Re: [PATCH] Revert "KVM: x86: WARN and reject loading KVM if NX is supported but not enabled"

[Paolo Bonzini]

On 09/07/21 19:21, Sean Christopherson wrote:
> On Thu, Jul 08, 2021, Paolo Bonzini wrote:
> > On 25/06/21 02:18, Sean Christopherson wrote:
> > > Let KVM load if EFER.NX=0 even if NX is supported, the analysis and
> > > testing (or lack thereof) for the non-PAE host case was garbage.
> > >
> > > If the kernel won't be using PAE paging, .Ldefault_entry in head_32.S
> > > skips over the entire EFER sequence.  Hopefully that can be changed in
> > > the future to allow KVM to require EFER.NX, but the motivation behind
> > > KVM's requirement isn't yet merged.  Reverting and revisiting the mess
> > > at a later date is by far the safest approach.
> > >
> > > This reverts commit 8bbed95d2cb6e5de8a342d761a89b0a04faed7be.
> > >
> > > Fixes: 8bbed95d2cb6 ("KVM: x86: WARN and reject loading KVM if NX is supported but not enabled")
> > > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > > ---
> > >
> > > Hopefully it's not too late to just drop the original patch...
> > >
> > >    arch/x86/kvm/x86.c | 3 ---
> > >    1 file changed, 3 deletions(-)
> > >
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 4a597aafe637..1cc02a3685d0 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -10981,9 +10981,6 @@ int kvm_arch_hardware_setup(void *opaque)
> > >    	int r;
> > >    	rdmsrl_safe(MSR_EFER, &host_efer);
> > > -	if (WARN_ON_ONCE(boot_cpu_has(X86_FEATURE_NX) &&
> > > -			 !(host_efer & EFER_NX)))
> > > -		return -EIO;
> > >    	if (boot_cpu_has(X86_FEATURE_XSAVES))
> > >    		rdmsrl(MSR_IA32_XSS, host_xss);
> >
> > So do we want this or "depends on X86_64 || X86_PAE"?
>
> Hmm, I'm leaning towards keeping !PAE support purely for testing the !PAE<->PAE
> MMU transitions for nested virtualization.  It's not much coverage, and the !PAE
> NPT horror is a much bigger testing gap (because KVM doesn't support it), but on
> the other hand setting EFER.NX for !PAE kernels appears to be trivial, e.g.
>
> diff --git a/arch/x86/kernel/head_32.S b/arch/x86/kernel/head_32.S
> index 67f590425d90..bfbea25a9fe8 100644
> --- a/arch/x86/kernel/head_32.S
> +++ b/arch/x86/kernel/head_32.S
> @@ -214,12 +214,6 @@ SYM_FUNC_START(startup_32_smp)
>          andl $~1,%edx                   # Ignore CPUID.FPU
>          jz .Lenable_paging              # No flags or only CPUID.FPU = no CR4
>
> -       movl pa(mmu_cr4_features),%eax
> -       movl %eax,%cr4
> -
> -       testb $X86_CR4_PAE, %al         # check if PAE is enabled
> -       jz .Lenable_paging
> -
>          /* Check if extended functions are implemented */
>          movl $0x80000000, %eax
>          cpuid
>
> My only hesitation is the risk of somehow breaking ancient CPUs by falling into
> the NX path.  Maybe try forcing EFER.NX=1 for !PAE, and fall back to requiring
> PAE if that gets NAK'd or needs to be reverted for whatever reason?

For now let's revert anyway.  Thanks!

Paolo