Re: [RFC PATCH 00/10] KVM: selftests: Add support for test-selectable ucall implementations

[Sean Christopherson]

On Wed, Jan 05, 2022, Michael Roth wrote:
> On Wed, Jan 05, 2022 at 05:43:21PM +0000, Sean Christopherson wrote:
> > Because it uses multiple VMs, and my rough sketch only allows for a single VM to
> > use ucall.  Though I suppose we could simply keep appending to the ucall list for
> > every VM.  The requirement would then be that all VMs are of the same type, i.e.
> > utilize the same ucall_ops.
> 
> Hmm, maybe I misread your patch. Not supporting multiple VMs was the
> reason I gave up on having the ucall structs allocated on-demand and
> went with requiring them to be passed as arguments to ucall().
> 
> I thought with your patch you had solved that by having each vm have it's
> own pool, via vm->ucall_list, and then mapping each pool into each guest
> separately via:
> 
>   ucall_init(vm):
>     ucall_list = vm->ucall_list
>     sync_global_to_guest(ucall_list).
> 
> then as long as that ucall_init() is done *after* the guest calls
> kvm_vm_elf_load(), it will end up with a 'ucall_list' global that points
> to it's own specific vm->ucall_list. Then on the test side it doesn't
> matter what the 'ucall_list' global is currently set to since you have
> the GPA and know what vm exited.
> 
> Or am I missing something there?

Ha, that was not at all intented.  But yes, it should work.  I'd rather be lucky
than good?

> Although even if that is the case, now that we're proposing doing the
> ucall_init() inside vm_create(), then we run the risk of a test calling
> kvm_vm_elf_load() after, which might clobber the guest's copy of
> ucall_list global if ucall_init() had since been called for another VM.
> But that could maybe be worked around by having whatever vm_create()
> variant we use also do the kvm_vm_elf_load() unconditionally as part of
> creation.

Will sync_global_to_guest() even work as intended if kvm_vm_elf_load() hasn't
been called?  If not, then sync_global_{to,from}_guest() should really assert if
the test hasn't been loaded.

As for ucall_init(), I think the best approach would be to make kvm_vm_elf_load()
a static and replace all calls with:

	kvm_vm_load_guest(vm);

where its implementation is:

  void kvm_vm_load_guest(struct kvm_vm *vm)
  {
  	kvm_vm_elf_load(vm, program_invocation_name);

	ucall_init(vm);
  }

The logic being that if a test creates a VM but never loads any code into the guest,
e.g. kvm_create_max_vcpus, then it _can't_ make ucalls.