Re: [PATCH 1/7] KVM: TDX: Add a place holder to handle TDX VM exit

[Binbin Wu <binbin.wu@linux.intel.com>]

On 12/16/2024 12:37 PM, Xiaoyao Li wrote:
> On 12/16/2024 8:54 AM, Binbin Wu wrote:
>>
>>
>>
>> On 12/13/2024 4:57 PM, Xiaoyao Li wrote:
>>> On 12/1/2024 11:53 AM, Binbin Wu wrote:
> ...
>>>
>>>>   }
>>>>     void tdx_load_mmu_pgd(struct kvm_vcpu *vcpu, hpa_t root_hpa, int pgd_level)
>>>> @@ -1135,6 +1215,88 @@ int tdx_sept_remove_private_spte(struct kvm *kvm, gfn_t gfn,
>>>>       return tdx_sept_drop_private_spte(kvm, gfn, level, pfn);
>>>>   }
>>>>   +int tdx_handle_exit(struct kvm_vcpu *vcpu, fastpath_t fastpath)
>>>> +{
>>>> +    struct vcpu_tdx *tdx = to_tdx(vcpu);
>>>> +    u64 vp_enter_ret = tdx->vp_enter_ret;
>>>> +    union vmx_exit_reason exit_reason;
>>>> +
>>>> +    if (fastpath != EXIT_FASTPATH_NONE)
>>>> +        return 1;
>>>> +
>>>> +    /*
>>>> +     * Handle TDX SW errors, including TDX_SEAMCALL_UD, TDX_SEAMCALL_GP and
>>>> +     * TDX_SEAMCALL_VMFAILINVALID.
>>>> +     */
>>>> +    if (unlikely((vp_enter_ret & TDX_SW_ERROR) == TDX_SW_ERROR)) {
>>>> +        KVM_BUG_ON(!kvm_rebooting, vcpu->kvm);
>>>> +        goto unhandled_exit;
>>>> +    }
>>>> +
>>>> +    /*
>>>> +     * Without off-TD debug enabled, failed_vmentry case must have
>>>> +     * TDX_NON_RECOVERABLE set.
>>>> +     */
>>>
>>> This comment is confusing. I'm not sure why it is put here. Below code does nothing with exit_reason.failed_vmentry.
>>
>> Because when failed_vmentry occurs, vp_enter_ret will have
>> TDX_NON_RECOVERABLE set, so it will be handled below.
>
> The words somehow is confusing, which to me is implying something like:
>
>     WARN_ON(!debug_td() && exit_reason.failed_vmentry &&
>                 !(vp_enter_ret & TDX_NON_RECOVERABLE))
The comment tried to say that the failed_vmentry case is also covered by
TDX_NON_RECOVERABLE since currently off-TD debug is not supported yet.
Since it's causing confusion, I'd like to add the code the handle the
failed_vmentry separately and aligned to VMX by using KVM_EXIT_FAIL_ENTRY
as suggested below.

>
> Besides, VMX returns KVM_EXIT_FAIL_ENTRY for vm-entry failure. So the question is why TDX cannot do it same way?
>
>>>
>>>> +    if (unlikely(vp_enter_ret & (TDX_ERROR | TDX_NON_RECOVERABLE))) {
>>>> +        /* Triple fault is non-recoverable. */
>>>> +        if (unlikely(tdx_check_exit_reason(vcpu, EXIT_REASON_TRIPLE_FAULT)))
>>>> +            return tdx_handle_triple_fault(vcpu);
>>>> +
>>>> +        kvm_pr_unimpl("TD vp_enter_ret 0x%llx, hkid 0x%x hkid pa 0x%llx\n",
>>>> +                  vp_enter_ret, to_kvm_tdx(vcpu->kvm)->hkid,
>>>> +                  set_hkid_to_hpa(0, to_kvm_tdx(vcpu->kvm)->hkid));
>>>
>>> It indeed needs clarification for the need of "hkid" and "hkid pa". Especially the "hkdi pa", which is the result of applying HKID of the current TD to a physical address 0. I cannot think of any reason why we need such info.
>> Yes, set_hkid_to_hpa(0, to_kvm_tdx(vcpu->kvm)->hkid) should be removed.
>> I didn't notice it.
>
> don't forget to justify why HKID is useful here. To me, HKID can be dropped as well.
OK.

>
>> Thanks!
>>
>>
>>>
>>>> +        goto unhandled_exit;
>>>> +    }
>>>> +
>>>> +    /* From now, the seamcall status should be TDX_SUCCESS. */
>>>> +    WARN_ON_ONCE((vp_enter_ret & TDX_SEAMCALL_STATUS_MASK) != TDX_SUCCESS);
>>>
>>> Is there any case that TDX_SUCCESS with additional non-zero information in the lower 32-bits? I thought TDX_SUCCESS is a whole 64- bit status code.
>> TDX status code uses the upper 32-bits.
>>
>> When the status code is TDX_SUCCESS and has a valid VMX exit reason, the lower
>> 32-bit is the VMX exit reason.
>>
>> You can refer to the TDX module ABI spec or interface_function_completion_status.json
>> from the intel-tdx-module-1.5-abi-table for details.
>>
>
> I see. (I asked a silly question that I even missed the normal Exit case)
>
>