Re: [PATCH v5 06/15] KVM: arm64: Add support for FEAT_HDBSS
From: Leonardo Bras
Date: Thu Nov 19 2026 - 12:27:42 EST
On Thu, Nov 19, 2026 at 03:52:37PM +0800, Tian Zheng wrote:
>
>
> On 10/30/2026 10:15 PM, Leonardo Bras wrote:
> > On Fri, Oct 30, 2026 at 10:53:57AM +0800, Tian Zheng wrote:
> > >
> > >
> > > On 10/27/2026 2:12 AM, Leonardo Bras wrote:
> > > > On Tue, Sep 29, 2026 at 06:36:46PM +0800, Tian Zheng wrote:
> > > > > diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
> > > > > index 32102c3912fa..12fa37328dd2 100644
> > > > > --- a/arch/arm64/kernel/cpufeature.c
> > > > > +++ b/arch/arm64/kernel/cpufeature.c
> > > > > @@ -2158,6 +2158,11 @@ static bool hvhe_possible(const struct arm64_cpu_capabilities *entry,
> > > > > return arm64_test_sw_feature_override(ARM64_SW_FEATURE_OVERRIDE_HVHE);
> > > > > }
> > > > >
> > > > > +static bool has_vhe_hdbss(const struct arm64_cpu_capabilities *entry, int scope)
> > > > > +{
> > > > > + return is_kernel_in_hyp_mode() && has_cpuid_feature(entry, scope);
> > > > > +}
> > > > > +
> > > >
> > >
> > > Hi leo,
> > >
> > > > IIUC you run this on every cpu in the system, to make sure it's matching
> > > > the feature. Is there any chance the kernel is in EL2 only on some of the
> > > > cpus here?
> > > >
> > >
> > > It's impossible for the system to have only some CPUs in EL2 and others
> > > not.
> > >
> > > > If not, maybe we can run it on system_supports_hdbss() instead, which would
> > > > be easier for the user to understand it only runs in EL2, on top of not
> > > > needing the above.
> > > >
> > >
> > > However, The question is whether a cap should mean "this system can use
> > > the feature" or "this hardware supports the feature". If the cap only
> > > tracks hardware support, users may be confused when the hardware
> > > supports HDBSS but the system runs in nVHE mode — they would mistakenly
> > > believe they can use the feature.
> > >
> > > So I think we should keep is_kernel_in_hyp_mode() in has_vhe_hdbss(),
> > > and add a comment in system_supports_hdbss() explaining this.
> > >
> > > Does that work for you? Looking forward to your reply.
> >
> > Well,
> > That's just my personal view of this:
> >
> > If we have a function called "has_hdbss" it should only report if the
> > system has the feature available.
> >
> > On the "system_supports_*" we have to figure out if the system itself
> > supports the feature, which checks the feature bits itself and could mean
> > checking for other features being there, such as hyp_mode.
> >
> > What do you think?
> >
> > Thx,
> > Leo
>
> Hi Leo,
Hi Tian,
> Sorry for the slow reply. I've been looking into how other cpufeatures
> are configured on arm64, on and off, and that's why it took a while.
I took some time to study it as well, and to be clear, my previous
understanding of this was incorrect, and it looks like the has_vhe_hdbss
(.matches callback) is only called once during boot, and once per hot-added
cpu.
> My reading is that detected: HDBSS should mean "this system can use
> HDBSS", not just "the silicon has it".
>
> Several caps already add conditions beyond the ID bits. CNP returns
> false for kdump and under ARM64_WORKAROUND_DISABLE_CNP, NV and pKVM
> match on kvm_get_mode(), and can_trap_icv_dir_el1() starts with
> is_hyp_mode_available(). So in practice a SYSTEM_FEATURE cap answers
> "can this kernel use it on this system".
Humm, or maybe "can this kernel, booted as it is, use it on this system"...
> HDBSS is EL2-only (HDBSSBR_EL2/HDBSSPROD_EL2, VTCR_EL2.HDBSS),
> implemented only under hyp/vhe, and used only by KVM, so "the hardware
> is there" is not the same as "this kernel can use it". That is why I'd
> keep the gate in has_vhe_hdbss() rather than moving it to the consumer.
>
> I'd be interested to hear how you see it.
I agree, that looks better this way.
I don't think I understand enough of what 'system' means in
system_supports_*(), but I think your version explains it enough based on
other similar functions.
Thanks!
Leo