Re: [PATCH for-4.21] x86/cpu: populate CPUID 0x1.edx features early for self-snoop detection

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Sep 25, 2025 at 09:03:06AM +0200, Jan Beulich wrote:
> On 24.09.2025 15:40, Roger Pau Monné wrote:
> > On Wed, Sep 24, 2025 at 11:50:02AM +0000, Andrew Cooper wrote:
> >> On 24/09/2025 4:00 am, Roger Pau Monne wrote:
> >>> Otherwise the check for the SS feature in
> >>> check_memory_type_self_snoop_errata() fails unconditionally, which leads 
> >>> to
> >>> X86_FEATURE_XEN_SELFSNOOP never being set.
> >>>
> >>> We could also avoid this by not doing the reset_cpuinfo() for the BSP in
> >>> identify_cpu(), because SS detection uses boot_cpu_data.
> >>
> >> Doesn't this, mean ...
> > 
> > Well, that's the reason for the rant here.  The reset at the top of
> > identify_cpu() has been there since 2005.  It's arguably to make sure
> > the BSP and the APs have the same empty state in the passed
> > cpuinfo_x86 struct, as for the BSP this would be already partially
> > initialized due to what's done in early_cpu_init().
> > 
> > The underlying question is whether we would rather prefer to not do
> > the reset for the BSP, but that would lead to differences in the
> > contents of cpuinfo_x86 struct between the BSP and the APs.  In the
> > past we have arranged for leaves needed early to be populated in
> > generic_identify(), like FEATURESET_e21a, hence the proposed patch
> > does that for FEATURESET_1d.
> > 
> >>>   However that
> >>> creates an imbalance on the state of the BSP versus the APs in the
> >>> identify_cpu() code.
> >>>
> >>> I've opted for the less controversial solution of populating FEATURESET_1d
> >>> in generic_identify(), as the value is already there.  The same is done 
> >>> for
> >>> the AMD faulting probe code.
> >>>
> >>> Fixes: f2663ca2e520 ("x86/cpu/intel: Clear cache self-snoop capability in 
> >>> CPUs with known errata")
> >>> Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> >>
> >> ... this Fixes tag is incorrect?
> > 
> > I think the Fixes tag is accurate; the code was OK before that change.
> > Nothing in c_early_init hooks depended on (some of) the x86_capability
> > fields being populated, which is required after the change.
> 
> I agree. Hence:
> Reviewed-by: Jan Beulich <jbeulich@xxxxxxxx>
> 
> I wonder though whether while there we wouldn't want to also store ecx if
> we already have it. (Really there is the question of whether we haven't
> other cpu_has_* uses which similarly come "too early".)

Yeah, I was about to do it, but it's not strictly needed for
c_early_init, and it's done anyway just after the call to
c_early_init.  I can set that field also, but then I would need to
tweak the comment ahead, something like:

        /*
         * Early init of Self Snoop support requires 0x1.edx, while
         * there also set 0x1.ecx as the value is already in context.
         */
        c->x86_capability[FEATURESET_1d] = edx;
        c->x86_capability[FEATURESET_1c] = ecx;

Thanks, Roger.