[PATCH v2 1/2] KVM: arm64: mixed-width check should be skipped for uninitialized vCPUs

[Marc Zyngier]

On Thu, 10 Feb 2022 05:31:49 +0000,
Reiji Watanabe <reijiw at google.com> wrote:
> 
> Hi Marc,
> 
> On Wed, Feb 9, 2022 at 4:04 AM Marc Zyngier <maz at kernel.org> wrote:
> >
> > Hi Reiji,
> >
> > On Wed, 09 Feb 2022 05:32:36 +0000,
> > Reiji Watanabe <reijiw at google.com> wrote:
> > >
> > > Hi Marc,
> > >
> > > On Tue, Feb 8, 2022 at 6:41 AM Marc Zyngier <maz at kernel.org> wrote:
> > > >
> > > > In [1], I suggested another approach that didn't require extra state,
> > > > and moved the existing checks under the kvm lock. What was wrong with
> > > > that approach?
> > >
> > > With that approach, even for a vcpu that has a broken set of features,
> > > which leads kvm_reset_vcpu() to fail for the vcpu, the vcpu->arch.features
> > > are checked by other vCPUs' vcpu_allowed_register_width() until the
> > > vcpu->arch.target is set to -1.
> > > Due to this, I would think some or possibly all vCPUs' kvm_reset_vcpu()
> > > may or may not fail (e.g. if userspace tries to configure vCPU#0 with
> > > 32bit EL1, and vCPU#1 and #2 with 64 bit EL1, KVM_ARM_VCPU_INIT
> > > for either vCPU#0, or both vCPU#1 and #2 should fail.  But, with that
> > > approach, it doesn't always work that way.  Instead, KVM_ARM_VCPU_INIT
> > > for all vCPUs could fail or KVM_ARM_VCPU_INIT for vCPU#0 and #1 could
> > > fail while the one for CPU#2 works).
> > > Also, even after the first KVM_RUN for vCPUs are already done,
> > > (the first) KVM_ARM_VCPU_INIT for another vCPU could cause the
> > > kvm_reset_vcpu() for those vCPUs to fail.
> > >
> > > I would think those behaviors are odd, and I wanted to avoid them.
> >
> > OK, fair enough. But then you need to remove most of the uses of
> > KVM_ARM_VCPU_EL1_32BIT so that it is only used as a userspace
> > interface and
> 
> Yes, I will.
> 
> > maybe not carried as part of the vcpu feature flag anymore.
> 
> At the first call of kvm_reset_vcpu() for the guest, the new kvm
> flag is not set yet. So, KVM_ARM_VCPU_EL1_32BIT will be needed
> by the function (unless we pass the flag as an argument for the
> function or by any other way).

Which is why I said 'maybe'. It's not a big deal if the flags stays,
but I don't want it evaluated further down the line. It is also pretty
similar to HCR_EL2.RW, which we already test with vcpu_el1_is_32bit().

Overall, we need to reduce that state to be as simple as possible.

> 
> > Also, we really should turn all these various bits in the kvm struct
> > into a set of flags. I have a patch posted there[1] for this, feel
> > free to pick it up.
> 
> Thank you for the suggestion. But, kvm->arch.el1_reg_width is not
> a binary because it needs to indicate an uninitialized state.  So, it
> won't fit perfectly with kvm->arch.flags, which is introduced by [1]
> as it is. Of course it's feasible by using 2 bits of the flags though...

2 bits is what I had in mind (one bit to indicate that it has already
been initialised, another to carry the actual width).

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.