Re: [PATCH v1 2/2] KVM: arm64: nv: update CPU register PAR_EL1 after 'at s*'

[Oliver Upton <oliver.upton@linux.dev>]

On Wed, Aug 06, 2025 at 02:17:55PM +0000, Volodymyr Babchuk wrote:
> Previously this code update only vCPU's in-memory value, which is good,
> but not enough, as there will be no context switch after exiting
> exception handler, so in-memory value will not get into actual
> register.
> 
> It worked good enough for VHE guests because KVM code tried fast path,
> which of course updated real PAR_EL1.
> 
> Signed-off-by: Volodymyr Babchuk <volodymyr_babchuk@epam.com>
> ---
>  arch/arm64/kvm/sys_regs.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
> index 7b8a0a6f26468..ab2b5e261d312 100644
> --- a/arch/arm64/kvm/sys_regs.c
> +++ b/arch/arm64/kvm/sys_regs.c
> @@ -3463,6 +3463,9 @@ static bool handle_at_s1e2(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
>  
>  	__kvm_at_s1e2(vcpu, op, p->regval);
>  
> +	/* No context switch happened, so we need to update PAR_EL1 manually */
> +	write_sysreg(vcpu_read_sys_reg(vcpu, PAR_EL1), par_el1);
> +

Ok, this had me thoroughly confused for a moment. The bug is actually in
kvm_write_sys_reg() which is supposed to update the sysreg value when
things are loaded on the CPU. __kvm_at_s1e2() is doing the right thing
by calling this accessor.

For registers like PAR_EL1 that don't have an EL2->EL1 mapping we assume
they belong to the EL1 context and thus are in-memory when in a hyp
context. TPIDR(RO)_EL0 is similarly affected.

This is a bit of an ugly hack, but something like the following should
get things working if you're able to test it:

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index ad2548477257..32f8d1de8f1a 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -149,6 +149,22 @@ static bool get_el2_to_el1_mapping(unsigned int reg,
 	}
 }
 
+/*
+ * Special-cased registers that do not have an ELx mapping and are always
+ * loaded on the CPU.
+ */
+static bool reg_has_elx_mapping(int reg)
+{
+	switch (reg) {
+	case TPIDR_EL0:
+	case TPIDRRO_EL0:
+	case PAR_EL1:
+		return false;
+	default:
+		return true;
+	}
+}
+
 u64 vcpu_read_sys_reg(const struct kvm_vcpu *vcpu, int reg)
 {
 	u64 val = 0x8badf00d8badf00d;
@@ -158,6 +174,9 @@ u64 vcpu_read_sys_reg(const struct kvm_vcpu *vcpu, int reg)
 	if (!vcpu_get_flag(vcpu, SYSREGS_ON_CPU))
 		goto memory_read;
 
+	if (!reg_has_elx_mapping(reg))
+		goto sysreg_read;
+
 	if (unlikely(get_el2_to_el1_mapping(reg, &el1r, &xlate))) {
 		if (!is_hyp_ctxt(vcpu))
 			goto memory_read;
@@ -204,6 +223,7 @@ u64 vcpu_read_sys_reg(const struct kvm_vcpu *vcpu, int reg)
 	if (unlikely(is_hyp_ctxt(vcpu)))
 		goto memory_read;
 
+sysreg_read:
 	if (__vcpu_read_sys_reg_from_cpu(reg, &val))
 		return val;
 
@@ -219,6 +239,9 @@ void vcpu_write_sys_reg(struct kvm_vcpu *vcpu, u64 val, int reg)
 	if (!vcpu_get_flag(vcpu, SYSREGS_ON_CPU))
 		goto memory_write;
 
+	if (!reg_has_elx_mapping(reg))
+		goto sysreg_write;
+
 	if (unlikely(get_el2_to_el1_mapping(reg, &el1r, &xlate))) {
 		if (!is_hyp_ctxt(vcpu))
 			goto memory_write;
@@ -259,6 +282,7 @@ void vcpu_write_sys_reg(struct kvm_vcpu *vcpu, u64 val, int reg)
 	if (unlikely(is_hyp_ctxt(vcpu)))
 		goto memory_write;
 
+sysreg_write:
 	if (__vcpu_write_sys_reg_to_cpu(val, reg))
 		return;
 
-- 
Thanks,
Oliver