Re: [PATCH] KVM: arm64: nv: Forward hvc traps if originated from nested VM

[Oliver Upton]

On Thu, Apr 10, 2025 at 06:52:45PM +0530, Ganapatrao Kulkarni wrote:
> 
> 
> On 10-04-2025 04:22 pm, Marc Zyngier wrote:
> > On Thu, 10 Apr 2025 11:20:24 +0100,
> > Ganapatrao Kulkarni <gankulkarni@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > 
> > > 
> > > 
> > > On 10-04-2025 12:49 pm, Marc Zyngier wrote:
> > > > On Thu, 10 Apr 2025 08:07:43 +0100,
> > > > Ganapatrao Kulkarni <gankulkarni@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > > > 
> > > > > It was discovered while trying selftest(smccc_filter) that the
> > > > > hvc trap is getting forwarded to guest hypervisor even if it is
> > > > > originated from itself.
> > > > > 
> > > > > HVC traps from guest hypervisor should be handled by the host
> > > > > hypervisor and traps originating from nested VM should be
> > > > > forwarded. Adding check to forward only if the hvc is trapped
> > > > > from the nested VM.
> > > > 
> > > > I disagree. HVC from EL2 must be routed to the same EL2. HVC from EL1
> > > > must be routed to the EL2 controlling EL1.
> > > 
> > > Thanks, Understood, In NV case, hvc has to be forwarded to L1
> > > irrespective of it origin (L1 or L2). Need to add hvc handler in the
> > > smccc_filter.c for the vm (when run as L1), so that it is handled and
> > > returns with required args set.
> > 
> > Why? This test checks under which conditions an HVC/SMC gets routed to
> > userspace. What does it even mean to test HVC if it doesn't make it
> > outside of the guest itself?
> 
> smccc_filter.c has 2 tests (test_filter_denied and test_filter_fwd_to_user),
> which runs the vm(guest_code).
> 
> I was trying to modify test_filter_denied to run in vEL2, which led to this
> patch/discussion. I agree, it does not makes sense to run this test for
> vEL2.
> 
> test_filter_fwd_to_user is not feasible to run in vEL2.
> Thanks for the feedback.
> 
> BTW, I could add hvc handler and run test_filter_denied in vEL2.

I would much prefer that the entire test be adapted to consider the EL of
the guest, only testing the behavior of SMCs for EL2.

Thanks,
Oliver