Re: [PATCH v3 3/3] x86/sev: Fix broken SNP support with KVM module built-in

["Kalra, Ashish" <ashish.kalra@amd.com>]

On 2/5/2025 1:31 PM, Sean Christopherson wrote:
> On Wed, Feb 05, 2025, Vasant Hegde wrote:
>> On 2/5/2025 8:47 PM, Sean Christopherson wrote:
>>> On Wed, Feb 05, 2025, Vasant Hegde wrote:
>>>>> @@ -3318,6 +3326,9 @@ static int __init iommu_go_to_state(enum iommu_init_state state)
>>>>>  		ret = state_next();
>>>>>  	}
>>>>>  
>>>>> +	if (ret && !amd_iommu_snp_en && cc_platform_has(CC_ATTR_HOST_SEV_SNP))
>>>>
>>>>
>>>> I think we should clear when `amd_iommu_snp_en` is true.
>>>
>>> That doesn't address the case where amd_iommu_prepare() fails, because amd_iommu_snp_en
>>> will be %false (its init value) and the RMP will be uninitialized, i.e.
>>> CC_ATTR_HOST_SEV_SNP will be incorrectly left set.
>>
>> You are right. I missed early failure scenarios :-(
>>
>>>
>>> And conversely, IMO clearing CC_ATTR_HOST_SEV_SNP after initializing the IOMMU
>>> and RMP is wrong as well.  Such a host is probably hosed regardless, but from
>>> the CPU's perspective, SNP is supported and enabled.
>>
>> So we don't want to clear  CC_ATTR_HOST_SEV_SNP after RMP initialization -OR-
>> clear for all failures?
> 
> I honestly don't know, because the answer largely depends on what happens with
> hardware.  I asked in an earlier version of this series if IOMMU initialization
> failure after the RMP is configured is even survivable.
> 

As i mentioned earlier and as part of this series and summarizing this again here:

- snp_rmptable_init() enables SNP support system-wide and that means the HW starts
doing RMP checks for memory accesses, but as RMP table is zeroed out initially, 
all memory is configured to be host/HV owned. 

It is only after SNP_INIT(_EX) that RMP table is configured and initialized with
HV_Fixed, firmware pages and stuff like IOMMU RMP enforcement is enabled. 

If the IOMMU initialization fails after IOMMU support on SNP check is completed
and host SNP is enabled, then SNP_INIT(_EX) will fail as IOMMUs need to be enabled
for SNP_INIT to succeed.

> For this series, I think it makes sense to match the existing behavior, unless
> someone from AMD can definitively state that we should do something different.
> And the existing behavior is that amd_iommu_snp_en and CC_ATTR_HOST_SEV_SNP will
> be left set if the IOMMU completes iommu_snp_enable(), and the kernel completes
> RMP setup.

Yes, that is true and this behavior is still consistent with this series.

Again to reiterate, if iommu_snp_enable() and host SNP enablement is successful,
any late IOMMU initialization failures should cause SNP_INIT to fail and that means
IOMMU RMP enforcement will never get enabled and RMP table will remain configured
for all memory marked as HV/host owned. 

Thanks,
Ashish