Re: [PATCH RFC] MIPS: KVM: role back pc in case of EMULATE_FAIL

[James Hogan]

On 07/05/15 13:51, Nicholas Mc Guire wrote:
> Currently if kvm_mips_complete_mmio_load() fails with EMULATE_FAIL it will
> not role back the pc nor will the caller handle this failure.
> 
> Signed-off-by: Nicholas Mc Guire <hofrat@xxxxxxxxx>
> ---
>  arch/mips/kvm/emulate.c |    4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> kvm_mips_complete_mmio_load is called only in arch/mips/kvm/mips.c without
> checking the return value to signal EMULATE_FAIL:
>  383         if (vcpu->mmio_needed) {
>  384                 if (!vcpu->mmio_is_write)
>  385                         kvm_mips_complete_mmio_load(vcpu, run);
>  386                 vcpu->mmio_needed = 0;
>  387         }
> 
> so maybe kvm_mips_complete_mmio_load should role back in case of failure 
> at arch/mips/kvm/emuilate.c:kvm_mips_complete_mmio_load()
> 2406         if (er == EMULATE_FAIL) {
> 2408                 return er;
> 
> something like the below patch - only based no looking at how EMULATE_FAIL
> is handled at other locations - not sure if this is appropriate here.
> 
> Patch was only compile tested msp71xx_defconfig + CONFIG_KVM=m
> 
> Patch is against 4.1-rc2 (localversion-next is -next-20150506)
> 
> diff --git a/arch/mips/kvm/emulate.c b/arch/mips/kvm/emulate.c
> index 2f0fc60..b58596b 100644
> --- a/arch/mips/kvm/emulate.c
> +++ b/arch/mips/kvm/emulate.c
> @@ -2403,8 +2403,10 @@ enum emulation_result kvm_mips_complete_mmio_load(struct kvm_vcpu *vcpu,
>  	 */
>  	curr_pc = vcpu->arch.pc;
>  	er = update_pc(vcpu, vcpu->arch.pending_load_cause);
> -	if (er == EMULATE_FAIL)
> +	if (er == EMULATE_FAIL) {
> +		vcpu->arch.pc = curr_pc;

If update_pc returns EMULATE_FAIL then vcpu->arch.pc won't have been
modified, so putting it back to the old value is redundant.

Actually, curr_pc can be dropped from this function since nothing else
can go wrong that would cause the PC to need rolling back. Effectively
kvm_mips_emulate_load() has omitted to update the PC since it'll only
return to userland to handle the MMIO and get the load data, so by the
time it gets back to kvm_mips_complete_mmio_load(), it already has
everything it needs to guarantee success.

Cheers
James


>  		return er;
> +	}
>  
>  	switch (run->mmio.len) {
>  	case 4:
>

Attachment: signature.asc
Description: OpenPGP digital signature