KVM: arm64: Fix CPU interface MMIO compatibility detection

In order to detect whether a GICv3 CPU interface is MMIO capable,
we switch ICC_SRE_EL1.SRE to 0 and check whether it sticks.

However, this is only possible if *ALL* of the HCR_EL2 interrupt
overrides are set, and the CPU is perfectly allowed to ignore
the write to ICC_SRE_EL1 otherwise. This leads KVM to pretend
that a whole bunch of ARMv8.0 CPUs aren't MMIO-capable, and
breaks VMs that should work correctly otherwise.

Fix this by setting IMO/FMO/IMO before touching ICC_SRE_EL1,
and clear them afterwards. This allows us to reliably detect
the CPU interface capabilities.

Tested-by: Shameerali Kolothum Thodi <[email protected]>
Fixes: 9739f6e ("KVM: arm64: Workaround firmware wrongly advertising GICv2-on-v3 compatibility")
Signed-off-by: Marc Zyngier <[email protected]>

arch/arm64/kvm/hyp/vgic-v3-sr.c

@@ -429,13 +429,22 @@ u64 __vgic_v3_get_gic_config(void)
 	if (has_vhe())
 		flags = local_daif_save();
 
+	/*
+	 * Table 11-2 "Permitted ICC_SRE_ELx.SRE settings" indicates
+	 * that to be able to set ICC_SRE_EL1.SRE to 0, all the
+	 * interrupt overrides must be set. You've got to love this.
+	 */
+	sysreg_clear_set(hcr_el2, 0, HCR_AMO | HCR_FMO | HCR_IMO);
+	isb();
 	write_gicreg(0, ICC_SRE_EL1);
 	isb();
 
 	val = read_gicreg(ICC_SRE_EL1);
 
 	write_gicreg(sre, ICC_SRE_EL1);
 	isb();
+	sysreg_clear_set(hcr_el2, HCR_AMO | HCR_FMO | HCR_IMO, 0);
+	isb();
 
 	if (has_vhe())
 		local_daif_restore(flags);