On Tue, Mar 24, 2026 at 04:47:15PM +0000, Marc Zyngier wrote: > Mark Brown <[email protected]> wrote:
> > On an aarch64 only system the 32 bit ID registers have UNDEFINED values. > > As a result set_id_regs skips tests for setting fields in these registers > > when testing an aarch64 only guest. This has the side effect of meaning > > that we don't record an expected value for these registers, meaning that > > when the subsequent tests for values being visible in guests and preserved > > over reset check the value they can spuriously fail. This can be seen by > > running on an emulated system with both NV and 32 bit enabled, NV will > > result in the guests created by the test program being 64 bit only but > > the 32 bit ID registers will have values. > I don't think papering over this problem is the right thing to do. > If the issue is that you have HW that has both NV and AArch32, then > KVM needs to be fixed to make the 32bit IDregs RAZ when NV is present > because that's not a configuration we support. Yes, I'm seeing this in practice - it hits by default with qemu which is rather more readily accessible than actual hardware with NV at this point and I was concerned that since these registers are explicitly UNKNOWN (sorry, a mistake in the commit log above there) rather than RAZ if FEAT_AA32 is not implemented there might be gotchas. I do see that there's some forcing for the case where the host doesn't support FEAT_AA32, I figured there must be some reason why it wasn't done based on the guest configuration. I'll add something to kvm_finalize_sys_regs(), it'll still need updates in set_id_regs since that'd make the 32 bit ID registers change value when the guest is run.
signature.asc
Description: PGP signature
