On VM entry, we disable access to the VFP registers in order to perform a lazy save/restore of these registers.
On VM exit, we restore access, test if we did enable them before, and save/restore the guest/host registers if necessary. In this sequence, the FPEXC register is always accessed, irrespective of the trapping configuration. If the guest didn't touch the VFP registers, then the HCPTR access has now enabled such access, but we're missing a barrier to ensure architectural execution of the new HCPTR configuration. If the HCPTR access has been delayed/reordered, the subsequent access to FPEXC will cause a trap, which we aren't prepared to handle at all. The fix is to introduce a barrier that only takes place if the guest hasn't accessed its view of the VFP registers, making the access to FPEXC safe. Signed-off-by: Marc Zyngier <[email protected]> --- arch/arm/kvm/interrupts.S | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/arch/arm/kvm/interrupts.S b/arch/arm/kvm/interrupts.S index 79caf79..3ac7aca 100644 --- a/arch/arm/kvm/interrupts.S +++ b/arch/arm/kvm/interrupts.S @@ -175,10 +175,13 @@ __kvm_vcpu_return: #ifdef CONFIG_VFPv3 @ Save floating point registers we if let guest use them. tst r2, #(HCPTR_TCP(10) | HCPTR_TCP(11)) - bne after_vfp_restore + beq 1f + + isb @ Force execution of HCPTR if we've just reenabled VFP access + b after_vfp_restore @ Switch VFP/NEON hardware state to the host's - add r7, vcpu, #VCPU_VFP_GUEST +1: add r7, vcpu, #VCPU_VFP_GUEST store_vfp_state r7 add r7, vcpu, #VCPU_VFP_HOST ldr r7, [r7] -- 2.1.4 _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
