Re: [PATCH v3 4/5] KVM: arm64: Convert TCR_EL2 to config-driven sanitisation

[Oliver Upton <oupton@kernel.org>]

Hi,

On Fri, Nov 06, 2026 at 05:30:57PM +0000, Wei-Lin Chang wrote:
> On Fri, Nov 06, 2026 at 11:34:47AM +0000, Mark Brown wrote:
> > On Thu, Oct 29, 2026 at 03:37:31PM -0700, Oliver Upton wrote:
> > > From: Wei-Lin Chang <weilin.chang@arm.com>
> > > 
> > > Feature dependencies of the fields in TCR_EL2 are extracted from
> > > Registers.json of the BSD licenced AARCHMRS
> > > (AARCHMRS_OPENSOURCE_A_profile_FAT-2026-06).
> > > 
> > > Places where we slightly deviate from the above:
> > > 
> > >  - Stick to the ARM ARM M.c requirement for TCR_EL2.TCMA*, which only
> > >    depends on FEAT_MTE2. This avoids updating ID_AA64PFR2_EL1 with new
> > >    VMTE* definitions that we don't know the meaning of, as they aren't
> > >    present in the ARM ARM yet.
> > > 
> > >  - Ignore TCR_EL2.DS's requirement of needing the effective value of
> > >    TCR2_EL2.D128 be 0. This requires runtime detection of TCR2_EL2.D128
> > >    and we don't have D128 support yet.
> > > 
> > >  - TCR_EL2.AS is dependent on FEAT_ASID16, this information is missing
> > >    from the json.
> > 
> > I'm seeing a regression in -next for the pre_fault_memory_test selftest
> > when run on qemu with EL2 and NV enabled which bisects to this commit in
> > -next:

Thanks for reporting this Mark. Next time can you also include the QEMU
command line that reproduces the issue?

> > # selftests: kvm: pre_fault_memory_test
> > <47>[ 1358.708301] systemd-journald[75]: Sent WATCHDOG=1 notification.
> > # Random seed: 0xe322a57
> > # Testing guest mode: PA-bits:52,  VA-bits:48,  4K pages
> > # Testing memory backing src type: anonymous
> > #
> > # ==== Test Assertion Failure ====
> > #   pre_fault_memory_test.c:260: false
> > #   pid=1888 tid=1888 errno=2 - No such file or directory
> > # sh: 1: addr2line: not found
> > #   Unknown ucall 0x0.
> > not ok 48 selftests: kvm: pre_fault_memory_test # exit=254
> > 
> > The NV pre fault test appears to run fine in the same run, and the test
> > passes when run on the same qemu version in vanilla VHE mode.
> > 
> > Full log:
> > 
> >   https://lava.sirena.org.uk/scheduler/job/196379#L2476
> > 
> 
> The selftests are running vCPUs in vEL2 with 52-bit page tables it seems
> like. NV doesn't support 52-bit and this commit enforces that by forcing
> vCPU TCR_EL2.DS (hardware TCR_EL1.DS) to 0, so bad things happen.
> 
> aarch64_get_supported_page_sizes() asks KVM for the preferred target,
> which does not include KVM_ARM_VCPU_HAS_EL2. Without
> KVM_ARM_VCPU_HAS_EL2, 52bit ipa is permitted if the host supports it.
> 
> Then with 52-bit PA deemed supported, the selftest builds page tables
> with the 52-bit format for that case.
> 
> Then when creating the actual vCPU to run, kvm_get_default_vcpu_target()
> includes KVM_ARM_VCPU_HAS_EL2.
> 
> One potential solution is make vm_supports_el2() return false when
> vm->pa_bits > 48.

Bit of a hack but gets the job done, do you have time to send a patch?
I'm sure we'll get around to doing LPA2 long before anything useful
actually shows up anyway.

Thanks,
Oliver