Re: [REGRESSION] from 74a0e79df68a8042fb84fd7207e57b70722cf825: VFIO PCI passthrough no longer works

[Tom Lendacky]

On 12/10/24 09:47, Sean Christopherson wrote:
> +Tom
> 
> On Tue, Dec 10, 2024, Simon Pilkington wrote:
>> Hi,
>>
>> With the aforementioned commit I am no longer able to use PCI passthrough to
>> a Windows guest on the X570 chipset with a 5950X CPU.
>>
>> The minimal reproducer for me is to attach a GPU to the VM and attempt to
>> start Windows setup from an iso image. The VM will apparently livelock at the
>> setup splash screen before the spinner appears as one of my CPU cores goes up
>> to 100% usage until I force off the VM. This could be very machine-specific
>> though.
> 
> Ugh.  Yeah, it's pretty much guaranteed to be CPU specific behavior.
> 
> Tom, any idea what the guest might be trying to do?  It probably doesn't matter
> in the end, it's not like KVM does anything with the value...

No clue. I do see that in Linux there is a zenbleed-related bit that is
used in DE_CFG:

  522b1d69219d ("x86/cpu/amd: Add a Zenbleed fix")

I wonder if it might be related to that.

Your suggestion below to see what bits are being requested might shed
some light on it.

Thanks,
Tom

> 
>> Reverting to the old XOR check fixes both 6.12.y stable and 6.13-rc2 for me.
>> Otherwise they're both bad. Can you please look into it? I can share the
>> config I used for test builds if it would help.
> 
> Can you run with the below to see what bits the guest is trying to set (or clear)?
> We could get the same info via tracepoints, but this will likely be faster/easier.
> 
> ---
>  arch/x86/kvm/svm/svm.c | 12 +++++++++---
>  1 file changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index dd15cc635655..5144d0283c9d 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -3195,11 +3195,14 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
>  	case MSR_AMD64_DE_CFG: {
>  		u64 supported_de_cfg;
>  
> -		if (svm_get_feature_msr(ecx, &supported_de_cfg))
> +		if (WARN_ON_ONCE(svm_get_feature_msr(ecx, &supported_de_cfg)))
>  			return 1;
>  
> -		if (data & ~supported_de_cfg)
> +		if (data & ~supported_de_cfg) {
> +			pr_warn("DE_CFG supported = %llx, WRMSR = %llx\n",
> +				supported_de_cfg, data);
>  			return 1;
> +		}
>  
>  		/*
>  		 * Don't let the guest change the host-programmed value.  The
> @@ -3207,8 +3210,11 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
>  		 * are completely unknown to KVM, and the one bit known to KVM
>  		 * is simply a reflection of hardware capabilities.
>  		 */
> -		if (!msr->host_initiated && data != svm->msr_decfg)
> +		if (!msr->host_initiated && data != svm->msr_decfg) {
> +			pr_warn("DE_CFG current = %llx, WRMSR = %llx\n",
> +				svm->msr_decfg, data);
>  			return 1;
> +		}
>  
>  		svm->msr_decfg = data;
>  		break;
> 
> base-commit: fac04efc5c793dccbd07e2d59af9f90b7fc0dca4