Re: [PATCH v12 4/4] x86/cpu: Clear feature bits whose dependencies were cleared

[H. Peter Anvin]

On 2026-03-27 08:11, Maciej Wieczor-Retman wrote:
> From: Maciej Wieczor-Retman <maciej.wieczor-retman@xxxxxxxxx>
> 
> After cpu_caps_cleared[] is initialized with DISABLED_MASK_INIT,
> features present in disabled bitmasks are cleared from x86_capability[].
> However features that depend on them and are not part of any disabled
> mask are not cleared by anything. They can trigger the warning in
> check_cpufeature_deps(), as before both features would show up as
> enabled even though they weren't. The uncleared features can also still
> falsely show up in /proc/cpuinfo.
> 
> Take such cases into account before applying the cpu_caps_cleared[]
> array.
> 
> +/*
> + * If the dependency was cleared through the disabled bitmasks while the
> + * feature wasn't it also needs to be cleared.
> + */
> +void clear_feature_disabled_deps(struct cpuinfo_x86 *c)
> +{
> +	const struct cpuid_dep *d;
> +
> +	for (d = cpuid_deps; d->feature; d++) {
> +		if (!DISABLED_MASK_BIT_SET(d->feature) && DISABLED_MASK_BIT_SET(d->depends))
> +			set_bit(d->feature, (unsigned long *)cpu_caps_cleared);
> +	}
> +}
> +

Thinking about it, if this is discovered at runtime, it is really too late,
because in a sense we have miscompiled the kernel with unwanted code (as we
should have compiled out these features just as with other DISABLED features.)

This ultimately reflects a failed dependency in Kconfig.cpufeatures, which may
have caused kconfig to do the Wrong Thing[TM]. So at the very least it might
be a good thing to print a message here saying Kconfig.cpufeatures should be
fixed.

The better option, which I don't know how difficult it would be, would be to
make the dependencies available to Kconfig. This sounds like something that
would fall in the scope of Ahmed's rework rather than this patchset, though
(Ahmed, would you agree?)

	-hpa