Re: [RESEND PATCH 05/32] vl: Add "sgx-epc" option to expose SGX EPC sections to guest

[Sean Christopherson <seanjc@google.com>]

On Tue, May 04, 2021, Paolo Bonzini wrote:
> On 04/05/21 02:09, Sean Christopherson wrote:
> > Is there a way to process "-device sgx-epc..." before vCPUs are realized?  The
> > ordering problem was the only reason I added a dedicated option.
> 
> If it's just CPUID, one possibility could be to mark the EPC sections
> specially in KVM_SET_USER_MEMORY_REGION and synthesize the leaves within
> KVM; or even look inside the VMA structs and detect EPC regions that way.

I experimented with those options, and a few others, and they all lack the
flexibility of making EPC just another memory backend.

For synthesizing CPUID within KVM:
  - Requires a vendor specific memory region flag for all architectures to work
    around a quirk of one userspace VMM.
  - Pushes a lot of complexity into KVM, e.g. KVM needs to update CPUID in
    response to memslot changes, and needs to query memslots in response to
    CPUID changes.
  - Does KVM or userspace define the section attributes, e.g. confidentiality,
    integrity, etc...?  If KVM, are they hardcoded to match the host?  What
    happens if a future Intel platform supports multiple EPC sections with
    different attributes?  If userspace, how does userspace communicate the
    attributes?
  - How does userspace know what KVM enumerated to the guest?  See the whole
    KVM_GET_CPUID2 fiasco...
  - Prevents userspace from enumerating EPC without a memslot, e.g. to trap on
    the first EPC access for tracking purposes.
 
For probing VMAs:
  - In addition to the above issues, requires MMU notifier integration to update
    CPUID in response to a VMA change.
  - Requires SGX subsystem to provide a helper to identify EPC VMAs.

In short, I feel very strongly that this is QEMU's problem to solve.

> Otherwise, the -M solution would work.
> 
> Paolo
> 
> >  From the changelog:
> > 
> >    Because SGX EPC is enumerated through CPUID, EPC "devices" need to be
> >    realized prior to realizing the vCPUs themselves, i.e. long before
> >    generic devices are parsed and realized.
> > 
> >    So even though EPC sections could be realized through the generic
> >    -devices command, they need to be created much earlier for them to
> >    actually be usable by the guest.
>