> > + if (kvm_read_cr3(vcpu) & X86_CR3_LAM_U57)
> > + addr = get_canonical(addr, 57);
> > + else if (kvm_read_cr3(vcpu) & X86_CR3_LAM_U48) {
> > + /*
> > + * If guest enabled 5-level paging and LAM_U48,
> > + * bit 47 should be 0, bit 48:56 contains meta
> > data
> > + * although bit 47:56 are valid 5-level address
> > + * bits.
> > + * If LAM_U48 and 4-level paging, bit47 is 0.
> > + */
> > + WARN_ON(addr & _BITUL(47));
>
> IIUC, LAM_47 userspace canonical checking rule requests "bit 63 ==
> bit 47 == 0"
> before sign-extened the address.
>
> if so looks it's guest's fault to not follow the LAM canonical
> checking rule,
> what's the behavior of such violation on bare metal, #GP ?
Spec (ISE 10.2) doesn't mention a #GP for this case. IIUC, those
overlap bits are zeroed.
Current native Kernel LAM only enables LAM_U57.
> The behavior
> shuld be same for guest instead of WARN_ON() on host, host does
> nothing wrong.
>
Agree that host did nothing wrong. So remove WARN_ON(), use a pr_warn()
instead? I here meant to warn that such case happens.
