On Tue, 2018-01-23 at 17:28 -0800, Dave Hansen wrote:
> On 01/23/2018 05:23 PM, Woodhouse, David wrote:
> > 
> > On Tue, 2018-01-23 at 10:43 -0800, Dave Hansen wrote:
> ...
> > 
> > > 
> > > > 
> > > >   /* Intel-defined CPU features, CPUID level 0x00000007:0 (EDX), word 18 */
> > > >   #define X86_FEATURE_AVX512_4VNNIW    (18*32+ 2) /* AVX-512 Neural Network Instructions */
> > > >   #define X86_FEATURE_AVX512_4FMAPS    (18*32+ 3) /* AVX-512 Multiply Accumulation Single precision */
> > > > +#define X86_FEATURE_SPEC_CTRL                (18*32+26) /* Speculation Control (IBRS + IBPB) */
> > > > +#define X86_FEATURE_STIBP            (18*32+27) /* Single Thread Indirect Branch Predictors */
> > > > +#define X86_FEATURE_ARCH_CAPABILITIES        (18*32+29) /* IA32_ARCH_CAPABILITIES MSR (Intel) */
> > > Should we be adding flags (STIBP) for which we currently have no user in
> > > the kernel?
> > They're in an existing word (now) so it costs us absolutely nothing to
> > do so. And they'll be exposed to KVM guests in imminent patches if
> > nothing else.
>
> Doesn't just defining it here generate something in the tables that then
> get exported in /proc/cpuinfo?  That's far from our most strict ABI, but
> a single #define here can be seen by users IIRC.

That's true, but still we're *working* on exposing and using these;
let's not go wild adding one feature at a time and having to tweak the
surrounding blacklist/enable/disable/expose logic at every step.