Re: [RFC PATCH v4 11/36] i386/tdx: Initialize TDX before creating TD vcpus

[Xiaoyao Li]

On 6/1/2022 3:54 PM, Gerd Hoffmann wrote:
> On Wed, Jun 01, 2022 at 03:20:46PM +0800, Xiaoyao Li wrote:
> > On 5/24/2022 2:57 PM, Gerd Hoffmann wrote:
> > >     Hi,
> > > Maybe it's a bit more work to add VM-scope initialization support to
> > > qemu.
> >
> > If just introducing VM-scope initialization to QEMU, it would be easy. What
> > matters is what needs to be done inside VM-scope initialization.
> >
> > For TDX, we need to settle down the features that configured for the TD.
> > Typically, the features are attributes of cpu object, parsed from "-cpu"
> > option and stored in cpu object.
>
> > 2) create a CPU object when initializing machine object and collect all the
> > info from "-cpu" and drop it in the end; then why not do it when creating
> > 1st vcpu like this patch.
>
> Do VM-scope tdx initialization late enough that cpu objects are already
> created at that point, so you can collect the info you need without a
> dummy cpu?

new CPU object is created during creating each vcpu. So we have to use 
mutex and flag to ensure VM-scope initialization is executed only once.

And it's werid to hook  VM-scope initialization in the middle of the 
vcpu creating phase to satisfy "late enough", so we choose to do it just 
before calling KVM API to initializing vcpu.

> I guess it could be helpful for the discussion when you can outine the
> 'big picture' for tdx initialization.  How does kvm accel setup look
> like without TDX, and what additional actions are needed for TDX?  What
> ordering requirements and other constrains exist?

To boot a TDX VM, it requires several changes/additional steps in the flow:

 1. specify the vm type KVM_X86_TDX_VM when creating VM with
    IOCTL(KVM_CREATE_VM);
        - When initializing KVM accel

 2. initialize VM scope configuration before creating any VCPU;

 3. initialize VCPU scope configuration;
        - done inside machine_init_done_notifier;

 4. initialize virtual firmware in guest private memory before vcpu 
running;
        - done inside machine_init_done_notifier;

 5. finalize the TD's measurement;
        - done inside machine init_done_notifier;


And we are discussing where to do step 2).

We can find from the code of tdx_pre_create_vcpu(), that it needs
cpuid entries[] and attributes as input to KVM.

  cpuid entries[] is set up by kvm_x86_arch_cpuid() mainly based on
  'CPUX86State *env'

  attributes.pks is retrieved from env->features[]
  and attributes.pmu is retrieved from x86cpu->enable_pmu

to make VM-socpe data is consistent with VCPU data, we do choose the 
point late enough to ensure all the info/configurations from VCPU are 
settle down, that just before calling KVM API to do VCPU-scope 
configuration.

> take care,
>    Gerd